Skip to content

examples : fix benchmark-matmult #1554

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Sep 20, 2023
Merged

examples : fix benchmark-matmult #1554

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 15 additions & 13 deletions examples/benchmark/benchmark-matmult.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,11 +17,11 @@
#include <algorithm>

float tensor_sum_elements(const ggml_tensor * tensor) {
float sum = 0;
if (tensor->type==GGML_TYPE_F32) {
double sum = 0;
if (tensor->type == GGML_TYPE_F32) {
for (int j = 0; j < tensor->ne[1]; j++) {
for (int k = 0; k < tensor->ne[0]; k++) {
sum += ((float *) tensor->data)[j*tensor->ne[0]+k];
sum += ((float *) tensor->data)[j*tensor->ne[0] + k];
}
}
}
Expand Down Expand Up @@ -110,12 +110,15 @@ int main(int argc, char ** argv) {

//printf("Memsize required = %i\n", sizex*sizex);

// TODO: perform the bench for all types or for a user specified type
const ggml_type qtype = GGML_TYPE_Q4_1;

size_t ctx_size = 0;
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
ctx_size += sizex*sizey*ggml_type_sizef(qtype);
ctx_size += sizex*sizey*ggml_type_sizef(qtype);
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
ctx_size += 1024*1024*16;
Expand Down Expand Up @@ -148,7 +151,7 @@ int main(int argc, char ** argv) {
struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
ggml_set_f32(m2, 2.0f);

printf("\n------ Test 1 - Matrix Mult via F32 code ------------------------------------------------------------------------------\n");
printf("\n------ Test 1 - Matrix Mult via F32 code\n");
// printf("Creating new tensor m11xm2\n");
struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);

Expand All @@ -165,17 +168,16 @@ int main(int argc, char ** argv) {

TENSOR_DUMP(gf.nodes[0]);

printf("\n------ Test 2 - Matrix Mult via Q4_0 code ------------------------------------------------------------------------------\n");
printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));

int32_t nelements = sizex*sizey;
int32_t ne[2] = { sizex, sizey };

std::vector<int64_t> hist_cur(1 << 4, 0);

// Set up a the benchmark matrices
// printf("Creating new tensor q11 & Running quantize\n");
struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
ggml_quantize_q4_0((const float *) m11->data, q11->data, nelements, ne[0], hist_cur.data());
struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements, hist_cur.data());

// Set up a the compute graph
// printf("Creating new tensor q31\n");
Expand All @@ -187,8 +189,8 @@ int main(int argc, char ** argv) {

// Set up a second graph computation to make sure we override the CPU cache lines
// printf("Creating new tensor q12 & Running quantize\n");
struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey);
ggml_quantize_q4_0((const float *) m12->data, q12->data, nelements, ne[0], hist_cur.data());
struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements, hist_cur.data());

// printf("Creating new tensor q32\n");
struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
Expand All @@ -206,7 +208,7 @@ int main(int argc, char ** argv) {
printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);


// Let's use the F32 result from above as a reference for the q4_0 multiplication
// Let's use the F32 result from above as a reference for the quantized multiplication
float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]);

printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
Expand Down