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

[gflegar]

Summary

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

[herhut]

Thanks for adding this.

[gflegar]

Also lower math.tan for ROCDL

mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp

populateOpPatterns<math::PowFOp>(converter, patterns, "__nv_powf",
                                    "__nv_pow");
   populateOpPatterns<math::RsqrtOp>(converter, patterns, "__nv_rsqrtf",
                                     "__nv_rsqrt");
   populateOpPatterns<math::SinOp>(converter, patterns, "__nv_sinf", "__nv_sin");
   populateOpPatterns<math::SqrtOp>(converter, patterns, "__nv_sqrtf",
                                    "__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>>
 mlir::createLowerGpuOpsToNVVMOpsPass(unsigned indexBitwidth) {
   return std::make_unique<LowerGpuOpsToNVVMOpsPass>(indexBitwidth);
 }

mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp

void mlir::populateGpuToROCDLConversionPatterns(
   populateOpPatterns<math::RsqrtOp>(converter, patterns, "__ocml_rsqrt_f32",
                                     "__ocml_rsqrt_f64");
   populateOpPatterns<math::SinOp>(converter, patterns, "__ocml_sin_f32",
                                   "__ocml_sin_f64");
   populateOpPatterns<math::SqrtOp>(converter, patterns, "__ocml_sqrt_f32",
                                    "__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>>
 mlir::createLowerGpuOpsToROCDLOpsPass(const std::string &chipset,
                                       unsigned indexBitwidth,
                                       bool useBarePtrCallConv,
                                       gpu::amd::Runtime runtime) {
   return std::make_unique<LowerGpuOpsToROCDLOpsPass>(
       chipset, indexBitwidth, useBarePtrCallConv, runtime);
 }

mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir

func.func @gpu_sin(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
     // CHECK: llvm.call @__nv_sin(%{{.*}}) : (f64) -> f64
     func.return %result32, %result64 : f32, f64
   }
 }
 
 // -----
 
 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
   func.func @gpu_tanh(%arg_f16 : f16, %arg_f32 : f32, %arg_f64 : f64) -> (f16, f32, f64) {
     %result16 = math.tanh %arg_f16 : f16
     // CHECK: llvm.fpext %{{.*}} : f16 to f32
     // CHECK-NEXT: llvm.call @__nv_tanhf(%{{.*}}) : (f32) -> f32
     // CHECK-NEXT: llvm.fptrunc %{{.*}} : f32 to f16

mlir/test/Conversion/GPUToROCDL/gpu-to-rocdl.mlir

func.func @gpu_sqrt(%arg_f16 : f16, %arg_f32 : f32, %arg_f64 : f64)
     // CHECK: llvm.call @__ocml_sqrt_f64(%{{.*}}) : (f64) -> f64
     func.return %result16, %result32, %result64 : f16, f32, f64
   }
 }
 
 // -----
 
 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
   func.func @gpu_tanh(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
     %result32 = math.tanh %arg_f32 : f32
     // CHECK: llvm.call @__ocml_tanh_f32(%{{.*}}) : (f32) -> f32
     %result64 = math.tanh %arg_f64 : f64
     // CHECK: llvm.call @__ocml_tanh_f64(%{{.*}}) : (f64) -> f64