⭐ 欢迎来到虫虫下载站! | 📦 资源下载 📁 资源专辑 ℹ️ 关于我们
⭐ 虫虫下载站

📄 gemm_kernel_g4.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 3 页
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#define PREA	r29#define PREC	r30	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)#if defined(TRMMKERNEL)	std	r19,  240(SP)	std	r18,  248(SP)#endif#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)#if defined(TRMMKERNEL)	stw	r19,  192(SP)	stw	r18,  196(SP)#endif#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#if defined(TRMMKERNEL)#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#endif#if defined(TRMMKERNEL) && !defined(LEFT)	neg	KK, OFFSET#endif	li	PREA, 8 * 8 * SIZE	li	PREC, 3 * SIZE	cmpwi	cr0, M, 0	ble	.L999	cmpwi	cr0, N, 0	ble	.L999	cmpwi	cr0, K, 0	ble	.L999	srawi.	J, N,  2	ble	.L40	.align 4#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.L10:	mr	CO1, C	add	CO2, C,  LDC	add	CO3, CO2, LDC	add	CO4, CO3, LDC#if defined(TRMMKERNEL) &&  defined(LEFT)	mr	KK, OFFSET#endif	lfs	f0,  FZERO 	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	.L20	.align 4.L11:#if defined(TRMMKERNEL)#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))	LFD	A1,  0 * SIZE(AO)	LFD	A2,  1 * SIZE(AO)	LFD	A3,  2 * SIZE(AO)	LFDU	A5,  4 * SIZE(AO)	LFD	B1,  0 * SIZE(B)	LFD	B2,  1 * SIZE(B)	LFD	B3,  2 * SIZE(B)	LFD	B4,  3 * SIZE(B)	mr	BO,  B#else	slwi	r0, KK, 2 + BASE_SHIFT	add	AO, AO, r0	add	BO, B,  r0	LFD	A1,  0 * SIZE(AO)	LFD	A2,  1 * SIZE(AO)	LFD	A3,  2 * SIZE(AO)	LFDU	A5,  4 * SIZE(AO)	LFD	B1,  0 * SIZE(BO)	LFD	B2,  1 * SIZE(BO)	LFD	B3,  2 * SIZE(BO)	LFD	B4,  3 * SIZE(BO)#endif	dcbtst	CO1, PREC	dcbtst	CO2, PREC	dcbtst	CO3, PREC	dcbtst	CO4, PREC#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	sub	TEMP, K, KK#elif defined(LEFT)	addi	TEMP, KK, 4#else	addi	TEMP, KK, 4#endif	srawi.	TEMP,  TEMP,  1	mtspr	CTR, TEMP	ble	.L15#else	LFD	A1,  0 * SIZE(AO)	LFD	A2,  1 * SIZE(AO)	LFD	A3,  2 * SIZE(AO)	LFDU	A5,  4 * SIZE(AO)	LFD	B1,  0 * SIZE(B)	LFD	B2,  1 * SIZE(B)	LFD	B3,  2 * SIZE(B)	LFD	B4,  3 * SIZE(B)	dcbtst	CO1, PREC	dcbtst	CO2, PREC	dcbtst	CO3, PREC	dcbtst	CO4, PREC	srawi.	r0,  K,  1	mtspr	CTR, r0	mr	BO,  B	ble	.L15#endif	.align 4.L12:	FMADD	f0,  A1, B1, f0	dcbt	AO, PREA	FMADD	f4,  A1, B2, f4	LFDU	B5,  4 * SIZE(BO)	FMADD	f8,  A1, B3, f8	dcbt	BO, PREA	FMADD	f12, A1, B4, f12	LFD	A4, -1 * SIZE(AO)	FMADD	f1,  A2, B1, f1	nop	FMADD	f5,  A2, B2, f5	LFD	B6,  1 * SIZE(BO)	FMADD	f9,  A2, B3, f9	LFDU	A1,  4 * SIZE(AO)	FMADD	f13, A2, B4, f13	nop	FMADD	f2,  A3, B1, f2	nop	FMADD	f6,  A3, B2, f6	LFD	B7,  2 * SIZE(BO)	FMADD	f10, A3, B3, f10	LFD	A2, -3 * SIZE(AO)	FMADD	f14, A3, B4, f14	nop	FMADD	f3,  A4, B1, f3	nop	FMADD	f7,  A4, B2, f7	LFD	B8,  3 * SIZE(BO)	FMADD	f11, A4, B3, f11	LFD	A3, -2 * SIZE(AO)	FMADD	f15, A4, B4, f15	nop	FMADD	f0,  A5, B5, f0#ifdef DOUBLE	dcbt	AO, PREA#else	nop#endif	FMADD	f4,  A5, B6, f4	LFDU	B1,  4 * SIZE(BO)	FMADD	f8,  A5, B7, f8#ifdef DOUBLE	dcbt	BO, PREA#else	nop#endif	FMADD	f12, A5, B8, f12	LFD	A4, -1 * SIZE(AO)	FMADD	f1,  A2, B5, f1	nop	FMADD	f5,  A2, B6, f5	LFD	B2,  1 * SIZE(BO)	FMADD	f9,  A2, B7, f9	LFDU	A5,  4 * SIZE(AO)	FMADD	f13, A2, B8, f13	nop	FMADD	f2,  A3, B5, f2	nop	FMADD	f6,  A3, B6, f6	LFD	B3,  2 * SIZE(BO)	FMADD	f10, A3, B7, f10	LFD	A2, -3 * SIZE(AO)	FMADD	f14, A3, B8, f14	nop	FMADD	f3,  A4, B5, f3	nop	FMADD	f7,  A4, B6, f7	LFD	B4,  3 * SIZE(BO)	FMADD	f11, A4, B7, f11	LFD	A3, -2 * SIZE(AO)	FMADD	f15, A4, B8, f15	bdnz	.L12	.align 4.L15:	addi	AO, AO, -4 * SIZE	lfd	f30,  ALPHA#if defined(TRMMKERNEL)#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	sub	TEMP, K, KK#elif defined(LEFT)	addi	TEMP, KK, 4#else	addi	TEMP, KK, 4#endif	andi.	TEMP,  TEMP,  1#else	andi.	r0,  K,  1#endif	ble+	.L18.L16:	LFD	A4,  3 * SIZE(AO)	FMADD	f0,  A1, B1, f0	FMADD	f4,  A1, B2, f4	FMADD	f8,  A1, B3, f8	FMADD	f12, A1, B4, f12	FMADD	f1,  A2, B1, f1	FMADD	f5,  A2, B2, f5	FMADD	f9,  A2, B3, f9	FMADD	f13, A2, B4, f13	FMADD	f2,  A3, B1, f2	FMADD	f6,  A3, B2, f6	FMADD	f10, A3, B3, f10	FMADD	f14, A3, B4, f14	FMADD	f3,  A4, B1, f3	FMADD	f7,  A4, B2, f7	FMADD	f11, A4, B3, f11	FMADD	f15, A4, B4, f15	addi	AO, AO, 4 * SIZE	addi	BO, BO, 4 * SIZE		.align 4.L18:#ifndef TRMMKERNEL	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	FMADD	f10, f10, f30, f18	FMADD	f11, f11, f30, f19	FMADD	f12, f12, f30, f20	FMADD	f13, f13, f30, f21	FMADD	f14, f14, f30, f22	FMADD	f15, f15, f30, f23#else	FMUL	f0,  f0, f30	FMUL	f1,  f1, f30	FMUL	f2,  f2, f30	FMUL	f3,  f3, f30	FMUL	f4,  f4, f30	FMUL	f5,  f5, f30	FMUL	f6,  f6, f30	FMUL	f7,  f7, f30	FMUL	f8,  f8,  f30	FMUL	f9,  f9,  f30	FMUL	f10, f10, f30	FMUL	f11, f11, f30	FMUL	f12, f12, f30	FMUL	f13, f13, f30	FMUL	f14, f14, f30	FMUL	f15, f15, f30#endif	STFD	f0,  0 * SIZE(CO1)	STFD	f1,  1 * SIZE(CO1)	STFD	f2,  2 * SIZE(CO1)	STFD	f3,  3 * SIZE(CO1)	lfs	f0,  FZERO 	fmr	f1,  f0	fmr	f2,  f0	fmr	f3,  f0	STFD	f4,  0 * SIZE(CO2)	fmr	f4,  f0	STFD	f5,  1 * SIZE(CO2)	fmr	f5,  f0	STFD	f6,  2 * SIZE(CO2)	fmr	f6,  f0	STFD	f7,  3 * SIZE(CO2)	fmr	f7,  f0	STFD	f8,  0 * SIZE(CO3)	fmr	f8,  f0	STFD	f9,  1 * SIZE(CO3)	fmr	f9,  f0	STFD	f10, 2 * SIZE(CO3)	fmr	f10, f0	STFD	f11, 3 * SIZE(CO3)	fmr	f11, f0	STFD	f12, 0 * SIZE(CO4)	fmr	f12, f0	STFD	f13, 1 * SIZE(CO4)	fmr	f13, f0	STFD	f14, 2 * SIZE(CO4)	fmr	f14, f0	STFD	f15, 3 * SIZE(CO4)	fmr	f15, f0	addi	CO1, CO1, 4 * SIZE	addi	CO2, CO2, 4 * SIZE	addi	CO3, CO3, 4 * SIZE	addi	CO4, CO4, 4 * SIZE	#ifdef TRMMKERNEL#if ( defined(LEFT) &&  defined(TRANSA)) || \    (!defined(LEFT) && !defined(TRANSA))	sub	TEMP, K, KK#ifdef LEFT	addi	TEMP, TEMP, -4#else	addi	TEMP, TEMP, -4#endif	slwi	TEMP, TEMP, 2 + BASE_SHIFT	add	AO, AO, TEMP	add	BO, BO, TEMP#endif#ifdef LEFT	addi	KK, KK, 4#endif#endif	addic.	I, I, -1	bgt+	.L11	.align 4.L20:	andi.	I,  M,  2	ble	.L30#if defined(TRMMKERNEL)#if ( defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))	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)	mr	BO,  B#else	slwi	r0,   KK, 1 + BASE_SHIFT	slwi	TEMP, KK, 2 + BASE_SHIFT	add	AO, AO, r0	add	BO, B,  TEMP	LFD	f16,  0 * SIZE(AO)	LFD	f17,  1 * SIZE(AO)	LFD	f18,  2 * SIZE(AO)	LFD	f19,  3 * SIZE(AO)	LFD	f20,  0 * SIZE(BO)	LFD	f21,  1 * SIZE(BO)	LFD	f22,  2 * SIZE(BO)	LFD	f23,  3 * SIZE(BO)	LFD	f24,  4 * SIZE(BO)	LFD	f25,  5 * SIZE(BO)	LFD	f26,  6 * SIZE(BO)	LFD	f27,  7 * SIZE(BO)#endif#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	sub	TEMP, K, KK#elif defined(LEFT)	addi	TEMP, KK, 2#else	addi	TEMP, KK, 4#endif	srawi.	TEMP,  TEMP,  2	mtspr	CTR, TEMP#else	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#endif	ble	.L25	.align 5.L22:	FMADD	f0,  f16, f20, f0	nop	FMADD	f1,  f17, f20, f1	LFD	f20,  8 * SIZE(BO)	FMADD	f4,  f16, f21, f4	nop	FMADD	f5,  f17, f21, f5	LFD	f21,  9 * SIZE(BO)	FMADD	f8,  f16, f22, f8	nop	FMADD	f9,  f17, f22, f9	LFD	f22, 10 * SIZE(BO)	FMADD	f12, f16, f23, f12	LFD	f16,  4 * SIZE(AO)	FMADD	f13, f17, f23, f13	LFD	f23, 11 * SIZE(BO)	FMADD	f2,  f18, f24, f2	LFD	f17,  5 * SIZE(AO)	FMADD	f3,  f19, f24, f3	LFD	f24, 12 * SIZE(BO)	FMADD	f6,  f18, f25, f6	nop	FMADD	f7,  f19, f25, f7	LFD	f25, 13 * SIZE(BO)	FMADD	f10, f18, f26, f10	nop	FMADD	f11, f19, f26, f11	LFD	f26, 14 * SIZE(BO)	FMADD	f14, f18, f27, f14	LFD	f18,  6 * SIZE(AO)	FMADD	f15, f19, f27, f15	LFD	f27, 15 * SIZE(BO)	FMADD	f0,  f16, f20, f0	LFD	f19,  7 * SIZE(AO)	FMADD	f1,  f17, f20, f1	LFDU	f20, 16 * SIZE(BO)	FMADD	f4,  f16, f21, f4	nop	FMADD	f5,  f17, f21, f5	LFD	f21,  1 * SIZE(BO)	FMADD	f8,  f16, f22, f8	nop	FMADD	f9,  f17, f22, f9	LFD	f22,  2 * SIZE(BO)	FMADD	f12, f16, f23, f12	LFDU	f16,  8 * SIZE(AO)	FMADD	f13, f17, f23, f13	LFD	f23,  3 * SIZE(BO)	FMADD	f2,  f18, f24, f2	LFD	f17,  1 * SIZE(AO)	FMADD	f3,  f19, f24, f3	LFD	f24,  4 * SIZE(BO)	FMADD	f6,  f18, f25, f6	nop	FMADD	f7,  f19, f25, f7	LFD	f25,  5 * SIZE(BO)	FMADD	f10, f18, f26, f10	nop	FMADD	f11, f19, f26, f11	LFD	f26,  6 * SIZE(BO)	FMADD	f14, f18, f27, f14	LFD	f18,  2 * SIZE(AO)	FMADD	f15, f19, f27, f15	LFD	f19,  3 * SIZE(AO)	LFD	f27,  7 * SIZE(BO)	bdnz	.L22	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 4.L25:	lfd	f30,  ALPHA#if   defined(TRMMKERNEL)#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	sub	TEMP, K, KK#elif defined(LEFT)	addi	TEMP, KK, 2#else	addi	TEMP, KK, 4#endif	andi.	TEMP,  TEMP,  3	mtspr	CTR, TEMP#else	andi.	r0,  K,  3	mtspr	CTR, r0#endif	ble+	.L28	.align 4.L26:	FMADD	f0,  f16, f20, f0	nop	FMADD	f1,  f17, f20, f1	LFDU	f20,  4 * SIZE(BO)	FMADD	f4,  f16, f21, f4	nop	FMADD	f5,  f17, f21, f5	LFD	f21,  1 * SIZE(BO)	FMADD	f8,  f16, f22, f8	nop	FMADD	f9,  f17, f22, f9	LFD	f22,  2 * SIZE(BO)	FMADD	f12, f16, f23, f12	LFDU	f16,  2 * SIZE(AO)	FMADD	f13, f17, f23, f13	LFD	f17,  1 * SIZE(AO)	LFD	f23,  3 * SIZE(BO)	bdnz	.L26	.align 4

⌨️ 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 + -