⭐ 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_ncopy_hummer_4.s

📁 Optimized GotoBLAS libraries
💻 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	M	r3#define	N	r4#define	A	r5#define	LDA	r6#define B	r7#define AO1	r8#define AO2	r9#define AO3	r10#define AO4	r11#define J	r12#define INC	r30#define INC2	r31#define c01	f0#define c02	f1#define c03	f2#define c04	f3#define c05	f4#define c06	f5#define c07	f6#define c08	f7#define c09	f8#define c10	f9#define c11	f10#define c12	f11#define c13	f12#define c14	f13#define c15	f14#define c16	f15#define	sel_p	f16#define	sel_s	f17#define c17	f18#define c18	f19	PROLOGUE	PROFCODE	li	r0, -16	stfpdux	f14, SP, r0	stfpdux	f15, SP, r0	stfpdux	f16, SP, r0	stfpdux	f17, SP, r0	stfpdux	f18, SP, r0	stfpdux	f19, SP, r0		stwu	r31,  -4(SP)	stwu	r30,  -4(SP)	lis	r9,   0x3f80	lis	r10,  0xbf80	stwu	r9,    -4(SP)	stwu	r10,   -4(SP)	stwu	r10,   -4(SP)	stwu	r9,    -4(SP)	slwi	LDA, LDA, BASE_SHIFT	cmpwi	cr0, M, 0	ble-	.L99	cmpwi	cr0, N, 0	ble-	.L99	andi.	r0, A,   2 * SIZE - 1	bne	.L100	andi.	r0, LDA, 2 * SIZE - 1	bne	.L100	li	r0, 8	addi	SP, SP, -8	lfpsux	sel_p, SP, r0	lfpsux	sel_s, SP, r0	li	INC,  1 * SIZE	li	INC2, 2 * SIZE	subi	A, A, 2 * SIZE	subi	B, B, 2 * SIZE	srawi.	J,  N,  2	ble	.L20	.align 4.L11:	mr	AO1, A	add	AO2, A,   LDA	add	AO3, AO2, LDA	add	AO4, AO3, LDA	add	A,   AO4, LDA	srawi.	r0,  M,  3	mtspr	CTR, r0	ble	.L15	.align 4.L12:	LFPDUX	c01,   AO1, INC2	LFXDUX	c05,   AO2, INC2	LFPDUX	c09,   AO3, INC2	LFXDUX	c13,   AO4, INC2	LFPDUX	c02,   AO1, INC2	LFXDUX	c06,   AO2, INC2	LFPDUX	c10,   AO3, INC2	LFXDUX	c14,   AO4, INC2	LFPDUX	c03,   AO1, INC2	LFXDUX	c07,   AO2, INC2	LFPDUX	c11,   AO3, INC2	LFXDUX	c15,   AO4, INC2	LFPDUX	c04,   AO1, INC2	LFXDUX	c08,   AO2, INC2	LFPDUX	c12,   AO3, INC2	LFXDUX	c16,   AO4, INC2	fpsel	c17, sel_p, c01, c05	fpsel	c18, sel_p, c09, c13	fpsel	c01, sel_s, c01, c05	fpsel	c05, sel_s, c09, c13	fpsel	c09, sel_p, c02, c06	fpsel	c13, sel_p, c10, c14	STFPDUX	c17,   B, INC2	fpsel	c02, sel_s, c02, c06 	STFPDUX	c18,   B, INC2	fpsel	c06, sel_s, c10, c14	STFXDUX	c01,   B, INC2	fpsel	c10, sel_p, c03, c07	STFXDUX	c05,   B, INC2	fpsel	c14, sel_p, c11, c15	STFPDUX	c09,   B, INC2	fpsel	c03, sel_s, c03, c07	STFPDUX	c13,   B, INC2	fpsel	c07, sel_s, c11, c15	STFXDUX	c02,   B, INC2	fpsel	c11, sel_p, c04, c08	STFXDUX	c06,   B, INC2	fpsel	c15, sel_p, c12, c16	STFPDUX	c10,   B, INC2	fpsel	c04, sel_s, c04, c08	STFPDUX	c14,   B, INC2	fpsel	c08, sel_s, c12, c16	STFXDUX	c03,   B, INC2	STFXDUX	c07,   B, INC2	STFPDUX	c11,   B, INC2	STFPDUX	c15,   B, INC2	STFXDUX	c04,   B, INC2	STFXDUX	c08,   B, INC2	bdnz	.L12	.align 4	.L15:	andi.	r0,  M,  7	ble	.L19	andi.	r0,  M,  4	beq	.L16	LFPDUX	c01,   AO1, INC2	LFXDUX	c05,   AO2, INC2	LFPDUX	c09,   AO3, INC2	LFXDUX	c13,   AO4, INC2	LFPDUX	c02,   AO1, INC2	LFXDUX	c06,   AO2, INC2	LFPDUX	c10,   AO3, INC2	LFXDUX	c14,   AO4, INC2	fpsel	c17, sel_p, c01, c05	fpsel	c18, sel_p, c09, c13	fpsel	c01, sel_s, c01, c05	fpsel	c05, sel_s, c09, c13	fpsel	c09, sel_p, c02, c06	fpsel	c13, sel_p, c10, c14	STFPDUX	c17,   B, INC2	fpsel	c02, sel_s, c02, c06	STFPDUX	c18,   B, INC2	fpsel	c06, sel_s, c10, c14	STFXDUX	c01,   B, INC2	STFXDUX	c05,   B, INC2	STFPDUX	c09,   B, INC2	STFPDUX	c13,   B, INC2	STFXDUX	c02,   B, INC2	STFXDUX	c06,   B, INC2	.align 4.L16:	andi.	r0,  M,  2	beq	.L17	LFPDUX	c01,   AO1, INC2	LFXDUX	c05,   AO2, INC2	LFPDUX	c09,   AO3, INC2	LFXDUX	c13,   AO4, INC2	fpsel	c17, sel_p, c01, c05	fpsel	c18, sel_p, c09, c13	fpsel	c01, sel_s, c01, c05	fpsel	c05, sel_s, c09, c13	STFPDUX	c17,   B, INC2	STFPDUX	c18,   B, INC2	STFXDUX	c01,   B, INC2	STFXDUX	c05,   B, INC2	.align 4.L17:	andi.	r0,  M,  1	beq	.L19	LFDUX	c01,   AO1, INC2	LFDUX	c02,   AO2, INC2	LFDUX	c03,   AO3, INC2	LFDUX	c04,   AO4, INC2	fsmfp	c01, c02	fsmfp	c03, c04	STFPDUX	c01,   B, INC2	STFPDUX	c03,   B, INC2	.align 4.L19:	addic.	J, J, -1	bgt	.L11	.align 4.L20:	andi.	J,  N,  2	ble	.L30	mr	AO1, A	add	AO2, A,   LDA	add	A,   AO2, LDA	srawi.	r0,  M,  3	mtspr	CTR, r0	ble	.L25	.align 4.L22:	LFPDUX	c01,   AO1, INC2	LFXDUX	c05,   AO2, INC2	LFPDUX	c02,   AO1, INC2	LFXDUX	c06,   AO2, INC2	LFPDUX	c03,   AO1, INC2	LFXDUX	c07,   AO2, INC2	LFPDUX	c04,   AO1, INC2	LFXDUX	c08,   AO2, INC2	fpsel	c17, sel_p, c01, c05	fpsel	c01, sel_s, c01, c05	fpsel	c09, sel_p, c02, c06	fpsel	c02, sel_s, c02, c06	fpsel	c10, sel_p, c03, c07	fpsel	c03, sel_s, c03, c07	STFPDUX	c17,   B, INC2	fpsel	c11, sel_p, c04, c08	STFXDUX	c01,   B, INC2	fpsel	c04, sel_s, c04, c08	STFPDUX	c09,   B, INC2	STFXDUX	c02,   B, INC2	STFPDUX	c10,   B, INC2	STFXDUX	c03,   B, INC2	STFPDUX	c11,   B, INC2	STFXDUX	c04,   B, INC2	bdnz	.L22	.align 4	.L25:	andi.	r0,  M,  7	ble	.L30	andi.	r0,  M,  4	beq	.L26	LFPDUX	c01,   AO1, INC2	LFXDUX	c05,   AO2, INC2	LFPDUX	c02,   AO1, INC2	LFXDUX	c06,   AO2, INC2	fpsel	c17, sel_p, c01, c05	fpsel	c01, sel_s, c01, c05	fpsel	c09, sel_p, c02, c06	fpsel	c02, sel_s, c02, c06	STFPDUX	c17,   B, INC2	STFXDUX	c01,   B, INC2	STFPDUX	c09,   B, INC2	STFXDUX	c02,   B, INC2	.align 4.L26:	andi.	r0,  M,  2	beq	.L27	LFPDUX	c01,   AO1, INC2	LFXDUX	c05,   AO2, INC2	fpsel	c17, sel_p, c01, c05	fpsel	c01, sel_s, c01, c05	STFPDUX	c17,   B, INC2	STFXDUX	c01,   B, INC2	.align 4.L27:	andi.	r0,  M,  1	beq	.L30	LFDUX	c01,   AO1, INC2	LFDUX	c02,   AO2, INC2	fsmfp	c01, c02	STFPDUX	c01,   B, INC2	.align 4.L30:	andi.	J,  N,  1	ble	.L99	mr	AO1, A	srawi.	r0,  M,  3	mtspr	CTR, r0	ble	.L35	.align 4.L32:	LFPDUX	c01,   AO1, INC2	LFPDUX	c02,   AO1, INC2	LFPDUX	c03,   AO1, INC2	LFPDUX	c04,   AO1, INC2	STFPDUX	c01,   B, INC2	STFPDUX	c02,   B, INC2	STFPDUX	c03,   B, INC2	STFPDUX	c04,   B, INC2	bdnz	.L32	.align 4	.L35:	andi.	r0,  M,  7	ble	.L99	andi.	r0,  M,  4	beq	.L36	LFPDUX	c01,   AO1, INC2	LFPDUX	c02,   AO1, INC2	STFPDUX	c01,   B, INC2	STFPDUX	c02,   B, INC2	.align 4.L36:	andi.	r0,  M,  2	beq	.L37	LFPDUX	c01,   AO1, INC2	STFPDUX	c01,   B, INC2	.align 4.L37:	andi.	r0,  M,  1	beq	.L99	LFDX	c01,   AO1, INC2	STFDX	c01,   B,  INC2	.align 4.L99:	addi	SP, SP, 4	lwzu	r30,   4(SP)	lwzu	r31,   4(SP)	subi	SP, SP, 12	li	r0, 16	lfpdux	f19, SP, r0	lfpdux	f18, SP, r0	lfpdux	f17, SP, r0	lfpdux	f16, SP, r0	lfpdux	f15, SP, r0	lfpdux	f14, SP, r0	addi	SP, SP, 16	blr	.align 4.L100:	li	INC,  1 * SIZE	li	INC2, 2 * SIZE	subi	A, A, 1 * SIZE	subi	B, B, 2 * SIZE	srawi.	J,  N,  2	ble	.L120	.align 4.L111:	mr	AO1, A	add	AO2, A,   LDA	add	AO3, AO2, LDA	add	AO4, AO3, LDA	add	A,   AO4, LDA	srawi.	r0,  M,  3	mtspr	CTR, r0	ble	.L115	.align 4.L112:	LFDUX	c01,   AO1, INC	LFDUX	c02,   AO1, INC	LFDUX	c03,   AO1, INC	LFDUX	c04,   AO1, INC	LFDUX	c09,   AO1, INC	LFDUX	c10,   AO1, INC	LFDUX	c11,   AO1, INC	LFDUX	c12,   AO1, INC	LFSDUX	c01,   AO2, INC	LFSDUX	c02,   AO2, INC	LFSDUX	c03,   AO2, INC	LFSDUX	c04,   AO2, INC	LFSDUX	c09,   AO2, INC	LFSDUX	c10,   AO2, INC	LFSDUX	c11,   AO2, INC	LFSDUX	c12,   AO2, INC	LFDUX	c05,   AO3, INC	LFDUX	c06,   AO3, INC	LFDUX	c07,   AO3, INC	LFDUX	c08,   AO3, INC	LFDUX	c13,   AO3, INC	LFDUX	c14,   AO3, INC	LFDUX	c15,   AO3, INC	LFDUX	c16,   AO3, INC	LFSDUX	c05,   AO4, INC	LFSDUX	c06,   AO4, INC	LFSDUX	c07,   AO4, INC	LFSDUX	c08,   AO4, INC	LFSDUX	c13,   AO4, INC	LFSDUX	c14,   AO4, INC	LFSDUX	c15,   AO4, INC	LFSDUX	c16,   AO4, INC	STFPDUX	c01,   B, INC2	STFPDUX	c05,   B, INC2	STFPDUX	c02,   B, INC2	STFPDUX	c06,   B, INC2	STFPDUX	c03,   B, INC2	STFPDUX	c07,   B, INC2	STFPDUX	c04,   B, INC2	STFPDUX	c08,   B, INC2	STFPDUX	c09,   B, INC2	STFPDUX	c13,   B, INC2	STFPDUX	c10,   B, INC2	STFPDUX	c14,   B, INC2	STFPDUX	c11,   B, INC2	STFPDUX	c15,   B, INC2	STFPDUX	c12,   B, INC2	STFPDUX	c16,   B, INC2	bdnz	.L112	.align 4	.L115:	andi.	r0,  M,  7	ble	.L119	andi.	r0,  M,  4	beq	.L116	LFDUX	c01,   AO1, INC	LFDUX	c02,   AO1, INC	LFDUX	c03,   AO1, INC	LFDUX	c04,   AO1, INC	LFSDUX	c01,   AO2, INC	LFSDUX	c02,   AO2, INC	LFSDUX	c03,   AO2, INC	LFSDUX	c04,   AO2, INC	LFDUX	c05,   AO3, INC	LFDUX	c06,   AO3, INC	LFDUX	c07,   AO3, INC	LFDUX	c08,   AO3, INC	LFSDUX	c05,   AO4, INC	LFSDUX	c06,   AO4, INC	LFSDUX	c07,   AO4, INC	LFSDUX	c08,   AO4, INC	STFPDUX	c01,   B, INC2	STFPDUX	c05,   B, INC2	STFPDUX	c02,   B, INC2	STFPDUX	c06,   B, INC2	STFPDUX	c03,   B, INC2	STFPDUX	c07,   B, INC2	STFPDUX	c04,   B, INC2	STFPDUX	c08,   B, INC2	.align 4.L116:	andi.	r0,  M,  2	beq	.L117	LFDUX	c01,   AO1, INC	LFDUX	c02,   AO1, INC	LFSDUX	c01,   AO2, INC	LFSDUX	c02,   AO2, INC	LFDUX	c05,   AO3, INC	LFDUX	c06,   AO3, INC	LFSDUX	c05,   AO4, INC	LFSDUX	c06,   AO4, INC	STFPDUX	c01,   B, INC2	STFPDUX	c05,   B, INC2	STFPDUX	c02,   B, INC2	STFPDUX	c06,   B, INC2	.align 4.L117:	andi.	r0,  M,  1	beq	.L119	LFDUX	c01,   AO1, INC	LFDUX	c05,   AO3, INC	nop	nop	LFSDUX	c01,   AO2, INC	LFSDUX	c05,   AO4, INC	STFPDUX	c01,   B, INC2	STFPDUX	c05,   B, INC2	.align 4.L119:	addic.	J, J, -1	bgt	.L111	.align 4.L120:	andi.	J,  N,  2	ble	.L130	mr	AO1, A	add	AO2, A,   LDA	add	A,   AO2, LDA	srawi.	r0,  M,  3	mtspr	CTR, r0	ble	.L125	.align 4.L122:	LFDUX	c01,   AO1, INC	LFDUX	c02,   AO1, INC	LFDUX	c03,   AO1, INC	LFDUX	c04,   AO1, INC	LFDUX	c09,   AO1, INC	LFDUX	c10,   AO1, INC	LFDUX	c11,   AO1, INC	LFDUX	c12,   AO1, INC	LFSDUX	c01,   AO2, INC	LFSDUX	c02,   AO2, INC	LFSDUX	c03,   AO2, INC	LFSDUX	c04,   AO2, INC	LFSDUX	c09,   AO2, INC	LFSDUX	c10,   AO2, INC	LFSDUX	c11,   AO2, INC	LFSDUX	c12,   AO2, INC	STFPDUX	c01,   B, INC2	STFPDUX	c02,   B, INC2	STFPDUX	c03,   B, INC2	STFPDUX	c04,   B, INC2	STFPDUX	c09,   B, INC2	STFPDUX	c10,   B, INC2	STFPDUX	c11,   B, INC2	STFPDUX	c12,   B, INC2	bdnz	.L122	.align 4	.L125:	andi.	r0,  M,  7	ble	.L130	andi.	r0,  M,  4	beq	.L126	LFDUX	c01,   AO1, INC	LFDUX	c02,   AO1, INC	LFDUX	c03,   AO1, INC	LFDUX	c04,   AO1, INC	LFSDUX	c01,   AO2, INC	LFSDUX	c02,   AO2, INC	LFSDUX	c03,   AO2, INC	LFSDUX	c04,   AO2, INC	STFPDUX	c01,   B, INC2	STFPDUX	c02,   B, INC2	STFPDUX	c03,   B, INC2	STFPDUX	c04,   B, INC2	.align 4.L126:	andi.	r0,  M,  2	beq	.L127	LFDUX	c01,   AO1, INC	LFDUX	c02,   AO1, INC	LFSDUX	c01,   AO2, INC	LFSDUX	c02,   AO2, INC	STFPDUX	c01,   B, INC2	STFPDUX	c02,   B, INC2	.align 4.L127:	andi.	r0,  M,  1	beq	.L130	LFDUX	c01,   AO1, INC	LFDUX	c02,   AO2, INC	fsmfp	c01, c02	STFPDUX	c01,   B, INC2	.align 4.L130:	andi.	J,  N,  1	ble	.L999	mr	AO1, A	srawi.	r0,  M,  3	mtspr	CTR, r0	ble	.L135	.align 4.L132:	LFDUX	c01,   AO1, INC	LFDUX	c02,   AO1, INC	LFDUX	c03,   AO1, INC	LFDUX	c04,   AO1, INC	LFDUX	c05,   AO1, INC	LFDUX	c06,   AO1, INC	LFDUX	c07,   AO1, INC	LFDUX	c08,   AO1, INC	fsmfp	c01, c02	fsmfp	c03, c04	fsmfp	c05, c06	fsmfp	c07, c08	STFPDUX	c01,   B, INC2	STFPDUX	c03,   B, INC2	STFPDUX	c05,   B, INC2	STFPDUX	c07,   B, INC2	bdnz	.L132	.align 4	.L135:	andi.	r0,  M,  7	ble	.L999	andi.	r0,  M,  4	beq	.L136	LFDUX	c01,   AO1, INC	LFDUX	c02,   AO1, INC	LFDUX	c03,   AO1, INC	LFDUX	c04,   AO1, INC	fsmfp	c01, c02	fsmfp	c03, c04	STFPDUX	c01,   B, INC2	STFPDUX	c03,   B, INC2	.align 4.L136:	andi.	r0,  M,  2	beq	.L137	LFDUX	c01,   AO1, INC	LFDUX	c02,   AO1, INC	fsmfp	c01, c02	STFPDUX	c01,   B, INC2	.align 4.L137:	andi.	r0,  M,  1	beq	.L999	LFDX	c01,   AO1, INC	STFDX	c01,   B,  INC2	.align 4.L999:	addi	SP, SP, 12	lwzu	r30,   4(SP)	lwzu	r31,   4(SP)	subi	SP, SP, 12	li	r0, 16	lfpdux	f19, SP, r0	lfpdux	f18, SP, r0	lfpdux	f17, SP, r0	lfpdux	f16, SP, r0	lfpdux	f15, SP, r0	lfpdux	f14, SP, r0	addi	SP, SP, 16	blr	EPILOGUE

⌨️ 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 + -