replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[mlir][tosa] Add clamp + clamp as single clamp canonicalization 

When 2 clamp ops are in a row, they can be canonicalized into a single clamp
that uses the most constrained range

Reviewed By: rsuderman

Differential Revision: https://reviews.llvm.org/D117934
This commit is contained in:
not-jenni 2022-01-21 16:16:29 -08:00 committed by Rob Suderman
parent 1613f8b8d7
commit 08574ce4d6
2 changed files with 41 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
View File

@ -526,9 +526,40 @@ struct ClampIsNoOp : public OpRewritePattern<tosa::ClampOp> {
}
};
struct ClampClampOptimization : public OpRewritePattern<tosa::ClampOp> {
using OpRewritePattern<tosa::ClampOp>::OpRewritePattern;
LogicalResult matchAndRewrite(tosa::ClampOp op,
PatternRewriter &rewriter) const override {
Value input = op.input();
Operation *definingOp = input.getDefiningOp();
if (!definingOp)
return failure();
if (tosa::ClampOp clampOp = dyn_cast<tosa::ClampOp>(definingOp)) {
auto min_fp = std::max(op.min_fp(), clampOp.min_fp()).convertToFloat();
auto max_fp = std::min(op.max_fp(), clampOp.max_fp()).convertToFloat();
auto min_int = std::max(op.min_int(), clampOp.min_int());
auto max_int = std::min(op.max_int(), clampOp.max_int());
rewriter.replaceOpWithNewOp<tosa::ClampOp>(
op, op.getType(), clampOp.input(),
rewriter.getI64IntegerAttr(min_int),
rewriter.getI64IntegerAttr(max_int), rewriter.getF32FloatAttr(min_fp),
rewriter.getF32FloatAttr(max_fp));
return success();
}
return failure();
}
};
void ClampOp::getCanonicalizationPatterns(OwningRewritePatternList &results,
MLIRContext *context) {
results.insert<ClampIsNoOp>(context);
results.insert<ClampClampOptimization>(context);
}
//===----------------------------------------------------------------------===//

10
mlir/test/Dialect/Tosa/canonicalize.mlir
View File

@ -98,6 +98,16 @@ func @clamp_uint8_is_noop(%arg0: tensor<4xui8>) -> tensor<4xui8> {
// -----
// CHECK-LABEL: @clamp_twice_is_single_clamp
func @clamp_twice_is_single_clamp(%arg0: tensor<4xi8>) -> tensor<4xi8> {
// CHECK: "tosa.clamp"(%arg0) {max_fp = 3.000000e+00 : f32, max_int = 2 : i64, min_fp = -3.000000e+00 : f32, min_int = -2 : i64}
%0 = "tosa.clamp"(%arg0) {max_fp = 3.0 : f32, max_int = 4 : i64, min_fp = -5.0 : f32, min_int = -2 : i64} : (tensor<4xi8>) -> tensor<4xi8>
%1 = "tosa.clamp"(%0) {max_fp = 5.0 : f32, max_int = 2 : i64, min_fp = -3.0 : f32, min_int = -4 : i64} : (tensor<4xi8>) -> tensor<4xi8>
return %1 : tensor<4xi8>
}
// -----
// CHECK-LABEL: @concat_fold
func @concat_fold(%arg0: tensor<?x1xf32>) -> tensor<?x1xf32> {
// CHECK: return %arg0