Merge branch 'develop' of...

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into add_fp16_wmma_conv_instance

library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp

@@ -11,14 +11,18 @@ namespace instance {
 static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
 
 void add_device_pool3d_fwd_ndhwc_f32_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, false>>>& instances)
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
 {
     add_device_operation_instances(
         instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
 }
 
 void add_device_pool3d_fwd_ndhwc_index_f32_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, true>>>& instances)
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, true>>>&
+        instances)
 {
     add_device_operation_instances(
         instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, true>{});

library/src/tensor_operation_instance/gpu/pool_fwd/pool_fwd_instance_common.hpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp

@@ -17,22 +17,9 @@ namespace instance {
 
 using I32   = int32_t;
 using F16   = ck::half_t;
+using BF16  = ck::bhalf_t;
 using F32   = float;
-
-template <typename InDataType,
-          typename OutDataType,
-          typename IndexDataType,
-          typename ComputeDataType,
-          ReduceTensorOp ReduceOpId,
-          bool OutputIndex>
-using device_pool2d_fwd_nhwc_instances =
-    // clang-format off
-    std::tuple <
-        DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
-        DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
-        DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
-               // clang-format on
-               >;
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
 
 template <typename InDataType,
           typename OutDataType,
@@ -43,9 +30,9 @@ template <typename InDataType,
 using device_pool3d_fwd_ndhwc_instances =
     // clang-format off
     std::tuple <
-        DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
-        DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
-        DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
+        DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
+        DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
+        DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
                // clang-format on
                >;
 

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

-add_instance_library(device_pool_fwd_instance
-    device_avg_pool2d_fwd_nhwc_f16_instance.cpp
-    device_avg_pool2d_fwd_nhwc_f32_instance.cpp
-    device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
-    device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
-    device_max_pool2d_fwd_nhwc_f16_instance.cpp
-    device_max_pool2d_fwd_nhwc_f32_instance.cpp
-    device_max_pool3d_fwd_ndhwc_f16_instance.cpp
-    device_max_pool3d_fwd_ndhwc_f32_instance.cpp
-)

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

 if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
-set(CONV2D_PERLAYER_QUANT_SRC
-    conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp
-    conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
-)
-
-set(CONV2D_PERCHANNEL_QUANT_SRC
-    conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp
-    conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
-)
-
-set(CONV2D_BIAS_PERLAYER_QUANT_SRC
-    conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp
-    conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
-)
-
-set(CONV2D_BIAS_PERCHANNEL_QUANT_SRC
-    conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp
-    conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
-)
 
+set(CONV2D_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp)
+set(CONV2D_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp)
+set(CONV2D_BIAS_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp)
+set(CONV2D_BIAS_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp)
 set(GEMM_QUANT_SRC
-    gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
-    gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
-    gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
-    gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
     gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
     gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
     gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
     gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
 )
+if(DL_KERNELS)
+    list(APPEND CONV2D_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp)
+    list(APPEND CONV2D_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp)
+    list(APPEND CONV2D_BIAS_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp)
+    list(APPEND CONV2D_BIAS_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp)
+    list(APPEND GEMM_QUANT_SRC
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp)
+endif()
 
 add_instance_library(device_quantization_instance
     ${CONV2D_PERLAYER_QUANT_SRC}

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

-add_instance_library(device_softmax_instance
-    device_softmax_f16_f16_instance_rank3_reduce1.cpp
+set(DEVICE_SOFTMAX_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_SOFTMAX_INSTANCES device_softmax_f16_f16_instance_rank3_reduce1.cpp
     device_softmax_f16_f16_instance_rank3_reduce2.cpp
     device_softmax_f16_f16_instance_rank3_reduce3.cpp
     device_softmax_f16_f16_instance_rank4_reduce1.cpp
     device_softmax_f16_f16_instance_rank4_reduce2.cpp
     device_softmax_f16_f16_instance_rank4_reduce3.cpp
-    device_softmax_f16_f16_instance_rank4_reduce4.cpp
-    device_softmax_f32_f32_instance_rank3_reduce1.cpp
+    device_softmax_f16_f16_instance_rank4_reduce4.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_SOFTMAX_INSTANCES device_softmax_f32_f32_instance_rank3_reduce1.cpp
     device_softmax_f32_f32_instance_rank3_reduce2.cpp
     device_softmax_f32_f32_instance_rank3_reduce3.cpp
     device_softmax_f32_f32_instance_rank4_reduce1.cpp
     device_softmax_f32_f32_instance_rank4_reduce2.cpp
     device_softmax_f32_f32_instance_rank4_reduce3.cpp
-    device_softmax_f32_f32_instance_rank4_reduce4.cpp
-)
+    device_softmax_f32_f32_instance_rank4_reduce4.cpp)
+endif()
+add_instance_library(device_softmax_instance ${DEVICE_SOFTMAX_INSTANCES})

library/src/utility/device_memory.cpp

@@ -10,20 +10,67 @@ DeviceMem::DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
     hip_check_error(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
 }
 
+void DeviceMem::Realloc(std::size_t mem_size)
+{
+    if(mpDeviceBuf)
+    {
+        hip_check_error(hipFree(mpDeviceBuf));
+    }
+    mMemSize = mem_size;
+    hip_check_error(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
+}
+
 void* DeviceMem::GetDeviceBuffer() const { return mpDeviceBuf; }
 
 std::size_t DeviceMem::GetBufferSize() const { return mMemSize; }
 
 void DeviceMem::ToDevice(const void* p) const
 {
-    hip_check_error(hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
+    if(mpDeviceBuf)
+    {
+        hip_check_error(
+            hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
+    }
+    else
+    {
+        throw std::runtime_error("ToDevice with an empty pointer");
+    }
+}
+
+void DeviceMem::ToDevice(const void* p, const std::size_t cpySize) const
+{
+    hip_check_error(hipMemcpy(mpDeviceBuf, const_cast<void*>(p), cpySize, hipMemcpyHostToDevice));
 }
 
 void DeviceMem::FromDevice(void* p) const
 {
+    if(mpDeviceBuf)
+    {
         hip_check_error(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
+    }
+    else
+    {
+        throw std::runtime_error("FromDevice with an empty pointer");
+    }
 }
 
-void DeviceMem::SetZero() const { hip_check_error(hipMemset(mpDeviceBuf, 0, mMemSize)); }
+void DeviceMem::FromDevice(void* p, const std::size_t cpySize) const
+{
+    hip_check_error(hipMemcpy(p, mpDeviceBuf, cpySize, hipMemcpyDeviceToHost));
+}
 
-DeviceMem::~DeviceMem() { hip_check_error(hipFree(mpDeviceBuf)); }
+void DeviceMem::SetZero() const
+{
+    if(mpDeviceBuf)
+    {
+        hip_check_error(hipMemset(mpDeviceBuf, 0, mMemSize));
+    }
+}
+
+DeviceMem::~DeviceMem()
+{
+    if(mpDeviceBuf)
+    {
+        hip_check_error(hipFree(mpDeviceBuf));
+    }
+}

profiler/README.md

@@ -144,7 +144,7 @@ GB/s: 127.947
 
 ## Profile grouped convolution backward weight kernels
 ```bash
-# arg1: tensor operation (grouped_conv_bwd_data: Grouped Convolution Backward Data)
+# arg1: tensor operation (grouped_conv_bwd_weight: Grouped Convolution Backward Weight)
 # arg2: data type (0: Input fp32, Weight fp32, Output fp32
 #                  1: Input fp16, Weight fp16, Output fp16
 #                  2: Input bf16, Weight fp32, Output bf16)
@@ -167,7 +167,7 @@ GB/s: 127.947
 # SplitK
 
  ################                   op   datatype  layout  verify  init  log  time  Ndims  G   N   K   C  Y  X  Hi  Wi  Sy  Sx  Dy  Dx  LeftPy  LeftPx  RightPy  RightPx  SplitK
-./bin/ckProfiler grouped_conv_bwd_data          1       0       1     1    0     1      2 32 256 256 512  3  3  28  28   1   1   1   1       1       0        0        0       1
+./bin/ckProfiler grouped_conv_bwd_weight          1       0       1     1    0     1      2 32 256 256 512  3  3  28  28   1   1   1   1       1       0        0        0       1
 
  ```
 
@@ -184,3 +184,41 @@ tflops: 95.337
 GB/s: 69.2301
 ```
 Note: This kernel use atomic add, this will cause output buffer to be accumulated multiple times, causing verification failure. To work around it, do not use CK's own timer and do verification at the same time.
+
+## Profile image to column kernels
+```bash
+# arg1: tensor operation (" OP_NAME ": " OP_DESC ")
+# arg2: data type (0: Input fp32, Weight fp32, Output fp32
+#                  1: Input fp16, Weight fp16, Output fp16
+#                  2: Input bf16, Weight bf16, Output bf16
+#                  3: Input int8, Weight int8, Output int8)
+# arg3: tensor layout (0: Input[N, Hi, Wi, C], Output[N * Ho * Wo, Y * X * C])
+# arg4: verification (0: no, 1: yes)
+# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
+# arg6: print tensor value (0: no; 1: yes)
+# arg7: time kernel (0: no, 1: yes)
+# Following arguments (depending on number of spatial dims):
+#  Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
+#  G, N, K, C, 
+#  <filter spatial dimensions>, (ie Y, X for 2D)
+#  <input image spatial dimensions>, (ie Hi, Wi for 2D)
+#  <strides>, (ie Sy, Sx for 2D)
+#  <dilations>, (ie Dy, Dx for 2D)
+#  <left padding>, (ie LeftPy, LeftPx for 2D)
+#  <right padding>, (ie RightPy, RightPx for 2D)
+
+ ################             op   datatype  layout  verify  init  log  time  Ndims  G   N   K   C  Y  X  Hi  Wi  Sy  Sx  Dy  Dx  LeftPy  LeftPx  RightPy  RightPx
+./bin/ckProfiler image_to_column          0       0       1     1    0     1      2  1 256   1 512  3  3   28  28   1   1   1   1        0       0       0        0
+
+ ```
+
+Result (MI210, FP32, NHWC)
+```
+input: dim 5, lengths {1, 256, 512, 28, 28}, strides {102760448, 401408, 1, 14336, 512}
+output: dim 2, lengths {173056, 4608}, strides {4608, 1}
+....
+Best configuration parameters:
+name: DeviceImageToColumn<128, 32, 64, 4>
+avg_time: 3.12326
+GB/s: 2042.59
+```

profiler/include/profiler/profile_pool2d_fwd_impl.hpp → profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp

@@ -6,41 +6,60 @@
 #include <iomanip>
 
 #include "ck/ck.hpp"
-#include "ck/library/tensor_operation_instance/gpu/pool2d_fwd.hpp"
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+#include "ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp"
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/utility/literals.hpp"
-#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp"
 
 namespace ck {
 namespace profiler {
 
-template <typename InDataType,
-          typename OutDataType,
+template <typename TensorLayout>
+std::vector<ck::index_t> f_tensor_strides_ncdhw(ck::index_t N_,
+                                                ck::index_t C_,
+                                                ck::index_t D,
+                                                ck::index_t H,
+                                                ck::index_t W,
+                                                TensorLayout layout)
+{
+    using namespace ck::literals;
+    (void)N_;
+    if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NDHWC>::value)
+        return {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_};
+    else
+        throw std::runtime_error("not supported yet");
+};
+
+template <typename DOutDataType,
+          typename DInDataType,
           typename ComputeDataType,
-          typename IndexDataType,
-          ck::ReduceTensorOp ReduceOpId,
-          bool PropagateNan,
-          bool OutputIndex>
-bool profile_pool2d_fwd_impl(int do_verification,
+          typename DOutLayout,
+          typename DInLayout>
+bool profile_avg_pool3d_bwd_impl(int do_verification,
                                  int init_method,
                                  bool do_log,
                                  bool time_kernel,
-                             std::vector<index_t> in_length, // NCHW
+                                 std::vector<index_t> in_length, // NCDHW
                                  std::vector<index_t> window_spatial_lengths,
                                  std::vector<index_t> window_strides,
+                                 std::vector<index_t> window_dilations,
                                  std::vector<index_t> input_left_pads,
                                  std::vector<index_t> input_right_pads)
 {
-    constexpr index_t InOutRank  = 4;
-    constexpr index_t WindowRank = 2;
+    constexpr index_t InOutRank  = 5;
+    constexpr index_t WindowRank = 3;
 
     if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
-       window_strides.size() != WindowRank || input_left_pads.size() != WindowRank ||
-       input_right_pads.size() != WindowRank)
+       window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
+       input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
+    {
+        std::cout << "Parameter is incorrect" << std::endl;
         return false;
+    }
 
     std::vector<index_t> out_length(InOutRank);
 
@@ -50,57 +69,52 @@ bool profile_pool2d_fwd_impl(int do_verification,
     out_length[0] = N;
     out_length[1] = C;
 
-    // Calculate Ho, Wo
+    // Calculate Do, Ho, Wo
     for(int i = 2; i < InOutRank; ++i)
     {
         auto pad1             = input_left_pads[i - 2];
         auto pad2             = input_right_pads[i - 2];
         auto windows_size     = window_spatial_lengths[i - 2];
         auto windows_stride   = window_strides[i - 2];
-        out_length[i]       = (in_length[i] + pad1 + pad2 - windows_size) / windows_stride + 1;
+        auto windows_dilation = window_dilations[i - 2];
+        auto eff              = (windows_size - 1) * windows_dilation + 1;
+        out_length[i]         = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
     }
 
-    int Hi = in_length[2];
-    int Wi = in_length[3];
-    int Ho = out_length[2];
-    int Wo = out_length[3];
+    int Di = in_length[2];
+    int Hi = in_length[3];
+    int Wi = in_length[4];
+    int Do = out_length[2];
+    int Ho = out_length[3];
+    int Wo = out_length[4];
 
     auto f_host_tensor_descriptor =
-        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W) {
+        [](std::size_t N_, std::size_t C_, std::size_t D, std::size_t H, std::size_t W) {
             using namespace ck::literals;
-            return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
-        };
 
-    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi));
-    Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo));
-    Tensor<IndexDataType> out_indices_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo));
+            return HostTensorDescriptor({N_, C_, D, H, W},
+                                        {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
+        };
 
-    Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo));
-    Tensor<IndexDataType> out_indices_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo));
+    Tensor<DOutDataType> dout_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<DInDataType> din_n_c_di_hi_wi_device(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+    Tensor<DInDataType> din_n_c_di_hi_wi_host(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
 
     switch(init_method)
     {
-    case 0: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{}); break;
-    case 1: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
-    default: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+    case 0: dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_1<DOutDataType>{}); break;
+    case 1: dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_2<DOutDataType>{-5, 5}); break;
+    default: dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_3<DOutDataType>{-0.5, 0.5});
     }
 
-    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
-    DeviceMem out_device_buf(sizeof(OutDataType) *
-                             out_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
-    DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
-                                     out_indices_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
+    DeviceMem dout_device_buf(sizeof(DOutDataType) * dout_n_c_do_ho_wo.mDesc.GetElementSpaceSize());
+    DeviceMem din_device_buf(sizeof(DInDataType) *
+                             din_n_c_di_hi_wi_device.mDesc.GetElementSpaceSize());
 
-    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    dout_device_buf.ToDevice(dout_n_c_do_ho_wo.mData.data());
 
-    // add device normalization instances
-    using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
-                                                                 WindowRank,
-                                                                 InDataType,
-                                                                 OutDataType,
-                                                                 IndexDataType,
-                                                                 ReduceOpId,
-                                                                 OutputIndex>;
+    using DeviceOp = ck::tensor_operation::device::
+        DeviceAvgPoolBwd<3, DOutDataType, DInDataType, DOutLayout, DInLayout>;
 
     // get device op instances
     const auto instance_ptrs =
@@ -115,26 +129,20 @@ bool profile_pool2d_fwd_impl(int do_verification,
 
     if(do_verification)
     {
-        using ReferenceInstance = ck::tensor_operation::host::ReferencePoolingFwd<InOutRank,
-                                                                                  WindowRank,
-                                                                                  InDataType,
-                                                                                  OutDataType,
-                                                                                  ComputeDataType,
-                                                                                  IndexDataType,
-                                                                                  ReduceOpId,
-                                                                                  PropagateNan,
-                                                                                  OutputIndex>;
-
-        ReferenceInstance ref;
-        auto ref_argument = ref.MakeArgument(in_n_c_hi_wi,
-                                             out_n_c_ho_wo_host,
-                                             out_indices_n_c_ho_wo_host,
+        using ReferencePoolingBwdInstance =
+            ck::tensor_operation::host::ReferenceAvgPoolBwd<3, DInDataType, DOutDataType>;
+
+        ReferencePoolingBwdInstance ref_pooling_bwd;
+        auto ref_pooling_bwd_argument = ref_pooling_bwd.MakeArgument(din_n_c_di_hi_wi_host,
+                                                                     dout_n_c_do_ho_wo,
                                                                      window_spatial_lengths,
                                                                      window_strides,
+                                                                     window_dilations,
                                                                      input_left_pads,
                                                                      input_right_pads);
-        auto ref_invoker  = ref.MakeInvoker();
-        ref_invoker.Run(ref_argument);
+
+        auto ref_invoker = ref_pooling_bwd.MakeInvoker();
+        ref_invoker.Run(ref_pooling_bwd_argument);
     }
 
     int num_kernel = 0;
@@ -142,19 +150,17 @@ bool profile_pool2d_fwd_impl(int do_verification,
     for(auto& inst_ptr : instance_ptrs)
     {
         auto argument_ptr = inst_ptr->MakeArgumentPointer(
-            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
-            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
-            static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
-            in_length,
+            static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            {N, C, Do, Ho, Wo},
+            {N, C, Di, Hi, Wi},
+            f_tensor_strides_ncdhw(N, C, Do, Ho, Wo, DOutLayout{}),
+            f_tensor_strides_ncdhw(N, C, Di, Hi, Wi, DInLayout{}),
             window_spatial_lengths,
-            out_length,
-            {C * Hi * Wi, 1, Wi * C, C},
-            {C * Ho * Wo, 1, Wo * C, C},
-            {C * Ho * Wo, 1, Wo * C, C},
             window_strides,
+            window_dilations,
             input_left_pads,
-            input_right_pads,
-            {2, 3});
+            input_right_pads);
 
         if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
         {
@@ -165,21 +171,20 @@ bool profile_pool2d_fwd_impl(int do_verification,
             if(time_kernel)
             {
                 std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
-                LogRange(std::cout << "input lengths = ", in_length, ", ") << std::endl;
+                LogRange(std::cout << "doutput lengths = ", out_length, ", ") << std::endl;
             }
 
             continue;
         }
 
-        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+        din_device_buf.SetZero();
 
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
         float avg_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
-        std::size_t num_bytes = in_n_c_hi_wi.mDesc.GetElementSize() * sizeof(InDataType) +
-                                out_n_c_ho_wo_host.mDesc.GetElementSize() * sizeof(OutDataType);
-
-        if constexpr(OutputIndex)
-            num_bytes += out_indices_n_c_ho_wo_host.mDesc.GetElementSize() * sizeof(IndexDataType);
+        std::size_t num_bytes =
+            dout_n_c_do_ho_wo.mDesc.GetElementSize() * sizeof(DOutDataType) +
+            din_n_c_di_hi_wi_device.mDesc.GetElementSize() * sizeof(DInDataType);
 
         float gb_per_sec = num_bytes / 1.E6 / avg_time;
 
@@ -196,44 +201,28 @@ bool profile_pool2d_fwd_impl(int do_verification,
 
         if(do_verification)
         {
-            out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
-
-            bool pass = ck::utils::check_err(out_n_c_ho_wo_device.mData,
-                                             out_n_c_ho_wo_host.mData,
+            din_device_buf.FromDevice(din_n_c_di_hi_wi_device.mData.data());
+            bool pass = ck::utils::check_err(din_n_c_di_hi_wi_device.mData,
+                                             din_n_c_di_hi_wi_host.mData,
                                              "Error: Incorrect results",
                                              1e-3,
                                              1e-3);
 
-            if constexpr(OutputIndex)
-            {
-                out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
-
-                pass = pass && ck::utils::check_err(out_indices_n_c_ho_wo_device,
-                                                    out_indices_n_c_ho_wo_host);
-            }
-
             if(do_log)
             {
-                LogRangeAsType<float>(std::cout << "in_n_c_hi_wi  : ", in_n_c_hi_wi.mData, ",")
-                    << std::endl;
                 LogRangeAsType<float>(
-                    std::cout << "out_n_c_ho_wo_host  : ", out_n_c_ho_wo_host.mData, ",")
-                    << std::endl;
-                LogRangeAsType<float>(
-                    std::cout << "out_n_c_ho_wo_device  : ", out_n_c_ho_wo_device.mData, ",")
+                    std::cout << "din_n_c_di_hi_wi_device: ", din_n_c_di_hi_wi_device.mData, ",")
                     << std::endl;
 
-                if constexpr(OutputIndex)
-                    LogRangeAsType<float>(std::cout << "out_indices_n_c_ho_wo_device  : ",
-                                          out_indices_n_c_ho_wo_device.mData,
-                                          ",")
+                LogRangeAsType<float>(
+                    std::cout << "din_n_c_di_hi_wi_host: ", din_n_c_di_hi_wi_host.mData, ",")
                     << std::endl;
             }
 
             if(!pass)
             {
                 std::cout << inst_ptr->GetTypeString() << " failed verification: ";
-                LogRange(std::cout << "lengths = [", in_length, ", ") << "]." << std::endl;
+                LogRange(std::cout << "doutput lengths = [", out_length, ", ") << "]." << std::endl;
                 return false;
             }
             else
@@ -246,7 +235,7 @@ bool profile_pool2d_fwd_impl(int do_verification,
 
     if(time_kernel)
     {
-        LogRange(std::cout << "length = ", in_length, ",") << std::endl;
+        LogRange(std::cout << "length = ", out_length, ",") << std::endl;
         std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
                   << best_instance_name << std::endl;
     }

profiler/include/profiler/profile_gemm_multiply_add_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout>
+bool profile_gemm_multiply_add_impl(int do_verification,
+                                    int init_method,
+                                    bool /*do_log*/,
+                                    bool time_kernel,
+                                    int M,
+                                    int N,
+                                    int K,
+                                    int StrideA,
+                                    int StrideB,
+                                    int StrideD0,
+                                    int StrideD1,
+                                    int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-1, 1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = MultiplyAdd;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                          BLayout,
+                                                          ck::Tuple<D0Layout, D1Layout>,
+                                                          ELayout,
+                                                          ADataType,
+                                                          BDataType,
+                                                          ck::Tuple<D0DataType, D1DataType>,
+                                                          EDataType,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profiler/profile_gemm_splitk_impl.hpp

@@ -94,7 +94,6 @@ bool profile_gemm_splitk_impl(int do_verification,
 
     a_device_buf.ToDevice(a_m_k.mData.data());
     b_device_buf.ToDevice(b_k_n.mData.data());
-    c_device_buf.SetZero();
 
     using DeviceOp = ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
                                                                     BLayout,
@@ -136,10 +135,23 @@ bool profile_gemm_splitk_impl(int do_verification,
     float best_ave_time   = 0;
     float best_tflops     = 0;
     float best_gb_per_sec = 0;
+    float best_kbatch     = 0;
 
     // profile device GEMM instances
     for(auto& op_ptr : op_ptrs)
     {
+        std::vector<int> kbatch_list = {1,  2,  4,  8,  12, 16,  20,  24,  32,  36,  40, 60,
+                                        64, 72, 80, 88, 96, 128, 144, 160, 176, 192, 256};
+
+        if(KBatch > 0)
+        {
+            kbatch_list = {KBatch};
+        }
+
+        for(std::size_t i = 0; i < kbatch_list.size(); i++)
+        {
+            auto kbatch_curr = kbatch_list[i];
+
             auto argument_ptr =
                 op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
                                             static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
@@ -153,15 +165,37 @@ bool profile_gemm_splitk_impl(int do_verification,
                                             a_element_op,
                                             b_element_op,
                                             c_element_op,
-                                        KBatch);
+                                            kbatch_curr);
 
             auto invoker_ptr = op_ptr->MakeInvokerPointer();
 
             if(op_ptr->IsSupportedArgument(argument_ptr.get()))
             {
+
                 // re-init C to zero before profiling next kernel
                 c_device_buf.SetZero();
 
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+
+                if(do_verification)
+                {
+                    c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+                    pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+
+                    if(do_log)
+                    {
+                        LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                        LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                            << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                            << std::endl;
+                    }
+                }
+
                 std::string op_name = op_ptr->GetTypeString();
 
                 float ave_time =
@@ -169,44 +203,46 @@ bool profile_gemm_splitk_impl(int do_verification,
 
                 std::size_t flop = std::size_t(2) * M * N * K;
 
-            std::size_t num_btype =
-                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+                std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                        sizeof(CDataType) * M * N;
 
                 float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
 
                 float gb_per_sec = num_btype / 1.E6 / ave_time;
 
-            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
-                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+                std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
+                          << " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
+                          << kbatch_curr << std::endl;
 
-            if(tflops > best_tflops)
+                // set softer tolerances for fp8
+                if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
+                             is_same_v<CDataType, f8_t>)
                 {
-                best_op_name    = op_name;
-                best_tflops     = tflops;
-                best_ave_time   = ave_time;
-                best_gb_per_sec = gb_per_sec;
+                    std::string msg = "Error: Incorrect results!";
+                    double rtol     = 1e-1;
+                    double atol     = 1e-1;
+                    pass            = pass & ck::utils::check_err(
+                                      c_m_n_device_result, c_m_n_host_result, msg, rtol, atol);
                 }
-
-            if(do_verification)
+                else
                 {
-                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
-
                     pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+                }
 
-                if(do_log)
+                if(tflops > best_tflops)
                 {
-                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
-                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
-                    LogRangeAsType<float>(std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
-                        << std::endl;
-                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
-                        << std::endl;
-                }
+                    best_op_name    = op_name;
+                    best_tflops     = tflops;
+                    best_ave_time   = ave_time;
+                    best_gb_per_sec = gb_per_sec;
+                    best_kbatch     = kbatch_curr;
                 }
             }
             else
             {
-            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+                std::cout << op_ptr->GetTypeString() << " does not support this problem"
+                          << std::endl;
+            }
         }
     }
 
@@ -246,7 +282,7 @@ bool profile_gemm_splitk_impl(int do_verification,
     }
 
     std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
-              << " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << KBatch
+              << " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << best_kbatch
               << " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
               << " GB/s, " << best_op_name << std::endl;
 

profiler/include/profiler/profile_gemm_streamk_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_streamk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_gemm_streamk_impl(int do_verification,
+                               int init_method,
+                               bool do_log,
+                               bool time_kernel,
+                               int M,
+                               int N,
+                               int K,
+                               int StrideA,
+                               int StrideB,
+                               int StrideC,
+                               uint32_t NumSKBlocks = 0xffffffff)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-3, 3});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    c_device_buf.ToDevice(c_m_n_device_result.mData.data());
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmStreamK<ALayout,
+                                                                     BLayout,
+                                                                     CLayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances, "
+              << (do_verification ? "with verification" : "without verification") << std::endl;
+
+    // Run reference GEMM
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device GEMM instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                        static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                        M,
+                                        N,
+                                        K,
+                                        StrideA,
+                                        StrideB,
+                                        StrideC,
+                                        a_element_op,
+                                        b_element_op,
+                                        c_element_op,
+                                        NumSKBlocks);
+        DeviceMem workspace;
+        std::size_t workspace_size = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+        if(workspace_size != 0)
+        {
+            workspace.Realloc(workspace_size);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+        }
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init C to zero before profiling next kernel
+            c_device_buf.SetZero();
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    if constexpr(is_same<CDataType, float>::value)
+    {
+        std::cout << "Best Perf for datatype = f32";
+    }
+    else if constexpr(is_same<CDataType, half_t>::value)
+    {
+        std::cout << "Best Perf for datatype = f16";
+    }
+    else if constexpr(is_same<CDataType, bhalf_t>::value)
+    {
+        std::cout << "Best Perf for datatype = bf16";
+    }
+    else if constexpr(is_same<CDataType, int8_t>::value)
+    {
+        std::cout << "Best Perf for datatype = int8";
+    }
+
+    if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " ALayout =  RowMajor";
+    }
+    else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " ALayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " BLayout =  RowMajor";
+    }
+    else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " BLayout =  ColumnMajor";
+    }
+
+    std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
+              << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_ave_time
+              << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp

@@ -136,10 +136,11 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
     // profile device Conv instances
     bool all_pass = true;
 
-    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> input_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> filter_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> output_lengths{};
     std::array<ck::index_t, NDimSpatial + 3> input_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> weights_strides{};
     std::array<ck::index_t, NDimSpatial + 3> output_strides{};
     std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
     std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
@@ -148,10 +149,11 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
 
     auto range_copy = [](const auto& from, auto to) { std::copy(begin(from), end(from), to); };
 
-    range_copy(conv_param.input_spatial_lengths_, begin(input_spatial_lengths));
-    range_copy(conv_param.filter_spatial_lengths_, begin(filter_spatial_lengths));
-    range_copy(conv_param.output_spatial_lengths_, begin(output_spatial_lengths));
+    range_copy(in_g_n_c_wis_desc.GetLengths(), begin(input_lengths));
     range_copy(in_g_n_c_wis_desc.GetStrides(), begin(input_strides));
+    range_copy(wei_g_k_c_xs_desc.GetLengths(), begin(filter_lengths));
+    range_copy(wei_g_k_c_xs_desc.GetStrides(), begin(weights_strides));
+    range_copy(out_g_n_k_wos_desc.GetLengths(), begin(output_lengths));
     range_copy(out_g_n_k_wos_desc.GetStrides(), begin(output_strides));
     range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
     range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
@@ -164,14 +166,11 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
             op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
                                         static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
                                         static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
-                                        conv_param.G_,
-                                        conv_param.N_,
-                                        conv_param.K_,
-                                        conv_param.C_,
-                                        input_spatial_lengths,
-                                        filter_spatial_lengths,
-                                        output_spatial_lengths,
+                                        input_lengths,
                                         input_strides,
+                                        filter_lengths,
+                                        weights_strides,
+                                        output_lengths,
                                         output_strides,
                                         conv_filter_strides,
                                         conv_filter_dilations,

profiler/include/profiler/profile_grouped_gemm_impl.hpp

@@ -70,6 +70,7 @@ bool profile_grouped_gemm_impl(int do_verification,
 
     std::vector<Tensor<ADataType>> a_m_k;
     std::vector<Tensor<BDataType>> b_k_n;
+    std::vector<Tensor<CDataType>> c_m_n_host_results;
     std::vector<Tensor<CDataType>> c_m_n_device_results;
 
     for(std::size_t i = 0; i < group_count; i++)
@@ -81,6 +82,9 @@ bool profile_grouped_gemm_impl(int do_verification,
 
         c_m_n_device_results.push_back(
             Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
+
+        c_m_n_host_results.push_back(
+            Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
 #if DEBUG_LOG
         std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
                   << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
@@ -137,7 +141,6 @@ bool profile_grouped_gemm_impl(int do_verification,
 
         a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
         b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
-        c_device_buf[i]->SetZero();
 
         gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
 
@@ -170,9 +173,36 @@ bool profile_grouped_gemm_impl(int do_verification,
     float best_ave_time   = 0;
     float best_tflops     = 0;
     float best_gb_per_sec = 0;
+    float best_kbatch     = 0;
 
     auto p_ds = std::vector<std::array<const void*, 0>>{};
 
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                    BDataType,
+                                                                                    CDataType,
+                                                                                    AccDataType,
+                                                                                    AElementOp,
+                                                                                    BElementOp,
+                                                                                    CElementOp>;
+
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
+                                                      b_k_n[i],
+                                                      c_m_n_host_results[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+        }
+    }
+
     // profile device GEMM instances
     for(auto& gemm_ptr : op_ptrs)
     {
@@ -193,10 +223,7 @@ bool profile_grouped_gemm_impl(int do_verification,
         gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
         std::string gemm_name = gemm_ptr->GetTypeString();
 
-        if(kbatch > 1)
-        {
-            using DeviceOpSplitK =
-                ck::tensor_operation::device::DeviceGroupedGemmSplitK<ALayout,
+        using DeviceOpSplitK = ck::tensor_operation::device::DeviceGroupedGemmSplitK<ALayout,
                                                                                      BLayout,
                                                                                      ck::Tuple<>,
                                                                                      CLayout,
@@ -208,45 +235,33 @@ bool profile_grouped_gemm_impl(int do_verification,
                                                                                      BElementOp,
                                                                                      CElementOp>;
 
-            if(dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get()) != nullptr)
+        // skip non-splitk grouped_gemm
+        if(dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get()) == nullptr)
         {
-                dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
-                    ->SetKBatchSize(argument_ptr.get(), kbatch);
-            }
+            continue;
         }
 
-        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
-        {
+        std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 48, 64};
 
-            float ave_time =
-                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
-
-            if(time_kernel)
+        if(kbatch > 0)
         {
-                std::size_t flop = 0, num_btype = 0;
-                for(std::size_t i = 0; i < gemm_descs.size(); i++)
-                {
-                    flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
-
-                    num_btype += sizeof(ADataType) * Ms[i] * Ks[i] +
-                                 sizeof(BDataType) * Ks[i] * Ns[i] +
-                                 sizeof(CDataType) * Ms[i] * Ns[i];
+            kbatch_list = {kbatch};
         }
 
-                float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+        for(std::size_t j = 0; j < kbatch_list.size(); j++)
+        {
 
-                float gb_per_sec = num_btype / 1.E6 / ave_time;
-                std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
-                          << " TFlops, " << gb_per_sec << " GB/s, " << gemm_name << std::endl;
+            auto kbatch_curr = kbatch_list[j];
 
-                if(tflops > best_tflops)
+            dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
+                ->SetKBatchSize(argument_ptr.get(), kbatch_curr);
+
+            if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
             {
-                    best_gemm_name  = gemm_name;
-                    best_tflops     = tflops;
-                    best_ave_time   = ave_time;
-                    best_gb_per_sec = gb_per_sec;
-                }
-            }
+                for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                    c_device_buf[i]->SetZero();
+
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
 
                 if(do_verification)
                 {
@@ -255,56 +270,33 @@ bool profile_grouped_gemm_impl(int do_verification,
                     {
 
                         c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
-                    c_device_buf[i]->SetZero();
-
-                    Tensor<CDataType> c_m_n_host_result(
-                        f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{}));
-
-                    using ReferenceGemmInstance =
-                        ck::tensor_operation::host::ReferenceGemm<ADataType,
-                                                                  BDataType,
-                                                                  CDataType,
-                                                                  AccDataType,
-                                                                  AElementOp,
-                                                                  BElementOp,
-                                                                  CElementOp>;
-
-                    auto ref_gemm    = ReferenceGemmInstance{};
-                    auto ref_invoker = ref_gemm.MakeInvoker();
-
-                    auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
-                                                              b_k_n[i],
-                                                              c_m_n_host_result,
-                                                              a_element_op,
-                                                              b_element_op,
-                                                              c_element_op);
 
-                    ref_invoker.Run(ref_argument);
-                    if(std::is_same_v<CDataType, ck::half_t> && kbatch > 1)
+                        if(std::is_same_v<CDataType, ck::half_t> && kbatch_curr > 1)
                         {
                             instance_pass =
                                 instance_pass && ck::utils::check_err(c_m_n_device_results[i],
-                                                                  c_m_n_host_result,
+                                                                      c_m_n_host_results[i],
                                                                       "Error: Incorrect results!",
                                                                       0.06);
                         }
                         else
                         {
                             instance_pass =
-                            instance_pass &&
-                            ck::utils::check_err(c_m_n_device_results[i], c_m_n_host_result);
+                                instance_pass && ck::utils::check_err(c_m_n_device_results[i],
+                                                                      c_m_n_host_results[i]);
                         }
 
                         if(do_log)
                         {
                             LogRangeAsType<float>(std::cout << "a : ", a_m_k[i].mData, ",")
                                 << std::endl;
-                        LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",") << std::endl;
+                            LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",")
+                                << std::endl;
                             LogRangeAsType<float>(
                                 std::cout << "c_device: ", c_m_n_device_results[i].mData, ",")
                                 << std::endl;
                             LogRangeAsType<float>(
-                            std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                                std::cout << "c_host  : ", c_m_n_host_results[i].mData, ",")
                                 << std::endl;
                         }
                     }
@@ -314,6 +306,38 @@ bool profile_grouped_gemm_impl(int do_verification,
 
                     pass = pass && instance_pass;
                 }
+
+                float ave_time =
+                    invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+                if(time_kernel)
+                {
+                    std::size_t flop = 0, num_btype = 0;
+                    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                    {
+                        flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
+
+                        num_btype += sizeof(ADataType) * Ms[i] * Ks[i] +
+                                     sizeof(BDataType) * Ks[i] * Ns[i] +
+                                     sizeof(CDataType) * Ms[i] * Ns[i];
+                    }
+
+                    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+                    float gb_per_sec = num_btype / 1.E6 / ave_time;
+                    std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
+                              << " TFlops, " << gb_per_sec << " GB/s, " << gemm_name << ", KBatch "
+                              << kbatch_curr << std::endl;
+
+                    if(tflops > best_tflops)
+                    {
+                        best_gemm_name  = gemm_name;
+                        best_tflops     = tflops;
+                        best_ave_time   = ave_time;
+                        best_gb_per_sec = gb_per_sec;
+                        best_kbatch     = kbatch_curr;
+                    }
+                }
             }
             else
             {
@@ -321,11 +345,13 @@ bool profile_grouped_gemm_impl(int do_verification,
                           << std::endl;
             }
         }
+    }
 
     if(time_kernel)
     {
         std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
-                  << best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
+                  << best_gb_per_sec << " GB/s, " << best_gemm_name << ", KBatch = " << best_kbatch
+                  << std::endl;
     }
 
     return pass;

profiler/include/profiler/profile_groupnorm_impl.hpp

@@ -139,6 +139,10 @@ bool profile_groupnorm_impl(int do_verification,
             continue;
         }
 
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+        DeviceMem workspace_dev(workspace_sz);
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
         auto invoker_ptr = inst_ptr->MakeInvokerPointer();
 
         float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});

profiler/include/profiler/profile_image_to_column_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+#include <limits>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_image_to_column.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp"
+#include "ck/library/tensor_operation_instance/gpu/image_to_column.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+template <index_t NDimSpatial,
+          typename InputLayout,
+          typename InputDataType,
+          typename OutputDataType>
+bool profile_image_to_column_impl(int do_verification,
+                                  int init_method,
+                                  bool do_log,
+                                  bool time_kernel,
+                                  const ck::utils::conv::ConvParam& conv_param)
+{
+    const ck::index_t NDoHoWo =
+        conv_param.N_ *
+        ck::accumulate_n<ck::index_t>(
+            conv_param.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+    const ck::index_t CZYX =
+        conv_param.C_ *
+        ck::accumulate_n<ck::index_t>(
+            conv_param.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+
+    const auto in_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InputLayout>(
+            conv_param);
+    const auto out_desc = HostTensorDescriptor({NDoHoWo, CZYX});
+
+    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> input_g_n_c_wis_strides{};
+    std::array<ck::index_t, 2> output_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_param.input_spatial_lengths_, input_spatial_lengths);
+    copy(conv_param.filter_spatial_lengths_, filter_spatial_lengths);
+    copy(conv_param.output_spatial_lengths_, output_spatial_lengths);
+    copy(in_desc.GetStrides(), input_g_n_c_wis_strides);
+    copy(out_desc.GetStrides(), output_m_k_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    Tensor<InputDataType> input(in_desc);
+    Tensor<OutputDataType> host_output(out_desc);
+    Tensor<OutputDataType> device_output(out_desc);
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "output: " << host_output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1: input.GenerateTensorValue(GeneratorTensor_2<InputDataType>{-5, 5}); break;
+    default: input.GenerateTensorValue(GeneratorTensor_3<InputDataType>{0.0, 1.0});
+    }
+
+    DeviceMem in_device_buf(sizeof(InputDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutputDataType) * device_output.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(input.mData.data());
+
+    // run reference op
+    if(do_verification)
+    {
+        auto ref_image_to_column = ck::tensor_operation::host::
+            ReferenceImageToColumn<NDimSpatial, InputLayout, InputDataType, OutputDataType>{};
+
+        auto ref_invoker  = ref_image_to_column.MakeInvoker();
+        auto ref_argument = ref_image_to_column.MakeArgument(input,
+                                                             host_output,
+                                                             conv_param.filter_spatial_lengths_,
+                                                             conv_param.conv_filter_strides_,
+                                                             conv_param.conv_filter_dilations_,
+                                                             conv_param.input_left_pads_,
+                                                             conv_param.input_right_pads_);
+
+        // init host output to zero
+        host_output.SetZero();
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    using DeviceOp = ck::tensor_operation::device::
+        DeviceImageToColumn<NDimSpatial, InputLayout, InputDataType, OutputDataType>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    bool pass                   = true;
+    bool is_supporting_instance = false;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<InputDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<OutputDataType*>(out_device_buf.GetDeviceBuffer()),
+            conv_param.N_,
+            conv_param.C_,
+            input_spatial_lengths,
+            filter_spatial_lengths,
+            output_spatial_lengths,
+            input_g_n_c_wis_strides,
+            output_m_k_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            is_supporting_instance = true;
+            // re-init output to zero before profiling next kernel
+            out_device_buf.SetZero();
+            std::string op_name = op_ptr->GetTypeString();
+            auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+            std::size_t num_btype =
+                NDoHoWo * CZYX * (sizeof(OutputDataType) + sizeof(InputDataType));
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(avg_time < best_avg_time)
+            {
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(device_output.mData.data());
+                pass = pass & ck::utils::check_err(device_output, host_output);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "input : ", input.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "host_output  : ", host_output.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "device_output: ", device_output.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return is_supporting_instance && pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profiler/profile_layernorm_impl.hpp

@@ -155,6 +155,10 @@ bool profile_layernorm_impl(int do_verification,
             continue;
         }
 
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+        DeviceMem workspace_dev(workspace_sz);
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
         auto invoker_ptr = inst_ptr->MakeInvokerPointer();
 
         float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});

profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+#include "ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          typename DOutDataType,
+          typename DInDataType,
+          bool PropagateNan>
+bool profile_max_pool3d_bwd_impl(int do_verification,
+                                 int init_method,
+                                 bool do_log,
+                                 bool time_kernel,
+                                 std::vector<index_t> in_length, // NCDHW
+                                 std::vector<index_t> window_spatial_lengths,
+                                 std::vector<index_t> window_strides,
+                                 std::vector<index_t> window_dilations,
+                                 std::vector<index_t> input_left_pads,
+                                 std::vector<index_t> input_right_pads)
+{
+    // AtomicAdd only support f32 for now. ComputeDataType must be float32
+    using ComputeDataType = float;
+
+    constexpr index_t InOutRank  = 5;
+    constexpr index_t WindowRank = 3;
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
+       window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
+       input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
+    {
+        std::cout << "Parameter is incorrect" << std::endl;
+        return false;
+    }
+
+    std::vector<index_t> out_length(InOutRank);
+
+    int N = in_length[0];
+    int C = in_length[1];
+
+    out_length[0] = N;
+    out_length[1] = C;
+
+    // Calculate Do, Ho, Wo
+    for(int i = 2; i < InOutRank; ++i)
+    {
+        auto pad1             = input_left_pads[i - 2];
+        auto pad2             = input_right_pads[i - 2];
+        auto windows_size     = window_spatial_lengths[i - 2];
+        auto windows_stride   = window_strides[i - 2];
+        auto windows_dilation = window_dilations[i - 2];
+        auto eff              = (windows_size - 1) * windows_dilation + 1;
+        out_length[i]         = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
+    }
+
+    int Di = in_length[2];
+    int Hi = in_length[3];
+    int Wi = in_length[4];
+    int Do = out_length[2];
+    int Ho = out_length[3];
+    int Wo = out_length[4];
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t D, std::size_t H, std::size_t W) {
+            using namespace ck::literals;
+
+            return HostTensorDescriptor({N_, C_, D, H, W},
+                                        {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
+        };
+
+    Tensor<InDataType> in_n_c_di_hi_wi(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+    Tensor<OutDataType> out_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<IndexDataType> out_indices_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<DOutDataType> dout_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<DInDataType> din_n_c_di_hi_wi_host(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+
+    Tensor<DInDataType> din_n_c_di_hi_wi_device(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+
+    switch(init_method)
+    {
+    case 0:
+        in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
+        dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_1<DOutDataType>{});
+        break;
+    case 1:
+        in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_2<DOutDataType>{-5, 5});
+        break;
+    default:
+        in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+        dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_3<DOutDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem indices_device_buf(sizeof(IndexDataType) *
+                                 out_indices_n_c_do_ho_wo.mDesc.GetElementSpaceSize());
+    DeviceMem dout_device_buf(sizeof(DOutDataType) * dout_n_c_do_ho_wo.mDesc.GetElementSpaceSize());
+    DeviceMem din_device_buf(sizeof(DInDataType) *
+                             din_n_c_di_hi_wi_device.mDesc.GetElementSpaceSize());
+
+    // Generate index data from forwarding
+    {
+        using ReferencePoolingFwdInstance =
+            ck::tensor_operation::host::ReferencePoolingFwd<InOutRank,
+                                                            WindowRank,
+                                                            InDataType,
+                                                            OutDataType,
+                                                            ComputeDataType,
+                                                            IndexDataType,
+                                                            ck::ReduceTensorOp::MAX,
+                                                            false,
+                                                            true>;
+
+        ReferencePoolingFwdInstance ref_pooling_fwd;
+        auto ref_pooling_fwd_argument = ref_pooling_fwd.MakeArgument(in_n_c_di_hi_wi,
+                                                                     out_n_c_do_ho_wo,
+                                                                     out_indices_n_c_do_ho_wo,
+                                                                     window_spatial_lengths,
+                                                                     window_strides,
+                                                                     window_dilations,
+                                                                     input_left_pads,
+                                                                     input_right_pads);
+        auto ref_pooling_fwd_invoker  = ref_pooling_fwd.MakeInvoker();
+        ref_pooling_fwd_invoker.Run(ref_pooling_fwd_argument);
+    }
+
+    indices_device_buf.ToDevice(out_indices_n_c_do_ho_wo.mData.data());
+    dout_device_buf.ToDevice(dout_n_c_do_ho_wo.mData.data());
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferencePoolingBwdInstance =
+            ck::tensor_operation::host::ReferenceMaxPoolBwd<DOutDataType,
+                                                            IndexDataType,
+                                                            ComputeDataType,
+                                                            DInDataType,
+                                                            PassThrough>;
+
+        ReferencePoolingBwdInstance ref_pooling_bwd;
+        auto ref_pooling_bwd_argument = ref_pooling_bwd.MakeArgument(
+            dout_n_c_do_ho_wo, out_indices_n_c_do_ho_wo, din_n_c_di_hi_wi_host, PassThrough{});
+        auto ref_invoker = ref_pooling_bwd.MakeInvoker();
+        ref_invoker.Run(ref_pooling_bwd_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            dout_n_c_do_ho_wo.mDesc.GetElementSpaceSize(),
+            din_n_c_di_hi_wi_device.mDesc.GetElementSpaceSize(),
+            window_spatial_lengths,
+            window_strides,
+            window_dilations);
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+                LogRange(std::cout << "doutput lengths = ", out_length, ", ") << std::endl;
+            }
+
+            continue;
+        }
+
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+        DeviceMem workspace_device_buf(workspace_sz);
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_device_buf.GetDeviceBuffer());
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+        float avg_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes =
+            dout_n_c_do_ho_wo.mDesc.GetElementSize() * sizeof(DOutDataType) +
+            out_indices_n_c_do_ho_wo.mDesc.GetElementSize() * sizeof(IndexDataType) +
+            din_n_c_di_hi_wi_device.mDesc.GetElementSize() * sizeof(DInDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            din_device_buf.FromDevice(din_n_c_di_hi_wi_device.mData.data());
+
+            bool pass = ck::utils::check_err(din_n_c_di_hi_wi_device.mData,
+                                             din_n_c_di_hi_wi_host.mData,
+                                             "Error: Incorrect results",
+                                             1e-3,
+                                             1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(
+                    std::cout << "out_indices_n_c_do_ho_wo: ", out_indices_n_c_do_ho_wo.mData, ",")
+                    << std::endl;
+
+                LogRangeAsType<float>(
+                    std::cout << "din_n_c_di_hi_wi_device: ", din_n_c_di_hi_wi_device.mData, ",")
+                    << std::endl;
+
+                LogRangeAsType<float>(
+                    std::cout << "din_n_c_di_hi_wi_host: ", din_n_c_di_hi_wi_host.mData, ",")
+                    << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "doutput lengths = [", out_length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", out_length, ",") << std::endl;
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profiler/profile_pool3d_fwd_impl.hpp

@@ -21,6 +21,8 @@ template <typename InDataType,
           typename OutDataType,
           typename ComputeDataType,
           typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
           ck::ReduceTensorOp ReduceOpId,
           bool PropagateNan,
           bool OutputIndex>
@@ -31,6 +33,7 @@ bool profile_pool3d_fwd_impl(int do_verification,
                              std::vector<index_t> in_length, // NCDHW
                              std::vector<index_t> window_spatial_lengths,
                              std::vector<index_t> window_strides,
+                             std::vector<index_t> window_dilations,
                              std::vector<index_t> input_left_pads,
                              std::vector<index_t> input_right_pads)
 {
@@ -38,8 +41,8 @@ bool profile_pool3d_fwd_impl(int do_verification,
     constexpr index_t WindowRank = 3;
 
     if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
-       window_strides.size() != WindowRank || input_left_pads.size() != WindowRank ||
-       input_right_pads.size() != WindowRank)
+       window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
+       input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
         return false;
 
     std::vector<index_t> out_length(InOutRank);
@@ -57,7 +60,9 @@ bool profile_pool3d_fwd_impl(int do_verification,
         auto pad2             = input_right_pads[i - 2];
         auto windows_size     = window_spatial_lengths[i - 2];
         auto windows_stride   = window_strides[i - 2];
-        out_length[i]       = (in_length[i] + pad1 + pad2 - windows_size) / windows_stride + 1;
+        auto windows_dilation = window_dilations[i - 2];
+        auto eff              = (windows_size - 1) * windows_dilation + 1;
+        out_length[i]         = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
     }
 
     int Di = in_length[2];
@@ -104,6 +109,8 @@ bool profile_pool3d_fwd_impl(int do_verification,
                                                                  InDataType,
                                                                  OutDataType,
                                                                  IndexDataType,
+                                                                 InLayout,
+                                                                 OutLayout,
                                                                  ReduceOpId,
                                                                  OutputIndex>;
 
@@ -136,6 +143,7 @@ bool profile_pool3d_fwd_impl(int do_verification,
                                              out_indices_n_c_do_ho_wo_host,
                                              window_spatial_lengths,
                                              window_strides,
+                                             window_dilations,
                                              input_left_pads,
                                              input_right_pads);
         auto ref_invoker  = ref.MakeInvoker();
@@ -157,6 +165,7 @@ bool profile_pool3d_fwd_impl(int do_verification,
             {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
             {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
             window_strides,
+            window_dilations,
             input_left_pads,
             input_right_pads,
             {2, 3, 4});

profiler/src/CMakeLists.txt

@@ -3,19 +3,12 @@ set(PROFILER_SOURCES
     profiler.cpp
     profile_gemm.cpp
     profile_gemm_splitk.cpp
-    profile_gemm_bilinear.cpp
     profile_gemm_bias_add_reduce.cpp
-    profile_gemm_add_add_fastgelu.cpp
     profile_gemm_add_multiply.cpp
-    profile_gemm_add_fastgelu.cpp
-    profile_gemm_add_relu_add_layernorm.cpp
-    profile_gemm_fastgelu.cpp
+    profile_gemm_multiply_add.cpp
     profile_gemm_reduce.cpp
     profile_batched_gemm.cpp
-    profile_batched_gemm_gemm.cpp
-    profile_batched_gemm_add_relu_gemm_add.cpp
     profile_batched_gemm_reduce.cpp
-    profile_grouped_gemm.cpp
     profile_conv_fwd.cpp
     profile_conv_fwd_bias_relu.cpp
     profile_conv_fwd_bias_relu_add.cpp
@@ -25,18 +18,33 @@ set(PROFILER_SOURCES
     profile_reduce.cpp
     profile_groupnorm.cpp
     profile_layernorm.cpp
-    profile_avg_pool2d_fwd.cpp
     profile_max_pool3d_fwd.cpp
+    profile_avg_pool3d_bwd.cpp
+    profile_max_pool3d_bwd.cpp
     profile_softmax.cpp
     profile_batchnorm_fwd.cpp
     profile_batchnorm_bwd.cpp
     profile_batchnorm_infer.cpp
-    profile_grouped_gemm_fastgelu.cpp
     profile_contraction_bilinear.cpp
     profile_contraction_scale.cpp
-    profile_batched_gemm_multi_d.cpp
     profile_grouped_conv_bwd_data.cpp
+    profile_image_to_column.cpp
 )
+if(DL_KERNELS)
+  list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
+  list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
+endif()
 
 set(PROFILER_EXECUTABLE ckProfiler)
 
@@ -46,19 +54,12 @@ target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
@@ -75,10 +76,27 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_instan
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool_fwd_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
+if(DL_KERNELS)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
+endif()
 rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)

profiler/src/profile_avg_pool3d_bwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_avg_pool3d_bwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct maxPoolbwdArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}},
+                                                                   {"wsize", {}},
+                                                                   {"wstride", {}},
+                                                                   {"wdilation", {}},
+                                                                   {"pad1", {}},
+                                                                   {"pad2", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help_avg_pool3d_bwd()
+{
+    std::cout << "arg1: data type (0: fp16; 1: fp32; 5: bf16)\n"
+              << "arg2: verification (0: no; 1: yes)\n"
+              << "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg4: print tensor value (0: no; 1: yes)\n"
+              << "arg5: time kernel (0=no, 1=yes)\n"
+              << "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n"
+              << "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n"
+              << "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n"
+              << "--wdilation: window dilation for DHW (e.g, --wdilation 1 1 1) \n"
+              << "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n"
+              << "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n"
+              << "eg: ckProfiler avg_pool3d_bwd 0 1 2 0 1 --length 2 32 30 30 30 --wsize 2 2 2 "
+                 "--wstride 2 2 2 --wdilation 1 1 1 --pad1 1 1 1 --pad2 1 1 1"
+              << std::endl;
+}
+
+int profile_avg_pool3d_bwd(int argc, char* argv[])
+{
+    ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
+    bool do_verification       = true;
+    int init_method            = 0;
+    bool do_log                = false;
+    bool time_kernel           = true;
+
+    std::vector<index_t> in_length = {2, 32, 30, 30, 30};
+    std::vector<index_t> wsize     = {2, 2, 2};
+    std::vector<index_t> wstride   = {2, 2, 2};
+    std::vector<index_t> wdilation = {1, 1, 1};
+    std::vector<index_t> pad1      = {1, 1, 1};
+    std::vector<index_t> pad2      = {1, 1, 1};
+
+    if(argc != 2 && argc != 33)
+    {
+        print_help_avg_pool3d_bwd();
+        return 0;
+    }
+    else if(argc == 33)
+    {
+        data_type       = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+        do_verification = std::stoi(argv[3]);
+        init_method     = std::stoi(argv[4]);
+        do_log          = std::stoi(argv[5]);
+        time_kernel     = std::stoi(argv[6]);
+
+        // parse the long options
+        maxPoolbwdArgParser arg_parser;
+        arg_parser(argc, argv);
+        in_length = arg_parser.long_opts["length"];
+        wsize     = arg_parser.long_opts["wsize"];
+        wstride   = arg_parser.long_opts["wstride"];
+        wdilation = arg_parser.long_opts["wdilation"];
+        pad1      = arg_parser.long_opts["pad1"];
+        pad2      = arg_parser.long_opts["pad2"];
+    }
+
+#ifdef CK_ENABLE_FP16
+    using F16 = ck::half_t;
+#endif
+#ifdef CK_ENABLE_BF16
+    using BF16 = ck::bhalf_t;
+#endif
+#ifdef CK_ENABLE_FP32
+    using F32 = float;
+#endif
+    using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+    if(false)
+        ;
+#ifdef CK_ENABLE_FP16
+    else if(data_type == ck::DataTypeEnum::Half)
+    {
+        ck::profiler::profile_avg_pool3d_bwd_impl<F16, F16, F16, NDHWC, NDHWC>(do_verification,
+                                                                               init_method,
+                                                                               do_log,
+                                                                               time_kernel,
+                                                                               in_length,
+                                                                               wsize,
+                                                                               wstride,
+                                                                               wdilation,
+                                                                               pad1,
+                                                                               pad2);
+    }
+#endif
+#ifdef CK_ENABLE_BF16
+    else if(data_type == ck::DataTypeEnum::BFloat16)
+    {
+        ck::profiler::profile_avg_pool3d_bwd_impl<BF16, BF16, BF16, NDHWC, NDHWC>(do_verification,
+                                                                                  init_method,
+                                                                                  do_log,
+                                                                                  time_kernel,
+                                                                                  in_length,
+                                                                                  wsize,
+                                                                                  wstride,
+                                                                                  wdilation,
+                                                                                  pad1,
+                                                                                  pad2);
+    }
+#endif
+#ifdef CK_ENABLE_FP32
+    else if(data_type == ck::DataTypeEnum::Float)
+    {
+        ck::profiler::profile_avg_pool3d_bwd_impl<F32, F32, F32, NDHWC, NDHWC>(do_verification,
+                                                                               init_method,
+                                                                               do_log,
+                                                                               time_kernel,
+                                                                               in_length,
+                                                                               wsize,
+                                                                               wstride,
+                                                                               wdilation,
+                                                                               pad1,
+                                                                               pad2);
+    }
+#endif
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("avg_pool3d_bwd", "max_pool bwd", profile_avg_pool3d_bwd);