[CPU] Speedup LM Head on Arm CPUs (#55352)

Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
Co-authored-by: Li, Jiang <jiang1.li@intel.com>
This commit is contained in:
Fadi Arafeh
2026-09-11 16:21:14 +08:00
committed by GitHub
co-authored by Li, Jiang <jiang1.li@intel.com>
parent 6376c601ed
commit e3f755b732
4 changed files with 37 additions and 6 deletions
+7 -4
View File
@@ -172,7 +172,8 @@ struct hash<MatMulPrimitiveHandler::ClassMatmulCacheKey> {
size_t operator()(
const MatMulPrimitiveHandler::ClassMatmulCacheKey& val) const {
return hash<dnnl_dim_t>()(val.b_n_size) ^ hash<dnnl_dim_t>()(val.b_k_size) ^
hash<int>()(static_cast<int>(val.b_type));
hash<int>()(static_cast<int>(val.b_type)) ^
hash<int>()(static_cast<int>(val.c_type));
}
};
@@ -202,7 +203,7 @@ bool operator==(const W8A8MatMulPrimitiveHandler::MSizeCacheKey& l,
bool operator==(const MatMulPrimitiveHandler::ClassMatmulCacheKey& l,
const MatMulPrimitiveHandler::ClassMatmulCacheKey& r) {
return l.b_n_size == r.b_n_size && l.b_k_size == r.b_k_size &&
l.b_type == r.b_type;
l.b_type == r.b_type && l.c_type == r.c_type;
}
bool operator==(const MatMulPrimitiveHandler::MSizeCacheKey& l,
@@ -482,8 +483,10 @@ void MatMulPrimitiveHandler::execute(ExecArgs& args) {
dnnl::matmul MatMulPrimitiveHandler::get_matmul_cache(
const MSizeCacheKey& key) {
if (m_size_cache_.get() == nullptr) {
ClassMatmulCacheKey class_key = {
.b_n_size = b_n_size_, .b_k_size = b_k_size_, .b_type = b_type_};
ClassMatmulCacheKey class_key = {.b_n_size = b_n_size_,
.b_k_size = b_k_size_,
.b_type = b_type_,
.c_type = c_type_};
m_size_cache_ =
get_matul_class_primitive_cache(class_key, primitive_cache_size_);
}