|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
/* Copyright 2009, 2010 The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/* All rights reserved. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* Redistribution and use in source and binary forms, with or */
|
|
kusano |
2b45e8 |
/* without modification, are permitted provided that the following */
|
|
kusano |
2b45e8 |
/* conditions are met: */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 1. Redistributions of source code must retain the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* 2. Redistributions in binary form must reproduce the above */
|
|
kusano |
2b45e8 |
/* copyright notice, this list of conditions and the following */
|
|
kusano |
2b45e8 |
/* disclaimer in the documentation and/or other materials */
|
|
kusano |
2b45e8 |
/* provided with the distribution. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
|
|
kusano |
2b45e8 |
/* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
|
|
kusano |
2b45e8 |
/* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
|
|
kusano |
2b45e8 |
/* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
|
|
kusano |
2b45e8 |
/* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
|
|
kusano |
2b45e8 |
/* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
|
|
kusano |
2b45e8 |
/* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
|
|
kusano |
2b45e8 |
/* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
|
|
kusano |
2b45e8 |
/* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
|
|
kusano |
2b45e8 |
/* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
|
|
kusano |
2b45e8 |
/* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
|
|
kusano |
2b45e8 |
/* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
|
|
kusano |
2b45e8 |
/* POSSIBILITY OF SUCH DAMAGE. */
|
|
kusano |
2b45e8 |
/* */
|
|
kusano |
2b45e8 |
/* The views and conclusions contained in the software and */
|
|
kusano |
2b45e8 |
/* documentation are those of the authors and should not be */
|
|
kusano |
2b45e8 |
/* interpreted as representing official policies, either expressed */
|
|
kusano |
2b45e8 |
/* or implied, of The University of Texas at Austin. */
|
|
kusano |
2b45e8 |
/*********************************************************************/
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define ASSEMBLER
|
|
kusano |
2b45e8 |
#include "common.h"
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define OLD_M %rdi
|
|
kusano |
2b45e8 |
#define OLD_N %rsi
|
|
kusano |
2b45e8 |
#define M %r13
|
|
kusano |
2b45e8 |
#define N %r14
|
|
kusano |
2b45e8 |
#define K %rdx
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define A %rcx
|
|
kusano |
2b45e8 |
#define B %r8
|
|
kusano |
2b45e8 |
#define C %r9
|
|
kusano |
2b45e8 |
#define LDC %r10
|
|
kusano |
2b45e8 |
#define I %r11
|
|
kusano |
2b45e8 |
#define AO %rdi
|
|
kusano |
2b45e8 |
#define BO %rsi
|
|
kusano |
2b45e8 |
#define CO1 %r15
|
|
kusano |
2b45e8 |
#define CO2 %rbp
|
|
kusano |
2b45e8 |
#define BB %r12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef WINDOWS_ABI
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define STACKSIZE 64
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define STACKSIZE 256
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define OLD_ALPHA_I 40 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
#define OLD_A 48 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
#define OLD_B 56 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
#define OLD_C 64 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
#define OLD_LDC 72 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
#define OLD_OFFSET 80 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define POSINV 0(%rsp)
|
|
kusano |
2b45e8 |
#define ALPHA_R 16(%rsp)
|
|
kusano |
2b45e8 |
#define ALPHA_I 32(%rsp)
|
|
kusano |
2b45e8 |
#define J 48(%rsp)
|
|
kusano |
2b45e8 |
#define OFFSET 56(%rsp)
|
|
kusano |
2b45e8 |
#define KK 64(%rsp)
|
|
kusano |
2b45e8 |
#define KKK 72(%rsp)
|
|
kusano |
2b45e8 |
#define BUFFER 256(%rsp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef OPTERON
|
|
kusano |
2b45e8 |
#define PREFETCH prefetch
|
|
kusano |
2b45e8 |
#define PREFETCHW prefetchw
|
|
kusano |
2b45e8 |
#define PREFETCHSIZE (8 * 9 + 4)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define RPREFETCHSIZE (8 * 7 + 4)
|
|
kusano |
2b45e8 |
#define WPREFETCHSIZE (8 * 8 + 4)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef GENERIC
|
|
kusano |
2b45e8 |
#define PREFETCH prefetcht0
|
|
kusano |
2b45e8 |
#define PREFETCHW prefetcht0
|
|
kusano |
2b45e8 |
#define PREFETCHSIZE (8 * 5 + 4)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define RPREFETCHSIZE (8 * 7 + 4)
|
|
kusano |
2b45e8 |
#define WPREFETCHSIZE (8 * 8 + 4)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef GENERIC
|
|
kusano |
2b45e8 |
#define KERNEL1(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1 ;\
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8 ;\
|
|
kusano |
2b45e8 |
movapd -16 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm1 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9 ;\
|
|
kusano |
2b45e8 |
movapd -14 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm5 ;\
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE + 1 * (xx) * SIZE(AO, %rax, 4) ;\
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm0 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm10 ;\
|
|
kusano |
2b45e8 |
movapd -12 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11 ;\
|
|
kusano |
2b45e8 |
movapd -8 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL2(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1 ;\
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm12 ;\
|
|
kusano |
2b45e8 |
movapd 0 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm1 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm13 ;\
|
|
kusano |
2b45e8 |
movapd -6 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm2 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm14 ;\
|
|
kusano |
2b45e8 |
movapd -4 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm15 ;\
|
|
kusano |
2b45e8 |
movapd -6 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL3(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm7 ;\
|
|
kusano |
2b45e8 |
addpd %xmm7, %xmm8 ;\
|
|
kusano |
2b45e8 |
movapd -8 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm7 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9 ;\
|
|
kusano |
2b45e8 |
movapd -6 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm4 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm10 ;\
|
|
kusano |
2b45e8 |
movapd -4 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm4, %xmm11 ;\
|
|
kusano |
2b45e8 |
movapd -4 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL4(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm7 ;\
|
|
kusano |
2b45e8 |
addpd %xmm7, %xmm12 ;\
|
|
kusano |
2b45e8 |
movapd 8 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm7 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm13 ;\
|
|
kusano |
2b45e8 |
movapd 2 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm6 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm14 ;\
|
|
kusano |
2b45e8 |
movapd 4 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 8) * SIZE + 1 * (xx) * SIZE(AO, %rax, 4) ;\
|
|
kusano |
2b45e8 |
addpd %xmm6, %xmm15 ;\
|
|
kusano |
2b45e8 |
movapd -2 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL5(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1 ;\
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8 ;\
|
|
kusano |
2b45e8 |
movapd 0 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm1 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9 ;\
|
|
kusano |
2b45e8 |
movapd 2 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd 6 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm0 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm10 ;\
|
|
kusano |
2b45e8 |
movapd 4 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11 ;\
|
|
kusano |
2b45e8 |
movapd 0 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL6(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1 ;\
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm12 ;\
|
|
kusano |
2b45e8 |
movapd 16 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm1 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm13 ;\
|
|
kusano |
2b45e8 |
movapd 10 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd 6 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm2 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm14 ;\
|
|
kusano |
2b45e8 |
movapd 12 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm15 ;\
|
|
kusano |
2b45e8 |
movapd 2 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL7(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm7 ;\
|
|
kusano |
2b45e8 |
addpd %xmm7, %xmm8 ;\
|
|
kusano |
2b45e8 |
movapd 8 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm7 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9 ;\
|
|
kusano |
2b45e8 |
movapd 10 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd 14 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm4 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm10 ;\
|
|
kusano |
2b45e8 |
movapd 12 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm4, %xmm11 ;\
|
|
kusano |
2b45e8 |
movapd 4 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL8(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm7 ;\
|
|
kusano |
2b45e8 |
addpd %xmm7, %xmm12 ;\
|
|
kusano |
2b45e8 |
movapd 24 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm7 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm13 ;\
|
|
kusano |
2b45e8 |
movapd 18 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd 14 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm6 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm14 ;\
|
|
kusano |
2b45e8 |
movapd 20 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm6, %xmm15 ;\
|
|
kusano |
2b45e8 |
movapd 6 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#define KERNEL1(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1 ;\
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8 ;\
|
|
kusano |
2b45e8 |
movapd -16 * SIZE + 2 * (xx) * SIZE(BO), %xmm1 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9 ;\
|
|
kusano |
2b45e8 |
movapd -14 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm5 ;\
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE + 1 * (xx) * SIZE(AO) ;\
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE + 2 * (xx) * SIZE(BO), %xmm0 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm10 ;\
|
|
kusano |
2b45e8 |
movapd -12 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11 ;\
|
|
kusano |
2b45e8 |
movapd -8 * SIZE + 1 * (xx) * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL2(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1 ;\
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm12 ;\
|
|
kusano |
2b45e8 |
movapd 0 * SIZE + 2 * (xx) * SIZE(BO), %xmm1 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm13 ;\
|
|
kusano |
2b45e8 |
movapd -6 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE + 2 * (xx) * SIZE(BO), %xmm2 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm14 ;\
|
|
kusano |
2b45e8 |
movapd -4 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm15 ;\
|
|
kusano |
2b45e8 |
movapd -6 * SIZE + 1 * (xx) * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL3(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm7 ;\
|
|
kusano |
2b45e8 |
addpd %xmm7, %xmm8 ;\
|
|
kusano |
2b45e8 |
movapd -8 * SIZE + 2 * (xx) * SIZE(BO), %xmm7 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9 ;\
|
|
kusano |
2b45e8 |
movapd -6 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE + 2 * (xx) * SIZE(BO), %xmm4 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm10 ;\
|
|
kusano |
2b45e8 |
movapd -4 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm4, %xmm11 ;\
|
|
kusano |
2b45e8 |
movapd -4 * SIZE + 1 * (xx) * SIZE(AO), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL4(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm7 ;\
|
|
kusano |
2b45e8 |
addpd %xmm7, %xmm12 ;\
|
|
kusano |
2b45e8 |
movapd 8 * SIZE + 2 * (xx) * SIZE(BO), %xmm7 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm13 ;\
|
|
kusano |
2b45e8 |
movapd 2 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE + 2 * (xx) * SIZE(BO), %xmm6 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm14 ;\
|
|
kusano |
2b45e8 |
movapd 4 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 8) * SIZE + 1 * (xx) * SIZE(AO) ;\
|
|
kusano |
2b45e8 |
addpd %xmm6, %xmm15 ;\
|
|
kusano |
2b45e8 |
movapd -2 * SIZE + 1 * (xx) * SIZE(AO), %xmm6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL5(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1 ;\
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8 ;\
|
|
kusano |
2b45e8 |
movapd 0 * SIZE + 2 * (xx) * SIZE(BO), %xmm1 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9 ;\
|
|
kusano |
2b45e8 |
movapd 2 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd 6 * SIZE + 2 * (xx) * SIZE(BO), %xmm0 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm10 ;\
|
|
kusano |
2b45e8 |
movapd 4 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11 ;\
|
|
kusano |
2b45e8 |
movapd 0 * SIZE + 1 * (xx) * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL6(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1 ;\
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm12 ;\
|
|
kusano |
2b45e8 |
movapd 16 * SIZE + 2 * (xx) * SIZE(BO), %xmm1 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm13 ;\
|
|
kusano |
2b45e8 |
movapd 10 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd 6 * SIZE + 2 * (xx) * SIZE(BO), %xmm2 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm14 ;\
|
|
kusano |
2b45e8 |
movapd 12 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm15 ;\
|
|
kusano |
2b45e8 |
movapd 2 * SIZE + 1 * (xx) * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL7(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm7 ;\
|
|
kusano |
2b45e8 |
addpd %xmm7, %xmm8 ;\
|
|
kusano |
2b45e8 |
movapd 8 * SIZE + 2 * (xx) * SIZE(BO), %xmm7 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9 ;\
|
|
kusano |
2b45e8 |
movapd 10 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd 14 * SIZE + 2 * (xx) * SIZE(BO), %xmm4 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm10 ;\
|
|
kusano |
2b45e8 |
movapd 12 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm4, %xmm11 ;\
|
|
kusano |
2b45e8 |
movapd 4 * SIZE + 1 * (xx) * SIZE(AO), %xmm4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL8(xx) \
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm7 ;\
|
|
kusano |
2b45e8 |
addpd %xmm7, %xmm12 ;\
|
|
kusano |
2b45e8 |
movapd 24 * SIZE + 2 * (xx) * SIZE(BO), %xmm7 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm3 ;\
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm13 ;\
|
|
kusano |
2b45e8 |
movapd 18 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm5 ;\
|
|
kusano |
2b45e8 |
mulpd 14 * SIZE + 2 * (xx) * SIZE(BO), %xmm6 ;\
|
|
kusano |
2b45e8 |
addpd %xmm5, %xmm14 ;\
|
|
kusano |
2b45e8 |
movapd 20 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
|
|
kusano |
2b45e8 |
addpd %xmm6, %xmm15 ;\
|
|
kusano |
2b45e8 |
movapd 6 * SIZE + 1 * (xx) * SIZE(AO), %xmm6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROLOGUE
|
|
kusano |
2b45e8 |
PROFCODE
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subq $STACKSIZE, %rsp
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq %rbx, 0(%rsp)
|
|
kusano |
2b45e8 |
movq %rbp, 8(%rsp)
|
|
kusano |
2b45e8 |
movq %r12, 16(%rsp)
|
|
kusano |
2b45e8 |
movq %r13, 24(%rsp)
|
|
kusano |
2b45e8 |
movq %r14, 32(%rsp)
|
|
kusano |
2b45e8 |
movq %r15, 40(%rsp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef WINDOWS_ABI
|
|
kusano |
2b45e8 |
movq %rdi, 48(%rsp)
|
|
kusano |
2b45e8 |
movq %rsi, 56(%rsp)
|
|
kusano |
2b45e8 |
movups %xmm6, 64(%rsp)
|
|
kusano |
2b45e8 |
movups %xmm7, 80(%rsp)
|
|
kusano |
2b45e8 |
movups %xmm8, 96(%rsp)
|
|
kusano |
2b45e8 |
movups %xmm9, 112(%rsp)
|
|
kusano |
2b45e8 |
movups %xmm10, 128(%rsp)
|
|
kusano |
2b45e8 |
movups %xmm11, 144(%rsp)
|
|
kusano |
2b45e8 |
movups %xmm12, 160(%rsp)
|
|
kusano |
2b45e8 |
movups %xmm13, 176(%rsp)
|
|
kusano |
2b45e8 |
movups %xmm14, 192(%rsp)
|
|
kusano |
2b45e8 |
movups %xmm15, 208(%rsp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq ARG1, OLD_M
|
|
kusano |
2b45e8 |
movq ARG2, OLD_N
|
|
kusano |
2b45e8 |
movq ARG3, K
|
|
kusano |
2b45e8 |
movq OLD_A, A
|
|
kusano |
2b45e8 |
movq OLD_B, B
|
|
kusano |
2b45e8 |
movq OLD_C, C
|
|
kusano |
2b45e8 |
movq OLD_LDC, LDC
|
|
kusano |
2b45e8 |
#ifdef TRMMKERNEL
|
|
kusano |
2b45e8 |
movsd OLD_OFFSET, %xmm12
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movaps %xmm3, %xmm0
|
|
kusano |
2b45e8 |
movsd OLD_ALPHA_I, %xmm1
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq 72(%rsp), LDC
|
|
kusano |
2b45e8 |
#ifdef TRMMKERNEL
|
|
kusano |
2b45e8 |
movsd 80(%rsp), %xmm12
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
EMMS
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq %rsp, %rbx # save old stack
|
|
kusano |
2b45e8 |
subq $256 + LOCAL_BUFFER_SIZE, %rsp
|
|
kusano |
2b45e8 |
andq $-4096, %rsp # align stack
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
STACK_TOUCHING
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq OLD_M, M
|
|
kusano |
2b45e8 |
movq OLD_N, N
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pcmpeqb %xmm7, %xmm7
|
|
kusano |
2b45e8 |
psllq $63, %xmm7 # Generate mask
|
|
kusano |
2b45e8 |
pxor %xmm10, %xmm10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm0, 0 + ALPHA_R
|
|
kusano |
2b45e8 |
movlpd %xmm0, 8 + ALPHA_R
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm1, 8 + ALPHA_I
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
movlpd %xmm1, 0 + ALPHA_I
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm10, 0 + POSINV
|
|
kusano |
2b45e8 |
movlpd %xmm7, 8 + POSINV
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef TRMMKERNEL
|
|
kusano |
2b45e8 |
movlpd %xmm12, OFFSET
|
|
kusano |
2b45e8 |
movlpd %xmm12, KK
|
|
kusano |
2b45e8 |
#ifndef LEFT
|
|
kusano |
2b45e8 |
negq KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subq $-16 * SIZE, A
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
salq $ZBASE_SHIFT, LDC
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq N, J
|
|
kusano |
2b45e8 |
sarq $1, J # j = (n >> 2)
|
|
kusano |
2b45e8 |
jle .L100
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L01:
|
|
kusano |
2b45e8 |
movq C, CO1 # coffset1 = c
|
|
kusano |
2b45e8 |
leaq (C, LDC, 1), CO2 # coffset2 = c + ldc
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(TRMMKERNEL) && defined(LEFT)
|
|
kusano |
2b45e8 |
movq OFFSET, %rax
|
|
kusano |
2b45e8 |
movq %rax, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq 16 * SIZE + BUFFER, BO
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
sarq $2, %rax
|
|
kusano |
2b45e8 |
jle .L03
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L02:
|
|
kusano |
2b45e8 |
PREFETCH (RPREFETCHSIZE + 0) * SIZE(B)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq 0 * SIZE(B), %mm0
|
|
kusano |
2b45e8 |
movq %mm0, -16 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm0, -15 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq 1 * SIZE(B), %mm1
|
|
kusano |
2b45e8 |
movq %mm1, -14 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm1, -13 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq 2 * SIZE(B), %mm2
|
|
kusano |
2b45e8 |
movq %mm2, -12 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm2, -11 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq 3 * SIZE(B), %mm3
|
|
kusano |
2b45e8 |
movq %mm3, -10 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm3, -9 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PREFETCHW (WPREFETCHSIZE + 0) * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq 4 * SIZE(B), %mm4
|
|
kusano |
2b45e8 |
movq %mm4, -8 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm4, -7 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq 5 * SIZE(B), %mm5
|
|
kusano |
2b45e8 |
movq %mm5, -6 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm5, -5 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PREFETCHW (WPREFETCHSIZE + 8) * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq 6 * SIZE(B), %mm6
|
|
kusano |
2b45e8 |
movq %mm6, -4 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm6, -3 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq 7 * SIZE(B), %mm7
|
|
kusano |
2b45e8 |
movq %mm7, -2 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm7, -1 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PREFETCH (RPREFETCHSIZE + 8) * SIZE(B)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq 8 * SIZE(B), %mm0
|
|
kusano |
2b45e8 |
movq %mm0, 0 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm0, 1 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq 9 * SIZE(B), %mm1
|
|
kusano |
2b45e8 |
movq %mm1, 2 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm1, 3 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq 10 * SIZE(B), %mm2
|
|
kusano |
2b45e8 |
movq %mm2, 4 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm2, 5 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq 11 * SIZE(B), %mm3
|
|
kusano |
2b45e8 |
movq %mm3, 6 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm3, 7 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PREFETCHW (WPREFETCHSIZE + 16) * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq 12 * SIZE(B), %mm4
|
|
kusano |
2b45e8 |
movq %mm4, 8 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm4, 9 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq 13 * SIZE(B), %mm5
|
|
kusano |
2b45e8 |
movq %mm5, 10 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm5, 11 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PREFETCHW (WPREFETCHSIZE + 24) * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq 14 * SIZE(B), %mm6
|
|
kusano |
2b45e8 |
movq %mm6, 12 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm6, 13 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq 15 * SIZE(B), %mm7
|
|
kusano |
2b45e8 |
movq %mm7, 14 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm7, 15 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $ 32 * SIZE, BO
|
|
kusano |
2b45e8 |
subq $-16 * SIZE, B
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jne .L02
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L03:
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
andq $3, %rax
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L05
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L04:
|
|
kusano |
2b45e8 |
movq 0 * SIZE(B), %mm0
|
|
kusano |
2b45e8 |
movq %mm0, -16 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm0, -15 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq 1 * SIZE(B), %mm1
|
|
kusano |
2b45e8 |
movq %mm1, -14 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm1, -13 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq 2 * SIZE(B), %mm2
|
|
kusano |
2b45e8 |
movq %mm2, -12 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm2, -11 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq 3 * SIZE(B), %mm3
|
|
kusano |
2b45e8 |
movq %mm3, -10 * SIZE(BO)
|
|
kusano |
2b45e8 |
movq %mm3, -9 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $ 4 * SIZE, B
|
|
kusano |
2b45e8 |
addq $ 8 * SIZE, BO
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jne .L04
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L05:
|
|
kusano |
2b45e8 |
movq A, AO # aoffset = a
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq (RPREFETCHSIZE + 0) * SIZE(B), BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq M, I
|
|
kusano |
2b45e8 |
sarq $1, I # i = (m >> 2)
|
|
kusano |
2b45e8 |
jle .L30
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L10:
|
|
kusano |
2b45e8 |
#if !defined(TRMMKERNEL) || \
|
|
kusano |
2b45e8 |
(defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
|
|
kusano |
2b45e8 |
(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq 16 * SIZE + BUFFER, BO
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
leaq 16 * SIZE + BUFFER, BO
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 4), AO
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 8), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm8, %xmm8
|
|
kusano |
2b45e8 |
PREFETCH 0 * SIZE(BB)
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
pxor %xmm9, %xmm9
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AO), %xmm4
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO), %xmm5
|
|
kusano |
2b45e8 |
pxor %xmm10, %xmm10
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(AO), %xmm6
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(BO), %xmm7
|
|
kusano |
2b45e8 |
pxor %xmm11, %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pxor %xmm12, %xmm12
|
|
kusano |
2b45e8 |
PREFETCHW 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
pxor %xmm13, %xmm13
|
|
kusano |
2b45e8 |
PREFETCHW 3 * SIZE(CO2)
|
|
kusano |
2b45e8 |
pxor %xmm14, %xmm14
|
|
kusano |
2b45e8 |
pxor %xmm15, %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
movq %rax, KKK
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#ifdef LEFT
|
|
kusano |
2b45e8 |
addq $2, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addq $2, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movq %rax, KKK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#ifndef GENERIC
|
|
kusano |
2b45e8 |
andq $-8, %rax
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 4), AO
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 8), BO
|
|
kusano |
2b45e8 |
negq %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
je .L15
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L12:
|
|
kusano |
2b45e8 |
KERNEL1(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
KERNEL1(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $8 * SIZE, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
je .L15
|
|
kusano |
2b45e8 |
KERNEL1(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
KERNEL1(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $8 * SIZE, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
je .L15
|
|
kusano |
2b45e8 |
KERNEL1(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
KERNEL1(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $8 * SIZE, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
je .L15
|
|
kusano |
2b45e8 |
KERNEL1(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
KERNEL1(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $8 * SIZE, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
je .L15
|
|
kusano |
2b45e8 |
KERNEL1(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
KERNEL1(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $8 * SIZE, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
je .L15
|
|
kusano |
2b45e8 |
KERNEL1(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
KERNEL1(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $8 * SIZE, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
je .L15
|
|
kusano |
2b45e8 |
KERNEL1(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
KERNEL1(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $8 * SIZE, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
je .L15
|
|
kusano |
2b45e8 |
KERNEL1(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
KERNEL1(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $8 * SIZE, %rax
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jl .L12
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L15:
|
|
kusano |
2b45e8 |
PREFETCH 8 * SIZE(BB)
|
|
kusano |
2b45e8 |
subq $-16 * SIZE, BB
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq KKK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
testq $4, %rax
|
|
kusano |
2b45e8 |
je .L16
|
|
kusano |
2b45e8 |
xorq %rax, %rax
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
KERNEL1(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $32 * SIZE, BO
|
|
kusano |
2b45e8 |
addq $16 * SIZE, AO
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
sarq $2, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
jle .L16
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L12:
|
|
kusano |
2b45e8 |
KERNEL1(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL2(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL3(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL4(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL5(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL6(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL7(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL8(16 * 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $ 32 * SIZE, BO
|
|
kusano |
2b45e8 |
subq $-16 * SIZE, AO
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jg .L12
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L16:
|
|
kusano |
2b45e8 |
movapd POSINV, %xmm5
|
|
kusano |
2b45e8 |
movapd ALPHA_R, %xmm6
|
|
kusano |
2b45e8 |
movapd ALPHA_I, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq KKK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andq $3, %rax # if (k & 1)
|
|
kusano |
2b45e8 |
je .L19
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 4), AO
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 8), BO
|
|
kusano |
2b45e8 |
negq %rax
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L17:
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BO, %rax, 8), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm9
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO, %rax, 8), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE(BO, %rax, 8), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BO, %rax, 8), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AO, %rax, 4), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BO, %rax, 8), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm13
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO, %rax, 8), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE(BO, %rax, 8), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm14
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(BO, %rax, 8), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm15
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(AO, %rax, 4), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $SIZE, %rax
|
|
kusano |
2b45e8 |
jl .L17
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L19:
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movlpd 0 * SIZE(CO1), %xmm0
|
|
kusano |
2b45e8 |
movhpd 1 * SIZE(CO1), %xmm0
|
|
kusano |
2b45e8 |
movlpd 2 * SIZE(CO1), %xmm2
|
|
kusano |
2b45e8 |
movhpd 3 * SIZE(CO1), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd 0 * SIZE(CO2), %xmm1
|
|
kusano |
2b45e8 |
movhpd 1 * SIZE(CO2), %xmm1
|
|
kusano |
2b45e8 |
movlpd 2 * SIZE(CO2), %xmm3
|
|
kusano |
2b45e8 |
movhpd 3 * SIZE(CO2), %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm9, %xmm9
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm11, %xmm11
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm15, %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
|
|
kusano |
2b45e8 |
defined(NR) || defined(NC) || defined(TR) || defined(TC)
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm9
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm11
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm13
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm15
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm8
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm10
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm14
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
|
|
kusano |
2b45e8 |
defined(RR) || defined(RC) || defined(CR) || defined(CC)
|
|
kusano |
2b45e8 |
subpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm11, %xmm10
|
|
kusano |
2b45e8 |
subpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
subpd %xmm15, %xmm14
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
addpd %xmm11, %xmm10
|
|
kusano |
2b45e8 |
addpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
addpd %xmm15, %xmm14
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x4e, %xmm8, %xmm9
|
|
kusano |
2b45e8 |
pshufd $0x4e, %xmm10, %xmm11
|
|
kusano |
2b45e8 |
pshufd $0x4e, %xmm12, %xmm13
|
|
kusano |
2b45e8 |
pshufd $0x4e, %xmm14, %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm10
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm13
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm14
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
addpd %xmm11, %xmm10
|
|
kusano |
2b45e8 |
addpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
addpd %xmm15, %xmm14
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm12
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm14
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movlpd %xmm12, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm12, 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm10, 0 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movhpd %xmm10, 1 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movlpd %xmm14, 2 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movhpd %xmm14, 3 * SIZE(CO2)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
|
|
kusano |
2b45e8 |
(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KKK, %rax
|
|
kusano |
2b45e8 |
leaq (,%rax, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 4), AO
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 8), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(TRMMKERNEL) && defined(LEFT)
|
|
kusano |
2b45e8 |
addq $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $4 * SIZE, CO1 # coffset += 4
|
|
kusano |
2b45e8 |
addq $4 * SIZE, CO2 # coffset += 4
|
|
kusano |
2b45e8 |
decq I # i --
|
|
kusano |
2b45e8 |
jg .L10
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L30:
|
|
kusano |
2b45e8 |
testq $1, M
|
|
kusano |
2b45e8 |
jle .L99
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if !defined(TRMMKERNEL) || \
|
|
kusano |
2b45e8 |
(defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
|
|
kusano |
2b45e8 |
(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq 16 * SIZE + BUFFER, BO
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
leaq 16 * SIZE + BUFFER, BO
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 2), AO
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 8), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm8, %xmm8
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
pxor %xmm9, %xmm9
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm10, %xmm10
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
pxor %xmm11, %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
movq %rax, KKK
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#ifdef LEFT
|
|
kusano |
2b45e8 |
addq $1, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addq $2, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movq %rax, KKK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarq $3, %rax
|
|
kusano |
2b45e8 |
je .L44
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L41:
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm9
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm8
|
|
kusano |
2b45e8 |
movapd -6 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm10
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 8) * SIZE(AO)
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm9
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd 6 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
movapd 16 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm8
|
|
kusano |
2b45e8 |
movapd 10 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9
|
|
kusano |
2b45e8 |
movapd 12 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulpd 14 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm10
|
|
kusano |
2b45e8 |
movapd 24 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd 18 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm9
|
|
kusano |
2b45e8 |
movapd 20 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
mulpd 22 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
movapd 32 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm11
|
|
kusano |
2b45e8 |
movapd -6 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm8
|
|
kusano |
2b45e8 |
movapd 26 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9
|
|
kusano |
2b45e8 |
movapd 28 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulpd 30 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm10
|
|
kusano |
2b45e8 |
movapd 40 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm11
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd 34 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm9
|
|
kusano |
2b45e8 |
movapd 36 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
mulpd 38 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
movapd 48 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm11
|
|
kusano |
2b45e8 |
movapd -2 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm8
|
|
kusano |
2b45e8 |
movapd 42 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9
|
|
kusano |
2b45e8 |
movapd 44 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulpd 46 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm10
|
|
kusano |
2b45e8 |
movapd 56 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm11
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subq $-16 * SIZE, AO
|
|
kusano |
2b45e8 |
addq $64 * SIZE, BO
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jne .L41
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L44:
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq KKK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andq $4, %rax
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L45
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm9
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm8
|
|
kusano |
2b45e8 |
movapd -6 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm10
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm9
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd 6 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
movapd 16 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm8
|
|
kusano |
2b45e8 |
movapd 10 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm9
|
|
kusano |
2b45e8 |
movapd 12 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulpd 14 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm10
|
|
kusano |
2b45e8 |
movapd 24 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $ 8 * SIZE, AO
|
|
kusano |
2b45e8 |
addq $32 * SIZE, BO
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L45:
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq KKK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movapd POSINV, %xmm5
|
|
kusano |
2b45e8 |
movapd ALPHA_R, %xmm6
|
|
kusano |
2b45e8 |
movapd ALPHA_I, %xmm7
|
|
kusano |
2b45e8 |
andq $3, %rax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L47
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L46:
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm9
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $2 * SIZE, AO
|
|
kusano |
2b45e8 |
addq $8 * SIZE, BO
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jg .L46
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L47:
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movlpd 0 * SIZE(CO1), %xmm0
|
|
kusano |
2b45e8 |
movhpd 1 * SIZE(CO1), %xmm0
|
|
kusano |
2b45e8 |
movlpd 0 * SIZE(CO2), %xmm1
|
|
kusano |
2b45e8 |
movhpd 1 * SIZE(CO2), %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm9, %xmm9
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm11, %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
|
|
kusano |
2b45e8 |
defined(NR) || defined(NC) || defined(TR) || defined(TC)
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm9
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm11
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm8
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm10
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
|
|
kusano |
2b45e8 |
defined(RR) || defined(RC) || defined(CR) || defined(CC)
|
|
kusano |
2b45e8 |
subpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm11, %xmm10
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
addpd %xmm11, %xmm10
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x4e, %xmm8, %xmm9
|
|
kusano |
2b45e8 |
pshufd $0x4e, %xmm10, %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm10
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
addpd %xmm11, %xmm10
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movlpd %xmm10, 0 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movhpd %xmm10, 1 * SIZE(CO2)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
|
|
kusano |
2b45e8 |
(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KKK, %rax
|
|
kusano |
2b45e8 |
leaq (,%rax, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 2), AO
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 8), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(TRMMKERNEL) && defined(LEFT)
|
|
kusano |
2b45e8 |
addq $1, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L99:
|
|
kusano |
2b45e8 |
#if defined(TRMMKERNEL) && !defined(LEFT)
|
|
kusano |
2b45e8 |
addl $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq (C, LDC, 2), C # c += 2 * ldc
|
|
kusano |
2b45e8 |
decq J # j --
|
|
kusano |
2b45e8 |
jg .L01
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L100:
|
|
kusano |
2b45e8 |
testq $1, N
|
|
kusano |
2b45e8 |
jle .L999
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L101:
|
|
kusano |
2b45e8 |
#if defined(TRMMKERNEL) && defined(LEFT)
|
|
kusano |
2b45e8 |
movq OFFSET, %rax
|
|
kusano |
2b45e8 |
movq %rax, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
/* Copying to Sub Buffer */
|
|
kusano |
2b45e8 |
leaq BUFFER, BO
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
sarq $2, %rax
|
|
kusano |
2b45e8 |
jle .L103
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L102:
|
|
kusano |
2b45e8 |
movlpd 0 * SIZE(B), %xmm8
|
|
kusano |
2b45e8 |
movlpd 1 * SIZE(B), %xmm9
|
|
kusano |
2b45e8 |
movlpd 2 * SIZE(B), %xmm10
|
|
kusano |
2b45e8 |
movlpd 3 * SIZE(B), %xmm11
|
|
kusano |
2b45e8 |
movlpd 4 * SIZE(B), %xmm12
|
|
kusano |
2b45e8 |
movlpd 5 * SIZE(B), %xmm13
|
|
kusano |
2b45e8 |
movlpd 6 * SIZE(B), %xmm14
|
|
kusano |
2b45e8 |
movlpd 7 * SIZE(B), %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm8, 0 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm8, 1 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm9, 2 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm9, 3 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm10, 4 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm10, 5 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm11, 6 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm11, 7 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm12, 8 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm12, 9 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm13, 10 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm13, 11 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm14, 12 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm14, 13 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm15, 14 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm15, 15 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subq $-16 * SIZE, BO
|
|
kusano |
2b45e8 |
addq $ 8 * SIZE, B
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jne .L102
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L103:
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
andq $3, %rax
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L105
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L104:
|
|
kusano |
2b45e8 |
movlpd 0 * SIZE(B), %xmm8
|
|
kusano |
2b45e8 |
movlpd 1 * SIZE(B), %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm8, 0 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm8, 1 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm9, 2 * SIZE(BO)
|
|
kusano |
2b45e8 |
movlpd %xmm9, 3 * SIZE(BO)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $4 * SIZE, BO
|
|
kusano |
2b45e8 |
addq $2 * SIZE, B
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jne .L104
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L105:
|
|
kusano |
2b45e8 |
movq C, CO1 # coffset1 = c
|
|
kusano |
2b45e8 |
movq A, AO # aoffset = a
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq M, I
|
|
kusano |
2b45e8 |
sarq $1, I # i = (m >> 2)
|
|
kusano |
2b45e8 |
jle .L130
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L110:
|
|
kusano |
2b45e8 |
#if !defined(TRMMKERNEL) || \
|
|
kusano |
2b45e8 |
(defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
|
|
kusano |
2b45e8 |
(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq 16 * SIZE + BUFFER, BO
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
leaq 16 * SIZE + BUFFER, BO
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 4), AO
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 4), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
pxor %xmm8, %xmm8
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
pxor %xmm9, %xmm9
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
pxor %xmm12, %xmm12
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
pxor %xmm13, %xmm13
|
|
kusano |
2b45e8 |
PREFETCHW 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
movq %rax, KKK
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#ifdef LEFT
|
|
kusano |
2b45e8 |
addq $2, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addq $1, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movq %rax, KKK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarq $2, %rax
|
|
kusano |
2b45e8 |
je .L112
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L111:
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
|
|
kusano |
2b45e8 |
mulpd -14 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -14 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm13
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm13
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 8) * SIZE(AO)
|
|
kusano |
2b45e8 |
mulpd -6 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm8
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm9
|
|
kusano |
2b45e8 |
movapd -6 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulpd -6 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm12
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm13
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm8
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm9
|
|
kusano |
2b45e8 |
movapd -2 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm12
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm13
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subq $-16 * SIZE, AO
|
|
kusano |
2b45e8 |
subq $-16 * SIZE, BO
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jne .L111
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L112:
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq KKK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movapd POSINV, %xmm5
|
|
kusano |
2b45e8 |
movapd ALPHA_R, %xmm6
|
|
kusano |
2b45e8 |
movapd ALPHA_I, %xmm7
|
|
kusano |
2b45e8 |
andq $3, %rax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L114
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L113:
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -14 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -14 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm13
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $4 * SIZE, AO # aoffset += 4
|
|
kusano |
2b45e8 |
addq $4 * SIZE, BO # boffset1 += 8
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jg .L113
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L114:
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movlpd 0 * SIZE(CO1), %xmm0
|
|
kusano |
2b45e8 |
movhpd 1 * SIZE(CO1), %xmm0
|
|
kusano |
2b45e8 |
movlpd 2 * SIZE(CO1), %xmm2
|
|
kusano |
2b45e8 |
movhpd 3 * SIZE(CO1), %xmm2
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm9, %xmm9
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
|
|
kusano |
2b45e8 |
defined(NR) || defined(NC) || defined(TR) || defined(TC)
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm9
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm13
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm8
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
|
|
kusano |
2b45e8 |
defined(RR) || defined(RC) || defined(CR) || defined(CC)
|
|
kusano |
2b45e8 |
subpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
addpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x4e, %xmm8, %xmm9
|
|
kusano |
2b45e8 |
pshufd $0x4e, %xmm12, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
addpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm12
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movlpd %xmm12, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm12, 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
|
|
kusano |
2b45e8 |
(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KKK, %rax
|
|
kusano |
2b45e8 |
leaq (,%rax, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 4), AO
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 4), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(TRMMKERNEL) && defined(LEFT)
|
|
kusano |
2b45e8 |
addq $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $4 * SIZE, CO1 # coffset += 4
|
|
kusano |
2b45e8 |
decq I # i --
|
|
kusano |
2b45e8 |
jg .L110
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L130:
|
|
kusano |
2b45e8 |
testq $1, M
|
|
kusano |
2b45e8 |
jle .L999
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if !defined(TRMMKERNEL) || \
|
|
kusano |
2b45e8 |
(defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
|
|
kusano |
2b45e8 |
(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq 16 * SIZE + BUFFER, BO
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
leaq 16 * SIZE + BUFFER, BO
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 2), AO
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 4), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
movapd -16 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pxor %xmm8, %xmm8
|
|
kusano |
2b45e8 |
pxor %xmm9, %xmm9
|
|
kusano |
2b45e8 |
pxor %xmm10, %xmm10
|
|
kusano |
2b45e8 |
pxor %xmm11, %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
movq %rax, KKK
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#ifdef LEFT
|
|
kusano |
2b45e8 |
addq $1, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addq $1, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movq %rax, KKK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarq $3, %rax
|
|
kusano |
2b45e8 |
je .L144
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L141:
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
|
|
kusano |
2b45e8 |
mulpd -14 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulpd -6 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm8
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm10
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 8) * SIZE(AO)
|
|
kusano |
2b45e8 |
mulpd 2 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm9
|
|
kusano |
2b45e8 |
movapd -6 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm1
|
|
kusano |
2b45e8 |
mulpd 6 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
movapd 16 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm11
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulpd 10 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm8
|
|
kusano |
2b45e8 |
movapd 12 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm9
|
|
kusano |
2b45e8 |
movapd -2 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
mulpd 14 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm10
|
|
kusano |
2b45e8 |
movapd 24 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm11
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subq $-16 * SIZE, AO
|
|
kusano |
2b45e8 |
subq $-32 * SIZE, BO
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jne .L141
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L144:
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq KKK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andq $4, %rax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L145
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -14 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -10 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm10
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulpd -6 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm8
|
|
kusano |
2b45e8 |
movapd -4 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
movapd -10 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm3
|
|
kusano |
2b45e8 |
mulpd -2 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm10
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
movapd -8 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $8 * SIZE, AO
|
|
kusano |
2b45e8 |
subq $-16 * SIZE, BO
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L145:
|
|
kusano |
2b45e8 |
movapd POSINV, %xmm5
|
|
kusano |
2b45e8 |
movapd ALPHA_R, %xmm6
|
|
kusano |
2b45e8 |
movapd ALPHA_I, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq KKK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andq $3, %rax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L148
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L146:
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
mulpd -14 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
movapd -12 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
movapd -14 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $2 * SIZE, AO # aoffset += 4
|
|
kusano |
2b45e8 |
addq $4 * SIZE, BO # boffset1 += 8
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jg .L146
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L148:
|
|
kusano |
2b45e8 |
addpd %xmm10, %xmm8
|
|
kusano |
2b45e8 |
addpd %xmm11, %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
movlpd 0 * SIZE(CO1), %xmm0
|
|
kusano |
2b45e8 |
movhpd 1 * SIZE(CO1), %xmm0
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm9, %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
|
|
kusano |
2b45e8 |
defined(NR) || defined(NC) || defined(TR) || defined(TC)
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm9
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
xorpd %xmm5, %xmm8
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
|
|
kusano |
2b45e8 |
defined(RR) || defined(RC) || defined(CR) || defined(CC)
|
|
kusano |
2b45e8 |
subpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
pshufd $0x4e, %xmm8, %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm6, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm7, %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addpd %xmm9, %xmm8
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef TRMMKERNEL
|
|
kusano |
2b45e8 |
addpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movlpd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L999:
|
|
kusano |
2b45e8 |
movq %rbx, %rsp
|
|
kusano |
2b45e8 |
EMMS
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq 0(%rsp), %rbx
|
|
kusano |
2b45e8 |
movq 8(%rsp), %rbp
|
|
kusano |
2b45e8 |
movq 16(%rsp), %r12
|
|
kusano |
2b45e8 |
movq 24(%rsp), %r13
|
|
kusano |
2b45e8 |
movq 32(%rsp), %r14
|
|
kusano |
2b45e8 |
movq 40(%rsp), %r15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef WINDOWS_ABI
|
|
kusano |
2b45e8 |
movq 48(%rsp), %rdi
|
|
kusano |
2b45e8 |
movq 56(%rsp), %rsi
|
|
kusano |
2b45e8 |
movups 64(%rsp), %xmm6
|
|
kusano |
2b45e8 |
movups 80(%rsp), %xmm7
|
|
kusano |
2b45e8 |
movups 96(%rsp), %xmm8
|
|
kusano |
2b45e8 |
movups 112(%rsp), %xmm9
|
|
kusano |
2b45e8 |
movups 128(%rsp), %xmm10
|
|
kusano |
2b45e8 |
movups 144(%rsp), %xmm11
|
|
kusano |
2b45e8 |
movups 160(%rsp), %xmm12
|
|
kusano |
2b45e8 |
movups 176(%rsp), %xmm13
|
|
kusano |
2b45e8 |
movups 192(%rsp), %xmm14
|
|
kusano |
2b45e8 |
movups 208(%rsp), %xmm15
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $STACKSIZE, %rsp
|
|
kusano |
2b45e8 |
ret
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
EPILOGUE
|