resolved conflicts

9f1b4276 · Jakub Piasecki · 711857c4 · c7010716 · 9f1b4276 · 9f1b4276
Commit 9f1b4276 authored Apr 04, 2024 by Jakub Piasecki
20 changed files
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -143,6 +143,22 @@ if(GPU_TARGETS)
 else()
    message("Building CK for the following targets: ${AMDGPU_TARGETS}")
 endif()
+
+if (GPU_TARGETS)
+    if (GPU_TARGETS MATCHES "gfx9")
+        add_definitions(-DCK_USE_XDL)
+        set(CK_USE_XDL "ON")
+    endif()
+    if (GPU_TARGETS MATCHES "gfx11")
+        add_definitions(-DCK_USE_WMMA)
+        set(CK_USE_WMMA "ON")
+    endif()
+else()
+    add_definitions(-DCK_USE_WMMA -DCK_USE_XDL)
+    set(CK_USE_XDL "ON")
+    set(CK_USE_WMMA "ON")
+endif()
+
 find_package(hip)
 # No assumption that HIP kernels are launched with uniform block size for backward compatibility
 # SWDEV-413293 and https://reviews.llvm.org/D155213

--- a/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
+++ b/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
-add_custom_target(client_gemm_fastgelu_examples)
+if(GPU_TARGETS MATCHES "gfx9")
+    add_custom_target(client_gemm_fastgelu_examples)

-add_executable(client_gemm_add_add_fastgelu gemm_add_add_fastgelu.cpp)
-target_link_libraries(client_gemm_add_add_fastgelu PRIVATE composable_kernel::device_gemm_operations)
+    add_executable(client_gemm_add_add_fastgelu gemm_add_add_fastgelu.cpp)
+    target_link_libraries(client_gemm_add_add_fastgelu PRIVATE composable_kernel::device_gemm_operations)

-add_executable(client_gemm_add_fastgelu gemm_add_fastgelu.cpp)
-target_link_libraries(client_gemm_add_fastgelu PRIVATE composable_kernel::device_gemm_operations)
+    add_executable(client_gemm_add_fastgelu gemm_add_fastgelu.cpp)
+    target_link_libraries(client_gemm_add_fastgelu PRIVATE composable_kernel::device_gemm_operations)

-add_executable(client_gemm_fastgelu gemm_fastgelu.cpp)
-target_link_libraries(client_gemm_fastgelu PRIVATE composable_kernel::device_gemm_operations)
+    add_executable(client_gemm_fastgelu gemm_fastgelu.cpp)
+    target_link_libraries(client_gemm_fastgelu PRIVATE composable_kernel::device_gemm_operations)

-add_dependencies(client_gemm_fastgelu_examples client_gemm_add_add_fastgelu client_gemm_add_fastgelu
+    add_dependencies(client_gemm_fastgelu_examples client_gemm_add_add_fastgelu client_gemm_add_fastgelu
                 client_gemm_fastgelu)

-add_custom_target(client_gemm_fastgelu_generic_examples)
+    add_custom_target(client_gemm_fastgelu_generic_examples)

-add_executable(client_gemm_add_add_fastgelu_generic gemm_add_add_fastgelu_generic.cpp)
-target_link_libraries(client_gemm_add_add_fastgelu_generic composable_kernel::device_gemm_operations)
+    add_executable(client_gemm_add_add_fastgelu_generic gemm_add_add_fastgelu_generic.cpp)
+    target_link_libraries(client_gemm_add_add_fastgelu_generic composable_kernel::device_gemm_operations)

-add_executable(client_gemm_add_fastgelu_generic gemm_add_fastgelu_generic.cpp)
-target_link_libraries(client_gemm_add_fastgelu_generic PRIVATE composable_kernel::device_gemm_operations)
+    add_executable(client_gemm_add_fastgelu_generic gemm_add_fastgelu_generic.cpp)
+    target_link_libraries(client_gemm_add_fastgelu_generic PRIVATE composable_kernel::device_gemm_operations)

-add_executable(client_gemm_fastgelu_generic gemm_fastgelu_generic.cpp)
-target_link_libraries(client_gemm_fastgelu_generic PRIVATE composable_kernel::device_gemm_operations)
+    add_executable(client_gemm_fastgelu_generic gemm_fastgelu_generic.cpp)
+    target_link_libraries(client_gemm_fastgelu_generic PRIVATE composable_kernel::device_gemm_operations)

-add_dependencies(client_gemm_fastgelu_generic_examples client_gemm_add_add_fastgelu_generic 
+    add_dependencies(client_gemm_fastgelu_generic_examples client_gemm_add_add_fastgelu_generic 
                 client_gemm_add_fastgelu_generic client_gemm_fastgelu_generic)
+endif()
--- a/client_example/03_gemm_layernorm/CMakeLists.txt
+++ b/client_example/03_gemm_layernorm/CMakeLists.txt
-add_executable(client_gemm_add_add_layernorm_naive gemm_add_add_layernorm_naive.cpp)
-target_link_libraries(client_gemm_add_add_layernorm_naive PRIVATE composable_kernel::device_gemm_operations composable_kernel::device_other_operations)
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_gemm_add_add_layernorm_naive gemm_add_add_layernorm_naive.cpp)
+    target_link_libraries(client_gemm_add_add_layernorm_naive PRIVATE composable_kernel::device_gemm_operations composable_kernel::device_other_operations)

-add_executable(client_gemm_add_relu_add_layernorm_welford gemm_add_relu_add_layernorm_welford.cpp)
-target_link_libraries(client_gemm_add_relu_add_layernorm_welford PRIVATE composable_kernel::device_gemm_operations composable_kernel::device_other_operations)
+    add_executable(client_gemm_add_relu_add_layernorm_welford gemm_add_relu_add_layernorm_welford.cpp)
+    target_link_libraries(client_gemm_add_relu_add_layernorm_welford PRIVATE composable_kernel::device_gemm_operations composable_kernel::device_other_operations)
+endif()
--- a/client_example/04_contraction/CMakeLists.txt
+++ b/client_example/04_contraction/CMakeLists.txt
-add_executable(client_contraction_scale_fp32 contraction_scale_fp32.cpp)
-target_link_libraries(client_contraction_scale_fp32 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_contraction_scale_fp32 contraction_scale_fp32.cpp)
+    target_link_libraries(client_contraction_scale_fp32 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)

-add_executable(client_contraction_bilinear_fp32 contraction_bilinear_fp32.cpp)
-target_link_libraries(client_contraction_bilinear_fp32 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
+    add_executable(client_contraction_bilinear_fp32 contraction_bilinear_fp32.cpp)
+    target_link_libraries(client_contraction_bilinear_fp32 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)

-add_executable(client_contraction_scale_fp64 contraction_scale_fp64.cpp)
-target_link_libraries(client_contraction_scale_fp64 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
+    add_executable(client_contraction_scale_fp64 contraction_scale_fp64.cpp)
+    target_link_libraries(client_contraction_scale_fp64 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)

-add_executable(client_contraction_bilinear_fp64 contraction_bilinear_fp64.cpp)
-target_link_libraries(client_contraction_bilinear_fp64 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
-
-add_executable(contraction_g1m2n3k1_add_xdl_fp16 contraction_g1m2n3k1_add_xdl_fp16.cpp)
-target_link_libraries(contraction_g1m2n3k1_add_xdl_fp16 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
+    add_executable(client_contraction_bilinear_fp64 contraction_bilinear_fp64.cpp)
+    target_link_libraries(client_contraction_bilinear_fp64 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)

+    add_executable(contraction_g1m2n3k1_add_xdl_fp16 contraction_g1m2n3k1_add_xdl_fp16.cpp)
+    target_link_libraries(contraction_g1m2n3k1_add_xdl_fp16 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
+endif()
--- a/client_example/07_grouped_convnd_fwd/CMakeLists.txt
+++ b/client_example/07_grouped_convnd_fwd/CMakeLists.txt
-add_executable(client_grouped_conv2d_fwd grouped_conv2d_fwd.cpp)
-target_link_libraries(client_grouped_conv2d_fwd PRIVATE composable_kernel::device_conv_operations)
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_grouped_conv2d_fwd grouped_conv2d_fwd.cpp)
+    target_link_libraries(client_grouped_conv2d_fwd PRIVATE composable_kernel::device_conv_operations)

-add_executable(client_grouped_conv1d_fwd grouped_conv1d_fwd.cpp)
-target_link_libraries(client_grouped_conv1d_fwd PRIVATE composable_kernel::device_conv_operations)
+    add_executable(client_grouped_conv1d_fwd grouped_conv1d_fwd.cpp)
+    target_link_libraries(client_grouped_conv1d_fwd PRIVATE composable_kernel::device_conv_operations)
+endif()
\ No newline at end of file
--- a/client_example/08_fused_attention/CMakeLists.txt
+++ b/client_example/08_fused_attention/CMakeLists.txt
-add_executable(client_fused_attention fused_attention.cpp)
-target_link_libraries(client_fused_attention PRIVATE composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_fused_attention fused_attention.cpp)
+    target_link_libraries(client_fused_attention PRIVATE composable_kernel::device_other_operations composable_kernel::device_gemm_operations)

-add_executable(client_fused_attention_bias fused_attention_bias.cpp)
-target_link_libraries(client_fused_attention_bias PRIVATE composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+    add_executable(client_fused_attention_bias fused_attention_bias.cpp)
+    target_link_libraries(client_fused_attention_bias PRIVATE composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+endif()
--- a/client_example/09_quantization/CMakeLists.txt
+++ b/client_example/09_quantization/CMakeLists.txt
-if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
-add_executable(client_conv2d_fwd_bias_tanh_perchannel_quantization conv2d_fwd_bias_tanh_perchannel_quantization.cpp)
-target_link_libraries(client_conv2d_fwd_bias_tanh_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+if(GPU_TARGETS MATCHES "gfx9" AND (DTYPES MATCHES "int8" OR NOT DEFINED DTYPES))
+    add_executable(client_conv2d_fwd_bias_tanh_perchannel_quantization conv2d_fwd_bias_tanh_perchannel_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_bias_tanh_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)

-add_executable(client_conv2d_fwd_bias_relu_perchannel_quantization conv2d_fwd_bias_relu_perchannel_quantization.cpp)
-target_link_libraries(client_conv2d_fwd_bias_relu_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+    add_executable(client_conv2d_fwd_bias_relu_perchannel_quantization conv2d_fwd_bias_relu_perchannel_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_bias_relu_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)

-add_executable(client_conv2d_fwd_bias_tanh_perlayer_quantization conv2d_fwd_bias_tanh_perlayer_quantization.cpp)
-target_link_libraries(client_conv2d_fwd_bias_tanh_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+    add_executable(client_conv2d_fwd_bias_tanh_perlayer_quantization conv2d_fwd_bias_tanh_perlayer_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_bias_tanh_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)

-add_executable(client_conv2d_fwd_bias_relu_perlayer_quantization conv2d_fwd_bias_relu_perlayer_quantization.cpp)
-target_link_libraries(client_conv2d_fwd_bias_relu_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+    add_executable(client_conv2d_fwd_bias_relu_perlayer_quantization conv2d_fwd_bias_relu_perlayer_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_bias_relu_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)

-add_executable(client_conv2d_fwd_perchannel_quantization conv2d_fwd_perchannel_quantization.cpp)
-target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+    add_executable(client_conv2d_fwd_perchannel_quantization conv2d_fwd_perchannel_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)

-add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp)
-target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+    add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp)
+    target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)

-add_executable(client_gemm_quantization gemm_quantization.cpp)
-target_link_libraries(client_gemm_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
+    add_executable(client_gemm_quantization gemm_quantization.cpp)
+    target_link_libraries(client_gemm_quantization PRIVATE composable_kernel::device_conv_operations composable_kernel::device_other_operations composable_kernel::device_gemm_operations)
 endif()
--- a/client_example/15_convnd_bwd_data/CMakeLists.txt
+++ b/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)
+if(GPU_TARGETS MATCHES "gfx9")
+    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_conv_operations)
-target_link_libraries(client_conv3d_bwd_data_fp32 PRIVATE composable_kernel::device_conv_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)
+endif()
--- a/client_example/15_gemm_add_multiply/CMakeLists.txt
+++ b/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_gemm_operations)
\ No newline at end of file
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_gemm_add_multiply gemm_add_multiply.cpp)
+    target_link_libraries(client_gemm_add_multiply PRIVATE composable_kernel::device_gemm_operations)
+endif()
--- a/client_example/16_convnd_fwd/CMakeLists.txt
+++ b/client_example/16_convnd_fwd/CMakeLists.txt
@@ -17,6 +17,11 @@ if((DTYPES MATCHES "bf8") OR NOT DEFINED DTYPES)
    target_link_libraries(client_conv3d_fwd_bf8 PRIVATE composable_kernel::device_conv_operations)
 endif()

+if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8") OR NOT DEFINED DTYPES)
+    add_executable(client_conv3d_fwd_fp8_bf8 conv3d_fwd_fp8_bf8.cpp)
+    target_link_libraries(client_conv3d_fwd_fp8_bf8 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_conv_operations)

--- a/client_example/16_convnd_fwd/common.hpp
+++ b/client_example/16_convnd_fwd/common.hpp
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.

 #include <cstdlib>
 #include <iomanip>
@@ -95,7 +95,8 @@ template <ck::index_t NumDimSpatial,
          typename WeiLayout,
          typename OutLayout,
          ck::index_t NumNonSpatialDim = 3,
-          typename ComputeType         = InDataType>
+          typename AComputeType        = InDataType,
+          typename BComputeType        = AComputeType>
 bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
                          std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
                          std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
@@ -186,7 +187,8 @@ bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialD
                                                                                   PassThrough,
                                                                                   PassThrough,
                                                                                   PassThrough,
-                                                                                   ComputeType>;
+                                                                                   AComputeType,
+                                                                                   BComputeType>;
    // get device op instances
    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
        DeviceOp>::GetInstances();

--- a/client_example/16_convnd_fwd/conv3d_fwd_fp8_bf8.cpp
+++ b/client_example/16_convnd_fwd/conv3d_fwd_fp8_bf8.cpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, 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::f8_t;
+using WeiDataType = ck::bf8_t;
+using OutDataType = ck::f8_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::f8_t;
+using BComputeType = ck::bf8_t;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+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;
+
+int main()
+{
+    return run_grouped_conv_fwd<NumDimSpatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InLayout,
+                                WeiLayout,
+                                OutLayout,
+                                3,
+                                AComputeType,
+                                BComputeType>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}
--- a/client_example/17_grouped_gemm_fastgelu/CMakeLists.txt
+++ b/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_gemm_operations)
\ No newline at end of file
+if(GPU_TARGETS MATCHES "gfx9")
+    add_executable(client_grouped_gemm_fastgelu grouped_gemm_fastgelu.cpp)
+    target_link_libraries(client_grouped_gemm_fastgelu PRIVATE composable_kernel::device_gemm_operations)
+endif()
--- a/client_example/20_splitk_gemm/CMakeLists.txt
+++ b/client_example/20_splitk_gemm/CMakeLists.txt
-if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
+if(GPU_TARGETS MATCHES "gfx9" AND ((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_gemm_operations)
 endif()
--- a/client_example/21_grouped_gemm_bias/CMakeLists.txt
+++ b/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_gemm_operations)
+if(GPU_TARGETS MATCHES "gfx9")
+    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_gemm_operations)
+endif()
--- a/client_example/22_grouped_gemm/CMakeLists.txt
+++ b/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_gemm_operations)
+if(GPU_TARGETS MATCHES "gfx9")
+    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_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_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_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_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_gemm_operations)

-add_executable(client_grouped_gemm_fixed_nk_bf16 grouped_gemm_fixed_nk_bf16.cpp)
-target_link_libraries(client_grouped_gemm_fixed_nk_bf16 PRIVATE composable_kernel::device_gemm_operations)
+    add_executable(client_grouped_gemm_fixed_nk_bf16 grouped_gemm_fixed_nk_bf16.cpp)
+    target_link_libraries(client_grouped_gemm_fixed_nk_bf16 PRIVATE composable_kernel::device_gemm_operations)
+endif()
--- a/client_example/24_grouped_conv_activation/CMakeLists.txt
+++ b/client_example/24_grouped_conv_activation/CMakeLists.txt
+if(GPU_TARGETS MATCHES "gfx9")
 # Fwd scaleadd scaleadd relu
 add_executable(client_grouped_convnd_fwd_scaleadd_scaleadd_relu_fp32 
               grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp32.cpp)
@@ -46,3 +47,4 @@ target_link_libraries(client_grouped_convnd_fwd_scale_fp16 PRIVATE composable_ke
 add_executable(client_grouped_convnd_bwd_data_scale_fp16
               grouped_convnd_bwd_data_scale/grouped_conv_bwd_data_scale_fp16.cpp)
 target_link_libraries(client_grouped_convnd_bwd_data_scale_fp16 PRIVATE composable_kernel::device_conv_operations)
+endif()
--- a/client_example/25_wrapper/CMakeLists.txt
+++ b/client_example/25_wrapper/CMakeLists.txt
@@ -2,9 +2,7 @@ add_executable(client_tensor_transform_using_wrapper tensor_transform_using_wrap
 target_link_libraries(client_tensor_transform_using_wrapper PRIVATE composable_kernel::device_other_operations)
 add_executable(client_wrapper_img2col wrapper_img2col.cpp)
 target_link_libraries(client_wrapper_img2col PRIVATE composable_kernel::device_other_operations)
-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR
-   GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR
-   GPU_TARGETS MATCHES "gfx942")
+if(GPU_TARGETS MATCHES "gfx9")
    add_executable(client_wrapper_basic_gemm wrapper_basic_gemm.cpp)
    target_link_libraries(client_wrapper_basic_gemm PRIVATE composable_kernel::device_other_operations)
    add_executable(client_wrapper_optimized_gemm wrapper_optimized_gemm.cpp)

--- a/client_example/CMakeLists.txt
+++ b/client_example/CMakeLists.txt
@@ -48,7 +48,10 @@ else()
    endif()
 endif()

-find_package(composable_kernel COMPONENTS device_other_operations device_gemm_operations device_conv_operations device_contraction_operations device_reduction_operations)
+find_package(composable_kernel COMPONENTS device_other_operations device_gemm_operations device_conv_operations  device_reduction_operations)
+if(GPU_TARGETS MATCHES "gfx9")
+    find_package(composable_kernel COMPONENTS device_contraction_operations)
+endif()
 find_package(hip REQUIRED PATHS /opt/rocm)
 message(STATUS "Build with HIP ${hip_VERSION}")


--- a/example/01_gemm/CMakeLists.txt
+++ b/example/01_gemm/CMakeLists.txt
@@ -27,11 +27,6 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)

 add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
-if(GPU_TARGETS MATCHES "gfx11")
-    add_custom_target(example_gemm_wmma)
-    add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
-    add_example_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
-endif()

 add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16)
@@ -47,8 +42,7 @@ if(USE_BITINT_EXTENSION_INT4)
    add_example_dependencies(example_gemm_xdl example_gemm_xdl_int4)
 endif(USE_BITINT_EXTENSION_INT4)

-# FIXME: re-enable this example as test when SWDEV-335738 is fixed
-add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
+add_example_executable(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp64)

 add_example_executable(example_gemm_xdl_streamk gemm_xdl_streamk.cpp)
@@ -75,3 +69,6 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_bf8)
 add_example_executable(example_gemm_xdl_fp16_fp8 gemm_xdl_fp16_fp8.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8)

+add_custom_target(example_gemm_wmma)
+add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
+add_example_dependencies(example_gemm_wmma example_gemm_wmma_fp16)