[CPU][Perf] Add fast vectorized exp impl from Arm Optimized Routines (#30068)

Signed-off-by: Ubuntu <ubuntu@ip-10-252-30-150.eu-west-1.compute.internal>
Signed-off-by: Elham Harirpoush <elham.harirpoush@arm.com>
Co-authored-by: Ubuntu <ubuntu@ip-10-252-30-150.eu-west-1.compute.internal>
This commit is contained in:
Elham
2025-12-05 13:09:20 +00:00
committed by GitHub
co-authored by Ubuntu
parent b7d85cf25c
commit 9843e332da
2 changed files with 50 additions and 13 deletions
-13
View File
@@ -1246,14 +1246,8 @@ class AttentionMainLoop {
// rescale sum and partial outputs
if (need_rescale) {
// compute rescale factor
#ifdef DEFINE_FAST_EXP
vec_op::FP32Vec16 rescale_factor_vec(rescale_factor);
rescale_factor_vec = fast_exp(rescale_factor_vec);
rescale_factor = rescale_factor_vec.get_last_elem();
#else
rescale_factor = std::exp(rescale_factor);
vec_op::FP32Vec16 rescale_factor_vec(rescale_factor);
#endif
// rescale sum
new_sum_val += rescale_factor * init_sum_val;
@@ -1889,15 +1883,8 @@ class AttentionMainLoop {
: curr_output_buffer;
float rescale_factor = final_max > curr_max ? curr_max - final_max
: final_max - curr_max;
#ifdef DEFINE_FAST_EXP
vec_op::FP32Vec16 rescale_factor_vec(rescale_factor);
rescale_factor_vec = fast_exp(rescale_factor_vec);
rescale_factor = rescale_factor_vec.get_last_elem();
#else
rescale_factor = std::exp(rescale_factor);
vec_op::FP32Vec16 rescale_factor_vec(rescale_factor);
#endif
local_sum[head_idx] = final_max > curr_max
? final_sum + rescale_factor * curr_sum