Merge remote-tracking branch 'origin/develop' into stream-k-initial-impl

Jenkinsfile

@@ -696,7 +696,7 @@ pipeline {
                     agent{ label rocmnode("gfx908 || gfx90a") }
                     environment{
                         setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940" """
-                        execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ 
+                        execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ 
                     }
                     steps{
                         Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')

client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp

@@ -172,18 +172,19 @@ int main()
             BLayout,
             CLayout>();
 
-    const auto normalize_ptrs =
-        ck::tensor_operation::device::instance::get_device_normalize_from_mean_meansquare_instances<
-            CDataType,
-            ReduceDataType,
-            ReduceDataType,
-            GammaDataType,
-            BetaDataType,
-            LayerNormOutDataType>();
-
     std::cout << "found " << gemm_reduce_ptrs.size()
               << " gemm_reduceMean_reduceSquareMean instances" << std::endl;
 
+    using NormalizeDeviceOp = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<CDataType, ReduceDataType, ReduceDataType, GammaDataType, BetaDataType>,
+        ck::Tuple<LayerNormOutDataType>,
+        ck::tensor_operation::element_wise::Normalize,
+        2>;
+
+    const auto normalize_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            NormalizeDeviceOp>::GetInstances();
+
     std::cout << "found " << normalize_ptrs.size() << " normalize instances" << std::endl;
 
     auto f_matrix_space_size =

client_example/06_softmax/softmax4d.cpp

@@ -53,12 +53,35 @@ int main(int argc, char* argv[])
     SimpleDeviceMem in(sizeof(InDataType) * num_elements);
     SimpleDeviceMem out(sizeof(OutDataType) * num_elements);
 
-    using DeviceOp = ck::tensor_operation::device::
-        DeviceSoftmax<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>;
+    using DeviceOp = ck::tensor_operation::device::DeviceSoftmax<InDataType,
+                                                                 AccDataType,
+                                                                 OutDataType,
+                                                                 PassThrough,
+                                                                 PassThrough,
+                                                                 Rank,
+                                                                 NumReduceDim>;
     // get device op instances
     const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
         DeviceOp>::GetInstances();
 
+    auto& generic_op_ptr = op_ptrs[0];
+
+    auto generic_argument_ptr = generic_op_ptr->MakeArgumentPointer(in_lengths,
+                                                                    in_strides,
+                                                                    reduce_dims,
+                                                                    alpha,
+                                                                    beta,
+                                                                    in.GetDeviceBuffer(),
+                                                                    out.GetDeviceBuffer(),
+                                                                    PassThrough{},
+                                                                    PassThrough{});
+
+    if(!generic_op_ptr->IsSupportedArgument(generic_argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "The generic kernel instance should be able to support any input shapes");
+    };
+
     std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
 
     std::string best_op_name;
@@ -74,11 +97,6 @@ int main(int argc, char* argv[])
     {
         auto& op_ptr = op_ptrs[i];
 
-        if(op_ptr->GetRank() != Rank || op_ptr->GetNumReduceDim() != NumReduceDim)
-        {
-            continue;
-        }
-
         auto argument_ptr   = op_ptr->MakeArgumentPointer(in_lengths,
                                                         in_strides,
                                                         reduce_dims,

client_example/16_convnd_fwd/common.hpp

@@ -141,14 +141,10 @@ bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialD
                 std::next(rbegin(in_strides)),
                 std::next(rbegin(in_strides), NumDimSpatial + 1));
 
-    std::rotate(
-        std::next(rbegin(wei_lengths)), std::next(rbegin(wei_lengths), 2), rend(wei_lengths));
     std::rotate(rbegin(wei_lengths),
                 std::next(rbegin(wei_lengths)),
                 std::next(rbegin(wei_lengths), NumDimSpatial + 1));
 
-    std::rotate(
-        std::next(rbegin(wei_strides)), std::next(rbegin(wei_strides), 2), rend(wei_strides));
     std::rotate(rbegin(wei_strides),
                 std::next(rbegin(wei_strides)),
                 std::next(rbegin(wei_strides), NumDimSpatial + 1));

client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp

@@ -11,7 +11,7 @@ using WeiDataType = ck::half_t;
 using OutDataType = ck::half_t;
 
 using InLayout  = ck::tensor_layout::convolution::NDHWGC;
-using WeiLayout = ck::tensor_layout::convolution::KZYXGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
 using OutLayout = ck::tensor_layout::convolution::NDHWGK;
 
 static constexpr ck::index_t NumDimSpatial = 3;
@@ -38,7 +38,7 @@ int main()
                                 InLayout,
                                 WeiLayout,
                                 OutLayout>(
-               {N, Di, Hi, Wi, G, C}, {K, Z, Y, X, G, C}, {N, Do, Ho, Wo, G, K})
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
                ? EXIT_SUCCESS
                : EXIT_FAILURE;
 }

client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp

@@ -11,7 +11,7 @@ using WeiDataType = float;
 using OutDataType = float;
 
 using InLayout  = ck::tensor_layout::convolution::NDHWGC;
-using WeiLayout = ck::tensor_layout::convolution::KZYXGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
 using OutLayout = ck::tensor_layout::convolution::NDHWGK;
 
 static constexpr ck::index_t NumDimSpatial = 3;
@@ -38,7 +38,7 @@ int main()
                                 InLayout,
                                 WeiLayout,
                                 OutLayout>(
-               {N, Di, Hi, Wi, G, C}, {K, Z, Y, X, G, C}, {N, Do, Ho, Wo, G, K})
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
                ? EXIT_SUCCESS
                : EXIT_FAILURE;
 }

client_example/18_groupnorm/groupnorm_swish.cpp

@@ -72,6 +72,30 @@ int main(int argc, char* argv[])
 
     std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
 
+    const auto& generic_op_ptr = op_ptrs[0];
+
+    auto generic_argument_ptr =
+        generic_op_ptr->MakeArgumentPointer({N, H, W, G, C},    // lengths
+                                            xy_strides,         // xStrides
+                                            gamma_beta_strides, // gammaStrides
+                                            gamma_beta_strides, // betaStrides
+                                            xy_strides,         // yStrides
+                                            {1, 2, 4},          // reduceDims
+                                            1e-6,
+                                            x_device_buf.GetDeviceBuffer(),
+                                            gamma_device_buf.GetDeviceBuffer(),
+                                            beta_device_buf.GetDeviceBuffer(),
+                                            y_device_buf.GetDeviceBuffer(),
+                                            nullptr,
+                                            nullptr,
+                                            Swish{});
+
+    if(!generic_op_ptr->IsSupportedArgument(generic_argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "The generic kernel instance should be able to support any input shapes");
+    };
+
     std::string best_op_name;
     bool found            = false;
     int best_op_id        = -1;

example/02_gemm_bilinear/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
+list(APPEND gpu_list1 gfx1100 gfx1101 gfx1102)
+list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
     add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp)
-endif()
-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+   set(target 1)
+ endif()
+endforeach()
+
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
     add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
-endif()
+   set(target 1)
+ endif()
+endforeach()

example/03_gemm_bias_relu/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
     add_example_executable(example_gemm_bias_relu_xdl_fp16 gemm_bias_relu_xdl_fp16.cpp)
-endif()
+   set(target 1)
\ No newline at end of file
+ endif()
+endforeach()

example/04_gemm_add_add_fastgelu/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
     add_custom_target(example_gemm_add_add_fastgelu_xdl)
 
     add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
@@ -16,4 +19,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M
         add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
     endif(USE_BITINT_EXTENSION_INT4)
     add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
-endif()
+   set(target 1)
\ No newline at end of file
+ endif()
+endforeach()
\ No newline at end of file

example/09_convnd_fwd/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
     add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
     add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
     add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
     add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
     # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
     add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
-endif()
+   set(target 1)
+ endif()
+endforeach()
+
 add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
 add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
 add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
-

example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
    add_custom_target(example_convnd_fwd_reduce_xdl)
    add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
    add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
@@ -12,4 +15,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M
       add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp)
       add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)
    endif(USE_BITINT_EXTENSION_INT4)
-endif()
+   set(target 1)
\ No newline at end of file
+ endif()
+endforeach()
\ No newline at end of file

example/14_gemm_quantization/CMakeLists.txt

@@ -2,7 +2,12 @@
 add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
 
 # xdlops
-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
    add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp)
    add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp)
-endif()
+   set(target 1)
\ No newline at end of file
+ endif()
+endforeach()
\ No newline at end of file

example/16_gemm_multi_d_multi_reduces/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
    add_custom_target(example_gemm_reduce_xdl)
    add_custom_target(example_gemm_reduce_xdl_max)
    add_custom_target(example_gemm_reduce_xdl_mean_meansquare)
@@ -39,4 +42,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M
       add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
       add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
    endif()
-endif()
+   set(target 1)
+ endif()
+endforeach()

example/17_convnd_bwd_data/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
    add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp)
    target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
-endif()
+   set(target 1)
+ endif()
+endforeach()
 add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
 target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)

example/18_batched_gemm_reduce/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
    add_example_executable(example_batched_gemm_reduce_xdl_fp16 batched_gemm_reduce_xdl_fp16.cpp)
-endif()
-
+   set(target 1)
+ endif()
+endforeach()

example/20_grouped_conv_bwd_weight/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
    add_custom_target(example_grouped_conv_bwd_weight)
 
    add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
@@ -6,7 +9,9 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M
 
    add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16
                                                  example_grouped_conv_bwd_weight_xdl_bf16)
-endif()
+   set(target 1)
+ endif()
+endforeach()
 
 add_custom_target(example_grouped_conv_bwd_weight_dl)
 

example/21_gemm_layernorm/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
    add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_welford_fp16 gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp)
    add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_naive_fp16 gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp)
    add_example_executable(example_gemm_layernorm_xdl_naive_fp16 gemm_layernorm_xdl_naive_fp16.cpp)
    add_example_executable(example_gemm_xdl_layernorm_naive_single_kernel_fp16 gemm_xdl_layernorm_naive_single_kernel_fp16.cpp)
-endif()
+   set(target 1)
+ endif()
+endforeach()

example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list2 gfx1100 gfx1101 gfx1102)
+
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
    add_custom_target(example_grouped_conv_fwd_multiple_d)
 
    add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
@@ -17,8 +22,15 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M
    endif() # USE_BITINT_EXTENSION_INT4
    add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
    add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
-endif()
-if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
-  add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
-  add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
-endif()
+   set(target 1)
+ endif()
+endforeach()
+
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
+   add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
+   add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
+   set(target 1)
+ endif()
+endforeach()

example/31_batched_gemm_gemm/CMakeLists.txt

-if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list2 gfx908 gfx90a)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
    add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
    add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
    add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)
-   if(NOT GPU_TARGETS MATCHES "gfx940")
-      add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
-   endif()
 
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp)
    endif(USE_BITINT_EXTENSION_INT4)
-endif()
+   set(target 1)
\ No newline at end of file
+ endif()
+endforeach()
+
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
+   add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
+   set(target 1)
+ endif()
+endforeach()
\ No newline at end of file