wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

sgemm_tcore: Write reference C before sim

Browse Source
This commit is contained in:
Hansung Kim
2024-06-07 10:39:11 -07:00
parent 985c5fc0dc
commit 856596cbb3
1 changed files with 12 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
tests/regression/sgemm_tcore/main.cpp
View File

@@ -108,14 +108,6 @@ int run_test(const kernel_arg_t& kernel_arg,
std::cout << "download destination buffer" << std::endl; std::cout << "download destination buffer" << std::endl;
RT_CHECK(vx_copy_from_dev(device, staging_buf.data(), kernel_arg.addr_c, buf_size)); RT_CHECK(vx_copy_from_dev(device, staging_buf.data(), kernel_arg.addr_c, buf_size));
std::ofstream ref_file("reference.c.bin", std::ios::binary | std::ios::out);
if (!ref_file) {
std::cerr << "error: failed to open reference.c.bin for writing\n";
exit(EXIT_FAILURE);
}
ref_file.write(reinterpret_cast<char *>(ref_data.data()), buf_size);
ref_file.close();
// verify result // verify result
std::cout << "verify result" << std::endl; std::cout << "verify result" << std::endl;
{ {
@@ -155,13 +147,22 @@ int main(int argc, char *argv[]) {
RT_CHECK(vx_dev_open(&device)); RT_CHECK(vx_dev_open(&device));
// FIXME: hardcoded // FIXME: hardcoded
uint32_t dim_m = 32; uint32_t dim_m = 128;
uint32_t dim_n = 32; uint32_t dim_n = 128;
uint32_t dim_k = 32; uint32_t dim_k = 128;
generate_source_matrix(dim_m, dim_n, dim_k); generate_source_matrix(dim_m, dim_n, dim_k);
generate_reference_matmul(dim_m, dim_n, dim_k); generate_reference_matmul(dim_m, dim_n, dim_k);
std::cout << "write reference output" << std::endl;
std::ofstream ref_file("reference.c.bin", std::ios::binary | std::ios::out);
if (!ref_file) {
std::cerr << "error: failed to open reference.c.bin for writing\n";
exit(EXIT_FAILURE);
}
ref_file.write(reinterpret_cast<char *>(ref_data.data()), ref_data.size() * sizeof(ref_data[0]));
ref_file.close();
uint32_t src_a_buf_size = src_a_data.size() * sizeof(src_a_data[0]); uint32_t src_a_buf_size = src_a_data.size() * sizeof(src_a_data[0]);
uint32_t src_b_buf_size = src_b_data.size() * sizeof(src_b_data[0]); uint32_t src_b_buf_size = src_b_data.size() * sizeof(src_b_data[0]);
uint32_t dst_buf_size = ref_data.size() * sizeof(src_a_data[0]); uint32_t dst_buf_size = ref_data.size() * sizeof(src_a_data[0]);