This is an archive of the discontinued LLVM Phabricator instance.

Differential D93011

Add rsqrt lowering from standard to ROCDL.
ClosedPublic

Authored by akuegel on Dec 10 2020, 1:48 AM.

Details

Reviewers
frgossen
herhut
Commits
rGada4c7a35171: Add rsqrt lowering from standard to ROCDL.
Summary

Add a lowering for rsqrt from standard dialect to ROCDL.

Diff Detail

Event Timeline

akuegel created this revision.Dec 10 2020, 1:48 AM
Herald added subscribers: teijeong, rdzhabarov, tatianashp and 15 others. · View Herald TranscriptDec 10 2020, 1:48 AM
akuegel requested review of this revision.Dec 10 2020, 1:48 AM
Herald added a reviewer: herhut. · View Herald TranscriptDec 10 2020, 1:48 AM
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: stephenneuendorffer, nicolasvasilache. · View Herald Transcript
Harbormaster completed remote builds in B81791: Diff 310801.Dec 10 2020, 2:09 AM
akuegel added a comment.Dec 11 2020, 3:52 AM

Friendly ping?

frgossen accepted this revision.Dec 11 2020, 4:04 AM

Thanks!

This revision is now accepted and ready to land.Dec 11 2020, 4:04 AM
This revision was landed with ongoing or failed builds.Dec 11 2020, 4:20 AM
Closed by commit rGada4c7a35171: Add rsqrt lowering from standard to ROCDL. (authored by akuegel). · Explain Why
This revision was automatically updated to reflect the committed changes.
akuegel added a commit: rGada4c7a35171: Add rsqrt lowering from standard to ROCDL..

Revision Contents

PathSize
mlir/
lib/
Conversion/
GPUToROCDL/
2 lines
test/
Conversion/
GPUToROCDL/
20 lines

Diff 311163

mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp

Show First 20 Lines • Show All 107 Lines • ▼ Show 20 Linesvoid mlir::populateGpuToROCDLConversionPatterns(
patterns.insert<OpToFuncCallLowering<FloorFOp>>(converter, "__ocml_floor_f32", patterns.insert<OpToFuncCallLowering<FloorFOp>>(converter, "__ocml_floor_f32",
"__ocml_floor_f64"); "__ocml_floor_f64");
patterns.insert<OpToFuncCallLowering<LogOp>>(converter, "__ocml_log_f32", patterns.insert<OpToFuncCallLowering<LogOp>>(converter, "__ocml_log_f32",
"__ocml_log_f64"); "__ocml_log_f64");
patterns.insert<OpToFuncCallLowering<Log10Op>>(converter, "__ocml_log10_f32", patterns.insert<OpToFuncCallLowering<Log10Op>>(converter, "__ocml_log10_f32",
"__ocml_log10_f64"); "__ocml_log10_f64");
patterns.insert<OpToFuncCallLowering<Log2Op>>(converter, "__ocml_log2_f32", patterns.insert<OpToFuncCallLowering<Log2Op>>(converter, "__ocml_log2_f32",
"__ocml_log2_f64"); "__ocml_log2_f64");
patterns.insert<OpToFuncCallLowering<RsqrtOp>>(converter, "__ocml_rsqrt_f32",
"__ocml_rsqrt_f64");
patterns.insert<OpToFuncCallLowering<SinOp>>(converter, "__ocml_sin_f32", patterns.insert<OpToFuncCallLowering<SinOp>>(converter, "__ocml_sin_f32",
"__ocml_sin_f64"); "__ocml_sin_f64");
patterns.insert<OpToFuncCallLowering<SqrtOp>>(converter, "__ocml_sqrt_f32", patterns.insert<OpToFuncCallLowering<SqrtOp>>(converter, "__ocml_sqrt_f32",
"__ocml_sqrt_f64"); "__ocml_sqrt_f64");
patterns.insert<OpToFuncCallLowering<TanhOp>>(converter, "__ocml_tanh_f32", patterns.insert<OpToFuncCallLowering<TanhOp>>(converter, "__ocml_tanh_f32",
"__ocml_tanh_f64"); "__ocml_tanh_f64");
} }
std::unique_ptr<OperationPass<gpu::GPUModuleOp>> std::unique_ptr<OperationPass<gpu::GPUModuleOp>>
mlir::createLowerGpuOpsToROCDLOpsPass(unsigned indexBitwidth) { mlir::createLowerGpuOpsToROCDLOpsPass(unsigned indexBitwidth) {
return std::make_unique<LowerGpuOpsToROCDLOpsPass>(indexBitwidth); return std::make_unique<LowerGpuOpsToROCDLOpsPass>(indexBitwidth);
} }

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

Show First 20 Lines • Show All 223 Lines • ▼ Show 20 Lines func @gpu_log2(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
// CHECK: llvm.call @__ocml_log2_f64(%{{.*}}) : (!llvm.double) -> !llvm.double // CHECK: llvm.call @__ocml_log2_f64(%{{.*}}) : (!llvm.double) -> !llvm.double
std.return %result32, %result64 : f32, f64 std.return %result32, %result64 : f32, f64
} }
} }
// ----- // -----
gpu.module @test_module { gpu.module @test_module {
// CHECK: llvm.func @__ocml_rsqrt_f32(!llvm.float) -> !llvm.float
// CHECK: llvm.func @__ocml_rsqrt_f64(!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 @__ocml_rsqrt_f32(%{{.*}}) : (!llvm.float) -> !llvm.float
// CHECK-NEXT: llvm.fptrunc %{{.*}} : !llvm.float to !llvm.half
%result32 = std.rsqrt %arg_f32 : f32
// CHECK: llvm.call @__ocml_rsqrt_f32(%{{.*}}) : (!llvm.float) -> !llvm.float
%result64 = std.rsqrt %arg_f64 : f64
// CHECK: llvm.call @__ocml_rsqrt_f64(%{{.*}}) : (!llvm.double) -> !llvm.double
std.return %result16, %result32, %result64 : f16, f32, f64
}
}
// -----
gpu.module @test_module {
// CHECK: llvm.func @__ocml_sqrt_f32(!llvm.float) -> !llvm.float // CHECK: llvm.func @__ocml_sqrt_f32(!llvm.float) -> !llvm.float
// CHECK: llvm.func @__ocml_sqrt_f64(!llvm.double) -> !llvm.double // CHECK: llvm.func @__ocml_sqrt_f64(!llvm.double) -> !llvm.double
// CHECK-LABEL: func @gpu_sqrt // CHECK-LABEL: func @gpu_sqrt
func @gpu_sqrt(%arg_f16 : f16, %arg_f32 : f32, %arg_f64 : f64) func @gpu_sqrt(%arg_f16 : f16, %arg_f32 : f32, %arg_f64 : f64)
-> (f16, f32, f64) { -> (f16, f32, f64) {
%result16 = std.sqrt %arg_f16 : f16 %result16 = std.sqrt %arg_f16 : f16
// CHECK: llvm.fpext %{{.*}} : !llvm.half to !llvm.float // CHECK: llvm.fpext %{{.*}} : !llvm.half to !llvm.float
// CHECK-NEXT: llvm.call @__ocml_sqrt_f32(%{{.*}}) : (!llvm.float) -> !llvm.float // CHECK-NEXT: llvm.call @__ocml_sqrt_f32(%{{.*}}) : (!llvm.float) -> !llvm.float
Show All 23 Lines