[AMDGPU] Optimize out s_barrier_signal/_wait #116993
Merged
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
Extend the optimization that converts s_barrier to wave_barrier (nop) when the number of work items is not larger than wave size. This handles the "split barrier" form of s_barrier where the barrier is represented by separate intrinsics (s_barrier_signal/s_barrier_wait). Note: the version where s_barrier is used in gfx12 (and later split) has the optimization already, but some front-ends may prefer to use split intrinsics and this is being addressed by the patch.
Member
|
@llvm/pr-subscribers-backend-amdgpu Author: Piotr Sobczak (piotrAMD) Changes[AMDGPU] Optimize out s_barrier_signal/_wait Extend the optimization that converts s_barrier to wave_barrier (nop) This handles the "split barrier" form of s_barrier where the barrier Full diff: https://github.com/llvm/llvm-project/pull/116993.diff 3 Files Affected:
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
index 3522ece24f1c45..f01f57de460c75 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -1843,8 +1843,9 @@ bool AMDGPUInstructionSelector::selectSBarrier(MachineInstr &MI) const {
}
}
- // On GFX12 lower s_barrier into s_barrier_signal_imm and s_barrier_wait
- if (STI.hasSplitBarriers()) {
+ Intrinsic::ID IntrinsicID = cast<GIntrinsic>(MI).getIntrinsicID();
+ if (STI.hasSplitBarriers() && IntrinsicID == Intrinsic::amdgcn_s_barrier) {
+ // On GFX12 lower s_barrier into s_barrier_signal_imm and s_barrier_wait
MachineBasicBlock *MBB = MI.getParent();
const DebugLoc &DL = MI.getDebugLoc();
BuildMI(*MBB, &MI, DL, TII.get(AMDGPU::S_BARRIER_SIGNAL_IMM))
@@ -2161,6 +2162,8 @@ bool AMDGPUInstructionSelector::selectG_INTRINSIC_W_SIDE_EFFECTS(
case Intrinsic::amdgcn_init_whole_wave:
return selectInitWholeWave(I);
case Intrinsic::amdgcn_s_barrier:
+ case Intrinsic::amdgcn_s_barrier_signal:
+ case Intrinsic::amdgcn_s_barrier_wait:
return selectSBarrier(I);
case Intrinsic::amdgcn_raw_buffer_load_lds:
case Intrinsic::amdgcn_raw_ptr_buffer_load_lds:
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 5b02f9bf80d3fc..4743ff0f65690e 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -9605,7 +9605,9 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
unsigned Opc = Done->isZero() ? AMDGPU::EXP : AMDGPU::EXP_DONE;
return SDValue(DAG.getMachineNode(Opc, DL, Op->getVTList(), Ops), 0);
}
- case Intrinsic::amdgcn_s_barrier: {
+ case Intrinsic::amdgcn_s_barrier:
+ case Intrinsic::amdgcn_s_barrier_signal:
+ case Intrinsic::amdgcn_s_barrier_wait: {
const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
if (getTargetMachine().getOptLevel() > CodeGenOptLevel::None) {
unsigned WGSize = ST.getFlatWorkGroupSizes(MF.getFunction()).second;
@@ -9615,8 +9617,8 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
0);
}
- // On GFX12 lower s_barrier into s_barrier_signal_imm and s_barrier_wait
- if (ST.hasSplitBarriers()) {
+ if (ST.hasSplitBarriers() && IntrinsicID == Intrinsic::amdgcn_s_barrier) {
+ // On GFX12 lower s_barrier into s_barrier_signal_imm and s_barrier_wait
SDValue K =
DAG.getTargetConstant(AMDGPU::Barrier::WORKGROUP, DL, MVT::i32);
SDValue BarSignal =
diff --git a/llvm/test/CodeGen/AMDGPU/barrier-elimination-gfx12.ll b/llvm/test/CodeGen/AMDGPU/barrier-elimination-gfx12.ll
new file mode 100644
index 00000000000000..42983b335f497c
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/barrier-elimination-gfx12.ll
@@ -0,0 +1,56 @@
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 < %s | FileCheck %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 < %s -global-isel | FileCheck %s
+
+; CHECK-LABEL: {{^}}signal_unknown_wgs:
+; CHECK: s_barrier_signal
+define amdgpu_kernel void @signal_unknown_wgs() {
+ tail call void @llvm.amdgcn.s.barrier.signal(i32 -1) #0
+ ret void
+}
+
+; CHECK-LABEL: {{^}}signal_flat_wgs_attr_32_128:
+; CHECK: s_barrier_signal
+define amdgpu_kernel void @signal_flat_wgs_attr_32_128() #1 {
+ tail call void @llvm.amdgcn.s.barrier.signal(i32 -1) #0
+ ret void
+}
+
+; CHECK-LABEL: {{^}}signal_flat_wgs_attr_32_64:
+; CHECK: :
+; CHECK-NEXT: ; wave barrier
+; CHECK-NEXT: s_endpgm
+define amdgpu_kernel void @signal_flat_wgs_attr_32_64() #2 {
+ tail call void @llvm.amdgcn.s.barrier.signal(i32 -1) #0
+ ret void
+}
+
+
+; CHECK-LABEL: {{^}}wait_unknown_wgs:
+; CHECK: s_barrier_wait
+define amdgpu_kernel void @wait_unknown_wgs() {
+ tail call void @llvm.amdgcn.s.barrier.wait(i16 -1) #0
+ ret void
+}
+
+; CHECK-LABEL: {{^}}wait_flat_wgs_attr_32_128:
+; CHECK: s_barrier_wait
+define amdgpu_kernel void @wait_flat_wgs_attr_32_128() #1 {
+ tail call void @llvm.amdgcn.s.barrier.wait(i16 -1) #0
+ ret void
+}
+
+; CHECK-LABEL: {{^}}wait_flat_wgs_attr_32_64:
+; CHECK: :
+; CHECK-NEXT: ; wave barrier
+; CHECK-NEXT: s_endpgm
+define amdgpu_kernel void @wait_flat_wgs_attr_32_64() #2 {
+ tail call void @llvm.amdgcn.s.barrier.wait(i16 -1) #0
+ ret void
+}
+
+declare void @llvm.amdgcn.s.barrier.signal(i32 immarg) #0
+declare void @llvm.amdgcn.s.barrier.wait(i16 immarg) #0
+
+attributes #0 = { convergent nounwind }
+attributes #1 = { nounwind "amdgpu-flat-work-group-size"="32,128" }
+attributes #2 = { nounwind "amdgpu-flat-work-group-size"="16,32" }
|
arsenm
reviewed
Nov 20, 2024
perlfu
reviewed
Nov 21, 2024
rovka
reviewed
Nov 21, 2024
shiltian
reviewed
Nov 21, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
7 participants
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.
Extend the optimization that converts s_barrier to wave_barrier (nop)
when the number of work items is not larger than wave size.
This handles the "split barrier" form of s_barrier where the barrier
is represented by separate intrinsics (s_barrier_signal/s_barrier_wait).
Note: the version where s_barrier is used in gfx12 (and later split)
has the optimization already, but some front-ends may prefer to use
split intrinsics and this is being addressed by the patch.