我编写了以下代码来缩放一组数字:
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include "immintrin.h"
void scale(struct problem_param prob_param, float* features)
{
int i,j,k;
for (j = 0; j < prob_param.Nr_ft; j++)
{
__m256 range_vec ,low_up_vec , low_vec,tmp_vec;
__m256 feat_min_vec, feat_vec;
unsigned count = prob_param.Size;
unsigned offset = j * prob_param.Size;
float feature_max = features[offset];
float feature_min = features[offset];
/*
* Look for min and max of each feature.
*/
for ( i = 1; i < prob_param.Size ; i++)
{
if (features[i + offset] > feature_max ) feature_max = features[i + offset];
if (features[i + offset] < feature_min ) feature_min = features[i + offset];
}
printf("feature : %u \t min = %f \t max = %f \n",j,feature_min,feature_max);
/*
* Set the range.
* Set constant vectors for the vector instructions.
*/
float range = feature_max - feature_min;
feat_min_vec = _mm256_set1_ps (feature_min);
range_vec = _mm256_set1_ps (range);
low_up_vec = _mm256_set1_ps (prob_param.upper_limit - prob_param.lower_limit);
low_vec = _mm256_set1_ps (prob_param.lower_limit);
/*
* Normalising
* -----------
* Head
*/
for ( i = 0; i < prob_param.Size && count >= 7 ; i+=8)
{
feat_vec = _mm256_load_ps(&features[i + offset]);
tmp_vec = _mm256_sub_ps(feat_vec,feat_min_vec);
tmp_vec = _mm256_mul_ps(tmp_vec,low_up_vec);
tmp_vec = _mm256_div_ps(tmp_vec,range_vec);
feat_vec = _mm256_add_ps(tmp_vec,low_vec);
_mm256_store_ps (&features[i + offset], feat_vec);
count -=8;
}
/*
* Normalising
* -----------
* Tail
*/
for ( k = i; k < prob_param.Size ; k++)
{
features[k + offset] = prob_param.lower_limit + (prob_param.upper_limit - prob_param.lower_limit) * (features[k + offset] - feature_min) / range;
}
}
这是负责缩放的函数,我这样称呼它:
#include <stdio.h>
#include <stdlib.h>
#include "data.h"
#include "common.h"
#define training_size 3089
#define number_features 4
#define low -1.0
#define up 1.0
float* feature_array;
int main()
{
struct problem_param pp;
pp.Size = training_size;
pp.Nr_ft = number_features;
pp.lower_limit = low;
pp.upper_limit = up;
posix_memalign((void **) &feature_array, 32, (size_t) training_size * number_features *sizeof(float));
scale(pp,feature_array);
return EXIT_SUCCESS;
}
我用MacBook Pro Core i5 Haswell测试了这段代码,它工作正常,但当我用华硕Core I7 Haswell测试时,它显示了加载的分段错误。我遗漏了什么吗?
最佳答案
offset
(因此i + offset
)的值并不总是8的倍数(在上面的示例中,它等于0、3089、6178、9267),因此您的加载和存储内部函数通常会不对齐。
最简单的解决方案是用_mm256_loadu_ps
代替_mm256_load_ps
,用_mm256_storeu_ps
代替_mm256_store_ps
。
至于为什么这看起来对你的MacBook Pro有效,我猜clang会在你背后生成未对齐的加载/存储指令,从而隐藏问题,直到你尝试在使用不同编译器的系统上运行代码。
更新:我刚刚通过编译和反汇编生成的代码(在带有macOS 10.13.4和Xcode 9.3.1的Haswell MacBook Pro上)验证了上述假设:
>>> vmovups (%r14,%r13,4), %ymm0
vsubps 192(%rsp), %ymm0, %ymm0 ## 32-byte Folded Reload
vmulps 448(%rsp), %ymm0, %ymm0 ## 32-byte Folded Reload
vdivps 384(%rsp), %ymm0, %ymm0 ## 32-byte Folded Reload
vaddps 416(%rsp), %ymm0, %ymm0 ## 32-byte Folded Reload
>>> vmovups %ymm0, (%r14,%r13,4)
注意使用
vmovups
而不是vmovaps
。关于c - 为什么load_ps()可在一台PC上运行而不能在另一台PC上运行?,我们在Stack Overflow上找到一个类似的问题:https://stackoverflow.com/questions/50469378/