zgemm_kernel_2x2_barcelona.s

Optimized GotoBLAS libraries

/*********************************************************************/
/*                                                                   */
/*             Optimized BLAS libraries                              */
/*                     By Kazushige Goto <[e-mail]>     */
/*                                                                   */
/* 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 OLD_M	%rdi
#define OLD_N	%rsi
#define M	%r13
#define N	%r14
#define K	%rdx

#define A	%rcx
#define B	%r8
#define C	%r9
#define LDC	%r10
#define I	%r11
#define AO	%rdi
#define BO	%rsi
#define	CO1	%rbp
#define CO2	%rbx
#define BB	%r12
#define J	%r15
	
#ifndef WINDOWS_ABI

#define STACKSIZE 96

#define ALPHA_R	 48(%rsp)
#define ALPHA_I	 56(%rsp)
#define OFFSET	 64(%rsp)
#define KK	 72(%rsp)
#define KKK	 80(%rsp)

#else

#define STACKSIZE 320

#define OLD_ALPHA_I	40 + STACKSIZE(%rsp)
#define OLD_A		48 + STACKSIZE(%rsp)
#define OLD_B		56 + STACKSIZE(%rsp)
#define OLD_C		64 + STACKSIZE(%rsp)
#define OLD_LDC		72 + STACKSIZE(%rsp)
#define OLD_OFFSET	80 + STACKSIZE(%rsp)

#define ALPHA_R	224(%rsp)
#define ALPHA_I	232(%rsp)
#define OFFSET	240(%rsp)
#define KK	248(%rsp)
#define KKK	256(%rsp)

#endif

#define PREFETCH     prefetch
#define PREFETCHSIZE  (8 *  7 + 0)

#define SPREFETCHSIZE (8 *  3 + 0)

#define movlpd	movsd
#define movapd	movups
#define movupd	movups
	
#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
#define ADD1	addpd
#define ADD2	addpd
#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
#define ADD1	subpd
#define ADD2	addpd
#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
#define ADD1	addpd
#define ADD2	subpd
#else
#define ADD1	subpd
#define ADD2	subpd
#endif

#define KERNEL1(xx) \
	mulpd	%xmm1, %xmm0 ;\
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm0, %xmm8 ;\
	movapd	%xmm2, %xmm0 ;\
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO, %rax, 4) ;\
	ADD1	%xmm1, %xmm12 ;\
	movddup	-14 * SIZE(BO, %rax, 4), %xmm1 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
	movapd	%xmm0, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	-13 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm1, %xmm0 ;\
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm0, %xmm10 ;\
	movapd	-12 * SIZE(AO, %rax, 4), %xmm0 ;\
	ADD1	%xmm1, %xmm14 ;\
	movddup	-12 * SIZE(BO, %rax, 4), %xmm1 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	-11 * SIZE(BO, %rax, 4), %xmm3 ;\
	movapd	%xmm0, %xmm2

#define KERNEL2(xx) \
	mulpd	%xmm1, %xmm0 ;\
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm0, %xmm8 ;\
	movapd	%xmm2, %xmm0 ;\
	ADD1	%xmm1, %xmm12 ;\
	movddup	-10 * SIZE(BO, %rax, 4), %xmm1 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
	movapd	%xmm0, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	 -9 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm1, %xmm0 ;\
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm0, %xmm10 ;\
	ADD1	%xmm1, %xmm14 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	 -7 * SIZE(BO, %rax, 4), %xmm3 ;\
/**/	movddup	  (BO, %rax, 4), %xmm1 ;\
	movapd	%xmm4, %xmm2

#define KERNEL3(xx) \
	mulpd	%xmm5, %xmm4 ;\
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm5 ;\
	ADD1	%xmm4, %xmm8 ;\
	movapd	%xmm2, %xmm4 ;\
	ADD1	%xmm5, %xmm12 ;\
	movddup	 -6 * SIZE(BO, %rax, 4), %xmm5 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
	movapd	%xmm4, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	 -5 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm5, %xmm4 ;\
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm5 ;\
	ADD1	%xmm4, %xmm10 ;\
	movapd	 -4 * SIZE(AO, %rax, 4), %xmm4 ;\
	ADD1	%xmm5, %xmm14 ;\
	movddup	 -4 * SIZE(BO, %rax, 4), %xmm5 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	 -3 * SIZE(BO, %rax, 4), %xmm3 ;\
	movapd	%xmm4, %xmm2

#define KERNEL4(xx) \
	mulpd	%xmm5, %xmm4 ;\
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm5 ;\
	ADD1	%xmm4, %xmm8 ;\
	movapd	%xmm2, %xmm4 ;\
	ADD1	%xmm5, %xmm12 ;\
	movddup	 -2 * SIZE(BO, %rax, 4), %xmm5 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
	movapd	%xmm4, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	 -1 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm5, %xmm4 ;\
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm5 ;\
/***/	movapd	  (AO, %rax, 4), %xmm6 ;\
	ADD1	%xmm4, %xmm10 ;\
	ADD1	%xmm5, %xmm14 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	  1 * SIZE(BO, %rax, 4), %xmm3 ;\
	movddup	  8 * SIZE(BO, %rax, 4), %xmm5 ;\
	movapd	%xmm6, %xmm2

#define KERNEL5(xx) \
	mulpd	%xmm1, %xmm6 ;\
	mulpd	  2 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm6, %xmm8 ;\
	movapd	%xmm2, %xmm6 ;\
	ADD1	%xmm1, %xmm12 ;\
	movddup	  2 * SIZE(BO, %rax, 4), %xmm1 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	  2 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
/**/	movapd	  8 * SIZE(AO, %rax, 4), %xmm7 ;\
	movapd	%xmm6, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	  3 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm1, %xmm6 ;\
	mulpd	  2 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm6, %xmm10 ;\
	movapd	  4 * SIZE(AO, %rax, 4), %xmm6 ;\
	ADD1	%xmm1, %xmm14 ;\
	movddup	  4 * SIZE(BO, %rax, 4), %xmm1 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	  2 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	  5 * SIZE(BO, %rax, 4), %xmm3 ;\
	movapd	%xmm6, %xmm2

#define KERNEL6(xx) \
	mulpd	%xmm1, %xmm6 ;\
	mulpd	  6 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm6, %xmm8 ;\
	movapd	%xmm2, %xmm6 ;\
	ADD1	%xmm1, %xmm12 ;\
	movddup	  6 * SIZE(BO, %rax, 4), %xmm1 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	  6 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
	movapd	%xmm6, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	  7 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm1, %xmm6 ;\
	mulpd	  6 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm6, %xmm10 ;\
/***/	movapd	 16 * SIZE(AO, %rax, 4), %xmm0 ;\
	ADD1	%xmm1, %xmm14 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	  6 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	  9 * SIZE(BO, %rax, 4), %xmm3 ;\
	movddup	 16 * SIZE(BO, %rax, 4), %xmm1 ;\
	movapd	%xmm7, %xmm2

#define KERNEL7(xx) \
	mulpd	%xmm5, %xmm7 ;\
	mulpd	 10 * SIZE(AO, %rax, 4), %xmm5 ;\
	ADD1	%xmm7, %xmm8 ;\
	movapd	%xmm2, %xmm7 ;\
	ADD1	%xmm5, %xmm12 ;\
	movddup	 10 * SIZE(BO, %rax, 4), %xmm5 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 10 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
	movapd	%xmm7, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	 11 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm5, %xmm7 ;\
	mulpd	 10 * SIZE(AO, %rax, 4), %xmm5 ;\
	ADD1	%xmm7, %xmm10 ;\
	movapd	 12 * SIZE(AO, %rax, 4), %xmm7 ;\
	ADD1	%xmm5, %xmm14 ;\
	movddup	 12 * SIZE(BO, %rax, 4), %xmm5 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 10 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	 13 * SIZE(BO, %rax, 4), %xmm3 ;\
	movapd	%xmm7, %xmm2

#define KERNEL8(xx) \
	mulpd	%xmm5, %xmm7 ;\
	mulpd	 14 * SIZE(AO, %rax, 4), %xmm5 ;\
	ADD1	%xmm7, %xmm8 ;\
	movapd	%xmm2, %xmm7 ;\
	ADD1	%xmm5, %xmm12 ;\
	movddup	 14 * SIZE(BO, %rax, 4), %xmm5 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 14 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
	movapd	%xmm7, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	 15 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm5, %xmm7 ;\
	mulpd	 14 * SIZE(AO, %rax, 4), %xmm5 ;\
	ADD1	%xmm7, %xmm10 ;\
	ADD1	%xmm5, %xmm14 ;\
/**/	movapd	 24 * SIZE(AO, %rax, 4), %xmm4 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 14 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	 17 * SIZE(BO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	movddup	 24 * SIZE(BO, %rax, 4), %xmm5 ;\
	movapd	%xmm0, %xmm2 ;\
	addq	$8 * SIZE, %rax
	
#define KERNEL_SUB1(xx) \
	mulpd	%xmm1, %xmm0 ;\
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm0, %xmm8 ;\
	movapd	%xmm2, %xmm0 ;\
	ADD1	%xmm1, %xmm12 ;\
	movddup	-14 * SIZE(BO, %rax, 4), %xmm1 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
	movapd	%xmm0, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	-13 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm1, %xmm0 ;\
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm0, %xmm10 ;\
	movapd	-12 * SIZE(AO, %rax, 4), %xmm0 ;\
	ADD1	%xmm1, %xmm14 ;\
	movddup	-12 * SIZE(BO, %rax, 4), %xmm1 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	-11 * SIZE(BO, %rax, 4), %xmm3 ;\
	movapd	%xmm0, %xmm2

#define KERNEL_SUB2(xx) \
	mulpd	%xmm1, %xmm0 ;\
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm0, %xmm8 ;\
	movapd	%xmm2, %xmm0 ;\
	ADD1	%xmm1, %xmm12 ;\
	movddup	-10 * SIZE(BO, %rax, 4), %xmm1 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
	movapd	%xmm0, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	 -9 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm1, %xmm0 ;\
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm1 ;\
	ADD1	%xmm0, %xmm10 ;\
	movapd	  (AO, %rax, 4), %xmm0 ;\
	ADD1	%xmm1, %xmm14 ;\
	movddup	  (BO, %rax, 4), %xmm1 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	 -7 * SIZE(BO, %rax, 4), %xmm3 ;\
	movapd	%xmm4, %xmm2

#define KERNEL_SUB3(xx) \
	mulpd	%xmm5, %xmm4 ;\
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm5 ;\
	ADD1	%xmm4, %xmm8 ;\
	movapd	%xmm2, %xmm4 ;\
	ADD1	%xmm5, %xmm12 ;\
	movddup	 -6 * SIZE(BO, %rax, 4), %xmm5 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
	movapd	%xmm4, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	 -5 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm5, %xmm4 ;\
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm5 ;\
	ADD1	%xmm4, %xmm10 ;\
	movapd	 -4 * SIZE(AO, %rax, 4), %xmm4 ;\
	ADD1	%xmm5, %xmm14 ;\
	movddup	 -4 * SIZE(BO, %rax, 4), %xmm5 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	 -3 * SIZE(BO, %rax, 4), %xmm3 ;\
	movapd	%xmm4, %xmm2

#define KERNEL_SUB4(xx) \
	mulpd	%xmm5, %xmm4 ;\
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm5 ;\
	ADD1	%xmm4, %xmm8 ;\
	movapd	%xmm2, %xmm4 ;\
	ADD1	%xmm5, %xmm12 ;\
	movddup	 -2 * SIZE(BO, %rax, 4), %xmm5 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm9 ;\
	movapd	%xmm4, %xmm2 ;\
	ADD2	%xmm3, %xmm13 ;\
	movddup	 -1 * SIZE(BO, %rax, 4), %xmm3 ;\
	mulpd	%xmm5, %xmm4 ;\
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm5 ;\
	ADD1	%xmm4, %xmm10 ;\
	ADD1	%xmm5, %xmm14 ;\
	mulpd	%xmm3, %xmm2 ;\
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm3 ;\
	ADD2	%xmm2, %xmm11 ;\
	ADD2	%xmm3, %xmm15 ;\
 	movddup	  1 * SIZE(BO, %rax, 4), %xmm3 ;\
	movapd	%xmm0, %xmm2

	PROLOGUE
	PROFCODE
	
	subq	$STACKSIZE, %rsp
	
	movq	%rbx,   (%rsp)
	movq	%rbp,  8(%rsp)
	movq	%r12, 16(%rsp)
	movq	%r13, 24(%rsp)
	movq	%r14, 32(%rsp)
	movq	%r15, 40(%rsp)

#ifdef WINDOWS_ABI
	movq	%rdi,    48(%rsp)
	movq	%rsi,    56(%rsp)
	movups	%xmm6,   64(%rsp)
	movups	%xmm7,   80(%rsp)
	movups	%xmm8,   96(%rsp)
	movups	%xmm9,  112(%rsp)
	movups	%xmm10, 128(%rsp)
	movups	%xmm11, 144(%rsp)
	movups	%xmm12, 160(%rsp)
	movups	%xmm13, 176(%rsp)
	movups	%xmm14, 192(%rsp)
	movups	%xmm15, 208(%rsp)

	movq	ARG1,      OLD_M
	movq	ARG2,      OLD_N
	movq	ARG3,      K
	movq	OLD_A,     A
	movq	OLD_B,     B
	movq	OLD_C,     C
	movq	OLD_LDC,   LDC
#ifdef TRMMKERNEL
	movsd	OLD_OFFSET, %xmm12
#endif
	movaps	%xmm3,       %xmm0
	movsd	OLD_ALPHA_I, %xmm1
#else
	movq	STACKSIZE +  8(%rsp), LDC
#ifdef TRMMKERNEL
	movsd	STACKSIZE + 16(%rsp), %xmm12
#endif

#endif

	movq	OLD_M, M
	movq	OLD_N, N

	movlpd	 %xmm0, ALPHA_R
	movlpd	 %xmm1, ALPHA_I

#ifdef TRMMKERNEL
	movlpd	%xmm12, OFFSET
	movlpd	%xmm12, KK
#ifndef LEFT
	negq	KK
#endif	
#endif