potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-21 19:33:03 +04:00
Code Packages Releases Wiki Activity

Merge pull request #29484 from velonica0:dnn-rvv-winograd-vlen-fix

Browse Source 
dnn: fix out-of-bounds access in RVV Winograd kernels when VLEN > 256
This commit is contained in:
Alexander Smorkalov
2026-07-10 11:02:30 +03:00
committed by GitHub
parent cbd227741f c86d642d19
commit bc647fe3ec
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.simd.hpp
+6 -3
View File
@@ -909,7 +909,10 @@ void impl_accum_F32(const uchar* inwptr_, const uchar* wptr_, uchar* outbuf_, in
const float* wptr = (const float*)wptr_; const float* wptr = (const float*)wptr_;
float* outbuf = (float*)outbuf_; float* outbuf = (float*)outbuf_;
CV_Assert(winoIblock == 6 && winoKblock == 4 && winoAtomF32 == 8); CV_Assert(winoIblock == 6 && winoKblock == 4 && winoAtomF32 == 8);
const size_t vl = __riscv_vsetvlmax_e32m1(); // 8 on VLEN=256 // One atom is winoAtomF32 == 8 floats, so vl must be 8 regardless of VLEN;
// vsetvlmax would return VLEN/32 (16 at VLEN=512, 32 at VLEN=1024) and make the
// atom loads below overlap their neighbours and the stores run past outbuf.
const size_t vl = __riscv_vsetvl_e32m1(8);
if (iblock > 3) { if (iblock > 3) {
for (int atom_id = 0; atom_id < winoNatomF32; atom_id++, for (int atom_id = 0; atom_id < winoNatomF32; atom_id++,
@@ -1039,7 +1042,7 @@ void impl_BtXB_8x8_F32(const float* inptr, int inpstep, uchar* outptr_, int Cg,
const int winoIblock, const int winoAtomF32) const int winoIblock, const int winoAtomF32)
{ {
float* outptr = (float*)outptr_; float* outptr = (float*)outptr_;
const size_t vl = __riscv_vsetvlmax_e32m1(); const size_t vl = __riscv_vsetvl_e32m1(8); // 8x8 tile: exactly 8 lanes, any VLEN
float tmp[8][8], tmp2[8][8]; float tmp[8][8], tmp2[8][8];
wino_bt8x8_rvv(inptr, inptr+inpstep, inptr+inpstep*2, inptr+inpstep*3, wino_bt8x8_rvv(inptr, inptr+inpstep, inptr+inpstep*2, inptr+inpstep*3,
@@ -1094,7 +1097,7 @@ void impl_AtXA_8x8_F32(const uchar* inptr_, int inpstep,
float bias, float minval, float maxval, bool ifMinMaxAct) float bias, float minval, float maxval, bool ifMinMaxAct)
{ {
const float* inptr = (const float*)inptr_; const float* inptr = (const float*)inptr_;
const size_t vl = __riscv_vsetvlmax_e32m1(); const size_t vl = __riscv_vsetvl_e32m1(8); // 8-wide winograd rows, any VLEN
const size_t vl6 = __riscv_vsetvl_e32m1(6); const size_t vl6 = __riscv_vsetvl_e32m1(6);
float tmp[8][8], tmp2[8][8]; float tmp[8][8], tmp2[8][8];