Merge pull request #4887 from chengduoZH/fix_im2col_kocf_for_sequence

Add up, down, left and right padding for im2col.

paddle/operators/conv2d_op.h

@@ -114,7 +114,7 @@ class GemmConv2DKernel : public framework::OpKernel<T> {
         // im2col
         Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
         im2col(context.device_context(), in_slice, col, strides[0], strides[1],
-               paddings[0], paddings[1]);
+               paddings[0], paddings[0], paddings[1], paddings[1]);
 
         // gemm
         Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
@@ -213,7 +213,8 @@ class GemmConvGrad2DKernel : public framework::OpKernel<T> {
           Tensor in_grad_slice =
               in_grad_batch.Slice(g * in_step, (g + 1) * in_step);
           col2im(context.device_context(), in_grad_slice, col, strides[0],
-                 strides[1], paddings[0], paddings[1]);
+                 strides[1], paddings[0], paddings[0], paddings[1],
+                 paddings[1]);
         }
       }
     }
@@ -235,7 +236,8 @@ class GemmConvGrad2DKernel : public framework::OpKernel<T> {
               out_grad_batch.Slice(g * out_step, (g + 1) * out_step);
           Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
           im2col(context.device_context(), in_slice, col, strides[0],
-                 strides[1], paddings[0], paddings[1]);
+                 strides[1], paddings[0], paddings[0], paddings[1],
+                 paddings[1]);
 
           // gemm
           Tensor filter_grad_slice =

paddle/operators/math/im2col.cc

@@ -29,8 +29,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
  public:
   void operator()(const platform::DeviceContext& context,
                   const framework::Tensor& im, framework::Tensor& col,
-                  int stride_height, int stride_width, int padding_height,
-                  int padding_width) {
+                  int stride_height, int stride_width, int padding_up,
+                  int padding_down, int padding_left, int padding_right) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
 
@@ -41,6 +41,22 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
     int filter_width = col.dims()[2];
     int output_height = col.dims()[3];
     int output_width = col.dims()[4];
+
+    PADDLE_ENFORCE_EQ(
+        (input_height + padding_up + padding_down - filter_height) /
+                stride_height +
+            1,
+        output_height,
+        "Output_height and padding(padding_up, padding_down) are "
+        "inconsistent.");
+    PADDLE_ENFORCE_EQ(
+        (input_width + padding_left + padding_right - filter_width) /
+                stride_width +
+            1,
+        output_width,
+        "output_width and padding(padding_left, padding_right) are "
+        "inconsistent.");
+
     int channels_col = input_channels * filter_height * filter_width;
 
     const T* im_data = im.data<T>();
@@ -52,16 +68,14 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
       int c_im = c / filter_width / filter_height;
       for (int h = 0; h < output_height; ++h) {
         for (int w = 0; w < output_width; ++w) {
-          int im_row_idx = h * stride_height + h_offset;
-          int im_col_idx = w * stride_width + w_offset;
-          if ((im_row_idx - padding_height) < 0 ||
-              (im_row_idx - padding_height) >= input_height ||
-              (im_col_idx - padding_width) < 0 ||
-              (im_col_idx - padding_width) >= input_width) {
+          int im_row_idx = h * stride_height + h_offset - padding_up;
+          int im_col_idx = w * stride_width + w_offset - padding_left;
+
+          if (im_row_idx < 0 || im_row_idx >= input_height || im_col_idx < 0 ||
+              im_col_idx >= input_width) {
             col_data[(c * output_height + h) * output_width + w] = T(0);
           } else {
-            im_row_idx += c_im * input_height - padding_height;
-            im_col_idx -= padding_width;
+            im_row_idx += c_im * input_height;
             col_data[(c * output_height + h) * output_width + w] =
                 im_data[im_row_idx * input_width + im_col_idx];
           }
@@ -82,7 +96,8 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
  public:
   void operator()(const platform::DeviceContext& context, framework::Tensor& im,
                   const framework::Tensor& col, int stride_height,
-                  int stride_width, int padding_height, int padding_width) {
+                  int stride_width, int padding_up, int padding_down,
+                  int padding_left, int padding_right) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
     int input_channels = im.dims()[0];
@@ -92,6 +107,22 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
     int filter_width = col.dims()[2];
     int output_height = col.dims()[3];
     int output_width = col.dims()[4];
+
+    PADDLE_ENFORCE_EQ(
+        (input_height + padding_up + padding_down - filter_height) /
+                stride_height +
+            1,
+        output_height,
+        "Output_height and padding(padding_up, padding_down) are "
+        "inconsistent.");
+    PADDLE_ENFORCE_EQ(
+        (input_width + padding_left + padding_right - filter_width) /
+                stride_width +
+            1,
+        output_width,
+        "output_width and padding(padding_left, padding_right) are "
+        "inconsistent.");
+
     int channels_col = input_channels * filter_height * filter_width;
 
     T* im_data = im.data<T>();
@@ -103,14 +134,12 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
       int c_im = c / filter_width / filter_height;
       for (int h = 0; h < output_height; ++h) {
         for (int w = 0; w < output_width; ++w) {
-          int im_row_idx = h * stride_height + h_offset;
-          int im_col_idx = w * stride_width + w_offset;
-          if ((im_row_idx - padding_height) >= 0 &&
-              (im_row_idx - padding_height) < input_height &&
-              (im_col_idx - padding_width) >= 0 &&
-              (im_col_idx - padding_width) < input_width) {
-            im_row_idx += c_im * input_height - padding_height;
-            im_col_idx -= padding_width;
+          int im_row_idx = h * stride_height + h_offset - padding_up;
+          int im_col_idx = w * stride_width + w_offset - padding_left;
+
+          if ((im_row_idx) >= 0 && (im_row_idx) < input_height &&
+              (im_col_idx) >= 0 && (im_col_idx) < input_width) {
+            im_row_idx += c_im * input_height;
             im_data[im_row_idx * input_width + im_col_idx] +=
                 col_data[(c * output_height + h) * output_width + w];
           }
@@ -140,8 +169,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
  public:
   void operator()(const platform::DeviceContext& context,
                   const framework::Tensor& im, framework::Tensor& col,
-                  int stride_height, int stride_width, int padding_height,
-                  int padding_width) {
+                  int stride_height, int stride_width, int padding_up,
+                  int padding_down, int padding_left, int padding_right) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
     int input_channels = im.dims()[0];
@@ -152,6 +181,21 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
     int output_height = col.dims()[0];
     int output_width = col.dims()[1];
 
+    PADDLE_ENFORCE_EQ(
+        (input_height + padding_up + padding_down - filter_height) /
+                stride_height +
+            1,
+        output_height,
+        "Output_height and padding(padding_up, padding_down) are "
+        "inconsistent.");
+    PADDLE_ENFORCE_EQ(
+        (input_width + padding_left + padding_right - filter_width) /
+                stride_width +
+            1,
+        output_width,
+        "output_width and padding(padding_left, padding_right) are "
+        "inconsistent.");
+
     const T* im_data = im.data<T>();
     T* col_data = col.data<T>();
 
@@ -163,10 +207,10 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
             for (int filter_col_idx = 0; filter_col_idx < filter_width;
                  ++filter_col_idx) {
               int im_row_offset =
-                  col_row_idx * stride_height + filter_row_idx - padding_height;
+                  col_row_idx * stride_height + filter_row_idx - padding_up;
               int im_col_offset =
-                  col_col_idx * stride_width + filter_col_idx - padding_width;
-              int col_offset = (((col_row_idx * output_width + col_col_idx) *
+                  col_col_idx * stride_width + filter_col_idx - padding_left;
+              int col_offset = ((((col_row_idx)*output_width + col_col_idx) *
                                      input_channels +
                                  channel) *
                                     filter_height +
@@ -201,7 +245,8 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
  public:
   void operator()(const platform::DeviceContext& context, framework::Tensor& im,
                   const framework::Tensor& col, int stride_height,
-                  int stride_width, int padding_height, int padding_width) {
+                  int stride_width, int padding_up, int padding_down,
+                  int padding_left, int padding_right) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
     int input_channels = im.dims()[0];
@@ -212,6 +257,21 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
     int output_height = col.dims()[0];
     int output_width = col.dims()[1];
 
+    PADDLE_ENFORCE_EQ(
+        (input_height + padding_up + padding_down - filter_height) /
+                stride_height +
+            1,
+        output_height,
+        "Output_height and padding(padding_up, padding_down) are "
+        "inconsistent.");
+    PADDLE_ENFORCE_EQ(
+        (input_width + padding_left + padding_right - filter_width) /
+                stride_width +
+            1,
+        output_width,
+        "output_width and padding(padding_left, padding_right) are "
+        "inconsistent.");
+
     T* im_data = im.data<T>();
     const T* col_data = col.data<T>();
 
@@ -223,9 +283,9 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
             for (int filter_col_idx = 0; filter_col_idx < filter_width;
                  ++filter_col_idx) {
               int im_row_offset =
-                  col_row_idx * stride_height + filter_row_idx - padding_height;
+                  col_row_idx * stride_height + filter_row_idx - padding_up;
               int im_col_offset =
-                  col_col_idx * stride_width + filter_col_idx - padding_width;
+                  col_col_idx * stride_width + filter_col_idx - padding_left;
               int col_offset = (((col_row_idx * output_width + col_col_idx) *
                                      input_channels +
                                  channel) *

paddle/operators/math/im2col.cu

@@ -66,8 +66,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
  public:
   void operator()(const platform::DeviceContext& context,
                   const framework::Tensor& im, framework::Tensor& col,
-                  int stride_height, int stride_width, int padding_height,
-                  int padding_width) {
+                  int stride_height, int stride_width, int padding_up,
+                  int padding_down, int padding_left, int padding_right) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
 
@@ -79,6 +79,15 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
     int output_height = col.dims()[3];
     int output_width = col.dims()[4];
 
+    PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) /
+                           stride_height +
+                       1 ==
+                   output_height);
+    PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) /
+                           stride_width +
+                       1 ==
+                   output_width);
+
     int num_outputs = input_channels * output_height * output_width;
     int blocks = (num_outputs + 1024 - 1) / 1024;
     int block_x = 512;
@@ -89,8 +98,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
                 reinterpret_cast<const platform::CUDADeviceContext&>(context)
                     .stream()>>>(
         im.data<T>(), num_outputs, input_height, input_width, filter_height,
-        filter_width, stride_height, stride_width, padding_height,
-        padding_width, output_height, output_width, col.data<T>());
+        filter_width, stride_height, stride_width, padding_up, padding_left,
+        output_height, output_width, col.data<T>());
   }
 };
 
@@ -152,7 +161,8 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
  public:
   void operator()(const platform::DeviceContext& context, framework::Tensor& im,
                   const framework::Tensor& col, int stride_height,
-                  int stride_width, int padding_height, int padding_width) {
+                  int stride_width, int padding_up, int padding_down,
+                  int padding_left, int padding_right) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
 
@@ -164,8 +174,18 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
     int output_height = col.dims()[3];
     int output_width = col.dims()[4];
 
-    size_t num_kernels = input_channels * (input_height + 2 * padding_height) *
-                         (input_width + 2 * padding_width);
+    PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) /
+                           stride_height +
+                       1 ==
+                   output_height);
+    PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) /
+                           stride_width +
+                       1 ==
+                   output_width);
+
+    size_t num_kernels = input_channels *
+                         (input_height + padding_up + padding_down) *
+                         (input_width + padding_left + padding_right);
 
     size_t blocks = (num_kernels + 1024 - 1) / 1024;
     size_t block_x = 512;
@@ -178,10 +198,10 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
     col2im<T><<<grid, threads, 0,
                 reinterpret_cast<const platform::CUDADeviceContext&>(context)
                     .stream()>>>(
-        num_kernels, col.data<T>(), input_height + 2 * padding_height,
-        input_width + 2 * padding_width, input_channels, filter_height,
-        filter_width, stride_height, stride_width, padding_height,
-        padding_width, output_height, output_width, im.data<T>());
+        num_kernels, col.data<T>(), input_height + padding_up + padding_down,
+        input_width + padding_left + padding_left, input_channels,
+        filter_height, filter_width, stride_height, stride_width, padding_up,
+        padding_left, output_height, output_width, im.data<T>());
   }
 };
 
@@ -238,8 +258,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
  public:
   void operator()(const platform::DeviceContext& context,
                   const framework::Tensor& im, framework::Tensor& col,
-                  int stride_height, int stride_width, int padding_height,
-                  int padding_width) {
+                  int stride_height, int stride_width, int padding_up,
+                  int padding_down, int padding_left, int padding_right) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
     int input_channels = im.dims()[0];
@@ -250,6 +270,15 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
     int output_height = col.dims()[0];
     int output_width = col.dims()[1];
 
+    PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) /
+                           stride_height +
+                       1 ==
+                   output_height);
+    PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) /
+                           stride_width +
+                       1 ==
+                   output_width);
+
     int block_dim_x = 0;
     int block_dim_y = 0;
     if (filter_height <= 4 && filter_width <= 4) {
@@ -274,8 +303,8 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
                    reinterpret_cast<const platform::CUDADeviceContext&>(context)
                        .stream()>>>(
         im.data<T>(), col.data<T>(), input_channels, input_height, input_width,
-        filter_height, filter_width, stride_height, stride_width,
-        padding_height, padding_width, output_height, output_width);
+        filter_height, filter_width, stride_height, stride_width, padding_up,
+        padding_left, output_height, output_width);
   }
 };
 
@@ -322,7 +351,8 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
  public:
   void operator()(const platform::DeviceContext& context, framework::Tensor& im,
                   const framework::Tensor& col, int stride_height,
-                  int stride_width, int padding_height, int padding_width) {
+                  int stride_width, int padding_up, int padding_down,
+                  int padding_left, int padding_right) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
     int input_channels = im.dims()[0];
@@ -333,6 +363,15 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
     int output_height = col.dims()[0];
     int output_width = col.dims()[1];
 
+    PADDLE_ENFORCE((input_height + padding_up + padding_down - filter_height) /
+                           stride_height +
+                       1 ==
+                   output_height);
+    PADDLE_ENFORCE((input_width + padding_left + padding_right - filter_width) /
+                           stride_width +
+                       1 ==
+                   output_width);
+
     int block_dim_x = 0;
     int block_dim_y = 0;
     if (filter_height <= 4 && filter_width <= 4) {
@@ -357,8 +396,8 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
                    reinterpret_cast<const platform::CUDADeviceContext&>(context)
                        .stream()>>>(
         im.data<T>(), col.data<T>(), input_channels, input_height, input_width,
-        filter_height, filter_width, stride_height, stride_width,
-        padding_height, padding_width, output_height, output_width);
+        filter_height, filter_width, stride_height, stride_width, padding_up,
+        padding_left, output_height, output_width);
   }
 };
 

paddle/operators/math/im2col.h

@@ -74,8 +74,8 @@ class Im2ColFunctor {
  public:
   void operator()(const platform::DeviceContext& context,
                   const framework::Tensor& im, framework::Tensor& col,
-                  int stride_height, int stride_width, int padding_height,
-                  int padding_width);
+                  int stride_height, int stride_width, int padding_up,
+                  int padding_down, int padding_left, int padding_right);
 };
 
 template <ColFormat Format, typename Place, typename T>
@@ -83,7 +83,8 @@ class Col2ImFunctor {
  public:
   void operator()(const platform::DeviceContext& context, framework::Tensor& im,
                   const framework::Tensor& col, int stride_height,
-                  int stride_width, int padding_height, int padding_width);
+                  int stride_width, int padding_up, int padding_down,
+                  int padding_left, int padding_right);
 };
 
 }  // namespace math

paddle/operators/math/im2col_test.cc

@@ -35,6 +35,12 @@ void testIm2col() {
    *
    * output_ocf = [0, 1, 3, 4
    *               1, 2, 4, 5]
+   *
+   * col2im_cfo = [0, 2, 2
+   *               3, 4, 5]
+   *
+   * col2im_ocf = [0, 2, 2
+   *               3, 4, 5]
    */
   int input_height = 2;
   int input_width = 3;
@@ -59,7 +65,7 @@ void testIm2col() {
         new paddle::platform::CUDADeviceContext(paddle::platform::GPUPlace());
 #else
     PADDLE_THROW("no GPU support");
-#endif  // PADDLE_ONLY_CPU
+#endif  // PADDLE_WITH_CUDA
   }
   if (paddle::platform::is_cpu_place(*place)) {
     input = input_tmp;
@@ -71,6 +77,7 @@ void testIm2col() {
   output_ocf.mutable_data<float>(
       {output_height, output_width, 1, filter_size, filter_size}, *place);
 
+  // Im2Col
   paddle::operators::math::Im2ColFunctor<
       paddle::operators::math::ColFormat::kCFO, Place, float>
       im2col;
@@ -78,8 +85,13 @@ void testIm2col() {
       paddle::operators::math::ColFormat::kOCF, Place, float>
       im2col_ocf;
 
-  im2col(*context, input, output_cfo, stride, stride, padding, padding);
-  im2col_ocf(*context, input, output_ocf, stride, stride, padding, padding);
+  im2col(*context, input, output_cfo, stride, stride, padding, padding, padding,
+         padding);
+  im2col_ocf(*context, input, output_ocf, stride, stride, padding, padding,
+             padding, padding);
+
+  float out_cfo_data[] = {0, 1, 1, 2, 3, 4, 4, 5};
+  float out_ocf_data[] = {0, 1, 3, 4, 1, 2, 4, 5};
 
   float* out_cfo_ptr;
   if (paddle::platform::is_cpu_place(*place)) {
@@ -88,14 +100,9 @@ void testIm2col() {
     output_tmp.CopyFrom(output_cfo, paddle::platform::CPUPlace(), *context);
     out_cfo_ptr = output_tmp.data<float>();
   }
-  EXPECT_EQ(out_cfo_ptr[0], 0);
-  EXPECT_EQ(out_cfo_ptr[1], 1);
-  EXPECT_EQ(out_cfo_ptr[2], 1);
-  EXPECT_EQ(out_cfo_ptr[3], 2);
-  EXPECT_EQ(out_cfo_ptr[4], 3);
-  EXPECT_EQ(out_cfo_ptr[5], 4);
-  EXPECT_EQ(out_cfo_ptr[6], 4);
-  EXPECT_EQ(out_cfo_ptr[7], 5);
+  for (int i = 0; i < 6; ++i) {
+    EXPECT_EQ(out_cfo_ptr[i], out_cfo_data[i]);
+  }
 
   float* out_ocf_ptr;
   if (paddle::platform::is_cpu_place(*place)) {
@@ -104,14 +111,60 @@ void testIm2col() {
     output_tmp.CopyFrom(output_ocf, paddle::platform::CPUPlace(), *context);
     out_ocf_ptr = output_tmp.data<float>();
   }
-  EXPECT_EQ(out_ocf_ptr[0], 0);
-  EXPECT_EQ(out_ocf_ptr[1], 1);
-  EXPECT_EQ(out_ocf_ptr[2], 3);
-  EXPECT_EQ(out_ocf_ptr[3], 4);
-  EXPECT_EQ(out_ocf_ptr[4], 1);
-  EXPECT_EQ(out_ocf_ptr[5], 2);
-  EXPECT_EQ(out_ocf_ptr[6], 4);
-  EXPECT_EQ(out_ocf_ptr[7], 5);
+  for (int i = 0; i < 6; ++i) {
+    EXPECT_EQ(out_ocf_ptr[i], out_ocf_data[i]);
+  }
+
+  // Col2Im: kCFO
+  paddle::operators::math::Col2ImFunctor<
+      paddle::operators::math::ColFormat::kCFO, Place, float>
+      col2im;
+  paddle::operators::math::Col2ImFunctor<
+      paddle::operators::math::ColFormat::kOCF, Place, float>
+      col2im_ocf;
+  float col2im_data[] = {0, 2, 2, 3, 8, 5};
+
+  memset(input_ptr, 0, 6 * sizeof(float));
+  if (paddle::platform::is_cpu_place(*place)) {
+    input = input_tmp;
+  } else {
+    input.CopyFrom(input_tmp, *place, *context);
+  }
+
+  col2im(*context, input, output_cfo, stride, stride, padding, padding, padding,
+         padding);
+
+  float* in_ptr;
+  if (paddle::platform::is_cpu_place(*place)) {
+    in_ptr = input.data<float>();
+  } else {
+    input_tmp.CopyFrom(input, paddle::platform::CPUPlace(), *context);
+    in_ptr = input_tmp.data<float>();
+  }
+  for (int i = 0; i < 6; ++i) {
+    EXPECT_EQ(in_ptr[i], col2im_data[i]);
+  }
+
+  // Col2Im: kOCF
+  memset(input_ptr, 0, 6 * sizeof(float));
+  if (paddle::platform::is_cpu_place(*place)) {
+    input = input_tmp;
+  } else {
+    input.CopyFrom(input_tmp, *place, *context);
+  }
+
+  col2im_ocf(*context, input, output_ocf, stride, stride, padding, padding,
+             padding, padding);
+
+  if (paddle::platform::is_cpu_place(*place)) {
+    in_ptr = input.data<float>();
+  } else {
+    input_tmp.CopyFrom(input, paddle::platform::CPUPlace(), *context);
+    in_ptr = input_tmp.data<float>();
+  }
+  for (int i = 0; i < 6; ++i) {
+    EXPECT_EQ(in_ptr[i], col2im_data[i]);
+  }
 }
 
 TEST(math, im2col) {