gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[codegen] some more optimizations

Browse Source
This commit is contained in:
Philippe Tillet
2019-07-18 19:39:40 -07:00
parent 71594da66f
commit 5215fb0424
12 changed files with 108 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
examples/cpp/dot.cpp
View File

@@ -10,37 +10,40 @@
int main() {
bool AT = true;
bool BT = false;
typedef float T;
std::string ty = "fp16";
size_t dt_nbytes = sizeof(T);
// initialize default compute device
auto context = triton::driver::backend::contexts::get_default();
// matrix multiplication parameters
int32_t M = 64, N = 128, K = 128;
std::vector<float> hc(M*N);
std::vector<float> rc(M*N);
std::vector<float> ha(M*K);
std::vector<float> hb(K*N);
int32_t M = 65536, N = 2048, K = 2048;
std::vector<T> hc(M*N);
std::vector<T> rc(M*N);
std::vector<T> ha(M*K);
std::vector<T> hb(K*N);
srand(0);
for(size_t i = 0; i < ha.size(); i++)
ha[i] = (float)rand()/RAND_MAX;
ha[i] = (T)rand()/RAND_MAX;
for(size_t i = 0; i < hb.size(); i++)
hb[i] = (float)rand()/RAND_MAX;
hb[i] = (T)rand()/RAND_MAX;
for(size_t i = 0; i < hc.size(); i++)
hc[i] = 0;
triton::driver::buffer* dc = triton::driver::buffer::create(context, hc.size()*4);
triton::driver::buffer* da = triton::driver::buffer::create(context, ha.size()*4);
triton::driver::buffer* db = triton::driver::buffer::create(context, hb.size()*4);
triton::driver::buffer* dc = triton::driver::buffer::create(context, hc.size()*dt_nbytes);
triton::driver::buffer* da = triton::driver::buffer::create(context, ha.size()*dt_nbytes);
triton::driver::buffer* db = triton::driver::buffer::create(context, hb.size()*dt_nbytes);
triton::driver::stream* stream = triton::driver::stream::create(context);
stream->write(da, true, 0, ha);
stream->write(db, true, 0, hb);
stream->write(dc, true, 0, hc);
stream->synchronize();
triton::dnn::gemm gemm(M, N, K, AT, BT, "fp32", "fp32", 4, 4);
gemm.enqueue(stream, {da, db, dc}, false);
stream->read(dc, true, 0, hc);
gemm.cpu_ref<float>(rc, ha, hb);
for(size_t i = 0; i < M*N; i++)
if(std::isnan(hc[i]) || std::abs(hc[i] - rc[i])/std::max(hc[i], rc[i]) > 1e-4){
std::cout << i << " " << hc[i] << " " << rc[i] << std::endl;
exit(EXIT_FAILURE);
}
std::cout << "Pass!" << std::endl;
triton::dnn::gemm gemm(M, N, K, AT, BT, ty, ty, 4, 4);
gemm.enqueue(stream, {da, db, dc}, true);
// stream->read(dc, true, 0, hc);
// gemm.cpu_ref<T>(rc, ha, hb);
// for(size_t i = 0; i < M*N; i++)
// if(std::isnan(hc[i]) || std::abs(hc[i] - rc[i])/std::max(hc[i], rc[i]) > 1e-4){
// std::cout << i << " " << hc[i] << " " << rc[i] << std::endl;
// exit(EXIT_FAILURE);
// }
// std::cout << "Pass!" << std::endl;
}