g++上证所内在困境——内在“饱和”带来的价值
我编写了一个简单的程序来实现 SSE 内在函数,用于计算两个大型(100000 个或更多元素)向量的内积。该程序比较了传统方法和使用内在函数计算内积的执行时间。一切都很顺利,直到我在计算内积的语句之前插入(只是为了好玩)一个内循环。在进一步讨论之前,先看一下代码:
//this is a sample Intrinsics program to compute inner product of two vectors and compare Intrinsics with traditional method of doing things.
#include <iostream>
#include <iomanip>
#include <xmmintrin.h>
#include <stdio.h>
#include <time.h>
#include <stdlib.h>
using namespace std;
typedef float v4sf __attribute__ ((vector_size(16)));
double innerProduct(float* arr1, int len1, float* arr2, int len2) { //assume len1 = len2.
float result = 0.0;
for(int i = 0; i < len1; i++) {
for(int j = 0; j < len1; j++) {
result += (arr1[i] * arr2[i]);
}
}
//float y = 1.23e+09;
//cout << "y = " << y << endl;
return result;
}
double sse_v4sf_innerProduct(float* arr1, int len1, float* arr2, int len2) { //assume that len1 = len2.
if(len1 != len2) {
cout << "Lengths not equal." << endl;
exit(1);
}
/*steps:
* 1. load a long-type (4 float) into a v4sf type data from both arrays.
* 2. multiply the two.
* 3. multiply the same and store result.
* 4. add this to previous results.
*/
v4sf arr1Data, arr2Data, prevSums, multVal, xyz;
//__builtin_ia32_xorps(prevSums, prevSums); //making it equal zero.
//can explicitly load 0 into prevSums using loadps or storeps (Check).
float temp[4] = {0.0, 0.0, 0.0, 0.0};
prevSums = __builtin_ia32_loadups(temp);
float result = 0.0;
for(int i = 0; i < (len1 - 3); i += 4) {
for(int j = 0; j < len1; j++) {
arr1Data = __builtin_ia32_loadups(&arr1[i]);
arr2Data = __builtin_ia32_loadups(&arr2[i]); //store the contents of two arrays.
multVal = __builtin_ia32_mulps(arr1Data, arr2Data); //multiply.
xyz = __builtin_ia32_addps(multVal, prevSums);
prevSums = xyz;
}
}
//prevSums will hold the sums of 4 32-bit floating point values taken at a time. Individual entries in prevSums also need to be added.
__builtin_ia32_storeups(temp, prevSums); //store prevSums into temp.
cout << "Values of temp:" << endl;
for(int i = 0; i < 4; i++)
cout << temp[i] << endl;
result += temp[0] + temp[1] + temp[2] + temp[3];
return result;
}
int main() {
clock_t begin, end;
int length = 100000;
float *arr1, *arr2;
double result_Conventional, result_Intrinsic;
// printStats("Allocating memory.");
arr1 = new float[length];
arr2 = new float[length];
// printStats("End allocation.");
srand(time(NULL)); //init random seed.
// printStats("Initializing array1 and array2");
begin = clock();
for(int i = 0; i < length; i++) {
// for(int j = 0; j < length; j++) {
// arr1[i] = rand() % 10 + 1;
arr1[i] = 2.5;
// arr2[i] = rand() % 10 - 1;
arr2[i] = 2.5;
// }
}
end = clock();
cout << "Time to initialize array1 and array2 = " << ((double) (end - begin)) / CLOCKS_PER_SEC << endl;
// printStats("Finished initialization.");
// printStats("Begin inner product conventionally.");
begin = clock();
result_Conventional = innerProduct(arr1, length, arr2, length);
end = clock();
cout << "Time to compute inner product conventionally = " << ((double) (end - begin)) / CLOCKS_PER_SEC << endl;
// printStats("End inner product conventionally.");
// printStats("Begin inner product using Intrinsics.");
begin = clock();
result_Intrinsic = sse_v4sf_innerProduct(arr1, length, arr2, length);
end = clock();
cout << "Time to compute inner product with intrinsics = " << ((double) (end - begin)) / CLOCKS_PER_SEC << endl;
//printStats("End inner product using Intrinsics.");
cout << "Results: " << endl;
cout << " result_Conventional = " << result_Conventional << endl;
cout << " result_Intrinsics = " << result_Intrinsic << endl;
return 0;
}
我使用以下 g++ 调用来构建此代码:
g++ -W -Wall -O2 -pedantic -march=i386 -msse intrinsics_SSE_innerProduct.C -o innerProduct
在这两个函数中,上面的每个循环总共运行 N^2 次。但是,考虑到 arr1 和 arr2(两个浮点向量)加载的值为 2.5,数组的长度为 100,000,两种情况下的结果应为 6.25e+10。我得到的结果是:
结果:
结果_常规 = 6.25e+10
result_Intrinsics = 5.36871e+08
这还不是全部。似乎从使用内在函数的函数返回的值在上述值处“饱和”。我尝试为数组的元素设置其他值以及不同的大小。但似乎数组内容大于 1.0 的任何值以及大于 1000 的任何大小都符合我们上面看到的相同值。
最初,我认为这可能是因为SSE中的所有操作都是浮点型的,但是浮点型应该能够存储e+08量级的数字。
我试图找出哪里可能出错,但似乎无法弄清楚。我使用的 g++ 版本:g++ (GCC) 4.4.1 20090725 (Red Hat 4.4.1-2)。
非常欢迎对此提供任何帮助。
谢谢,
斯里拉姆。
I wrote a simple program to implement SSE intrinsics for computing the inner product of two large (100000 or more elements) vectors. The program compares the execution time for both, inner product computed the conventional way and using intrinsics. Everything works out fine, until I insert (just for the fun of it) an inner loop before the statement that computes the inner product. Before I go further, here is the code:
//this is a sample Intrinsics program to compute inner product of two vectors and compare Intrinsics with traditional method of doing things.
#include <iostream>
#include <iomanip>
#include <xmmintrin.h>
#include <stdio.h>
#include <time.h>
#include <stdlib.h>
using namespace std;
typedef float v4sf __attribute__ ((vector_size(16)));
double innerProduct(float* arr1, int len1, float* arr2, int len2) { //assume len1 = len2.
float result = 0.0;
for(int i = 0; i < len1; i++) {
for(int j = 0; j < len1; j++) {
result += (arr1[i] * arr2[i]);
}
}
//float y = 1.23e+09;
//cout << "y = " << y << endl;
return result;
}
double sse_v4sf_innerProduct(float* arr1, int len1, float* arr2, int len2) { //assume that len1 = len2.
if(len1 != len2) {
cout << "Lengths not equal." << endl;
exit(1);
}
/*steps:
* 1. load a long-type (4 float) into a v4sf type data from both arrays.
* 2. multiply the two.
* 3. multiply the same and store result.
* 4. add this to previous results.
*/
v4sf arr1Data, arr2Data, prevSums, multVal, xyz;
//__builtin_ia32_xorps(prevSums, prevSums); //making it equal zero.
//can explicitly load 0 into prevSums using loadps or storeps (Check).
float temp[4] = {0.0, 0.0, 0.0, 0.0};
prevSums = __builtin_ia32_loadups(temp);
float result = 0.0;
for(int i = 0; i < (len1 - 3); i += 4) {
for(int j = 0; j < len1; j++) {
arr1Data = __builtin_ia32_loadups(&arr1[i]);
arr2Data = __builtin_ia32_loadups(&arr2[i]); //store the contents of two arrays.
multVal = __builtin_ia32_mulps(arr1Data, arr2Data); //multiply.
xyz = __builtin_ia32_addps(multVal, prevSums);
prevSums = xyz;
}
}
//prevSums will hold the sums of 4 32-bit floating point values taken at a time. Individual entries in prevSums also need to be added.
__builtin_ia32_storeups(temp, prevSums); //store prevSums into temp.
cout << "Values of temp:" << endl;
for(int i = 0; i < 4; i++)
cout << temp[i] << endl;
result += temp[0] + temp[1] + temp[2] + temp[3];
return result;
}
int main() {
clock_t begin, end;
int length = 100000;
float *arr1, *arr2;
double result_Conventional, result_Intrinsic;
// printStats("Allocating memory.");
arr1 = new float[length];
arr2 = new float[length];
// printStats("End allocation.");
srand(time(NULL)); //init random seed.
// printStats("Initializing array1 and array2");
begin = clock();
for(int i = 0; i < length; i++) {
// for(int j = 0; j < length; j++) {
// arr1[i] = rand() % 10 + 1;
arr1[i] = 2.5;
// arr2[i] = rand() % 10 - 1;
arr2[i] = 2.5;
// }
}
end = clock();
cout << "Time to initialize array1 and array2 = " << ((double) (end - begin)) / CLOCKS_PER_SEC << endl;
// printStats("Finished initialization.");
// printStats("Begin inner product conventionally.");
begin = clock();
result_Conventional = innerProduct(arr1, length, arr2, length);
end = clock();
cout << "Time to compute inner product conventionally = " << ((double) (end - begin)) / CLOCKS_PER_SEC << endl;
// printStats("End inner product conventionally.");
// printStats("Begin inner product using Intrinsics.");
begin = clock();
result_Intrinsic = sse_v4sf_innerProduct(arr1, length, arr2, length);
end = clock();
cout << "Time to compute inner product with intrinsics = " << ((double) (end - begin)) / CLOCKS_PER_SEC << endl;
//printStats("End inner product using Intrinsics.");
cout << "Results: " << endl;
cout << " result_Conventional = " << result_Conventional << endl;
cout << " result_Intrinsics = " << result_Intrinsic << endl;
return 0;
}
I use the following g++ invocation to build this:
g++ -W -Wall -O2 -pedantic -march=i386 -msse intrinsics_SSE_innerProduct.C -o innerProduct
Each of the loops above, in both the functions, runs a total of N^2 times. However, given that arr1 and arr2 (the two floating point vectors) are loaded with a value 2.5, the length of the array is 100,000, the result in both cases should be 6.25e+10. The results I get are:
Results:
result_Conventional = 6.25e+10
result_Intrinsics = 5.36871e+08
This is not all. It seems that the value returned from the function that uses intrinsics "saturates" at the value above. I tried putting other values for the elements of the array and different sizes too. But it seems that any value above 1.0 for the array contents and any size above 1000 meets with the same value we see above.
Initially, I thought it might be because all operations within SSE are in floating point, but floating point should be able to store a number that is of the order of e+08.
I am trying to see where I could be going wrong but cannot seem to figure it out. I am using g++ version: g++ (GCC) 4.4.1 20090725 (Red Hat 4.4.1-2).
Any help on this is most welcome.
Thanks,
Sriram.
如果你对这篇内容有疑问,欢迎到本站社区发帖提问 参与讨论,获取更多帮助,或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱,如果其他用户或者作者回复了您的评论,将不能在第一时间通知您!
发布评论
评论(1)
您遇到的问题是,虽然
float
可以存储 6.25e+10,但它只有几个有效数字的精度。这意味着,当您通过一次一点地添加大量小数字来构建一个大数字时,您会达到一个点,即较小的数字小于较大数字中的最低精度数字,因此将其相加没有效果。
至于为什么在非内在版本中没有得到这种行为,很可能
result
变量被保存在一个寄存器中,该寄存器使用比浮点的实际存储更高的精度,因此它是在循环的每次迭代中不会被截断为float
的精度。您必须查看生成的汇编代码才能确定。The problem that you are having is that while a
float
can store 6.25e+10, it only has a few significant digits of precision.This means that when you are building a large number by adding lots of small numbers together a bit at a time, you reach a point where the smaller number is smaller than the lowest precision digit in the larger number so adding it up has no effect.
As to why you are not getting this behaviour in the non-intrinsic version, it is likely that
result
variable is being held in a register which uses a higher precision that the actual storage of a float so it is not being truncated to the precision of afloat
on every iteration of the loop. You would have to look at the generated assembler code to be sure.