Skip to content

OPT: Optimize indexing using dynamic thread block sizes. #3111

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
9prady9 merged 1 commit into from
Mar 23, 2021
Merged

OPT: Optimize indexing using dynamic thread block sizes. #3111

Show file tree
Hide file tree
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
14 changes: 9 additions & 5 deletions src/backend/cuda/kernel/index.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,13 +21,17 @@ namespace kernel {

template<typename T>
void index(Param<T> out, CParam<T> in, const IndexKernelParam& p) {
constexpr int THREADS_X = 32;
constexpr int THREADS_Y = 8;

auto index = common::getKernel("cuda::index", {index_cuh_src},
{TemplateTypename<T>()});

const dim3 threads(THREADS_X, THREADS_Y);
dim3 threads;
switch (out.dims[1]) {
case 1: threads.y = 1; break;
case 2: threads.y = 2; break;
case 3:
case 4: threads.y = 4; break;
default: threads.y = 8; break;
}
threads.x = static_cast<unsigned>(256.f / threads.y);

int blks_x = divup(out.dims[0], threads.x);
int blks_y = divup(out.dims[1], threads.y);
Expand Down
25 changes: 16 additions & 9 deletions src/backend/opencl/kernel/index.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,23 +31,30 @@ typedef struct {
template<typename T>
void index(Param out, const Param in, const IndexKernelParam_t& p,
cl::Buffer* bPtr[4]) {
constexpr int THREADS_X = 32;
constexpr int THREADS_Y = 8;

std::vector<std::string> options = {
DefineKeyValue(T, dtype_traits<T>::getName()),
};
options.emplace_back(getTypeBuildDefinition<T>());

auto index = common::getKernel("indexKernel", {index_cl_src},
auto index = common::getKernel("indexKernel", {index_cl_src},
{TemplateTypename<T>()}, options);
cl::NDRange local(THREADS_X, THREADS_Y);
int threads_x = 256;
int threads_y = 1;
cl::NDRange local(threads_x, threads_y);
switch (out.info.dims[1]) {
case 1: threads_y = 1; break;
case 2: threads_y = 2; break;
case 3:
case 4: threads_y = 4; break;
default: threads_y = 8; break;
}
threads_x = static_cast<unsigned>(256.f / threads_y);

int blk_x = divup(out.info.dims[0], THREADS_X);
int blk_y = divup(out.info.dims[1], THREADS_Y);
int blk_x = divup(out.info.dims[0], local[0]);
int blk_y = divup(out.info.dims[1], local[1]);

cl::NDRange global(blk_x * out.info.dims[2] * THREADS_X,
blk_y * out.info.dims[3] * THREADS_Y);
cl::NDRange global(blk_x * out.info.dims[2] * local[0],
blk_y * out.info.dims[3] * local[1]);

index(cl::EnqueueArgs(getQueue(), global, local), *out.data, out.info,
*in.data, in.info, p, *bPtr[0], *bPtr[1], *bPtr[2], *bPtr[3], blk_x,
Expand Down