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

llama : fix duplicate symbols + refactor example benchmark

Browse source
This commit is contained in:
Georgi Gerganov 2023-07-06 21:18:42 +03:00
parent 8fdf86dd25
commit 9c9bdaf0b8
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
2 changed files with 15 additions and 25 deletions
Download patch file Download diff file Expand all files Collapse all files

38
examples/benchmark/benchmark-matmult.cpp
View file

@ -20,6 +20,17 @@
#pragma warning(disable: 4244 4267) // possible loss of data #pragma warning(disable: 4244 4267) // possible loss of data
#endif #endif
void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
if (plan.work_size > 0) {
buf.resize(plan.work_size);
plan.work_data = buf.data();
}
ggml_graph_compute(graph, &plan);
}
float tensor_sum_elements(const ggml_tensor * tensor) { float tensor_sum_elements(const ggml_tensor * tensor) {
float sum = 0; float sum = 0;
if (tensor->type==GGML_TYPE_F32) { if (tensor->type==GGML_TYPE_F32) {
@ -166,14 +177,7 @@ int main(int argc, char ** argv) {
std::vector<uint8_t> work_buffer; std::vector<uint8_t> work_buffer;
{ ggml_graph_compute_helper(work_buffer, &gf, benchmark_params.n_threads);
ggml_cplan pf = ggml_graph_plan(&gf, benchmark_params.n_threads);
if (pf.work_size > 0) {
work_buffer.resize(pf.work_size);
pf.work_data = work_buffer.data();
}
ggml_graph_compute(&gf, &pf);
}
TENSOR_DUMP(gf.nodes[0]); TENSOR_DUMP(gf.nodes[0]);
@ -227,14 +231,7 @@ int main(int argc, char ** argv) {
long long int start = ggml_time_us(); long long int start = ggml_time_us();
//printf("Running ggml_graph_compute\n"); //printf("Running ggml_graph_compute\n");
{ ggml_graph_compute_helper(work_buffer, &gf31, benchmark_params.n_threads);
ggml_cplan pf31 = ggml_graph_plan(&gf31, benchmark_params.n_threads);
if (pf31.work_size > 0) {
work_buffer.resize(pf31.work_size);
pf31.work_data = work_buffer.data();
}
ggml_graph_compute(&gf31, &pf31);
}
long long int stop = ggml_time_us(); long long int stop = ggml_time_us();
long long int usec = stop-start; long long int usec = stop-start;
@ -267,14 +264,7 @@ int main(int argc, char ** argv) {
} }
// Running a different graph computation to make sure we override the CPU cache lines // Running a different graph computation to make sure we override the CPU cache lines
{ ggml_graph_compute_helper(work_buffer, &gf32, benchmark_params.n_threads);
ggml_cplan pf32 = ggml_graph_plan(&gf32, benchmark_params.n_threads);
if (pf32.work_size > 0) {
work_buffer.resize(pf32.work_size);
pf32.work_data = work_buffer.data();
}
ggml_graph_compute(&gf32, &pf32);
}
} }
printf("\n"); printf("\n");
printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations)); printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));

2
llama.cpp
View file

@ -83,7 +83,7 @@ void llama_nop(struct ggml_tensor * tensor) { // don't offload by default
// ggml helpers // ggml helpers
// //
void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) { static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
struct ggml_cplan plan = ggml_graph_plan(graph, n_threads); struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
if (plan.work_size > 0) { if (plan.work_size > 0) {