我有一些带有SSE2指令的c ++函数。问题是我在使用microsoft visual c ++编译此代码时遇到以下链接器错误：
函数＆＃34; void * __cdecl中引用的未解析的外部符号_m_empty 过程（无效*）＆＃34;
当我评论_m_empty时，我会收到运行时错误！ 但它应该用于MMX指令，不应该吗？
#include "mex.h"
#include <pthread.h>
#include <emmintrin.h>
#include <stdint.h>
#include <stdlib.h>
#define malloc_aligned(a,b) _aligned_malloc(a,b)
#define IS_ALIGNED(ptr) ((((uintptr_t)(ptr)) & 0xF) == 0)
#define NUM_FEATURES 32
#define __attribute__(A) /* do nothing */
/*
* This code is used for computing filter responses. It computes the
* response of a set of filters with a feature map.
*
* Multithreaded version.
*/
struct thread_data {
float *A;
float *B;
double *C;
mxArray *mxC;
const mwSize *A_dims;
const mwSize *B_dims;
mwSize C_dims[2];
};
// convolve A and B
void *process(void *thread_arg) {
thread_data *args = (thread_data *)thread_arg;
float *A = args->A;
float *B = args->B;
double *C = args->C;
const mwSize *A_dims = args->A_dims;
const mwSize *B_dims = args->B_dims;
const mwSize *C_dims = args->C_dims;
__m128 a,b,c;
double *dst = C;
for (int x = 0; x < C_dims[1]; x++) {
for (int y = 0; y < C_dims[0]; y++) {
__m128 v = _mm_setzero_ps();
const float *A_src = A + y*NUM_FEATURES + x*A_dims[0]*NUM_FEATURES;
const float *B_src = B;
for (int xp = 0; xp < B_dims[1]; xp++) {
const float *A_off = A_src;
const float *B_off = B_src;
for (int yp = 0; yp < B_dims[0]; yp++) {
a = _mm_load_ps(A_off+0);
b = _mm_load_ps(B_off+0);
c = _mm_mul_ps(a, b);
v = _mm_add_ps(v, c);
a = _mm_load_ps(A_off+4);
b = _mm_load_ps(B_off+4);
c = _mm_mul_ps(a, b);
v = _mm_add_ps(v, c);
a = _mm_load_ps(A_off+8);
b = _mm_load_ps(B_off+8);
c = _mm_mul_ps(a, b);
v = _mm_add_ps(v, c);
a = _mm_load_ps(A_off+12);
b = _mm_load_ps(B_off+12);
c = _mm_mul_ps(a, b);
v = _mm_add_ps(v, c);
a = _mm_load_ps(A_off+16);
b = _mm_load_ps(B_off+16);
c = _mm_mul_ps(a, b);
v = _mm_add_ps(v, c);
a = _mm_load_ps(A_off+20);
b = _mm_load_ps(B_off+20);
c = _mm_mul_ps(a, b);
v = _mm_add_ps(v, c);
a = _mm_load_ps(A_off+24);
b = _mm_load_ps(B_off+24);
c = _mm_mul_ps(a, b);
v = _mm_add_ps(v, c);
a = _mm_load_ps(A_off+28);
b = _mm_load_ps(B_off+28);
c = _mm_mul_ps(a, b);
v = _mm_add_ps(v, c);
// N.B. Unroll me more/less if you change NUM_FEATURES
A_off += NUM_FEATURES;
B_off += NUM_FEATURES;
}
A_src += A_dims[0]*NUM_FEATURES;
B_src += B_dims[0]*NUM_FEATURES;
}
// buf[] must be 16-byte aligned
float buf[4] __attribute__ ((aligned (16)));
_mm_store_ps(buf, v);
_mm_empty();
*(dst++) = buf[0]+buf[1]+buf[2]+buf[3];
}
}
pthread_exit(NULL);
return 0;
}
float *prepare(float *in, const int *dims) {
float *F = (float *)malloc_aligned(16, dims[0]*dims[1]*NUM_FEATURES*sizeof(float));
// Sanity check that memory is aligned
if (!IS_ALIGNED(F))
mexErrMsgTxt("Memory not aligned");
float *p = F;
for (int x = 0; x < dims[1]; x++) {
for (int y = 0; y < dims[0]; y++) {
for (int f = 0; f < dims[2]; f++)
*(p++) = in[y + f*dims[0]*dims[1] + x*dims[0]];
for (int f = dims[2]; f < NUM_FEATURES; f++)
*(p++) = 0;
}
}
return F;
}
// matlab entry point
// C = fconv(A, cell of B, start, end);
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) {
if (nrhs != 4)
mexErrMsgTxt("Wrong number of inputs");
if (nlhs != 1)
mexErrMsgTxt("Wrong number of outputs");
// get A
const mxArray *mxA = prhs[0];
if (mxGetNumberOfDimensions(mxA) != 3 ||
mxGetClassID(mxA) != mxSINGLE_CLASS)
mexErrMsgTxt("Invalid input: A");
// get B and start/end
const mxArray *cellB = prhs[1];
mwSize num_bs = mxGetNumberOfElements(cellB);
int start = (int)mxGetScalar(prhs[2]) - 1;
int end = (int)mxGetScalar(prhs[3]) - 1;
if (start < 0 || end >= num_bs || start > end)
mexErrMsgTxt("Invalid input: start/end");
int len = end-start+1;
// start threads
thread_data *td = (thread_data *)mxCalloc(len, sizeof(thread_data));
pthread_t *ts = (pthread_t *)mxCalloc(len, sizeof(pthread_t));
const mwSize *A_dims = mxGetDimensions(mxA);
float *A = prepare((float *)mxGetPr(mxA), A_dims);
for (int i = 0; i < len; i++) {
const mxArray *mxB = mxGetCell(cellB, i+start);
td[i].A_dims = A_dims;
td[i].A = A;
td[i].B_dims = mxGetDimensions(mxB);
td[i].B = prepare((float *)mxGetPr(mxB), td[i].B_dims);
if (mxGetNumberOfDimensions(mxB) != 3 ||
mxGetClassID(mxB) != mxSINGLE_CLASS ||
td[i].A_dims[2] != td[i].B_dims[2])
mexErrMsgTxt("Invalid input: B");
// compute size of output
int height = td[i].A_dims[0] - td[i].B_dims[0] + 1;
int width = td[i].A_dims[1] - td[i].B_dims[1] + 1;
if (height < 1 || width < 1)
mexErrMsgTxt("Invalid input: B should be smaller than A");
td[i].C_dims[0] = height;
td[i].C_dims[1] = width;
td[i].mxC = mxCreateNumericArray(2, td[i].C_dims, mxDOUBLE_CLASS, mxREAL);
td[i].C = (double *)mxGetPr(td[i].mxC);
if (pthread_create(&ts[i], NULL, process, (void *)&td[i]))
mexErrMsgTxt("Error creating thread");
}
// wait for the treads to finish and set return values
void *status;
plhs[0] = mxCreateCellMatrix(1, len);
for (int i = 0; i < len; i++) {
pthread_join(ts[i], &status);
mxSetCell(plhs[0], i, td[i].mxC);
free(td[i].B);
}
mxFree(td);
mxFree(ts);
free(A);
}
答案 0 :(得分：2)
根据this link，MMX未针对x64实施。使用成熟的SSE2 n x64。
答案 1 :(得分：1)
我认为代码中的_mm_empty会变成对_m_empty的引用，因为they're synonyms，而您的构建环境仍然有一些带有#define _mm_empty _m_empty或其他内容的标头。< / p>
但奇怪的是，您的构建环境实际上并未提供内在的定义。有关于它被隐式声明的编译器警告吗？这很奇怪，因为我认为完全缺乏MMX支持意味着_mm_empty / _m_empty等效性也不存在。
运行时错误可能与此无关。就像Paul R在评论中指出的那样，你假设如果你可以得到未经修改的源代码来编译，它就会起作用。情况可能并非如此，因为_mm_empty看起来对于对这个问题发表评论的x86 asm专家（包括我）来说是不必要的。
我认为Paul R猜测你可能在某个地方有一个未对齐的指针听起来很合理。任何这些数组的对齐是否依赖于sizeof指针？如果struct thread_data类似于：
struct thread_data {
some_type *ptr1;
some_type *ptr2;
int a;
int b;
float A[1024];
float B[1024];
...
};
然后32位构建将具有16B对齐的阵列，但64位构建不会。
因此调试运行时错误，并找出它是什么。 如果您告诉我们的所有内容都是“运行时错误”，我们无法帮助您。如果您首先告诉我们它的内容，我们可以告诉您这种方式或其他方式是否可能与删除_mm_empty。
答案 2 :(得分：0)
感谢哈罗德，保罗和彼得，我发现了这个问题！你是对的，运行时错误与_mm_empty无关！问题是_aligned_malloc输入参数顺序。当我交换输入时，运行时错误消失了。
另一个错误是free（）函数。 _aligned_free（）必须用于释放对齐的内存。
正如哈罗德所说，我改变了主循环以使用三个独立的v。如果我做错了，请纠正我。现在程序（不是函数！）运行速度提高了300ms（2.4s - > 2.1s）。
void *process(void *thread_arg) {
thread_data *args = (thread_data *)thread_arg;
float *A = args->A;
float *B = args->B;
double *C = args->C;
const mwSize *A_dims = args->A_dims;
const mwSize *B_dims = args->B_dims;
const mwSize *C_dims = args->C_dims;
__m128 a,b,c;
double *dst = C;
for (int x = 0; x < C_dims[1]; x++) {
for (int y = 0; y < C_dims[0]; y++) {
__m128 v = _mm_setzero_ps(), v1 = _mm_setzero_ps(), v2 = _mm_setzero_ps(), v3 = _mm_setzero_ps();
const float *A_src = A + y*NUM_FEATURES + x*A_dims[0]*NUM_FEATURES;
const float *B_src = B;
for (int xp = 0; xp < B_dims[1]; xp++) {
const float *A_off = A_src;
const float *B_off = B_src;
for (int yp = 0; yp < B_dims[0]; yp++) {
a = _mm_load_ps(A_off+0);
b = _mm_load_ps(B_off+0);
c = _mm_mul_ps(a, b);
v1 = _mm_add_ps(v1, c);
a = _mm_load_ps(A_off+4);
b = _mm_load_ps(B_off+4);
c = _mm_mul_ps(a, b);
v2 = _mm_add_ps(v2, c);
a = _mm_load_ps(A_off+8);
b = _mm_load_ps(B_off+8);
c = _mm_mul_ps(a, b);
v3 = _mm_add_ps(v3, c);
a = _mm_load_ps(A_off+12);
b = _mm_load_ps(B_off+12);
c = _mm_mul_ps(a, b);
v1 = _mm_add_ps(v1, c);
a = _mm_load_ps(A_off+16);
b = _mm_load_ps(B_off+16);
c = _mm_mul_ps(a, b);
v2 = _mm_add_ps(v2, c);
a = _mm_load_ps(A_off+20);
b = _mm_load_ps(B_off+20);
c = _mm_mul_ps(a, b);
v3 = _mm_add_ps(v3, c);
a = _mm_load_ps(A_off+24);
b = _mm_load_ps(B_off+24);
c = _mm_mul_ps(a, b);
v1 = _mm_add_ps(v1, c);
a = _mm_load_ps(A_off+28);
b = _mm_load_ps(B_off+28);
c = _mm_mul_ps(a, b);
v2 = _mm_add_ps(v2, c);
// N.B. Unroll me more/less if you change NUM_FEATURES
A_off += NUM_FEATURES;
B_off += NUM_FEATURES;
}
A_src += A_dims[0]*NUM_FEATURES;
B_src += B_dims[0]*NUM_FEATURES;
}
v = _mm_add_ps(v, v1);
v = _mm_add_ps(v, v2);
v = _mm_add_ps(v, v3);
// buf[] must be 16-byte aligned
__declspec(align(16)) float buf[4];
_mm_store_ps(buf, v);
*(dst++) = buf[0]+buf[1]+buf[2]+buf[3];
}
}
pthread_exit(NULL);
return 0;
}