Skip to content
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

Fix maximum number of threads per SM #1405

Merged
merged 2 commits into from
Sep 22, 2023
Merged

Fix maximum number of threads per SM #1405

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c552a1a
In the mdw local operator code add limits for the max threads per SM.
hummingtree Sep 20, 2023
73723a9
Apply clang-format.
hummingtree Sep 20, 2023
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
23 changes: 23 additions & 0 deletions include/targets/cuda/device.in.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,7 @@ namespace quda

/**
@brief A constexpr function to returns the maximum dyanmic shared memory per block.
See https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#feature-availability
*/
constexpr int maximum_dynamic_shared_memory()
{
Expand Down Expand Up @@ -37,5 +38,27 @@ namespace quda
#endif
}
}

/**
@brief A constexpr function to return the maximum number of resident threads per SM.
See https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#feature-availability
*/
constexpr unsigned int maximum_resident_threads()
{
#if (__COMPUTE_CAPABILITY__ < 750)
return 2048;
#elif (__COMPUTE_CAPABILITY__ == 750)
return 1024;
#elif (__COMPUTE_CAPABILITY__ == 800)
return 2048;
#elif ((__COMPUTE_CAPABILITY__ > 800) && (__COMPUTE_CAPABILITY__ < 900))
return 1536;
#elif (__COMPUTE_CAPABILITY__ == 900)
return 2048;
#else
return 0;
#endif
}

} // namespace device
} // namespace quda
24 changes: 20 additions & 4 deletions lib/dslash_mdw_fused_impl.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,10 @@
#include <kernels/dslash_mdw_fused.cuh>
#include <dslash_mdw_fused.hpp>

#ifdef QUDA_MMA_AVAILABLE
#include <device.hpp>
#endif

namespace quda
{

Expand Down Expand Up @@ -87,7 +91,11 @@ namespace quda

int blockStep() const { return 16; }
int blockMin() const { return 16; }
unsigned int maxBlockSize(const TuneParam &) const { return 32; }
unsigned int maxBlockSize(const TuneParam &param) const
{
unsigned int m = std::min(device::max_threads_per_block() / (param.block.y * param.block.z), 32u);
return std::min(m, device::maximum_resident_threads() / (param.block.y * param.block.z * param.aux.y));
}

int gridStep() const { return device::processor_count(); }
unsigned int maxGridSize() const { return (volume_4d_cb_active + blockMin() - 1) / blockMin(); }
Expand Down Expand Up @@ -116,7 +124,9 @@ namespace quda
param.aux.x++;
aux_advanced = true;
} else {
if (param.aux.y < 3) { // second see if aux.y
if (param.aux.y < 3
&& (param.aux.y + 1) * param.block.x * param.block.y
<= device::maximum_resident_threads()) { // second see if aux.y
param.aux.y++;
aux_advanced = true;
param.aux.x = 0;
Expand Down Expand Up @@ -182,8 +192,14 @@ namespace quda
template <int block_dim_x, int min_blocks, bool reload, MdwfFusedDslashType type>
void apply(const TuneParam &tp, const qudaStream_t &stream)
{
launch_cuda<FusedMobiusDslash>(tp, stream, Arg<type, Ls, block_dim_x, min_blocks, reload>
(out, in, U, y, x, m_f, m_5, b_5, c_5, parity, shift, halo_shift));
if constexpr (block_dim_x * Ls * min_blocks <= device::maximum_resident_threads()) {
launch_cuda<FusedMobiusDslash>(tp, stream,
Arg<type, Ls, block_dim_x, min_blocks, reload>(out, in, U, y, x, m_f, m_5, b_5,
c_5, parity, shift, halo_shift));
} else {
errorQuda("Maximum number of resident_threads reached: %d * %d * %d > %u\n", block_dim_x, Ls, min_blocks,
device::maximum_resident_threads());
}
}

template <int block_dim_x, bool reload, MdwfFusedDslashType type>
Expand Down