Skip to content

Commit 68574f7

Browse files
jsjodinjsjodin
committed
Address review comments
1 parent ee7245e commit 68574f7

File tree

3 files changed

+80
-80
lines changed

3 files changed

+80
-80
lines changed

‎mlir/include/mlir/Dialect/Math/IR/MathOps.td‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -550,10 +550,10 @@ def Math_ErfOp : Math_FloatUnaryOp<"erf"> {
550550
def Math_ErfcOp : Math_FloatUnaryOp<"erfc"> {
551551
let summary = "complementary error function of the specified value";
552552
let description = [{
553-
The `erfc` operation computes the complementary error function.
554-
It takes one operand of floating point type (i.e., scalar, tensor or
555-
vector) and returns one result of the same type.
556-
It has no standard attributes.
553+
The `erfc` operation computes the complementary error function, defined
554+
as 1-erf(x). It takes one operand of floating point type (i.e., scalar,
555+
tensor or vector) and returns one result of the same type. It has no
556+
standard attributes.
557557

558558
Example:
559559

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

Lines changed: 7 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -1125,6 +1125,12 @@ ErfPolynomialApproximation::matchAndRewrite(math::ErfOp op,
11251125
// Approximates erfc(x) with p((x - 2) / (x + 2)), where p is a 9 degree
11261126
// polynomial.This approximation is based on the following stackoverflow post:
11271127
// https://stackoverflow.com/questions/35966695/vectorizable-implementation-of-complementary-error-function-erfcf
1128+
// The stackoverflow post is in turn based on:
1129+
// M. M. Shepherd and J. G. Laframboise, "Chebyshev Approximation of
1130+
// (1+2x)exp(x^2)erfc x in 0 <= x < INF", Mathematics of Computation, Vol. 36,
1131+
// No. 153, January 1981, pp. 249-253.
1132+
//
1133+
// Maximum error: 2.65 ulps
11281134
LogicalResult
11291135
ErfcPolynomialApproximation::matchAndRewrite(math::ErfcOp op,
11301136
PatternRewriter &rewriter) const {
@@ -1150,11 +1156,7 @@ ErfcPolynomialApproximation::matchAndRewrite(math::ErfcOp op,
11501156
Value posInf = bcast(f32FromBits(builder, 0x7f800000u));
11511157
Value clampVal = bcast(floatCst(builder, 10.0546875f, et));
11521158

1153-
// Get abs(x)
1154-
Value isNegativeArg =
1155-
builder.create<arith::CmpFOp>(arith::CmpFPredicate::OLT, x, zero);
1156-
Value negArg = builder.create<arith::NegFOp>(x);
1157-
Value a = builder.create<arith::SelectOp>(isNegativeArg, negArg, x);
1159+
Value a = builder.create<math::AbsFOp>(x);
11581160
Value p = builder.create<arith::AddFOp>(a, pos2);
11591161
Value r = builder.create<arith::DivFOp>(one, p);
11601162
Value q = builder.create<math::FmaOp>(neg4, r, one);

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

Lines changed: 69 additions & 71 deletions
Original file line numberDiff line numberDiff line change
@@ -116,77 +116,75 @@ func.func @erf_scalar(%arg0: f32) -> f32 {
116116
// CHECK-DAG: %[[cst_26:.*]] = arith.constant 2.000000e+00 : f32
117117
// CHECK-DAG: %[[cst_27:.*]] = arith.constant 0x7F800000 : f32
118118
// CHECK-DAG: %[[cst_28:.*]] = arith.constant 10.0546875 : f32
119-
// CHECK: %[[val_0:.*]] = arith.cmpf olt, %[[val_arg0]], %[[cst_21]] : f32
120-
// CHECK-NEXT: %[[val_1:.*]] = arith.negf %[[val_arg0]] : f32
121-
// CHECK-NEXT: %[[val_2:.*]] = arith.select %[[val_0]], %[[val_1]], %[[val_arg0]] : f32
122-
// CHECK-NEXT: %[[val_3:.*]] = arith.addf %[[val_2]], %[[cst_26]] : f32
123-
// CHECK-NEXT: %[[val_4:.*]] = arith.divf %[[cst_22]], %[[val_3]] : f32
124-
// CHECK-NEXT: %[[val_5:.*]] = math.fma %[[cst_24]], %[[val_4]], %[[cst_22]] : f32
125-
// CHECK-NEXT: %[[val_6:.*]] = arith.addf %[[val_5]], %[[cst_22]] : f32
126-
// CHECK-NEXT: %[[val_7:.*]] = math.fma %[[val_6]], %[[cst_25]], %[[val_2]] : f32
127-
// CHECK-NEXT: %[[val_8:.*]] = arith.negf %[[val_2]] : f32
128-
// CHECK-NEXT: %[[val_9:.*]] = math.fma %[[val_8]], %[[val_5]], %[[val_7]] : f32
129-
// CHECK-NEXT: %[[val_10:.*]] = math.fma %[[val_4]], %[[val_9]], %[[val_5]] : f32
130-
// CHECK-NEXT: %[[val_11:.*]] = math.fma %[[cst_20]], %[[val_10]], %[[cst_19]] : f32
131-
// CHECK-NEXT: %[[val_12:.*]] = math.fma %[[val_11]], %[[val_10]], %[[cst_18]] : f32
132-
// CHECK-NEXT: %[[val_13:.*]] = math.fma %[[val_12]], %[[val_10]], %[[cst_17]] : f32
133-
// CHECK-NEXT: %[[val_14:.*]] = math.fma %[[val_13]], %[[val_10]], %[[cst_16]] : f32
134-
// CHECK-NEXT: %[[val_15:.*]] = math.fma %[[val_14]], %[[val_10]], %[[cst_15]] : f32
135-
// CHECK-NEXT: %[[val_16:.*]] = math.fma %[[val_15]], %[[val_10]], %[[cst_14]] : f32
136-
// CHECK-NEXT: %[[val_17:.*]] = math.fma %[[val_16]], %[[val_10]], %[[cst_13]] : f32
137-
// CHECK-NEXT: %[[val_18:.*]] = math.fma %[[val_17]], %[[val_10]], %[[cst_12]] : f32
138-
// CHECK-NEXT: %[[val_19:.*]] = math.fma %[[val_18]], %[[val_10]], %[[cst_11]] : f32
139-
// CHECK-NEXT: %[[val_20:.*]] = math.fma %[[cst_26]], %[[val_2]], %[[cst_22]] : f32
140-
// CHECK-NEXT: %[[val_21:.*]] = arith.divf %[[cst_22]], %[[val_20]] : f32
141-
// CHECK-NEXT: %[[val_22:.*]] = math.fma %[[val_19]], %[[val_21]], %[[val_21]] : f32
142-
// CHECK-NEXT: %[[val_23:.*]] = arith.negf %[[val_2]] : f32
143-
// CHECK-NEXT: %[[val_24:.*]] = math.fma %[[val_22]], %[[val_23]], %[[cst_23]] : f32
144-
// CHECK-NEXT: %[[val_25:.*]] = arith.subf %[[val_19]], %[[val_22]] : f32
145-
// CHECK-NEXT: %[[val_26:.*]] = math.fma %[[val_24]], %[[cst_26]], %[[val_25]] : f32
146-
// CHECK-NEXT: %[[val_27:.*]] = math.fma %[[val_26]], %[[val_21]], %[[val_22]] : f32
147-
// CHECK-NEXT: %[[val_28:.*]] = arith.mulf %[[val_2]], %[[val_2]] : f32
148-
// CHECK-NEXT: %[[val_29:.*]] = arith.negf %[[val_28]] : f32
149-
// CHECK-NEXT: %[[val_30:.*]] = arith.cmpf uge, %[[val_29]], %[[cst_2]] : f32
150-
// CHECK-NEXT: %[[val_31:.*]] = arith.select %[[val_30]], %[[val_29]], %[[cst_2]] : f32
151-
// CHECK-NEXT: %[[val_32:.*]] = arith.cmpf ule, %[[val_31]], %[[cst_1]] : f32
152-
// CHECK-NEXT: %[[val_33:.*]] = arith.select %[[val_32]], %[[val_31]], %[[cst_1]] : f32
153-
// CHECK-NEXT: %[[val_34:.*]] = math.fma %[[val_33]], %[[cst_10]], %[[cst_23]] : f32
154-
// CHECK-NEXT: %[[val_35:.*]] = math.floor %[[val_34]] : f32
155-
// CHECK-NEXT: %[[val_36:.*]] = arith.cmpf uge, %[[val_35]], %[[cst_0]] : f32
156-
// CHECK-NEXT: %[[val_37:.*]] = arith.select %[[val_36]], %[[val_35]], %[[cst_0]] : f32
157-
// CHECK-NEXT: %[[val_38:.*]] = arith.cmpf ule, %[[val_37]], %[[cst]] : f32
158-
// CHECK-NEXT: %[[val_39:.*]] = arith.select %[[val_38]], %[[val_37]], %[[cst]] : f32
159-
// CHECK-NEXT: %[[val_40:.*]] = math.fma %[[cst_9]], %[[val_39]], %[[val_33]] : f32
160-
// CHECK-NEXT: %[[val_41:.*]] = math.fma %[[cst_8]], %[[val_39]], %[[val_40]] : f32
161-
// CHECK-NEXT: %[[val_42:.*]] = math.fma %[[val_41]], %[[cst_7]], %[[cst_6]] : f32
162-
// CHECK-NEXT: %[[val_43:.*]] = math.fma %[[val_42]], %[[val_41]], %[[cst_5]] : f32
163-
// CHECK-NEXT: %[[val_44:.*]] = math.fma %[[val_43]], %[[val_41]], %[[cst_4]] : f32
164-
// CHECK-NEXT: %[[val_45:.*]] = math.fma %[[val_44]], %[[val_41]], %[[cst_3]] : f32
165-
// CHECK-NEXT: %[[val_46:.*]] = math.fma %[[val_45]], %[[val_41]], %[[cst_23]] : f32
166-
// CHECK-NEXT: %[[val_47:.*]] = arith.mulf %[[val_41]], %[[val_41]] : f32
167-
// CHECK-NEXT: %[[val_48:.*]] = math.fma %[[val_46]], %[[val_47]], %[[val_41]] : f32
168-
// CHECK-NEXT: %[[val_49:.*]] = arith.addf %[[val_48]], %[[cst_22]] : f32
169-
// CHECK-NEXT: %[[val_50:.*]] = arith.fptosi %[[val_39]] : f32 to i32
170-
// CHECK-NEXT: %[[val_51:.*]] = arith.addi %[[val_50]], %[[c127_i32]] : i32
171-
// CHECK-NEXT: %[[val_52:.*]] = arith.shli %[[val_51]], %[[c23_i32]] : i32
172-
// CHECK-NEXT: %[[val_53:.*]] = arith.bitcast %[[val_52]] : i32 to f32
173-
// CHECK-NEXT: %[[val_54:.*]] = arith.mulf %[[val_49]], %[[val_53]] : f32
174-
// CHECK-NEXT: %[[val_55:.*]] = arith.negf %[[val_2]] : f32
175-
// CHECK-NEXT: %[[val_56:.*]] = math.fma %[[val_55]], %[[val_2]], %[[val_28]] : f32
176-
// CHECK-NEXT: %[[val_57:.*]] = arith.mulf %[[val_27]], %[[val_54]] : f32
177-
// CHECK-NEXT: %[[val_58:.*]] = arith.mulf %[[val_57]], %[[val_56]] : f32
178-
// CHECK-NEXT: %[[val_59:.*]] = math.fma %[[val_27]], %[[val_54]], %[[val_58]] : f32
179-
// CHECK-NEXT: %[[val_60:.*]] = arith.cmpf olt, %[[val_2]], %[[cst_27]] : f32
180-
// CHECK-NEXT: %[[val_61:.*]] = arith.xori %[[val_60]], %[[cst_true]] : i1
181-
// CHECK-NEXT: %[[val_62:.*]] = arith.addf %[[val_arg0]], %[[val_arg0]] : f32
182-
// CHECK-NEXT: %[[val_63:.*]] = arith.select %[[val_61]], %[[val_62]], %[[val_59]] : f32
183-
// CHECK-NEXT: %[[val_64:.*]] = arith.cmpf ogt, %[[val_2]], %[[cst_28]] : f32
184-
// CHECK-NEXT: %[[val_65:.*]] = arith.select %[[val_64]], %[[cst_21]], %[[val_63]] : f32
185-
// CHECK-NEXT: %[[val_66:.*]] = arith.cmpf olt, %[[val_arg0]], %[[cst_21]] : f32
186-
// CHECK-NEXT: %[[val_67:.*]] = arith.subf %[[cst_26]], %[[val_65]] : f32
187-
// CHECK-NEXT: %[[val_68:.*]] = arith.select %[[val_66]], %[[val_67]], %[[val_65]] : f32
188-
// CHECK-NEXT: return %[[val_68]] : f32
189-
// CHECK-NEXT: }
119+
// CHECK: %[[val_2:.*]] = math.absf %[[val_arg0]] : f32
120+
// CHECK-NEXT: %[[val_3:.*]] = arith.addf %[[val_2]], %[[cst_26]] : f32
121+
// CHECK-NEXT: %[[val_4:.*]] = arith.divf %[[cst_22]], %[[val_3]] : f32
122+
// CHECK-NEXT: %[[val_5:.*]] = math.fma %[[cst_24]], %[[val_4]], %[[cst_22]] : f32
123+
// CHECK-NEXT: %[[val_6:.*]] = arith.addf %[[val_5]], %[[cst_22]] : f32
124+
// CHECK-NEXT: %[[val_7:.*]] = math.fma %[[val_6]], %[[cst_25]], %[[val_2]] : f32
125+
// CHECK-NEXT: %[[val_8:.*]] = arith.negf %[[val_2]] : f32
126+
// CHECK-NEXT: %[[val_9:.*]] = math.fma %[[val_8]], %[[val_5]], %[[val_7]] : f32
127+
// CHECK-NEXT: %[[val_10:.*]] = math.fma %[[val_4]], %[[val_9]], %[[val_5]] : f32
128+
// CHECK-NEXT: %[[val_11:.*]] = math.fma %[[cst_20]], %[[val_10]], %[[cst_19]] : f32
129+
// CHECK-NEXT: %[[val_12:.*]] = math.fma %[[val_11]], %[[val_10]], %[[cst_18]] : f32
130+
// CHECK-NEXT: %[[val_13:.*]] = math.fma %[[val_12]], %[[val_10]], %[[cst_17]] : f32
131+
// CHECK-NEXT: %[[val_14:.*]] = math.fma %[[val_13]], %[[val_10]], %[[cst_16]] : f32
132+
// CHECK-NEXT: %[[val_15:.*]] = math.fma %[[val_14]], %[[val_10]], %[[cst_15]] : f32
133+
// CHECK-NEXT: %[[val_16:.*]] = math.fma %[[val_15]], %[[val_10]], %[[cst_14]] : f32
134+
// CHECK-NEXT: %[[val_17:.*]] = math.fma %[[val_16]], %[[val_10]], %[[cst_13]] : f32
135+
// CHECK-NEXT: %[[val_18:.*]] = math.fma %[[val_17]], %[[val_10]], %[[cst_12]] : f32
136+
// CHECK-NEXT: %[[val_19:.*]] = math.fma %[[val_18]], %[[val_10]], %[[cst_11]] : f32
137+
// CHECK-NEXT: %[[val_20:.*]] = math.fma %[[cst_26]], %[[val_2]], %[[cst_22]] : f32
138+
// CHECK-NEXT: %[[val_21:.*]] = arith.divf %[[cst_22]], %[[val_20]] : f32
139+
// CHECK-NEXT: %[[val_22:.*]] = math.fma %[[val_19]], %[[val_21]], %[[val_21]] : f32
140+
// CHECK-NEXT: %[[val_23:.*]] = arith.negf %[[val_2]] : f32
141+
// CHECK-NEXT: %[[val_24:.*]] = math.fma %[[val_22]], %[[val_23]], %[[cst_23]] : f32
142+
// CHECK-NEXT: %[[val_25:.*]] = arith.subf %[[val_19]], %[[val_22]] : f32
143+
// CHECK-NEXT: %[[val_26:.*]] = math.fma %[[val_24]], %[[cst_26]], %[[val_25]] : f32
144+
// CHECK-NEXT: %[[val_27:.*]] = math.fma %[[val_26]], %[[val_21]], %[[val_22]] : f32
145+
// CHECK-NEXT: %[[val_28:.*]] = arith.mulf %[[val_2]], %[[val_2]] : f32
146+
// CHECK-NEXT: %[[val_29:.*]] = arith.negf %[[val_28]] : f32
147+
// CHECK-NEXT: %[[val_30:.*]] = arith.cmpf uge, %[[val_29]], %[[cst_2]] : f32
148+
// CHECK-NEXT: %[[val_31:.*]] = arith.select %[[val_30]], %[[val_29]], %[[cst_2]] : f32
149+
// CHECK-NEXT: %[[val_32:.*]] = arith.cmpf ule, %[[val_31]], %[[cst_1]] : f32
150+
// CHECK-NEXT: %[[val_33:.*]] = arith.select %[[val_32]], %[[val_31]], %[[cst_1]] : f32
151+
// CHECK-NEXT: %[[val_34:.*]] = math.fma %[[val_33]], %[[cst_10]], %[[cst_23]] : f32
152+
// CHECK-NEXT: %[[val_35:.*]] = math.floor %[[val_34]] : f32
153+
// CHECK-NEXT: %[[val_36:.*]] = arith.cmpf uge, %[[val_35]], %[[cst_0]] : f32
154+
// CHECK-NEXT: %[[val_37:.*]] = arith.select %[[val_36]], %[[val_35]], %[[cst_0]] : f32
155+
// CHECK-NEXT: %[[val_38:.*]] = arith.cmpf ule, %[[val_37]], %[[cst]] : f32
156+
// CHECK-NEXT: %[[val_39:.*]] = arith.select %[[val_38]], %[[val_37]], %[[cst]] : f32
157+
// CHECK-NEXT: %[[val_40:.*]] = math.fma %[[cst_9]], %[[val_39]], %[[val_33]] : f32
158+
// CHECK-NEXT: %[[val_41:.*]] = math.fma %[[cst_8]], %[[val_39]], %[[val_40]] : f32
159+
// CHECK-NEXT: %[[val_42:.*]] = math.fma %[[val_41]], %[[cst_7]], %[[cst_6]] : f32
160+
// CHECK-NEXT: %[[val_43:.*]] = math.fma %[[val_42]], %[[val_41]], %[[cst_5]] : f32
161+
// CHECK-NEXT: %[[val_44:.*]] = math.fma %[[val_43]], %[[val_41]], %[[cst_4]] : f32
162+
// CHECK-NEXT: %[[val_45:.*]] = math.fma %[[val_44]], %[[val_41]], %[[cst_3]] : f32
163+
// CHECK-NEXT: %[[val_46:.*]] = math.fma %[[val_45]], %[[val_41]], %[[cst_23]] : f32
164+
// CHECK-NEXT: %[[val_47:.*]] = arith.mulf %[[val_41]], %[[val_41]] : f32
165+
// CHECK-NEXT: %[[val_48:.*]] = math.fma %[[val_46]], %[[val_47]], %[[val_41]] : f32
166+
// CHECK-NEXT: %[[val_49:.*]] = arith.addf %[[val_48]], %[[cst_22]] : f32
167+
// CHECK-NEXT: %[[val_50:.*]] = arith.fptosi %[[val_39]] : f32 to i32
168+
// CHECK-NEXT: %[[val_51:.*]] = arith.addi %[[val_50]], %[[c127_i32]] : i32
169+
// CHECK-NEXT: %[[val_52:.*]] = arith.shli %[[val_51]], %[[c23_i32]] : i32
170+
// CHECK-NEXT: %[[val_53:.*]] = arith.bitcast %[[val_52]] : i32 to f32
171+
// CHECK-NEXT: %[[val_54:.*]] = arith.mulf %[[val_49]], %[[val_53]] : f32
172+
// CHECK-NEXT: %[[val_55:.*]] = arith.negf %[[val_2]] : f32
173+
// CHECK-NEXT: %[[val_56:.*]] = math.fma %[[val_55]], %[[val_2]], %[[val_28]] : f32
174+
// CHECK-NEXT: %[[val_57:.*]] = arith.mulf %[[val_27]], %[[val_54]] : f32
175+
// CHECK-NEXT: %[[val_58:.*]] = arith.mulf %[[val_57]], %[[val_56]] : f32
176+
// CHECK-NEXT: %[[val_59:.*]] = math.fma %[[val_27]], %[[val_54]], %[[val_58]] : f32
177+
// CHECK-NEXT: %[[val_60:.*]] = arith.cmpf olt, %[[val_2]], %[[cst_27]] : f32
178+
// CHECK-NEXT: %[[val_61:.*]] = arith.xori %[[val_60]], %[[cst_true]] : i1
179+
// CHECK-NEXT: %[[val_62:.*]] = arith.addf %[[val_arg0]], %[[val_arg0]] : f32
180+
// CHECK-NEXT: %[[val_63:.*]] = arith.select %[[val_61]], %[[val_62]], %[[val_59]] : f32
181+
// CHECK-NEXT: %[[val_64:.*]] = arith.cmpf ogt, %[[val_2]], %[[cst_28]] : f32
182+
// CHECK-NEXT: %[[val_65:.*]] = arith.select %[[val_64]], %[[cst_21]], %[[val_63]] : f32
183+
// CHECK-NEXT: %[[val_66:.*]] = arith.cmpf olt, %[[val_arg0]], %[[cst_21]] : f32
184+
// CHECK-NEXT: %[[val_67:.*]] = arith.subf %[[cst_26]], %[[val_65]] : f32
185+
// CHECK-NEXT: %[[val_68:.*]] = arith.select %[[val_66]], %[[val_67]], %[[val_65]] : f32
186+
// CHECK-NEXT: return %[[val_68]] : f32
187+
// CHECK-NEXT: }
190188

191189
func.func @erfc_scalar(%arg0: f32) -> f32 {
192190
%0 = math.erfc %arg0 : f32

0 commit comments

Comments
 (0)