Merge pull request #23112 from zihaomu:fix_x86_winograd

modules/dnn/src/layers/fast_convolution/fast_convolution.avx2.cpp

@@ -119,7 +119,7 @@ void convBlock_AVX2(int np, const float* a, const float* b, float* c, int ldc, b
 void _fx_winograd_accum_f32(const float* inwptr, const float* wptr,
                        float* outbuf, int Cg, int iblock)
 {
-    CV_Assert(_FX_WINO_IBLOCK == 6 && _FX_WINO_KBLOCK == 4);// && _FX_WINO_ATOM_F32 == 8);
+    CV_Assert(_FX_WINO_IBLOCK == 6 && _FX_WINO_KBLOCK == 4 && _FX_WINO_ATOM_F32 == 8);
     if (iblock > 3)
     {
         for (int atom_id = 0; atom_id < _FX_WINO_NATOMS_F32; atom_id++,

modules/dnn/src/layers/fast_convolution/fast_convolution.cpp

@@ -105,6 +105,12 @@ Ptr<FastConv> initFastConv(
         conv->conv_type = _FX_CONV_TYPE_GENERIC;
 #endif
 
+#if CV_TRY_AVX2
+    // Disabel Winograd when CV_TRY_AVX2 is true, but conv->useAVX2 is false.
+    if (conv->conv_type == _FX_CONV_TYPE_WINOGRAD3X3 && !conv->useAVX2)
+        conv->conv_type = _FX_CONV_TYPE_GENERIC;
+#endif
+
     Mat weightsMat = _weightsMat.getMat();
     auto wShape = shape(weightsMat);
     const size_t wstep = weightsMat.step1();
@@ -257,7 +263,7 @@ Ptr<FastConv> initFastConv(
     // we can always read MR elements starting from any valid index
     {
         int k = 0, nbias = K + VEC_ALIGN;
-        conv->biasBuf.reserve(nbias);
+        conv->biasBuf.resize(nbias);
         float* biasBufPtr = conv->biasBuf.data();
         for(; k < K; k++)
             biasBufPtr[k] = srcBias ? srcBias[k] : 0.f;

modules/dnn/src/layers/fast_convolution/winograd_3x3s1_f63.cpp

@@ -22,7 +22,7 @@ _fx_winograd_accum_f32(const float* inwptr, const float* wptr,
                        float* outbuf, int Cg, int iblock)
  {
 #if CV_NEON && CV_NEON_AARCH64
-    CV_Assert(_FX_WINO_IBLOCK == 6 && _FX_WINO_KBLOCK == 4);
+    CV_Assert(_FX_WINO_IBLOCK == 6 && _FX_WINO_KBLOCK == 4 && _FX_WINO_ATOM_F32 == 4);
     if (iblock > 3)
     {
         for (int atom_id = 0; atom_id < _FX_WINO_NATOMS_F32; atom_id++,
@@ -144,7 +144,7 @@ _fx_winograd_accum_f32(const float* inwptr, const float* wptr,
         }
     }
 #elif CV_SIMD128
-    CV_Assert(_FX_WINO_IBLOCK == 3 && _FX_WINO_KBLOCK == 4);
+    CV_Assert(_FX_WINO_IBLOCK == 3 && _FX_WINO_KBLOCK == 4 && _FX_WINO_ATOM_F32 == 4);
     for (int atom_id = 0; atom_id < _FX_WINO_NATOMS_F32; atom_id++,
             outbuf += _FX_WINO_ATOM_F32)
     {