potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-31 00:03:03 +04:00
Code Packages Releases Wiki Activity

Merge pull request #29601 from Teddy-Yangjiale:rvv-fast-attn

Browse Source 
dnn: enable the existing SIMD attention-softmax on scalable RVV
This commit is contained in:
Alexander Smorkalov
2026-07-26 12:06:45 +03:00
committed by GitHub
parent 18b39f1f34 73c580723c
commit 4ba2752f36
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/cpu_kernels/fast_attn.cpp
+3 -3
View File
@@ -102,7 +102,7 @@ void run_fused_softmax(
const int tmax = is_causal ? std::min(past_seq_len + tq + 1, seq_len_kv) : seq_len_kv;
float maxVal = -FLT_MAX;
int tk = 0;
#if CV_SIMD
#if (CV_SIMD || CV_SIMD_SCALABLE)
const int w = VTraits<v_float32>::vlanes();
v_float32 v_max_val = vx_setall_f32(maxVal);
v_float32 v_softcap = vx_setall_f32(softcap);
@@ -173,7 +173,7 @@ void run_fused_softmax(
if (do_softmax) {
float sum = 0.f;
tk = 0;
#if CV_SIMD
#if (CV_SIMD || CV_SIMD_SCALABLE)
v_float32 v_sum = vx_setzero_f32();
v_float32 v_max_val_shift = vx_setall_f32(maxVal);
for (; tk <= seq_len_kv - w; tk += w) {
@@ -192,7 +192,7 @@ void run_fused_softmax(
float inv_sum = 1.f / sum;
tk = 0;
#if CV_SIMD
#if (CV_SIMD || CV_SIMD_SCALABLE)
v_float32 v_inv_sum = vx_setall_f32(inv_sum);
for (; tk <= seq_len_kv - w; tk += w) {
v_float32 v_val = vx_load(&data[offset + tk]);