[clang] 78f92c3 - [openmp][amdgpu] Initial gfx10 offloading implementation
Jon Chesterfield via cfe-commits
cfe-commits at lists.llvm.org
Fri Aug 27 04:34:22 PDT 2021
Author: Jon Chesterfield
Date: 2021-08-27T12:34:03+01:00
New Revision: 78f92c38101fd1f6788500b3362d3c9c28213bc0
URL: https://github.com/llvm/llvm-project/commit/78f92c38101fd1f6788500b3362d3c9c28213bc0
DIFF: https://github.com/llvm/llvm-project/commit/78f92c38101fd1f6788500b3362d3c9c28213bc0.diff
LOG: [openmp][amdgpu] Initial gfx10 offloading implementation
Lets wavefront size be 32 for amdgpu openmp, as well as 64.
Fixes up as little as possible to pass that through the libraries. This change
is end to end, as opposed to updating clang/devicertl/plugin separately. It can
be broken up for review/commit if preferred. Posting as-is so that others with
a gfx10 can try it out. It works roughly as well as gfx9 for me, but there are
probably bugs remaining as well as the todo: for letting grid values vary more.
Reviewed By: ronlieb
Differential Revision: https://reviews.llvm.org/D108708
Added:
Modified:
clang/lib/Basic/Targets/AMDGPU.h
llvm/include/llvm/Frontend/OpenMP/OMPGridValues.h
openmp/libomptarget/deviceRTLs/amdgcn/CMakeLists.txt
openmp/libomptarget/deviceRTLs/amdgcn/src/target_impl.h
openmp/libomptarget/plugins/amdgpu/src/rtl.cpp
Removed:
################################################################################
diff --git a/clang/lib/Basic/Targets/AMDGPU.h b/clang/lib/Basic/Targets/AMDGPU.h
index e791a83f38ae7..700b76452eea2 100644
--- a/clang/lib/Basic/Targets/AMDGPU.h
+++ b/clang/lib/Basic/Targets/AMDGPU.h
@@ -371,7 +371,14 @@ class LLVM_LIBRARY_VISIBILITY AMDGPUTargetInfo final : public TargetInfo {
}
const llvm::omp::GV &getGridValue() const override {
- return llvm::omp::AMDGPUGridValues;
+ switch (WavefrontSize) {
+ case 32:
+ return llvm::omp::getAMDGPUGridValues<32>();
+ case 64:
+ return llvm::omp::getAMDGPUGridValues<64>();
+ default:
+ llvm_unreachable("getGridValue not implemented for this wavesize");
+ }
}
/// \returns Target specific vtbl ptr address space.
diff --git a/llvm/include/llvm/Frontend/OpenMP/OMPGridValues.h b/llvm/include/llvm/Frontend/OpenMP/OMPGridValues.h
index f5674e443b67e..89f5de229b3b1 100644
--- a/llvm/include/llvm/Frontend/OpenMP/OMPGridValues.h
+++ b/llvm/include/llvm/Frontend/OpenMP/OMPGridValues.h
@@ -81,7 +81,7 @@ struct GV {
};
/// For AMDGPU GPUs
-static constexpr GV AMDGPUGridValues = {
+static constexpr GV AMDGPUGridValues64 = {
256, // GV_Slot_Size
64, // GV_Warp_Size
128, // GV_Max_Teams
@@ -90,6 +90,20 @@ static constexpr GV AMDGPUGridValues = {
256, // GV_Default_WG_Size
};
+static constexpr GV AMDGPUGridValues32 = {
+ 256, // GV_Slot_Size
+ 32, // GV_Warp_Size
+ 128, // GV_Max_Teams
+ 896, // GV_SimpleBufferSize
+ 1024, // GV_Max_WG_Size,
+ 256, // GV_Default_WG_Size
+};
+
+template <unsigned wavesize> constexpr const GV &getAMDGPUGridValues() {
+ static_assert(wavesize == 32 || wavesize == 64, "");
+ return wavesize == 32 ? AMDGPUGridValues32 : AMDGPUGridValues64;
+}
+
/// For Nvidia GPUs
static constexpr GV NVPTXGridValues = {
256, // GV_Slot_Size
diff --git a/openmp/libomptarget/deviceRTLs/amdgcn/CMakeLists.txt b/openmp/libomptarget/deviceRTLs/amdgcn/CMakeLists.txt
index 7d10b1edabe04..a0aa471594f70 100644
--- a/openmp/libomptarget/deviceRTLs/amdgcn/CMakeLists.txt
+++ b/openmp/libomptarget/deviceRTLs/amdgcn/CMakeLists.txt
@@ -107,7 +107,7 @@ else()
endif()
# create libraries
-set(mcpus gfx700 gfx701 gfx801 gfx803 gfx900 gfx902 gfx906 gfx908)
+set(mcpus gfx700 gfx701 gfx801 gfx803 gfx900 gfx902 gfx906 gfx908 gfx1010 gfx1030 gfx1031)
if (DEFINED LIBOMPTARGET_AMDGCN_GFXLIST)
set(mcpus ${LIBOMPTARGET_AMDGCN_GFXLIST})
endif()
diff --git a/openmp/libomptarget/deviceRTLs/amdgcn/src/target_impl.h b/openmp/libomptarget/deviceRTLs/amdgcn/src/target_impl.h
index 5396590d46e00..f4755996c86aa 100644
--- a/openmp/libomptarget/deviceRTLs/amdgcn/src/target_impl.h
+++ b/openmp/libomptarget/deviceRTLs/amdgcn/src/target_impl.h
@@ -38,7 +38,7 @@ typedef uint64_t __kmpc_impl_lanemask_t;
#include "llvm/Frontend/OpenMP/OMPGridValues.h"
INLINE constexpr const llvm::omp::GV &getGridValue() {
- return llvm::omp::AMDGPUGridValues;
+ return llvm::omp::getAMDGPUGridValues<__AMDGCN_WAVEFRONT_SIZE>();
}
////////////////////////////////////////////////////////////////////////////////
diff --git a/openmp/libomptarget/plugins/amdgpu/src/rtl.cpp b/openmp/libomptarget/plugins/amdgpu/src/rtl.cpp
index 15f67cea1ea98..2b131a2784925 100644
--- a/openmp/libomptarget/plugins/amdgpu/src/rtl.cpp
+++ b/openmp/libomptarget/plugins/amdgpu/src/rtl.cpp
@@ -435,8 +435,9 @@ struct EnvironmentVariables {
int MaxTeamsDefault;
};
+template <uint32_t wavesize>
static constexpr const llvm::omp::GV &getGridValue() {
- return llvm::omp::AMDGPUGridValues;
+ return llvm::omp::getAMDGPUGridValues<wavesize>();
}
/// Class containing all the device information
@@ -507,10 +508,24 @@ class RTLDeviceInfoTy {
static const unsigned HardTeamLimit =
(1 << 16) - 1; // 64K needed to fit in uint16
static const int DefaultNumTeams = 128;
- static const int Max_Teams = getGridValue().GV_Max_Teams;
- static const int Warp_Size = getGridValue().GV_Warp_Size;
- static const int Max_WG_Size = getGridValue().GV_Max_WG_Size;
- static const int Default_WG_Size = getGridValue().GV_Default_WG_Size;
+
+ // These need to be per-device since
diff erent devices can have
diff erent
+ // wave sizes, but are currently the same number for each so that refactor
+ // can be postponed.
+ static_assert(getGridValue<32>().GV_Max_Teams ==
+ getGridValue<64>().GV_Max_Teams,
+ "");
+ static const int Max_Teams = getGridValue<64>().GV_Max_Teams;
+
+ static_assert(getGridValue<32>().GV_Max_WG_Size ==
+ getGridValue<64>().GV_Max_WG_Size,
+ "");
+ static const int Max_WG_Size = getGridValue<64>().GV_Max_WG_Size;
+
+ static_assert(getGridValue<32>().GV_Default_WG_Size ==
+ getGridValue<64>().GV_Default_WG_Size,
+ "");
+ static const int Default_WG_Size = getGridValue<64>().GV_Default_WG_Size;
using MemcpyFunc = hsa_status_t (*)(hsa_signal_t, void *, const void *,
size_t size, hsa_agent_t,
@@ -1059,8 +1074,9 @@ int32_t __tgt_rtl_init_device(int device_id) {
DP("Queried wavefront size: %d\n", wavefront_size);
DeviceInfo.WarpSize[device_id] = wavefront_size;
} else {
- DP("Default wavefront size: %d\n", getGridValue().GV_Warp_Size);
- DeviceInfo.WarpSize[device_id] = getGridValue().GV_Warp_Size;
+ // TODO: Burn the wavefront size into the code object
+ DP("Warning: Unknown wavefront size, assuming 64\n");
+ DeviceInfo.WarpSize[device_id] = 64;
}
// Adjust teams to the env variables
@@ -1885,9 +1901,10 @@ struct launchVals {
int WorkgroupSize;
int GridSize;
};
-launchVals getLaunchVals(EnvironmentVariables Env, int ConstWGSize,
- int ExecutionMode, int num_teams, int thread_limit,
- uint64_t loop_tripcount, int DeviceNumTeams) {
+launchVals getLaunchVals(int WarpSize, EnvironmentVariables Env,
+ int ConstWGSize, int ExecutionMode, int num_teams,
+ int thread_limit, uint64_t loop_tripcount,
+ int DeviceNumTeams) {
int threadsPerGroup = RTLDeviceInfoTy::Default_WG_Size;
int num_groups = 0;
@@ -1900,7 +1917,7 @@ launchVals getLaunchVals(EnvironmentVariables Env, int ConstWGSize,
if (print_kernel_trace & STARTUP_DETAILS) {
DP("RTLDeviceInfoTy::Max_Teams: %d\n", RTLDeviceInfoTy::Max_Teams);
DP("Max_Teams: %d\n", Max_Teams);
- DP("RTLDeviceInfoTy::Warp_Size: %d\n", RTLDeviceInfoTy::Warp_Size);
+ DP("RTLDeviceInfoTy::Warp_Size: %d\n", WarpSize);
DP("RTLDeviceInfoTy::Max_WG_Size: %d\n", RTLDeviceInfoTy::Max_WG_Size);
DP("RTLDeviceInfoTy::Default_WG_Size: %d\n",
RTLDeviceInfoTy::Default_WG_Size);
@@ -1913,8 +1930,8 @@ launchVals getLaunchVals(EnvironmentVariables Env, int ConstWGSize,
threadsPerGroup = thread_limit;
DP("Setting threads per block to requested %d\n", thread_limit);
if (ExecutionMode == GENERIC) { // Add master warp for GENERIC
- threadsPerGroup += RTLDeviceInfoTy::Warp_Size;
- DP("Adding master wavefront: +%d threads\n", RTLDeviceInfoTy::Warp_Size);
+ threadsPerGroup += WarpSize;
+ DP("Adding master wavefront: +%d threads\n", WarpSize);
}
if (threadsPerGroup > RTLDeviceInfoTy::Max_WG_Size) { // limit to max
threadsPerGroup = RTLDeviceInfoTy::Max_WG_Size;
@@ -1950,7 +1967,7 @@ launchVals getLaunchVals(EnvironmentVariables Env, int ConstWGSize,
// So we only handle constant thread_limits.
if (threadsPerGroup >
RTLDeviceInfoTy::Default_WG_Size) // 256 < threadsPerGroup <= 1024
- // Should we round threadsPerGroup up to nearest RTLDeviceInfoTy::Warp_Size
+ // Should we round threadsPerGroup up to nearest WarpSize
// here?
num_groups = (Max_Teams * RTLDeviceInfoTy::Max_WG_Size) / threadsPerGroup;
@@ -2099,12 +2116,13 @@ int32_t __tgt_rtl_run_target_team_region_locked(
/*
* Set limit based on ThreadsPerGroup and GroupsPerDevice
*/
- launchVals LV = getLaunchVals(DeviceInfo.Env, KernelInfo->ConstWGSize,
- KernelInfo->ExecutionMode,
- num_teams, // From run_region arg
- thread_limit, // From run_region arg
- loop_tripcount, // From run_region arg
- DeviceInfo.NumTeams[KernelInfo->device_id]);
+ launchVals LV =
+ getLaunchVals(DeviceInfo.WarpSize[device_id], DeviceInfo.Env,
+ KernelInfo->ConstWGSize, KernelInfo->ExecutionMode,
+ num_teams, // From run_region arg
+ thread_limit, // From run_region arg
+ loop_tripcount, // From run_region arg
+ DeviceInfo.NumTeams[KernelInfo->device_id]);
const int GridSize = LV.GridSize;
const int WorkgroupSize = LV.WorkgroupSize;
More information about the cfe-commits
mailing list