TI中文支持网
TI专业的中文技术问题搜集分享网站
当前位置：TI中文支持网 > 数字信号处理器 (DSP) & ARM® 微处理器 > C6000™多核 > 正文

下面的程序开优化后，运行结果出错，求助

2013-07-25 分类：C6000™多核 阅读(184)

下面的程序是来自TI  Optimizing Loops on the C66x DSP 文档中的26页上的4*4矩阵乘法，我们把_cmatmpyr1指令改为_cmatmpy和一些移位，来完成 10位量化定点复数（Signed Complex 16-bit (16-bit real/16-bit Imaginary)）输入和输出的4*4 矩阵乘法。

我在不开优化时，结果正确，但是现在开优化（-o2）之后结果却是错误的，麻烦给我看看是为什么呢？

/*
* MATRIX_Mult4by4.c
*
* Created on: 2013-6-3
* Author: net
*/
#include "c6x.h"
#define MATRIXSIZE 4
#define NUMOFPOINT 120*14

void MATRIX_Mult4by4(int M, int *restrict input_mat1, int *restrict input_mat2, int *restrict output_mat)
{
_nassert((int) input_mat1 % 8 == 0);
_nassert((int) input_mat2 % 8 == 0);
_nassert((int) output_mat % 8 == 0);
long long * input_vec1;
long long * input_vec2;
long long * input_vec3;
long long * input_vec4;
long long * inputMatPtr1;
long long * inputMatPtr2;
long long * inputMatPtr3;
long long * inputMatPtr4;
long long llinputV1;
long long llinputV2;
long long llinputV3;
long long llinputV4;
long long llinputM1;
long long llinputM2;

__x128_t inputMat;

__x128_t acc1_128;
__x128_t acc2_128;
__x128_t acc3_128;
__x128_t acc4_128;
__x128_t acc5_128;
__x128_t acc6_128;
__x128_t acc7_128;
__x128_t acc8_128;

int* acc1_32;
int* acc2_32;
int* acc3_32;
int* acc4_32;
int* acc5_32;
int* acc6_32;
int* acc7_32;
int* acc8_32;

long long acc1;
long long acc2;
long long acc3;
long long acc4;
long long acc5;
long long acc6;
long long acc7;
long long acc8;

int mm;

#pragma MUST_ITERATE(NUMOFPOINT,NUMOFPOINT,2)
for ( mm = 0; mm < M; mm++ ) //M 4×4 matrix multiplications in the loop
{
input_vec1 =(long long *) &input_mat1[mm * MATRIXSIZE * MATRIXSIZE + 0 * MATRIXSIZE];
input_vec2 =(long long *) &input_mat1[mm * MATRIXSIZE * MATRIXSIZE + 1 * MATRIXSIZE];
input_vec3 =(long long *) &input_mat1[mm * MATRIXSIZE * MATRIXSIZE + 2 * MATRIXSIZE];
input_vec4 =(long long *) &input_mat1[mm * MATRIXSIZE * MATRIXSIZE + 3 * MATRIXSIZE];
inputMatPtr1=(long long*) &input_mat2[mm * MATRIXSIZE * MATRIXSIZE + 0 * MATRIXSIZE];
inputMatPtr2=(long long*) &input_mat2[mm * MATRIXSIZE * MATRIXSIZE + 1 * MATRIXSIZE];
inputMatPtr3=(long long*) &input_mat2[mm * MATRIXSIZE * MATRIXSIZE + 2 * MATRIXSIZE];
inputMatPtr4=(long long*) &input_mat2[mm * MATRIXSIZE * MATRIXSIZE + 3 * MATRIXSIZE];
/* (4×4) X (4×4) */
llinputV1 =_amem8(input_vec1++ );
llinputV2 =_amem8(input_vec2++ );
llinputV3 =_amem8(input_vec3++ );
llinputV4 =_amem8(input_vec4++ );
llinputM1 =_amem8(inputMatPtr1++ );
llinputM2 =_amem8(inputMatPtr2++ );
// #ifdef _LITTLE_ENDIAN
inputMat = _llto128(llinputM2, llinputM1);
// #else
// inputMat = _llto128(llinputM1, llinputM2);
// #endif
acc1_128 = _cmatmpy(llinputV1, inputMat);
acc2_128 = _cmatmpy(llinputV2, inputMat);
acc3_128 = _cmatmpy(llinputV3, inputMat);
acc4_128 = _cmatmpy(llinputV4, inputMat);
acc1_32 = (int *)&acc1_128;
acc2_32 = (int *)&acc2_128;
acc3_32 = (int *)&acc3_128;
acc4_32 = (int *)&acc4_128;
acc1 = _itoll(_pack2(_amem4(acc1_32++)>>10,_amem4(acc1_32++)>>10),_pack2(_amem4(acc1_32++)>>10,_amem4(acc1_32++)>>10));
acc2 = _itoll(_pack2(_amem4(acc2_32++)>>10,_amem4(acc2_32++)>>10),_pack2(_amem4(acc2_32++)>>10,_amem4(acc2_32++)>>10));
acc3 = _itoll(_pack2(_amem4(acc3_32++)>>10,_amem4(acc3_32++)>>10),_pack2(_amem4(acc3_32++)>>10,_amem4(acc3_32++)>>10));
acc4 = _itoll(_pack2(_amem4(acc4_32++)>>10,_amem4(acc4_32++)>>10),_pack2(_amem4(acc4_32++)>>10,_amem4(acc4_32++)>>10));
llinputM1 = _amem8(inputMatPtr1++ );
llinputM2 = _amem8(inputMatPtr2++ );
// #ifdef _LITTLE_ENDIAN
inputMat = _llto128(llinputM2, llinputM1);
// #else
// inputMat = _llto128(llinputM1, llinputM2);
// #endif
acc5_128 = _cmatmpy(llinputV1, inputMat);
acc6_128 = _cmatmpy(llinputV2, inputMat);
acc7_128 = _cmatmpy(llinputV3, inputMat);
acc8_128 = _cmatmpy(llinputV4, inputMat);
acc5_32 = (int *)&acc5_128;
acc6_32 = (int *)&acc6_128;
acc7_32 = (int *)&acc7_128;
acc8_32 = (int *)&acc8_128;
acc5 = _itoll(_pack2(_amem4(acc5_32++)>>10,_amem4(acc5_32++)>>10),_pack2(_amem4(acc5_32++)>>10,_amem4(acc5_32++)>>10));
acc6 = _itoll(_pack2(_amem4(acc6_32++)>>10,_amem4(acc6_32++)>>10),_pack2(_amem4(acc6_32++)>>10,_amem4(acc6_32++)>>10));
acc7 = _itoll(_pack2(_amem4(acc7_32++)>>10,_amem4(acc7_32++)>>10),_pack2(_amem4(acc7_32++)>>10,_amem4(acc7_32++)>>10));
acc8 = _itoll(_pack2(_amem4(acc8_32++)>>10,_amem4(acc8_32++)>>10),_pack2(_amem4(acc8_32++)>>10,_amem4(acc8_32++)>>10));
llinputV1 =_amem8(input_vec1++ );
llinputV2 =_amem8(input_vec2++ );
llinputV3 =_amem8(input_vec3++ );
llinputV4 =_amem8(input_vec4++ );
llinputM1=_amem8(inputMatPtr3++ );
llinputM2=_amem8(inputMatPtr4++ );
// #ifdef _LITTLE_ENDIAN
inputMat = _llto128(llinputM2, llinputM1);
// #else
// inputMat = _llto128(llinputM1, llinputM2);
// #endif
// acc1 = _dadd2(_cmatmpyr1(llinputV1, inputMat), acc1);
// acc2 = _dadd2(_cmatmpyr1(llinputV2, inputMat), acc2);
// acc3 = _dadd2(_cmatmpyr1(llinputV3, inputMat), acc3);
// acc4 = _dadd2(_cmatmpyr1(llinputV4, inputMat), acc4);
acc1_128 = _cmatmpy(llinputV1, inputMat);
acc2_128 = _cmatmpy(llinputV2, inputMat);
acc3_128 = _cmatmpy(llinputV3, inputMat);
acc4_128 = _cmatmpy(llinputV4, inputMat);
acc1_32 = (int *)&acc1_128;
acc2_32 = (int *)&acc2_128;
acc3_32 = (int *)&acc3_128;
acc4_32 = (int *)&acc4_128;
acc1 = _dadd2( _itoll(_pack2(_amem4(acc1_32++)>>10,_amem4(acc1_32++)>>10),_pack2(_amem4(acc1_32++)>>10,_amem4(acc1_32++)>>10)),acc1);
acc2 = _dadd2( _itoll(_pack2(_amem4(acc2_32++)>>10,_amem4(acc2_32++)>>10),_pack2(_amem4(acc2_32++)>>10,_amem4(acc2_32++)>>10)),acc2);
acc3 = _dadd2( _itoll(_pack2(_amem4(acc3_32++)>>10,_amem4(acc3_32++)>>10),_pack2(_amem4(acc3_32++)>>10,_amem4(acc3_32++)>>10)),acc3);
acc4 = _dadd2( _itoll(_pack2(_amem4(acc4_32++)>>10,_amem4(acc4_32++)>>10),_pack2(_amem4(acc4_32++)>>10,_amem4(acc4_32++)>>10)),acc4);
llinputM1=_amem8(inputMatPtr3++ );
llinputM2=_amem8(inputMatPtr4++ );
// #ifdef _LITTLE_ENDIAN
inputMat = _llto128(llinputM2, llinputM1);
// #else
// inputMat = _llto128(llinputM1, llinputM2);
// #endif
// acc5 = _dadd2(_cmatmpyr1(llinputV1, inputMat), acc5);
// acc6 = _dadd2(_cmatmpyr1(llinputV2, inputMat), acc6);
// acc7 = _dadd2(_cmatmpyr1(llinputV3, inputMat), acc7);
// acc8 = _dadd2(_cmatmpyr1(llinputV4, inputMat), acc8);
acc5_128 = _cmatmpy(llinputV1, inputMat);
acc6_128 = _cmatmpy(llinputV2, inputMat);
acc7_128 = _cmatmpy(llinputV3, inputMat);
acc8_128 = _cmatmpy(llinputV4, inputMat);
acc5_32 = (int *)&acc5_128;
acc6_32 = (int *)&acc6_128;
acc7_32 = (int *)&acc7_128;
acc8_32 = (int *)&acc8_128;
acc5 = _dadd2( _itoll(_pack2(_amem4(acc5_32++)>>10,_amem4(acc5_32++)>>10),_pack2(_amem4(acc5_32++)>>10,_amem4(acc5_32++)>>10)),acc5);
acc6 = _dadd2( _itoll(_pack2(_amem4(acc6_32++)>>10,_amem4(acc6_32++)>>10),_pack2(_amem4(acc6_32++)>>10,_amem4(acc6_32++)>>10)),acc6);
acc7 = _dadd2( _itoll(_pack2(_amem4(acc7_32++)>>10,_amem4(acc7_32++)>>10),_pack2(_amem4(acc7_32++)>>10,_amem4(acc7_32++)>>10)),acc7);
acc8 = _dadd2( _itoll(_pack2(_amem4(acc8_32++)>>10,_amem4(acc8_32++)>>10),_pack2(_amem4(acc8_32++)>>10,_amem4(acc8_32++)>>10)),acc8);
_amem8(&output_mat[mm * MATRIXSIZE * MATRIXSIZE + 0 * MATRIXSIZE + 0]) = acc1;
_amem8(&output_mat[mm * MATRIXSIZE * MATRIXSIZE + 0 * MATRIXSIZE + 2]) = acc5;
_amem8(&output_mat[mm * MATRIXSIZE * MATRIXSIZE + 1 * MATRIXSIZE + 0]) = acc2;
_amem8(&output_mat[mm * MATRIXSIZE * MATRIXSIZE + 1 * MATRIXSIZE + 2]) = acc6;
_amem8(&output_mat[mm * MATRIXSIZE * MATRIXSIZE + 2 * MATRIXSIZE + 0]) = acc3;
_amem8(&output_mat[mm * MATRIXSIZE * MATRIXSIZE + 2 * MATRIXSIZE + 2]) = acc7;
_amem8(&output_mat[mm * MATRIXSIZE * MATRIXSIZE + 3 * MATRIXSIZE + 0]) = acc4;
_amem8(&output_mat[mm * MATRIXSIZE * MATRIXSIZE + 3 * MATRIXSIZE + 2]) = acc8;
}
}

wenhao zhang：

附件是源代码

Andy Yin1：

回复 wenhao zhang:

建议将你的代码细化成多个模块，分模块打开优化看看是哪个模块影响的，还有就是可以比较一下优化前后的汇编程序，这个确实不太好查。

赞(0)
未经允许不得转载：TI中文支持网 » 下面的程序开优化后，运行结果出错，求助
分享到： 更多 (0)
标签：

相关推荐

热门标签

ECTE (3265)#希望得到TI的工程师回答 (1566)C2000WARE (918)DSP (853)希望得到TI工程师回答# (853)盼望能得到TI的人回答 (753)UniFlash (751)TMS320F28335 (750)电源管理 (748)TMS320C6678 (720)【紧急+重要】 (703)技术询问 (691)cc1310 (618)TDA4VM (612)电源 (599)Controlsuite (576)AWR2243 (531)6678 (530)AWR1642 (516)电池管理 (495)【紧急】+重要 (492)sysbios (486)AM5728 (486)CC2640R2F (448)AM3352 (446)#DCA1000EVM (432)放大器 (417)TMS320C6748 (410)bq40z50 (403)28335 (400)

热门文章