This is an archive of the discontinued LLVM Phabricator instance.

Differential D141015

[mlir][tosa] Remove clamping behavior in `tosa.cast` for integer truncation
ClosedPublic

Authored by mariecwhite on Jan 4 2023, 2:49 PM.

Details

Reviewers
rsuderman
Commits
rG76dc9a853add: [mlir][tosa] Remove clamping behavior in `tosa.cast` for integer truncation

Diff Detail

Event Timeline

mariecwhite created this revision.Jan 4 2023, 2:49 PM
Herald added a project: Restricted Project. · View Herald TranscriptJan 4 2023, 2:49 PM
Herald added subscribers: hanchung, Moerafaat, zero9178 and 26 others. · View Herald Transcript
mariecwhite requested review of this revision.Jan 4 2023, 2:49 PM
Herald added a project: Restricted Project. · View Herald TranscriptJan 4 2023, 2:49 PM
Herald added subscribers: stephenneuendorffer, nicolasvasilache. · View Herald Transcript
mariecwhite added a reviewer: rsuderman.Jan 4 2023, 2:49 PM
rsuderman accepted this revision.Jan 4 2023, 2:50 PM
This revision is now accepted and ready to land.Jan 4 2023, 2:50 PM
Harbormaster completed remote builds in B205780: Diff 486400.Jan 4 2023, 3:07 PM
Closed by commit rG76dc9a853add: [mlir][tosa] Remove clamping behavior in `tosa.cast` for integer truncation (authored by mariecwhite, committed by rsuderman). · Explain WhyJan 4 2023, 3:11 PM
This revision was automatically updated to reflect the committed changes.
rsuderman added a commit: rG76dc9a853add: [mlir][tosa] Remove clamping behavior in `tosa.cast` for integer truncation.

Revision Contents

PathSize
mlir/
lib/
Conversion/
TosaToLinalg/
15 lines
test/
Conversion/
TosaToLinalg/
6 lines

Diff 486400

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

Show First 20 Lines • Show All 506 Lines • ▼ Show 20 Lines if (srcTy.isa<IntegerType>() && dstTy.isInteger(1)) {
args.front(), zero); args.front(), zero);
} }
if (srcTy.isa<IntegerType>() && dstTy.isa<IntegerType>() && bitExtend) if (srcTy.isa<IntegerType>() && dstTy.isa<IntegerType>() && bitExtend)
return rewriter.create<arith::ExtSIOp>(loc, resultTypes, args, return rewriter.create<arith::ExtSIOp>(loc, resultTypes, args,
std::nullopt); std::nullopt);
if (srcTy.isa<IntegerType>() && dstTy.isa<IntegerType>() && !bitExtend) { if (srcTy.isa<IntegerType>() && dstTy.isa<IntegerType>() && !bitExtend) {
auto intMin = rewriter.create<arith::ConstantIntOp>( return rewriter.create<arith::TruncIOp>(loc, dstTy, args[0]);
loc,
APInt::getSignedMinValue(dstTy.getIntOrFloatBitWidth())
.getSExtValue(),
srcTy.getIntOrFloatBitWidth());
auto intMax = rewriter.create<arith::ConstantIntOp>(
loc,
APInt::getSignedMaxValue(dstTy.getIntOrFloatBitWidth())
.getSExtValue(),
srcTy.getIntOrFloatBitWidth());
auto clamped = clampIntHelper(loc, args[0], intMin, intMax, rewriter);
return rewriter.create<arith::TruncIOp>(loc, dstTy, clamped);
} }
} }
(void)rewriter.notifyMatchFailure( (void)rewriter.notifyMatchFailure(
op, "unhandled op for linalg body calculation for elementwise op"); op, "unhandled op for linalg body calculation for elementwise op");
return nullptr; return nullptr;
} }
▲ Show 20 Lines • Show All 1,852 Lines • Show Last 20 Lines

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir

Show First 20 Lines • Show All 394 Lines • ▼ Show 20 Linesfunc.func @test_simple_i32(%arg0: tensor<1xi32>) -> () {
%18 = "tosa.minimum"(%0, %1) : (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32> %18 = "tosa.minimum"(%0, %1) : (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>
// CHECK: linalg.generic // CHECK: linalg.generic
// CHECK: arith.cmpi // CHECK: arith.cmpi
// CHECK: select // CHECK: select
%19 = "tosa.clamp"(%0) {min_int = 1 : i64, max_int = 5 : i64, min_fp = 1.0 : f32, max_fp = 5.0 : f32} : (tensor<1xi32>) -> tensor<1xi32> %19 = "tosa.clamp"(%0) {min_int = 1 : i64, max_int = 5 : i64, min_fp = 1.0 : f32, max_fp = 5.0 : f32} : (tensor<1xi32>) -> tensor<1xi32>
// CHECK: linalg.generic // CHECK: linalg.generic
// CHECK: arith.constant -32768
// CHECK: arith.constant 32767
// CHECK: arith.cmpi slt
// CHECK: select
// CHECK: arith.cmpi slt
// CHECK: select
// CHECK: arith.trunci // CHECK: arith.trunci
%20 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xi16> %20 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xi16>
// CHECK: linalg.generic // CHECK: linalg.generic
// CHECK: arith.extsi // CHECK: arith.extsi
%21 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xi64> %21 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xi64>
// CHECK: linalg.generic // CHECK: linalg.generic
▲ Show 20 Lines • Show All 1,146 Lines • Show Last 20 Lines