Skip to content

perf[gpu]: reduce register pressure in dyn dispatch #7489

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
0ax1 merged 2 commits into from
Apr 16, 2026
+6 −4
Merged

perf[gpu]: reduce register pressure in dyn dispatch #7489

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
84ec712
perf: reduce register pressure in dyn dispatch
0ax1 Apr 16, 2026
ac24aaf
Merge branch 'develop' into ad/cap-values-per-tile
0ax1 Apr 16, 2026
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
10 changes: 6 additions & 4 deletions vortex-cuda/kernels/src/dynamic_dispatch.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -279,7 +279,8 @@ __device__ void execute_output_stage(T *__restrict output,
char *__restrict smem,
uint64_t block_start,
uint32_t block_len) {
constexpr uint32_t VALUES_PER_TILE = 32 / sizeof(T);
// Cap at 4 values per thread per tile to minimise register pressure.
constexpr uint32_t VALUES_PER_TILE = (32 / sizeof(T)) < 4 ? (32 / sizeof(T)) : 4;
const uint32_t tile_size = blockDim.x * VALUES_PER_TILE;
const auto &src = stage.source;
const void *raw_input = reinterpret_cast<const void *>(stage.input_ptr);
Expand Down Expand Up @@ -472,9 +473,10 @@ dynamic_dispatch(T *__restrict output, uint64_t array_len, const uint8_t *__rest
// matters is load_element(), which dispatches on the per-op PTypeTag to
// sign-extend or zero-extend when widening a narrow source to T.
#define GENERATE_KERNEL(suffix, Type) \
extern "C" __global__ void dynamic_dispatch_##suffix(Type *__restrict output, \
uint64_t array_len, \
const uint8_t *__restrict packed_plan) { \
extern "C" __global__ void __launch_bounds__(BLOCK_SIZE, 32) \
dynamic_dispatch_##suffix(Type *__restrict output, \
uint64_t array_len, \
const uint8_t *__restrict packed_plan) { \
dynamic_dispatch<Type>(output, array_len, packed_plan); \
}

Expand Down
Loading