Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[mlir][amdgpu] Support for 8bit extf for 0d vector type #126102

Merged
merged 1 commit into from
Feb 6, 2025
Merged

[mlir][amdgpu] Support for 8bit extf for 0d vector type #126102

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 6 additions & 3 deletions mlir/lib/Conversion/ArithToAMDGPU/ArithToAMDGPU.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,20 +102,23 @@ void ExtFOnFloat8RewritePattern::rewrite(arith::ExtFOp op,
return rewriter.replaceOp(op, result);
}
int64_t numElements = inType.getNumElements();

Value zero = rewriter.create<arith::ConstantOp>(
loc, outElemType, rewriter.getFloatAttr(outElemType, 0.0));
VectorType outType = cast<VectorType>(op.getOut().getType());

if (inType.getShape().empty()) {
Value zerodSplat =
rewriter.createOrFold<vector::SplatOp>(loc, outType, zero);
Value scalarIn =
rewriter.create<vector::ExtractOp>(loc, in, ArrayRef<int64_t>{});
// Recurse to send the 0-D vector case to the 1-D vector case
Value scalarExt =
rewriter.create<arith::ExtFOp>(loc, outElemType, scalarIn);
Value result = rewriter.create<vector::InsertOp>(loc, scalarExt, zero,
Value result = rewriter.create<vector::InsertOp>(loc, scalarExt, zerodSplat,
ArrayRef<int64_t>{});
return rewriter.replaceOp(op, result);
}

VectorType outType = cast<VectorType>(op.getOut().getType());
VectorType flatTy = VectorType::get(SmallVector<int64_t>{numElements},
outType.getElementType());
Value result = rewriter.createOrFold<vector::SplatOp>(loc, flatTy, zero);
Expand Down
14 changes: 13 additions & 1 deletion mlir/test/Conversion/ArithToAMDGPU/8-bit-floats.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,19 @@ func.func @scalar_ext(%v: f8E5M2FNUZ) -> f16 {
return %w : f16
}

// No 0-D test because arith.extf hasn't been extended to support it.
// -----

// CHECK-LABEL: func.func @vector_zero_d(
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: vector<f8E5M2FNUZ>) -> vector<f32>
// CHECK: %[[CONST:.+]] = arith.constant dense<0.000000e+00> : vector<f32>
// CHECK: %[[EXTRACT:.+]] = vector.extract %[[ARG0]][] : f8E5M2FNUZ from vector<f8E5M2FNUZ>
// CHECK: %[[CONVERT:.+]] = amdgpu.ext_packed_fp8 %[[EXTRACT]][0] : f8E5M2FNUZ to f32
// CHECK: %[[RESULT:.+]] = vector.insert %[[CONVERT]], %[[CONST]] [] : f32 into vector<f32>
// CHECK: return %[[RESULT]] : vector<f32>
func.func @vector_zero_d(%v: vector<f8E5M2FNUZ>) -> vector<f32> {
%w = arith.extf %v : vector<f8E5M2FNUZ> to vector<f32>
return %w : vector<f32>
}

// -----

Expand Down