[OpenMP] Replace use of target address space with <gpuintrin.h> local #126119
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-llvm-transforms @llvm/pr-subscribers-offload Author: Joseph Huber (jhuber6) ChangesSummary: Full diff: https://github.com/llvm/llvm-project/pull/126119.diff 7 Files Affected:
diff --git a/offload/DeviceRTL/include/DeviceTypes.h b/offload/DeviceRTL/include/DeviceTypes.h
index 308109b0749f05b..395d72eafbf4054 100644
--- a/offload/DeviceRTL/include/DeviceTypes.h
+++ b/offload/DeviceRTL/include/DeviceTypes.h
@@ -12,6 +12,7 @@
#ifndef OMPTARGET_TYPES_H
#define OMPTARGET_TYPES_H
+#include <gpuintrin.h>
#include <stddef.h>
#include <stdint.h>
@@ -155,19 +156,6 @@ typedef enum omp_allocator_handle_t {
#define __PRAGMA(STR) _Pragma(#STR)
#define OMP_PRAGMA(STR) __PRAGMA(omp STR)
-#define SHARED(NAME) \
- [[clang::address_space(3)]] NAME [[clang::loader_uninitialized]];
-
-// TODO: clang should use address space 5 for omp_thread_mem_alloc, but right
-// now that's not the case.
-#define THREAD_LOCAL(NAME) \
- [[clang::address_space(5)]] NAME [[clang::loader_uninitialized]]
-
-// TODO: clang should use address space 4 for omp_const_mem_alloc, maybe it
-// does?
-#define CONSTANT(NAME) \
- [[clang::address_space(4)]] NAME [[clang::loader_uninitialized]]
-
///}
#endif
diff --git a/offload/DeviceRTL/src/Configuration.cpp b/offload/DeviceRTL/src/Configuration.cpp
index a2dfa4a02a09478..796e9ee254f3ac8 100644
--- a/offload/DeviceRTL/src/Configuration.cpp
+++ b/offload/DeviceRTL/src/Configuration.cpp
@@ -28,8 +28,8 @@ using namespace ompx;
// This variable should be visible to the plugin so we override the default
// hidden visibility.
[[gnu::used, gnu::retain, gnu::weak,
- gnu::visibility("protected")]] DeviceEnvironmentTy
- CONSTANT(__omp_rtl_device_environment);
+ gnu::visibility("protected")]] DeviceEnvironmentTy __gpu_constant
+ __omp_rtl_device_environment;
uint32_t config::getAssumeTeamsOversubscription() {
return __omp_rtl_assume_teams_oversubscription;
diff --git a/offload/DeviceRTL/src/Mapping.cpp b/offload/DeviceRTL/src/Mapping.cpp
index a0c0f6721a84cce..bce62ec9546bc6a 100644
--- a/offload/DeviceRTL/src/Mapping.cpp
+++ b/offload/DeviceRTL/src/Mapping.cpp
@@ -308,7 +308,7 @@ uint32_t mapping::getNumberOfProcessorElements() {
// TODO: This is a workaround for initialization coming from kernels outside of
// the TU. We will need to solve this more correctly in the future.
-[[gnu::weak]] int SHARED(IsSPMDMode);
+[[clang::loader_uninitialized]] [[gnu::weak]] int __gpu_local IsSPMDMode;
void mapping::init(bool IsSPMD) {
if (mapping::isInitialThreadInLevel0(IsSPMD))
diff --git a/offload/DeviceRTL/src/Reduction.cpp b/offload/DeviceRTL/src/Reduction.cpp
index 25f34005532f7c6..3b5f911451cb7fa 100644
--- a/offload/DeviceRTL/src/Reduction.cpp
+++ b/offload/DeviceRTL/src/Reduction.cpp
@@ -71,16 +71,16 @@ static int32_t nvptx_parallel_reduce_nowait(void *reduce_data,
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.
- //
+ //
+ // 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 __has_builtin(__nvvm_reflect)
if (__nvvm_reflect("__CUDA_ARCH") >= 700) {
@@ -196,8 +196,8 @@ int32_t __kmpc_nvptx_teams_reduce_nowait_v2(
uint32_t NumThreads = omp_get_num_threads();
uint32_t TeamId = omp_get_team_num();
uint32_t NumTeams = omp_get_num_teams();
- static unsigned SHARED(Bound);
- static unsigned SHARED(ChunkTeamCount);
+ [[clang::loader_uninitialized]] static unsigned __gpu_local Bound;
+ [[clang::loader_uninitialized]] static unsigned __gpu_local ChunkTeamCount;
// Block progress for teams greater than the current upper
// limit. We always only allow a number of teams less or equal
diff --git a/offload/DeviceRTL/src/State.cpp b/offload/DeviceRTL/src/State.cpp
index 89edb4802198c9c..0981b33dfdd4dce 100644
--- a/offload/DeviceRTL/src/State.cpp
+++ b/offload/DeviceRTL/src/State.cpp
@@ -32,11 +32,13 @@ using namespace ompx;
[[clang::address_space(3)]] DynamicSharedBuffer[];
/// The kernel environment passed to the init method by the compiler.
-static KernelEnvironmentTy *SHARED(KernelEnvironmentPtr);
+[[clang::loader_uninitialized]] static KernelEnvironmentTy *__gpu_local
+ KernelEnvironmentPtr;
/// The kernel launch environment passed as argument to the kernel by the
/// runtime.
-static KernelLaunchEnvironmentTy *SHARED(KernelLaunchEnvironmentPtr);
+[[clang::loader_uninitialized]] static KernelLaunchEnvironmentTy *__gpu_local
+ KernelLaunchEnvironmentPtr;
///}
@@ -108,7 +110,8 @@ static_assert(state::SharedScratchpadSize / mapping::MaxThreadsPerTeam <= 256,
"Shared scratchpad of this size not supported yet.");
/// The allocation of a single shared memory scratchpad.
-static SharedMemorySmartStackTy SHARED(SharedMemorySmartStack);
+[[clang::loader_uninitialized]] static SharedMemorySmartStackTy __gpu_local
+ SharedMemorySmartStack;
void SharedMemorySmartStackTy::init(bool IsSPMD) {
Usage[mapping::getThreadIdInBlock()] = 0;
@@ -220,8 +223,10 @@ void state::TeamStateTy::assertEqual(TeamStateTy &Other) const {
ASSERT(HasThreadState == Other.HasThreadState, nullptr);
}
-state::TeamStateTy SHARED(ompx::state::TeamState);
-state::ThreadStateTy **SHARED(ompx::state::ThreadStates);
+[[clang::loader_uninitialized]] state::TeamStateTy __gpu_local
+ ompx::state::TeamState;
+[[clang::loader_uninitialized]] state::ThreadStateTy **__gpu_local
+ ompx::state::ThreadStates;
namespace {
diff --git a/offload/DeviceRTL/src/Synchronization.cpp b/offload/DeviceRTL/src/Synchronization.cpp
index a5090b96560c8bc..55d6d75dfa7e61c 100644
--- a/offload/DeviceRTL/src/Synchronization.cpp
+++ b/offload/DeviceRTL/src/Synchronization.cpp
@@ -69,7 +69,7 @@ uint32_t atomicInc(uint32_t *A, uint32_t V, atomic::OrderingTy Ordering,
}
}
-uint32_t SHARED(namedBarrierTracker);
+[[clang::loader_uninitialized]] uint32_t __gpu_local namedBarrierTracker;
void namedBarrierInit() {
// Don't have global ctors, and shared memory is not zero init
diff --git a/offload/DeviceRTL/src/Workshare.cpp b/offload/DeviceRTL/src/Workshare.cpp
index b1f037a11bddf1f..9e34b94fe4b208c 100644
--- a/offload/DeviceRTL/src/Workshare.cpp
+++ b/offload/DeviceRTL/src/Workshare.cpp
@@ -45,7 +45,7 @@ struct DynamicScheduleTracker {
#define LAST_CHUNK 2
// TODO: This variable is a hack inherited from the old runtime.
-static uint64_t SHARED(Cnt);
+[[clang::loader_uninitialized]] static uint64_t __gpu_local Cnt;
template <typename T, typename ST> struct omptarget_nvptx_LoopSupport {
////////////////////////////////////////////////////////////////////////////////
@@ -457,7 +457,8 @@ template <typename T, typename ST> struct omptarget_nvptx_LoopSupport {
//
// __kmpc_dispatch_deinit
//
-static DynamicScheduleTracker **SHARED(ThreadDST);
+[[clang::loader_uninitialized]] static DynamicScheduleTracker **__gpu_local
+ ThreadDST;
// Create a new DST, link the current one, and define the new as current.
static DynamicScheduleTracker *pushDST() {
|
shiltian
reviewed
Feb 7, 2025
No it just pulls out this part to make it easier to land in pieces. |
|
I did not see issues on this PR in local testing. Will take a closer look later today. |
jhuber6
force-pushed
the
Local
branch
2 times, most recently
from
8248a82 to
d83539b
Compare
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
Summary: This definition is more portable since it defines the correct value for the target. I got rid of the helper mostly because I think it's easy enough to use now that it's a type and being explicit about what's undef or poison is good.
ronlieb
approved these changes
Feb 9, 2025
searlmc1
pushed a commit
to ROCm/llvm-project
that referenced
this pull request
Feb 9, 2025
Icohedron
pushed a commit
to Icohedron/llvm-project
that referenced
this pull request
Feb 11, 2025
…llvm#126119) Summary: This definition is more portable since it defines the correct value for the target. I got rid of the helper mostly because I think it's easy enough to use now that it's a type and being explicit about what's `undef` or `poison` is good.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
This definition is more portable since it defines the correct value for
the target. I got rid of the helper mostly because I think it's easy
enough to use now that it's a type and being explicit about what's
undeforpoisonis good.