This is an archive of the discontinued LLVM Phabricator instance.

Differential D112011

[mlir] Update approximation range for Tanh operation
ClosedPublic

Authored by ezhulenev on Oct 18 2021, 9:06 AM.

Details

Reviewers
cota
Commits
rGbf32bb7e056c: [mlir] Update approximation range for Tanh operation
Summary

Use wider range for approximating Tanh to match results computed in Eigen with AVX.

Diff Detail

Event Timeline

ezhulenev created this revision.Oct 18 2021, 9:06 AM
Herald added subscribers: wenzhicui, wrengr, Chia-hungDuan and 19 others. · View Herald TranscriptOct 18 2021, 9:06 AM
ezhulenev requested review of this revision.Oct 18 2021, 9:06 AM
Herald added a project: Restricted Project. · View Herald TranscriptOct 18 2021, 9:06 AM
Herald added subscribers: stephenneuendorffer, nicolasvasilache. · View Herald Transcript
ezhulenev edited the summary of this revision. (Show Details)Oct 18 2021, 9:09 AM
ezhulenev added a reviewer: cota.
cota accepted this revision.Oct 18 2021, 9:27 AM
This revision is now accepted and ready to land.Oct 18 2021, 9:27 AM
Harbormaster completed remote builds in B129372: Diff 380434.Oct 18 2021, 9:45 AM
Closed by commit rGbf32bb7e056c: [mlir] Update approximation range for Tanh operation (authored by ezhulenev). · Explain WhyOct 18 2021, 10:57 AM
This revision was automatically updated to reflect the committed changes.
ezhulenev added a commit: rGbf32bb7e056c: [mlir] Update approximation range for Tanh operation.

Revision Contents

PathSize
mlir/
lib/
Dialect/
Math/
Transforms/
4 lines
test/
Dialect/
Math/
6 lines

Diff 380480

mlir/lib/Dialect/Math/Transforms/PolynomialApproximation.cpp

Show First 20 Lines • Show All 203 Lines • ▼ Show 20 Lines if (!width.hasValue())
return rewriter.notifyMatchFailure(op, "unsupported operand type"); return rewriter.notifyMatchFailure(op, "unsupported operand type");
ImplicitLocOpBuilder builder(op->getLoc(), rewriter); ImplicitLocOpBuilder builder(op->getLoc(), rewriter);
auto bcast = [&](Value value) -> Value { auto bcast = [&](Value value) -> Value {
return broadcast(builder, value, *width); return broadcast(builder, value, *width);
}; };
// Clamp operand into [plusClamp, minusClamp] range. // Clamp operand into [plusClamp, minusClamp] range.
Value minusClamp = bcast(f32Cst(builder, -7.9053111076354980f)); Value minusClamp = bcast(f32Cst(builder, -7.99881172180175781f));
Value plusClamp = bcast(f32Cst(builder, 7.90531110763549805f)); Value plusClamp = bcast(f32Cst(builder, 7.99881172180175781f));
Value x = clamp(builder, op.operand(), minusClamp, plusClamp); Value x = clamp(builder, op.operand(), minusClamp, plusClamp);
// Mask for tiny values that are approximated with `operand`. // Mask for tiny values that are approximated with `operand`.
Value tiny = bcast(f32Cst(builder, 0.0004f)); Value tiny = bcast(f32Cst(builder, 0.0004f));
Value tinyMask = builder.create<arith::CmpFOp>( Value tinyMask = builder.create<arith::CmpFOp>(
arith::CmpFPredicate::OLT, builder.create<math::AbsOp>(op.operand()), arith::CmpFPredicate::OLT, builder.create<math::AbsOp>(op.operand()),
tiny); tiny);
▲ Show 20 Lines • Show All 564 Lines • Show Last 20 Lines

mlir/test/Dialect/Math/polynomial-approximation.mlir

Show First 20 Lines • Show All 242 Lines • ▼ Show 20 Lines
func @log1p_vector(%arg0: vector<8xf32>) -> vector<8xf32> { func @log1p_vector(%arg0: vector<8xf32>) -> vector<8xf32> {
%0 = math.log1p %arg0 : vector<8xf32> %0 = math.log1p %arg0 : vector<8xf32>
return %0 : vector<8xf32> return %0 : vector<8xf32>
} }
// CHECK-LABEL: func @tanh_scalar( // CHECK-LABEL: func @tanh_scalar(
// CHECK-SAME: %[[VAL_0:.*]]: f32) -> f32 { // CHECK-SAME: %[[VAL_0:.*]]: f32) -> f32 {
// CHECK: %[[VAL_1:.*]] = arith.constant -7.90531111 : f32 // CHECK: %[[VAL_1:.*]] = arith.constant -7.99881172 : f32
// CHECK: %[[VAL_2:.*]] = arith.constant 7.90531111 : f32 // CHECK: %[[VAL_2:.*]] = arith.constant 7.99881172 : f32
// CHECK: %[[VAL_3:.*]] = arith.constant 4.000000e-04 : f32 // CHECK: %[[VAL_3:.*]] = arith.constant 4.000000e-04 : f32
// CHECK: %[[VAL_4:.*]] = arith.constant 0.00489352457 : f32 // CHECK: %[[VAL_4:.*]] = arith.constant 0.00489352457 : f32
// CHECK: %[[VAL_5:.*]] = arith.constant 6.37261954E-4 : f32 // CHECK: %[[VAL_5:.*]] = arith.constant 6.37261954E-4 : f32
// CHECK: %[[VAL_6:.*]] = arith.constant 1.48572235E-5 : f32 // CHECK: %[[VAL_6:.*]] = arith.constant 1.48572235E-5 : f32
// CHECK: %[[VAL_7:.*]] = arith.constant 5.12229725E-8 : f32 // CHECK: %[[VAL_7:.*]] = arith.constant 5.12229725E-8 : f32
// CHECK: %[[VAL_8:.*]] = arith.constant -8.60467184E-11 : f32 // CHECK: %[[VAL_8:.*]] = arith.constant -8.60467184E-11 : f32
// CHECK: %[[VAL_9:.*]] = arith.constant 2.00018794E-13 : f32 // CHECK: %[[VAL_9:.*]] = arith.constant 2.00018794E-13 : f32
// CHECK: %[[VAL_10:.*]] = arith.constant -2.76076837E-16 : f32 // CHECK: %[[VAL_10:.*]] = arith.constant -2.76076837E-16 : f32
Show All 24 Lines
// CHECK: } // CHECK: }
func @tanh_scalar(%arg0: f32) -> f32 { func @tanh_scalar(%arg0: f32) -> f32 {
%0 = math.tanh %arg0 : f32 %0 = math.tanh %arg0 : f32
return %0 : f32 return %0 : f32
} }
// CHECK-LABEL: func @tanh_vector( // CHECK-LABEL: func @tanh_vector(
// CHECK-SAME: %[[VAL_0:.*]]: vector<8xf32>) -> vector<8xf32> { // CHECK-SAME: %[[VAL_0:.*]]: vector<8xf32>) -> vector<8xf32> {
// CHECK: %[[VAL_1:.*]] = arith.constant dense<-7.90531111> : vector<8xf32> // CHECK: %[[VAL_1:.*]] = arith.constant dense<-7.99881172> : vector<8xf32>
// CHECK-NOT: tanh // CHECK-NOT: tanh
// CHECK-COUNT-2: select // CHECK-COUNT-2: select
// CHECK: %[[VAL_33:.*]] = select // CHECK: %[[VAL_33:.*]] = select
// CHECK: return %[[VAL_33]] : vector<8xf32> // CHECK: return %[[VAL_33]] : vector<8xf32>
// CHECK: } // CHECK: }
func @tanh_vector(%arg0: vector<8xf32>) -> vector<8xf32> { func @tanh_vector(%arg0: vector<8xf32>) -> vector<8xf32> {
%0 = math.tanh %arg0 : vector<8xf32> %0 = math.tanh %arg0 : vector<8xf32>
return %0 : vector<8xf32> return %0 : vector<8xf32>
} }