Skip to content

[ROCm] Unroll loads in global_reduce #2554

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
merged 1 commit into from
Aug 22, 2025
Merged

[ROCm] Unroll loads in global_reduce #2554

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
18 changes: 18 additions & 0 deletions aten/src/ATen/native/cuda/Reduce.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -822,6 +822,23 @@ struct ReduceOp {
} else {
index_t input_offset = threadIdx.y;
index_t step = blockDim.y;
#ifdef USE_ROCM // Prefetch loads to better hide their latency
#define PRFCH 4
for (; input_offset < config.ctas_per_output; input_offset += step*PRFCH) {
arg_vec_t next[PRFCH];
#pragma unroll
for (int u = 0; (u < PRFCH) && (input_offset + u*step < config.ctas_per_output); u++) {
index_t idx = config.staging_memory_offset(input_offset + u*step);
next[u] = reduce_buffer[idx];
}
for (int u = 0; (u < PRFCH) && (input_offset + u*step < config.ctas_per_output); u++) {
#pragma unroll
for (int i = 0; i < output_vec_size; i++) {
value[i] = ops.combine(value[i], next[u][i]);
}
}
}
#else
for (; input_offset < config.ctas_per_output; input_offset += step) {
index_t idx = config.staging_memory_offset(input_offset);
arg_vec_t next = reduce_buffer[idx];
Expand All @@ -830,6 +847,7 @@ struct ReduceOp {
value[i] = ops.combine(value[i], next[i]);
}
}
#endif
}
value = block_y_reduce<output_vec_size>(value, shared_memory);
if (config.should_block_x_reduce()) {
Expand Down