Init Sandybridge codes based on Nehalem.

This commit is contained in:
Xianyi Zhang
2012-03-30 20:01:03 +08:00
parent 5cbbc496b0
commit 19a48b82cf
57 changed files with 309 additions and 45 deletions
+59
View File
@@ -0,0 +1,59 @@
SGEMMKERNEL = gemm_kernel_4x8_nehalem.S
SGEMMINCOPY = gemm_ncopy_4.S
SGEMMITCOPY = gemm_tcopy_4.S
SGEMMONCOPY = ../generic/gemm_ncopy_8.c
SGEMMOTCOPY = ../generic/gemm_tcopy_8.c
SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX)
SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)
DGEMMKERNEL = gemm_kernel_2x8_nehalem.S
DGEMMINCOPY = dgemm_ncopy_2.S
DGEMMITCOPY = dgemm_tcopy_2.S
DGEMMONCOPY = dgemm_ncopy_8.S
DGEMMOTCOPY = dgemm_tcopy_8.S
DGEMMINCOPYOBJ = dgemm_incopy$(TSUFFIX).$(SUFFIX)
DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX)
DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
CGEMMKERNEL = zgemm_kernel_2x4_nehalem.S
CGEMMINCOPY = zgemm_ncopy_2.S
CGEMMITCOPY = zgemm_tcopy_2.S
CGEMMONCOPY = ../generic/zgemm_ncopy_4.c
CGEMMOTCOPY = ../generic/zgemm_tcopy_4.c
CGEMMINCOPYOBJ = cgemm_incopy$(TSUFFIX).$(SUFFIX)
CGEMMITCOPYOBJ = cgemm_itcopy$(TSUFFIX).$(SUFFIX)
CGEMMONCOPYOBJ = cgemm_oncopy$(TSUFFIX).$(SUFFIX)
CGEMMOTCOPYOBJ = cgemm_otcopy$(TSUFFIX).$(SUFFIX)
ZGEMMKERNEL = zgemm_kernel_1x4_nehalem.S
ZGEMMINCOPY = zgemm_ncopy_1.S
ZGEMMITCOPY = zgemm_tcopy_1.S
ZGEMMONCOPY = ../generic/zgemm_ncopy_4.c
ZGEMMOTCOPY = ../generic/zgemm_tcopy_4.c
ZGEMMINCOPYOBJ = zgemm_incopy$(TSUFFIX).$(SUFFIX)
ZGEMMITCOPYOBJ = zgemm_itcopy$(TSUFFIX).$(SUFFIX)
ZGEMMONCOPYOBJ = zgemm_oncopy$(TSUFFIX).$(SUFFIX)
ZGEMMOTCOPYOBJ = zgemm_otcopy$(TSUFFIX).$(SUFFIX)
STRSMKERNEL_LN = trsm_kernel_LN_4x8_nehalem.S
STRSMKERNEL_LT = trsm_kernel_LT_4x8_nehalem.S
STRSMKERNEL_RN = trsm_kernel_LT_4x8_nehalem.S
STRSMKERNEL_RT = trsm_kernel_RT_4x8_nehalem.S
DTRSMKERNEL_LN = trsm_kernel_LN_2x8_nehalem.S
DTRSMKERNEL_LT = trsm_kernel_LT_2x8_nehalem.S
DTRSMKERNEL_RN = trsm_kernel_LT_2x8_nehalem.S
DTRSMKERNEL_RT = trsm_kernel_RT_2x8_nehalem.S
CTRSMKERNEL_LN = ztrsm_kernel_LN_2x4_nehalem.S
CTRSMKERNEL_LT = ztrsm_kernel_LT_2x4_nehalem.S
CTRSMKERNEL_RN = ztrsm_kernel_LT_2x4_nehalem.S
CTRSMKERNEL_RT = ztrsm_kernel_RT_2x4_nehalem.S
ZTRSMKERNEL_LN = ztrsm_kernel_LT_1x4_nehalem.S
ZTRSMKERNEL_LT = ztrsm_kernel_LT_1x4_nehalem.S
ZTRSMKERNEL_RN = ztrsm_kernel_LT_1x4_nehalem.S
ZTRSMKERNEL_RT = ztrsm_kernel_RT_1x4_nehalem.S
CGEMM3MKERNEL = zgemm3m_kernel_4x8_nehalem.S
ZGEMM3MKERNEL = zgemm3m_kernel_2x8_nehalem.S
+6
View File
@@ -45,6 +45,12 @@
#define PREFETCHW prefetcht0
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE 16
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#endif
#ifndef MOVAPS
#define MOVAPS movaps
#endif
+1 -1
View File
@@ -45,7 +45,7 @@
#define PREFETCHW prefetcht0
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCHSIZE 16
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
+6
View File
@@ -45,6 +45,12 @@
#define PREFETCHW prefetcht0
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE 12
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#endif
#ifndef MOVAPS
#define MOVAPS movaps
#endif
+7
View File
@@ -52,6 +52,13 @@
#define MOVUPS_A movups
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE 12
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define MOVUPS_A movups
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON)
#define PREFETCHSIZE 16
#define PREFETCH prefetcht0
+6
View File
@@ -51,6 +51,12 @@
#define MOVUPS_A movups
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE 12
#define PREFETCH prefetcht0
#define MOVUPS_A movups
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON)
#define PREFETCHSIZE 16
#define PREFETCH prefetcht0
+7
View File
@@ -46,6 +46,13 @@
#define MOVUPS_A movups
#endif
#ifdef SANDYBRIDGE
#define PREFETCHSIZE 16
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define MOVUPS_A movups
#endif
#ifdef MOVUPS_A
#define MOVUPS_A1(OFF, ADDR, REGS) MOVUPS_A OFF(ADDR), REGS
#define MOVUPS_A2(OFF, ADDR, BASE, SCALE, REGS) MOVUPS_A OFF(ADDR, BASE, SCALE), REGS
+7
View File
@@ -46,6 +46,13 @@
#define PREFETCHW prefetcht0
#endif
#if defined(SANDYBRIDGE)
#define RPREFETCHSIZE 12
#define WPREFETCHSIZE (RPREFETCHSIZE * 2)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#endif
#ifndef WINDOWS_ABI
#define M ARG1 /* rdi */
+1 -1
View File
@@ -46,7 +46,7 @@
#define PREFETCHW prefetcht0
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define RPREFETCHSIZE 12
#define WPREFETCHSIZE (RPREFETCHSIZE * 4)
#define PREFETCH prefetcht0
+7
View File
@@ -46,6 +46,13 @@
#define PREFETCHW prefetcht0
#endif
#if defined(SANDYBRIDGE)
#define RPREFETCHSIZE 12
#define WPREFETCHSIZE (RPREFETCHSIZE * 2)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#endif
#ifndef WINDOWS_ABI
#define M ARG1 /* rdi */
+1 -1
View File
@@ -46,7 +46,7 @@
#define PREFETCHW prefetcht0
#endif
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM)
#if defined(CORE2) || defined(PENRYN) || defined(DUNNINGTON) || defined(NEHALEM) || defined(SANDYBRIDGE)
#define RPREFETCHSIZE 12
#define WPREFETCHSIZE (RPREFETCHSIZE * 4)
#define PREFETCH prefetcht0
+1 -1
View File
@@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 12)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 12)
+1 -1
View File
@@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 12)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 12)
+1 -1
View File
@@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 12)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 12)
+1 -1
View File
@@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 12)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 24)
+1 -1
View File
@@ -685,7 +685,7 @@
cmpq $2 * SIZE, INCX
jne .L120
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM)
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM) && !defined(SANDYBRIDGE)
pshufd $0, %xmm0, %xmm14
pshufd $0, %xmm1, %xmm1
+2 -2
View File
@@ -55,7 +55,7 @@
#include "l1param.h"
#if defined(NEHALEM) || defined(PENRYN) || defined(DUNNINGTON) || defined(BARCELONA) || defined(NANO)
#if defined(NEHALEM) || defined(PENRYN) || defined(DUNNINGTON) || defined(BARCELONA) || defined(NANO) || defined(SANDYBRIDGE)
#define USE_PSHUFD
#else
#define USE_PSHUFD_HALF
@@ -803,7 +803,7 @@
cmpq $2 * SIZE, INCX
jne .L220
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM)
#if defined(ALIGNED_ACCESS) && !defined(NEHALEM) && !defined(SANDYBRIDGE)
movddup %xmm0, %xmm14
pxor %xmm15, %xmm15
+1 -1
View File
@@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 24)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 24)
+1 -1
View File
@@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 24)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 24)
+1 -1
View File
@@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 24)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 24)
+1 -1
View File
@@ -57,7 +57,7 @@
#define PREFETCHSIZE (16 * 24)
#endif
#ifdef NEHALEM
#if defined(NEHALEM) || defined(SANDYBRIDGE)
#define PREFETCH prefetcht0
#define PREFETCHW prefetcht0
#define PREFETCHSIZE (16 * 24)