zgemm_kernel_4x2_sse3.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 M	%rdi
#define N	%rsi
#define K	%rdx

#define A	%rcx
#define B	%r8
#define C	%r9
#define LDC	%r10

#define I	%r11
#define AO	%r12
#define BO	%r13
#define	CO1	%r14
#define CO2	%r15
#define BB	%rbp
	
#ifndef WINDOWS_ABI

#define STACKSIZE 64

#else

#define STACKSIZE 256

#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)

#endif

#define ALPHA_R	  0(%rsp)
#define ALPHA_I	 16(%rsp)
#define J	 32(%rsp)
#define OFFSET	 40(%rsp)
#define KK	 48(%rsp)
#define KKK	 56(%rsp)
#define BUFFER	128(%rsp)
	
#define PREFETCH     prefetcht0
#define PREFETCHSIZE  320

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(RN) || defined(RT) || defined(CN) || defined(CT)
#define ADDSUB	addps
#else
#define ADDSUB	subps
#endif

#define KERNEL1(address) \
	mulps	%xmm8, %xmm9; \
	PREFETCH  (PREFETCHSIZE +  0) * SIZE + (address) * 2 * SIZE(AO); \
	addps	%xmm9, %xmm0; \
	movshdup  0 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	ADDSUB	%xmm9, %xmm1; \
	movsldup  4 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	addps	%xmm9, %xmm2; \
	movshdup  4 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	movaps	  4 * SIZE + (address) * 2 * SIZE(AO), %xmm8; \
	ADDSUB	%xmm9, %xmm3; \
	movsldup  0 * SIZE + (address) * 2 * SIZE(BO), %xmm9

#define KERNEL2(address) \
	mulps	%xmm8, %xmm9; \
	addps	%xmm9, %xmm4; \
	movshdup  0 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	ADDSUB	%xmm9, %xmm5; \
	movsldup  4 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	addps	%xmm9, %xmm6; \
	movshdup  4 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	movaps	  8 * SIZE + (address) * 2 * SIZE(AO), %xmm8; \
	ADDSUB	%xmm9, %xmm7; \
	movsldup  8 * SIZE + (address) * 2 * SIZE(BO), %xmm9

#define KERNEL3(address) \
	mulps	%xmm8, %xmm9; \
	addps	%xmm9, %xmm0; \
	movshdup  8 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	ADDSUB	%xmm9, %xmm1; \
	movsldup 12 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	addps	%xmm9, %xmm2; \
	movshdup 12 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	movaps	 12 * SIZE + (address) * 2 * SIZE(AO), %xmm8; \
	ADDSUB	%xmm9, %xmm3; \
	movsldup  8 * SIZE + (address) * 2 * SIZE(BO), %xmm9

#define KERNEL4(address) \
	mulps	%xmm8, %xmm9; \
	addps	%xmm9, %xmm4; \
	movshdup  8 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	ADDSUB	%xmm9, %xmm5; \
	movsldup 12 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	addps	%xmm9, %xmm6; \
	movshdup 12 * SIZE + (address) * 2 * SIZE(BO), %xmm9; \
	mulps	%xmm8, %xmm9; \
	movaps	 64 * SIZE + (address) * 2 * SIZE(AO), %xmm8; \
	ADDSUB	%xmm9, %xmm7; \
	movsldup 64 * SIZE + (address) * 2 * SIZE(BO), %xmm9

#define KERNEL5(address) \
	mulps	%xmm10, %xmm11; \
	addps	%xmm11, %xmm0; \
	movshdup 16 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	ADDSUB	%xmm11, %xmm1; \
	movsldup 20 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	addps	%xmm11, %xmm2; \
	movshdup 20 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	movaps	 20 * SIZE + (address) * 2 * SIZE(AO), %xmm10; \
	ADDSUB	%xmm11, %xmm3; \
	movsldup 16 * SIZE + (address) * 2 * SIZE(BO), %xmm11

#define KERNEL6(address) \
	mulps	%xmm10, %xmm11; \
	addps	%xmm11, %xmm4; \
	movshdup 16 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	ADDSUB	%xmm11, %xmm5; \
	movsldup 20 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	addps	%xmm11, %xmm6; \
	movshdup 20 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	movaps	 24 * SIZE + (address) * 2 * SIZE(AO), %xmm10; \
	ADDSUB	%xmm11, %xmm7; \
	movsldup 24 * SIZE + (address) * 2 * SIZE(BO), %xmm11

#define KERNEL7(address) \
	mulps	%xmm10, %xmm11; \
	addps	%xmm11, %xmm0; \
	movshdup 24 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	ADDSUB	%xmm11, %xmm1; \
	movsldup 28 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	addps	%xmm11, %xmm2; \
	movshdup 28 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	movaps	 28 * SIZE + (address) * 2 * SIZE(AO), %xmm10; \
	ADDSUB	%xmm11, %xmm3; \
	movsldup 24 * SIZE + (address) * 2 * SIZE(BO), %xmm11

#define KERNEL8(address) \
	mulps	%xmm10, %xmm11; \
	addps	%xmm11, %xmm4; \
	movshdup 24 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	ADDSUB	%xmm11, %xmm5; \
	movsldup 28 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	addps	%xmm11, %xmm6; \
	movshdup 28 * SIZE + (address) * 2 * SIZE(BO), %xmm11; \
	mulps	%xmm10, %xmm11; \
	movaps	 80 * SIZE + (address) * 2 * SIZE(AO), %xmm10; \
	ADDSUB	%xmm11, %xmm7; \
	movsldup 80 * SIZE + (address) * 2 * SIZE(BO), %xmm11

#define KERNEL9(address) \
	mulps	%xmm12, %xmm13; \
	PREFETCH  (PREFETCHSIZE + 32) * SIZE + (address) * 2 * SIZE(AO); \
	addps	%xmm13, %xmm0; \
	movshdup 32 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
	ADDSUB	%xmm13, %xmm1; \
	movsldup 36 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
	addps	%xmm13, %xmm2; \
	movshdup 36 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
	movaps	 36 * SIZE + (address) * 2 * SIZE(AO), %xmm12; \
	ADDSUB	%xmm13, %xmm3; \
	movsldup 32 * SIZE + (address) * 2 * SIZE(BO), %xmm13

#define KERNEL10(address) \
	mulps	%xmm12, %xmm13; \
	addps	%xmm13, %xmm4; \
	movshdup 32 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
	ADDSUB	%xmm13, %xmm5; \
	movsldup 36 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
	addps	%xmm13, %xmm6; \
	movshdup 36 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
	movaps	 40 * SIZE + (address) * 2 * SIZE(AO), %xmm12; \
	ADDSUB	%xmm13, %xmm7; \
	movsldup 40 * SIZE + (address) * 2 * SIZE(BO), %xmm13

#define KERNEL11(address) \
	mulps	%xmm12, %xmm13; \
	addps	%xmm13, %xmm0; \
	movshdup 40 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
	ADDSUB	%xmm13, %xmm1; \
	movsldup 44 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
	addps	%xmm13, %xmm2; \
	movshdup 44 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
	movaps	 44 * SIZE + (address) * 2 * SIZE(AO), %xmm12; \
	ADDSUB	%xmm13, %xmm3; \
	movsldup 40 * SIZE + (address) * 2 * SIZE(BO), %xmm13

#define KERNEL12(address) \
	mulps	%xmm12, %xmm13; \
	addps	%xmm13, %xmm4; \
	movshdup 40 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
	ADDSUB	%xmm13, %xmm5; \
	movsldup 44 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
	addps	%xmm13, %xmm6; \
	movshdup 44 * SIZE + (address) * 2 * SIZE(BO), %xmm13; \
	mulps	%xmm12, %xmm13; \
 	movaps	 96 * SIZE + (address) * 2 * SIZE(AO), %xmm12; \
	ADDSUB	%xmm13, %xmm7; \
	movsldup 96 * SIZE + (address) * 2 * SIZE(BO), %xmm13

#define KERNEL13(address) \
	mulps	%xmm14, %xmm15; \
	addps	%xmm15, %xmm0; \
	movshdup 48 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	ADDSUB	%xmm15, %xmm1; \
	movsldup 52 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	addps	%xmm15, %xmm2; \
	movshdup 52 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	movaps	 52 * SIZE + (address) * 2 * SIZE(AO), %xmm14; \
	ADDSUB	%xmm15, %xmm3; \
	movsldup 48 * SIZE + (address) * 2 * SIZE(BO), %xmm15

#define KERNEL14(address) \
	mulps	%xmm14, %xmm15; \
	addps	%xmm15, %xmm4; \
	movshdup 48 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	ADDSUB	%xmm15, %xmm5; \
	movsldup 52 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	addps	%xmm15, %xmm6; \
	movshdup 52 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	movaps	 56 * SIZE + (address) * 2 * SIZE(AO), %xmm14; \
	ADDSUB	%xmm15, %xmm7; \
	movsldup 56 * SIZE + (address) * 2 * SIZE(BO), %xmm15

#define KERNEL15(address) \
	mulps	%xmm14, %xmm15; \
	addps	%xmm15, %xmm0; \
	movshdup 56 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	ADDSUB	%xmm15, %xmm1; \
	movsldup 60 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	addps	%xmm15, %xmm2; \
	movshdup 60 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	movaps	 60 * SIZE + (address) * 2 * SIZE(AO), %xmm14; \
	ADDSUB	%xmm15, %xmm3; \
	movsldup 56 * SIZE + (address) * 2 * SIZE(BO), %xmm15

#define KERNEL16(address) \
	mulps	%xmm14, %xmm15; \
	addps	%xmm15, %xmm4; \
	movshdup 56 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	ADDSUB	%xmm15, %xmm5; \
	movsldup 60 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	addps	%xmm15, %xmm6; \
	movshdup 60 * SIZE + (address) * 2 * SIZE(BO), %xmm15; \
	mulps	%xmm14, %xmm15; \
	movaps	112 * SIZE + (address) * 2 * SIZE(AO), %xmm14; \
	ADDSUB	%xmm15, %xmm7; \
	movsldup 112 * SIZE + (address) * 2 * SIZE(BO), %xmm15


	PROLOGUE
	PROFCODE
	
	subq	$STACKSIZE, %rsp
	
	movq	%rbx,  0(%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,      M
	movq	ARG2,      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, %xmm4
#endif
	movaps	%xmm3,       %xmm0
	movsd	OLD_ALPHA_I, %xmm1
#else
	movq	72(%rsp), LDC
#ifdef TRMMKERNEL
	movsd	80(%rsp), %xmm4
#endif

#endif

	movq	%rsp, %rbx	# save old stack
	subq	$128 + GEMM_Q * 32, %rsp
	andq	$-4096, %rsp	# align stack

#ifdef WINDOWS_ABI
#if GEMM_Q > 128
	movq	$0,  4096 * 1(%rsp)
#endif
	movq	$0,  4096 * 0(%rsp)
#endif

	pxor	%xmm15, %xmm15
	cmpeqps	%xmm15, %xmm15
	pslld	$31, %xmm15	# Generate mask
	pxor	%xmm2, %xmm2

	shufps	$0, %xmm0, %xmm0
	movaps	 %xmm0,  0 + ALPHA_R
	
	movss	 %xmm1,  4 + ALPHA_I
	movss	 %xmm1, 12 + ALPHA_I
	xorps	 %xmm15, %xmm1
	movss	 %xmm1,  0 + ALPHA_I
	movss	 %xmm1,  8 + ALPHA_I

#ifdef TRMMKERNEL
	movsd	%xmm4, OFFSET
	movsd	%xmm4, KK
#ifndef LEFT
	negq	KK
#endif	
#endif

	salq	$ZBASE_SHIFT, LDC
	movq	N,  J
	sarq	$1, J		# j = (n >> 2)
	jle	.L40
	ALIGN_4

.L01:
#if defined(TRMMKERNEL) && defined(LEFT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif	

/* Copying to Sub Buffer */
	leaq	BUFFER, BO
	
	movq	K, %rax
	sarq	$2, %rax
	jle	.L03
	ALIGN_4
	
.L02:
	movddup	  0 * SIZE(B), %xmm0
	movddup	  2 * SIZE(B), %xmm1
	movddup	  4 * SIZE(B), %xmm2
	movddup	  6 * SIZE(B), %xmm3
	movddup	  8 * SIZE(B), %xmm4
	movddup	 10 * SIZE(B), %xmm5
	movddup	 12 * SIZE(B), %xmm6
	movddup	 14 * SIZE(B), %xmm7

	movaps	%xmm0,  0 * SIZE(BO)
	movaps	%xmm1,  4 * SIZE(BO)
	movaps	%xmm2,  8 * SIZE(BO)
	movaps	%xmm3, 12 * SIZE(BO)
	movaps	%xmm4, 16 * SIZE(BO)
	movaps	%xmm5, 20 * SIZE(BO)
	movaps	%xmm6, 24 * SIZE(BO)
	movaps	%xmm7, 28 * SIZE(BO)

	prefetcht1	128 * SIZE(BO)
	prefetcht0	112 * SIZE(B)

 	addq	$16 * SIZE, B
	addq	$32 * SIZE, BO
	decq	%rax
	jne	.L02
	ALIGN_4

.L03:
	movq	K, %rax
	andq	$3, %rax
	BRANCH
	jle	.L10
	ALIGN_4

.L04:
	movddup	 0 * SIZE(B), %xmm0
	movddup	 2 * SIZE(B), %xmm1

	movaps	%xmm0,  0 * SIZE(BO)
	movaps	%xmm1,  4 * SIZE(BO)

	addq	$4 * SIZE, B
	addq	$8 * SIZE, BO
	decq	%rax
	jne	.L04
	ALIGN_4
	
.L10:
	movq	C, CO1			# coffset1 = c
	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc
	movq	A, AO		# aoffset = a

	leaq	112 * SIZE(B), BB

	movq	M,  I
	sarq	$2, I		# i = (m >> 2)
	jle	.L20
	ALIGN_4

.L11:
	prefetcht0	 0 * SIZE(BB)
	subq	   $-8 * SIZE, BB

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leaq	BUFFER, BO
#else
	leaq	BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax,   8), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 4), BO
#endif	

	movaps	  0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movaps	 16 * SIZE(AO), %xmm10
	pxor	%xmm1, %xmm1
	movaps	 32 * SIZE(AO), %xmm12
	pxor	%xmm2, %xmm2
	movaps	 48 * SIZE(AO), %xmm14
	pxor	%xmm3, %xmm3

	movsldup  0 * SIZE(BO), %xmm9
	pxor	%xmm4, %xmm4
	movsldup 16 * SIZE(BO), %xmm11
	pxor	%xmm5, %xmm5
	movsldup 32 * SIZE(BO), %xmm13
	pxor	%xmm6, %xmm6
	movsldup 48 * SIZE(BO), %xmm15
	pxor	%xmm7, %xmm7

	prefetchnta     8 * SIZE(CO1)
	prefetchnta     8 * SIZE(CO2)

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK	
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$4, %rax
#else
	addq	$2, %rax
#endif
	movq	%rax, KKK
#endif
#if 1
	andq	$-8, %rax
	salq	$4, %rax
	je	.L15
	
.L1X:
	KERNEL1 (32  *  0)
	KERNEL2 (32  *  0)
	KERNEL3 (32  *  0)
	KERNEL4 (32  *  0)
	KERNEL5 (32  *  0)
	KERNEL6 (32  *  0)
	KERNEL7 (32  *  0)
	KERNEL8 (32  *  0)
	KERNEL9 (32  *  0)
	KERNEL10(32  *  0)
	KERNEL11(32  *  0)
	KERNEL12(32  *  0)
	KERNEL13(32  *  0)
	KERNEL14(32  *  0)
	KERNEL15(32  *  0)
	KERNEL16(32  *  0)
	cmpq	$128 *  1, %rax
	jle	.L12
	KERNEL1 (32  *  1)
	KERNEL2 (32  *  1)
	KERNEL3 (32  *  1)
	KERNEL4 (32  *  1)
	KERNEL5 (32  *  1)
	KERNEL6 (32  *  1)
	KERNEL7 (32  *  1)
	KERNEL8 (32  *  1)
	KERNEL9 (32  *  1)
	KERNEL10(32  *  1)
	KERNEL11(32  *  1)
	KERNEL12(32  *  1)
	KERNEL13(32  *  1)
	KERNEL14(32  *  1)
	KERNEL15(32  *  1)
	KERNEL16(32  *  1)
	cmpq	$128 *  2, %rax
	jle	.L12
	KERNEL1 (32  *  2)
	KERNEL2 (32  *  2)
	KERNEL3 (32  *  2)
	KERNEL4 (32  *  2)
	KERNEL5 (32  *  2)
	KERNEL6 (32  *  2)
	KERNEL7 (32  *  2)
	KERNEL8 (32  *  2)
	KERNEL9 (32  *  2)
	KERNEL10(32  *  2)
	KERNEL11(32  *  2)
	KERNEL12(32  *  2)
	KERNEL13(32  *  2)
	KERNEL14(32  *  2)
	KERNEL15(32  *  2)
	KERNEL16(32  *  2)
	cmpq	$128 *  3, %rax
	jle	.L12
	KERNEL1 (32  *  3)
	KERNEL2 (32  *  3)
	KERNEL3 (32  *  3)
	KERNEL4 (32  *  3)
	KERNEL5 (32  *  3)
	KERNEL6 (32  *  3)
	KERNEL7 (32  *  3)
	KERNEL8 (32  *  3)
	KERNEL9 (32  *  3)
	KERNEL10(32  *  3)
	KERNEL11(32  *  3)
	KERNEL12(32  *  3)
	KERNEL13(32  *  3)
	KERNEL14(32  *  3)
	KERNEL15(32  *  3)
	KERNEL16(32  *  3)
	cmpq	$128 *  4, %rax
	jle	.L12
	KERNEL1 (32  *  4)
	KERNEL2 (32  *  4)
	KERNEL3 (32  *  4)
	KERNEL4 (32  *  4)
	KERNEL5 (32  *  4)
	KERNEL6 (32  *  4)
	KERNEL7 (32  *  4)
	KERNEL8 (32  *  4)
	KERNEL9 (32  *  4)
	KERNEL10(32  *  4)
	KERNEL11(32  *  4)
	KERNEL12(32  *  4)
	KERNEL13(32  *  4)
	KERNEL14(32  *  4)
	KERNEL15(32  *  4)
	KERNEL16(32  *  4)
	cmpq	$128 *  5, %rax
	jle	.L12
	KERNEL1 (32  *  5)
	KERNEL2 (32  *  5)
	KERNEL3 (32  *  5)
	KERNEL4 (32  *  5)
	KERNEL5 (32  *  5)
	KERNEL6 (32  *  5)
	KERNEL7 (32  *  5)
	KERNEL8 (32  *  5)
	KERNEL9 (32  *  5)
	KERNEL10(32  *  5)
	KERNEL11(32  *  5)
	KERNEL12(32  *  5)
	KERNEL13(32  *  5)
	KERNEL14(32  *  5)
	KERNEL15(32  *  5)
	KERNEL16(32  *  5)
	cmpq	$128 *  6, %rax
	jle	.L12
	KERNEL1 (32  *  6)
	KERNEL2 (32  *  6)
	KERNEL3 (32  *  6)
	KERNEL4 (32  *  6)
	KERNEL5 (32  *  6)
	KERNEL6 (32  *  6)
	KERNEL7 (32  *  6)
	KERNEL8 (32  *  6)
	KERNEL9 (32  *  6)
	KERNEL10(32  *  6)
	KERNEL11(32  *  6)
	KERNEL12(32  *  6)
	KERNEL13(32  *  6)
	KERNEL14(32  *  6)
	KERNEL15(32  *  6)
	KERNEL16(32  *  6)
	cmpq	$128 *  7, %rax
	jle	.L12
	KERNEL1 (32  *  7)
	KERNEL2 (32  *  7)
	KERNEL3 (32  *  7)
	KERNEL4 (32  *  7)
	KERNEL5 (32  *  7)
	KERNEL6 (32  *  7)
	KERNEL7 (32  *  7)
	KERNEL8 (32  *  7)
	KERNEL9 (32  *  7)
	KERNEL10(32  *  7)
	KERNEL11(32  *  7)
	KERNEL12(32  *  7)
	KERNEL13(32  *  7)
	KERNEL14(32  *  7)
	KERNEL15(32  *  7)
	KERNEL16(32  *  7)

	addq	$64 * 8  * SIZE, AO
	addq	$64 * 8  * SIZE, BO
	subq	$128 * 8, %rax
	jg	.L1X

.L12:
	leaq	(AO, %rax, 2), AO	# * 16
	leaq	(BO, %rax, 2), BO	# * 64
#else
	sarq	$3, %rax
	je	.L15
	ALIGN_4

.L12:
	KERNEL1 (32  *  0)
	KERNEL2 (32  *  0)
	KERNEL3 (32  *  0)
	KERNEL4 (32  *  0)