Add column to image kernel (#930)

* Add column to image kernel * Minor fixes for dtypes and client examples * Disable tests for disabled dtypes * Disable add instances functions for disabled data types * Minor stylistic fixes * Revert "Disable add instances functions for disabled data types" This reverts commit 728b869563. * Instances reduction * Add comments in device_column_to_image_impl * Update changelog and Copyrights * Improve changelog [ROCm/composable_kernel commit: e2243a4d1e]
2026-05-17 11:30:02 +00:00 · 2023-09-27 17:19:06 +02:00
parent e21f21022c
commit 9bc92adde3
41 changed files with 3040 additions and 1201 deletions
--- a/library/src/tensor_operation_instance/gpu/column_to_image/CMakeLists.txt
+++ b/library/src/tensor_operation_instance/gpu/column_to_image/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_instance_library(device_column_to_image_instance
+   device_column_to_image_nhwc_1d_instance.cpp
+   device_column_to_image_nhwc_2d_instance.cpp
+   device_column_to_image_nhwc_3d_instance.cpp
+)
--- a/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_1d_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_1d_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using namespace ck::conv_tensor_rearrange_op;
+
+void add_device_column_to_image_nwc_1d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, BF16, BF16, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_BF16
+    add_device_operation_instances(instances, device_column_to_image_bf16_instances<1, GNWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+void add_device_column_to_image_nwc_1d_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, F16, F16, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_FP16
+    add_device_operation_instances(instances, device_column_to_image_f16_instances<1, GNWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+void add_device_column_to_image_nwc_1d_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, F32, F32, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_FP32
+    add_device_operation_instances(instances, device_column_to_image_f32_instances<1, GNWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+void add_device_column_to_image_nwc_1d_i8_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, int8_t, int8_t, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_INT8
+    add_device_operation_instances(instances, device_column_to_image_i8_instances<1, GNWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
--- a/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_2d_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_2d_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using namespace ck::conv_tensor_rearrange_op;
+
+void add_device_column_to_image_nhwc_2d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, BF16, BF16, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_BF16
+    add_device_operation_instances(instances, device_column_to_image_bf16_instances<2, GNHWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+void add_device_column_to_image_nhwc_2d_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, F16, F16, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_FP16
+    add_device_operation_instances(instances, device_column_to_image_f16_instances<2, GNHWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+void add_device_column_to_image_nhwc_2d_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, F32, F32, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_FP32
+    add_device_operation_instances(instances, device_column_to_image_f32_instances<2, GNHWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+void add_device_column_to_image_nhwc_2d_i8_instances(
+    std::vector<
+        std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, int8_t, int8_t, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_INT8
+    add_device_operation_instances(instances, device_column_to_image_i8_instances<2, GNHWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
--- a/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_3d_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_3d_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using namespace ck::conv_tensor_rearrange_op;
+
+void add_device_column_to_image_ndhwc_3d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, BF16, BF16, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_BF16
+    add_device_operation_instances(instances, device_column_to_image_bf16_instances<3, GNDHWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+void add_device_column_to_image_ndhwc_3d_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, F16, F16, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_FP16
+    add_device_operation_instances(instances, device_column_to_image_f16_instances<3, GNDHWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+void add_device_column_to_image_ndhwc_3d_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, F32, F32, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_FP32
+    add_device_operation_instances(instances, device_column_to_image_f32_instances<3, GNDHWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+void add_device_column_to_image_ndhwc_3d_i8_instances(
+    std::vector<
+        std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, int8_t, int8_t, ColumnToImage>>>&
+        instances)
+{
+#ifdef CK_ENABLE_INT8
+    add_device_operation_instances(instances, device_column_to_image_i8_instances<3, GNDHWC>{});
+#else
+    ignore = instances;
+#endif
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
--- a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_1d_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_1d_instance.cpp
@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.

-#include "ck/library/tensor_operation_instance/gpu/image_to_column/device_image_to_column_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp"
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"

 namespace ck {
@@ -9,28 +9,50 @@ namespace tensor_operation {
 namespace device {
 namespace instance {

-void add_device_image_to_column_nhwc_1d_bf16_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, BF16, BF16>>>& instances)
+using namespace ck::conv_tensor_rearrange_op;
+
+void add_device_image_to_column_nwc_1d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, BF16, BF16, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_BF16
    add_device_operation_instances(instances, device_image_to_column_bf16_instances<1, GNWC>{});
+#else
+    ignore = instances;
+#endif
 }

-void add_device_image_to_column_nhwc_1d_f16_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, F16, F16>>>& instances)
+void add_device_image_to_column_nwc_1d_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, F16, F16, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_FP16
    add_device_operation_instances(instances, device_image_to_column_f16_instances<1, GNWC>{});
+#else
+    ignore = instances;
+#endif
 }

-void add_device_image_to_column_nhwc_1d_f32_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, F32, F32>>>& instances)
+void add_device_image_to_column_nwc_1d_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, F32, F32, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_FP32
    add_device_operation_instances(instances, device_image_to_column_f32_instances<1, GNWC>{});
+#else
+    ignore = instances;
+#endif
 }

-void add_device_image_to_column_nhwc_1d_i8_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, int8_t, int8_t>>>& instances)
+void add_device_image_to_column_nwc_1d_i8_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<1, GNWC, int8_t, int8_t, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_INT8
    add_device_operation_instances(instances, device_image_to_column_i8_instances<1, GNWC>{});
+#else
+    ignore = instances;
+#endif
 }

 } // namespace instance
--- a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_2d_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_2d_instance.cpp
@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.

-#include "ck/library/tensor_operation_instance/gpu/image_to_column/device_image_to_column_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp"
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"

 namespace ck {
@@ -9,28 +9,51 @@ namespace tensor_operation {
 namespace device {
 namespace instance {

+using namespace ck::conv_tensor_rearrange_op;
+
 void add_device_image_to_column_nhwc_2d_bf16_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, BF16, BF16>>>& instances)
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, BF16, BF16, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_BF16
    add_device_operation_instances(instances, device_image_to_column_bf16_instances<2, GNHWC>{});
+#else
+    ignore = instances;
+#endif
 }

 void add_device_image_to_column_nhwc_2d_f16_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, F16, F16>>>& instances)
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, F16, F16, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_FP16
    add_device_operation_instances(instances, device_image_to_column_f16_instances<2, GNHWC>{});
+#else
+    ignore = instances;
+#endif
 }

 void add_device_image_to_column_nhwc_2d_f32_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, F32, F32>>>& instances)
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, F32, F32, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_FP32
    add_device_operation_instances(instances, device_image_to_column_f32_instances<2, GNHWC>{});
+#else
+    ignore = instances;
+#endif
 }

 void add_device_image_to_column_nhwc_2d_i8_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, int8_t, int8_t>>>& instances)
+    std::vector<
+        std::unique_ptr<DeviceConvTensorRearrange<2, GNHWC, int8_t, int8_t, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_INT8
    add_device_operation_instances(instances, device_image_to_column_i8_instances<2, GNHWC>{});
+#else
+    ignore = instances;
+#endif
 }

 } // namespace instance
--- a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_3d_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_3d_instance.cpp
@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.

-#include "ck/library/tensor_operation_instance/gpu/image_to_column/device_image_to_column_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp"
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"

 namespace ck {
@@ -9,28 +9,51 @@ namespace tensor_operation {
 namespace device {
 namespace instance {

-void add_device_image_to_column_nhwc_3d_bf16_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, BF16, BF16>>>& instances)
+using namespace ck::conv_tensor_rearrange_op;
+
+void add_device_image_to_column_ndhwc_3d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, BF16, BF16, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_BF16
    add_device_operation_instances(instances, device_image_to_column_bf16_instances<3, GNDHWC>{});
+#else
+    ignore = instances;
+#endif
 }

-void add_device_image_to_column_nhwc_3d_f16_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, F16, F16>>>& instances)
+void add_device_image_to_column_ndhwc_3d_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, F16, F16, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_FP16
    add_device_operation_instances(instances, device_image_to_column_f16_instances<3, GNDHWC>{});
+#else
+    ignore = instances;
+#endif
 }

-void add_device_image_to_column_nhwc_3d_f32_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, F32, F32>>>& instances)
+void add_device_image_to_column_ndhwc_3d_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, F32, F32, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_FP32
    add_device_operation_instances(instances, device_image_to_column_f32_instances<3, GNDHWC>{});
+#else
+    ignore = instances;
+#endif
 }

-void add_device_image_to_column_nhwc_3d_i8_instances(
-    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, int8_t, int8_t>>>& instances)
+void add_device_image_to_column_ndhwc_3d_i8_instances(
+    std::vector<
+        std::unique_ptr<DeviceConvTensorRearrange<3, GNDHWC, int8_t, int8_t, ImageToColumn>>>&
+        instances)
 {
+#ifdef CK_ENABLE_INT8
    add_device_operation_instances(instances, device_image_to_column_i8_instances<3, GNDHWC>{});
+#else
+    ignore = instances;
+#endif
 }

 } // namespace instance