Revert padding in winograd input transform, it looks faster than padding in advance

src/operators/kernel/central-arm-func/conv_arm_func.h

@@ -130,8 +130,8 @@ inline void WinogradConv3x3(const ConvParam<CPU> &param) {
 
   auto winograd_pad = [&](int width, int pad) {
     int output_tile = tile - kernel + 1;
-    //    int tiles = (width + pad - kernel) / output_tile + 1;
-    //    return (tiles - 1) * output_tile + tile - width;
+    // int tiles = (width + pad - kernel) / output_tile + 1;
+    // return (tiles - 1) * output_tile + tile - width;
     int pad_width = (width + 2 * pad - kernel) / output_tile * output_tile;
     return pad_width + tile - width;
   };
@@ -141,8 +141,10 @@ inline void WinogradConv3x3(const ConvParam<CPU> &param) {
   for (int i = 0; i < batch_size; ++i) {
     Tensor in_batch = input->Slice(i, i + 1);
     Tensor out_batch = output->Slice(i, i + 1);
-    int pad_bottom = winograd_pad(in_batch.dims()[2], paddings[0]);
-    int pad_right = winograd_pad(in_batch.dims()[3], paddings[1]);
+    // int pad_bottom = winograd_pad(in_batch.dims()[2], paddings[0]);
+    // int pad_right = winograd_pad(in_batch.dims()[3], paddings[1]);
+    int pad_bottom = paddings[0];
+    int pad_right = paddings[1];
     if (paddings[0] || paddings[1] || pad_bottom || pad_right) {
       framework::DDim pad_shape = in_batch.dims();
       pad_shape[2] += paddings[0] + pad_bottom;

src/operators/math/winograd/winograd_transform_f6k3.cpp

@@ -336,6 +336,23 @@ void winograd_transform_input<8, 3>(const framework::Tensor &input,
   memset(outptr, 0, output->numel() * sizeof(float));
 
   const float *inptr = input.data<float>();
+  int inter_h = (height - 2) / 6;
+  int inter_w = (width - 2) / 6;
+  int remain_h = height - (inter_h * 6);
+  int remain_w = width - (inter_w * 6);
+  framework::Tensor input_pad;
+  if (remain_h > 2 || remain_w > 2) {
+    inter_h += (remain_h > 2);
+    inter_w += (remain_w > 2);
+    height = (inter_h - 1) * 6 + 8;
+    width = (inter_w - 1) * 6 + 8;
+    framework::DDim input_shape =
+        framework::make_ddim(std::vector<int>{1, channel, height, width});
+    PadFunctor<CPU, float> pad;
+    inptr = input_pad.mutable_data<float>(input_shape);
+    pad(input, 0, height - input.dims()[2], 0, width - input.dims()[3],
+        &input_pad);
+  }
   size_t image_size = height * width;
   const float transform_matrix[8] = {5.25f, -5.f,   -4.25f, -2.5f,
                                      2.f,   -1.25f, 0.5f,   0.25f};