Skip to content

Commit 176bfa8

Browse files
ita9naiwaita9naiwa
committed
Update ExpandPatterns.cpp
1 parent 2696e4f commit 176bfa8

File tree

2 files changed

+59
-63
lines changed

2 files changed

+59
-63
lines changed

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

Lines changed: 17 additions & 17 deletions
Original file line numberDiff line numberDiff line change
@@ -311,40 +311,40 @@ static LogicalResult convertFPowIOp(math::FPowIOp op,
311311
return success();
312312
}
313313

314-
// Converts Powf(float a, float b) (meaning a^b) to exp^(b * ln(a))
314+
// Converts Powf(float a, float b) (meaning a^b) to exp^(b * ln(|a|))
315+
// * sign(a)^b
315316
static LogicalResult convertPowfOp(math::PowFOp op, PatternRewriter &rewriter) {
316317
ImplicitLocOpBuilder b(op->getLoc(), rewriter);
317318
Value operandA = op.getOperand(0);
318319
Value operandB = op.getOperand(1);
319320
Type opType = operandA.getType();
320321
Value zero = createFloatConst(op->getLoc(), opType, 0.00, rewriter);
321322
Value one = createFloatConst(op->getLoc(), opType, 1.00, rewriter);
322-
Value two = createFloatConst(op->getLoc(), opType, 2.00, rewriter);
323323
Value negOne = createFloatConst(op->getLoc(), opType, -1.00, rewriter);
324-
Value opASquared = b.create<arith::MulFOp>(opType, operandA, operandA);
325-
Value opBHalf = b.create<arith::DivFOp>(opType, operandB, two);
324+
Value two = createFloatConst(op->getLoc(), opType, 2.00, rewriter);
326325

327-
Value logA = b.create<math::LogOp>(opType, opASquared);
328-
Value mult = b.create<arith::MulFOp>(opType, opBHalf, logA);
326+
Value absA = b.create<math::AbsFOp>(opType, operandA);
327+
Value isNegative =
328+
b.create<arith::CmpFOp>(arith::CmpFPredicate::OLT, operandA, zero);
329+
Value signA =
330+
b.create<arith::SelectOp>(op->getLoc(), isNegative, negOne, one);
331+
Value logA = b.create<math::LogOp>(opType, absA);
332+
Value mult = b.create<arith::MulFOp>(opType, operandB, logA);
329333
Value expResult = b.create<math::ExpOp>(opType, mult);
330-
Value negExpResult = b.create<arith::MulFOp>(opType, expResult, negOne);
331334
Value remainder = b.create<arith::RemFOp>(opType, operandB, two);
332-
Value negCheck =
333-
b.create<arith::CmpFOp>(arith::CmpFPredicate::OLT, operandA, zero);
334-
Value oddPower =
335+
Value isOdd =
335336
b.create<arith::CmpFOp>(arith::CmpFPredicate::ONE, remainder, zero);
336-
Value oddAndNeg = b.create<arith::AndIOp>(op->getLoc(), oddPower, negCheck);
337+
Value signedExpResult = b.create<arith::SelectOp>(
338+
op->getLoc(), isOdd, b.create<arith::MulFOp>(opType, expResult, signA),
339+
expResult);
337340

338-
// First, we select between the exp value and the adjusted value for odd
339-
// powers of negatives. Then, we ensure that one is produced if `b` is zero.
340341
// This corresponds to `libm` behavior, even for `0^0`. Without this check,
341342
// `exp(0 * ln(0)) = exp(0 *-inf) = exp(-nan) = -nan`.
342343
Value zeroCheck =
343344
b.create<arith::CmpFOp>(arith::CmpFPredicate::OEQ, operandB, zero);
344-
Value res = b.create<arith::SelectOp>(op->getLoc(), oddAndNeg, negExpResult,
345-
expResult);
346-
res = b.create<arith::SelectOp>(op->getLoc(), zeroCheck, one, res);
347-
rewriter.replaceOp(op, res);
345+
Value finalResult =
346+
b.create<arith::SelectOp>(op->getLoc(), zeroCheck, one, signedExpResult);
347+
rewriter.replaceOp(op, finalResult);
348348
return success();
349349
}
350350

‎mlir/test/Dialect/Math/expand-math.mlir

Lines changed: 42 additions & 46 deletions
Original file line numberDiff line numberDiff line change
@@ -202,24 +202,23 @@ func.func @roundf_func(%a: f32) -> f32 {
202202

203203
// CHECK-LABEL: func @powf_func
204204
// CHECK-SAME: ([[ARG0:%.+]]: f64, [[ARG1:%.+]]: f64)
205-
func.func @powf_func(%a: f64, %b: f64) ->f64 {
205+
func.func @powf_func(%a: f64, %b: f64) -> f64 {
206206
// CHECK-DAG: [[CST0:%.+]] = arith.constant 0.000000e+00
207207
// CHECK-DAG: [[CST1:%.+]] = arith.constant 1.0
208-
// CHECK-DAG: [[TWO:%.+]] = arith.constant 2.000000e+00
209-
// CHECK-DAG: [[NEGONE:%.+]] = arith.constant -1.000000e+00
210-
// CHECK-DAG: [[SQR:%.+]] = arith.mulf [[ARG0]], [[ARG0]]
211-
// CHECK-DAG: [[HALF:%.+]] = arith.divf [[ARG1]], [[TWO]]
212-
// CHECK-DAG: [[LOG:%.+]] = math.log [[SQR]]
213-
// CHECK-DAG: [[MULT:%.+]] = arith.mulf [[HALF]], [[LOG]]
214-
// CHECK-DAG: [[EXPR:%.+]] = math.exp [[MULT]]
215-
// CHECK-DAG: [[NEGEXPR:%.+]] = arith.mulf [[EXPR]], [[NEGONE]]
216-
// CHECK-DAG: [[REMF:%.+]] = arith.remf [[ARG1]], [[TWO]]
217-
// CHECK-DAG: [[CMPNEG:%.+]] = arith.cmpf olt, [[ARG0]]
218-
// CHECK-DAG: [[CMPZERO:%.+]] = arith.cmpf one, [[REMF]]
219-
// CHECK-DAG: [[AND:%.+]] = arith.andi [[CMPZERO]], [[CMPNEG]]
220-
// CHECK-DAG: [[CMPZERO:%.+]] = arith.cmpf oeq, [[ARG1]], [[CST0]]
221-
// CHECK-DAG: [[SEL:%.+]] = arith.select [[AND]], [[NEGEXPR]], [[EXPR]]
222-
// CHECK-DAG: [[SEL1:%.+]] = arith.select [[CMPZERO]], [[CST1]], [[SEL]]
208+
// CHECK-DAG: [[CSTNEG1:%.+]] = arith.constant -1.000000e+00
209+
// CHECK-DAG: [[CSTTWO:%.+]] = arith.constant 2.000000e+00
210+
// CHECK: [[ABSA:%.+]] = math.absf [[ARG0]]
211+
// CHECK: [[ISNEG:%.+]] = arith.cmpf olt, [[ARG0]], [[CST0]]
212+
// CHECK: [[SIGNA:%.+]] = arith.select [[ISNEG]], [[CSTNEG1]], [[CST1]]
213+
// CHECK: [[LOGA:%.+]] = math.log [[ABSA]]
214+
// CHECK: [[MULB:%.+]] = arith.mulf [[ARG1]], [[LOGA]]
215+
// CHECK: [[EXP:%.+]] = math.exp [[MULB]]
216+
// CHECK: [[REM:%.+]] = arith.remf [[ARG1]], [[CSTTWO]]
217+
// CHECK: [[CMPF:%.+]] = arith.cmpf one, [[REM]], [[CST0]]
218+
// CHECK: [[ABMUL:%.+]] = arith.mulf [[EXP]], [[SIGNA]]
219+
// CHECK: [[SEL0:%.+]] = arith.select [[CMPF]], [[ABMUL]], [[EXP]]
220+
// CHECK: [[CMPF2:%.+]] = arith.cmpf oeq, [[ARG1]], [[CST0]]
221+
// CHECK: [[SEL1:%.+]] = arith.select [[CMPF2]], [[CST1]], [[SEL0]]
223222
// CHECK: return [[SEL1]]
224223
%ret = math.powf %a, %b : f64
225224
return %ret : f64
@@ -602,26 +601,24 @@ func.func @math_fpowi_to_powf_tensor(%0 : tensor<8xf32>, %1: tensor<8xi32>) -> t
602601
return %2 : tensor<8xf32>
603602
}
604603
// CHECK-SAME: (%[[ARG0:.*]]: tensor<8xf32>, %[[ARG1:.*]]: tensor<8xi32>) -> tensor<8xf32> {
605-
// CHECK-DAG: %[[CSTNEG1:.*]] = arith.constant dense<-1.000000e+00> : tensor<8xf32>
606604
// CHECK-DAG: %[[CST2:.*]] = arith.constant dense<2.000000e+00> : tensor<8xf32>
607-
// CHECK-DAG: %[[CST0:.*]] = arith.constant dense<0.000000e+00> : tensor<8xf32>
605+
// CHECK-DAG: %[[CSTNEG1:.*]] = arith.constant dense<-1.000000e+00> : tensor<8xf32>
608606
// CHECK-DAG: %[[CST1:.+]] = arith.constant dense<1.000000e+00> : tensor<8xf32>
609-
// CHECK: %[[TOFP:.*]] = arith.sitofp %[[ARG1]] : tensor<8xi32> to tensor<8xf32>
610-
// CHECK: %[[SQ:.*]] = arith.mulf %[[ARG0]], %[[ARG0]] : tensor<8xf32>
611-
// CHECK: %[[DIV:.*]] = arith.divf %[[TOFP]], %[[CST2]] : tensor<8xf32>
612-
// CHECK: %[[LG:.*]] = math.log %[[SQ]] : tensor<8xf32>
613-
// CHECK: %[[MUL:.*]] = arith.mulf %[[DIV]], %[[LG]] : tensor<8xf32>
614-
// CHECK: %[[EXP:.*]] = math.exp %[[MUL]] : tensor<8xf32>
615-
// CHECK: %[[MUL1:.*]] = arith.mulf %[[EXP]], %[[CSTNEG1]] : tensor<8xf32>
616-
// CHECK: %[[REM:.*]] = arith.remf %[[TOFP]], %[[CST2]] : tensor<8xf32>
617-
// CHECK: %[[CMPF:.*]] = arith.cmpf olt, %[[ARG0]], %[[CST0]] : tensor<8xf32>
618-
// CHECK: %[[CMPF1:.*]] = arith.cmpf one, %[[REM]], %[[CST0]] : tensor<8xf32>
619-
// CHECK: %[[AND:.*]] = arith.andi %[[CMPF1]], %[[CMPF]] : tensor<8xi1>
620-
// CHECK: %[[CMPZERO:.*]] = arith.cmpf oeq, %[[TOFP]], %[[CST0]]
621-
// CHECK: %[[SEL:.*]] = arith.select %[[AND]], %[[MUL1]], %[[EXP]] : tensor<8xi1>, tensor<8xf32>
622-
// CHECK: %[[SEL1:.+]] = arith.select %[[CMPZERO]], %[[CST1]], %[[SEL]]
623-
// CHECK: return %[[SEL1]] : tensor<8xf32>
624-
607+
// CHECK-DAG: %[[CST0:.*]] = arith.constant dense<0.000000e+00> : tensor<8xf32>
608+
// CHECK: %[[TOFP:.*]] = arith.sitofp %[[ARG1]] : tensor<8xi32> to tensor<8xf32>
609+
// CHECK: %[[ABSA:.*]] = math.absf %[[ARG0]] : tensor<8xf32>
610+
// CHECK: %[[ISNEG:.*]] = arith.cmpf olt, %[[ARG0]], %[[CST0]] : tensor<8xf32>
611+
// CHECK: %[[SIGNA:.*]] = arith.select %[[ISNEG]], %[[CSTNEG1]], %[[CST1]] : tensor<8xi1>, tensor<8xf32>
612+
// CHECK: %[[LOGA:.*]] = math.log %[[ABSA]] : tensor<8xf32>
613+
// CHECK: %[[MULA:.*]] = arith.mulf %[[TOFP]], %[[LOGA]] : tensor<8xf32>
614+
// CHECK: %[[EXPA:.*]] = math.exp %[[MULA]] : tensor<8xf32>
615+
// CHECK: %[[REM:.*]] = arith.remf %[[TOFP]], %[[CST2]] : tensor<8xf32>
616+
// CHECK: %[[CMPF:.*]] = arith.cmpf one, %[[REM]], %[[CST0]] : tensor<8xf32>
617+
// CHECK: %[[ABMUL:.*]] = arith.mulf %[[EXPA]], %[[SIGNA]] : tensor<8xf32>
618+
// CHECK: %[[SEL0:.*]] = arith.select %[[CMPF]], %[[ABMUL]], %[[EXPA]] : tensor<8xi1>, tensor<8xf32>
619+
// CHECK: %[[CMPZERO:.*]] = arith.cmpf oeq, %[[TOFP]], %[[CST0]] : tensor<8xf32>
620+
// CHECK: %[[SEL1:.*]] = arith.select %[[CMPZERO]], %[[CST1]], %[[SEL0]] : tensor<8xi1>, tensor<8xf32>
621+
// CHECK: return %[[SEL1]]
625622
// -----
626623

627624
// CHECK-LABEL: func.func @math_fpowi_to_powf_scalar
@@ -635,19 +632,18 @@ func.func @math_fpowi_to_powf_scalar(%0 : f32, %1: i64) -> f32 {
635632
// CHECK-DAG: %[[CST0:.*]] = arith.constant 0.000000e+00 : f32
636633
// CHECK-DAG: %[[CST1:.+]] = arith.constant 1.000000e+00 : f32
637634
// CHECK: %[[TOFP:.*]] = arith.sitofp %[[ARG1]] : i64 to f32
638-
// CHECK: %[[SQ:.*]] = arith.mulf %[[ARG0]], %[[ARG0]] : f32
639-
// CHECK: %[[DIV:.*]] = arith.divf %[[TOFP]], %[[CST2]] : f32
640-
// CHECK: %[[LG:.*]] = math.log %[[SQ]] : f32
641-
// CHECK: %[[MUL:.*]] = arith.mulf %[[DIV]], %[[LG]] : f32
642-
// CHECK: %[[EXP:.*]] = math.exp %[[MUL]] : f32
643-
// CHECK: %[[MUL1:.*]] = arith.mulf %[[EXP]], %[[CSTNEG1]] : f32
635+
// CHECK: %[[ABSA:.*]] = math.absf %[[ARG0]] : f32
636+
// CHECK: %[[ISNEG:.*]] = arith.cmpf olt, %[[ARG0]], %[[CST0]] : f32
637+
// CHECK: %[[SIGNA:.*]] = arith.select %[[ISNEG]], %[[CSTNEG1]], %[[CST1]] : f32
638+
// CHECK: %[[LOGA:.*]] = math.log %[[ABSA]] : f32
639+
// CHECK: %[[MULA:.*]] = arith.mulf %[[TOFP]], %[[LOGA]] : f32
640+
// CHECK: %[[EXPA:.*]] = math.exp %[[MULA]] : f32
644641
// CHECK: %[[REM:.*]] = arith.remf %[[TOFP]], %[[CST2]] : f32
645-
// CHECK: %[[CMPF:.*]] = arith.cmpf olt, %[[ARG0]], %[[CST0]] : f32
646-
// CHECK: %[[CMPF1:.*]] = arith.cmpf one, %[[REM]], %[[CST0]] : f32
647-
// CHECK: %[[AND:.*]] = arith.andi %[[CMPF1]], %[[CMPF]] : i1
648-
// CHECK: %[[CMPZERO:.*]] = arith.cmpf oeq, %[[TOFP]], %[[CST0]]
649-
// CHECK: %[[SEL:.*]] = arith.select %[[AND]], %[[MUL1]], %[[EXP]] : f32
650-
// CHECK: %[[SEL1:.+]] = arith.select %[[CMPZERO]], %[[CST1]], %[[SEL]]
642+
// CHECK: %[[CMPF:.*]] = arith.cmpf one, %[[REM]], %[[CST0]] : f32
643+
// CHECK: %[[ABMUL:.*]] = arith.mulf %[[EXPA]], %[[SIGNA]] : f32
644+
// CHECK: %[[SEL0:.*]] = arith.select %[[CMPF]], %[[ABMUL]], %[[EXPA]] : f32
645+
// CHECK: %[[CMPZERO:.*]] = arith.cmpf oeq, %[[TOFP]], %[[CST0]] : f32
646+
// CHECK: %[[SEL1:.*]] = arith.select %[[CMPZERO]], %[[CST1]], %[[SEL0]] : f32
651647
// CHECK: return %[[SEL1]] : f32
652648

653649
// -----

0 commit comments

Comments
 (0)