fixed merge

example/62_conv_fwd_activ/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
-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_activ_xdl)
-      # Sigmoid
-      add_example_executable(example_convnd_fwd_xdl_sigmoid_fp16 convnd_fwd_xdl_sigmoid_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_sigmoid_fp16)
-      # Tanh
-      add_example_executable(example_convnd_fwd_xdl_tanh_fp16 convnd_fwd_xdl_tanh_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_tanh_fp16)
-      # Relu
-      add_example_executable(example_convnd_fwd_xdl_relu_fp16 convnd_fwd_xdl_relu_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_relu_fp16)
-      # SoftRelu
-      add_example_executable(example_convnd_fwd_xdl_softrelu_fp16 convnd_fwd_xdl_softrelu_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_softrelu_fp16)
-      # Abs
-      add_example_executable(example_convnd_fwd_xdl_abs_fp16 convnd_fwd_xdl_abs_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_abs_fp16)
-      # Pow
-      add_example_executable(example_convnd_fwd_xdl_pow_fp16 convnd_fwd_xdl_pow_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_pow_fp16)
-      # Clipped Relu
-      add_example_executable(example_convnd_fwd_xdl_clippedrelu_fp16 convnd_fwd_xdl_clippedrelu_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_clippedrelu_fp16)
-      # Leaky Relu
-      add_example_executable(example_convnd_fwd_xdl_leakyrelu_fp16 convnd_fwd_xdl_leakyrelu_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_leakyrelu_fp16)
-      # Elu
-      add_example_executable(example_convnd_fwd_xdl_elu_fp16 convnd_fwd_xdl_elu_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_elu_fp16)
-      # ScaleAdd on A and B
-      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_fp16 multi_AB/conv_fwd_xdl_scaleadd_ab_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_fp16)
-      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_fp32 multi_AB/conv_fwd_xdl_scaleadd_ab_fp32.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_fp32)
-      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_bf16 multi_AB/conv_fwd_xdl_scaleadd_ab_bf16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_bf16)
-      add_example_executable(example_conv_fwd_xdl_scaleadd_ab_int8 multi_AB/conv_fwd_xdl_scaleadd_ab_int8.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_conv_fwd_xdl_scaleadd_ab_int8)
-      # ScaleAdd ScaleAdd Relu
-      add_example_executable(example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16 convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16)
-      add_example_executable(example_convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16 convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16.cpp)
-      add_example_dependencies(example_convnd_fwd_activ_xdl example_convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16)
-   set(target 1)
- endif()
-endforeach()

include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp

@@ -537,7 +537,6 @@ struct BlockwiseGemmWMMA
     // permutation
     static constexpr index_t A_KRow = AEnableLds ? 1 : 2;
     static constexpr index_t B_KRow = BEnableLds ? 1 : 2;
-
     static constexpr index_t A_K1   = ABlockDesc{}.GetLength(I5);
     static constexpr index_t B_K1   = BBlockDesc{}.GetLength(I5);