colin
/
mlir-hlo
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[MLIR][KernelGen] Add chlo.erfc lowerings for f64 

PiperOrigin-RevId: 352993223
This commit is contained in:
A. Unique TensorFlower 2021-01-21 04:41:21 -08:00 committed by TensorFlow MLIR Team
parent bec2e625a2
commit 1a37078132
2 changed files with 219 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

56
lib/Dialect/mhlo/transforms/chlo_legalize_to_hlo.cc
View File

@ -96,6 +96,8 @@ Value MaterializePolynomialApproximation(ConversionPatternRewriter &rewriter,
// This implementation is based on Cephes. // This implementation is based on Cephes.
Value MaterializeErfcApproximationF64ForMagnituteGEOne( Value MaterializeErfcApproximationF64ForMagnituteGEOne(
ConversionPatternRewriter &rewriter, Location loc, Value x) { ConversionPatternRewriter &rewriter, Location loc, Value x) {
assert(x.getType().cast<ShapedType>().getElementType().isF64() &&
"expect f64 element type");
const double kMaxlog = 7.09782712893383996843E2; const double kMaxlog = 7.09782712893383996843E2;
const std::vector<double> kErfcPCoefficients{ const std::vector<double> kErfcPCoefficients{
2.46196981473530512524E-10, 5.64189564831068821977E-1, 2.46196981473530512524E-10, 5.64189564831068821977E-1,
@ -177,6 +179,8 @@ Value MaterializeErfcApproximationF64ForMagnituteGEOne(
// This implementation is based on Cephes. // This implementation is based on Cephes.
Value MaterializeErfApproximationF64ForMagnituteLEOne( Value MaterializeErfApproximationF64ForMagnituteLEOne(
ConversionPatternRewriter &rewriter, Location loc, Value x) { ConversionPatternRewriter &rewriter, Location loc, Value x) {
assert(x.getType().cast<ShapedType>().getElementType().isF64() &&
"expect f64 element type");
const std::vector<double> kErfTCoefficients{ const std::vector<double> kErfTCoefficients{
9.60497373987051638749E0, 9.00260197203842689217E1, 9.60497373987051638749E0, 9.00260197203842689217E1,
2.23200534594684319226E3, 7.00332514112805075473E3, 2.23200534594684319226E3, 7.00332514112805075473E3,
@ -200,7 +204,7 @@ Value MaterializeErfApproximationF64ForMagnituteLEOne(
// This implementation is based on Cephes. // This implementation is based on Cephes.
Value MaterializeErfApproximationF64(ConversionPatternRewriter &rewriter, Value MaterializeErfApproximationF64(ConversionPatternRewriter &rewriter,
Location loc, Value x) { Location loc, Value x) {
assert(x.getType().cast<RankedTensorType>().getElementType().isF64() && assert(x.getType().cast<ShapedType>().getElementType().isF64() &&
"expect f64 element type"); "expect f64 element type");
// Rely on erf approximation for |x| < 1 // Rely on erf approximation for |x| < 1
@ -224,10 +228,36 @@ Value MaterializeErfApproximationF64(ConversionPatternRewriter &rewriter,
erfc_based_approx); erfc_based_approx);
} }
Value MaterializeErfcApproximationF64(ConversionPatternRewriter &rewriter,
Location loc, Value x) {
assert(x.getType().cast<ShapedType>().getElementType().isF64() &&
"expect f64 element type");
// Rely on erfc approximation for |x| >= 1
// erfc(x) = erfc_approx(x)
Value erfc_approx =
MaterializeErfcApproximationF64ForMagnituteGEOne(rewriter, loc, x);
// Rely on erf approximation for |x| < 1 and materialize erfc as
// erfc(x) = 1 - erf_approx(x)
Value one = chlo::getConstantLike(rewriter, loc, 1.0, x);
Value erf_approx =
MaterializeErfApproximationF64ForMagnituteLEOne(rewriter, loc, x);
Value erf_based_approx = rewriter.create<mhlo::SubOp>(loc, one, erf_approx);
// Materialize approximation selection based on argument.
Value abs_x = rewriter.create<mhlo::AbsOp>(loc, x);
const StringAttr kLT = rewriter.getStringAttr(
mhlo::stringifyComparisonDirection(mhlo::ComparisonDirection::LT));
Value abs_x_lt_one = rewriter.create<mhlo::CompareOp>(loc, abs_x, one, kLT);
return rewriter.create<mhlo::SelectOp>(loc, abs_x_lt_one, erf_based_approx,
erfc_approx);
}
// This is the same approximation as used in Eigen. // This is the same approximation as used in Eigen.
Value MaterializeErfApproximationF32(ConversionPatternRewriter &rewriter, Value MaterializeErfApproximationF32(ConversionPatternRewriter &rewriter,
Location loc, Value operand) { Location loc, Value operand) {
assert(operand.getType().cast<RankedTensorType>().getElementType().isF32() && assert(operand.getType().cast<ShapedType>().getElementType().isF32() &&
"expect f32 element type"); "expect f32 element type");
const std::vector<float> kAlpha{ const std::vector<float> kAlpha{
-2.72614225801306e-10f, 2.77068142495902e-08f, -2.10102402082508e-06f, -2.72614225801306e-10f, 2.77068142495902e-08f, -2.10102402082508e-06f,
@ -264,7 +294,7 @@ struct ConvertErfOp : public OpConversionPattern<ErfOp> {
Location loc = op.getLoc(); Location loc = op.getLoc();
ErfOp::Adaptor transformed(operands); ErfOp::Adaptor transformed(operands);
Value x = transformed.operand(); Value x = transformed.operand();
Type ty = x.getType().cast<RankedTensorType>().getElementType(); Type ty = x.getType().cast<ShapedType>().getElementType();
// For now, we support only f64, f32, and f16. // For now, we support only f64, f32, and f16.
if (!ty.isF64() && !ty.isF32() && !ty.isF16()) return failure(); if (!ty.isF64() && !ty.isF32() && !ty.isF16()) return failure();
@ -292,6 +322,24 @@ struct ConvertErfOp : public OpConversionPattern<ErfOp> {
} }
}; };
struct ConvertErfcOp : public OpConversionPattern<ErfcOp> {
using OpConversionPattern<ErfcOp>::OpConversionPattern;
LogicalResult matchAndRewrite(
ErfcOp op, ArrayRef<Value> operands,
ConversionPatternRewriter &rewriter) const override {
Location loc = op.getLoc();
ErfcOp::Adaptor transformed(operands);
Value x = transformed.operand();
Type ty = x.getType().cast<ShapedType>().getElementType();
// For now, we support only f64.
if (!ty.isF64()) return failure();
rewriter.replaceOp(op, MaterializeErfcApproximationF64(rewriter, loc, x));
return success();
}
};
// Converts binary ops that statically are determined to not broadcast directly // Converts binary ops that statically are determined to not broadcast directly
// to the corresponding mhlo non-broadcasting op. // to the corresponding mhlo non-broadcasting op.
template <typename ChloOpTy, typename HloOpTy, typename Adaptor> template <typename ChloOpTy, typename HloOpTy, typename Adaptor>
@ -443,7 +491,7 @@ void PopulateLegalizeChloToHloPatterns(MLIRContext *context,
context, patterns, 5); context, patterns, 5);
// Other patterns. // Other patterns.
patterns->insert<ConvertConstantLikeOp, ConvertErfOp>(context); patterns->insert<ConvertConstantLikeOp, ConvertErfOp, ConvertErfcOp>(context);
} }
} // namespace chlo } // namespace chlo

167
tests/chlo_legalize_to_mhlo.mlir
View File

@ -366,3 +366,170 @@ func @acosh(%arg: tensor<f16>) -> tensor<f16> {
%1 = "chlo.acosh"(%arg) : (tensor<f16>) -> tensor<f16> %1 = "chlo.acosh"(%arg) : (tensor<f16>) -> tensor<f16>
return %1 : tensor<f16> return %1 : tensor<f16>
} }
// CHECK-LABEL: @erfc_f64
// CHECK-SAME: %[[ARG:.*]]: tensor<f64>
func @erfc_f64(%arg : tensor<f64>) -> tensor<f64> {
// CHECK-NEXT: %[[TMP_0:.*]] = mhlo.multiply %[[ARG]], %[[ARG]]
// CHECK-NEXT: %[[TMP_1:.*]] = "mhlo.negate"(%[[TMP_0]])
// CHECK-NEXT: %[[TMP_2:.*]] = "mhlo.exponential"(%[[TMP_1]])
// CHECK-NEXT: %[[TMP_3:.*]] = "mhlo.abs"(%[[ARG]])
// CHECK-NEXT: %[[TMP_4:.*]] = mhlo.constant dense<0.000000e+00>
// CHECK-NEXT: %[[TMP_5:.*]] = mhlo.multiply %[[TMP_4]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_6:.*]] = mhlo.constant dense<2.4619698147353052E-10>
// CHECK-NEXT: %[[TMP_7:.*]] = mhlo.add %[[TMP_5]], %[[TMP_6]]
// CHECK-NEXT: %[[TMP_8:.*]] = mhlo.multiply %[[TMP_7]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_9:.*]] = mhlo.constant dense<0.56418956483106886>
// CHECK-NEXT: %[[TMP_10:.*]] = mhlo.add %[[TMP_8]], %[[TMP_9]]
// CHECK-NEXT: %[[TMP_11:.*]] = mhlo.multiply %[[TMP_10]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_12:.*]] = mhlo.constant dense<7.4632105644226989>
// CHECK-NEXT: %[[TMP_13:.*]] = mhlo.add %[[TMP_11]], %[[TMP_12]]
// CHECK-NEXT: %[[TMP_14:.*]] = mhlo.multiply %[[TMP_13]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_15:.*]] = mhlo.constant dense<48.637197098568137>
// CHECK-NEXT: %[[TMP_16:.*]] = mhlo.add %[[TMP_14]], %[[TMP_15]]
// CHECK-NEXT: %[[TMP_17:.*]] = mhlo.multiply %[[TMP_16]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_18:.*]] = mhlo.constant dense<196.5208329560771>
// CHECK-NEXT: %[[TMP_19:.*]] = mhlo.add %[[TMP_17]], %[[TMP_18]]
// CHECK-NEXT: %[[TMP_20:.*]] = mhlo.multiply %[[TMP_19]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_21:.*]] = mhlo.constant dense<526.44519499547732>
// CHECK-NEXT: %[[TMP_22:.*]] = mhlo.add %[[TMP_20]], %[[TMP_21]]
// CHECK-NEXT: %[[TMP_23:.*]] = mhlo.multiply %[[TMP_22]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_24:.*]] = mhlo.constant dense<934.52852717195765>
// CHECK-NEXT: %[[TMP_25:.*]] = mhlo.add %[[TMP_23]], %[[TMP_24]]
// CHECK-NEXT: %[[TMP_26:.*]] = mhlo.multiply %[[TMP_25]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_27:.*]] = mhlo.constant dense<1027.5518868951572>
// CHECK-NEXT: %[[TMP_28:.*]] = mhlo.add %[[TMP_26]], %[[TMP_27]]
// CHECK-NEXT: %[[TMP_29:.*]] = mhlo.multiply %[[TMP_28]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_30:.*]] = mhlo.constant dense<557.53533536939938>
// CHECK-NEXT: %[[TMP_31:.*]] = mhlo.add %[[TMP_29]], %[[TMP_30]]
// CHECK-NEXT: %[[TMP_32:.*]] = mhlo.multiply %[[TMP_2]], %[[TMP_31]]
// CHECK-NEXT: %[[TMP_33:.*]] = mhlo.constant dense<0.000000e+00>
// CHECK-NEXT: %[[TMP_34:.*]] = mhlo.multiply %[[TMP_33]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_35:.*]] = mhlo.constant dense<1.000000e+00>
// CHECK-NEXT: %[[TMP_36:.*]] = mhlo.add %[[TMP_34]], %[[TMP_35]]
// CHECK-NEXT: %[[TMP_37:.*]] = mhlo.multiply %[[TMP_36]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_38:.*]] = mhlo.constant dense<13.228195115474499>
// CHECK-NEXT: %[[TMP_39:.*]] = mhlo.add %[[TMP_37]], %[[TMP_38]]
// CHECK-NEXT: %[[TMP_40:.*]] = mhlo.multiply %[[TMP_39]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_41:.*]] = mhlo.constant dense<86.707214088598973>
// CHECK-NEXT: %[[TMP_42:.*]] = mhlo.add %[[TMP_40]], %[[TMP_41]]
// CHECK-NEXT: %[[TMP_43:.*]] = mhlo.multiply %[[TMP_42]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_44:.*]] = mhlo.constant dense<354.93777888781989>
// CHECK-NEXT: %[[TMP_45:.*]] = mhlo.add %[[TMP_43]], %[[TMP_44]]
// CHECK-NEXT: %[[TMP_46:.*]] = mhlo.multiply %[[TMP_45]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_47:.*]] = mhlo.constant dense<975.70850174320549>
// CHECK-NEXT: %[[TMP_48:.*]] = mhlo.add %[[TMP_46]], %[[TMP_47]]
// CHECK-NEXT: %[[TMP_49:.*]] = mhlo.multiply %[[TMP_48]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_50:.*]] = mhlo.constant dense<1823.9091668790973>
// CHECK-NEXT: %[[TMP_51:.*]] = mhlo.add %[[TMP_49]], %[[TMP_50]]
// CHECK-NEXT: %[[TMP_52:.*]] = mhlo.multiply %[[TMP_51]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_53:.*]] = mhlo.constant dense<2246.3376081871097>
// CHECK-NEXT: %[[TMP_54:.*]] = mhlo.add %[[TMP_52]], %[[TMP_53]]
// CHECK-NEXT: %[[TMP_55:.*]] = mhlo.multiply %[[TMP_54]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_56:.*]] = mhlo.constant dense<1656.6630919416134>
// CHECK-NEXT: %[[TMP_57:.*]] = mhlo.add %[[TMP_55]], %[[TMP_56]]
// CHECK-NEXT: %[[TMP_58:.*]] = mhlo.multiply %[[TMP_57]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_59:.*]] = mhlo.constant dense<557.53534081772773>
// CHECK-NEXT: %[[TMP_60:.*]] = mhlo.add %[[TMP_58]], %[[TMP_59]]
// CHECK-NEXT: %[[TMP_61:.*]] = mhlo.divide %[[TMP_32]], %[[TMP_60]]
// CHECK-NEXT: %[[TMP_62:.*]] = mhlo.constant dense<0.000000e+00>
// CHECK-NEXT: %[[TMP_63:.*]] = mhlo.multiply %[[TMP_62]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_64:.*]] = mhlo.constant dense<0.56418958354775506>
// CHECK-NEXT: %[[TMP_65:.*]] = mhlo.add %[[TMP_63]], %[[TMP_64]]
// CHECK-NEXT: %[[TMP_66:.*]] = mhlo.multiply %[[TMP_65]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_67:.*]] = mhlo.constant dense<1.275366707599781>
// CHECK-NEXT: %[[TMP_68:.*]] = mhlo.add %[[TMP_66]], %[[TMP_67]]
// CHECK-NEXT: %[[TMP_69:.*]] = mhlo.multiply %[[TMP_68]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_70:.*]] = mhlo.constant dense<5.0190504225118051>
// CHECK-NEXT: %[[TMP_71:.*]] = mhlo.add %[[TMP_69]], %[[TMP_70]]
// CHECK-NEXT: %[[TMP_72:.*]] = mhlo.multiply %[[TMP_71]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_73:.*]] = mhlo.constant dense<6.160210979930536>
// CHECK-NEXT: %[[TMP_74:.*]] = mhlo.add %[[TMP_72]], %[[TMP_73]]
// CHECK-NEXT: %[[TMP_75:.*]] = mhlo.multiply %[[TMP_74]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_76:.*]] = mhlo.constant dense<7.4097426995044895>
// CHECK-NEXT: %[[TMP_77:.*]] = mhlo.add %[[TMP_75]], %[[TMP_76]]
// CHECK-NEXT: %[[TMP_78:.*]] = mhlo.multiply %[[TMP_77]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_79:.*]] = mhlo.constant dense<2.9788666537210022>
// CHECK-NEXT: %[[TMP_80:.*]] = mhlo.add %[[TMP_78]], %[[TMP_79]]
// CHECK-NEXT: %[[TMP_81:.*]] = mhlo.multiply %[[TMP_2]], %[[TMP_80]]
// CHECK-NEXT: %[[TMP_82:.*]] = mhlo.constant dense<0.000000e+00>
// CHECK-NEXT: %[[TMP_83:.*]] = mhlo.multiply %[[TMP_82]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_84:.*]] = mhlo.constant dense<1.000000e+00>
// CHECK-NEXT: %[[TMP_85:.*]] = mhlo.add %[[TMP_83]], %[[TMP_84]]
// CHECK-NEXT: %[[TMP_86:.*]] = mhlo.multiply %[[TMP_85]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_87:.*]] = mhlo.constant dense<2.2605286322011726>
// CHECK-NEXT: %[[TMP_88:.*]] = mhlo.add %[[TMP_86]], %[[TMP_87]]
// CHECK-NEXT: %[[TMP_89:.*]] = mhlo.multiply %[[TMP_88]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_90:.*]] = mhlo.constant dense<9.3960352493800147>
// CHECK-NEXT: %[[TMP_91:.*]] = mhlo.add %[[TMP_89]], %[[TMP_90]]
// CHECK-NEXT: %[[TMP_92:.*]] = mhlo.multiply %[[TMP_91]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_93:.*]] = mhlo.constant dense<12.048953980809666>
// CHECK-NEXT: %[[TMP_94:.*]] = mhlo.add %[[TMP_92]], %[[TMP_93]]
// CHECK-NEXT: %[[TMP_95:.*]] = mhlo.multiply %[[TMP_94]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_96:.*]] = mhlo.constant dense<17.081445074756591>
// CHECK-NEXT: %[[TMP_97:.*]] = mhlo.add %[[TMP_95]], %[[TMP_96]]
// CHECK-NEXT: %[[TMP_98:.*]] = mhlo.multiply %[[TMP_97]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_99:.*]] = mhlo.constant dense<9.6089680906328585>
// CHECK-NEXT: %[[TMP_100:.*]] = mhlo.add %[[TMP_98]], %[[TMP_99]]
// CHECK-NEXT: %[[TMP_101:.*]] = mhlo.multiply %[[TMP_100]], %[[TMP_3]]
// CHECK-NEXT: %[[TMP_102:.*]] = mhlo.constant dense<3.3690764510008151>
// CHECK-NEXT: %[[TMP_103:.*]] = mhlo.add %[[TMP_101]], %[[TMP_102]]
// CHECK-NEXT: %[[TMP_104:.*]] = mhlo.divide %[[TMP_81]], %[[TMP_103]]
// CHECK-NEXT: %[[TMP_105:.*]] = mhlo.constant dense<8.000000e+00>
// CHECK-NEXT: %[[TMP_106:.*]] = "mhlo.compare"(%[[TMP_3]], %[[TMP_105]]) {comparison_direction = "LT"}
// CHECK-NEXT: %[[TMP_107:.*]] = "mhlo.select"(%[[TMP_106]], %[[TMP_61]], %[[TMP_104]])
// CHECK-NEXT: %[[TMP_108:.*]] = mhlo.constant dense<-709.78271289338397>
// CHECK-NEXT: %[[TMP_109:.*]] = "mhlo.compare"(%[[TMP_1]], %[[TMP_108]]) {comparison_direction = "LT"}
// CHECK-NEXT: %[[TMP_110:.*]] = mhlo.constant dense<0.000000e+00>
// CHECK-NEXT: %[[TMP_111:.*]] = "mhlo.select"(%[[TMP_109]], %[[TMP_110]], %[[TMP_107]])
// CHECK-NEXT: %[[TMP_113:.*]] = "mhlo.compare"(%[[ARG]], %[[TMP_110]]) {comparison_direction = "LT"}
// CHECK-NEXT: %[[TMP_114:.*]] = mhlo.constant dense<2.000000e+00>
// CHECK-NEXT: %[[TMP_115:.*]] = mhlo.subtract %[[TMP_114]], %[[TMP_111]]
// CHECK-NEXT: %[[TMP_116:.*]] = "mhlo.select"(%[[TMP_113]], %[[TMP_115]], %[[TMP_111]])
// CHECK-NEXT: %[[TMP_117:.*]] = mhlo.constant dense<1.000000e+00>
// CHECK-NEXT: %[[TMP_118:.*]] = mhlo.multiply %[[ARG]], %[[ARG]]
// CHECK-NEXT: %[[TMP_119:.*]] = mhlo.constant dense<0.000000e+00>
// CHECK-NEXT: %[[TMP_120:.*]] = mhlo.multiply %[[TMP_119]], %[[TMP_118]]
// CHECK-NEXT: %[[TMP_121:.*]] = mhlo.constant dense<9.6049737398705161>
// CHECK-NEXT: %[[TMP_122:.*]] = mhlo.add %[[TMP_120]], %[[TMP_121]]
// CHECK-NEXT: %[[TMP_123:.*]] = mhlo.multiply %[[TMP_122]], %[[TMP_118]]
// CHECK-NEXT: %[[TMP_124:.*]] = mhlo.constant dense<90.026019720384269>
// CHECK-NEXT: %[[TMP_125:.*]] = mhlo.add %[[TMP_123]], %[[TMP_124]]
// CHECK-NEXT: %[[TMP_126:.*]] = mhlo.multiply %[[TMP_125]], %[[TMP_118]]
// CHECK-NEXT: %[[TMP_127:.*]] = mhlo.constant dense<2232.0053459468431>
// CHECK-NEXT: %[[TMP_128:.*]] = mhlo.add %[[TMP_126]], %[[TMP_127]]
// CHECK-NEXT: %[[TMP_129:.*]] = mhlo.multiply %[[TMP_128]], %[[TMP_118]]
// CHECK-NEXT: %[[TMP_130:.*]] = mhlo.constant dense<7003.3251411280507>
// CHECK-NEXT: %[[TMP_131:.*]] = mhlo.add %[[TMP_129]], %[[TMP_130]]
// CHECK-NEXT: %[[TMP_132:.*]] = mhlo.multiply %[[TMP_131]], %[[TMP_118]]
// CHECK-NEXT: %[[TMP_133:.*]] = mhlo.constant dense<55592.301301039493>
// CHECK-NEXT: %[[TMP_134:.*]] = mhlo.add %[[TMP_132]], %[[TMP_133]]
// CHECK-NEXT: %[[TMP_135:.*]] = mhlo.multiply %[[ARG]], %[[TMP_134]]
// CHECK-NEXT: %[[TMP_136:.*]] = mhlo.constant dense<0.000000e+00>
// CHECK-NEXT: %[[TMP_137:.*]] = mhlo.multiply %[[TMP_136]], %[[TMP_118]]
// CHECK-NEXT: %[[TMP_138:.*]] = mhlo.constant dense<1.000000e+00>
// CHECK-NEXT: %[[TMP_139:.*]] = mhlo.add %[[TMP_137]], %[[TMP_138]]
// CHECK-NEXT: %[[TMP_140:.*]] = mhlo.multiply %[[TMP_139]], %[[TMP_118]]
// CHECK-NEXT: %[[TMP_141:.*]] = mhlo.constant dense<33.561714164750313>
// CHECK-NEXT: %[[TMP_142:.*]] = mhlo.add %[[TMP_140]], %[[TMP_141]]
// CHECK-NEXT: %[[TMP_143:.*]] = mhlo.multiply %[[TMP_142]], %[[TMP_118]]
// CHECK-NEXT: %[[TMP_144:.*]] = mhlo.constant dense<521.35794978015269>
// CHECK-NEXT: %[[TMP_145:.*]] = mhlo.add %[[TMP_143]], %[[TMP_144]]
// CHECK-NEXT: %[[TMP_146:.*]] = mhlo.multiply %[[TMP_145]], %[[TMP_118]]
// CHECK-NEXT: %[[TMP_147:.*]] = mhlo.constant dense<4594.3238297098014>
// CHECK-NEXT: %[[TMP_148:.*]] = mhlo.add %[[TMP_146]], %[[TMP_147]]
// CHECK-NEXT: %[[TMP_149:.*]] = mhlo.multiply %[[TMP_148]], %[[TMP_118]]
// CHECK-NEXT: %[[TMP_150:.*]] = mhlo.constant dense<22629.000061389095>
// CHECK-NEXT: %[[TMP_151:.*]] = mhlo.add %[[TMP_149]], %[[TMP_150]]
// CHECK-NEXT: %[[TMP_152:.*]] = mhlo.multiply %[[TMP_151]], %[[TMP_118]]
// CHECK-NEXT: %[[TMP_153:.*]] = mhlo.constant dense<49267.394260863592>
// CHECK-NEXT: %[[TMP_154:.*]] = mhlo.add %[[TMP_152]], %[[TMP_153]]
// CHECK-NEXT: %[[TMP_155:.*]] = mhlo.divide %[[TMP_135]], %[[TMP_154]]
// CHECK-NEXT: %[[TMP_156:.*]] = mhlo.subtract %[[TMP_117]], %[[TMP_155]]
// CHECK-NEXT: %[[TMP_157:.*]] = "mhlo.abs"(%[[ARG]])
// CHECK-NEXT: %[[TMP_159:.*]] = "mhlo.compare"(%[[TMP_157]], %[[TMP_117]]) {comparison_direction = "LT"}
// CHECK-NEXT: %[[RESULT:.*]] = "mhlo.select"(%[[TMP_159]], %[[TMP_156]], %[[TMP_116]])
// CHECK-NEXT: return %[[RESULT]]
%1 = "chlo.erfc"(%arg) : (tensor<f64>) -> tensor<f64>
return %1 : tensor<f64>
}