[Hardware][XPU] Register matmul and linear batch-invariant kernels for XPU (#49209)

Signed-off-by: tzielinski-habana <tomasz.zielinski@intel.com>
Signed-off-by: Tomasz Zielinski <85164140+tzielinski-habana@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Tomasz Zielinski
2026-09-02 13:31:38 +00:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Claude Opus 5 Harry Mellor
parent 7894394b07
commit ffe3bb3c72
7 changed files with 272 additions and 33 deletions
@@ -327,3 +327,53 @@ def _get_matmul_config(
"num_warps": m_config.num_warps,
"num_stages": m_config.num_stages,
}
def _get_descriptor_matmul_config(
M: int, N: int, K: int, dtype: torch.dtype
) -> dict[str, int]:
"""Select a config for the XPU descriptor matmul from shape and dtype.
The default 128×128×64 config is optimized for large square-ish GEMMs but
wastes resources on skinny shapes (small M during decode, or very tall/thin
matrices during prefill). Shape-dependent tuning closes the gap with oneDNN.
Only BLOCK_SIZE_M/N and the launch parameters vary with M; dtype alone
decides BLOCK_SIZE_K, so the K-reduction order never depends on batch size.
"""
# fp32 uses smaller BLOCK_SIZE_K due to register pressure
block_k = 32 if dtype == torch.float32 else 64
if M <= 16:
# Decode: M=1-16. Tiny M means most of a 128-row tile is wasted.
# Use small M-block, wide N-block to maximize useful work per tile.
return {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": block_k,
"GROUP_SIZE_M": 1,
"num_stages": 4,
"num_warps": 8,
}
elif M <= 64:
# Small batch decode or very short prefill.
return {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": block_k,
"GROUP_SIZE_M": 4,
"num_stages": 4,
"num_warps": 8,
}
else:
# Medium and large prefill (M > 64). 64×128 tiles provide the best
# balance of register pressure vs parallelism on Intel XPU.
# M=2048, N=4096 → 32×32 = 1024 tiles, well above ~160 compute units.
return {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": block_k,
"GROUP_SIZE_M": 8,
"num_stages": 3,
"num_warps": 8,
}