gonum.org/v1/gonum@v0.14.0/lapack/gonum/dgeqrf.go (about) 1 // Copyright ©2015 The Gonum Authors. All rights reserved. 2 // Use of this source code is governed by a BSD-style 3 // license that can be found in the LICENSE file. 4 5 package gonum 6 7 import ( 8 "gonum.org/v1/gonum/blas" 9 "gonum.org/v1/gonum/lapack" 10 ) 11 12 // Dgeqrf computes the QR factorization of the m×n matrix A using a blocked 13 // algorithm. See the documentation for Dgeqr2 for a description of the 14 // parameters at entry and exit. 15 // 16 // work is temporary storage, and lwork specifies the usable memory length. 17 // The length of work must be at least max(1, lwork) and lwork must be -1 18 // or at least n, otherwise this function will panic. 19 // Dgeqrf is a blocked QR factorization, but the block size is limited 20 // by the temporary space available. If lwork == -1, instead of performing Dgeqrf, 21 // the optimal work length will be stored into work[0]. 22 // 23 // tau must have length at least min(m,n), and this function will panic otherwise. 24 func (impl Implementation) Dgeqrf(m, n int, a []float64, lda int, tau, work []float64, lwork int) { 25 switch { 26 case m < 0: 27 panic(mLT0) 28 case n < 0: 29 panic(nLT0) 30 case lda < max(1, n): 31 panic(badLdA) 32 case lwork < max(1, n) && lwork != -1: 33 panic(badLWork) 34 case len(work) < max(1, lwork): 35 panic(shortWork) 36 } 37 38 // Quick return if possible. 39 k := min(m, n) 40 if k == 0 { 41 work[0] = 1 42 return 43 } 44 45 // nb is the optimal blocksize, i.e. the number of columns transformed at a time. 46 nb := impl.Ilaenv(1, "DGEQRF", " ", m, n, -1, -1) 47 if lwork == -1 { 48 work[0] = float64(n * nb) 49 return 50 } 51 52 if len(a) < (m-1)*lda+n { 53 panic(shortA) 54 } 55 if len(tau) < k { 56 panic(shortTau) 57 } 58 59 nbmin := 2 // Minimal block size. 60 var nx int // Use unblocked (unless changed in the next for loop) 61 iws := n 62 // Only consider blocked if the suggested block size is > 1 and the 63 // number of rows or columns is sufficiently large. 64 if 1 < nb && nb < k { 65 // nx is the block size at which the code switches from blocked 66 // to unblocked. 67 nx = max(0, impl.Ilaenv(3, "DGEQRF", " ", m, n, -1, -1)) 68 if k > nx { 69 iws = n * nb 70 if lwork < iws { 71 // Not enough workspace to use the optimal block 72 // size. Get the minimum block size instead. 73 nb = lwork / n 74 nbmin = max(2, impl.Ilaenv(2, "DGEQRF", " ", m, n, -1, -1)) 75 } 76 } 77 } 78 79 // Compute QR using a blocked algorithm. 80 var i int 81 if nbmin <= nb && nb < k && nx < k { 82 ldwork := nb 83 for i = 0; i < k-nx; i += nb { 84 ib := min(k-i, nb) 85 // Compute the QR factorization of the current block. 86 impl.Dgeqr2(m-i, ib, a[i*lda+i:], lda, tau[i:], work) 87 if i+ib < n { 88 // Form the triangular factor of the block reflector and apply Hᵀ 89 // In Dlarft, work becomes the T matrix. 90 impl.Dlarft(lapack.Forward, lapack.ColumnWise, m-i, ib, 91 a[i*lda+i:], lda, 92 tau[i:], 93 work, ldwork) 94 impl.Dlarfb(blas.Left, blas.Trans, lapack.Forward, lapack.ColumnWise, 95 m-i, n-i-ib, ib, 96 a[i*lda+i:], lda, 97 work, ldwork, 98 a[i*lda+i+ib:], lda, 99 work[ib*ldwork:], ldwork) 100 } 101 } 102 } 103 // Call unblocked code on the remaining columns. 104 if i < k { 105 impl.Dgeqr2(m-i, n-i, a[i*lda+i:], lda, tau[i:], work) 106 } 107 work[0] = float64(iws) 108 }