[openmp] Use llvm GridValues from devicertl

Add include path to the cmakefiles and set the target_impl enums
from the llvm constants instead of copying the values.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D108391

openmp/libomptarget/DeviceRTL/CMakeLists.txt

@@ -21,6 +21,12 @@ if (NOT (LIBOMPTARGET_DEP_CUDA_FOUND OR LIBOMPTARGET_BUILD_NVPTX_BCLIB))
   return()
 endif()
 
+if (NOT LIBOMPTARGET_LLVM_INCLUDE_DIRS)
+  libomptarget_say("Not building device RTL: Missing definition for LIBOMPTARGET_LLVM_INCLUDE_DIRS")
+  return()
+endif()
+
+
 # Check if we can create an LLVM bitcode implementation of the runtime library
 # that could be inlined in the user application. For that we need to find
 # a Clang compiler capable of compiling our CUDA files to LLVM bitcode and
@@ -132,6 +138,10 @@ set(src_files
 set(clang_opt_flags -O1 -mllvm -openmp-opt-disable -DSHARED_SCRATCHPAD_SIZE=2048)
 set(link_opt_flags  -O1        -openmp-opt-disable)
 
+# Prepend -I to each list element
+set (LIBOMPTARGET_LLVM_INCLUDE_DIRS_DEVICERTL "${LIBOMPTARGET_LLVM_INCLUDE_DIRS}")
+list(TRANSFORM LIBOMPTARGET_LLVM_INCLUDE_DIRS_DEVICERTL PREPEND "-I")
+
 # Set flags for LLVM Bitcode compilation.
 set(bc_flags -S -x c++ -std=c++17
               ${clang_opt_flags}
@@ -141,6 +151,7 @@ set(bc_flags -S -x c++ -std=c++17
              -fopenmp -fopenmp-cuda-mode -Xclang -fopenmp-is-device
              -Xclang -target-feature -Xclang +ptx61
              -I${include_directory}
+             ${LIBOMPTARGET_LLVM_INCLUDE_DIRS_DEVICERTL}
 )
 
 if(${LIBOMPTARGET_DEVICE_DEBUG})

openmp/libomptarget/DeviceRTL/src/Mapping.cpp

@@ -16,6 +16,8 @@
 
 #pragma omp declare target
 
+#include "llvm/Frontend/OpenMP/OMPGridValues.h"
+
 using namespace _OMP;
 
 namespace _OMP {
@@ -26,6 +28,10 @@ namespace impl {
 ///{
 #pragma omp begin declare variant match(device = {arch(amdgcn)})
 
+constexpr const llvm::omp::GV &getGridValue() {
+  return llvm::omp::AMDGPUGridValues;
+}
+
 uint32_t getGridDim(uint32_t n, uint16_t d) {
   uint32_t q = n / d;
   return q + (n > q * d);
@@ -86,8 +92,6 @@ uint32_t getWarpId() {
   return mapping::getThreadIdInBlock() / mapping::getWarpSize();
 }
 
-uint32_t getWarpSize() { return 64; }
-
 uint32_t getNumberOfWarpsInBlock() {
   return mapping::getBlockSize() / mapping::getWarpSize();
 }
@@ -101,6 +105,10 @@ uint32_t getNumberOfWarpsInBlock() {
 #pragma omp begin declare variant match(                                       \
     device = {arch(nvptx, nvptx64)}, implementation = {extension(match_any)})
 
+constexpr const llvm::omp::GV &getGridValue() {
+  return llvm::omp::NVPTXGridValues;
+}
+
 LaneMaskTy activemask() {
   unsigned int Mask;
   asm("activemask.b32 %0;" : "=r"(Mask));
@@ -144,8 +152,6 @@ uint32_t getWarpId() {
   return mapping::getThreadIdInBlock() / mapping::getWarpSize();
 }
 
-uint32_t getWarpSize() { return 32; }
-
 uint32_t getNumberOfWarpsInBlock() {
   return (mapping::getBlockSize() + mapping::getWarpSize() - 1) /
          mapping::getWarpSize();
@@ -154,6 +160,8 @@ uint32_t getNumberOfWarpsInBlock() {
 #pragma omp end declare variant
 ///}
 
+uint32_t getWarpSize() { return getGridValue().GV_Warp_Size; }
+
 } // namespace impl
 } // namespace _OMP
 

openmp/libomptarget/deviceRTLs/amdgcn/CMakeLists.txt

@@ -18,6 +18,12 @@ if (NOT LIBOMPTARGET_BUILD_AMDGCN_BCLIB)
   return()
 endif()
 
+if (NOT LIBOMPTARGET_LLVM_INCLUDE_DIRS)
+  libomptarget_say("Not building AMDGCN device RTL: Missing definition for LIBOMPTARGET_LLVM_INCLUDE_DIRS")
+  return()
+endif()
+
+
 # Copied from nvptx CMakeLists
 if(CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "x86_64")
   set(aux_triple x86_64-unknown-linux-gnu)
@@ -103,6 +109,10 @@ if (DEFINED LIBOMPTARGET_AMDGCN_GFXLIST)
   set(mcpus ${LIBOMPTARGET_AMDGCN_GFXLIST})
 endif()
 
+# Prepend -I to each list element
+set (LIBOMPTARGET_LLVM_INCLUDE_DIRS_AMDGCN "${LIBOMPTARGET_LLVM_INCLUDE_DIRS}")
+list(TRANSFORM LIBOMPTARGET_LLVM_INCLUDE_DIRS_AMDGCN PREPEND "-I")
+
 macro(add_cuda_bc_library)
   set(cu_cmd ${CLANG_TOOL}
     -xc++
@@ -123,7 +133,8 @@ macro(add_cuda_bc_library)
     ${CUDA_DEBUG}
     -I${CMAKE_CURRENT_SOURCE_DIR}/src
     -I${devicertl_base_directory}/common/include
-    -I${devicertl_base_directory})
+    -I${devicertl_base_directory}
+    ${LIBOMPTARGET_LLVM_INCLUDE_DIRS_AMDGCN})
 
   set(bc1_files)
 

openmp/libomptarget/deviceRTLs/amdgcn/src/target_impl.h

@@ -31,6 +31,12 @@ typedef uint64_t __kmpc_impl_lanemask_t;
 #define NOINLINE __attribute__((noinline))
 #define ALIGN(N) __attribute__((aligned(N)))
 
+#include "llvm/Frontend/OpenMP/OMPGridValues.h"
+
+INLINE constexpr const llvm::omp::GV &getGridValue() {
+  return llvm::omp::AMDGPUGridValues;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 // Kernel options
 ////////////////////////////////////////////////////////////////////////////////
@@ -38,9 +44,8 @@ typedef uint64_t __kmpc_impl_lanemask_t;
 ////////////////////////////////////////////////////////////////////////////////
 // The following def must match the absolute limit hardwired in the host RTL
 // max number of threads per team
-#define MAX_THREADS_PER_TEAM 1024
-
-#define WARPSIZE 64
+enum { MAX_THREADS_PER_TEAM = getGridValue().GV_Max_WG_Size };
+enum { WARPSIZE = getGridValue().GV_Warp_Size };
 
 // Maximum number of omp state objects per SM allocated statically in global
 // memory.
@@ -52,11 +57,11 @@ typedef uint64_t __kmpc_impl_lanemask_t;
 // Data sharing related quantities, need to match what is used in the compiler.
 enum DATA_SHARING_SIZES {
   // The size reserved for data in a shared memory slot.
-  DS_Slot_Size = 256,
+  DS_Slot_Size = getGridValue().GV_Slot_Size,
   // The slot size that should be reserved for a working warp.
-  DS_Worker_Warp_Slot_Size = WARPSIZE * DS_Slot_Size,
+  DS_Worker_Warp_Slot_Size = getGridValue().warpSlotSize(),
   // The maximum number of warps in use
-  DS_Max_Warp_Number = 16,
+  DS_Max_Warp_Number = getGridValue().maxWarpNumber(),
 };
 
 enum : __kmpc_impl_lanemask_t {

openmp/libomptarget/deviceRTLs/nvptx/CMakeLists.txt

@@ -19,6 +19,11 @@ if (NOT (LIBOMPTARGET_DEP_CUDA_FOUND OR LIBOMPTARGET_BUILD_NVPTX_BCLIB))
   return()
 endif()
 
+if (NOT LIBOMPTARGET_LLVM_INCLUDE_DIRS)
+  libomptarget_say("Not building NVPTX device RTL: Missing definition for LIBOMPTARGET_LLVM_INCLUDE_DIRS")
+  return()
+endif()
+
 # Check if we can create an LLVM bitcode implementation of the runtime library
 # that could be inlined in the user application. For that we need to find
 # a Clang compiler capable of compiling our CUDA files to LLVM bitcode and
@@ -151,6 +156,10 @@ set(cuda_src_files
   src/target_impl.cu
 )
 
+# Prepend -I to each list element
+set (LIBOMPTARGET_LLVM_INCLUDE_DIRS_NVPTX "${LIBOMPTARGET_LLVM_INCLUDE_DIRS}")
+list(TRANSFORM LIBOMPTARGET_LLVM_INCLUDE_DIRS_NVPTX PREPEND "-I")
+
 # Set flags for LLVM Bitcode compilation.
 set(bc_flags -S -x c++ -O1 -std=c++14
              -mllvm -openmp-opt-disable
@@ -162,7 +171,8 @@ set(bc_flags -S -x c++ -O1 -std=c++14
              -D__CUDACC__
              -I${devicertl_base_directory}
              -I${devicertl_common_directory}/include
-             -I${devicertl_nvptx_directory}/src)
+             -I${devicertl_nvptx_directory}/src
+             ${LIBOMPTARGET_LLVM_INCLUDE_DIRS_NVPTX})
 
 if(${LIBOMPTARGET_NVPTX_DEBUG})
   list(APPEND bc_flags -DOMPTARGET_NVPTX_DEBUG=-1 -g)

openmp/libomptarget/deviceRTLs/nvptx/src/target_impl.h

@@ -24,6 +24,12 @@ typedef uint32_t __kmpc_impl_lanemask_t;
 #define NOINLINE __attribute__((noinline))
 #define ALIGN(N) __attribute__((aligned(N)))
 
+#include "llvm/Frontend/OpenMP/OMPGridValues.h"
+
+INLINE constexpr const llvm::omp::GV &getGridValue() {
+  return llvm::omp::NVPTXGridValues;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 // Kernel options
 ////////////////////////////////////////////////////////////////////////////////
@@ -31,9 +37,8 @@ typedef uint32_t __kmpc_impl_lanemask_t;
 ////////////////////////////////////////////////////////////////////////////////
 // The following def must match the absolute limit hardwired in the host RTL
 // max number of threads per team
-#define MAX_THREADS_PER_TEAM 1024
-
-#define WARPSIZE 32
+enum { MAX_THREADS_PER_TEAM = getGridValue().GV_Max_WG_Size };
+enum { WARPSIZE = getGridValue().GV_Warp_Size };
 
 // Maximum number of omp state objects per SM allocated statically in global
 // memory.
@@ -64,11 +69,11 @@ typedef uint32_t __kmpc_impl_lanemask_t;
 // Data sharing related quantities, need to match what is used in the compiler.
 enum DATA_SHARING_SIZES {
   // The size reserved for data in a shared memory slot.
-  DS_Slot_Size = 256,
+  DS_Slot_Size = getGridValue().GV_Slot_Size,
   // The slot size that should be reserved for a working warp.
-  DS_Worker_Warp_Slot_Size = WARPSIZE * DS_Slot_Size,
+  DS_Worker_Warp_Slot_Size = getGridValue().warpSlotSize(),
   // The maximum number of warps in use
-  DS_Max_Warp_Number = 32,
+  DS_Max_Warp_Number = getGridValue().maxWarpNumber(),
 };
 
 enum : __kmpc_impl_lanemask_t {