match targets to tensile targets (#90) (#92)

* clarify targets and force them to match tensile
* tensile hipclang also needs the target compile flags

Co-authored-by: Eiden Yoshida <[email protected]>

Author: TorreZuk

CMakeLists.txt

@@ -203,14 +203,23 @@ if( CMAKE_CXX_COMPILER MATCHES ".*/hcc$" )
   find_package( hcc REQUIRED CONFIG PATHS ${ROCM_PATH} /opt/rocm )
   find_package( hip REQUIRED CONFIG PATHS ${ROCM_PATH} /opt/rocm )
 endif( )
+if( CMAKE_CXX_COMPILER MATCHES ".*/hipcc$" )
+  find_package( hip REQUIRED CONFIG PATHS ${ROCM_PATH} /opt/rocm )
+endif( )
 
 # CMake list of machine targets
-if( Tensile_ARCHITECTURE STREQUAL "all" )
-  set( gpus "gfx803;gfx900;gfx906;gfx908")
+if( Tensile_ARCHITECTURE )
+  if( Tensile_ARCHITECTURE STREQUAL "all" )
+    set( gpus "gfx803;gfx900;gfx906;gfx908")
+  else()
+    set( gpus ${Tensile_ARCHITECTURE})
+  endif()
+  # We must match tensile if architecture is set
+  set( ROCBLAS_GPUS ${gpus} )
 else()
-  set( gpus ${Tensile_ARCHITECTURE})
-endif()
-set( AMDGPU_TARGETS ${gpus} CACHE STRING "List of specific machine types for library to target" )
+  set( ROCBLAS_GPUS ${AMDGPU_TARGETS} )
+endif( )
+message(STATUS "Specific GPU TARGET(s): " ${ROCBLAS_GPUS})
 
 add_subdirectory( library )
 

library/src/CMakeLists.txt

@@ -42,7 +42,7 @@ if( BUILD_WITH_TENSILE )
   endif( )
 
   set( Tensile_RUNTIME_LANGUAGE "HIP" )
-  message( STATUS "AMDGPU_TARGETS=${AMDGPU_TARGETS}" )
+  message( STATUS "ROCBLAS_GPUS=${ROCBLAS_GPUS}" )
   if( BUILD_WITH_TENSILE_HOST )
     #TODO update when this feature has been validated
     #set( PACKAGE_TENSILE_LIBRARY ON )
@@ -84,7 +84,7 @@ if( BUILD_WITH_TENSILE )
     # Remove following when hcc is fixed; hcc emits following spurious warning ROCm v1.6.1
     # "clang-5.0: warning: argument unused during compilation: '-isystem /opt/rocm/include'"
     # target_compile_options( Tensile PRIVATE -Wno-unused-command-line-argument -fno-gpu-rdc -Wno-deprecated-declarations)
-    foreach( target ${AMDGPU_TARGETS} )
+    foreach( target ${ROCBLAS_GPUS} )
       if( BUILD_WITH_TENSILE_HOST )
           # Remove following when hcc is fixed; hcc emits following spurious warning ROCm v1.6.1
           # "clang-5.0: warning: argument unused during compilation: '-isystem /opt/rocm/include'"
@@ -99,6 +99,16 @@ if( BUILD_WITH_TENSILE )
     endforeach( )
   endif( )
 
+  if( CMAKE_CXX_COMPILER MATCHES ".*/hipcc$" )
+    foreach( target ${ROCBLAS_GPUS} )
+      if( BUILD_WITH_TENSILE_HOST )
+        target_compile_options( TensileHost PRIVATE --amdgpu-target=${target} )
+      else()
+        target_compile_options( Tensile PRIVATE --amdgpu-target=${target} )
+      endif()
+    endforeach( )
+  endif( )
+
   if( ROCBLAS_SHARED_LIBS )
     set( BUILD_SHARED_LIBS ON )
     if( BUILD_WITH_TENSILE_HOST )
@@ -326,8 +336,16 @@ if( CMAKE_CXX_COMPILER MATCHES ".*/hcc$" )
   # Remove following when hcc is fixed; hcc emits following spurious warning ROCm v1.6.1
   # "clang-5.0: warning: argument unused during compilation: '-isystem /opt/rocm/include'"
   target_compile_options( rocblas PRIVATE -Wno-unused-command-line-argument -fno-gpu-rdc )
-  foreach( target ${AMDGPU_TARGETS} )
+  foreach( target ${ROCBLAS_GPUS} )
+    target_compile_options( rocblas PRIVATE --amdgpu-target=${target} )
+  endforeach( )
+endif( )
+
+if( CMAKE_CXX_COMPILER MATCHES ".*/hipcc$" )
+  foreach( target ${ROCBLAS_GPUS} )
+    message( STATUS "TARGET: " ${target} )
     target_compile_options( rocblas PRIVATE --amdgpu-target=${target} )
+    target_link_libraries( rocblas PRIVATE --amdgpu-target=${target} )
   endforeach( )
 endif( )