Determine tile size for matrix multiplication with AVX - c

Question

I wrote the code below for tiled matrix multiplication with AVX.
#include <x86intrin.h>
#include <immintrin.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
void fill(double *mat, int n) {
register int i;
register int j;
for (i = 0; i < n; ++i) {
for (j = 0; j < n; ++j) {
mat[i * n + j] = (double) (i * n + j);
}
}
}
void blocked_mmul_avx(double *A, double *B, double *C, int n) {
register int row, block, chunk, sub_chunk;
int block_size = 4;
for (row = 0; row < n; row++)
for (block = 0; block < n; block += block_size) {
register __m256d c0 = _mm256_setzero_pd();
for (chunk = 0; chunk < n; chunk += block_size) {
for (sub_chunk = 0; sub_chunk < block_size; sub_chunk++) {
register __m256d a = _mm256_broadcast_sd(
&A[row * n + chunk + sub_chunk]);
register __m256d b = _mm256_loadu_pd(
&B[chunk * n + sub_chunk * n + block]);
c0 += a * b;
}
}
_mm256_storeu_pd(&C[row * n + block], c0);
}
}
int main() {
int n = 4096; //n = 512, 1024, 2048, 4096
double *A;
double *B;
double *C;
A = (double*) malloc(n * n * sizeof(double));
B = (double*) malloc(n * n * sizeof(double));
C = (double*) malloc(n * n * sizeof(double));
srand(0);
fill(A, n); // it is a function to load A and B with random numbers
fill(B, n);
printf("n = %d\n\n", n);
clock_t t0 = clock();
blocked_mmul_avx(A, B, C, n);
clock_t t1 = clock();
printf("Execution Time = %0.3f s \n\n", (float) (t1 - t0) / CLOCKS_PER_SEC);
free(A);
free(B);
free(C);
return 0;
}
If I set size of the block to 4, the code works correctly. But if I increase it to 16, 1/4 of entries of output matrix is correct. I don't know why this is happening??
I checked and there are 16 YMM registers so I think block size can be increased to 16. But my experiment shows the otherwise.
Any idea why this is happening?