feat(modules): migrate to go modules and bump go version 1.14.4

- migrate to go module
- bump go version 1.14.4

Signed-off-by: prateekpandey14 <prateek.pandey@mayadata.io>

vendor/gonum.org/v1/gonum/AUTHORS

@@ -0,0 +1,89 @@
+# This is the official list of gonum authors for copyright purposes.
+# This file is distinct from the CONTRIBUTORS files.
+# See the latter for an explanation.
+
+# Names should be added to this file as
+#	Name or Organization <email address>
+# The email address is not required for organizations.
+
+# Please keep the list sorted.
+
+Alexander Egurnov <alexander.egurnov@gmail.com>
+Bill Gray <wgray@gogray.com>
+Bill Noon <noon.bill@gmail.com>
+Brendan Tracey <tracey.brendan@gmail.com>
+Brent Pedersen <bpederse@gmail.com>
+Chad Kunde <kunde21@gmail.com>
+Chih-Wei Chang <bert.cwchang@gmail.com>
+Chris Tessum <ctessum@gmail.com>
+Christophe Meessen <christophe.meessen@gmail.com>
+Clayton Northey <clayton.northey@gmail.com>
+Dan Kortschak <dan.kortschak@adelaide.edu.au> <dan@kortschak.io>
+Daniel Fireman <danielfireman@gmail.com>
+David Samborski <bloggingarrow@gmail.com>
+Davor Kapsa <davor.kapsa@gmail.com>
+DeepMind Technologies
+Dezmond Goff <goff.dezmond@gmail.com>
+Egon Elbre <egonelbre@gmail.com>
+Ekaterina Efimova <katerina.efimova@gmail.com>
+Ethan Burns <burns.ethan@gmail.com>
+Evert Lammerts <evert.lammerts@gmail.com>
+Facundo Gaich <facugaich@gmail.com>
+Fazlul Shahriar <fshahriar@gmail.com>
+Francesc Campoy <campoy@golang.org>
+Google Inc
+Gustaf Johansson <gustaf@pinon.se>
+Iakov Davydov <iakov.davydov@unil.ch>
+Igor Mikushkin <igor.mikushkin@gmail.com>
+Iskander Sharipov <quasilyte@gmail.com>
+Jalem Raj Rohit <jrajrohit33@gmail.com>
+James Bell <james@stellentus.com>
+James Bowman <james.edward.bowman@gmail.com>
+James Holmes <32bitkid@gmail.com>
+Janne Snabb <snabb@epipe.com>
+Jeff Juozapaitis <jjjuozap@email.arizona.edu>
+Jeremy Atkinson <jchatkinson@gmail.com>
+Jonas Kahler <jonas@derkahler.de>
+Jonas Schulze <jonas.schulze@ovgu.de>
+Jonathan J Lawlor <jonathan.lawlor@gmail.com>
+Jonathan Schroeder <jd.schroeder@gmail.com>
+Joseph Watson <jtwatson@linux-consulting.us>
+Josh Wilson <josh.craig.wilson@gmail.com>
+Julien Roland <juroland@gmail.com>
+Kai Trukenmüller <ktye78@gmail.com>
+Kent English <kent.english@gmail.com>
+Kevin C. Zimmerman <kevinczimmerman@gmail.com>
+Kirill Motkov <motkov.kirill@gmail.com>
+Konstantin Shaposhnikov <k.shaposhnikov@gmail.com>
+Leonid Kneller <recondite.matter@gmail.com>
+Lyron Winderbaum <lyron.winderbaum@student.adelaide.edu.au>
+Martin Diz <github@martindiz.com.ar>
+Matthieu Di Mercurio <matthieu.dimercurio@gmail.com>
+Max Halford <maxhalford25@gmail.com>
+MinJae Kwon <k239507@gmail.com>
+Nick Potts <nick@the-potts.com>
+Olivier Wulveryck <olivier.wulveryck@gmail.com>
+Or Rikon <rikonor@gmail.com>
+Pontus Melke <pontusmelke@gmail.com>
+Renée French
+Rishi Desai <desai.rishi1@gmail.com>
+Robin Eklind <r.eklind.87@gmail.com>
+Sam Zaydel <szaydel@gmail.com>
+Samuel Kelemen <Samuel@Kelemen.us>
+Saran Ahluwalia <ahlusar.ahluwalia@gmail.com>
+Scott Holden <scott@sshconnection.com>
+Sebastien Binet <seb.binet@gmail.com>
+Shawn Smith <shawnpsmith@gmail.com>
+source{d} <hello@sourced.tech>
+Spencer Lyon <spencerlyon2@gmail.com>
+Steve McCoy <mccoyst@gmail.com>
+Taesu Pyo <pyotaesu@gmail.com>
+Takeshi Yoneda <cz.rk.t0415y.g@gmail.com>
+The University of Adelaide
+The University of Minnesota
+The University of Washington
+Thomas Berg <tomfuture@gmail.com>
+Tobin Harding <me@tobin.cc>
+Vincent Thiery <vjmthiery@gmail.com>
+Vladimír Chalupecký <vladimir.chalupecky@gmail.com>
+Yevgeniy Vahlis <evahlis@gmail.com>

vendor/gonum.org/v1/gonum/CONTRIBUTORS

@@ -0,0 +1,91 @@
+# This is the official list of people who can contribute
+# (and typically have contributed) code to the gonum
+# repository.
+#
+# The AUTHORS file lists the copyright holders; this file
+# lists people.  For example, Google employees would be listed here
+# but not in AUTHORS, because Google would hold the copyright.
+#
+# When adding J Random Contributor's name to this file,
+# either J's name or J's organization's name should be
+# added to the AUTHORS file.
+#
+# Names should be added to this file like so:
+#     Name <email address>
+#
+# Please keep the list sorted.
+
+Alexander Egurnov <alexander.egurnov@gmail.com>
+Andrew Brampton <brampton@gmail.com>
+Bill Gray <wgray@gogray.com>
+Bill Noon <noon.bill@gmail.com>
+Brendan Tracey <tracey.brendan@gmail.com>
+Brent Pedersen <bpederse@gmail.com>
+Chad Kunde <kunde21@gmail.com>
+Chih-Wei Chang <bert.cwchang@gmail.com>
+Chris Tessum <ctessum@gmail.com>
+Christophe Meessen <christophe.meessen@gmail.com>
+Clayton Northey <clayton.northey@gmail.com>
+Dan Kortschak <dan.kortschak@adelaide.edu.au> <dan@kortschak.io>
+Daniel Fireman <danielfireman@gmail.com>
+David Samborski <bloggingarrow@gmail.com>
+Davor Kapsa <davor.kapsa@gmail.com>
+Dezmond Goff <goff.dezmond@gmail.com>
+Egon Elbre <egonelbre@gmail.com>
+Ekaterina Efimova <katerina.efimova@gmail.com>
+Ethan Burns <burns.ethan@gmail.com>
+Evert Lammerts <evert.lammerts@gmail.com>
+Facundo Gaich <facugaich@gmail.com>
+Fazlul Shahriar <fshahriar@gmail.com>
+Francesc Campoy <campoy@golang.org>
+Gustaf Johansson <gustaf@pinon.se>
+Iakov Davydov <iakov.davydov@unil.ch>
+Igor Mikushkin <igor.mikushkin@gmail.com>
+Iskander Sharipov <quasilyte@gmail.com>
+Jalem Raj Rohit <jrajrohit33@gmail.com>
+James Bell <james@stellentus.com>
+James Bowman <james.edward.bowman@gmail.com>
+James Holmes <32bitkid@gmail.com>
+Janne Snabb <snabb@epipe.com>
+Jeff Juozapaitis <jjjuozap@email.arizona.edu>
+Jeremy Atkinson <jchatkinson@gmail.com>
+Jonas Kahler <jonas@derkahler.de>
+Jonas Schulze <jonas.schulze@ovgu.de>
+Jonathan J Lawlor <jonathan.lawlor@gmail.com>
+Jonathan Schroeder <jd.schroeder@gmail.com>
+Joseph Watson <jtwatson@linux-consulting.us>
+Josh Wilson <josh.craig.wilson@gmail.com>
+Julien Roland <juroland@gmail.com>
+Kai Trukenmüller <ktye78@gmail.com>
+Kent English <kent.english@gmail.com>
+Kevin C. Zimmerman <kevinczimmerman@gmail.com>
+Kirill Motkov <motkov.kirill@gmail.com>
+Konstantin Shaposhnikov <k.shaposhnikov@gmail.com>
+Leonid Kneller <recondite.matter@gmail.com>
+Lyron Winderbaum <lyron.winderbaum@student.adelaide.edu.au>
+Martin Diz <github@martindiz.com.ar>
+Matthieu Di Mercurio <matthieu.dimercurio@gmail.com>
+Max Halford <maxhalford25@gmail.com>
+MinJae Kwon <k239507@gmail.com>
+Nick Potts <nick@the-potts.com>
+Olivier Wulveryck <olivier.wulveryck@gmail.com>
+Or Rikon <rikonor@gmail.com>
+Pontus Melke <pontusmelke@gmail.com>
+Renée French
+Rishi Desai <desai.rishi1@gmail.com>
+Robin Eklind <r.eklind.87@gmail.com>
+Sam Zaydel <szaydel@gmail.com>
+Samuel Kelemen <Samuel@Kelemen.us>
+Saran Ahluwalia <ahlusar.ahluwalia@gmail.com>
+Scott Holden <scott@sshconnection.com>
+Sebastien Binet <seb.binet@gmail.com>
+Shawn Smith <shawnpsmith@gmail.com>
+Spencer Lyon <spencerlyon2@gmail.com>
+Steve McCoy <mccoyst@gmail.com>
+Taesu Pyo <pyotaesu@gmail.com>
+Takeshi Yoneda <cz.rk.t0415y.g@gmail.com>
+Thomas Berg <tomfuture@gmail.com>
+Tobin Harding <me@tobin.cc>
+Vincent Thiery <vjmthiery@gmail.com>
+Vladimír Chalupecký <vladimir.chalupecky@gmail.com>
+Yevgeniy Vahlis <evahlis@gmail.com>

vendor/gonum.org/v1/gonum/LICENSE

@@ -0,0 +1,23 @@
+Copyright ©2013 The Gonum Authors. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+    * Neither the name of the gonum project nor the names of its authors and
+      contributors may be used to endorse or promote products derived from this
+      software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

vendor/gonum.org/v1/gonum/blas/README.md

@@ -0,0 +1,47 @@
+# Gonum BLAS [![GoDoc](https://godoc.org/gonum.org/v1/gonum/blas?status.svg)](https://godoc.org/gonum.org/v1/gonum/blas)
+
+A collection of packages to provide BLAS functionality for the [Go programming
+language](http://golang.org)
+
+## Installation
+```sh
+  go get gonum.org/v1/gonum/blas/...
+```
+
+## Packages
+
+### blas
+
+Defines [BLAS API](http://www.netlib.org/blas/blast-forum/cinterface.pdf) split in several
+interfaces.
+
+### blas/gonum
+
+Go implementation of the BLAS API (incomplete, implements the `float32` and `float64` API).
+
+### blas/blas64 and blas/blas32
+
+Wrappers for an implementation of the double (i.e., `float64`) and single (`float32`)
+precision real parts of the BLAS API.
+
+```Go
+package main
+
+import (
+	"fmt"
+
+	"gonum.org/v1/gonum/blas/blas64"
+)
+
+func main() {
+	v := blas64.Vector{Inc: 1, Data: []float64{1, 1, 1}}
+	fmt.Println("v has length:", blas64.Nrm2(len(v.Data), v))
+}
+```
+
+### blas/cblas128 and blas/cblas64
+
+Wrappers for an implementation of the double (i.e., `complex128`) and single (`complex64`) 
+precision complex parts of the blas API.
+
+Currently blas/cblas64 and blas/cblas128 require gonum.org/v1/netlib/blas.

vendor/gonum.org/v1/gonum/blas/blas.go

@@ -0,0 +1,283 @@
+// Copyright ©2013 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//go:generate ./conversions.bash
+
+package blas
+
+// Flag constants indicate Givens transformation H matrix state.
+type Flag int
+
+const (
+	Identity    Flag = -2 // H is the identity matrix; no rotation is needed.
+	Rescaling   Flag = -1 // H specifies rescaling.
+	OffDiagonal Flag = 0  // Off-diagonal elements of H are non-unit.
+	Diagonal    Flag = 1  // Diagonal elements of H are non-unit.
+)
+
+// SrotmParams contains Givens transformation parameters returned
+// by the Float32 Srotm method.
+type SrotmParams struct {
+	Flag
+	H [4]float32 // Column-major 2 by 2 matrix.
+}
+
+// DrotmParams contains Givens transformation parameters returned
+// by the Float64 Drotm method.
+type DrotmParams struct {
+	Flag
+	H [4]float64 // Column-major 2 by 2 matrix.
+}
+
+// Transpose specifies the transposition operation of a matrix.
+type Transpose byte
+
+const (
+	NoTrans   Transpose = 'N'
+	Trans     Transpose = 'T'
+	ConjTrans Transpose = 'C'
+)
+
+// Uplo specifies whether a matrix is upper or lower triangular.
+type Uplo byte
+
+const (
+	Upper Uplo = 'U'
+	Lower Uplo = 'L'
+	All   Uplo = 'A'
+)
+
+// Diag specifies whether a matrix is unit triangular.
+type Diag byte
+
+const (
+	NonUnit Diag = 'N'
+	Unit    Diag = 'U'
+)
+
+// Side specifies from which side a multiplication operation is performed.
+type Side byte
+
+const (
+	Left  Side = 'L'
+	Right Side = 'R'
+)
+
+// Float32 implements the single precision real BLAS routines.
+type Float32 interface {
+	Float32Level1
+	Float32Level2
+	Float32Level3
+}
+
+// Float32Level1 implements the single precision real BLAS Level 1 routines.
+type Float32Level1 interface {
+	Sdsdot(n int, alpha float32, x []float32, incX int, y []float32, incY int) float32
+	Dsdot(n int, x []float32, incX int, y []float32, incY int) float64
+	Sdot(n int, x []float32, incX int, y []float32, incY int) float32
+	Snrm2(n int, x []float32, incX int) float32
+	Sasum(n int, x []float32, incX int) float32
+	Isamax(n int, x []float32, incX int) int
+	Sswap(n int, x []float32, incX int, y []float32, incY int)
+	Scopy(n int, x []float32, incX int, y []float32, incY int)
+	Saxpy(n int, alpha float32, x []float32, incX int, y []float32, incY int)
+	Srotg(a, b float32) (c, s, r, z float32)
+	Srotmg(d1, d2, b1, b2 float32) (p SrotmParams, rd1, rd2, rb1 float32)
+	Srot(n int, x []float32, incX int, y []float32, incY int, c, s float32)
+	Srotm(n int, x []float32, incX int, y []float32, incY int, p SrotmParams)
+	Sscal(n int, alpha float32, x []float32, incX int)
+}
+
+// Float32Level2 implements the single precision real BLAS Level 2 routines.
+type Float32Level2 interface {
+	Sgemv(tA Transpose, m, n int, alpha float32, a []float32, lda int, x []float32, incX int, beta float32, y []float32, incY int)
+	Sgbmv(tA Transpose, m, n, kL, kU int, alpha float32, a []float32, lda int, x []float32, incX int, beta float32, y []float32, incY int)
+	Strmv(ul Uplo, tA Transpose, d Diag, n int, a []float32, lda int, x []float32, incX int)
+	Stbmv(ul Uplo, tA Transpose, d Diag, n, k int, a []float32, lda int, x []float32, incX int)
+	Stpmv(ul Uplo, tA Transpose, d Diag, n int, ap []float32, x []float32, incX int)
+	Strsv(ul Uplo, tA Transpose, d Diag, n int, a []float32, lda int, x []float32, incX int)
+	Stbsv(ul Uplo, tA Transpose, d Diag, n, k int, a []float32, lda int, x []float32, incX int)
+	Stpsv(ul Uplo, tA Transpose, d Diag, n int, ap []float32, x []float32, incX int)
+	Ssymv(ul Uplo, n int, alpha float32, a []float32, lda int, x []float32, incX int, beta float32, y []float32, incY int)
+	Ssbmv(ul Uplo, n, k int, alpha float32, a []float32, lda int, x []float32, incX int, beta float32, y []float32, incY int)
+	Sspmv(ul Uplo, n int, alpha float32, ap []float32, x []float32, incX int, beta float32, y []float32, incY int)
+	Sger(m, n int, alpha float32, x []float32, incX int, y []float32, incY int, a []float32, lda int)
+	Ssyr(ul Uplo, n int, alpha float32, x []float32, incX int, a []float32, lda int)
+	Sspr(ul Uplo, n int, alpha float32, x []float32, incX int, ap []float32)
+	Ssyr2(ul Uplo, n int, alpha float32, x []float32, incX int, y []float32, incY int, a []float32, lda int)
+	Sspr2(ul Uplo, n int, alpha float32, x []float32, incX int, y []float32, incY int, a []float32)
+}
+
+// Float32Level3 implements the single precision real BLAS Level 3 routines.
+type Float32Level3 interface {
+	Sgemm(tA, tB Transpose, m, n, k int, alpha float32, a []float32, lda int, b []float32, ldb int, beta float32, c []float32, ldc int)
+	Ssymm(s Side, ul Uplo, m, n int, alpha float32, a []float32, lda int, b []float32, ldb int, beta float32, c []float32, ldc int)
+	Ssyrk(ul Uplo, t Transpose, n, k int, alpha float32, a []float32, lda int, beta float32, c []float32, ldc int)
+	Ssyr2k(ul Uplo, t Transpose, n, k int, alpha float32, a []float32, lda int, b []float32, ldb int, beta float32, c []float32, ldc int)
+	Strmm(s Side, ul Uplo, tA Transpose, d Diag, m, n int, alpha float32, a []float32, lda int, b []float32, ldb int)
+	Strsm(s Side, ul Uplo, tA Transpose, d Diag, m, n int, alpha float32, a []float32, lda int, b []float32, ldb int)
+}
+
+// Float64 implements the single precision real BLAS routines.
+type Float64 interface {
+	Float64Level1
+	Float64Level2
+	Float64Level3
+}
+
+// Float64Level1 implements the double precision real BLAS Level 1 routines.
+type Float64Level1 interface {
+	Ddot(n int, x []float64, incX int, y []float64, incY int) float64
+	Dnrm2(n int, x []float64, incX int) float64
+	Dasum(n int, x []float64, incX int) float64
+	Idamax(n int, x []float64, incX int) int
+	Dswap(n int, x []float64, incX int, y []float64, incY int)
+	Dcopy(n int, x []float64, incX int, y []float64, incY int)
+	Daxpy(n int, alpha float64, x []float64, incX int, y []float64, incY int)
+	Drotg(a, b float64) (c, s, r, z float64)
+	Drotmg(d1, d2, b1, b2 float64) (p DrotmParams, rd1, rd2, rb1 float64)
+	Drot(n int, x []float64, incX int, y []float64, incY int, c float64, s float64)
+	Drotm(n int, x []float64, incX int, y []float64, incY int, p DrotmParams)
+	Dscal(n int, alpha float64, x []float64, incX int)
+}
+
+// Float64Level2 implements the double precision real BLAS Level 2 routines.
+type Float64Level2 interface {
+	Dgemv(tA Transpose, m, n int, alpha float64, a []float64, lda int, x []float64, incX int, beta float64, y []float64, incY int)
+	Dgbmv(tA Transpose, m, n, kL, kU int, alpha float64, a []float64, lda int, x []float64, incX int, beta float64, y []float64, incY int)
+	Dtrmv(ul Uplo, tA Transpose, d Diag, n int, a []float64, lda int, x []float64, incX int)
+	Dtbmv(ul Uplo, tA Transpose, d Diag, n, k int, a []float64, lda int, x []float64, incX int)
+	Dtpmv(ul Uplo, tA Transpose, d Diag, n int, ap []float64, x []float64, incX int)
+	Dtrsv(ul Uplo, tA Transpose, d Diag, n int, a []float64, lda int, x []float64, incX int)
+	Dtbsv(ul Uplo, tA Transpose, d Diag, n, k int, a []float64, lda int, x []float64, incX int)
+	Dtpsv(ul Uplo, tA Transpose, d Diag, n int, ap []float64, x []float64, incX int)
+	Dsymv(ul Uplo, n int, alpha float64, a []float64, lda int, x []float64, incX int, beta float64, y []float64, incY int)
+	Dsbmv(ul Uplo, n, k int, alpha float64, a []float64, lda int, x []float64, incX int, beta float64, y []float64, incY int)
+	Dspmv(ul Uplo, n int, alpha float64, ap []float64, x []float64, incX int, beta float64, y []float64, incY int)
+	Dger(m, n int, alpha float64, x []float64, incX int, y []float64, incY int, a []float64, lda int)
+	Dsyr(ul Uplo, n int, alpha float64, x []float64, incX int, a []float64, lda int)
+	Dspr(ul Uplo, n int, alpha float64, x []float64, incX int, ap []float64)
+	Dsyr2(ul Uplo, n int, alpha float64, x []float64, incX int, y []float64, incY int, a []float64, lda int)
+	Dspr2(ul Uplo, n int, alpha float64, x []float64, incX int, y []float64, incY int, a []float64)
+}
+
+// Float64Level3 implements the double precision real BLAS Level 3 routines.
+type Float64Level3 interface {
+	Dgemm(tA, tB Transpose, m, n, k int, alpha float64, a []float64, lda int, b []float64, ldb int, beta float64, c []float64, ldc int)
+	Dsymm(s Side, ul Uplo, m, n int, alpha float64, a []float64, lda int, b []float64, ldb int, beta float64, c []float64, ldc int)
+	Dsyrk(ul Uplo, t Transpose, n, k int, alpha float64, a []float64, lda int, beta float64, c []float64, ldc int)
+	Dsyr2k(ul Uplo, t Transpose, n, k int, alpha float64, a []float64, lda int, b []float64, ldb int, beta float64, c []float64, ldc int)
+	Dtrmm(s Side, ul Uplo, tA Transpose, d Diag, m, n int, alpha float64, a []float64, lda int, b []float64, ldb int)
+	Dtrsm(s Side, ul Uplo, tA Transpose, d Diag, m, n int, alpha float64, a []float64, lda int, b []float64, ldb int)
+}
+
+// Complex64 implements the single precision complex BLAS routines.
+type Complex64 interface {
+	Complex64Level1
+	Complex64Level2
+	Complex64Level3
+}
+
+// Complex64Level1 implements the single precision complex BLAS Level 1 routines.
+type Complex64Level1 interface {
+	Cdotu(n int, x []complex64, incX int, y []complex64, incY int) (dotu complex64)
+	Cdotc(n int, x []complex64, incX int, y []complex64, incY int) (dotc complex64)
+	Scnrm2(n int, x []complex64, incX int) float32
+	Scasum(n int, x []complex64, incX int) float32
+	Icamax(n int, x []complex64, incX int) int
+	Cswap(n int, x []complex64, incX int, y []complex64, incY int)
+	Ccopy(n int, x []complex64, incX int, y []complex64, incY int)
+	Caxpy(n int, alpha complex64, x []complex64, incX int, y []complex64, incY int)
+	Cscal(n int, alpha complex64, x []complex64, incX int)
+	Csscal(n int, alpha float32, x []complex64, incX int)
+}
+
+// Complex64Level2 implements the single precision complex BLAS routines Level 2 routines.
+type Complex64Level2 interface {
+	Cgemv(tA Transpose, m, n int, alpha complex64, a []complex64, lda int, x []complex64, incX int, beta complex64, y []complex64, incY int)
+	Cgbmv(tA Transpose, m, n, kL, kU int, alpha complex64, a []complex64, lda int, x []complex64, incX int, beta complex64, y []complex64, incY int)
+	Ctrmv(ul Uplo, tA Transpose, d Diag, n int, a []complex64, lda int, x []complex64, incX int)
+	Ctbmv(ul Uplo, tA Transpose, d Diag, n, k int, a []complex64, lda int, x []complex64, incX int)
+	Ctpmv(ul Uplo, tA Transpose, d Diag, n int, ap []complex64, x []complex64, incX int)
+	Ctrsv(ul Uplo, tA Transpose, d Diag, n int, a []complex64, lda int, x []complex64, incX int)
+	Ctbsv(ul Uplo, tA Transpose, d Diag, n, k int, a []complex64, lda int, x []complex64, incX int)
+	Ctpsv(ul Uplo, tA Transpose, d Diag, n int, ap []complex64, x []complex64, incX int)
+	Chemv(ul Uplo, n int, alpha complex64, a []complex64, lda int, x []complex64, incX int, beta complex64, y []complex64, incY int)
+	Chbmv(ul Uplo, n, k int, alpha complex64, a []complex64, lda int, x []complex64, incX int, beta complex64, y []complex64, incY int)
+	Chpmv(ul Uplo, n int, alpha complex64, ap []complex64, x []complex64, incX int, beta complex64, y []complex64, incY int)
+	Cgeru(m, n int, alpha complex64, x []complex64, incX int, y []complex64, incY int, a []complex64, lda int)
+	Cgerc(m, n int, alpha complex64, x []complex64, incX int, y []complex64, incY int, a []complex64, lda int)
+	Cher(ul Uplo, n int, alpha float32, x []complex64, incX int, a []complex64, lda int)
+	Chpr(ul Uplo, n int, alpha float32, x []complex64, incX int, a []complex64)
+	Cher2(ul Uplo, n int, alpha complex64, x []complex64, incX int, y []complex64, incY int, a []complex64, lda int)
+	Chpr2(ul Uplo, n int, alpha complex64, x []complex64, incX int, y []complex64, incY int, ap []complex64)
+}
+
+// Complex64Level3 implements the single precision complex BLAS Level 3 routines.
+type Complex64Level3 interface {
+	Cgemm(tA, tB Transpose, m, n, k int, alpha complex64, a []complex64, lda int, b []complex64, ldb int, beta complex64, c []complex64, ldc int)
+	Csymm(s Side, ul Uplo, m, n int, alpha complex64, a []complex64, lda int, b []complex64, ldb int, beta complex64, c []complex64, ldc int)
+	Csyrk(ul Uplo, t Transpose, n, k int, alpha complex64, a []complex64, lda int, beta complex64, c []complex64, ldc int)
+	Csyr2k(ul Uplo, t Transpose, n, k int, alpha complex64, a []complex64, lda int, b []complex64, ldb int, beta complex64, c []complex64, ldc int)
+	Ctrmm(s Side, ul Uplo, tA Transpose, d Diag, m, n int, alpha complex64, a []complex64, lda int, b []complex64, ldb int)
+	Ctrsm(s Side, ul Uplo, tA Transpose, d Diag, m, n int, alpha complex64, a []complex64, lda int, b []complex64, ldb int)
+	Chemm(s Side, ul Uplo, m, n int, alpha complex64, a []complex64, lda int, b []complex64, ldb int, beta complex64, c []complex64, ldc int)
+	Cherk(ul Uplo, t Transpose, n, k int, alpha float32, a []complex64, lda int, beta float32, c []complex64, ldc int)
+	Cher2k(ul Uplo, t Transpose, n, k int, alpha complex64, a []complex64, lda int, b []complex64, ldb int, beta float32, c []complex64, ldc int)
+}
+
+// Complex128 implements the double precision complex BLAS routines.
+type Complex128 interface {
+	Complex128Level1
+	Complex128Level2
+	Complex128Level3
+}
+
+// Complex128Level1 implements the double precision complex BLAS Level 1 routines.
+type Complex128Level1 interface {
+	Zdotu(n int, x []complex128, incX int, y []complex128, incY int) (dotu complex128)
+	Zdotc(n int, x []complex128, incX int, y []complex128, incY int) (dotc complex128)
+	Dznrm2(n int, x []complex128, incX int) float64
+	Dzasum(n int, x []complex128, incX int) float64
+	Izamax(n int, x []complex128, incX int) int
+	Zswap(n int, x []complex128, incX int, y []complex128, incY int)
+	Zcopy(n int, x []complex128, incX int, y []complex128, incY int)
+	Zaxpy(n int, alpha complex128, x []complex128, incX int, y []complex128, incY int)
+	Zscal(n int, alpha complex128, x []complex128, incX int)
+	Zdscal(n int, alpha float64, x []complex128, incX int)
+}
+
+// Complex128Level2 implements the double precision complex BLAS Level 2 routines.
+type Complex128Level2 interface {
+	Zgemv(tA Transpose, m, n int, alpha complex128, a []complex128, lda int, x []complex128, incX int, beta complex128, y []complex128, incY int)
+	Zgbmv(tA Transpose, m, n int, kL int, kU int, alpha complex128, a []complex128, lda int, x []complex128, incX int, beta complex128, y []complex128, incY int)
+	Ztrmv(ul Uplo, tA Transpose, d Diag, n int, a []complex128, lda int, x []complex128, incX int)
+	Ztbmv(ul Uplo, tA Transpose, d Diag, n, k int, a []complex128, lda int, x []complex128, incX int)
+	Ztpmv(ul Uplo, tA Transpose, d Diag, n int, ap []complex128, x []complex128, incX int)
+	Ztrsv(ul Uplo, tA Transpose, d Diag, n int, a []complex128, lda int, x []complex128, incX int)
+	Ztbsv(ul Uplo, tA Transpose, d Diag, n, k int, a []complex128, lda int, x []complex128, incX int)
+	Ztpsv(ul Uplo, tA Transpose, d Diag, n int, ap []complex128, x []complex128, incX int)
+	Zhemv(ul Uplo, n int, alpha complex128, a []complex128, lda int, x []complex128, incX int, beta complex128, y []complex128, incY int)
+	Zhbmv(ul Uplo, n, k int, alpha complex128, a []complex128, lda int, x []complex128, incX int, beta complex128, y []complex128, incY int)
+	Zhpmv(ul Uplo, n int, alpha complex128, ap []complex128, x []complex128, incX int, beta complex128, y []complex128, incY int)
+	Zgeru(m, n int, alpha complex128, x []complex128, incX int, y []complex128, incY int, a []complex128, lda int)
+	Zgerc(m, n int, alpha complex128, x []complex128, incX int, y []complex128, incY int, a []complex128, lda int)
+	Zher(ul Uplo, n int, alpha float64, x []complex128, incX int, a []complex128, lda int)
+	Zhpr(ul Uplo, n int, alpha float64, x []complex128, incX int, a []complex128)
+	Zher2(ul Uplo, n int, alpha complex128, x []complex128, incX int, y []complex128, incY int, a []complex128, lda int)
+	Zhpr2(ul Uplo, n int, alpha complex128, x []complex128, incX int, y []complex128, incY int, ap []complex128)
+}
+
+// Complex128Level3 implements the double precision complex BLAS Level 3 routines.
+type Complex128Level3 interface {
+	Zgemm(tA, tB Transpose, m, n, k int, alpha complex128, a []complex128, lda int, b []complex128, ldb int, beta complex128, c []complex128, ldc int)
+	Zsymm(s Side, ul Uplo, m, n int, alpha complex128, a []complex128, lda int, b []complex128, ldb int, beta complex128, c []complex128, ldc int)
+	Zsyrk(ul Uplo, t Transpose, n, k int, alpha complex128, a []complex128, lda int, beta complex128, c []complex128, ldc int)
+	Zsyr2k(ul Uplo, t Transpose, n, k int, alpha complex128, a []complex128, lda int, b []complex128, ldb int, beta complex128, c []complex128, ldc int)
+	Ztrmm(s Side, ul Uplo, tA Transpose, d Diag, m, n int, alpha complex128, a []complex128, lda int, b []complex128, ldb int)
+	Ztrsm(s Side, ul Uplo, tA Transpose, d Diag, m, n int, alpha complex128, a []complex128, lda int, b []complex128, ldb int)
+	Zhemm(s Side, ul Uplo, m, n int, alpha complex128, a []complex128, lda int, b []complex128, ldb int, beta complex128, c []complex128, ldc int)
+	Zherk(ul Uplo, t Transpose, n, k int, alpha float64, a []complex128, lda int, beta float64, c []complex128, ldc int)
+	Zher2k(ul Uplo, t Transpose, n, k int, alpha complex128, a []complex128, lda int, b []complex128, ldb int, beta float64, c []complex128, ldc int)
+}

vendor/gonum.org/v1/gonum/blas/blas64/blas64.go

@@ -0,0 +1,469 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package blas64
+
+import (
+	"gonum.org/v1/gonum/blas"
+	"gonum.org/v1/gonum/blas/gonum"
+)
+
+var blas64 blas.Float64 = gonum.Implementation{}
+
+// Use sets the BLAS float64 implementation to be used by subsequent BLAS calls.
+// The default implementation is
+// gonum.org/v1/gonum/blas/gonum.Implementation.
+func Use(b blas.Float64) {
+	blas64 = b
+}
+
+// Implementation returns the current BLAS float64 implementation.
+//
+// Implementation allows direct calls to the current the BLAS float64 implementation
+// giving finer control of parameters.
+func Implementation() blas.Float64 {
+	return blas64
+}
+
+// Vector represents a vector with an associated element increment.
+type Vector struct {
+	N    int
+	Data []float64
+	Inc  int
+}
+
+// General represents a matrix using the conventional storage scheme.
+type General struct {
+	Rows, Cols int
+	Data       []float64
+	Stride     int
+}
+
+// Band represents a band matrix using the band storage scheme.
+type Band struct {
+	Rows, Cols int
+	KL, KU     int
+	Data       []float64
+	Stride     int
+}
+
+// Triangular represents a triangular matrix using the conventional storage scheme.
+type Triangular struct {
+	Uplo   blas.Uplo
+	Diag   blas.Diag
+	N      int
+	Data   []float64
+	Stride int
+}
+
+// TriangularBand represents a triangular matrix using the band storage scheme.
+type TriangularBand struct {
+	Uplo   blas.Uplo
+	Diag   blas.Diag
+	N, K   int
+	Data   []float64
+	Stride int
+}
+
+// TriangularPacked represents a triangular matrix using the packed storage scheme.
+type TriangularPacked struct {
+	Uplo blas.Uplo
+	Diag blas.Diag
+	N    int
+	Data []float64
+}
+
+// Symmetric represents a symmetric matrix using the conventional storage scheme.
+type Symmetric struct {
+	Uplo   blas.Uplo
+	N      int
+	Data   []float64
+	Stride int
+}
+
+// SymmetricBand represents a symmetric matrix using the band storage scheme.
+type SymmetricBand struct {
+	Uplo   blas.Uplo
+	N, K   int
+	Data   []float64
+	Stride int
+}
+
+// SymmetricPacked represents a symmetric matrix using the packed storage scheme.
+type SymmetricPacked struct {
+	Uplo blas.Uplo
+	N    int
+	Data []float64
+}
+
+// Level 1
+
+const (
+	negInc    = "blas64: negative vector increment"
+	badLength = "blas64: vector length mismatch"
+)
+
+// Dot computes the dot product of the two vectors:
+//  \sum_i x[i]*y[i].
+func Dot(x, y Vector) float64 {
+	if x.N != y.N {
+		panic(badLength)
+	}
+	return blas64.Ddot(x.N, x.Data, x.Inc, y.Data, y.Inc)
+}
+
+// Nrm2 computes the Euclidean norm of the vector x:
+//  sqrt(\sum_i x[i]*x[i]).
+//
+// Nrm2 will panic if the vector increment is negative.
+func Nrm2(x Vector) float64 {
+	if x.Inc < 0 {
+		panic(negInc)
+	}
+	return blas64.Dnrm2(x.N, x.Data, x.Inc)
+}
+
+// Asum computes the sum of the absolute values of the elements of x:
+//  \sum_i |x[i]|.
+//
+// Asum will panic if the vector increment is negative.
+func Asum(x Vector) float64 {
+	if x.Inc < 0 {
+		panic(negInc)
+	}
+	return blas64.Dasum(x.N, x.Data, x.Inc)
+}
+
+// Iamax returns the index of an element of x with the largest absolute value.
+// If there are multiple such indices the earliest is returned.
+// Iamax returns -1 if n == 0.
+//
+// Iamax will panic if the vector increment is negative.
+func Iamax(x Vector) int {
+	if x.Inc < 0 {
+		panic(negInc)
+	}
+	return blas64.Idamax(x.N, x.Data, x.Inc)
+}
+
+// Swap exchanges the elements of the two vectors:
+//  x[i], y[i] = y[i], x[i] for all i.
+func Swap(x, y Vector) {
+	if x.N != y.N {
+		panic(badLength)
+	}
+	blas64.Dswap(x.N, x.Data, x.Inc, y.Data, y.Inc)
+}
+
+// Copy copies the elements of x into the elements of y:
+//  y[i] = x[i] for all i.
+// Copy requires that the lengths of x and y match and will panic otherwise.
+func Copy(x, y Vector) {
+	if x.N != y.N {
+		panic(badLength)
+	}
+	blas64.Dcopy(x.N, x.Data, x.Inc, y.Data, y.Inc)
+}
+
+// Axpy adds x scaled by alpha to y:
+//  y[i] += alpha*x[i] for all i.
+func Axpy(alpha float64, x, y Vector) {
+	if x.N != y.N {
+		panic(badLength)
+	}
+	blas64.Daxpy(x.N, alpha, x.Data, x.Inc, y.Data, y.Inc)
+}
+
+// Rotg computes the parameters of a Givens plane rotation so that
+//  ⎡ c s⎤   ⎡a⎤   ⎡r⎤
+//  ⎣-s c⎦ * ⎣b⎦ = ⎣0⎦
+// where a and b are the Cartesian coordinates of a given point.
+// c, s, and r are defined as
+//  r = ±Sqrt(a^2 + b^2),
+//  c = a/r, the cosine of the rotation angle,
+//  s = a/r, the sine of the rotation angle,
+// and z is defined such that
+//  if |a| > |b|,        z = s,
+//  otherwise if c != 0, z = 1/c,
+//  otherwise            z = 1.
+func Rotg(a, b float64) (c, s, r, z float64) {
+	return blas64.Drotg(a, b)
+}
+
+// Rotmg computes the modified Givens rotation. See
+// http://www.netlib.org/lapack/explore-html/df/deb/drotmg_8f.html
+// for more details.
+func Rotmg(d1, d2, b1, b2 float64) (p blas.DrotmParams, rd1, rd2, rb1 float64) {
+	return blas64.Drotmg(d1, d2, b1, b2)
+}
+
+// Rot applies a plane transformation to n points represented by the vectors x
+// and y:
+//  x[i] =  c*x[i] + s*y[i],
+//  y[i] = -s*x[i] + c*y[i], for all i.
+func Rot(x, y Vector, c, s float64) {
+	if x.N != y.N {
+		panic(badLength)
+	}
+	blas64.Drot(x.N, x.Data, x.Inc, y.Data, y.Inc, c, s)
+}
+
+// Rotm applies the modified Givens rotation to n points represented by the
+// vectors x and y.
+func Rotm(x, y Vector, p blas.DrotmParams) {
+	if x.N != y.N {
+		panic(badLength)
+	}
+	blas64.Drotm(x.N, x.Data, x.Inc, y.Data, y.Inc, p)
+}
+
+// Scal scales the vector x by alpha:
+//  x[i] *= alpha for all i.
+//
+// Scal will panic if the vector increment is negative.
+func Scal(alpha float64, x Vector) {
+	if x.Inc < 0 {
+		panic(negInc)
+	}
+	blas64.Dscal(x.N, alpha, x.Data, x.Inc)
+}
+
+// Level 2
+
+// Gemv computes
+//  y = alpha * A * x + beta * y,   if t == blas.NoTrans,
+//  y = alpha * A^T * x + beta * y, if t == blas.Trans or blas.ConjTrans,
+// where A is an m×n dense matrix, x and y are vectors, and alpha and beta are scalars.
+func Gemv(t blas.Transpose, alpha float64, a General, x Vector, beta float64, y Vector) {
+	blas64.Dgemv(t, a.Rows, a.Cols, alpha, a.Data, a.Stride, x.Data, x.Inc, beta, y.Data, y.Inc)
+}
+
+// Gbmv computes
+//  y = alpha * A * x + beta * y,   if t == blas.NoTrans,
+//  y = alpha * A^T * x + beta * y, if t == blas.Trans or blas.ConjTrans,
+// where A is an m×n band matrix, x and y are vectors, and alpha and beta are scalars.
+func Gbmv(t blas.Transpose, alpha float64, a Band, x Vector, beta float64, y Vector) {
+	blas64.Dgbmv(t, a.Rows, a.Cols, a.KL, a.KU, alpha, a.Data, a.Stride, x.Data, x.Inc, beta, y.Data, y.Inc)
+}
+
+// Trmv computes
+//  x = A * x,   if t == blas.NoTrans,
+//  x = A^T * x, if t == blas.Trans or blas.ConjTrans,
+// where A is an n×n triangular matrix, and x is a vector.
+func Trmv(t blas.Transpose, a Triangular, x Vector) {
+	blas64.Dtrmv(a.Uplo, t, a.Diag, a.N, a.Data, a.Stride, x.Data, x.Inc)
+}
+
+// Tbmv computes
+//  x = A * x,   if t == blas.NoTrans,
+//  x = A^T * x, if t == blas.Trans or blas.ConjTrans,
+// where A is an n×n triangular band matrix, and x is a vector.
+func Tbmv(t blas.Transpose, a TriangularBand, x Vector) {
+	blas64.Dtbmv(a.Uplo, t, a.Diag, a.N, a.K, a.Data, a.Stride, x.Data, x.Inc)
+}
+
+// Tpmv computes
+//  x = A * x,   if t == blas.NoTrans,
+//  x = A^T * x, if t == blas.Trans or blas.ConjTrans,
+// where A is an n×n triangular matrix in packed format, and x is a vector.
+func Tpmv(t blas.Transpose, a TriangularPacked, x Vector) {
+	blas64.Dtpmv(a.Uplo, t, a.Diag, a.N, a.Data, x.Data, x.Inc)
+}
+
+// Trsv solves
+//  A * x = b,   if t == blas.NoTrans,
+//  A^T * x = b, if t == blas.Trans or blas.ConjTrans,
+// where A is an n×n triangular matrix, and x and b are vectors.
+//
+// At entry to the function, x contains the values of b, and the result is
+// stored in-place into x.
+//
+// No test for singularity or near-singularity is included in this
+// routine. Such tests must be performed before calling this routine.
+func Trsv(t blas.Transpose, a Triangular, x Vector) {
+	blas64.Dtrsv(a.Uplo, t, a.Diag, a.N, a.Data, a.Stride, x.Data, x.Inc)
+}
+
+// Tbsv solves
+//  A * x = b,   if t == blas.NoTrans,
+//  A^T * x = b, if t == blas.Trans or blas.ConjTrans,
+// where A is an n×n triangular band matrix, and x and b are vectors.
+//
+// At entry to the function, x contains the values of b, and the result is
+// stored in place into x.
+//
+// No test for singularity or near-singularity is included in this
+// routine. Such tests must be performed before calling this routine.
+func Tbsv(t blas.Transpose, a TriangularBand, x Vector) {
+	blas64.Dtbsv(a.Uplo, t, a.Diag, a.N, a.K, a.Data, a.Stride, x.Data, x.Inc)
+}
+
+// Tpsv solves
+//  A * x = b,   if t == blas.NoTrans,
+//  A^T * x = b, if t == blas.Trans or blas.ConjTrans,
+// where A is an n×n triangular matrix in packed format, and x and b are
+// vectors.
+//
+// At entry to the function, x contains the values of b, and the result is
+// stored in place into x.
+//
+// No test for singularity or near-singularity is included in this
+// routine. Such tests must be performed before calling this routine.
+func Tpsv(t blas.Transpose, a TriangularPacked, x Vector) {
+	blas64.Dtpsv(a.Uplo, t, a.Diag, a.N, a.Data, x.Data, x.Inc)
+}
+
+// Symv computes
+//    y = alpha * A * x + beta * y,
+// where A is an n×n symmetric matrix, x and y are vectors, and alpha and
+// beta are scalars.
+func Symv(alpha float64, a Symmetric, x Vector, beta float64, y Vector) {
+	blas64.Dsymv(a.Uplo, a.N, alpha, a.Data, a.Stride, x.Data, x.Inc, beta, y.Data, y.Inc)
+}
+
+// Sbmv performs
+//  y = alpha * A * x + beta * y,
+// where A is an n×n symmetric band matrix, x and y are vectors, and alpha
+// and beta are scalars.
+func Sbmv(alpha float64, a SymmetricBand, x Vector, beta float64, y Vector) {
+	blas64.Dsbmv(a.Uplo, a.N, a.K, alpha, a.Data, a.Stride, x.Data, x.Inc, beta, y.Data, y.Inc)
+}
+
+// Spmv performs
+//    y = alpha * A * x + beta * y,
+// where A is an n×n symmetric matrix in packed format, x and y are vectors,
+// and alpha and beta are scalars.
+func Spmv(alpha float64, a SymmetricPacked, x Vector, beta float64, y Vector) {
+	blas64.Dspmv(a.Uplo, a.N, alpha, a.Data, x.Data, x.Inc, beta, y.Data, y.Inc)
+}
+
+// Ger performs a rank-1 update
+//  A += alpha * x * y^T,
+// where A is an m×n dense matrix, x and y are vectors, and alpha is a scalar.
+func Ger(alpha float64, x, y Vector, a General) {
+	blas64.Dger(a.Rows, a.Cols, alpha, x.Data, x.Inc, y.Data, y.Inc, a.Data, a.Stride)
+}
+
+// Syr performs a rank-1 update
+//  A += alpha * x * x^T,
+// where A is an n×n symmetric matrix, x is a vector, and alpha is a scalar.
+func Syr(alpha float64, x Vector, a Symmetric) {
+	blas64.Dsyr(a.Uplo, a.N, alpha, x.Data, x.Inc, a.Data, a.Stride)
+}
+
+// Spr performs the rank-1 update
+//  A += alpha * x * x^T,
+// where A is an n×n symmetric matrix in packed format, x is a vector, and
+// alpha is a scalar.
+func Spr(alpha float64, x Vector, a SymmetricPacked) {
+	blas64.Dspr(a.Uplo, a.N, alpha, x.Data, x.Inc, a.Data)
+}
+
+// Syr2 performs a rank-2 update
+//  A += alpha * x * y^T + alpha * y * x^T,
+// where A is a symmetric n×n matrix, x and y are vectors, and alpha is a scalar.
+func Syr2(alpha float64, x, y Vector, a Symmetric) {
+	blas64.Dsyr2(a.Uplo, a.N, alpha, x.Data, x.Inc, y.Data, y.Inc, a.Data, a.Stride)
+}
+
+// Spr2 performs a rank-2 update
+//  A += alpha * x * y^T + alpha * y * x^T,
+// where A is an n×n symmetric matrix in packed format, x and y are vectors,
+// and alpha is a scalar.
+func Spr2(alpha float64, x, y Vector, a SymmetricPacked) {
+	blas64.Dspr2(a.Uplo, a.N, alpha, x.Data, x.Inc, y.Data, y.Inc, a.Data)
+}
+
+// Level 3
+
+// Gemm computes
+//  C = alpha * A * B + beta * C,
+// where A, B, and C are dense matrices, and alpha and beta are scalars.
+// tA and tB specify whether A or B are transposed.
+func Gemm(tA, tB blas.Transpose, alpha float64, a, b General, beta float64, c General) {
+	var m, n, k int
+	if tA == blas.NoTrans {
+		m, k = a.Rows, a.Cols
+	} else {
+		m, k = a.Cols, a.Rows
+	}
+	if tB == blas.NoTrans {
+		n = b.Cols
+	} else {
+		n = b.Rows
+	}
+	blas64.Dgemm(tA, tB, m, n, k, alpha, a.Data, a.Stride, b.Data, b.Stride, beta, c.Data, c.Stride)
+}
+
+// Symm performs
+//  C = alpha * A * B + beta * C, if s == blas.Left,
+//  C = alpha * B * A + beta * C, if s == blas.Right,
+// where A is an n×n or m×m symmetric matrix, B and C are m×n matrices, and
+// alpha is a scalar.
+func Symm(s blas.Side, alpha float64, a Symmetric, b General, beta float64, c General) {
+	var m, n int
+	if s == blas.Left {
+		m, n = a.N, b.Cols
+	} else {
+		m, n = b.Rows, a.N
+	}
+	blas64.Dsymm(s, a.Uplo, m, n, alpha, a.Data, a.Stride, b.Data, b.Stride, beta, c.Data, c.Stride)
+}
+
+// Syrk performs a symmetric rank-k update
+//  C = alpha * A * A^T + beta * C, if t == blas.NoTrans,
+//  C = alpha * A^T * A + beta * C, if t == blas.Trans or blas.ConjTrans,
+// where C is an n×n symmetric matrix, A is an n×k matrix if t == blas.NoTrans and
+// a k×n matrix otherwise, and alpha and beta are scalars.
+func Syrk(t blas.Transpose, alpha float64, a General, beta float64, c Symmetric) {
+	var n, k int
+	if t == blas.NoTrans {
+		n, k = a.Rows, a.Cols
+	} else {
+		n, k = a.Cols, a.Rows
+	}
+	blas64.Dsyrk(c.Uplo, t, n, k, alpha, a.Data, a.Stride, beta, c.Data, c.Stride)
+}
+
+// Syr2k performs a symmetric rank-2k update
+//  C = alpha * A * B^T + alpha * B * A^T + beta * C, if t == blas.NoTrans,
+//  C = alpha * A^T * B + alpha * B^T * A + beta * C, if t == blas.Trans or blas.ConjTrans,
+// where C is an n×n symmetric matrix, A and B are n×k matrices if t == NoTrans
+// and k×n matrices otherwise, and alpha and beta are scalars.
+func Syr2k(t blas.Transpose, alpha float64, a, b General, beta float64, c Symmetric) {
+	var n, k int
+	if t == blas.NoTrans {
+		n, k = a.Rows, a.Cols
+	} else {
+		n, k = a.Cols, a.Rows
+	}
+	blas64.Dsyr2k(c.Uplo, t, n, k, alpha, a.Data, a.Stride, b.Data, b.Stride, beta, c.Data, c.Stride)
+}
+
+// Trmm performs
+//  B = alpha * A * B,   if tA == blas.NoTrans and s == blas.Left,
+//  B = alpha * A^T * B, if tA == blas.Trans or blas.ConjTrans, and s == blas.Left,
+//  B = alpha * B * A,   if tA == blas.NoTrans and s == blas.Right,
+//  B = alpha * B * A^T, if tA == blas.Trans or blas.ConjTrans, and s == blas.Right,
+// where A is an n×n or m×m triangular matrix, B is an m×n matrix, and alpha is
+// a scalar.
+func Trmm(s blas.Side, tA blas.Transpose, alpha float64, a Triangular, b General) {
+	blas64.Dtrmm(s, a.Uplo, tA, a.Diag, b.Rows, b.Cols, alpha, a.Data, a.Stride, b.Data, b.Stride)
+}
+
+// Trsm solves
+//  A * X = alpha * B,   if tA == blas.NoTrans and s == blas.Left,
+//  A^T * X = alpha * B, if tA == blas.Trans or blas.ConjTrans, and s == blas.Left,
+//  X * A = alpha * B,   if tA == blas.NoTrans and s == blas.Right,
+//  X * A^T = alpha * B, if tA == blas.Trans or blas.ConjTrans, and s == blas.Right,
+// where A is an n×n or m×m triangular matrix, X and B are m×n matrices, and
+// alpha is a scalar.
+//
+// At entry to the function, X contains the values of B, and the result is
+// stored in-place into X.
+//
+// No check is made that A is invertible.
+func Trsm(s blas.Side, tA blas.Transpose, alpha float64, a Triangular, b General) {
+	blas64.Dtrsm(s, a.Uplo, tA, a.Diag, b.Rows, b.Cols, alpha, a.Data, a.Stride, b.Data, b.Stride)
+}

vendor/gonum.org/v1/gonum/blas/blas64/conv.go

@@ -0,0 +1,277 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package blas64
+
+import "gonum.org/v1/gonum/blas"
+
+// GeneralCols represents a matrix using the conventional column-major storage scheme.
+type GeneralCols General
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions as a and have adequate backing
+// data storage.
+func (t GeneralCols) From(a General) {
+	if t.Rows != a.Rows || t.Cols != a.Cols {
+		panic("blas64: mismatched dimension")
+	}
+	if len(t.Data) < (t.Cols-1)*t.Stride+t.Rows {
+		panic("blas64: short data slice")
+	}
+	for i := 0; i < a.Rows; i++ {
+		for j, v := range a.Data[i*a.Stride : i*a.Stride+a.Cols] {
+			t.Data[i+j*t.Stride] = v
+		}
+	}
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions as a and have adequate backing
+// data storage.
+func (t General) From(a GeneralCols) {
+	if t.Rows != a.Rows || t.Cols != a.Cols {
+		panic("blas64: mismatched dimension")
+	}
+	if len(t.Data) < (t.Rows-1)*t.Stride+t.Cols {
+		panic("blas64: short data slice")
+	}
+	for j := 0; j < a.Cols; j++ {
+		for i, v := range a.Data[j*a.Stride : j*a.Stride+a.Rows] {
+			t.Data[i*t.Stride+j] = v
+		}
+	}
+}
+
+// TriangularCols represents a matrix using the conventional column-major storage scheme.
+type TriangularCols Triangular
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, uplo and diag as a and have
+// adequate backing data storage.
+func (t TriangularCols) From(a Triangular) {
+	if t.N != a.N {
+		panic("blas64: mismatched dimension")
+	}
+	if t.Uplo != a.Uplo {
+		panic("blas64: mismatched BLAS uplo")
+	}
+	if t.Diag != a.Diag {
+		panic("blas64: mismatched BLAS diag")
+	}
+	switch a.Uplo {
+	default:
+		panic("blas64: bad BLAS uplo")
+	case blas.Upper:
+		for i := 0; i < a.N; i++ {
+			for j := i; j < a.N; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	case blas.Lower:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j <= i; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	case blas.All:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j < a.N; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	}
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, uplo and diag as a and have
+// adequate backing data storage.
+func (t Triangular) From(a TriangularCols) {
+	if t.N != a.N {
+		panic("blas64: mismatched dimension")
+	}
+	if t.Uplo != a.Uplo {
+		panic("blas64: mismatched BLAS uplo")
+	}
+	if t.Diag != a.Diag {
+		panic("blas64: mismatched BLAS diag")
+	}
+	switch a.Uplo {
+	default:
+		panic("blas64: bad BLAS uplo")
+	case blas.Upper:
+		for i := 0; i < a.N; i++ {
+			for j := i; j < a.N; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	case blas.Lower:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j <= i; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	case blas.All:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j < a.N; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	}
+}
+
+// BandCols represents a matrix using the band column-major storage scheme.
+type BandCols Band
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions and bandwidth as a and have
+// adequate backing data storage.
+func (t BandCols) From(a Band) {
+	if t.Rows != a.Rows || t.Cols != a.Cols {
+		panic("blas64: mismatched dimension")
+	}
+	if t.KL != a.KL || t.KU != a.KU {
+		panic("blas64: mismatched bandwidth")
+	}
+	if a.Stride < a.KL+a.KU+1 {
+		panic("blas64: short stride for source")
+	}
+	if t.Stride < t.KL+t.KU+1 {
+		panic("blas64: short stride for destination")
+	}
+	for i := 0; i < a.Rows; i++ {
+		for j := max(0, i-a.KL); j < min(i+a.KU+1, a.Cols); j++ {
+			t.Data[i+t.KU-j+j*t.Stride] = a.Data[j+a.KL-i+i*a.Stride]
+		}
+	}
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions and bandwidth as a and have
+// adequate backing data storage.
+func (t Band) From(a BandCols) {
+	if t.Rows != a.Rows || t.Cols != a.Cols {
+		panic("blas64: mismatched dimension")
+	}
+	if t.KL != a.KL || t.KU != a.KU {
+		panic("blas64: mismatched bandwidth")
+	}
+	if a.Stride < a.KL+a.KU+1 {
+		panic("blas64: short stride for source")
+	}
+	if t.Stride < t.KL+t.KU+1 {
+		panic("blas64: short stride for destination")
+	}
+	for j := 0; j < a.Cols; j++ {
+		for i := max(0, j-a.KU); i < min(j+a.KL+1, a.Rows); i++ {
+			t.Data[j+a.KL-i+i*a.Stride] = a.Data[i+t.KU-j+j*t.Stride]
+		}
+	}
+}
+
+// TriangularBandCols represents a symmetric matrix using the band column-major storage scheme.
+type TriangularBandCols TriangularBand
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, bandwidth and uplo as a and
+// have adequate backing data storage.
+func (t TriangularBandCols) From(a TriangularBand) {
+	if t.N != a.N {
+		panic("blas64: mismatched dimension")
+	}
+	if t.K != a.K {
+		panic("blas64: mismatched bandwidth")
+	}
+	if a.Stride < a.K+1 {
+		panic("blas64: short stride for source")
+	}
+	if t.Stride < t.K+1 {
+		panic("blas64: short stride for destination")
+	}
+	if t.Uplo != a.Uplo {
+		panic("blas64: mismatched BLAS uplo")
+	}
+	if t.Diag != a.Diag {
+		panic("blas64: mismatched BLAS diag")
+	}
+	dst := BandCols{
+		Rows: t.N, Cols: t.N,
+		Stride: t.Stride,
+		Data:   t.Data,
+	}
+	src := Band{
+		Rows: a.N, Cols: a.N,
+		Stride: a.Stride,
+		Data:   a.Data,
+	}
+	switch a.Uplo {
+	default:
+		panic("blas64: bad BLAS uplo")
+	case blas.Upper:
+		dst.KU = t.K
+		src.KU = a.K
+	case blas.Lower:
+		dst.KL = t.K
+		src.KL = a.K
+	}
+	dst.From(src)
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, bandwidth and uplo as a and
+// have adequate backing data storage.
+func (t TriangularBand) From(a TriangularBandCols) {
+	if t.N != a.N {
+		panic("blas64: mismatched dimension")
+	}
+	if t.K != a.K {
+		panic("blas64: mismatched bandwidth")
+	}
+	if a.Stride < a.K+1 {
+		panic("blas64: short stride for source")
+	}
+	if t.Stride < t.K+1 {
+		panic("blas64: short stride for destination")
+	}
+	if t.Uplo != a.Uplo {
+		panic("blas64: mismatched BLAS uplo")
+	}
+	if t.Diag != a.Diag {
+		panic("blas64: mismatched BLAS diag")
+	}
+	dst := Band{
+		Rows: t.N, Cols: t.N,
+		Stride: t.Stride,
+		Data:   t.Data,
+	}
+	src := BandCols{
+		Rows: a.N, Cols: a.N,
+		Stride: a.Stride,
+		Data:   a.Data,
+	}
+	switch a.Uplo {
+	default:
+		panic("blas64: bad BLAS uplo")
+	case blas.Upper:
+		dst.KU = t.K
+		src.KU = a.K
+	case blas.Lower:
+		dst.KL = t.K
+		src.KL = a.K
+	}
+	dst.From(src)
+}
+
+func min(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
+
+func max(a, b int) int {
+	if a > b {
+		return a
+	}
+	return b
+}

vendor/gonum.org/v1/gonum/blas/blas64/conv_symmetric.go

@@ -0,0 +1,153 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package blas64
+
+import "gonum.org/v1/gonum/blas"
+
+// SymmetricCols represents a matrix using the conventional column-major storage scheme.
+type SymmetricCols Symmetric
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions and uplo as a and have adequate
+// backing data storage.
+func (t SymmetricCols) From(a Symmetric) {
+	if t.N != a.N {
+		panic("blas64: mismatched dimension")
+	}
+	if t.Uplo != a.Uplo {
+		panic("blas64: mismatched BLAS uplo")
+	}
+	switch a.Uplo {
+	default:
+		panic("blas64: bad BLAS uplo")
+	case blas.Upper:
+		for i := 0; i < a.N; i++ {
+			for j := i; j < a.N; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	case blas.Lower:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j <= i; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	}
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions and uplo as a and have adequate
+// backing data storage.
+func (t Symmetric) From(a SymmetricCols) {
+	if t.N != a.N {
+		panic("blas64: mismatched dimension")
+	}
+	if t.Uplo != a.Uplo {
+		panic("blas64: mismatched BLAS uplo")
+	}
+	switch a.Uplo {
+	default:
+		panic("blas64: bad BLAS uplo")
+	case blas.Upper:
+		for i := 0; i < a.N; i++ {
+			for j := i; j < a.N; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	case blas.Lower:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j <= i; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	}
+}
+
+// SymmetricBandCols represents a symmetric matrix using the band column-major storage scheme.
+type SymmetricBandCols SymmetricBand
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, bandwidth and uplo as a and
+// have adequate backing data storage.
+func (t SymmetricBandCols) From(a SymmetricBand) {
+	if t.N != a.N {
+		panic("blas64: mismatched dimension")
+	}
+	if t.K != a.K {
+		panic("blas64: mismatched bandwidth")
+	}
+	if a.Stride < a.K+1 {
+		panic("blas64: short stride for source")
+	}
+	if t.Stride < t.K+1 {
+		panic("blas64: short stride for destination")
+	}
+	if t.Uplo != a.Uplo {
+		panic("blas64: mismatched BLAS uplo")
+	}
+	dst := BandCols{
+		Rows: t.N, Cols: t.N,
+		Stride: t.Stride,
+		Data:   t.Data,
+	}
+	src := Band{
+		Rows: a.N, Cols: a.N,
+		Stride: a.Stride,
+		Data:   a.Data,
+	}
+	switch a.Uplo {
+	default:
+		panic("blas64: bad BLAS uplo")
+	case blas.Upper:
+		dst.KU = t.K
+		src.KU = a.K
+	case blas.Lower:
+		dst.KL = t.K
+		src.KL = a.K
+	}
+	dst.From(src)
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, bandwidth and uplo as a and
+// have adequate backing data storage.
+func (t SymmetricBand) From(a SymmetricBandCols) {
+	if t.N != a.N {
+		panic("blas64: mismatched dimension")
+	}
+	if t.K != a.K {
+		panic("blas64: mismatched bandwidth")
+	}
+	if a.Stride < a.K+1 {
+		panic("blas64: short stride for source")
+	}
+	if t.Stride < t.K+1 {
+		panic("blas64: short stride for destination")
+	}
+	if t.Uplo != a.Uplo {
+		panic("blas64: mismatched BLAS uplo")
+	}
+	dst := Band{
+		Rows: t.N, Cols: t.N,
+		Stride: t.Stride,
+		Data:   t.Data,
+	}
+	src := BandCols{
+		Rows: a.N, Cols: a.N,
+		Stride: a.Stride,
+		Data:   a.Data,
+	}
+	switch a.Uplo {
+	default:
+		panic("blas64: bad BLAS uplo")
+	case blas.Upper:
+		dst.KU = t.K
+		src.KU = a.K
+	case blas.Lower:
+		dst.KL = t.K
+		src.KL = a.K
+	}
+	dst.From(src)
+}

vendor/gonum.org/v1/gonum/blas/blas64/doc.go

@@ -0,0 +1,6 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package blas64 provides a simple interface to the float64 BLAS API.
+package blas64 // import "gonum.org/v1/gonum/blas/blas64"

vendor/gonum.org/v1/gonum/blas/cblas128/cblas128.go

@@ -0,0 +1,508 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package cblas128
+
+import (
+	"gonum.org/v1/gonum/blas"
+	"gonum.org/v1/gonum/blas/gonum"
+)
+
+var cblas128 blas.Complex128 = gonum.Implementation{}
+
+// Use sets the BLAS complex128 implementation to be used by subsequent BLAS calls.
+// The default implementation is
+// gonum.org/v1/gonum/blas/gonum.Implementation.
+func Use(b blas.Complex128) {
+	cblas128 = b
+}
+
+// Implementation returns the current BLAS complex128 implementation.
+//
+// Implementation allows direct calls to the current the BLAS complex128 implementation
+// giving finer control of parameters.
+func Implementation() blas.Complex128 {
+	return cblas128
+}
+
+// Vector represents a vector with an associated element increment.
+type Vector struct {
+	Inc  int
+	Data []complex128
+}
+
+// General represents a matrix using the conventional storage scheme.
+type General struct {
+	Rows, Cols int
+	Stride     int
+	Data       []complex128
+}
+
+// Band represents a band matrix using the band storage scheme.
+type Band struct {
+	Rows, Cols int
+	KL, KU     int
+	Stride     int
+	Data       []complex128
+}
+
+// Triangular represents a triangular matrix using the conventional storage scheme.
+type Triangular struct {
+	N      int
+	Stride int
+	Data   []complex128
+	Uplo   blas.Uplo
+	Diag   blas.Diag
+}
+
+// TriangularBand represents a triangular matrix using the band storage scheme.
+type TriangularBand struct {
+	N, K   int
+	Stride int
+	Data   []complex128
+	Uplo   blas.Uplo
+	Diag   blas.Diag
+}
+
+// TriangularPacked represents a triangular matrix using the packed storage scheme.
+type TriangularPacked struct {
+	N    int
+	Data []complex128
+	Uplo blas.Uplo
+	Diag blas.Diag
+}
+
+// Symmetric represents a symmetric matrix using the conventional storage scheme.
+type Symmetric struct {
+	N      int
+	Stride int
+	Data   []complex128
+	Uplo   blas.Uplo
+}
+
+// SymmetricBand represents a symmetric matrix using the band storage scheme.
+type SymmetricBand struct {
+	N, K   int
+	Stride int
+	Data   []complex128
+	Uplo   blas.Uplo
+}
+
+// SymmetricPacked represents a symmetric matrix using the packed storage scheme.
+type SymmetricPacked struct {
+	N    int
+	Data []complex128
+	Uplo blas.Uplo
+}
+
+// Hermitian represents an Hermitian matrix using the conventional storage scheme.
+type Hermitian Symmetric
+
+// HermitianBand represents an Hermitian matrix using the band storage scheme.
+type HermitianBand SymmetricBand
+
+// HermitianPacked represents an Hermitian matrix using the packed storage scheme.
+type HermitianPacked SymmetricPacked
+
+// Level 1
+
+const negInc = "cblas128: negative vector increment"
+
+// Dotu computes the dot product of the two vectors without
+// complex conjugation:
+//  x^T * y.
+func Dotu(n int, x, y Vector) complex128 {
+	return cblas128.Zdotu(n, x.Data, x.Inc, y.Data, y.Inc)
+}
+
+// Dotc computes the dot product of the two vectors with
+// complex conjugation:
+//  x^H * y.
+func Dotc(n int, x, y Vector) complex128 {
+	return cblas128.Zdotc(n, x.Data, x.Inc, y.Data, y.Inc)
+}
+
+// Nrm2 computes the Euclidean norm of the vector x:
+//  sqrt(\sum_i x[i] * x[i]).
+//
+// Nrm2 will panic if the vector increment is negative.
+func Nrm2(n int, x Vector) float64 {
+	if x.Inc < 0 {
+		panic(negInc)
+	}
+	return cblas128.Dznrm2(n, x.Data, x.Inc)
+}
+
+// Asum computes the sum of magnitudes of the real and imaginary parts of
+// elements of the vector x:
+//  \sum_i (|Re x[i]| + |Im x[i]|).
+//
+// Asum will panic if the vector increment is negative.
+func Asum(n int, x Vector) float64 {
+	if x.Inc < 0 {
+		panic(negInc)
+	}
+	return cblas128.Dzasum(n, x.Data, x.Inc)
+}
+
+// Iamax returns the index of an element of x with the largest sum of
+// magnitudes of the real and imaginary parts (|Re x[i]|+|Im x[i]|).
+// If there are multiple such indices, the earliest is returned.
+//
+// Iamax returns -1 if n == 0.
+//
+// Iamax will panic if the vector increment is negative.
+func Iamax(n int, x Vector) int {
+	if x.Inc < 0 {
+		panic(negInc)
+	}
+	return cblas128.Izamax(n, x.Data, x.Inc)
+}
+
+// Swap exchanges the elements of two vectors:
+//  x[i], y[i] = y[i], x[i] for all i.
+func Swap(n int, x, y Vector) {
+	cblas128.Zswap(n, x.Data, x.Inc, y.Data, y.Inc)
+}
+
+// Copy copies the elements of x into the elements of y:
+//  y[i] = x[i] for all i.
+func Copy(n int, x, y Vector) {
+	cblas128.Zcopy(n, x.Data, x.Inc, y.Data, y.Inc)
+}
+
+// Axpy computes
+//  y = alpha * x + y,
+// where x and y are vectors, and alpha is a scalar.
+func Axpy(n int, alpha complex128, x, y Vector) {
+	cblas128.Zaxpy(n, alpha, x.Data, x.Inc, y.Data, y.Inc)
+}
+
+// Scal computes
+//  x = alpha * x,
+// where x is a vector, and alpha is a scalar.
+//
+// Scal will panic if the vector increment is negative.
+func Scal(n int, alpha complex128, x Vector) {
+	if x.Inc < 0 {
+		panic(negInc)
+	}
+	cblas128.Zscal(n, alpha, x.Data, x.Inc)
+}
+
+// Dscal computes
+//  x = alpha * x,
+// where x is a vector, and alpha is a real scalar.
+//
+// Dscal will panic if the vector increment is negative.
+func Dscal(n int, alpha float64, x Vector) {
+	if x.Inc < 0 {
+		panic(negInc)
+	}
+	cblas128.Zdscal(n, alpha, x.Data, x.Inc)
+}
+
+// Level 2
+
+// Gemv computes
+//  y = alpha * A * x + beta * y,   if t == blas.NoTrans,
+//  y = alpha * A^T * x + beta * y, if t == blas.Trans,
+//  y = alpha * A^H * x + beta * y, if t == blas.ConjTrans,
+// where A is an m×n dense matrix, x and y are vectors, and alpha and beta are
+// scalars.
+func Gemv(t blas.Transpose, alpha complex128, a General, x Vector, beta complex128, y Vector) {
+	cblas128.Zgemv(t, a.Rows, a.Cols, alpha, a.Data, a.Stride, x.Data, x.Inc, beta, y.Data, y.Inc)
+}
+
+// Gbmv computes
+//  y = alpha * A * x + beta * y,   if t == blas.NoTrans,
+//  y = alpha * A^T * x + beta * y, if t == blas.Trans,
+//  y = alpha * A^H * x + beta * y, if t == blas.ConjTrans,
+// where A is an m×n band matrix, x and y are vectors, and alpha and beta are
+// scalars.
+func Gbmv(t blas.Transpose, alpha complex128, a Band, x Vector, beta complex128, y Vector) {
+	cblas128.Zgbmv(t, a.Rows, a.Cols, a.KL, a.KU, alpha, a.Data, a.Stride, x.Data, x.Inc, beta, y.Data, y.Inc)
+}
+
+// Trmv computes
+//  x = A * x,   if t == blas.NoTrans,
+//  x = A^T * x, if t == blas.Trans,
+//  x = A^H * x, if t == blas.ConjTrans,
+// where A is an n×n triangular matrix, and x is a vector.
+func Trmv(t blas.Transpose, a Triangular, x Vector) {
+	cblas128.Ztrmv(a.Uplo, t, a.Diag, a.N, a.Data, a.Stride, x.Data, x.Inc)
+}
+
+// Tbmv computes
+//  x = A * x,   if t == blas.NoTrans,
+//  x = A^T * x, if t == blas.Trans,
+//  x = A^H * x, if t == blas.ConjTrans,
+// where A is an n×n triangular band matrix, and x is a vector.
+func Tbmv(t blas.Transpose, a TriangularBand, x Vector) {
+	cblas128.Ztbmv(a.Uplo, t, a.Diag, a.N, a.K, a.Data, a.Stride, x.Data, x.Inc)
+}
+
+// Tpmv computes
+//  x = A * x,   if t == blas.NoTrans,
+//  x = A^T * x, if t == blas.Trans,
+//  x = A^H * x, if t == blas.ConjTrans,
+// where A is an n×n triangular matrix in packed format, and x is a vector.
+func Tpmv(t blas.Transpose, a TriangularPacked, x Vector) {
+	cblas128.Ztpmv(a.Uplo, t, a.Diag, a.N, a.Data, x.Data, x.Inc)
+}
+
+// Trsv solves
+//  A * x = b,   if t == blas.NoTrans,
+//  A^T * x = b, if t == blas.Trans,
+//  A^H * x = b, if t == blas.ConjTrans,
+// where A is an n×n triangular matrix and x is a vector.
+//
+// At entry to the function, x contains the values of b, and the result is
+// stored in-place into x.
+//
+// No test for singularity or near-singularity is included in this
+// routine. Such tests must be performed before calling this routine.
+func Trsv(t blas.Transpose, a Triangular, x Vector) {
+	cblas128.Ztrsv(a.Uplo, t, a.Diag, a.N, a.Data, a.Stride, x.Data, x.Inc)
+}
+
+// Tbsv solves
+//  A * x = b,   if t == blas.NoTrans,
+//  A^T * x = b, if t == blas.Trans,
+//  A^H * x = b, if t == blas.ConjTrans,
+// where A is an n×n triangular band matrix, and x is a vector.
+//
+// At entry to the function, x contains the values of b, and the result is
+// stored in-place into x.
+//
+// No test for singularity or near-singularity is included in this
+// routine. Such tests must be performed before calling this routine.
+func Tbsv(t blas.Transpose, a TriangularBand, x Vector) {
+	cblas128.Ztbsv(a.Uplo, t, a.Diag, a.N, a.K, a.Data, a.Stride, x.Data, x.Inc)
+}
+
+// Tpsv solves
+//  A * x = b,   if t == blas.NoTrans,
+//  A^T * x = b, if t == blas.Trans,
+//  A^H * x = b, if t == blas.ConjTrans,
+// where A is an n×n triangular matrix in packed format and x is a vector.
+//
+// At entry to the function, x contains the values of b, and the result is
+// stored in-place into x.
+//
+// No test for singularity or near-singularity is included in this
+// routine. Such tests must be performed before calling this routine.
+func Tpsv(t blas.Transpose, a TriangularPacked, x Vector) {
+	cblas128.Ztpsv(a.Uplo, t, a.Diag, a.N, a.Data, x.Data, x.Inc)
+}
+
+// Hemv computes
+//  y = alpha * A * x + beta * y,
+// where A is an n×n Hermitian matrix, x and y are vectors, and alpha and
+// beta are scalars.
+func Hemv(alpha complex128, a Hermitian, x Vector, beta complex128, y Vector) {
+	cblas128.Zhemv(a.Uplo, a.N, alpha, a.Data, a.Stride, x.Data, x.Inc, beta, y.Data, y.Inc)
+}
+
+// Hbmv performs
+//  y = alpha * A * x + beta * y,
+// where A is an n×n Hermitian band matrix, x and y are vectors, and alpha
+// and beta are scalars.
+func Hbmv(alpha complex128, a HermitianBand, x Vector, beta complex128, y Vector) {
+	cblas128.Zhbmv(a.Uplo, a.N, a.K, alpha, a.Data, a.Stride, x.Data, x.Inc, beta, y.Data, y.Inc)
+}
+
+// Hpmv performs
+//  y = alpha * A * x + beta * y,
+// where A is an n×n Hermitian matrix in packed format, x and y are vectors,
+// and alpha and beta are scalars.
+func Hpmv(alpha complex128, a HermitianPacked, x Vector, beta complex128, y Vector) {
+	cblas128.Zhpmv(a.Uplo, a.N, alpha, a.Data, x.Data, x.Inc, beta, y.Data, y.Inc)
+}
+
+// Geru performs a rank-1 update
+//  A += alpha * x * y^T,
+// where A is an m×n dense matrix, x and y are vectors, and alpha is a scalar.
+func Geru(alpha complex128, x, y Vector, a General) {
+	cblas128.Zgeru(a.Rows, a.Cols, alpha, x.Data, x.Inc, y.Data, y.Inc, a.Data, a.Stride)
+}
+
+// Gerc performs a rank-1 update
+//  A += alpha * x * y^H,
+// where A is an m×n dense matrix, x and y are vectors, and alpha is a scalar.
+func Gerc(alpha complex128, x, y Vector, a General) {
+	cblas128.Zgerc(a.Rows, a.Cols, alpha, x.Data, x.Inc, y.Data, y.Inc, a.Data, a.Stride)
+}
+
+// Her performs a rank-1 update
+//  A += alpha * x * y^T,
+// where A is an m×n Hermitian matrix, x and y are vectors, and alpha is a scalar.
+func Her(alpha float64, x Vector, a Hermitian) {
+	cblas128.Zher(a.Uplo, a.N, alpha, x.Data, x.Inc, a.Data, a.Stride)
+}
+
+// Hpr performs a rank-1 update
+//  A += alpha * x * x^H,
+// where A is an n×n Hermitian matrix in packed format, x is a vector, and
+// alpha is a scalar.
+func Hpr(alpha float64, x Vector, a HermitianPacked) {
+	cblas128.Zhpr(a.Uplo, a.N, alpha, x.Data, x.Inc, a.Data)
+}
+
+// Her2 performs a rank-2 update
+//  A += alpha * x * y^H + conj(alpha) * y * x^H,
+// where A is an n×n Hermitian matrix, x and y are vectors, and alpha is a scalar.
+func Her2(alpha complex128, x, y Vector, a Hermitian) {
+	cblas128.Zher2(a.Uplo, a.N, alpha, x.Data, x.Inc, y.Data, y.Inc, a.Data, a.Stride)
+}
+
+// Hpr2 performs a rank-2 update
+//  A += alpha * x * y^H + conj(alpha) * y * x^H,
+// where A is an n×n Hermitian matrix in packed format, x and y are vectors,
+// and alpha is a scalar.
+func Hpr2(alpha complex128, x, y Vector, a HermitianPacked) {
+	cblas128.Zhpr2(a.Uplo, a.N, alpha, x.Data, x.Inc, y.Data, y.Inc, a.Data)
+}
+
+// Level 3
+
+// Gemm computes
+//  C = alpha * A * B + beta * C,
+// where A, B, and C are dense matrices, and alpha and beta are scalars.
+// tA and tB specify whether A or B are transposed or conjugated.
+func Gemm(tA, tB blas.Transpose, alpha complex128, a, b General, beta complex128, c General) {
+	var m, n, k int
+	if tA == blas.NoTrans {
+		m, k = a.Rows, a.Cols
+	} else {
+		m, k = a.Cols, a.Rows
+	}
+	if tB == blas.NoTrans {
+		n = b.Cols
+	} else {
+		n = b.Rows
+	}
+	cblas128.Zgemm(tA, tB, m, n, k, alpha, a.Data, a.Stride, b.Data, b.Stride, beta, c.Data, c.Stride)
+}
+
+// Symm performs
+//  C = alpha * A * B + beta * C, if s == blas.Left,
+//  C = alpha * B * A + beta * C, if s == blas.Right,
+// where A is an n×n or m×m symmetric matrix, B and C are m×n matrices, and
+// alpha and beta are scalars.
+func Symm(s blas.Side, alpha complex128, a Symmetric, b General, beta complex128, c General) {
+	var m, n int
+	if s == blas.Left {
+		m, n = a.N, b.Cols
+	} else {
+		m, n = b.Rows, a.N
+	}
+	cblas128.Zsymm(s, a.Uplo, m, n, alpha, a.Data, a.Stride, b.Data, b.Stride, beta, c.Data, c.Stride)
+}
+
+// Syrk performs a symmetric rank-k update
+//  C = alpha * A * A^T + beta * C, if t == blas.NoTrans,
+//  C = alpha * A^T * A + beta * C, if t == blas.Trans,
+// where C is an n×n symmetric matrix, A is an n×k matrix if t == blas.NoTrans
+// and a k×n matrix otherwise, and alpha and beta are scalars.
+func Syrk(t blas.Transpose, alpha complex128, a General, beta complex128, c Symmetric) {
+	var n, k int
+	if t == blas.NoTrans {
+		n, k = a.Rows, a.Cols
+	} else {
+		n, k = a.Cols, a.Rows
+	}
+	cblas128.Zsyrk(c.Uplo, t, n, k, alpha, a.Data, a.Stride, beta, c.Data, c.Stride)
+}
+
+// Syr2k performs a symmetric rank-2k update
+//  C = alpha * A * B^T + alpha * B * A^T + beta * C, if t == blas.NoTrans,
+//  C = alpha * A^T * B + alpha * B^T * A + beta * C, if t == blas.Trans,
+// where C is an n×n symmetric matrix, A and B are n×k matrices if
+// t == blas.NoTrans and k×n otherwise, and alpha and beta are scalars.
+func Syr2k(t blas.Transpose, alpha complex128, a, b General, beta complex128, c Symmetric) {
+	var n, k int
+	if t == blas.NoTrans {
+		n, k = a.Rows, a.Cols
+	} else {
+		n, k = a.Cols, a.Rows
+	}
+	cblas128.Zsyr2k(c.Uplo, t, n, k, alpha, a.Data, a.Stride, b.Data, b.Stride, beta, c.Data, c.Stride)
+}
+
+// Trmm performs
+//  B = alpha * A * B,   if tA == blas.NoTrans and s == blas.Left,
+//  B = alpha * A^T * B, if tA == blas.Trans and s == blas.Left,
+//  B = alpha * A^H * B, if tA == blas.ConjTrans and s == blas.Left,
+//  B = alpha * B * A,   if tA == blas.NoTrans and s == blas.Right,
+//  B = alpha * B * A^T, if tA == blas.Trans and s == blas.Right,
+//  B = alpha * B * A^H, if tA == blas.ConjTrans and s == blas.Right,
+// where A is an n×n or m×m triangular matrix, B is an m×n matrix, and alpha is
+// a scalar.
+func Trmm(s blas.Side, tA blas.Transpose, alpha complex128, a Triangular, b General) {
+	cblas128.Ztrmm(s, a.Uplo, tA, a.Diag, b.Rows, b.Cols, alpha, a.Data, a.Stride, b.Data, b.Stride)
+}
+
+// Trsm solves
+//  A * X = alpha * B,   if tA == blas.NoTrans and s == blas.Left,
+//  A^T * X = alpha * B, if tA == blas.Trans and s == blas.Left,
+//  A^H * X = alpha * B, if tA == blas.ConjTrans and s == blas.Left,
+//  X * A = alpha * B,   if tA == blas.NoTrans and s == blas.Right,
+//  X * A^T = alpha * B, if tA == blas.Trans and s == blas.Right,
+//  X * A^H = alpha * B, if tA == blas.ConjTrans and s == blas.Right,
+// where A is an n×n or m×m triangular matrix, X and B are m×n matrices, and
+// alpha is a scalar.
+//
+// At entry to the function, b contains the values of B, and the result is
+// stored in-place into b.
+//
+// No check is made that A is invertible.
+func Trsm(s blas.Side, tA blas.Transpose, alpha complex128, a Triangular, b General) {
+	cblas128.Ztrsm(s, a.Uplo, tA, a.Diag, b.Rows, b.Cols, alpha, a.Data, a.Stride, b.Data, b.Stride)
+}
+
+// Hemm performs
+//  C = alpha * A * B + beta * C, if s == blas.Left,
+//  C = alpha * B * A + beta * C, if s == blas.Right,
+// where A is an n×n or m×m Hermitian matrix, B and C are m×n matrices, and
+// alpha and beta are scalars.
+func Hemm(s blas.Side, alpha complex128, a Hermitian, b General, beta complex128, c General) {
+	var m, n int
+	if s == blas.Left {
+		m, n = a.N, b.Cols
+	} else {
+		m, n = b.Rows, a.N
+	}
+	cblas128.Zhemm(s, a.Uplo, m, n, alpha, a.Data, a.Stride, b.Data, b.Stride, beta, c.Data, c.Stride)
+}
+
+// Herk performs the Hermitian rank-k update
+//  C = alpha * A * A^H + beta*C, if t == blas.NoTrans,
+//  C = alpha * A^H * A + beta*C, if t == blas.ConjTrans,
+// where C is an n×n Hermitian matrix, A is an n×k matrix if t == blas.NoTrans
+// and a k×n matrix otherwise, and alpha and beta are scalars.
+func Herk(t blas.Transpose, alpha float64, a General, beta float64, c Hermitian) {
+	var n, k int
+	if t == blas.NoTrans {
+		n, k = a.Rows, a.Cols
+	} else {
+		n, k = a.Cols, a.Rows
+	}
+	cblas128.Zherk(c.Uplo, t, n, k, alpha, a.Data, a.Stride, beta, c.Data, c.Stride)
+}
+
+// Her2k performs the Hermitian rank-2k update
+//  C = alpha * A * B^H + conj(alpha) * B * A^H + beta * C, if t == blas.NoTrans,
+//  C = alpha * A^H * B + conj(alpha) * B^H * A + beta * C, if t == blas.ConjTrans,
+// where C is an n×n Hermitian matrix, A and B are n×k matrices if t == NoTrans
+// and k×n matrices otherwise, and alpha and beta are scalars.
+func Her2k(t blas.Transpose, alpha complex128, a, b General, beta float64, c Hermitian) {
+	var n, k int
+	if t == blas.NoTrans {
+		n, k = a.Rows, a.Cols
+	} else {
+		n, k = a.Cols, a.Rows
+	}
+	cblas128.Zher2k(c.Uplo, t, n, k, alpha, a.Data, a.Stride, b.Data, b.Stride, beta, c.Data, c.Stride)
+}

vendor/gonum.org/v1/gonum/blas/cblas128/conv.go

@@ -0,0 +1,279 @@
+// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.
+
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package cblas128
+
+import "gonum.org/v1/gonum/blas"
+
+// GeneralCols represents a matrix using the conventional column-major storage scheme.
+type GeneralCols General
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions as a and have adequate backing
+// data storage.
+func (t GeneralCols) From(a General) {
+	if t.Rows != a.Rows || t.Cols != a.Cols {
+		panic("cblas128: mismatched dimension")
+	}
+	if len(t.Data) < (t.Cols-1)*t.Stride+t.Rows {
+		panic("cblas128: short data slice")
+	}
+	for i := 0; i < a.Rows; i++ {
+		for j, v := range a.Data[i*a.Stride : i*a.Stride+a.Cols] {
+			t.Data[i+j*t.Stride] = v
+		}
+	}
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions as a and have adequate backing
+// data storage.
+func (t General) From(a GeneralCols) {
+	if t.Rows != a.Rows || t.Cols != a.Cols {
+		panic("cblas128: mismatched dimension")
+	}
+	if len(t.Data) < (t.Rows-1)*t.Stride+t.Cols {
+		panic("cblas128: short data slice")
+	}
+	for j := 0; j < a.Cols; j++ {
+		for i, v := range a.Data[j*a.Stride : j*a.Stride+a.Rows] {
+			t.Data[i*t.Stride+j] = v
+		}
+	}
+}
+
+// TriangularCols represents a matrix using the conventional column-major storage scheme.
+type TriangularCols Triangular
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, uplo and diag as a and have
+// adequate backing data storage.
+func (t TriangularCols) From(a Triangular) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	if t.Diag != a.Diag {
+		panic("cblas128: mismatched BLAS diag")
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		for i := 0; i < a.N; i++ {
+			for j := i; j < a.N; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	case blas.Lower:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j <= i; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	case blas.All:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j < a.N; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	}
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, uplo and diag as a and have
+// adequate backing data storage.
+func (t Triangular) From(a TriangularCols) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	if t.Diag != a.Diag {
+		panic("cblas128: mismatched BLAS diag")
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		for i := 0; i < a.N; i++ {
+			for j := i; j < a.N; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	case blas.Lower:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j <= i; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	case blas.All:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j < a.N; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	}
+}
+
+// BandCols represents a matrix using the band column-major storage scheme.
+type BandCols Band
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions and bandwidth as a and have
+// adequate backing data storage.
+func (t BandCols) From(a Band) {
+	if t.Rows != a.Rows || t.Cols != a.Cols {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.KL != a.KL || t.KU != a.KU {
+		panic("cblas128: mismatched bandwidth")
+	}
+	if a.Stride < a.KL+a.KU+1 {
+		panic("cblas128: short stride for source")
+	}
+	if t.Stride < t.KL+t.KU+1 {
+		panic("cblas128: short stride for destination")
+	}
+	for i := 0; i < a.Rows; i++ {
+		for j := max(0, i-a.KL); j < min(i+a.KU+1, a.Cols); j++ {
+			t.Data[i+t.KU-j+j*t.Stride] = a.Data[j+a.KL-i+i*a.Stride]
+		}
+	}
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions and bandwidth as a and have
+// adequate backing data storage.
+func (t Band) From(a BandCols) {
+	if t.Rows != a.Rows || t.Cols != a.Cols {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.KL != a.KL || t.KU != a.KU {
+		panic("cblas128: mismatched bandwidth")
+	}
+	if a.Stride < a.KL+a.KU+1 {
+		panic("cblas128: short stride for source")
+	}
+	if t.Stride < t.KL+t.KU+1 {
+		panic("cblas128: short stride for destination")
+	}
+	for j := 0; j < a.Cols; j++ {
+		for i := max(0, j-a.KU); i < min(j+a.KL+1, a.Rows); i++ {
+			t.Data[j+a.KL-i+i*a.Stride] = a.Data[i+t.KU-j+j*t.Stride]
+		}
+	}
+}
+
+// TriangularBandCols represents a symmetric matrix using the band column-major storage scheme.
+type TriangularBandCols TriangularBand
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, bandwidth and uplo as a and
+// have adequate backing data storage.
+func (t TriangularBandCols) From(a TriangularBand) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.K != a.K {
+		panic("cblas128: mismatched bandwidth")
+	}
+	if a.Stride < a.K+1 {
+		panic("cblas128: short stride for source")
+	}
+	if t.Stride < t.K+1 {
+		panic("cblas128: short stride for destination")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	if t.Diag != a.Diag {
+		panic("cblas128: mismatched BLAS diag")
+	}
+	dst := BandCols{
+		Rows: t.N, Cols: t.N,
+		Stride: t.Stride,
+		Data:   t.Data,
+	}
+	src := Band{
+		Rows: a.N, Cols: a.N,
+		Stride: a.Stride,
+		Data:   a.Data,
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		dst.KU = t.K
+		src.KU = a.K
+	case blas.Lower:
+		dst.KL = t.K
+		src.KL = a.K
+	}
+	dst.From(src)
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, bandwidth and uplo as a and
+// have adequate backing data storage.
+func (t TriangularBand) From(a TriangularBandCols) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.K != a.K {
+		panic("cblas128: mismatched bandwidth")
+	}
+	if a.Stride < a.K+1 {
+		panic("cblas128: short stride for source")
+	}
+	if t.Stride < t.K+1 {
+		panic("cblas128: short stride for destination")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	if t.Diag != a.Diag {
+		panic("cblas128: mismatched BLAS diag")
+	}
+	dst := Band{
+		Rows: t.N, Cols: t.N,
+		Stride: t.Stride,
+		Data:   t.Data,
+	}
+	src := BandCols{
+		Rows: a.N, Cols: a.N,
+		Stride: a.Stride,
+		Data:   a.Data,
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		dst.KU = t.K
+		src.KU = a.K
+	case blas.Lower:
+		dst.KL = t.K
+		src.KL = a.K
+	}
+	dst.From(src)
+}
+
+func min(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
+
+func max(a, b int) int {
+	if a > b {
+		return a
+	}
+	return b
+}

vendor/gonum.org/v1/gonum/blas/cblas128/conv_hermitian.go

@@ -0,0 +1,155 @@
+// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.
+
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package cblas128
+
+import "gonum.org/v1/gonum/blas"
+
+// HermitianCols represents a matrix using the conventional column-major storage scheme.
+type HermitianCols Hermitian
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions and uplo as a and have adequate
+// backing data storage.
+func (t HermitianCols) From(a Hermitian) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		for i := 0; i < a.N; i++ {
+			for j := i; j < a.N; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	case blas.Lower:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j <= i; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	}
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions and uplo as a and have adequate
+// backing data storage.
+func (t Hermitian) From(a HermitianCols) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		for i := 0; i < a.N; i++ {
+			for j := i; j < a.N; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	case blas.Lower:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j <= i; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	}
+}
+
+// HermitianBandCols represents an Hermitian matrix using the band column-major storage scheme.
+type HermitianBandCols HermitianBand
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, bandwidth and uplo as a and
+// have adequate backing data storage.
+func (t HermitianBandCols) From(a HermitianBand) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.K != a.K {
+		panic("cblas128: mismatched bandwidth")
+	}
+	if a.Stride < a.K+1 {
+		panic("cblas128: short stride for source")
+	}
+	if t.Stride < t.K+1 {
+		panic("cblas128: short stride for destination")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	dst := BandCols{
+		Rows: t.N, Cols: t.N,
+		Stride: t.Stride,
+		Data:   t.Data,
+	}
+	src := Band{
+		Rows: a.N, Cols: a.N,
+		Stride: a.Stride,
+		Data:   a.Data,
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		dst.KU = t.K
+		src.KU = a.K
+	case blas.Lower:
+		dst.KL = t.K
+		src.KL = a.K
+	}
+	dst.From(src)
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, bandwidth and uplo as a and
+// have adequate backing data storage.
+func (t HermitianBand) From(a HermitianBandCols) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.K != a.K {
+		panic("cblas128: mismatched bandwidth")
+	}
+	if a.Stride < a.K+1 {
+		panic("cblas128: short stride for source")
+	}
+	if t.Stride < t.K+1 {
+		panic("cblas128: short stride for destination")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	dst := Band{
+		Rows: t.N, Cols: t.N,
+		Stride: t.Stride,
+		Data:   t.Data,
+	}
+	src := BandCols{
+		Rows: a.N, Cols: a.N,
+		Stride: a.Stride,
+		Data:   a.Data,
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		dst.KU = t.K
+		src.KU = a.K
+	case blas.Lower:
+		dst.KL = t.K
+		src.KL = a.K
+	}
+	dst.From(src)
+}

vendor/gonum.org/v1/gonum/blas/cblas128/conv_symmetric.go

@@ -0,0 +1,155 @@
+// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.
+
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package cblas128
+
+import "gonum.org/v1/gonum/blas"
+
+// SymmetricCols represents a matrix using the conventional column-major storage scheme.
+type SymmetricCols Symmetric
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions and uplo as a and have adequate
+// backing data storage.
+func (t SymmetricCols) From(a Symmetric) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		for i := 0; i < a.N; i++ {
+			for j := i; j < a.N; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	case blas.Lower:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j <= i; j++ {
+				t.Data[i+j*t.Stride] = a.Data[i*a.Stride+j]
+			}
+		}
+	}
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions and uplo as a and have adequate
+// backing data storage.
+func (t Symmetric) From(a SymmetricCols) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		for i := 0; i < a.N; i++ {
+			for j := i; j < a.N; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	case blas.Lower:
+		for i := 0; i < a.N; i++ {
+			for j := 0; j <= i; j++ {
+				t.Data[i*t.Stride+j] = a.Data[i+j*a.Stride]
+			}
+		}
+	}
+}
+
+// SymmetricBandCols represents a symmetric matrix using the band column-major storage scheme.
+type SymmetricBandCols SymmetricBand
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, bandwidth and uplo as a and
+// have adequate backing data storage.
+func (t SymmetricBandCols) From(a SymmetricBand) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.K != a.K {
+		panic("cblas128: mismatched bandwidth")
+	}
+	if a.Stride < a.K+1 {
+		panic("cblas128: short stride for source")
+	}
+	if t.Stride < t.K+1 {
+		panic("cblas128: short stride for destination")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	dst := BandCols{
+		Rows: t.N, Cols: t.N,
+		Stride: t.Stride,
+		Data:   t.Data,
+	}
+	src := Band{
+		Rows: a.N, Cols: a.N,
+		Stride: a.Stride,
+		Data:   a.Data,
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		dst.KU = t.K
+		src.KU = a.K
+	case blas.Lower:
+		dst.KL = t.K
+		src.KL = a.K
+	}
+	dst.From(src)
+}
+
+// From fills the receiver with elements from a. The receiver
+// must have the same dimensions, bandwidth and uplo as a and
+// have adequate backing data storage.
+func (t SymmetricBand) From(a SymmetricBandCols) {
+	if t.N != a.N {
+		panic("cblas128: mismatched dimension")
+	}
+	if t.K != a.K {
+		panic("cblas128: mismatched bandwidth")
+	}
+	if a.Stride < a.K+1 {
+		panic("cblas128: short stride for source")
+	}
+	if t.Stride < t.K+1 {
+		panic("cblas128: short stride for destination")
+	}
+	if t.Uplo != a.Uplo {
+		panic("cblas128: mismatched BLAS uplo")
+	}
+	dst := Band{
+		Rows: t.N, Cols: t.N,
+		Stride: t.Stride,
+		Data:   t.Data,
+	}
+	src := BandCols{
+		Rows: a.N, Cols: a.N,
+		Stride: a.Stride,
+		Data:   a.Data,
+	}
+	switch a.Uplo {
+	default:
+		panic("cblas128: bad BLAS uplo")
+	case blas.Upper:
+		dst.KU = t.K
+		src.KU = a.K
+	case blas.Lower:
+		dst.KL = t.K
+		src.KL = a.K
+	}
+	dst.From(src)
+}

vendor/gonum.org/v1/gonum/blas/cblas128/doc.go

@@ -0,0 +1,6 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package cblas128 provides a simple interface to the complex128 BLAS API.
+package cblas128 // import "gonum.org/v1/gonum/blas/cblas128"

vendor/gonum.org/v1/gonum/blas/conversions.bash

@@ -0,0 +1,159 @@
+#!/usr/bin/env bash
+
+# Copyright ©2017 The Gonum Authors. All rights reserved.
+# Use of this source code is governed by a BSD-style
+# license that can be found in the LICENSE file.
+
+# Generate code for blas32.
+echo Generating blas32/conv.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > blas32/conv.go
+cat blas64/conv.go \
+| gofmt -r 'float64 -> float32' \
+\
+| sed -e 's/blas64/blas32/' \
+\
+>> blas32/conv.go
+
+echo Generating blas32/conv_test.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > blas32/conv_test.go
+cat blas64/conv_test.go \
+| gofmt -r 'float64 -> float32' \
+\
+| sed -e 's/blas64/blas32/' \
+      -e 's_"math"_math "gonum.org/v1/gonum/internal/math32"_' \
+\
+>> blas32/conv_test.go
+
+echo Generating blas32/conv_symmetric.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > blas32/conv_symmetric.go
+cat blas64/conv_symmetric.go \
+| gofmt -r 'float64 -> float32' \
+\
+| sed -e 's/blas64/blas32/' \
+\
+>> blas32/conv_symmetric.go
+
+echo Generating blas32/conv_symmetric_test.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > blas32/conv_symmetric_test.go
+cat blas64/conv_symmetric_test.go \
+| gofmt -r 'float64 -> float32' \
+\
+| sed -e 's/blas64/blas32/' \
+      -e 's_"math"_math "gonum.org/v1/gonum/internal/math32"_' \
+\
+>> blas32/conv_symmetric_test.go
+
+
+# Generate code for cblas128.
+echo Generating cblas128/conv.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > cblas128/conv.go
+cat blas64/conv.go \
+| gofmt -r 'float64 -> complex128' \
+\
+| sed -e 's/blas64/cblas128/' \
+\
+>> cblas128/conv.go
+
+echo Generating cblas128/conv_test.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > cblas128/conv_test.go
+cat blas64/conv_test.go \
+| gofmt -r 'float64 -> complex128' \
+\
+| sed -e 's/blas64/cblas128/' \
+      -e 's_"math"_math "math/cmplx"_' \
+\
+>> cblas128/conv_test.go
+
+echo Generating cblas128/conv_symmetric.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > cblas128/conv_symmetric.go
+cat blas64/conv_symmetric.go \
+| gofmt -r 'float64 -> complex128' \
+\
+| sed -e 's/blas64/cblas128/' \
+\
+>> cblas128/conv_symmetric.go
+
+echo Generating cblas128/conv_symmetric_test.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > cblas128/conv_symmetric_test.go
+cat blas64/conv_symmetric_test.go \
+| gofmt -r 'float64 -> complex128' \
+\
+| sed -e 's/blas64/cblas128/' \
+      -e 's_"math"_math "math/cmplx"_' \
+\
+>> cblas128/conv_symmetric_test.go
+
+echo Generating cblas128/conv_hermitian.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > cblas128/conv_hermitian.go
+cat blas64/conv_symmetric.go \
+| gofmt -r 'float64 -> complex128' \
+\
+| sed -e 's/blas64/cblas128/' \
+      -e 's/Symmetric/Hermitian/g' \
+      -e 's/a symmetric/an Hermitian/g' \
+      -e 's/symmetric/hermitian/g' \
+      -e 's/Sym/Herm/g' \
+\
+>> cblas128/conv_hermitian.go
+
+echo Generating cblas128/conv_hermitian_test.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > cblas128/conv_hermitian_test.go
+cat blas64/conv_symmetric_test.go \
+| gofmt -r 'float64 -> complex128' \
+\
+| sed -e 's/blas64/cblas128/' \
+      -e 's/Symmetric/Hermitian/g' \
+      -e 's/a symmetric/an Hermitian/g' \
+      -e 's/symmetric/hermitian/g' \
+      -e 's/Sym/Herm/g' \
+      -e 's_"math"_math "math/cmplx"_' \
+\
+>> cblas128/conv_hermitian_test.go
+
+
+# Generate code for cblas64.
+echo Generating cblas64/conv.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > cblas64/conv.go
+cat blas64/conv.go \
+| gofmt -r 'float64 -> complex64' \
+\
+| sed -e 's/blas64/cblas64/' \
+\
+>> cblas64/conv.go
+
+echo Generating cblas64/conv_test.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > cblas64/conv_test.go
+cat blas64/conv_test.go \
+| gofmt -r 'float64 -> complex64' \
+\
+| sed -e 's/blas64/cblas64/' \
+      -e 's_"math"_math "gonum.org/v1/gonum/internal/cmplx64"_' \
+\
+>> cblas64/conv_test.go
+
+echo Generating cblas64/conv_hermitian.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > cblas64/conv_hermitian.go
+cat blas64/conv_symmetric.go \
+| gofmt -r 'float64 -> complex64' \
+\
+| sed -e 's/blas64/cblas64/' \
+      -e 's/Symmetric/Hermitian/g' \
+      -e 's/a symmetric/an Hermitian/g' \
+      -e 's/symmetric/hermitian/g' \
+      -e 's/Sym/Herm/g' \
+\
+>> cblas64/conv_hermitian.go
+
+echo Generating cblas64/conv_hermitian_test.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas”; DO NOT EDIT.\n' > cblas64/conv_hermitian_test.go
+cat blas64/conv_symmetric_test.go \
+| gofmt -r 'float64 -> complex64' \
+\
+| sed -e 's/blas64/cblas64/' \
+      -e 's/Symmetric/Hermitian/g' \
+      -e 's/a symmetric/an Hermitian/g' \
+      -e 's/symmetric/hermitian/g' \
+      -e 's/Sym/Herm/g' \
+      -e 's_"math"_math "gonum.org/v1/gonum/internal/cmplx64"_' \
+\
+>> cblas64/conv_hermitian_test.go

vendor/gonum.org/v1/gonum/blas/doc.go

@@ -0,0 +1,108 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+/*
+Package blas provides interfaces for the BLAS linear algebra standard.
+
+All methods must perform appropriate parameter checking and panic if
+provided parameters that do not conform to the requirements specified
+by the BLAS standard.
+
+Quick Reference Guide to the BLAS from http://www.netlib.org/lapack/lug/node145.html
+
+This version is modified to remove the "order" option. All matrix operations are
+on row-order matrices.
+
+Level 1 BLAS
+
+	        dim scalar vector   vector   scalars              5-element prefixes
+	                                                          struct
+
+	_rotg (                                      a, b )                S, D
+	_rotmg(                              d1, d2, a, b )                S, D
+	_rot  ( n,         x, incX, y, incY,               c, s )          S, D
+	_rotm ( n,         x, incX, y, incY,                      param )  S, D
+	_swap ( n,         x, incX, y, incY )                              S, D, C, Z
+	_scal ( n,  alpha, x, incX )                                       S, D, C, Z, Cs, Zd
+	_copy ( n,         x, incX, y, incY )                              S, D, C, Z
+	_axpy ( n,  alpha, x, incX, y, incY )                              S, D, C, Z
+	_dot  ( n,         x, incX, y, incY )                              S, D, Ds
+	_dotu ( n,         x, incX, y, incY )                              C, Z
+	_dotc ( n,         x, incX, y, incY )                              C, Z
+	__dot ( n,  alpha, x, incX, y, incY )                              Sds
+	_nrm2 ( n,         x, incX )                                       S, D, Sc, Dz
+	_asum ( n,         x, incX )                                       S, D, Sc, Dz
+	I_amax( n,         x, incX )                                       s, d, c, z
+
+Level 2 BLAS
+
+	        options                   dim   b-width scalar matrix  vector   scalar vector   prefixes
+
+	_gemv (        trans,      m, n,         alpha, a, lda, x, incX, beta,  y, incY ) S, D, C, Z
+	_gbmv (        trans,      m, n, kL, kU, alpha, a, lda, x, incX, beta,  y, incY ) S, D, C, Z
+	_hemv ( uplo,                 n,         alpha, a, lda, x, incX, beta,  y, incY ) C, Z
+	_hbmv ( uplo,                 n, k,      alpha, a, lda, x, incX, beta,  y, incY ) C, Z
+	_hpmv ( uplo,                 n,         alpha, ap,     x, incX, beta,  y, incY ) C, Z
+	_symv ( uplo,                 n,         alpha, a, lda, x, incX, beta,  y, incY ) S, D
+	_sbmv ( uplo,                 n, k,      alpha, a, lda, x, incX, beta,  y, incY ) S, D
+	_spmv ( uplo,                 n,         alpha, ap,     x, incX, beta,  y, incY ) S, D
+	_trmv ( uplo, trans, diag,    n,                a, lda, x, incX )                 S, D, C, Z
+	_tbmv ( uplo, trans, diag,    n, k,             a, lda, x, incX )                 S, D, C, Z
+	_tpmv ( uplo, trans, diag,    n,                ap,     x, incX )                 S, D, C, Z
+	_trsv ( uplo, trans, diag,    n,                a, lda, x, incX )                 S, D, C, Z
+	_tbsv ( uplo, trans, diag,    n, k,             a, lda, x, incX )                 S, D, C, Z
+	_tpsv ( uplo, trans, diag,    n,                ap,     x, incX )                 S, D, C, Z
+
+	        options                   dim   scalar vector   vector   matrix  prefixes
+
+	_ger  (                    m, n, alpha, x, incX, y, incY, a, lda ) S, D
+	_geru (                    m, n, alpha, x, incX, y, incY, a, lda ) C, Z
+	_gerc (                    m, n, alpha, x, incX, y, incY, a, lda ) C, Z
+	_her  ( uplo,                 n, alpha, x, incX,          a, lda ) C, Z
+	_hpr  ( uplo,                 n, alpha, x, incX,          ap )     C, Z
+	_her2 ( uplo,                 n, alpha, x, incX, y, incY, a, lda ) C, Z
+	_hpr2 ( uplo,                 n, alpha, x, incX, y, incY, ap )     C, Z
+	_syr  ( uplo,                 n, alpha, x, incX,          a, lda ) S, D
+	_spr  ( uplo,                 n, alpha, x, incX,          ap )     S, D
+	_syr2 ( uplo,                 n, alpha, x, incX, y, incY, a, lda ) S, D
+	_spr2 ( uplo,                 n, alpha, x, incX, y, incY, ap )     S, D
+
+Level 3 BLAS
+
+	        options                                 dim      scalar matrix  matrix  scalar matrix  prefixes
+
+	_gemm (             transA, transB,      m, n, k, alpha, a, lda, b, ldb, beta,  c, ldc ) S, D, C, Z
+	_symm ( side, uplo,                      m, n,    alpha, a, lda, b, ldb, beta,  c, ldc ) S, D, C, Z
+	_hemm ( side, uplo,                      m, n,    alpha, a, lda, b, ldb, beta,  c, ldc ) C, Z
+	_syrk (       uplo, trans,                  n, k, alpha, a, lda,         beta,  c, ldc ) S, D, C, Z
+	_herk (       uplo, trans,                  n, k, alpha, a, lda,         beta,  c, ldc ) C, Z
+	_syr2k(       uplo, trans,                  n, k, alpha, a, lda, b, ldb, beta,  c, ldc ) S, D, C, Z
+	_her2k(       uplo, trans,                  n, k, alpha, a, lda, b, ldb, beta,  c, ldc ) C, Z
+	_trmm ( side, uplo, transA,        diag, m, n,    alpha, a, lda, b, ldb )                S, D, C, Z
+	_trsm ( side, uplo, transA,        diag, m, n,    alpha, a, lda, b, ldb )                S, D, C, Z
+
+Meaning of prefixes
+
+	S - float32	C - complex64
+	D - float64	Z - complex128
+
+Matrix types
+
+	GE - GEneral 		GB - General Band
+	SY - SYmmetric 		SB - Symmetric Band 	SP - Symmetric Packed
+	HE - HErmitian 		HB - Hermitian Band 	HP - Hermitian Packed
+	TR - TRiangular 	TB - Triangular Band 	TP - Triangular Packed
+
+Options
+
+	trans 	= NoTrans, Trans, ConjTrans
+	uplo 	= Upper, Lower
+	diag 	= Nonunit, Unit
+	side 	= Left, Right (A or op(A) on the left, or A or op(A) on the right)
+
+For real matrices, Trans and ConjTrans have the same meaning.
+For Hermitian matrices, trans = Trans is not allowed.
+For complex symmetric matrices, trans = ConjTrans is not allowed.
+*/
+package blas // import "gonum.org/v1/gonum/blas"

vendor/gonum.org/v1/gonum/blas/gonum/dgemm.go

@@ -0,0 +1,314 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	"runtime"
+	"sync"
+
+	"gonum.org/v1/gonum/blas"
+	"gonum.org/v1/gonum/internal/asm/f64"
+)
+
+// Dgemm performs one of the matrix-matrix operations
+//  C = alpha * A * B + beta * C
+//  C = alpha * A^T * B + beta * C
+//  C = alpha * A * B^T + beta * C
+//  C = alpha * A^T * B^T + beta * C
+// where A is an m×k or k×m dense matrix, B is an n×k or k×n dense matrix, C is
+// an m×n matrix, and alpha and beta are scalars. tA and tB specify whether A or
+// B are transposed.
+func (Implementation) Dgemm(tA, tB blas.Transpose, m, n, k int, alpha float64, a []float64, lda int, b []float64, ldb int, beta float64, c []float64, ldc int) {
+	switch tA {
+	default:
+		panic(badTranspose)
+	case blas.NoTrans, blas.Trans, blas.ConjTrans:
+	}
+	switch tB {
+	default:
+		panic(badTranspose)
+	case blas.NoTrans, blas.Trans, blas.ConjTrans:
+	}
+	if m < 0 {
+		panic(mLT0)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	if k < 0 {
+		panic(kLT0)
+	}
+	aTrans := tA == blas.Trans || tA == blas.ConjTrans
+	if aTrans {
+		if lda < max(1, m) {
+			panic(badLdA)
+		}
+	} else {
+		if lda < max(1, k) {
+			panic(badLdA)
+		}
+	}
+	bTrans := tB == blas.Trans || tB == blas.ConjTrans
+	if bTrans {
+		if ldb < max(1, k) {
+			panic(badLdB)
+		}
+	} else {
+		if ldb < max(1, n) {
+			panic(badLdB)
+		}
+	}
+	if ldc < max(1, n) {
+		panic(badLdC)
+	}
+
+	// Quick return if possible.
+	if m == 0 || n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if aTrans {
+		if len(a) < (k-1)*lda+m {
+			panic(shortA)
+		}
+	} else {
+		if len(a) < (m-1)*lda+k {
+			panic(shortA)
+		}
+	}
+	if bTrans {
+		if len(b) < (n-1)*ldb+k {
+			panic(shortB)
+		}
+	} else {
+		if len(b) < (k-1)*ldb+n {
+			panic(shortB)
+		}
+	}
+	if len(c) < (m-1)*ldc+n {
+		panic(shortC)
+	}
+
+	// Quick return if possible.
+	if (alpha == 0 || k == 0) && beta == 1 {
+		return
+	}
+
+	// scale c
+	if beta != 1 {
+		if beta == 0 {
+			for i := 0; i < m; i++ {
+				ctmp := c[i*ldc : i*ldc+n]
+				for j := range ctmp {
+					ctmp[j] = 0
+				}
+			}
+		} else {
+			for i := 0; i < m; i++ {
+				ctmp := c[i*ldc : i*ldc+n]
+				for j := range ctmp {
+					ctmp[j] *= beta
+				}
+			}
+		}
+	}
+
+	dgemmParallel(aTrans, bTrans, m, n, k, a, lda, b, ldb, c, ldc, alpha)
+}
+
+func dgemmParallel(aTrans, bTrans bool, m, n, k int, a []float64, lda int, b []float64, ldb int, c []float64, ldc int, alpha float64) {
+	// dgemmParallel computes a parallel matrix multiplication by partitioning
+	// a and b into sub-blocks, and updating c with the multiplication of the sub-block
+	// In all cases,
+	// A = [ 	A_11	A_12 ... 	A_1j
+	//			A_21	A_22 ...	A_2j
+	//				...
+	//			A_i1	A_i2 ...	A_ij]
+	//
+	// and same for B. All of the submatrix sizes are blockSize×blockSize except
+	// at the edges.
+	//
+	// In all cases, there is one dimension for each matrix along which
+	// C must be updated sequentially.
+	// Cij = \sum_k Aik Bki,	(A * B)
+	// Cij = \sum_k Aki Bkj,	(A^T * B)
+	// Cij = \sum_k Aik Bjk,	(A * B^T)
+	// Cij = \sum_k Aki Bjk,	(A^T * B^T)
+	//
+	// This code computes one {i, j} block sequentially along the k dimension,
+	// and computes all of the {i, j} blocks concurrently. This
+	// partitioning allows Cij to be updated in-place without race-conditions.
+	// Instead of launching a goroutine for each possible concurrent computation,
+	// a number of worker goroutines are created and channels are used to pass
+	// available and completed cases.
+	//
+	// http://alexkr.com/docs/matrixmult.pdf is a good reference on matrix-matrix
+	// multiplies, though this code does not copy matrices to attempt to eliminate
+	// cache misses.
+
+	maxKLen := k
+	parBlocks := blocks(m, blockSize) * blocks(n, blockSize)
+	if parBlocks < minParBlock {
+		// The matrix multiplication is small in the dimensions where it can be
+		// computed concurrently. Just do it in serial.
+		dgemmSerial(aTrans, bTrans, m, n, k, a, lda, b, ldb, c, ldc, alpha)
+		return
+	}
+
+	nWorkers := runtime.GOMAXPROCS(0)
+	if parBlocks < nWorkers {
+		nWorkers = parBlocks
+	}
+	// There is a tradeoff between the workers having to wait for work
+	// and a large buffer making operations slow.
+	buf := buffMul * nWorkers
+	if buf > parBlocks {
+		buf = parBlocks
+	}
+
+	sendChan := make(chan subMul, buf)
+
+	// Launch workers. A worker receives an {i, j} submatrix of c, and computes
+	// A_ik B_ki (or the transposed version) storing the result in c_ij. When the
+	// channel is finally closed, it signals to the waitgroup that it has finished
+	// computing.
+	var wg sync.WaitGroup
+	for i := 0; i < nWorkers; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for sub := range sendChan {
+				i := sub.i
+				j := sub.j
+				leni := blockSize
+				if i+leni > m {
+					leni = m - i
+				}
+				lenj := blockSize
+				if j+lenj > n {
+					lenj = n - j
+				}
+
+				cSub := sliceView64(c, ldc, i, j, leni, lenj)
+
+				// Compute A_ik B_kj for all k
+				for k := 0; k < maxKLen; k += blockSize {
+					lenk := blockSize
+					if k+lenk > maxKLen {
+						lenk = maxKLen - k
+					}
+					var aSub, bSub []float64
+					if aTrans {
+						aSub = sliceView64(a, lda, k, i, lenk, leni)
+					} else {
+						aSub = sliceView64(a, lda, i, k, leni, lenk)
+					}
+					if bTrans {
+						bSub = sliceView64(b, ldb, j, k, lenj, lenk)
+					} else {
+						bSub = sliceView64(b, ldb, k, j, lenk, lenj)
+					}
+					dgemmSerial(aTrans, bTrans, leni, lenj, lenk, aSub, lda, bSub, ldb, cSub, ldc, alpha)
+				}
+			}
+		}()
+	}
+
+	// Send out all of the {i, j} subblocks for computation.
+	for i := 0; i < m; i += blockSize {
+		for j := 0; j < n; j += blockSize {
+			sendChan <- subMul{
+				i: i,
+				j: j,
+			}
+		}
+	}
+	close(sendChan)
+	wg.Wait()
+}
+
+// dgemmSerial is serial matrix multiply
+func dgemmSerial(aTrans, bTrans bool, m, n, k int, a []float64, lda int, b []float64, ldb int, c []float64, ldc int, alpha float64) {
+	switch {
+	case !aTrans && !bTrans:
+		dgemmSerialNotNot(m, n, k, a, lda, b, ldb, c, ldc, alpha)
+		return
+	case aTrans && !bTrans:
+		dgemmSerialTransNot(m, n, k, a, lda, b, ldb, c, ldc, alpha)
+		return
+	case !aTrans && bTrans:
+		dgemmSerialNotTrans(m, n, k, a, lda, b, ldb, c, ldc, alpha)
+		return
+	case aTrans && bTrans:
+		dgemmSerialTransTrans(m, n, k, a, lda, b, ldb, c, ldc, alpha)
+		return
+	default:
+		panic("unreachable")
+	}
+}
+
+// dgemmSerial where neither a nor b are transposed
+func dgemmSerialNotNot(m, n, k int, a []float64, lda int, b []float64, ldb int, c []float64, ldc int, alpha float64) {
+	// This style is used instead of the literal [i*stride +j]) is used because
+	// approximately 5 times faster as of go 1.3.
+	for i := 0; i < m; i++ {
+		ctmp := c[i*ldc : i*ldc+n]
+		for l, v := range a[i*lda : i*lda+k] {
+			tmp := alpha * v
+			if tmp != 0 {
+				f64.AxpyUnitary(tmp, b[l*ldb:l*ldb+n], ctmp)
+			}
+		}
+	}
+}
+
+// dgemmSerial where neither a is transposed and b is not
+func dgemmSerialTransNot(m, n, k int, a []float64, lda int, b []float64, ldb int, c []float64, ldc int, alpha float64) {
+	// This style is used instead of the literal [i*stride +j]) is used because
+	// approximately 5 times faster as of go 1.3.
+	for l := 0; l < k; l++ {
+		btmp := b[l*ldb : l*ldb+n]
+		for i, v := range a[l*lda : l*lda+m] {
+			tmp := alpha * v
+			if tmp != 0 {
+				ctmp := c[i*ldc : i*ldc+n]
+				f64.AxpyUnitary(tmp, btmp, ctmp)
+			}
+		}
+	}
+}
+
+// dgemmSerial where neither a is not transposed and b is
+func dgemmSerialNotTrans(m, n, k int, a []float64, lda int, b []float64, ldb int, c []float64, ldc int, alpha float64) {
+	// This style is used instead of the literal [i*stride +j]) is used because
+	// approximately 5 times faster as of go 1.3.
+	for i := 0; i < m; i++ {
+		atmp := a[i*lda : i*lda+k]
+		ctmp := c[i*ldc : i*ldc+n]
+		for j := 0; j < n; j++ {
+			ctmp[j] += alpha * f64.DotUnitary(atmp, b[j*ldb:j*ldb+k])
+		}
+	}
+}
+
+// dgemmSerial where both are transposed
+func dgemmSerialTransTrans(m, n, k int, a []float64, lda int, b []float64, ldb int, c []float64, ldc int, alpha float64) {
+	// This style is used instead of the literal [i*stride +j]) is used because
+	// approximately 5 times faster as of go 1.3.
+	for l := 0; l < k; l++ {
+		for i, v := range a[l*lda : l*lda+m] {
+			tmp := alpha * v
+			if tmp != 0 {
+				ctmp := c[i*ldc : i*ldc+n]
+				f64.AxpyInc(tmp, b[l:], ctmp, uintptr(n), uintptr(ldb), 1, 0, 0)
+			}
+		}
+	}
+}
+
+func sliceView64(a []float64, lda, i, j, r, c int) []float64 {
+	return a[i*lda+j : (i+r-1)*lda+j+c]
+}

vendor/gonum.org/v1/gonum/blas/gonum/doc.go

@@ -0,0 +1,88 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Ensure changes made to blas/native are reflected in blas/cgo where relevant.
+
+/*
+Package gonum is a Go implementation of the BLAS API. This implementation
+panics when the input arguments are invalid as per the standard, for example
+if a vector increment is zero. Note that the treatment of NaN values
+is not specified, and differs among the BLAS implementations.
+gonum.org/v1/gonum/blas/blas64 provides helpful wrapper functions to the BLAS
+interface. The rest of this text describes the layout of the data for the input types.
+
+Note that in the function documentation, x[i] refers to the i^th element
+of the vector, which will be different from the i^th element of the slice if
+incX != 1.
+
+See http://www.netlib.org/lapack/explore-html/d4/de1/_l_i_c_e_n_s_e_source.html
+for more license information.
+
+Vector arguments are effectively strided slices. They have two input arguments,
+a number of elements, n, and an increment, incX. The increment specifies the
+distance between elements of the vector. The actual Go slice may be longer
+than necessary.
+The increment may be positive or negative, except in functions with only
+a single vector argument where the increment may only be positive. If the increment
+is negative, s[0] is the last element in the slice. Note that this is not the same
+as counting backward from the end of the slice, as len(s) may be longer than
+necessary. So, for example, if n = 5 and incX = 3, the elements of s are
+	[0 * * 1 * * 2 * * 3 * * 4 * * * ...]
+where ∗ elements are never accessed. If incX = -3, the same elements are
+accessed, just in reverse order (4, 3, 2, 1, 0).
+
+Dense matrices are specified by a number of rows, a number of columns, and a stride.
+The stride specifies the number of entries in the slice between the first element
+of successive rows. The stride must be at least as large as the number of columns
+but may be longer.
+	[a00 ... a0n a0* ... a1stride-1 a21 ... amn am* ... amstride-1]
+Thus, dense[i*ld + j] refers to the {i, j}th element of the matrix.
+
+Symmetric and triangular matrices (non-packed) are stored identically to Dense,
+except that only elements in one triangle of the matrix are accessed.
+
+Packed symmetric and packed triangular matrices are laid out with the entries
+condensed such that all of the unreferenced elements are removed. So, the upper triangular
+matrix
+  [
+    1  2  3
+    0  4  5
+    0  0  6
+  ]
+and the lower-triangular matrix
+  [
+    1  0  0
+    2  3  0
+    4  5  6
+  ]
+will both be compacted as [1 2 3 4 5 6]. The (i, j) element of the original
+dense matrix can be found at element i*n - (i-1)*i/2 + j for upper triangular,
+and at element i * (i+1) /2 + j for lower triangular.
+
+Banded matrices are laid out in a compact format, constructed by removing the
+zeros in the rows and aligning the diagonals. For example, the matrix
+  [
+    1  2  3  0  0  0
+    4  5  6  7  0  0
+    0  8  9 10 11  0
+    0  0 12 13 14 15
+    0  0  0 16 17 18
+    0  0  0  0 19 20
+  ]
+
+implicitly becomes (∗ entries are never accessed)
+  [
+     *  1  2  3
+     4  5  6  7
+     8  9 10 11
+    12 13 14 15
+    16 17 18  *
+    19 20  *  *
+  ]
+which is given to the BLAS routine as [∗ 1 2 3 4 ...].
+
+See http://www.crest.iu.edu/research/mtl/reference/html/banded.html
+for more information
+*/
+package gonum // import "gonum.org/v1/gonum/blas/gonum"

vendor/gonum.org/v1/gonum/blas/gonum/errors.go

@@ -0,0 +1,35 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+// Panic strings used during parameter checks.
+// This list is duplicated in netlib/blas/netlib. Keep in sync.
+const (
+	zeroIncX = "blas: zero x index increment"
+	zeroIncY = "blas: zero y index increment"
+
+	mLT0  = "blas: m < 0"
+	nLT0  = "blas: n < 0"
+	kLT0  = "blas: k < 0"
+	kLLT0 = "blas: kL < 0"
+	kULT0 = "blas: kU < 0"
+
+	badUplo      = "blas: illegal triangle"
+	badTranspose = "blas: illegal transpose"
+	badDiag      = "blas: illegal diagonal"
+	badSide      = "blas: illegal side"
+	badFlag      = "blas: illegal rotm flag"
+
+	badLdA = "blas: bad leading dimension of A"
+	badLdB = "blas: bad leading dimension of B"
+	badLdC = "blas: bad leading dimension of C"
+
+	shortX  = "blas: insufficient length of x"
+	shortY  = "blas: insufficient length of y"
+	shortAP = "blas: insufficient length of ap"
+	shortA  = "blas: insufficient length of a"
+	shortB  = "blas: insufficient length of b"
+	shortC  = "blas: insufficient length of c"
+)

vendor/gonum.org/v1/gonum/blas/gonum/gemv.go

@@ -0,0 +1,190 @@
+// Copyright ©2018 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	"gonum.org/v1/gonum/blas"
+	"gonum.org/v1/gonum/internal/asm/f32"
+	"gonum.org/v1/gonum/internal/asm/f64"
+)
+
+// TODO(Kunde21):  Merge these methods back into level2double/level2single when Sgemv assembly kernels are merged into f32.
+
+// Dgemv computes
+//  y = alpha * A * x + beta * y    if tA = blas.NoTrans
+//  y = alpha * A^T * x + beta * y  if tA = blas.Trans or blas.ConjTrans
+// where A is an m×n dense matrix, x and y are vectors, and alpha and beta are scalars.
+func (Implementation) Dgemv(tA blas.Transpose, m, n int, alpha float64, a []float64, lda int, x []float64, incX int, beta float64, y []float64, incY int) {
+	if tA != blas.NoTrans && tA != blas.Trans && tA != blas.ConjTrans {
+		panic(badTranspose)
+	}
+	if m < 0 {
+		panic(mLT0)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	if lda < max(1, n) {
+		panic(badLdA)
+	}
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	// Set up indexes
+	lenX := m
+	lenY := n
+	if tA == blas.NoTrans {
+		lenX = n
+		lenY = m
+	}
+
+	// Quick return if possible
+	if m == 0 || n == 0 {
+		return
+	}
+
+	if (incX > 0 && (lenX-1)*incX >= len(x)) || (incX < 0 && (1-lenX)*incX >= len(x)) {
+		panic(shortX)
+	}
+	if (incY > 0 && (lenY-1)*incY >= len(y)) || (incY < 0 && (1-lenY)*incY >= len(y)) {
+		panic(shortY)
+	}
+	if len(a) < lda*(m-1)+n {
+		panic(shortA)
+	}
+
+	// Quick return if possible
+	if alpha == 0 && beta == 1 {
+		return
+	}
+
+	if alpha == 0 {
+		// First form y = beta * y
+		if incY > 0 {
+			Implementation{}.Dscal(lenY, beta, y, incY)
+		} else {
+			Implementation{}.Dscal(lenY, beta, y, -incY)
+		}
+		return
+	}
+
+	// Form y = alpha * A * x + y
+	if tA == blas.NoTrans {
+		f64.GemvN(uintptr(m), uintptr(n), alpha, a, uintptr(lda), x, uintptr(incX), beta, y, uintptr(incY))
+		return
+	}
+	// Cases where a is transposed.
+	f64.GemvT(uintptr(m), uintptr(n), alpha, a, uintptr(lda), x, uintptr(incX), beta, y, uintptr(incY))
+}
+
+// Sgemv computes
+//  y = alpha * A * x + beta * y    if tA = blas.NoTrans
+//  y = alpha * A^T * x + beta * y  if tA = blas.Trans or blas.ConjTrans
+// where A is an m×n dense matrix, x and y are vectors, and alpha and beta are scalars.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Sgemv(tA blas.Transpose, m, n int, alpha float32, a []float32, lda int, x []float32, incX int, beta float32, y []float32, incY int) {
+	if tA != blas.NoTrans && tA != blas.Trans && tA != blas.ConjTrans {
+		panic(badTranspose)
+	}
+	if m < 0 {
+		panic(mLT0)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	if lda < max(1, n) {
+		panic(badLdA)
+	}
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+
+	// Quick return if possible.
+	if m == 0 || n == 0 {
+		return
+	}
+
+	// Set up indexes
+	lenX := m
+	lenY := n
+	if tA == blas.NoTrans {
+		lenX = n
+		lenY = m
+	}
+	if (incX > 0 && (lenX-1)*incX >= len(x)) || (incX < 0 && (1-lenX)*incX >= len(x)) {
+		panic(shortX)
+	}
+	if (incY > 0 && (lenY-1)*incY >= len(y)) || (incY < 0 && (1-lenY)*incY >= len(y)) {
+		panic(shortY)
+	}
+	if len(a) < lda*(m-1)+n {
+		panic(shortA)
+	}
+
+	// Quick return if possible.
+	if alpha == 0 && beta == 1 {
+		return
+	}
+
+	// First form y = beta * y
+	if incY > 0 {
+		Implementation{}.Sscal(lenY, beta, y, incY)
+	} else {
+		Implementation{}.Sscal(lenY, beta, y, -incY)
+	}
+
+	if alpha == 0 {
+		return
+	}
+
+	var kx, ky int
+	if incX < 0 {
+		kx = -(lenX - 1) * incX
+	}
+	if incY < 0 {
+		ky = -(lenY - 1) * incY
+	}
+
+	// Form y = alpha * A * x + y
+	if tA == blas.NoTrans {
+		if incX == 1 && incY == 1 {
+			for i := 0; i < m; i++ {
+				y[i] += alpha * f32.DotUnitary(a[lda*i:lda*i+n], x[:n])
+			}
+			return
+		}
+		iy := ky
+		for i := 0; i < m; i++ {
+			y[iy] += alpha * f32.DotInc(x, a[lda*i:lda*i+n], uintptr(n), uintptr(incX), 1, uintptr(kx), 0)
+			iy += incY
+		}
+		return
+	}
+	// Cases where a is transposed.
+	if incX == 1 && incY == 1 {
+		for i := 0; i < m; i++ {
+			tmp := alpha * x[i]
+			if tmp != 0 {
+				f32.AxpyUnitaryTo(y, tmp, a[lda*i:lda*i+n], y[:n])
+			}
+		}
+		return
+	}
+	ix := kx
+	for i := 0; i < m; i++ {
+		tmp := alpha * x[ix]
+		if tmp != 0 {
+			f32.AxpyInc(tmp, a[lda*i:lda*i+n], y, uintptr(n), 1, uintptr(incY), 0, uintptr(ky))
+		}
+		ix += incX
+	}
+}

vendor/gonum.org/v1/gonum/blas/gonum/gonum.go

@@ -0,0 +1,58 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//go:generate ./single_precision.bash
+
+package gonum
+
+import (
+	"math"
+
+	"gonum.org/v1/gonum/internal/math32"
+)
+
+type Implementation struct{}
+
+// [SD]gemm behavior constants. These are kept here to keep them out of the
+// way during single precision code genration.
+const (
+	blockSize   = 64 // b x b matrix
+	minParBlock = 4  // minimum number of blocks needed to go parallel
+	buffMul     = 4  // how big is the buffer relative to the number of workers
+)
+
+// subMul is a common type shared by [SD]gemm.
+type subMul struct {
+	i, j int // index of block
+}
+
+func max(a, b int) int {
+	if a > b {
+		return a
+	}
+	return b
+}
+
+func min(a, b int) int {
+	if a > b {
+		return b
+	}
+	return a
+}
+
+// blocks returns the number of divisions of the dimension length with the given
+// block size.
+func blocks(dim, bsize int) int {
+	return (dim + bsize - 1) / bsize
+}
+
+// dcabs1 returns |real(z)|+|imag(z)|.
+func dcabs1(z complex128) float64 {
+	return math.Abs(real(z)) + math.Abs(imag(z))
+}
+
+// scabs1 returns |real(z)|+|imag(z)|.
+func scabs1(z complex64) float32 {
+	return math32.Abs(real(z)) + math32.Abs(imag(z))
+}

vendor/gonum.org/v1/gonum/blas/gonum/level1cmplx128.go

@@ -0,0 +1,445 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	"math"
+
+	"gonum.org/v1/gonum/blas"
+	"gonum.org/v1/gonum/internal/asm/c128"
+)
+
+var _ blas.Complex128Level1 = Implementation{}
+
+// Dzasum returns the sum of the absolute values of the elements of x
+//  \sum_i |Re(x[i])| + |Im(x[i])|
+// Dzasum returns 0 if incX is negative.
+func (Implementation) Dzasum(n int, x []complex128, incX int) float64 {
+	if n < 0 {
+		panic(nLT0)
+	}
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return 0
+	}
+	var sum float64
+	if incX == 1 {
+		if len(x) < n {
+			panic(shortX)
+		}
+		for _, v := range x[:n] {
+			sum += dcabs1(v)
+		}
+		return sum
+	}
+	if (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	for i := 0; i < n; i++ {
+		v := x[i*incX]
+		sum += dcabs1(v)
+	}
+	return sum
+}
+
+// Dznrm2 computes the Euclidean norm of the complex vector x,
+//  ‖x‖_2 = sqrt(\sum_i x[i] * conj(x[i])).
+// This function returns 0 if incX is negative.
+func (Implementation) Dznrm2(n int, x []complex128, incX int) float64 {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return 0
+	}
+	if n < 1 {
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	if (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	var (
+		scale float64
+		ssq   float64 = 1
+	)
+	if incX == 1 {
+		for _, v := range x[:n] {
+			re, im := math.Abs(real(v)), math.Abs(imag(v))
+			if re != 0 {
+				if re > scale {
+					ssq = 1 + ssq*(scale/re)*(scale/re)
+					scale = re
+				} else {
+					ssq += (re / scale) * (re / scale)
+				}
+			}
+			if im != 0 {
+				if im > scale {
+					ssq = 1 + ssq*(scale/im)*(scale/im)
+					scale = im
+				} else {
+					ssq += (im / scale) * (im / scale)
+				}
+			}
+		}
+		if math.IsInf(scale, 1) {
+			return math.Inf(1)
+		}
+		return scale * math.Sqrt(ssq)
+	}
+	for ix := 0; ix < n*incX; ix += incX {
+		re, im := math.Abs(real(x[ix])), math.Abs(imag(x[ix]))
+		if re != 0 {
+			if re > scale {
+				ssq = 1 + ssq*(scale/re)*(scale/re)
+				scale = re
+			} else {
+				ssq += (re / scale) * (re / scale)
+			}
+		}
+		if im != 0 {
+			if im > scale {
+				ssq = 1 + ssq*(scale/im)*(scale/im)
+				scale = im
+			} else {
+				ssq += (im / scale) * (im / scale)
+			}
+		}
+	}
+	if math.IsInf(scale, 1) {
+		return math.Inf(1)
+	}
+	return scale * math.Sqrt(ssq)
+}
+
+// Izamax returns the index of the first element of x having largest |Re(·)|+|Im(·)|.
+// Izamax returns -1 if n is 0 or incX is negative.
+func (Implementation) Izamax(n int, x []complex128, incX int) int {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		// Return invalid index.
+		return -1
+	}
+	if n < 1 {
+		if n == 0 {
+			// Return invalid index.
+			return -1
+		}
+		panic(nLT0)
+	}
+	if len(x) <= (n-1)*incX {
+		panic(shortX)
+	}
+	idx := 0
+	max := dcabs1(x[0])
+	if incX == 1 {
+		for i, v := range x[1:n] {
+			absV := dcabs1(v)
+			if absV > max {
+				max = absV
+				idx = i + 1
+			}
+		}
+		return idx
+	}
+	ix := incX
+	for i := 1; i < n; i++ {
+		absV := dcabs1(x[ix])
+		if absV > max {
+			max = absV
+			idx = i
+		}
+		ix += incX
+	}
+	return idx
+}
+
+// Zaxpy adds alpha times x to y:
+//  y[i] += alpha * x[i] for all i
+func (Implementation) Zaxpy(n int, alpha complex128, x []complex128, incX int, y []complex128, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && (n-1)*incX >= len(x)) || (incX < 0 && (1-n)*incX >= len(x)) {
+		panic(shortX)
+	}
+	if (incY > 0 && (n-1)*incY >= len(y)) || (incY < 0 && (1-n)*incY >= len(y)) {
+		panic(shortY)
+	}
+	if alpha == 0 {
+		return
+	}
+	if incX == 1 && incY == 1 {
+		c128.AxpyUnitary(alpha, x[:n], y[:n])
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (1 - n) * incX
+	}
+	if incY < 0 {
+		iy = (1 - n) * incY
+	}
+	c128.AxpyInc(alpha, x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy))
+}
+
+// Zcopy copies the vector x to vector y.
+func (Implementation) Zcopy(n int, x []complex128, incX int, y []complex128, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && (n-1)*incX >= len(x)) || (incX < 0 && (1-n)*incX >= len(x)) {
+		panic(shortX)
+	}
+	if (incY > 0 && (n-1)*incY >= len(y)) || (incY < 0 && (1-n)*incY >= len(y)) {
+		panic(shortY)
+	}
+	if incX == 1 && incY == 1 {
+		copy(y[:n], x[:n])
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	for i := 0; i < n; i++ {
+		y[iy] = x[ix]
+		ix += incX
+		iy += incY
+	}
+}
+
+// Zdotc computes the dot product
+//  x^H · y
+// of two complex vectors x and y.
+func (Implementation) Zdotc(n int, x []complex128, incX int, y []complex128, incY int) complex128 {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n <= 0 {
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	if incX == 1 && incY == 1 {
+		if len(x) < n {
+			panic(shortX)
+		}
+		if len(y) < n {
+			panic(shortY)
+		}
+		return c128.DotcUnitary(x[:n], y[:n])
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	if ix >= len(x) || (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if iy >= len(y) || (n-1)*incY >= len(y) {
+		panic(shortY)
+	}
+	return c128.DotcInc(x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy))
+}
+
+// Zdotu computes the dot product
+//  x^T · y
+// of two complex vectors x and y.
+func (Implementation) Zdotu(n int, x []complex128, incX int, y []complex128, incY int) complex128 {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n <= 0 {
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	if incX == 1 && incY == 1 {
+		if len(x) < n {
+			panic(shortX)
+		}
+		if len(y) < n {
+			panic(shortY)
+		}
+		return c128.DotuUnitary(x[:n], y[:n])
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	if ix >= len(x) || (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if iy >= len(y) || (n-1)*incY >= len(y) {
+		panic(shortY)
+	}
+	return c128.DotuInc(x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy))
+}
+
+// Zdscal scales the vector x by a real scalar alpha.
+// Zdscal has no effect if incX < 0.
+func (Implementation) Zdscal(n int, alpha float64, x []complex128, incX int) {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return
+	}
+	if (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if alpha == 0 {
+		if incX == 1 {
+			x = x[:n]
+			for i := range x {
+				x[i] = 0
+			}
+			return
+		}
+		for ix := 0; ix < n*incX; ix += incX {
+			x[ix] = 0
+		}
+		return
+	}
+	if incX == 1 {
+		x = x[:n]
+		for i, v := range x {
+			x[i] = complex(alpha*real(v), alpha*imag(v))
+		}
+		return
+	}
+	for ix := 0; ix < n*incX; ix += incX {
+		v := x[ix]
+		x[ix] = complex(alpha*real(v), alpha*imag(v))
+	}
+}
+
+// Zscal scales the vector x by a complex scalar alpha.
+// Zscal has no effect if incX < 0.
+func (Implementation) Zscal(n int, alpha complex128, x []complex128, incX int) {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return
+	}
+	if (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if alpha == 0 {
+		if incX == 1 {
+			x = x[:n]
+			for i := range x {
+				x[i] = 0
+			}
+			return
+		}
+		for ix := 0; ix < n*incX; ix += incX {
+			x[ix] = 0
+		}
+		return
+	}
+	if incX == 1 {
+		c128.ScalUnitary(alpha, x[:n])
+		return
+	}
+	c128.ScalInc(alpha, x, uintptr(n), uintptr(incX))
+}
+
+// Zswap exchanges the elements of two complex vectors x and y.
+func (Implementation) Zswap(n int, x []complex128, incX int, y []complex128, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && (n-1)*incX >= len(x)) || (incX < 0 && (1-n)*incX >= len(x)) {
+		panic(shortX)
+	}
+	if (incY > 0 && (n-1)*incY >= len(y)) || (incY < 0 && (1-n)*incY >= len(y)) {
+		panic(shortY)
+	}
+	if incX == 1 && incY == 1 {
+		x = x[:n]
+		for i, v := range x {
+			x[i], y[i] = y[i], v
+		}
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	for i := 0; i < n; i++ {
+		x[ix], y[iy] = y[iy], x[ix]
+		ix += incX
+		iy += incY
+	}
+}

vendor/gonum.org/v1/gonum/blas/gonum/level1cmplx64.go

@@ -0,0 +1,467 @@
+// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.
+
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	math "gonum.org/v1/gonum/internal/math32"
+
+	"gonum.org/v1/gonum/blas"
+	"gonum.org/v1/gonum/internal/asm/c64"
+)
+
+var _ blas.Complex64Level1 = Implementation{}
+
+// Scasum returns the sum of the absolute values of the elements of x
+//  \sum_i |Re(x[i])| + |Im(x[i])|
+// Scasum returns 0 if incX is negative.
+//
+// Complex64 implementations are autogenerated and not directly tested.
+func (Implementation) Scasum(n int, x []complex64, incX int) float32 {
+	if n < 0 {
+		panic(nLT0)
+	}
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return 0
+	}
+	var sum float32
+	if incX == 1 {
+		if len(x) < n {
+			panic(shortX)
+		}
+		for _, v := range x[:n] {
+			sum += scabs1(v)
+		}
+		return sum
+	}
+	if (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	for i := 0; i < n; i++ {
+		v := x[i*incX]
+		sum += scabs1(v)
+	}
+	return sum
+}
+
+// Scnrm2 computes the Euclidean norm of the complex vector x,
+//  ‖x‖_2 = sqrt(\sum_i x[i] * conj(x[i])).
+// This function returns 0 if incX is negative.
+//
+// Complex64 implementations are autogenerated and not directly tested.
+func (Implementation) Scnrm2(n int, x []complex64, incX int) float32 {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return 0
+	}
+	if n < 1 {
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	if (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	var (
+		scale float32
+		ssq   float32 = 1
+	)
+	if incX == 1 {
+		for _, v := range x[:n] {
+			re, im := math.Abs(real(v)), math.Abs(imag(v))
+			if re != 0 {
+				if re > scale {
+					ssq = 1 + ssq*(scale/re)*(scale/re)
+					scale = re
+				} else {
+					ssq += (re / scale) * (re / scale)
+				}
+			}
+			if im != 0 {
+				if im > scale {
+					ssq = 1 + ssq*(scale/im)*(scale/im)
+					scale = im
+				} else {
+					ssq += (im / scale) * (im / scale)
+				}
+			}
+		}
+		if math.IsInf(scale, 1) {
+			return math.Inf(1)
+		}
+		return scale * math.Sqrt(ssq)
+	}
+	for ix := 0; ix < n*incX; ix += incX {
+		re, im := math.Abs(real(x[ix])), math.Abs(imag(x[ix]))
+		if re != 0 {
+			if re > scale {
+				ssq = 1 + ssq*(scale/re)*(scale/re)
+				scale = re
+			} else {
+				ssq += (re / scale) * (re / scale)
+			}
+		}
+		if im != 0 {
+			if im > scale {
+				ssq = 1 + ssq*(scale/im)*(scale/im)
+				scale = im
+			} else {
+				ssq += (im / scale) * (im / scale)
+			}
+		}
+	}
+	if math.IsInf(scale, 1) {
+		return math.Inf(1)
+	}
+	return scale * math.Sqrt(ssq)
+}
+
+// Icamax returns the index of the first element of x having largest |Re(·)|+|Im(·)|.
+// Icamax returns -1 if n is 0 or incX is negative.
+//
+// Complex64 implementations are autogenerated and not directly tested.
+func (Implementation) Icamax(n int, x []complex64, incX int) int {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		// Return invalid index.
+		return -1
+	}
+	if n < 1 {
+		if n == 0 {
+			// Return invalid index.
+			return -1
+		}
+		panic(nLT0)
+	}
+	if len(x) <= (n-1)*incX {
+		panic(shortX)
+	}
+	idx := 0
+	max := scabs1(x[0])
+	if incX == 1 {
+		for i, v := range x[1:n] {
+			absV := scabs1(v)
+			if absV > max {
+				max = absV
+				idx = i + 1
+			}
+		}
+		return idx
+	}
+	ix := incX
+	for i := 1; i < n; i++ {
+		absV := scabs1(x[ix])
+		if absV > max {
+			max = absV
+			idx = i
+		}
+		ix += incX
+	}
+	return idx
+}
+
+// Caxpy adds alpha times x to y:
+//  y[i] += alpha * x[i] for all i
+//
+// Complex64 implementations are autogenerated and not directly tested.
+func (Implementation) Caxpy(n int, alpha complex64, x []complex64, incX int, y []complex64, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && (n-1)*incX >= len(x)) || (incX < 0 && (1-n)*incX >= len(x)) {
+		panic(shortX)
+	}
+	if (incY > 0 && (n-1)*incY >= len(y)) || (incY < 0 && (1-n)*incY >= len(y)) {
+		panic(shortY)
+	}
+	if alpha == 0 {
+		return
+	}
+	if incX == 1 && incY == 1 {
+		c64.AxpyUnitary(alpha, x[:n], y[:n])
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (1 - n) * incX
+	}
+	if incY < 0 {
+		iy = (1 - n) * incY
+	}
+	c64.AxpyInc(alpha, x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy))
+}
+
+// Ccopy copies the vector x to vector y.
+//
+// Complex64 implementations are autogenerated and not directly tested.
+func (Implementation) Ccopy(n int, x []complex64, incX int, y []complex64, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && (n-1)*incX >= len(x)) || (incX < 0 && (1-n)*incX >= len(x)) {
+		panic(shortX)
+	}
+	if (incY > 0 && (n-1)*incY >= len(y)) || (incY < 0 && (1-n)*incY >= len(y)) {
+		panic(shortY)
+	}
+	if incX == 1 && incY == 1 {
+		copy(y[:n], x[:n])
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	for i := 0; i < n; i++ {
+		y[iy] = x[ix]
+		ix += incX
+		iy += incY
+	}
+}
+
+// Cdotc computes the dot product
+//  x^H · y
+// of two complex vectors x and y.
+//
+// Complex64 implementations are autogenerated and not directly tested.
+func (Implementation) Cdotc(n int, x []complex64, incX int, y []complex64, incY int) complex64 {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n <= 0 {
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	if incX == 1 && incY == 1 {
+		if len(x) < n {
+			panic(shortX)
+		}
+		if len(y) < n {
+			panic(shortY)
+		}
+		return c64.DotcUnitary(x[:n], y[:n])
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	if ix >= len(x) || (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if iy >= len(y) || (n-1)*incY >= len(y) {
+		panic(shortY)
+	}
+	return c64.DotcInc(x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy))
+}
+
+// Cdotu computes the dot product
+//  x^T · y
+// of two complex vectors x and y.
+//
+// Complex64 implementations are autogenerated and not directly tested.
+func (Implementation) Cdotu(n int, x []complex64, incX int, y []complex64, incY int) complex64 {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n <= 0 {
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	if incX == 1 && incY == 1 {
+		if len(x) < n {
+			panic(shortX)
+		}
+		if len(y) < n {
+			panic(shortY)
+		}
+		return c64.DotuUnitary(x[:n], y[:n])
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	if ix >= len(x) || (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if iy >= len(y) || (n-1)*incY >= len(y) {
+		panic(shortY)
+	}
+	return c64.DotuInc(x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy))
+}
+
+// Csscal scales the vector x by a real scalar alpha.
+// Csscal has no effect if incX < 0.
+//
+// Complex64 implementations are autogenerated and not directly tested.
+func (Implementation) Csscal(n int, alpha float32, x []complex64, incX int) {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return
+	}
+	if (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if alpha == 0 {
+		if incX == 1 {
+			x = x[:n]
+			for i := range x {
+				x[i] = 0
+			}
+			return
+		}
+		for ix := 0; ix < n*incX; ix += incX {
+			x[ix] = 0
+		}
+		return
+	}
+	if incX == 1 {
+		x = x[:n]
+		for i, v := range x {
+			x[i] = complex(alpha*real(v), alpha*imag(v))
+		}
+		return
+	}
+	for ix := 0; ix < n*incX; ix += incX {
+		v := x[ix]
+		x[ix] = complex(alpha*real(v), alpha*imag(v))
+	}
+}
+
+// Cscal scales the vector x by a complex scalar alpha.
+// Cscal has no effect if incX < 0.
+//
+// Complex64 implementations are autogenerated and not directly tested.
+func (Implementation) Cscal(n int, alpha complex64, x []complex64, incX int) {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return
+	}
+	if (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if alpha == 0 {
+		if incX == 1 {
+			x = x[:n]
+			for i := range x {
+				x[i] = 0
+			}
+			return
+		}
+		for ix := 0; ix < n*incX; ix += incX {
+			x[ix] = 0
+		}
+		return
+	}
+	if incX == 1 {
+		c64.ScalUnitary(alpha, x[:n])
+		return
+	}
+	c64.ScalInc(alpha, x, uintptr(n), uintptr(incX))
+}
+
+// Cswap exchanges the elements of two complex vectors x and y.
+//
+// Complex64 implementations are autogenerated and not directly tested.
+func (Implementation) Cswap(n int, x []complex64, incX int, y []complex64, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && (n-1)*incX >= len(x)) || (incX < 0 && (1-n)*incX >= len(x)) {
+		panic(shortX)
+	}
+	if (incY > 0 && (n-1)*incY >= len(y)) || (incY < 0 && (1-n)*incY >= len(y)) {
+		panic(shortY)
+	}
+	if incX == 1 && incY == 1 {
+		x = x[:n]
+		for i, v := range x {
+			x[i], y[i] = y[i], v
+		}
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	for i := 0; i < n; i++ {
+		x[ix], y[iy] = y[iy], x[ix]
+		ix += incX
+		iy += incY
+	}
+}

vendor/gonum.org/v1/gonum/blas/gonum/level1float32.go

@@ -0,0 +1,644 @@
+// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.
+
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	math "gonum.org/v1/gonum/internal/math32"
+
+	"gonum.org/v1/gonum/blas"
+	"gonum.org/v1/gonum/internal/asm/f32"
+)
+
+var _ blas.Float32Level1 = Implementation{}
+
+// Snrm2 computes the Euclidean norm of a vector,
+//  sqrt(\sum_i x[i] * x[i]).
+// This function returns 0 if incX is negative.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Snrm2(n int, x []float32, incX int) float32 {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return 0
+	}
+	if len(x) <= (n-1)*incX {
+		panic(shortX)
+	}
+	if n < 2 {
+		if n == 1 {
+			return math.Abs(x[0])
+		}
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	var (
+		scale      float32 = 0
+		sumSquares float32 = 1
+	)
+	if incX == 1 {
+		x = x[:n]
+		for _, v := range x {
+			if v == 0 {
+				continue
+			}
+			absxi := math.Abs(v)
+			if math.IsNaN(absxi) {
+				return math.NaN()
+			}
+			if scale < absxi {
+				sumSquares = 1 + sumSquares*(scale/absxi)*(scale/absxi)
+				scale = absxi
+			} else {
+				sumSquares = sumSquares + (absxi/scale)*(absxi/scale)
+			}
+		}
+		if math.IsInf(scale, 1) {
+			return math.Inf(1)
+		}
+		return scale * math.Sqrt(sumSquares)
+	}
+	for ix := 0; ix < n*incX; ix += incX {
+		val := x[ix]
+		if val == 0 {
+			continue
+		}
+		absxi := math.Abs(val)
+		if math.IsNaN(absxi) {
+			return math.NaN()
+		}
+		if scale < absxi {
+			sumSquares = 1 + sumSquares*(scale/absxi)*(scale/absxi)
+			scale = absxi
+		} else {
+			sumSquares = sumSquares + (absxi/scale)*(absxi/scale)
+		}
+	}
+	if math.IsInf(scale, 1) {
+		return math.Inf(1)
+	}
+	return scale * math.Sqrt(sumSquares)
+}
+
+// Sasum computes the sum of the absolute values of the elements of x.
+//  \sum_i |x[i]|
+// Sasum returns 0 if incX is negative.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Sasum(n int, x []float32, incX int) float32 {
+	var sum float32
+	if n < 0 {
+		panic(nLT0)
+	}
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return 0
+	}
+	if len(x) <= (n-1)*incX {
+		panic(shortX)
+	}
+	if incX == 1 {
+		x = x[:n]
+		for _, v := range x {
+			sum += math.Abs(v)
+		}
+		return sum
+	}
+	for i := 0; i < n; i++ {
+		sum += math.Abs(x[i*incX])
+	}
+	return sum
+}
+
+// Isamax returns the index of an element of x with the largest absolute value.
+// If there are multiple such indices the earliest is returned.
+// Isamax returns -1 if n == 0.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Isamax(n int, x []float32, incX int) int {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return -1
+	}
+	if len(x) <= (n-1)*incX {
+		panic(shortX)
+	}
+	if n < 2 {
+		if n == 1 {
+			return 0
+		}
+		if n == 0 {
+			return -1 // Netlib returns invalid index when n == 0.
+		}
+		panic(nLT0)
+	}
+	idx := 0
+	max := math.Abs(x[0])
+	if incX == 1 {
+		for i, v := range x[:n] {
+			absV := math.Abs(v)
+			if absV > max {
+				max = absV
+				idx = i
+			}
+		}
+		return idx
+	}
+	ix := incX
+	for i := 1; i < n; i++ {
+		v := x[ix]
+		absV := math.Abs(v)
+		if absV > max {
+			max = absV
+			idx = i
+		}
+		ix += incX
+	}
+	return idx
+}
+
+// Sswap exchanges the elements of two vectors.
+//  x[i], y[i] = y[i], x[i] for all i
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Sswap(n int, x []float32, incX int, y []float32, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && len(x) <= (n-1)*incX) || (incX < 0 && len(x) <= (1-n)*incX) {
+		panic(shortX)
+	}
+	if (incY > 0 && len(y) <= (n-1)*incY) || (incY < 0 && len(y) <= (1-n)*incY) {
+		panic(shortY)
+	}
+	if incX == 1 && incY == 1 {
+		x = x[:n]
+		for i, v := range x {
+			x[i], y[i] = y[i], v
+		}
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	for i := 0; i < n; i++ {
+		x[ix], y[iy] = y[iy], x[ix]
+		ix += incX
+		iy += incY
+	}
+}
+
+// Scopy copies the elements of x into the elements of y.
+//  y[i] = x[i] for all i
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Scopy(n int, x []float32, incX int, y []float32, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && len(x) <= (n-1)*incX) || (incX < 0 && len(x) <= (1-n)*incX) {
+		panic(shortX)
+	}
+	if (incY > 0 && len(y) <= (n-1)*incY) || (incY < 0 && len(y) <= (1-n)*incY) {
+		panic(shortY)
+	}
+	if incX == 1 && incY == 1 {
+		copy(y[:n], x[:n])
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	for i := 0; i < n; i++ {
+		y[iy] = x[ix]
+		ix += incX
+		iy += incY
+	}
+}
+
+// Saxpy adds alpha times x to y
+//  y[i] += alpha * x[i] for all i
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Saxpy(n int, alpha float32, x []float32, incX int, y []float32, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && len(x) <= (n-1)*incX) || (incX < 0 && len(x) <= (1-n)*incX) {
+		panic(shortX)
+	}
+	if (incY > 0 && len(y) <= (n-1)*incY) || (incY < 0 && len(y) <= (1-n)*incY) {
+		panic(shortY)
+	}
+	if alpha == 0 {
+		return
+	}
+	if incX == 1 && incY == 1 {
+		f32.AxpyUnitary(alpha, x[:n], y[:n])
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	f32.AxpyInc(alpha, x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy))
+}
+
+// Srotg computes the plane rotation
+//   _    _      _ _       _ _
+//  |  c s |    | a |     | r |
+//  | -s c |  * | b |   = | 0 |
+//   ‾    ‾      ‾ ‾       ‾ ‾
+// where
+//  r = ±√(a^2 + b^2)
+//  c = a/r, the cosine of the plane rotation
+//  s = b/r, the sine of the plane rotation
+//
+// NOTE: There is a discrepancy between the reference implementation and the BLAS
+// technical manual regarding the sign for r when a or b are zero.
+// Srotg agrees with the definition in the manual and other
+// common BLAS implementations.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Srotg(a, b float32) (c, s, r, z float32) {
+	if b == 0 && a == 0 {
+		return 1, 0, a, 0
+	}
+	absA := math.Abs(a)
+	absB := math.Abs(b)
+	aGTb := absA > absB
+	r = math.Hypot(a, b)
+	if aGTb {
+		r = math.Copysign(r, a)
+	} else {
+		r = math.Copysign(r, b)
+	}
+	c = a / r
+	s = b / r
+	if aGTb {
+		z = s
+	} else if c != 0 { // r == 0 case handled above
+		z = 1 / c
+	} else {
+		z = 1
+	}
+	return
+}
+
+// Srotmg computes the modified Givens rotation. See
+// http://www.netlib.org/lapack/explore-html/df/deb/drotmg_8f.html
+// for more details.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Srotmg(d1, d2, x1, y1 float32) (p blas.SrotmParams, rd1, rd2, rx1 float32) {
+	// The implementation of Drotmg used here is taken from Hopkins 1997
+	// Appendix A: https://doi.org/10.1145/289251.289253
+	// with the exception of the gam constants below.
+
+	const (
+		gam    = 4096.0
+		gamsq  = gam * gam
+		rgamsq = 1.0 / gamsq
+	)
+
+	if d1 < 0 {
+		p.Flag = blas.Rescaling // Error state.
+		return p, 0, 0, 0
+	}
+
+	if d2 == 0 || y1 == 0 {
+		p.Flag = blas.Identity
+		return p, d1, d2, x1
+	}
+
+	var h11, h12, h21, h22 float32
+	if (d1 == 0 || x1 == 0) && d2 > 0 {
+		p.Flag = blas.Diagonal
+		h12 = 1
+		h21 = -1
+		x1 = y1
+		d1, d2 = d2, d1
+	} else {
+		p2 := d2 * y1
+		p1 := d1 * x1
+		q2 := p2 * y1
+		q1 := p1 * x1
+		if math.Abs(q1) > math.Abs(q2) {
+			p.Flag = blas.OffDiagonal
+			h11 = 1
+			h22 = 1
+			h21 = -y1 / x1
+			h12 = p2 / p1
+			u := 1 - h12*h21
+			if u <= 0 {
+				p.Flag = blas.Rescaling // Error state.
+				return p, 0, 0, 0
+			}
+
+			d1 /= u
+			d2 /= u
+			x1 *= u
+		} else {
+			if q2 < 0 {
+				p.Flag = blas.Rescaling // Error state.
+				return p, 0, 0, 0
+			}
+
+			p.Flag = blas.Diagonal
+			h21 = -1
+			h12 = 1
+			h11 = p1 / p2
+			h22 = x1 / y1
+			u := 1 + h11*h22
+			d1, d2 = d2/u, d1/u
+			x1 = y1 * u
+		}
+	}
+
+	for d1 <= rgamsq && d1 != 0 {
+		p.Flag = blas.Rescaling
+		d1 = (d1 * gam) * gam
+		x1 /= gam
+		h11 /= gam
+		h12 /= gam
+	}
+	for d1 > gamsq {
+		p.Flag = blas.Rescaling
+		d1 = (d1 / gam) / gam
+		x1 *= gam
+		h11 *= gam
+		h12 *= gam
+	}
+
+	for math.Abs(d2) <= rgamsq && d2 != 0 {
+		p.Flag = blas.Rescaling
+		d2 = (d2 * gam) * gam
+		h21 /= gam
+		h22 /= gam
+	}
+	for math.Abs(d2) > gamsq {
+		p.Flag = blas.Rescaling
+		d2 = (d2 / gam) / gam
+		h21 *= gam
+		h22 *= gam
+	}
+
+	switch p.Flag {
+	case blas.Diagonal:
+		p.H = [4]float32{0: h11, 3: h22}
+	case blas.OffDiagonal:
+		p.H = [4]float32{1: h21, 2: h12}
+	case blas.Rescaling:
+		p.H = [4]float32{h11, h21, h12, h22}
+	default:
+		panic(badFlag)
+	}
+
+	return p, d1, d2, x1
+}
+
+// Srot applies a plane transformation.
+//  x[i] = c * x[i] + s * y[i]
+//  y[i] = c * y[i] - s * x[i]
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Srot(n int, x []float32, incX int, y []float32, incY int, c float32, s float32) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && len(x) <= (n-1)*incX) || (incX < 0 && len(x) <= (1-n)*incX) {
+		panic(shortX)
+	}
+	if (incY > 0 && len(y) <= (n-1)*incY) || (incY < 0 && len(y) <= (1-n)*incY) {
+		panic(shortY)
+	}
+	if incX == 1 && incY == 1 {
+		x = x[:n]
+		for i, vx := range x {
+			vy := y[i]
+			x[i], y[i] = c*vx+s*vy, c*vy-s*vx
+		}
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	for i := 0; i < n; i++ {
+		vx := x[ix]
+		vy := y[iy]
+		x[ix], y[iy] = c*vx+s*vy, c*vy-s*vx
+		ix += incX
+		iy += incY
+	}
+}
+
+// Srotm applies the modified Givens rotation to the 2×n matrix.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Srotm(n int, x []float32, incX int, y []float32, incY int, p blas.SrotmParams) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n <= 0 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && len(x) <= (n-1)*incX) || (incX < 0 && len(x) <= (1-n)*incX) {
+		panic(shortX)
+	}
+	if (incY > 0 && len(y) <= (n-1)*incY) || (incY < 0 && len(y) <= (1-n)*incY) {
+		panic(shortY)
+	}
+
+	if p.Flag == blas.Identity {
+		return
+	}
+
+	switch p.Flag {
+	case blas.Rescaling:
+		h11 := p.H[0]
+		h12 := p.H[2]
+		h21 := p.H[1]
+		h22 := p.H[3]
+		if incX == 1 && incY == 1 {
+			x = x[:n]
+			for i, vx := range x {
+				vy := y[i]
+				x[i], y[i] = vx*h11+vy*h12, vx*h21+vy*h22
+			}
+			return
+		}
+		var ix, iy int
+		if incX < 0 {
+			ix = (-n + 1) * incX
+		}
+		if incY < 0 {
+			iy = (-n + 1) * incY
+		}
+		for i := 0; i < n; i++ {
+			vx := x[ix]
+			vy := y[iy]
+			x[ix], y[iy] = vx*h11+vy*h12, vx*h21+vy*h22
+			ix += incX
+			iy += incY
+		}
+	case blas.OffDiagonal:
+		h12 := p.H[2]
+		h21 := p.H[1]
+		if incX == 1 && incY == 1 {
+			x = x[:n]
+			for i, vx := range x {
+				vy := y[i]
+				x[i], y[i] = vx+vy*h12, vx*h21+vy
+			}
+			return
+		}
+		var ix, iy int
+		if incX < 0 {
+			ix = (-n + 1) * incX
+		}
+		if incY < 0 {
+			iy = (-n + 1) * incY
+		}
+		for i := 0; i < n; i++ {
+			vx := x[ix]
+			vy := y[iy]
+			x[ix], y[iy] = vx+vy*h12, vx*h21+vy
+			ix += incX
+			iy += incY
+		}
+	case blas.Diagonal:
+		h11 := p.H[0]
+		h22 := p.H[3]
+		if incX == 1 && incY == 1 {
+			x = x[:n]
+			for i, vx := range x {
+				vy := y[i]
+				x[i], y[i] = vx*h11+vy, -vx+vy*h22
+			}
+			return
+		}
+		var ix, iy int
+		if incX < 0 {
+			ix = (-n + 1) * incX
+		}
+		if incY < 0 {
+			iy = (-n + 1) * incY
+		}
+		for i := 0; i < n; i++ {
+			vx := x[ix]
+			vy := y[iy]
+			x[ix], y[iy] = vx*h11+vy, -vx+vy*h22
+			ix += incX
+			iy += incY
+		}
+	}
+}
+
+// Sscal scales x by alpha.
+//  x[i] *= alpha
+// Sscal has no effect if incX < 0.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Sscal(n int, alpha float32, x []float32, incX int) {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if alpha == 0 {
+		if incX == 1 {
+			x = x[:n]
+			for i := range x {
+				x[i] = 0
+			}
+			return
+		}
+		for ix := 0; ix < n*incX; ix += incX {
+			x[ix] = 0
+		}
+		return
+	}
+	if incX == 1 {
+		f32.ScalUnitary(alpha, x[:n])
+		return
+	}
+	f32.ScalInc(alpha, x, uintptr(n), uintptr(incX))
+}

vendor/gonum.org/v1/gonum/blas/gonum/level1float32_dsdot.go

@@ -0,0 +1,53 @@
+// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.
+
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	"gonum.org/v1/gonum/internal/asm/f32"
+)
+
+// Dsdot computes the dot product of the two vectors
+//  \sum_i x[i]*y[i]
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Dsdot(n int, x []float32, incX int, y []float32, incY int) float64 {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n <= 0 {
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	if incX == 1 && incY == 1 {
+		if len(x) < n {
+			panic(shortX)
+		}
+		if len(y) < n {
+			panic(shortY)
+		}
+		return f32.DdotUnitary(x[:n], y[:n])
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	if ix >= len(x) || ix+(n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if iy >= len(y) || iy+(n-1)*incY >= len(y) {
+		panic(shortY)
+	}
+	return f32.DdotInc(x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy))
+}

vendor/gonum.org/v1/gonum/blas/gonum/level1float32_sdot.go

@@ -0,0 +1,53 @@
+// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.
+
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	"gonum.org/v1/gonum/internal/asm/f32"
+)
+
+// Sdot computes the dot product of the two vectors
+//  \sum_i x[i]*y[i]
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Sdot(n int, x []float32, incX int, y []float32, incY int) float32 {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n <= 0 {
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	if incX == 1 && incY == 1 {
+		if len(x) < n {
+			panic(shortX)
+		}
+		if len(y) < n {
+			panic(shortY)
+		}
+		return f32.DotUnitary(x[:n], y[:n])
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	if ix >= len(x) || ix+(n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if iy >= len(y) || iy+(n-1)*incY >= len(y) {
+		panic(shortY)
+	}
+	return f32.DotInc(x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy))
+}

vendor/gonum.org/v1/gonum/blas/gonum/level1float32_sdsdot.go

@@ -0,0 +1,53 @@
+// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.
+
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	"gonum.org/v1/gonum/internal/asm/f32"
+)
+
+// Sdsdot computes the dot product of the two vectors plus a constant
+//  alpha + \sum_i x[i]*y[i]
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Sdsdot(n int, alpha float32, x []float32, incX int, y []float32, incY int) float32 {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n <= 0 {
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	if incX == 1 && incY == 1 {
+		if len(x) < n {
+			panic(shortX)
+		}
+		if len(y) < n {
+			panic(shortY)
+		}
+		return alpha + float32(f32.DdotUnitary(x[:n], y[:n]))
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	if ix >= len(x) || ix+(n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if iy >= len(y) || iy+(n-1)*incY >= len(y) {
+		panic(shortY)
+	}
+	return alpha + float32(f32.DdotInc(x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy)))
+}

vendor/gonum.org/v1/gonum/blas/gonum/level1float64.go

@@ -0,0 +1,620 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	"math"
+
+	"gonum.org/v1/gonum/blas"
+	"gonum.org/v1/gonum/internal/asm/f64"
+)
+
+var _ blas.Float64Level1 = Implementation{}
+
+// Dnrm2 computes the Euclidean norm of a vector,
+//  sqrt(\sum_i x[i] * x[i]).
+// This function returns 0 if incX is negative.
+func (Implementation) Dnrm2(n int, x []float64, incX int) float64 {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return 0
+	}
+	if len(x) <= (n-1)*incX {
+		panic(shortX)
+	}
+	if n < 2 {
+		if n == 1 {
+			return math.Abs(x[0])
+		}
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	var (
+		scale      float64 = 0
+		sumSquares float64 = 1
+	)
+	if incX == 1 {
+		x = x[:n]
+		for _, v := range x {
+			if v == 0 {
+				continue
+			}
+			absxi := math.Abs(v)
+			if math.IsNaN(absxi) {
+				return math.NaN()
+			}
+			if scale < absxi {
+				sumSquares = 1 + sumSquares*(scale/absxi)*(scale/absxi)
+				scale = absxi
+			} else {
+				sumSquares = sumSquares + (absxi/scale)*(absxi/scale)
+			}
+		}
+		if math.IsInf(scale, 1) {
+			return math.Inf(1)
+		}
+		return scale * math.Sqrt(sumSquares)
+	}
+	for ix := 0; ix < n*incX; ix += incX {
+		val := x[ix]
+		if val == 0 {
+			continue
+		}
+		absxi := math.Abs(val)
+		if math.IsNaN(absxi) {
+			return math.NaN()
+		}
+		if scale < absxi {
+			sumSquares = 1 + sumSquares*(scale/absxi)*(scale/absxi)
+			scale = absxi
+		} else {
+			sumSquares = sumSquares + (absxi/scale)*(absxi/scale)
+		}
+	}
+	if math.IsInf(scale, 1) {
+		return math.Inf(1)
+	}
+	return scale * math.Sqrt(sumSquares)
+}
+
+// Dasum computes the sum of the absolute values of the elements of x.
+//  \sum_i |x[i]|
+// Dasum returns 0 if incX is negative.
+func (Implementation) Dasum(n int, x []float64, incX int) float64 {
+	var sum float64
+	if n < 0 {
+		panic(nLT0)
+	}
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return 0
+	}
+	if len(x) <= (n-1)*incX {
+		panic(shortX)
+	}
+	if incX == 1 {
+		x = x[:n]
+		for _, v := range x {
+			sum += math.Abs(v)
+		}
+		return sum
+	}
+	for i := 0; i < n; i++ {
+		sum += math.Abs(x[i*incX])
+	}
+	return sum
+}
+
+// Idamax returns the index of an element of x with the largest absolute value.
+// If there are multiple such indices the earliest is returned.
+// Idamax returns -1 if n == 0.
+func (Implementation) Idamax(n int, x []float64, incX int) int {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return -1
+	}
+	if len(x) <= (n-1)*incX {
+		panic(shortX)
+	}
+	if n < 2 {
+		if n == 1 {
+			return 0
+		}
+		if n == 0 {
+			return -1 // Netlib returns invalid index when n == 0.
+		}
+		panic(nLT0)
+	}
+	idx := 0
+	max := math.Abs(x[0])
+	if incX == 1 {
+		for i, v := range x[:n] {
+			absV := math.Abs(v)
+			if absV > max {
+				max = absV
+				idx = i
+			}
+		}
+		return idx
+	}
+	ix := incX
+	for i := 1; i < n; i++ {
+		v := x[ix]
+		absV := math.Abs(v)
+		if absV > max {
+			max = absV
+			idx = i
+		}
+		ix += incX
+	}
+	return idx
+}
+
+// Dswap exchanges the elements of two vectors.
+//  x[i], y[i] = y[i], x[i] for all i
+func (Implementation) Dswap(n int, x []float64, incX int, y []float64, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && len(x) <= (n-1)*incX) || (incX < 0 && len(x) <= (1-n)*incX) {
+		panic(shortX)
+	}
+	if (incY > 0 && len(y) <= (n-1)*incY) || (incY < 0 && len(y) <= (1-n)*incY) {
+		panic(shortY)
+	}
+	if incX == 1 && incY == 1 {
+		x = x[:n]
+		for i, v := range x {
+			x[i], y[i] = y[i], v
+		}
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	for i := 0; i < n; i++ {
+		x[ix], y[iy] = y[iy], x[ix]
+		ix += incX
+		iy += incY
+	}
+}
+
+// Dcopy copies the elements of x into the elements of y.
+//  y[i] = x[i] for all i
+func (Implementation) Dcopy(n int, x []float64, incX int, y []float64, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && len(x) <= (n-1)*incX) || (incX < 0 && len(x) <= (1-n)*incX) {
+		panic(shortX)
+	}
+	if (incY > 0 && len(y) <= (n-1)*incY) || (incY < 0 && len(y) <= (1-n)*incY) {
+		panic(shortY)
+	}
+	if incX == 1 && incY == 1 {
+		copy(y[:n], x[:n])
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	for i := 0; i < n; i++ {
+		y[iy] = x[ix]
+		ix += incX
+		iy += incY
+	}
+}
+
+// Daxpy adds alpha times x to y
+//  y[i] += alpha * x[i] for all i
+func (Implementation) Daxpy(n int, alpha float64, x []float64, incX int, y []float64, incY int) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && len(x) <= (n-1)*incX) || (incX < 0 && len(x) <= (1-n)*incX) {
+		panic(shortX)
+	}
+	if (incY > 0 && len(y) <= (n-1)*incY) || (incY < 0 && len(y) <= (1-n)*incY) {
+		panic(shortY)
+	}
+	if alpha == 0 {
+		return
+	}
+	if incX == 1 && incY == 1 {
+		f64.AxpyUnitary(alpha, x[:n], y[:n])
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	f64.AxpyInc(alpha, x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy))
+}
+
+// Drotg computes the plane rotation
+//   _    _      _ _       _ _
+//  |  c s |    | a |     | r |
+//  | -s c |  * | b |   = | 0 |
+//   ‾    ‾      ‾ ‾       ‾ ‾
+// where
+//  r = ±√(a^2 + b^2)
+//  c = a/r, the cosine of the plane rotation
+//  s = b/r, the sine of the plane rotation
+//
+// NOTE: There is a discrepancy between the reference implementation and the BLAS
+// technical manual regarding the sign for r when a or b are zero.
+// Drotg agrees with the definition in the manual and other
+// common BLAS implementations.
+func (Implementation) Drotg(a, b float64) (c, s, r, z float64) {
+	if b == 0 && a == 0 {
+		return 1, 0, a, 0
+	}
+	absA := math.Abs(a)
+	absB := math.Abs(b)
+	aGTb := absA > absB
+	r = math.Hypot(a, b)
+	if aGTb {
+		r = math.Copysign(r, a)
+	} else {
+		r = math.Copysign(r, b)
+	}
+	c = a / r
+	s = b / r
+	if aGTb {
+		z = s
+	} else if c != 0 { // r == 0 case handled above
+		z = 1 / c
+	} else {
+		z = 1
+	}
+	return
+}
+
+// Drotmg computes the modified Givens rotation. See
+// http://www.netlib.org/lapack/explore-html/df/deb/drotmg_8f.html
+// for more details.
+func (Implementation) Drotmg(d1, d2, x1, y1 float64) (p blas.DrotmParams, rd1, rd2, rx1 float64) {
+	// The implementation of Drotmg used here is taken from Hopkins 1997
+	// Appendix A: https://doi.org/10.1145/289251.289253
+	// with the exception of the gam constants below.
+
+	const (
+		gam    = 4096.0
+		gamsq  = gam * gam
+		rgamsq = 1.0 / gamsq
+	)
+
+	if d1 < 0 {
+		p.Flag = blas.Rescaling // Error state.
+		return p, 0, 0, 0
+	}
+
+	if d2 == 0 || y1 == 0 {
+		p.Flag = blas.Identity
+		return p, d1, d2, x1
+	}
+
+	var h11, h12, h21, h22 float64
+	if (d1 == 0 || x1 == 0) && d2 > 0 {
+		p.Flag = blas.Diagonal
+		h12 = 1
+		h21 = -1
+		x1 = y1
+		d1, d2 = d2, d1
+	} else {
+		p2 := d2 * y1
+		p1 := d1 * x1
+		q2 := p2 * y1
+		q1 := p1 * x1
+		if math.Abs(q1) > math.Abs(q2) {
+			p.Flag = blas.OffDiagonal
+			h11 = 1
+			h22 = 1
+			h21 = -y1 / x1
+			h12 = p2 / p1
+			u := 1 - h12*h21
+			if u <= 0 {
+				p.Flag = blas.Rescaling // Error state.
+				return p, 0, 0, 0
+			}
+
+			d1 /= u
+			d2 /= u
+			x1 *= u
+		} else {
+			if q2 < 0 {
+				p.Flag = blas.Rescaling // Error state.
+				return p, 0, 0, 0
+			}
+
+			p.Flag = blas.Diagonal
+			h21 = -1
+			h12 = 1
+			h11 = p1 / p2
+			h22 = x1 / y1
+			u := 1 + h11*h22
+			d1, d2 = d2/u, d1/u
+			x1 = y1 * u
+		}
+	}
+
+	for d1 <= rgamsq && d1 != 0 {
+		p.Flag = blas.Rescaling
+		d1 = (d1 * gam) * gam
+		x1 /= gam
+		h11 /= gam
+		h12 /= gam
+	}
+	for d1 > gamsq {
+		p.Flag = blas.Rescaling
+		d1 = (d1 / gam) / gam
+		x1 *= gam
+		h11 *= gam
+		h12 *= gam
+	}
+
+	for math.Abs(d2) <= rgamsq && d2 != 0 {
+		p.Flag = blas.Rescaling
+		d2 = (d2 * gam) * gam
+		h21 /= gam
+		h22 /= gam
+	}
+	for math.Abs(d2) > gamsq {
+		p.Flag = blas.Rescaling
+		d2 = (d2 / gam) / gam
+		h21 *= gam
+		h22 *= gam
+	}
+
+	switch p.Flag {
+	case blas.Diagonal:
+		p.H = [4]float64{0: h11, 3: h22}
+	case blas.OffDiagonal:
+		p.H = [4]float64{1: h21, 2: h12}
+	case blas.Rescaling:
+		p.H = [4]float64{h11, h21, h12, h22}
+	default:
+		panic(badFlag)
+	}
+
+	return p, d1, d2, x1
+}
+
+// Drot applies a plane transformation.
+//  x[i] = c * x[i] + s * y[i]
+//  y[i] = c * y[i] - s * x[i]
+func (Implementation) Drot(n int, x []float64, incX int, y []float64, incY int, c float64, s float64) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && len(x) <= (n-1)*incX) || (incX < 0 && len(x) <= (1-n)*incX) {
+		panic(shortX)
+	}
+	if (incY > 0 && len(y) <= (n-1)*incY) || (incY < 0 && len(y) <= (1-n)*incY) {
+		panic(shortY)
+	}
+	if incX == 1 && incY == 1 {
+		x = x[:n]
+		for i, vx := range x {
+			vy := y[i]
+			x[i], y[i] = c*vx+s*vy, c*vy-s*vx
+		}
+		return
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	for i := 0; i < n; i++ {
+		vx := x[ix]
+		vy := y[iy]
+		x[ix], y[iy] = c*vx+s*vy, c*vy-s*vx
+		ix += incX
+		iy += incY
+	}
+}
+
+// Drotm applies the modified Givens rotation to the 2×n matrix.
+func (Implementation) Drotm(n int, x []float64, incX int, y []float64, incY int, p blas.DrotmParams) {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n <= 0 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (incX > 0 && len(x) <= (n-1)*incX) || (incX < 0 && len(x) <= (1-n)*incX) {
+		panic(shortX)
+	}
+	if (incY > 0 && len(y) <= (n-1)*incY) || (incY < 0 && len(y) <= (1-n)*incY) {
+		panic(shortY)
+	}
+
+	if p.Flag == blas.Identity {
+		return
+	}
+
+	switch p.Flag {
+	case blas.Rescaling:
+		h11 := p.H[0]
+		h12 := p.H[2]
+		h21 := p.H[1]
+		h22 := p.H[3]
+		if incX == 1 && incY == 1 {
+			x = x[:n]
+			for i, vx := range x {
+				vy := y[i]
+				x[i], y[i] = vx*h11+vy*h12, vx*h21+vy*h22
+			}
+			return
+		}
+		var ix, iy int
+		if incX < 0 {
+			ix = (-n + 1) * incX
+		}
+		if incY < 0 {
+			iy = (-n + 1) * incY
+		}
+		for i := 0; i < n; i++ {
+			vx := x[ix]
+			vy := y[iy]
+			x[ix], y[iy] = vx*h11+vy*h12, vx*h21+vy*h22
+			ix += incX
+			iy += incY
+		}
+	case blas.OffDiagonal:
+		h12 := p.H[2]
+		h21 := p.H[1]
+		if incX == 1 && incY == 1 {
+			x = x[:n]
+			for i, vx := range x {
+				vy := y[i]
+				x[i], y[i] = vx+vy*h12, vx*h21+vy
+			}
+			return
+		}
+		var ix, iy int
+		if incX < 0 {
+			ix = (-n + 1) * incX
+		}
+		if incY < 0 {
+			iy = (-n + 1) * incY
+		}
+		for i := 0; i < n; i++ {
+			vx := x[ix]
+			vy := y[iy]
+			x[ix], y[iy] = vx+vy*h12, vx*h21+vy
+			ix += incX
+			iy += incY
+		}
+	case blas.Diagonal:
+		h11 := p.H[0]
+		h22 := p.H[3]
+		if incX == 1 && incY == 1 {
+			x = x[:n]
+			for i, vx := range x {
+				vy := y[i]
+				x[i], y[i] = vx*h11+vy, -vx+vy*h22
+			}
+			return
+		}
+		var ix, iy int
+		if incX < 0 {
+			ix = (-n + 1) * incX
+		}
+		if incY < 0 {
+			iy = (-n + 1) * incY
+		}
+		for i := 0; i < n; i++ {
+			vx := x[ix]
+			vy := y[iy]
+			x[ix], y[iy] = vx*h11+vy, -vx+vy*h22
+			ix += incX
+			iy += incY
+		}
+	}
+}
+
+// Dscal scales x by alpha.
+//  x[i] *= alpha
+// Dscal has no effect if incX < 0.
+func (Implementation) Dscal(n int, alpha float64, x []float64, incX int) {
+	if incX < 1 {
+		if incX == 0 {
+			panic(zeroIncX)
+		}
+		return
+	}
+	if n < 1 {
+		if n == 0 {
+			return
+		}
+		panic(nLT0)
+	}
+	if (n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if alpha == 0 {
+		if incX == 1 {
+			x = x[:n]
+			for i := range x {
+				x[i] = 0
+			}
+			return
+		}
+		for ix := 0; ix < n*incX; ix += incX {
+			x[ix] = 0
+		}
+		return
+	}
+	if incX == 1 {
+		f64.ScalUnitary(alpha, x[:n])
+		return
+	}
+	f64.ScalInc(alpha, x, uintptr(n), uintptr(incX))
+}

vendor/gonum.org/v1/gonum/blas/gonum/level1float64_ddot.go

@@ -0,0 +1,49 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	"gonum.org/v1/gonum/internal/asm/f64"
+)
+
+// Ddot computes the dot product of the two vectors
+//  \sum_i x[i]*y[i]
+func (Implementation) Ddot(n int, x []float64, incX int, y []float64, incY int) float64 {
+	if incX == 0 {
+		panic(zeroIncX)
+	}
+	if incY == 0 {
+		panic(zeroIncY)
+	}
+	if n <= 0 {
+		if n == 0 {
+			return 0
+		}
+		panic(nLT0)
+	}
+	if incX == 1 && incY == 1 {
+		if len(x) < n {
+			panic(shortX)
+		}
+		if len(y) < n {
+			panic(shortY)
+		}
+		return f64.DotUnitary(x[:n], y[:n])
+	}
+	var ix, iy int
+	if incX < 0 {
+		ix = (-n + 1) * incX
+	}
+	if incY < 0 {
+		iy = (-n + 1) * incY
+	}
+	if ix >= len(x) || ix+(n-1)*incX >= len(x) {
+		panic(shortX)
+	}
+	if iy >= len(y) || iy+(n-1)*incY >= len(y) {
+		panic(shortY)
+	}
+	return f64.DotInc(x, y, uintptr(n), uintptr(incX), uintptr(incY), uintptr(ix), uintptr(iy))
+}

vendor/gonum.org/v1/gonum/blas/gonum/level3float32.go

@@ -0,0 +1,876 @@
+// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.
+
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	"gonum.org/v1/gonum/blas"
+	"gonum.org/v1/gonum/internal/asm/f32"
+)
+
+var _ blas.Float32Level3 = Implementation{}
+
+// Strsm solves one of the matrix equations
+//  A * X = alpha * B    if tA == blas.NoTrans and side == blas.Left
+//  A^T * X = alpha * B  if tA == blas.Trans or blas.ConjTrans, and side == blas.Left
+//  X * A = alpha * B    if tA == blas.NoTrans and side == blas.Right
+//  X * A^T = alpha * B  if tA == blas.Trans or blas.ConjTrans, and side == blas.Right
+// where A is an n×n or m×m triangular matrix, X and B are m×n matrices, and alpha is a
+// scalar.
+//
+// At entry to the function, X contains the values of B, and the result is
+// stored in-place into X.
+//
+// No check is made that A is invertible.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Strsm(s blas.Side, ul blas.Uplo, tA blas.Transpose, d blas.Diag, m, n int, alpha float32, a []float32, lda int, b []float32, ldb int) {
+	if s != blas.Left && s != blas.Right {
+		panic(badSide)
+	}
+	if ul != blas.Lower && ul != blas.Upper {
+		panic(badUplo)
+	}
+	if tA != blas.NoTrans && tA != blas.Trans && tA != blas.ConjTrans {
+		panic(badTranspose)
+	}
+	if d != blas.NonUnit && d != blas.Unit {
+		panic(badDiag)
+	}
+	if m < 0 {
+		panic(mLT0)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	k := n
+	if s == blas.Left {
+		k = m
+	}
+	if lda < max(1, k) {
+		panic(badLdA)
+	}
+	if ldb < max(1, n) {
+		panic(badLdB)
+	}
+
+	// Quick return if possible.
+	if m == 0 || n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if len(a) < lda*(k-1)+k {
+		panic(shortA)
+	}
+	if len(b) < ldb*(m-1)+n {
+		panic(shortB)
+	}
+
+	if alpha == 0 {
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			for j := range btmp {
+				btmp[j] = 0
+			}
+		}
+		return
+	}
+	nonUnit := d == blas.NonUnit
+	if s == blas.Left {
+		if tA == blas.NoTrans {
+			if ul == blas.Upper {
+				for i := m - 1; i >= 0; i-- {
+					btmp := b[i*ldb : i*ldb+n]
+					if alpha != 1 {
+						f32.ScalUnitary(alpha, btmp)
+					}
+					for ka, va := range a[i*lda+i+1 : i*lda+m] {
+						if va != 0 {
+							k := ka + i + 1
+							f32.AxpyUnitary(-va, b[k*ldb:k*ldb+n], btmp)
+						}
+					}
+					if nonUnit {
+						tmp := 1 / a[i*lda+i]
+						f32.ScalUnitary(tmp, btmp)
+					}
+				}
+				return
+			}
+			for i := 0; i < m; i++ {
+				btmp := b[i*ldb : i*ldb+n]
+				if alpha != 1 {
+					f32.ScalUnitary(alpha, btmp)
+				}
+				for k, va := range a[i*lda : i*lda+i] {
+					if va != 0 {
+						f32.AxpyUnitary(-va, b[k*ldb:k*ldb+n], btmp)
+					}
+				}
+				if nonUnit {
+					tmp := 1 / a[i*lda+i]
+					f32.ScalUnitary(tmp, btmp)
+				}
+			}
+			return
+		}
+		// Cases where a is transposed
+		if ul == blas.Upper {
+			for k := 0; k < m; k++ {
+				btmpk := b[k*ldb : k*ldb+n]
+				if nonUnit {
+					tmp := 1 / a[k*lda+k]
+					f32.ScalUnitary(tmp, btmpk)
+				}
+				for ia, va := range a[k*lda+k+1 : k*lda+m] {
+					if va != 0 {
+						i := ia + k + 1
+						f32.AxpyUnitary(-va, btmpk, b[i*ldb:i*ldb+n])
+					}
+				}
+				if alpha != 1 {
+					f32.ScalUnitary(alpha, btmpk)
+				}
+			}
+			return
+		}
+		for k := m - 1; k >= 0; k-- {
+			btmpk := b[k*ldb : k*ldb+n]
+			if nonUnit {
+				tmp := 1 / a[k*lda+k]
+				f32.ScalUnitary(tmp, btmpk)
+			}
+			for i, va := range a[k*lda : k*lda+k] {
+				if va != 0 {
+					f32.AxpyUnitary(-va, btmpk, b[i*ldb:i*ldb+n])
+				}
+			}
+			if alpha != 1 {
+				f32.ScalUnitary(alpha, btmpk)
+			}
+		}
+		return
+	}
+	// Cases where a is to the right of X.
+	if tA == blas.NoTrans {
+		if ul == blas.Upper {
+			for i := 0; i < m; i++ {
+				btmp := b[i*ldb : i*ldb+n]
+				if alpha != 1 {
+					f32.ScalUnitary(alpha, btmp)
+				}
+				for k, vb := range btmp {
+					if vb == 0 {
+						continue
+					}
+					if nonUnit {
+						btmp[k] /= a[k*lda+k]
+					}
+					f32.AxpyUnitary(-btmp[k], a[k*lda+k+1:k*lda+n], btmp[k+1:n])
+				}
+			}
+			return
+		}
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			if alpha != 1 {
+				f32.ScalUnitary(alpha, btmp)
+			}
+			for k := n - 1; k >= 0; k-- {
+				if btmp[k] == 0 {
+					continue
+				}
+				if nonUnit {
+					btmp[k] /= a[k*lda+k]
+				}
+				f32.AxpyUnitary(-btmp[k], a[k*lda:k*lda+k], btmp[:k])
+			}
+		}
+		return
+	}
+	// Cases where a is transposed.
+	if ul == blas.Upper {
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			for j := n - 1; j >= 0; j-- {
+				tmp := alpha*btmp[j] - f32.DotUnitary(a[j*lda+j+1:j*lda+n], btmp[j+1:])
+				if nonUnit {
+					tmp /= a[j*lda+j]
+				}
+				btmp[j] = tmp
+			}
+		}
+		return
+	}
+	for i := 0; i < m; i++ {
+		btmp := b[i*ldb : i*ldb+n]
+		for j := 0; j < n; j++ {
+			tmp := alpha*btmp[j] - f32.DotUnitary(a[j*lda:j*lda+j], btmp[:j])
+			if nonUnit {
+				tmp /= a[j*lda+j]
+			}
+			btmp[j] = tmp
+		}
+	}
+}
+
+// Ssymm performs one of the matrix-matrix operations
+//  C = alpha * A * B + beta * C  if side == blas.Left
+//  C = alpha * B * A + beta * C  if side == blas.Right
+// where A is an n×n or m×m symmetric matrix, B and C are m×n matrices, and alpha
+// is a scalar.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Ssymm(s blas.Side, ul blas.Uplo, m, n int, alpha float32, a []float32, lda int, b []float32, ldb int, beta float32, c []float32, ldc int) {
+	if s != blas.Right && s != blas.Left {
+		panic(badSide)
+	}
+	if ul != blas.Lower && ul != blas.Upper {
+		panic(badUplo)
+	}
+	if m < 0 {
+		panic(mLT0)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	k := n
+	if s == blas.Left {
+		k = m
+	}
+	if lda < max(1, k) {
+		panic(badLdA)
+	}
+	if ldb < max(1, n) {
+		panic(badLdB)
+	}
+	if ldc < max(1, n) {
+		panic(badLdC)
+	}
+
+	// Quick return if possible.
+	if m == 0 || n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if len(a) < lda*(k-1)+k {
+		panic(shortA)
+	}
+	if len(b) < ldb*(m-1)+n {
+		panic(shortB)
+	}
+	if len(c) < ldc*(m-1)+n {
+		panic(shortC)
+	}
+
+	// Quick return if possible.
+	if alpha == 0 && beta == 1 {
+		return
+	}
+
+	if alpha == 0 {
+		if beta == 0 {
+			for i := 0; i < m; i++ {
+				ctmp := c[i*ldc : i*ldc+n]
+				for j := range ctmp {
+					ctmp[j] = 0
+				}
+			}
+			return
+		}
+		for i := 0; i < m; i++ {
+			ctmp := c[i*ldc : i*ldc+n]
+			for j := 0; j < n; j++ {
+				ctmp[j] *= beta
+			}
+		}
+		return
+	}
+
+	isUpper := ul == blas.Upper
+	if s == blas.Left {
+		for i := 0; i < m; i++ {
+			atmp := alpha * a[i*lda+i]
+			btmp := b[i*ldb : i*ldb+n]
+			ctmp := c[i*ldc : i*ldc+n]
+			for j, v := range btmp {
+				ctmp[j] *= beta
+				ctmp[j] += atmp * v
+			}
+
+			for k := 0; k < i; k++ {
+				var atmp float32
+				if isUpper {
+					atmp = a[k*lda+i]
+				} else {
+					atmp = a[i*lda+k]
+				}
+				atmp *= alpha
+				f32.AxpyUnitary(atmp, b[k*ldb:k*ldb+n], ctmp)
+			}
+			for k := i + 1; k < m; k++ {
+				var atmp float32
+				if isUpper {
+					atmp = a[i*lda+k]
+				} else {
+					atmp = a[k*lda+i]
+				}
+				atmp *= alpha
+				f32.AxpyUnitary(atmp, b[k*ldb:k*ldb+n], ctmp)
+			}
+		}
+		return
+	}
+	if isUpper {
+		for i := 0; i < m; i++ {
+			for j := n - 1; j >= 0; j-- {
+				tmp := alpha * b[i*ldb+j]
+				var tmp2 float32
+				atmp := a[j*lda+j+1 : j*lda+n]
+				btmp := b[i*ldb+j+1 : i*ldb+n]
+				ctmp := c[i*ldc+j+1 : i*ldc+n]
+				for k, v := range atmp {
+					ctmp[k] += tmp * v
+					tmp2 += btmp[k] * v
+				}
+				c[i*ldc+j] *= beta
+				c[i*ldc+j] += tmp*a[j*lda+j] + alpha*tmp2
+			}
+		}
+		return
+	}
+	for i := 0; i < m; i++ {
+		for j := 0; j < n; j++ {
+			tmp := alpha * b[i*ldb+j]
+			var tmp2 float32
+			atmp := a[j*lda : j*lda+j]
+			btmp := b[i*ldb : i*ldb+j]
+			ctmp := c[i*ldc : i*ldc+j]
+			for k, v := range atmp {
+				ctmp[k] += tmp * v
+				tmp2 += btmp[k] * v
+			}
+			c[i*ldc+j] *= beta
+			c[i*ldc+j] += tmp*a[j*lda+j] + alpha*tmp2
+		}
+	}
+}
+
+// Ssyrk performs one of the symmetric rank-k operations
+//  C = alpha * A * A^T + beta * C  if tA == blas.NoTrans
+//  C = alpha * A^T * A + beta * C  if tA == blas.Trans or tA == blas.ConjTrans
+// where A is an n×k or k×n matrix, C is an n×n symmetric matrix, and alpha and
+// beta are scalars.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Ssyrk(ul blas.Uplo, tA blas.Transpose, n, k int, alpha float32, a []float32, lda int, beta float32, c []float32, ldc int) {
+	if ul != blas.Lower && ul != blas.Upper {
+		panic(badUplo)
+	}
+	if tA != blas.Trans && tA != blas.NoTrans && tA != blas.ConjTrans {
+		panic(badTranspose)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	if k < 0 {
+		panic(kLT0)
+	}
+	row, col := k, n
+	if tA == blas.NoTrans {
+		row, col = n, k
+	}
+	if lda < max(1, col) {
+		panic(badLdA)
+	}
+	if ldc < max(1, n) {
+		panic(badLdC)
+	}
+
+	// Quick return if possible.
+	if n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if len(a) < lda*(row-1)+col {
+		panic(shortA)
+	}
+	if len(c) < ldc*(n-1)+n {
+		panic(shortC)
+	}
+
+	if alpha == 0 {
+		if beta == 0 {
+			if ul == blas.Upper {
+				for i := 0; i < n; i++ {
+					ctmp := c[i*ldc+i : i*ldc+n]
+					for j := range ctmp {
+						ctmp[j] = 0
+					}
+				}
+				return
+			}
+			for i := 0; i < n; i++ {
+				ctmp := c[i*ldc : i*ldc+i+1]
+				for j := range ctmp {
+					ctmp[j] = 0
+				}
+			}
+			return
+		}
+		if ul == blas.Upper {
+			for i := 0; i < n; i++ {
+				ctmp := c[i*ldc+i : i*ldc+n]
+				for j := range ctmp {
+					ctmp[j] *= beta
+				}
+			}
+			return
+		}
+		for i := 0; i < n; i++ {
+			ctmp := c[i*ldc : i*ldc+i+1]
+			for j := range ctmp {
+				ctmp[j] *= beta
+			}
+		}
+		return
+	}
+	if tA == blas.NoTrans {
+		if ul == blas.Upper {
+			for i := 0; i < n; i++ {
+				ctmp := c[i*ldc+i : i*ldc+n]
+				atmp := a[i*lda : i*lda+k]
+				if beta == 0 {
+					for jc := range ctmp {
+						j := jc + i
+						ctmp[jc] = alpha * f32.DotUnitary(atmp, a[j*lda:j*lda+k])
+					}
+				} else {
+					for jc, vc := range ctmp {
+						j := jc + i
+						ctmp[jc] = vc*beta + alpha*f32.DotUnitary(atmp, a[j*lda:j*lda+k])
+					}
+				}
+			}
+			return
+		}
+		for i := 0; i < n; i++ {
+			ctmp := c[i*ldc : i*ldc+i+1]
+			atmp := a[i*lda : i*lda+k]
+			if beta == 0 {
+				for j := range ctmp {
+					ctmp[j] = alpha * f32.DotUnitary(a[j*lda:j*lda+k], atmp)
+				}
+			} else {
+				for j, vc := range ctmp {
+					ctmp[j] = vc*beta + alpha*f32.DotUnitary(a[j*lda:j*lda+k], atmp)
+				}
+			}
+		}
+		return
+	}
+	// Cases where a is transposed.
+	if ul == blas.Upper {
+		for i := 0; i < n; i++ {
+			ctmp := c[i*ldc+i : i*ldc+n]
+			if beta == 0 {
+				for j := range ctmp {
+					ctmp[j] = 0
+				}
+			} else if beta != 1 {
+				for j := range ctmp {
+					ctmp[j] *= beta
+				}
+			}
+			for l := 0; l < k; l++ {
+				tmp := alpha * a[l*lda+i]
+				if tmp != 0 {
+					f32.AxpyUnitary(tmp, a[l*lda+i:l*lda+n], ctmp)
+				}
+			}
+		}
+		return
+	}
+	for i := 0; i < n; i++ {
+		ctmp := c[i*ldc : i*ldc+i+1]
+		if beta != 1 {
+			for j := range ctmp {
+				ctmp[j] *= beta
+			}
+		}
+		for l := 0; l < k; l++ {
+			tmp := alpha * a[l*lda+i]
+			if tmp != 0 {
+				f32.AxpyUnitary(tmp, a[l*lda:l*lda+i+1], ctmp)
+			}
+		}
+	}
+}
+
+// Ssyr2k performs one of the symmetric rank 2k operations
+//  C = alpha * A * B^T + alpha * B * A^T + beta * C  if tA == blas.NoTrans
+//  C = alpha * A^T * B + alpha * B^T * A + beta * C  if tA == blas.Trans or tA == blas.ConjTrans
+// where A and B are n×k or k×n matrices, C is an n×n symmetric matrix, and
+// alpha and beta are scalars.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Ssyr2k(ul blas.Uplo, tA blas.Transpose, n, k int, alpha float32, a []float32, lda int, b []float32, ldb int, beta float32, c []float32, ldc int) {
+	if ul != blas.Lower && ul != blas.Upper {
+		panic(badUplo)
+	}
+	if tA != blas.Trans && tA != blas.NoTrans && tA != blas.ConjTrans {
+		panic(badTranspose)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	if k < 0 {
+		panic(kLT0)
+	}
+	row, col := k, n
+	if tA == blas.NoTrans {
+		row, col = n, k
+	}
+	if lda < max(1, col) {
+		panic(badLdA)
+	}
+	if ldb < max(1, col) {
+		panic(badLdB)
+	}
+	if ldc < max(1, n) {
+		panic(badLdC)
+	}
+
+	// Quick return if possible.
+	if n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if len(a) < lda*(row-1)+col {
+		panic(shortA)
+	}
+	if len(b) < ldb*(row-1)+col {
+		panic(shortB)
+	}
+	if len(c) < ldc*(n-1)+n {
+		panic(shortC)
+	}
+
+	if alpha == 0 {
+		if beta == 0 {
+			if ul == blas.Upper {
+				for i := 0; i < n; i++ {
+					ctmp := c[i*ldc+i : i*ldc+n]
+					for j := range ctmp {
+						ctmp[j] = 0
+					}
+				}
+				return
+			}
+			for i := 0; i < n; i++ {
+				ctmp := c[i*ldc : i*ldc+i+1]
+				for j := range ctmp {
+					ctmp[j] = 0
+				}
+			}
+			return
+		}
+		if ul == blas.Upper {
+			for i := 0; i < n; i++ {
+				ctmp := c[i*ldc+i : i*ldc+n]
+				for j := range ctmp {
+					ctmp[j] *= beta
+				}
+			}
+			return
+		}
+		for i := 0; i < n; i++ {
+			ctmp := c[i*ldc : i*ldc+i+1]
+			for j := range ctmp {
+				ctmp[j] *= beta
+			}
+		}
+		return
+	}
+	if tA == blas.NoTrans {
+		if ul == blas.Upper {
+			for i := 0; i < n; i++ {
+				atmp := a[i*lda : i*lda+k]
+				btmp := b[i*ldb : i*ldb+k]
+				ctmp := c[i*ldc+i : i*ldc+n]
+				for jc := range ctmp {
+					j := i + jc
+					var tmp1, tmp2 float32
+					binner := b[j*ldb : j*ldb+k]
+					for l, v := range a[j*lda : j*lda+k] {
+						tmp1 += v * btmp[l]
+						tmp2 += atmp[l] * binner[l]
+					}
+					ctmp[jc] *= beta
+					ctmp[jc] += alpha * (tmp1 + tmp2)
+				}
+			}
+			return
+		}
+		for i := 0; i < n; i++ {
+			atmp := a[i*lda : i*lda+k]
+			btmp := b[i*ldb : i*ldb+k]
+			ctmp := c[i*ldc : i*ldc+i+1]
+			for j := 0; j <= i; j++ {
+				var tmp1, tmp2 float32
+				binner := b[j*ldb : j*ldb+k]
+				for l, v := range a[j*lda : j*lda+k] {
+					tmp1 += v * btmp[l]
+					tmp2 += atmp[l] * binner[l]
+				}
+				ctmp[j] *= beta
+				ctmp[j] += alpha * (tmp1 + tmp2)
+			}
+		}
+		return
+	}
+	if ul == blas.Upper {
+		for i := 0; i < n; i++ {
+			ctmp := c[i*ldc+i : i*ldc+n]
+			if beta != 1 {
+				for j := range ctmp {
+					ctmp[j] *= beta
+				}
+			}
+			for l := 0; l < k; l++ {
+				tmp1 := alpha * b[l*ldb+i]
+				tmp2 := alpha * a[l*lda+i]
+				btmp := b[l*ldb+i : l*ldb+n]
+				if tmp1 != 0 || tmp2 != 0 {
+					for j, v := range a[l*lda+i : l*lda+n] {
+						ctmp[j] += v*tmp1 + btmp[j]*tmp2
+					}
+				}
+			}
+		}
+		return
+	}
+	for i := 0; i < n; i++ {
+		ctmp := c[i*ldc : i*ldc+i+1]
+		if beta != 1 {
+			for j := range ctmp {
+				ctmp[j] *= beta
+			}
+		}
+		for l := 0; l < k; l++ {
+			tmp1 := alpha * b[l*ldb+i]
+			tmp2 := alpha * a[l*lda+i]
+			btmp := b[l*ldb : l*ldb+i+1]
+			if tmp1 != 0 || tmp2 != 0 {
+				for j, v := range a[l*lda : l*lda+i+1] {
+					ctmp[j] += v*tmp1 + btmp[j]*tmp2
+				}
+			}
+		}
+	}
+}
+
+// Strmm performs one of the matrix-matrix operations
+//  B = alpha * A * B    if tA == blas.NoTrans and side == blas.Left
+//  B = alpha * A^T * B  if tA == blas.Trans or blas.ConjTrans, and side == blas.Left
+//  B = alpha * B * A    if tA == blas.NoTrans and side == blas.Right
+//  B = alpha * B * A^T  if tA == blas.Trans or blas.ConjTrans, and side == blas.Right
+// where A is an n×n or m×m triangular matrix, B is an m×n matrix, and alpha is a scalar.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Strmm(s blas.Side, ul blas.Uplo, tA blas.Transpose, d blas.Diag, m, n int, alpha float32, a []float32, lda int, b []float32, ldb int) {
+	if s != blas.Left && s != blas.Right {
+		panic(badSide)
+	}
+	if ul != blas.Lower && ul != blas.Upper {
+		panic(badUplo)
+	}
+	if tA != blas.NoTrans && tA != blas.Trans && tA != blas.ConjTrans {
+		panic(badTranspose)
+	}
+	if d != blas.NonUnit && d != blas.Unit {
+		panic(badDiag)
+	}
+	if m < 0 {
+		panic(mLT0)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	k := n
+	if s == blas.Left {
+		k = m
+	}
+	if lda < max(1, k) {
+		panic(badLdA)
+	}
+	if ldb < max(1, n) {
+		panic(badLdB)
+	}
+
+	// Quick return if possible.
+	if m == 0 || n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if len(a) < lda*(k-1)+k {
+		panic(shortA)
+	}
+	if len(b) < ldb*(m-1)+n {
+		panic(shortB)
+	}
+
+	if alpha == 0 {
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			for j := range btmp {
+				btmp[j] = 0
+			}
+		}
+		return
+	}
+
+	nonUnit := d == blas.NonUnit
+	if s == blas.Left {
+		if tA == blas.NoTrans {
+			if ul == blas.Upper {
+				for i := 0; i < m; i++ {
+					tmp := alpha
+					if nonUnit {
+						tmp *= a[i*lda+i]
+					}
+					btmp := b[i*ldb : i*ldb+n]
+					f32.ScalUnitary(tmp, btmp)
+					for ka, va := range a[i*lda+i+1 : i*lda+m] {
+						k := ka + i + 1
+						if va != 0 {
+							f32.AxpyUnitary(alpha*va, b[k*ldb:k*ldb+n], btmp)
+						}
+					}
+				}
+				return
+			}
+			for i := m - 1; i >= 0; i-- {
+				tmp := alpha
+				if nonUnit {
+					tmp *= a[i*lda+i]
+				}
+				btmp := b[i*ldb : i*ldb+n]
+				f32.ScalUnitary(tmp, btmp)
+				for k, va := range a[i*lda : i*lda+i] {
+					if va != 0 {
+						f32.AxpyUnitary(alpha*va, b[k*ldb:k*ldb+n], btmp)
+					}
+				}
+			}
+			return
+		}
+		// Cases where a is transposed.
+		if ul == blas.Upper {
+			for k := m - 1; k >= 0; k-- {
+				btmpk := b[k*ldb : k*ldb+n]
+				for ia, va := range a[k*lda+k+1 : k*lda+m] {
+					i := ia + k + 1
+					btmp := b[i*ldb : i*ldb+n]
+					if va != 0 {
+						f32.AxpyUnitary(alpha*va, btmpk, btmp)
+					}
+				}
+				tmp := alpha
+				if nonUnit {
+					tmp *= a[k*lda+k]
+				}
+				if tmp != 1 {
+					f32.ScalUnitary(tmp, btmpk)
+				}
+			}
+			return
+		}
+		for k := 0; k < m; k++ {
+			btmpk := b[k*ldb : k*ldb+n]
+			for i, va := range a[k*lda : k*lda+k] {
+				btmp := b[i*ldb : i*ldb+n]
+				if va != 0 {
+					f32.AxpyUnitary(alpha*va, btmpk, btmp)
+				}
+			}
+			tmp := alpha
+			if nonUnit {
+				tmp *= a[k*lda+k]
+			}
+			if tmp != 1 {
+				f32.ScalUnitary(tmp, btmpk)
+			}
+		}
+		return
+	}
+	// Cases where a is on the right
+	if tA == blas.NoTrans {
+		if ul == blas.Upper {
+			for i := 0; i < m; i++ {
+				btmp := b[i*ldb : i*ldb+n]
+				for k := n - 1; k >= 0; k-- {
+					tmp := alpha * btmp[k]
+					if tmp == 0 {
+						continue
+					}
+					btmp[k] = tmp
+					if nonUnit {
+						btmp[k] *= a[k*lda+k]
+					}
+					f32.AxpyUnitary(tmp, a[k*lda+k+1:k*lda+n], btmp[k+1:n])
+				}
+			}
+			return
+		}
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			for k := 0; k < n; k++ {
+				tmp := alpha * btmp[k]
+				if tmp == 0 {
+					continue
+				}
+				btmp[k] = tmp
+				if nonUnit {
+					btmp[k] *= a[k*lda+k]
+				}
+				f32.AxpyUnitary(tmp, a[k*lda:k*lda+k], btmp[:k])
+			}
+		}
+		return
+	}
+	// Cases where a is transposed.
+	if ul == blas.Upper {
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			for j, vb := range btmp {
+				tmp := vb
+				if nonUnit {
+					tmp *= a[j*lda+j]
+				}
+				tmp += f32.DotUnitary(a[j*lda+j+1:j*lda+n], btmp[j+1:n])
+				btmp[j] = alpha * tmp
+			}
+		}
+		return
+	}
+	for i := 0; i < m; i++ {
+		btmp := b[i*ldb : i*ldb+n]
+		for j := n - 1; j >= 0; j-- {
+			tmp := btmp[j]
+			if nonUnit {
+				tmp *= a[j*lda+j]
+			}
+			tmp += f32.DotUnitary(a[j*lda:j*lda+j], btmp[:j])
+			btmp[j] = alpha * tmp
+		}
+	}
+}

vendor/gonum.org/v1/gonum/blas/gonum/level3float64.go

@@ -0,0 +1,864 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	"gonum.org/v1/gonum/blas"
+	"gonum.org/v1/gonum/internal/asm/f64"
+)
+
+var _ blas.Float64Level3 = Implementation{}
+
+// Dtrsm solves one of the matrix equations
+//  A * X = alpha * B    if tA == blas.NoTrans and side == blas.Left
+//  A^T * X = alpha * B  if tA == blas.Trans or blas.ConjTrans, and side == blas.Left
+//  X * A = alpha * B    if tA == blas.NoTrans and side == blas.Right
+//  X * A^T = alpha * B  if tA == blas.Trans or blas.ConjTrans, and side == blas.Right
+// where A is an n×n or m×m triangular matrix, X and B are m×n matrices, and alpha is a
+// scalar.
+//
+// At entry to the function, X contains the values of B, and the result is
+// stored in-place into X.
+//
+// No check is made that A is invertible.
+func (Implementation) Dtrsm(s blas.Side, ul blas.Uplo, tA blas.Transpose, d blas.Diag, m, n int, alpha float64, a []float64, lda int, b []float64, ldb int) {
+	if s != blas.Left && s != blas.Right {
+		panic(badSide)
+	}
+	if ul != blas.Lower && ul != blas.Upper {
+		panic(badUplo)
+	}
+	if tA != blas.NoTrans && tA != blas.Trans && tA != blas.ConjTrans {
+		panic(badTranspose)
+	}
+	if d != blas.NonUnit && d != blas.Unit {
+		panic(badDiag)
+	}
+	if m < 0 {
+		panic(mLT0)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	k := n
+	if s == blas.Left {
+		k = m
+	}
+	if lda < max(1, k) {
+		panic(badLdA)
+	}
+	if ldb < max(1, n) {
+		panic(badLdB)
+	}
+
+	// Quick return if possible.
+	if m == 0 || n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if len(a) < lda*(k-1)+k {
+		panic(shortA)
+	}
+	if len(b) < ldb*(m-1)+n {
+		panic(shortB)
+	}
+
+	if alpha == 0 {
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			for j := range btmp {
+				btmp[j] = 0
+			}
+		}
+		return
+	}
+	nonUnit := d == blas.NonUnit
+	if s == blas.Left {
+		if tA == blas.NoTrans {
+			if ul == blas.Upper {
+				for i := m - 1; i >= 0; i-- {
+					btmp := b[i*ldb : i*ldb+n]
+					if alpha != 1 {
+						f64.ScalUnitary(alpha, btmp)
+					}
+					for ka, va := range a[i*lda+i+1 : i*lda+m] {
+						if va != 0 {
+							k := ka + i + 1
+							f64.AxpyUnitary(-va, b[k*ldb:k*ldb+n], btmp)
+						}
+					}
+					if nonUnit {
+						tmp := 1 / a[i*lda+i]
+						f64.ScalUnitary(tmp, btmp)
+					}
+				}
+				return
+			}
+			for i := 0; i < m; i++ {
+				btmp := b[i*ldb : i*ldb+n]
+				if alpha != 1 {
+					f64.ScalUnitary(alpha, btmp)
+				}
+				for k, va := range a[i*lda : i*lda+i] {
+					if va != 0 {
+						f64.AxpyUnitary(-va, b[k*ldb:k*ldb+n], btmp)
+					}
+				}
+				if nonUnit {
+					tmp := 1 / a[i*lda+i]
+					f64.ScalUnitary(tmp, btmp)
+				}
+			}
+			return
+		}
+		// Cases where a is transposed
+		if ul == blas.Upper {
+			for k := 0; k < m; k++ {
+				btmpk := b[k*ldb : k*ldb+n]
+				if nonUnit {
+					tmp := 1 / a[k*lda+k]
+					f64.ScalUnitary(tmp, btmpk)
+				}
+				for ia, va := range a[k*lda+k+1 : k*lda+m] {
+					if va != 0 {
+						i := ia + k + 1
+						f64.AxpyUnitary(-va, btmpk, b[i*ldb:i*ldb+n])
+					}
+				}
+				if alpha != 1 {
+					f64.ScalUnitary(alpha, btmpk)
+				}
+			}
+			return
+		}
+		for k := m - 1; k >= 0; k-- {
+			btmpk := b[k*ldb : k*ldb+n]
+			if nonUnit {
+				tmp := 1 / a[k*lda+k]
+				f64.ScalUnitary(tmp, btmpk)
+			}
+			for i, va := range a[k*lda : k*lda+k] {
+				if va != 0 {
+					f64.AxpyUnitary(-va, btmpk, b[i*ldb:i*ldb+n])
+				}
+			}
+			if alpha != 1 {
+				f64.ScalUnitary(alpha, btmpk)
+			}
+		}
+		return
+	}
+	// Cases where a is to the right of X.
+	if tA == blas.NoTrans {
+		if ul == blas.Upper {
+			for i := 0; i < m; i++ {
+				btmp := b[i*ldb : i*ldb+n]
+				if alpha != 1 {
+					f64.ScalUnitary(alpha, btmp)
+				}
+				for k, vb := range btmp {
+					if vb == 0 {
+						continue
+					}
+					if nonUnit {
+						btmp[k] /= a[k*lda+k]
+					}
+					f64.AxpyUnitary(-btmp[k], a[k*lda+k+1:k*lda+n], btmp[k+1:n])
+				}
+			}
+			return
+		}
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			if alpha != 1 {
+				f64.ScalUnitary(alpha, btmp)
+			}
+			for k := n - 1; k >= 0; k-- {
+				if btmp[k] == 0 {
+					continue
+				}
+				if nonUnit {
+					btmp[k] /= a[k*lda+k]
+				}
+				f64.AxpyUnitary(-btmp[k], a[k*lda:k*lda+k], btmp[:k])
+			}
+		}
+		return
+	}
+	// Cases where a is transposed.
+	if ul == blas.Upper {
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			for j := n - 1; j >= 0; j-- {
+				tmp := alpha*btmp[j] - f64.DotUnitary(a[j*lda+j+1:j*lda+n], btmp[j+1:])
+				if nonUnit {
+					tmp /= a[j*lda+j]
+				}
+				btmp[j] = tmp
+			}
+		}
+		return
+	}
+	for i := 0; i < m; i++ {
+		btmp := b[i*ldb : i*ldb+n]
+		for j := 0; j < n; j++ {
+			tmp := alpha*btmp[j] - f64.DotUnitary(a[j*lda:j*lda+j], btmp[:j])
+			if nonUnit {
+				tmp /= a[j*lda+j]
+			}
+			btmp[j] = tmp
+		}
+	}
+}
+
+// Dsymm performs one of the matrix-matrix operations
+//  C = alpha * A * B + beta * C  if side == blas.Left
+//  C = alpha * B * A + beta * C  if side == blas.Right
+// where A is an n×n or m×m symmetric matrix, B and C are m×n matrices, and alpha
+// is a scalar.
+func (Implementation) Dsymm(s blas.Side, ul blas.Uplo, m, n int, alpha float64, a []float64, lda int, b []float64, ldb int, beta float64, c []float64, ldc int) {
+	if s != blas.Right && s != blas.Left {
+		panic(badSide)
+	}
+	if ul != blas.Lower && ul != blas.Upper {
+		panic(badUplo)
+	}
+	if m < 0 {
+		panic(mLT0)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	k := n
+	if s == blas.Left {
+		k = m
+	}
+	if lda < max(1, k) {
+		panic(badLdA)
+	}
+	if ldb < max(1, n) {
+		panic(badLdB)
+	}
+	if ldc < max(1, n) {
+		panic(badLdC)
+	}
+
+	// Quick return if possible.
+	if m == 0 || n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if len(a) < lda*(k-1)+k {
+		panic(shortA)
+	}
+	if len(b) < ldb*(m-1)+n {
+		panic(shortB)
+	}
+	if len(c) < ldc*(m-1)+n {
+		panic(shortC)
+	}
+
+	// Quick return if possible.
+	if alpha == 0 && beta == 1 {
+		return
+	}
+
+	if alpha == 0 {
+		if beta == 0 {
+			for i := 0; i < m; i++ {
+				ctmp := c[i*ldc : i*ldc+n]
+				for j := range ctmp {
+					ctmp[j] = 0
+				}
+			}
+			return
+		}
+		for i := 0; i < m; i++ {
+			ctmp := c[i*ldc : i*ldc+n]
+			for j := 0; j < n; j++ {
+				ctmp[j] *= beta
+			}
+		}
+		return
+	}
+
+	isUpper := ul == blas.Upper
+	if s == blas.Left {
+		for i := 0; i < m; i++ {
+			atmp := alpha * a[i*lda+i]
+			btmp := b[i*ldb : i*ldb+n]
+			ctmp := c[i*ldc : i*ldc+n]
+			for j, v := range btmp {
+				ctmp[j] *= beta
+				ctmp[j] += atmp * v
+			}
+
+			for k := 0; k < i; k++ {
+				var atmp float64
+				if isUpper {
+					atmp = a[k*lda+i]
+				} else {
+					atmp = a[i*lda+k]
+				}
+				atmp *= alpha
+				f64.AxpyUnitary(atmp, b[k*ldb:k*ldb+n], ctmp)
+			}
+			for k := i + 1; k < m; k++ {
+				var atmp float64
+				if isUpper {
+					atmp = a[i*lda+k]
+				} else {
+					atmp = a[k*lda+i]
+				}
+				atmp *= alpha
+				f64.AxpyUnitary(atmp, b[k*ldb:k*ldb+n], ctmp)
+			}
+		}
+		return
+	}
+	if isUpper {
+		for i := 0; i < m; i++ {
+			for j := n - 1; j >= 0; j-- {
+				tmp := alpha * b[i*ldb+j]
+				var tmp2 float64
+				atmp := a[j*lda+j+1 : j*lda+n]
+				btmp := b[i*ldb+j+1 : i*ldb+n]
+				ctmp := c[i*ldc+j+1 : i*ldc+n]
+				for k, v := range atmp {
+					ctmp[k] += tmp * v
+					tmp2 += btmp[k] * v
+				}
+				c[i*ldc+j] *= beta
+				c[i*ldc+j] += tmp*a[j*lda+j] + alpha*tmp2
+			}
+		}
+		return
+	}
+	for i := 0; i < m; i++ {
+		for j := 0; j < n; j++ {
+			tmp := alpha * b[i*ldb+j]
+			var tmp2 float64
+			atmp := a[j*lda : j*lda+j]
+			btmp := b[i*ldb : i*ldb+j]
+			ctmp := c[i*ldc : i*ldc+j]
+			for k, v := range atmp {
+				ctmp[k] += tmp * v
+				tmp2 += btmp[k] * v
+			}
+			c[i*ldc+j] *= beta
+			c[i*ldc+j] += tmp*a[j*lda+j] + alpha*tmp2
+		}
+	}
+}
+
+// Dsyrk performs one of the symmetric rank-k operations
+//  C = alpha * A * A^T + beta * C  if tA == blas.NoTrans
+//  C = alpha * A^T * A + beta * C  if tA == blas.Trans or tA == blas.ConjTrans
+// where A is an n×k or k×n matrix, C is an n×n symmetric matrix, and alpha and
+// beta are scalars.
+func (Implementation) Dsyrk(ul blas.Uplo, tA blas.Transpose, n, k int, alpha float64, a []float64, lda int, beta float64, c []float64, ldc int) {
+	if ul != blas.Lower && ul != blas.Upper {
+		panic(badUplo)
+	}
+	if tA != blas.Trans && tA != blas.NoTrans && tA != blas.ConjTrans {
+		panic(badTranspose)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	if k < 0 {
+		panic(kLT0)
+	}
+	row, col := k, n
+	if tA == blas.NoTrans {
+		row, col = n, k
+	}
+	if lda < max(1, col) {
+		panic(badLdA)
+	}
+	if ldc < max(1, n) {
+		panic(badLdC)
+	}
+
+	// Quick return if possible.
+	if n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if len(a) < lda*(row-1)+col {
+		panic(shortA)
+	}
+	if len(c) < ldc*(n-1)+n {
+		panic(shortC)
+	}
+
+	if alpha == 0 {
+		if beta == 0 {
+			if ul == blas.Upper {
+				for i := 0; i < n; i++ {
+					ctmp := c[i*ldc+i : i*ldc+n]
+					for j := range ctmp {
+						ctmp[j] = 0
+					}
+				}
+				return
+			}
+			for i := 0; i < n; i++ {
+				ctmp := c[i*ldc : i*ldc+i+1]
+				for j := range ctmp {
+					ctmp[j] = 0
+				}
+			}
+			return
+		}
+		if ul == blas.Upper {
+			for i := 0; i < n; i++ {
+				ctmp := c[i*ldc+i : i*ldc+n]
+				for j := range ctmp {
+					ctmp[j] *= beta
+				}
+			}
+			return
+		}
+		for i := 0; i < n; i++ {
+			ctmp := c[i*ldc : i*ldc+i+1]
+			for j := range ctmp {
+				ctmp[j] *= beta
+			}
+		}
+		return
+	}
+	if tA == blas.NoTrans {
+		if ul == blas.Upper {
+			for i := 0; i < n; i++ {
+				ctmp := c[i*ldc+i : i*ldc+n]
+				atmp := a[i*lda : i*lda+k]
+				if beta == 0 {
+					for jc := range ctmp {
+						j := jc + i
+						ctmp[jc] = alpha * f64.DotUnitary(atmp, a[j*lda:j*lda+k])
+					}
+				} else {
+					for jc, vc := range ctmp {
+						j := jc + i
+						ctmp[jc] = vc*beta + alpha*f64.DotUnitary(atmp, a[j*lda:j*lda+k])
+					}
+				}
+			}
+			return
+		}
+		for i := 0; i < n; i++ {
+			ctmp := c[i*ldc : i*ldc+i+1]
+			atmp := a[i*lda : i*lda+k]
+			if beta == 0 {
+				for j := range ctmp {
+					ctmp[j] = alpha * f64.DotUnitary(a[j*lda:j*lda+k], atmp)
+				}
+			} else {
+				for j, vc := range ctmp {
+					ctmp[j] = vc*beta + alpha*f64.DotUnitary(a[j*lda:j*lda+k], atmp)
+				}
+			}
+		}
+		return
+	}
+	// Cases where a is transposed.
+	if ul == blas.Upper {
+		for i := 0; i < n; i++ {
+			ctmp := c[i*ldc+i : i*ldc+n]
+			if beta == 0 {
+				for j := range ctmp {
+					ctmp[j] = 0
+				}
+			} else if beta != 1 {
+				for j := range ctmp {
+					ctmp[j] *= beta
+				}
+			}
+			for l := 0; l < k; l++ {
+				tmp := alpha * a[l*lda+i]
+				if tmp != 0 {
+					f64.AxpyUnitary(tmp, a[l*lda+i:l*lda+n], ctmp)
+				}
+			}
+		}
+		return
+	}
+	for i := 0; i < n; i++ {
+		ctmp := c[i*ldc : i*ldc+i+1]
+		if beta != 1 {
+			for j := range ctmp {
+				ctmp[j] *= beta
+			}
+		}
+		for l := 0; l < k; l++ {
+			tmp := alpha * a[l*lda+i]
+			if tmp != 0 {
+				f64.AxpyUnitary(tmp, a[l*lda:l*lda+i+1], ctmp)
+			}
+		}
+	}
+}
+
+// Dsyr2k performs one of the symmetric rank 2k operations
+//  C = alpha * A * B^T + alpha * B * A^T + beta * C  if tA == blas.NoTrans
+//  C = alpha * A^T * B + alpha * B^T * A + beta * C  if tA == blas.Trans or tA == blas.ConjTrans
+// where A and B are n×k or k×n matrices, C is an n×n symmetric matrix, and
+// alpha and beta are scalars.
+func (Implementation) Dsyr2k(ul blas.Uplo, tA blas.Transpose, n, k int, alpha float64, a []float64, lda int, b []float64, ldb int, beta float64, c []float64, ldc int) {
+	if ul != blas.Lower && ul != blas.Upper {
+		panic(badUplo)
+	}
+	if tA != blas.Trans && tA != blas.NoTrans && tA != blas.ConjTrans {
+		panic(badTranspose)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	if k < 0 {
+		panic(kLT0)
+	}
+	row, col := k, n
+	if tA == blas.NoTrans {
+		row, col = n, k
+	}
+	if lda < max(1, col) {
+		panic(badLdA)
+	}
+	if ldb < max(1, col) {
+		panic(badLdB)
+	}
+	if ldc < max(1, n) {
+		panic(badLdC)
+	}
+
+	// Quick return if possible.
+	if n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if len(a) < lda*(row-1)+col {
+		panic(shortA)
+	}
+	if len(b) < ldb*(row-1)+col {
+		panic(shortB)
+	}
+	if len(c) < ldc*(n-1)+n {
+		panic(shortC)
+	}
+
+	if alpha == 0 {
+		if beta == 0 {
+			if ul == blas.Upper {
+				for i := 0; i < n; i++ {
+					ctmp := c[i*ldc+i : i*ldc+n]
+					for j := range ctmp {
+						ctmp[j] = 0
+					}
+				}
+				return
+			}
+			for i := 0; i < n; i++ {
+				ctmp := c[i*ldc : i*ldc+i+1]
+				for j := range ctmp {
+					ctmp[j] = 0
+				}
+			}
+			return
+		}
+		if ul == blas.Upper {
+			for i := 0; i < n; i++ {
+				ctmp := c[i*ldc+i : i*ldc+n]
+				for j := range ctmp {
+					ctmp[j] *= beta
+				}
+			}
+			return
+		}
+		for i := 0; i < n; i++ {
+			ctmp := c[i*ldc : i*ldc+i+1]
+			for j := range ctmp {
+				ctmp[j] *= beta
+			}
+		}
+		return
+	}
+	if tA == blas.NoTrans {
+		if ul == blas.Upper {
+			for i := 0; i < n; i++ {
+				atmp := a[i*lda : i*lda+k]
+				btmp := b[i*ldb : i*ldb+k]
+				ctmp := c[i*ldc+i : i*ldc+n]
+				for jc := range ctmp {
+					j := i + jc
+					var tmp1, tmp2 float64
+					binner := b[j*ldb : j*ldb+k]
+					for l, v := range a[j*lda : j*lda+k] {
+						tmp1 += v * btmp[l]
+						tmp2 += atmp[l] * binner[l]
+					}
+					ctmp[jc] *= beta
+					ctmp[jc] += alpha * (tmp1 + tmp2)
+				}
+			}
+			return
+		}
+		for i := 0; i < n; i++ {
+			atmp := a[i*lda : i*lda+k]
+			btmp := b[i*ldb : i*ldb+k]
+			ctmp := c[i*ldc : i*ldc+i+1]
+			for j := 0; j <= i; j++ {
+				var tmp1, tmp2 float64
+				binner := b[j*ldb : j*ldb+k]
+				for l, v := range a[j*lda : j*lda+k] {
+					tmp1 += v * btmp[l]
+					tmp2 += atmp[l] * binner[l]
+				}
+				ctmp[j] *= beta
+				ctmp[j] += alpha * (tmp1 + tmp2)
+			}
+		}
+		return
+	}
+	if ul == blas.Upper {
+		for i := 0; i < n; i++ {
+			ctmp := c[i*ldc+i : i*ldc+n]
+			if beta != 1 {
+				for j := range ctmp {
+					ctmp[j] *= beta
+				}
+			}
+			for l := 0; l < k; l++ {
+				tmp1 := alpha * b[l*ldb+i]
+				tmp2 := alpha * a[l*lda+i]
+				btmp := b[l*ldb+i : l*ldb+n]
+				if tmp1 != 0 || tmp2 != 0 {
+					for j, v := range a[l*lda+i : l*lda+n] {
+						ctmp[j] += v*tmp1 + btmp[j]*tmp2
+					}
+				}
+			}
+		}
+		return
+	}
+	for i := 0; i < n; i++ {
+		ctmp := c[i*ldc : i*ldc+i+1]
+		if beta != 1 {
+			for j := range ctmp {
+				ctmp[j] *= beta
+			}
+		}
+		for l := 0; l < k; l++ {
+			tmp1 := alpha * b[l*ldb+i]
+			tmp2 := alpha * a[l*lda+i]
+			btmp := b[l*ldb : l*ldb+i+1]
+			if tmp1 != 0 || tmp2 != 0 {
+				for j, v := range a[l*lda : l*lda+i+1] {
+					ctmp[j] += v*tmp1 + btmp[j]*tmp2
+				}
+			}
+		}
+	}
+}
+
+// Dtrmm performs one of the matrix-matrix operations
+//  B = alpha * A * B    if tA == blas.NoTrans and side == blas.Left
+//  B = alpha * A^T * B  if tA == blas.Trans or blas.ConjTrans, and side == blas.Left
+//  B = alpha * B * A    if tA == blas.NoTrans and side == blas.Right
+//  B = alpha * B * A^T  if tA == blas.Trans or blas.ConjTrans, and side == blas.Right
+// where A is an n×n or m×m triangular matrix, B is an m×n matrix, and alpha is a scalar.
+func (Implementation) Dtrmm(s blas.Side, ul blas.Uplo, tA blas.Transpose, d blas.Diag, m, n int, alpha float64, a []float64, lda int, b []float64, ldb int) {
+	if s != blas.Left && s != blas.Right {
+		panic(badSide)
+	}
+	if ul != blas.Lower && ul != blas.Upper {
+		panic(badUplo)
+	}
+	if tA != blas.NoTrans && tA != blas.Trans && tA != blas.ConjTrans {
+		panic(badTranspose)
+	}
+	if d != blas.NonUnit && d != blas.Unit {
+		panic(badDiag)
+	}
+	if m < 0 {
+		panic(mLT0)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	k := n
+	if s == blas.Left {
+		k = m
+	}
+	if lda < max(1, k) {
+		panic(badLdA)
+	}
+	if ldb < max(1, n) {
+		panic(badLdB)
+	}
+
+	// Quick return if possible.
+	if m == 0 || n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if len(a) < lda*(k-1)+k {
+		panic(shortA)
+	}
+	if len(b) < ldb*(m-1)+n {
+		panic(shortB)
+	}
+
+	if alpha == 0 {
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			for j := range btmp {
+				btmp[j] = 0
+			}
+		}
+		return
+	}
+
+	nonUnit := d == blas.NonUnit
+	if s == blas.Left {
+		if tA == blas.NoTrans {
+			if ul == blas.Upper {
+				for i := 0; i < m; i++ {
+					tmp := alpha
+					if nonUnit {
+						tmp *= a[i*lda+i]
+					}
+					btmp := b[i*ldb : i*ldb+n]
+					f64.ScalUnitary(tmp, btmp)
+					for ka, va := range a[i*lda+i+1 : i*lda+m] {
+						k := ka + i + 1
+						if va != 0 {
+							f64.AxpyUnitary(alpha*va, b[k*ldb:k*ldb+n], btmp)
+						}
+					}
+				}
+				return
+			}
+			for i := m - 1; i >= 0; i-- {
+				tmp := alpha
+				if nonUnit {
+					tmp *= a[i*lda+i]
+				}
+				btmp := b[i*ldb : i*ldb+n]
+				f64.ScalUnitary(tmp, btmp)
+				for k, va := range a[i*lda : i*lda+i] {
+					if va != 0 {
+						f64.AxpyUnitary(alpha*va, b[k*ldb:k*ldb+n], btmp)
+					}
+				}
+			}
+			return
+		}
+		// Cases where a is transposed.
+		if ul == blas.Upper {
+			for k := m - 1; k >= 0; k-- {
+				btmpk := b[k*ldb : k*ldb+n]
+				for ia, va := range a[k*lda+k+1 : k*lda+m] {
+					i := ia + k + 1
+					btmp := b[i*ldb : i*ldb+n]
+					if va != 0 {
+						f64.AxpyUnitary(alpha*va, btmpk, btmp)
+					}
+				}
+				tmp := alpha
+				if nonUnit {
+					tmp *= a[k*lda+k]
+				}
+				if tmp != 1 {
+					f64.ScalUnitary(tmp, btmpk)
+				}
+			}
+			return
+		}
+		for k := 0; k < m; k++ {
+			btmpk := b[k*ldb : k*ldb+n]
+			for i, va := range a[k*lda : k*lda+k] {
+				btmp := b[i*ldb : i*ldb+n]
+				if va != 0 {
+					f64.AxpyUnitary(alpha*va, btmpk, btmp)
+				}
+			}
+			tmp := alpha
+			if nonUnit {
+				tmp *= a[k*lda+k]
+			}
+			if tmp != 1 {
+				f64.ScalUnitary(tmp, btmpk)
+			}
+		}
+		return
+	}
+	// Cases where a is on the right
+	if tA == blas.NoTrans {
+		if ul == blas.Upper {
+			for i := 0; i < m; i++ {
+				btmp := b[i*ldb : i*ldb+n]
+				for k := n - 1; k >= 0; k-- {
+					tmp := alpha * btmp[k]
+					if tmp == 0 {
+						continue
+					}
+					btmp[k] = tmp
+					if nonUnit {
+						btmp[k] *= a[k*lda+k]
+					}
+					f64.AxpyUnitary(tmp, a[k*lda+k+1:k*lda+n], btmp[k+1:n])
+				}
+			}
+			return
+		}
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			for k := 0; k < n; k++ {
+				tmp := alpha * btmp[k]
+				if tmp == 0 {
+					continue
+				}
+				btmp[k] = tmp
+				if nonUnit {
+					btmp[k] *= a[k*lda+k]
+				}
+				f64.AxpyUnitary(tmp, a[k*lda:k*lda+k], btmp[:k])
+			}
+		}
+		return
+	}
+	// Cases where a is transposed.
+	if ul == blas.Upper {
+		for i := 0; i < m; i++ {
+			btmp := b[i*ldb : i*ldb+n]
+			for j, vb := range btmp {
+				tmp := vb
+				if nonUnit {
+					tmp *= a[j*lda+j]
+				}
+				tmp += f64.DotUnitary(a[j*lda+j+1:j*lda+n], btmp[j+1:n])
+				btmp[j] = alpha * tmp
+			}
+		}
+		return
+	}
+	for i := 0; i < m; i++ {
+		btmp := b[i*ldb : i*ldb+n]
+		for j := n - 1; j >= 0; j-- {
+			tmp := btmp[j]
+			if nonUnit {
+				tmp *= a[j*lda+j]
+			}
+			tmp += f64.DotUnitary(a[j*lda:j*lda+j], btmp[:j])
+			btmp[j] = alpha * tmp
+		}
+	}
+}

vendor/gonum.org/v1/gonum/blas/gonum/sgemm.go

@@ -0,0 +1,318 @@
+// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.
+
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package gonum
+
+import (
+	"runtime"
+	"sync"
+
+	"gonum.org/v1/gonum/blas"
+	"gonum.org/v1/gonum/internal/asm/f32"
+)
+
+// Sgemm performs one of the matrix-matrix operations
+//  C = alpha * A * B + beta * C
+//  C = alpha * A^T * B + beta * C
+//  C = alpha * A * B^T + beta * C
+//  C = alpha * A^T * B^T + beta * C
+// where A is an m×k or k×m dense matrix, B is an n×k or k×n dense matrix, C is
+// an m×n matrix, and alpha and beta are scalars. tA and tB specify whether A or
+// B are transposed.
+//
+// Float32 implementations are autogenerated and not directly tested.
+func (Implementation) Sgemm(tA, tB blas.Transpose, m, n, k int, alpha float32, a []float32, lda int, b []float32, ldb int, beta float32, c []float32, ldc int) {
+	switch tA {
+	default:
+		panic(badTranspose)
+	case blas.NoTrans, blas.Trans, blas.ConjTrans:
+	}
+	switch tB {
+	default:
+		panic(badTranspose)
+	case blas.NoTrans, blas.Trans, blas.ConjTrans:
+	}
+	if m < 0 {
+		panic(mLT0)
+	}
+	if n < 0 {
+		panic(nLT0)
+	}
+	if k < 0 {
+		panic(kLT0)
+	}
+	aTrans := tA == blas.Trans || tA == blas.ConjTrans
+	if aTrans {
+		if lda < max(1, m) {
+			panic(badLdA)
+		}
+	} else {
+		if lda < max(1, k) {
+			panic(badLdA)
+		}
+	}
+	bTrans := tB == blas.Trans || tB == blas.ConjTrans
+	if bTrans {
+		if ldb < max(1, k) {
+			panic(badLdB)
+		}
+	} else {
+		if ldb < max(1, n) {
+			panic(badLdB)
+		}
+	}
+	if ldc < max(1, n) {
+		panic(badLdC)
+	}
+
+	// Quick return if possible.
+	if m == 0 || n == 0 {
+		return
+	}
+
+	// For zero matrix size the following slice length checks are trivially satisfied.
+	if aTrans {
+		if len(a) < (k-1)*lda+m {
+			panic(shortA)
+		}
+	} else {
+		if len(a) < (m-1)*lda+k {
+			panic(shortA)
+		}
+	}
+	if bTrans {
+		if len(b) < (n-1)*ldb+k {
+			panic(shortB)
+		}
+	} else {
+		if len(b) < (k-1)*ldb+n {
+			panic(shortB)
+		}
+	}
+	if len(c) < (m-1)*ldc+n {
+		panic(shortC)
+	}
+
+	// Quick return if possible.
+	if (alpha == 0 || k == 0) && beta == 1 {
+		return
+	}
+
+	// scale c
+	if beta != 1 {
+		if beta == 0 {
+			for i := 0; i < m; i++ {
+				ctmp := c[i*ldc : i*ldc+n]
+				for j := range ctmp {
+					ctmp[j] = 0
+				}
+			}
+		} else {
+			for i := 0; i < m; i++ {
+				ctmp := c[i*ldc : i*ldc+n]
+				for j := range ctmp {
+					ctmp[j] *= beta
+				}
+			}
+		}
+	}
+
+	sgemmParallel(aTrans, bTrans, m, n, k, a, lda, b, ldb, c, ldc, alpha)
+}
+
+func sgemmParallel(aTrans, bTrans bool, m, n, k int, a []float32, lda int, b []float32, ldb int, c []float32, ldc int, alpha float32) {
+	// dgemmParallel computes a parallel matrix multiplication by partitioning
+	// a and b into sub-blocks, and updating c with the multiplication of the sub-block
+	// In all cases,
+	// A = [ 	A_11	A_12 ... 	A_1j
+	//			A_21	A_22 ...	A_2j
+	//				...
+	//			A_i1	A_i2 ...	A_ij]
+	//
+	// and same for B. All of the submatrix sizes are blockSize×blockSize except
+	// at the edges.
+	//
+	// In all cases, there is one dimension for each matrix along which
+	// C must be updated sequentially.
+	// Cij = \sum_k Aik Bki,	(A * B)
+	// Cij = \sum_k Aki Bkj,	(A^T * B)
+	// Cij = \sum_k Aik Bjk,	(A * B^T)
+	// Cij = \sum_k Aki Bjk,	(A^T * B^T)
+	//
+	// This code computes one {i, j} block sequentially along the k dimension,
+	// and computes all of the {i, j} blocks concurrently. This
+	// partitioning allows Cij to be updated in-place without race-conditions.
+	// Instead of launching a goroutine for each possible concurrent computation,
+	// a number of worker goroutines are created and channels are used to pass
+	// available and completed cases.
+	//
+	// http://alexkr.com/docs/matrixmult.pdf is a good reference on matrix-matrix
+	// multiplies, though this code does not copy matrices to attempt to eliminate
+	// cache misses.
+
+	maxKLen := k
+	parBlocks := blocks(m, blockSize) * blocks(n, blockSize)
+	if parBlocks < minParBlock {
+		// The matrix multiplication is small in the dimensions where it can be
+		// computed concurrently. Just do it in serial.
+		sgemmSerial(aTrans, bTrans, m, n, k, a, lda, b, ldb, c, ldc, alpha)
+		return
+	}
+
+	nWorkers := runtime.GOMAXPROCS(0)
+	if parBlocks < nWorkers {
+		nWorkers = parBlocks
+	}
+	// There is a tradeoff between the workers having to wait for work
+	// and a large buffer making operations slow.
+	buf := buffMul * nWorkers
+	if buf > parBlocks {
+		buf = parBlocks
+	}
+
+	sendChan := make(chan subMul, buf)
+
+	// Launch workers. A worker receives an {i, j} submatrix of c, and computes
+	// A_ik B_ki (or the transposed version) storing the result in c_ij. When the
+	// channel is finally closed, it signals to the waitgroup that it has finished
+	// computing.
+	var wg sync.WaitGroup
+	for i := 0; i < nWorkers; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for sub := range sendChan {
+				i := sub.i
+				j := sub.j
+				leni := blockSize
+				if i+leni > m {
+					leni = m - i
+				}
+				lenj := blockSize
+				if j+lenj > n {
+					lenj = n - j
+				}
+
+				cSub := sliceView32(c, ldc, i, j, leni, lenj)
+
+				// Compute A_ik B_kj for all k
+				for k := 0; k < maxKLen; k += blockSize {
+					lenk := blockSize
+					if k+lenk > maxKLen {
+						lenk = maxKLen - k
+					}
+					var aSub, bSub []float32
+					if aTrans {
+						aSub = sliceView32(a, lda, k, i, lenk, leni)
+					} else {
+						aSub = sliceView32(a, lda, i, k, leni, lenk)
+					}
+					if bTrans {
+						bSub = sliceView32(b, ldb, j, k, lenj, lenk)
+					} else {
+						bSub = sliceView32(b, ldb, k, j, lenk, lenj)
+					}
+					sgemmSerial(aTrans, bTrans, leni, lenj, lenk, aSub, lda, bSub, ldb, cSub, ldc, alpha)
+				}
+			}
+		}()
+	}
+
+	// Send out all of the {i, j} subblocks for computation.
+	for i := 0; i < m; i += blockSize {
+		for j := 0; j < n; j += blockSize {
+			sendChan <- subMul{
+				i: i,
+				j: j,
+			}
+		}
+	}
+	close(sendChan)
+	wg.Wait()
+}
+
+// sgemmSerial is serial matrix multiply
+func sgemmSerial(aTrans, bTrans bool, m, n, k int, a []float32, lda int, b []float32, ldb int, c []float32, ldc int, alpha float32) {
+	switch {
+	case !aTrans && !bTrans:
+		sgemmSerialNotNot(m, n, k, a, lda, b, ldb, c, ldc, alpha)
+		return
+	case aTrans && !bTrans:
+		sgemmSerialTransNot(m, n, k, a, lda, b, ldb, c, ldc, alpha)
+		return
+	case !aTrans && bTrans:
+		sgemmSerialNotTrans(m, n, k, a, lda, b, ldb, c, ldc, alpha)
+		return
+	case aTrans && bTrans:
+		sgemmSerialTransTrans(m, n, k, a, lda, b, ldb, c, ldc, alpha)
+		return
+	default:
+		panic("unreachable")
+	}
+}
+
+// sgemmSerial where neither a nor b are transposed
+func sgemmSerialNotNot(m, n, k int, a []float32, lda int, b []float32, ldb int, c []float32, ldc int, alpha float32) {
+	// This style is used instead of the literal [i*stride +j]) is used because
+	// approximately 5 times faster as of go 1.3.
+	for i := 0; i < m; i++ {
+		ctmp := c[i*ldc : i*ldc+n]
+		for l, v := range a[i*lda : i*lda+k] {
+			tmp := alpha * v
+			if tmp != 0 {
+				f32.AxpyUnitary(tmp, b[l*ldb:l*ldb+n], ctmp)
+			}
+		}
+	}
+}
+
+// sgemmSerial where neither a is transposed and b is not
+func sgemmSerialTransNot(m, n, k int, a []float32, lda int, b []float32, ldb int, c []float32, ldc int, alpha float32) {
+	// This style is used instead of the literal [i*stride +j]) is used because
+	// approximately 5 times faster as of go 1.3.
+	for l := 0; l < k; l++ {
+		btmp := b[l*ldb : l*ldb+n]
+		for i, v := range a[l*lda : l*lda+m] {
+			tmp := alpha * v
+			if tmp != 0 {
+				ctmp := c[i*ldc : i*ldc+n]
+				f32.AxpyUnitary(tmp, btmp, ctmp)
+			}
+		}
+	}
+}
+
+// sgemmSerial where neither a is not transposed and b is
+func sgemmSerialNotTrans(m, n, k int, a []float32, lda int, b []float32, ldb int, c []float32, ldc int, alpha float32) {
+	// This style is used instead of the literal [i*stride +j]) is used because
+	// approximately 5 times faster as of go 1.3.
+	for i := 0; i < m; i++ {
+		atmp := a[i*lda : i*lda+k]
+		ctmp := c[i*ldc : i*ldc+n]
+		for j := 0; j < n; j++ {
+			ctmp[j] += alpha * f32.DotUnitary(atmp, b[j*ldb:j*ldb+k])
+		}
+	}
+}
+
+// sgemmSerial where both are transposed
+func sgemmSerialTransTrans(m, n, k int, a []float32, lda int, b []float32, ldb int, c []float32, ldc int, alpha float32) {
+	// This style is used instead of the literal [i*stride +j]) is used because
+	// approximately 5 times faster as of go 1.3.
+	for l := 0; l < k; l++ {
+		for i, v := range a[l*lda : l*lda+m] {
+			tmp := alpha * v
+			if tmp != 0 {
+				ctmp := c[i*ldc : i*ldc+n]
+				f32.AxpyInc(tmp, b[l:], ctmp, uintptr(n), uintptr(ldb), 1, 0, 0)
+			}
+		}
+	}
+}
+
+func sliceView32(a []float32, lda, i, j, r, c int) []float32 {
+	return a[i*lda+j : (i+r-1)*lda+j+c]
+}

vendor/gonum.org/v1/gonum/blas/gonum/single_precision.bash

@@ -0,0 +1,218 @@
+#!/usr/bin/env bash
+
+# Copyright ©2015 The Gonum Authors. All rights reserved.
+# Use of this source code is governed by a BSD-style
+# license that can be found in the LICENSE file.
+
+WARNINGF32='//\
+// Float32 implementations are autogenerated and not directly tested.\
+'
+WARNINGC64='//\
+// Complex64 implementations are autogenerated and not directly tested.\
+'
+
+# Level1 routines.
+
+echo Generating level1float32.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.\n' > level1float32.go
+cat level1float64.go \
+| gofmt -r 'blas.Float64Level1 -> blas.Float32Level1' \
+\
+| gofmt -r 'float64 -> float32' \
+| gofmt -r 'blas.DrotmParams -> blas.SrotmParams' \
+\
+| gofmt -r 'f64.AxpyInc -> f32.AxpyInc' \
+| gofmt -r 'f64.AxpyUnitary -> f32.AxpyUnitary' \
+| gofmt -r 'f64.DotUnitary -> f32.DotUnitary' \
+| gofmt -r 'f64.ScalInc -> f32.ScalInc' \
+| gofmt -r 'f64.ScalUnitary -> f32.ScalUnitary' \
+\
+| sed -e "s_^\(func (Implementation) \)D\(.*\)\$_$WARNINGF32\1S\2_" \
+      -e 's_^// D_// S_' \
+      -e "s_^\(func (Implementation) \)Id\(.*\)\$_$WARNINGF32\1Is\2_" \
+      -e 's_^// Id_// Is_' \
+      -e 's_"gonum.org/v1/gonum/internal/asm/f64"_"gonum.org/v1/gonum/internal/asm/f32"_' \
+      -e 's_"math"_math "gonum.org/v1/gonum/internal/math32"_' \
+>> level1float32.go
+
+echo Generating level1cmplx64.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.\n' > level1cmplx64.go
+cat level1cmplx128.go \
+| gofmt -r 'blas.Complex128Level1 -> blas.Complex64Level1' \
+\
+| gofmt -r 'float64 -> float32' \
+| gofmt -r 'complex128 -> complex64' \
+\
+| gofmt -r 'c128.AxpyInc -> c64.AxpyInc' \
+| gofmt -r 'c128.AxpyUnitary -> c64.AxpyUnitary' \
+| gofmt -r 'c128.DotcInc -> c64.DotcInc' \
+| gofmt -r 'c128.DotcUnitary -> c64.DotcUnitary' \
+| gofmt -r 'c128.DotuInc -> c64.DotuInc' \
+| gofmt -r 'c128.DotuUnitary -> c64.DotuUnitary' \
+| gofmt -r 'c128.ScalInc -> c64.ScalInc' \
+| gofmt -r 'c128.ScalUnitary -> c64.ScalUnitary' \
+| gofmt -r 'dcabs1 -> scabs1' \
+\
+| sed -e "s_^\(func (Implementation) \)Zdot\(.*\)\$_$WARNINGC64\1Cdot\2_" \
+      -e 's_^// Zdot_// Cdot_' \
+      -e "s_^\(func (Implementation) \)Zdscal\(.*\)\$_$WARNINGC64\1Csscal\2_" \
+      -e 's_^// Zdscal_// Csscal_' \
+      -e "s_^\(func (Implementation) \)Z\(.*\)\$_$WARNINGC64\1C\2_" \
+      -e 's_^// Z_// C_' \
+      -e "s_^\(func (Implementation) \)Iz\(.*\)\$_$WARNINGC64\1Ic\2_" \
+      -e 's_^// Iz_// Ic_' \
+      -e "s_^\(func (Implementation) \)Dz\(.*\)\$_$WARNINGC64\1Sc\2_" \
+      -e 's_^// Dz_// Sc_' \
+      -e 's_"gonum.org/v1/gonum/internal/asm/c128"_"gonum.org/v1/gonum/internal/asm/c64"_' \
+      -e 's_"math"_math "gonum.org/v1/gonum/internal/math32"_' \
+>> level1cmplx64.go
+
+echo Generating level1float32_sdot.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.\n' > level1float32_sdot.go
+cat level1float64_ddot.go \
+| gofmt -r 'float64 -> float32' \
+\
+| gofmt -r 'f64.DotInc -> f32.DotInc' \
+| gofmt -r 'f64.DotUnitary -> f32.DotUnitary' \
+\
+| sed -e "s_^\(func (Implementation) \)D\(.*\)\$_$WARNINGF32\1S\2_" \
+      -e 's_^// D_// S_' \
+      -e 's_"gonum.org/v1/gonum/internal/asm/f64"_"gonum.org/v1/gonum/internal/asm/f32"_' \
+>> level1float32_sdot.go
+
+echo Generating level1float32_dsdot.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.\n' > level1float32_dsdot.go
+cat level1float64_ddot.go \
+| gofmt -r '[]float64 -> []float32' \
+\
+| gofmt -r 'f64.DotInc -> f32.DdotInc' \
+| gofmt -r 'f64.DotUnitary -> f32.DdotUnitary' \
+\
+| sed -e "s_^\(func (Implementation) \)D\(.*\)\$_$WARNINGF32\1Ds\2_" \
+      -e 's_^// D_// Ds_' \
+      -e 's_"gonum.org/v1/gonum/internal/asm/f64"_"gonum.org/v1/gonum/internal/asm/f32"_' \
+>> level1float32_dsdot.go
+
+echo Generating level1float32_sdsdot.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.\n' > level1float32_sdsdot.go
+cat level1float64_ddot.go \
+| gofmt -r 'float64 -> float32' \
+\
+| gofmt -r 'f64.DotInc(x, y, f(n), f(incX), f(incY), f(ix), f(iy)) -> alpha + float32(f32.DdotInc(x, y, f(n), f(incX), f(incY), f(ix), f(iy)))' \
+| gofmt -r 'f64.DotUnitary(a, b) -> alpha + float32(f32.DdotUnitary(a, b))' \
+\
+| sed -e "s_^\(func (Implementation) \)D\(.*\)\$_$WARNINGF32\1Sds\2_" \
+      -e 's_^// D\(.*\)$_// Sds\1 plus a constant_' \
+      -e 's_\\sum_alpha + \\sum_' \
+      -e 's/n int/n int, alpha float32/' \
+      -e 's_"gonum.org/v1/gonum/internal/asm/f64"_"gonum.org/v1/gonum/internal/asm/f32"_' \
+>> level1float32_sdsdot.go
+
+
+# Level2 routines.
+
+echo Generating level2float32.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.\n' > level2float32.go
+cat level2float64.go \
+| gofmt -r 'blas.Float64Level2 -> blas.Float32Level2' \
+\
+| gofmt -r 'float64 -> float32' \
+\
+| gofmt -r 'f64.AxpyInc -> f32.AxpyInc' \
+| gofmt -r 'f64.AxpyIncTo -> f32.AxpyIncTo' \
+| gofmt -r 'f64.AxpyUnitary -> f32.AxpyUnitary' \
+| gofmt -r 'f64.AxpyUnitaryTo -> f32.AxpyUnitaryTo' \
+| gofmt -r 'f64.DotInc -> f32.DotInc' \
+| gofmt -r 'f64.DotUnitary -> f32.DotUnitary' \
+| gofmt -r 'f64.ScalInc -> f32.ScalInc' \
+| gofmt -r 'f64.ScalUnitary -> f32.ScalUnitary' \
+| gofmt -r 'f64.Ger -> f32.Ger' \
+\
+| sed -e "s_^\(func (Implementation) \)D\(.*\)\$_$WARNINGF32\1S\2_" \
+      -e 's_^// D_// S_' \
+      -e 's_"gonum.org/v1/gonum/internal/asm/f64"_"gonum.org/v1/gonum/internal/asm/f32"_' \
+>> level2float32.go
+
+echo Generating level2cmplx64.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.\n' > level2cmplx64.go
+cat level2cmplx128.go \
+| gofmt -r 'blas.Complex128Level2 -> blas.Complex64Level2' \
+\
+| gofmt -r 'complex128 -> complex64' \
+| gofmt -r 'float64 -> float32' \
+\
+| gofmt -r 'c128.AxpyInc -> c64.AxpyInc' \
+| gofmt -r 'c128.AxpyUnitary -> c64.AxpyUnitary' \
+| gofmt -r 'c128.DotuInc -> c64.DotuInc' \
+| gofmt -r 'c128.DotuUnitary -> c64.DotuUnitary' \
+| gofmt -r 'c128.ScalInc -> c64.ScalInc' \
+| gofmt -r 'c128.ScalUnitary -> c64.ScalUnitary' \
+\
+| sed -e "s_^\(func (Implementation) \)Z\(.*\)\$_$WARNINGC64\1C\2_" \
+      -e 's_^// Z_// C_' \
+      -e 's_"gonum.org/v1/gonum/internal/asm/c128"_"gonum.org/v1/gonum/internal/asm/c64"_' \
+      -e 's_"math/cmplx"_cmplx "gonum.org/v1/gonum/internal/cmplx64"_' \
+>> level2cmplx64.go
+
+# Level3 routines.
+
+echo Generating level3float32.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.\n' > level3float32.go
+cat level3float64.go \
+| gofmt -r 'blas.Float64Level3 -> blas.Float32Level3' \
+\
+| gofmt -r 'float64 -> float32' \
+\
+| gofmt -r 'f64.AxpyUnitaryTo -> f32.AxpyUnitaryTo' \
+| gofmt -r 'f64.AxpyUnitary -> f32.AxpyUnitary' \
+| gofmt -r 'f64.DotUnitary -> f32.DotUnitary' \
+| gofmt -r 'f64.ScalUnitary -> f32.ScalUnitary' \
+\
+| sed -e "s_^\(func (Implementation) \)D\(.*\)\$_$WARNINGF32\1S\2_" \
+      -e 's_^// D_// S_' \
+      -e 's_"gonum.org/v1/gonum/internal/asm/f64"_"gonum.org/v1/gonum/internal/asm/f32"_' \
+>> level3float32.go
+
+echo Generating sgemm.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.\n' > sgemm.go
+cat dgemm.go \
+| gofmt -r 'float64 -> float32' \
+| gofmt -r 'sliceView64 -> sliceView32' \
+\
+| gofmt -r 'dgemmParallel -> sgemmParallel' \
+| gofmt -r 'computeNumBlocks64 -> computeNumBlocks32' \
+| gofmt -r 'dgemmSerial -> sgemmSerial' \
+| gofmt -r 'dgemmSerialNotNot -> sgemmSerialNotNot' \
+| gofmt -r 'dgemmSerialTransNot -> sgemmSerialTransNot' \
+| gofmt -r 'dgemmSerialNotTrans -> sgemmSerialNotTrans' \
+| gofmt -r 'dgemmSerialTransTrans -> sgemmSerialTransTrans' \
+\
+| gofmt -r 'f64.AxpyInc -> f32.AxpyInc' \
+| gofmt -r 'f64.AxpyUnitary -> f32.AxpyUnitary' \
+| gofmt -r 'f64.DotUnitary -> f32.DotUnitary' \
+\
+| sed -e "s_^\(func (Implementation) \)D\(.*\)\$_$WARNINGF32\1S\2_" \
+      -e 's_^// D_// S_' \
+      -e 's_^// d_// s_' \
+      -e 's_"gonum.org/v1/gonum/internal/asm/f64"_"gonum.org/v1/gonum/internal/asm/f32"_' \
+>> sgemm.go
+
+echo Generating level3cmplx64.go
+echo -e '// Code generated by "go generate gonum.org/v1/gonum/blas/gonum”; DO NOT EDIT.\n' > level3cmplx64.go
+cat level3cmplx128.go \
+| gofmt -r 'blas.Complex128Level3 -> blas.Complex64Level3' \
+\
+| gofmt -r 'float64 -> float32' \
+| gofmt -r 'complex128 -> complex64' \
+\
+| gofmt -r 'c128.ScalUnitary -> c64.ScalUnitary' \
+| gofmt -r 'c128.DscalUnitary -> c64.SscalUnitary' \
+| gofmt -r 'c128.DotcUnitary -> c64.DotcUnitary' \
+| gofmt -r 'c128.AxpyUnitary -> c64.AxpyUnitary' \
+| gofmt -r 'c128.DotuUnitary -> c64.DotuUnitary' \
+\
+| sed -e "s_^\(func (Implementation) \)Z\(.*\)\$_$WARNINGC64\1C\2_" \
+      -e 's_^// Z_// C_' \
+      -e 's_"gonum.org/v1/gonum/internal/asm/c128"_"gonum.org/v1/gonum/internal/asm/c64"_' \
+      -e 's_"math/cmplx"_cmplx "gonum.org/v1/gonum/internal/cmplx64"_' \
+>> level3cmplx64.go

vendor/gonum.org/v1/gonum/floats/README.md

@@ -0,0 +1,4 @@
+# Gonum floats [![GoDoc](https://godoc.org/gonum.org/v1/gonum/floats?status.svg)](https://godoc.org/gonum.org/v1/gonum/floats)
+
+Package floats provides a set of helper routines for dealing with slices of float64.
+The functions avoid allocations to allow for use within tight loops without garbage collection overhead.

vendor/gonum.org/v1/gonum/floats/doc.go

@@ -0,0 +1,11 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package floats provides a set of helper routines for dealing with slices
+// of float64. The functions avoid allocations to allow for use within tight
+// loops without garbage collection overhead.
+//
+// The convention used is that when a slice is being modified in place, it has
+// the name dst.
+package floats // import "gonum.org/v1/gonum/floats"

vendor/gonum.org/v1/gonum/floats/floats.go

@@ -0,0 +1,933 @@
+// Copyright ©2013 The Gonum Authors. All rights reserved.
+// Use of this code is governed by a BSD-style
+// license that can be found in the LICENSE file
+
+package floats
+
+import (
+	"errors"
+	"math"
+	"sort"
+	"strconv"
+
+	"gonum.org/v1/gonum/internal/asm/f64"
+)
+
+// Add adds, element-wise, the elements of s and dst, and stores in dst.
+// Panics if the lengths of dst and s do not match.
+func Add(dst, s []float64) {
+	if len(dst) != len(s) {
+		panic("floats: length of the slices do not match")
+	}
+	f64.AxpyUnitaryTo(dst, 1, s, dst)
+}
+
+// AddTo adds, element-wise, the elements of s and t and
+// stores the result in dst. Panics if the lengths of s, t and dst do not match.
+func AddTo(dst, s, t []float64) []float64 {
+	if len(s) != len(t) {
+		panic("floats: length of adders do not match")
+	}
+	if len(dst) != len(s) {
+		panic("floats: length of destination does not match length of adder")
+	}
+	f64.AxpyUnitaryTo(dst, 1, s, t)
+	return dst
+}
+
+// AddConst adds the scalar c to all of the values in dst.
+func AddConst(c float64, dst []float64) {
+	f64.AddConst(c, dst)
+}
+
+// AddScaled performs dst = dst + alpha * s.
+// It panics if the lengths of dst and s are not equal.
+func AddScaled(dst []float64, alpha float64, s []float64) {
+	if len(dst) != len(s) {
+		panic("floats: length of destination and source to not match")
+	}
+	f64.AxpyUnitaryTo(dst, alpha, s, dst)
+}
+
+// AddScaledTo performs dst = y + alpha * s, where alpha is a scalar,
+// and dst, y and s are all slices.
+// It panics if the lengths of dst, y, and s are not equal.
+//
+// At the return of the function, dst[i] = y[i] + alpha * s[i]
+func AddScaledTo(dst, y []float64, alpha float64, s []float64) []float64 {
+	if len(dst) != len(s) || len(dst) != len(y) {
+		panic("floats: lengths of slices do not match")
+	}
+	f64.AxpyUnitaryTo(dst, alpha, s, y)
+	return dst
+}
+
+// argsort is a helper that implements sort.Interface, as used by
+// Argsort.
+type argsort struct {
+	s    []float64
+	inds []int
+}
+
+func (a argsort) Len() int {
+	return len(a.s)
+}
+
+func (a argsort) Less(i, j int) bool {
+	return a.s[i] < a.s[j]
+}
+
+func (a argsort) Swap(i, j int) {
+	a.s[i], a.s[j] = a.s[j], a.s[i]
+	a.inds[i], a.inds[j] = a.inds[j], a.inds[i]
+}
+
+// Argsort sorts the elements of dst while tracking their original order.
+// At the conclusion of Argsort, dst will contain the original elements of dst
+// but sorted in increasing order, and inds will contain the original position
+// of the elements in the slice such that dst[i] = origDst[inds[i]].
+// It panics if the lengths of dst and inds do not match.
+func Argsort(dst []float64, inds []int) {
+	if len(dst) != len(inds) {
+		panic("floats: length of inds does not match length of slice")
+	}
+	for i := range dst {
+		inds[i] = i
+	}
+
+	a := argsort{s: dst, inds: inds}
+	sort.Sort(a)
+}
+
+// Count applies the function f to every element of s and returns the number
+// of times the function returned true.
+func Count(f func(float64) bool, s []float64) int {
+	var n int
+	for _, val := range s {
+		if f(val) {
+			n++
+		}
+	}
+	return n
+}
+
+// CumProd finds the cumulative product of the first i elements in
+// s and puts them in place into the ith element of the
+// destination dst. A panic will occur if the lengths of arguments
+// do not match.
+//
+// At the return of the function, dst[i] = s[i] * s[i-1] * s[i-2] * ...
+func CumProd(dst, s []float64) []float64 {
+	if len(dst) != len(s) {
+		panic("floats: length of destination does not match length of the source")
+	}
+	if len(dst) == 0 {
+		return dst
+	}
+	return f64.CumProd(dst, s)
+}
+
+// CumSum finds the cumulative sum of the first i elements in
+// s and puts them in place into the ith element of the
+// destination dst. A panic will occur if the lengths of arguments
+// do not match.
+//
+// At the return of the function, dst[i] = s[i] + s[i-1] + s[i-2] + ...
+func CumSum(dst, s []float64) []float64 {
+	if len(dst) != len(s) {
+		panic("floats: length of destination does not match length of the source")
+	}
+	if len(dst) == 0 {
+		return dst
+	}
+	return f64.CumSum(dst, s)
+}
+
+// Distance computes the L-norm of s - t. See Norm for special cases.
+// A panic will occur if the lengths of s and t do not match.
+func Distance(s, t []float64, L float64) float64 {
+	if len(s) != len(t) {
+		panic("floats: slice lengths do not match")
+	}
+	if len(s) == 0 {
+		return 0
+	}
+	var norm float64
+	if L == 2 {
+		for i, v := range s {
+			diff := t[i] - v
+			norm = math.Hypot(norm, diff)
+		}
+		return norm
+	}
+	if L == 1 {
+		for i, v := range s {
+			norm += math.Abs(t[i] - v)
+		}
+		return norm
+	}
+	if math.IsInf(L, 1) {
+		for i, v := range s {
+			absDiff := math.Abs(t[i] - v)
+			if absDiff > norm {
+				norm = absDiff
+			}
+		}
+		return norm
+	}
+	for i, v := range s {
+		norm += math.Pow(math.Abs(t[i]-v), L)
+	}
+	return math.Pow(norm, 1/L)
+}
+
+// Div performs element-wise division dst / s
+// and stores the value in dst. It panics if the
+// lengths of s and t are not equal.
+func Div(dst, s []float64) {
+	if len(dst) != len(s) {
+		panic("floats: slice lengths do not match")
+	}
+	f64.Div(dst, s)
+}
+
+// DivTo performs element-wise division s / t
+// and stores the value in dst. It panics if the
+// lengths of s, t, and dst are not equal.
+func DivTo(dst, s, t []float64) []float64 {
+	if len(s) != len(t) || len(dst) != len(t) {
+		panic("floats: slice lengths do not match")
+	}
+	return f64.DivTo(dst, s, t)
+}
+
+// Dot computes the dot product of s1 and s2, i.e.
+// sum_{i = 1}^N s1[i]*s2[i].
+// A panic will occur if lengths of arguments do not match.
+func Dot(s1, s2 []float64) float64 {
+	if len(s1) != len(s2) {
+		panic("floats: lengths of the slices do not match")
+	}
+	return f64.DotUnitary(s1, s2)
+}
+
+// Equal returns true if the slices have equal lengths and
+// all elements are numerically identical.
+func Equal(s1, s2 []float64) bool {
+	if len(s1) != len(s2) {
+		return false
+	}
+	for i, val := range s1 {
+		if s2[i] != val {
+			return false
+		}
+	}
+	return true
+}
+
+// EqualApprox returns true if the slices have equal lengths and
+// all element pairs have an absolute tolerance less than tol or a
+// relative tolerance less than tol.
+func EqualApprox(s1, s2 []float64, tol float64) bool {
+	if len(s1) != len(s2) {
+		return false
+	}
+	for i, a := range s1 {
+		if !EqualWithinAbsOrRel(a, s2[i], tol, tol) {
+			return false
+		}
+	}
+	return true
+}
+
+// EqualFunc returns true if the slices have the same lengths
+// and the function returns true for all element pairs.
+func EqualFunc(s1, s2 []float64, f func(float64, float64) bool) bool {
+	if len(s1) != len(s2) {
+		return false
+	}
+	for i, val := range s1 {
+		if !f(val, s2[i]) {
+			return false
+		}
+	}
+	return true
+}
+
+// EqualWithinAbs returns true if a and b have an absolute
+// difference of less than tol.
+func EqualWithinAbs(a, b, tol float64) bool {
+	return a == b || math.Abs(a-b) <= tol
+}
+
+const minNormalFloat64 = 2.2250738585072014e-308
+
+// EqualWithinRel returns true if the difference between a and b
+// is not greater than tol times the greater value.
+func EqualWithinRel(a, b, tol float64) bool {
+	if a == b {
+		return true
+	}
+	delta := math.Abs(a - b)
+	if delta <= minNormalFloat64 {
+		return delta <= tol*minNormalFloat64
+	}
+	// We depend on the division in this relationship to identify
+	// infinities (we rely on the NaN to fail the test) otherwise
+	// we compare Infs of the same sign and evaluate Infs as equal
+	// independent of sign.
+	return delta/math.Max(math.Abs(a), math.Abs(b)) <= tol
+}
+
+// EqualWithinAbsOrRel returns true if a and b are equal to within
+// the absolute tolerance.
+func EqualWithinAbsOrRel(a, b, absTol, relTol float64) bool {
+	if EqualWithinAbs(a, b, absTol) {
+		return true
+	}
+	return EqualWithinRel(a, b, relTol)
+}
+
+// EqualWithinULP returns true if a and b are equal to within
+// the specified number of floating point units in the last place.
+func EqualWithinULP(a, b float64, ulp uint) bool {
+	if a == b {
+		return true
+	}
+	if math.IsNaN(a) || math.IsNaN(b) {
+		return false
+	}
+	if math.Signbit(a) != math.Signbit(b) {
+		return math.Float64bits(math.Abs(a))+math.Float64bits(math.Abs(b)) <= uint64(ulp)
+	}
+	return ulpDiff(math.Float64bits(a), math.Float64bits(b)) <= uint64(ulp)
+}
+
+func ulpDiff(a, b uint64) uint64 {
+	if a > b {
+		return a - b
+	}
+	return b - a
+}
+
+// EqualLengths returns true if all of the slices have equal length,
+// and false otherwise. Returns true if there are no input slices.
+func EqualLengths(slices ...[]float64) bool {
+	// This length check is needed: http://play.golang.org/p/sdty6YiLhM
+	if len(slices) == 0 {
+		return true
+	}
+	l := len(slices[0])
+	for i := 1; i < len(slices); i++ {
+		if len(slices[i]) != l {
+			return false
+		}
+	}
+	return true
+}
+
+// Find applies f to every element of s and returns the indices of the first
+// k elements for which the f returns true, or all such elements
+// if k < 0.
+// Find will reslice inds to have 0 length, and will append
+// found indices to inds.
+// If k > 0 and there are fewer than k elements in s satisfying f,
+// all of the found elements will be returned along with an error.
+// At the return of the function, the input inds will be in an undetermined state.
+func Find(inds []int, f func(float64) bool, s []float64, k int) ([]int, error) {
+	// inds is also returned to allow for calling with nil
+
+	// Reslice inds to have zero length
+	inds = inds[:0]
+
+	// If zero elements requested, can just return
+	if k == 0 {
+		return inds, nil
+	}
+
+	// If k < 0, return all of the found indices
+	if k < 0 {
+		for i, val := range s {
+			if f(val) {
+				inds = append(inds, i)
+			}
+		}
+		return inds, nil
+	}
+
+	// Otherwise, find the first k elements
+	nFound := 0
+	for i, val := range s {
+		if f(val) {
+			inds = append(inds, i)
+			nFound++
+			if nFound == k {
+				return inds, nil
+			}
+		}
+	}
+	// Finished iterating over the loop, which means k elements were not found
+	return inds, errors.New("floats: insufficient elements found")
+}
+
+// HasNaN returns true if the slice s has any values that are NaN and false
+// otherwise.
+func HasNaN(s []float64) bool {
+	for _, v := range s {
+		if math.IsNaN(v) {
+			return true
+		}
+	}
+	return false
+}
+
+// LogSpan returns a set of n equally spaced points in log space between,
+// l and u where N is equal to len(dst). The first element of the
+// resulting dst will be l and the final element of dst will be u.
+// Panics if len(dst) < 2
+// Note that this call will return NaNs if either l or u are negative, and
+// will return all zeros if l or u is zero.
+// Also returns the mutated slice dst, so that it can be used in range, like:
+//
+//     for i, x := range LogSpan(dst, l, u) { ... }
+func LogSpan(dst []float64, l, u float64) []float64 {
+	Span(dst, math.Log(l), math.Log(u))
+	for i := range dst {
+		dst[i] = math.Exp(dst[i])
+	}
+	return dst
+}
+
+// LogSumExp returns the log of the sum of the exponentials of the values in s.
+// Panics if s is an empty slice.
+func LogSumExp(s []float64) float64 {
+	// Want to do this in a numerically stable way which avoids
+	// overflow and underflow
+	// First, find the maximum value in the slice.
+	maxval := Max(s)
+	if math.IsInf(maxval, 0) {
+		// If it's infinity either way, the logsumexp will be infinity as well
+		// returning now avoids NaNs
+		return maxval
+	}
+	var lse float64
+	// Compute the sumexp part
+	for _, val := range s {
+		lse += math.Exp(val - maxval)
+	}
+	// Take the log and add back on the constant taken out
+	return math.Log(lse) + maxval
+}
+
+// Max returns the maximum value in the input slice. If the slice is empty, Max will panic.
+func Max(s []float64) float64 {
+	return s[MaxIdx(s)]
+}
+
+// MaxIdx returns the index of the maximum value in the input slice. If several
+// entries have the maximum value, the first such index is returned. If the slice
+// is empty, MaxIdx will panic.
+func MaxIdx(s []float64) int {
+	if len(s) == 0 {
+		panic("floats: zero slice length")
+	}
+	max := math.NaN()
+	var ind int
+	for i, v := range s {
+		if math.IsNaN(v) {
+			continue
+		}
+		if v > max || math.IsNaN(max) {
+			max = v
+			ind = i
+		}
+	}
+	return ind
+}
+
+// Min returns the maximum value in the input slice. If the slice is empty, Min will panic.
+func Min(s []float64) float64 {
+	return s[MinIdx(s)]
+}
+
+// MinIdx returns the index of the minimum value in the input slice. If several
+// entries have the maximum value, the first such index is returned. If the slice
+// is empty, MinIdx will panic.
+func MinIdx(s []float64) int {
+	if len(s) == 0 {
+		panic("floats: zero slice length")
+	}
+	min := math.NaN()
+	var ind int
+	for i, v := range s {
+		if math.IsNaN(v) {
+			continue
+		}
+		if v < min || math.IsNaN(min) {
+			min = v
+			ind = i
+		}
+	}
+	return ind
+}
+
+// Mul performs element-wise multiplication between dst
+// and s and stores the value in dst. Panics if the
+// lengths of s and t are not equal.
+func Mul(dst, s []float64) {
+	if len(dst) != len(s) {
+		panic("floats: slice lengths do not match")
+	}
+	for i, val := range s {
+		dst[i] *= val
+	}
+}
+
+// MulTo performs element-wise multiplication between s
+// and t and stores the value in dst. Panics if the
+// lengths of s, t, and dst are not equal.
+func MulTo(dst, s, t []float64) []float64 {
+	if len(s) != len(t) || len(dst) != len(t) {
+		panic("floats: slice lengths do not match")
+	}
+	for i, val := range t {
+		dst[i] = val * s[i]
+	}
+	return dst
+}
+
+const (
+	nanBits = 0x7ff8000000000000
+	nanMask = 0xfff8000000000000
+)
+
+// NaNWith returns an IEEE 754 "quiet not-a-number" value with the
+// payload specified in the low 51 bits of payload.
+// The NaN returned by math.NaN has a bit pattern equal to NaNWith(1).
+func NaNWith(payload uint64) float64 {
+	return math.Float64frombits(nanBits | (payload &^ nanMask))
+}
+
+// NaNPayload returns the lowest 51 bits payload of an IEEE 754 "quiet
+// not-a-number". For values of f other than quiet-NaN, NaNPayload
+// returns zero and false.
+func NaNPayload(f float64) (payload uint64, ok bool) {
+	b := math.Float64bits(f)
+	if b&nanBits != nanBits {
+		return 0, false
+	}
+	return b &^ nanMask, true
+}
+
+// NearestIdx returns the index of the element in s
+// whose value is nearest to v.  If several such
+// elements exist, the lowest index is returned.
+// NearestIdx panics if len(s) == 0.
+func NearestIdx(s []float64, v float64) int {
+	if len(s) == 0 {
+		panic("floats: zero length slice")
+	}
+	switch {
+	case math.IsNaN(v):
+		return 0
+	case math.IsInf(v, 1):
+		return MaxIdx(s)
+	case math.IsInf(v, -1):
+		return MinIdx(s)
+	}
+	var ind int
+	dist := math.NaN()
+	for i, val := range s {
+		newDist := math.Abs(v - val)
+		// A NaN distance will not be closer.
+		if math.IsNaN(newDist) {
+			continue
+		}
+		if newDist < dist || math.IsNaN(dist) {
+			dist = newDist
+			ind = i
+		}
+	}
+	return ind
+}
+
+// NearestIdxForSpan return the index of a hypothetical vector created
+// by Span with length n and bounds l and u whose value is closest
+// to v. That is, NearestIdxForSpan(n, l, u, v) is equivalent to
+// Nearest(Span(make([]float64, n),l,u),v) without an allocation.
+// NearestIdxForSpan panics if n is less than two.
+func NearestIdxForSpan(n int, l, u float64, v float64) int {
+	if n <= 1 {
+		panic("floats: span must have length >1")
+	}
+	if math.IsNaN(v) {
+		return 0
+	}
+
+	// Special cases for Inf and NaN.
+	switch {
+	case math.IsNaN(l) && !math.IsNaN(u):
+		return n - 1
+	case math.IsNaN(u):
+		return 0
+	case math.IsInf(l, 0) && math.IsInf(u, 0):
+		if l == u {
+			return 0
+		}
+		if n%2 == 1 {
+			if !math.IsInf(v, 0) {
+				return n / 2
+			}
+			if math.Copysign(1, v) == math.Copysign(1, l) {
+				return 0
+			}
+			return n/2 + 1
+		}
+		if math.Copysign(1, v) == math.Copysign(1, l) {
+			return 0
+		}
+		return n / 2
+	case math.IsInf(l, 0):
+		if v == l {
+			return 0
+		}
+		return n - 1
+	case math.IsInf(u, 0):
+		if v == u {
+			return n - 1
+		}
+		return 0
+	case math.IsInf(v, -1):
+		if l <= u {
+			return 0
+		}
+		return n - 1
+	case math.IsInf(v, 1):
+		if u <= l {
+			return 0
+		}
+		return n - 1
+	}
+
+	// Special cases for v outside (l, u) and (u, l).
+	switch {
+	case l < u:
+		if v <= l {
+			return 0
+		}
+		if v >= u {
+			return n - 1
+		}
+	case l > u:
+		if v >= l {
+			return 0
+		}
+		if v <= u {
+			return n - 1
+		}
+	default:
+		return 0
+	}
+
+	// Can't guarantee anything about exactly halfway between
+	// because of floating point weirdness.
+	return int((float64(n)-1)/(u-l)*(v-l) + 0.5)
+}
+
+// Norm returns the L norm of the slice S, defined as
+// (sum_{i=1}^N s[i]^L)^{1/L}
+// Special cases:
+// L = math.Inf(1) gives the maximum absolute value.
+// Does not correctly compute the zero norm (use Count).
+func Norm(s []float64, L float64) float64 {
+	// Should this complain if L is not positive?
+	// Should this be done in log space for better numerical stability?
+	//	would be more cost
+	//	maybe only if L is high?
+	if len(s) == 0 {
+		return 0
+	}
+	if L == 2 {
+		twoNorm := math.Abs(s[0])
+		for i := 1; i < len(s); i++ {
+			twoNorm = math.Hypot(twoNorm, s[i])
+		}
+		return twoNorm
+	}
+	var norm float64
+	if L == 1 {
+		for _, val := range s {
+			norm += math.Abs(val)
+		}
+		return norm
+	}
+	if math.IsInf(L, 1) {
+		for _, val := range s {
+			norm = math.Max(norm, math.Abs(val))
+		}
+		return norm
+	}
+	for _, val := range s {
+		norm += math.Pow(math.Abs(val), L)
+	}
+	return math.Pow(norm, 1/L)
+}
+
+// ParseWithNA converts the string s to a float64 in v.
+// If s equals missing, w is returned as 0, otherwise 1.
+func ParseWithNA(s, missing string) (v, w float64, err error) {
+	if s == missing {
+		return 0, 0, nil
+	}
+	v, err = strconv.ParseFloat(s, 64)
+	if err == nil {
+		w = 1
+	}
+	return v, w, err
+}
+
+// Prod returns the product of the elements of the slice.
+// Returns 1 if len(s) = 0.
+func Prod(s []float64) float64 {
+	prod := 1.0
+	for _, val := range s {
+		prod *= val
+	}
+	return prod
+}
+
+// Reverse reverses the order of elements in the slice.
+func Reverse(s []float64) {
+	for i, j := 0, len(s)-1; i < j; i, j = i+1, j-1 {
+		s[i], s[j] = s[j], s[i]
+	}
+}
+
+// Round returns the half away from zero rounded value of x with prec precision.
+//
+// Special cases are:
+// 	Round(±0) = +0
+// 	Round(±Inf) = ±Inf
+// 	Round(NaN) = NaN
+func Round(x float64, prec int) float64 {
+	if x == 0 {
+		// Make sure zero is returned
+		// without the negative bit set.
+		return 0
+	}
+	// Fast path for positive precision on integers.
+	if prec >= 0 && x == math.Trunc(x) {
+		return x
+	}
+	pow := math.Pow10(prec)
+	intermed := x * pow
+	if math.IsInf(intermed, 0) {
+		return x
+	}
+	if x < 0 {
+		x = math.Ceil(intermed - 0.5)
+	} else {
+		x = math.Floor(intermed + 0.5)
+	}
+
+	if x == 0 {
+		return 0
+	}
+
+	return x / pow
+}
+
+// RoundEven returns the half even rounded value of x with prec precision.
+//
+// Special cases are:
+// 	RoundEven(±0) = +0
+// 	RoundEven(±Inf) = ±Inf
+// 	RoundEven(NaN) = NaN
+func RoundEven(x float64, prec int) float64 {
+	if x == 0 {
+		// Make sure zero is returned
+		// without the negative bit set.
+		return 0
+	}
+	// Fast path for positive precision on integers.
+	if prec >= 0 && x == math.Trunc(x) {
+		return x
+	}
+	pow := math.Pow10(prec)
+	intermed := x * pow
+	if math.IsInf(intermed, 0) {
+		return x
+	}
+	if isHalfway(intermed) {
+		correction, _ := math.Modf(math.Mod(intermed, 2))
+		intermed += correction
+		if intermed > 0 {
+			x = math.Floor(intermed)
+		} else {
+			x = math.Ceil(intermed)
+		}
+	} else {
+		if x < 0 {
+			x = math.Ceil(intermed - 0.5)
+		} else {
+			x = math.Floor(intermed + 0.5)
+		}
+	}
+
+	if x == 0 {
+		return 0
+	}
+
+	return x / pow
+}
+
+func isHalfway(x float64) bool {
+	_, frac := math.Modf(x)
+	frac = math.Abs(frac)
+	return frac == 0.5 || (math.Nextafter(frac, math.Inf(-1)) < 0.5 && math.Nextafter(frac, math.Inf(1)) > 0.5)
+}
+
+// Same returns true if the input slices have the same length and the all elements
+// have the same value with NaN treated as the same.
+func Same(s, t []float64) bool {
+	if len(s) != len(t) {
+		return false
+	}
+	for i, v := range s {
+		w := t[i]
+		if v != w && !(math.IsNaN(v) && math.IsNaN(w)) {
+			return false
+		}
+	}
+	return true
+}
+
+// Scale multiplies every element in dst by the scalar c.
+func Scale(c float64, dst []float64) {
+	if len(dst) > 0 {
+		f64.ScalUnitary(c, dst)
+	}
+}
+
+// ScaleTo multiplies the elements in s by c and stores the result in dst.
+func ScaleTo(dst []float64, c float64, s []float64) []float64 {
+	if len(dst) != len(s) {
+		panic("floats: lengths of slices do not match")
+	}
+	if len(dst) > 0 {
+		f64.ScalUnitaryTo(dst, c, s)
+	}
+	return dst
+}
+
+// Span returns a set of N equally spaced points between l and u, where N
+// is equal to the length of the destination. The first element of the destination
+// is l, the final element of the destination is u.
+//
+// Panics if len(dst) < 2.
+//
+// Span also returns the mutated slice dst, so that it can be used in range expressions,
+// like:
+//
+//     for i, x := range Span(dst, l, u) { ... }
+func Span(dst []float64, l, u float64) []float64 {
+	n := len(dst)
+	if n < 2 {
+		panic("floats: destination must have length >1")
+	}
+
+	// Special cases for Inf and NaN.
+	switch {
+	case math.IsNaN(l):
+		for i := range dst[:len(dst)-1] {
+			dst[i] = math.NaN()
+		}
+		dst[len(dst)-1] = u
+		return dst
+	case math.IsNaN(u):
+		for i := range dst[1:] {
+			dst[i+1] = math.NaN()
+		}
+		dst[0] = l
+		return dst
+	case math.IsInf(l, 0) && math.IsInf(u, 0):
+		for i := range dst[:len(dst)/2] {
+			dst[i] = l
+			dst[len(dst)-i-1] = u
+		}
+		if len(dst)%2 == 1 {
+			if l != u {
+				dst[len(dst)/2] = 0
+			} else {
+				dst[len(dst)/2] = l
+			}
+		}
+		return dst
+	case math.IsInf(l, 0):
+		for i := range dst[:len(dst)-1] {
+			dst[i] = l
+		}
+		dst[len(dst)-1] = u
+		return dst
+	case math.IsInf(u, 0):
+		for i := range dst[1:] {
+			dst[i+1] = u
+		}
+		dst[0] = l
+		return dst
+	}
+
+	step := (u - l) / float64(n-1)
+	for i := range dst {
+		dst[i] = l + step*float64(i)
+	}
+	return dst
+}
+
+// Sub subtracts, element-wise, the elements of s from dst. Panics if
+// the lengths of dst and s do not match.
+func Sub(dst, s []float64) {
+	if len(dst) != len(s) {
+		panic("floats: length of the slices do not match")
+	}
+	f64.AxpyUnitaryTo(dst, -1, s, dst)
+}
+
+// SubTo subtracts, element-wise, the elements of t from s and
+// stores the result in dst. Panics if the lengths of s, t and dst do not match.
+func SubTo(dst, s, t []float64) []float64 {
+	if len(s) != len(t) {
+		panic("floats: length of subtractor and subtractee do not match")
+	}
+	if len(dst) != len(s) {
+		panic("floats: length of destination does not match length of subtractor")
+	}
+	f64.AxpyUnitaryTo(dst, -1, t, s)
+	return dst
+}
+
+// Sum returns the sum of the elements of the slice.
+func Sum(s []float64) float64 {
+	return f64.Sum(s)
+}
+
+// Within returns the first index i where s[i] <= v < s[i+1]. Within panics if:
+//  - len(s) < 2
+//  - s is not sorted
+func Within(s []float64, v float64) int {
+	if len(s) < 2 {
+		panic("floats: slice length less than 2")
+	}
+	if !sort.Float64sAreSorted(s) {
+		panic("floats: input slice not sorted")
+	}
+	if v < s[0] || v >= s[len(s)-1] || math.IsNaN(v) {
+		return -1
+	}
+	for i, f := range s[1:] {
+		if v < f {
+			return i
+		}
+	}
+	return -1
+}

vendor/gonum.org/v1/gonum/graph/.gitignore

@@ -0,0 +1 @@
+test.out

vendor/gonum.org/v1/gonum/graph/README.md

@@ -0,0 +1,3 @@
+# Gonum graph [![GoDoc](https://godoc.org/gonum.org/v1/gonum/graph?status.svg)](https://godoc.org/gonum.org/v1/gonum/graph)
+
+This is a generalized graph package for the Go language.

vendor/gonum.org/v1/gonum/graph/doc.go

@@ -0,0 +1,9 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package graph defines graph interfaces.
+//
+// Routines to test contract compliance by user implemented graph types
+// are available in gonum.org/v1/gonum/graph/testgraph.
+package graph // import "gonum.org/v1/gonum/graph"

vendor/gonum.org/v1/gonum/graph/graph.go

@@ -0,0 +1,282 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package graph
+
+// Node is a graph node. It returns a graph-unique integer ID.
+type Node interface {
+	ID() int64
+}
+
+// Edge is a graph edge. In directed graphs, the direction of the
+// edge is given from -> to, otherwise the edge is semantically
+// unordered.
+type Edge interface {
+	// From returns the from node of the edge.
+	From() Node
+
+	// To returns the to node of the edge.
+	To() Node
+
+	// ReversedEdge returns an edge that has
+	// the end points of the receiver swapped.
+	ReversedEdge() Edge
+}
+
+// WeightedEdge is a weighted graph edge. In directed graphs, the direction
+// of the edge is given from -> to, otherwise the edge is semantically
+// unordered.
+type WeightedEdge interface {
+	Edge
+	Weight() float64
+}
+
+// Graph is a generalized graph.
+type Graph interface {
+	// Node returns the node with the given ID if it exists
+	// in the graph, and nil otherwise.
+	Node(id int64) Node
+
+	// Nodes returns all the nodes in the graph.
+	//
+	// Nodes must not return nil.
+	Nodes() Nodes
+
+	// From returns all nodes that can be reached directly
+	// from the node with the given ID.
+	//
+	// From must not return nil.
+	From(id int64) Nodes
+
+	// HasEdgeBetween returns whether an edge exists between
+	// nodes with IDs xid and yid without considering direction.
+	HasEdgeBetween(xid, yid int64) bool
+
+	// Edge returns the edge from u to v, with IDs uid and vid,
+	// if such an edge exists and nil otherwise. The node v
+	// must be directly reachable from u as defined by the
+	// From method.
+	Edge(uid, vid int64) Edge
+}
+
+// Weighted is a weighted graph.
+type Weighted interface {
+	Graph
+
+	// WeightedEdge returns the weighted edge from u to v
+	// with IDs uid and vid if such an edge exists and
+	// nil otherwise. The node v must be directly
+	// reachable from u as defined by the From method.
+	WeightedEdge(uid, vid int64) WeightedEdge
+
+	// Weight returns the weight for the edge between
+	// x and y with IDs xid and yid if Edge(xid, yid)
+	// returns a non-nil Edge.
+	// If x and y are the same node or there is no
+	// joining edge between the two nodes the weight
+	// value returned is implementation dependent.
+	// Weight returns true if an edge exists between
+	// x and y or if x and y have the same ID, false
+	// otherwise.
+	Weight(xid, yid int64) (w float64, ok bool)
+}
+
+// Undirected is an undirected graph.
+type Undirected interface {
+	Graph
+
+	// EdgeBetween returns the edge between nodes x and y
+	// with IDs xid and yid.
+	EdgeBetween(xid, yid int64) Edge
+}
+
+// WeightedUndirected is a weighted undirected graph.
+type WeightedUndirected interface {
+	Weighted
+
+	// WeightedEdgeBetween returns the edge between nodes
+	// x and y with IDs xid and yid.
+	WeightedEdgeBetween(xid, yid int64) WeightedEdge
+}
+
+// Directed is a directed graph.
+type Directed interface {
+	Graph
+
+	// HasEdgeFromTo returns whether an edge exists
+	// in the graph from u to v with IDs uid and vid.
+	HasEdgeFromTo(uid, vid int64) bool
+
+	// To returns all nodes that can reach directly
+	// to the node with the given ID.
+	//
+	// To must not return nil.
+	To(id int64) Nodes
+}
+
+// WeightedDirected is a weighted directed graph.
+type WeightedDirected interface {
+	Weighted
+
+	// HasEdgeFromTo returns whether an edge exists
+	// in the graph from u to v with the IDs uid and
+	// vid.
+	HasEdgeFromTo(uid, vid int64) bool
+
+	// To returns all nodes that can reach directly
+	// to the node with the given ID.
+	//
+	// To must not return nil.
+	To(id int64) Nodes
+}
+
+// NodeAdder is an interface for adding arbitrary nodes to a graph.
+type NodeAdder interface {
+	// NewNode returns a new Node with a unique
+	// arbitrary ID.
+	NewNode() Node
+
+	// AddNode adds a node to the graph. AddNode panics if
+	// the added node ID matches an existing node ID.
+	AddNode(Node)
+}
+
+// NodeRemover is an interface for removing nodes from a graph.
+type NodeRemover interface {
+	// RemoveNode removes the node with the given ID
+	// from the graph, as well as any edges attached
+	// to it. If the node is not in the graph it is
+	// a no-op.
+	RemoveNode(id int64)
+}
+
+// EdgeAdder is an interface for adding edges to a graph.
+type EdgeAdder interface {
+	// NewEdge returns a new Edge from the source to the destination node.
+	NewEdge(from, to Node) Edge
+
+	// SetEdge adds an edge from one node to another.
+	// If the graph supports node addition the nodes
+	// will be added if they do not exist, otherwise
+	// SetEdge will panic.
+	// The behavior of an EdgeAdder when the IDs
+	// returned by e.From() and e.To() are equal is
+	// implementation-dependent.
+	// Whether e, e.From() and e.To() are stored
+	// within the graph is implementation dependent.
+	SetEdge(e Edge)
+}
+
+// WeightedEdgeAdder is an interface for adding edges to a graph.
+type WeightedEdgeAdder interface {
+	// NewWeightedEdge returns a new WeightedEdge from
+	// the source to the destination node.
+	NewWeightedEdge(from, to Node, weight float64) WeightedEdge
+
+	// SetWeightedEdge adds an edge from one node to
+	// another. If the graph supports node addition
+	// the nodes will be added if they do not exist,
+	// otherwise SetWeightedEdge will panic.
+	// The behavior of a WeightedEdgeAdder when the IDs
+	// returned by e.From() and e.To() are equal is
+	// implementation-dependent.
+	// Whether e, e.From() and e.To() are stored
+	// within the graph is implementation dependent.
+	SetWeightedEdge(e WeightedEdge)
+}
+
+// EdgeRemover is an interface for removing nodes from a graph.
+type EdgeRemover interface {
+	// RemoveEdge removes the edge with the given end
+	// IDs, leaving the terminal nodes. If the edge
+	// does not exist it is a no-op.
+	RemoveEdge(fid, tid int64)
+}
+
+// Builder is a graph that can have nodes and edges added.
+type Builder interface {
+	NodeAdder
+	EdgeAdder
+}
+
+// WeightedBuilder is a graph that can have nodes and weighted edges added.
+type WeightedBuilder interface {
+	NodeAdder
+	WeightedEdgeAdder
+}
+
+// UndirectedBuilder is an undirected graph builder.
+type UndirectedBuilder interface {
+	Undirected
+	Builder
+}
+
+// UndirectedWeightedBuilder is an undirected weighted graph builder.
+type UndirectedWeightedBuilder interface {
+	Undirected
+	WeightedBuilder
+}
+
+// DirectedBuilder is a directed graph builder.
+type DirectedBuilder interface {
+	Directed
+	Builder
+}
+
+// DirectedWeightedBuilder is a directed weighted graph builder.
+type DirectedWeightedBuilder interface {
+	Directed
+	WeightedBuilder
+}
+
+// Copy copies nodes and edges as undirected edges from the source to the destination
+// without first clearing the destination. Copy will panic if a node ID in the source
+// graph matches a node ID in the destination.
+//
+// If the source is undirected and the destination is directed both directions will
+// be present in the destination after the copy is complete.
+func Copy(dst Builder, src Graph) {
+	nodes := src.Nodes()
+	for nodes.Next() {
+		dst.AddNode(nodes.Node())
+	}
+	nodes.Reset()
+	for nodes.Next() {
+		u := nodes.Node()
+		uid := u.ID()
+		to := src.From(uid)
+		for to.Next() {
+			v := to.Node()
+			dst.SetEdge(src.Edge(uid, v.ID()))
+		}
+	}
+}
+
+// CopyWeighted copies nodes and edges as undirected edges from the source to the destination
+// without first clearing the destination. Copy will panic if a node ID in the source
+// graph matches a node ID in the destination.
+//
+// If the source is undirected and the destination is directed both directions will
+// be present in the destination after the copy is complete.
+//
+// If the source is a directed graph, the destination is undirected, and a fundamental
+// cycle exists with two nodes where the edge weights differ, the resulting destination
+// graph's edge weight between those nodes is undefined. If there is a defined function
+// to resolve such conflicts, an UndirectWeighted may be used to do this.
+func CopyWeighted(dst WeightedBuilder, src Weighted) {
+	nodes := src.Nodes()
+	for nodes.Next() {
+		dst.AddNode(nodes.Node())
+	}
+	nodes.Reset()
+	for nodes.Next() {
+		u := nodes.Node()
+		uid := u.ID()
+		to := src.From(uid)
+		for to.Next() {
+			v := to.Node()
+			dst.SetWeightedEdge(src.WeightedEdge(uid, v.ID()))
+		}
+	}
+}

vendor/gonum.org/v1/gonum/graph/internal/linear/doc.go

@@ -0,0 +1,6 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package linear provides common linear data structures.
+package linear // import "gonum.org/v1/gonum/graph/internal/linear"

vendor/gonum.org/v1/gonum/graph/internal/linear/linear.go

@@ -0,0 +1,73 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package linear
+
+import (
+	"gonum.org/v1/gonum/graph"
+)
+
+// NodeStack implements a LIFO stack of graph.Node.
+type NodeStack []graph.Node
+
+// Len returns the number of graph.Nodes on the stack.
+func (s *NodeStack) Len() int { return len(*s) }
+
+// Pop returns the last graph.Node on the stack and removes it
+// from the stack.
+func (s *NodeStack) Pop() graph.Node {
+	v := *s
+	v, n := v[:len(v)-1], v[len(v)-1]
+	*s = v
+	return n
+}
+
+// Push adds the node n to the stack at the last position.
+func (s *NodeStack) Push(n graph.Node) { *s = append(*s, n) }
+
+// NodeQueue implements a FIFO queue.
+type NodeQueue struct {
+	head int
+	data []graph.Node
+}
+
+// Len returns the number of graph.Nodes in the queue.
+func (q *NodeQueue) Len() int { return len(q.data) - q.head }
+
+// Enqueue adds the node n to the back of the queue.
+func (q *NodeQueue) Enqueue(n graph.Node) {
+	if len(q.data) == cap(q.data) && q.head > 0 {
+		l := q.Len()
+		copy(q.data, q.data[q.head:])
+		q.head = 0
+		q.data = append(q.data[:l], n)
+	} else {
+		q.data = append(q.data, n)
+	}
+}
+
+// Dequeue returns the graph.Node at the front of the queue and
+// removes it from the queue.
+func (q *NodeQueue) Dequeue() graph.Node {
+	if q.Len() == 0 {
+		panic("queue: empty queue")
+	}
+
+	var n graph.Node
+	n, q.data[q.head] = q.data[q.head], nil
+	q.head++
+
+	if q.Len() == 0 {
+		q.head = 0
+		q.data = q.data[:0]
+	}
+
+	return n
+}
+
+// Reset clears the queue for reuse.
+func (q *NodeQueue) Reset() {
+	q.head = 0
+	q.data = q.data[:0]
+}

vendor/gonum.org/v1/gonum/graph/internal/ordered/doc.go

@@ -0,0 +1,6 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package ordered provides common sort ordering types.
+package ordered // import "gonum.org/v1/gonum/graph/internal/ordered"

vendor/gonum.org/v1/gonum/graph/internal/ordered/sort.go

@@ -0,0 +1,93 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package ordered
+
+import "gonum.org/v1/gonum/graph"
+
+// ByID implements the sort.Interface sorting a slice of graph.Node
+// by ID.
+type ByID []graph.Node
+
+func (n ByID) Len() int           { return len(n) }
+func (n ByID) Less(i, j int) bool { return n[i].ID() < n[j].ID() }
+func (n ByID) Swap(i, j int)      { n[i], n[j] = n[j], n[i] }
+
+// BySliceValues implements the sort.Interface sorting a slice of
+// []int64 lexically by the values of the []int64.
+type BySliceValues [][]int64
+
+func (c BySliceValues) Len() int { return len(c) }
+func (c BySliceValues) Less(i, j int) bool {
+	a, b := c[i], c[j]
+	l := len(a)
+	if len(b) < l {
+		l = len(b)
+	}
+	for k, v := range a[:l] {
+		if v < b[k] {
+			return true
+		}
+		if v > b[k] {
+			return false
+		}
+	}
+	return len(a) < len(b)
+}
+func (c BySliceValues) Swap(i, j int) { c[i], c[j] = c[j], c[i] }
+
+// BySliceIDs implements the sort.Interface sorting a slice of
+// []graph.Node lexically by the IDs of the []graph.Node.
+type BySliceIDs [][]graph.Node
+
+func (c BySliceIDs) Len() int { return len(c) }
+func (c BySliceIDs) Less(i, j int) bool {
+	a, b := c[i], c[j]
+	l := len(a)
+	if len(b) < l {
+		l = len(b)
+	}
+	for k, v := range a[:l] {
+		if v.ID() < b[k].ID() {
+			return true
+		}
+		if v.ID() > b[k].ID() {
+			return false
+		}
+	}
+	return len(a) < len(b)
+}
+func (c BySliceIDs) Swap(i, j int) { c[i], c[j] = c[j], c[i] }
+
+// Int64s implements the sort.Interface sorting a slice of
+// int64.
+type Int64s []int64
+
+func (s Int64s) Len() int           { return len(s) }
+func (s Int64s) Less(i, j int) bool { return s[i] < s[j] }
+func (s Int64s) Swap(i, j int)      { s[i], s[j] = s[j], s[i] }
+
+// Reverse reverses the order of nodes.
+func Reverse(nodes []graph.Node) {
+	for i, j := 0, len(nodes)-1; i < j; i, j = i+1, j-1 {
+		nodes[i], nodes[j] = nodes[j], nodes[i]
+	}
+}
+
+// LinesByIDs implements the sort.Interface sorting a slice of graph.LinesByIDs
+// lexically by the From IDs, then by the To IDs, finally by the Line IDs.
+type LinesByIDs []graph.Line
+
+func (n LinesByIDs) Len() int { return len(n) }
+func (n LinesByIDs) Less(i, j int) bool {
+	a, b := n[i], n[j]
+	if a.From().ID() != b.From().ID() {
+		return a.From().ID() < b.From().ID()
+	}
+	if a.To().ID() != b.To().ID() {
+		return a.To().ID() < b.To().ID()
+	}
+	return n[i].ID() < n[j].ID()
+}
+func (n LinesByIDs) Swap(i, j int) { n[i], n[j] = n[j], n[i] }

vendor/gonum.org/v1/gonum/graph/internal/set/doc.go

@@ -0,0 +1,6 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package set provides integer and graph.Node sets.
+package set // import "gonum.org/v1/gonum/graph/internal/set"

vendor/gonum.org/v1/gonum/graph/internal/set/same.go

@@ -0,0 +1,36 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !appengine,!safe
+
+package set
+
+import "unsafe"
+
+// same determines whether two sets are backed by the same store. In the
+// current implementation using hash maps it makes use of the fact that
+// hash maps are passed as a pointer to a runtime Hmap struct. A map is
+// not seen by the runtime as a pointer though, so we use unsafe to get
+// the maps' pointer values to compare.
+func same(a, b Nodes) bool {
+	return *(*uintptr)(unsafe.Pointer(&a)) == *(*uintptr)(unsafe.Pointer(&b))
+}
+
+// intsSame determines whether two sets are backed by the same store. In the
+// current implementation using hash maps it makes use of the fact that
+// hash maps are passed as a pointer to a runtime Hmap struct. A map is
+// not seen by the runtime as a pointer though, so we use unsafe to get
+// the maps' pointer values to compare.
+func intsSame(a, b Ints) bool {
+	return *(*uintptr)(unsafe.Pointer(&a)) == *(*uintptr)(unsafe.Pointer(&b))
+}
+
+// int64sSame determines whether two sets are backed by the same store. In the
+// current implementation using hash maps it makes use of the fact that
+// hash maps are passed as a pointer to a runtime Hmap struct. A map is
+// not seen by the runtime as a pointer though, so we use unsafe to get
+// the maps' pointer values to compare.
+func int64sSame(a, b Int64s) bool {
+	return *(*uintptr)(unsafe.Pointer(&a)) == *(*uintptr)(unsafe.Pointer(&b))
+}

vendor/gonum.org/v1/gonum/graph/internal/set/same_appengine.go

@@ -0,0 +1,36 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build appengine safe
+
+package set
+
+import "reflect"
+
+// same determines whether two sets are backed by the same store. In the
+// current implementation using hash maps it makes use of the fact that
+// hash maps are passed as a pointer to a runtime Hmap struct. A map is
+// not seen by the runtime as a pointer though, so we use reflect to get
+// the maps' pointer values to compare.
+func same(a, b Nodes) bool {
+	return reflect.ValueOf(a).Pointer() == reflect.ValueOf(b).Pointer()
+}
+
+// intsSame determines whether two sets are backed by the same store. In the
+// current implementation using hash maps it makes use of the fact that
+// hash maps are passed as a pointer to a runtime Hmap struct. A map is
+// not seen by the runtime as a pointer though, so we use reflect to get
+// the maps' pointer values to compare.
+func intsSame(a, b Ints) bool {
+	return reflect.ValueOf(a).Pointer() == reflect.ValueOf(b).Pointer()
+}
+
+// int64sSame determines whether two sets are backed by the same store. In the
+// current implementation using hash maps it makes use of the fact that
+// hash maps are passed as a pointer to a runtime Hmap struct. A map is
+// not seen by the runtime as a pointer though, so we use reflect to get
+// the maps' pointer values to compare.
+func int64sSame(a, b Int64s) bool {
+	return reflect.ValueOf(a).Pointer() == reflect.ValueOf(b).Pointer()
+}

vendor/gonum.org/v1/gonum/graph/internal/set/set.go

@@ -0,0 +1,228 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package set
+
+import "gonum.org/v1/gonum/graph"
+
+// Ints is a set of int identifiers.
+type Ints map[int]struct{}
+
+// The simple accessor methods for Ints are provided to allow ease of
+// implementation change should the need arise.
+
+// Add inserts an element into the set.
+func (s Ints) Add(e int) {
+	s[e] = struct{}{}
+}
+
+// Has reports the existence of the element in the set.
+func (s Ints) Has(e int) bool {
+	_, ok := s[e]
+	return ok
+}
+
+// Remove deletes the specified element from the set.
+func (s Ints) Remove(e int) {
+	delete(s, e)
+}
+
+// Count reports the number of elements stored in the set.
+func (s Ints) Count() int {
+	return len(s)
+}
+
+// IntsEqual reports set equality between the parameters. Sets are equal if
+// and only if they have the same elements.
+func IntsEqual(a, b Ints) bool {
+	if intsSame(a, b) {
+		return true
+	}
+
+	if len(a) != len(b) {
+		return false
+	}
+
+	for e := range a {
+		if _, ok := b[e]; !ok {
+			return false
+		}
+	}
+
+	return true
+}
+
+// Int64s is a set of int64 identifiers.
+type Int64s map[int64]struct{}
+
+// The simple accessor methods for Ints are provided to allow ease of
+// implementation change should the need arise.
+
+// Add inserts an element into the set.
+func (s Int64s) Add(e int64) {
+	s[e] = struct{}{}
+}
+
+// Has reports the existence of the element in the set.
+func (s Int64s) Has(e int64) bool {
+	_, ok := s[e]
+	return ok
+}
+
+// Remove deletes the specified element from the set.
+func (s Int64s) Remove(e int64) {
+	delete(s, e)
+}
+
+// Count reports the number of elements stored in the set.
+func (s Int64s) Count() int {
+	return len(s)
+}
+
+// Int64sEqual reports set equality between the parameters. Sets are equal if
+// and only if they have the same elements.
+func Int64sEqual(a, b Int64s) bool {
+	if int64sSame(a, b) {
+		return true
+	}
+
+	if len(a) != len(b) {
+		return false
+	}
+
+	for e := range a {
+		if _, ok := b[e]; !ok {
+			return false
+		}
+	}
+
+	return true
+}
+
+// Nodes is a set of nodes keyed in their integer identifiers.
+type Nodes map[int64]graph.Node
+
+// NewNodes returns a new Nodes.
+func NewNodes() Nodes {
+	return make(Nodes)
+}
+
+// NewNodes returns a new Nodes with the given size hint, n.
+func NewNodesSize(n int) Nodes {
+	return make(Nodes, n)
+}
+
+// The simple accessor methods for Nodes are provided to allow ease of
+// implementation change should the need arise.
+
+// Add inserts an element into the set.
+func (s Nodes) Add(n graph.Node) {
+	s[n.ID()] = n
+}
+
+// Remove deletes the specified element from the set.
+func (s Nodes) Remove(e graph.Node) {
+	delete(s, e.ID())
+}
+
+// Count returns the number of element in the set.
+func (s Nodes) Count() int {
+	return len(s)
+}
+
+// Has reports the existence of the elements in the set.
+func (s Nodes) Has(n graph.Node) bool {
+	_, ok := s[n.ID()]
+	return ok
+}
+
+// CloneNodes returns a clone of src.
+func CloneNodes(src Nodes) Nodes {
+	dst := make(Nodes, len(src))
+	for e, n := range src {
+		dst[e] = n
+	}
+	return dst
+}
+
+// Equal reports set equality between the parameters. Sets are equal if
+// and only if they have the same elements.
+func Equal(a, b Nodes) bool {
+	if same(a, b) {
+		return true
+	}
+
+	if len(a) != len(b) {
+		return false
+	}
+
+	for e := range a {
+		if _, ok := b[e]; !ok {
+			return false
+		}
+	}
+
+	return true
+}
+
+// UnionOfNodes returns the union of a and b.
+//
+// The union of two sets, a and b, is the set containing all the
+// elements of each, for instance:
+//
+//     {a,b,c} UNION {d,e,f} = {a,b,c,d,e,f}
+//
+// Since sets may not have repetition, unions of two sets that overlap
+// do not contain repeat elements, that is:
+//
+//     {a,b,c} UNION {b,c,d} = {a,b,c,d}
+//
+func UnionOfNodes(a, b Nodes) Nodes {
+	if same(a, b) {
+		return CloneNodes(a)
+	}
+
+	dst := make(Nodes)
+	for e, n := range a {
+		dst[e] = n
+	}
+	for e, n := range b {
+		dst[e] = n
+	}
+
+	return dst
+}
+
+// IntersectionOfNodes returns the intersection of a and b.
+//
+// The intersection of two sets, a and b, is the set containing all
+// the elements shared between the two sets, for instance:
+//
+//     {a,b,c} INTERSECT {b,c,d} = {b,c}
+//
+// The intersection between a set and itself is itself, and thus
+// effectively a copy operation:
+//
+//     {a,b,c} INTERSECT {a,b,c} = {a,b,c}
+//
+// The intersection between two sets that share no elements is the empty
+// set:
+//
+//     {a,b,c} INTERSECT {d,e,f} = {}
+//
+func IntersectionOfNodes(a, b Nodes) Nodes {
+	if same(a, b) {
+		return CloneNodes(a)
+	}
+	dst := make(Nodes)
+	if len(a) > len(b) {
+		a, b = b, a
+	}
+	for e, n := range a {
+		if _, ok := b[e]; ok {
+			dst[e] = n
+		}
+	}
+	return dst
+}

vendor/gonum.org/v1/gonum/graph/internal/uid/uid.go

@@ -0,0 +1,54 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package uid implements unique ID provision for graphs.
+package uid
+
+import "gonum.org/v1/gonum/graph/internal/set"
+
+// Max is the maximum value of int64.
+const Max = int64(^uint64(0) >> 1)
+
+// Set implements available ID storage.
+type Set struct {
+	maxID      int64
+	used, free set.Int64s
+}
+
+// NewSet returns a new Set. The returned value should not be passed except by pointer.
+func NewSet() Set {
+	return Set{maxID: -1, used: make(set.Int64s), free: make(set.Int64s)}
+}
+
+// NewID returns a new unique ID. The ID returned is not considered used
+// until passed in a call to use.
+func (s *Set) NewID() int64 {
+	for id := range s.free {
+		return id
+	}
+	if s.maxID != Max {
+		return s.maxID + 1
+	}
+	for id := int64(0); id <= s.maxID+1; id++ {
+		if !s.used.Has(id) {
+			return id
+		}
+	}
+	panic("unreachable")
+}
+
+// Use adds the id to the used IDs in the Set.
+func (s *Set) Use(id int64) {
+	s.used.Add(id)
+	s.free.Remove(id)
+	if id > s.maxID {
+		s.maxID = id
+	}
+}
+
+// Release frees the id for reuse.
+func (s *Set) Release(id int64) {
+	s.free.Add(id)
+	s.used.Remove(id)
+}

vendor/gonum.org/v1/gonum/graph/iterator/doc.go

@@ -0,0 +1,9 @@
+// Copyright ©2018 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package iterator provides node, edge and line iterators.
+//
+// The iterators provided satisfy the graph.Nodes, graph.Edges and
+// graph.Lines interfaces.
+package iterator

vendor/gonum.org/v1/gonum/graph/iterator/edges.go

@@ -0,0 +1,131 @@
+// Copyright ©2018 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package iterator
+
+import "gonum.org/v1/gonum/graph"
+
+// OrderedEdges implements the graph.Edges and graph.EdgeSlicer interfaces.
+// The iteration order of OrderedEdges is the order of edges passed to
+// NewEdgeIterator.
+type OrderedEdges struct {
+	idx   int
+	edges []graph.Edge
+}
+
+// NewOrderedEdges returns an OrderedEdges initialized with the provided edges.
+func NewOrderedEdges(edges []graph.Edge) *OrderedEdges {
+	return &OrderedEdges{idx: -1, edges: edges}
+}
+
+// Len returns the remaining number of edges to be iterated over.
+func (e *OrderedEdges) Len() int {
+	if e.idx >= len(e.edges) {
+		return 0
+	}
+	if e.idx <= 0 {
+		return len(e.edges)
+	}
+	return len(e.edges[e.idx:])
+}
+
+// Next returns whether the next call of Edge will return a valid edge.
+func (e *OrderedEdges) Next() bool {
+	if uint(e.idx)+1 < uint(len(e.edges)) {
+		e.idx++
+		return true
+	}
+	e.idx = len(e.edges)
+	return false
+}
+
+// Edge returns the current edge of the iterator. Next must have been
+// called prior to a call to Edge.
+func (e *OrderedEdges) Edge() graph.Edge {
+	if e.idx >= len(e.edges) || e.idx < 0 {
+		return nil
+	}
+	return e.edges[e.idx]
+}
+
+// EdgeSlice returns all the remaining edges in the iterator and advances
+// the iterator.
+func (e *OrderedEdges) EdgeSlice() []graph.Edge {
+	if e.idx >= len(e.edges) {
+		return nil
+	}
+	idx := e.idx
+	if idx == -1 {
+		idx = 0
+	}
+	e.idx = len(e.edges)
+	return e.edges[idx:]
+}
+
+// Reset returns the iterator to its initial state.
+func (e *OrderedEdges) Reset() {
+	e.idx = -1
+}
+
+// OrderedWeightedEdges implements the graph.Edges and graph.EdgeSlicer interfaces.
+// The iteration order of OrderedWeightedEdges is the order of edges passed to
+// NewEdgeIterator.
+type OrderedWeightedEdges struct {
+	idx   int
+	edges []graph.WeightedEdge
+}
+
+// NewOrderedWeightedEdges returns an OrderedWeightedEdges initialized with the provided edges.
+func NewOrderedWeightedEdges(edges []graph.WeightedEdge) *OrderedWeightedEdges {
+	return &OrderedWeightedEdges{idx: -1, edges: edges}
+}
+
+// Len returns the remaining number of edges to be iterated over.
+func (e *OrderedWeightedEdges) Len() int {
+	if e.idx >= len(e.edges) {
+		return 0
+	}
+	if e.idx <= 0 {
+		return len(e.edges)
+	}
+	return len(e.edges[e.idx:])
+}
+
+// Next returns whether the next call of WeightedEdge will return a valid edge.
+func (e *OrderedWeightedEdges) Next() bool {
+	if uint(e.idx)+1 < uint(len(e.edges)) {
+		e.idx++
+		return true
+	}
+	e.idx = len(e.edges)
+	return false
+}
+
+// WeightedEdge returns the current edge of the iterator. Next must have been
+// called prior to a call to WeightedEdge.
+func (e *OrderedWeightedEdges) WeightedEdge() graph.WeightedEdge {
+	if e.idx >= len(e.edges) || e.idx < 0 {
+		return nil
+	}
+	return e.edges[e.idx]
+}
+
+// WeightedEdgeSlice returns all the remaining edges in the iterator and advances
+// the iterator.
+func (e *OrderedWeightedEdges) WeightedEdgeSlice() []graph.WeightedEdge {
+	if e.idx >= len(e.edges) {
+		return nil
+	}
+	idx := e.idx
+	if idx == -1 {
+		idx = 0
+	}
+	e.idx = len(e.edges)
+	return e.edges[idx:]
+}
+
+// Reset returns the iterator to its initial state.
+func (e *OrderedWeightedEdges) Reset() {
+	e.idx = -1
+}

vendor/gonum.org/v1/gonum/graph/iterator/lines.go

@@ -0,0 +1,131 @@
+// Copyright ©2018 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package iterator
+
+import "gonum.org/v1/gonum/graph"
+
+// OrderedLines implements the graph.Lines and graph.LineSlicer interfaces.
+// The iteration order of OrderedLines is the order of lines passed to
+// NewLineIterator.
+type OrderedLines struct {
+	idx   int
+	lines []graph.Line
+}
+
+// NewOrderedLines returns an OrderedLines initialized with the provided lines.
+func NewOrderedLines(lines []graph.Line) *OrderedLines {
+	return &OrderedLines{idx: -1, lines: lines}
+}
+
+// Len returns the remaining number of lines to be iterated over.
+func (e *OrderedLines) Len() int {
+	if e.idx >= len(e.lines) {
+		return 0
+	}
+	if e.idx <= 0 {
+		return len(e.lines)
+	}
+	return len(e.lines[e.idx:])
+}
+
+// Next returns whether the next call of Line will return a valid line.
+func (e *OrderedLines) Next() bool {
+	if uint(e.idx)+1 < uint(len(e.lines)) {
+		e.idx++
+		return true
+	}
+	e.idx = len(e.lines)
+	return false
+}
+
+// Line returns the current line of the iterator. Next must have been
+// called prior to a call to Line.
+func (e *OrderedLines) Line() graph.Line {
+	if e.idx >= len(e.lines) || e.idx < 0 {
+		return nil
+	}
+	return e.lines[e.idx]
+}
+
+// LineSlice returns all the remaining lines in the iterator and advances
+// the iterator.
+func (e *OrderedLines) LineSlice() []graph.Line {
+	if e.idx >= len(e.lines) {
+		return nil
+	}
+	idx := e.idx
+	if idx == -1 {
+		idx = 0
+	}
+	e.idx = len(e.lines)
+	return e.lines[idx:]
+}
+
+// Reset returns the iterator to its initial state.
+func (e *OrderedLines) Reset() {
+	e.idx = -1
+}
+
+// OrderedWeightedLines implements the graph.Lines and graph.LineSlicer interfaces.
+// The iteration order of OrderedWeightedLines is the order of lines passed to
+// NewLineIterator.
+type OrderedWeightedLines struct {
+	idx   int
+	lines []graph.WeightedLine
+}
+
+// NewWeightedLineIterator returns an OrderedWeightedLines initialized with the provided lines.
+func NewOrderedWeightedLines(lines []graph.WeightedLine) *OrderedWeightedLines {
+	return &OrderedWeightedLines{idx: -1, lines: lines}
+}
+
+// Len returns the remaining number of lines to be iterated over.
+func (e *OrderedWeightedLines) Len() int {
+	if e.idx >= len(e.lines) {
+		return 0
+	}
+	if e.idx <= 0 {
+		return len(e.lines)
+	}
+	return len(e.lines[e.idx:])
+}
+
+// Next returns whether the next call of WeightedLine will return a valid line.
+func (e *OrderedWeightedLines) Next() bool {
+	if uint(e.idx)+1 < uint(len(e.lines)) {
+		e.idx++
+		return true
+	}
+	e.idx = len(e.lines)
+	return false
+}
+
+// WeightedLine returns the current line of the iterator. Next must have been
+// called prior to a call to WeightedLine.
+func (e *OrderedWeightedLines) WeightedLine() graph.WeightedLine {
+	if e.idx >= len(e.lines) || e.idx < 0 {
+		return nil
+	}
+	return e.lines[e.idx]
+}
+
+// WeightedLineSlice returns all the remaining lines in the iterator and advances
+// the iterator.
+func (e *OrderedWeightedLines) WeightedLineSlice() []graph.WeightedLine {
+	if e.idx >= len(e.lines) {
+		return nil
+	}
+	idx := e.idx
+	if idx == -1 {
+		idx = 0
+	}
+	e.idx = len(e.lines)
+	return e.lines[idx:]
+}
+
+// Reset returns the iterator to its initial state.
+func (e *OrderedWeightedLines) Reset() {
+	e.idx = -1
+}

vendor/gonum.org/v1/gonum/graph/iterator/nodes.go

@@ -0,0 +1,125 @@
+// Copyright ©2018 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package iterator
+
+import "gonum.org/v1/gonum/graph"
+
+// OrderedNodes implements the graph.Nodes and graph.NodeSlicer interfaces.
+// The iteration order of OrderedNodes is the order of nodes passed to
+// NewNodeIterator.
+type OrderedNodes struct {
+	idx   int
+	nodes []graph.Node
+}
+
+// NewOrderedNodes returns a OrderedNodes initialized with the provided nodes.
+func NewOrderedNodes(nodes []graph.Node) *OrderedNodes {
+	return &OrderedNodes{idx: -1, nodes: nodes}
+}
+
+// Len returns the remaining number of nodes to be iterated over.
+func (n *OrderedNodes) Len() int {
+	if n.idx >= len(n.nodes) {
+		return 0
+	}
+	if n.idx <= 0 {
+		return len(n.nodes)
+	}
+	return len(n.nodes[n.idx:])
+}
+
+// Next returns whether the next call of Node will return a valid node.
+func (n *OrderedNodes) Next() bool {
+	if uint(n.idx)+1 < uint(len(n.nodes)) {
+		n.idx++
+		return true
+	}
+	n.idx = len(n.nodes)
+	return false
+}
+
+// Node returns the current node of the iterator. Next must have been
+// called prior to a call to Node.
+func (n *OrderedNodes) Node() graph.Node {
+	if n.idx >= len(n.nodes) || n.idx < 0 {
+		return nil
+	}
+	return n.nodes[n.idx]
+}
+
+// NodeSlice returns all the remaining nodes in the iterator and advances
+// the iterator.
+func (n *OrderedNodes) NodeSlice() []graph.Node {
+	if n.idx >= len(n.nodes) {
+		return nil
+	}
+	idx := n.idx
+	if idx == -1 {
+		idx = 0
+	}
+	n.idx = len(n.nodes)
+	return n.nodes[idx:]
+}
+
+// Reset returns the iterator to its initial state.
+func (n *OrderedNodes) Reset() {
+	n.idx = -1
+}
+
+// ImplicitNodes implements the graph.Nodes interface for a set of nodes over
+// a contiguous ID range.
+type ImplicitNodes struct {
+	beg, end int
+	curr     int
+	newNode  func(id int) graph.Node
+}
+
+// NewImplicitNodes returns a new implicit node iterator spanning nodes in [beg,end).
+// The provided new func maps the id to a graph.Node. NewImplicitNodes will panic
+// if beg is greater than end.
+func NewImplicitNodes(beg, end int, new func(id int) graph.Node) *ImplicitNodes {
+	if beg > end {
+		panic("iterator: invalid range")
+	}
+	return &ImplicitNodes{beg: beg, end: end, curr: beg - 1, newNode: new}
+}
+
+// Len returns the remaining number of nodes to be iterated over.
+func (n *ImplicitNodes) Len() int {
+	return n.end - n.curr - 1
+}
+
+// Next returns whether the next call of Node will return a valid node.
+func (n *ImplicitNodes) Next() bool {
+	if n.curr == n.end {
+		return false
+	}
+	n.curr++
+	return n.curr < n.end
+}
+
+// Node returns the current node of the iterator. Next must have been
+// called prior to a call to Node.
+func (n *ImplicitNodes) Node() graph.Node {
+	if n.Len() == -1 || n.curr < n.beg {
+		return nil
+	}
+	return n.newNode(n.curr)
+}
+
+// Reset returns the iterator to its initial state.
+func (n *ImplicitNodes) Reset() {
+	n.curr = n.beg - 1
+}
+
+// NodeSlice returns all the remaining nodes in the iterator and advances
+// the iterator.
+func (n *ImplicitNodes) NodeSlice() []graph.Node {
+	nodes := make([]graph.Node, 0, n.Len())
+	for n.curr++; n.curr < n.end; n.curr++ {
+		nodes = append(nodes, n.newNode(n.curr))
+	}
+	return nodes
+}

vendor/gonum.org/v1/gonum/graph/multigraph.go

@@ -0,0 +1,198 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package graph
+
+// Line is an edge in a multigraph. A Line returns an ID that must
+// distinguish Lines sharing Node end points.
+type Line interface {
+	// From returns the from node of the edge.
+	From() Node
+
+	// To returns the to node of the edge.
+	To() Node
+
+	// ReversedLine returns a line that has the
+	// end points of the receiver swapped.
+	ReversedLine() Line
+
+	// ID returns the unique ID for the Line.
+	ID() int64
+}
+
+// WeightedLine is a weighted multigraph edge.
+type WeightedLine interface {
+	Line
+	Weight() float64
+}
+
+// Multigraph is a generalized multigraph.
+type Multigraph interface {
+	// Node returns the node with the given ID if it exists
+	// in the multigraph, and nil otherwise.
+	Node(id int64) Node
+
+	// Nodes returns all the nodes in the multigraph.
+	//
+	// Nodes must not return nil.
+	Nodes() Nodes
+
+	// From returns all nodes that can be reached directly
+	// from the node with the given ID.
+	//
+	// From must not return nil.
+	From(id int64) Nodes
+
+	// HasEdgeBetween returns whether an edge exists between
+	// nodes with IDs xid and yid without considering direction.
+	HasEdgeBetween(xid, yid int64) bool
+
+	// Lines returns the lines from u to v, with IDs uid and
+	// vid, if any such lines exist and nil otherwise. The
+	// node v must be directly reachable from u as defined by
+	// the From method.
+	//
+	// Lines must not return nil.
+	Lines(uid, vid int64) Lines
+}
+
+// WeightedMultigraph is a weighted multigraph.
+type WeightedMultigraph interface {
+	Multigraph
+
+	// WeightedLines returns the weighted lines from u to v
+	// with IDs uid and vid if any such lines exist and nil
+	// otherwise. The node v must be directly reachable
+	// from u as defined by the From method.
+	//
+	// WeightedLines must not return nil.
+	WeightedLines(uid, vid int64) WeightedLines
+}
+
+// UndirectedMultigraph is an undirected multigraph.
+type UndirectedMultigraph interface {
+	Multigraph
+
+	// LinesBetween returns the lines between nodes x and y
+	// with IDs xid and yid.
+	//
+	// LinesBetween must not return nil.
+	LinesBetween(xid, yid int64) Lines
+}
+
+// WeightedUndirectedMultigraph is a weighted undirected multigraph.
+type WeightedUndirectedMultigraph interface {
+	WeightedMultigraph
+
+	// WeightedLinesBetween returns the lines between nodes
+	// x and y with IDs xid and yid.
+	//
+	// WeightedLinesBetween must not return nil.
+	WeightedLinesBetween(xid, yid int64) WeightedLines
+}
+
+// DirectedMultigraph is a directed multigraph.
+type DirectedMultigraph interface {
+	Multigraph
+
+	// HasEdgeFromTo returns whether an edge exists
+	// in the multigraph from u to v with IDs uid
+	// and vid.
+	HasEdgeFromTo(uid, vid int64) bool
+
+	// To returns all nodes that can reach directly
+	// to the node with the given ID.
+	//
+	// To must not return nil.
+	To(id int64) Nodes
+}
+
+// WeightedDirectedMultigraph is a weighted directed multigraph.
+type WeightedDirectedMultigraph interface {
+	WeightedMultigraph
+
+	// HasEdgeFromTo returns whether an edge exists
+	// in the multigraph from u to v with IDs uid
+	// and vid.
+	HasEdgeFromTo(uid, vid int64) bool
+
+	// To returns all nodes that can reach directly
+	// to the node with the given ID.
+	//
+	// To must not return nil.
+	To(id int64) Nodes
+}
+
+// LineAdder is an interface for adding lines to a multigraph.
+type LineAdder interface {
+	// NewLine returns a new Line from the source to the destination node.
+	NewLine(from, to Node) Line
+
+	// SetLine adds a Line from one node to another.
+	// If the multigraph supports node addition the nodes
+	// will be added if they do not exist, otherwise
+	// SetLine will panic.
+	// Whether l, l.From() and l.To() are stored
+	// within the graph is implementation dependent.
+	SetLine(l Line)
+}
+
+// WeightedLineAdder is an interface for adding lines to a multigraph.
+type WeightedLineAdder interface {
+	// NewWeightedLine returns a new WeightedLine from
+	// the source to the destination node.
+	NewWeightedLine(from, to Node, weight float64) WeightedLine
+
+	// SetWeightedLine adds a weighted line from one node
+	// to another. If the multigraph supports node addition
+	// the nodes will be added if they do not exist,
+	// otherwise SetWeightedLine will panic.
+	// Whether l, l.From() and l.To() are stored
+	// within the graph is implementation dependent.
+	SetWeightedLine(l WeightedLine)
+}
+
+// LineRemover is an interface for removing lines from a multigraph.
+type LineRemover interface {
+	// RemoveLine removes the line with the given end
+	// and line IDs, leaving the terminal nodes. If
+	// the line does not exist it is a no-op.
+	RemoveLine(fid, tid, id int64)
+}
+
+// MultigraphBuilder is a multigraph that can have nodes and lines added.
+type MultigraphBuilder interface {
+	NodeAdder
+	LineAdder
+}
+
+// WeightedMultigraphBuilder is a multigraph that can have nodes and weighted lines added.
+type WeightedMultigraphBuilder interface {
+	NodeAdder
+	WeightedLineAdder
+}
+
+// UndirectedMultgraphBuilder is an undirected multigraph builder.
+type UndirectedMultigraphBuilder interface {
+	UndirectedMultigraph
+	MultigraphBuilder
+}
+
+// UndirectedWeightedMultigraphBuilder is an undirected weighted multigraph builder.
+type UndirectedWeightedMultigraphBuilder interface {
+	UndirectedMultigraph
+	WeightedMultigraphBuilder
+}
+
+// DirectedMultigraphBuilder is a directed multigraph builder.
+type DirectedMultigraphBuilder interface {
+	DirectedMultigraph
+	MultigraphBuilder
+}
+
+// DirectedWeightedMultigraphBuilder is a directed weighted multigraph builder.
+type DirectedWeightedMultigraphBuilder interface {
+	DirectedMultigraph
+	WeightedMultigraphBuilder
+}

vendor/gonum.org/v1/gonum/graph/nodes_edges.go

@@ -0,0 +1,300 @@
+// Copyright ©2018 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package graph
+
+// Iterator is an item iterator.
+type Iterator interface {
+	// Next advances the iterator and returns whether
+	// the next call to the item method will return a
+	// non-nil item.
+	//
+	// Next should be called prior to any call to the
+	// iterator's item retrieval method after the
+	// iterator has been obtained or reset.
+	//
+	// The order of iteration is implementation
+	// dependent.
+	Next() bool
+
+	// Len returns the number of items remaining in the
+	// iterator.
+	//
+	// If the number of items in the iterator is unknown,
+	// too large to materialize or too costly to calculate
+	// then Len may return a negative value.
+	// In this case the consuming function must be able
+	// to operate on the items of the iterator directly
+	// without materializing the items into a slice.
+	// The magnitude of a negative length has
+	// implementation-dependent semantics.
+	Len() int
+
+	// Reset returns the iterator to its start position.
+	Reset()
+}
+
+// Nodes is a Node iterator.
+type Nodes interface {
+	Iterator
+
+	// Node returns the current Node from the iterator.
+	Node() Node
+}
+
+// NodeSlicer wraps the NodeSlice method.
+type NodeSlicer interface {
+	// NodeSlice returns the set of nodes remaining
+	// to be iterated by a Nodes iterator.
+	// The holder of the iterator may arbitrarily
+	// change elements in the returned slice, but
+	// those changes may be reflected to other
+	// iterators.
+	NodeSlice() []Node
+}
+
+// NodesOf returns it.Len() nodes from it. If it is a NodeSlicer, the NodeSlice method
+// is used to obtain the nodes. It is safe to pass a nil Nodes to NodesOf.
+//
+// If the Nodes has an indeterminate length, NodesOf will panic.
+func NodesOf(it Nodes) []Node {
+	if it == nil {
+		return nil
+	}
+	len := it.Len()
+	switch {
+	case len == 0:
+		return nil
+	case len < 0:
+		panic("graph: called NodesOf on indeterminate iterator")
+	}
+	switch it := it.(type) {
+	case NodeSlicer:
+		return it.NodeSlice()
+	}
+	n := make([]Node, 0, len)
+	for it.Next() {
+		n = append(n, it.Node())
+	}
+	return n
+}
+
+// Edges is an Edge iterator.
+type Edges interface {
+	Iterator
+
+	// Edge returns the current Edge from the iterator.
+	Edge() Edge
+}
+
+// EdgeSlicer wraps the EdgeSlice method.
+type EdgeSlicer interface {
+	// EdgeSlice returns the set of edges remaining
+	// to be iterated by an Edges iterator.
+	// The holder of the iterator may arbitrarily
+	// change elements in the returned slice, but
+	// those changes may be reflected to other
+	// iterators.
+	EdgeSlice() []Edge
+}
+
+// EdgesOf returns it.Len() nodes from it. If it is an EdgeSlicer, the EdgeSlice method is used
+// to obtain the edges. It is safe to pass a nil Edges to EdgesOf.
+//
+// If the Edges has an indeterminate length, EdgesOf will panic.
+func EdgesOf(it Edges) []Edge {
+	if it == nil {
+		return nil
+	}
+	len := it.Len()
+	switch {
+	case len == 0:
+		return nil
+	case len < 0:
+		panic("graph: called EdgesOf on indeterminate iterator")
+	}
+	switch it := it.(type) {
+	case EdgeSlicer:
+		return it.EdgeSlice()
+	}
+	e := make([]Edge, 0, len)
+	for it.Next() {
+		e = append(e, it.Edge())
+	}
+	return e
+}
+
+// WeightedEdges is a WeightedEdge iterator.
+type WeightedEdges interface {
+	Iterator
+
+	// Edge returns the current Edge from the iterator.
+	WeightedEdge() WeightedEdge
+}
+
+// WeightedEdgeSlicer wraps the WeightedEdgeSlice method.
+type WeightedEdgeSlicer interface {
+	// EdgeSlice returns the set of edges remaining
+	// to be iterated by an Edges iterator.
+	// The holder of the iterator may arbitrarily
+	// change elements in the returned slice, but
+	// those changes may be reflected to other
+	// iterators.
+	WeightedEdgeSlice() []WeightedEdge
+}
+
+// WeightedEdgesOf returns it.Len() weighted edge from it. If it is a WeightedEdgeSlicer, the
+// WeightedEdgeSlice method is used to obtain the edges. It is safe to pass a nil WeightedEdges
+// to WeightedEdgesOf.
+//
+// If the WeightedEdges has an indeterminate length, WeightedEdgesOf will panic.
+func WeightedEdgesOf(it WeightedEdges) []WeightedEdge {
+	if it == nil {
+		return nil
+	}
+	len := it.Len()
+	switch {
+	case len == 0:
+		return nil
+	case len < 0:
+		panic("graph: called WeightedEdgesOf on indeterminate iterator")
+	}
+	switch it := it.(type) {
+	case WeightedEdgeSlicer:
+		return it.WeightedEdgeSlice()
+	}
+	e := make([]WeightedEdge, 0, len)
+	for it.Next() {
+		e = append(e, it.WeightedEdge())
+	}
+	return e
+}
+
+// Lines is a Line iterator.
+type Lines interface {
+	Iterator
+
+	// Line returns the current Line from the iterator.
+	Line() Line
+}
+
+// LineSlicer wraps the LineSlice method.
+type LineSlicer interface {
+	// LineSlice returns the set of lines remaining
+	// to be iterated by an Lines iterator.
+	// The holder of the iterator may arbitrarily
+	// change elements in the returned slice, but
+	// those changes may be reflected to other
+	// iterators.
+	LineSlice() []Line
+}
+
+// LinesOf returns it.Len() nodes from it. If it is a LineSlicer, the LineSlice method is used
+// to obtain the lines. It is safe to pass a nil Lines to LinesOf.
+//
+// If the Lines has an indeterminate length, LinesOf will panic.
+func LinesOf(it Lines) []Line {
+	if it == nil {
+		return nil
+	}
+	len := it.Len()
+	switch {
+	case len == 0:
+		return nil
+	case len < 0:
+		panic("graph: called LinesOf on indeterminate iterator")
+	}
+	switch it := it.(type) {
+	case LineSlicer:
+		return it.LineSlice()
+	}
+	l := make([]Line, 0, len)
+	for it.Next() {
+		l = append(l, it.Line())
+	}
+	return l
+}
+
+// WeightedLines is a WeightedLine iterator.
+type WeightedLines interface {
+	Iterator
+
+	// Line returns the current Line from the iterator.
+	WeightedLine() WeightedLine
+}
+
+// WeightedLineSlicer wraps the WeightedLineSlice method.
+type WeightedLineSlicer interface {
+	// LineSlice returns the set of lines remaining
+	// to be iterated by an Lines iterator.
+	// The holder of the iterator may arbitrarily
+	// change elements in the returned slice, but
+	// those changes may be reflected to other
+	// iterators.
+	WeightedLineSlice() []WeightedLine
+}
+
+// WeightedLinesOf returns it.Len() weighted line from it. If it is a WeightedLineSlicer, the
+// WeightedLineSlice method is used to obtain the lines. It is safe to pass a nil WeightedLines
+// to WeightedLinesOf.
+//
+// If the WeightedLines has an indeterminate length, WeightedLinesOf will panic.
+func WeightedLinesOf(it WeightedLines) []WeightedLine {
+	if it == nil {
+		return nil
+	}
+	len := it.Len()
+	switch {
+	case len == 0:
+		return nil
+	case len < 0:
+		panic("graph: called WeightedLinesOf on indeterminate iterator")
+	}
+	switch it := it.(type) {
+	case WeightedLineSlicer:
+		return it.WeightedLineSlice()
+	}
+	l := make([]WeightedLine, 0, len)
+	for it.Next() {
+		l = append(l, it.WeightedLine())
+	}
+	return l
+}
+
+// Empty is an empty set of nodes, edges or lines. It should be used when
+// a graph returns a zero-length Iterator. Empty implements the slicer
+// interfaces for nodes, edges and lines, returning nil for each of these.
+const Empty = nothing
+
+var (
+	_ Iterator           = Empty
+	_ Nodes              = Empty
+	_ NodeSlicer         = Empty
+	_ Edges              = Empty
+	_ EdgeSlicer         = Empty
+	_ WeightedEdges      = Empty
+	_ WeightedEdgeSlicer = Empty
+	_ Lines              = Empty
+	_ LineSlicer         = Empty
+	_ WeightedLines      = Empty
+	_ WeightedLineSlicer = Empty
+)
+
+const nothing = empty(true)
+
+type empty bool
+
+func (empty) Next() bool                        { return false }
+func (empty) Len() int                          { return 0 }
+func (empty) Reset()                            {}
+func (empty) Node() Node                        { return nil }
+func (empty) NodeSlice() []Node                 { return nil }
+func (empty) Edge() Edge                        { return nil }
+func (empty) EdgeSlice() []Edge                 { return nil }
+func (empty) WeightedEdge() WeightedEdge        { return nil }
+func (empty) WeightedEdgeSlice() []WeightedEdge { return nil }
+func (empty) Line() Line                        { return nil }
+func (empty) LineSlice() []Line                 { return nil }
+func (empty) WeightedLine() WeightedLine        { return nil }
+func (empty) WeightedLineSlice() []WeightedLine { return nil }

vendor/gonum.org/v1/gonum/graph/simple/dense_directed_matrix.go

@@ -0,0 +1,301 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package simple
+
+import (
+	"sort"
+
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/ordered"
+	"gonum.org/v1/gonum/graph/iterator"
+	"gonum.org/v1/gonum/mat"
+)
+
+var (
+	dm *DirectedMatrix
+
+	_ graph.Graph        = dm
+	_ graph.Directed     = dm
+	_ edgeSetter         = dm
+	_ weightedEdgeSetter = dm
+)
+
+// DirectedMatrix represents a directed graph using an adjacency
+// matrix such that all IDs are in a contiguous block from 0 to n-1.
+// Edges are stored implicitly as an edge weight, so edges stored in
+// the graph are not recoverable.
+type DirectedMatrix struct {
+	mat   *mat.Dense
+	nodes []graph.Node
+
+	self   float64
+	absent float64
+}
+
+// NewDirectedMatrix creates a directed dense graph with n nodes.
+// All edges are initialized with the weight given by init. The self parameter
+// specifies the cost of self connection, and absent specifies the weight
+// returned for absent edges.
+func NewDirectedMatrix(n int, init, self, absent float64) *DirectedMatrix {
+	matrix := make([]float64, n*n)
+	if init != 0 {
+		for i := range matrix {
+			matrix[i] = init
+		}
+	}
+	for i := 0; i < len(matrix); i += n + 1 {
+		matrix[i] = self
+	}
+	return &DirectedMatrix{
+		mat:    mat.NewDense(n, n, matrix),
+		self:   self,
+		absent: absent,
+	}
+}
+
+// NewDirectedMatrixFrom creates a directed dense graph with the given nodes.
+// The IDs of the nodes must be contiguous from 0 to len(nodes)-1, but may
+// be in any order. If IDs are not contiguous NewDirectedMatrixFrom will panic.
+// All edges are initialized with the weight given by init. The self parameter
+// specifies the cost of self connection, and absent specifies the weight
+// returned for absent edges.
+func NewDirectedMatrixFrom(nodes []graph.Node, init, self, absent float64) *DirectedMatrix {
+	sort.Sort(ordered.ByID(nodes))
+	for i, n := range nodes {
+		if int64(i) != n.ID() {
+			panic("simple: non-contiguous node IDs")
+		}
+	}
+	g := NewDirectedMatrix(len(nodes), init, self, absent)
+	g.nodes = nodes
+	return g
+}
+
+// Edge returns the edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+func (g *DirectedMatrix) Edge(uid, vid int64) graph.Edge {
+	return g.WeightedEdge(uid, vid)
+}
+
+// Edges returns all the edges in the graph.
+func (g *DirectedMatrix) Edges() graph.Edges {
+	var edges []graph.Edge
+	r, _ := g.mat.Dims()
+	for i := 0; i < r; i++ {
+		for j := 0; j < r; j++ {
+			if i == j {
+				continue
+			}
+			if w := g.mat.At(i, j); !isSame(w, g.absent) {
+				edges = append(edges, WeightedEdge{F: g.Node(int64(i)), T: g.Node(int64(j)), W: w})
+			}
+		}
+	}
+	if len(edges) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedEdges(edges)
+}
+
+// From returns all nodes in g that can be reached directly from n.
+func (g *DirectedMatrix) From(id int64) graph.Nodes {
+	if !g.has(id) {
+		return graph.Empty
+	}
+	var nodes []graph.Node
+	_, c := g.mat.Dims()
+	for j := 0; j < c; j++ {
+		if int64(j) == id {
+			continue
+		}
+		// id is not greater than maximum int by this point.
+		if !isSame(g.mat.At(int(id), j), g.absent) {
+			nodes = append(nodes, g.Node(int64(j)))
+		}
+	}
+	if len(nodes) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedNodes(nodes)
+}
+
+// HasEdgeBetween returns whether an edge exists between nodes x and y without
+// considering direction.
+func (g *DirectedMatrix) HasEdgeBetween(xid, yid int64) bool {
+	if !g.has(xid) {
+		return false
+	}
+	if !g.has(yid) {
+		return false
+	}
+	// xid and yid are not greater than maximum int by this point.
+	return xid != yid && (!isSame(g.mat.At(int(xid), int(yid)), g.absent) || !isSame(g.mat.At(int(yid), int(xid)), g.absent))
+}
+
+// HasEdgeFromTo returns whether an edge exists in the graph from u to v.
+func (g *DirectedMatrix) HasEdgeFromTo(uid, vid int64) bool {
+	if !g.has(uid) {
+		return false
+	}
+	if !g.has(vid) {
+		return false
+	}
+	// uid and vid are not greater than maximum int by this point.
+	return uid != vid && !isSame(g.mat.At(int(uid), int(vid)), g.absent)
+}
+
+// Matrix returns the mat.Matrix representation of the graph. The orientation
+// of the matrix is such that the matrix entry at G_{ij} is the weight of the edge
+// from node i to node j.
+func (g *DirectedMatrix) Matrix() mat.Matrix {
+	// Prevent alteration of dimensions of the returned matrix.
+	m := *g.mat
+	return &m
+}
+
+// Node returns the node with the given ID if it exists in the graph,
+// and nil otherwise.
+func (g *DirectedMatrix) Node(id int64) graph.Node {
+	if !g.has(id) {
+		return nil
+	}
+	if g.nodes == nil {
+		return Node(id)
+	}
+	return g.nodes[id]
+}
+
+// Nodes returns all the nodes in the graph.
+func (g *DirectedMatrix) Nodes() graph.Nodes {
+	if g.nodes != nil {
+		nodes := make([]graph.Node, len(g.nodes))
+		copy(nodes, g.nodes)
+		return iterator.NewOrderedNodes(nodes)
+	}
+	r, _ := g.mat.Dims()
+	// Matrix graphs must have at least one node.
+	return iterator.NewImplicitNodes(0, r, newSimpleNode)
+}
+
+// RemoveEdge removes the edge with the given end point nodes from the graph, leaving the terminal
+// nodes. If the edge does not exist it is a no-op.
+func (g *DirectedMatrix) RemoveEdge(fid, tid int64) {
+	if !g.has(fid) {
+		return
+	}
+	if !g.has(tid) {
+		return
+	}
+	// fid and tid are not greater than maximum int by this point.
+	g.mat.Set(int(fid), int(tid), g.absent)
+}
+
+// SetEdge sets e, an edge from one node to another with unit weight. If the ends of the edge
+// are not in g or the edge is a self loop, SetEdge panics. SetEdge will store the nodes of
+// e in the graph if it was initialized with NewDirectedMatrixFrom.
+func (g *DirectedMatrix) SetEdge(e graph.Edge) {
+	g.setWeightedEdge(e, 1)
+}
+
+// SetWeightedEdge sets e, an edge from one node to another. If the ends of the edge are not in g
+// or the edge is a self loop, SetWeightedEdge panics. SetWeightedEdge will store the nodes of
+// e in the graph if it was initialized with NewDirectedMatrixFrom.
+func (g *DirectedMatrix) SetWeightedEdge(e graph.WeightedEdge) {
+	g.setWeightedEdge(e, e.Weight())
+}
+
+func (g *DirectedMatrix) setWeightedEdge(e graph.Edge, weight float64) {
+	from := e.From()
+	fid := from.ID()
+	to := e.To()
+	tid := to.ID()
+	if fid == tid {
+		panic("simple: set illegal edge")
+	}
+	if int64(int(fid)) != fid {
+		panic("simple: unavailable from node ID for dense graph")
+	}
+	if int64(int(tid)) != tid {
+		panic("simple: unavailable to node ID for dense graph")
+	}
+	if g.nodes != nil {
+		g.nodes[fid] = from
+		g.nodes[tid] = to
+	}
+	// fid and tid are not greater than maximum int by this point.
+	g.mat.Set(int(fid), int(tid), weight)
+}
+
+// To returns all nodes in g that can reach directly to n.
+func (g *DirectedMatrix) To(id int64) graph.Nodes {
+	if !g.has(id) {
+		return graph.Empty
+	}
+	var nodes []graph.Node
+	r, _ := g.mat.Dims()
+	for i := 0; i < r; i++ {
+		if int64(i) == id {
+			continue
+		}
+		// id is not greater than maximum int by this point.
+		if !isSame(g.mat.At(i, int(id)), g.absent) {
+			nodes = append(nodes, g.Node(int64(i)))
+		}
+	}
+	if len(nodes) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedNodes(nodes)
+}
+
+// Weight returns the weight for the edge between x and y if Edge(x, y) returns a non-nil Edge.
+// If x and y are the same node or there is no joining edge between the two nodes the weight
+// value returned is either the graph's absent or self value. Weight returns true if an edge
+// exists between x and y or if x and y have the same ID, false otherwise.
+func (g *DirectedMatrix) Weight(xid, yid int64) (w float64, ok bool) {
+	if xid == yid {
+		return g.self, true
+	}
+	if g.HasEdgeFromTo(xid, yid) {
+		// xid and yid are not greater than maximum int by this point.
+		return g.mat.At(int(xid), int(yid)), true
+	}
+	return g.absent, false
+}
+
+// WeightedEdge returns the weighted edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+func (g *DirectedMatrix) WeightedEdge(uid, vid int64) graph.WeightedEdge {
+	if g.HasEdgeFromTo(uid, vid) {
+		// xid and yid are not greater than maximum int by this point.
+		return WeightedEdge{F: g.Node(uid), T: g.Node(vid), W: g.mat.At(int(uid), int(vid))}
+	}
+	return nil
+}
+
+// WeightedEdges returns all the edges in the graph.
+func (g *DirectedMatrix) WeightedEdges() graph.WeightedEdges {
+	var edges []graph.WeightedEdge
+	r, _ := g.mat.Dims()
+	for i := 0; i < r; i++ {
+		for j := 0; j < r; j++ {
+			if i == j {
+				continue
+			}
+			if w := g.mat.At(i, j); !isSame(w, g.absent) {
+				edges = append(edges, WeightedEdge{F: g.Node(int64(i)), T: g.Node(int64(j)), W: w})
+			}
+		}
+	}
+	if len(edges) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedWeightedEdges(edges)
+}
+
+func (g *DirectedMatrix) has(id int64) bool {
+	r, _ := g.mat.Dims()
+	return 0 <= id && id < int64(r)
+}

vendor/gonum.org/v1/gonum/graph/simple/dense_undirected_matrix.go

@@ -0,0 +1,268 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package simple
+
+import (
+	"sort"
+
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/ordered"
+	"gonum.org/v1/gonum/graph/iterator"
+	"gonum.org/v1/gonum/mat"
+)
+
+var (
+	um *UndirectedMatrix
+
+	_ graph.Graph        = um
+	_ graph.Undirected   = um
+	_ edgeSetter         = um
+	_ weightedEdgeSetter = um
+)
+
+// UndirectedMatrix represents an undirected graph using an adjacency
+// matrix such that all IDs are in a contiguous block from 0 to n-1.
+// Edges are stored implicitly as an edge weight, so edges stored in
+// the graph are not recoverable.
+type UndirectedMatrix struct {
+	mat   *mat.SymDense
+	nodes []graph.Node
+
+	self   float64
+	absent float64
+}
+
+// NewUndirectedMatrix creates an undirected dense graph with n nodes.
+// All edges are initialized with the weight given by init. The self parameter
+// specifies the cost of self connection, and absent specifies the weight
+// returned for absent edges.
+func NewUndirectedMatrix(n int, init, self, absent float64) *UndirectedMatrix {
+	matrix := make([]float64, n*n)
+	if init != 0 {
+		for i := range matrix {
+			matrix[i] = init
+		}
+	}
+	for i := 0; i < len(matrix); i += n + 1 {
+		matrix[i] = self
+	}
+	return &UndirectedMatrix{
+		mat:    mat.NewSymDense(n, matrix),
+		self:   self,
+		absent: absent,
+	}
+}
+
+// NewUndirectedMatrixFrom creates an undirected dense graph with the given nodes.
+// The IDs of the nodes must be contiguous from 0 to len(nodes)-1, but may
+// be in any order. If IDs are not contiguous NewUndirectedMatrixFrom will panic.
+// All edges are initialized with the weight given by init. The self parameter
+// specifies the cost of self connection, and absent specifies the weight
+// returned for absent edges.
+func NewUndirectedMatrixFrom(nodes []graph.Node, init, self, absent float64) *UndirectedMatrix {
+	sort.Sort(ordered.ByID(nodes))
+	for i, n := range nodes {
+		if int64(i) != n.ID() {
+			panic("simple: non-contiguous node IDs")
+		}
+	}
+	g := NewUndirectedMatrix(len(nodes), init, self, absent)
+	g.nodes = nodes
+	return g
+}
+
+// Edge returns the edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+func (g *UndirectedMatrix) Edge(uid, vid int64) graph.Edge {
+	return g.WeightedEdgeBetween(uid, vid)
+}
+
+// EdgeBetween returns the edge between nodes x and y.
+func (g *UndirectedMatrix) EdgeBetween(uid, vid int64) graph.Edge {
+	return g.WeightedEdgeBetween(uid, vid)
+}
+
+// Edges returns all the edges in the graph.
+func (g *UndirectedMatrix) Edges() graph.Edges {
+	var edges []graph.Edge
+	r, _ := g.mat.Dims()
+	for i := 0; i < r; i++ {
+		for j := i + 1; j < r; j++ {
+			if w := g.mat.At(i, j); !isSame(w, g.absent) {
+				edges = append(edges, WeightedEdge{F: g.Node(int64(i)), T: g.Node(int64(j)), W: w})
+			}
+		}
+	}
+	if len(edges) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedEdges(edges)
+}
+
+// From returns all nodes in g that can be reached directly from n.
+func (g *UndirectedMatrix) From(id int64) graph.Nodes {
+	if !g.has(id) {
+		return graph.Empty
+	}
+	var nodes []graph.Node
+	r := g.mat.Symmetric()
+	for i := 0; i < r; i++ {
+		if int64(i) == id {
+			continue
+		}
+		// id is not greater than maximum int by this point.
+		if !isSame(g.mat.At(int(id), i), g.absent) {
+			nodes = append(nodes, g.Node(int64(i)))
+		}
+	}
+	if len(nodes) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedNodes(nodes)
+}
+
+// HasEdgeBetween returns whether an edge exists between nodes x and y.
+func (g *UndirectedMatrix) HasEdgeBetween(uid, vid int64) bool {
+	if !g.has(uid) {
+		return false
+	}
+	if !g.has(vid) {
+		return false
+	}
+	// uid and vid are not greater than maximum int by this point.
+	return uid != vid && !isSame(g.mat.At(int(uid), int(vid)), g.absent)
+}
+
+// Matrix returns the mat.Matrix representation of the graph.
+func (g *UndirectedMatrix) Matrix() mat.Matrix {
+	// Prevent alteration of dimensions of the returned matrix.
+	m := *g.mat
+	return &m
+}
+
+// Node returns the node with the given ID if it exists in the graph,
+// and nil otherwise.
+func (g *UndirectedMatrix) Node(id int64) graph.Node {
+	if !g.has(id) {
+		return nil
+	}
+	if g.nodes == nil {
+		return Node(id)
+	}
+	return g.nodes[id]
+}
+
+// Nodes returns all the nodes in the graph.
+func (g *UndirectedMatrix) Nodes() graph.Nodes {
+	if g.nodes != nil {
+		nodes := make([]graph.Node, len(g.nodes))
+		copy(nodes, g.nodes)
+		return iterator.NewOrderedNodes(nodes)
+	}
+	r := g.mat.Symmetric()
+	// Matrix graphs must have at least one node.
+	return iterator.NewImplicitNodes(0, r, newSimpleNode)
+}
+
+// RemoveEdge removes the edge with the given end point IDs from the graph, leaving the terminal
+// nodes. If the edge does not exist it is a no-op.
+func (g *UndirectedMatrix) RemoveEdge(fid, tid int64) {
+	if !g.has(fid) {
+		return
+	}
+	if !g.has(tid) {
+		return
+	}
+	// fid and tid are not greater than maximum int by this point.
+	g.mat.SetSym(int(fid), int(tid), g.absent)
+}
+
+// SetEdge sets e, an edge from one node to another with unit weight. If the ends of the edge are
+// not in g or the edge is a self loop, SetEdge panics. SetEdge will store the nodes of
+// e in the graph if it was initialized with NewUndirectedMatrixFrom.
+func (g *UndirectedMatrix) SetEdge(e graph.Edge) {
+	g.setWeightedEdge(e, 1)
+}
+
+// SetWeightedEdge sets e, an edge from one node to another. If the ends of the edge are not in g
+// or the edge is a self loop, SetWeightedEdge panics. SetWeightedEdge will store the nodes of
+// e in the graph if it was initialized with NewUndirectedMatrixFrom.
+func (g *UndirectedMatrix) SetWeightedEdge(e graph.WeightedEdge) {
+	g.setWeightedEdge(e, e.Weight())
+}
+
+func (g *UndirectedMatrix) setWeightedEdge(e graph.Edge, weight float64) {
+	from := e.From()
+	fid := from.ID()
+	to := e.To()
+	tid := to.ID()
+	if fid == tid {
+		panic("simple: set illegal edge")
+	}
+	if int64(int(fid)) != fid {
+		panic("simple: unavailable from node ID for dense graph")
+	}
+	if int64(int(tid)) != tid {
+		panic("simple: unavailable to node ID for dense graph")
+	}
+	if g.nodes != nil {
+		g.nodes[fid] = from
+		g.nodes[tid] = to
+	}
+	// fid and tid are not greater than maximum int by this point.
+	g.mat.SetSym(int(fid), int(tid), weight)
+}
+
+// Weight returns the weight for the edge between x and y if Edge(x, y) returns a non-nil Edge.
+// If x and y are the same node or there is no joining edge between the two nodes the weight
+// value returned is either the graph's absent or self value. Weight returns true if an edge
+// exists between x and y or if x and y have the same ID, false otherwise.
+func (g *UndirectedMatrix) Weight(xid, yid int64) (w float64, ok bool) {
+	if xid == yid {
+		return g.self, true
+	}
+	if g.HasEdgeBetween(xid, yid) {
+		// xid and yid are not greater than maximum int by this point.
+		return g.mat.At(int(xid), int(yid)), true
+	}
+	return g.absent, false
+}
+
+// WeightedEdge returns the weighted edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+func (g *UndirectedMatrix) WeightedEdge(uid, vid int64) graph.WeightedEdge {
+	return g.WeightedEdgeBetween(uid, vid)
+}
+
+// WeightedEdgeBetween returns the weighted edge between nodes x and y.
+func (g *UndirectedMatrix) WeightedEdgeBetween(uid, vid int64) graph.WeightedEdge {
+	if g.HasEdgeBetween(uid, vid) {
+		// uid and vid are not greater than maximum int by this point.
+		return WeightedEdge{F: g.Node(uid), T: g.Node(vid), W: g.mat.At(int(uid), int(vid))}
+	}
+	return nil
+}
+
+// WeightedEdges returns all the edges in the graph.
+func (g *UndirectedMatrix) WeightedEdges() graph.WeightedEdges {
+	var edges []graph.WeightedEdge
+	r, _ := g.mat.Dims()
+	for i := 0; i < r; i++ {
+		for j := i + 1; j < r; j++ {
+			if w := g.mat.At(i, j); !isSame(w, g.absent) {
+				edges = append(edges, WeightedEdge{F: g.Node(int64(i)), T: g.Node(int64(j)), W: w})
+			}
+		}
+	}
+	if len(edges) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedWeightedEdges(edges)
+}
+
+func (g *UndirectedMatrix) has(id int64) bool {
+	r := g.mat.Symmetric()
+	return 0 <= id && id < int64(r)
+}

vendor/gonum.org/v1/gonum/graph/simple/directed.go

@@ -0,0 +1,235 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package simple
+
+import (
+	"fmt"
+
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/uid"
+	"gonum.org/v1/gonum/graph/iterator"
+)
+
+var (
+	dg *DirectedGraph
+
+	_ graph.Graph       = dg
+	_ graph.Directed    = dg
+	_ graph.NodeAdder   = dg
+	_ graph.NodeRemover = dg
+	_ graph.EdgeAdder   = dg
+	_ graph.EdgeRemover = dg
+)
+
+// DirectedGraph implements a generalized directed graph.
+type DirectedGraph struct {
+	nodes map[int64]graph.Node
+	from  map[int64]map[int64]graph.Edge
+	to    map[int64]map[int64]graph.Edge
+
+	nodeIDs uid.Set
+}
+
+// NewDirectedGraph returns a DirectedGraph.
+func NewDirectedGraph() *DirectedGraph {
+	return &DirectedGraph{
+		nodes: make(map[int64]graph.Node),
+		from:  make(map[int64]map[int64]graph.Edge),
+		to:    make(map[int64]map[int64]graph.Edge),
+
+		nodeIDs: uid.NewSet(),
+	}
+}
+
+// AddNode adds n to the graph. It panics if the added node ID matches an existing node ID.
+func (g *DirectedGraph) AddNode(n graph.Node) {
+	if _, exists := g.nodes[n.ID()]; exists {
+		panic(fmt.Sprintf("simple: node ID collision: %d", n.ID()))
+	}
+	g.nodes[n.ID()] = n
+	g.from[n.ID()] = make(map[int64]graph.Edge)
+	g.to[n.ID()] = make(map[int64]graph.Edge)
+	g.nodeIDs.Use(n.ID())
+}
+
+// Edge returns the edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+func (g *DirectedGraph) Edge(uid, vid int64) graph.Edge {
+	edge, ok := g.from[uid][vid]
+	if !ok {
+		return nil
+	}
+	return edge
+}
+
+// Edges returns all the edges in the graph.
+func (g *DirectedGraph) Edges() graph.Edges {
+	var edges []graph.Edge
+	for _, u := range g.nodes {
+		for _, e := range g.from[u.ID()] {
+			edges = append(edges, e)
+		}
+	}
+	if len(edges) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedEdges(edges)
+}
+
+// From returns all nodes in g that can be reached directly from n.
+func (g *DirectedGraph) From(id int64) graph.Nodes {
+	if _, ok := g.from[id]; !ok {
+		return graph.Empty
+	}
+
+	from := make([]graph.Node, len(g.from[id]))
+	i := 0
+	for vid := range g.from[id] {
+		from[i] = g.nodes[vid]
+		i++
+	}
+	if len(from) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedNodes(from)
+}
+
+// HasEdgeBetween returns whether an edge exists between nodes x and y without
+// considering direction.
+func (g *DirectedGraph) HasEdgeBetween(xid, yid int64) bool {
+	if _, ok := g.from[xid][yid]; ok {
+		return true
+	}
+	_, ok := g.from[yid][xid]
+	return ok
+}
+
+// HasEdgeFromTo returns whether an edge exists in the graph from u to v.
+func (g *DirectedGraph) HasEdgeFromTo(uid, vid int64) bool {
+	if _, ok := g.from[uid][vid]; !ok {
+		return false
+	}
+	return true
+}
+
+// NewEdge returns a new Edge from the source to the destination node.
+func (g *DirectedGraph) NewEdge(from, to graph.Node) graph.Edge {
+	return &Edge{F: from, T: to}
+}
+
+// NewNode returns a new unique Node to be added to g. The Node's ID does
+// not become valid in g until the Node is added to g.
+func (g *DirectedGraph) NewNode() graph.Node {
+	if len(g.nodes) == 0 {
+		return Node(0)
+	}
+	if int64(len(g.nodes)) == uid.Max {
+		panic("simple: cannot allocate node: no slot")
+	}
+	return Node(g.nodeIDs.NewID())
+}
+
+// Node returns the node with the given ID if it exists in the graph,
+// and nil otherwise.
+func (g *DirectedGraph) Node(id int64) graph.Node {
+	return g.nodes[id]
+}
+
+// Nodes returns all the nodes in the graph.
+func (g *DirectedGraph) Nodes() graph.Nodes {
+	if len(g.nodes) == 0 {
+		return graph.Empty
+	}
+	nodes := make([]graph.Node, len(g.nodes))
+	i := 0
+	for _, n := range g.nodes {
+		nodes[i] = n
+		i++
+	}
+	return iterator.NewOrderedNodes(nodes)
+}
+
+// RemoveEdge removes the edge with the given end point IDs from the graph, leaving the terminal
+// nodes. If the edge does not exist it is a no-op.
+func (g *DirectedGraph) RemoveEdge(fid, tid int64) {
+	if _, ok := g.nodes[fid]; !ok {
+		return
+	}
+	if _, ok := g.nodes[tid]; !ok {
+		return
+	}
+
+	delete(g.from[fid], tid)
+	delete(g.to[tid], fid)
+}
+
+// RemoveNode removes the node with the given ID from the graph, as well as any edges attached
+// to it. If the node is not in the graph it is a no-op.
+func (g *DirectedGraph) RemoveNode(id int64) {
+	if _, ok := g.nodes[id]; !ok {
+		return
+	}
+	delete(g.nodes, id)
+
+	for from := range g.from[id] {
+		delete(g.to[from], id)
+	}
+	delete(g.from, id)
+
+	for to := range g.to[id] {
+		delete(g.from[to], id)
+	}
+	delete(g.to, id)
+
+	g.nodeIDs.Release(id)
+}
+
+// SetEdge adds e, an edge from one node to another. If the nodes do not exist, they are added
+// and are set to the nodes of the edge otherwise.
+// It will panic if the IDs of the e.From and e.To are equal.
+func (g *DirectedGraph) SetEdge(e graph.Edge) {
+	var (
+		from = e.From()
+		fid  = from.ID()
+		to   = e.To()
+		tid  = to.ID()
+	)
+
+	if fid == tid {
+		panic("simple: adding self edge")
+	}
+
+	if _, ok := g.nodes[fid]; !ok {
+		g.AddNode(from)
+	} else {
+		g.nodes[fid] = from
+	}
+	if _, ok := g.nodes[tid]; !ok {
+		g.AddNode(to)
+	} else {
+		g.nodes[tid] = to
+	}
+
+	g.from[fid][tid] = e
+	g.to[tid][fid] = e
+}
+
+// To returns all nodes in g that can reach directly to n.
+func (g *DirectedGraph) To(id int64) graph.Nodes {
+	if _, ok := g.from[id]; !ok {
+		return graph.Empty
+	}
+
+	to := make([]graph.Node, len(g.to[id]))
+	i := 0
+	for uid := range g.to[id] {
+		to[i] = g.nodes[uid]
+		i++
+	}
+	if len(to) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedNodes(to)
+}

vendor/gonum.org/v1/gonum/graph/simple/doc.go

@@ -0,0 +1,9 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package simple provides a suite of simple graph implementations satisfying
+// the gonum/graph interfaces.
+//
+// All types in simple return the graph.Empty value for empty iterators.
+package simple // import "gonum.org/v1/gonum/graph/simple"

vendor/gonum.org/v1/gonum/graph/simple/simple.go

@@ -0,0 +1,72 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package simple
+
+import (
+	"math"
+
+	"gonum.org/v1/gonum/graph"
+)
+
+// Node is a simple graph node.
+type Node int64
+
+// ID returns the ID number of the node.
+func (n Node) ID() int64 {
+	return int64(n)
+}
+
+func newSimpleNode(id int) graph.Node {
+	return Node(id)
+}
+
+// Edge is a simple graph edge.
+type Edge struct {
+	F, T graph.Node
+}
+
+// From returns the from-node of the edge.
+func (e Edge) From() graph.Node { return e.F }
+
+// To returns the to-node of the edge.
+func (e Edge) To() graph.Node { return e.T }
+
+// ReversedLine returns a new Edge with the F and T fields
+// swapped.
+func (e Edge) ReversedEdge() graph.Edge { return Edge{F: e.T, T: e.F} }
+
+// WeightedEdge is a simple weighted graph edge.
+type WeightedEdge struct {
+	F, T graph.Node
+	W    float64
+}
+
+// From returns the from-node of the edge.
+func (e WeightedEdge) From() graph.Node { return e.F }
+
+// To returns the to-node of the edge.
+func (e WeightedEdge) To() graph.Node { return e.T }
+
+// ReversedLine returns a new Edge with the F and T fields
+// swapped. The weight of the new Edge is the same as
+// the weight of the receiver.
+func (e WeightedEdge) ReversedEdge() graph.Edge { return WeightedEdge{F: e.T, T: e.F, W: e.W} }
+
+// Weight returns the weight of the edge.
+func (e WeightedEdge) Weight() float64 { return e.W }
+
+// isSame returns whether two float64 values are the same where NaN values
+// are equalable.
+func isSame(a, b float64) bool {
+	return a == b || (math.IsNaN(a) && math.IsNaN(b))
+}
+
+type edgeSetter interface {
+	SetEdge(e graph.Edge)
+}
+
+type weightedEdgeSetter interface {
+	SetWeightedEdge(e graph.WeightedEdge)
+}

vendor/gonum.org/v1/gonum/graph/simple/undirected.go

@@ -0,0 +1,216 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package simple
+
+import (
+	"fmt"
+
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/uid"
+	"gonum.org/v1/gonum/graph/iterator"
+)
+
+var (
+	ug *UndirectedGraph
+
+	_ graph.Graph       = ug
+	_ graph.Undirected  = ug
+	_ graph.NodeAdder   = ug
+	_ graph.NodeRemover = ug
+	_ graph.EdgeAdder   = ug
+	_ graph.EdgeRemover = ug
+)
+
+// UndirectedGraph implements a generalized undirected graph.
+type UndirectedGraph struct {
+	nodes map[int64]graph.Node
+	edges map[int64]map[int64]graph.Edge
+
+	nodeIDs uid.Set
+}
+
+// NewUndirectedGraph returns an UndirectedGraph.
+func NewUndirectedGraph() *UndirectedGraph {
+	return &UndirectedGraph{
+		nodes: make(map[int64]graph.Node),
+		edges: make(map[int64]map[int64]graph.Edge),
+
+		nodeIDs: uid.NewSet(),
+	}
+}
+
+// AddNode adds n to the graph. It panics if the added node ID matches an existing node ID.
+func (g *UndirectedGraph) AddNode(n graph.Node) {
+	if _, exists := g.nodes[n.ID()]; exists {
+		panic(fmt.Sprintf("simple: node ID collision: %d", n.ID()))
+	}
+	g.nodes[n.ID()] = n
+	g.edges[n.ID()] = make(map[int64]graph.Edge)
+	g.nodeIDs.Use(n.ID())
+}
+
+// Edge returns the edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+func (g *UndirectedGraph) Edge(uid, vid int64) graph.Edge {
+	return g.EdgeBetween(uid, vid)
+}
+
+// EdgeBetween returns the edge between nodes x and y.
+func (g *UndirectedGraph) EdgeBetween(xid, yid int64) graph.Edge {
+	edge, ok := g.edges[xid][yid]
+	if !ok {
+		return nil
+	}
+	if edge.From().ID() == xid {
+		return edge
+	}
+	return edge.ReversedEdge()
+}
+
+// Edges returns all the edges in the graph.
+func (g *UndirectedGraph) Edges() graph.Edges {
+	if len(g.edges) == 0 {
+		return graph.Empty
+	}
+	var edges []graph.Edge
+	seen := make(map[[2]int64]struct{})
+	for _, u := range g.edges {
+		for _, e := range u {
+			uid := e.From().ID()
+			vid := e.To().ID()
+			if _, ok := seen[[2]int64{uid, vid}]; ok {
+				continue
+			}
+			seen[[2]int64{uid, vid}] = struct{}{}
+			seen[[2]int64{vid, uid}] = struct{}{}
+			edges = append(edges, e)
+		}
+	}
+	if len(edges) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedEdges(edges)
+}
+
+// From returns all nodes in g that can be reached directly from n.
+func (g *UndirectedGraph) From(id int64) graph.Nodes {
+	if _, ok := g.nodes[id]; !ok {
+		return graph.Empty
+	}
+
+	nodes := make([]graph.Node, len(g.edges[id]))
+	i := 0
+	for from := range g.edges[id] {
+		nodes[i] = g.nodes[from]
+		i++
+	}
+	if len(nodes) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedNodes(nodes)
+}
+
+// HasEdgeBetween returns whether an edge exists between nodes x and y.
+func (g *UndirectedGraph) HasEdgeBetween(xid, yid int64) bool {
+	_, ok := g.edges[xid][yid]
+	return ok
+}
+
+// NewEdge returns a new Edge from the source to the destination node.
+func (g *UndirectedGraph) NewEdge(from, to graph.Node) graph.Edge {
+	return &Edge{F: from, T: to}
+}
+
+// NewNode returns a new unique Node to be added to g. The Node's ID does
+// not become valid in g until the Node is added to g.
+func (g *UndirectedGraph) NewNode() graph.Node {
+	if len(g.nodes) == 0 {
+		return Node(0)
+	}
+	if int64(len(g.nodes)) == uid.Max {
+		panic("simple: cannot allocate node: no slot")
+	}
+	return Node(g.nodeIDs.NewID())
+}
+
+// Node returns the node with the given ID if it exists in the graph,
+// and nil otherwise.
+func (g *UndirectedGraph) Node(id int64) graph.Node {
+	return g.nodes[id]
+}
+
+// Nodes returns all the nodes in the graph.
+func (g *UndirectedGraph) Nodes() graph.Nodes {
+	if len(g.nodes) == 0 {
+		return graph.Empty
+	}
+	nodes := make([]graph.Node, len(g.nodes))
+	i := 0
+	for _, n := range g.nodes {
+		nodes[i] = n
+		i++
+	}
+	return iterator.NewOrderedNodes(nodes)
+}
+
+// RemoveEdge removes the edge with the given end IDs from the graph, leaving the terminal nodes.
+// If the edge does not exist it is a no-op.
+func (g *UndirectedGraph) RemoveEdge(fid, tid int64) {
+	if _, ok := g.nodes[fid]; !ok {
+		return
+	}
+	if _, ok := g.nodes[tid]; !ok {
+		return
+	}
+
+	delete(g.edges[fid], tid)
+	delete(g.edges[tid], fid)
+}
+
+// RemoveNode removes the node with the given ID from the graph, as well as any edges attached
+// to it. If the node is not in the graph it is a no-op.
+func (g *UndirectedGraph) RemoveNode(id int64) {
+	if _, ok := g.nodes[id]; !ok {
+		return
+	}
+	delete(g.nodes, id)
+
+	for from := range g.edges[id] {
+		delete(g.edges[from], id)
+	}
+	delete(g.edges, id)
+
+	g.nodeIDs.Release(id)
+}
+
+// SetEdge adds e, an edge from one node to another. If the nodes do not exist, they are added
+// and are set to the nodes of the edge otherwise.
+// It will panic if the IDs of the e.From and e.To are equal.
+func (g *UndirectedGraph) SetEdge(e graph.Edge) {
+	var (
+		from = e.From()
+		fid  = from.ID()
+		to   = e.To()
+		tid  = to.ID()
+	)
+
+	if fid == tid {
+		panic("simple: adding self edge")
+	}
+
+	if _, ok := g.nodes[fid]; !ok {
+		g.AddNode(from)
+	} else {
+		g.nodes[fid] = from
+	}
+	if _, ok := g.nodes[tid]; !ok {
+		g.AddNode(to)
+	} else {
+		g.nodes[tid] = to
+	}
+
+	g.edges[fid][tid] = e
+	g.edges[tid][fid] = e
+}

vendor/gonum.org/v1/gonum/graph/simple/weighted_directed.go

@@ -0,0 +1,279 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package simple
+
+import (
+	"fmt"
+
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/uid"
+	"gonum.org/v1/gonum/graph/iterator"
+)
+
+var (
+	wdg *WeightedDirectedGraph
+
+	_ graph.Graph             = wdg
+	_ graph.Weighted          = wdg
+	_ graph.Directed          = wdg
+	_ graph.WeightedDirected  = wdg
+	_ graph.NodeAdder         = wdg
+	_ graph.NodeRemover       = wdg
+	_ graph.WeightedEdgeAdder = wdg
+	_ graph.EdgeRemover       = wdg
+)
+
+// WeightedDirectedGraph implements a generalized weighted directed graph.
+type WeightedDirectedGraph struct {
+	nodes map[int64]graph.Node
+	from  map[int64]map[int64]graph.WeightedEdge
+	to    map[int64]map[int64]graph.WeightedEdge
+
+	self, absent float64
+
+	nodeIDs uid.Set
+}
+
+// NewWeightedDirectedGraph returns a WeightedDirectedGraph with the specified self and absent
+// edge weight values.
+func NewWeightedDirectedGraph(self, absent float64) *WeightedDirectedGraph {
+	return &WeightedDirectedGraph{
+		nodes: make(map[int64]graph.Node),
+		from:  make(map[int64]map[int64]graph.WeightedEdge),
+		to:    make(map[int64]map[int64]graph.WeightedEdge),
+
+		self:   self,
+		absent: absent,
+
+		nodeIDs: uid.NewSet(),
+	}
+}
+
+// AddNode adds n to the graph. It panics if the added node ID matches an existing node ID.
+func (g *WeightedDirectedGraph) AddNode(n graph.Node) {
+	if _, exists := g.nodes[n.ID()]; exists {
+		panic(fmt.Sprintf("simple: node ID collision: %d", n.ID()))
+	}
+	g.nodes[n.ID()] = n
+	g.from[n.ID()] = make(map[int64]graph.WeightedEdge)
+	g.to[n.ID()] = make(map[int64]graph.WeightedEdge)
+	g.nodeIDs.Use(n.ID())
+}
+
+// Edge returns the edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+func (g *WeightedDirectedGraph) Edge(uid, vid int64) graph.Edge {
+	return g.WeightedEdge(uid, vid)
+}
+
+// Edges returns all the edges in the graph.
+func (g *WeightedDirectedGraph) Edges() graph.Edges {
+	var edges []graph.Edge
+	for _, u := range g.nodes {
+		for _, e := range g.from[u.ID()] {
+			edges = append(edges, e)
+		}
+	}
+	if len(edges) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedEdges(edges)
+}
+
+// From returns all nodes in g that can be reached directly from n.
+func (g *WeightedDirectedGraph) From(id int64) graph.Nodes {
+	if _, ok := g.from[id]; !ok {
+		return graph.Empty
+	}
+
+	from := make([]graph.Node, len(g.from[id]))
+	i := 0
+	for vid := range g.from[id] {
+		from[i] = g.nodes[vid]
+		i++
+	}
+	if len(from) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedNodes(from)
+}
+
+// HasEdgeBetween returns whether an edge exists between nodes x and y without
+// considering direction.
+func (g *WeightedDirectedGraph) HasEdgeBetween(xid, yid int64) bool {
+	if _, ok := g.from[xid][yid]; ok {
+		return true
+	}
+	_, ok := g.from[yid][xid]
+	return ok
+}
+
+// HasEdgeFromTo returns whether an edge exists in the graph from u to v.
+func (g *WeightedDirectedGraph) HasEdgeFromTo(uid, vid int64) bool {
+	if _, ok := g.from[uid][vid]; !ok {
+		return false
+	}
+	return true
+}
+
+// NewNode returns a new unique Node to be added to g. The Node's ID does
+// not become valid in g until the Node is added to g.
+func (g *WeightedDirectedGraph) NewNode() graph.Node {
+	if len(g.nodes) == 0 {
+		return Node(0)
+	}
+	if int64(len(g.nodes)) == uid.Max {
+		panic("simple: cannot allocate node: no slot")
+	}
+	return Node(g.nodeIDs.NewID())
+}
+
+// NewWeightedEdge returns a new weighted edge from the source to the destination node.
+func (g *WeightedDirectedGraph) NewWeightedEdge(from, to graph.Node, weight float64) graph.WeightedEdge {
+	return &WeightedEdge{F: from, T: to, W: weight}
+}
+
+// Node returns the node with the given ID if it exists in the graph,
+// and nil otherwise.
+func (g *WeightedDirectedGraph) Node(id int64) graph.Node {
+	return g.nodes[id]
+}
+
+// Nodes returns all the nodes in the graph.
+func (g *WeightedDirectedGraph) Nodes() graph.Nodes {
+	if len(g.from) == 0 {
+		return graph.Empty
+	}
+	nodes := make([]graph.Node, len(g.nodes))
+	i := 0
+	for _, n := range g.nodes {
+		nodes[i] = n
+		i++
+	}
+	return iterator.NewOrderedNodes(nodes)
+}
+
+// RemoveEdge removes the edge with the given end point IDs from the graph, leaving the terminal
+// nodes. If the edge does not exist it is a no-op.
+func (g *WeightedDirectedGraph) RemoveEdge(fid, tid int64) {
+	if _, ok := g.nodes[fid]; !ok {
+		return
+	}
+	if _, ok := g.nodes[tid]; !ok {
+		return
+	}
+
+	delete(g.from[fid], tid)
+	delete(g.to[tid], fid)
+}
+
+// RemoveNode removes the node with the given ID from the graph, as well as any edges attached
+// to it. If the node is not in the graph it is a no-op.
+func (g *WeightedDirectedGraph) RemoveNode(id int64) {
+	if _, ok := g.nodes[id]; !ok {
+		return
+	}
+	delete(g.nodes, id)
+
+	for from := range g.from[id] {
+		delete(g.to[from], id)
+	}
+	delete(g.from, id)
+
+	for to := range g.to[id] {
+		delete(g.from[to], id)
+	}
+	delete(g.to, id)
+
+	g.nodeIDs.Release(id)
+}
+
+// SetWeightedEdge adds a weighted edge from one node to another. If the nodes do not exist, they are added
+// and are set to the nodes of the edge otherwise.
+// It will panic if the IDs of the e.From and e.To are equal.
+func (g *WeightedDirectedGraph) SetWeightedEdge(e graph.WeightedEdge) {
+	var (
+		from = e.From()
+		fid  = from.ID()
+		to   = e.To()
+		tid  = to.ID()
+	)
+
+	if fid == tid {
+		panic("simple: adding self edge")
+	}
+
+	if _, ok := g.nodes[fid]; !ok {
+		g.AddNode(from)
+	} else {
+		g.nodes[fid] = from
+	}
+	if _, ok := g.nodes[tid]; !ok {
+		g.AddNode(to)
+	} else {
+		g.nodes[tid] = to
+	}
+
+	g.from[fid][tid] = e
+	g.to[tid][fid] = e
+}
+
+// To returns all nodes in g that can reach directly to n.
+func (g *WeightedDirectedGraph) To(id int64) graph.Nodes {
+	if _, ok := g.from[id]; !ok {
+		return graph.Empty
+	}
+
+	to := make([]graph.Node, len(g.to[id]))
+	i := 0
+	for uid := range g.to[id] {
+		to[i] = g.nodes[uid]
+		i++
+	}
+	if len(to) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedNodes(to)
+}
+
+// Weight returns the weight for the edge between x and y if Edge(x, y) returns a non-nil Edge.
+// If x and y are the same node or there is no joining edge between the two nodes the weight
+// value returned is either the graph's absent or self value. Weight returns true if an edge
+// exists between x and y or if x and y have the same ID, false otherwise.
+func (g *WeightedDirectedGraph) Weight(xid, yid int64) (w float64, ok bool) {
+	if xid == yid {
+		return g.self, true
+	}
+	if to, ok := g.from[xid]; ok {
+		if e, ok := to[yid]; ok {
+			return e.Weight(), true
+		}
+	}
+	return g.absent, false
+}
+
+// WeightedEdge returns the weighted edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+func (g *WeightedDirectedGraph) WeightedEdge(uid, vid int64) graph.WeightedEdge {
+	edge, ok := g.from[uid][vid]
+	if !ok {
+		return nil
+	}
+	return edge
+}
+
+// WeightedEdges returns all the weighted edges in the graph.
+func (g *WeightedDirectedGraph) WeightedEdges() graph.WeightedEdges {
+	var edges []graph.WeightedEdge
+	for _, u := range g.nodes {
+		for _, e := range g.from[u.ID()] {
+			edges = append(edges, e)
+		}
+	}
+	if len(edges) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedWeightedEdges(edges)
+}

vendor/gonum.org/v1/gonum/graph/simple/weighted_undirected.go

@@ -0,0 +1,273 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package simple
+
+import (
+	"fmt"
+
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/uid"
+	"gonum.org/v1/gonum/graph/iterator"
+)
+
+var (
+	wug *WeightedUndirectedGraph
+
+	_ graph.Graph              = wug
+	_ graph.Weighted           = wug
+	_ graph.Undirected         = wug
+	_ graph.WeightedUndirected = wug
+	_ graph.NodeAdder          = wug
+	_ graph.NodeRemover        = wug
+	_ graph.WeightedEdgeAdder  = wug
+	_ graph.EdgeRemover        = wug
+)
+
+// WeightedUndirectedGraph implements a generalized weighted undirected graph.
+type WeightedUndirectedGraph struct {
+	nodes map[int64]graph.Node
+	edges map[int64]map[int64]graph.WeightedEdge
+
+	self, absent float64
+
+	nodeIDs uid.Set
+}
+
+// NewWeightedUndirectedGraph returns an WeightedUndirectedGraph with the specified self and absent
+// edge weight values.
+func NewWeightedUndirectedGraph(self, absent float64) *WeightedUndirectedGraph {
+	return &WeightedUndirectedGraph{
+		nodes: make(map[int64]graph.Node),
+		edges: make(map[int64]map[int64]graph.WeightedEdge),
+
+		self:   self,
+		absent: absent,
+
+		nodeIDs: uid.NewSet(),
+	}
+}
+
+// AddNode adds n to the graph. It panics if the added node ID matches an existing node ID.
+func (g *WeightedUndirectedGraph) AddNode(n graph.Node) {
+	if _, exists := g.nodes[n.ID()]; exists {
+		panic(fmt.Sprintf("simple: node ID collision: %d", n.ID()))
+	}
+	g.nodes[n.ID()] = n
+	g.edges[n.ID()] = make(map[int64]graph.WeightedEdge)
+	g.nodeIDs.Use(n.ID())
+}
+
+// Edge returns the edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+func (g *WeightedUndirectedGraph) Edge(uid, vid int64) graph.Edge {
+	return g.WeightedEdgeBetween(uid, vid)
+}
+
+// EdgeBetween returns the edge between nodes x and y.
+func (g *WeightedUndirectedGraph) EdgeBetween(xid, yid int64) graph.Edge {
+	return g.WeightedEdgeBetween(xid, yid)
+}
+
+// Edges returns all the edges in the graph.
+func (g *WeightedUndirectedGraph) Edges() graph.Edges {
+	if len(g.edges) == 0 {
+		return graph.Empty
+	}
+	var edges []graph.Edge
+	seen := make(map[[2]int64]struct{})
+	for _, u := range g.edges {
+		for _, e := range u {
+			uid := e.From().ID()
+			vid := e.To().ID()
+			if _, ok := seen[[2]int64{uid, vid}]; ok {
+				continue
+			}
+			seen[[2]int64{uid, vid}] = struct{}{}
+			seen[[2]int64{vid, uid}] = struct{}{}
+			edges = append(edges, e)
+		}
+	}
+	if len(edges) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedEdges(edges)
+}
+
+// From returns all nodes in g that can be reached directly from n.
+func (g *WeightedUndirectedGraph) From(id int64) graph.Nodes {
+	if _, ok := g.nodes[id]; !ok {
+		return graph.Empty
+	}
+
+	nodes := make([]graph.Node, len(g.edges[id]))
+	i := 0
+	for from := range g.edges[id] {
+		nodes[i] = g.nodes[from]
+		i++
+	}
+	if len(nodes) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedNodes(nodes)
+}
+
+// HasEdgeBetween returns whether an edge exists between nodes x and y.
+func (g *WeightedUndirectedGraph) HasEdgeBetween(xid, yid int64) bool {
+	_, ok := g.edges[xid][yid]
+	return ok
+}
+
+// NewNode returns a new unique Node to be added to g. The Node's ID does
+// not become valid in g until the Node is added to g.
+func (g *WeightedUndirectedGraph) NewNode() graph.Node {
+	if len(g.nodes) == 0 {
+		return Node(0)
+	}
+	if int64(len(g.nodes)) == uid.Max {
+		panic("simple: cannot allocate node: no slot")
+	}
+	return Node(g.nodeIDs.NewID())
+}
+
+// NewWeightedEdge returns a new weighted edge from the source to the destination node.
+func (g *WeightedUndirectedGraph) NewWeightedEdge(from, to graph.Node, weight float64) graph.WeightedEdge {
+	return &WeightedEdge{F: from, T: to, W: weight}
+}
+
+// Node returns the node with the given ID if it exists in the graph,
+// and nil otherwise.
+func (g *WeightedUndirectedGraph) Node(id int64) graph.Node {
+	return g.nodes[id]
+}
+
+// Nodes returns all the nodes in the graph.
+func (g *WeightedUndirectedGraph) Nodes() graph.Nodes {
+	if len(g.nodes) == 0 {
+		return graph.Empty
+	}
+	nodes := make([]graph.Node, len(g.nodes))
+	i := 0
+	for _, n := range g.nodes {
+		nodes[i] = n
+		i++
+	}
+	return iterator.NewOrderedNodes(nodes)
+}
+
+// RemoveEdge removes the edge with the given end point IDs from the graph, leaving the terminal
+// nodes. If the edge does not exist  it is a no-op.
+func (g *WeightedUndirectedGraph) RemoveEdge(fid, tid int64) {
+	if _, ok := g.nodes[fid]; !ok {
+		return
+	}
+	if _, ok := g.nodes[tid]; !ok {
+		return
+	}
+
+	delete(g.edges[fid], tid)
+	delete(g.edges[tid], fid)
+}
+
+// RemoveNode removes the node with the given ID from the graph, as well as any edges attached
+// to it. If the node is not in the graph it is a no-op.
+func (g *WeightedUndirectedGraph) RemoveNode(id int64) {
+	if _, ok := g.nodes[id]; !ok {
+		return
+	}
+	delete(g.nodes, id)
+
+	for from := range g.edges[id] {
+		delete(g.edges[from], id)
+	}
+	delete(g.edges, id)
+
+	g.nodeIDs.Release(id)
+}
+
+// SetWeightedEdge adds a weighted edge from one node to another. If the nodes do not exist, they are added
+// and are set to the nodes of the edge otherwise.
+// It will panic if the IDs of the e.From and e.To are equal.
+func (g *WeightedUndirectedGraph) SetWeightedEdge(e graph.WeightedEdge) {
+	var (
+		from = e.From()
+		fid  = from.ID()
+		to   = e.To()
+		tid  = to.ID()
+	)
+
+	if fid == tid {
+		panic("simple: adding self edge")
+	}
+
+	if _, ok := g.nodes[fid]; !ok {
+		g.AddNode(from)
+	} else {
+		g.nodes[fid] = from
+	}
+	if _, ok := g.nodes[tid]; !ok {
+		g.AddNode(to)
+	} else {
+		g.nodes[tid] = to
+	}
+
+	g.edges[fid][tid] = e
+	g.edges[tid][fid] = e
+}
+
+// Weight returns the weight for the edge between x and y if Edge(x, y) returns a non-nil Edge.
+// If x and y are the same node or there is no joining edge between the two nodes the weight
+// value returned is either the graph's absent or self value. Weight returns true if an edge
+// exists between x and y or if x and y have the same ID, false otherwise.
+func (g *WeightedUndirectedGraph) Weight(xid, yid int64) (w float64, ok bool) {
+	if xid == yid {
+		return g.self, true
+	}
+	if n, ok := g.edges[xid]; ok {
+		if e, ok := n[yid]; ok {
+			return e.Weight(), true
+		}
+	}
+	return g.absent, false
+}
+
+// WeightedEdge returns the weighted edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+func (g *WeightedUndirectedGraph) WeightedEdge(uid, vid int64) graph.WeightedEdge {
+	return g.WeightedEdgeBetween(uid, vid)
+}
+
+// WeightedEdgeBetween returns the weighted edge between nodes x and y.
+func (g *WeightedUndirectedGraph) WeightedEdgeBetween(xid, yid int64) graph.WeightedEdge {
+	edge, ok := g.edges[xid][yid]
+	if !ok {
+		return nil
+	}
+	if edge.From().ID() == xid {
+		return edge
+	}
+	return edge.ReversedEdge().(graph.WeightedEdge)
+}
+
+// WeightedEdges returns all the weighted edges in the graph.
+func (g *WeightedUndirectedGraph) WeightedEdges() graph.WeightedEdges {
+	var edges []graph.WeightedEdge
+	seen := make(map[[2]int64]struct{})
+	for _, u := range g.edges {
+		for _, e := range u {
+			uid := e.From().ID()
+			vid := e.To().ID()
+			if _, ok := seen[[2]int64{uid, vid}]; ok {
+				continue
+			}
+			seen[[2]int64{uid, vid}] = struct{}{}
+			seen[[2]int64{vid, uid}] = struct{}{}
+			edges = append(edges, e)
+		}
+	}
+	if len(edges) == 0 {
+		return graph.Empty
+	}
+	return iterator.NewOrderedWeightedEdges(edges)
+}

vendor/gonum.org/v1/gonum/graph/topo/bron_kerbosch.go

@@ -0,0 +1,250 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package topo
+
+import (
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/ordered"
+	"gonum.org/v1/gonum/graph/internal/set"
+)
+
+// DegeneracyOrdering returns the degeneracy ordering and the k-cores of
+// the undirected graph g.
+func DegeneracyOrdering(g graph.Undirected) (order []graph.Node, cores [][]graph.Node) {
+	order, offsets := degeneracyOrdering(g)
+
+	ordered.Reverse(order)
+	cores = make([][]graph.Node, len(offsets))
+	offset := len(order)
+	for i, n := range offsets {
+		cores[i] = order[offset-n : offset]
+		offset -= n
+	}
+	return order, cores
+}
+
+// KCore returns the k-core of the undirected graph g with nodes in an
+// optimal ordering for the coloring number.
+func KCore(k int, g graph.Undirected) []graph.Node {
+	order, offsets := degeneracyOrdering(g)
+
+	var offset int
+	for _, n := range offsets[:k] {
+		offset += n
+	}
+	core := make([]graph.Node, len(order)-offset)
+	copy(core, order[offset:])
+	return core
+}
+
+// degeneracyOrdering is the common code for DegeneracyOrdering and KCore. It
+// returns l, the nodes of g in optimal ordering for coloring number and
+// s, a set of relative offsets into l for each k-core, where k is an index
+// into s.
+func degeneracyOrdering(g graph.Undirected) (l []graph.Node, s []int) {
+	nodes := graph.NodesOf(g.Nodes())
+
+	// The algorithm used here is essentially as described at
+	// http://en.wikipedia.org/w/index.php?title=Degeneracy_%28graph_theory%29&oldid=640308710
+
+	// Initialize an output list L in return parameters.
+
+	// Compute a number d_v for each vertex v in G,
+	// the number of neighbors of v that are not already in L.
+	// Initially, these numbers are just the degrees of the vertices.
+	dv := make(map[int64]int, len(nodes))
+	var (
+		maxDegree  int
+		neighbours = make(map[int64][]graph.Node)
+	)
+	for _, n := range nodes {
+		id := n.ID()
+		adj := graph.NodesOf(g.From(id))
+		neighbours[id] = adj
+		dv[id] = len(adj)
+		if len(adj) > maxDegree {
+			maxDegree = len(adj)
+		}
+	}
+
+	// Initialize an array D such that D[i] contains a list of the
+	// vertices v that are not already in L for which d_v = i.
+	d := make([][]graph.Node, maxDegree+1)
+	for _, n := range nodes {
+		deg := dv[n.ID()]
+		d[deg] = append(d[deg], n)
+	}
+
+	// Initialize k to 0.
+	k := 0
+	// Repeat n times:
+	s = []int{0}
+	for range nodes {
+		// Scan the array cells D[0], D[1], ... until
+		// finding an i for which D[i] is nonempty.
+		var (
+			i  int
+			di []graph.Node
+		)
+		for i, di = range d {
+			if len(di) != 0 {
+				break
+			}
+		}
+
+		// Set k to max(k,i).
+		if i > k {
+			k = i
+			s = append(s, make([]int, k-len(s)+1)...)
+		}
+
+		// Select a vertex v from D[i]. Add v to the
+		// beginning of L and remove it from D[i].
+		var v graph.Node
+		v, d[i] = di[len(di)-1], di[:len(di)-1]
+		l = append(l, v)
+		s[k]++
+		delete(dv, v.ID())
+
+		// For each neighbor w of v not already in L,
+		// subtract one from d_w and move w to the
+		// cell of D corresponding to the new value of d_w.
+		for _, w := range neighbours[v.ID()] {
+			dw, ok := dv[w.ID()]
+			if !ok {
+				continue
+			}
+			for i, n := range d[dw] {
+				if n.ID() == w.ID() {
+					d[dw][i], d[dw] = d[dw][len(d[dw])-1], d[dw][:len(d[dw])-1]
+					dw--
+					d[dw] = append(d[dw], w)
+					break
+				}
+			}
+			dv[w.ID()] = dw
+		}
+	}
+
+	return l, s
+}
+
+// BronKerbosch returns the set of maximal cliques of the undirected graph g.
+func BronKerbosch(g graph.Undirected) [][]graph.Node {
+	nodes := graph.NodesOf(g.Nodes())
+
+	// The algorithm used here is essentially BronKerbosch3 as described at
+	// http://en.wikipedia.org/w/index.php?title=Bron%E2%80%93Kerbosch_algorithm&oldid=656805858
+
+	p := set.NewNodesSize(len(nodes))
+	for _, n := range nodes {
+		p.Add(n)
+	}
+	x := set.NewNodes()
+	var bk bronKerbosch
+	order, _ := degeneracyOrdering(g)
+	ordered.Reverse(order)
+	for _, v := range order {
+		neighbours := graph.NodesOf(g.From(v.ID()))
+		nv := set.NewNodesSize(len(neighbours))
+		for _, n := range neighbours {
+			nv.Add(n)
+		}
+		bk.maximalCliquePivot(g, []graph.Node{v}, set.IntersectionOfNodes(p, nv), set.IntersectionOfNodes(x, nv))
+		p.Remove(v)
+		x.Add(v)
+	}
+	return bk
+}
+
+type bronKerbosch [][]graph.Node
+
+func (bk *bronKerbosch) maximalCliquePivot(g graph.Undirected, r []graph.Node, p, x set.Nodes) {
+	if len(p) == 0 && len(x) == 0 {
+		*bk = append(*bk, r)
+		return
+	}
+
+	neighbours := bk.choosePivotFrom(g, p, x)
+	nu := set.NewNodesSize(len(neighbours))
+	for _, n := range neighbours {
+		nu.Add(n)
+	}
+	for _, v := range p {
+		if nu.Has(v) {
+			continue
+		}
+		vid := v.ID()
+		neighbours := graph.NodesOf(g.From(vid))
+		nv := set.NewNodesSize(len(neighbours))
+		for _, n := range neighbours {
+			nv.Add(n)
+		}
+
+		var found bool
+		for _, n := range r {
+			if n.ID() == vid {
+				found = true
+				break
+			}
+		}
+		var sr []graph.Node
+		if !found {
+			sr = append(r[:len(r):len(r)], v)
+		}
+
+		bk.maximalCliquePivot(g, sr, set.IntersectionOfNodes(p, nv), set.IntersectionOfNodes(x, nv))
+		p.Remove(v)
+		x.Add(v)
+	}
+}
+
+func (*bronKerbosch) choosePivotFrom(g graph.Undirected, p, x set.Nodes) (neighbors []graph.Node) {
+	// TODO(kortschak): Investigate the impact of pivot choice that maximises
+	// |p ⋂ neighbours(u)| as a function of input size. Until then, leave as
+	// compile time option.
+	if !tomitaTanakaTakahashi {
+		for _, n := range p {
+			return graph.NodesOf(g.From(n.ID()))
+		}
+		for _, n := range x {
+			return graph.NodesOf(g.From(n.ID()))
+		}
+		panic("bronKerbosch: empty set")
+	}
+
+	var (
+		max   = -1
+		pivot graph.Node
+	)
+	maxNeighbors := func(s set.Nodes) {
+	outer:
+		for _, u := range s {
+			nb := graph.NodesOf(g.From(u.ID()))
+			c := len(nb)
+			if c <= max {
+				continue
+			}
+			for n := range nb {
+				if _, ok := p[int64(n)]; ok {
+					continue
+				}
+				c--
+				if c <= max {
+					continue outer
+				}
+			}
+			max = c
+			pivot = u
+			neighbors = nb
+		}
+	}
+	maxNeighbors(p)
+	maxNeighbors(x)
+	if pivot == nil {
+		panic("bronKerbosch: empty set")
+	}
+	return neighbors
+}

vendor/gonum.org/v1/gonum/graph/topo/clique_graph.go

@@ -0,0 +1,111 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package topo
+
+import (
+	"sort"
+
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/ordered"
+	"gonum.org/v1/gonum/graph/internal/set"
+)
+
+// Builder is a pure topological graph construction type.
+type Builder interface {
+	AddNode(graph.Node)
+	SetEdge(graph.Edge)
+}
+
+// CliqueGraph builds the clique graph of g in dst using Clique and CliqueGraphEdge
+// nodes and edges. The nodes returned by calls to Nodes on the nodes and edges of
+// the constructed graph are the cliques and the common nodes between cliques
+// respectively. The dst graph is not cleared.
+func CliqueGraph(dst Builder, g graph.Undirected) {
+	cliques := BronKerbosch(g)
+
+	// Construct a consistent view of cliques in g. Sorting costs
+	// us a little, but not as much as the cliques themselves.
+	for _, c := range cliques {
+		sort.Sort(ordered.ByID(c))
+	}
+	sort.Sort(ordered.BySliceIDs(cliques))
+
+	cliqueNodes := make(cliqueNodeSets, len(cliques))
+	for id, c := range cliques {
+		s := set.NewNodesSize(len(c))
+		for _, n := range c {
+			s.Add(n)
+		}
+		ns := &nodeSet{Clique: Clique{id: int64(id), nodes: c}, nodes: s}
+		dst.AddNode(ns.Clique)
+		for _, n := range c {
+			nid := n.ID()
+			cliqueNodes[nid] = append(cliqueNodes[nid], ns)
+		}
+	}
+
+	for _, cliques := range cliqueNodes {
+		for i, uc := range cliques {
+			for _, vc := range cliques[i+1:] {
+				// Retain the nodes that contribute to the
+				// edge between the cliques.
+				var edgeNodes []graph.Node
+				switch 1 {
+				case len(uc.Clique.nodes):
+					edgeNodes = []graph.Node{uc.Clique.nodes[0]}
+				case len(vc.Clique.nodes):
+					edgeNodes = []graph.Node{vc.Clique.nodes[0]}
+				default:
+					for _, n := range set.IntersectionOfNodes(uc.nodes, vc.nodes) {
+						edgeNodes = append(edgeNodes, n)
+					}
+					sort.Sort(ordered.ByID(edgeNodes))
+				}
+
+				dst.SetEdge(CliqueGraphEdge{from: uc.Clique, to: vc.Clique, nodes: edgeNodes})
+			}
+		}
+	}
+}
+
+type cliqueNodeSets map[int64][]*nodeSet
+
+type nodeSet struct {
+	Clique
+	nodes set.Nodes
+}
+
+// Clique is a node in a clique graph.
+type Clique struct {
+	id    int64
+	nodes []graph.Node
+}
+
+// ID returns the node ID.
+func (n Clique) ID() int64 { return n.id }
+
+// Nodes returns the nodes in the clique.
+func (n Clique) Nodes() []graph.Node { return n.nodes }
+
+// CliqueGraphEdge is an edge in a clique graph.
+type CliqueGraphEdge struct {
+	from, to Clique
+	nodes    []graph.Node
+}
+
+// From returns the from node of the edge.
+func (e CliqueGraphEdge) From() graph.Node { return e.from }
+
+// To returns the to node of the edge.
+func (e CliqueGraphEdge) To() graph.Node { return e.to }
+
+// ReversedEdge returns a new CliqueGraphEdge with
+// the edge end points swapped. The nodes of the
+// new edge are shared with the receiver.
+func (e CliqueGraphEdge) ReversedEdge() graph.Edge { e.from, e.to = e.to, e.from; return e }
+
+// Nodes returns the common nodes in the cliques of the underlying graph
+// corresponding to the from and to nodes in the clique graph.
+func (e CliqueGraphEdge) Nodes() []graph.Node { return e.nodes }

vendor/gonum.org/v1/gonum/graph/topo/doc.go

@@ -0,0 +1,6 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package topo provides graph topology analysis functions.
+package topo // import "gonum.org/v1/gonum/graph/topo"

vendor/gonum.org/v1/gonum/graph/topo/johnson_cycles.go

@@ -0,0 +1,285 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package topo
+
+import (
+	"sort"
+
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/ordered"
+	"gonum.org/v1/gonum/graph/internal/set"
+	"gonum.org/v1/gonum/graph/iterator"
+)
+
+// johnson implements Johnson's "Finding all the elementary
+// circuits of a directed graph" algorithm. SIAM J. Comput. 4(1):1975.
+//
+// Comments in the johnson methods are kept in sync with the comments
+// and labels from the paper.
+type johnson struct {
+	adjacent johnsonGraph // SCC adjacency list.
+	b        []set.Ints   // Johnson's "B-list".
+	blocked  []bool
+	s        int
+
+	stack []graph.Node
+
+	result [][]graph.Node
+}
+
+// DirectedCyclesIn returns the set of elementary cycles in the graph g.
+func DirectedCyclesIn(g graph.Directed) [][]graph.Node {
+	jg := johnsonGraphFrom(g)
+	j := johnson{
+		adjacent: jg,
+		b:        make([]set.Ints, len(jg.orig)),
+		blocked:  make([]bool, len(jg.orig)),
+	}
+
+	// len(j.nodes) is the order of g.
+	for j.s < len(j.adjacent.orig)-1 {
+		// We use the previous SCC adjacency to reduce the work needed.
+		sccs := TarjanSCC(j.adjacent.subgraph(j.s))
+		// A_k = adjacency structure of strong component K with least
+		//       vertex in subgraph of G induced by {s, s+1, ... ,n}.
+		j.adjacent = j.adjacent.sccSubGraph(sccs, 2) // Only allow SCCs with >= 2 vertices.
+		if j.adjacent.order() == 0 {
+			break
+		}
+
+		// s = least vertex in V_k
+		if s := j.adjacent.leastVertexIndex(); s < j.s {
+			j.s = s
+		}
+		for i, v := range j.adjacent.orig {
+			if !j.adjacent.nodes.Has(v.ID()) {
+				continue
+			}
+			if len(j.adjacent.succ[v.ID()]) > 0 {
+				j.blocked[i] = false
+				j.b[i] = make(set.Ints)
+			}
+		}
+		//L3:
+		_ = j.circuit(j.s)
+		j.s++
+	}
+
+	return j.result
+}
+
+// circuit is the CIRCUIT sub-procedure in the paper.
+func (j *johnson) circuit(v int) bool {
+	f := false
+	n := j.adjacent.orig[v]
+	j.stack = append(j.stack, n)
+	j.blocked[v] = true
+
+	//L1:
+	for w := range j.adjacent.succ[n.ID()] {
+		w := j.adjacent.indexOf(w)
+		if w == j.s {
+			// Output circuit composed of stack followed by s.
+			r := make([]graph.Node, len(j.stack)+1)
+			copy(r, j.stack)
+			r[len(r)-1] = j.adjacent.orig[j.s]
+			j.result = append(j.result, r)
+			f = true
+		} else if !j.blocked[w] {
+			if j.circuit(w) {
+				f = true
+			}
+		}
+	}
+
+	//L2:
+	if f {
+		j.unblock(v)
+	} else {
+		for w := range j.adjacent.succ[n.ID()] {
+			j.b[j.adjacent.indexOf(w)].Add(v)
+		}
+	}
+	j.stack = j.stack[:len(j.stack)-1]
+
+	return f
+}
+
+// unblock is the UNBLOCK sub-procedure in the paper.
+func (j *johnson) unblock(u int) {
+	j.blocked[u] = false
+	for w := range j.b[u] {
+		j.b[u].Remove(w)
+		if j.blocked[w] {
+			j.unblock(w)
+		}
+	}
+}
+
+// johnsonGraph is an edge list representation of a graph with helpers
+// necessary for Johnson's algorithm
+type johnsonGraph struct {
+	// Keep the original graph nodes and a
+	// look-up to into the non-sparse
+	// collection of potentially sparse IDs.
+	orig  []graph.Node
+	index map[int64]int
+
+	nodes set.Int64s
+	succ  map[int64]set.Int64s
+}
+
+// johnsonGraphFrom returns a deep copy of the graph g.
+func johnsonGraphFrom(g graph.Directed) johnsonGraph {
+	nodes := graph.NodesOf(g.Nodes())
+	sort.Sort(ordered.ByID(nodes))
+	c := johnsonGraph{
+		orig:  nodes,
+		index: make(map[int64]int, len(nodes)),
+
+		nodes: make(set.Int64s, len(nodes)),
+		succ:  make(map[int64]set.Int64s),
+	}
+	for i, u := range nodes {
+		uid := u.ID()
+		c.index[uid] = i
+		for _, v := range graph.NodesOf(g.From(uid)) {
+			if c.succ[uid] == nil {
+				c.succ[uid] = make(set.Int64s)
+				c.nodes.Add(uid)
+			}
+			c.nodes.Add(v.ID())
+			c.succ[uid].Add(v.ID())
+		}
+	}
+	return c
+}
+
+// order returns the order of the graph.
+func (g johnsonGraph) order() int { return g.nodes.Count() }
+
+// indexOf returns the index of the retained node for the given node ID.
+func (g johnsonGraph) indexOf(id int64) int {
+	return g.index[id]
+}
+
+// leastVertexIndex returns the index into orig of the least vertex.
+func (g johnsonGraph) leastVertexIndex() int {
+	for _, v := range g.orig {
+		if g.nodes.Has(v.ID()) {
+			return g.indexOf(v.ID())
+		}
+	}
+	panic("johnsonCycles: empty set")
+}
+
+// subgraph returns a subgraph of g induced by {s, s+1, ... , n}. The
+// subgraph is destructively generated in g.
+func (g johnsonGraph) subgraph(s int) johnsonGraph {
+	sn := g.orig[s].ID()
+	for u, e := range g.succ {
+		if u < sn {
+			g.nodes.Remove(u)
+			delete(g.succ, u)
+			continue
+		}
+		for v := range e {
+			if v < sn {
+				g.succ[u].Remove(v)
+			}
+		}
+	}
+	return g
+}
+
+// sccSubGraph returns the graph of the tarjan's strongly connected
+// components with each SCC containing at least min vertices.
+// sccSubGraph returns nil if there is no SCC with at least min
+// members.
+func (g johnsonGraph) sccSubGraph(sccs [][]graph.Node, min int) johnsonGraph {
+	if len(g.nodes) == 0 {
+		g.nodes = nil
+		g.succ = nil
+		return g
+	}
+	sub := johnsonGraph{
+		orig:  g.orig,
+		index: g.index,
+		nodes: make(set.Int64s),
+		succ:  make(map[int64]set.Int64s),
+	}
+
+	var n int
+	for _, scc := range sccs {
+		if len(scc) < min {
+			continue
+		}
+		n++
+		for _, u := range scc {
+			for _, v := range scc {
+				if _, ok := g.succ[u.ID()][v.ID()]; ok {
+					if sub.succ[u.ID()] == nil {
+						sub.succ[u.ID()] = make(set.Int64s)
+						sub.nodes.Add(u.ID())
+					}
+					sub.nodes.Add(v.ID())
+					sub.succ[u.ID()].Add(v.ID())
+				}
+			}
+		}
+	}
+	if n == 0 {
+		g.nodes = nil
+		g.succ = nil
+		return g
+	}
+
+	return sub
+}
+
+// Nodes is required to satisfy Tarjan.
+func (g johnsonGraph) Nodes() graph.Nodes {
+	n := make([]graph.Node, 0, len(g.nodes))
+	for id := range g.nodes {
+		n = append(n, johnsonGraphNode(id))
+	}
+	return iterator.NewOrderedNodes(n)
+}
+
+// Successors is required to satisfy Tarjan.
+func (g johnsonGraph) From(id int64) graph.Nodes {
+	adj := g.succ[id]
+	if len(adj) == 0 {
+		return graph.Empty
+	}
+	succ := make([]graph.Node, 0, len(adj))
+	for id := range adj {
+		succ = append(succ, johnsonGraphNode(id))
+	}
+	return iterator.NewOrderedNodes(succ)
+}
+
+func (johnsonGraph) Has(int64) bool {
+	panic("topo: unintended use of johnsonGraph")
+}
+func (johnsonGraph) Node(int64) graph.Node {
+	panic("topo: unintended use of johnsonGraph")
+}
+func (johnsonGraph) HasEdgeBetween(_, _ int64) bool {
+	panic("topo: unintended use of johnsonGraph")
+}
+func (johnsonGraph) Edge(_, _ int64) graph.Edge {
+	panic("topo: unintended use of johnsonGraph")
+}
+func (johnsonGraph) HasEdgeFromTo(_, _ int64) bool {
+	panic("topo: unintended use of johnsonGraph")
+}
+func (johnsonGraph) To(int64) graph.Nodes {
+	panic("topo: unintended use of johnsonGraph")
+}
+
+type johnsonGraphNode int64
+
+func (n johnsonGraphNode) ID() int64 { return int64(n) }

vendor/gonum.org/v1/gonum/graph/topo/non_tomita_choice.go

@@ -0,0 +1,9 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !tomita
+
+package topo
+
+const tomitaTanakaTakahashi = false

vendor/gonum.org/v1/gonum/graph/topo/paton_cycles.go

@@ -0,0 +1,83 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package topo
+
+import (
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/linear"
+	"gonum.org/v1/gonum/graph/internal/set"
+)
+
+// UndirectedCyclesIn returns a set of cycles that forms a cycle basis in the graph g.
+// Any cycle in g can be constructed as a symmetric difference of its elements.
+func UndirectedCyclesIn(g graph.Undirected) [][]graph.Node {
+	// From "An algorithm for finding a fundamental set of cycles of a graph"
+	// https://doi.org/10.1145/363219.363232
+
+	var cycles [][]graph.Node
+	done := make(set.Int64s)
+	var tree linear.NodeStack
+	nodes := g.Nodes()
+	for nodes.Next() {
+		n := nodes.Node()
+		id := n.ID()
+		if done.Has(id) {
+			continue
+		}
+		done.Add(id)
+
+		tree = tree[:0]
+		tree.Push(n)
+		from := sets{id: set.Int64s{}}
+		to := map[int64]graph.Node{id: n}
+
+		for tree.Len() != 0 {
+			u := tree.Pop()
+			uid := u.ID()
+			adj := from[uid]
+			for _, v := range graph.NodesOf(g.From(uid)) {
+				vid := v.ID()
+				switch {
+				case uid == vid:
+					cycles = append(cycles, []graph.Node{u})
+				case !from.has(vid):
+					done.Add(vid)
+					to[vid] = u
+					tree.Push(v)
+					from.add(uid, vid)
+				case !adj.Has(vid):
+					c := []graph.Node{v, u}
+					adj := from[vid]
+					p := to[uid]
+					for !adj.Has(p.ID()) {
+						c = append(c, p)
+						p = to[p.ID()]
+					}
+					c = append(c, p, c[0])
+					cycles = append(cycles, c)
+					adj.Add(uid)
+				}
+			}
+		}
+	}
+
+	return cycles
+}
+
+type sets map[int64]set.Int64s
+
+func (s sets) add(uid, vid int64) {
+	e, ok := s[vid]
+	if !ok {
+		e = make(set.Int64s)
+		s[vid] = e
+	}
+	e.Add(uid)
+}
+
+func (s sets) has(uid int64) bool {
+	_, ok := s[uid]
+	return ok
+}

vendor/gonum.org/v1/gonum/graph/topo/tarjan.go

@@ -0,0 +1,199 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package topo
+
+import (
+	"fmt"
+	"sort"
+
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/ordered"
+	"gonum.org/v1/gonum/graph/internal/set"
+)
+
+// Unorderable is an error containing sets of unorderable graph.Nodes.
+type Unorderable [][]graph.Node
+
+// Error satisfies the error interface.
+func (e Unorderable) Error() string {
+	const maxNodes = 10
+	var n int
+	for _, c := range e {
+		n += len(c)
+	}
+	if n > maxNodes {
+		// Don't return errors that are too long.
+		return fmt.Sprintf("topo: no topological ordering: %d nodes in %d cyclic components", n, len(e))
+	}
+	return fmt.Sprintf("topo: no topological ordering: cyclic components: %v", [][]graph.Node(e))
+}
+
+func lexical(nodes []graph.Node) { sort.Sort(ordered.ByID(nodes)) }
+
+// Sort performs a topological sort of the directed graph g returning the 'from' to 'to'
+// sort order. If a topological ordering is not possible, an Unorderable error is returned
+// listing cyclic components in g with each cyclic component's members sorted by ID. When
+// an Unorderable error is returned, each cyclic component's topological position within
+// the sorted nodes is marked with a nil graph.Node.
+func Sort(g graph.Directed) (sorted []graph.Node, err error) {
+	sccs := TarjanSCC(g)
+	return sortedFrom(sccs, lexical)
+}
+
+// SortStabilized performs a topological sort of the directed graph g returning the 'from'
+// to 'to' sort order, or the order defined by the in place order sort function where there
+// is no unambiguous topological ordering. If a topological ordering is not possible, an
+// Unorderable error is returned listing cyclic components in g with each cyclic component's
+// members sorted by the provided order function. If order is nil, nodes are ordered lexically
+// by node ID. When an Unorderable error is returned, each cyclic component's topological
+// position within the sorted nodes is marked with a nil graph.Node.
+func SortStabilized(g graph.Directed, order func([]graph.Node)) (sorted []graph.Node, err error) {
+	if order == nil {
+		order = lexical
+	}
+	sccs := tarjanSCCstabilized(g, order)
+	return sortedFrom(sccs, order)
+}
+
+func sortedFrom(sccs [][]graph.Node, order func([]graph.Node)) ([]graph.Node, error) {
+	sorted := make([]graph.Node, 0, len(sccs))
+	var sc Unorderable
+	for _, s := range sccs {
+		if len(s) != 1 {
+			order(s)
+			sc = append(sc, s)
+			sorted = append(sorted, nil)
+			continue
+		}
+		sorted = append(sorted, s[0])
+	}
+	var err error
+	if sc != nil {
+		for i, j := 0, len(sc)-1; i < j; i, j = i+1, j-1 {
+			sc[i], sc[j] = sc[j], sc[i]
+		}
+		err = sc
+	}
+	ordered.Reverse(sorted)
+	return sorted, err
+}
+
+// TarjanSCC returns the strongly connected components of the graph g using Tarjan's algorithm.
+//
+// A strongly connected component of a graph is a set of vertices where it's possible to reach any
+// vertex in the set from any other (meaning there's a cycle between them.)
+//
+// Generally speaking, a directed graph where the number of strongly connected components is equal
+// to the number of nodes is acyclic, unless you count reflexive edges as a cycle (which requires
+// only a little extra testing.)
+//
+func TarjanSCC(g graph.Directed) [][]graph.Node {
+	return tarjanSCCstabilized(g, nil)
+}
+
+func tarjanSCCstabilized(g graph.Directed, order func([]graph.Node)) [][]graph.Node {
+	nodes := graph.NodesOf(g.Nodes())
+	var succ func(id int64) []graph.Node
+	if order == nil {
+		succ = func(id int64) []graph.Node {
+			return graph.NodesOf(g.From(id))
+		}
+	} else {
+		order(nodes)
+		ordered.Reverse(nodes)
+
+		succ = func(id int64) []graph.Node {
+			to := graph.NodesOf(g.From(id))
+			order(to)
+			ordered.Reverse(to)
+			return to
+		}
+	}
+
+	t := tarjan{
+		succ: succ,
+
+		indexTable: make(map[int64]int, len(nodes)),
+		lowLink:    make(map[int64]int, len(nodes)),
+		onStack:    make(set.Int64s),
+	}
+	for _, v := range nodes {
+		if t.indexTable[v.ID()] == 0 {
+			t.strongconnect(v)
+		}
+	}
+	return t.sccs
+}
+
+// tarjan implements Tarjan's strongly connected component finding
+// algorithm. The implementation is from the pseudocode at
+//
+// http://en.wikipedia.org/wiki/Tarjan%27s_strongly_connected_components_algorithm?oldid=642744644
+//
+type tarjan struct {
+	succ func(id int64) []graph.Node
+
+	index      int
+	indexTable map[int64]int
+	lowLink    map[int64]int
+	onStack    set.Int64s
+
+	stack []graph.Node
+
+	sccs [][]graph.Node
+}
+
+// strongconnect is the strongconnect function described in the
+// wikipedia article.
+func (t *tarjan) strongconnect(v graph.Node) {
+	vID := v.ID()
+
+	// Set the depth index for v to the smallest unused index.
+	t.index++
+	t.indexTable[vID] = t.index
+	t.lowLink[vID] = t.index
+	t.stack = append(t.stack, v)
+	t.onStack.Add(vID)
+
+	// Consider successors of v.
+	for _, w := range t.succ(vID) {
+		wID := w.ID()
+		if t.indexTable[wID] == 0 {
+			// Successor w has not yet been visited; recur on it.
+			t.strongconnect(w)
+			t.lowLink[vID] = min(t.lowLink[vID], t.lowLink[wID])
+		} else if t.onStack.Has(wID) {
+			// Successor w is in stack s and hence in the current SCC.
+			t.lowLink[vID] = min(t.lowLink[vID], t.indexTable[wID])
+		}
+	}
+
+	// If v is a root node, pop the stack and generate an SCC.
+	if t.lowLink[vID] == t.indexTable[vID] {
+		// Start a new strongly connected component.
+		var (
+			scc []graph.Node
+			w   graph.Node
+		)
+		for {
+			w, t.stack = t.stack[len(t.stack)-1], t.stack[:len(t.stack)-1]
+			t.onStack.Remove(w.ID())
+			// Add w to current strongly connected component.
+			scc = append(scc, w)
+			if w.ID() == vID {
+				break
+			}
+		}
+		// Output the current strongly connected component.
+		t.sccs = append(t.sccs, scc)
+	}
+}
+
+func min(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}

vendor/gonum.org/v1/gonum/graph/topo/tomita_choice.go

@@ -0,0 +1,9 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build tomita
+
+package topo
+
+const tomitaTanakaTakahashi = true

vendor/gonum.org/v1/gonum/graph/topo/topo.go

@@ -0,0 +1,68 @@
+// Copyright ©2014 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package topo
+
+import (
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/traverse"
+)
+
+// IsPathIn returns whether path is a path in g.
+//
+// As special cases, IsPathIn returns true for a zero length path or for
+// a path of length 1 when the node in path exists in the graph.
+func IsPathIn(g graph.Graph, path []graph.Node) bool {
+	switch len(path) {
+	case 0:
+		return true
+	case 1:
+		return g.Node(path[0].ID()) != nil
+	default:
+		var canReach func(uid, vid int64) bool
+		switch g := g.(type) {
+		case graph.Directed:
+			canReach = g.HasEdgeFromTo
+		default:
+			canReach = g.HasEdgeBetween
+		}
+
+		for i, u := range path[:len(path)-1] {
+			if !canReach(u.ID(), path[i+1].ID()) {
+				return false
+			}
+		}
+		return true
+	}
+}
+
+// PathExistsIn returns whether there is a path in g starting at from extending
+// to to.
+//
+// PathExistsIn exists as a helper function. If many tests for path existence
+// are being performed, other approaches will be more efficient.
+func PathExistsIn(g graph.Graph, from, to graph.Node) bool {
+	var t traverse.BreadthFirst
+	return t.Walk(g, from, func(n graph.Node, _ int) bool { return n.ID() == to.ID() }) != nil
+}
+
+// ConnectedComponents returns the connected components of the undirected graph g.
+func ConnectedComponents(g graph.Undirected) [][]graph.Node {
+	var (
+		w  traverse.DepthFirst
+		c  []graph.Node
+		cc [][]graph.Node
+	)
+	during := func(n graph.Node) {
+		c = append(c, n)
+	}
+	after := func() {
+		cc = append(cc, []graph.Node(nil))
+		cc[len(cc)-1] = append(cc[len(cc)-1], c...)
+		c = c[:0]
+	}
+	w.WalkAll(g, nil, after, during)
+
+	return cc
+}

vendor/gonum.org/v1/gonum/graph/traverse/doc.go

@@ -0,0 +1,6 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package traverse provides basic graph traversal primitives.
+package traverse // import "gonum.org/v1/gonum/graph/traverse"

vendor/gonum.org/v1/gonum/graph/traverse/traverse.go

@@ -0,0 +1,231 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package traverse
+
+import (
+	"gonum.org/v1/gonum/graph"
+	"gonum.org/v1/gonum/graph/internal/linear"
+	"gonum.org/v1/gonum/graph/internal/set"
+)
+
+var _ Graph = graph.Graph(nil)
+
+// Graph is the subset of graph.Graph necessary for graph traversal.
+type Graph interface {
+	// From returns all nodes that can be reached directly
+	// from the node with the given ID.
+	From(id int64) graph.Nodes
+
+	// Edge returns the edge from u to v, with IDs uid and vid,
+	// if such an edge exists and nil otherwise. The node v
+	// must be directly reachable from u as defined by
+	// the From method.
+	Edge(uid, vid int64) graph.Edge
+}
+
+// BreadthFirst implements stateful breadth-first graph traversal.
+type BreadthFirst struct {
+	// Visit is called on all nodes on their first visit.
+	Visit func(graph.Node)
+
+	// Traverse is called on all edges that may be traversed
+	// during the walk. This includes edges that would hop to
+	// an already visited node.
+	//
+	// The value returned by Traverse determines whether
+	// an edge can be traversed during the walk.
+	Traverse func(graph.Edge) bool
+
+	queue   linear.NodeQueue
+	visited set.Int64s
+}
+
+// Walk performs a breadth-first traversal of the graph g starting from the given node,
+// depending on the Traverse field and the until parameter if they are non-nil.
+// The traversal follows edges for which Traverse(edge) is true and returns the first node
+// for which until(node, depth) is true. During the traversal, if the Visit field is
+// non-nil, it is called with each node the first time it is visited.
+func (b *BreadthFirst) Walk(g Graph, from graph.Node, until func(n graph.Node, d int) bool) graph.Node {
+	if b.visited == nil {
+		b.visited = make(set.Int64s)
+	}
+	b.queue.Enqueue(from)
+	if b.Visit != nil && !b.visited.Has(from.ID()) {
+		b.Visit(from)
+	}
+	b.visited.Add(from.ID())
+
+	var (
+		depth     int
+		children  int
+		untilNext = 1
+	)
+	for b.queue.Len() > 0 {
+		t := b.queue.Dequeue()
+		if until != nil && until(t, depth) {
+			return t
+		}
+		tid := t.ID()
+		to := g.From(tid)
+		for to.Next() {
+			n := to.Node()
+			nid := n.ID()
+			if b.Traverse != nil && !b.Traverse(g.Edge(tid, nid)) {
+				continue
+			}
+			if b.visited.Has(nid) {
+				continue
+			}
+			if b.Visit != nil {
+				b.Visit(n)
+			}
+			b.visited.Add(nid)
+			children++
+			b.queue.Enqueue(n)
+		}
+		if untilNext--; untilNext == 0 {
+			depth++
+			untilNext = children
+			children = 0
+		}
+	}
+
+	return nil
+}
+
+// WalkAll calls Walk for each unvisited node of the graph g using edges independent
+// of their direction. The functions before and after are called prior to commencing
+// and after completing each walk if they are non-nil respectively. The function
+// during is called on each node as it is traversed.
+func (b *BreadthFirst) WalkAll(g graph.Undirected, before, after func(), during func(graph.Node)) {
+	b.Reset()
+	nodes := g.Nodes()
+	for nodes.Next() {
+		from := nodes.Node()
+		if b.Visited(from) {
+			continue
+		}
+		if before != nil {
+			before()
+		}
+		b.Walk(g, from, func(n graph.Node, _ int) bool {
+			if during != nil {
+				during(n)
+			}
+			return false
+		})
+		if after != nil {
+			after()
+		}
+	}
+}
+
+// Visited returned whether the node n was visited during a traverse.
+func (b *BreadthFirst) Visited(n graph.Node) bool {
+	return b.visited.Has(n.ID())
+}
+
+// Reset resets the state of the traverser for reuse.
+func (b *BreadthFirst) Reset() {
+	b.queue.Reset()
+	b.visited = nil
+}
+
+// DepthFirst implements stateful depth-first graph traversal.
+type DepthFirst struct {
+	// Visit is called on all nodes on their first visit.
+	Visit func(graph.Node)
+
+	// Traverse is called on all edges that may be traversed
+	// during the walk. This includes edges that would hop to
+	// an already visited node.
+	//
+	// The value returned by Traverse determines whether an
+	// edge can be traversed during the walk.
+	Traverse func(graph.Edge) bool
+
+	stack   linear.NodeStack
+	visited set.Int64s
+}
+
+// Walk performs a depth-first traversal of the graph g starting from the given node,
+// depending on the Traverse field and the until parameter if they are non-nil.
+// The traversal follows edges for which Traverse(edge) is true and returns the first node
+// for which until(node) is true. During the traversal, if the Visit field is non-nil, it
+// is called with each node the first time it is visited.
+func (d *DepthFirst) Walk(g Graph, from graph.Node, until func(graph.Node) bool) graph.Node {
+	if d.visited == nil {
+		d.visited = make(set.Int64s)
+	}
+	d.stack.Push(from)
+	if d.Visit != nil && !d.visited.Has(from.ID()) {
+		d.Visit(from)
+	}
+	d.visited.Add(from.ID())
+
+	for d.stack.Len() > 0 {
+		t := d.stack.Pop()
+		if until != nil && until(t) {
+			return t
+		}
+		tid := t.ID()
+		to := g.From(tid)
+		for to.Next() {
+			n := to.Node()
+			nid := n.ID()
+			if d.Traverse != nil && !d.Traverse(g.Edge(tid, nid)) {
+				continue
+			}
+			if d.visited.Has(nid) {
+				continue
+			}
+			if d.Visit != nil {
+				d.Visit(n)
+			}
+			d.visited.Add(nid)
+			d.stack.Push(n)
+		}
+	}
+
+	return nil
+}
+
+// WalkAll calls Walk for each unvisited node of the graph g using edges independent
+// of their direction. The functions before and after are called prior to commencing
+// and after completing each walk if they are non-nil respectively. The function
+// during is called on each node as it is traversed.
+func (d *DepthFirst) WalkAll(g graph.Undirected, before, after func(), during func(graph.Node)) {
+	d.Reset()
+	nodes := g.Nodes()
+	for nodes.Next() {
+		from := nodes.Node()
+		if d.Visited(from) {
+			continue
+		}
+		if before != nil {
+			before()
+		}
+		d.Walk(g, from, func(n graph.Node) bool {
+			if during != nil {
+				during(n)
+			}
+			return false
+		})
+		if after != nil {
+			after()
+		}
+	}
+}
+
+// Visited returned whether the node n was visited during a traverse.
+func (d *DepthFirst) Visited(n graph.Node) bool {
+	return d.visited.Has(n.ID())
+}
+
+// Reset resets the state of the traverser for reuse.
+func (d *DepthFirst) Reset() {
+	d.stack = d.stack[:0]
+	d.visited = nil
+}

vendor/gonum.org/v1/gonum/graph/undirect.go

@@ -0,0 +1,270 @@
+// Copyright ©2015 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package graph
+
+// Undirect converts a directed graph to an undirected graph.
+type Undirect struct {
+	G Directed
+}
+
+var _ Undirected = Undirect{}
+
+// Node returns the node with the given ID if it exists in the graph,
+// and nil otherwise.
+func (g Undirect) Node(id int64) Node { return g.G.Node(id) }
+
+// Nodes returns all the nodes in the graph.
+func (g Undirect) Nodes() Nodes { return g.G.Nodes() }
+
+// From returns all nodes in g that can be reached directly from u.
+func (g Undirect) From(uid int64) Nodes {
+	return newNodeFilterIterator(g.G.From(uid), g.G.To(uid))
+}
+
+// HasEdgeBetween returns whether an edge exists between nodes x and y.
+func (g Undirect) HasEdgeBetween(xid, yid int64) bool { return g.G.HasEdgeBetween(xid, yid) }
+
+// Edge returns the edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+// If an edge exists, the Edge returned is an EdgePair. The weight of
+// the edge is determined by applying the Merge func to the weights of the
+// edges between u and v.
+func (g Undirect) Edge(uid, vid int64) Edge { return g.EdgeBetween(uid, vid) }
+
+// EdgeBetween returns the edge between nodes x and y. If an edge exists, the
+// Edge returned is an EdgePair. The weight of the edge is determined by
+// applying the Merge func to the weights of edges between x and y.
+func (g Undirect) EdgeBetween(xid, yid int64) Edge {
+	fe := g.G.Edge(xid, yid)
+	re := g.G.Edge(yid, xid)
+	if fe == nil && re == nil {
+		return nil
+	}
+
+	return EdgePair{fe, re}
+}
+
+// UndirectWeighted converts a directed weighted graph to an undirected weighted graph,
+// resolving edge weight conflicts.
+type UndirectWeighted struct {
+	G WeightedDirected
+
+	// Absent is the value used to
+	// represent absent edge weights
+	// passed to Merge if the reverse
+	// edge is present.
+	Absent float64
+
+	// Merge defines how discordant edge
+	// weights in G are resolved. A merge
+	// is performed if at least one edge
+	// exists between the nodes being
+	// considered. The edges corresponding
+	// to the two weights are also passed,
+	// in the same order.
+	// The order of weight parameters
+	// passed to Merge is not defined, so
+	// the function should be commutative.
+	// If Merge is nil, the arithmetic
+	// mean is used to merge weights.
+	Merge func(x, y float64, xe, ye Edge) float64
+}
+
+var (
+	_ Undirected         = UndirectWeighted{}
+	_ WeightedUndirected = UndirectWeighted{}
+)
+
+// Node returns the node with the given ID if it exists in the graph,
+// and nil otherwise.
+func (g UndirectWeighted) Node(id int64) Node { return g.G.Node(id) }
+
+// Nodes returns all the nodes in the graph.
+func (g UndirectWeighted) Nodes() Nodes { return g.G.Nodes() }
+
+// From returns all nodes in g that can be reached directly from u.
+func (g UndirectWeighted) From(uid int64) Nodes {
+	return newNodeFilterIterator(g.G.From(uid), g.G.To(uid))
+}
+
+// HasEdgeBetween returns whether an edge exists between nodes x and y.
+func (g UndirectWeighted) HasEdgeBetween(xid, yid int64) bool { return g.G.HasEdgeBetween(xid, yid) }
+
+// Edge returns the edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+// If an edge exists, the Edge returned is an EdgePair. The weight of
+// the edge is determined by applying the Merge func to the weights of the
+// edges between u and v.
+func (g UndirectWeighted) Edge(uid, vid int64) Edge { return g.WeightedEdgeBetween(uid, vid) }
+
+// WeightedEdge returns the weighted edge from u to v if such an edge exists and nil otherwise.
+// The node v must be directly reachable from u as defined by the From method.
+// If an edge exists, the Edge returned is an EdgePair. The weight of
+// the edge is determined by applying the Merge func to the weights of the
+// edges between u and v.
+func (g UndirectWeighted) WeightedEdge(uid, vid int64) WeightedEdge {
+	return g.WeightedEdgeBetween(uid, vid)
+}
+
+// EdgeBetween returns the edge between nodes x and y. If an edge exists, the
+// Edge returned is an EdgePair. The weight of the edge is determined by
+// applying the Merge func to the weights of edges between x and y.
+func (g UndirectWeighted) EdgeBetween(xid, yid int64) Edge {
+	return g.WeightedEdgeBetween(xid, yid)
+}
+
+// WeightedEdgeBetween returns the weighted edge between nodes x and y. If an edge exists, the
+// Edge returned is an EdgePair. The weight of the edge is determined by
+// applying the Merge func to the weights of edges between x and y.
+func (g UndirectWeighted) WeightedEdgeBetween(xid, yid int64) WeightedEdge {
+	fe := g.G.Edge(xid, yid)
+	re := g.G.Edge(yid, xid)
+	if fe == nil && re == nil {
+		return nil
+	}
+
+	f, ok := g.G.Weight(xid, yid)
+	if !ok {
+		f = g.Absent
+	}
+	r, ok := g.G.Weight(yid, xid)
+	if !ok {
+		r = g.Absent
+	}
+
+	var w float64
+	if g.Merge == nil {
+		w = (f + r) / 2
+	} else {
+		w = g.Merge(f, r, fe, re)
+	}
+	return WeightedEdgePair{EdgePair: [2]Edge{fe, re}, W: w}
+}
+
+// Weight returns the weight for the edge between x and y if Edge(x, y) returns a non-nil Edge.
+// If x and y are the same node the internal node weight is returned. If there is no joining
+// edge between the two nodes the weight value returned is zero. Weight returns true if an edge
+// exists between x and y or if x and y have the same ID, false otherwise.
+func (g UndirectWeighted) Weight(xid, yid int64) (w float64, ok bool) {
+	fe := g.G.Edge(xid, yid)
+	re := g.G.Edge(yid, xid)
+
+	f, fOk := g.G.Weight(xid, yid)
+	if !fOk {
+		f = g.Absent
+	}
+	r, rOK := g.G.Weight(yid, xid)
+	if !rOK {
+		r = g.Absent
+	}
+	ok = fOk || rOK
+
+	if g.Merge == nil {
+		return (f + r) / 2, ok
+	}
+	return g.Merge(f, r, fe, re), ok
+}
+
+// EdgePair is an opposed pair of directed edges.
+type EdgePair [2]Edge
+
+// From returns the from node of the first non-nil edge, or nil.
+func (e EdgePair) From() Node {
+	if e[0] != nil {
+		return e[0].From()
+	} else if e[1] != nil {
+		return e[1].From()
+	}
+	return nil
+}
+
+// To returns the to node of the first non-nil edge, or nil.
+func (e EdgePair) To() Node {
+	if e[0] != nil {
+		return e[0].To()
+	} else if e[1] != nil {
+		return e[1].To()
+	}
+	return nil
+}
+
+// ReversedEdge returns a new Edge with the end point of the
+// edges in the pair swapped.
+func (e EdgePair) ReversedEdge() Edge {
+	if e[0] != nil {
+		e[0] = e[0].ReversedEdge()
+	}
+	if e[1] != nil {
+		e[1] = e[1].ReversedEdge()
+	}
+	return e
+}
+
+// WeightedEdgePair is an opposed pair of directed edges.
+type WeightedEdgePair struct {
+	EdgePair
+	W float64
+}
+
+// ReversedEdge returns a new Edge with the end point of the
+// edges in the pair swapped.
+func (e WeightedEdgePair) ReversedEdge() Edge {
+	e.EdgePair = e.EdgePair.ReversedEdge().(EdgePair)
+	return e
+}
+
+// Weight returns the merged edge weights of the two edges.
+func (e WeightedEdgePair) Weight() float64 { return e.W }
+
+// nodeFilterIterator combines two Nodes to produce a single stream of
+// unique nodes.
+type nodeFilterIterator struct {
+	a, b Nodes
+
+	// unique indicates the node in b with the key ID is unique.
+	unique map[int64]bool
+}
+
+func newNodeFilterIterator(a, b Nodes) *nodeFilterIterator {
+	n := nodeFilterIterator{a: a, b: b, unique: make(map[int64]bool)}
+	for n.b.Next() {
+		n.unique[n.b.Node().ID()] = true
+	}
+	n.b.Reset()
+	for n.a.Next() {
+		n.unique[n.a.Node().ID()] = false
+	}
+	n.a.Reset()
+	return &n
+}
+
+func (n *nodeFilterIterator) Len() int {
+	return len(n.unique)
+}
+
+func (n *nodeFilterIterator) Next() bool {
+	n.Len()
+	if n.a.Next() {
+		return true
+	}
+	for n.b.Next() {
+		if n.unique[n.b.Node().ID()] {
+			return true
+		}
+	}
+	return false
+}
+
+func (n *nodeFilterIterator) Node() Node {
+	if n.a.Len() != 0 {
+		return n.a.Node()
+	}
+	return n.b.Node()
+}
+
+func (n *nodeFilterIterator) Reset() {
+	n.a.Reset()
+	n.b.Reset()
+}

vendor/gonum.org/v1/gonum/internal/asm/c128/axpyinc_amd64.s

@@ -0,0 +1,134 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+// MOVDDUP X2, X3
+#define MOVDDUP_X2_X3 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xDA
+// MOVDDUP X4, X5
+#define MOVDDUP_X4_X5 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xEC
+// MOVDDUP X6, X7
+#define MOVDDUP_X6_X7 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xFE
+// MOVDDUP X8, X9
+#define MOVDDUP_X8_X9 BYTE $0xF2; BYTE $0x45; BYTE $0x0F; BYTE $0x12; BYTE $0xC8
+
+// ADDSUBPD X2, X3
+#define ADDSUBPD_X2_X3 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xDA
+// ADDSUBPD X4, X5
+#define ADDSUBPD_X4_X5 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xEC
+// ADDSUBPD X6, X7
+#define ADDSUBPD_X6_X7 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xFE
+// ADDSUBPD X8, X9
+#define ADDSUBPD_X8_X9 BYTE $0x66; BYTE $0x45; BYTE $0x0F; BYTE $0xD0; BYTE $0xC8
+
+// func AxpyInc(alpha complex128, x, y []complex128, n, incX, incY, ix, iy uintptr)
+TEXT ·AxpyInc(SB), NOSPLIT, $0
+	MOVQ   x_base+16(FP), SI // SI = &x
+	MOVQ   y_base+40(FP), DI // DI = &y
+	MOVQ   n+64(FP), CX      // CX = n
+	CMPQ   CX, $0            // if n==0 { return }
+	JE     axpyi_end
+	MOVQ   ix+88(FP), R8     // R8 = ix  // Load the first index
+	SHLQ   $4, R8            // R8 *= sizeof(complex128)
+	MOVQ   iy+96(FP), R9     // R9 = iy
+	SHLQ   $4, R9            // R9 *= sizeof(complex128)
+	LEAQ   (SI)(R8*1), SI    // SI = &(x[ix])
+	LEAQ   (DI)(R9*1), DI    // DI = &(y[iy])
+	MOVQ   DI, DX            // DX = DI      // Separate Read/Write pointers
+	MOVQ   incX+72(FP), R8   // R8 = incX
+	SHLQ   $4, R8            // R8 *= sizeof(complex128)
+	MOVQ   incY+80(FP), R9   // R9 = iy
+	SHLQ   $4, R9            // R9 *= sizeof(complex128)
+	MOVUPS alpha+0(FP), X0   // X0 = { imag(a), real(a) }
+	MOVAPS X0, X1
+	SHUFPD $0x1, X1, X1      // X1 = { real(a), imag(a) }
+	MOVAPS X0, X10           // Copy X0 and X1 for pipelining
+	MOVAPS X1, X11
+	MOVQ   CX, BX
+	ANDQ   $3, CX            // CX = n % 4
+	SHRQ   $2, BX            // BX = floor( n / 4 )
+	JZ     axpyi_tail        // if BX == 0 { goto axpyi_tail }
+
+axpyi_loop: // do {
+	MOVUPS (SI), X2       // X_i = { imag(x[i]), real(x[i]) }
+	MOVUPS (SI)(R8*1), X4
+	LEAQ   (SI)(R8*2), SI // SI = &(SI[incX*2])
+	MOVUPS (SI), X6
+	MOVUPS (SI)(R8*1), X8
+
+	// X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_X2_X3
+	MOVDDUP_X4_X5
+	MOVDDUP_X6_X7
+	MOVDDUP_X8_X9
+
+	// X_i = { imag(x[i]), imag(x[i]) }
+	SHUFPD $0x3, X2, X2
+	SHUFPD $0x3, X4, X4
+	SHUFPD $0x3, X6, X6
+	SHUFPD $0x3, X8, X8
+
+	// X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	// X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+	MULPD X1, X2
+	MULPD X0, X3
+	MULPD X11, X4
+	MULPD X10, X5
+	MULPD X1, X6
+	MULPD X0, X7
+	MULPD X11, X8
+	MULPD X10, X9
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDSUBPD_X4_X5
+	ADDSUBPD_X6_X7
+	ADDSUBPD_X8_X9
+
+	// X_(i+1) = { imag(result[i]) + imag(y[i]), real(result[i]) + real(y[i]) }
+	ADDPD  (DX), X3
+	ADDPD  (DX)(R9*1), X5
+	LEAQ   (DX)(R9*2), DX // DX = &(DX[incY*2])
+	ADDPD  (DX), X7
+	ADDPD  (DX)(R9*1), X9
+	MOVUPS X3, (DI)       // dst[i] = X_(i+1)
+	MOVUPS X5, (DI)(R9*1)
+	LEAQ   (DI)(R9*2), DI
+	MOVUPS X7, (DI)
+	MOVUPS X9, (DI)(R9*1)
+	LEAQ   (SI)(R8*2), SI // SI = &(SI[incX*2])
+	LEAQ   (DX)(R9*2), DX // DX = &(DX[incY*2])
+	LEAQ   (DI)(R9*2), DI // DI = &(DI[incY*2])
+	DECQ   BX
+	JNZ    axpyi_loop     // } while --BX > 0
+	CMPQ   CX, $0         // if CX == 0 { return }
+	JE     axpyi_end
+
+axpyi_tail: // do {
+	MOVUPS (SI), X2     // X_i = { imag(x[i]), real(x[i]) }
+	MOVDDUP_X2_X3       // X_(i+1) = { real(x[i], real(x[i]) }
+	SHUFPD $0x3, X2, X2 // X_i = { imag(x[i]), imag(x[i]) }
+	MULPD  X1, X2       // X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	MULPD  X0, X3       // X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+
+	// X_(i+1) = { imag(result[i]) + imag(y[i]), real(result[i]) + real(y[i]) }
+	ADDPD  (DI), X3
+	MOVUPS X3, (DI)   // y[i] = X_i
+	ADDQ   R8, SI     // SI = &(SI[incX])
+	ADDQ   R9, DI     // DI = &(DI[incY])
+	LOOP   axpyi_tail // } while --CX > 0
+
+axpyi_end:
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/axpyincto_amd64.s

@@ -0,0 +1,141 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+// MOVDDUP X2, X3
+#define MOVDDUP_X2_X3 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xDA
+// MOVDDUP X4, X5
+#define MOVDDUP_X4_X5 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xEC
+// MOVDDUP X6, X7
+#define MOVDDUP_X6_X7 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xFE
+// MOVDDUP X8, X9
+#define MOVDDUP_X8_X9 BYTE $0xF2; BYTE $0x45; BYTE $0x0F; BYTE $0x12; BYTE $0xC8
+
+// ADDSUBPD X2, X3
+#define ADDSUBPD_X2_X3 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xDA
+// ADDSUBPD X4, X5
+#define ADDSUBPD_X4_X5 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xEC
+// ADDSUBPD X6, X7
+#define ADDSUBPD_X6_X7 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xFE
+// ADDSUBPD X8, X9
+#define ADDSUBPD_X8_X9 BYTE $0x66; BYTE $0x45; BYTE $0x0F; BYTE $0xD0; BYTE $0xC8
+
+// func AxpyIncTo(dst []complex128, incDst, idst uintptr, alpha complex128, x, y []complex128, n, incX, incY, ix, iy uintptr)
+TEXT ·AxpyIncTo(SB), NOSPLIT, $0
+	MOVQ   dst_base+0(FP), DI // DI = &dst
+	MOVQ   x_base+56(FP), SI  // SI = &x
+	MOVQ   y_base+80(FP), DX  // DX = &y
+	MOVQ   n+104(FP), CX      // CX = n
+	CMPQ   CX, $0             // if n==0 { return }
+	JE     axpyi_end
+	MOVQ   ix+128(FP), R8     // R8 = ix  // Load the first index
+	SHLQ   $4, R8             // R8 *= sizeof(complex128)
+	MOVQ   iy+136(FP), R9     // R9 = iy
+	SHLQ   $4, R9             // R9 *= sizeof(complex128)
+	MOVQ   idst+32(FP), R10   // R10 = idst
+	SHLQ   $4, R10            // R10 *= sizeof(complex128)
+	LEAQ   (SI)(R8*1), SI     // SI = &(x[ix])
+	LEAQ   (DX)(R9*1), DX     // DX = &(y[iy])
+	LEAQ   (DI)(R10*1), DI    // DI = &(dst[idst])
+	MOVQ   incX+112(FP), R8   // R8 = incX
+	SHLQ   $4, R8             // R8 *= sizeof(complex128)
+	MOVQ   incY+120(FP), R9   // R9 = incY
+	SHLQ   $4, R9             // R9 *= sizeof(complex128)
+	MOVQ   incDst+24(FP), R10 // R10 = incDst
+	SHLQ   $4, R10            // R10 *= sizeof(complex128)
+	MOVUPS alpha+40(FP), X0   // X0 = { imag(a), real(a) }
+	MOVAPS X0, X1
+	SHUFPD $0x1, X1, X1       // X1 = { real(a), imag(a) }
+	MOVAPS X0, X10            // Copy X0 and X1 for pipelining
+	MOVAPS X1, X11
+	MOVQ   CX, BX
+	ANDQ   $3, CX             // CX = n % 4
+	SHRQ   $2, BX             // BX = floor( n / 4 )
+	JZ     axpyi_tail         // if BX == 0 { goto axpyi_tail }
+
+axpyi_loop: // do {
+	MOVUPS (SI), X2       // X_i = { imag(x[i]), real(x[i]) }
+	MOVUPS (SI)(R8*1), X4
+	LEAQ   (SI)(R8*2), SI // SI = &(SI[incX*2])
+
+	MOVUPS (SI), X6
+	MOVUPS (SI)(R8*1), X8
+
+	// X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_X2_X3
+	MOVDDUP_X4_X5
+	MOVDDUP_X6_X7
+	MOVDDUP_X8_X9
+
+	// X_i = { imag(x[i]), imag(x[i]) }
+	SHUFPD $0x3, X2, X2
+	SHUFPD $0x3, X4, X4
+	SHUFPD $0x3, X6, X6
+	SHUFPD $0x3, X8, X8
+
+	// X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	// X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+	MULPD X1, X2
+	MULPD X0, X3
+	MULPD X11, X4
+	MULPD X10, X5
+	MULPD X1, X6
+	MULPD X0, X7
+	MULPD X11, X8
+	MULPD X10, X9
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDSUBPD_X4_X5
+	ADDSUBPD_X6_X7
+	ADDSUBPD_X8_X9
+
+	// X_(i+1) = { imag(result[i]) + imag(y[i]), real(result[i]) + real(y[i]) }
+	ADDPD  (DX), X3
+	ADDPD  (DX)(R9*1), X5
+	LEAQ   (DX)(R9*2), DX  // DX = &(DX[incY*2])
+	ADDPD  (DX), X7
+	ADDPD  (DX)(R9*1), X9
+	MOVUPS X3, (DI)        // dst[i] = X_(i+1)
+	MOVUPS X5, (DI)(R10*1)
+	LEAQ   (DI)(R10*2), DI
+	MOVUPS X7, (DI)
+	MOVUPS X9, (DI)(R10*1)
+	LEAQ   (SI)(R8*2), SI  // SI = &(SI[incX*2])
+	LEAQ   (DX)(R9*2), DX  // DX = &(DX[incY*2])
+	LEAQ   (DI)(R10*2), DI // DI = &(DI[incDst*2])
+	DECQ   BX
+	JNZ    axpyi_loop      // } while --BX > 0
+	CMPQ   CX, $0          // if CX == 0 { return }
+	JE     axpyi_end
+
+axpyi_tail: // do {
+	MOVUPS (SI), X2     // X_i = { imag(x[i]), real(x[i]) }
+	MOVDDUP_X2_X3       // X_(i+1) = { real(x[i], real(x[i]) }
+	SHUFPD $0x3, X2, X2 // X_i = { imag(x[i]), imag(x[i]) }
+	MULPD  X1, X2       // X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	MULPD  X0, X3       // X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+
+	// X_(i+1) = { imag(result[i]) + imag(y[i]), real(result[i]) + real(y[i]) }
+	ADDPD  (DX), X3
+	MOVUPS X3, (DI)   // y[i] X_(i+1)
+	ADDQ   R8, SI     // SI += incX
+	ADDQ   R9, DX     // DX += incY
+	ADDQ   R10, DI    // DI += incDst
+	LOOP   axpyi_tail // } while --CX > 0
+
+axpyi_end:
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/axpyunitary_amd64.s

@@ -0,0 +1,122 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+// MOVDDUP X2, X3
+#define MOVDDUP_X2_X3 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xDA
+// MOVDDUP X4, X5
+#define MOVDDUP_X4_X5 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xEC
+// MOVDDUP X6, X7
+#define MOVDDUP_X6_X7 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xFE
+// MOVDDUP X8, X9
+#define MOVDDUP_X8_X9 BYTE $0xF2; BYTE $0x45; BYTE $0x0F; BYTE $0x12; BYTE $0xC8
+
+// ADDSUBPD X2, X3
+#define ADDSUBPD_X2_X3 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xDA
+// ADDSUBPD X4, X5
+#define ADDSUBPD_X4_X5 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xEC
+// ADDSUBPD X6, X7
+#define ADDSUBPD_X6_X7 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xFE
+// ADDSUBPD X8, X9
+#define ADDSUBPD_X8_X9 BYTE $0x66; BYTE $0x45; BYTE $0x0F; BYTE $0xD0; BYTE $0xC8
+
+// func AxpyUnitary(alpha complex128, x, y []complex128)
+TEXT ·AxpyUnitary(SB), NOSPLIT, $0
+	MOVQ    x_base+16(FP), SI // SI = &x
+	MOVQ    y_base+40(FP), DI // DI = &y
+	MOVQ    x_len+24(FP), CX  // CX = min( len(x), len(y) )
+	CMPQ    y_len+48(FP), CX
+	CMOVQLE y_len+48(FP), CX
+	CMPQ    CX, $0            // if CX == 0 { return }
+	JE      caxy_end
+	PXOR    X0, X0            // Clear work registers and cache-align loop
+	PXOR    X1, X1
+	MOVUPS  alpha+0(FP), X0   // X0 = { imag(a), real(a) }
+	MOVAPS  X0, X1
+	SHUFPD  $0x1, X1, X1      // X1 = { real(a), imag(a) }
+	XORQ    AX, AX            // i = 0
+	MOVAPS  X0, X10           // Copy X0 and X1 for pipelining
+	MOVAPS  X1, X11
+	MOVQ    CX, BX
+	ANDQ    $3, CX            // CX = n % 4
+	SHRQ    $2, BX            // BX = floor( n / 4 )
+	JZ      caxy_tail         // if BX == 0 { goto caxy_tail }
+
+caxy_loop: // do {
+	MOVUPS (SI)(AX*8), X2   // X_i = { imag(x[i]), real(x[i]) }
+	MOVUPS 16(SI)(AX*8), X4
+	MOVUPS 32(SI)(AX*8), X6
+	MOVUPS 48(SI)(AX*8), X8
+
+	// X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_X2_X3
+	MOVDDUP_X4_X5
+	MOVDDUP_X6_X7
+	MOVDDUP_X8_X9
+
+	// X_i = { imag(x[i]), imag(x[i]) }
+	SHUFPD $0x3, X2, X2
+	SHUFPD $0x3, X4, X4
+	SHUFPD $0x3, X6, X6
+	SHUFPD $0x3, X8, X8
+
+	// X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	// X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+	MULPD X1, X2
+	MULPD X0, X3
+	MULPD X11, X4
+	MULPD X10, X5
+	MULPD X1, X6
+	MULPD X0, X7
+	MULPD X11, X8
+	MULPD X10, X9
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDSUBPD_X4_X5
+	ADDSUBPD_X6_X7
+	ADDSUBPD_X8_X9
+
+	// X_(i+1) = { imag(result[i]) + imag(y[i]), real(result[i]) + real(y[i]) }
+	ADDPD  (DI)(AX*8), X3
+	ADDPD  16(DI)(AX*8), X5
+	ADDPD  32(DI)(AX*8), X7
+	ADDPD  48(DI)(AX*8), X9
+	MOVUPS X3, (DI)(AX*8)   // y[i] = X_(i+1)
+	MOVUPS X5, 16(DI)(AX*8)
+	MOVUPS X7, 32(DI)(AX*8)
+	MOVUPS X9, 48(DI)(AX*8)
+	ADDQ   $8, AX           // i += 8
+	DECQ   BX
+	JNZ    caxy_loop        // } while --BX > 0
+	CMPQ   CX, $0           // if CX == 0 { return }
+	JE     caxy_end
+
+caxy_tail: // do {
+	MOVUPS (SI)(AX*8), X2 // X_i = { imag(x[i]), real(x[i]) }
+	MOVDDUP_X2_X3         // X_(i+1) = { real(x[i], real(x[i]) }
+	SHUFPD $0x3, X2, X2   // X_i = { imag(x[i]), imag(x[i]) }
+	MULPD  X1, X2         // X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	MULPD  X0, X3         // X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+
+	// X_(i+1) = { imag(result[i]) + imag(y[i]), real(result[i]) + real(y[i]) }
+	ADDPD  (DI)(AX*8), X3
+	MOVUPS X3, (DI)(AX*8) // y[i] = X_(i+1)
+	ADDQ   $2, AX         // i += 2
+	LOOP   caxy_tail      // }  while --CX > 0
+
+caxy_end:
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/axpyunitaryto_amd64.s

@@ -0,0 +1,123 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+// MOVDDUP X2, X3
+#define MOVDDUP_X2_X3 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xDA
+// MOVDDUP X4, X5
+#define MOVDDUP_X4_X5 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xEC
+// MOVDDUP X6, X7
+#define MOVDDUP_X6_X7 BYTE $0xF2; BYTE $0x0F; BYTE $0x12; BYTE $0xFE
+// MOVDDUP X8, X9
+#define MOVDDUP_X8_X9 BYTE $0xF2; BYTE $0x45; BYTE $0x0F; BYTE $0x12; BYTE $0xC8
+
+// ADDSUBPD X2, X3
+#define ADDSUBPD_X2_X3 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xDA
+// ADDSUBPD X4, X5
+#define ADDSUBPD_X4_X5 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xEC
+// ADDSUBPD X6, X7
+#define ADDSUBPD_X6_X7 BYTE $0x66; BYTE $0x0F; BYTE $0xD0; BYTE $0xFE
+// ADDSUBPD X8, X9
+#define ADDSUBPD_X8_X9 BYTE $0x66; BYTE $0x45; BYTE $0x0F; BYTE $0xD0; BYTE $0xC8
+
+// func AxpyUnitaryTo(dst []complex128, alpha complex64, x, y []complex128)
+TEXT ·AxpyUnitaryTo(SB), NOSPLIT, $0
+	MOVQ    dst_base+0(FP), DI // DI = &dst
+	MOVQ    x_base+40(FP), SI  // SI = &x
+	MOVQ    y_base+64(FP), DX  // DX = &y
+	MOVQ    x_len+48(FP), CX   // CX = min( len(x), len(y), len(dst) )
+	CMPQ    y_len+72(FP), CX
+	CMOVQLE y_len+72(FP), CX
+	CMPQ    dst_len+8(FP), CX
+	CMOVQLE dst_len+8(FP), CX
+	CMPQ    CX, $0             // if CX == 0 { return }
+	JE      caxy_end
+	MOVUPS  alpha+24(FP), X0   // X0 = { imag(a), real(a) }
+	MOVAPS  X0, X1
+	SHUFPD  $0x1, X1, X1       // X1 = { real(a), imag(a) }
+	XORQ    AX, AX             // i = 0
+	MOVAPS  X0, X10            // Copy X0 and X1 for pipelining
+	MOVAPS  X1, X11
+	MOVQ    CX, BX
+	ANDQ    $3, CX             // CX = n % 4
+	SHRQ    $2, BX             // BX = floor( n / 4 )
+	JZ      caxy_tail          // if BX == 0 { goto caxy_tail }
+
+caxy_loop: // do {
+	MOVUPS (SI)(AX*8), X2   // X_i = { imag(x[i]), real(x[i]) }
+	MOVUPS 16(SI)(AX*8), X4
+	MOVUPS 32(SI)(AX*8), X6
+	MOVUPS 48(SI)(AX*8), X8
+
+	// X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_X2_X3 // Load and duplicate imag elements (xi, xi)
+	MOVDDUP_X4_X5
+	MOVDDUP_X6_X7
+	MOVDDUP_X8_X9
+
+	// X_i = { imag(x[i]), imag(x[i]) }
+	SHUFPD $0x3, X2, X2 // duplicate real elements (xr, xr)
+	SHUFPD $0x3, X4, X4
+	SHUFPD $0x3, X6, X6
+	SHUFPD $0x3, X8, X8
+
+	// X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	// X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+	MULPD X1, X2
+	MULPD X0, X3
+	MULPD X11, X4
+	MULPD X10, X5
+	MULPD X1, X6
+	MULPD X0, X7
+	MULPD X11, X8
+	MULPD X10, X9
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDSUBPD_X4_X5
+	ADDSUBPD_X6_X7
+	ADDSUBPD_X8_X9
+
+	// X_(i+1) = { imag(result[i]) + imag(y[i]), real(result[i]) + real(y[i]) }
+	ADDPD  (DX)(AX*8), X3
+	ADDPD  16(DX)(AX*8), X5
+	ADDPD  32(DX)(AX*8), X7
+	ADDPD  48(DX)(AX*8), X9
+	MOVUPS X3, (DI)(AX*8)   // y[i] = X_(i+1)
+	MOVUPS X5, 16(DI)(AX*8)
+	MOVUPS X7, 32(DI)(AX*8)
+	MOVUPS X9, 48(DI)(AX*8)
+	ADDQ   $8, AX           // i += 8
+	DECQ   BX
+	JNZ    caxy_loop        // } while --BX > 0
+	CMPQ   CX, $0           // if CX == 0 { return }
+	JE     caxy_end
+
+caxy_tail: // Same calculation, but read in values to avoid trampling memory
+	MOVUPS (SI)(AX*8), X2 // X_i = { imag(x[i]), real(x[i]) }
+	MOVDDUP_X2_X3         // X_(i+1) = { real(x[i], real(x[i]) }
+	SHUFPD $0x3, X2, X2   // X_i = { imag(x[i]), imag(x[i]) }
+	MULPD  X1, X2         // X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	MULPD  X0, X3         // X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+
+	// X_(i+1) = { imag(result[i]) + imag(y[i]), real(result[i]) + real(y[i]) }
+	ADDPD  (DX)(AX*8), X3
+	MOVUPS X3, (DI)(AX*8) // y[i] = X_(i+1)
+	ADDQ   $2, AX         // i += 2
+	LOOP   caxy_tail      // }  while --CX > 0
+
+caxy_end:
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/doc.go

@@ -0,0 +1,6 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package c128 provides complex128 vector primitives.
+package c128 // import "gonum.org/v1/gonum/internal/asm/c128"

vendor/gonum.org/v1/gonum/internal/asm/c128/dotcinc_amd64.s

@@ -0,0 +1,153 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+#define MOVDDUP_XPTR__X3    LONG $0x1E120FF2 // MOVDDUP (SI), X3
+#define MOVDDUP_XPTR_INCX__X5    LONG $0x120F42F2; WORD $0x062C // MOVDDUP (SI)(R8*1), X5
+#define MOVDDUP_XPTR_INCX_2__X7    LONG $0x120F42F2; WORD $0x463C // MOVDDUP (SI)(R8*2), X7
+#define MOVDDUP_XPTR_INCx3X__X9    LONG $0x120F46F2; WORD $0x0E0C // MOVDDUP (SI)(R9*1), X9
+
+#define MOVDDUP_8_XPTR__X2    LONG $0x56120FF2; BYTE $0x08 // MOVDDUP 8(SI), X2
+#define MOVDDUP_8_XPTR_INCX__X4    LONG $0x120F42F2; WORD $0x0664; BYTE $0x08 // MOVDDUP 8(SI)(R8*1), X4
+#define MOVDDUP_8_XPTR_INCX_2__X6    LONG $0x120F42F2; WORD $0x4674; BYTE $0x08 // MOVDDUP 8(SI)(R8*2), X6
+#define MOVDDUP_8_XPTR_INCx3X__X8    LONG $0x120F46F2; WORD $0x0E44; BYTE $0x08 // MOVDDUP 8(SI)(R9*1), X8
+
+#define ADDSUBPD_X2_X3    LONG $0xDAD00F66 // ADDSUBPD X2, X3
+#define ADDSUBPD_X4_X5    LONG $0xECD00F66 // ADDSUBPD X4, X5
+#define ADDSUBPD_X6_X7    LONG $0xFED00F66 // ADDSUBPD X6, X7
+#define ADDSUBPD_X8_X9    LONG $0xD00F4566; BYTE $0xC8 // ADDSUBPD X8, X9
+
+#define X_PTR SI
+#define Y_PTR DI
+#define LEN CX
+#define TAIL BX
+#define SUM X0
+#define P_SUM X1
+#define INC_X R8
+#define INCx3_X R9
+#define INC_Y R10
+#define INCx3_Y R11
+#define NEG1 X15
+#define P_NEG1 X14
+
+// func DotcInc(x, y []complex128, n, incX, incY, ix, iy uintptr) (sum complex128)
+TEXT ·DotcInc(SB), NOSPLIT, $0
+	MOVQ   x_base+0(FP), X_PTR       // X_PTR = &x
+	MOVQ   y_base+24(FP), Y_PTR      // Y_PTR = &y
+	MOVQ   n+48(FP), LEN             // LEN = n
+	PXOR   SUM, SUM                  // SUM = 0
+	CMPQ   LEN, $0                   // if LEN == 0 { return }
+	JE     dot_end
+	PXOR   P_SUM, P_SUM              // P_SUM = 0
+	MOVQ   ix+72(FP), INC_X          // INC_X = ix * sizeof(complex128)
+	SHLQ   $4, INC_X
+	MOVQ   iy+80(FP), INC_Y          // INC_Y = iy * sizeof(complex128)
+	SHLQ   $4, INC_Y
+	LEAQ   (X_PTR)(INC_X*1), X_PTR   // X_PTR = &(X_PTR[ix])
+	LEAQ   (Y_PTR)(INC_Y*1), Y_PTR   // Y_PTR = &(Y_PTR[iy])
+	MOVQ   incX+56(FP), INC_X        // INC_X = incX
+	SHLQ   $4, INC_X                 // INC_X *=  sizeof(complex128)
+	MOVQ   incY+64(FP), INC_Y        // INC_Y = incY
+	SHLQ   $4, INC_Y                 // INC_Y *=  sizeof(complex128)
+	MOVSD  $(-1.0), NEG1
+	SHUFPD $0, NEG1, NEG1            // { -1, -1 }
+	MOVQ   LEN, TAIL
+	ANDQ   $3, TAIL                  // TAIL = n % 4
+	SHRQ   $2, LEN                   // LEN = floor( n / 4 )
+	JZ     dot_tail                  // if n <= 4 { goto dot_tail }
+	MOVAPS NEG1, P_NEG1              // Copy NEG1 to P_NEG1 for pipelining
+	LEAQ   (INC_X)(INC_X*2), INCx3_X // INCx3_X = 3 * incX * sizeof(complex128)
+	LEAQ   (INC_Y)(INC_Y*2), INCx3_Y // INCx3_Y = 3 * incY * sizeof(complex128)
+
+dot_loop: // do {
+	MOVDDUP_XPTR__X3        // X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_XPTR_INCX__X5
+	MOVDDUP_XPTR_INCX_2__X7
+	MOVDDUP_XPTR_INCx3X__X9
+
+	MOVDDUP_8_XPTR__X2        // X_i = { imag(x[i]), imag(x[i]) }
+	MOVDDUP_8_XPTR_INCX__X4
+	MOVDDUP_8_XPTR_INCX_2__X6
+	MOVDDUP_8_XPTR_INCx3X__X8
+
+	// X_i = { -imag(x[i]), -imag(x[i]) }
+	MULPD NEG1, X2
+	MULPD P_NEG1, X4
+	MULPD NEG1, X6
+	MULPD P_NEG1, X8
+
+	// X_j = { imag(y[i]), real(y[i]) }
+	MOVUPS (Y_PTR), X10
+	MOVUPS (Y_PTR)(INC_Y*1), X11
+	MOVUPS (Y_PTR)(INC_Y*2), X12
+	MOVUPS (Y_PTR)(INCx3_Y*1), X13
+
+	// X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+	MULPD X10, X3
+	MULPD X11, X5
+	MULPD X12, X7
+	MULPD X13, X9
+
+	// X_j     = { real(y[i]), imag(y[i]) }
+	SHUFPD $0x1, X10, X10
+	SHUFPD $0x1, X11, X11
+	SHUFPD $0x1, X12, X12
+	SHUFPD $0x1, X13, X13
+
+	// X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	MULPD X10, X2
+	MULPD X11, X4
+	MULPD X12, X6
+	MULPD X13, X8
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDSUBPD_X4_X5
+	ADDSUBPD_X6_X7
+	ADDSUBPD_X8_X9
+
+	// psum += result[i]
+	ADDPD X3, SUM
+	ADDPD X5, P_SUM
+	ADDPD X7, SUM
+	ADDPD X9, P_SUM
+
+	LEAQ (X_PTR)(INC_X*4), X_PTR // X_PTR = &(X_PTR[incX*4])
+	LEAQ (Y_PTR)(INC_Y*4), Y_PTR // Y_PTR = &(Y_PTR[incY*4])
+
+	DECQ  LEN
+	JNZ   dot_loop   // } while --LEN > 0
+	ADDPD P_SUM, SUM // sum += psum
+	CMPQ  TAIL, $0   // if TAIL == 0 { return }
+	JE    dot_end
+
+dot_tail: // do {
+	MOVDDUP_XPTR__X3      // X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_8_XPTR__X2    // X_i = { imag(x[i]), imag(x[i]) }
+	MULPD  NEG1, X2       // X_i     = { -imag(x[i])          , -imag(x[i])           }
+	MOVUPS (Y_PTR), X10   // X_j     = {  imag(y[i])          ,  real(y[i])           }
+	MULPD  X10, X3        // X_(i+1) = {  imag(a) * real(x[i]),  real(a) * real(x[i]) }
+	SHUFPD $0x1, X10, X10 // X_j     = {  real(y[i])          ,  imag(y[i])           }
+	MULPD  X10, X2        // X_i     = {  real(a) * imag(x[i]),  imag(a) * imag(x[i]) }
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDPD X3, SUM      // sum += result[i]
+	ADDQ  INC_X, X_PTR // X_PTR += incX
+	ADDQ  INC_Y, Y_PTR // Y_PTR += incY
+	DECQ  TAIL
+	JNZ   dot_tail     // }  while --TAIL > 0
+
+dot_end:
+	MOVUPS SUM, sum+88(FP)
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/dotcunitary_amd64.s

@@ -0,0 +1,143 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+#define MOVDDUP_XPTR_IDX_8__X3    LONG $0x1C120FF2; BYTE $0xC6 // MOVDDUP (SI)(AX*8), X3
+#define MOVDDUP_16_XPTR_IDX_8__X5    LONG $0x6C120FF2; WORD $0x10C6 // MOVDDUP 16(SI)(AX*8), X5
+#define MOVDDUP_32_XPTR_IDX_8__X7    LONG $0x7C120FF2; WORD $0x20C6 // MOVDDUP 32(SI)(AX*8), X7
+#define MOVDDUP_48_XPTR_IDX_8__X9    LONG $0x120F44F2; WORD $0xC64C; BYTE $0x30 // MOVDDUP 48(SI)(AX*8), X9
+
+#define MOVDDUP_XPTR_IIDX_8__X2    LONG $0x14120FF2; BYTE $0xD6 // MOVDDUP (SI)(DX*8), X2
+#define MOVDDUP_16_XPTR_IIDX_8__X4    LONG $0x64120FF2; WORD $0x10D6 // MOVDDUP 16(SI)(DX*8), X4
+#define MOVDDUP_32_XPTR_IIDX_8__X6    LONG $0x74120FF2; WORD $0x20D6 // MOVDDUP 32(SI)(DX*8), X6
+#define MOVDDUP_48_XPTR_IIDX_8__X8    LONG $0x120F44F2; WORD $0xD644; BYTE $0x30 // MOVDDUP 48(SI)(DX*8), X8
+
+#define ADDSUBPD_X2_X3    LONG $0xDAD00F66 // ADDSUBPD X2, X3
+#define ADDSUBPD_X4_X5    LONG $0xECD00F66 // ADDSUBPD X4, X5
+#define ADDSUBPD_X6_X7    LONG $0xFED00F66 // ADDSUBPD X6, X7
+#define ADDSUBPD_X8_X9    LONG $0xD00F4566; BYTE $0xC8 // ADDSUBPD X8, X9
+
+#define X_PTR SI
+#define Y_PTR DI
+#define LEN CX
+#define TAIL BX
+#define SUM X0
+#define P_SUM X1
+#define IDX AX
+#define I_IDX DX
+#define NEG1 X15
+#define P_NEG1 X14
+
+// func DotcUnitary(x, y []complex128) (sum complex128)
+TEXT ·DotcUnitary(SB), NOSPLIT, $0
+	MOVQ    x_base+0(FP), X_PTR  // X_PTR = &x
+	MOVQ    y_base+24(FP), Y_PTR // Y_PTR = &y
+	MOVQ    x_len+8(FP), LEN     // LEN = min( len(x), len(y) )
+	CMPQ    y_len+32(FP), LEN
+	CMOVQLE y_len+32(FP), LEN
+	PXOR    SUM, SUM             // sum = 0
+	CMPQ    LEN, $0              // if LEN == 0 { return }
+	JE      dot_end
+	XORPS   P_SUM, P_SUM         // psum = 0
+	MOVSD   $(-1.0), NEG1
+	SHUFPD  $0, NEG1, NEG1       // { -1, -1 }
+	XORQ    IDX, IDX             // i := 0
+	MOVQ    $1, I_IDX            // j := 1
+	MOVQ    LEN, TAIL
+	ANDQ    $3, TAIL             // TAIL = floor( TAIL / 4 )
+	SHRQ    $2, LEN              // LEN = TAIL % 4
+	JZ      dot_tail             // if LEN == 0 { goto dot_tail }
+
+	MOVAPS NEG1, P_NEG1 // Copy NEG1 to P_NEG1 for pipelining
+
+dot_loop: // do {
+	MOVDDUP_XPTR_IDX_8__X3    // X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_16_XPTR_IDX_8__X5
+	MOVDDUP_32_XPTR_IDX_8__X7
+	MOVDDUP_48_XPTR_IDX_8__X9
+
+	MOVDDUP_XPTR_IIDX_8__X2    // X_i = { imag(x[i]), imag(x[i]) }
+	MOVDDUP_16_XPTR_IIDX_8__X4
+	MOVDDUP_32_XPTR_IIDX_8__X6
+	MOVDDUP_48_XPTR_IIDX_8__X8
+
+	// X_i = { -imag(x[i]), -imag(x[i]) }
+	MULPD NEG1, X2
+	MULPD P_NEG1, X4
+	MULPD NEG1, X6
+	MULPD P_NEG1, X8
+
+	// X_j = { imag(y[i]), real(y[i]) }
+	MOVUPS (Y_PTR)(IDX*8), X10
+	MOVUPS 16(Y_PTR)(IDX*8), X11
+	MOVUPS 32(Y_PTR)(IDX*8), X12
+	MOVUPS 48(Y_PTR)(IDX*8), X13
+
+	// X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+	MULPD X10, X3
+	MULPD X11, X5
+	MULPD X12, X7
+	MULPD X13, X9
+
+	// X_j = { real(y[i]), imag(y[i]) }
+	SHUFPD $0x1, X10, X10
+	SHUFPD $0x1, X11, X11
+	SHUFPD $0x1, X12, X12
+	SHUFPD $0x1, X13, X13
+
+	// X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	MULPD X10, X2
+	MULPD X11, X4
+	MULPD X12, X6
+	MULPD X13, X8
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDSUBPD_X4_X5
+	ADDSUBPD_X6_X7
+	ADDSUBPD_X8_X9
+
+	// psum += result[i]
+	ADDPD X3, SUM
+	ADDPD X5, P_SUM
+	ADDPD X7, SUM
+	ADDPD X9, P_SUM
+
+	ADDQ  $8, IDX    // IDX += 8
+	ADDQ  $8, I_IDX  // I_IDX += 8
+	DECQ  LEN
+	JNZ   dot_loop   // } while --LEN > 0
+	ADDPD P_SUM, SUM // sum += psum
+	CMPQ  TAIL, $0   // if TAIL == 0 { return }
+	JE    dot_end
+
+dot_tail: // do {
+	MOVDDUP_XPTR_IDX_8__X3     // X_(i+1) = {  real(x[i])          ,  real(x[i])           }
+	MOVDDUP_XPTR_IIDX_8__X2    // X_i     = {  imag(x[i])          ,  imag(x[i])           }
+	MULPD  NEG1, X2            // X_i     = { -imag(x[i])          , -imag(x[i])           }
+	MOVUPS (Y_PTR)(IDX*8), X10 // X_j     = {  imag(y[i])          ,  real(y[i])           }
+	MULPD  X10, X3             // X_(i+1) = {  imag(a) * real(x[i]),  real(a) * real(x[i]) }
+	SHUFPD $0x1, X10, X10      // X_j     = {  real(y[i])          ,  imag(y[i])           }
+	MULPD  X10, X2             // X_i     = {  real(a) * imag(x[i]),  imag(a) * imag(x[i]) }
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDPD X3, SUM   // SUM += result[i]
+	ADDQ  $2, IDX   // IDX += 2
+	ADDQ  $2, I_IDX // I_IDX += 2
+	DECQ  TAIL
+	JNZ   dot_tail  // }  while --TAIL > 0
+
+dot_end:
+	MOVUPS SUM, sum+48(FP)
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/dotuinc_amd64.s

@@ -0,0 +1,141 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+#define MOVDDUP_XPTR__X3    LONG $0x1E120FF2 // MOVDDUP (SI), X3
+#define MOVDDUP_XPTR_INCX__X5    LONG $0x120F42F2; WORD $0x062C // MOVDDUP (SI)(R8*1), X5
+#define MOVDDUP_XPTR_INCX_2__X7    LONG $0x120F42F2; WORD $0x463C // MOVDDUP (SI)(R8*2), X7
+#define MOVDDUP_XPTR_INCx3X__X9    LONG $0x120F46F2; WORD $0x0E0C // MOVDDUP (SI)(R9*1), X9
+
+#define MOVDDUP_8_XPTR__X2    LONG $0x56120FF2; BYTE $0x08 // MOVDDUP 8(SI), X2
+#define MOVDDUP_8_XPTR_INCX__X4    LONG $0x120F42F2; WORD $0x0664; BYTE $0x08 // MOVDDUP 8(SI)(R8*1), X4
+#define MOVDDUP_8_XPTR_INCX_2__X6    LONG $0x120F42F2; WORD $0x4674; BYTE $0x08 // MOVDDUP 8(SI)(R8*2), X6
+#define MOVDDUP_8_XPTR_INCx3X__X8    LONG $0x120F46F2; WORD $0x0E44; BYTE $0x08 // MOVDDUP 8(SI)(R9*1), X8
+
+#define ADDSUBPD_X2_X3    LONG $0xDAD00F66 // ADDSUBPD X2, X3
+#define ADDSUBPD_X4_X5    LONG $0xECD00F66 // ADDSUBPD X4, X5
+#define ADDSUBPD_X6_X7    LONG $0xFED00F66 // ADDSUBPD X6, X7
+#define ADDSUBPD_X8_X9    LONG $0xD00F4566; BYTE $0xC8 // ADDSUBPD X8, X9
+
+#define X_PTR SI
+#define Y_PTR DI
+#define LEN CX
+#define TAIL BX
+#define SUM X0
+#define P_SUM X1
+#define INC_X R8
+#define INCx3_X R9
+#define INC_Y R10
+#define INCx3_Y R11
+
+// func DotuInc(x, y []complex128, n, incX, incY, ix, iy uintptr) (sum complex128)
+TEXT ·DotuInc(SB), NOSPLIT, $0
+	MOVQ x_base+0(FP), X_PTR       // X_PTR = &x
+	MOVQ y_base+24(FP), Y_PTR      // Y_PTR = &y
+	MOVQ n+48(FP), LEN             // LEN = n
+	PXOR SUM, SUM                  // sum = 0
+	CMPQ LEN, $0                   // if LEN == 0 { return }
+	JE   dot_end
+	MOVQ ix+72(FP), INC_X          // INC_X = ix * sizeof(complex128)
+	SHLQ $4, INC_X
+	MOVQ iy+80(FP), INC_Y          // INC_Y = iy * sizeof(complex128)
+	SHLQ $4, INC_Y
+	LEAQ (X_PTR)(INC_X*1), X_PTR   // X_PTR = &(X_PTR[ix])
+	LEAQ (Y_PTR)(INC_Y*1), Y_PTR   // Y_PTR = &(Y_PTR[iy])
+	MOVQ incX+56(FP), INC_X        // INC_X = incX
+	SHLQ $4, INC_X                 // INC_X *=  sizeof(complex128)
+	MOVQ incY+64(FP), INC_Y        // INC_Y = incY
+	SHLQ $4, INC_Y                 // INC_Y *=  sizeof(complex128)
+	MOVQ LEN, TAIL
+	ANDQ $3, TAIL                  // LEN = LEN % 4
+	SHRQ $2, LEN                   // LEN = floor( LEN / 4 )
+	JZ   dot_tail                  // if LEN <= 4 { goto dot_tail }
+	PXOR P_SUM, P_SUM              // psum = 0
+	LEAQ (INC_X)(INC_X*2), INCx3_X // INCx3_X = 3 * incX * sizeof(complex128)
+	LEAQ (INC_Y)(INC_Y*2), INCx3_Y // INCx3_Y = 3 * incY * sizeof(complex128)
+
+dot_loop: // do {
+	MOVDDUP_XPTR__X3        // X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_XPTR_INCX__X5
+	MOVDDUP_XPTR_INCX_2__X7
+	MOVDDUP_XPTR_INCx3X__X9
+
+	MOVDDUP_8_XPTR__X2        // X_i = { imag(x[i]), imag(x[i]) }
+	MOVDDUP_8_XPTR_INCX__X4
+	MOVDDUP_8_XPTR_INCX_2__X6
+	MOVDDUP_8_XPTR_INCx3X__X8
+
+	// X_j = { imag(y[i]), real(y[i]) }
+	MOVUPS (Y_PTR), X10
+	MOVUPS (Y_PTR)(INC_Y*1), X11
+	MOVUPS (Y_PTR)(INC_Y*2), X12
+	MOVUPS (Y_PTR)(INCx3_Y*1), X13
+
+	// X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+	MULPD X10, X3
+	MULPD X11, X5
+	MULPD X12, X7
+	MULPD X13, X9
+
+	// X_j     = { real(y[i]), imag(y[i]) }
+	SHUFPD $0x1, X10, X10
+	SHUFPD $0x1, X11, X11
+	SHUFPD $0x1, X12, X12
+	SHUFPD $0x1, X13, X13
+
+	// X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	MULPD X10, X2
+	MULPD X11, X4
+	MULPD X12, X6
+	MULPD X13, X8
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDSUBPD_X4_X5
+	ADDSUBPD_X6_X7
+	ADDSUBPD_X8_X9
+
+	// psum += result[i]
+	ADDPD X3, SUM
+	ADDPD X5, P_SUM
+	ADDPD X7, SUM
+	ADDPD X9, P_SUM
+
+	LEAQ (X_PTR)(INC_X*4), X_PTR // X_PTR = &(X_PTR[incX*4])
+	LEAQ (Y_PTR)(INC_Y*4), Y_PTR // Y_PTR = &(Y_PTR[incY*4])
+
+	DECQ  LEN
+	JNZ   dot_loop   // } while --BX > 0
+	ADDPD P_SUM, SUM // sum += psum
+	CMPQ  TAIL, $0   // if TAIL == 0 { return }
+	JE    dot_end
+
+dot_tail: // do {
+	MOVDDUP_XPTR__X3      // X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_8_XPTR__X2    // X_i = { imag(x[i]), imag(x[i]) }
+	MOVUPS (Y_PTR), X10   // X_j     = {  imag(y[i])          ,  real(y[i])           }
+	MULPD  X10, X3        // X_(i+1) = {  imag(a) * real(x[i]),  real(a) * real(x[i]) }
+	SHUFPD $0x1, X10, X10 // X_j     = {  real(y[i])          ,  imag(y[i])           }
+	MULPD  X10, X2        // X_i     = {  real(a) * imag(x[i]),  imag(a) * imag(x[i]) }
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDPD X3, SUM      // sum += result[i]
+	ADDQ  INC_X, X_PTR // X_PTR += incX
+	ADDQ  INC_Y, Y_PTR // Y_PTR += incY
+	DECQ  TAIL         // --TAIL
+	JNZ   dot_tail     // }  while TAIL > 0
+
+dot_end:
+	MOVUPS SUM, sum+88(FP)
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/dotuunitary_amd64.s

@@ -0,0 +1,130 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+#define MOVDDUP_XPTR_IDX_8__X3    LONG $0x1C120FF2; BYTE $0xC6 // MOVDDUP (SI)(AX*8), X3
+#define MOVDDUP_16_XPTR_IDX_8__X5    LONG $0x6C120FF2; WORD $0x10C6 // MOVDDUP 16(SI)(AX*8), X5
+#define MOVDDUP_32_XPTR_IDX_8__X7    LONG $0x7C120FF2; WORD $0x20C6 // MOVDDUP 32(SI)(AX*8), X7
+#define MOVDDUP_48_XPTR_IDX_8__X9    LONG $0x120F44F2; WORD $0xC64C; BYTE $0x30 // MOVDDUP 48(SI)(AX*8), X9
+
+#define MOVDDUP_XPTR_IIDX_8__X2    LONG $0x14120FF2; BYTE $0xD6 // MOVDDUP (SI)(DX*8), X2
+#define MOVDDUP_16_XPTR_IIDX_8__X4    LONG $0x64120FF2; WORD $0x10D6 // MOVDDUP 16(SI)(DX*8), X4
+#define MOVDDUP_32_XPTR_IIDX_8__X6    LONG $0x74120FF2; WORD $0x20D6 // MOVDDUP 32(SI)(DX*8), X6
+#define MOVDDUP_48_XPTR_IIDX_8__X8    LONG $0x120F44F2; WORD $0xD644; BYTE $0x30 // MOVDDUP 48(SI)(DX*8), X8
+
+#define ADDSUBPD_X2_X3    LONG $0xDAD00F66 // ADDSUBPD X2, X3
+#define ADDSUBPD_X4_X5    LONG $0xECD00F66 // ADDSUBPD X4, X5
+#define ADDSUBPD_X6_X7    LONG $0xFED00F66 // ADDSUBPD X6, X7
+#define ADDSUBPD_X8_X9    LONG $0xD00F4566; BYTE $0xC8 // ADDSUBPD X8, X9
+
+#define X_PTR SI
+#define Y_PTR DI
+#define LEN CX
+#define TAIL BX
+#define SUM X0
+#define P_SUM X1
+#define IDX AX
+#define I_IDX DX
+
+// func DotuUnitary(x, y []complex128) (sum complex128)
+TEXT ·DotuUnitary(SB), NOSPLIT, $0
+	MOVQ    x_base+0(FP), X_PTR  // X_PTR = &x
+	MOVQ    y_base+24(FP), Y_PTR // Y_PTR = &y
+	MOVQ    x_len+8(FP), LEN     // LEN = min( len(x), len(y) )
+	CMPQ    y_len+32(FP), LEN
+	CMOVQLE y_len+32(FP), LEN
+	PXOR    SUM, SUM             // SUM = 0
+	CMPQ    LEN, $0              // if LEN == 0 { return }
+	JE      dot_end
+	PXOR    P_SUM, P_SUM         // P_SUM = 0
+	XORQ    IDX, IDX             // IDX = 0
+	MOVQ    $1, DX               // j = 1
+	MOVQ    LEN, TAIL
+	ANDQ    $3, TAIL             // TAIL = floor( LEN / 4 )
+	SHRQ    $2, LEN              // LEN = LEN % 4
+	JZ      dot_tail             // if LEN == 0 { goto dot_tail }
+
+dot_loop: // do {
+	MOVDDUP_XPTR_IDX_8__X3    // X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_16_XPTR_IDX_8__X5
+	MOVDDUP_32_XPTR_IDX_8__X7
+	MOVDDUP_48_XPTR_IDX_8__X9
+
+	MOVDDUP_XPTR_IIDX_8__X2    // X_i = { imag(x[i]), imag(x[i]) }
+	MOVDDUP_16_XPTR_IIDX_8__X4
+	MOVDDUP_32_XPTR_IIDX_8__X6
+	MOVDDUP_48_XPTR_IIDX_8__X8
+
+	// X_j = { imag(y[i]), real(y[i]) }
+	MOVUPS (Y_PTR)(IDX*8), X10
+	MOVUPS 16(Y_PTR)(IDX*8), X11
+	MOVUPS 32(Y_PTR)(IDX*8), X12
+	MOVUPS 48(Y_PTR)(IDX*8), X13
+
+	// X_(i+1) = { imag(a) * real(x[i]), real(a) * real(x[i])  }
+	MULPD X10, X3
+	MULPD X11, X5
+	MULPD X12, X7
+	MULPD X13, X9
+
+	// X_j = { real(y[i]), imag(y[i]) }
+	SHUFPD $0x1, X10, X10
+	SHUFPD $0x1, X11, X11
+	SHUFPD $0x1, X12, X12
+	SHUFPD $0x1, X13, X13
+
+	// X_i     = { real(a) * imag(x[i]), imag(a) * imag(x[i])  }
+	MULPD X10, X2
+	MULPD X11, X4
+	MULPD X12, X6
+	MULPD X13, X8
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDSUBPD_X4_X5
+	ADDSUBPD_X6_X7
+	ADDSUBPD_X8_X9
+
+	// psum += result[i]
+	ADDPD X3, SUM
+	ADDPD X5, P_SUM
+	ADDPD X7, SUM
+	ADDPD X9, P_SUM
+
+	ADDQ  $8, IDX    // IDX += 8
+	ADDQ  $8, I_IDX  // I_IDX += 8
+	DECQ  LEN
+	JNZ   dot_loop   // } while --LEN > 0
+	ADDPD P_SUM, SUM // SUM += P_SUM
+	CMPQ  TAIL, $0   // if TAIL == 0 { return }
+	JE    dot_end
+
+dot_tail: // do {
+	MOVDDUP_XPTR_IDX_8__X3     // X_(i+1) = { real(x[i]            , real(x[i])            }
+	MOVDDUP_XPTR_IIDX_8__X2    // X_i     = { imag(x[i])           , imag(x[i])            }
+	MOVUPS (Y_PTR)(IDX*8), X10 // X_j     = {  imag(y[i])          ,  real(y[i])           }
+	MULPD  X10, X3             // X_(i+1) = {  imag(a) * real(x[i]),  real(a) * real(x[i]) }
+	SHUFPD $0x1, X10, X10      // X_j     = {  real(y[i])          ,  imag(y[i])           }
+	MULPD  X10, X2             // X_i     = {  real(a) * imag(x[i]),  imag(a) * imag(x[i]) }
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):  real(a)*real(x[i]) - imag(a)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDPD X3, SUM   // psum += result[i]
+	ADDQ  $2, IDX   // IDX += 2
+	ADDQ  $2, I_IDX // I_IDX += 2
+	DECQ  TAIL      // --TAIL
+	JNZ   dot_tail  // }  while TAIL > 0
+
+dot_end:
+	MOVUPS SUM, sum+48(FP)
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/dscalinc_amd64.s

@@ -0,0 +1,69 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+#define SRC SI
+#define DST SI
+#define LEN CX
+#define TAIL BX
+#define INC R9
+#define INC3 R10
+#define ALPHA X0
+#define ALPHA_2 X1
+
+#define MOVDDUP_ALPHA    LONG $0x44120FF2; WORD $0x0824 // MOVDDUP 8(SP), X0
+
+// func DscalInc(alpha float64, x []complex128, n, inc uintptr)
+TEXT ·DscalInc(SB), NOSPLIT, $0
+	MOVQ x_base+8(FP), SRC // SRC = &x
+	MOVQ n+32(FP), LEN     // LEN = n
+	CMPQ LEN, $0           // if LEN == 0 { return }
+	JE   dscal_end
+
+	MOVDDUP_ALPHA             // ALPHA = alpha
+	MOVQ   inc+40(FP), INC    // INC = inc
+	SHLQ   $4, INC            // INC = INC * sizeof(complex128)
+	LEAQ   (INC)(INC*2), INC3 // INC3 = 3 * INC
+	MOVUPS ALPHA, ALPHA_2     // Copy ALPHA and ALPHA_2 for pipelining
+	MOVQ   LEN, TAIL          // TAIL = LEN
+	SHRQ   $2, LEN            // LEN = floor( n / 4 )
+	JZ     dscal_tail         // if LEN == 0 { goto dscal_tail }
+
+dscal_loop: // do {
+	MOVUPS (SRC), X2         // X_i = x[i]
+	MOVUPS (SRC)(INC*1), X3
+	MOVUPS (SRC)(INC*2), X4
+	MOVUPS (SRC)(INC3*1), X5
+
+	MULPD ALPHA, X2   // X_i *= ALPHA
+	MULPD ALPHA_2, X3
+	MULPD ALPHA, X4
+	MULPD ALPHA_2, X5
+
+	MOVUPS X2, (DST)         // x[i] = X_i
+	MOVUPS X3, (DST)(INC*1)
+	MOVUPS X4, (DST)(INC*2)
+	MOVUPS X5, (DST)(INC3*1)
+
+	LEAQ (SRC)(INC*4), SRC // SRC += INC*4
+	DECQ LEN
+	JNZ  dscal_loop        // } while --LEN > 0
+
+dscal_tail:
+	ANDQ $3, TAIL  // TAIL = TAIL % 4
+	JE   dscal_end // if TAIL == 0 { return }
+
+dscal_tail_loop: // do {
+	MOVUPS (SRC), X2       // X_i = x[i]
+	MULPD  ALPHA, X2       // X_i *= ALPHA
+	MOVUPS X2, (DST)       // x[i] = X_i
+	ADDQ   INC, SRC        // SRC += INC
+	DECQ   TAIL
+	JNZ    dscal_tail_loop // } while --TAIL > 0
+
+dscal_end:
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/dscalunitary_amd64.s

@@ -0,0 +1,66 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+#define SRC SI
+#define DST SI
+#define LEN CX
+#define IDX AX
+#define TAIL BX
+#define ALPHA X0
+#define ALPHA_2 X1
+
+#define MOVDDUP_ALPHA    LONG $0x44120FF2; WORD $0x0824 // MOVDDUP 8(SP), X0
+
+// func DscalUnitary(alpha float64, x []complex128)
+TEXT ·DscalUnitary(SB), NOSPLIT, $0
+	MOVQ x_base+8(FP), SRC // SRC = &x
+	MOVQ x_len+16(FP), LEN // LEN = len(x)
+	CMPQ LEN, $0           // if LEN == 0 { return }
+	JE   dscal_end
+
+	MOVDDUP_ALPHA         // ALPHA = alpha
+	XORQ   IDX, IDX       // IDX = 0
+	MOVUPS ALPHA, ALPHA_2 // Copy ALPHA to ALPHA_2 for pipelining
+	MOVQ   LEN, TAIL      // TAIL = LEN
+	SHRQ   $2, LEN        // LEN = floor( n / 4 )
+	JZ     dscal_tail     // if LEN == 0 { goto dscal_tail }
+
+dscal_loop: // do {
+	MOVUPS (SRC)(IDX*8), X2   // X_i = x[i]
+	MOVUPS 16(SRC)(IDX*8), X3
+	MOVUPS 32(SRC)(IDX*8), X4
+	MOVUPS 48(SRC)(IDX*8), X5
+
+	MULPD ALPHA, X2   // X_i *= ALPHA
+	MULPD ALPHA_2, X3
+	MULPD ALPHA, X4
+	MULPD ALPHA_2, X5
+
+	MOVUPS X2, (DST)(IDX*8)   // x[i] = X_i
+	MOVUPS X3, 16(DST)(IDX*8)
+	MOVUPS X4, 32(DST)(IDX*8)
+	MOVUPS X5, 48(DST)(IDX*8)
+
+	ADDQ $8, IDX    // IDX += 8
+	DECQ LEN
+	JNZ  dscal_loop // } while --LEN > 0
+
+dscal_tail:
+	ANDQ $3, TAIL  // TAIL = TAIL % 4
+	JZ   dscal_end // if TAIL == 0 { return }
+
+dscal_tail_loop: // do {
+	MOVUPS (SRC)(IDX*8), X2 // X_i = x[i]
+	MULPD  ALPHA, X2        // X_i *= ALPHA
+	MOVUPS X2, (DST)(IDX*8) // x[i] = X_i
+	ADDQ   $2, IDX          // IDX += 2
+	DECQ   TAIL
+	JNZ    dscal_tail_loop  // } while --TAIL > 0
+
+dscal_end:
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/scal.go

@@ -0,0 +1,31 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package c128
+
+// ScalUnitaryTo is
+//  for i, v := range x {
+//  	dst[i] = alpha * v
+//  }
+func ScalUnitaryTo(dst []complex128, alpha complex128, x []complex128) {
+	for i, v := range x {
+		dst[i] = alpha * v
+	}
+}
+
+// ScalIncTo is
+//  var idst, ix uintptr
+//  for i := 0; i < int(n); i++ {
+//  	dst[idst] = alpha * x[ix]
+//  	ix += incX
+//  	idst += incDst
+//  }
+func ScalIncTo(dst []complex128, incDst uintptr, alpha complex128, x []complex128, n, incX uintptr) {
+	var idst, ix uintptr
+	for i := 0; i < int(n); i++ {
+		dst[idst] = alpha * x[ix]
+		ix += incX
+		idst += incDst
+	}
+}

vendor/gonum.org/v1/gonum/internal/asm/c128/scalUnitary_amd64.s

@@ -0,0 +1,116 @@
+// Copyright ©2017 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+#define SRC SI
+#define DST SI
+#define LEN CX
+#define IDX AX
+#define TAIL BX
+#define ALPHA X0
+#define ALPHA_C X1
+#define ALPHA2 X10
+#define ALPHA_C2 X11
+
+#define MOVDDUP_X2_X3    LONG $0xDA120FF2 // MOVDDUP X2, X3
+#define MOVDDUP_X4_X5    LONG $0xEC120FF2 // MOVDDUP X4, X5
+#define MOVDDUP_X6_X7    LONG $0xFE120FF2 // MOVDDUP X6, X7
+#define MOVDDUP_X8_X9    LONG $0x120F45F2; BYTE $0xC8 // MOVDDUP X8, X9
+
+#define ADDSUBPD_X2_X3    LONG $0xDAD00F66 // ADDSUBPD X2, X3
+#define ADDSUBPD_X4_X5    LONG $0xECD00F66 // ADDSUBPD X4, X5
+#define ADDSUBPD_X6_X7    LONG $0xFED00F66 // ADDSUBPD X6, X7
+#define ADDSUBPD_X8_X9    LONG $0xD00F4566; BYTE $0xC8 // ADDSUBPD X8, X9
+
+// func ScalUnitary(alpha complex128, x []complex128)
+TEXT ·ScalUnitary(SB), NOSPLIT, $0
+	MOVQ x_base+16(FP), SRC // SRC = &x
+	MOVQ x_len+24(FP), LEN  // LEN = len(x)
+	CMPQ LEN, $0            // if LEN == 0 { return }
+	JE   scal_end
+
+	MOVUPS alpha+0(FP), ALPHA     // ALPHA = { imag(alpha), real(alpha) }
+	MOVAPS ALPHA, ALPHA_C
+	SHUFPD $0x1, ALPHA_C, ALPHA_C // ALPHA_C = { real(alpha), imag(alpha) }
+
+	XORQ   IDX, IDX          // IDX = 0
+	MOVAPS ALPHA, ALPHA2     // Copy ALPHA and ALPHA_C for pipelining
+	MOVAPS ALPHA_C, ALPHA_C2
+	MOVQ   LEN, TAIL
+	SHRQ   $2, LEN           // LEN = floor( n / 4 )
+	JZ     scal_tail         // if BX == 0 { goto scal_tail }
+
+scal_loop: // do {
+	MOVUPS (SRC)(IDX*8), X2   // X_i = { imag(x[i]), real(x[i]) }
+	MOVUPS 16(SRC)(IDX*8), X4
+	MOVUPS 32(SRC)(IDX*8), X6
+	MOVUPS 48(SRC)(IDX*8), X8
+
+	// X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_X2_X3
+	MOVDDUP_X4_X5
+	MOVDDUP_X6_X7
+	MOVDDUP_X8_X9
+
+	// X_i = { imag(x[i]), imag(x[i]) }
+	SHUFPD $0x3, X2, X2
+	SHUFPD $0x3, X4, X4
+	SHUFPD $0x3, X6, X6
+	SHUFPD $0x3, X8, X8
+
+	// X_i     = { real(ALPHA) * imag(x[i]), imag(ALPHA) * imag(x[i])  }
+	// X_(i+1) = { imag(ALPHA) * real(x[i]), real(ALPHA) * real(x[i])  }
+	MULPD ALPHA_C, X2
+	MULPD ALPHA, X3
+	MULPD ALPHA_C2, X4
+	MULPD ALPHA2, X5
+	MULPD ALPHA_C, X6
+	MULPD ALPHA, X7
+	MULPD ALPHA_C2, X8
+	MULPD ALPHA2, X9
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(ALPHA)*real(x[i]) + real(ALPHA)*imag(x[i]),
+	//	real(result[i]):  real(ALPHA)*real(x[i]) - imag(ALPHA)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDSUBPD_X4_X5
+	ADDSUBPD_X6_X7
+	ADDSUBPD_X8_X9
+
+	MOVUPS X3, (DST)(IDX*8)   // x[i] = X_(i+1)
+	MOVUPS X5, 16(DST)(IDX*8)
+	MOVUPS X7, 32(DST)(IDX*8)
+	MOVUPS X9, 48(DST)(IDX*8)
+	ADDQ   $8, IDX            // IDX += 8
+	DECQ   LEN
+	JNZ    scal_loop          // } while --LEN > 0
+
+scal_tail:
+	ANDQ $3, TAIL // TAIL = TAIL % 4
+	JZ   scal_end // if TAIL == 0 { return }
+
+scal_tail_loop: // do {
+	MOVUPS (SRC)(IDX*8), X2 // X_i = { imag(x[i]), real(x[i]) }
+	MOVDDUP_X2_X3           // X_(i+1) = { real(x[i], real(x[i]) }
+	SHUFPD $0x3, X2, X2     // X_i = { imag(x[i]), imag(x[i]) }
+	MULPD  ALPHA_C, X2      // X_i     = { real(ALPHA) * imag(x[i]), imag(ALPHA) * imag(x[i])  }
+	MULPD  ALPHA, X3        // X_(i+1) = { imag(ALPHA) * real(x[i]), real(ALPHA) * real(x[i])  }
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(ALPHA)*real(x[i]) + real(ALPHA)*imag(x[i]),
+	//	real(result[i]):  real(ALPHA)*real(x[i]) - imag(ALPHA)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+
+	MOVUPS X3, (DST)(IDX*8) // x[i] = X_(i+1)
+	ADDQ   $2, IDX          // IDX += 2
+	DECQ   TAIL
+	JNZ    scal_tail_loop   // }  while --LEN > 0
+
+scal_end:
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/scalinc_amd64.s

@@ -0,0 +1,121 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+#define SRC SI
+#define DST SI
+#define LEN CX
+#define TAIL BX
+#define INC R9
+#define INC3 R10
+#define ALPHA X0
+#define ALPHA_C X1
+#define ALPHA2 X10
+#define ALPHA_C2 X11
+
+#define MOVDDUP_X2_X3    LONG $0xDA120FF2 // MOVDDUP X2, X3
+#define MOVDDUP_X4_X5    LONG $0xEC120FF2 // MOVDDUP X4, X5
+#define MOVDDUP_X6_X7    LONG $0xFE120FF2 // MOVDDUP X6, X7
+#define MOVDDUP_X8_X9    LONG $0x120F45F2; BYTE $0xC8 // MOVDDUP X8, X9
+
+#define ADDSUBPD_X2_X3    LONG $0xDAD00F66 // ADDSUBPD X2, X3
+#define ADDSUBPD_X4_X5    LONG $0xECD00F66 // ADDSUBPD X4, X5
+#define ADDSUBPD_X6_X7    LONG $0xFED00F66 // ADDSUBPD X6, X7
+#define ADDSUBPD_X8_X9    LONG $0xD00F4566; BYTE $0xC8 // ADDSUBPD X8, X9
+
+// func ScalInc(alpha complex128, x []complex128, n, inc uintptr)
+TEXT ·ScalInc(SB), NOSPLIT, $0
+	MOVQ x_base+16(FP), SRC // SRC = &x
+	MOVQ n+40(FP), LEN      // LEN = len(x)
+	CMPQ LEN, $0
+	JE   scal_end           // if LEN == 0 { return }
+
+	MOVQ inc+48(FP), INC    // INC = inc
+	SHLQ $4, INC            // INC = INC * sizeof(complex128)
+	LEAQ (INC)(INC*2), INC3 // INC3 = 3 * INC
+
+	MOVUPS alpha+0(FP), ALPHA     // ALPHA = { imag(alpha), real(alpha) }
+	MOVAPS ALPHA, ALPHA_C
+	SHUFPD $0x1, ALPHA_C, ALPHA_C // ALPHA_C = { real(alpha), imag(alpha) }
+
+	MOVAPS ALPHA, ALPHA2     // Copy ALPHA and ALPHA_C for pipelining
+	MOVAPS ALPHA_C, ALPHA_C2
+	MOVQ   LEN, TAIL
+	SHRQ   $2, LEN           // LEN = floor( n / 4 )
+	JZ     scal_tail         // if BX == 0 { goto scal_tail }
+
+scal_loop: // do {
+	MOVUPS (SRC), X2         // X_i = { imag(x[i]), real(x[i]) }
+	MOVUPS (SRC)(INC*1), X4
+	MOVUPS (SRC)(INC*2), X6
+	MOVUPS (SRC)(INC3*1), X8
+
+	// X_(i+1) = { real(x[i], real(x[i]) }
+	MOVDDUP_X2_X3
+	MOVDDUP_X4_X5
+	MOVDDUP_X6_X7
+	MOVDDUP_X8_X9
+
+	// X_i = { imag(x[i]), imag(x[i]) }
+	SHUFPD $0x3, X2, X2
+	SHUFPD $0x3, X4, X4
+	SHUFPD $0x3, X6, X6
+	SHUFPD $0x3, X8, X8
+
+	// X_i     = { real(ALPHA) * imag(x[i]), imag(ALPHA) * imag(x[i])  }
+	// X_(i+1) = { imag(ALPHA) * real(x[i]), real(ALPHA) * real(x[i])  }
+	MULPD ALPHA_C, X2
+	MULPD ALPHA, X3
+	MULPD ALPHA_C2, X4
+	MULPD ALPHA2, X5
+	MULPD ALPHA_C, X6
+	MULPD ALPHA, X7
+	MULPD ALPHA_C2, X8
+	MULPD ALPHA2, X9
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(ALPHA)*real(x[i]) + real(ALPHA)*imag(x[i]),
+	//	real(result[i]):  real(ALPHA)*real(x[i]) - imag(ALPHA)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+	ADDSUBPD_X4_X5
+	ADDSUBPD_X6_X7
+	ADDSUBPD_X8_X9
+
+	MOVUPS X3, (DST)         // x[i] = X_(i+1)
+	MOVUPS X5, (DST)(INC*1)
+	MOVUPS X7, (DST)(INC*2)
+	MOVUPS X9, (DST)(INC3*1)
+
+	LEAQ (SRC)(INC*4), SRC // SRC = &(SRC[inc*4])
+	DECQ LEN
+	JNZ  scal_loop         // } while --BX > 0
+
+scal_tail:
+	ANDQ $3, TAIL // TAIL = TAIL % 4
+	JE   scal_end // if TAIL == 0 { return }
+
+scal_tail_loop: // do {
+	MOVUPS (SRC), X2    // X_i = { imag(x[i]), real(x[i]) }
+	MOVDDUP_X2_X3       // X_(i+1) = { real(x[i], real(x[i]) }
+	SHUFPD $0x3, X2, X2 // X_i = { imag(x[i]), imag(x[i]) }
+	MULPD  ALPHA_C, X2  // X_i     = { real(ALPHA) * imag(x[i]), imag(ALPHA) * imag(x[i])  }
+	MULPD  ALPHA, X3    // X_(i+1) = { imag(ALPHA) * real(x[i]), real(ALPHA) * real(x[i])  }
+
+	// X_(i+1) = {
+	//	imag(result[i]):  imag(ALPHA)*real(x[i]) + real(ALPHA)*imag(x[i]),
+	//	real(result[i]):  real(ALPHA)*real(x[i]) - imag(ALPHA)*imag(x[i])
+	//  }
+	ADDSUBPD_X2_X3
+
+	MOVUPS X3, (DST)      // x[i] = X_i
+	ADDQ   INC, SRC       // SRC = &(SRC[incX])
+	DECQ   TAIL
+	JNZ    scal_tail_loop // } while --TAIL > 0
+
+scal_end:
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c128/stubs_amd64.go

@@ -0,0 +1,96 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !noasm,!appengine,!safe
+
+package c128
+
+// AxpyUnitary is
+//  for i, v := range x {
+//  	y[i] += alpha * v
+//  }
+func AxpyUnitary(alpha complex128, x, y []complex128)
+
+// AxpyUnitaryTo is
+//  for i, v := range x {
+//  	dst[i] = alpha*v + y[i]
+//  }
+func AxpyUnitaryTo(dst []complex128, alpha complex128, x, y []complex128)
+
+// AxpyInc is
+//  for i := 0; i < int(n); i++ {
+//  	y[iy] += alpha * x[ix]
+//  	ix += incX
+//  	iy += incY
+//  }
+func AxpyInc(alpha complex128, x, y []complex128, n, incX, incY, ix, iy uintptr)
+
+// AxpyIncTo is
+//  for i := 0; i < int(n); i++ {
+//  	dst[idst] = alpha*x[ix] + y[iy]
+//  	ix += incX
+//  	iy += incY
+//  	idst += incDst
+//  }
+func AxpyIncTo(dst []complex128, incDst, idst uintptr, alpha complex128, x, y []complex128, n, incX, incY, ix, iy uintptr)
+
+// DscalUnitary is
+//  for i, v := range x {
+//  	x[i] = complex(real(v)*alpha, imag(v)*alpha)
+//  }
+func DscalUnitary(alpha float64, x []complex128)
+
+// DscalInc is
+//  var ix uintptr
+//  for i := 0; i < int(n); i++ {
+//  	x[ix] = complex(real(x[ix])*alpha, imag(x[ix])*alpha)
+//  	ix += inc
+//  }
+func DscalInc(alpha float64, x []complex128, n, inc uintptr)
+
+// ScalInc is
+//  var ix uintptr
+//  for i := 0; i < int(n); i++ {
+//  	x[ix] *= alpha
+//  	ix += incX
+//  }
+func ScalInc(alpha complex128, x []complex128, n, inc uintptr)
+
+// ScalUnitary is
+//  for i := range x {
+//  	x[i] *= alpha
+//  }
+func ScalUnitary(alpha complex128, x []complex128)
+
+// DotcUnitary is
+//  for i, v := range x {
+//  	sum += y[i] * cmplx.Conj(v)
+//  }
+//  return sum
+func DotcUnitary(x, y []complex128) (sum complex128)
+
+// DotcInc is
+//  for i := 0; i < int(n); i++ {
+//  	sum += y[iy] * cmplx.Conj(x[ix])
+//  	ix += incX
+//  	iy += incY
+//  }
+//  return sum
+func DotcInc(x, y []complex128, n, incX, incY, ix, iy uintptr) (sum complex128)
+
+// DotuUnitary is
+//  for i, v := range x {
+//  	sum += y[i] * v
+//  }
+//  return sum
+func DotuUnitary(x, y []complex128) (sum complex128)
+
+// DotuInc is
+//  for i := 0; i < int(n); i++ {
+//  	sum += y[iy] * x[ix]
+//  	ix += incX
+//  	iy += incY
+//  }
+//  return sum
+func DotuInc(x, y []complex128, n, incX, incY, ix, iy uintptr) (sum complex128)

vendor/gonum.org/v1/gonum/internal/asm/c128/stubs_noasm.go

@@ -0,0 +1,163 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !amd64 noasm appengine safe
+
+package c128
+
+import "math/cmplx"
+
+// AxpyUnitary is
+//  for i, v := range x {
+//  	y[i] += alpha * v
+//  }
+func AxpyUnitary(alpha complex128, x, y []complex128) {
+	for i, v := range x {
+		y[i] += alpha * v
+	}
+}
+
+// AxpyUnitaryTo is
+//  for i, v := range x {
+//  	dst[i] = alpha*v + y[i]
+//  }
+func AxpyUnitaryTo(dst []complex128, alpha complex128, x, y []complex128) {
+	for i, v := range x {
+		dst[i] = alpha*v + y[i]
+	}
+}
+
+// AxpyInc is
+//  for i := 0; i < int(n); i++ {
+//  	y[iy] += alpha * x[ix]
+//  	ix += incX
+//  	iy += incY
+//  }
+func AxpyInc(alpha complex128, x, y []complex128, n, incX, incY, ix, iy uintptr) {
+	for i := 0; i < int(n); i++ {
+		y[iy] += alpha * x[ix]
+		ix += incX
+		iy += incY
+	}
+}
+
+// AxpyIncTo is
+//  for i := 0; i < int(n); i++ {
+//  	dst[idst] = alpha*x[ix] + y[iy]
+//  	ix += incX
+//  	iy += incY
+//  	idst += incDst
+//  }
+func AxpyIncTo(dst []complex128, incDst, idst uintptr, alpha complex128, x, y []complex128, n, incX, incY, ix, iy uintptr) {
+	for i := 0; i < int(n); i++ {
+		dst[idst] = alpha*x[ix] + y[iy]
+		ix += incX
+		iy += incY
+		idst += incDst
+	}
+}
+
+// DscalUnitary is
+//  for i, v := range x {
+//  	x[i] = complex(real(v)*alpha, imag(v)*alpha)
+//  }
+func DscalUnitary(alpha float64, x []complex128) {
+	for i, v := range x {
+		x[i] = complex(real(v)*alpha, imag(v)*alpha)
+	}
+}
+
+// DscalInc is
+//  var ix uintptr
+//  for i := 0; i < int(n); i++ {
+//  	x[ix] = complex(real(x[ix])*alpha, imag(x[ix])*alpha)
+//  	ix += inc
+//  }
+func DscalInc(alpha float64, x []complex128, n, inc uintptr) {
+	var ix uintptr
+	for i := 0; i < int(n); i++ {
+		x[ix] = complex(real(x[ix])*alpha, imag(x[ix])*alpha)
+		ix += inc
+	}
+}
+
+// ScalInc is
+//  var ix uintptr
+//  for i := 0; i < int(n); i++ {
+//  	x[ix] *= alpha
+//  	ix += incX
+//  }
+func ScalInc(alpha complex128, x []complex128, n, inc uintptr) {
+	var ix uintptr
+	for i := 0; i < int(n); i++ {
+		x[ix] *= alpha
+		ix += inc
+	}
+}
+
+// ScalUnitary is
+//  for i := range x {
+//  	x[i] *= alpha
+//  }
+func ScalUnitary(alpha complex128, x []complex128) {
+	for i := range x {
+		x[i] *= alpha
+	}
+}
+
+// DotcUnitary is
+//  for i, v := range x {
+//  	sum += y[i] * cmplx.Conj(v)
+//  }
+//  return sum
+func DotcUnitary(x, y []complex128) (sum complex128) {
+	for i, v := range x {
+		sum += y[i] * cmplx.Conj(v)
+	}
+	return sum
+}
+
+// DotcInc is
+//  for i := 0; i < int(n); i++ {
+//  	sum += y[iy] * cmplx.Conj(x[ix])
+//  	ix += incX
+//  	iy += incY
+//  }
+//  return sum
+func DotcInc(x, y []complex128, n, incX, incY, ix, iy uintptr) (sum complex128) {
+	for i := 0; i < int(n); i++ {
+		sum += y[iy] * cmplx.Conj(x[ix])
+		ix += incX
+		iy += incY
+	}
+	return sum
+}
+
+// DotuUnitary is
+//  for i, v := range x {
+//  	sum += y[i] * v
+//  }
+//  return sum
+func DotuUnitary(x, y []complex128) (sum complex128) {
+	for i, v := range x {
+		sum += y[i] * v
+	}
+	return sum
+}
+
+// DotuInc is
+//  for i := 0; i < int(n); i++ {
+//  	sum += y[iy] * x[ix]
+//  	ix += incX
+//  	iy += incY
+//  }
+//  return sum
+func DotuInc(x, y []complex128, n, incX, incY, ix, iy uintptr) (sum complex128) {
+	for i := 0; i < int(n); i++ {
+		sum += y[iy] * x[ix]
+		ix += incX
+		iy += incY
+	}
+	return sum
+}

vendor/gonum.org/v1/gonum/internal/asm/c64/axpyinc_amd64.s

@@ -0,0 +1,151 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+// MOVSHDUP X3, X2
+#define MOVSHDUP_X3_X2 BYTE $0xF3; BYTE $0x0F; BYTE $0x16; BYTE $0xD3
+// MOVSLDUP X3, X3
+#define MOVSLDUP_X3_X3 BYTE $0xF3; BYTE $0x0F; BYTE $0x12; BYTE $0xDB
+// ADDSUBPS X2, X3
+#define ADDSUBPS_X2_X3 BYTE $0xF2; BYTE $0x0F; BYTE $0xD0; BYTE $0xDA
+
+// MOVSHDUP X5, X4
+#define MOVSHDUP_X5_X4 BYTE $0xF3; BYTE $0x0F; BYTE $0x16; BYTE $0xE5
+// MOVSLDUP X5, X5
+#define MOVSLDUP_X5_X5 BYTE $0xF3; BYTE $0x0F; BYTE $0x12; BYTE $0xED
+// ADDSUBPS X4, X5
+#define ADDSUBPS_X4_X5 BYTE $0xF2; BYTE $0x0F; BYTE $0xD0; BYTE $0xEC
+
+// MOVSHDUP X7, X6
+#define MOVSHDUP_X7_X6 BYTE $0xF3; BYTE $0x0F; BYTE $0x16; BYTE $0xF7
+// MOVSLDUP X7, X7
+#define MOVSLDUP_X7_X7 BYTE $0xF3; BYTE $0x0F; BYTE $0x12; BYTE $0xFF
+// ADDSUBPS X6, X7
+#define ADDSUBPS_X6_X7 BYTE $0xF2; BYTE $0x0F; BYTE $0xD0; BYTE $0xFE
+
+// MOVSHDUP X9, X8
+#define MOVSHDUP_X9_X8 BYTE $0xF3; BYTE $0x45; BYTE $0x0F; BYTE $0x16; BYTE $0xC1
+// MOVSLDUP X9, X9
+#define MOVSLDUP_X9_X9 BYTE $0xF3; BYTE $0x45; BYTE $0x0F; BYTE $0x12; BYTE $0xC9
+// ADDSUBPS X8, X9
+#define ADDSUBPS_X8_X9 BYTE $0xF2; BYTE $0x45; BYTE $0x0F; BYTE $0xD0; BYTE $0xC8
+
+// func AxpyInc(alpha complex64, x, y []complex64, n, incX, incY, ix, iy uintptr)
+TEXT ·AxpyInc(SB), NOSPLIT, $0
+	MOVQ   x_base+8(FP), SI  // SI = &x
+	MOVQ   y_base+32(FP), DI // DI = &y
+	MOVQ   n+56(FP), CX      // CX = n
+	CMPQ   CX, $0            // if n==0 { return }
+	JE     axpyi_end
+	MOVQ   ix+80(FP), R8     // R8 = ix
+	MOVQ   iy+88(FP), R9     // R9 = iy
+	LEAQ   (SI)(R8*8), SI    // SI = &(x[ix])
+	LEAQ   (DI)(R9*8), DI    // DI = &(y[iy])
+	MOVQ   DI, DX            // DX = DI    // Read/Write pointers
+	MOVQ   incX+64(FP), R8   // R8 = incX
+	SHLQ   $3, R8            // R8 *= sizeof(complex64)
+	MOVQ   incY+72(FP), R9   // R9 = incY
+	SHLQ   $3, R9            // R9 *= sizeof(complex64)
+	MOVSD  alpha+0(FP), X0   // X0 = { 0, 0, imag(a), real(a) }
+	MOVAPS X0, X1
+	SHUFPS $0x11, X1, X1     // X1 = { 0, 0, real(a), imag(a) }
+	MOVAPS X0, X10           // Copy X0 and X1 for pipelining
+	MOVAPS X1, X11
+	MOVQ   CX, BX
+	ANDQ   $3, CX            // CX = n % 4
+	SHRQ   $2, BX            // BX = floor( n / 4 )
+	JZ     axpyi_tail        // if BX == 0 { goto axpyi_tail }
+
+axpyi_loop: // do {
+	MOVSD (SI), X3       // X_i = { imag(x[i+1]), real(x[i+1]) }
+	MOVSD (SI)(R8*1), X5
+	LEAQ  (SI)(R8*2), SI // SI = &(SI[incX*2])
+	MOVSD (SI), X7
+	MOVSD (SI)(R8*1), X9
+
+	// X_(i-1) = { imag(x[i]), imag(x[i]) }
+	MOVSHDUP_X3_X2
+	MOVSHDUP_X5_X4
+	MOVSHDUP_X7_X6
+	MOVSHDUP_X9_X8
+
+	// X_i = { real(x[i]), real(x[i]) }
+	MOVSLDUP_X3_X3
+	MOVSLDUP_X5_X5
+	MOVSLDUP_X7_X7
+	MOVSLDUP_X9_X9
+
+	// X_(i-1) = {  real(a) * imag(x[i]),   imag(a) * imag(x[i]) }
+	// X_i     = {  imag(a) * real(x[i]),   real(a) * real(x[i])  }
+	MULPS X1, X2
+	MULPS X0, X3
+	MULPS X11, X4
+	MULPS X10, X5
+	MULPS X1, X6
+	MULPS X0, X7
+	MULPS X11, X8
+	MULPS X10, X9
+
+	// X_i = {
+	//	imag(result[i]):   imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):   real(a)*real(x[i]) - imag(a)*imag(x[i]),
+	//  }
+	ADDSUBPS_X2_X3
+	ADDSUBPS_X4_X5
+	ADDSUBPS_X6_X7
+	ADDSUBPS_X8_X9
+
+	// X_i = { imag(result[i]) + imag(y[i]), real(result[i]) + real(y[i]) }
+	MOVSD (DX), X2
+	MOVSD (DX)(R9*1), X4
+	LEAQ  (DX)(R9*2), DX // DX = &(DX[incY*2])
+	MOVSD (DX), X6
+	MOVSD (DX)(R9*1), X8
+	ADDPS X2, X3
+	ADDPS X4, X5
+	ADDPS X6, X7
+	ADDPS X8, X9
+
+	MOVSD X3, (DI)       // y[i] = X_i
+	MOVSD X5, (DI)(R9*1)
+	LEAQ  (DI)(R9*2), DI // DI = &(DI[incDst])
+	MOVSD X7, (DI)
+	MOVSD X9, (DI)(R9*1)
+	LEAQ  (SI)(R8*2), SI // SI = &(SI[incX*2])
+	LEAQ  (DX)(R9*2), DX // DX = &(DX[incY*2])
+	LEAQ  (DI)(R9*2), DI // DI = &(DI[incDst])
+	DECQ  BX
+	JNZ   axpyi_loop     // }  while --BX > 0
+	CMPQ  CX, $0         // if CX == 0 { return }
+	JE    axpyi_end
+
+axpyi_tail: // do {
+	MOVSD (SI), X3 // X_i = { imag(x[i+1]), real(x[i+1]) }
+	MOVSHDUP_X3_X2 // X_(i-1) = { real(x[i]), real(x[i]) }
+	MOVSLDUP_X3_X3 // X_i = { imag(x[i]), imag(x[i]) }
+
+	// X_i     = { imag(a) * real(x[i]),  real(a) * real(x[i]) }
+	// X_(i-1) = { real(a) * imag(x[i]),  imag(a) * imag(x[i]) }
+	MULPS X1, X2
+	MULPS X0, X3
+
+	// X_i = {
+	//	imag(result[i]):   imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):   real(a)*real(x[i]) - imag(a)*imag(x[i]),
+	//  }
+	ADDSUBPS_X2_X3 // (ai*x1r+ar*x1i, ar*x1r-ai*x1i)
+
+	// X_i = { imag(result[i]) + imag(y[i]),  real(result[i]) + real(y[i])  }
+	MOVSD (DI), X4
+	ADDPS X4, X3
+	MOVSD X3, (DI)   // y[i] = X_i
+	ADDQ  R8, SI     // SI += incX
+	ADDQ  R9, DI     // DI += incY
+	LOOP  axpyi_tail // } while --CX > 0
+
+axpyi_end:
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c64/axpyincto_amd64.s

@@ -0,0 +1,156 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+// MOVSHDUP X3, X2
+#define MOVSHDUP_X3_X2 BYTE $0xF3; BYTE $0x0F; BYTE $0x16; BYTE $0xD3
+// MOVSLDUP X3, X3
+#define MOVSLDUP_X3_X3 BYTE $0xF3; BYTE $0x0F; BYTE $0x12; BYTE $0xDB
+// ADDSUBPS X2, X3
+#define ADDSUBPS_X2_X3 BYTE $0xF2; BYTE $0x0F; BYTE $0xD0; BYTE $0xDA
+
+// MOVSHDUP X5, X4
+#define MOVSHDUP_X5_X4 BYTE $0xF3; BYTE $0x0F; BYTE $0x16; BYTE $0xE5
+// MOVSLDUP X5, X5
+#define MOVSLDUP_X5_X5 BYTE $0xF3; BYTE $0x0F; BYTE $0x12; BYTE $0xED
+// ADDSUBPS X4, X5
+#define ADDSUBPS_X4_X5 BYTE $0xF2; BYTE $0x0F; BYTE $0xD0; BYTE $0xEC
+
+// MOVSHDUP X7, X6
+#define MOVSHDUP_X7_X6 BYTE $0xF3; BYTE $0x0F; BYTE $0x16; BYTE $0xF7
+// MOVSLDUP X7, X7
+#define MOVSLDUP_X7_X7 BYTE $0xF3; BYTE $0x0F; BYTE $0x12; BYTE $0xFF
+// ADDSUBPS X6, X7
+#define ADDSUBPS_X6_X7 BYTE $0xF2; BYTE $0x0F; BYTE $0xD0; BYTE $0xFE
+
+// MOVSHDUP X9, X8
+#define MOVSHDUP_X9_X8 BYTE $0xF3; BYTE $0x45; BYTE $0x0F; BYTE $0x16; BYTE $0xC1
+// MOVSLDUP X9, X9
+#define MOVSLDUP_X9_X9 BYTE $0xF3; BYTE $0x45; BYTE $0x0F; BYTE $0x12; BYTE $0xC9
+// ADDSUBPS X8, X9
+#define ADDSUBPS_X8_X9 BYTE $0xF2; BYTE $0x45; BYTE $0x0F; BYTE $0xD0; BYTE $0xC8
+
+// func AxpyIncTo(dst []complex64, incDst, idst uintptr, alpha complex64, x, y []complex64, n, incX, incY, ix, iy uintptr)
+TEXT ·AxpyIncTo(SB), NOSPLIT, $0
+	MOVQ   dst_base+0(FP), DI // DI = &dst
+	MOVQ   x_base+48(FP), SI  // SI = &x
+	MOVQ   y_base+72(FP), DX  // DX = &y
+	MOVQ   n+96(FP), CX       // CX = n
+	CMPQ   CX, $0             // if n==0 { return }
+	JE     axpyi_end
+	MOVQ   ix+120(FP), R8     // Load the first index
+	MOVQ   iy+128(FP), R9
+	MOVQ   idst+32(FP), R10
+	LEAQ   (SI)(R8*8), SI     // SI = &(x[ix])
+	LEAQ   (DX)(R9*8), DX     // DX = &(y[iy])
+	LEAQ   (DI)(R10*8), DI    // DI = &(dst[idst])
+	MOVQ   incX+104(FP), R8   // Incrementors*8 for easy iteration (ADDQ)
+	SHLQ   $3, R8
+	MOVQ   incY+112(FP), R9
+	SHLQ   $3, R9
+	MOVQ   incDst+24(FP), R10
+	SHLQ   $3, R10
+	MOVSD  alpha+40(FP), X0   // X0 = { 0, 0, imag(a), real(a) }
+	MOVAPS X0, X1
+	SHUFPS $0x11, X1, X1      // X1 = { 0, 0, real(a), imag(a) }
+	MOVAPS X0, X10            // Copy X0 and X1 for pipelining
+	MOVAPS X1, X11
+	MOVQ   CX, BX
+	ANDQ   $3, CX             // CX = n % 4
+	SHRQ   $2, BX             // BX = floor( n / 4 )
+	JZ     axpyi_tail         // if BX == 0 { goto axpyi_tail }
+
+axpyi_loop: // do {
+	MOVSD (SI), X3       // X_i = { imag(x[i]), real(x[i]) }
+	MOVSD (SI)(R8*1), X5
+	LEAQ  (SI)(R8*2), SI // SI = &(SI[incX*2])
+	MOVSD (SI), X7
+	MOVSD (SI)(R8*1), X9
+
+	// X_(i-1) = { imag(x[i]), imag(x[i]) }
+	MOVSHDUP_X3_X2
+	MOVSHDUP_X5_X4
+	MOVSHDUP_X7_X6
+	MOVSHDUP_X9_X8
+
+	// X_i = { real(x[i]), real(x[i]) }
+	MOVSLDUP_X3_X3
+	MOVSLDUP_X5_X5
+	MOVSLDUP_X7_X7
+	MOVSLDUP_X9_X9
+
+	// X_(i-1) = {  real(a) * imag(x[i]),   imag(a) * imag(x[i]) }
+	// X_i     = {  imag(a) * real(x[i]),   real(a) * real(x[i])  }
+	MULPS X1, X2
+	MULPS X0, X3
+	MULPS X11, X4
+	MULPS X10, X5
+	MULPS X1, X6
+	MULPS X0, X7
+	MULPS X11, X8
+	MULPS X10, X9
+
+	// X_i = {
+	//	imag(result[i]):   imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):   real(a)*real(x[i]) - imag(a)*imag(x[i]),
+	//  }
+	ADDSUBPS_X2_X3
+	ADDSUBPS_X4_X5
+	ADDSUBPS_X6_X7
+	ADDSUBPS_X8_X9
+
+	// X_i = { imag(result[i]) + imag(y[i]), real(result[i]) + real(y[i]) }
+	MOVSD (DX), X2
+	MOVSD (DX)(R9*1), X4
+	LEAQ  (DX)(R9*2), DX // DX = &(DX[incY*2])
+	MOVSD (DX), X6
+	MOVSD (DX)(R9*1), X8
+	ADDPS X2, X3
+	ADDPS X4, X5
+	ADDPS X6, X7
+	ADDPS X8, X9
+
+	MOVSD X3, (DI)        // y[i] = X_i
+	MOVSD X5, (DI)(R10*1)
+	LEAQ  (DI)(R10*2), DI // DI = &(DI[incDst])
+	MOVSD X7, (DI)
+	MOVSD X9, (DI)(R10*1)
+	LEAQ  (SI)(R8*2), SI  // SI = &(SI[incX*2])
+	LEAQ  (DX)(R9*2), DX  // DX = &(DX[incY*2])
+	LEAQ  (DI)(R10*2), DI // DI = &(DI[incDst])
+	DECQ  BX
+	JNZ   axpyi_loop      // } while --BX > 0
+	CMPQ  CX, $0          // if CX == 0 { return }
+	JE    axpyi_end
+
+axpyi_tail:
+	MOVSD (SI), X3 // X_i     = { imag(x[i]), real(x[i]) }
+	MOVSHDUP_X3_X2 // X_(i-1) = { imag(x[i]), imag(x[i]) }
+	MOVSLDUP_X3_X3 // X_i     = { real(x[i]), real(x[i]) }
+
+	// X_i     = { imag(a) * real(x[i]),  real(a) * real(x[i]) }
+	// X_(i-1) = { real(a) * imag(x[i]),  imag(a) * imag(x[i]) }
+	MULPS X1, X2
+	MULPS X0, X3
+
+	// X_i = {
+	//	imag(result[i]):   imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):   real(a)*real(x[i]) - imag(a)*imag(x[i]),
+	//  }
+	ADDSUBPS_X2_X3
+
+	// X_i = { imag(result[i]) + imag(y[i]),  real(result[i]) + real(y[i])  }
+	MOVSD (DX), X4
+	ADDPS X4, X3
+	MOVSD X3, (DI)   // y[i] = X_i
+	ADDQ  R8, SI     // SI += incX
+	ADDQ  R9, DX     // DX += incY
+	ADDQ  R10, DI    // DI += incDst
+	LOOP  axpyi_tail // } while --CX > 0
+
+axpyi_end:
+	RET

vendor/gonum.org/v1/gonum/internal/asm/c64/axpyunitary_amd64.s

@@ -0,0 +1,160 @@
+// Copyright ©2016 The Gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+//+build !noasm,!appengine,!safe
+
+#include "textflag.h"
+
+// MOVSHDUP X3, X2
+#define MOVSHDUP_X3_X2 BYTE $0xF3; BYTE $0x0F; BYTE $0x16; BYTE $0xD3
+// MOVSLDUP X3, X3
+#define MOVSLDUP_X3_X3 BYTE $0xF3; BYTE $0x0F; BYTE $0x12; BYTE $0xDB
+// ADDSUBPS X2, X3
+#define ADDSUBPS_X2_X3 BYTE $0xF2; BYTE $0x0F; BYTE $0xD0; BYTE $0xDA
+
+// MOVSHDUP X5, X4
+#define MOVSHDUP_X5_X4 BYTE $0xF3; BYTE $0x0F; BYTE $0x16; BYTE $0xE5
+// MOVSLDUP X5, X5
+#define MOVSLDUP_X5_X5 BYTE $0xF3; BYTE $0x0F; BYTE $0x12; BYTE $0xED
+// ADDSUBPS X4, X5
+#define ADDSUBPS_X4_X5 BYTE $0xF2; BYTE $0x0F; BYTE $0xD0; BYTE $0xEC
+
+// MOVSHDUP X7, X6
+#define MOVSHDUP_X7_X6 BYTE $0xF3; BYTE $0x0F; BYTE $0x16; BYTE $0xF7
+// MOVSLDUP X7, X7
+#define MOVSLDUP_X7_X7 BYTE $0xF3; BYTE $0x0F; BYTE $0x12; BYTE $0xFF
+// ADDSUBPS X6, X7
+#define ADDSUBPS_X6_X7 BYTE $0xF2; BYTE $0x0F; BYTE $0xD0; BYTE $0xFE
+
+// MOVSHDUP X9, X8
+#define MOVSHDUP_X9_X8 BYTE $0xF3; BYTE $0x45; BYTE $0x0F; BYTE $0x16; BYTE $0xC1
+// MOVSLDUP X9, X9
+#define MOVSLDUP_X9_X9 BYTE $0xF3; BYTE $0x45; BYTE $0x0F; BYTE $0x12; BYTE $0xC9
+// ADDSUBPS X8, X9
+#define ADDSUBPS_X8_X9 BYTE $0xF2; BYTE $0x45; BYTE $0x0F; BYTE $0xD0; BYTE $0xC8
+
+// func AxpyUnitary(alpha complex64, x, y []complex64)
+TEXT ·AxpyUnitary(SB), NOSPLIT, $0
+	MOVQ    x_base+8(FP), SI  // SI = &x
+	MOVQ    y_base+32(FP), DI // DI = &y
+	MOVQ    x_len+16(FP), CX  // CX = min( len(x), len(y) )
+	CMPQ    y_len+40(FP), CX
+	CMOVQLE y_len+40(FP), CX
+	CMPQ    CX, $0            // if CX == 0 { return }
+	JE      caxy_end
+	PXOR    X0, X0            // Clear work registers and cache-align loop
+	PXOR    X1, X1
+	MOVSD   alpha+0(FP), X0   // X0 = { 0, 0, imag(a), real(a) }
+	SHUFPD  $0, X0, X0        // X0  = { imag(a), real(a), imag(a), real(a) }
+	MOVAPS  X0, X1
+	SHUFPS  $0x11, X1, X1     // X1 = { real(a), imag(a), real(a), imag(a) }
+	XORQ    AX, AX            // i = 0
+	MOVQ    DI, BX            // Align on 16-byte boundary for ADDPS
+	ANDQ    $15, BX           // BX = &y & 15
+	JZ      caxy_no_trim      // if BX == 0 { goto caxy_no_trim }
+
+	// Trim first value in unaligned buffer
+	XORPS X2, X2         // Clear work registers and cache-align loop
+	XORPS X3, X3
+	XORPS X4, X4
+	MOVSD (SI)(AX*8), X3 // X3 = { imag(x[i]), real(x[i]) }
+	MOVSHDUP_X3_X2       // X2 = { imag(x[i]), imag(x[i]) }
+	MOVSLDUP_X3_X3       // X3 = { real(x[i]), real(x[i]) }
+	MULPS X1, X2         // X2 = { real(a) * imag(x[i]), imag(a) * imag(x[i]) }
+	MULPS X0, X3         // X3 = { imag(a) * real(x[i]), real(a) * real(x[i]) }
+
+	// X3 = { imag(a)*real(x[i]) + real(a)*imag(x[i]), real(a)*real(x[i]) - imag(a)*imag(x[i]) }
+	ADDSUBPS_X2_X3
+	MOVSD (DI)(AX*8), X4 // X3 += y[i]
+	ADDPS X4, X3
+	MOVSD X3, (DI)(AX*8) // y[i]  = X3
+	INCQ  AX             // i++
+	DECQ  CX             // --CX
+	JZ    caxy_end       // if CX == 0 { return }
+
+caxy_no_trim:
+	MOVAPS X0, X10   // Copy X0 and X1 for pipelineing
+	MOVAPS X1, X11
+	MOVQ   CX, BX
+	ANDQ   $7, CX    // CX = n % 8
+	SHRQ   $3, BX    // BX = floor( n / 8 )
+	JZ     caxy_tail // if BX == 0 { goto caxy_tail }
+
+caxy_loop: // do {
+	// X_i = { imag(x[i]), real(x[i]), imag(x[i+1]), real(x[i+1]) }
+	MOVUPS (SI)(AX*8), X3
+	MOVUPS 16(SI)(AX*8), X5
+	MOVUPS 32(SI)(AX*8), X7
+	MOVUPS 48(SI)(AX*8), X9
+
+	// X_(i-1) = { imag(x[i]), imag(x[i]), imag(x[i]+1), imag(x[i]+1) }
+	MOVSHDUP_X3_X2
+	MOVSHDUP_X5_X4
+	MOVSHDUP_X7_X6
+	MOVSHDUP_X9_X8
+
+	// X_i = { real(x[i]), real(x[i]), real(x[i+1]), real(x[i+1]) }
+	MOVSLDUP_X3_X3
+	MOVSLDUP_X5_X5
+	MOVSLDUP_X7_X7
+	MOVSLDUP_X9_X9
+
+	// X_i     = {  imag(a) * real(x[i]),   real(a) * real(x[i]),
+	// 		imag(a) * real(x[i+1]), real(a) * real(x[i+1])  }
+	// X_(i-1) = {  real(a) * imag(x[i]),   imag(a) * imag(x[i]),
+	//		real(a) * imag(x[i+1]), imag(a) * imag(x[i+1])  }
+	MULPS X1, X2
+	MULPS X0, X3
+	MULPS X11, X4
+	MULPS X10, X5
+	MULPS X1, X6
+	MULPS X0, X7
+	MULPS X11, X8
+	MULPS X10, X9
+
+	// X_i = {
+	//	imag(result[i]):   imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	real(result[i]):   real(a)*real(x[i]) - imag(a)*imag(x[i]),
+	//	imag(result[i+1]): imag(a)*real(x[i+1]) + real(a)*imag(x[i+1]),
+	//	real(result[i+1]): real(a)*real(x[i+1]) - imag(a)*imag(x[i+1]),
+	//  }
+	ADDSUBPS_X2_X3
+	ADDSUBPS_X4_X5
+	ADDSUBPS_X6_X7
+	ADDSUBPS_X8_X9
+
+	// X_i = { imag(result[i])   + imag(y[i]),   real(result[i])   + real(y[i]),
+	//	   imag(result[i+1]) + imag(y[i+1]), real(result[i+1]) + real(y[i+1])  }
+	ADDPS  (DI)(AX*8), X3
+	ADDPS  16(DI)(AX*8), X5
+	ADDPS  32(DI)(AX*8), X7
+	ADDPS  48(DI)(AX*8), X9
+	MOVUPS X3, (DI)(AX*8)   // y[i:i+1] = X_i
+	MOVUPS X5, 16(DI)(AX*8)
+	MOVUPS X7, 32(DI)(AX*8)
+	MOVUPS X9, 48(DI)(AX*8)
+	ADDQ   $8, AX           // i += 8
+	DECQ   BX               // --BX
+	JNZ    caxy_loop        // }  while BX > 0
+	CMPQ   CX, $0           // if CX == 0  { return }
+	JE     caxy_end
+
+caxy_tail: // do {
+	MOVSD (SI)(AX*8), X3 // X3 = { imag(x[i]), real(x[i]) }
+	MOVSHDUP_X3_X2       // X2 = { imag(x[i]), imag(x[i]) }
+	MOVSLDUP_X3_X3       // X3 = { real(x[i]), real(x[i]) }
+	MULPS X1, X2         // X2 = { real(a) * imag(x[i]), imag(a) * imag(x[i]) }
+	MULPS X0, X3         // X3 = { imag(a) * real(x[i]), real(a) * real(x[i]) }
+
+	// X3 = { imag(a)*real(x[i]) + real(a)*imag(x[i]),
+	//	  real(a)*real(x[i]) - imag(a)*imag(x[i])   }
+	ADDSUBPS_X2_X3
+	MOVSD (DI)(AX*8), X4 // X3 += y[i]
+	ADDPS X4, X3
+	MOVSD X3, (DI)(AX*8) // y[i]  = X3
+	INCQ  AX             // ++i
+	LOOP  caxy_tail      // } while --CX > 0
+
+caxy_end:
+	RET