Skip to content

Commit 4570bdb

Browse files
cirosantillicirosantilli
committed
Matmul make minimal unit tests work again
1 parent db069b1 commit 4570bdb

3 files changed

Lines changed: 68 additions & 55 deletions

File tree

‎opencl/configure‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1 @@
1+
sudo apt-get install libblas-dev libatlas-base-dev libopenblas-dev nvidia-opencl-dev

‎opencl/matmul.c‎

Lines changed: 65 additions & 55 deletions
Original file line numberDiff line numberDiff line change
@@ -57,6 +57,18 @@ typedef struct {
5757
} Cache;
5858
typedef void (*MatMul)(const F *A, const F *B, F *C, size_t n, Cache *cache);
5959

60+
void cl_buf_init(Cache *cache, size_t mat_sizeof) {
61+
cache->buf_a = clCreateBuffer(cache->common.context, CL_MEM_READ_ONLY, mat_sizeof, NULL, NULL);
62+
cache->buf_b = clCreateBuffer(cache->common.context, CL_MEM_READ_ONLY, mat_sizeof, NULL, NULL);
63+
cache->buf_c = clCreateBuffer(cache->common.context, CL_MEM_WRITE_ONLY, mat_sizeof, NULL, NULL);
64+
}
65+
66+
void cl_buf_deinit(Cache *cache) {
67+
clReleaseMemObject(cache->buf_a);
68+
clReleaseMemObject(cache->buf_b);
69+
clReleaseMemObject(cache->buf_c);
70+
}
71+
6072
/* No, this was not created for debugging, my code is flawless from the first try. */
6173
void mat_print(const F *A, size_t n) {
6274
size_t i, j;
@@ -574,56 +586,62 @@ int main(int argc, char **argv) {
574586
max_runtime = 1.0;
575587
}
576588

589+
common_init(&(cache.common), NULL);
590+
577591
/* Unit test 2x2. */
578592
{
579-
/*const F A[] = {*/
580-
/*1.0, 2.0,*/
581-
/*3.0, 4.0*/
582-
/*};*/
583-
/*const F B[] = {*/
584-
/*5.0, 6.0,*/
585-
/*7.0, 8.0*/
586-
/*};*/
587-
/*enum N { n = 2 };*/
588-
/*F C[n*n];*/
589-
/*const F C_ref[] = {*/
590-
/*19.0, 22.0,*/
591-
/*43.0, 50.0*/
592-
/*};*/
593-
/*for (f = 0; f < sizeof(mat_mul_funcs)/sizeof(mat_mul_funcs[0]); ++f) {*/
594-
/*mat_zero(C, n);*/
595-
/*mat_mul_funcs[f](A, B, C, n, &cache);*/
596-
/*mat_assert_eq(C, C_ref, n);*/
597-
/*}*/
593+
const F A[] = {
594+
1.0, 2.0,
595+
3.0, 4.0
596+
};
597+
const F B[] = {
598+
5.0, 6.0,
599+
7.0, 8.0
600+
};
601+
enum N { n = 2 };
602+
F C[n*n];
603+
const F C_ref[] = {
604+
19.0, 22.0,
605+
43.0, 50.0
606+
};
607+
cl_buf_init(&cache, n * n * sizeof(F));
608+
for (f = 0; f < sizeof(mat_mul_funcs)/sizeof(mat_mul_funcs[0]); ++f) {
609+
mat_zero(C, n);
610+
mat_mul_funcs[f](A, B, C, n, &cache);
611+
mat_assert_eq(C, C_ref, n);
612+
}
613+
cl_buf_deinit(&cache);
598614
}
599615

600616
/* Unit test 4x4. */
601617
{
602-
/*const F A[] = {*/
603-
/*1.0, 2.0, 3.0, 4.0,*/
604-
/*5.0, 6.0, 7.0, 8.0,*/
605-
/*9.0, 10.0, 11.0, 12.0,*/
606-
/*13.0, 14.0, 15.0, 16.0,*/
607-
/*};*/
608-
/*const F B[] = {*/
609-
/*17.0, 18.0, 19.0, 20.0,*/
610-
/*21.0, 22.0, 23.0, 24.0,*/
611-
/*25.0, 26.0, 27.0, 28.0,*/
612-
/*29.0, 30.0, 31.0, 32.0,*/
613-
/*};*/
614-
/*const F C_ref[] = {*/
615-
/*250.0, 260.0, 270.0, 280.0,*/
616-
/*618.0, 644.0, 670.0, 696.0,*/
617-
/*986.0, 1028.0, 1070.0, 1112.0,*/
618-
/*1354.0, 1412.0, 1470.0, 1528.0,*/
619-
/*};*/
620-
/*enum N { n = 4 };*/
621-
/*F C[n*n];*/
622-
/*for (f = 0; f < NELEMS(mat_mul_funcs); ++f) {*/
623-
/*mat_zero(C, n);*/
624-
/*mat_mul_funcs[f](A, B, C, n, &cache);*/
625-
/*mat_assert_eq(C, C_ref, n);*/
626-
/*}*/
618+
const F A[] = {
619+
1.0, 2.0, 3.0, 4.0,
620+
5.0, 6.0, 7.0, 8.0,
621+
9.0, 10.0, 11.0, 12.0,
622+
13.0, 14.0, 15.0, 16.0,
623+
};
624+
const F B[] = {
625+
17.0, 18.0, 19.0, 20.0,
626+
21.0, 22.0, 23.0, 24.0,
627+
25.0, 26.0, 27.0, 28.0,
628+
29.0, 30.0, 31.0, 32.0,
629+
};
630+
const F C_ref[] = {
631+
250.0, 260.0, 270.0, 280.0,
632+
618.0, 644.0, 670.0, 696.0,
633+
986.0, 1028.0, 1070.0, 1112.0,
634+
1354.0, 1412.0, 1470.0, 1528.0,
635+
};
636+
enum N { n = 4 };
637+
F C[n*n];
638+
cl_buf_init(&cache, n * n * sizeof(F));
639+
for (f = 0; f < NELEMS(mat_mul_funcs); ++f) {
640+
mat_zero(C, n);
641+
mat_mul_funcs[f](A, B, C, n, &cache);
642+
mat_assert_eq(C, C_ref, n);
643+
}
644+
cl_buf_deinit(&cache);
627645
}
628646

629647
/* Benchmarks. */
@@ -634,7 +652,6 @@ int main(int argc, char **argv) {
634652
size_t n = 2;
635653

636654
puts("#matmul");
637-
common_init(&(cache.common), NULL);
638655
done = 0;
639656
while(1) {
640657
printf("%zu ", (size_t)log2(n));
@@ -652,11 +669,7 @@ int main(int argc, char **argv) {
652669
mat_rand(A, n);
653670
mat_rand(B, n);
654671

655-
/* CL setup. */
656-
cache.buf_a = clCreateBuffer(cache.common.context, CL_MEM_READ_ONLY, mat_sizeof, NULL, NULL);
657-
cache.buf_b = clCreateBuffer(cache.common.context, CL_MEM_READ_ONLY, mat_sizeof, NULL, NULL);
658-
cache.buf_c = clCreateBuffer(cache.common.context, CL_MEM_WRITE_ONLY, mat_sizeof, NULL, NULL);
659-
672+
cl_buf_init(&cache, mat_sizeof);
660673
first = 1;
661674
for (f = 0; f < NELEMS(mat_mul_funcs); ++f) {
662675
if (func_done[f]) {
@@ -687,16 +700,13 @@ int main(int argc, char **argv) {
687700
break;
688701
n *= 2;
689702

690-
/* CPU teardown. */
703+
/* CPU deinit. */
691704
free(A);
692705
free(B);
693706
free(C);
694707
free(C_ref);
695708

696-
/* GPU teardown. */
697-
clReleaseMemObject(cache.buf_a);
698-
clReleaseMemObject(cache.buf_b);
699-
clReleaseMemObject(cache.buf_c);
709+
cl_buf_deinit(&cache);
700710
}
701711
common_deinit(&cache.common);
702712
}

‎sdl/common.h‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -24,10 +24,12 @@ double common_get_secs(void) {
2424
const double COMMON_FPS_GRANULARITY_S = 0.5;
2525
double common_fps_last_time_s;
2626
unsigned int common_fps_nframes;
27+
2728
void common_fps_init() {
2829
common_fps_nframes = 0;
2930
common_fps_last_time_s = common_get_secs();
3031
}
32+
3133
void common_fps_update_and_print() {
3234
double dt, current_time_s;
3335
current_time_s = common_get_secs();

0 commit comments

Comments
 (0)