This is an archive of the discontinued LLVM Phabricator instance.

Differential D143008

[mlir] Add support for f8 data types to LLVM dialect types
ClosedPublic

Authored by ezhulenev on Jan 31 2023, 12:16 PM.

Details

Reviewers
ftynse
nicolasvasilache
dcaballe
reedwm
mpaszkowski
Commits
rGfaf697e49bad: [mlir] Add support for f8 data types to LLVM dialect types
Summary

This change allows using fp8 pointers when exporting to LLVM, because we anyway export them as opaque pointers, however full support of fp8 types is not yet implemented on the LLVM side.

Diff Detail

Event Timeline

ezhulenev created this revision.Jan 31 2023, 12:16 PM
Herald added a reviewer: ftynse. · View Herald TranscriptJan 31 2023, 12:16 PM
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: Moerafaat, zero9178, bzcheeseman and 22 others. · View Herald Transcript
ezhulenev requested review of this revision.Jan 31 2023, 12:16 PM
Herald added a reviewer: nicolasvasilache. · View Herald TranscriptJan 31 2023, 12:16 PM
Herald added a reviewer: dcaballe. · View Herald Transcript
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: stephenneuendorffer, nicolasvasilache. · View Herald Transcript
ezhulenev updated this revision to Diff 493707.Jan 31 2023, 12:26 PM

Do not add f8 to supported floating point types

ezhulenev edited the summary of this revision. (Show Details)Jan 31 2023, 12:29 PM
Herald added a reviewer: mpaszkowski. · View Herald TranscriptJan 31 2023, 12:29 PM
ezhulenev edited reviewers, added: reedwm; removed: mpaszkowski.Jan 31 2023, 12:30 PM
Herald added a reviewer: mpaszkowski. · View Herald TranscriptJan 31 2023, 12:30 PM
Harbormaster completed remote builds in B211056: Diff 493707.Jan 31 2023, 1:08 PM
ftynse added a comment.Feb 1 2023, 2:40 AM

Could you add a test?

ezhulenev updated this revision to Diff 494064.Feb 1 2023, 1:49 PM

Add a test for lowering f8 pointers to LLVM IR

Harbormaster completed remote builds in B211308: Diff 494064.Feb 1 2023, 3:06 PM
ezhulenev updated this revision to Diff 498359.Feb 17 2023, 7:12 AM

Rebased

ezhulenev updated this revision to Diff 498360.Feb 17 2023, 7:14 AM

Remove accidental change

Harbormaster completed remote builds in B214406: Diff 498360.Feb 17 2023, 7:43 AM
ezhulenev updated this revision to Diff 503789.Mar 9 2023, 8:28 AM

Rebased

Harbormaster completed remote builds in B218409: Diff 503789.Mar 9 2023, 9:15 AM
ezhulenev updated this revision to Diff 503873.Mar 9 2023, 11:59 AM

Rebased

Harbormaster completed remote builds in B218469: Diff 503873.Mar 9 2023, 1:50 PM
ezhulenev updated this revision to Diff 503994.Mar 9 2023, 5:47 PM

Rebased

Harbormaster completed remote builds in B218568: Diff 503994.Mar 9 2023, 6:26 PM
This revision was not accepted when it landed; it landed in state Needs Review.Mar 9 2023, 6:52 PM
Closed by commit rGfaf697e49bad: [mlir] Add support for f8 data types to LLVM dialect types (authored by ezhulenev). · Explain Why
This revision was automatically updated to reflect the committed changes.
ezhulenev added a commit: rGfaf697e49bad: [mlir] Add support for f8 data types to LLVM dialect types.

Revision Contents

PathSize
mlir/
lib/
Dialect/
LLVMIR/
IR/
4 lines
test/
Target/
LLVMIR/
5 lines

Diff 504006

mlir/lib/Dialect/LLVMIR/IR/LLVMTypes.cpp

Show First 20 Lines • Show All 725 Lines • ▼ Show 20 Linesbool mlir::LLVM::isCompatibleOuterType(Type type) {
// clang-format off // clang-format off
if (type.isa< if (type.isa<
BFloat16Type, BFloat16Type,
Float16Type, Float16Type,
Float32Type, Float32Type,
Float64Type, Float64Type,
Float80Type, Float80Type,
Float128Type, Float128Type,
Float8E4M3FNType,
Float8E5M2Type,
LLVMArrayType, LLVMArrayType,
LLVMFunctionType, LLVMFunctionType,
LLVMLabelType, LLVMLabelType,
LLVMMetadataType, LLVMMetadataType,
LLVMPPCFP128Type, LLVMPPCFP128Type,
LLVMPointerType, LLVMPointerType,
LLVMStructType, LLVMStructType,
LLVMTokenType, LLVMTokenType,
▲ Show 20 Lines • Show All 54 Lines • ▼ Show 20 Lines bool result =
}) })
.Case< .Case<
BFloat16Type, BFloat16Type,
Float16Type, Float16Type,
Float32Type, Float32Type,
Float64Type, Float64Type,
Float80Type, Float80Type,
Float128Type, Float128Type,
Float8E4M3FNType,
Float8E5M2Type,
LLVMLabelType, LLVMLabelType,
LLVMMetadataType, LLVMMetadataType,
LLVMPPCFP128Type, LLVMPPCFP128Type,
LLVMTokenType, LLVMTokenType,
LLVMVoidType, LLVMVoidType,
LLVMX86MMXType LLVMX86MMXType
>([](Type) { return true; }) >([](Type) { return true; })
// clang-format on // clang-format on
▲ Show 20 Lines • Show All 183 Lines • Show Last 20 Lines

mlir/test/Target/LLVMIR/llvmir.mlir

Show First 20 Lines • Show All 938 Lines • ▼ Show 20 Lines
// CHECK-LABEL: @vector_splat_nonzero_scalable // CHECK-LABEL: @vector_splat_nonzero_scalable
llvm.func @vector_splat_nonzero_scalable() -> vector<[4]xf32> { llvm.func @vector_splat_nonzero_scalable() -> vector<[4]xf32> {
// CHECK: ret <vscale x 4 x float> shufflevector (<vscale x 4 x float> insertelement (<vscale x 4 x float> poison, float 1.000000e+00, i64 0), <vscale x 4 x float> poison, <vscale x 4 x i32> zeroinitializer) // CHECK: ret <vscale x 4 x float> shufflevector (<vscale x 4 x float> insertelement (<vscale x 4 x float> poison, float 1.000000e+00, i64 0), <vscale x 4 x float> poison, <vscale x 4 x i32> zeroinitializer)
%0 = llvm.mlir.constant(dense<1.000000e+00> : vector<[4]xf32>) : vector<[4]xf32> %0 = llvm.mlir.constant(dense<1.000000e+00> : vector<[4]xf32>) : vector<[4]xf32>
llvm.return %0 : vector<[4]xf32> llvm.return %0 : vector<[4]xf32>
} }
// CHECK-LABEL: @f8_ptrs(ptr {{%.*}}, ptr {{%.*}})
llvm.func @f8_ptrs(%arg0: !llvm.ptr<f8E5M2>, %arg1: !llvm.ptr<f8E4M3FN>) {
llvm.return
}
// CHECK-LABEL: @ops // CHECK-LABEL: @ops
llvm.func @ops(%arg0: f32, %arg1: f32, %arg2: i32, %arg3: i32) -> !llvm.struct<(f32, i32)> { llvm.func @ops(%arg0: f32, %arg1: f32, %arg2: i32, %arg3: i32) -> !llvm.struct<(f32, i32)> {
// CHECK-NEXT: fsub float %0, %1 // CHECK-NEXT: fsub float %0, %1
%0 = llvm.fsub %arg0, %arg1 : f32 %0 = llvm.fsub %arg0, %arg1 : f32
// CHECK-NEXT: %6 = sub i32 %2, %3 // CHECK-NEXT: %6 = sub i32 %2, %3
%1 = llvm.sub %arg2, %arg3 : i32 %1 = llvm.sub %arg2, %arg3 : i32
// CHECK-NEXT: %7 = icmp slt i32 %2, %6 // CHECK-NEXT: %7 = icmp slt i32 %2, %6
%2 = llvm.icmp "slt" %arg2, %1 : i32 %2 = llvm.icmp "slt" %arg2, %1 : i32
▲ Show 20 Lines • Show All 1,212 Lines • Show Last 20 Lines