[mlir][nvvm] Add conversion for math.erfc

[clementval]

Add missing pattern to convert math.erfc operation to __nv_erfcf or __nv_erfc function call.

[llvmbot]

@llvm/pr-subscribers-mlir

Author: Valentin Clement (バレンタイン クレメン) (clementval)

Changes

Add missing pattern to convert math.erfc operation to __nv_erfcf or __nv_erfc function call.

---

Full diff: https://github.com/llvm/llvm-project/pull/129329.diff

2 Files Affected:

• (modified) mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp (+2)
• (modified) mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir (+13)

diff --git a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
index 61b73f546b5da..159a6afd4a917 100644
--- a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
+++ b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
@@ -540,6 +540,8 @@ void mlir::populateLibDeviceConversionPatterns(
                                    "__nv_cosh");
   populateOpPatterns<math::ErfOp>(converter, patterns, benefit, "__nv_erff",
                                   "__nv_erf");
+  populateOpPatterns<math::ErfcOp>(converter, patterns, benefit, "__nv_erfcf",
+                                   "__nv_erfc");
   populateOpPatterns<math::ExpOp>(converter, patterns, benefit, "__nv_expf",
                                   "__nv_exp", "__nv_fast_expf");
   populateOpPatterns<math::Exp2Op>(converter, patterns, benefit, "__nv_exp2f",
diff --git a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
index 7b5b11ec02724..14594cd6badb1 100644
--- a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
+++ b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
@@ -1095,3 +1095,16 @@ gpu.module @test_module_54 {
     return %0, %1, %2, %3, %4, %5 : i1, i1, i1, i1, i1, i1
   }
 }
+
+gpu.module @test_module_55 {
+  // CHECK: llvm.func @__nv_erfcf(f32) -> f32
+  // CHECK: llvm.func @__nv_erfc(f64) -> f64
+  // CHECK-LABEL: func @gpu_erf
+  func.func @gpu_erfc(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
+    %result32 = math.erfc %arg_f32 : f32
+    // CHECK: llvm.call @__nv_erfcf(%{{.*}}) : (f32) -> f32
+    %result64 = math.erfc %arg_f64 : f64
+    // CHECK: llvm.call @__nv_erfc(%{{.*}}) : (f64) -> f64
+    func.return %result32, %result64 : f32, f64
+  }
+}

[llvmbot]

@llvm/pr-subscribers-mlir-gpu

Author: Valentin Clement (バレンタイン クレメン) (clementval)

Changes

Add missing pattern to convert math.erfc operation to __nv_erfcf or __nv_erfc function call.

---

Full diff: https://github.com/llvm/llvm-project/pull/129329.diff

2 Files Affected:

• (modified) mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp (+2)
• (modified) mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir (+13)

diff --git a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
index 61b73f546b5da..159a6afd4a917 100644
--- a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
+++ b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
@@ -540,6 +540,8 @@ void mlir::populateLibDeviceConversionPatterns(
                                    "__nv_cosh");
   populateOpPatterns<math::ErfOp>(converter, patterns, benefit, "__nv_erff",
                                   "__nv_erf");
+  populateOpPatterns<math::ErfcOp>(converter, patterns, benefit, "__nv_erfcf",
+                                   "__nv_erfc");
   populateOpPatterns<math::ExpOp>(converter, patterns, benefit, "__nv_expf",
                                   "__nv_exp", "__nv_fast_expf");
   populateOpPatterns<math::Exp2Op>(converter, patterns, benefit, "__nv_exp2f",
diff --git a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
index 7b5b11ec02724..14594cd6badb1 100644
--- a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
+++ b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
@@ -1095,3 +1095,16 @@ gpu.module @test_module_54 {
     return %0, %1, %2, %3, %4, %5 : i1, i1, i1, i1, i1, i1
   }
 }
+
+gpu.module @test_module_55 {
+  // CHECK: llvm.func @__nv_erfcf(f32) -> f32
+  // CHECK: llvm.func @__nv_erfc(f64) -> f64
+  // CHECK-LABEL: func @gpu_erf
+  func.func @gpu_erfc(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
+    %result32 = math.erfc %arg_f32 : f32
+    // CHECK: llvm.call @__nv_erfcf(%{{.*}}) : (f32) -> f32
+    %result64 = math.erfc %arg_f64 : f64
+    // CHECK: llvm.call @__nv_erfc(%{{.*}}) : (f64) -> f64
+    func.return %result32, %result64 : f32, f64
+  }
+}