WIP: Upgrade to OpenBLAS 0.3.13 (JuliaLang#39216)

* Use OpenBLAS 0.3.13 Bumping to support xcode/clang 12 which was addressed in OpenBLAS 0.3.11 * Use OpenBLAS 0.3.13+1 * Add openblas-exshift patch for src build * Update LinearAlgebra doctests for Linux * non-ambiguous ordering in eigen and eigvals test (JuliaLang#39767) add missing sortby's Co-authored-by: Pablo San-Jose <[email protected]>
Liozou · Feb 20, 2021 · 3129a5b · 3129a5b
1 parent d524f21
commit 3129a5b
Show file tree

Hide file tree

Showing 11 changed files with 302 additions and 149 deletions.
diff --git a/deps/Versions.make b/deps/Versions.make
@@ -75,7 +75,7 @@ OBJCONV_JLL_NAME := Objconv
 OBJCONV_JLL_VER := 2.49.1+0
 
 # OpenBLAS
-OPENBLAS_VER := 0.3.10
+OPENBLAS_VER := 0.3.13
 OPENBLAS_JLL_NAME := OpenBLAS
 
 # OpenLibm

diff --git a/deps/blas.mk b/deps/blas.mk
@@ -103,7 +103,12 @@ $(BUILDDIR)/$(OPENBLAS_SRC_DIR)/openblas-ofast-power.patch-applied: $(BUILDDIR)/
  patch -p1 -f < $(SRCDIR)/patches/openblas-ofast-power.patch
  echo 1 > $@
 
-$(BUILDDIR)/$(OPENBLAS_SRC_DIR)/build-configured: $(BUILDDIR)/$(OPENBLAS_SRC_DIR)/openblas-ofast-power.patch-applied
+$(BUILDDIR)/$(OPENBLAS_SRC_DIR)/openblas-exshift.patch-applied: $(BUILDDIR)/$(OPENBLAS_SRC_DIR)/openblas-ofast-power.patch-applied
+ cd $(BUILDDIR)/$(OPENBLAS_SRC_DIR) && \
+ patch -p1 -f < $(SRCDIR)/patches/openblas-exshift.patch
+ echo 1 > $@
+
+$(BUILDDIR)/$(OPENBLAS_SRC_DIR)/build-configured: $(BUILDDIR)/$(OPENBLAS_SRC_DIR)/openblas-exshift.patch-applied
  echo 1 > $@
 
 $(BUILDDIR)/$(OPENBLAS_SRC_DIR)/build-compiled: $(BUILDDIR)/$(OPENBLAS_SRC_DIR)/build-configured

diff --git a/deps/checksums/openblas b/deps/checksums/openblas
diff --git a/deps/openblas.version b/deps/openblas.version
@@ -1,2 +1,2 @@
-OPENBLAS_BRANCH=v0.3.10
-OPENBLAS_SHA1=63b03efc2af332c88b86d4fd8079d00f4b439adf
+OPENBLAS_BRANCH=v0.3.13
+OPENBLAS_SHA1=d2b11c47774b9216660e76e2fc67e87079f26fa1
diff --git a/deps/patches/openblas-exshift.patch b/deps/patches/openblas-exshift.patch
@@ -0,0 +1,149 @@
+commit c4b5abbe43d7c22215ef36ef4f7c1413c975678c
+Author: Martin Kroeker <[email protected]>
+Date: Fri Jan 29 10:45:36 2021 +0100
+
+ fix data type
+
+commit f87842483eee9d158f44d51d4c09662c3cff7526
+Author: Martin Kroeker <[email protected]>
+Date: Fri Jan 29 09:56:12 2021 +0100
+
+ fix calculation of non-exceptional shift (from Reference-LAPACK PR 477)
+
+commit 856bc365338f7559639f341d76ca8746d1628ee5
+Author: Martin Kroeker <[email protected]>
+Date: Wed Jan 27 13:41:45 2021 +0100
+
+ Add exceptional shift to fix rare convergence problems
+
+---
+diff --git a/lapack-netlib/SRC/chgeqz.f b/lapack-netlib/SRC/chgeqz.f
+index 73d35621..4725e716 100644
+--- a/lapack-netlib/SRC/chgeqz.f
++++ b/lapack-netlib/SRC/chgeqz.f
+@@ -320,12 +320,13 @@
+ $ C, SAFMIN, TEMP, TEMP2, TEMPR, ULP
+ COMPLEX ABI22, AD11, AD12, AD21, AD22, CTEMP, CTEMP2,
+ $ CTEMP3, ESHIFT, RTDISC, S, SHIFT, SIGNBC, T1,
+- $ U12, X
++ $ U12, X, ABI12, Y
+ * ..
+ * .. External Functions ..
++ COMPLEX CLADIV
+ LOGICAL LSAME
+ REAL CLANHS, SLAMCH
+- EXTERNAL LSAME, CLANHS, SLAMCH
++ EXTERNAL CLADIV, LLSAME, CLANHS, SLAMCH
+ * ..
+ * .. External Subroutines ..
+ EXTERNAL CLARTG, CLASET, CROT, CSCAL, XERBLA
+@@ -729,22 +730,34 @@
+ AD22 = ( ASCALE*H( ILAST, ILAST ) ) /
+ $ ( BSCALE*T( ILAST, ILAST ) )
+ ABI22 = AD22 - U12*AD21
++ ABI12 = AD12 - U12*AD11
+ *
+- T1 = HALF*( AD11+ABI22 )
+- RTDISC = SQRT( T1**2+AD12*AD21-AD11*AD22 )
+- TEMP = REAL( T1-ABI22 )*REAL( RTDISC ) +
+- $ AIMAG( T1-ABI22 )*AIMAG( RTDISC )
+- IF( TEMP.LE.ZERO ) THEN
+- SHIFT = T1 + RTDISC
+- ELSE
+- SHIFT = T1 - RTDISC
++ SHIFT = ABI22
++ CTEMP = SQRT( ABI12 )*SQRT( AD21 )
++ TEMP = ABS1( CTEMP )
++ IF( CTEMP.NE.ZERO ) THEN
++ X = HALF*( AD11-SHIFT )
++ TEMP2 = ABS1( X )
++ TEMP = MAX( TEMP, ABS1( X ) )
++ Y = TEMP*SQRT( ( X / TEMP )**2+( CTEMP / TEMP )**2 )
++ IF( TEMP2.GT.ZERO ) THEN
++ IF( REAL( X / TEMP2 )*REAL( Y )+
++ $ AIMAG( X / TEMP2 )*AIMAG( Y ).LT.ZERO )Y = -Y
++ END IF
++ SHIFT = SHIFT - CTEMP*CLADIV( CTEMP, ( X+Y ) )
+ END IF
+ ELSE
+ *
+ * Exceptional shift. Chosen for no particularly good reason.
+ *
+- ESHIFT = ESHIFT + (ASCALE*H(ILAST,ILAST-1))/
+- $ (BSCALE*T(ILAST-1,ILAST-1))
++ IF( ( IITER / 20 )*20.EQ.IITER .AND. 
++ $ BSCALE*ABS1(T( ILAST, ILAST )).GT.SAFMIN ) THEN
++ ESHIFT = ESHIFT + ( ASCALE*H( ILAST,
++ $ ILAST ) )/( BSCALE*T( ILAST, ILAST ) )
++ ELSE
++ ESHIFT = ESHIFT + ( ASCALE*H( ILAST,
++ $ ILAST-1 ) )/( BSCALE*T( ILAST-1, ILAST-1 ) )
++ END IF
+ SHIFT = ESHIFT
+ END IF
+ *
+diff --git a/lapack-netlib/SRC/zhgeqz.f b/lapack-netlib/SRC/zhgeqz.f
+index b51cba4f..b28ae47a 100644
+--- a/lapack-netlib/SRC/zhgeqz.f
++++ b/lapack-netlib/SRC/zhgeqz.f
+@@ -320,12 +320,13 @@
+ $ C, SAFMIN, TEMP, TEMP2, TEMPR, ULP
+ COMPLEX*16 ABI22, AD11, AD12, AD21, AD22, CTEMP, CTEMP2,
+ $ CTEMP3, ESHIFT, RTDISC, S, SHIFT, SIGNBC, T1,
+- $ U12, X
++ $ U12, X, ABI12, Y
+ * ..
+ * .. External Functions ..
++ COMPLEX*16 ZLADIV
+ LOGICAL LSAME
+ DOUBLE PRECISION DLAMCH, ZLANHS
+- EXTERNAL LSAME, DLAMCH, ZLANHS
++ EXTERNAL ZLADIV, LSAME, DLAMCH, ZLANHS
+ * ..
+ * .. External Subroutines ..
+ EXTERNAL XERBLA, ZLARTG, ZLASET, ZROT, ZSCAL
+@@ -730,22 +731,34 @@
+ AD22 = ( ASCALE*H( ILAST, ILAST ) ) /
+ $ ( BSCALE*T( ILAST, ILAST ) )
+ ABI22 = AD22 - U12*AD21
++ ABI12 = AD12 - U12*AD11
+ *
+- T1 = HALF*( AD11+ABI22 )
+- RTDISC = SQRT( T1**2+AD12*AD21-AD11*AD22 )
+- TEMP = DBLE( T1-ABI22 )*DBLE( RTDISC ) +
+- $ DIMAG( T1-ABI22 )*DIMAG( RTDISC )
+- IF( TEMP.LE.ZERO ) THEN
+- SHIFT = T1 + RTDISC
+- ELSE
+- SHIFT = T1 - RTDISC
++ SHIFT = ABI22
++ CTEMP = SQRT( ABI12 )*SQRT( AD21 )
++ TEMP = ABS1( CTEMP )
++ IF( CTEMP.NE.ZERO ) THEN
++ X = HALF*( AD11-SHIFT )
++ TEMP2 = ABS1( X )
++ TEMP = MAX( TEMP, ABS1( X ) )
++ Y = TEMP*SQRT( ( X / TEMP )**2+( CTEMP / TEMP )**2 )
++ IF( TEMP2.GT.ZERO ) THEN
++ IF( DBLE( X / TEMP2 )*DBLE( Y )+
++ $ DIMAG( X / TEMP2 )*DIMAG( Y ).LT.ZERO )Y = -Y
++ END IF
++ SHIFT = SHIFT - CTEMP*ZLADIV( CTEMP, ( X+Y ) )
+ END IF
+ ELSE
+ *
+ * Exceptional shift. Chosen for no particularly good reason.
+ *
+- ESHIFT = ESHIFT + (ASCALE*H(ILAST,ILAST-1))/
+- $ (BSCALE*T(ILAST-1,ILAST-1))
++ IF( ( IITER / 20 )*20.EQ.IITER .AND. 
++ $ BSCALE*ABS1(T( ILAST, ILAST )).GT.SAFMIN ) THEN
++ ESHIFT = ESHIFT + ( ASCALE*H( ILAST,
++ $ ILAST ) )/( BSCALE*T( ILAST, ILAST ) )
++ ELSE
++ ESHIFT = ESHIFT + ( ASCALE*H( ILAST,
++ $ ILAST-1 ) )/( BSCALE*T( ILAST-1, ILAST-1 ) )
++ END IF
+ SHIFT = ESHIFT
+ END IF
+ *
diff --git a/deps/patches/openblas-ofast-power.patch b/deps/patches/openblas-ofast-power.patch
@@ -1,33 +1,32 @@
- Makefile.power | 8 ++++----
- 1 file changed, 4 insertions(+), 4 deletions(-)
+ Makefile.power | 6 +++---
+ 1 file changed, 3 insertions(+), 3 deletions(-)
 
 diff --git a/Makefile.power b/Makefile.power
-index 24d8aa8a..e53a243a 100644
+index c7e97229..8426e816 100644
 --- a/Makefile.power
 +++ b/Makefile.power
-@@ -11,20 +11,20 @@ endif
-
- ifeq ($(CORE), POWER9)
- ifeq ($(USE_OPENMP), 1)
--COMMON_OPT += -Ofast -mcpu=power9 -mtune=power9 -mvsx -malign-power -DUSE_OPENMP -fno-fast-math -fopenmp
-+COMMON_OPT += -mcpu=power9 -mtune=power9 -mvsx -malign-power -DUSE_OPENMP -fno-fast-math -fopenmp
- FCOMMON_OPT += -O2 -frecursive -mcpu=power9 -mtune=power9 -malign-power -DUSE_OPENMP -fno-fast-math -fopenmp
- else
--COMMON_OPT += -Ofast -mcpu=power9 -mtune=power9 -mvsx -malign-power -fno-fast-math
-+COMMON_OPT += -mcpu=power9 -mtune=power9 -mvsx -malign-power -fno-fast-math
- FCOMMON_OPT += -O2 -frecursive -mcpu=power9 -mtune=power9 -malign-power -fno-fast-math
+@@ -10,13 +10,13 @@ USE_OPENMP = 1
  endif
+
+ ifeq ($(CORE), POWER10)
+-CCOMMON_OPT += -Ofast -mcpu=power10 -mtune=power10 -mvsx -fno-fast-math
++CCOMMON_OPT += -mcpu=power10 -mtune=power10 -mvsx -fno-fast-math
+ FCOMMON_OPT += -O2 -frecursive -mcpu=power10 -mtune=power10 -fno-fast-math
  endif
 
+ ifeq ($(CORE), POWER9)
+ ifneq ($(C_COMPILER), PGI)
+-CCOMMON_OPT += -Ofast -mvsx -fno-fast-math
++CCOMMON_OPT += -mvsx -fno-fast-math
+ ifeq ($(C_COMPILER), GCC)
+ ifneq ($(GCCVERSIONGT4), 1)
+ $(warning your compiler is too old to fully support POWER9, getting a newer version of gcc is recommended)
+@@ -49,7 +49,7 @@ endif
+
  ifeq ($(CORE), POWER8)
- ifeq ($(USE_OPENMP), 1)
--COMMON_OPT += -Ofast -mcpu=power8 -mtune=power8 -mvsx -malign-power -DUSE_OPENMP -fno-fast-math -fopenmp
-+COMMON_OPT += -mcpu=power8 -mtune=power8 -mvsx -malign-power -DUSE_OPENMP -fno-fast-math -fopenmp
- FCOMMON_OPT += -O2 -frecursive -mcpu=power8 -mtune=power8 -malign-power -DUSE_OPENMP -fno-fast-math -fopenmp
+ ifneq ($(C_COMPILER), PGI)
+-CCOMMON_OPT += -Ofast -mcpu=power8 -mtune=power8 -mvsx -fno-fast-math
++CCOMMON_OPT += -mcpu=power8 -mtune=power8 -mvsx -fno-fast-math
  else
--COMMON_OPT += -Ofast -mcpu=power8 -mtune=power8 -mvsx -malign-power -fno-fast-math
-+COMMON_OPT += -mcpu=power8 -mtune=power8 -mvsx -malign-power -fno-fast-math
- FCOMMON_OPT += -O2 -frecursive -mcpu=power8 -mtune=power8 -malign-power -fno-fast-math
+ CCOMMON_OPT += -fast -Mvect=simd -Mcache_align
  endif
- endif
-
diff --git a/deps/patches/openblas-winexit.patch b/deps/patches/openblas-winexit.patch
@@ -3,14 +3,14 @@ From: Keno Fischer <[email protected]>
 Date: Sat, 14 Mar 2020 12:05:19 +0100
 
 ---
- driver/others/memory.c | 131 +----------------------------------------
+ driver/others/memory.c | 131 +------------------------------------------------
  1 file changed, 2 insertions(+), 129 deletions(-)
 
 diff --git a/driver/others/memory.c b/driver/others/memory.c
-index 62a5a021..23f8fe65 100644
+index ba2bb55b..bf6b5529 100644
 --- a/driver/others/memory.c
 +++ b/driver/others/memory.c
-@@ -1510,7 +1510,7 @@ void CONSTRUCTOR gotoblas_init(void) {
+@@ -1534,7 +1534,7 @@ void CONSTRUCTOR gotoblas_init(void) {
 
  }
 
@@ -19,7 +19,7 @@ index 62a5a021..23f8fe65 100644
 
  if (gotoblas_initialized == 0) return;
 
-@@ -1547,74 +1547,12 @@ void DESTRUCTOR gotoblas_quit(void) {
+@@ -1571,74 +1571,12 @@ void DESTRUCTOR gotoblas_quit(void) {
  #endif
  }
 
@@ -57,8 +57,8 @@ index 62a5a021..23f8fe65 100644
 -*/
 -static int on_process_term(void)
 -{
-- gotoblas_quit();
-- return 0;
+- gotoblas_quit();
+- return 0;
 -}
  #ifdef _WIN64
  #pragma comment(linker, "/INCLUDE:_tls_used")
@@ -94,7 +94,7 @@ index 62a5a021..23f8fe65 100644
  #if (defined(C_PGI) || (!defined(C_SUN) && defined(F_INTERFACE_SUN))) && (defined(ARCH_X86) || defined(ARCH_X86_64))
  /* Don't call me; this is just work around for PGI / Sun bug */
  void gotoblas_dummy_for_PGI(void) {
-@@ -3104,7 +3042,7 @@ void CONSTRUCTOR gotoblas_init(void) {
+@@ -3136,7 +3074,7 @@ void CONSTRUCTOR gotoblas_init(void) {
 
  }
 
@@ -103,7 +103,7 @@ index 62a5a021..23f8fe65 100644
 
  if (gotoblas_initialized == 0) return;
 
-@@ -3133,71 +3071,6 @@ void DESTRUCTOR gotoblas_quit(void) {
+@@ -3165,71 +3103,6 @@ void DESTRUCTOR gotoblas_quit(void) {
  #endif
  }
 
@@ -138,8 +138,8 @@ index 62a5a021..23f8fe65 100644
 -*/
 -static int on_process_term(void)
 -{
-- gotoblas_quit();
-- return 0;
+- gotoblas_quit();
+- return 0;
 -}
 -#ifdef _WIN64
 -#pragma comment(linker, "/INCLUDE:_tls_used")
@@ -175,4 +175,3 @@ index 62a5a021..23f8fe65 100644
  #if (defined(C_PGI) || (!defined(C_SUN) && defined(F_INTERFACE_SUN))) && (defined(ARCH_X86) || defined(ARCH_X86_64))
  /* Don't call me; this is just work around for PGI / Sun bug */
  void gotoblas_dummy_for_PGI(void) {
-
diff --git a/stdlib/LinearAlgebra/src/LinearAlgebra.jl b/stdlib/LinearAlgebra/src/LinearAlgebra.jl
@@ -282,14 +282,14 @@ julia> ldiv!(Y, qr(A), X);
 julia> Y
 3-element Vector{Float64}:
  0.7128099173553719
- -0.051652892561983674
- 0.10020661157024757
+ -0.051652892561983806
+ 0.10020661157024781
 
 julia> A\\X
 3-element Vector{Float64}:
  0.7128099173553719
- -0.05165289256198333
- 0.10020661157024785
+ -0.05165289256198342
+ 0.1002066115702479
 ```
 """
 ldiv!(Y, A, B)
@@ -319,14 +319,14 @@ julia> ldiv!(qr(A), X);
 julia> X
 3-element Vector{Float64}:
  0.7128099173553719
- -0.051652892561983674
- 0.10020661157024757
+ -0.051652892561983806
+ 0.10020661157024781
 
 julia> A\\Y
 3-element Vector{Float64}:
  0.7128099173553719
- -0.05165289256198333
- 0.10020661157024785
+ -0.05165289256198342
+ 0.1002066115702479
 ```
 """
 ldiv!(A, B)

diff --git a/stdlib/LinearAlgebra/src/hessenberg.jl b/stdlib/LinearAlgebra/src/hessenberg.jl
@@ -402,9 +402,9 @@ Q factor:
  0.0 -0.707107 0.707107
 H factor:
 3×3 UpperHessenberg{Float64, Matrix{Float64}}:
- 4.0 -11.3137  -1.41421
- -5.65685 5.0  2.0
- ⋅ -8.88178e-16 1.0
+ 4.0 -11.3137 -1.41421
+ -5.65685 5.0 2.0
+ ⋅ -1.0444e-15 1.0
 
 julia> F.Q * F.H * F.Q'
 3×3 Matrix{Float64}:

diff --git a/stdlib/LinearAlgebra/test/eigen.jl b/stdlib/LinearAlgebra/test/eigen.jl
@@ -82,14 +82,15 @@ aimg = randn(n,n)/2
  a1_nsg = view(a, 1:n1, 1:n1)
  a2_nsg = view(a, n1+1:n2, n1+1:n2)
  end
- f = eigen(a1_nsg, a2_nsg)
+ sortfunc = x -> real(x) + imag(x)
+ f = eigen(a1_nsg, a2_nsg; sortby = sortfunc)
  @test a1_nsg*f.vectors ≈ (a2_nsg*f.vectors) * Diagonal(f.values)
- @test f.values ≈ eigvals(a1_nsg, a2_nsg)
- @test prod(f.values) ≈ prod(eigvals(a1_nsg/a2_nsg)) atol=50000ε
- @test eigvecs(a1_nsg, a2_nsg) == f.vectors
+ @test f.values ≈ eigvals(a1_nsg, a2_nsg; sortby = sortfunc)
+ @test prod(f.values) ≈ prod(eigvals(a1_nsg/a2_nsg, sortby = sortfunc)) atol=50000ε
+ @test eigvecs(a1_nsg, a2_nsg; sortby = sortfunc) == f.vectors
  @test_throws ErrorException f.Z
 
- d,v = eigen(a1_nsg, a2_nsg)
+ d,v = eigen(a1_nsg, a2_nsg; sortby = sortfunc)
  @test d == f.values
  @test v == f.vectors
  end

diff --git a/stdlib/OpenBLAS_jll/Project.toml b/stdlib/OpenBLAS_jll/Project.toml
@@ -1,6 +1,6 @@
 name = "OpenBLAS_jll"
 uuid = "4536629a-c528-5b80-bd46-f80d51c5b363"
-version = "0.3.10+3"
+version = "0.3.13+1"
 
 [deps]
 CompilerSupportLibraries_jll = "e66e0078-7015-5450-92f7-15fbd957f2ae"