pkgs/development/rocm-modules/composable_kernel/fix-empty-offload-targets.diff at r-updates

tjh.dev / nixpkgs

fork atom

nixpkgs mirror (for testing) github.com/NixOS/nixpkgs

nix

fork atom

nixpkgs / pkgs / development / rocm-modules / composable_kernel / fix-empty-offload-targets.diff

at r-updates 23 lines 1.3 kB view raw

wrap content

 1diff --git a/library/src/tensor_operation_instance/gpu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/CMakeLists.txt
 2index 172f6681b8..d3ddbb2f15 100644
 3--- a/library/src/tensor_operation_instance/gpu/CMakeLists.txt
 4+++ b/library/src/tensor_operation_instance/gpu/CMakeLists.txt
 5@@ -154,9 +154,15 @@ function(add_instance_library INSTANCE_NAME)
 6                 list(FILTER INST_TARGETS INCLUDE REGEX "gfx12")
 7             endif()
 8             set(offload_targets)
 9-            foreach(target IN LISTS INST_TARGETS)
10-                string(APPEND offload_targets "--offload-arch=${target} ")
11-            endforeach()
12+            if(NOT INST_TARGETS)
13+                # No valid GPU targets for this source, compile for 90a as fallback
14+                # so the add_device…instances and hip_fatbin… symbols still exist
15+                set(offload_targets "--offload-arch=gfx90a")
16+            else()
17+                foreach(target IN LISTS INST_TARGETS)
18+                    string(APPEND offload_targets "--offload-arch=${target} ")
19+                endforeach()
20+            endif()
21             set_source_files_properties(${source} PROPERTIES COMPILE_FLAGS ${offload_targets})
22             list(APPEND INST_OBJ ${source})
23         endforeach()