📄 trsm_kernel_ppc440_lt.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#define OFFSET r10#else#define A r7#define B r8#define C r9#define LDC r10#define OFFSET r6#endif#endif#if defined(_AIX) || defined(__APPLE__)#if !defined(__64BIT__) && defined(DOUBLE)#define A r8#define B r9#define C r10#define LDC r7#define OFFSET r6#else#define A r7#define B r8#define C r9#define LDC r10#define OFFSET r6#endif#endif#define AORIG r18#define TEMP r19#define KK r20#define I r21#define J r22#define AO r23#define BO r24#define CO1 r25#define CO2 r26#define CO3 r27#define CO4 r28#ifndef NEEDPARAM 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) std r20, 232(SP) std r19, 240(SP) std r18, 248(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) stw r20, 188(SP) stw r19, 192(SP) stw r18, 196(SP)#endif 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#if defined(linux) && defined(__64BIT__) ld OFFSET, 112 + STACKSIZE(SP)#endif#if defined(_AIX) || defined(__APPLE__)#ifdef __64BIT__ ld OFFSET, 112 + STACKSIZE(SP)#else#ifdef DOUBLE lwz OFFSET, 60 + STACKSIZE(SP)#else lwz OFFSET, 56 + STACKSIZE(SP)#endif#endif#endif#define A1 f16#define A2 f17#define A3 f18#define A4 f19#define A5 f20#define A6 f21#define B1 f22#define B2 f23#define B3 f24#define B4 f25#define B5 f26#define B6 f27#define B7 f28#define B8 f29#define B9 f30#define B10 f31#ifdef LN mullw r0, M, K slwi r0, r0, BASE_SHIFT add A, A, r0 slwi r0, M, BASE_SHIFT add C, C, r0#endif#ifdef RN neg KK, OFFSET#endif#ifdef RT mullw r0, N, K slwi r0, r0, BASE_SHIFT add B, B, r0 mullw r0, N, LDC add C, C, r0 sub KK, N, OFFSET#endif cmpwi cr0, M, 0 ble .L999 cmpwi cr0, N, 0 ble .L999 cmpwi cr0, K, 0 ble .L999 lfs f0, FZERO srawi. J, N, 2 ble .L40 .align 4.L10:#ifdef RT slwi r0, K, 2 + BASE_SHIFT sub B, B, r0 slwi r0, LDC, 2 sub C, C, r0#endif mr CO1, C add CO2, C, LDC add CO3, CO2, LDC add CO4, CO3, LDC#ifdef LN add KK, M, OFFSET#endif#ifdef LT mr KK, OFFSET#endif 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#if defined(LN) || defined(RT) mr AORIG, A#else mr AO, A#endif#ifndef RT add C, CO4, LDC#endif ble .L20 .align 4.L11:#if defined(LT) || defined(RN) LFD A1, 0 * SIZE(AO) LFD A2, 1 * SIZE(AO) LFD A4, 4 * SIZE(AO) LFD A5, 8 * SIZE(AO) LFD B1, 0 * SIZE(B) LFD B2, 1 * SIZE(B) LFD B3, 2 * SIZE(B) LFD B4, 3 * SIZE(B) LFD B5, 4 * SIZE(B) LFD B6, 8 * SIZE(B) LFD B7, 12 * SIZE(B) srawi. r0, KK, 2 mtspr CTR, r0 mr BO, B#else#ifdef LN slwi r0, K, 2 + BASE_SHIFT sub AORIG, AORIG, r0#endif slwi TEMP, KK, 2 + BASE_SHIFT add AO, AORIG, TEMP add BO, B, TEMP sub TEMP, K, KK LFD A1, 0 * SIZE(AO) LFD A2, 1 * SIZE(AO) LFD A4, 4 * SIZE(AO) LFD A5, 8 * SIZE(AO) LFD B1, 0 * SIZE(BO) LFD B2, 1 * SIZE(BO) LFD B3, 2 * SIZE(BO) LFD B4, 3 * SIZE(BO) LFD B5, 4 * SIZE(BO) LFD B6, 8 * SIZE(BO) LFD B7, 12 * SIZE(BO) srawi. r0, TEMP, 2 mtspr CTR, r0#endif ble .L15 .align 4.L12: FMADD f0, A1, B1, f0 LFD A3, 2 * SIZE(AO) FMADD f4, A1, B2, f4 LFD A6, 12 * SIZE(AO) FMADD f8, A1, B3, f8 nop FMADD f12, A1, B4, f12 nop FMADD f1, A2, B1, f1 LFD A1, 3 * SIZE(AO) FMADD f5, A2, B2, f5 nop FMADD f9, A2, B3, f9 nop FMADD f13, A2, B4, f13 nop FMADD f2, A3, B1, f2 nop FMADD f6, A3, B2, f6 LFD B8, 5 * SIZE(BO) FMADD f10, A3, B3, f10 LFD B9, 6 * SIZE(BO) FMADD f14, A3, B4, f14 LFD B10, 7 * SIZE(BO) FMADD f3, A1, B1, f3 LFD A2, 5 * SIZE(AO) FMADD f7, A1, B2, f7 LFD B1, 16 * SIZE(BO) FMADD f11, A1, B3, f11 nop FMADD f15, A1, B4, f15 nop FMADD f0, A4, B5, f0 LFD A3, 6 * SIZE(AO) FMADD f4, A4, B8, f4 LFD A1, 16 * SIZE(AO) FMADD f8, A4, B9, f8 nop FMADD f12, A4, B10, f12 nop FMADD f1, A2, B5, f1 LFD A4, 7 * SIZE(AO) FMADD f5, A2, B8, f5 nop FMADD f9, A2, B9, f9 nop FMADD f13, A2, B10, f13 nop FMADD f2, A3, B5, f2 nop FMADD f6, A3, B8, f6 LFD B2, 9 * SIZE(BO) FMADD f10, A3, B9, f10 LFD B3, 10 * SIZE(BO) FMADD f14, A3, B10, f14 LFD B4, 11 * SIZE(BO) FMADD f3, A4, B5, f3 LFD A2, 9 * SIZE(AO) FMADD f7, A4, B8, f7 LFD B5, 20 * SIZE(BO) FMADD f11, A4, B9, f11 nop FMADD f15, A4, B10, f15 nop FMADD f0, A5, B6, f0 LFD A3, 10 * SIZE(AO) FMADD f4, A5, B2, f4 LFD A4, 20 * SIZE(AO) FMADD f8, A5, B3, f8 nop FMADD f12, A5, B4, f12 nop FMADD f1, A2, B6, f1 LFD A5, 11 * SIZE(AO) FMADD f5, A2, B2, f5 nop FMADD f9, A2, B3, f9 nop FMADD f13, A2, B4, f13 nop FMADD f2, A3, B6, f2 nop FMADD f6, A3, B2, f6 LFD B8, 13 * SIZE(BO) FMADD f10, A3, B3, f10 LFD B9, 14 * SIZE(BO) FMADD f14, A3, B4, f14 LFD B10,15 * SIZE(BO) FMADD f3, A5, B6, f3 LFD A2, 13 * SIZE(AO) FMADD f7, A5, B2, f7 LFD B6, 24 * SIZE(BO) FMADD f11, A5, B3, f11 nop FMADD f15, A5, B4, f15 nop FMADD f0, A6, B7, f0 LFD A3, 14 * SIZE(AO) FMADD f4, A6, B8, f4 LFD A5, 24 * SIZE(AO) FMADD f8, A6, B9, f8 nop FMADD f12, A6, B10, f12 nop FMADD f1, A2, B7, f1 LFD A6, 15 * SIZE(AO) FMADD f5, A2, B8, f5 nop FMADD f9, A2, B9, f9 nop FMADD f13, A2, B10, f13 nop FMADD f2, A3, B7, f2 addi AO, AO, 16 * SIZE FMADD f6, A3, B8, f6 LFD B2, 17 * SIZE(BO) FMADD f10, A3, B9, f10 LFD B3, 18 * SIZE(BO) FMADD f14, A3, B10, f14 LFD B4, 19 * SIZE(BO) FMADD f3, A6, B7, f3 LFD A2, 1 * SIZE(AO) FMADD f7, A6, B8, f7 LFD B7, 28 * SIZE(BO) FMADD f11, A6, B9, f11 addi BO, BO, 16 * SIZE FMADD f15, A6, B10, f15 bdnz .L12 .align 4.L15:#if defined(LT) || defined(RN) andi. r0, KK, 3#else andi. r0, TEMP, 3#endif mtspr CTR, r0 ble+ .L18 .align 4.L16: FMADD f0, A1, B1, f0 LFD A3, 2 * SIZE(AO) FMADD f4, A1, B2, f4 FMADD f8, A1, B3, f8 FMADD f12, A1, B4, f12 LFD A4, 3 * SIZE(AO) FMADD f1, A2, B1, f1 FMADD f5, A2, B2, f5 FMADD f9, A2, B3, f9 FMADD f13, A2, B4, f13 LFDU A1, 4 * SIZE(AO) FMADD f2, A3, B1, f2 FMADD f6, A3, B2, f6 FMADD f10, A3, B3, f10 FMADD f14, A3, B4, f14 LFD A2, 1 * SIZE(AO) FMADD f3, A4, B1, f3 LFDU B1, 4 * SIZE(BO) FMADD f7, A4, B2, f7 LFD B2, 1 * SIZE(BO) FMADD f11, A4, B3, f11 LFD B3, 2 * SIZE(BO) FMADD f15, A4, B4, f15 LFD B4, 3 * SIZE(BO) bdnz .L16 .align 4.L18:#if defined(LN) || defined(RT) subi r0, KK, 4 slwi r0, r0, 2 + BASE_SHIFT add AO, AORIG, r0 add BO, B, r0#endif#if defined(LN) || defined(LT) LFD f16, 0 * SIZE(BO) LFD f17, 1 * SIZE(BO) LFD f18, 2 * SIZE(BO) LFD f19, 3 * SIZE(BO) LFD f20, 4 * SIZE(BO) LFD f21, 5 * SIZE(BO) LFD f22, 6 * SIZE(BO) LFD f23, 7 * SIZE(BO) LFD f24, 8 * SIZE(BO) LFD f25, 9 * SIZE(BO) LFD f26, 10 * SIZE(BO) LFD f27, 11 * SIZE(BO) LFD f28, 12 * SIZE(BO) LFD f29, 13 * SIZE(BO) LFD f30, 14 * SIZE(BO) LFD f31, 15 * SIZE(BO) FSUB f0, f16, f0 FSUB f4, f17, f4 FSUB f8, f18, f8 FSUB f12, f19, f12 FSUB f1, f20, f1 FSUB f5, f21, f5 FSUB f9, f22, f9 FSUB f13, f23, f13 FSUB f2, f24, f2 FSUB f6, f25, f6 FSUB f10, f26, f10 FSUB f14, f27, f14 FSUB f3, f28, f3 FSUB f7, f29, f7 FSUB f11, f30, f11 FSUB f15, f31, f15#else LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) LFD f18, 2 * SIZE(AO) LFD f19, 3 * SIZE(AO) LFD f20, 4 * SIZE(AO) LFD f21, 5 * SIZE(AO) LFD f22, 6 * SIZE(AO) LFD f23, 7 * SIZE(AO) LFD f24, 8 * SIZE(AO) LFD f25, 9 * SIZE(AO) LFD f26, 10 * SIZE(AO) LFD f27, 11 * SIZE(AO) LFD f28, 12 * SIZE(AO) LFD f29, 13 * SIZE(AO) LFD f30, 14 * SIZE(AO) LFD f31, 15 * SIZE(AO) FSUB f0, f16, f0 FSUB f1, f17, f1 FSUB f2, f18, f2 FSUB f3, f19, f3 FSUB f4, f20, f4 FSUB f5, f21, f5 FSUB f6, f22, f6 FSUB f7, f23, f7 FSUB f8, f24, f8 FSUB f9, f25, f9 FSUB f10, f26, f10 FSUB f11, f27, f11 FSUB f12, f28, f12 FSUB f13, f29, f13 FSUB f14, f30, f14 FSUB f15, f31, f15#endif#ifdef LN LFD f16, 15 * SIZE(AO) LFD f17, 14 * SIZE(AO) LFD f18, 13 * SIZE(AO) LFD f19, 12 * SIZE(AO) FMUL f3, f16, f3 FMUL f7, f16, f7 FMUL f11, f16, f11 FMUL f15, f16, f15 FNMSUB f2, f17, f3, f2 FNMSUB f6, f17, f7, f6 FNMSUB f10, f17, f11, f10 FNMSUB f14, f17, f15, f14 FNMSUB f1, f18, f3, f1 FNMSUB f5, f18, f7, f5 FNMSUB f9, f18, f11, f9 FNMSUB f13, f18, f15, f13 FNMSUB f0, f19, f3, f0 FNMSUB f4, f19, f7, f4 FNMSUB f8, f19, f11, f8 FNMSUB f12, f19, f15, f12 LFD f16, 10 * SIZE(AO) LFD f17, 9 * SIZE(AO) LFD f18, 8 * SIZE(AO) LFD f19, 5 * SIZE(AO) FMUL f2, f16, f2 FMUL f6, f16, f6 FMUL f10, f16, f10 FMUL f14, f16, f14 LFD f20, 4 * SIZE(AO) LFD f21, 0 * SIZE(AO) FNMSUB f1, f17, f2, f1 FNMSUB f5, f17, f6, f5 FNMSUB f9, f17, f10, f9 FNMSUB f13, f17, f14, f13 FNMSUB f0, f18, f2, f0 FNMSUB f4, f18, f6, f4 FNMSUB f8, f18, f10, f8 FNMSUB f12, f18, f14, f12 FMUL f1, f19, f1 FMUL f5, f19, f5 FMUL f9, f19, f9 FMUL f13, f19, f13 FNMSUB f0, f20, f1, f0 FNMSUB f4, f20, f5, f4 FNMSUB f8, f20, f9, f8 FNMSUB f12, f20, f13, f12 FMUL f0, f21, f0 FMUL f4, f21, f4 FMUL f8, f21, f8 FMUL f12, f21, f12#endif#ifdef LT LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) LFD f18, 2 * SIZE(AO) LFD f19, 3 * SIZE(AO) FMUL f0, f16, f0 FMUL f4, f16, f4 FMUL f8, f16, f8 FMUL f12, f16, f12 FNMSUB f1, f17, f0, f1 FNMSUB f5, f17, f4, f5 FNMSUB f9, f17, f8, f9 FNMSUB f13, f17, f12, f13 FNMSUB f2, f18, f0, f2 FNMSUB f6, f18, f4, f6 FNMSUB f10, f18, f8, f10 FNMSUB f14, f18, f12, f14 FNMSUB f3, f19, f0, f3 FNMSUB f7, f19, f4, f7 FNMSUB f11, f19, f8, f11 FNMSUB f15, f19, f12, f15 LFD f16, 5 * SIZE(AO) LFD f17, 6 * SIZE(AO) LFD f18, 7 * SIZE(AO) LFD f19, 10 * SIZE(AO) FMUL f1, f16, f1 FMUL f5, f16, f5 FMUL f9, f16, f9 FMUL f13, f16, f13 LFD f20, 11 * SIZE(AO) LFD f21, 15 * SIZE(AO) FNMSUB f2, f17, f1, f2 FNMSUB f6, f17, f5, f6 FNMSUB f10, f17, f9, f10 FNMSUB f14, f17, f13, f14 FNMSUB f3, f18, f1, f3 FNMSUB f7, f18, f5, f7 FNMSUB f11, f18, f9, f11 FNMSUB f15, f18, f13, f15 FMUL f2, f19, f2 FMUL f6, f19, f6 FMUL f10, f19, f10 FMUL f14, f19, f14 FNMSUB f3, f20, f2, f3 FNMSUB f7, f20, f6, f7 FNMSUB f11, f20, f10, f11 FNMSUB f15, f20, f14, f15 FMUL f3, f21, f3 FMUL f7, f21, f7 FMUL f11, f21, f11 FMUL f15, f21, f15#endif#ifdef RN LFD f16, 0 * SIZE(BO) LFD f17, 1 * SIZE(BO) LFD f18, 2 * SIZE(BO) LFD f19, 3 * SIZE(BO) FMUL f0, f16, f0 FMUL f1, f16, f1 FMUL f2, f16, f2 FMUL f3, f16, f3 FNMSUB f4, f17, f0, f4 FNMSUB f5, f17, f1, f5 FNMSUB f6, f17, f2, f6 FNMSUB f7, f17, f3, f7 FNMSUB f8, f18, f0, f8 FNMSUB f9, f18, f1, f9 FNMSUB f10, f18, f2, f10 FNMSUB f11, f18, f3, f11 FNMSUB f12, f19, f0, f12 FNMSUB f13, f19, f1, f13 FNMSUB f14, f19, f2, f14 FNMSUB f15, f19, f3, f15 LFD f16, 5 * SIZE(BO) LFD f17, 6 * SIZE(BO) LFD f18, 7 * SIZE(BO) LFD f19, 10 * SIZE(BO) FMUL f4, f16, f4 FMUL f5, f16, f5 FMUL f6, f16, f6 FMUL f7, f16, f7 LFD f20, 11 * SIZE(BO) LFD f21, 15 * SIZE(BO) FNMSUB f8, f17, f4, f8 FNMSUB f9, f17, f5, f9 FNMSUB f10, f17, f6, f10 FNMSUB f11, f17, f7, f11 FNMSUB f12, f18, f4, f12 FNMSUB f13, f18, f5, f13 FNMSUB f14, f18, f6, f14 FNMSUB f15, f18, f7, f15 FMUL f8, f19, f8 FMUL f9, f19, f9 FMUL f10, f19, f10 FMUL f11, f19, f11 FNMSUB f12, f20, f8, f12 FNMSUB f13, f20, f9, f13 FNMSUB f14, f20, f10, f14 FNMSUB f15, f20, f11, f15 FMUL f12, f21, f12 FMUL f13, f21, f13 FMUL f14, f21, f14 FMUL f15, f21, f15#endif#ifdef RT LFD f16, 15 * SIZE(BO) LFD f17, 14 * SIZE(BO) LFD f18, 13 * SIZE(BO) LFD f19, 12 * SIZE(BO) FMUL f12, f16, f12 FMUL f13, f16, f13 FMUL f14, f16, f14 FMUL f15, f16, f15 FNMSUB f8, f17, f12, f8 FNMSUB f9, f17, f13, f9 FNMSUB f10, f17, f14, f10 FNMSUB f11, f17, f15, f11 FNMSUB f4, f18, f12, f4 FNMSUB f5, f18, f13, f5 FNMSUB f6, f18, f14, f6 FNMSUB f7, f18, f15, f7 FNMSUB f0, f19, f12, f0 FNMSUB f1, f19, f13, f1 FNMSUB f2, f19, f14, f2 FNMSUB f3, f19, f15, f3 LFD f16, 10 * SIZE(BO) LFD f17, 9 * SIZE(BO) LFD f18, 8 * SIZE(BO) LFD f19, 5 * SIZE(BO) FMUL f8, f16, f8 FMUL f9, f16, f9 FMUL f10, f16, f10 FMUL f11, f16, f11 LFD f20, 4 * SIZE(BO) LFD f21, 0 * SIZE(BO) FNMSUB f4, f17, f8, f4 FNMSUB f5, f17, f9, f5 FNMSUB f6, f17, f10, f6 FNMSUB f7, f17, f11, f7 FNMSUB f0, f18, f8, f0 FNMSUB f1, f18, f9, f1 FNMSUB f2, f18, f10, f2 FNMSUB f3, f18, f11, f3 FMUL f4, f19, f4 FMUL f5, f19, f5 FMUL f6, f19, f6 FMUL f7, f19, f7 FNMSUB f0, f20, f4, f0 FNMSUB f1, f20, f5, f1 FNMSUB f2, f20, f6, f2 FNMSUB f3, f20, f7, f3 FMUL f0, f21, f0 FMUL f1, f21, f1
⌨️ 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 + -