Merge pull request #266 from scchan/gfx900_option

add gfx900 option, remove gpu code names
ROCm · Mar 16, 2017 · 810d137 · 810d137
2 parents f81fbb1 + 5417b23
commit 810d137
Show file tree

Hide file tree

Showing 6 changed files with 11 additions and 46 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -283,30 +283,8 @@ foreach(long_form IN LISTS tmp)
   list(APPEND HSA_AMDGPU_GPU_TARGET ${short_form})
 endforeach()
 
-# Map to LC-consumable name.
-foreach(AMDGPU_GPU_TARGET IN LISTS HSA_AMDGPU_GPU_TARGET)
-  # AMDGPU_TARGET_LIST holds the values which are used by the compiler scripts.
-  if (AMDGPU_GPU_TARGET STREQUAL "gfx700")
-    list(APPEND AMDGPU_TARGET_LIST "kaveri")
-  elseif (AMDGPU_GPU_TARGET STREQUAL "gfx701")
-    list(APPEND AMDGPU_TARGET_LIST "hawaii")
-  elseif (AMDGPU_GPU_TARGET STREQUAL "gfx801")
-    list(APPEND AMDGPU_TARGET_LIST "carrizo")
-  elseif (AMDGPU_GPU_TARGET STREQUAL "gfx802")
-    list(APPEND AMDGPU_TARGET_LIST "tonga")
-  elseif (AMDGPU_GPU_TARGET STREQUAL "gfx803")
-    list(APPEND AMDGPU_TARGET_LIST "fiji")
-  elseif (AMDGPU_GPU_TARGET STREQUAL "gfx900")
-    list(APPEND AMDGPU_TARGET_LIST "gfx900")
-  elseif (AMDGPU_GPU_TARGET STREQUAL "gfx901")
-    list(APPEND AMDGPU_TARGET_LIST "gfx901")
-  else ()
-    message(FATAL_ERROR "${AMDGPU_GPU_TARGET} is not a valid AMDPU target!")
-  endif ()
-endforeach(AMDGPU_GPU_TARGET)
-
 # join list
-string(REPLACE ";" " " AMDGPU_TARGET "${AMDGPU_TARGET_LIST}")
+string(REPLACE ";" " " AMDGPU_TARGET "${HSA_AMDGPU_GPU_TARGET}")
 MESSAGE(STATUS "HCC configured with AMDGPU targets: ${AMDGPU_TARGET}")
 
 #################

diff --git a/clang b/clang
diff --git a/lib/clamp-device.in b/lib/clamp-device.in
@@ -134,19 +134,19 @@ fi
 HCC_EXTRA_ARCH_FILE=""
 
 # select appropriate ROCm-Device-Libs per AMDGPU_TARGET
-if [ $AMDGPU_TARGET == "kaveri" ]; then
+if [ $AMDGPU_TARGET == "gfx700" ]; then
   OCLC_ISA_VERSION_LIB="$ROCM_LIB/oclc_isa_version_700.amdgcn.bc"
   HCC_EXTRA_ARCH_FILE=$HCC_EXTRA_LIBRARIES_GFX700
-elif [ $AMDGPU_TARGET == "hawaii" ]; then
+elif [ $AMDGPU_TARGET == "gfx701" ]; then
   OCLC_ISA_VERSION_LIB="$ROCM_LIB/oclc_isa_version_701.amdgcn.bc"
   HCC_EXTRA_ARCH_FILE=$HCC_EXTRA_LIBRARIES_GFX701
-elif [ $AMDGPU_TARGET == "carrizo" ]; then
+elif [ $AMDGPU_TARGET == "gfx801" ]; then
   OCLC_ISA_VERSION_LIB="$ROCM_LIB/oclc_isa_version_801.amdgcn.bc"
   HCC_EXTRA_ARCH_FILE=$HCC_EXTRA_LIBRARIES_GFX801
-elif [ $AMDGPU_TARGET == "tonga" ]; then
+elif [ $AMDGPU_TARGET == "gfx802" ]; then
   OCLC_ISA_VERSION_LIB="$ROCM_LIB/oclc_isa_version_802.amdgcn.bc"
   HCC_EXTRA_ARCH_FILE=$HCC_EXTRA_LIBRARIES_GFX802
-elif [ $AMDGPU_TARGET == "fiji" ]; then
+elif [ $AMDGPU_TARGET == "gfx803" ]; then
   OCLC_ISA_VERSION_LIB="$ROCM_LIB/oclc_isa_version_803.amdgcn.bc"
   HCC_EXTRA_ARCH_FILE=$HCC_EXTRA_LIBRARIES_GFX803
 elif [ $AMDGPU_TARGET == "gfx900" ]; then

diff --git a/lib/clamp-link.in b/lib/clamp-link.in
@@ -327,23 +327,10 @@ if [ -n "$HCC_AMDGPU_TARGET" ]; then
   IFS=','
   HCC_AMDGPU_TARGET_TOKENS=($HCC_AMDGPU_TARGET)
   unset IFS
-
   # parse individual token, override AMDGPU_TARGET_ARRAY
   iter=0
   for HCC_AMDGPU_TARGET_TOKEN in ${HCC_AMDGPU_TARGET_TOKENS[@]}; do
-    # map ISA version string to GPU family
-    if [ "$HCC_AMDGPU_TARGET_TOKEN" = "gfx701" ]; then
-      AMDGPU_TARGET_ARRAY[$iter]="hawaii"
-    elif [ "$HCC_AMDGPU_TARGET_TOKEN" = "gfx801" ]; then
-      AMDGPU_TARGET_ARRAY[$iter]="carrizo"
-    elif [ "$HCC_AMDGPU_TARGET_TOKEN" = "gfx802" ]; then
-      AMDGPU_TARGET_ARRAY[$iter]="tonga"
-    elif [ "$HCC_AMDGPU_TARGET_TOKEN" = "gfx803" ]; then
-      AMDGPU_TARGET_ARRAY[$iter]="fiji"
-    else
-      # not mappable, simply pass to backend and wish it accepts
-      AMDGPU_TARGET_ARRAY[$iter]=$HCC_AMDGPU_TARGET
-    fi
+    AMDGPU_TARGET_ARRAY[$iter]=$HCC_AMDGPU_TARGET
     ((iter++))
   done
 fi

diff --git a/tests/benchEmptyKernel/bench.cpp b/tests/benchEmptyKernel/bench.cpp
@@ -97,7 +97,7 @@ void explicit_launch_null_kernel(const grid_launch_parm *lp, const Kernel &k)
 void time_dispatch_hsa_kernel(std::string testName, const grid_launch_parm *lp, const char *nullkernel_hsaco_dir)
 {
   std::string nullkernel_hsaco(nullkernel_hsaco_dir);
-  nullkernel_hsaco += "/nullkernel-fiji.hsaco";
+  nullkernel_hsaco += "/nullkernel-gfx803.hsaco";
   Kernel k = load_hsaco(lp->av, nullkernel_hsaco.c_str(), KERNEL_NAME);
   std::chrono::time_point<std::chrono::high_resolution_clock> start, end;
 

diff --git a/tests/benchEmptyKernel/nullkernel.cpp b/tests/benchEmptyKernel/nullkernel.cpp
@@ -1,5 +1,5 @@
-// RUN: %hc --amdgpu-target=gfx701 --amdgpu-target=gfx801 --amdgpu-target=gfx802 --amdgpu-target=gfx803 -fPIC -Wl,-Bsymbolic -shared %S/nullkernel.cpp -o %T/nullkernel
-// RUN: %extractkernel -i %T/nullkernel
+// RUN: %hc --amdgpu-target=gfx801 --amdgpu-target=gfx802 --amdgpu-target=gfx803 -fPIC -Wl,-Bsymbolic -shared %S/nullkernel.cpp -o %T/nullkernel
+// RUN: HCC_HOME=%llvm_libs_dir/../../ %extractkernel -i %T/nullkernel
 
 #include "hc.hpp"
 #include "grid_launch.hpp"