vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : move CPU backend to a separate file (#10144)

Browse source
This commit is contained in:
Diego Devesa 2024-11-03 19:34:08 +01:00 • committed by GitHub
parent 08828a6d7d
commit 9f40989351
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
32 changed files with 14747 additions and 19345 deletions
Download patch file Download diff file Expand all files Collapse all files

6
tests/test-quantize-perf.cpp
View file

@ -1,6 +1,7 @@
// Benchmark quantization specific functions on synthetic data
#include "ggml.h"
#include "ggml-cpu.h"
#undef NDEBUG
#include <algorithm>
@ -271,6 +272,7 @@ int main(int argc, char * argv[]) {
for (int i = 0; i < GGML_TYPE_COUNT; i++) {
ggml_type type = (ggml_type) i;
const auto * qfns = ggml_get_type_traits(type);
const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
if (!params.include_types.empty() && ggml_type_name(type) && std::find(params.include_types.begin(), params.include_types.end(), ggml_type_name(type)) == params.include_types.end()) {
continue;
}
@ -328,7 +330,7 @@ int main(int argc, char * argv[]) {
for (size_t size : params.test_sizes) {
printf(" %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
auto quantize_fn = [&](void) -> float {
const auto * vdot = ggml_get_type_traits(qfns->vec_dot_type);
const auto * vdot = ggml_get_type_traits(qfns_cpu->vec_dot_type);
vdot->from_float(test_data1, test_q1, size);
return test_q1[0];
};
@ -346,7 +348,7 @@ int main(int argc, char * argv[]) {
printf(" %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
auto quantize_fn = [&](void) -> float {
float result;
qfns->vec_dot(size, &result, 0, test_q1, 0, test_q2, 0, 1);
qfns_cpu->vec_dot(size, &result, 0, test_q1, 0, test_q2, 0, 1);
return result;
};
size_t quantized_size = ggml_row_size(type, size);