mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-11 08:48:45 +00:00
9cf49cd322
[CK TILE] Increase default kPerXdl for grouped convolution instances (#7465) ## Summary Increases the default `kPerXdl` used in CK Tile grouped convolution instance generation for forward, backward-data, and backward-weight operations. ### Changes in `generate_instances.py` - **Larger default `kPerXdl` for all fp16/bf16 tile sizes**: `get_k_mfma()` now returns `32` for `m/nPerXdl = 16` and `16` for `m/nPerXdl = 32`. - **Cap `kPerXdl` to `kPerBlock`**: All three parsers (`parse_fwd_instances`, `parse_bwd_weight_instances`, `parse_bwd_data_instances`) now clamp the computed value with `min(..., k_per_block)` to prevent generating invalid instances where `kPerXdl > kPerBlock`. ### Expected impact Higher `kPerXdl` increases the number of MFMA instructions issued per warp per inner-loop iteration, improving arithmetic intensity and reducing pipeline stall overhead for memory-bound shapes.