From 7da9f64ed0f466156d4c41abf28fff89b2f41b54 Mon Sep 17 00:00:00 2001 From: apoorva Date: Wed, 11 Jun 2025 11:18:51 +0000 Subject: [PATCH] Adding instances and executables --- .../src/tensor_operation_instance/gpu/gemm_add/CMakeLists.txt | 2 ++ profiler/src/CMakeLists.txt | 2 ++ 2 files changed, 4 insertions(+) diff --git a/library/src/tensor_operation_instance/gpu/gemm_add/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add/CMakeLists.txt index 5f8a418b4a..46a2e56726 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_add/CMakeLists.txt @@ -4,4 +4,6 @@ add_instance_library(device_gemm_add_instance device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp device_gemm_add_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp device_gemm_add_wmma_c_shuffle_bf16_f16_bf16_bf16_mk_kn_mn_mn_instance.cpp + device_gemm_add_wmma_c_shuffle_v3_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp + device_gemm_add_wmma_c_shuffle_v3_bf16_f16_bf16_bf16_mk_kn_mn_mn_instance.cpp ) diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 2eb1038357..7c381a4093 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -65,6 +65,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") list(APPEND PROFILER_OPS profile_batched_gemm.cpp) list(APPEND PROFILER_OPS profile_batched_gemm_reduce.cpp) list(APPEND PROFILER_OPS profile_gemm_add_multiply.cpp) + list(APPEND PROFILER_OPS profile_gemm_add.cpp) list(APPEND PROFILER_OPS profile_gemm_bias_add_reduce.cpp) list(APPEND PROFILER_OPS profile_gemm_splitk.cpp) list(APPEND PROFILER_OPS profile_gemm_b_scale.cpp) @@ -179,6 +180,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") list(APPEND DEVICE_INSTANCES device_gemm_universal_reduce_instance) list(APPEND DEVICE_INSTANCES device_gemm_universal_streamk_instance) list(APPEND DEVICE_INSTANCES device_gemm_add_multiply_instance) + list(APPEND DEVICE_INSTANCES device_gemm_add_instance) list(APPEND DEVICE_INSTANCES device_gemm_reduce_instance) list(APPEND DEVICE_INSTANCES device_gemm_bias_add_reduce_instance) list(APPEND DEVICE_INSTANCES device_conv2d_fwd_instance)