ICC 15.0 generates slower AVX codes than SSSE3 codes

[Leo_W_]

#include <stdio.h>
#include <stdlib.h>
#include <string.h>

int matrix_rot(uint8_t* src, uint8_t* des, int w, int h) {
	for (int y = 0; y < h; ++y) {
		const int stride = y * w + h;
		for (int x = 0; x < w; ++x) {
			des[x*h+y] = src[stride-x];
		}
	}
	return 0;
}

int matrix_dump(uint8_t* buf, int w, int h) {
	for (int y = 0; y < h; y++) {
		for (int x = 0; x < w; x++) {
			printf("%d ", buf[y*w+x]);
		}
		printf("\n");
	}
	return 0;
}

int main(int argc, char** argv) {
	if (argc != 3) return -1;
	int w = strtol(argv[1], NULL, 10);
	int h = strtol(argv[2], NULL, 10);

	uint8_t *bmp = malloc(w*h);
	uint8_t *dst = malloc(w*h);

	srand(12);
	int c = rand();
	printf("%d\n", c);
	for (int i = 0; i < c; i++) {
		matrix_rot(bmp, dst, w, h); 
		matrix_rot(bmp, dst, w, h); 
	}

	matrix_dump(bmp, 4, 3);
	matrix_dump(dst, 3, 4);
	free(bmp);
	free(dst);

	return 0;
}

ICC 15.0 generates slower AVX/SSE4.2 codes than SSSE3 codes.

The SSSE3 compilation command line and result:

$ icc -std=c99 -xSSSE3 -O3 -o rot rot.c 
$ time ./rot 320 240                   
201684
0 0 0 0 
0 0 0 0 
0 0 0 0 
0 0 0 
0 0 0 
0 0 0 
0 0 0 
./rot 320 240  10.06s user 0.01s system 99% cpu 10.077 total

The AVX compilation command line and result:

$ icc -std=c99 -xSSE4.2 -O3 -o rot rot.c
$ time ./rot 320 240                    
201684
0 0 0 0 
0 0 0 0 
0 0 0 0 
0 0 0 
0 0 0 
0 0 0 
0 0 0 
./rot 320 240  16.64s user 0.00s system 99% cpu 16.650 total

As you can see, AVX version runs 60% slower than the SSSE3 version.

[TimP]

You'd probably want to compare opt-report4 or vec-report.  I didn't think AVX offered any new facilities for these data types; if AVX option removes optimizations, it seems a deficiency.  If there is any new AVX optimization, you would want to try the tests with 32-byte data alignment.

it may not be surprising if AVX doesn't offer an advantage on strided stores, or if the AVX option brings vectorization which wasn't there before, that it isn't advantageous.   Does one of the options result in loop interchange?

I was able to compile after adding #include stdint.h

Apparently, the vectorization strategy changes from "outer loop vectorized" to one with "heavy-overhead vector operations."

So, addition of #pragma omp simd on the outer loop (so as to return in effect to outer loop vectorization) shows a slight advantage for AVX over SSSE3; a larger advantage for AVX2 (on a second warm cache execution) with aligned_malloc.