delay cvt to block_gemm

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

@@ -37,7 +37,9 @@ template <index_t BlockSize,
           index_t NPerXDL,
           index_t MRepeat,
           index_t NRepeat,
-          index_t KPack>
+          index_t KPack,
+          typename ComputeTypeA = FloatA,
+          typename ComputeTypeB = ComputeTypeA>
 struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
 {
     static constexpr auto I0 = Number<0>{};
@@ -59,7 +61,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
     static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
     static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
 
-    static constexpr auto xdlops_gemm = XdlopsGemm<FloatA, MPerXDL, NPerXDL, KPack, FloatB>{};
+    static constexpr auto xdlops_gemm =
+        XdlopsGemm<ComputeTypeA, MPerXDL, NPerXDL, KPack, ComputeTypeB>{};
 
     static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
 
@@ -295,9 +298,9 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
                         const BBlockBuffer& b_block_buf,
                         CThreadBuffer& c_thread_buf) const
     {
-        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
             a_thread_desc_.GetElementSpaceSize());
-        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeB>(
             b_thread_desc_.GetElementSpaceSize());
 
         static_for<0, MRepeat, 1>{}([&](auto m0) {
@@ -319,20 +322,20 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
                                    b_thread_buf);
 
                 static_for<0, KPerThread, KPack>{}([&](auto k) {
-                    vector_type<FloatA, KPack> a_thread_vec;
-                    vector_type<FloatB, KPack> b_thread_vec;
+                    vector_type<ComputeTypeA, KPack> a_thread_vec;
+                    vector_type<ComputeTypeB, KPack> b_thread_vec;
 
                     static_for<0, KPack, 1>{}([&](auto i) {
-                        a_thread_vec.template AsType<FloatA>()(i) = a_thread_buf
+                        a_thread_vec.template AsType<ComputeTypeA>()(i) = a_thread_buf
                             [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
-                        b_thread_vec.template AsType<FloatB>()(i) = b_thread_buf
+                        b_thread_vec.template AsType<ComputeTypeB>()(i) = b_thread_buf
                             [Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
                     });
 
                     using mfma_input_type_a =
-                        typename vector_type<FloatA, xdlops_gemm.K1PerXdlops>::type;
+                        typename vector_type<ComputeTypeA, xdlops_gemm.K1PerXdlops>::type;
                     using mfma_input_type_b =
-                        typename vector_type<FloatB, xdlops_gemm.K1PerXdlops>::type;
+                        typename vector_type<ComputeTypeB, xdlops_gemm.K1PerXdlops>::type;
 
                     constexpr index_t c_offset =
                         c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
@@ -360,7 +363,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
         make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
 
     using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
-                                                         FloatA,
+                                                         ComputeTypeA,
                                                          decltype(a_block_desc_m0_m1_m2_k),
                                                          decltype(a_thread_desc_),
                                                          Sequence<1, 1, 1, KPerThread>,
@@ -370,7 +373,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
                                                          A_K1>;
 
     using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
-                                                         FloatB,
+                                                         ComputeTypeB,
                                                          decltype(b_block_desc_n0_n1_n2_k),
                                                          decltype(b_thread_desc_),
                                                          Sequence<1, 1, 1, KPerThread>,
@@ -398,7 +401,9 @@ template <index_t BlockSize,
           index_t MRepeat,
           index_t NRepeat,
           index_t KPack,
-          index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS>
+          index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS,
+          typename ComputeTypeA  = FloatA,
+          typename ComputeTypeB  = ComputeTypeA>
 struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
     : public BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
                                                                  FloatA,
@@ -446,9 +451,9 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
                         const BBlockBuffer& b_block_buf,
                         CThreadBuffer& c_thread_buf) const
     {
-        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
             a_thread_desc_.GetElementSpaceSize());
-        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeB>(
             b_thread_desc_.GetElementSpaceSize());
 
         static_for<0, KPerThread, KPerInnerLoop>{}([&](auto k) {
@@ -485,22 +490,22 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
             static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
                 static_for<0, MRepeat, 1>{}([&](auto m0) {
                     static_for<0, NRepeat, 1>{}([&](auto n0) {
-                        vector_type<FloatA, KPack> a_thread_vec;
-                        vector_type<FloatB, KPack> b_thread_vec;
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
 
                         static_for<0, KPack, 1>{}([&](auto i) {
-                            a_thread_vec.template AsType<FloatA>()(i) =
+                            a_thread_vec.template AsType<ComputeTypeA>()(i) =
                                 a_thread_buf[Number<a_thread_desc_.CalculateOffset(
                                     make_tuple(m0, 0, 0, k_ + i))>{}];
-                            b_thread_vec.template AsType<FloatB>()(i) =
+                            b_thread_vec.template AsType<ComputeTypeB>()(i) =
                                 b_thread_buf[Number<b_thread_desc_.CalculateOffset(
                                     make_tuple(n0, 0, 0, k_ + i))>{}];
                         });
 
                         using mfma_input_type_a =
-                            typename vector_type<FloatA, xdlops_gemm.K1PerXdlops>::type;
+                            typename vector_type<ComputeTypeA, xdlops_gemm.K1PerXdlops>::type;
                         using mfma_input_type_b =
-                            typename vector_type<FloatB, xdlops_gemm.K1PerXdlops>::type;
+                            typename vector_type<ComputeTypeB, xdlops_gemm.K1PerXdlops>::type;
 
                         constexpr index_t c_offset =
                             c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
@@ -550,7 +555,7 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
         make_tuple(Number<NRepeat>{}, I1, I1, Number<KPerInnerLoop>{}));
 
     using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
-                                                         FloatA,
+                                                         ComputeTypeA,
                                                          decltype(a_block_desc_m0_m1_m2_k),
                                                          decltype(a_thread_desc_),
                                                          Sequence<1, 1, 1, KPerInnerLoop>,
@@ -560,7 +565,7 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
                                                          A_K1>;
 
     using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
-                                                         FloatB,
+                                                         ComputeTypeB,
                                                          decltype(b_block_desc_n0_n1_n2_k),
                                                          decltype(b_thread_desc_),
                                                          Sequence<1, 1, 1, KPerInnerLoop>,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp

@@ -400,7 +400,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
         constexpr auto c_block_size =
             GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock().GetElementSpaceSize();
 
-        return math::max((a_block_space_size + b_block_space_size) * sizeof(ComputeType),
+        return math::max(a_block_space_size * sizeof(FloatA) + b_block_space_size * sizeof(FloatB),
                          c_block_size * sizeof(FloatC));
     }
 
@@ -755,7 +755,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
                                                 ABlockTransferThreadClusterLengths_K0_M_K1,
                                                 ABlockTransferThreadClusterArrangeOrder,
                                                 FloatA,
-                                                ComputeType,
+                                                FloatA,
                                                 decltype(a_b_k0_m_k1_grid_desc),
                                                 decltype(a_b_k0_m_k1_block_desc),
                                                 ABlockTransferSrcAccessOrder,
@@ -785,7 +785,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
                                                 BBlockTransferThreadClusterLengths_K0_N_K1,
                                                 BBlockTransferThreadClusterArrangeOrder,
                                                 FloatB,
-                                                ComputeType,
+                                                FloatB,
                                                 decltype(b_b_k0_n_k1_grid_desc),
                                                 decltype(b_b_k0_n_k1_block_desc),
                                                 BBlockTransferSrcAccessOrder,
@@ -815,8 +815,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
 
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            ComputeType, // ComputeType A
-            ComputeType, // ComputeType B
+            FloatA,
+            FloatB,
             FloatAcc,
             decltype(a_k0_m_k1_block_desc),
             decltype(b_k0_n_k1_block_desc),
@@ -833,8 +833,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
         constexpr auto a_block_space_size =
             math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
 
-        ComputeType* p_a_block = static_cast<ComputeType*>(p_shared_block);
-        ComputeType* p_b_block = static_cast<ComputeType*>(p_shared_block) + a_block_space_size;
+        auto* p_a_block = static_cast<FloatA*>(p_shared_block);
+        auto* p_b_block = static_cast<FloatB*>(p_shared_block) + a_block_space_size;
 
         constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
         constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);

profiler/src/CMakeLists.txt

 # ckProfiler
 set(PROFILER_SOURCES
     profiler.cpp
-    profile_gemm.cpp
-    profile_gemm_splitk.cpp
-    profile_gemm_bias_add_reduce.cpp
-    profile_gemm_add_multiply.cpp
-    profile_gemm_multiply_add.cpp
-    profile_gemm_reduce.cpp
-    profile_batched_gemm.cpp
-    profile_batched_gemm_reduce.cpp
-    profile_conv_fwd.cpp
-    profile_conv_fwd_bias_relu.cpp
-    profile_conv_fwd_bias_relu_add.cpp
-    profile_conv_bwd_data.cpp
-    profile_grouped_conv_fwd.cpp
-    profile_grouped_conv_bwd_weight.cpp
-    profile_reduce.cpp
-    profile_groupnorm_fwd.cpp
-    profile_layernorm_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_conv_bwd_data.cpp
-    profile_conv_tensor_rearrange.cpp
+    # profile_gemm.cpp
+     profile_gemm_splitk.cpp
+    # profile_gemm_bias_add_reduce.cpp
+    # profile_gemm_add_multiply.cpp
+    # profile_gemm_multiply_add.cpp
+    # profile_gemm_reduce.cpp
+    # profile_batched_gemm.cpp
+    # profile_batched_gemm_reduce.cpp
+    # profile_conv_fwd.cpp
+    # profile_conv_fwd_bias_relu.cpp
+    # profile_conv_fwd_bias_relu_add.cpp
+    # profile_conv_bwd_data.cpp
+    # profile_grouped_conv_fwd.cpp
+    # profile_grouped_conv_bwd_weight.cpp
+    # profile_reduce.cpp
+    # profile_groupnorm_fwd.cpp
+    # profile_layernorm_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_conv_bwd_data.cpp
+    # profile_conv_tensor_rearrange.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()
-
-if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
-  list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
-  list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
-endif()
+#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()
+#
+#if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+#  list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
+#  list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
+#endif()
 
 set(PROFILER_EXECUTABLE ckProfiler)
 
@@ -57,60 +57,60 @@ add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES})
 target_compile_options(${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(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
-target_link_libraries(${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_reduce_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(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
-target_link_libraries(${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(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
-target_link_libraries(${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)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
-
-if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
-  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
-  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
-endif()
-
-
-
-if(DL_KERNELS)
-  target_link_libraries(${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()
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
+#target_link_libraries(${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_reduce_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(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
+#target_link_libraries(${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(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
+#target_link_libraries(${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)
+#target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
+#
+#if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+#  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
+#  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
+#endif()
+#
+#
+#
+#if(DL_KERNELS)
+#  target_link_libraries(${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)