diff options
| author | Frederik Gossen <frgossen@google.com> | 2020-12-08 13:57:54 +0100 |
|---|---|---|
| committer | Frederik Gossen <frgossen@google.com> | 2020-12-08 14:33:58 +0100 |
| commit | bb7d43e7d5f683fc3e7109072610dc8d176a3bf8 (patch) | |
| tree | 982d37699162ac0a01e843dae06a51fe04c575ec /mlir | |
| parent | 083e035c47f6c73084ecf5ab7f41cddca19ce332 (diff) | |
| download | llvm-project-bb7d43e7d5f683fc3e7109072610dc8d176a3bf8.tar.gz | |
Add rsqrt lowering from standard to NVVM
Differential Revision: https://reviews.llvm.org/D92838
Diffstat (limited to 'mlir')
| -rw-r--r-- | mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp | 2 | ||||
| -rw-r--r-- | mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir | 20 |
2 files changed, 22 insertions, 0 deletions
diff --git a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp index 61f97d44e125..4b7bb6e193a4 100644 --- a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp +++ b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp @@ -183,6 +183,8 @@ void mlir::populateGpuToNVVMConversionPatterns( "__nv_log10"); patterns.insert<OpToFuncCallLowering<Log2Op>>(converter, "__nv_log2f", "__nv_log2"); + patterns.insert<OpToFuncCallLowering<RsqrtOp>>(converter, "__nv_rsqrtf", + "__nv_rsqrt"); patterns.insert<OpToFuncCallLowering<SinOp>>(converter, "__nv_sinf", "__nv_sin"); patterns.insert<OpToFuncCallLowering<TanhOp>>(converter, "__nv_tanhf", diff --git a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir index 8e4e4515c638..347cc48daa20 100644 --- a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir +++ b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir @@ -295,6 +295,26 @@ gpu.module @test_module { // ----- +gpu.module @test_module { + // CHECK: llvm.func @__nv_rsqrtf(!llvm.float) -> !llvm.float + // CHECK: llvm.func @__nv_rsqrt(!llvm.double) -> !llvm.double + // CHECK-LABEL: func @gpu_rsqrt + func @gpu_rsqrt(%arg_f16 : f16, %arg_f32 : f32, %arg_f64 : f64) + -> (f16, f32, f64) { + %result16 = std.rsqrt %arg_f16 : f16 + // CHECK: llvm.fpext %{{.*}} : !llvm.half to !llvm.float + // CHECK-NEXT: llvm.call @__nv_rsqrtf(%{{.*}}) : (!llvm.float) -> !llvm.float + // CHECK-NEXT: llvm.fptrunc %{{.*}} : !llvm.float to !llvm.half + %result32 = std.rsqrt %arg_f32 : f32 + // CHECK: llvm.call @__nv_rsqrtf(%{{.*}}) : (!llvm.float) -> !llvm.float + %result64 = std.rsqrt %arg_f64 : f64 + // CHECK: llvm.call @__nv_rsqrt(%{{.*}}) : (!llvm.double) -> !llvm.double + std.return %result16, %result32, %result64 : f16, f32, f64 + } +} + +// ----- + // Test that we handled properly operation with SymbolTable other than module op gpu.module @test_module { "test.symbol_scope"() ({ |