|
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 M %rdi
|
|
kusano |
2b45e8 |
#define N %rsi
|
|
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 |
|
|
kusano |
2b45e8 |
#define I %r11
|
|
kusano |
2b45e8 |
#define J %r12
|
|
kusano |
2b45e8 |
#define AO %r13
|
|
kusano |
2b45e8 |
#define BO %r14
|
|
kusano |
2b45e8 |
#define CO1 %r15
|
|
kusano |
2b45e8 |
#define CO2 %rbx
|
|
kusano |
2b45e8 |
#define KK %rbp
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef WINDOWS_ABI
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define STACKSIZE 128
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define OLD_LDC 8 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
#define OLD_OFFSET 16 + STACKSIZE(%rsp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define OFFSET 48(%rsp)
|
|
kusano |
2b45e8 |
#define KKK 56(%rsp)
|
|
kusano |
2b45e8 |
#define AORIG 64(%rsp)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define STACKSIZE 256
|
|
kusano |
2b45e8 |
|
|
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 |
#define OFFSET 224(%rsp)
|
|
kusano |
2b45e8 |
#define KKK 232(%rsp)
|
|
kusano |
2b45e8 |
#define AORIG 240(%rsp)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define PREFETCH prefetcht1
|
|
kusano |
2b45e8 |
#define PREFETCHSIZE (16 * 12 + 3)
|
|
kusano |
2b45e8 |
#define PREFETCH_R (4 * 4 + 0)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
#define ADD1 addpd
|
|
kusano |
2b45e8 |
#define ADD2 addpd
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#define ADD1 subpd
|
|
kusano |
2b45e8 |
#define ADD2 addpd
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL1(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE + (address) * 2 * SIZE(AO);\
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0;\
|
|
kusano |
2b45e8 |
movddup 1 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1;\
|
|
kusano |
2b45e8 |
movddup 2 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm2;\
|
|
kusano |
2b45e8 |
movddup 3 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
movapd 2 * SIZE + (address) * 2 * SIZE(AO), %xmm8;\
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm3;\
|
|
kusano |
2b45e8 |
movddup 0 * SIZE + (address) * 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL2(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm4;\
|
|
kusano |
2b45e8 |
movddup 1 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm5;\
|
|
kusano |
2b45e8 |
movddup 2 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm6;\
|
|
kusano |
2b45e8 |
movddup 3 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
movapd 4 * SIZE + (address) * 2 * SIZE(AO), %xmm8;\
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm7;\
|
|
kusano |
2b45e8 |
movddup 4 * SIZE + (address) * 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL3(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0;\
|
|
kusano |
2b45e8 |
movddup 5 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1;\
|
|
kusano |
2b45e8 |
movddup 6 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm2;\
|
|
kusano |
2b45e8 |
movddup 7 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
movapd 6 * SIZE + (address) * 2 * SIZE(AO), %xmm8;\
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm3;\
|
|
kusano |
2b45e8 |
movddup 4 * SIZE + (address) * 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL4(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm4;\
|
|
kusano |
2b45e8 |
movddup 5 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm5;\
|
|
kusano |
2b45e8 |
movddup 6 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm6;\
|
|
kusano |
2b45e8 |
movddup 7 * SIZE + (address) * 2 * SIZE(BO), %xmm9;\
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9;\
|
|
kusano |
2b45e8 |
movapd 32 * SIZE + (address) * 2 * SIZE(AO), %xmm8;\
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm7;\
|
|
kusano |
2b45e8 |
movddup 32 * SIZE + (address) * 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL5(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0;\
|
|
kusano |
2b45e8 |
movddup 9 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1;\
|
|
kusano |
2b45e8 |
movddup 10 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm2;\
|
|
kusano |
2b45e8 |
movddup 11 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
movapd 10 * SIZE + (address) * 2 * SIZE(AO), %xmm10;\
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm3;\
|
|
kusano |
2b45e8 |
movddup 8 * SIZE + (address) * 2 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL6(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm4;\
|
|
kusano |
2b45e8 |
movddup 9 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm5;\
|
|
kusano |
2b45e8 |
movddup 10 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm6;\
|
|
kusano |
2b45e8 |
movddup 11 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
movapd 12 * SIZE + (address) * 2 * SIZE(AO), %xmm10;\
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm7;\
|
|
kusano |
2b45e8 |
movddup 12 * SIZE + (address) * 2 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL7(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0;\
|
|
kusano |
2b45e8 |
movddup 13 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1;\
|
|
kusano |
2b45e8 |
movddup 14 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm2;\
|
|
kusano |
2b45e8 |
movddup 15 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
movapd 14 * SIZE + (address) * 2 * SIZE(AO), %xmm10;\
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm3;\
|
|
kusano |
2b45e8 |
movddup 12 * SIZE + (address) * 2 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL8(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm4;\
|
|
kusano |
2b45e8 |
movddup 13 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm5;\
|
|
kusano |
2b45e8 |
movddup 14 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm6;\
|
|
kusano |
2b45e8 |
movddup 15 * SIZE + (address) * 2 * SIZE(BO), %xmm11;\
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11;\
|
|
kusano |
2b45e8 |
movapd 40 * SIZE + (address) * 2 * SIZE(AO), %xmm10;\
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm7;\
|
|
kusano |
2b45e8 |
movddup 40 * SIZE + (address) * 2 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL9(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 16) * SIZE + (address) * 2 * SIZE(AO);\
|
|
kusano |
2b45e8 |
ADD1 %xmm13, %xmm0;\
|
|
kusano |
2b45e8 |
movddup 17 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
ADD2 %xmm13, %xmm1;\
|
|
kusano |
2b45e8 |
movddup 18 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
ADD1 %xmm13, %xmm2;\
|
|
kusano |
2b45e8 |
movddup 19 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
movapd 18 * SIZE + (address) * 2 * SIZE(AO), %xmm12;\
|
|
kusano |
2b45e8 |
ADD2 %xmm13, %xmm3;\
|
|
kusano |
2b45e8 |
movddup 16 * SIZE + (address) * 2 * SIZE(BO), %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL10(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
ADD1 %xmm13, %xmm4;\
|
|
kusano |
2b45e8 |
movddup 17 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
ADD2 %xmm13, %xmm5;\
|
|
kusano |
2b45e8 |
movddup 18 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
ADD1 %xmm13, %xmm6;\
|
|
kusano |
2b45e8 |
movddup 19 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
movapd 20 * SIZE + (address) * 2 * SIZE(AO), %xmm12;\
|
|
kusano |
2b45e8 |
ADD2 %xmm13, %xmm7;\
|
|
kusano |
2b45e8 |
movddup 20 * SIZE + (address) * 2 * SIZE(BO), %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL11(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
ADD1 %xmm13, %xmm0;\
|
|
kusano |
2b45e8 |
movddup 21 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
ADD2 %xmm13, %xmm1;\
|
|
kusano |
2b45e8 |
movddup 22 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
ADD1 %xmm13, %xmm2;\
|
|
kusano |
2b45e8 |
movddup 23 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
movapd 22 * SIZE + (address) * 2 * SIZE(AO), %xmm12;\
|
|
kusano |
2b45e8 |
ADD2 %xmm13, %xmm3;\
|
|
kusano |
2b45e8 |
movddup 20 * SIZE + (address) * 2 * SIZE(BO), %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL12(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
ADD1 %xmm13, %xmm4;\
|
|
kusano |
2b45e8 |
movddup 21 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
ADD2 %xmm13, %xmm5;\
|
|
kusano |
2b45e8 |
movddup 22 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
ADD1 %xmm13, %xmm6;\
|
|
kusano |
2b45e8 |
movddup 23 * SIZE + (address) * 2 * SIZE(BO), %xmm13;\
|
|
kusano |
2b45e8 |
mulpd %xmm12, %xmm13;\
|
|
kusano |
2b45e8 |
movapd 48 * SIZE + (address) * 2 * SIZE(AO), %xmm12;\
|
|
kusano |
2b45e8 |
ADD2 %xmm13, %xmm7;\
|
|
kusano |
2b45e8 |
movddup 48 * SIZE + (address) * 2 * SIZE(BO), %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL13(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD1 %xmm15, %xmm0;\
|
|
kusano |
2b45e8 |
movddup 25 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD2 %xmm15, %xmm1;\
|
|
kusano |
2b45e8 |
movddup 26 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD1 %xmm15, %xmm2;\
|
|
kusano |
2b45e8 |
movddup 27 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
movapd 26 * SIZE + (address) * 2 * SIZE(AO), %xmm14;\
|
|
kusano |
2b45e8 |
ADD2 %xmm15, %xmm3;\
|
|
kusano |
2b45e8 |
movddup 24 * SIZE + (address) * 2 * SIZE(BO), %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL14(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD1 %xmm15, %xmm4;\
|
|
kusano |
2b45e8 |
movddup 25 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD2 %xmm15, %xmm5;\
|
|
kusano |
2b45e8 |
movddup 26 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD1 %xmm15, %xmm6;\
|
|
kusano |
2b45e8 |
movddup 27 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
movapd 28 * SIZE + (address) * 2 * SIZE(AO), %xmm14;\
|
|
kusano |
2b45e8 |
ADD2 %xmm15, %xmm7;\
|
|
kusano |
2b45e8 |
movddup 28 * SIZE + (address) * 2 * SIZE(BO), %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL15(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD1 %xmm15, %xmm0;\
|
|
kusano |
2b45e8 |
movddup 29 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD2 %xmm15, %xmm1;\
|
|
kusano |
2b45e8 |
movddup 30 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD1 %xmm15, %xmm2;\
|
|
kusano |
2b45e8 |
movddup 31 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
movapd 30 * SIZE + (address) * 2 * SIZE(AO), %xmm14;\
|
|
kusano |
2b45e8 |
ADD2 %xmm15, %xmm3;\
|
|
kusano |
2b45e8 |
movddup 28 * SIZE + (address) * 2 * SIZE(BO), %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define KERNEL16(address) \
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD1 %xmm15, %xmm4;\
|
|
kusano |
2b45e8 |
movddup 29 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD2 %xmm15, %xmm5;\
|
|
kusano |
2b45e8 |
movddup 30 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
ADD1 %xmm15, %xmm6;\
|
|
kusano |
2b45e8 |
movddup 31 * SIZE + (address) * 2 * SIZE(BO), %xmm15;\
|
|
kusano |
2b45e8 |
mulpd %xmm14, %xmm15;\
|
|
kusano |
2b45e8 |
movapd 56 * SIZE + (address) * 2 * SIZE(AO), %xmm14;\
|
|
kusano |
2b45e8 |
ADD2 %xmm15, %xmm7;\
|
|
kusano |
2b45e8 |
movddup 56 * SIZE + (address) * 2 * SIZE(BO), %xmm15
|
|
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, M
|
|
kusano |
2b45e8 |
movq ARG2, 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 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq OLD_LDC, LDC
|
|
kusano |
2b45e8 |
movq OLD_OFFSET, KK
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq KK, OFFSET
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
salq $ZBASE_SHIFT, LDC
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movq M, %rax
|
|
kusano |
2b45e8 |
salq $ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
addq %rax, C
|
|
kusano |
2b45e8 |
imulq K, %rax
|
|
kusano |
2b45e8 |
addq %rax, A
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movq N, %rax
|
|
kusano |
2b45e8 |
salq $ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
imulq K, %rax
|
|
kusano |
2b45e8 |
addq %rax, B
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq N, %rax
|
|
kusano |
2b45e8 |
imulq LDC, %rax
|
|
kusano |
2b45e8 |
addq %rax, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
negq KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movq N, KK
|
|
kusano |
2b45e8 |
subq OFFSET, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
testq $1, N
|
|
kusano |
2b45e8 |
jle .L100
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L101:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq A, AO
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq A, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
salq $0 + ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
subq %rax, B
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subq LDC, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
movq C, CO1 # coffset1 = c
|
|
kusano |
2b45e8 |
#ifndef RT
|
|
kusano |
2b45e8 |
addq LDC, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movq OFFSET, KK
|
|
kusano |
2b45e8 |
addq M, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movq OFFSET, KK
|
|
kusano |
2b45e8 |
#endif
|
|
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 |
#ifdef LN
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
salq $1 + ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
subq %rax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
movq AORIG, AO
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 4), AO
|
|
kusano |
2b45e8 |
leaq (B, %rax, 2), BO
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq B, BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm0
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
pxor %xmm1, %xmm1
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
movddup 8 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
prefetchnta 4 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarq $3, %rax
|
|
kusano |
2b45e8 |
je .L112
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L111:
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm4
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm5
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 3 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 6 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm4
|
|
kusano |
2b45e8 |
movddup 3 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 16 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm5
|
|
kusano |
2b45e8 |
movddup 4 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 5 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
movapd 10 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 4 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm4
|
|
kusano |
2b45e8 |
movddup 5 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
movapd 12 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm5
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
movapd 14 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm4
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
movapd 40 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm5
|
|
kusano |
2b45e8 |
movddup 16 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 16) * SIZE(AO)
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0
|
|
kusano |
2b45e8 |
movddup 9 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
movapd 18 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1
|
|
kusano |
2b45e8 |
movddup 8 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm4
|
|
kusano |
2b45e8 |
movddup 9 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
movapd 20 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm5
|
|
kusano |
2b45e8 |
movddup 10 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0
|
|
kusano |
2b45e8 |
movddup 11 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
movapd 22 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1
|
|
kusano |
2b45e8 |
movddup 10 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm4
|
|
kusano |
2b45e8 |
movddup 11 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
movapd 24 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm5
|
|
kusano |
2b45e8 |
movddup 12 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0
|
|
kusano |
2b45e8 |
movddup 13 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
movapd 26 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1
|
|
kusano |
2b45e8 |
movddup 12 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm4
|
|
kusano |
2b45e8 |
movddup 13 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
movapd 28 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm5
|
|
kusano |
2b45e8 |
movddup 14 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0
|
|
kusano |
2b45e8 |
movddup 15 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
movapd 30 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1
|
|
kusano |
2b45e8 |
movddup 14 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm4
|
|
kusano |
2b45e8 |
movddup 15 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
movapd 32 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm5
|
|
kusano |
2b45e8 |
movddup 24 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $32 * SIZE, AO
|
|
kusano |
2b45e8 |
addq $16 * SIZE, BO
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jne .L111
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L112:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andq $7, %rax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L114
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L113:
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm4
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $4 * SIZE, AO # aoffset += 4
|
|
kusano |
2b45e8 |
addq $2 * SIZE, BO # boffset1 += 8
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jg .L113
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L114:
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm1, %xmm1
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm5, %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
addsubpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
addsubpd %xmm5, %xmm4
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addsubpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addsubpd %xmm4, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subq $2, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
subq $1, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq AORIG, AO
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 4), AO
|
|
kusano |
2b45e8 |
leaq (B, %rax, 2), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BO), %xmm8
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(AO), %xmm9
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if (defined(LN) || defined(LT)) && !defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm9
|
|
kusano |
2b45e8 |
#elif (defined(LN) || defined(LT)) && defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm5, %xmm9
|
|
kusano |
2b45e8 |
#elif (defined(RN) || defined(RT)) && !defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm9
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addsubpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
addsubpd %xmm5, %xmm9
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
pcmpeqb %xmm7, %xmm7
|
|
kusano |
2b45e8 |
psllq $63, %xmm7
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(AO), %xmm1
|
|
kusano |
2b45e8 |
movddup 4 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
movddup 5 * SIZE(AO), %xmm3
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(AO), %xmm4
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(AO), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm9
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm13
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
subpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(AO), %xmm1
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
movddup 3 * SIZE(AO), %xmm3
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(AO), %xmm4
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(AO), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm13
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
subpd %xmm12, %xmm9
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm9
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm9
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm4
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm9
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subq $4 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm9, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm9, 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm9, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm9, 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd %xmm8, 0 * SIZE(BO)
|
|
kusano |
2b45e8 |
movapd %xmm9, 2 * SIZE(BO)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd %xmm8, 0 * SIZE(AO)
|
|
kusano |
2b45e8 |
movapd %xmm9, 2 * SIZE(AO)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addq $4 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
leaq (,%rax, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 4), AO
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 2), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subq $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addq $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
salq $1 + ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
addq %rax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
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 .L149
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
salq $0 + ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
subq %rax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
movq AORIG, AO
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 2), AO
|
|
kusano |
2b45e8 |
leaq (B, %rax, 2), BO
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq B, BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm0
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
pxor %xmm1, %xmm1
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
pxor %xmm2, %xmm2
|
|
kusano |
2b45e8 |
movddup 8 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
pxor %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarq $3, %rax
|
|
kusano |
2b45e8 |
je .L142
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L141:
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm2
|
|
kusano |
2b45e8 |
movddup 3 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm3
|
|
kusano |
2b45e8 |
movddup 4 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 5 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 6 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm2
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 16 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm3
|
|
kusano |
2b45e8 |
movddup 16 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0
|
|
kusano |
2b45e8 |
movddup 9 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
movapd 10 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1
|
|
kusano |
2b45e8 |
movddup 10 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm2
|
|
kusano |
2b45e8 |
movddup 11 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
movapd 12 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm3
|
|
kusano |
2b45e8 |
movddup 12 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0
|
|
kusano |
2b45e8 |
movddup 13 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
movapd 14 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1
|
|
kusano |
2b45e8 |
movddup 14 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm2
|
|
kusano |
2b45e8 |
movddup 15 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
movapd 24 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm3
|
|
kusano |
2b45e8 |
movddup 24 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $16 * SIZE, AO
|
|
kusano |
2b45e8 |
addq $16 * SIZE, BO
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jne .L141
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L142:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andq $7, %rax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L144
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L143:
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $2 * SIZE, AO # aoffset += 4
|
|
kusano |
2b45e8 |
addq $2 * SIZE, BO # boffset1 += 8
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jg .L143
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L144:
|
|
kusano |
2b45e8 |
addpd %xmm2, %xmm0
|
|
kusano |
2b45e8 |
addpd %xmm3, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm1, %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
addsubpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addsubpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
subq $1, %rax
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq AORIG, AO
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 2), AO
|
|
kusano |
2b45e8 |
leaq (B, %rax, 2), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BO), %xmm8
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if (defined(LN) || defined(LT)) && !defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
#elif (defined(LN) || defined(LT)) && defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
#elif (defined(RN) || defined(RT)) && !defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addsubpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
pcmpeqb %xmm7, %xmm7
|
|
kusano |
2b45e8 |
psllq $63, %xmm7
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(AO), %xmm4
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(AO), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(AO), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm4
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subq $2 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd %xmm8, 0 * SIZE(BO)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd %xmm8, 0 * SIZE(AO)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addq $2 * SIZE, CO1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
leaq (,%rax, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 2), AO
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 2), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subq $1, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addq $1, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
salq $0 + ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
addq %rax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L149:
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
leaq (, K, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (B, %rax, 2), B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq BO, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
addq $1, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
subq $1, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ALIGN_3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L100:
|
|
kusano |
2b45e8 |
movq N, J
|
|
kusano |
2b45e8 |
sarq $1, J # j = (n >> 2)
|
|
kusano |
2b45e8 |
jle .L999
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L01:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq A, AO
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq A, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
salq $1 + ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
subq %rax, B
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq (, LDC, 2), %rax
|
|
kusano |
2b45e8 |
subq %rax, C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq C, CO1 # coffset1 = c
|
|
kusano |
2b45e8 |
leaq (C, LDC, 1), CO2 # coffset2 = c + ldc
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef RT
|
|
kusano |
2b45e8 |
leaq (C, LDC, 2), C
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movq OFFSET, KK
|
|
kusano |
2b45e8 |
addq M, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movq OFFSET, KK
|
|
kusano |
2b45e8 |
#endif
|
|
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 |
#ifdef LN
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
salq $1 + ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
subq %rax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
movq AORIG, AO
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 4), AO
|
|
kusano |
2b45e8 |
leaq (B, %rax, 4), BO
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq B, BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm0
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
pxor %xmm1, %xmm1
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
pxor %xmm2, %xmm2
|
|
kusano |
2b45e8 |
movddup 8 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
pxor %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd 16 * SIZE(AO), %xmm12
|
|
kusano |
2b45e8 |
movddup 16 * SIZE(BO), %xmm13
|
|
kusano |
2b45e8 |
movapd 24 * SIZE(AO), %xmm14
|
|
kusano |
2b45e8 |
movddup 24 * SIZE(BO), %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
prefetchnta 4 * SIZE(CO1)
|
|
kusano |
2b45e8 |
pxor %xmm4, %xmm4
|
|
kusano |
2b45e8 |
prefetchnta 4 * SIZE(CO2)
|
|
kusano |
2b45e8 |
pxor %xmm5, %xmm5
|
|
kusano |
2b45e8 |
pxor %xmm6, %xmm6
|
|
kusano |
2b45e8 |
pxor %xmm7, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andq $-8, %rax
|
|
kusano |
2b45e8 |
salq $4, %rax
|
|
kusano |
2b45e8 |
je .L12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L1X:
|
|
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 |
KERNEL9 (16 * 0)
|
|
kusano |
2b45e8 |
KERNEL10(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL11(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL12(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL13(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL14(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL15(16 * 0)
|
|
kusano |
2b45e8 |
KERNEL16(16 * 0)
|
|
kusano |
2b45e8 |
cmpq $128 * 1, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
jle .L11
|
|
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 |
KERNEL9 (16 * 1)
|
|
kusano |
2b45e8 |
KERNEL10(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL11(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL12(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL13(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL14(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL15(16 * 1)
|
|
kusano |
2b45e8 |
KERNEL16(16 * 1)
|
|
kusano |
2b45e8 |
cmpq $128 * 2, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
jle .L11
|
|
kusano |
2b45e8 |
KERNEL1 (16 * 2)
|
|
kusano |
2b45e8 |
KERNEL2 (16 * 2)
|
|
kusano |
2b45e8 |
KERNEL3 (16 * 2)
|
|
kusano |
2b45e8 |
KERNEL4 (16 * 2)
|
|
kusano |
2b45e8 |
KERNEL5 (16 * 2)
|
|
kusano |
2b45e8 |
KERNEL6 (16 * 2)
|
|
kusano |
2b45e8 |
KERNEL7 (16 * 2)
|
|
kusano |
2b45e8 |
KERNEL8 (16 * 2)
|
|
kusano |
2b45e8 |
KERNEL9 (16 * 2)
|
|
kusano |
2b45e8 |
KERNEL10(16 * 2)
|
|
kusano |
2b45e8 |
KERNEL11(16 * 2)
|
|
kusano |
2b45e8 |
KERNEL12(16 * 2)
|
|
kusano |
2b45e8 |
KERNEL13(16 * 2)
|
|
kusano |
2b45e8 |
KERNEL14(16 * 2)
|
|
kusano |
2b45e8 |
KERNEL15(16 * 2)
|
|
kusano |
2b45e8 |
KERNEL16(16 * 2)
|
|
kusano |
2b45e8 |
cmpq $128 * 3, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
jle .L11
|
|
kusano |
2b45e8 |
KERNEL1 (16 * 3)
|
|
kusano |
2b45e8 |
KERNEL2 (16 * 3)
|
|
kusano |
2b45e8 |
KERNEL3 (16 * 3)
|
|
kusano |
2b45e8 |
KERNEL4 (16 * 3)
|
|
kusano |
2b45e8 |
KERNEL5 (16 * 3)
|
|
kusano |
2b45e8 |
KERNEL6 (16 * 3)
|
|
kusano |
2b45e8 |
KERNEL7 (16 * 3)
|
|
kusano |
2b45e8 |
KERNEL8 (16 * 3)
|
|
kusano |
2b45e8 |
KERNEL9 (16 * 3)
|
|
kusano |
2b45e8 |
KERNEL10(16 * 3)
|
|
kusano |
2b45e8 |
KERNEL11(16 * 3)
|
|
kusano |
2b45e8 |
KERNEL12(16 * 3)
|
|
kusano |
2b45e8 |
KERNEL13(16 * 3)
|
|
kusano |
2b45e8 |
KERNEL14(16 * 3)
|
|
kusano |
2b45e8 |
KERNEL15(16 * 3)
|
|
kusano |
2b45e8 |
KERNEL16(16 * 3)
|
|
kusano |
2b45e8 |
cmpq $128 * 4, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
jle .L11
|
|
kusano |
2b45e8 |
KERNEL1 (16 * 4)
|
|
kusano |
2b45e8 |
KERNEL2 (16 * 4)
|
|
kusano |
2b45e8 |
KERNEL3 (16 * 4)
|
|
kusano |
2b45e8 |
KERNEL4 (16 * 4)
|
|
kusano |
2b45e8 |
KERNEL5 (16 * 4)
|
|
kusano |
2b45e8 |
KERNEL6 (16 * 4)
|
|
kusano |
2b45e8 |
KERNEL7 (16 * 4)
|
|
kusano |
2b45e8 |
KERNEL8 (16 * 4)
|
|
kusano |
2b45e8 |
KERNEL9 (16 * 4)
|
|
kusano |
2b45e8 |
KERNEL10(16 * 4)
|
|
kusano |
2b45e8 |
KERNEL11(16 * 4)
|
|
kusano |
2b45e8 |
KERNEL12(16 * 4)
|
|
kusano |
2b45e8 |
KERNEL13(16 * 4)
|
|
kusano |
2b45e8 |
KERNEL14(16 * 4)
|
|
kusano |
2b45e8 |
KERNEL15(16 * 4)
|
|
kusano |
2b45e8 |
KERNEL16(16 * 4)
|
|
kusano |
2b45e8 |
cmpq $128 * 5, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
jle .L11
|
|
kusano |
2b45e8 |
KERNEL1 (16 * 5)
|
|
kusano |
2b45e8 |
KERNEL2 (16 * 5)
|
|
kusano |
2b45e8 |
KERNEL3 (16 * 5)
|
|
kusano |
2b45e8 |
KERNEL4 (16 * 5)
|
|
kusano |
2b45e8 |
KERNEL5 (16 * 5)
|
|
kusano |
2b45e8 |
KERNEL6 (16 * 5)
|
|
kusano |
2b45e8 |
KERNEL7 (16 * 5)
|
|
kusano |
2b45e8 |
KERNEL8 (16 * 5)
|
|
kusano |
2b45e8 |
KERNEL9 (16 * 5)
|
|
kusano |
2b45e8 |
KERNEL10(16 * 5)
|
|
kusano |
2b45e8 |
KERNEL11(16 * 5)
|
|
kusano |
2b45e8 |
KERNEL12(16 * 5)
|
|
kusano |
2b45e8 |
KERNEL13(16 * 5)
|
|
kusano |
2b45e8 |
KERNEL14(16 * 5)
|
|
kusano |
2b45e8 |
KERNEL15(16 * 5)
|
|
kusano |
2b45e8 |
KERNEL16(16 * 5)
|
|
kusano |
2b45e8 |
cmpq $128 * 6, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
jle .L11
|
|
kusano |
2b45e8 |
KERNEL1 (16 * 6)
|
|
kusano |
2b45e8 |
KERNEL2 (16 * 6)
|
|
kusano |
2b45e8 |
KERNEL3 (16 * 6)
|
|
kusano |
2b45e8 |
KERNEL4 (16 * 6)
|
|
kusano |
2b45e8 |
KERNEL5 (16 * 6)
|
|
kusano |
2b45e8 |
KERNEL6 (16 * 6)
|
|
kusano |
2b45e8 |
KERNEL7 (16 * 6)
|
|
kusano |
2b45e8 |
KERNEL8 (16 * 6)
|
|
kusano |
2b45e8 |
KERNEL9 (16 * 6)
|
|
kusano |
2b45e8 |
KERNEL10(16 * 6)
|
|
kusano |
2b45e8 |
KERNEL11(16 * 6)
|
|
kusano |
2b45e8 |
KERNEL12(16 * 6)
|
|
kusano |
2b45e8 |
KERNEL13(16 * 6)
|
|
kusano |
2b45e8 |
KERNEL14(16 * 6)
|
|
kusano |
2b45e8 |
KERNEL15(16 * 6)
|
|
kusano |
2b45e8 |
KERNEL16(16 * 6)
|
|
kusano |
2b45e8 |
cmpq $128 * 7, %rax
|
|
kusano |
2b45e8 |
NOBRANCH
|
|
kusano |
2b45e8 |
jle .L11
|
|
kusano |
2b45e8 |
KERNEL1 (16 * 7)
|
|
kusano |
2b45e8 |
KERNEL2 (16 * 7)
|
|
kusano |
2b45e8 |
KERNEL3 (16 * 7)
|
|
kusano |
2b45e8 |
KERNEL4 (16 * 7)
|
|
kusano |
2b45e8 |
KERNEL5 (16 * 7)
|
|
kusano |
2b45e8 |
KERNEL6 (16 * 7)
|
|
kusano |
2b45e8 |
KERNEL7 (16 * 7)
|
|
kusano |
2b45e8 |
KERNEL8 (16 * 7)
|
|
kusano |
2b45e8 |
KERNEL9 (16 * 7)
|
|
kusano |
2b45e8 |
KERNEL10(16 * 7)
|
|
kusano |
2b45e8 |
KERNEL11(16 * 7)
|
|
kusano |
2b45e8 |
KERNEL12(16 * 7)
|
|
kusano |
2b45e8 |
KERNEL13(16 * 7)
|
|
kusano |
2b45e8 |
KERNEL14(16 * 7)
|
|
kusano |
2b45e8 |
KERNEL15(16 * 7)
|
|
kusano |
2b45e8 |
KERNEL16(16 * 7)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $32 * 8 * SIZE, AO
|
|
kusano |
2b45e8 |
addq $32 * 8 * SIZE, BO
|
|
kusano |
2b45e8 |
subq $128 * 8, %rax
|
|
kusano |
2b45e8 |
jg .L1X
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L11:
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 2), AO # * 16
|
|
kusano |
2b45e8 |
leaq (BO, %rax, 2), BO # * 64
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L12:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andq $7, %rax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
je .L14
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L13:
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm2
|
|
kusano |
2b45e8 |
movddup 3 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm3
|
|
kusano |
2b45e8 |
movddup 4 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm4
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm5
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm6
|
|
kusano |
2b45e8 |
movddup 3 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm7
|
|
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 .L13
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L14:
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm1, %xmm1
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm3, %xmm3
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm5, %xmm5
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm7, %xmm7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
addsubpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
addsubpd %xmm3, %xmm2
|
|
kusano |
2b45e8 |
addsubpd %xmm5, %xmm4
|
|
kusano |
2b45e8 |
addsubpd %xmm7, %xmm6
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addsubpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addsubpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
addsubpd %xmm4, %xmm5
|
|
kusano |
2b45e8 |
addsubpd %xmm6, %xmm7
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
subq $2, %rax
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq AORIG, AO
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 4), AO
|
|
kusano |
2b45e8 |
leaq (B, %rax, 4), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BO), %xmm8
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(BO), %xmm10
|
|
kusano |
2b45e8 |
movapd 6 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(AO), %xmm9
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
movapd 6 * SIZE(AO), %xmm11
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if (defined(LN) || defined(LT)) && !defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm2, %xmm9
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm10
|
|
kusano |
2b45e8 |
subpd %xmm6, %xmm11
|
|
kusano |
2b45e8 |
#elif (defined(LN) || defined(LT)) && defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm3, %xmm9
|
|
kusano |
2b45e8 |
subpd %xmm5, %xmm10
|
|
kusano |
2b45e8 |
subpd %xmm7, %xmm11
|
|
kusano |
2b45e8 |
#elif (defined(RN) || defined(RT)) && !defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm4, %xmm9
|
|
kusano |
2b45e8 |
subpd %xmm2, %xmm10
|
|
kusano |
2b45e8 |
subpd %xmm6, %xmm11
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addsubpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
addsubpd %xmm5, %xmm9
|
|
kusano |
2b45e8 |
addsubpd %xmm3, %xmm10
|
|
kusano |
2b45e8 |
addsubpd %xmm7, %xmm11
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
pcmpeqb %xmm7, %xmm7
|
|
kusano |
2b45e8 |
psllq $63, %xmm7
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(AO), %xmm1
|
|
kusano |
2b45e8 |
movddup 4 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
movddup 5 * SIZE(AO), %xmm3
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(AO), %xmm4
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(AO), %xmm5
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
movddup 4 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
movddup 5 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm4
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm10, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm11, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm10
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm10
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm10, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm10, %xmm13
|
|
kusano |
2b45e8 |
movapd %xmm11, %xmm14
|
|
kusano |
2b45e8 |
movapd %xmm11, %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm15, %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm14
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm13
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
addsubpd %xmm15, %xmm14
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm14, %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm9
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(AO), %xmm1
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(AO), %xmm2
|
|
kusano |
2b45e8 |
movddup 3 * SIZE(AO), %xmm3
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(AO), %xmm4
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(AO), %xmm5
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
movddup 3 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(BO), %xmm4
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(BO), %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm13
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm14
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm15, %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm14
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm13
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
addsubpd %xmm15, %xmm14
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subpd %xmm12, %xmm10
|
|
kusano |
2b45e8 |
subpd %xmm14, %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm10, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm11, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm10
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm10
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm11
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subq $4 * SIZE, CO1
|
|
kusano |
2b45e8 |
subq $4 * SIZE, CO2
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm10, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm10, 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd %xmm9, 0 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movhpd %xmm9, 1 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movsd %xmm11, 2 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movhpd %xmm11, 3 * SIZE(CO2)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movsd %xmm9, 2 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm9, 3 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd %xmm10, 0 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movhpd %xmm10, 1 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movsd %xmm11, 2 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movhpd %xmm11, 3 * SIZE(CO2)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd %xmm8, 0 * SIZE(BO)
|
|
kusano |
2b45e8 |
movapd %xmm9, 2 * SIZE(BO)
|
|
kusano |
2b45e8 |
movapd %xmm10, 4 * SIZE(BO)
|
|
kusano |
2b45e8 |
movapd %xmm11, 6 * SIZE(BO)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd %xmm8, 0 * SIZE(AO)
|
|
kusano |
2b45e8 |
movapd %xmm9, 2 * SIZE(AO)
|
|
kusano |
2b45e8 |
movapd %xmm10, 4 * SIZE(AO)
|
|
kusano |
2b45e8 |
movapd %xmm11, 6 * SIZE(AO)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addq $4 * SIZE, CO1
|
|
kusano |
2b45e8 |
addq $4 * SIZE, CO2
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq 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 |
#ifdef LN
|
|
kusano |
2b45e8 |
subq $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addq $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
salq $1 + ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
addq %rax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
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 |
#ifdef LN
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
salq $0 + ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
subq %rax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
movq AORIG, AO
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 2), AO
|
|
kusano |
2b45e8 |
leaq (B, %rax, 4), BO
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq B, BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
pxor %xmm0, %xmm0
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
pxor %xmm1, %xmm1
|
|
kusano |
2b45e8 |
movapd 8 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
pxor %xmm2, %xmm2
|
|
kusano |
2b45e8 |
movddup 8 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
pxor %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
sarq $3, %rax
|
|
kusano |
2b45e8 |
je .L42
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L41:
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm2
|
|
kusano |
2b45e8 |
movddup 3 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm3
|
|
kusano |
2b45e8 |
movddup 4 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 5 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm2
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 4 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm3
|
|
kusano |
2b45e8 |
movddup 16 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0
|
|
kusano |
2b45e8 |
movddup 9 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1
|
|
kusano |
2b45e8 |
movddup 10 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm2
|
|
kusano |
2b45e8 |
movddup 11 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
movapd 6 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm3
|
|
kusano |
2b45e8 |
movddup 12 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0
|
|
kusano |
2b45e8 |
movddup 13 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1
|
|
kusano |
2b45e8 |
movddup 14 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm2
|
|
kusano |
2b45e8 |
movddup 15 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm11
|
|
kusano |
2b45e8 |
movapd 16 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm3
|
|
kusano |
2b45e8 |
movddup 24 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 17 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 18 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm2
|
|
kusano |
2b45e8 |
movddup 19 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
movapd 10 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm3
|
|
kusano |
2b45e8 |
movddup 20 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 21 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 22 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm2
|
|
kusano |
2b45e8 |
movddup 23 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm9
|
|
kusano |
2b45e8 |
movapd 12 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm3
|
|
kusano |
2b45e8 |
movddup 32 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0
|
|
kusano |
2b45e8 |
movddup 25 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1
|
|
kusano |
2b45e8 |
movddup 26 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm2
|
|
kusano |
2b45e8 |
movddup 27 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
movapd 14 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm3
|
|
kusano |
2b45e8 |
movddup 28 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm0
|
|
kusano |
2b45e8 |
movddup 29 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm1
|
|
kusano |
2b45e8 |
movddup 30 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
ADD1 %xmm11, %xmm2
|
|
kusano |
2b45e8 |
movddup 31 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
mulpd %xmm10, %xmm11
|
|
kusano |
2b45e8 |
movapd 24 * SIZE(AO), %xmm10
|
|
kusano |
2b45e8 |
ADD2 %xmm11, %xmm3
|
|
kusano |
2b45e8 |
movddup 40 * SIZE(BO), %xmm11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addq $16 * SIZE, AO
|
|
kusano |
2b45e8 |
addq $32 * SIZE, BO
|
|
kusano |
2b45e8 |
decq %rax
|
|
kusano |
2b45e8 |
jne .L41
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L42:
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq KK, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
andq $7, %rax # if (k & 1)
|
|
kusano |
2b45e8 |
BRANCH
|
|
kusano |
2b45e8 |
jle .L44
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L43:
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm1
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
ADD1 %xmm9, %xmm2
|
|
kusano |
2b45e8 |
movddup 3 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm8, %xmm9
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
ADD2 %xmm9, %xmm3
|
|
kusano |
2b45e8 |
movddup 4 * SIZE(BO), %xmm9
|
|
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 .L43
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L44:
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm1, %xmm1
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm3, %xmm3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef CONJ
|
|
kusano |
2b45e8 |
addsubpd %xmm1, %xmm0
|
|
kusano |
2b45e8 |
addsubpd %xmm3, %xmm2
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addsubpd %xmm0, %xmm1
|
|
kusano |
2b45e8 |
addsubpd %xmm2, %xmm3
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(RT)
|
|
kusano |
2b45e8 |
movq KK, %rax
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subq $1, %rax
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
subq $2, %rax
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
leaq (, %rax, SIZE), %rax
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movq AORIG, AO
|
|
kusano |
2b45e8 |
leaq (AO, %rax, 2), AO
|
|
kusano |
2b45e8 |
leaq (B, %rax, 4), BO
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(BO), %xmm8
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(BO), %xmm9
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd 0 * SIZE(AO), %xmm8
|
|
kusano |
2b45e8 |
movapd 2 * SIZE(AO), %xmm9
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if (defined(LN) || defined(LT)) && !defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm2, %xmm9
|
|
kusano |
2b45e8 |
#elif (defined(LN) || defined(LT)) && defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm3, %xmm9
|
|
kusano |
2b45e8 |
#elif (defined(RN) || defined(RT)) && !defined(CONJ)
|
|
kusano |
2b45e8 |
subpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
subpd %xmm2, %xmm9
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
addsubpd %xmm1, %xmm8
|
|
kusano |
2b45e8 |
addsubpd %xmm3, %xmm9
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
pcmpeqb %xmm7, %xmm7
|
|
kusano |
2b45e8 |
psllq $63, %xmm7
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(AO), %xmm4
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(AO), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm9
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(AO), %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(AO), %xmm1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm13
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm9
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
movddup 2 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
movddup 3 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(BO), %xmm4
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(BO), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subpd %xmm12, %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm9
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movddup 6 * SIZE(BO), %xmm0
|
|
kusano |
2b45e8 |
movddup 7 * SIZE(BO), %xmm1
|
|
kusano |
2b45e8 |
movddup 4 * SIZE(BO), %xmm2
|
|
kusano |
2b45e8 |
movddup 5 * SIZE(BO), %xmm3
|
|
kusano |
2b45e8 |
movddup 0 * SIZE(BO), %xmm4
|
|
kusano |
2b45e8 |
movddup 1 * SIZE(BO), %xmm5
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef CONJ
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm1
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm3
|
|
kusano |
2b45e8 |
xorpd %xmm7, %xmm5
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm0, %xmm9
|
|
kusano |
2b45e8 |
mulpd %xmm1, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm12
|
|
kusano |
2b45e8 |
movapd %xmm9, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm13, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm2, %xmm12
|
|
kusano |
2b45e8 |
mulpd %xmm3, %xmm13
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm13, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
subpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movapd %xmm8, %xmm12
|
|
kusano |
2b45e8 |
SHUFPD_1 %xmm12, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mulpd %xmm4, %xmm8
|
|
kusano |
2b45e8 |
mulpd %xmm5, %xmm12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
addsubpd %xmm12, %xmm8
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
subq $2 * SIZE, CO1
|
|
kusano |
2b45e8 |
subq $2 * SIZE, CO2
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movsd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd %xmm9, 0 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movhpd %xmm9, 1 * SIZE(CO2)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movsd %xmm8, 0 * SIZE(CO1)
|
|
kusano |
2b45e8 |
movhpd %xmm8, 1 * SIZE(CO1)
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
movsd %xmm9, 0 * SIZE(CO2)
|
|
kusano |
2b45e8 |
movhpd %xmm9, 1 * SIZE(CO2)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LN) || defined(LT)
|
|
kusano |
2b45e8 |
movapd %xmm8, 0 * SIZE(BO)
|
|
kusano |
2b45e8 |
movapd %xmm9, 2 * SIZE(BO)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
movapd %xmm8, 0 * SIZE(AO)
|
|
kusano |
2b45e8 |
movapd %xmm9, 2 * SIZE(AO)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifndef LN
|
|
kusano |
2b45e8 |
addq $2 * SIZE, CO1
|
|
kusano |
2b45e8 |
addq $2 * SIZE, CO2
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
subq 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 |
#ifdef LN
|
|
kusano |
2b45e8 |
subq $1, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef LT
|
|
kusano |
2b45e8 |
addq $1, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
movq K, %rax
|
|
kusano |
2b45e8 |
salq $0 + ZBASE_SHIFT, %rax
|
|
kusano |
2b45e8 |
addq %rax, AORIG
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ALIGN_4
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L99:
|
|
kusano |
2b45e8 |
#ifdef LN
|
|
kusano |
2b45e8 |
leaq (, K, SIZE), %rax
|
|
kusano |
2b45e8 |
leaq (B, %rax, 4), B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#if defined(LT) || defined(RN)
|
|
kusano |
2b45e8 |
movq BO, B
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RN
|
|
kusano |
2b45e8 |
addq $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef RT
|
|
kusano |
2b45e8 |
subq $2, KK
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
decq J # j --
|
|
kusano |
2b45e8 |
jg .L01
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L999:
|
|
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
|