ztrsm_kernel_ln_2x2_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 J	%r12
#define AO	%r13
#define BO	%r14
#define	CO1	%r15
#define CO2	%rbx
#define KK	%rbp
	
#ifndef WINDOWS_ABI

#define STACKSIZE 128

#define OLD_LDC		 8 + STACKSIZE(%rsp)
#define OLD_OFFSET	16 + STACKSIZE(%rsp)

#define OFFSET	 48(%rsp)
#define KKK	 56(%rsp)
#define AORIG    64(%rsp)

#else

#define STACKSIZE 256

#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 OFFSET	 224(%rsp)
#define KKK	 232(%rsp)
#define AORIG    240(%rsp)
#endif

#define PREFETCH     prefetcht1
#define PREFETCHSIZE (16 * 12 + 3)
#define PREFETCH_R    (4 *  4 + 0)

#ifndef CONJ
#define ADD1	  addpd
#define ADD2	  addpd
#else
#define ADD1	  subpd
#define ADD2	  addpd
#endif

#define KERNEL1(address) \
	mulpd	%xmm8, %xmm9;\
	PREFETCH  (PREFETCHSIZE +  0) * SIZE + (address) * 2 * SIZE(AO);\
	ADD1	%xmm9, %xmm0;\
	movddup	 1 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	ADD2	%xmm9, %xmm1;\
	movddup	 2 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	ADD1	%xmm9, %xmm2;\
	movddup	 3 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	movapd	 2 * SIZE + (address) * 2 * SIZE(AO), %xmm8;\
	ADD2	%xmm9, %xmm3;\
	movddup	 0 * SIZE + (address) * 2 * SIZE(BO), %xmm9

#define KERNEL2(address) \
	mulpd	%xmm8, %xmm9;\
	ADD1	%xmm9, %xmm4;\
	movddup	 1 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	ADD2	%xmm9, %xmm5;\
	movddup	 2 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	ADD1	%xmm9, %xmm6;\
	movddup	 3 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	movapd	 4 * SIZE + (address) * 2 * SIZE(AO), %xmm8;\
	ADD2	%xmm9, %xmm7;\
	movddup	 4 * SIZE + (address) * 2 * SIZE(BO), %xmm9

#define KERNEL3(address) \
	mulpd	%xmm8, %xmm9;\
	ADD1	%xmm9, %xmm0;\
	movddup	 5 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	ADD2	%xmm9, %xmm1;\
	movddup	 6 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	ADD1	%xmm9, %xmm2;\
	movddup	 7 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	movapd	 6 * SIZE + (address) * 2 * SIZE(AO), %xmm8;\
	ADD2	%xmm9, %xmm3;\
	movddup	 4 * SIZE + (address) * 2 * SIZE(BO), %xmm9

#define KERNEL4(address) \
	mulpd	%xmm8, %xmm9;\
	ADD1	%xmm9, %xmm4;\
	movddup	 5 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	ADD2	%xmm9, %xmm5;\
	movddup	 6 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	ADD1	%xmm9, %xmm6;\
	movddup	 7 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
	mulpd	%xmm8, %xmm9;\
	movapd	32 * SIZE + (address) * 2 * SIZE(AO), %xmm8;\
	ADD2	%xmm9, %xmm7;\
	movddup	32 * SIZE + (address) * 2 * SIZE(BO), %xmm9

#define KERNEL5(address) \
	mulpd	%xmm10, %xmm11;\
	ADD1	%xmm11, %xmm0;\
	movddup	 9 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	ADD2	%xmm11, %xmm1;\
	movddup	10 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	ADD1	%xmm11, %xmm2;\
	movddup	11 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	movapd	10 * SIZE + (address) * 2 * SIZE(AO), %xmm10;\
	ADD2	%xmm11, %xmm3;\
	movddup	 8 * SIZE + (address) * 2 * SIZE(BO), %xmm11

#define KERNEL6(address) \
	mulpd	%xmm10, %xmm11;\
	ADD1	%xmm11, %xmm4;\
	movddup	 9 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	ADD2	%xmm11, %xmm5;\
	movddup	10 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	ADD1	%xmm11, %xmm6;\
	movddup	11 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	movapd	12 * SIZE + (address) * 2 * SIZE(AO), %xmm10;\
	ADD2	%xmm11, %xmm7;\
	movddup	12 * SIZE + (address) * 2 * SIZE(BO), %xmm11

#define KERNEL7(address) \
	mulpd	%xmm10, %xmm11;\
	ADD1	%xmm11, %xmm0;\
	movddup	13 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	ADD2	%xmm11, %xmm1;\
	movddup	14 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	ADD1	%xmm11, %xmm2;\
	movddup	15 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	movapd	14 * SIZE + (address) * 2 * SIZE(AO), %xmm10;\
	ADD2	%xmm11, %xmm3;\
	movddup	12 * SIZE + (address) * 2 * SIZE(BO), %xmm11

#define KERNEL8(address) \
	mulpd	%xmm10, %xmm11;\
	ADD1	%xmm11, %xmm4;\
	movddup	13 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	ADD2	%xmm11, %xmm5;\
	movddup	14 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	ADD1	%xmm11, %xmm6;\
	movddup	15 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
	mulpd	%xmm10, %xmm11;\
	movapd	40 * SIZE + (address) * 2 * SIZE(AO), %xmm10;\
	ADD2	%xmm11, %xmm7;\
	movddup	40 * SIZE + (address) * 2 * SIZE(BO), %xmm11

#define KERNEL9(address) \
	mulpd	%xmm12, %xmm13;\
	PREFETCH  (PREFETCHSIZE + 16) * SIZE + (address) * 2 * SIZE(AO);\
	ADD1	%xmm13, %xmm0;\
	movddup	17 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	ADD2	%xmm13, %xmm1;\
	movddup	18 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	ADD1	%xmm13, %xmm2;\
	movddup	19 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	movapd	18 * SIZE + (address) * 2 * SIZE(AO), %xmm12;\
	ADD2	%xmm13, %xmm3;\
	movddup	16 * SIZE + (address) * 2 * SIZE(BO), %xmm13

#define KERNEL10(address) \
	mulpd	%xmm12, %xmm13;\
	ADD1	%xmm13, %xmm4;\
	movddup	17 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	ADD2	%xmm13, %xmm5;\
	movddup	18 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	ADD1	%xmm13, %xmm6;\
	movddup	19 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	movapd	20 * SIZE + (address) * 2 * SIZE(AO), %xmm12;\
	ADD2	%xmm13, %xmm7;\
	movddup	20 * SIZE + (address) * 2 * SIZE(BO), %xmm13

#define KERNEL11(address) \
	mulpd	%xmm12, %xmm13;\
	ADD1	%xmm13, %xmm0;\
	movddup	21 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	ADD2	%xmm13, %xmm1;\
	movddup	22 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	ADD1	%xmm13, %xmm2;\
	movddup	23 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	movapd	22 * SIZE + (address) * 2 * SIZE(AO), %xmm12;\
	ADD2	%xmm13, %xmm3;\
	movddup	20 * SIZE + (address) * 2 * SIZE(BO), %xmm13

#define KERNEL12(address) \
	mulpd	%xmm12, %xmm13;\
	ADD1	%xmm13, %xmm4;\
	movddup 21 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	ADD2	%xmm13, %xmm5;\
	movddup	22 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	ADD1	%xmm13, %xmm6;\
	movddup	23 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
	mulpd	%xmm12, %xmm13;\
	movapd	48 * SIZE + (address) * 2 * SIZE(AO), %xmm12;\
	ADD2	%xmm13, %xmm7;\
	movddup	48 * SIZE + (address) * 2 * SIZE(BO), %xmm13

#define KERNEL13(address) \
	mulpd	%xmm14, %xmm15;\
	ADD1	%xmm15, %xmm0;\
	movddup	25 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	ADD2	%xmm15, %xmm1;\
	movddup	26 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	ADD1	%xmm15, %xmm2;\
	movddup	27 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	movapd	26 * SIZE + (address) * 2 * SIZE(AO), %xmm14;\
	ADD2	%xmm15, %xmm3;\
	movddup	24 * SIZE + (address) * 2 * SIZE(BO), %xmm15

#define KERNEL14(address) \
	mulpd	%xmm14, %xmm15;\
	ADD1	%xmm15, %xmm4;\
	movddup	25 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	ADD2	%xmm15, %xmm5;\
	movddup	26 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	ADD1	%xmm15, %xmm6;\
	movddup	27 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	movapd	28 * SIZE + (address) * 2 * SIZE(AO), %xmm14;\
	ADD2	%xmm15, %xmm7;\
	movddup	28 * SIZE + (address) * 2 * SIZE(BO), %xmm15

#define KERNEL15(address) \
	mulpd	%xmm14, %xmm15;\
	ADD1	%xmm15, %xmm0;\
	movddup	29 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	ADD2	%xmm15, %xmm1;\
	movddup	30 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	ADD1	%xmm15, %xmm2;\
	movddup	31 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	movapd	30 * SIZE + (address) * 2 * SIZE(AO), %xmm14;\
	ADD2	%xmm15, %xmm3;\
	movddup	28 * SIZE + (address) * 2 * SIZE(BO), %xmm15

#define KERNEL16(address) \
	mulpd	%xmm14, %xmm15;\
	ADD1	%xmm15, %xmm4;\
	movddup	29 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	ADD2	%xmm15, %xmm5;\
	movddup	30 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	ADD1	%xmm15, %xmm6;\
	movddup	31 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
	mulpd	%xmm14, %xmm15;\
	movapd	56 * SIZE + (address) * 2 * SIZE(AO), %xmm14;\
	ADD2	%xmm15, %xmm7;\
	movddup	56 * 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
#endif

	movq	OLD_LDC,    LDC
	movq	OLD_OFFSET, KK

	movq	KK, OFFSET
	
	salq	$ZBASE_SHIFT, LDC

#ifdef LN
       movq	M, %rax
       salq	$ZBASE_SHIFT, %rax
       addq	%rax, C
       imulq	K, %rax
       addq	%rax, A
#endif

#ifdef RT
       movq	N, %rax
       salq	$ZBASE_SHIFT, %rax
       imulq	K, %rax
       addq	%rax, B

       movq	N, %rax
       imulq	LDC, %rax
       addq	%rax, C
#endif

#ifdef RN
	negq	KK
#endif	

#ifdef RT
       movq	N, KK
       subq	OFFSET, KK
#endif

	movq	N,  J
	sarq	$1, J		# j = (n >> 2)
	jle	.L100
	ALIGN_4

.L01:
#if defined(LT) || defined(RN)
	movq	A, AO
#else
	movq	A, AORIG
#endif

#ifdef RT
       movq	K, %rax
       salq	$1 + ZBASE_SHIFT, %rax
       subq	%rax, B
       
       leaq	(, LDC, 2), %rax
       subq	%rax, C
#endif

	movq	C, CO1			# coffset1 = c
	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc

#ifndef RT
	leaq	(C, LDC, 2), C
#endif

#ifdef LN
	movq	OFFSET, KK
	addq	M, KK
#endif	
#ifdef LT
	movq	OFFSET, KK
#endif

	testq	$1, M
	jle	.L30

#ifdef LN
       movq	K, %rax
       salq	$0 + ZBASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	movq	AORIG, AO
	leaq	(AO, %rax, 2), AO
	leaq	(B,  %rax, 4), BO
#else
	movq	B, BO
#endif	

	movapd	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movddup	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movapd	 8 * SIZE(AO), %xmm10
	pxor	%xmm2, %xmm2
	movddup	 8 * SIZE(BO), %xmm11
	pxor	%xmm3, %xmm3

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$3, %rax
	je	.L42

.L41:
	mulpd	%xmm8, %xmm9
	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)
	ADD1	%xmm9, %xmm0
	movddup	 1 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	ADD2	%xmm9, %xmm1
	movddup	 2 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	ADD1	%xmm9, %xmm2
	movddup	 3 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 2 * SIZE(AO), %xmm8
	ADD2	%xmm9, %xmm3
	movddup	 4 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	ADD1	%xmm9, %xmm0
	movddup	 5 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	ADD2	%xmm9, %xmm1
	movddup	 6 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	ADD1	%xmm9, %xmm2
	movddup	 7 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 4 * SIZE(AO), %xmm8
	ADD2	%xmm9, %xmm3
	movddup	16 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm11
	ADD1	%xmm11, %xmm0
	movddup	 9 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	ADD2	%xmm11, %xmm1
	movddup	10 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	ADD1	%xmm11, %xmm2
	movddup	11 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movapd	 6 * SIZE(AO), %xmm8
	ADD2	%xmm11, %xmm3
	movddup	12 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	ADD1	%xmm11, %xmm0
	movddup	13 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	ADD2	%xmm11, %xmm1
	movddup	14 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	ADD1	%xmm11, %xmm2
	movddup	15 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movapd	16 * SIZE(AO), %xmm8
	ADD2	%xmm11, %xmm3
	movddup	24 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm9
	ADD1	%xmm9, %xmm0
	movddup	17 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	ADD2	%xmm9, %xmm1
	movddup	18 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	ADD1	%xmm9, %xmm2
	movddup	19 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	movapd	10 * SIZE(AO), %xmm10
	ADD2	%xmm9, %xmm3
	movddup	20 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	ADD1	%xmm9, %xmm0
	movddup	21 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	ADD2	%xmm9, %xmm1
	movddup	22 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	ADD1	%xmm9, %xmm2
	movddup	23 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	movapd	12 * SIZE(AO), %xmm10
	ADD2	%xmm9, %xmm3
	movddup	32 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm11
	ADD1	%xmm11, %xmm0
	movddup	25 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	ADD2	%xmm11, %xmm1
	movddup	26 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	ADD1	%xmm11, %xmm2
	movddup	27 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movapd	14 * SIZE(AO), %xmm10
	ADD2	%xmm11, %xmm3
	movddup	28 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	ADD1	%xmm11, %xmm0
	movddup	29 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	ADD2	%xmm11, %xmm1
	movddup	30 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	ADD1	%xmm11, %xmm2
	movddup	31 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movapd	24 * SIZE(AO), %xmm10
	ADD2	%xmm11, %xmm3
	movddup	40 * SIZE(BO), %xmm11

	addq   $16 * SIZE, AO
	addq   $32 * SIZE, BO
	decq   %rax
	jne    .L41

.L42:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$7, %rax		# if (k & 1)
	BRANCH
	jle .L44

.L43:
	mulpd	%xmm8, %xmm9
	ADD1	%xmm9, %xmm0
	movddup	 1 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	ADD2	%xmm9, %xmm1
	movddup	 2 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	ADD1	%xmm9, %xmm2
	movddup	 3 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 2 * SIZE(AO), %xmm8
	ADD2	%xmm9, %xmm3
	movddup	 4 * SIZE(BO), %xmm9

	addq	$2 * SIZE, AO		# aoffset  += 4
	addq	$4 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L43
	ALIGN_4

.L44:
	SHUFPD_1 %xmm1, %xmm1
	SHUFPD_1 %xmm3, %xmm3

#ifndef CONJ
	addsubpd	%xmm1, %xmm0
	addsubpd	%xmm3, %xmm2
#else
	addsubpd	%xmm0, %xmm1
	addsubpd	%xmm2, %xmm3
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$1, %rax
#else
	subq	$2, %rax
#endif

	leaq	(, %rax, SIZE), %rax

	movq	AORIG, AO
	leaq	(AO, %rax, 2), AO
	leaq	(B,  %rax, 4), BO
#endif

#if defined(LN) || defined(LT)
	movapd	 0 * SIZE(BO), %xmm8
	movapd	 2 * SIZE(BO), %xmm9
#else
	movapd	 0 * SIZE(AO), %xmm8
	movapd	 2 * SIZE(AO), %xmm9
#endif

#if   (defined(LN) || defined(LT)) && !defined(CONJ)
	subpd	%xmm0,  %xmm8
	subpd	%xmm2,  %xmm9
#elif (defined(LN) || defined(LT)) &&  defined(CONJ)
	subpd	%xmm1,  %xmm8
	subpd	%xmm3,  %xmm9
#elif (defined(RN) || defined(RT)) && !defined(CONJ)
	subpd	%xmm0,  %xmm8
	subpd	%xmm2,  %xmm9
#else
	addsubpd %xmm1, %xmm8
	addsubpd %xmm3, %xmm9
#endif

#ifdef CONJ
	pcmpeqb	%xmm7,  %xmm7
	psllq	$63,    %xmm7
#endif

#ifdef LN
	movddup	 0 * SIZE(AO), %xmm4
	movddup	 1 * SIZE(AO), %xmm5

#ifdef CONJ
	xorpd	%xmm7, %xmm5
#endif

	movapd	%xmm8,  %xmm12
	movapd	%xmm9,  %xmm13
	SHUFPD_1 %xmm12, %xmm12
	SHUFPD_1 %xmm13, %xmm13

	mulpd	 %xmm4, %xmm8
	mulpd	 %xmm4, %xmm9
	mulpd	 %xmm5, %xmm12
	mulpd	 %xmm5, %xmm13

	addsubpd %xmm12, %xmm8
	addsubpd %xmm13, %xmm9
#endif

#ifdef LT
	movddup	 0 * SIZE(AO), %xmm0
	movddup	 1 * SIZE(AO), %xmm1

#ifdef CONJ
	xorpd	%xmm7, %xmm1
#endif

	movapd	%xmm8,  %xmm12
	movapd	%xmm9,  %xmm13
	SHUFPD_1 %xmm12, %xmm12
	SHUFPD_1 %xmm13, %xmm13

	mulpd	 %xmm0, %xmm8
	mulpd	 %xmm0, %xmm9
	mulpd	 %xmm1, %xmm12
	mulpd	 %xmm1, %xmm13

	addsubpd %xmm12, %xmm8
	addsubpd %xmm13, %xmm9
#endif

#ifdef RN
	movddup	 0 * SIZE(BO), %xmm0
	movddup	 1 * SIZE(BO), %xmm1
	movddup	 2 * SIZE(BO), %xmm2
	movddup	 3 * SIZE(BO), %xmm3
	movddup	 6 * SIZE(BO), %xmm4
	movddup	 7 * SIZE(BO), %xmm5

#ifdef CONJ
	xorpd	%xmm7, %xmm1
	xorpd	%xmm7, %xmm3
	xorpd	%xmm7, %xmm5
#endif

	movapd	%xmm8,  %xmm12
	SHUFPD_1 %xmm12, %xmm12

	mulpd	 %xmm0, %xmm8
	mulpd	 %xmm1, %xmm12

	addsubpd %xmm12, %xmm8