[OpenMP] Use generic IR for the OpenMP DeviceRTL

jhuber6 · jhuber6 · commit 4c710e49eea9 · 2024-12-07T13:55:43.000-06:00
Summary:
We previously built this for every single architecture to deal with
incompatibility. This patch updates it to use the 'generic' IR that
`libc` and other projects use. Who knows if this will have any
side-effects, probably worth testing more but it passes the tests I
expect to pass on my side.
diff --git a/clang/lib/Driver/ToolChains/CommonArgs.cpp b/clang/lib/Driver/ToolChains/CommonArgs.cpp
@@ -2800,8 +2800,7 @@ void tools::addOpenMPDeviceRTL(const Driver &D,
                         : options::OPT_libomptarget_nvptx_bc_path_EQ;
 
   StringRef ArchPrefix = Triple.isAMDGCN() ? "amdgpu" : "nvptx";
-  std::string LibOmpTargetName =
-      ("libomptarget-" + ArchPrefix + "-" + BitcodeSuffix + ".bc").str();
+  std::string LibOmpTargetName = ("libomptarget-" + ArchPrefix + ".bc").str();
 
   // First check whether user specifies bc library
   if (const Arg *A = DriverArgs.getLastArg(LibomptargetBCPathOpt)) {
diff --git a/clang/lib/Driver/ToolChains/Cuda.cpp b/clang/lib/Driver/ToolChains/Cuda.cpp
@@ -851,7 +851,6 @@ void CudaToolChain::addClangTargetOptions(
   HostTC.addClangTargetOptions(DriverArgs, CC1Args, DeviceOffloadingKind);
 
   StringRef GpuArch = DriverArgs.getLastArgValue(options::OPT_march_EQ);
-  assert(!GpuArch.empty() && "Must have an explicit GPU arch.");
   assert((DeviceOffloadingKind == Action::OFK_OpenMP ||
           DeviceOffloadingKind == Action::OFK_Cuda) &&
          "Only OpenMP or CUDA offloading kinds are supported for NVIDIA GPUs.");
diff --git a/offload/DeviceRTL/CMakeLists.txt b/offload/DeviceRTL/CMakeLists.txt
@@ -42,38 +42,6 @@ set(devicertl_base_directory ${CMAKE_CURRENT_SOURCE_DIR})
 set(include_directory ${devicertl_base_directory}/include)
 set(source_directory ${devicertl_base_directory}/src)
 
-set(all_amdgpu_architectures "gfx700;gfx701;gfx801;gfx803;gfx900;gfx902;gfx906"
-                             "gfx908;gfx90a;gfx90c;gfx940;gfx941;gfx942;gfx950;gfx1010"
-                             "gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035"
-                             "gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150"
-                             "gfx1151;gfx1152;gfx1153")
-set(all_nvptx_architectures "sm_35;sm_37;sm_50;sm_52;sm_53;sm_60;sm_61;sm_62"
-                            "sm_70;sm_72;sm_75;sm_80;sm_86;sm_87;sm_89;sm_90")
-set(all_gpu_architectures
-    "${all_amdgpu_architectures};${all_nvptx_architectures}")
-
-set(LIBOMPTARGET_DEVICE_ARCHITECTURES "all" CACHE STRING
-    "List of device architectures to be used to compile the OpenMP DeviceRTL.")
-
-if(LIBOMPTARGET_DEVICE_ARCHITECTURES STREQUAL "all")
-  set(LIBOMPTARGET_DEVICE_ARCHITECTURES ${all_gpu_architectures})
-elseif(LIBOMPTARGET_DEVICE_ARCHITECTURES STREQUAL "amdgpu")
-  set(LIBOMPTARGET_DEVICE_ARCHITECTURES ${all_amdgpu_architectures})
-elseif(LIBOMPTARGET_DEVICE_ARCHITECTURES STREQUAL "nvptx")
-  set(LIBOMPTARGET_DEVICE_ARCHITECTURES ${all_nvptx_architectures})
-elseif(LIBOMPTARGET_DEVICE_ARCHITECTURES STREQUAL "auto" OR
-       LIBOMPTARGET_DEVICE_ARCHITECTURES STREQUAL "native")
-  if(NOT LIBOMPTARGET_NVPTX_ARCH AND NOT LIBOMPTARGET_AMDGPU_ARCH)
-    message(FATAL_ERROR
-      "Could not find 'amdgpu-arch' and 'nvptx-arch' tools required for 'auto'")
-  elseif(NOT LIBOMPTARGET_FOUND_NVIDIA_GPU AND NOT LIBOMPTARGET_FOUND_AMDGPU_GPU)
-    message(FATAL_ERROR "No AMD or NVIDIA GPU found on the system when using 'auto'")
-  endif()
-  set(LIBOMPTARGET_DEVICE_ARCHITECTURES
-      "${LIBOMPTARGET_NVPTX_DETECTED_ARCH_LIST};${LIBOMPTARGET_AMDGPU_DETECTED_ARCH_LIST}")
-endif()
-list(REMOVE_DUPLICATES LIBOMPTARGET_DEVICE_ARCHITECTURES)
-
 set(include_files
   ${include_directory}/Allocator.h
   ${include_directory}/Configuration.h
@@ -141,20 +109,21 @@ set(bc_flags -c -foffload-lto -std=c++17 -fvisibility=hidden
 
 # first create an object target
 add_library(omptarget.devicertl.all_objs OBJECT IMPORTED)
-function(compileDeviceRTLLibrary target_cpu target_name target_triple)
+function(compileDeviceRTLLibrary target_name target_triple)
   set(target_bc_flags ${ARGN})
 
   set(bc_files "")
   foreach(src ${src_files})
     get_filename_component(infile ${src} ABSOLUTE)
     get_filename_component(outfile ${src} NAME)
-    set(outfile "${outfile}-${target_cpu}.bc")
+    set(outfile "${outfile}-${target_name}.bc")
     set(depfile "${outfile}.d")
 
     add_custom_command(OUTPUT ${outfile}
       COMMAND ${CLANG_TOOL}
       ${bc_flags}
-      --offload-arch=${target_cpu}
+      -fopenmp-targets=${target_triple}
+      -Xopenmp-target=${target_triple} -march=
       ${target_bc_flags}
       -MD -MF ${depfile}
       ${infile} -o ${outfile}
@@ -177,7 +146,7 @@ function(compileDeviceRTLLibrary target_cpu target_name target_triple)
     list(APPEND bc_files ${outfile})
   endforeach()
 
-  set(bclib_name "libomptarget-${target_name}-${target_cpu}.bc")
+  set(bclib_name "libomptarget-${target_name}.bc")
 
   # Link to a bitcode library.
   add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/linked_${bclib_name}
@@ -217,7 +186,7 @@ function(compileDeviceRTLLibrary target_cpu target_name target_triple)
       APPEND)
   endif()
 
-  set(bclib_target_name "omptarget-${target_name}-${target_cpu}-bc")
+  set(bclib_target_name "omptarget-${target_name}-bc")
   add_custom_target(${bclib_target_name} DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name})
 
   # Copy library to destination.
@@ -239,7 +208,7 @@ function(compileDeviceRTLLibrary target_cpu target_name target_triple)
   # Package the bitcode in the bitcode and embed it in an ELF for the static library
   add_custom_command(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/packaged_${bclib_name}
       COMMAND ${PACKAGER_TOOL} -o ${CMAKE_CURRENT_BINARY_DIR}/packaged_${bclib_name}
-        "--image=file=${CMAKE_CURRENT_BINARY_DIR}/${bclib_name},${target_feature},triple=${target_triple},arch=${target_cpu},kind=openmp"
+        "--image=file=${CMAKE_CURRENT_BINARY_DIR}/${bclib_name},${target_feature},triple=${target_triple},arch=generic,kind=openmp"
       DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${bclib_name}
       COMMENT "Packaging LLVM offloading binary ${bclib_name}.out"
   )
@@ -249,14 +218,14 @@ function(compileDeviceRTLLibrary target_cpu target_name target_triple)
       APPEND)
   endif()
 
-  set(output_name "${CMAKE_CURRENT_BINARY_DIR}/devicertl-${target_name}-${target_cpu}.o")
+  set(output_name "${CMAKE_CURRENT_BINARY_DIR}/devicertl-${target_name}.o")
   add_custom_command(OUTPUT ${output_name}
     COMMAND ${CLANG_TOOL} --std=c++17 -c -nostdlib
             -Xclang -fembed-offload-object=${CMAKE_CURRENT_BINARY_DIR}/packaged_${bclib_name}
             -o ${output_name}
             ${source_directory}/Stub.cpp
     DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/packaged_${bclib_name} ${source_directory}/Stub.cpp
-    COMMENT "Embedding LLVM offloading binary in devicertl-${target_name}-${target_cpu}.o"
+    COMMENT "Embedding LLVM offloading binary in devicertl-${target_name}.o"
     VERBATIM
   )
   if(TARGET clang)
@@ -269,11 +238,11 @@ function(compileDeviceRTLLibrary target_cpu target_name target_triple)
   set_property(TARGET omptarget.devicertl.all_objs APPEND PROPERTY IMPORTED_OBJECTS ${output_name})
 
   if (CMAKE_EXPORT_COMPILE_COMMANDS)
-    set(ide_target_name omptarget-ide-${target_name}-${target_cpu})
+    set(ide_target_name omptarget-ide-${target_name})
     add_library(${ide_target_name} STATIC EXCLUDE_FROM_ALL ${src_files})
     target_compile_options(${ide_target_name} PRIVATE
-      -fopenmp --offload-arch=${target_cpu} -fopenmp-cuda-mode
-      -mllvm -openmp-opt-disable
+      -fopenmp-targets=${target_triple} -Xopenmp-target=${target_triple} -march=
+      -fopenmp -fopenmp-cuda-mode -mllvm -openmp-opt-disable
       -foffload-lto -fvisibility=hidden --offload-device-only
       -nocudalib -nogpulib -nogpuinc -nostdlibinc -Wno-unknown-cuda-version
     )
@@ -288,18 +257,11 @@ function(compileDeviceRTLLibrary target_cpu target_name target_triple)
   endif()
 endfunction()
 
-# Generate a Bitcode library for all the gpu architectures the user requested.
-add_custom_target(omptarget.devicertl.nvptx)
 add_custom_target(omptarget.devicertl.amdgpu)
-foreach(gpu_arch ${LIBOMPTARGET_DEVICE_ARCHITECTURES})
-  if("${gpu_arch}" IN_LIST all_amdgpu_architectures)
-    compileDeviceRTLLibrary(${gpu_arch} amdgpu amdgcn-amd-amdhsa -Xclang -mcode-object-version=none)
-  elseif("${gpu_arch}" IN_LIST all_nvptx_architectures)
-    compileDeviceRTLLibrary(${gpu_arch} nvptx nvptx64-nvidia-cuda --cuda-feature=+ptx63)
-  else()
-    message(FATAL_ERROR "Unknown GPU architecture '${gpu_arch}'")
-  endif()
-endforeach()
+compileDeviceRTLLibrary(amdgpu amdgcn-amd-amdhsa -Xclang -mcode-object-version=none)
+
+add_custom_target(omptarget.devicertl.nvptx)
+compileDeviceRTLLibrary(nvptx nvptx64-nvidia-cuda --cuda-feature=+ptx63)
 
 # Archive all the object files generated above into a static library
 add_library(omptarget.devicertl STATIC)
diff --git a/offload/DeviceRTL/src/Reduction.cpp b/offload/DeviceRTL/src/Reduction.cpp
@@ -44,7 +44,6 @@ void gpu_irregular_warp_reduce(void *reduce_data, ShuffleReductFnTy shflFct,
   }
 }
 
-#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 700
 static uint32_t gpu_irregular_simd_reduce(void *reduce_data,
                                           ShuffleReductFnTy shflFct) {
   uint32_t size, remote_id, physical_lane_id;
@@ -63,7 +62,6 @@ static uint32_t gpu_irregular_simd_reduce(void *reduce_data,
   } while (logical_lane_id % 2 == 0 && size > 1);
   return (logical_lane_id == 0);
 }
-#endif
 
 static int32_t nvptx_parallel_reduce_nowait(void *reduce_data,
                                             ShuffleReductFnTy shflFct,
@@ -74,49 +72,53 @@ static int32_t nvptx_parallel_reduce_nowait(void *reduce_data,
   uint32_t NumThreads = omp_get_num_threads();
   if (NumThreads == 1)
     return 1;
-    /*
-     * This reduce function handles reduction within a team. It handles
-     * parallel regions in both L1 and L2 parallelism levels. It also
-     * supports Generic, SPMD, and NoOMP modes.
-     *
-     * 1. Reduce within a warp.
-     * 2. Warp master copies value to warp 0 via shared memory.
-     * 3. Warp 0 reduces to a single value.
-     * 4. The reduced value is available in the thread that returns 1.
-     */
-
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
-  uint32_t WarpsNeeded =
-      (NumThreads + mapping::getWarpSize() - 1) / mapping::getWarpSize();
-  uint32_t WarpId = mapping::getWarpIdInBlock();
-
-  // Volta execution model:
-  // For the Generic execution mode a parallel region either has 1 thread and
-  // beyond that, always a multiple of 32. For the SPMD execution mode we may
-  // have any number of threads.
-  if ((NumThreads % mapping::getWarpSize() == 0) || (WarpId < WarpsNeeded - 1))
-    gpu_regular_warp_reduce(reduce_data, shflFct);
-  else if (NumThreads > 1) // Only SPMD execution mode comes thru this case.
-    gpu_irregular_warp_reduce(reduce_data, shflFct,
-                              /*LaneCount=*/NumThreads % mapping::getWarpSize(),
-                              /*LaneId=*/mapping::getThreadIdInBlock() %
-                                  mapping::getWarpSize());
 
-  // When we have more than [mapping::getWarpSize()] number of threads
-  // a block reduction is performed here.
   //
-  // Only L1 parallel region can enter this if condition.
-  if (NumThreads > mapping::getWarpSize()) {
-    // Gather all the reduced values from each warp
-    // to the first warp.
-    cpyFct(reduce_data, WarpsNeeded);
+  // This reduce function handles reduction within a team. It handles
+  // parallel regions in both L1 and L2 parallelism levels. It also
+  // supports Generic, SPMD, and NoOMP modes.
+  //
+  // 1. Reduce within a warp.
+  // 2. Warp master copies value to warp 0 via shared memory.
+  // 3. Warp 0 reduces to a single value.
+  // 4. The reduced value is available in the thread that returns 1.
+  //
 
-    if (WarpId == 0)
-      gpu_irregular_warp_reduce(reduce_data, shflFct, WarpsNeeded,
-                                BlockThreadId);
+#if __has_builtin(__nvvm_reflect)
+  if (__nvvm_reflect("__CUDA_ARCH") >= 700) {
+    uint32_t WarpsNeeded =
+        (NumThreads + mapping::getWarpSize() - 1) / mapping::getWarpSize();
+    uint32_t WarpId = mapping::getWarpIdInBlock();
+
+    // Volta execution model:
+    // For the Generic execution mode a parallel region either has 1 thread and
+    // beyond that, always a multiple of 32. For the SPMD execution mode we may
+    // have any number of threads.
+    if ((NumThreads % mapping::getWarpSize() == 0) ||
+        (WarpId < WarpsNeeded - 1))
+      gpu_regular_warp_reduce(reduce_data, shflFct);
+    else if (NumThreads > 1) // Only SPMD execution mode comes thru this case.
+      gpu_irregular_warp_reduce(
+          reduce_data, shflFct,
+          /*LaneCount=*/NumThreads % mapping::getWarpSize(),
+          /*LaneId=*/mapping::getThreadIdInBlock() % mapping::getWarpSize());
+
+    // When we have more than [mapping::getWarpSize()] number of threads
+    // a block reduction is performed here.
+    //
+    // Only L1 parallel region can enter this if condition.
+    if (NumThreads > mapping::getWarpSize()) {
+      // Gather all the reduced values from each warp
+      // to the first warp.
+      cpyFct(reduce_data, WarpsNeeded);
+
+      if (WarpId == 0)
+        gpu_irregular_warp_reduce(reduce_data, shflFct, WarpsNeeded,
+                                  BlockThreadId);
+    }
+    return BlockThreadId == 0;
   }
-  return BlockThreadId == 0;
-#else
+#endif
   __kmpc_impl_lanemask_t Liveness = mapping::activemask();
   if (Liveness == lanes::All) // Full warp
     gpu_regular_warp_reduce(reduce_data, shflFct);
@@ -150,10 +152,9 @@ static int32_t nvptx_parallel_reduce_nowait(void *reduce_data,
     return BlockThreadId == 0;
   }
 
-  // Get the OMP thread Id. This is different from BlockThreadId in the case of
-  // an L2 parallel region.
+  // Get the OMP thread Id. This is different from BlockThreadId in the case
+  // of an L2 parallel region.
   return BlockThreadId == 0;
-#endif // __CUDA_ARCH__ >= 700
 }
 
 uint32_t roundToWarpsize(uint32_t s) {
