Merge branch 'develop' into amd-develop

client_example/11_grouped_conv_bwd_weight/CMakeLists.txt

@@ -2,8 +2,10 @@ add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_f
 add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp)
 add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp)
 add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp)
+add_executable(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp)
 
-target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
-target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
-target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
-target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
+target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
+target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_conv_operations)
+target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_conv_operations)
+target_link_libraries(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE composable_kernel::device_conv_operations)

client_example/11_grouped_conv_bwd_weight/common.hpp

@@ -85,7 +85,9 @@ template <ck::index_t NumDimSpatial,
           typename OutDataType,
           typename InLayout,
           typename WeiLayout,
-          typename OutLayout>
+          typename OutLayout,
+          typename AComputeType = InDataType,
+          typename BComputeType = AComputeType>
 bool run_grouped_conv_bwd_weight(
     const std::array<ck::index_t, NumDimSpatial + 3>& input_lengths,
     const std::array<ck::index_t, NumDimSpatial + 3>& input_strides,
@@ -113,7 +115,9 @@ bool run_grouped_conv_bwd_weight(
                                                                               OutDataType,
                                                                               PassThrough,
                                                                               PassThrough,
-                                                                              PassThrough>;
+                                                                              PassThrough,
+                                                                              AComputeType,
+                                                                              BComputeType>;
     // get device op instances
     const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
         DeviceOp>::GetInstances();

client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+using AComputeType = ck::bf8_t;
+using BComputeType = ck::f8_t;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 8;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 128;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
+    N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
+    K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
+    N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+int main()
+{
+    return run_grouped_conv_bwd_weight<NumDimSpatial,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout,
+                                       AComputeType,
+                                       BComputeType>(input_lengths,
+                                                     input_strides,
+                                                     filter_lengths,
+                                                     weights_strides,
+                                                     output_lengths,
+                                                     output_strides,
+                                                     conv_filter_strides,
+                                                     conv_filter_dilations,
+                                                     input_left_pads,
+                                                     input_right_pads)
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}

client_example/12_elementwise_normalization/CMakeLists.txt

 add_executable(client_elementwise_layernorm2d elementwise_layernorm2d.cpp)
-target_link_libraries(client_elementwise_layernorm2d PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_elementwise_layernorm2d PRIVATE composable_kernel::device_other_operations)

client_example/13_batchnorm/CMakeLists.txt

 add_executable(client_batchnorm_fwd_nhwc batchnorm_fwd_nhwc.cpp)
 add_executable(client_batchnorm_bwd_nhwc batchnorm_bwd_nhwc.cpp)
 add_executable(client_batchnorm_infer_nhwc batchnorm_infer_nhwc.cpp)
-target_link_libraries(client_batchnorm_fwd_nhwc PRIVATE composable_kernel::device_operations)
-target_link_libraries(client_batchnorm_bwd_nhwc PRIVATE composable_kernel::device_operations)
-target_link_libraries(client_batchnorm_infer_nhwc PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_batchnorm_fwd_nhwc PRIVATE composable_kernel::device_other_operations)
+target_link_libraries(client_batchnorm_bwd_nhwc PRIVATE composable_kernel::device_other_operations)
+target_link_libraries(client_batchnorm_infer_nhwc PRIVATE composable_kernel::device_other_operations)

client_example/14_instance_id/CMakeLists.txt

 add_executable(client_batchnorm_fwd_instance_id batchnorm_fwd_instance_id.cpp)
-target_link_libraries(client_batchnorm_fwd_instance_id PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_batchnorm_fwd_instance_id PRIVATE composable_kernel::device_other_operations)

client_example/15_convnd_bwd_data/CMakeLists.txt

 add_executable(client_conv3d_bwd_data_fp16 conv3d_bwd_data_fp16.cpp)
 add_executable(client_conv3d_bwd_data_fp32 conv3d_bwd_data_fp32.cpp)
 
-target_link_libraries(client_conv3d_bwd_data_fp16 PRIVATE composable_kernel::device_operations)
-target_link_libraries(client_conv3d_bwd_data_fp32 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_conv3d_bwd_data_fp16 PRIVATE composable_kernel::device_conv_operations)
+target_link_libraries(client_conv3d_bwd_data_fp32 PRIVATE composable_kernel::device_conv_operations)

client_example/15_gemm_add_multiply/CMakeLists.txt

 
 add_executable(client_gemm_add_multiply gemm_add_multiply.cpp)
-target_link_libraries(client_gemm_add_multiply PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_gemm_add_multiply PRIVATE composable_kernel::device_gemm_operations)
\ No newline at end of file

client_example/15_reduce/CMakeLists.txt

 add_executable(client_reduce_nhwc_c reduce_nhwc_c.cpp)
-target_link_libraries(client_reduce_nhwc_c PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_reduce_nhwc_c PRIVATE composable_kernel::device_reduction_operations)

client_example/16_convnd_fwd/CMakeLists.txt

 if((DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
     add_executable(client_conv3d_fwd_fp16 conv3d_fwd_fp16.cpp)
-    target_link_libraries(client_conv3d_fwd_fp16 PRIVATE composable_kernel::device_operations)
+    target_link_libraries(client_conv3d_fwd_fp16 PRIVATE composable_kernel::device_conv_operations)
 
 endif()
 
 if((DTYPES MATCHES "fp8") OR NOT DEFINED DTYPES)
     add_executable(client_conv3d_fwd_fp16_comp_fp8 conv3d_fwd_fp16_comp_fp8.cpp)
-    target_link_libraries(client_conv3d_fwd_fp16_comp_fp8 PRIVATE composable_kernel::device_operations)
+    target_link_libraries(client_conv3d_fwd_fp16_comp_fp8 PRIVATE composable_kernel::device_conv_operations)
 endif()
 
 if((DTYPES MATCHES "fp32") OR NOT DEFINED DTYPES)
     add_executable(client_conv3d_fwd_fp32 conv3d_fwd_fp32.cpp)
-    target_link_libraries(client_conv3d_fwd_fp32 PRIVATE composable_kernel::device_operations)
+    target_link_libraries(client_conv3d_fwd_fp32 PRIVATE composable_kernel::device_conv_operations)
 endif()

client_example/16_convnd_fwd/common.hpp

@@ -11,7 +11,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
@@ -174,19 +174,19 @@ bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialD
     std::size_t flop      = GetFlops<NumDimSpatial>(out_lengths, wei_lengths);
     std::size_t num_bytes = in_mem_size + wei_mem_size + out_mem_size;
 
-    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
-                                                                                 InLayout,
-                                                                                 WeiLayout,
-                                                                                 ck::Tuple<>,
-                                                                                 OutLayout,
-                                                                                 InDataType,
-                                                                                 WeiDataType,
-                                                                                 ck::Tuple<>,
-                                                                                 OutDataType,
-                                                                                 PassThrough,
-                                                                                 PassThrough,
-                                                                                 PassThrough,
-                                                                                 ComputeType>;
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
+                                                                                   InLayout,
+                                                                                   WeiLayout,
+                                                                                   ck::Tuple<>,
+                                                                                   OutLayout,
+                                                                                   InDataType,
+                                                                                   WeiDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   OutDataType,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   ComputeType>;
     // get device op instances
     const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
         DeviceOp>::GetInstances();

client_example/17_grouped_gemm_fastgelu/CMakeLists.txt

 add_executable(client_grouped_gemm_fastgelu grouped_gemm_fastgelu.cpp)
-target_link_libraries(client_grouped_gemm_fastgelu PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_gemm_fastgelu PRIVATE composable_kernel::device_gemm_operations)
\ No newline at end of file

client_example/18_groupnorm/CMakeLists.txt

 add_executable(client_groupnorm_swish groupnorm_swish.cpp)
-target_link_libraries(client_groupnorm_swish PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_groupnorm_swish PRIVATE composable_kernel::device_other_operations)

client_example/19_pool/CMakeLists.txt

 add_executable(client_max_pool2d_fwd max_pool2d_fwd.cpp)
-target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_other_operations)
 
 add_executable(client_max_pool2d_bwd max_pool2d_bwd.cpp)
-target_link_libraries(client_max_pool2d_bwd PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_max_pool2d_bwd PRIVATE composable_kernel::device_other_operations)
 
 add_executable(client_avg_pool3d_fwd avg_pool3d_fwd.cpp)
-target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_other_operations)
 
 add_executable(client_avg_pool3d_bwd avg_pool3d_bwd.cpp)
-target_link_libraries(client_avg_pool3d_bwd PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_avg_pool3d_bwd PRIVATE composable_kernel::device_other_operations)

client_example/20_splitk_gemm/CMakeLists.txt

 if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
   add_executable(client_splitK_gemm splitK_gemm_fp16_f8.cpp)
-  target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_operations)
+  target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_gemm_operations)
 endif()

client_example/21_grouped_gemm_bias/CMakeLists.txt

 add_executable(client_grouped_gemm_fixed_nk_bias_fp16 grouped_gemm_fixed_nk_bias_fp16.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_bias_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_gemm_fixed_nk_bias_fp16 PRIVATE composable_kernel::device_gemm_operations)

client_example/22_grouped_gemm/CMakeLists.txt

 add_executable(client_grouped_gemm_fixed_nk_fp16 grouped_gemm_fixed_nk_fp16.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_gemm_fixed_nk_fp16 PRIVATE composable_kernel::device_gemm_operations)
 
 add_executable(client_grouped_gemm_fixed_nk_fp8 grouped_gemm_fixed_nk_fp8.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_fp8 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_gemm_fixed_nk_fp8 PRIVATE composable_kernel::device_gemm_operations)
 
 add_executable(client_grouped_gemm_fixed_nk_i8 grouped_gemm_fixed_nk_i8.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_i8 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_gemm_fixed_nk_i8 PRIVATE composable_kernel::device_gemm_operations)

client_example/22_im2col_col2im/CMakeLists.txt

 add_executable(client_image_to_column image_to_column.cpp)
-target_link_libraries(client_image_to_column PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_image_to_column PRIVATE composable_kernel::device_other_operations)
 
 add_executable(client_column_to_image column_to_image.cpp)
-target_link_libraries(client_column_to_image PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_column_to_image PRIVATE composable_kernel::device_other_operations)

client_example/23_elementwise_transpose/CMakeLists.txt

+add_executable(client_elementwise_transpose3d elementwise_transpose_3d.cpp)
+target_link_libraries(client_elementwise_transpose3d PRIVATE composable_kernel::device_other_operations)

client_example/23_elementwise_transpose/elementwise_transpose_3d.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/transpose_3d.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    const int N = 16;
+    const int C = 8;
+    const int D = 8;
+    const int H = 8;
+    const int W = 8;
+
+    std::vector<std::size_t> ncdhw = {N, C, D, H, W};
+    std::vector<std::size_t> nchwd = {N, C, H, W, D};
+    auto size                      = N * C * D * H * W;
+
+    std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
+    std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
+    std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, C, H, W, D
+
+    SimpleDeviceMem a_dev_buf(sizeof(ADataType) * size);
+    SimpleDeviceMem b_dev_buf(sizeof(BDataType) * size);
+
+    std::array<const void*, 1> input = {a_dev_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_dev_buf.GetDeviceBuffer()};
+
+    using DeviceElementwisePermuteInstance = ck::tensor_operation::device::
+        DeviceElementwise<ck::Tuple<ADataType>, ck::Tuple<BDataType>, PassThrough, 5>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceElementwisePermuteInstance>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte =
+                sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
+                sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}