📄 gemm_kernel_power3.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" #ifndef __64BIT__#define LOAD lwz#else#define LOAD ld#endif#ifdef __64BIT__#define STACKSIZE 320#define ALPHA 296(SP)#define FZERO 304(SP)#else#define STACKSIZE 240#define ALPHA 224(SP)#define FZERO 232(SP)#endif#define M r3#define N r4#define K r5#ifdef linux#ifndef __64BIT__#define A r6#define B r7#define C r8#define LDC r9#else#define A r7#define B r8#define C r9#define LDC r10#endif#endif#if defined(_AIX) || defined(__APPLE__)#if !defined(__64BIT__) && defined(DOUBLE)#define A r8#define B r9#define C r10#define LDC r7#else#define A r7#define B r8#define C r9#define LDC r10#endif#endif#define I r21#define J r22#define AO r23#define BO r24#define CO1 r25#define CO2 r26#define CO3 r27#define CO4 r28#define PREA r29#define PREB r30#define PREC r31#ifndef NEEDPARAM#ifndef DOUBLE#include "../sparam.h"#else#include "../dparam.h"#endif PROLOGUE PROFCODE addi SP, SP, -STACKSIZE li r0, 0 stfd f14, 0(SP) stfd f15, 8(SP) stfd f16, 16(SP) stfd f17, 24(SP) stfd f18, 32(SP) stfd f19, 40(SP) stfd f20, 48(SP) stfd f21, 56(SP) stfd f22, 64(SP) stfd f23, 72(SP) stfd f24, 80(SP) stfd f25, 88(SP) stfd f26, 96(SP) stfd f27, 104(SP) stfd f28, 112(SP) stfd f29, 120(SP) stfd f30, 128(SP) stfd f31, 136(SP)#ifdef __64BIT__ std r31, 144(SP) std r30, 152(SP) std r29, 160(SP) std r28, 168(SP) std r27, 176(SP) std r26, 184(SP) std r25, 192(SP) std r24, 200(SP) std r23, 208(SP) std r22, 216(SP) std r21, 224(SP)#else stw r31, 144(SP) stw r30, 148(SP) stw r29, 152(SP) stw r28, 156(SP) stw r27, 160(SP) stw r26, 164(SP) stw r25, 168(SP) stw r24, 172(SP) stw r23, 176(SP) stw r22, 180(SP) stw r21, 184(SP)#endif stfd f1, ALPHA stw r0, FZERO#if defined(_AIX) || defined(__APPLE__)#if !defined(__64BIT__) && defined(DOUBLE) lwz LDC, 56 + STACKSIZE(SP)#endif#endif slwi LDC, LDC, BASE_SHIFT cmpwi cr0, M, 0 ble LL(999) cmpwi cr0, N, 0 ble LL(999) cmpwi cr0, K, 0 ble LL(999)#ifndef PREFETCHTEST li PREA, (16 * 5 * SIZE + 16) li PREB, (16 * 5 * SIZE + 16) li PREC, 4 * SIZE#else#ifdef linux#ifndef __64BIT__ mr PREA, r10 lwz PREB, 8 + STACKSIZE(SP) lwz PREC, 12 + STACKSIZE(SP)#else ld PREA, 112 + STACKSIZE(SP) ld PREB, 120 + STACKSIZE(SP) ld PREC, 128 + STACKSIZE(SP)#endif#endif#if defined(_AIX) || defined(__APPLE__)#ifdef __64BIT__ ld PREA, 112 + STACKSIZE(SP) ld PREB, 120 + STACKSIZE(SP) ld PREC, 128 + STACKSIZE(SP)#else#ifdef DOUBLE lwz PREA, 60 + STACKSIZE(SP) lwz PREB, 64 + STACKSIZE(SP) lwz PREC, 68 + STACKSIZE(SP)#else lwz PREA, 56 + STACKSIZE(SP) lwz PREB, 60 + STACKSIZE(SP) lwz PREC, 64 + STACKSIZE(SP)#endif#endif#endif#endif lfs f0, FZERO srawi. J, N, 2 ble LL(40) .align 4LL(10): mr CO1, C add CO2, C, LDC add CO3, CO2, LDC add CO4, CO3, LDC fmr f1, f0 fmr f2, f0 fmr f3, f0 fmr f4, f0 fmr f5, f0 fmr f6, f0 fmr f7, f0 fmr f8, f0 fmr f9, f0 fmr f10, f0 fmr f11, f0 fmr f12, f0 fmr f13, f0 fmr f14, f0 fmr f15, f0 srawi. I, M, 2 mr AO, A add C, CO4, LDC ble LL(20) .align 4LL(11): LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) LFD f18, 2 * SIZE(AO) LFD f19, 3 * SIZE(AO) LFD f20, 0 * SIZE(B) LFD f21, 1 * SIZE(B) LFD f22, 2 * SIZE(B) LFD f23, 3 * SIZE(B) LFD f24, 4 * SIZE(AO) LFD f25, 5 * SIZE(AO) LFD f26, 6 * SIZE(AO) LFD f27, 7 * SIZE(AO)#if 0 PREFETCH_C1 PREFETCH_C2 PREFETCH_C3 PREFETCH_C4#endif srawi. r0, K, 2 mtspr CTR, r0 mr BO, B ble LL(15) .align 4LL(12): fmadd f0, f16, f20, f0 fmadd f4, f16, f21, f4 LFD f28, 4 * SIZE(BO) fmadd f8, f16, f22, f8 fmadd f12, f16, f23, f12 LFD f16, 8 * SIZE(AO) fmadd f1, f17, f20, f1 fmadd f5, f17, f21, f5 LFD f29, 5 * SIZE(BO) fmadd f9, f17, f22, f9 fmadd f13, f17, f23, f13 LFD f17, 9 * SIZE(AO) fmadd f2, f18, f20, f2 fmadd f6, f18, f21, f6 LFD f30, 6 * SIZE(BO) fmadd f10, f18, f22, f10 fmadd f14, f18, f23, f14 LFD f18, 10 * SIZE(AO) fmadd f3, f19, f20, f3 fmadd f7, f19, f21, f7 LFD f31, 7 * SIZE(BO) fmadd f11, f19, f22, f11 fmadd f15, f19, f23, f15 LFD f19, 11 * SIZE(AO) fmadd f0, f24, f28, f0 fmadd f4, f24, f29, f4 LFD f20, 8 * SIZE(BO) fmadd f8, f24, f30, f8 fmadd f12, f24, f31, f12 LFD f24, 12 * SIZE(AO) fmadd f1, f25, f28, f1 fmadd f5, f25, f29, f5 LFD f21, 9 * SIZE(BO) fmadd f9, f25, f30, f9 fmadd f13, f25, f31, f13 LFD f25, 13 * SIZE(AO) fmadd f2, f26, f28, f2 fmadd f6, f26, f29, f6 LFD f22, 10 * SIZE(BO) fmadd f10, f26, f30, f10 fmadd f14, f26, f31, f14 LFD f26, 14 * SIZE(AO) fmadd f3, f27, f28, f3 fmadd f7, f27, f29, f7 LFD f23, 11 * SIZE(BO) fmadd f11, f27, f30, f11 fmadd f15, f27, f31, f15 LFD f27, 15 * SIZE(AO) fmadd f0, f16, f20, f0 fmadd f4, f16, f21, f4 LFD f28, 12 * SIZE(BO) fmadd f8, f16, f22, f8 fmadd f12, f16, f23, f12 LFDU f16, 16 * SIZE(AO) fmadd f1, f17, f20, f1 fmadd f5, f17, f21, f5 LFD f29, 13 * SIZE(BO) fmadd f9, f17, f22, f9 fmadd f13, f17, f23, f13 LFD f17, 1 * SIZE(AO) fmadd f2, f18, f20, f2 fmadd f6, f18, f21, f6 LFD f30, 14 * SIZE(BO) fmadd f10, f18, f22, f10 fmadd f14, f18, f23, f14 LFD f18, 2 * SIZE(AO) fmadd f3, f19, f20, f3 fmadd f7, f19, f21, f7 LFD f31, 15 * SIZE(BO) fmadd f11, f19, f22, f11 fmadd f15, f19, f23, f15 LFD f19, 3 * SIZE(AO) fmadd f0, f24, f28, f0 fmadd f4, f24, f29, f4 LFDU f20, 16 * SIZE(BO) fmadd f8, f24, f30, f8 fmadd f12, f24, f31, f12 LFD f24, 4 * SIZE(AO) fmadd f1, f25, f28, f1 fmadd f5, f25, f29, f5 LFD f21, 1 * SIZE(BO) fmadd f9, f25, f30, f9 fmadd f13, f25, f31, f13 LFD f25, 5 * SIZE(AO) fmadd f2, f26, f28, f2 fmadd f6, f26, f29, f6 LFD f22, 2 * SIZE(BO) fmadd f10, f26, f30, f10 fmadd f14, f26, f31, f14 LFD f26, 6 * SIZE(AO) fmadd f3, f27, f28, f3 fmadd f7, f27, f29, f7 LFD f23, 3 * SIZE(BO) fmadd f11, f27, f30, f11 fmadd f15, f27, f31, f15 LFD f27, 7 * SIZE(AO) bdnz LL(12) .align 4LL(15): andi. r0, K, 3 lfd f30, ALPHA lfs f31, FZERO mtspr CTR, r0 ble+ LL(18) .align 4LL(16): fmadd f0, f16, f20, f0 fmadd f4, f16, f21, f4 fmadd f8, f16, f22, f8 fmadd f12, f16, f23, f12 LFD f16, 4 * SIZE(AO) fmadd f1, f17, f20, f1 fmadd f5, f17, f21, f5 fmadd f9, f17, f22, f9 fmadd f13, f17, f23, f13 LFD f17, 5 * SIZE(AO) fmadd f2, f18, f20, f2 fmadd f6, f18, f21, f6 fmadd f10, f18, f22, f10 fmadd f14, f18, f23, f14 LFD f18, 6 * SIZE(AO) fmadd f3, f19, f20, f3 LFD f20, 4 * SIZE(BO) fmadd f7, f19, f21, f7 LFD f21, 5 * SIZE(BO) fmadd f11, f19, f22, f11 LFD f22, 6 * SIZE(BO) fmadd f15, f19, f23, f15 LFD f19, 7 * SIZE(AO) LFD f23, 7 * SIZE(BO) addi BO, BO, 4 * SIZE addi AO, AO, 4 * SIZE bdnz LL(16) .align 4LL(18): LFD f16, 0 * SIZE(CO1) LFD f17, 1 * SIZE(CO1) LFD f18, 2 * SIZE(CO1) LFD f19, 3 * SIZE(CO1) LFD f20, 0 * SIZE(CO2) LFD f21, 1 * SIZE(CO2) LFD f22, 2 * SIZE(CO2) LFD f23, 3 * SIZE(CO2) fmadd f0, f0, f30, f16 LFD f16, 0 * SIZE(CO3) fmadd f1, f1, f30, f17 LFD f17, 1 * SIZE(CO3) fmadd f2, f2, f30, f18 LFD f18, 2 * SIZE(CO3) fmadd f3, f3, f30, f19 LFD f19, 3 * SIZE(CO3) fmadd f4, f4, f30, f20 LFD f20, 0 * SIZE(CO4) fmadd f5, f5, f30, f21 LFD f21, 1 * SIZE(CO4) fmadd f6, f6, f30, f22 LFD f22, 2 * SIZE(CO4) fmadd f7, f7, f30, f23 LFD f23, 3 * SIZE(CO4) fmadd f8, f8, f30, f16 fmadd f9, f9, f30, f17 STFD f0, 0 * SIZE(CO1) fmadd f10, f10, f30, f18 fmadd f11, f11, f30, f19 STFD f1, 1 * SIZE(CO1) fmadd f12, f12, f30, f20 fmadd f13, f13, f30, f21 STFD f2, 2 * SIZE(CO1) fmadd f14, f14, f30, f22 fmadd f15, f15, f30, f23 STFD f3, 3 * SIZE(CO1) STFD f4, 0 * SIZE(CO2) fmr f0, f31 fmr f1, f31 STFD f5, 1 * SIZE(CO2) fmr f2, f31 fmr f3, f31 STFD f6, 2 * SIZE(CO2) fmr f4, f31 fmr f5, f31 STFD f7, 3 * SIZE(CO2) fmr f6, f31 fmr f7, f31 STFD f8, 0 * SIZE(CO3) STFD f9, 1 * SIZE(CO3) addi CO1, CO1, 4 * SIZE fmr f8, f31 fmr f9, f31 STFD f10, 2 * SIZE(CO3) STFD f11, 3 * SIZE(CO3) addi CO2, CO2, 4 * SIZE fmr f10, f31 fmr f11, f31 STFD f12, 0 * SIZE(CO4) STFD f13, 1 * SIZE(CO4) addi CO3, CO3, 4 * SIZE fmr f12, f31 fmr f13, f31 STFD f14, 2 * SIZE(CO4) STFD f15, 3 * SIZE(CO4) addi CO4, CO4, 4 * SIZE fmr f14, f31 fmr f15, f31 addic. I, I, -1 bgt+ LL(11) .align 4LL(20): andi. I, M, 2 ble LL(30) LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) LFD f18, 2 * SIZE(AO) LFD f19, 3 * SIZE(AO) LFD f20, 0 * SIZE(B) LFD f21, 1 * SIZE(B) LFD f22, 2 * SIZE(B) LFD f23, 3 * SIZE(B) LFD f24, 4 * SIZE(B) LFD f25, 5 * SIZE(B) LFD f26, 6 * SIZE(B) LFD f27, 7 * SIZE(B) srawi. r0, K, 2 mtspr CTR, r0 mr BO, B ble LL(25) .align 5LL(22): fmadd f0, f16, f20, f0 fmadd f1, f17, f20, f1 fmadd f4, f16, f21, f4 fmadd f5, f17, f21, f5 fmadd f8, f16, f22, f8 fmadd f9, f17, f22, f9 fmadd f12, f16, f23, f12 fmadd f13, f17, f23, f13 LFD f20, 8 * SIZE(BO) LFD f21, 9 * SIZE(BO) LFD f22, 10 * SIZE(BO) LFD f23, 11 * SIZE(BO) fmadd f2, f18, f24, f2 fmadd f3, f19, f24, f3 fmadd f6, f18, f25, f6 fmadd f7, f19, f25, f7 fmadd f10, f18, f26, f10 fmadd f11, f19, f26, f11 fmadd f14, f18, f27, f14 fmadd f15, f19, f27, f15 LFD f16, 4 * SIZE(AO) LFD f17, 5 * SIZE(AO) LFD f18, 6 * SIZE(AO) LFD f19, 7 * SIZE(AO) fmadd f0, f16, f20, f0 fmadd f1, f17, f20, f1 fmadd f4, f16, f21, f4 fmadd f5, f17, f21, f5 LFD f24, 12 * SIZE(BO) LFD f25, 13 * SIZE(BO) LFD f26, 14 * SIZE(BO) LFD f27, 15 * SIZE(BO) fmadd f8, f16, f22, f8 fmadd f9, f17, f22, f9 fmadd f12, f16, f23, f12 fmadd f13, f17, f23, f13 LFD f20, 16 * SIZE(BO) LFD f21, 17 * SIZE(BO) LFD f22, 18 * SIZE(BO) LFD f23, 19 * SIZE(BO) fmadd f2, f18, f24, f2 fmadd f3, f19, f24, f3 fmadd f6, f18, f25, f6 fmadd f7, f19, f25, f7 fmadd f10, f18, f26, f10 fmadd f11, f19, f26, f11 fmadd f14, f18, f27, f14 fmadd f15, f19, f27, f15 LFD f16, 8 * SIZE(AO) LFD f17, 9 * SIZE(AO) LFD f18, 10 * SIZE(AO) LFD f19, 11 * SIZE(AO) LFD f24, 20 * SIZE(BO) LFD f25, 21 * SIZE(BO) LFD f26, 22 * SIZE(BO) LFD f27, 23 * SIZE(BO) addi AO, AO, 8 * SIZE addi BO, BO, 16 * SIZE PREFETCH_B bdnz LL(22) fadd f0, f2, f0 fadd f1, f3, f1 fadd f4, f6, f4 fadd f5, f7, f5 fadd f8, f10, f8 fadd f9, f11, f9 fadd f12, f14, f12 fadd f13, f15, f13 .align 4LL(25): lfd f30, ALPHA andi. r0, K, 3 mtspr CTR, r0 ble+ LL(28) .align 4LL(26): fmadd f0, f16, f20, f0 fmadd f1, f17, f20, f1 fmadd f4, f16, f21, f4 fmadd f5, f17, f21, f5 fmadd f8, f16, f22, f8 fmadd f9, f17, f22, f9 fmadd f12, f16, f23, f12 fmadd f13, f17, f23, f13 LFD f16, 2 * SIZE(AO) LFD f17, 3 * SIZE(AO) LFD f20, 4 * SIZE(BO) LFD f21, 5 * SIZE(BO) LFD f22, 6 * SIZE(BO) LFD f23, 7 * SIZE(BO) addi BO, BO, 4 * SIZE addi AO, AO, 2 * SIZE bdnz LL(26) .align 4LL(28): LFD f16, 0 * SIZE(CO1) LFD f17, 1 * SIZE(CO1) LFD f18, 0 * SIZE(CO2) LFD f19, 1 * SIZE(CO2) fmadd f0, f0, f30, f16 fmadd f1, f1, f30, f17 fmadd f4, f4, f30, f18 fmadd f5, f5, f30, f19 LFD f20, 0 * SIZE(CO3) LFD f21, 1 * SIZE(CO3) LFD f22, 0 * SIZE(CO4) LFD f23, 1 * SIZE(CO4) fmadd f8, f8, f30, f20 fmadd f9, f9, f30, f21 fmadd f12, f12, f30, f22 fmadd f13, f13, f30, f23 STFD f0, 0 * SIZE(CO1) STFD f1, 1 * SIZE(CO1) STFD f4, 0 * SIZE(CO2) STFD f5, 1 * SIZE(CO2) lfs f0, FZERO fmr f1, f0 fmr f2, f0 fmr f3, f0 STFD f8, 0 * SIZE(CO3) STFD f9, 1 * SIZE(CO3) STFD f12, 0 * SIZE(CO4) STFD f13, 1 * SIZE(CO4) fmr f4, f0 fmr f5, f0 fmr f6, f0 fmr f7, f0 fmr f8, f0 fmr f9, f0 fmr f10, f0 fmr f11, f0 fmr f12, f0 fmr f13, f0 fmr f14, f0 fmr f15, f0 addi CO1, CO1, 2 * SIZE addi CO2, CO2, 2 * SIZE addi CO3, CO3, 2 * SIZE addi CO4, CO4, 2 * SIZE .align 4LL(30): andi. I, M, 1 ble LL(39) LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) LFD f18, 2 * SIZE(AO) LFD f19, 3 * SIZE(AO) LFD f20, 0 * SIZE(B) LFD f21, 1 * SIZE(B) LFD f22, 2 * SIZE(B) LFD f23, 3 * SIZE(B) LFD f24, 4 * SIZE(B) LFD f25, 5 * SIZE(B) LFD f26, 6 * SIZE(B) LFD f27, 7 * SIZE(B) srawi. r0, K, 2 mtspr CTR, r0 mr BO, B ble LL(35) .align 5LL(32): fmadd f0, f16, f20, f0 fmadd f4, f16, f21, f4 fmadd f8, f16, f22, f8 fmadd f12, f16, f23, f12 LFD f20, 8 * SIZE(BO) LFD f21, 9 * SIZE(BO) LFD f22, 10 * SIZE(BO) LFD f23, 11 * SIZE(BO) fmadd f1, f17, f24, f1 fmadd f5, f17, f25, f5 fmadd f9, f17, f26, f9 fmadd f13, f17, f27, f13 LFD f24, 12 * SIZE(BO) LFD f25, 13 * SIZE(BO) LFD f26, 14 * SIZE(BO) LFD f27, 15 * SIZE(BO) fmadd f0, f18, f20, f0 fmadd f4, f18, f21, f4 fmadd f8, f18, f22, f8 fmadd f12, f18, f23, f12 LFD f20, 16 * SIZE(BO) LFD f21, 17 * SIZE(BO) LFD f22, 18 * SIZE(BO) LFD f23, 19 * SIZE(BO) fmadd f1, f19, f24, f1 fmadd f5, f19, f25, f5 fmadd f9, f19, f26, f9 fmadd f13, f19, f27, f13 LFD f16, 4 * SIZE(AO) LFD f17, 5 * SIZE(AO) LFD f18, 6 * SIZE(AO) LFD f19, 7 * SIZE(AO) LFD f24, 20 * SIZE(BO) LFD f25, 21 * SIZE(BO) LFD f26, 22 * SIZE(BO) LFD f27, 23 * SIZE(BO) addi AO, AO, 4 * SIZE addi BO, BO, 16 * SIZE PREFETCH_B bdnz LL(32) fadd f0, f1, f0 fadd f4, f5, f4 fadd f8, f9, f8 fadd f12, f13, f12 .align 4LL(35): lfd f30, ALPHA andi. r0, K, 3 mtspr CTR, r0 ble+ LL(38) .align 4LL(36): fmadd f0, f16, f20, f0 fmadd f4, f16, f21, f4 fmadd f8, f16, f22, f8 fmadd f12, f16, f23, f12 LFD f16, 1 * SIZE(AO) LFD f20, 4 * SIZE(BO) LFD f21, 5 * SIZE(BO) LFD f22, 6 * SIZE(BO) LFD f23, 7 * SIZE(BO) addi BO, BO, 4 * SIZE addi AO, AO, 1 * SIZE bdnz LL(36) .align 4LL(38): LFD f16, 0 * SIZE(CO1) LFD f18, 0 * SIZE(CO2) LFD f20, 0 * SIZE(CO3) LFD f22, 0 * SIZE(CO4)
⌨️ 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 + -