Lower math.tan to __nv_tan[f] / __ocml_tan_f{32|64}

At present math.tan fails to lower for NVVM and ROCDL.

Differential Revision: https://reviews.llvm.org/D141505

diff --git a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
index 5050f6b..fe578f7 100644 (file)
--- a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
+++ b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
@@ -294,6 +294,7 @@ void mlir::populateGpuToNVVMConversionPatterns(LLVMTypeConverter &converter,
                                    "__nv_sqrt");
   populateOpPatterns<math::TanhOp>(converter, patterns, "__nv_tanhf",
                                    "__nv_tanh");
+  populateOpPatterns<math::TanOp>(converter, patterns, "__nv_tanf", "__nv_tan");
 }
 
 std::unique_ptr<OperationPass<gpu::GPUModuleOp>>

diff --git a/mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp b/mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp
index fead289..0182886 100644 (file)
--- a/mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp
+++ b/mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp
@@ -256,6 +256,8 @@ void mlir::populateGpuToROCDLConversionPatterns(
                                    "__ocml_sqrt_f64");
   populateOpPatterns<math::TanhOp>(converter, patterns, "__ocml_tanh_f32",
                                    "__ocml_tanh_f64");
+  populateOpPatterns<math::TanOp>(converter, patterns, "__ocml_tan_f32",
+                                  "__ocml_tan_f64");
 }
 
 std::unique_ptr<OperationPass<gpu::GPUModuleOp>>

diff --git a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
index 40035fa..4a7e6f5 100644 (file)
--- a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
+++ b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
@@ -344,6 +344,25 @@ gpu.module @test_module {
 // -----
 
 gpu.module @test_module {
+  // CHECK: llvm.func @__nv_tanf(f32) -> f32
+  // CHECK: llvm.func @__nv_tan(f64) -> f64
+  // CHECK-LABEL: func @gpu_tan
+  func.func @gpu_tan(%arg_f16 : f16, %arg_f32 : f32, %arg_f64 : f64) -> (f16, f32, f64) {
+    %result16 = math.tan %arg_f16 : f16
+    // CHECK: llvm.fpext %{{.*}} : f16 to f32
+    // CHECK-NEXT: llvm.call @__nv_tanf(%{{.*}}) : (f32) -> f32
+    // CHECK-NEXT: llvm.fptrunc %{{.*}} : f32 to f16
+    %result32 = math.tan %arg_f32 : f32
+    // CHECK: llvm.call @__nv_tanf(%{{.*}}) : (f32) -> f32
+    %result64 = math.tan %arg_f64 : f64
+    // CHECK: llvm.call @__nv_tan(%{{.*}}) : (f64) -> f64
+    func.return %result16, %result32, %result64 : f16, f32, f64
+  }
+}
+
+// -----
+
+gpu.module @test_module {
   // CHECK: llvm.func @__nv_tanhf(f32) -> f32
   // CHECK: llvm.func @__nv_tanh(f64) -> f64
   // CHECK-LABEL: func @gpu_tanh

diff --git a/mlir/test/Conversion/GPUToROCDL/gpu-to-rocdl.mlir b/mlir/test/Conversion/GPUToROCDL/gpu-to-rocdl.mlir
index 0403cdb..ccf8e47 100644 (file)
--- a/mlir/test/Conversion/GPUToROCDL/gpu-to-rocdl.mlir
+++ b/mlir/test/Conversion/GPUToROCDL/gpu-to-rocdl.mlir
@@ -363,6 +363,21 @@ gpu.module @test_module {
 // -----
 
 gpu.module @test_module {
+  // CHECK: llvm.func @__ocml_tan_f32(f32) -> f32
+  // CHECK: llvm.func @__ocml_tan_f64(f64) -> f64
+  // CHECK-LABEL: func @gpu_tan
+  func.func @gpu_tan(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
+    %result32 = math.tan %arg_f32 : f32
+    // CHECK: llvm.call @__ocml_tan_f32(%{{.*}}) : (f32) -> f32
+    %result64 = math.tan %arg_f64 : f64
+    // CHECK: llvm.call @__ocml_tan_f64(%{{.*}}) : (f64) -> f64
+    func.return %result32, %result64 : f32, f64
+  }
+}
+
+// -----
+
+gpu.module @test_module {
   // CHECK: llvm.func @__ocml_tanh_f32(f32) -> f32
   // CHECK: llvm.func @__ocml_tanh_f64(f64) -> f64
   // CHECK-LABEL: func @gpu_tanh