Add two stage grouped conv bwd weight kernel (#1280)

client_example/11_grouped_conv_bwd_weight/common.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <cstdlib>
 #include <iomanip>
@@ -160,6 +160,10 @@ bool run_grouped_conv_bwd_weight(
         auto invoker_ptr    = op_ptr->MakeInvokerPointer();
         std::string op_name = op_ptr->GetTypeString();
 
+        const std::size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+        SimpleDeviceMem workspace_dev(workspace_sz);
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
         if(op_ptr->IsSupportedArgument(argument_ptr.get()))
         {
             float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using namespace ck::tensor_layout::convolution;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec>
+using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances = std::tuple<
+    // clang-format off
+        //#########################################|              Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|              Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     4,  8,   16,   16,    1,    1,  S<1, 4, 8,  1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              1,              4,      true,  S<1, 4, 8,  1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              1,              4,      true,           1,           1,   S<1, 8, 1, 8>,               1>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp

@@ -352,6 +352,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 {
                     add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
                         op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+                        op_ptrs);
                 }
 #endif
 #ifdef CK_ENABLE_BF16
@@ -419,6 +421,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 {
                     add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
                         op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+                        op_ptrs);
                 }
 #endif
 #ifdef CK_ENABLE_BF16

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_xdl.inc

@@ -113,6 +113,18 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
                                                            PassThrough,
                                                            PassThrough,
                                                            PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
 #endif
 #ifdef CK_ENABLE_FP32
 void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
@@ -192,6 +204,18 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances
                                                            PassThrough,
                                                            PassThrough,
                                                            PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
 #endif
 #ifdef CK_ENABLE_FP32
 void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt

@@ -5,7 +5,8 @@ set(GROUPED_CONV2D_BWD_WEIGHT
     xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
     xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
     xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
-    xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp)
+    xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp)
 
 if(DL_KERNELS)
     list(APPEND GROUPED_CONV2D_BWD_WEIGHT

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances<
+            2,
+            NHWGC,
+            GKYXC,
+            NHWGK,
+            ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances<
+            2,
+            NHWGC,
+            GKYXC,
+            NHWGK,
+            ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt

@@ -5,7 +5,8 @@ set(GROUPED_CONV3D_BWD_WEIGHT
     xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
     xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
     xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
-    xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp)
+    xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
+    xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp)
 
 if(DL_KERNELS)
     list(APPEND GROUPED_CONV3D_BWD_WEIGHT

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances<
+            3,
+            NDHWGC,
+            GKZYXC,
+            NDHWGK,
+            ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances<
+            3,
+            NDHWGC,
+            GKZYXC,
+            NDHWGK,
+            ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp

@@ -188,6 +188,10 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
                                         out_element_op,
                                         split_k);
 
+        const std::size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+        DeviceMem workspace_dev(workspace_sz);
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
         if(op_ptr->IsSupportedArgument(argument_ptr.get()))
         {
             // using atomic add, so need to reset input