[CPU] Added faster exp routine for lower precision data types. (#38112)

Signed-off-by: Anna Mayne <anna.mayne@arm.com>
Co-authored-by: Fadi Arafeh <fadi.arafeh@arm.com>
Co-authored-by: Li, Jiang <jiang1.li@intel.com>
This commit is contained in:
almayne
2026-04-23 13:14:44 +00:00
committed by GitHub
co-authored by Fadi Arafeh Li, Jiang <jiang1.li@intel.com>
parent 2196bac135
commit 2f314bc5e6
2 changed files with 76 additions and 6 deletions
+28 -4
View File
@@ -1152,7 +1152,11 @@ class AttentionMainLoop {
bool use_sink) {
#ifdef DEFINE_FAST_EXP
DEFINE_FAST_EXP
bool constexpr IsReducedPrecision =
std::is_same_v<query_t, c10::BFloat16> ||
std::is_same_v<query_t, c10::Half>;
#endif
using prob_buffer_vec_t = typename VecTypeTrait<prob_buffer_t>::vec_t;
static_assert(sizeof(prob_buffer_t) <= sizeof(logits_buffer_t));
@@ -1201,8 +1205,17 @@ class AttentionMainLoop {
vec = vec - max_vec;
// compute exp
#ifdef DEFINE_FAST_EXP
vec = fast_exp(vec);
#if defined(DEFINE_FAST_EXP)
#ifdef __aarch64__
if constexpr (IsReducedPrecision) {
vec = fast_exp_f16(vec);
} else
#endif
{
vec = fast_exp(vec);
}
prob_buffer_vec_t output_vec(vec);
output_vec.save(curr_prob_buffer_iter);
#else
@@ -1258,7 +1271,11 @@ class AttentionMainLoop {
int32_t kv_tile_token_num, float softcap_scale) {
#ifdef DEFINE_FAST_EXP
DEFINE_FAST_EXP
bool constexpr IsReducedPrecision =
std::is_same_v<query_t, c10::BFloat16> ||
std::is_same_v<query_t, c10::Half>;
#endif
float inv_softcap_scale = 1.0 / softcap_scale;
vec_op::FP32Vec16 softcap_scale_vec(softcap_scale);
vec_op::FP32Vec16 inv_softcap_scale_vec(inv_softcap_scale);
@@ -1272,8 +1289,15 @@ class AttentionMainLoop {
vec_op::FP32Vec16 vec(curr_logits_buffer_iter);
vec = vec * inv_softcap_scale_vec;
#ifdef DEFINE_FAST_EXP
vec = fast_exp(vec);
#if defined(DEFINE_FAST_EXP)
#ifdef __aarch64__
if constexpr (IsReducedPrecision) {
vec = fast_exp_f16(vec);
} else
#endif
{
vec = fast_exp(vec);
}
vec_op::FP32Vec16 inv_vec = ones_vec / vec;
vec = (vec - inv_vec) / (vec + inv_vec);
#else