📄 gemm_kernel_2x4_3dnow.s
字号:
/*********************************************************************//* *//* Optimized BLAS libraries *//* By Kazushige Goto <kgoto@tacc.utexas.edu> *//* *//* Copyright (c) The University of Texas, 2005. All rights reserved. *//* UNIVERSITY EXPRESSLY DISCLAIMS ANY AND ALL WARRANTIES CONCERNING *//* THIS SOFTWARE AND DOCUMENTATION, INCLUDING ANY WARRANTIES OF *//* MERCHANTABILITY, FITNESS FOR ANY PARTICULAR PURPOSE, *//* NON-INFRINGEMENT AND WARRANTIES OF PERFORMANCE, AND ANY WARRANTY *//* THAT MIGHT OTHERWISE ARISE FROM COURSE OF DEALING OR USAGE OF *//* TRADE. NO WARRANTY IS EITHER EXPRESS OR IMPLIED WITH RESPECT TO *//* THE USE OF THE SOFTWARE OR DOCUMENTATION. *//* Under no circumstances shall University be liable for incidental, *//* special, indirect, direct or consequential damages or loss of *//* profits, interruption of business, or related expenses which may *//* arise from use of Software or Documentation, including but not *//* limited to those resulting from defects in Software and/or *//* Documentation, or loss or inaccuracy of data of any kind. *//*********************************************************************/#define ASSEMBLER#include "common.h"#define STACK 16#define ARGS 0 #define OLD_M 4 + STACK + ARGS(%esi)#define OLD_N 8 + STACK + ARGS(%esi)#define OLD_K 12 + STACK + ARGS(%esi)#define OLD_ALPHA 16 + STACK + ARGS(%esi)#define OLD_A 20 + STACK + ARGS(%esi)#define OLD_B 24 + STACK + ARGS(%esi)#define OLD_C 28 + STACK + ARGS(%esi)#define OLD_LDC 32 + STACK + ARGS(%esi)#define OLD_OFFSET 36 + STACK + ARGS(%esi)#define ALPHA 0(%esp)#define K 8(%esp)#define N 12(%esp)#define M 16(%esp)#define A 20(%esp)#define C 24(%esp)#define J 28(%esp)#define OLD_STACK 32(%esp)#define OFFSET 36(%esp)#define KK 40(%esp)#define KKK 44(%esp)#define BUFFER 64(%esp)#define AA %edx#define BB %ecx#define PREFETCHSIZE (16 * 2 + 6)#define LOCAL_BUFFER_SIZE GEMM_Q * GEMM_UNROLL_N * COMPSIZE * 16#define AOFFSET -32#define BOFFSET 128/* A hint of scheduling is received from following URLhttps://sourceforge.net/mailarchive/forum.php?forum_id=426&max_rows=25&style=flat&viewmonth=200309&viewday=11*/ PROLOGUE pushl %ebp pushl %edi pushl %esi pushl %ebx PROFCODE EMMS movl %esp, %esi # save old stack subl $128 + LOCAL_BUFFER_SIZE, %esp movl OLD_M, %ebx andl $-1024, %esp # align stack#ifdef WINDOWS_ABI#if LOCAL_BUFFER_SIZE > 12288 movl $0, 4096 * 3(%esp)#endif#if LOCAL_BUFFER_SIZE > 8192 movl $0, 4096 * 2(%esp)#endif#if LOCAL_BUFFER_SIZE > 4096 movl $0, 4096 * 1(%esp)#endif movl $0, 4096 * 0(%esp)#endif movl OLD_N, %eax movl OLD_K, %ecx movl OLD_A, %edx movd OLD_ALPHA, %mm3 movl %ebx, M movl %eax, N movl %ecx, K subl $AOFFSET * SIZE, %edx movl %edx, A movl %esi, OLD_STACK movl OLD_B, %edi movl OLD_C, %ebx punpckldq %mm3, %mm3 movq %mm3, ALPHA movl %ebx, C movl OLD_LDC, %ebp leal (, %ebp, SIZE), %ebp#ifdef TRMMKERNEL movl OLD_OFFSET, %eax movl %eax, OFFSET#ifndef LEFT negl %eax movl %eax, KK#endif#endif movl N, %eax sarl $2, %eax movl %eax, J jle .L30 ALIGN_3.L01:/* Copying to Sub Buffer */ leal BUFFER, %ecx#if defined(TRMMKERNEL) && defined(LEFT) movl OFFSET, %eax movl %eax, KK#endif movl K, %eax sarl $2, %eax jle .L03 ALIGN_3 .L02: movd 0 * SIZE(%edi), %mm0 movd 1 * SIZE(%edi), %mm1 movd 2 * SIZE(%edi), %mm2 movd 3 * SIZE(%edi), %mm3 movd 4 * SIZE(%edi), %mm4 movd 5 * SIZE(%edi), %mm5 movd 6 * SIZE(%edi), %mm6 movd 7 * SIZE(%edi), %mm7 prefetchnta 72 * SIZE(%edi) punpckldq %mm0, %mm0 punpckldq %mm1, %mm1 punpckldq %mm2, %mm2 punpckldq %mm3, %mm3 punpckldq %mm4, %mm4 punpckldq %mm5, %mm5 punpckldq %mm6, %mm6 punpckldq %mm7, %mm7 movq %mm0, 0 * SIZE(%ecx) movq %mm1, 2 * SIZE(%ecx) movq %mm2, 4 * SIZE(%ecx) movq %mm3, 6 * SIZE(%ecx) movq %mm4, 8 * SIZE(%ecx) movq %mm5, 10 * SIZE(%ecx) movq %mm6, 12 * SIZE(%ecx) movq %mm7, 14 * SIZE(%ecx) movd 8 * SIZE(%edi), %mm0 movd 9 * SIZE(%edi), %mm1 movd 10 * SIZE(%edi), %mm2 movd 11 * SIZE(%edi), %mm3 movd 12 * SIZE(%edi), %mm4 movd 13 * SIZE(%edi), %mm5 movd 14 * SIZE(%edi), %mm6 movd 15 * SIZE(%edi), %mm7 punpckldq %mm0, %mm0 punpckldq %mm1, %mm1 punpckldq %mm2, %mm2 punpckldq %mm3, %mm3 punpckldq %mm4, %mm4 punpckldq %mm5, %mm5 punpckldq %mm6, %mm6 punpckldq %mm7, %mm7 movq %mm0, 16 * SIZE(%ecx) movq %mm1, 18 * SIZE(%ecx) movq %mm2, 20 * SIZE(%ecx) movq %mm3, 22 * SIZE(%ecx) movq %mm4, 24 * SIZE(%ecx) movq %mm5, 26 * SIZE(%ecx) movq %mm6, 28 * SIZE(%ecx) movq %mm7, 30 * SIZE(%ecx) addl $16 * SIZE, %edi addl $32 * SIZE, %ecx decl %eax jne .L02.L03: movl K, %eax andl $3, %eax BRANCH jle .L10 ALIGN_2.L04: movd 0 * SIZE(%edi), %mm0 movd 1 * SIZE(%edi), %mm1 movd 2 * SIZE(%edi), %mm2 movd 3 * SIZE(%edi), %mm3 punpckldq %mm0, %mm0 punpckldq %mm1, %mm1 punpckldq %mm2, %mm2 punpckldq %mm3, %mm3 movq %mm0, 0 * SIZE(%ecx) movq %mm1, 2 * SIZE(%ecx) movq %mm2, 4 * SIZE(%ecx) movq %mm3, 6 * SIZE(%ecx) addl $4 * SIZE, %edi addl $8 * SIZE, %ecx decl %eax jne .L04 ALIGN_4 .L10: movl C, %esi # coffset = c movl A, %edx # aoffset = a movl M, %ebx sarl $1, %ebx # i = (m >> 2) jle .L20 ALIGN_4.L11: leal - BOFFSET * SIZE + BUFFER, BB#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))#else movl KK, %eax leal (, %eax, SIZE), %eax leal (AA, %eax, 2), AA leal (BB, %eax, 8), BB#endif movq ( 0 + AOFFSET) * SIZE(AA), %mm0 pxor %mm4, %mm4 movq ( 16 + AOFFSET) * SIZE(AA), %mm1 pxor %mm5, %mm5 PADDING movq ( 0 + BOFFSET) * SIZE(BB), %mm2 pxor %mm6, %mm6 PADDING movq ( 16 + BOFFSET) * SIZE(BB), %mm3 pxor %mm7, %mm7 leal (%ebp, %ebp, 2), %eax prefetchw 2 * SIZE(%esi) prefetchw 2 * SIZE(%esi, %ebp) prefetchw 2 * SIZE(%esi, %ebp, 2) prefetchw 2 * SIZE(%esi, %eax)#ifndef TRMMKERNEL movl K, %eax#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) movl K, %eax subl KK, %eax movl %eax, KKK #else movl KK, %eax#ifdef LEFT addl $2, %eax#else addl $4, %eax#endif movl %eax, KKK#endif sarl $4, %eax je .L15 ALIGN_4.L12: pfmul %mm0, %mm2 pfadd %mm2, %mm4 PADDING movq ( 2 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm0, %mm2 pfadd %mm2, %mm5 PADDING movq ( 4 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm0, %mm2 pfadd %mm2, %mm6 PADDING prefetch (PREFETCHSIZE + 0) * SIZE(AA) PADDING movq ( 8 + BOFFSET) * SIZE(BB), %mm2 pfmul ( 6 + BOFFSET) * SIZE(BB), %mm0 pfadd %mm0, %mm7 movq ( 2 + AOFFSET) * SIZE(AA), %mm0 pfmul %mm0, %mm2 pfadd %mm2, %mm4 PADDING movq ( 10 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm0, %mm2 pfadd %mm2, %mm5 PADDING movq ( 12 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm0, %mm2 pfadd %mm2, %mm6 PADDING movq ( 32 + BOFFSET) * SIZE(BB), %mm2 pfmul ( 14 + BOFFSET) * SIZE(BB), %mm0 pfadd %mm0, %mm7 movq ( 4 + AOFFSET) * SIZE(AA), %mm0 pfmul %mm0, %mm3 pfadd %mm3, %mm4 PADDING movq ( 18 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm0, %mm3 pfadd %mm3, %mm5 PADDING movq ( 20 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm0, %mm3 pfadd %mm3, %mm6 PADDING movq ( 24 + BOFFSET) * SIZE(BB), %mm3 pfmul ( 22 + BOFFSET) * SIZE(BB), %mm0 pfadd %mm0, %mm7 movq ( 6 + AOFFSET) * SIZE(AA), %mm0 pfmul %mm0, %mm3 pfadd %mm3, %mm4 PADDING movq ( 26 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm0, %mm3 pfadd %mm3, %mm5 PADDING movq ( 28 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm0, %mm3 pfadd %mm3, %mm6 PADDING movq ( 48 + BOFFSET) * SIZE(BB), %mm3 pfmul ( 30 + BOFFSET) * SIZE(BB), %mm0 pfadd %mm0, %mm7 movq ( 8 + AOFFSET) * SIZE(AA), %mm0 pfmul %mm0, %mm2 pfadd %mm2, %mm4 PADDING movq ( 34 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm0, %mm2 pfadd %mm2, %mm5 PADDING movq ( 36 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm0, %mm2 pfadd %mm2, %mm6 PADDING movq ( 40 + BOFFSET) * SIZE(BB), %mm2 pfmul ( 38 + BOFFSET) * SIZE(BB), %mm0 pfadd %mm0, %mm7 movq ( 10 + AOFFSET) * SIZE(AA), %mm0 pfmul %mm0, %mm2 pfadd %mm2, %mm4 PADDING movq ( 42 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm0, %mm2 pfadd %mm2, %mm5 PADDING movq ( 44 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm0, %mm2 pfadd %mm2, %mm6 PADDING movq ( 64 + BOFFSET) * SIZE(BB), %mm2 pfmul ( 46 + BOFFSET) * SIZE(BB), %mm0 pfadd %mm0, %mm7 movq ( 12 + AOFFSET) * SIZE(AA), %mm0 pfmul %mm0, %mm3 pfadd %mm3, %mm4 PADDING movq ( 50 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm0, %mm3 pfadd %mm3, %mm5 PADDING movq ( 52 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm0, %mm3 pfadd %mm3, %mm6 PADDING movq ( 56 + BOFFSET) * SIZE(BB), %mm3 pfmul ( 54 + BOFFSET) * SIZE(BB), %mm0 pfadd %mm0, %mm7 movq ( 14 + AOFFSET) * SIZE(AA), %mm0 pfmul %mm0, %mm3 pfadd %mm3, %mm4 PADDING movq ( 58 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm0, %mm3 pfadd %mm3, %mm5 PADDING movq ( 60 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm0, %mm3 pfadd %mm3, %mm6 PADDING movq ( 80 + BOFFSET) * SIZE(BB), %mm3 pfmul ( 62 + BOFFSET) * SIZE(BB), %mm0 pfadd %mm0, %mm7 movq ( 32 + AOFFSET) * SIZE(AA), %mm0 pfmul %mm1, %mm2 pfadd %mm2, %mm4 PADDING movq ( 66 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm1, %mm2 pfadd %mm2, %mm5 PADDING movq ( 68 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm1, %mm2 pfadd %mm2, %mm6 PADDING movq ( 72 + BOFFSET) * SIZE(BB), %mm2 pfmul ( 70 + BOFFSET) * SIZE(BB), %mm1 pfadd %mm1, %mm7 movq ( 18 + AOFFSET) * SIZE(AA), %mm1 pfmul %mm1, %mm2 pfadd %mm2, %mm4 PADDING movq ( 74 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm1, %mm2 pfadd %mm2, %mm5 PADDING movq ( 76 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm1, %mm2 pfadd %mm2, %mm6 PADDING movq ( 96 + BOFFSET) * SIZE(BB), %mm2 pfmul ( 78 + BOFFSET) * SIZE(BB), %mm1 pfadd %mm1, %mm7 movq ( 20 + AOFFSET) * SIZE(AA), %mm1 pfmul %mm1, %mm3 pfadd %mm3, %mm4 PADDING movq ( 82 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm1, %mm3 pfadd %mm3, %mm5 PADDING movq ( 84 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm1, %mm3 pfadd %mm3, %mm6 PADDING movq ( 88 + BOFFSET) * SIZE(BB), %mm3 pfmul ( 86 + BOFFSET) * SIZE(BB), %mm1 pfadd %mm1, %mm7 movq ( 22 + AOFFSET) * SIZE(AA), %mm1 pfmul %mm1, %mm3 pfadd %mm3, %mm4 PADDING movq ( 90 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm1, %mm3 pfadd %mm3, %mm5 PADDING movq ( 92 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm1, %mm3 pfadd %mm3, %mm6 PADDING movq (112 + BOFFSET) * SIZE(BB), %mm3 pfmul ( 94 + BOFFSET) * SIZE(BB), %mm1 pfadd %mm1, %mm7 movq ( 24 + AOFFSET) * SIZE(AA), %mm1 pfmul %mm1, %mm2 pfadd %mm2, %mm4 PADDING movq ( 98 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm1, %mm2 pfadd %mm2, %mm5 PADDING movq (100 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm1, %mm2 pfadd %mm2, %mm6 PADDING movq (104 + BOFFSET) * SIZE(BB), %mm2 pfmul (102 + BOFFSET) * SIZE(BB), %mm1 pfadd %mm1, %mm7 movq ( 26 + AOFFSET) * SIZE(AA), %mm1 pfmul %mm1, %mm2 pfadd %mm2, %mm4 PADDING movq (106 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm1, %mm2 pfadd %mm2, %mm5 PADDING movq (108 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm1, %mm2 pfadd %mm2, %mm6 PADDING movq (128 + BOFFSET) * SIZE(BB), %mm2 pfmul (110 + BOFFSET) * SIZE(BB), %mm1 pfadd %mm1, %mm7 movq ( 28 + AOFFSET) * SIZE(AA), %mm1 pfmul %mm1, %mm3 pfadd %mm3, %mm4 PADDING movq (114 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm1, %mm3 pfadd %mm3, %mm5 PADDING movq (116 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm1, %mm3 pfadd %mm3, %mm6 PADDING movq (120 + BOFFSET) * SIZE(BB), %mm3 pfmul (118 + BOFFSET) * SIZE(BB), %mm1 pfadd %mm1, %mm7 movq ( 30 + AOFFSET) * SIZE(AA), %mm1 pfmul %mm1, %mm3 pfadd %mm3, %mm4 PADDING movq (122 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm1, %mm3 pfadd %mm3, %mm5 PADDING movq (124 + BOFFSET) * SIZE(BB), %mm3 pfmul %mm1, %mm3 pfadd %mm3, %mm6 PADDING movq (144 + BOFFSET) * SIZE(BB), %mm3 pfmul (126 + BOFFSET) * SIZE(BB), %mm1 pfadd %mm1, %mm7 movq ( 48 + AOFFSET) * SIZE(AA), %mm1 subl $-32 * SIZE, AA addl $128 * SIZE, BB decl %eax jne .L12 ALIGN_3.L15: movq ALPHA, %mm3#ifndef TRMMKERNEL movl K, %eax#else movl KKK, %eax#endif andl $15, %eax BRANCH je .L18 ALIGN_3.L16: pfmul %mm0, %mm2 pfadd %mm2, %mm4 PADDING movq ( 2 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm0, %mm2 pfadd %mm2, %mm5 PADDING movq ( 4 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm0, %mm2 pfadd %mm2, %mm6 PADDING movq ( 8 + BOFFSET) * SIZE(BB), %mm2 pfmul ( 6 + BOFFSET) * SIZE(BB), %mm0 pfadd %mm0, %mm7 movq ( 2 + AOFFSET) * SIZE(AA), %mm0 addl $2 * SIZE, AA addl $8 * SIZE, BB decl %eax jg .L16 ALIGN_3.L18: leal (%ebp, %ebp, 2), %eax#ifndef TRMMKERNEL pfmul %mm3, %mm4 pfadd 0 * SIZE(%esi), %mm4 pfmul %mm3, %mm5 pfadd 0 * SIZE(%esi, %ebp, 1), %mm5 pfmul %mm3, %mm6 pfadd 0 * SIZE(%esi, %ebp, 2), %mm6 pfmul %mm3, %mm7 pfadd 0 * SIZE(%esi, %eax, 1), %mm7#else pfmul %mm3, %mm4 pfmul %mm3, %mm5 pfmul %mm3, %mm6 pfmul %mm3, %mm7#endif movq %mm4, 0 * SIZE(%esi) movq %mm5, 0 * SIZE(%esi, %ebp, 1) movq %mm6, 0 * SIZE(%esi, %ebp, 2) movq %mm7, 0 * SIZE(%esi, %eax, 1)#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) movl K, %eax subl KKK, %eax leal (,%eax, SIZE), %eax leal (AA, %eax, 2), AA leal (BB, %eax, 8), BB#endif#if defined(TRMMKERNEL) && defined(LEFT) addl $2, KK#endif addl $2 * SIZE, %esi # coffset += 2 decl %ebx # i -- jg .L11 ALIGN_4.L20: movl M, %ebx testl $1, %ebx # i = (m >> 2) jle .L29 ALIGN_4.L21: leal - BOFFSET * SIZE + BUFFER, BB#if !defined(TRMMKERNEL) || \ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))#else movl KK, %eax leal (, %eax, SIZE), %eax leal (AA, %eax, 1), AA leal (BB, %eax, 8), BB#endif movq ( 0 + AOFFSET) * SIZE(AA), %mm0 pxor %mm4, %mm4 movq ( 8 + AOFFSET) * SIZE(AA), %mm1 pxor %mm5, %mm5 PADDING movq ( 0 + BOFFSET) * SIZE(BB), %mm2 pxor %mm6, %mm6 PADDING movq ( 16 + BOFFSET) * SIZE(BB), %mm3 pxor %mm7, %mm7#ifndef TRMMKERNEL movl K, %eax#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) movl K, %eax subl KK, %eax movl %eax, KKK #else movl KK, %eax#ifdef LEFT addl $1, %eax#else addl $4, %eax#endif movl %eax, KKK#endif sarl $4, %eax je .L25 ALIGN_4.L22: pfmul %mm0, %mm2 pfadd %mm2, %mm4 PADDING movd ( 2 + BOFFSET) * SIZE(BB), %mm2 pfmul %mm0, %mm2⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -