Trying to start working with SIMD instructions, but having problems with error (0xC0000005) in C

Question

I'm trying to learn to work with SIMD instructions in C. I decided to start working with SSE. I am using Windows 8.1 and am coding with Codeblocks in a Windows environment. My CodeBlocks settings are set to use an Intel i7 (which I have), so SSE is enabled. My program compiles just fine, but I get an error as soon as runtime hits. The error is "(0xC0000005)". Here is my code:

#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <xmmintrin.h>

float ScalarSSE(float *m1, float *m2) {

    float prod;
    int i;
    __m128 X, Y, Z;

    for(i=0; i<5; i+=4) {
        X = _mm_load_ps(&*m1);
        Y = _mm_load_ps(&*m2);
        X = _mm_mul_ps(X, Y);
        Z = _mm_add_ps(X, Z);
    }

    for(i=0; i<4; i++) {
        prod += Z[i];
    }

    return prod;

}

int main() {

    int i;
    float *s1 = calloc(1,sizeof (float));
    float *s2 = calloc(1,sizeof (float));

    for(i=0; i<100; i++) {
        *s1 = 2;
        *s2 = 2;
        float scalar_product_sse = ScalarSSE(s1, s2);
    }

    printf("Done");

    free (s1);
    free (s2);

}

I cannot use debug mode as I am not in a project nor know how to open up a project in CodeBlocks (it gives me errors :( )

I am wondering how to make this work without any errors. Thanks!

Answer 1

Here you are allocating space for one float:

float *s1 = calloc(1,sizeof (float));

You are then accessing more than one float in ScalarSSE

Maybe something like this will help, although determining the exact number to use would be better:

float *s1 = calloc(100,sizeof (float));
float *s2 = calloc(100,sizeof (float));

Additionally:

for(i=0; i<100; i++) {
    *s1 = 2;
    *s2 = 2;
    float scalar_product_sse = ScalarSSE(s1, s2);
}

This does the same thing 100 times - is that what you intend?

"(0xC0000005)" means 'Access violation' -- basically you have tried to access memory you are not allowed to touch (or doesn't exist)

Answer 2

The code had so many issues that I'll just state some of them, that most likely were causing the access violation and provide a working example, and then you can try to modify it as for your needs, since it is really hard to understand what you are trying to achieve.

1) Memory alignment - _mm_load_ps takes a 16 byte aligned address, you must use a proper memory allocation function where you can specify the alignment, like _aligned_malloc or wrapping the floats in some kind of aligned struct like in the example

2) You are using uninitialized value of Z and prod so results after the additions would be undefined.

3) as mentioned in the comments _mm_load_ps needs a vector of 4 floats, so allocating 1 float as you are doing is invalid, if you actually wanted to load only one value see _mm_load1_ps, that loads one float into all four values of __m128 type, but it is not possible to understand what the intial idea was because of the for loop.

Here is an example, based on your code and without the strange for loops, that works, I hope it is enough to build upon to achieve what you wanted:

#include <xmmintrin.h>
#include <stdio.h>
#include <malloc.h>

// This ensures the proper alignment
__declspec(align(16)) struct float_vector
{
    float a, b, c, d;
};

float ScalarSSE(float *m1, float *m2) {

    float prod = 0;
    int i;
    __m128 X, Y, Z;

    X = _mm_load_ps(m1); // This loads 4 values, so the pointer should point to an array of at least 4 floats
    Y = _mm_load_ps(m2); // This loads 4 values, so the pointer should point to an array of at least 4 floats
    Z = _mm_mul_ps(X, Y);

    for(i=0; i<4; i++) {
        prod += (reinterpret_cast<float*>(&Z))[i];
    }

    return prod;

}

int main() {

    float_vector s1 = { 2.0, 2.0, 2.0, 2.0 };
    float_vector s2 = { 2.0, 2.0, 2.0, 2.0 };

    float scalar_product_sse = ScalarSSE(&s1.a, &s2.a);
    printf("Done, result: %f\n", scalar_product_sse);

    return 0;
}

And it actually gives a valid result - (2*2) + (2*2) + (2*2) + (2*2) = 16.