Merge pull request #7213 from JuliaLang/vs/openblas

WIP: Bump OpenBLAS to 0.2.9 and LAPACK to 3.5.0

Author: ViralBShah

base/linalg/matmul.jl

@@ -203,23 +203,18 @@ function copytri!(A::StridedMatrix, uplo::Char, conjugate::Bool=false)
     A
 end
 
-if Base.blas_vendor() == :openblas
-    ## Avoid calling BLAS.gemv! when OpenBLAS is being used until #6941 is fixed.
-    gemv!{T<:BlasFloat}(y::StridedVector{T}, tA::Char, A::StridedMatrix{T}, x::StridedVector{T}) = generic_matvecmul!(y, tA, A, x)
-else
-    function gemv!{T<:BlasFloat}(y::StridedVector{T}, tA::Char, A::StridedMatrix{T}, x::StridedVector{T})
-        stride(A, 1)==1 || return generic_matvecmul!(y, tA, A, x)
-        if tA != 'N'
-            (nA, mA) = size(A)
-        else
-            (mA, nA) = size(A)
-        end
-        nA==length(x) || throw(DimensionMismatch(""))
-        mA==length(y) || throw(DimensionMismatch(""))
-        mA == 0 && return zeros(T, 0)
-        nA == 0 && return zeros(T, mA)
-        return BLAS.gemv!(tA, one(T), A, x, zero(T), y)
+function gemv!{T<:BlasFloat}(y::StridedVector{T}, tA::Char, A::StridedMatrix{T}, x::StridedVector{T})
+    stride(A, 1)==1 || return generic_matvecmul!(y, tA, A, x)
+    if tA != 'N'
+        (nA, mA) = size(A)
+    else
+        (mA, nA) = size(A)
     end
+    nA==length(x) || throw(DimensionMismatch(""))
+    mA==length(y) || throw(DimensionMismatch(""))
+    mA == 0 && return zeros(T, 0)
+    nA == 0 && return zeros(T, mA)
+    return BLAS.gemv!(tA, one(T), A, x, zero(T), y)
 end
 
 function syrk_wrapper!{T<:BlasFloat}(C::StridedMatrix{T}, tA::Char, A::StridedMatrix{T})

deps/Makefile

@@ -799,13 +799,6 @@ openblas-$(OPENBLAS_VER)/config.status: openblas-$(OPENBLAS_VER).tar.gz
 	mkdir -p openblas-$(OPENBLAS_VER) && \
 	$(TAR) -C openblas-$(OPENBLAS_VER) --strip-components 1 -xf $<
 	perl -i -ple 's/^\s*(EXTRALIB\s*\+=\s*-lSystemStubs)\s*$$/# $$1/g' openblas-$(OPENBLAS_VER)/Makefile.system
-ifeq  ($(OPENBLAS_VER),v0.2.8)
-	cd openblas-$(OPENBLAS_VER)/lapack-netlib/SRC && \
-		patch < ../../../dlasd4-lapack-3.4.2.patch && \
-		patch < ../../../slasd4-lapack-3.4.2.patch
-	cd openblas-$(OPENBLAS_VER)/exports && \
-		patch < ../../openblas-link-osx109.patch
-endif
 	echo 1 > $@
 $(OPENBLAS_OBJ_SOURCE): openblas-$(OPENBLAS_VER)/config.status
 	$(MAKE) -C openblas-$(OPENBLAS_VER) $(OPENBLAS_BUILD_OPTS) || (echo "*** Clean the OpenBLAS build with 'make -C deps clean-openblas'. Rebuild with 'make OPENBLAS_USE_THREAD=0 if OpenBLAS had trouble linking libpthread.so, and with 'make OPENBLAS_TARGET_ARCH=NEHALEM' if there were errors building SandyBridge support. Both these options can also be used simultaneously. ***" && false)
@@ -930,7 +923,6 @@ lapack-$(LAPACK_VER).tgz:
 lapack-$(LAPACK_VER)/Makefile: lapack-$(LAPACK_VER).tgz
 	$(JLCHECKSUM) $<
 	$(TAR) zxf $<
-	cd lapack-$(LAPACK_VER)/SRC && patch < ../../dlasd4-lapack-3.4.2.patch && patch < ../../slasd4-lapack-3.4.2.patch
 	touch -c $@
 ifeq ($(USE_SYSTEM_BLAS), 0)
 lapack-$(LAPACK_VER)/liblapack.a: | $(OPENBLAS_OBJ_TARGET)

deps/Versions.make

@@ -3,12 +3,8 @@ LLVM_LIB_SUFFIX =
 PCRE_VER = 8.31
 GRISU_VER = 1.1.1
 DSFMT_VER = 2.2
-ifeq ($(OS),WINNT)
-OPENBLAS_VER = v0.2.9.rc1
-else
-OPENBLAS_VER = v0.2.8
-endif
-LAPACK_VER = 3.4.2
+OPENBLAS_VER = v0.2.9
+LAPACK_VER = 3.5.0
 ARPACK_VER = 3.1.5
 FFTW_VER = 3.3.3
 SUITESPARSE_VER = 4.2.1

deps/checksums/lapack-3.5.0.tgz/md5

@@ -0,0 +1 @@
+b1d3e3e425b2e44a06760ff173104bdf

deps/checksums/lapack-3.5.0.tgz/sha512

@@ -0,0 +1 @@
+b948a0a0db032bda455ec4f519d4e89e4c29e29cecb5b6258ca61f68faaeeac9fdf4ece5c39ffcd0154c5505facbc392c7d09c8348b1d60bdd2685153ab2543f

deps/checksums/openblas-v0.2.9.rc1.tar.gz/md5

@@ -0,0 +1 @@
+b3af3bed5f353d1fc024cab1b836ac3e

deps/checksums/openblas-v0.2.9.rc1.tar.gz/sha512

@@ -0,0 +1 @@
+26675d3f66121e1be1b1ab149883afb392fa40f313380416d2e56093cf88c64e5d36dab366c734b9d5e9a4bd49085235cb73d819e9b559e924039cd2b42aa2f9

deps/checksums/openblas-v0.2.9.tar.gz/md5

@@ -266,3 +266,7 @@ debug && println("Matrix square root")
     end
 end
 end
+
+#6941
+#@test (ones(10^7,4)*ones(4))[3] == 4.0
+