[mlir] lower min/maxnum to libdevice calls #127323
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
Introduce lowering from arith.minnum/maxxnum operations to the corresponding Nvidia libdevice calls. This requires to reorder pattern population methods so that the libdevice-targeting patterns are prioritized over default patterns targeting LLVM IR intrinsics from the Arith dialect. The tests are placed into a separate file as the existing gpu-to-nvvm.mlir files has a mode that forces Arith dialect operations to be preserved as is without using a separate FileCheck tag to differentiate.
|
@llvm/pr-subscribers-mlir-gpu Author: Oleksandr "Alex" Zinenko (ftynse) ChangesIntroduce lowering from arith.minnum/maxxnum operations to the corresponding Nvidia libdevice calls. This requires to reorder pattern population methods so that the libdevice-targeting patterns are prioritized over default patterns targeting LLVM IR intrinsics from the Arith dialect. The tests are placed into a separate file as the existing gpu-to-nvvm.mlir files has a mode that forces Arith dialect operations to be preserved as is without using a separate FileCheck tag to differentiate. Full diff: https://github.com/llvm/llvm-project/pull/127323.diff 2 Files Affected:
diff --git a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
index 35330f870e6ae..c1a4d31861d3b 100644
--- a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
+++ b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
@@ -378,6 +378,8 @@ struct LowerGpuOpsToNVVMOpsPass final
RewritePatternSet llvmPatterns(m.getContext());
LLVMConversionTarget target(getContext());
+ populateGpuToNVVMConversionPatterns(converter, llvmPatterns);
+
llvm::SmallDenseSet<StringRef> allowedDialectsSet(allowedDialects.begin(),
allowedDialects.end());
for (Dialect *dialect : getContext().getLoadedDialects()) {
@@ -407,7 +409,6 @@ struct LowerGpuOpsToNVVMOpsPass final
llvmPatterns);
}
- populateGpuToNVVMConversionPatterns(converter, llvmPatterns);
populateGpuWMMAToNVVMConversionPatterns(converter, llvmPatterns);
if (this->hasRedux)
populateGpuSubgroupReduceOpLoweringPattern(converter, llvmPatterns);
@@ -552,6 +553,11 @@ void mlir::populateGpuToNVVMConversionPatterns(
populateOpPatterns<arith::RemFOp>(converter, patterns, "__nv_fmodf",
"__nv_fmod");
+ populateOpPatterns<arith::MaxNumFOp>(converter, patterns, "__nv_fmaxf",
+ "__nv_fmax");
+ populateOpPatterns<arith::MinNumFOp>(converter, patterns, "__nv_fminf",
+ "__nv_fmin");
+
populateIntOpPatterns<math::AbsIOp>(converter, patterns, "__nv_abs");
populateOpPatterns<math::AbsFOp>(converter, patterns, "__nv_fabsf",
"__nv_fabs");
diff --git a/mlir/test/Conversion/GPUToNVVM/gpu-arith-ops-to-nvvm.mlir b/mlir/test/Conversion/GPUToNVVM/gpu-arith-ops-to-nvvm.mlir
new file mode 100644
index 0000000000000..2b1c5a7fef8fe
--- /dev/null
+++ b/mlir/test/Conversion/GPUToNVVM/gpu-arith-ops-to-nvvm.mlir
@@ -0,0 +1,21 @@
+// RUN: mlir-opt %s -convert-gpu-to-nvvm -split-input-file | FileCheck %s
+
+gpu.module @test_module_54 {
+ // CHECK: llvm.func @__nv_fmaxf(f32, f32) -> f32
+ // CHECK: llvm.func @__nv_fminf(f32, f32) -> f32
+ // CHECK: llvm.func @__nv_fmax(f64, f64) -> f64
+ // CHECK: llvm.func @__nv_fmin(f64, f64) -> f64
+ // CHECK-LABEL: @gpu_fminmax
+ func.func @gpu_fminmax(%arg1_f32: f32, %arg2_f32: f32, %arg1_f64: f64, %arg2_f64: f64)
+ -> (f32, f32, f64, f64) {
+ // CHECK: llvm.call @__nv_fmaxf
+ %max_f32 = arith.maxnumf %arg1_f32, %arg2_f32 : f32
+ // CHECK: llvm.call @__nv_fminf
+ %min_f32 = arith.minnumf %arg1_f32, %arg2_f32 : f32
+ // CHECK: llvm.call @__nv_fmax(
+ %max_f64 = arith.maxnumf %arg1_f64, %arg2_f64 : f64
+ // CHECK: llvm.call @__nv_fmin(
+ %min_f64 = arith.minnumf %arg1_f64, %arg2_f64 : f64
+ return %max_f32, %min_f32, %max_f64, %min_f64 : f32, f32, f64, f64
+ }
+}
|
|
@llvm/pr-subscribers-mlir Author: Oleksandr "Alex" Zinenko (ftynse) ChangesIntroduce lowering from arith.minnum/maxxnum operations to the corresponding Nvidia libdevice calls. This requires to reorder pattern population methods so that the libdevice-targeting patterns are prioritized over default patterns targeting LLVM IR intrinsics from the Arith dialect. The tests are placed into a separate file as the existing gpu-to-nvvm.mlir files has a mode that forces Arith dialect operations to be preserved as is without using a separate FileCheck tag to differentiate. Full diff: https://github.com/llvm/llvm-project/pull/127323.diff 2 Files Affected:
diff --git a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
index 35330f870e6ae..c1a4d31861d3b 100644
--- a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
+++ b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
@@ -378,6 +378,8 @@ struct LowerGpuOpsToNVVMOpsPass final
RewritePatternSet llvmPatterns(m.getContext());
LLVMConversionTarget target(getContext());
+ populateGpuToNVVMConversionPatterns(converter, llvmPatterns);
+
llvm::SmallDenseSet<StringRef> allowedDialectsSet(allowedDialects.begin(),
allowedDialects.end());
for (Dialect *dialect : getContext().getLoadedDialects()) {
@@ -407,7 +409,6 @@ struct LowerGpuOpsToNVVMOpsPass final
llvmPatterns);
}
- populateGpuToNVVMConversionPatterns(converter, llvmPatterns);
populateGpuWMMAToNVVMConversionPatterns(converter, llvmPatterns);
if (this->hasRedux)
populateGpuSubgroupReduceOpLoweringPattern(converter, llvmPatterns);
@@ -552,6 +553,11 @@ void mlir::populateGpuToNVVMConversionPatterns(
populateOpPatterns<arith::RemFOp>(converter, patterns, "__nv_fmodf",
"__nv_fmod");
+ populateOpPatterns<arith::MaxNumFOp>(converter, patterns, "__nv_fmaxf",
+ "__nv_fmax");
+ populateOpPatterns<arith::MinNumFOp>(converter, patterns, "__nv_fminf",
+ "__nv_fmin");
+
populateIntOpPatterns<math::AbsIOp>(converter, patterns, "__nv_abs");
populateOpPatterns<math::AbsFOp>(converter, patterns, "__nv_fabsf",
"__nv_fabs");
diff --git a/mlir/test/Conversion/GPUToNVVM/gpu-arith-ops-to-nvvm.mlir b/mlir/test/Conversion/GPUToNVVM/gpu-arith-ops-to-nvvm.mlir
new file mode 100644
index 0000000000000..2b1c5a7fef8fe
--- /dev/null
+++ b/mlir/test/Conversion/GPUToNVVM/gpu-arith-ops-to-nvvm.mlir
@@ -0,0 +1,21 @@
+// RUN: mlir-opt %s -convert-gpu-to-nvvm -split-input-file | FileCheck %s
+
+gpu.module @test_module_54 {
+ // CHECK: llvm.func @__nv_fmaxf(f32, f32) -> f32
+ // CHECK: llvm.func @__nv_fminf(f32, f32) -> f32
+ // CHECK: llvm.func @__nv_fmax(f64, f64) -> f64
+ // CHECK: llvm.func @__nv_fmin(f64, f64) -> f64
+ // CHECK-LABEL: @gpu_fminmax
+ func.func @gpu_fminmax(%arg1_f32: f32, %arg2_f32: f32, %arg1_f64: f64, %arg2_f64: f64)
+ -> (f32, f32, f64, f64) {
+ // CHECK: llvm.call @__nv_fmaxf
+ %max_f32 = arith.maxnumf %arg1_f32, %arg2_f32 : f32
+ // CHECK: llvm.call @__nv_fminf
+ %min_f32 = arith.minnumf %arg1_f32, %arg2_f32 : f32
+ // CHECK: llvm.call @__nv_fmax(
+ %max_f64 = arith.maxnumf %arg1_f64, %arg2_f64 : f64
+ // CHECK: llvm.call @__nv_fmin(
+ %min_f64 = arith.minnumf %arg1_f64, %arg2_f64 : f64
+ return %max_f32, %min_f32, %max_f64, %min_f64 : f32, f32, f64, f64
+ }
+}
|
wsmoses
approved these changes
Feb 15, 2025
sivan-shani
pushed a commit
to sivan-shani/llvm-project
that referenced
this pull request
Feb 24, 2025
Introduce lowering from arith.minnum/maxxnum operations to the corresponding Nvidia libdevice calls. This requires to reorder pattern population methods so that the libdevice-targeting patterns are prioritized over default patterns targeting LLVM IR intrinsics from the Arith dialect. The tests are placed into a separate file as the existing gpu-to-nvvm.mlir files has a mode that forces Arith dialect operations to be preserved as is without using a separate FileCheck tag to differentiate. Co-authored-by: William Moses <[email protected]>
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.
Introduce lowering from arith.minnum/maxxnum operations to the corresponding Nvidia libdevice calls. This requires to reorder pattern population methods so that the libdevice-targeting patterns are prioritized over default patterns targeting LLVM IR intrinsics from the Arith dialect. The tests are placed into a separate file as the existing gpu-to-nvvm.mlir files has a mode that forces Arith dialect operations to be preserved as is without using a separate FileCheck tag to differentiate.