Merge pull request #21426 from alalek:dnn_simd_unaligned_weights_fix

modules/dnn/src/layers/convolution_layer.cpp

@@ -347,7 +347,9 @@ public:
         if (!blobs.empty())
         {
             Mat wm = blobs[0].reshape(1, numOutput);
-            if( wm.step1() % VEC_ALIGN != 0 )
+            if ((wm.step1() % VEC_ALIGN != 0) ||
+                !isAligned<VEC_ALIGN * sizeof(float)>(wm.data)
+            )
             {
                 int newcols = (int)alignSize(wm.step1(), VEC_ALIGN);
                 Mat wm_buffer = Mat(numOutput, newcols, wm.type());
@@ -1299,7 +1301,6 @@ public:
                                 }
                             }
                         }
-
                         // now compute dot product of the weights
                         // and im2row-transformed part of the tensor
                     #if CV_TRY_AVX512_SKX

modules/dnn/src/layers/layers_common.simd.hpp

@@ -81,6 +81,8 @@ void fastConv( const float* weights, size_t wstep, const float* bias,
                int blockSize, int vecsize, int vecsize_aligned,
                const float* relu, bool initOutput )
 {
+    CV_Assert(isAligned<32>(weights));
+
     int outCn = outShape[1];
     size_t outPlaneSize = outShape[2]*outShape[3];
     float r0 = 1.f, r1 = 1.f, r2 = 1.f;