Attachment 927916 Details for Bug 1080073 – atlas.3.10.2-ppc64le_abiv2_step3.patch

[patch] atlas.3.10.2-ppc64le_abiv2_step3.patch

atlas.3.10.2-ppc64le_abiv2_step3.patch (text/plain), 6.60 KB, created by Michel Normand on 2014-08-18 14:17:39 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Michel Normand

Created: 2014-08-18 14:17:39 UTC

Size: 6.60 KB

patch

obsolete

>From: Michel Normand <normand@linux.vnet.ibm.com>
>Subject: atlas.3.10.2 ppc64le abiv2 step3
>Date: Tue, 29 Jul 2014 15:33:18 +0200
>
>atlas.3.10.2 ppc64le abiv2 step3
>* change offsets of parameters read from stack to avoid some segfaults.
>  (values changes 120 => 104 and 128 => 112 identified by gdb investigation)
>
>Despite this step3 patch there are two Remaining problems for ppc64le archi:
>* TODO: still have seg-faults in console during build/check
>but is not critical (without make check) and rpm are generated on fedora.
>unable to investigate because of problem tracked by issue 950
>https://sourceforge.net/p/math-atlas/support-requests/950/
>
>* TODO: make check failure because xsslvtst execution failure
>related to vector assembly code that assumes big-endian env
>as written in ATL_cmm4x4x128_av.c and ATL_smm4x4x128_av.c.
>Would need significant work to support little-endian as per
>endianess comments of all PowerPC vector instructions in:
>https://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/FBFA164F824370F987256D6A006F424D/$file/vector_simd_pem.ppc.2005AUG23.pdf
>
>Signed-off-by: Michel Normand <normand@linux.vnet.ibm.com>
>---
> tune/blas/gemm/CASES/ATL_cmm4x4x128_av.c |    7 +++++++
> tune/blas/gemm/CASES/ATL_dmm4x4x2pf_av.c |    7 +++++++
> tune/blas/gemm/CASES/ATL_dmm4x4x32_ppc.c |    7 +++++++
> tune/blas/gemm/CASES/ATL_dmm4x4x80_ppc.c |   17 ++++++++++++++++-
> tune/blas/gemm/CASES/ATL_smm4x4x128_av.c |   21 +++++++++++++++++++++
> 5 files changed, 58 insertions(+), 1 deletion(-)
>
>Index: ATLAS/tune/blas/gemm/CASES/ATL_dmm4x4x2pf_av.c
>===================================================================
>--- ATLAS.orig/tune/blas/gemm/CASES/ATL_dmm4x4x2pf_av.c
>+++ ATLAS/tune/blas/gemm/CASES/ATL_dmm4x4x2pf_av.c
>@@ -405,8 +405,15 @@ Mjoin(_,ATL_USERMM):
>  */
> #ifdef ATL_GAS_LINUX_PPC
>    #ifdef ATL_USE64BITS
>+      #if _CALL_ELF == 2
>+      /* ABIv2 */
>+        ld      pC0, 104(r1)
>+        ld      incCn, 112(r1)
>+      #else
>+      /* ABIv1 */
> 	ld 	pC0, 120(r1)
> 	ld 	incCn, 128(r1)
>+      #endif
>    #else
> 	lwz	incCn, FSIZE+8(r1)
>    #endif
>Index: ATLAS/tune/blas/gemm/CASES/ATL_dmm4x4x32_ppc.c
>===================================================================
>--- ATLAS.orig/tune/blas/gemm/CASES/ATL_dmm4x4x32_ppc.c
>+++ ATLAS/tune/blas/gemm/CASES/ATL_dmm4x4x32_ppc.c
>@@ -324,8 +324,15 @@ ATL_USERMM:
> #endif
> 
> #ifdef ATL_USE64BITS
>+#if _CALL_ELF == 2
>+/* ABIv2 */
>+        ld      pC0, 104(r1)
>+        ld      incCn, 112(r1)
>+#else
>+/* ABIv1 */
>         ld      pC0, 120(r1)
>         ld      incCn, 128(r1)
>+#endif
> #elif defined(ATL_AS_OSX_PPC) || defined(ATL_AS_AIX_PPC)
>         lwz     pC0, 68(r1)
>         lwz     incCn,  72(r1)
>Index: ATLAS/tune/blas/gemm/CASES/ATL_dmm4x4x80_ppc.c
>===================================================================
>--- ATLAS.orig/tune/blas/gemm/CASES/ATL_dmm4x4x80_ppc.c
>+++ ATLAS/tune/blas/gemm/CASES/ATL_dmm4x4x80_ppc.c
>@@ -170,13 +170,21 @@ void ATL_USERMM(const int M, const int N
>                 const TYPE beta, TYPE *C, const int ldc)
>                                   (r10)    8(r1)
> *******************************************************************************
>-64 bit ABIs:
>+64 bit ABIv1s:
>                          r3           r4           r5             r6/f1
> void ATL_USERMM(const int M, const int N, const int K, const TYPE alpha,
>                            r7             r8             r9            r10
>                 const TYPE *A, const int lda, const TYPE *B, const int ldb,
>                              f2   120(r1)        128(r1)
>                 const TYPE beta, TYPE *C, const int ldc)
>+
>+64 bit ABIv2s:
>+                         r3           r4           r5             r6/f1
>+void ATL_USERMM(const int M, const int N, const int K, const TYPE alpha,
>+                           r7             r8             r9            r10
>+                const TYPE *A, const int lda, const TYPE *B, const int ldb,
>+                             f2   104(r1)        112(r1)
>+                const TYPE beta, TYPE *C, const int ldc)
> #endif
> #ifdef ATL_AS_AIX_PPC
>         .csect .text[PR]
>@@ -259,8 +267,15 @@ ATL_USERMM:
> 
> 
> #if defined (ATL_USE64BITS)
>+#if _CALL_ELF == 2
>+/* ABIv2 */
>+        ld      pC0, 104(r1)
>+        ld      incCn, 112(r1)
>+#else
>+/* ABIv1 */
>         ld      pC0, 120(r1)
>         ld      incCn, 128(r1)
>+#endif
> #elif defined(ATL_AS_OSX_PPC) || defined(ATL_AS_AIX_PPC)
>         lwz     pC0, 68(r1)
>         lwz     incCn,  72(r1)
>Index: ATLAS/tune/blas/gemm/CASES/ATL_smm4x4x128_av.c
>===================================================================
>--- ATLAS.orig/tune/blas/gemm/CASES/ATL_smm4x4x128_av.c
>+++ ATLAS/tune/blas/gemm/CASES/ATL_smm4x4x128_av.c
>@@ -221,8 +221,15 @@ ATL_USERMM:
>  *      kernel instead
>  */
> #if defined (ATL_USE64BITS)
>+#if _CALL_ELF == 2
>+/* ABIv2 */
>+        ld      r10, 104(r1)
>+        ld      r5, 112(r1)
>+#else
>+/* ABIv1 */
>         ld      r10, 120(r1)
>         ld      r5, 128(r1)
>+#endif
> #elif defined(ATL_AS_OSX_PPC)
>         lwz     r10, 60(r1)
>         lwz     r5,  64(r1)
>@@ -285,8 +292,15 @@ ATL_USERMM:
>         eqv     r0, r0, r0      /* all 1s */
>         ATL_WriteVRSAVE(r0)     /* signal we use all vector regs */
> #if defined (ATL_USE64BITS)
>+#if _CALL_ELF == 2
>+        /* ABIv2 */
>+        ld      pC0, FSIZE+104(r1)
>+        ld      ldc, FSIZE+112(r1)
>+#else
>+        /* ABIv1 */
>         ld      pC0, FSIZE+120(r1)
>         ld      ldc, FSIZE+128(r1)
>+#endif
> #elif defined(ATL_AS_OSX_PPC)
>         lwz     pC0, FSIZE+60(r1)
>         lwz     ldc,  FSIZE+64(r1)
>@@ -4258,8 +4272,15 @@ UNALIGNED_C:
>         eqv     r0, r0, r0      /* all 1s */
>         ATL_WriteVRSAVE(r0)     /* signal we use all vector regs */
> #if defined (ATL_USE64BITS)
>+#if _CALL_ELF == 2
>+        /* ABIv2 */
>+        ld      pC0, FSIZE+104(r1)
>+        ld      ldc, FSIZE+112(r1)
>+#else
>+        /* ABIv1 */
>         ld      pC0, FSIZE+120(r1)
>         ld      ldc, FSIZE+128(r1)
>+#endif
> #elif defined(ATL_AS_OSX_PPC)
>         lwz     pC0, FSIZE+60(r1)
>         lwz     ldc,  FSIZE+64(r1)
>Index: ATLAS/tune/blas/gemm/CASES/ATL_cmm4x4x128_av.c
>===================================================================
>--- ATLAS.orig/tune/blas/gemm/CASES/ATL_cmm4x4x128_av.c
>+++ ATLAS/tune/blas/gemm/CASES/ATL_cmm4x4x128_av.c
>@@ -258,8 +258,15 @@ ATL_USERMM:
>         eqv     r0, r0, r0      /* all 1s */
>         ATL_WriteVRSAVE(r0)     /* signal we use all vector regs */
> #if defined (ATL_USE64BITS)
>+#if _CALL_ELF == 2
>+/* ABIv2 */
>+        ld      pC0, FSIZE+104(r1)
>+        ld      ldc, FSIZE+112(r1)
>+#else
>+/* ABIv1 */
>         ld      pC0, FSIZE+120(r1)
>         ld      ldc, FSIZE+128(r1)
>+#endif
> #elif defined(ATL_AS_OSX_PPC)
>         lwz     pC0, FSIZE+60(r1)
>         lwz     ldc,  FSIZE+64(r1)

Actions: View | Diff

Attachments on bug 1080073: 878111 | 898654 | 898655 | 898656 | 898657 | 898658 | 898659 | 917277 | 917280 | 917284 | 919654 | 927912 | 927913 | 927915 | 927916 | 939224 | 939225 | 939357 | 940404 | 942268 | 942576 | 942578 | 943255