Merge remote-tracking branch 'origin/develop' into aosewski/gemm_tile_loop

example/35_splitK_gemm/CMakeLists.txt

@@ -3,25 +3,28 @@ set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)
    add_custom_target(example_splitK_gemm_xdl)
-   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
-    add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
+
+  add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
+  if(result EQUAL 0)
     add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32)
    endif()
-   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
+  add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
+  if(result EQUAL 0)
     add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
    endif()
-   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_splitK_gemm_xdl_bfp16 splitK_gemm_xdl_bfp16.cpp)
-    add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bfp16)
+  add_example_executable(example_splitK_gemm_xdl_bf16 splitK_gemm_xdl_bf16.cpp)
+  if(result EQUAL 0)
+    add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bf16)
    endif()
-   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
-    add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
+  add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
+  if(result EQUAL 0)
     add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int8)
    endif()
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
-      add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)
+      if(result EQUAL 0)
+        add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)
+      endif()
    endif()
    set(target 1)
  endif()

example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)
-endif()
+add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)

example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list_wmma gfx1100 gfx1101 gfx1102)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
+ if(gpu IN_LIST gpu_list_xdl AND target EQUAL 0)
    add_custom_target(example_grouped_conv_bwd_data)
-   add_example_executable(example_grouped_conv_bwd_data_fp16 grouped_conv_bwd_data_fp16.cpp)
-   add_example_executable(example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp)
+   add_example_executable(example_grouped_conv_bwd_data_xdl_fp16 grouped_conv_bwd_data_xdl_fp16.cpp)
+   if(result EQUAL 0)
+      add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_xdl_fp16)
+   endif()
+   add_example_executable(example_grouped_conv_bwd_data_bias_relu_xdl_fp16 grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp)
+   if(result EQUAL 0)
+     add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_xdl_fp16)
+   endif()
+   set(target 1)
+ endif()
+endforeach()
 
-   add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_fp16)
-   add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list_wmma AND target EQUAL 0)
+   add_custom_target(example_grouped_conv_bwd_data)
+   add_example_executable(example_grouped_conv_bwd_data_wmma_fp16 grouped_conv_bwd_data_wmma_fp16.cpp)
+   if(result EQUAL 0)
+      add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_wmma_fp16)
+   endif()
    set(target 1)
  endif()
 endforeach()
-endif()

example/38_grouped_conv_bwd_data_multiple_d/common.hpp

@@ -10,7 +10,6 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 

example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_fp16.cpp → example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp

 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
 #include "common.hpp"
 
 using OutDataType      = FP16;

example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp"
+#include "common.hpp"
+
+using OutDataType      = FP16;
+using WeiDataType      = FP16;
+using AccDataType      = FP32;
+using CShuffleDataType = FP16;
+using DsDataType       = ck::Tuple<>;
+using InDataType       = FP16;
+
+using OutLayout = ck::tensor_layout::convolution::GNHWK;
+using WeiLayout = ck::tensor_layout::convolution::GKYXC;
+using DsLayout  = ck::Tuple<>;
+using InLayout  = ck::tensor_layout::convolution::GNHWC;
+
+using OutElementOp = PassThrough;
+using WeiElementOp = PassThrough;
+using InElementOp  = PassThrough;
+
+// clang-format off
+using DeviceConvInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
+        //|    NumDim|        A|         B|       Ds|       E|        AData|        BData|    AccData|          CShuffle|     DsData|       EData|           A|           B|          CDE|       ConvForward| Block|  MPer|  NPer| K0Per| K1|  MPer| NPer| MRepeat| NRepeat|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //|   Spatial|   Layout|    Layout|   Layout|  Layout|         Type|         Type|       Type|          DataType|       Type|        Type| Elementwise| Elementwise|  Elementwise|    Specialization|  Size| Block| Block| Block|   |  WMMA| WMMA|        |        |   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //|          |         |          |         |        |             |             |           |                  |           |            |   Operation|   Operation|    Operation|                  |      |      |      |      |   |      |     |        |        | Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //|          |         |          |         |        |             |             |           |                  |           |            |            |            |             |                  |      |      |      |      |   |      |     |        |        |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+           <        2,OutLayout, WeiLayout, DsLayout, InLayout, OutDataType,  WeiDataType, AccDataType, CShuffleDataType, DsDataType,  InDataType, OutElementOp, WeiElementOp, InElementOp, ConvBwdDataDefault, 128,    64,    64,     4,  8,    16,   16,       2,       2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,               1>;
+// clang-format on
+
+#include "run_grouped_conv_bwd_data_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_bwd_data_example(argc, argv); }

example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_fp16.cpp → example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_xdl_fp16.cpp

 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
 #include "common.hpp"
 
 using OutDataType      = FP16;

example/39_permute/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_custom_target(example_permute)
-
-    add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
-    add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
-    add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
+add_custom_target(example_permute)
 
+add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
+if(result EQUAL 0)
     add_dependencies(example_permute example_permute_1xHxW_fp16)
+endif()
+add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
+if(result EQUAL 0)
     add_dependencies(example_permute example_permute_NxHxW_fp16)
+endif()
+add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
+if(result EQUAL 0)
     add_dependencies(example_permute example_permute_HxWx4_fp16)
 endif()

example/40_conv2d_fwd_quantization/CMakeLists.txt

-if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
@@ -11,7 +10,6 @@ foreach(gpu IN LISTS GPU_TARGETS)
  endif()
 endforeach()
 
-  if(DL_KERNELS)
   # Conv perlayer quantization
   add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
   # Conv perchannel quantization
@@ -24,5 +22,3 @@ endforeach()
   add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
   # Conv + bias + tanh perchannel quantization
   add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
-  endif()
-endif()
\ No newline at end of file

example/41_grouped_conv_conv_fwd/CMakeLists.txt

@@ -3,15 +3,9 @@ list(APPEND gpu_list2 gfx908 gfx90a)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
-   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
-      add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
-   endif()
-   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-      add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
-   endif()
-   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
-      add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
-   endif()
+   add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
+   add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
+   add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp)
    endif(USE_BITINT_EXTENSION_INT4)
@@ -20,7 +14,5 @@ foreach(gpu IN LISTS GPU_TARGETS)
 endforeach()
 
 if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
-   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
-      add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
-   endif()
+   add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
 endif()

example/42_groupnorm/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
-    add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
-    add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
-endif()
+add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
+add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
+add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)

example/43_splitk_gemm_bias_e_permute/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
-endif()
-if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
-    add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
-endif()
+add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
+add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)

example/44_elementwise_permute/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
-    add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
-endif()
+add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
+add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)

example/46_gemm_add_multiply/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    if(DL_KERNELS)
-        add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
-    endif()
-    add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
-endif()
+add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
+add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)

example/48_pool3d_fwd/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
-endif()
+add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)

example/49_maxpool2d_bwd/CMakeLists.txt

-if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
-endif()
-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
-endif()
-if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
-    add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)
-endif()
+add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
+add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
+add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)

example/50_put_element/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_put_element_fp16 put_element_fp16.cpp)
-endif()
+add_example_executable(example_put_element_fp16 put_element_fp16.cpp)

example/52_image_to_column/CMakeLists.txt → example/52_im2col_col2im/CMakeLists.txt

@@ -2,9 +2,11 @@ list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)
-   add_custom_target(example_image_to_column)
+   add_custom_target(example_im2col_col2im)
    add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)
-   add_dependencies(example_image_to_column example_image_to_column_f32)
+   add_dependencies(example_im2col_col2im example_image_to_column_f32)
+   add_example_executable(example_column_to_image_f32 column_to_image_f32.cpp)
+   add_dependencies(example_im2col_col2im example_column_to_image_f32)
    set(target 1)
  endif()
 endforeach()

example/52_im2col_col2im/column_to_image_f32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using InDataType  = FP32; // ck::bhalf_t;//FP32;
+using OutDataType = FP32; // ck::bhalf_t;//FP32;
+
+using ImLayout        = ck::tensor_layout::convolution::GNHWC;
+using ColumnToImageOp = ck::conv_tensor_rearrange_op::ColumnToImage;
+
+// clang-format off
+using DeviceColToImgInstance = ck::tensor_operation::device::DeviceColumnToImageImpl
+        //#####################|        Num| ImLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+                              < NDimSpatial, ImLayout, InDataType, OutDataType,   256,   128,   128, S<16, 16>,     1>;
+// clang-format on
+
+bool RunColumnToImage(const ExecutionConfig& config, const ck::utils::conv::ConvParam& conv_params)
+{
+
+    const auto N = conv_params.N_;
+    const auto C = conv_params.C_;
+
+    const ck::index_t NDoHoWo =
+        N * ck::accumulate_n<ck::index_t>(
+                conv_params.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+    const ck::index_t CZYX =
+        C * ck::accumulate_n<ck::index_t>(
+                conv_params.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+
+    const auto in_desc = HostTensorDescriptor({NDoHoWo, CZYX});
+    const auto out_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ImLayout>(conv_params);
+
+    std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> image_g_n_c_wis_strides{};
+    std::array<ck::index_t, 2> gemm_m_k_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+
+    copy(conv_params.input_spatial_lengths_, input_spatial_lengths);
+    copy(conv_params.filter_spatial_lengths_, filter_spatial_lengths);
+    copy(conv_params.output_spatial_lengths_, output_spatial_lengths);
+    copy(in_desc.GetStrides(), gemm_m_k_strides);
+    copy(out_desc.GetStrides(), image_g_n_c_wis_strides);
+    copy(conv_params.conv_filter_strides_, conv_filter_strides);
+    copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_params.input_left_pads_, input_left_pads);
+    copy(conv_params.input_right_pads_, input_right_pads);
+
+    Tensor<InDataType> in(in_desc);
+    Tensor<OutDataType> out_device(out_desc);
+    Tensor<OutDataType> out_host(out_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "out: " << out_device.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1: in.GenerateTensorValue(GeneratorTensor_2<InDataType>{1, 2}); break;
+    default: in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+
+    // reset input to zero
+    out_device_buf.SetZero();
+
+    static_assert(std::is_default_constructible_v<DeviceColToImgInstance>);
+
+    // do conv
+    auto col2img  = DeviceColToImgInstance{};
+    auto invoker  = col2img.MakeInvoker();
+    auto argument = col2img.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                         out_device_buf.GetDeviceBuffer(),
+                                         N,
+                                         C,
+                                         input_spatial_lengths,
+                                         filter_spatial_lengths,
+                                         output_spatial_lengths,
+                                         image_g_n_c_wis_strides,
+                                         gemm_m_k_strides,
+                                         conv_filter_strides,
+                                         conv_filter_dilations,
+                                         input_left_pads,
+                                         input_right_pads);
+
+    if(!col2img.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_col2img with the specified compilation parameters does "
+                     "not support this col2img problem"
+                  << std::endl;
+
+        return false;
+    }
+
+    float ave_time        = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+    std::size_t num_btype = NDoHoWo * CZYX * (sizeof(OutDataType) + sizeof(InDataType));
+    float gb_per_sec      = num_btype / 1.E6 / ave_time;
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_column_to_image = ck::tensor_operation::host::
+            ReferenceColumnToImage<NDimSpatial, ImLayout, InDataType, OutDataType>();
+
+        auto ref_invoker = ref_column_to_image.MakeInvoker();
+
+        auto ref_argument = ref_column_to_image.MakeArgument(in,
+                                                             out_host,
+                                                             conv_params.filter_spatial_lengths_,
+                                                             conv_params.conv_filter_strides_,
+                                                             conv_params.conv_filter_dilations_,
+                                                             conv_params.input_left_pads_,
+                                                             conv_params.input_right_pads_);
+
+        if(!ref_column_to_image.IsSupportedArgument(&ref_argument))
+        {
+            std::cerr << "wrong! ref_col2img with the specified compilation parameters does "
+                         "not support this col2img problem"
+                      << std::endl;
+            return false;
+        }
+
+        ref_invoker.Run(ref_argument);
+        out_device_buf.FromDevice(out_device.mData.data());
+        return ck::utils::check_err(out_device.mData, out_host.mData);
+    }
+
+    return true;
+}
+
+int RunColumnToImageExample(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_params = DefaultConvParams;
+
+    if(!parse_cmd_args(argc, argv, config, conv_params))
+    {
+        return EXIT_FAILURE;
+    }
+
+    if(conv_params.num_dim_spatial_ != NDimSpatial)
+    {
+        std::cerr << "unsupported # of spatial dimensions" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    return !RunColumnToImage(config, conv_params);
+}
+
+int main(int argc, char* argv[]) { return RunColumnToImageExample(argc, argv); }