[perf][cpu] Accelerate BF16 GELU with LUT impl on Arm CPUs (#37469)

Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Fadi Arafeh
2026-04-15 22:26:17 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 18013df6ae
commit 445b7093fd
9 changed files with 240 additions and 2 deletions
+12
View File
@@ -85,6 +85,9 @@ at::Tensor int4_scaled_mm_cpu(at::Tensor& x, at::Tensor& w, at::Tensor& w_zeros,
at::Tensor& w_scales,
std::optional<at::Tensor> bias);
void activation_lut_bf16(torch::Tensor& out, torch::Tensor& input,
const std::string& activation);
torch::Tensor get_scheduler_metadata(
const int64_t num_req, const int64_t num_heads_q,
const int64_t num_heads_kv, const int64_t head_dim,
@@ -231,6 +234,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
ops.def("gelu_quick(Tensor! out, Tensor input) -> ()");
ops.impl("gelu_quick", torch::kCPU, &gelu_quick);
#if (defined(__aarch64__) && !defined(__APPLE__))
ops.def(
"activation_lut_bf16(Tensor! out, Tensor input, str activation)"
" -> ()");
ops.impl("activation_lut_bf16", torch::kCPU, &activation_lut_bf16);
#endif // (defined(__aarch64__) && !defined(__APPLE__))
// Layernorm
// Apply Root Mean Square (RMS) Normalization to the input tensor.
ops.def(