• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 #include "tensorflow/compiler/xla/client/lib/math.h"
17 
18 #include <cmath>
19 
20 #include "tensorflow/compiler/xla/client/lib/arithmetic.h"
21 #include "tensorflow/compiler/xla/client/lib/constants.h"
22 #include "tensorflow/compiler/xla/client/lib/loops.h"
23 #include "tensorflow/compiler/xla/client/xla_builder.h"
24 #include "tensorflow/compiler/xla/primitive_util.h"
25 #include "tensorflow/compiler/xla/shape_util.h"
26 #include "tensorflow/compiler/xla/status_macros.h"
27 
28 namespace xla {
29 namespace {
30 
31 // Evaluate the polynomial given `x` and coefficients in decreasing order.
32 template <typename FP>
EvaluatePolynomial(XlaOp x,absl::Span<const FP> coefficients)33 XlaOp EvaluatePolynomial(XlaOp x, absl::Span<const FP> coefficients) {
34   static_assert(std::is_floating_point<FP>::value,
35                 "Template-argument 'FP' must be a floating-point type");
36   XlaOp poly = ScalarLike(x, 0.0);
37   for (FP c : coefficients) {
38     poly = poly * x + ScalarLike(x, c);
39   }
40   return poly;
41 }
42 
43 // Evaluate the chebyshev polynomial given `x` and coefficients in decreasing
44 // order.
45 template <typename FP>
EvaluateChebyshevPolynomial(XlaOp x,absl::Span<const FP> coefficients)46 XlaOp EvaluateChebyshevPolynomial(XlaOp x, absl::Span<const FP> coefficients) {
47   static_assert(std::is_floating_point<FP>::value,
48                 "Template-argument 'FP' must be a floating-point type");
49   XlaOp b0 = ScalarLike(x, 0.0);
50   XlaOp b1 = ScalarLike(x, 0.0);
51   XlaOp b2 = ScalarLike(x, 0.0);
52   for (FP c : coefficients) {
53     b2 = b1;
54     b1 = b0;
55     b0 = x * b1 - b2 + ScalarLike(x, c);
56   }
57   return ScalarLike(x, 0.5) * (b0 - b2);
58 }
59 
60 }  // namespace
61 
62 // Returns operation(operand), except if `operand` is one of the types in
63 // upcast_types, in which case first converts it to F32, and then converts the
64 // result down to the original type.
DoWithUpcastToF32(XlaOp operand,absl::Span<const PrimitiveType> upcast_types,const std::function<XlaOp (XlaOp)> & operation)65 static XlaOp DoWithUpcastToF32(XlaOp operand,
66                                absl::Span<const PrimitiveType> upcast_types,
67                                const std::function<XlaOp(XlaOp)>& operation) {
68   auto& b = *operand.builder();
69   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
70     TF_ASSIGN_OR_RETURN(auto shape, b.GetShape(operand));
71     PrimitiveType elem_ty = shape.element_type();
72     bool needs_upcast = absl::c_linear_search(upcast_types, elem_ty);
73 
74     if (needs_upcast) {
75       operand = ConvertElementType(operand, F32);
76     }
77     XlaOp result = operation(operand);
78     if (needs_upcast) {
79       result = ConvertElementType(result, elem_ty);
80     }
81     return result;
82   });
83 }
84 
85 // TODO(jlebar): Use this function in more places in this file to restrict the
86 // domain of other functions.
EnsureOperandIsRealFp(absl::string_view op_name,XlaOp operand)87 static Status EnsureOperandIsRealFp(absl::string_view op_name, XlaOp operand) {
88   auto& b = *operand.builder();
89   TF_ASSIGN_OR_RETURN(auto shape, b.GetShape(operand));
90   auto elem_ty = shape.element_type();
91   if (!primitive_util::IsFloatingPointType(elem_ty)) {
92     return InvalidArgument(
93         "Operands to %s must be real-valued floating-point, but got %s",
94         op_name, PrimitiveType_Name(elem_ty));
95   }
96   return Status::OK();
97 }
98 
IsPosInf(XlaOp operand)99 XlaOp IsPosInf(XlaOp operand) {
100   auto& b = *operand.builder();
101   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
102     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("IsPosInf", operand));
103     TF_ASSIGN_OR_RETURN(auto shape, b.GetShape(operand));
104     // Note that this is only correct for floating-point types.  If we wanted it
105     // to be correct for all types, we'd need to Gt(MaxFiniteValue).
106     return Eq(operand, MaxValue(&b, shape.element_type()));
107   });
108 }
109 
IsNegInf(XlaOp operand)110 XlaOp IsNegInf(XlaOp operand) {
111   auto& b = *operand.builder();
112   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
113     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("IsNegInf", operand));
114     TF_ASSIGN_OR_RETURN(auto shape, b.GetShape(operand));
115     // Note that this is only correct for floating-point types.  If we wanted it
116     // to be correct for all types, we'd need to Lt(MinFiniteValue).
117     return Eq(operand, MinValue(&b, shape.element_type()));
118   });
119 }
120 
IsInf(XlaOp operand)121 XlaOp IsInf(XlaOp operand) {
122   auto& b = *operand.builder();
123   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
124     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("IsInf", operand));
125     return IsPosInf(Abs(operand));
126   });
127 }
128 
IsNan(XlaOp operand)129 XlaOp IsNan(XlaOp operand) {
130   auto& b = *operand.builder();
131   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
132     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("IsNan", operand));
133     return Ne(operand, operand);
134   });
135 }
136 
IsNegZero(XlaOp operand)137 XlaOp IsNegZero(XlaOp operand) {
138   auto& b = *operand.builder();
139   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
140     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("IsNegZero", operand));
141     TF_ASSIGN_OR_RETURN(auto shape, b.GetShape(operand));
142 
143     // The bitwise representation of -0 in bfloat16 and IEEE 754 is 0x80...0
144     // (sign bit on, all other bits off).
145     switch (shape.element_type()) {
146       case F64:
147         return Eq(BitcastConvertType(operand, U64),
148                   ConstantR0WithType(&b, U64, uint64{1} << 63));
149       case F32:
150         return Eq(BitcastConvertType(operand, U32),
151                   ConstantR0WithType(&b, U32, uint32{1} << 31));
152       case F16:
153       case BF16:
154         // Not all XLA backends handle U16 well, so we convert to F32/U32.
155         // TODO(jlebar): It would be nice if we could stay in (B)F16/U16 for
156         // backends that *do* support it.
157         return Eq(BitcastConvertType(ConvertElementType(operand, F32), U32),
158                   ConstantR0WithType(&b, U32, uint32{1} << 31));
159       default:
160         LOG(FATAL) << "Expected real fp type.";
161     }
162   });
163 }
164 
Square(XlaOp operand)165 XlaOp Square(XlaOp operand) { return operand * operand; }
166 
Reciprocal(XlaOp operand)167 XlaOp Reciprocal(XlaOp operand) { return ScalarLike(operand, 1.0) / operand; }
168 
169 // Computes an approximation of the error function complement (1 - erf(x)).
170 //
171 // Precondition: abs(x) >= 1.  Otherwise, use ErfImpl.
172 //
173 // This follows Cephes's f32 implementation of erfc.
ErfcImpl32(XlaOp x)174 static XlaOp ErfcImpl32(XlaOp x) {
175   // Coefficients for erfc(f32), from Cephes.
176   const double kMaxlog = 88.72283905206835;
177   // erfc(x) = exp(-x^2) P(1/x^2), 1 < x < 2
178   static const std::array<float, 9> kErfcPCoefficient{
179       +2.326819970068386E-2, -1.387039388740657E-1, +3.687424674597105E-1,
180       -5.824733027278666E-1, +6.210004621745983E-1, -4.944515323274145E-1,
181       +3.404879937665872E-1, -2.741127028184656E-1, +5.638259427386472E-1,
182   };
183   // erfc(x) = exp(-x^2) R(1/x^2), 2 <= x < kMaxlog
184   static const std::array<float, 8> kErfcRCoefficient{
185       -1.047766399936249E+1, +1.297719955372516E+1, -7.495518717768503E+0,
186       +2.921019019210786E+0, -1.015265279202700E+0, +4.218463358204948E-1,
187       -2.820767439740514E-1, +5.641895067754075E-1,
188   };
189   XlaOp abs_x = Abs(x);
190   XlaOp z = Exp(-x * x);
191   XlaOp q = ScalarLike(x, 1) / abs_x;
192   XlaOp y = q * q;
193   XlaOp p = Select(Lt(abs_x, ScalarLike(x, 2.0)),
194                    EvaluatePolynomial<float>(y, kErfcPCoefficient),
195                    EvaluatePolynomial<float>(y, kErfcRCoefficient));
196   y = z * q * p;
197   XlaOp y_clamp = Select(Lt(z, ScalarLike(x, -kMaxlog)), ScalarLike(x, 0), y);
198   return Select(Lt(x, ScalarLike(x, 0)), ScalarLike(x, 2.0) - y_clamp, y_clamp);
199 }
200 
201 // Compute a polynomial approximation of the error function.
202 //
203 // Precondition: abs(x) <= 1.  Otherwise, use ErfcImpl.
204 //
205 // This follows Cephes's f32 implementation of erf.
ErfImpl32Cephes(XlaOp x)206 static XlaOp ErfImpl32Cephes(XlaOp x) {
207   // Coefficients for by erf(f32), from Cephes.
208   //
209   // erf(x) = x P(x^2), 0 < x < 1
210   static const std::array<float, 7> kErfTCoefficient{
211       +7.853861353153693E-5, -8.010193625184903E-4, +5.188327685732524E-3,
212       -2.685381193529856E-2, +1.128358514861418E-1, -3.761262582423300E-1,
213       +1.128379165726710E+0,
214   };
215   return x * EvaluatePolynomial<float>(x * x, kErfTCoefficient);
216 }
217 
ErfcImpl64(XlaOp x)218 static XlaOp ErfcImpl64(XlaOp x) {
219   // Coefficients for erfc(f64), from Cephes.
220   const double kMaxlog = 7.09782712893383996843E2;
221   // erfc(x) = exp(-x^2) P(|x|) / Q(|x|), 1 < x < 8
222   static const std::array<double, 9> kErfcPCoefficient{
223       2.46196981473530512524E-10, 5.64189564831068821977E-1,
224       7.46321056442269912687E0,   4.86371970985681366614E1,
225       1.96520832956077098242E2,   5.26445194995477358631E2,
226       9.34528527171957607540E2,   1.02755188689515710272E3,
227       5.57535335369399327526E2};
228   static const std::array<double, 9> kErfcQCoefficient{
229       1.00000000000000000000E0, 1.32281951154744992508E1,
230       8.67072140885989742329E1, 3.54937778887819891062E2,
231       9.75708501743205489753E2, 1.82390916687909736289E3,
232       2.24633760818710981792E3, 1.65666309194161350182E3,
233       5.57535340817727675546E2};
234 
235   // erfc(x) = exp(-x^2) R(|x|) / S(|x|), 8 <= x < kMaxlog
236   static const std::array<double, 6> kErfcRCoefficient{
237       5.64189583547755073984E-1, 1.27536670759978104416E0,
238       5.01905042251180477414E0,  6.16021097993053585195E0,
239       7.40974269950448939160E0,  2.97886665372100240670E0};
240   static const std::array<double, 7> kErfcSCoefficient{
241       1.00000000000000000000E0, 2.26052863220117276590E0,
242       9.39603524938001434673E0, 1.20489539808096656605E1,
243       1.70814450747565897222E1, 9.60896809063285878198E0,
244       3.36907645100081516050E0};
245 
246   XlaOp z = -x * x;
247   XlaOp abs_x = Abs(x);
248   XlaOp y =
249       Select(Lt(abs_x, ScalarLike(x, 8.0)),
250              Exp(z) * EvaluatePolynomial<double>(abs_x, kErfcPCoefficient) /
251                  EvaluatePolynomial<double>(abs_x, kErfcQCoefficient),
252              Exp(z) * EvaluatePolynomial<double>(abs_x, kErfcRCoefficient) /
253                  EvaluatePolynomial<double>(abs_x, kErfcSCoefficient));
254   XlaOp y_clamp = Select(Lt(z, ScalarLike(x, -kMaxlog)), ScalarLike(x, 0), y);
255   return Select(Lt(x, ScalarLike(x, 0)), ScalarLike(x, 2.0) - y_clamp, y_clamp);
256 }
257 
258 // Compute a polynomial approximation of the error function.
259 //
260 // Precondition: abs(x) <= 1.  Otherwise, use ErfcImpl.
ErfImpl64(XlaOp x)261 static XlaOp ErfImpl64(XlaOp x) {
262   // Coefficients for by erf(f64), from Cephes.
263   //
264   // erf(x) = x T(x^2) / U(x^2), 0 < x < 1
265   static std::array<double, 5> kErfTCoefficient{
266       9.60497373987051638749E0, 9.00260197203842689217E1,
267       2.23200534594684319226E3, 7.00332514112805075473E3,
268       5.55923013010394962768E4};
269   static std::array<double, 6> kErfUCoefficient{
270       1.00000000000000000000E0, 3.35617141647503099647E1,
271       5.21357949780152679795E2, 4.59432382970980127987E3,
272       2.26290000613890934246E4, 4.92673942608635921086E4};
273   XlaOp z = x * x;
274   return x * EvaluatePolynomial<double>(z, kErfTCoefficient) /
275          EvaluatePolynomial<double>(z, kErfUCoefficient);
276 }
277 
Erfc(XlaOp x)278 XlaOp Erfc(XlaOp x) {
279   auto& b = *x.builder();
280   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
281     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("Erfc", x));
282     TF_ASSIGN_OR_RETURN(auto shape, b.GetShape(x));
283     // erfc(x) =
284     //   erfc_impl(x)           if x > 1
285     //   1 - erf_impl(x)        otherwise
286     if (shape.element_type() == F64) {
287       return Select(Gt(Abs(x), ScalarLike(x, 1)), ErfcImpl64(x),
288                     ScalarLike(x, 1) - ErfImpl64(x));
289     }
290     // Erf(c)Impl don't have enough precision when run with bf16 intermediates
291     // (not surprising!), so upcast to f32 in this case.
292     return DoWithUpcastToF32(x, {BF16, F16}, [](XlaOp x) {
293       return Select(Gt(Abs(x), ScalarLike(x, 1)), ErfcImpl32(x),
294                     ScalarLike(x, 1) - ErfImpl32Cephes(x));
295     });
296   });
297 }
298 
299 // Compute a polynomial approximation of the error function.
300 // This is the same approximation used by Eigen.
ErfImpl32(XlaOp x)301 static XlaOp ErfImpl32(XlaOp x) {
302   static const std::array<float, 7> kAlpha{
303       -2.72614225801306e-10f, 2.77068142495902e-08f,  -2.10102402082508e-06f,
304       -5.69250639462346e-05f, -7.34990630326855e-04f, -2.95459980854025e-03f,
305       -1.60960333262415e-02f,
306   };
307 
308   static const std::array<float, 5> kBeta{
309       -1.45660718464996e-05f, -2.13374055278905e-04f, -1.68282697438203e-03f,
310       -7.37332916720468e-03f, -1.42647390514189e-02f,
311   };
312 
313   x = Clamp(ScalarLike(x, -4.f), x, ScalarLike(x, 4.f));
314   auto x2 = x * x;
315   return x * EvaluatePolynomial<float>(x2, kAlpha) /
316          EvaluatePolynomial<float>(x2, kBeta);
317 }
318 
Erf(XlaOp x)319 XlaOp Erf(XlaOp x) {
320   auto& b = *x.builder();
321   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
322     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("Erf", x));
323     TF_ASSIGN_OR_RETURN(auto shape, b.GetShape(x));
324     // erf(x) =
325     //   erf_impl(x)            if x < 1
326     //   1 - erfc_impl(x)       otherwise
327     if (shape.element_type() == F64) {
328       return Select(Lt(Abs(x), ScalarLike(x, 1)), ErfImpl64(x),
329                     ScalarLike(x, 1) - ErfcImpl64(x));
330     }
331     // Erf(c)Impl don't have enough precision when run with bf16 intermediates
332     // (not surprising!), so upcast to f32 in this case.
333     return DoWithUpcastToF32(x, {BF16, F16},
334                              [](XlaOp x) { return ErfImpl32(x); });
335   });
336 }
337 
338 namespace {
339 
340 // Approximation for the inverse error function from
341 //   Giles, M., "Approximating the erfinv function".
342 // The approximation has the form:
343 //   w = -log((1 - x) * (1 + x))
344 //   if ( w < 5 ) {
345 //     w = w - 2.5
346 //     p = sum_{i=1}^n lq[i]*w^i
347 //   } else {
348 //     w = sqrt(w) - 3
349 //     p = sum_{i=1}^n gq[i]*w^i
350 //   }
351 //   return p*x
ErfInv32(XlaOp x)352 XlaOp ErfInv32(XlaOp x) {
353   constexpr int kDegree = 9;
354   constexpr std::array<float, 9> w_less_than_5_constants = {
355       2.81022636e-08f,  3.43273939e-07f, -3.5233877e-06f,
356       -4.39150654e-06f, 0.00021858087f,  -0.00125372503f,
357       -0.00417768164f,  0.246640727f,    1.50140941f};
358   constexpr std::array<float, 9> w_greater_than_5_constants = {
359       -0.000200214257f, 0.000100950558f, 0.00134934322f,
360       -0.00367342844f,  0.00573950773f,  -0.0076224613f,
361       0.00943887047f,   1.00167406f,     2.83297682f};
362 
363   // Compute logarithm of (1+arg) using log1p(arg) which is more precise than
364   // log(1+arg) when arg is close to zero. For more details, see
365   // https://en.cppreference.com/w/cpp/numeric/math/log1p
366   auto w = -Log1p(-x * x);
367 
368   auto lt = Lt(w, ScalarLike(x, 5.0));
369   auto coefficient = [&](int i) {
370     return Select(lt, FullLike(x, w_less_than_5_constants[i]),
371                   FullLike(x, w_greater_than_5_constants[i]));
372   };
373   w = Select(lt, w - ScalarLike(x, 2.5), Sqrt(w) - ScalarLike(x, 3.0));
374   auto p = coefficient(0);
375   for (int i = 1; i < kDegree; ++i) {
376     p = coefficient(i) + p * w;
377   }
378 
379   // Result modulo edge cases.
380   XlaOp result = p * x;
381 
382   // Handle edge cases, namely erfinv(+/-1) = +/-inf.  (The above computation is
383   // indeterminate, and can give nan or -/+inf.)
384   auto& b = *x.builder();
385   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
386     TF_ASSIGN_OR_RETURN(Shape shape, b.GetShape(x));
387     return Select(Eq(Abs(x), ScalarLike(x, 1)),
388                   x * MaxValue(&b, shape.element_type()), result);
389   });
390 }
391 
ErfInv64(XlaOp x)392 XlaOp ErfInv64(XlaOp x) {
393   constexpr std::array<double, 23> w_less_than_6_25_constants = {
394       -3.6444120640178196996e-21, -1.685059138182016589e-19,
395       1.2858480715256400167e-18,  1.115787767802518096e-17,
396       -1.333171662854620906e-16,  2.0972767875968561637e-17,
397       6.6376381343583238325e-15,  -4.0545662729752068639e-14,
398       -8.1519341976054721522e-14, 2.6335093153082322977e-12,
399       -1.2975133253453532498e-11, -5.4154120542946279317e-11,
400       1.051212273321532285e-09,   -4.1126339803469836976e-09,
401       -2.9070369957882005086e-08, 4.2347877827932403518e-07,
402       -1.3654692000834678645e-06, -1.3882523362786468719e-05,
403       0.0001867342080340571352,   -0.00074070253416626697512,
404       -0.0060336708714301490533,  0.24015818242558961693,
405       1.6536545626831027356};
406   constexpr std::array<double, 19> w_less_than_16_constants = {
407       2.2137376921775787049e-09,  9.0756561938885390979e-08,
408       -2.7517406297064545428e-07, 1.8239629214389227755e-08,
409       1.5027403968909827627e-06,  -4.013867526981545969e-06,
410       2.9234449089955446044e-06,  1.2475304481671778723e-05,
411       -4.7318229009055733981e-05, 6.8284851459573175448e-05,
412       2.4031110387097893999e-05,  -0.0003550375203628474796,
413       0.00095328937973738049703,  -0.0016882755560235047313,
414       0.0024914420961078508066,   -0.0037512085075692412107,
415       0.005370914553590063617,    1.0052589676941592334,
416       3.0838856104922207635,
417   };
418   constexpr std::array<double, 17> w_greater_than_16_constants = {
419       -2.7109920616438573243e-11, -2.5556418169965252055e-10,
420       1.5076572693500548083e-09,  -3.7894654401267369937e-09,
421       7.6157012080783393804e-09,  -1.4960026627149240478e-08,
422       2.9147953450901080826e-08,  -6.7711997758452339498e-08,
423       2.2900482228026654717e-07,  -9.9298272942317002539e-07,
424       4.5260625972231537039e-06,  -1.9681778105531670567e-05,
425       7.5995277030017761139e-05,  -0.00021503011930044477347,
426       -0.00013871931833623122026, 1.0103004648645343977,
427       4.8499064014085844221,
428   };
429   // Compute logarithm of (1+arg) using log1p(arg) which is more precise than
430   // log(1+arg) when arg is close to zero. For more details, see
431   // https://en.cppreference.com/w/cpp/numeric/math/log1p
432   auto w = -Log1p(-x * x);
433 
434   auto lt_6_25 = Lt(w, ScalarLike(x, 6.25));
435   auto lt_16 = Lt(w, ScalarLike(x, 16));
436   auto coefficient = [&](int i) {
437     auto c = FullLike(x, w_less_than_6_25_constants[i]);
438     if (i < 19) {
439       c = Select(lt_6_25, c, FullLike(x, w_less_than_16_constants[i]));
440     }
441     if (i < 17) {
442       c = Select(lt_16, c, FullLike(x, w_greater_than_16_constants[i]));
443     }
444     return c;
445   };
446   auto sqrt_w = Sqrt(w);
447   w = Select(lt_6_25, w - ScalarLike(x, 3.125),
448              sqrt_w - Select(lt_16, ScalarLike(x, 3.25), ScalarLike(x, 5.0)));
449   auto p = coefficient(0);
450   for (int i = 1; i < 17; ++i) {
451     p = coefficient(i) + p * w;
452   }
453   for (int i = 17; i < 19; ++i) {
454     p = Select(lt_16, coefficient(i) + p * w, p);
455   }
456   for (int i = 19; i < 23; ++i) {
457     p = Select(lt_6_25, coefficient(i) + p * w, p);
458   }
459   // Result modulo edge cases.
460   XlaOp result = p * x;
461 
462   // Handle edge cases, namely erfinv(+/-1) = +/-inf.  (The above computation is
463   // indeterminate, and can give nan or -/+inf.)
464   auto& b = *x.builder();
465   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
466     TF_ASSIGN_OR_RETURN(Shape shape, b.GetShape(x));
467     return Select(Eq(Abs(x), ScalarLike(x, 1)),
468                   x * MaxValue(&b, shape.element_type()), result);
469   });
470 }
471 
472 }  // namespace
473 
ErfInv(XlaOp x)474 XlaOp ErfInv(XlaOp x) {
475   auto& b = *x.builder();
476   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
477     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("ErfInv", x));
478     TF_ASSIGN_OR_RETURN(auto shape, b.GetShape(x));
479     if (shape.element_type() == F64) {
480       return ErfInv64(x);
481     }
482     return DoWithUpcastToF32(x, {BF16, F16},
483                              [](XlaOp x) { return ErfInv32(x); });
484   });
485 }
486 
487 namespace {
488 // Coefficients for the Lanczos approximation of the gamma function. The
489 // coefficients are uniquely determined by the choice of g and n (kLanczosGamma
490 // and kLanczosCoefficients.size() + 1). The coefficients below correspond to
491 // [7, 9]. [5, 7], [7, 9], [9, 10], and [607/128.0, 15] were evaluated and [7,
492 // 9] seemed to be the least sensitive to the quality of the log function. In
493 // particular, [5, 7] is the only choice where -1.5e-5 <= lgamma(2) <= 1.5e-5
494 // for a particularly inaccurate log function.
495 static constexpr double kLanczosGamma = 7;  // aka g
496 static constexpr double kBaseLanczosCoeff = 0.99999999999980993227684700473478;
497 static constexpr std::array<double, 8> kLanczosCoefficients = {
498     676.520368121885098567009190444019, -1259.13921672240287047156078755283,
499     771.3234287776530788486528258894,   -176.61502916214059906584551354,
500     12.507343278686904814458936853,     -0.13857109526572011689554707,
501     9.984369578019570859563e-6,         1.50563273514931155834e-7};
502 }  // namespace
503 
504 // Compute the Lgamma function using Lanczos' approximation from "A Precision
505 // Approximation of the Gamma Function". SIAM Journal on Numerical Analysis
506 // series B. Vol. 1:
507 // lgamma(z + 1) = (log(2) + log(pi)) / 2 + (z + 1/2) * log(t(z)) - t(z) + A(z)
508 // t(z) = z + kLanczosGamma + 1/2
509 // A(z) = kBaseLanczosCoeff + sigma(k = 1, n, kLanczosCoefficients[i] / (z + k))
Lgamma(XlaOp input)510 XlaOp Lgamma(XlaOp input) {
511   auto do_it = [](XlaOp input) {
512     XlaOp one_half = ScalarLike(input, 0.5);
513     XlaOp one = ScalarLike(input, 1);
514 
515     XlaOp pi = ScalarLike(input, M_PI);
516     XlaOp log_pi = ScalarLike(input, std::log(M_PI));
517     XlaOp log_sqrt_two_pi =
518         ScalarLike(input, (std::log(2) + std::log(M_PI)) / 2);
519 
520     XlaOp lanczos_gamma_plus_one_half = ScalarLike(input, kLanczosGamma + 0.5);
521     XlaOp log_lanczos_gamma_plus_one_half =
522         ScalarLike(input, std::log(kLanczosGamma + 0.5));
523 
524     XlaOp base_lanczos_coeff = ScalarLike(input, kBaseLanczosCoeff);
525 
526     // If the input is less than 0.5 use Euler's reflection formula:
527     // gamma(x) = pi / (sin(pi * x) * gamma(1 - x))
528     XlaOp need_to_reflect = Lt(input, one_half);
529     XlaOp z = Select(need_to_reflect, -input, input - one);
530 
531     XlaOp x = base_lanczos_coeff;
532     for (int i = 0, end = kLanczosCoefficients.size(); i < end; ++i) {
533       XlaOp lanczos_coefficient = ScalarLike(input, kLanczosCoefficients[i]);
534       XlaOp index = ScalarLike(input, i);
535       x = x + lanczos_coefficient / (z + index + one);
536     }
537 
538     // To improve accuracy on platforms with less-precise log implementations,
539     // compute log(lanczos_gamma_plus_one_half) at compile time and use log1p on
540     // the device.
541     // log(t) = log(kLanczosGamma + 0.5 + z)
542     //        = log(kLanczosGamma + 0.5) + log1p(z / (kLanczosGamma + 0.5))
543     XlaOp t = lanczos_gamma_plus_one_half + z;
544     XlaOp log_t = log_lanczos_gamma_plus_one_half +
545                   Log1p(z / lanczos_gamma_plus_one_half);
546 
547     // Compute the final result (modulo reflection).  t(z) may be large, and we
548     // need to be careful not to overflow to infinity in the first term of
549     //
550     //   (z + 1/2) * log(t(z)) - t(z).
551     //
552     // Therefore we compute this as
553     //
554     //   (z + 1/2 - t(z) / log(t(z))) * log(t(z)).
555     //
556     XlaOp log_y = log_sqrt_two_pi + (z + one_half - t / log_t) * log_t + Log(x);
557 
558     // Compute the reflected value, used when x < 0.5:
559     //
560     //   lgamma(x) = log(pi) - lgamma(1-x) - log(abs(sin(pi * x))).
561     //
562     // (The abs is because lgamma is the log of the absolute value of the gamma
563     // function.)
564     //
565     // We have to be careful when computing the final term above. gamma(x) goes
566     // to +/-inf at every integer x < 0, and this is controlled by the
567     // sin(pi * x) term.  The slope is large, so precision is particularly
568     // important.
569     //
570     // Because abs(sin(pi * x)) has period 1, we can equivalently use
571     // abs(sin(pi * frac(x))), where frac(x) is the fractional part of x.  This
572     // is more numerically accurate: It doesn't overflow to inf like pi * x can,
573     // and if x is an integer, it evaluates to 0 exactly, which is significant
574     // because we then take the log of this value, and log(0) is inf.
575     //
576     // We don't have a frac(x) primitive in XLA and computing it is tricky, but
577     // because abs(sin(pi * x)) = abs(sin(pi * abs(x))), it's good enough for
578     // our purposes to use abs(frac(x)) = abs(x) - floor(abs(x)).
579     //
580     // Furthermore, pi * abs(frac(x)) loses precision when abs(frac(x)) is close
581     // to 1.  To remedy this, we can use the fact that sin(pi * x) in the domain
582     // [0, 1] is symmetric across the line Y=0.5.
583     //
584     XlaOp abs_input = Abs(input);
585     XlaOp abs_frac_input = abs_input - Floor(abs_input);
586     // Convert values of abs_frac_input > 0.5 to (1 - frac_input) to improve
587     // precision of pi * abs_frac_input for values of abs_frac_input close to 1.
588     XlaOp reduced_frac_input =
589         Select(Gt(abs_frac_input, ScalarLike(abs_frac_input, 0.5)),
590                ScalarLike(abs_frac_input, 1) - abs_frac_input, abs_frac_input);
591     XlaOp reflection_denom = Log(Sin(pi * reduced_frac_input));
592 
593     // Avoid computing -inf - inf, which is nan.  If reflection_denom is +/-inf,
594     // then it "wins" and the result is +/-inf.
595     XlaOp reflection =
596         Select(IsFinite(reflection_denom), log_pi - reflection_denom - log_y,
597                -reflection_denom);
598     XlaOp result = Select(need_to_reflect, reflection, log_y);
599 
600     // lgamma(+/-inf) = +inf.
601     XlaOp inf_bcast = FullLike(input, std::numeric_limits<float>::infinity());
602     return Select(IsInf(input), inf_bcast, result);
603   };
604 
605   auto& b = *input.builder();
606   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
607     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("Lgamma", input));
608     // F16 and BF16 don't provide sufficient precision for intermediate results
609     // here (although it's better than you might expect!), so do the
610     // computations in F32.
611     return DoWithUpcastToF32(input, {BF16, F16}, do_it);
612   });
613 }
614 
615 // Computes an approximation of the lbeta function which is equivalent to
616 // log(abs(Beta(a, b))) but avoids overflow by computing it with lgamma.
Lbeta(XlaOp a,XlaOp b)617 static XlaOp Lbeta(XlaOp a, XlaOp b) {
618   // Beta(a, b) can be computed using Gamma as per
619   // http://dlmf.nist.gov/5.12.E1 as follows:
620   //   Beta(a, b) = (Gamma(a) * Gamma(b)) / Gamma(a + b)
621   //
622   // To avoid overflow, we compute in the log domain.
623   //
624   // As per http://dlmf.nist.gov/4.8.E2 we can transform:
625   //   Log(a * b)
626   // into:
627   //   Log(a) + Log(b)
628   //
629   // Likewise, per https://dlmf.nist.gov/4.8.E4, we can turn:
630   //   Log(a - b)
631   // into:
632   //   Log(a) - Log(b)
633   //
634   // This means that we can compute Log(Beta(a, b)) by:
635   //   Log(Gamma(a)) + Log(Gamma(b)) - Log(Gamma(a + b))
636   return Lgamma(a) + Lgamma(b) - Lgamma(a + b);
637 }
638 
639 // Compute the Digamma function using Lanczos' approximation from "A Precision
640 // Approximation of the Gamma Function". SIAM Journal on Numerical Analysis
641 // series B. Vol. 1:
642 // digamma(z + 1) = log(t(z)) + A'(z) / A(z) - kLanczosGamma / t(z)
643 // t(z) = z + kLanczosGamma + 1/2
644 // A(z) = kBaseLanczosCoeff + sigma(k = 1, n, kLanczosCoefficients[i] / (z + k))
645 // A'(z) = sigma(k = 1, n, kLanczosCoefficients[i] / (z + k) / (z + k))
Digamma(XlaOp input)646 XlaOp Digamma(XlaOp input) {
647   auto do_it = [](XlaOp input) {
648     XlaOp zero = ScalarLike(input, 0);
649     XlaOp one_half = ScalarLike(input, 0.5);
650     XlaOp one = ScalarLike(input, 1);
651 
652     XlaOp pi = ScalarLike(input, M_PI);
653 
654     XlaOp lanczos_gamma = ScalarLike(input, kLanczosGamma);
655     XlaOp lanczos_gamma_plus_one_half = ScalarLike(input, kLanczosGamma + 0.5);
656     XlaOp log_lanczos_gamma_plus_one_half =
657         ScalarLike(input, std::log(kLanczosGamma + 0.5));
658 
659     XlaOp base_lanczos_coeff = ScalarLike(input, kBaseLanczosCoeff);
660 
661     // If the input is less than 0.5 use Euler's reflection formula:
662     // digamma(x) = digamma(1 - x) - pi * cot(pi * x)
663     XlaOp need_to_reflect = Lt(input, one_half);
664     XlaOp z = Select(need_to_reflect, -input, input - one);
665 
666     XlaOp num = zero;
667     XlaOp denom = base_lanczos_coeff;
668     for (int i = 0, end = kLanczosCoefficients.size(); i < end; ++i) {
669       XlaOp lanczos_coefficient = ScalarLike(input, kLanczosCoefficients[i]);
670       XlaOp index = ScalarLike(input, i);
671       num = num - lanczos_coefficient / ((z + index + one) * (z + index + one));
672       denom = denom + lanczos_coefficient / (z + index + one);
673     }
674 
675     // To improve accuracy on platforms with less-precise log implementations,
676     // compute log(lanczos_gamma_plus_one_half) at compile time and use log1p on
677     // the device.
678     // log(t) = log(kLanczosGamma + 0.5 + z)
679     //        = log(kLanczosGamma + 0.5) + log1p(z / (kLanczosGamma + 0.5))
680     XlaOp t = lanczos_gamma_plus_one_half + z;
681     XlaOp log_t = log_lanczos_gamma_plus_one_half +
682                   Log1p(z / lanczos_gamma_plus_one_half);
683 
684     XlaOp y = log_t + num / denom - lanczos_gamma / t;
685 
686     // We need to be careful how we compute cot(pi * input) below: For
687     // near-integral values of `input`, pi * input can lose precision.
688     //
689     // Input is already known to be less than 0.5 (otherwise we don't have to
690     // reflect).  We shift values smaller than -0.5 into the range [-.5, .5] to
691     // increase precision of pi * input and the resulting cotangent.
692     XlaOp reduced_input = input + Abs(Floor(input + ScalarLike(input, 0.5)));
693     XlaOp reflection =
694         y - pi * Cos(pi * reduced_input) / Sin(pi * reduced_input);
695     XlaOp real_result = Select(need_to_reflect, reflection, y);
696 
697     // Digamma has poles at negative integers and zero; return nan for those.
698     return Select(And(Le(input, zero), Eq(input, Floor(input))),
699                   FullLike(input, std::numeric_limits<float>::quiet_NaN()),
700                   real_result);
701   };
702 
703   auto& b = *input.builder();
704   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
705     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("Digamma", input));
706     return DoWithUpcastToF32(input, {BF16, F16}, do_it);
707   });
708 }
709 
710 // Incomplete gamma functions
711 
712 namespace {
713 
714 enum kIgammaMode { VALUE, DERIVATIVE, SAMPLE_DERIVATIVE };
715 
716 // Helper function for computing Igamma using a power series.
717 template <kIgammaMode mode>
IgammaSeries(XlaOp ax,XlaOp x,XlaOp a,XlaOp enabled,xla::PrimitiveType type)718 XlaOp IgammaSeries(XlaOp ax, XlaOp x, XlaOp a, XlaOp enabled,
719                    xla::PrimitiveType type) {
720   // vals: (enabled, r, c, ans, x)
721   // 'enabled' is a predication mask that says for which elements we should
722   // execute the loop body. Disabled elements have no effect in the loop body.
723   // TODO(phawkins): in general this isn't an optimal implementation on any
724   // backend. For example, on GPU, we should probably vectorize to the warp
725   // size, and then run independent loops for each warp's worth of
726   // data.
727   auto cond = [&](absl::Span<const XlaOp> vals,
728                   XlaBuilder* builder) -> StatusOr<XlaOp> {
729     XlaOp enabled = vals[0];
730     return Any(enabled);
731   };
732   auto body = [&](absl::Span<const XlaOp> vals,
733                   XlaBuilder* builder) -> StatusOr<std::vector<XlaOp>> {
734     XlaOp enabled = vals[0];
735     XlaOp r = vals[1];
736     XlaOp c = vals[2];
737     XlaOp ans = vals[3];
738     XlaOp x = vals[4];
739     XlaOp dc_da = vals[5];
740     XlaOp dans_da = vals[6];
741 
742     r = r + ScalarLike(r, 1);
743     dc_da = dc_da * (x / r) + (ScalarLike(r, -1) * c * x) / (r * r);
744     dans_da = dans_da + dc_da;
745     c = c * (x / r);
746     ans = ans + c;
747     XlaOp conditional;
748     if (mode == VALUE) {
749       conditional = And(enabled, Gt(c / ans, Epsilon(builder, type)));
750     } else {
751       conditional =
752           And(enabled, Gt(Abs(dc_da / dans_da), Epsilon(builder, type)));
753     }
754 
755     return std::vector<XlaOp>{
756         conditional,
757         Select(enabled, r, vals[1]),
758         Select(enabled, c, vals[2]),
759         Select(enabled, ans, vals[3]),
760         Select(enabled, x, vals[4]),
761         Select(enabled, dc_da, vals[5]),
762         Select(enabled, dans_da, vals[6]),
763     };
764   };
765   auto& b = *ax.builder();
766   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
767     std::vector<XlaOp> vals = {
768         enabled,        a, FullLike(a, 1), FullLike(a, 1), x, FullLike(a, 0),
769         FullLike(a, 0),
770     };
771 
772     TF_ASSIGN_OR_RETURN(vals, WhileLoopHelper(cond, body, vals, "igamma", &b));
773     XlaOp ans = vals[3];
774     XlaOp dans_da = vals[6];
775     if (mode == VALUE) {
776       return (ans * ax) / a;
777     }
778 
779     XlaOp dlogax_da = Log(x) - Digamma(a + ScalarLike(a, 1));
780 
781     switch (mode) {
782       case DERIVATIVE:
783         return ax * (ans * dlogax_da + dans_da) / a;
784       case SAMPLE_DERIVATIVE:
785       default:
786         return -(dans_da + ans * dlogax_da) * x / a;
787     }
788   });
789 }
790 
791 // Helper function for computing Igammac using a continued fraction.
792 template <kIgammaMode mode>
IgammacContinuedFraction(XlaOp ax,XlaOp x,XlaOp a,XlaOp enabled,xla::PrimitiveType type)793 XlaOp IgammacContinuedFraction(XlaOp ax, XlaOp x, XlaOp a, XlaOp enabled,
794                                xla::PrimitiveType type) {
795   // vals: enabled, ans, t, y, z, c, pkm1, qkm1, pkm2, qkm2
796   auto cond = [&](absl::Span<const XlaOp> vals,
797                   XlaBuilder* builder) -> StatusOr<XlaOp> {
798     XlaOp enabled = vals[0];
799     XlaOp c = vals[5];
800     return And(Lt(c, ScalarLike(c, 2000)), Any(enabled));
801   };
802   auto body = [&](absl::Span<const XlaOp> vals,
803                   XlaBuilder* builder) -> StatusOr<std::vector<XlaOp>> {
804     XlaOp enabled = vals[0];
805     XlaOp ans = vals[1];
806     XlaOp t = vals[2];
807     XlaOp y = vals[3];
808     XlaOp z = vals[4];
809     XlaOp c = vals[5];
810     XlaOp pkm1 = vals[6];
811     XlaOp qkm1 = vals[7];
812     XlaOp pkm2 = vals[8];
813     XlaOp qkm2 = vals[9];
814 
815     XlaOp dpkm2_da = vals[10];
816     XlaOp dqkm2_da = vals[11];
817     XlaOp dpkm1_da = vals[12];
818     XlaOp dqkm1_da = vals[13];
819     XlaOp dans_da = vals[14];
820 
821     c = c + ScalarLike(c, 1);
822     y = y + ScalarLike(y, 1);
823     z = z + ScalarLike(z, 2);
824     XlaOp yc = y * c;
825     XlaOp pk = pkm1 * z - pkm2 * yc;
826     XlaOp qk = qkm1 * z - qkm2 * yc;
827     XlaOp qk_is_nonzero = Ne(qk, ScalarLike(qk, 0));
828     XlaOp r = pk / qk;
829 
830     t = Select(qk_is_nonzero, Abs((ans - r) / r), FullLike(t, 1));
831     ans = Select(qk_is_nonzero, r, ans);
832 
833     XlaOp dpk_da = dpkm1_da * z - pkm1 - dpkm2_da * yc + pkm2 * c;
834     XlaOp dqk_da = dqkm1_da * z - qkm1 - dqkm2_da * yc + qkm2 * c;
835     XlaOp dans_da_new =
836         Select(qk_is_nonzero, (dpk_da - ans * dqk_da) / qk, dans_da);
837     XlaOp grad_conditional =
838         Select(qk_is_nonzero, Abs(dans_da_new - dans_da), FullLike(dans_da, 1));
839 
840     pkm2 = pkm1;
841     pkm1 = pk;
842     qkm2 = qkm1;
843     qkm1 = qk;
844 
845     dpkm2_da = dpkm1_da;
846     dqkm2_da = dqkm1_da;
847     dpkm1_da = dpk_da;
848     dqkm1_da = dqk_da;
849 
850     XlaOp rescale = Gt(Abs(pk), Reciprocal(Epsilon(builder, type)));
851     pkm2 = Select(rescale, pkm2 * Epsilon(builder, type), pkm2);
852     pkm1 = Select(rescale, pkm1 * Epsilon(builder, type), pkm1);
853     qkm2 = Select(rescale, qkm2 * Epsilon(builder, type), qkm2);
854     qkm1 = Select(rescale, qkm1 * Epsilon(builder, type), qkm1);
855 
856     dpkm2_da = Select(rescale, dpkm2_da * Epsilon(builder, type), dpkm2_da);
857     dqkm2_da = Select(rescale, dqkm2_da * Epsilon(builder, type), dqkm2_da);
858     dpkm1_da = Select(rescale, dpkm1_da * Epsilon(builder, type), dpkm1_da);
859     dqkm1_da = Select(rescale, dqkm1_da * Epsilon(builder, type), dqkm1_da);
860 
861     XlaOp conditional;
862     if (mode == VALUE) {
863       conditional = And(enabled, Gt(t, Epsilon(builder, type)));
864     } else {
865       conditional = And(enabled, Gt(grad_conditional, Epsilon(builder, type)));
866     }
867 
868     return std::vector<XlaOp>{conditional,
869                               Select(enabled, ans, vals[1]),
870                               Select(enabled, t, vals[2]),
871                               Select(enabled, y, vals[3]),
872                               Select(enabled, z, vals[4]),
873                               c,
874                               Select(enabled, pkm1, vals[6]),
875                               Select(enabled, qkm1, vals[7]),
876                               Select(enabled, pkm2, vals[8]),
877                               Select(enabled, qkm2, vals[9]),
878                               Select(enabled, dpkm2_da, vals[10]),
879                               Select(enabled, dqkm2_da, vals[11]),
880                               Select(enabled, dpkm1_da, vals[12]),
881                               Select(enabled, dqkm1_da, vals[13]),
882                               Select(enabled, dans_da_new, vals[14])};
883   };
884 
885   auto& b = *ax.builder();
886   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
887     XlaOp y = ScalarLike(a, 1) - a;
888     XlaOp z = x + y + ScalarLike(x, 1);
889     XlaOp c = ScalarLike(x, 0);
890     XlaOp pkm2 = FullLike(x, 1);
891     XlaOp qkm2 = x;
892     XlaOp pkm1 = x + ScalarLike(x, 1);
893     XlaOp qkm1 = z * x;
894     XlaOp ans = pkm1 / qkm1;
895     XlaOp t = FullLike(x, 1);
896     XlaOp dpkm2_da = FullLike(x, 0);
897     XlaOp dqkm2_da = FullLike(x, 0);
898     XlaOp dpkm1_da = FullLike(x, 0);
899     XlaOp dqkm1_da = -x;
900     XlaOp dans_da = (dpkm1_da - ans * dqkm1_da) / qkm1;
901     std::vector<XlaOp> vals = {enabled,  ans,      t,        y,        z,
902                                c,        pkm1,     qkm1,     pkm2,     qkm2,
903                                dpkm2_da, dqkm2_da, dpkm1_da, dqkm1_da, dans_da};
904 
905     TF_ASSIGN_OR_RETURN(vals, WhileLoopHelper(cond, body, vals, "igammac", &b));
906     ans = vals[1];
907     if (mode == VALUE) {
908       return ans * ax;
909     }
910 
911     dans_da = vals[14];
912     XlaOp dlogax_da = Log(x) - Digamma(a);
913 
914     switch (mode) {
915       case DERIVATIVE:
916         return ax * (ans * dlogax_da + dans_da);
917       case SAMPLE_DERIVATIVE:
918       default:
919         return -(dans_da + ans * dlogax_da) * x;
920     }
921   });
922 }
923 
924 }  // namespace
925 
Igamma(XlaOp a,XlaOp x)926 XlaOp Igamma(XlaOp a, XlaOp x) {
927   auto& b = *a.builder();
928   auto doit = [&b](XlaOp a, XlaOp x, PrimitiveType type) -> XlaOp {
929     XlaOp is_nan = Or(IsNan(a), IsNan(x));
930     XlaOp x_is_zero = Eq(x, ScalarLike(x, 0));
931     XlaOp domain_error = Or(Lt(x, ScalarLike(x, 0)), Le(a, ScalarLike(a, 0)));
932     XlaOp use_igammac = And(Gt(x, ScalarLike(x, 1)), Gt(x, a));
933     XlaOp ax = a * Log(x) - x - Lgamma(a);
934     XlaOp underflow = Lt(ax, -Log(MaxFiniteValue(&b, type)));
935     ax = Exp(ax);
936     XlaOp enabled = Not(Or(Or(Or(x_is_zero, domain_error), underflow), is_nan));
937     const double nan = std::numeric_limits<double>::quiet_NaN();
938     XlaOp output = Select(
939         use_igammac,
940         ScalarLike(a, 1) - IgammacContinuedFraction<VALUE>(
941                                ax, x, a, And(enabled, use_igammac), type),
942         IgammaSeries<VALUE>(ax, x, a, And(enabled, Not(use_igammac)), type));
943     output = Select(x_is_zero, ZerosLike(output), output);
944     output = Select(Or(domain_error, is_nan), FullLike(a, nan), output);
945     return output;
946   };
947   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
948     TF_ASSIGN_OR_RETURN(auto a_shape, b.GetShape(a));
949     TF_ASSIGN_OR_RETURN(auto x_shape, b.GetShape(x));
950     if (a_shape != x_shape) {
951       return InvalidArgument(
952           "Arguments to Igamma must have equal shapes and types; got %s and %s",
953           a_shape.ToString(), x_shape.ToString());
954     }
955     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("Igamma", a));
956     PrimitiveType a_x_type = a_shape.element_type();
957     bool needs_upcast =
958         a_shape.element_type() == F16 || a_shape.element_type() == BF16;
959 
960     if (needs_upcast) {
961       a = ConvertElementType(a, F32);
962       x = ConvertElementType(x, F32);
963       a_x_type = F32;
964     }
965     XlaOp result = doit(a, x, a_x_type);
966     if (needs_upcast) {
967       result = ConvertElementType(result, a_shape.element_type());
968     }
969     return result;
970   });
971 }
972 
IgammaGradA(XlaOp a,XlaOp x)973 XlaOp IgammaGradA(XlaOp a, XlaOp x) {
974   auto& b = *a.builder();
975   auto doit = [&b](XlaOp a, XlaOp x, PrimitiveType type) -> XlaOp {
976     XlaOp is_nan = Or(IsNan(a), IsNan(x));
977     XlaOp x_is_zero = Eq(x, ScalarLike(x, 0));
978     XlaOp domain_error = Or(Lt(x, ScalarLike(x, 0)), Le(a, ScalarLike(a, 0)));
979     XlaOp use_igammac = And(Gt(x, ScalarLike(x, 1)), Gt(x, a));
980     XlaOp ax = a * Log(x) - x - Lgamma(a);
981     XlaOp underflow = Lt(ax, -Log(MaxFiniteValue(&b, type)));
982     ax = Exp(ax);
983     XlaOp enabled = Not(Or(Or(Or(x_is_zero, domain_error), underflow), is_nan));
984     const double nan = std::numeric_limits<double>::quiet_NaN();
985     XlaOp output = Select(use_igammac,
986                           -IgammacContinuedFraction<DERIVATIVE>(
987                               ax, x, a, And(enabled, use_igammac), type),
988                           IgammaSeries<DERIVATIVE>(
989                               ax, x, a, And(enabled, Not(use_igammac)), type));
990     output = Select(x_is_zero, ZerosLike(output), output);
991     output = Select(Or(domain_error, is_nan), FullLike(a, nan), output);
992     return output;
993   };
994   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
995     TF_ASSIGN_OR_RETURN(auto a_shape, b.GetShape(a));
996     TF_ASSIGN_OR_RETURN(auto x_shape, b.GetShape(x));
997     if (a_shape != x_shape) {
998       return InvalidArgument(
999           "Arguments to IgammaGradA must have equal shapes and types; got %s "
1000           "and %s",
1001           a_shape.ToString(), x_shape.ToString());
1002     }
1003     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("IgammaGradA", a));
1004     bool needs_upcast =
1005         a_shape.element_type() == F16 || a_shape.element_type() == BF16;
1006 
1007     if (needs_upcast) {
1008       a = ConvertElementType(a, F32);
1009       x = ConvertElementType(x, F32);
1010     }
1011     XlaOp result = doit(a, x, a_shape.element_type());
1012     if (needs_upcast) {
1013       result = ConvertElementType(result, a_shape.element_type());
1014     }
1015     return result;
1016   });
1017 }
1018 
1019 // Gradient of Gamma sample from Gamma(a, 1) with respect to `a`.
RandomGammaGrad(XlaOp a,XlaOp x)1020 XlaOp RandomGammaGrad(XlaOp a, XlaOp x) {
1021   auto& b = *a.builder();
1022   auto doit = [&b](XlaOp a, XlaOp x, PrimitiveType type) -> XlaOp {
1023     XlaOp is_nan = Or(IsNan(a), IsNan(x));
1024     XlaOp x_is_zero = Eq(x, ScalarLike(x, 0));
1025     XlaOp domain_error = Or(Lt(x, ScalarLike(x, 0)), Le(a, ScalarLike(a, 0)));
1026     XlaOp use_igammac = And(Gt(x, ScalarLike(x, 1)), Gt(x, a));
1027     XlaOp ax = a * Log(x) - x - Lgamma(a);
1028     XlaOp underflow = Lt(ax, -Log(MaxFiniteValue(&b, type)));
1029     ax = Exp(ax);
1030     XlaOp enabled = Not(Or(Or(Or(x_is_zero, domain_error), underflow), is_nan));
1031     const double nan = std::numeric_limits<double>::quiet_NaN();
1032     XlaOp output = Select(use_igammac,
1033                           -IgammacContinuedFraction<SAMPLE_DERIVATIVE>(
1034                               ax, x, a, And(enabled, use_igammac), type),
1035                           IgammaSeries<SAMPLE_DERIVATIVE>(
1036                               ax, x, a, And(enabled, Not(use_igammac)), type));
1037     output = Select(x_is_zero, ZerosLike(output), output);
1038     output = Select(Or(domain_error, is_nan), FullLike(a, nan), output);
1039     return output;
1040   };
1041   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1042     TF_ASSIGN_OR_RETURN(auto a_shape, b.GetShape(a));
1043     TF_ASSIGN_OR_RETURN(auto x_shape, b.GetShape(x));
1044     if (a_shape != x_shape) {
1045       return InvalidArgument(
1046           "Arguments to RandomGammaGrad must have equal shapes and types; got "
1047           "%s and %s",
1048           a_shape.ToString(), x_shape.ToString());
1049     }
1050     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("RandomGammaGrad", a));
1051     bool needs_upcast =
1052         a_shape.element_type() == F16 || a_shape.element_type() == BF16;
1053 
1054     if (needs_upcast) {
1055       a = ConvertElementType(a, F32);
1056       x = ConvertElementType(x, F32);
1057     }
1058     XlaOp result = doit(a, x, a_shape.element_type());
1059     if (needs_upcast) {
1060       result = ConvertElementType(result, a_shape.element_type());
1061     }
1062     return result;
1063   });
1064 }
1065 
Igammac(XlaOp a,XlaOp x)1066 XlaOp Igammac(XlaOp a, XlaOp x) {
1067   auto& b = *a.builder();
1068   auto doit = [&b](XlaOp a, XlaOp x, PrimitiveType type) -> XlaOp {
1069     XlaOp out_of_range = Or(Le(x, ScalarLike(x, 0)), Le(a, ScalarLike(a, 0)));
1070     XlaOp use_igamma = Or(Lt(x, ScalarLike(x, 1)), Lt(x, a));
1071     XlaOp ax = a * Log(x) - x - Lgamma(a);
1072     XlaOp underflow = Lt(ax, -Log(MaxFiniteValue(&b, type)));
1073     XlaOp enabled = Not(Or(out_of_range, underflow));
1074     ax = Exp(ax);
1075     XlaOp result =
1076         Select(use_igamma,
1077                ScalarLike(a, 1) - IgammaSeries<VALUE>(
1078                                       ax, x, a, And(enabled, use_igamma), type),
1079                IgammacContinuedFraction<VALUE>(
1080                    ax, x, a, And(enabled, Not(use_igamma)), type));
1081     return Select(out_of_range, FullLike(a, 1), result);
1082   };
1083   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1084     TF_ASSIGN_OR_RETURN(auto a_shape, b.GetShape(a));
1085     TF_ASSIGN_OR_RETURN(auto x_shape, b.GetShape(x));
1086     if (a_shape != x_shape) {
1087       return InvalidArgument(
1088           "Arguments to Igammac must have equal shapes and types; "
1089           "got %s and %s",
1090           a_shape.ToString(), x_shape.ToString());
1091     }
1092     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("Igammac", a));
1093     PrimitiveType a_x_type = a_shape.element_type();
1094     bool needs_upcast =
1095         a_shape.element_type() == F16 || a_shape.element_type() == BF16;
1096 
1097     if (needs_upcast) {
1098       a = ConvertElementType(a, F32);
1099       x = ConvertElementType(x, F32);
1100       a_x_type = F32;
1101     }
1102     XlaOp result = doit(a, x, a_x_type);
1103     if (needs_upcast) {
1104       result = ConvertElementType(result, a_shape.element_type());
1105     }
1106     return result;
1107   });
1108 }
1109 // Implements Banker's rounding: numbers that are equidistant between two
1110 // integers are rounded towards even.
RoundToEven(XlaOp x)1111 XlaOp RoundToEven(XlaOp x) {
1112   auto& b = *x.builder();
1113   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1114     // Reject non-real non-fp inputs (What does it even mean to round a complex
1115     // number?  Do you round each component equally?  In that case, you should
1116     // just ask for that explicitly.)
1117     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("RoundToEven", x));
1118 
1119     auto half = ScalarLike(x, 0.5);
1120     auto one = ScalarLike(x, 1.0);
1121     auto two = ScalarLike(x, 2.0);
1122 
1123     auto round_val = Floor(x);
1124     auto fraction = x - round_val;
1125     auto nearest_even_int = round_val - two * Floor(half * x);
1126     auto is_odd = Eq(nearest_even_int, one);
1127     return Select(Or(Gt(fraction, half), And(Eq(fraction, half), is_odd)),
1128                   round_val + one, round_val);
1129   });
1130 }
1131 
1132 // Trigonometric functions.
1133 
1134 // acos(x) = 2 * atan(sqrt(1 - x^2) / (1 + x)) if x != -1
1135 //           pi                                if x == -1
1136 // For complex:
1137 // acos(x) = -(i * log(x + i * sqrt((1 + x) * (1 - x))))
Acos(XlaOp x)1138 XlaOp Acos(XlaOp x) {
1139   XlaBuilder* b = x.builder();
1140   return b->ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1141     TF_ASSIGN_OR_RETURN(auto shape, b->GetShape(x));
1142 
1143     if (primitive_util::IsComplexType(shape.element_type())) {
1144       auto one = ScalarLike(x, 1);
1145       auto imag_one = Complex(
1146           Zero(b, primitive_util::ComplexComponentType(shape.element_type())),
1147           One(b, primitive_util::ComplexComponentType(shape.element_type())));
1148 
1149       auto result =
1150           Neg(imag_one * Log(x + imag_one * Sqrt((one + x) * (one - x))));
1151       return result;
1152     }
1153     return Select(Ne(x, FullLike(x, -1)),
1154                   ScalarLike(x, 2.0) * Atan2(Sqrt(ScalarLike(x, 1.0) - x * x),
1155                                              ScalarLike(x, 1.0) + x),
1156                   FullLike(x, M_PI));
1157   });
1158 }
1159 
1160 // asin(x) = 2 * atan(x / (1 + sqrt(1 - x^2)))
Asin(XlaOp x)1161 XlaOp Asin(XlaOp x) {
1162   return ScalarLike(x, 2.0) *
1163          Atan2(x, ScalarLike(x, 1.0) + Sqrt(ScalarLike(x, 1.0) - x * x));
1164 }
1165 
Atan(XlaOp x)1166 XlaOp Atan(XlaOp x) { return Atan2(x, ScalarLike(x, 1.0)); }
1167 
Tan(XlaOp x)1168 XlaOp Tan(XlaOp x) {
1169   return DoWithUpcastToF32(x, {F16}, [](XlaOp x) { return Sin(x) / Cos(x); });
1170 }
1171 
1172 // Hyperbolic trigonometric functions.
1173 
1174 // acosh(x) = log(x + sqrt(x^2 - 1))      if x >= -1
1175 //          = log(x + sqrt((x+1)*(x-1)))
1176 // acosh(x) = nan                         if x < -1
1177 //
1178 // If x^2 will overflow, we approximate sqrt(x^2 - 1) == x and compute as
1179 // log(2*x) = log(2) + log(x).  (Note this works because negative x never
1180 // overflows; x < -1 simply yields nan.  This is quite different than asinh!)
Acosh(XlaOp x)1181 XlaOp Acosh(XlaOp x) {
1182   XlaBuilder* b = x.builder();
1183   return b->ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1184     TF_ASSIGN_OR_RETURN(auto shape, b->GetShape(x));
1185 
1186     auto one = ScalarLike(x, 1);
1187     auto neg_one = ScalarLike(x, -1);
1188     auto nan = FullLike(x, std::numeric_limits<float>::quiet_NaN());
1189 
1190     // return
1191     //
1192     //   nan                        if x < -1
1193     //   log(x) + log(2)            if x >= sqrt_max_value
1194     //   log(x + sqrt((x+1)*(x-1))) otherwise
1195     //
1196     // TODO(jlebar): For now, we ignore the question of overflow if x is a
1197     // complex type, because we don't yet have exhaustive tests for complex trig
1198     // functions.
1199     auto naive_result = Log(x + Sqrt((x + one) * (x - one)));
1200     if (primitive_util::IsComplexType(shape.element_type())) {
1201       return naive_result;
1202     }
1203     auto overflow_result = Log(x) + Log(ScalarLike(x, 2));
1204 
1205     auto sqrt_max_value = Sqrt(MaxFiniteValue(b, shape.element_type()));
1206     return Select(Lt(x, neg_one), nan,
1207                   Select(Ge(x, sqrt_max_value), overflow_result, naive_result));
1208   });
1209 }
1210 
1211 // asinh(x) = log(x + sqrt(x^2 + 1))
1212 //
1213 // If x^2 will overflow and x is positive, we can approximate x + sqrt(x^2 + 1)
1214 // as 2*x and return log(2) + log(x).
1215 //
1216 // If x is negative, the above would give us some trouble; we can't approximate
1217 // the result as x + abs(x) = 0!  But we're saved by the fact that asinh(-x) =
1218 // -asinh(x).
Asinh(XlaOp x)1219 XlaOp Asinh(XlaOp x) {
1220   XlaBuilder* b = x.builder();
1221   auto do_it = [&](XlaOp x) -> StatusOr<XlaOp> {
1222     TF_ASSIGN_OR_RETURN(auto shape, b->GetShape(x));
1223     auto one = ScalarLike(x, 1);
1224 
1225     // Let a = abs(x).  Compute
1226     //
1227     //   y = log(a + sqrt(a*a + 1))  if a < sqrt_max_value, or
1228     //   y = log(a) + log(2)         otherwise
1229     //
1230     // and then return
1231     //
1232     //   y * sign(x).
1233     //
1234     // TODO(jlebar): For now, we ignore the question of overflow if x is a
1235     // complex type, because we don't yet have exhaustive tests for complex trig
1236     // functions.
1237     if (primitive_util::IsComplexType(shape.element_type())) {
1238       return Log(x + Sqrt(x * x + one));
1239     }
1240     // For small x, sqrt(x**2 + 1) will evaluate to 1 due to floating point
1241     // arithmetic. However, we would like to retain the low order term of this,
1242     // which is around 0.5 * x**2 using a binomial expansion.
1243     // Let z = sqrt(a**2 + 1)
1244     // log(a + sqrt(a**2 + 1)) =
1245     // log((a + sqrt(a**2 + 1)) * (1 + sqrt(a**2 + 1)) / (1 + sqrt(a**2 + 1))) =
1246     // log((a + a**2 + 1 + a * z + z) / (1 + z)) =
1247     // log(1 + a + a**2 / (1 + z)) =
1248     // log(1 + a + a ** 2 / (1 + sqrt(a**2 + 1)))
1249     // This rewrite retains the lower order term.
1250     auto a = Abs(x);
1251     auto small_result = Log1p(a + a * a / (one + Sqrt(a * a + one)));
1252     auto naive_result = Log(a + Sqrt(a * a + one));
1253     auto overflow_result = Log(Abs(a)) + Log(ScalarLike(a, 2));
1254     auto sqrt_max_value = Sqrt(MaxFiniteValue(b, shape.element_type()));
1255     return Sign(x) * Select(Ge(a, sqrt_max_value), overflow_result,
1256                             Select(Le(a, one), small_result, naive_result));
1257   };
1258   // These upcasts are not strictly necessary on all platforms to get within our
1259   // error tolerances, so we could relax this if it ever mattered.
1260   return DoWithUpcastToF32(x, {BF16, F16}, [&](XlaOp x) {
1261     return b->ReportErrorOrReturn(do_it(x));
1262   });
1263 }
1264 
1265 // atanh(x) = 0.5 * log((1 + x) / (1 - x)) if abs(x) <= 1
1266 // atanh(x) = nan                          otherwise
Atanh(XlaOp x)1267 XlaOp Atanh(XlaOp x) {
1268   XlaBuilder* b = x.builder();
1269   auto do_it = [&](XlaOp x) -> StatusOr<XlaOp> {
1270     TF_ASSIGN_OR_RETURN(auto shape, b->GetShape(x));
1271     auto naive_result = (Log1p(x) - Log1p(-x)) * ScalarLike(x, 0.5);
1272 
1273     // TODO(jlebar): For now, we ignore the nan edge case for complex inputs,
1274     // because we don't yet have exhaustive tests for complex trig functions.
1275     if (primitive_util::IsComplexType(shape.element_type())) {
1276       return naive_result;
1277     }
1278 
1279     auto nan = FullLike(x, std::numeric_limits<float>::quiet_NaN());
1280     return Select(Gt(Abs(x), ScalarLike(x, 1)), nan, naive_result);
1281   };
1282   return DoWithUpcastToF32(x, {BF16}, [&](XlaOp x) {  //
1283     return b->ReportErrorOrReturn(do_it(x));
1284   });
1285 }
1286 
1287 // Cosh(x) = (e^x + e^-x) / 2
1288 //         = e^(x + log(1/2)) + e^(-x + log(1/2)).
1289 //
1290 // The second formulation avoids overflowing when e^x = inf but (e^x)/2 is not
1291 // inf.
1292 //
1293 // This incorrectly overflows to inf for two f32 input values, namely
1294 // +/-89.4159851, due to rounding error when computing x +/- log(1/2).  The
1295 // correct answer of 3.40281961e+38 (0x7f7fffec) is very close to max-float, so
1296 // we deem this acceptable.
Cosh(XlaOp x)1297 XlaOp Cosh(XlaOp x) {
1298   return DoWithUpcastToF32(x, {BF16, F16}, [](XlaOp x) {
1299     auto log_one_half = Log(ScalarLike(x, 0.5));
1300     return Exp(x + log_one_half) + Exp(-x + log_one_half);
1301   });
1302 }
1303 
1304 // Sinh(x) = (e^x - e^-x) / 2
1305 //         = e^(x + log(1/2)) - e^(-x + log(1/2)).
1306 //
1307 // The second formulation avoids overflowing when e^x = inf but (e^x)/2 is not
1308 // inf.
1309 //
1310 // This incorrectly overflows to +/-inf for two f32 input values, namely
1311 // +/-89.4159851, due to rounding error when computing x +/- log(1/2).  The
1312 // correct answer of 3.40281961e+38 (0x7f7fffec) is very close to max-float, so
1313 // we deem this acceptable.
Sinh(XlaOp x)1314 XlaOp Sinh(XlaOp x) {
1315   XlaBuilder* b = x.builder();
1316   auto do_it = [&](XlaOp x) -> StatusOr<XlaOp> {
1317     TF_ASSIGN_OR_RETURN(auto shape, b->GetShape(x));
1318     auto one_half = ScalarLike(x, 0.5);
1319     auto log_one_half = Log(ScalarLike(x, 0.5));
1320     auto large_sinh_result = Exp(x + log_one_half) - Exp(-x + log_one_half);
1321 
1322     if (primitive_util::IsComplexType(shape.element_type())) {
1323       return large_sinh_result;
1324     }
1325 
1326     // Here we use e^x = e^(x / 2) * e^(x / 2). This avoids overflow for large
1327     // values of x.
1328 
1329     // For smaller x, we get unwanted cancellations of e^x - e^-x, resulting in
1330     // 0.
1331     // Rewrite this to avoid that. We use expm1(x) because that preserves the
1332     // first order term of the taylor series of e^x.
1333     // (e^(x) - e^(-x)) / 2. =
1334     // (e^(x) - 1 + 1 - e^(-x)) / 2.
1335     // (expm1(x) + (e^(x) - 1) / e^x) / 2.
1336     // (expm1(x) + expm1(x) / (expm1(x) + 1)) / 2.
1337     auto expm1 = Expm1(x);
1338     auto one = ScalarLike(x, 1.);
1339     auto small_sinh_result = one_half * (expm1 + expm1 / (expm1 + one));
1340     return Select(Lt(Abs(x), one), small_sinh_result, large_sinh_result);
1341   };
1342   return DoWithUpcastToF32(x, {BF16, F16}, [&](XlaOp x) {
1343     return b->ReportErrorOrReturn(do_it(x));
1344   });
1345 }
1346 
MaybeConjugate(XlaOp x,bool conjugate)1347 XlaOp MaybeConjugate(XlaOp x, bool conjugate) {
1348   XlaBuilder* builder = x.builder();
1349   return builder->ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1350     TF_ASSIGN_OR_RETURN(Shape shape, builder->GetShape(x));
1351     auto perform_conj =
1352         primitive_util::IsComplexType(shape.element_type()) && conjugate;
1353     return perform_conj ? Conj(x) : x;
1354   });
1355 }
1356 
NextAfter(XlaOp from,XlaOp to)1357 XlaOp NextAfter(XlaOp from, XlaOp to) {
1358   auto builder = from.builder();
1359   return builder->ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1360     TF_ASSIGN_OR_RETURN(auto shape, builder->GetShape(from));
1361     int bitwidth = primitive_util::BitWidth(shape.element_type());
1362     auto int_type = primitive_util::UnsignedIntegralTypeForBitWidth(bitwidth);
1363     auto from_as_int = BitcastConvertType(from, int_type);
1364     auto to_as_int = BitcastConvertType(to, int_type);
1365 
1366     // The result is NaN if either "from" or "to" are NaN.
1367     auto from_is_nan = Ne(from, from);
1368     auto to_is_nan = Ne(to, to);
1369     auto nan_input = Or(from_is_nan, to_is_nan);
1370     auto result_for_nan =
1371         Broadcast(ScalarLike(from, std::numeric_limits<double>::quiet_NaN()),
1372                   shape.dimensions());
1373     result_for_nan = BitcastConvertType(result_for_nan, int_type);
1374 
1375     // The sign bit is the MSB.
1376     const int64 sign_mask = int64{1} << (bitwidth - 1);
1377     // Discard the sign bit to make the result non-negative.
1378     auto from_abs = And(from_as_int, ScalarLike(from_as_int, ~sign_mask));
1379     auto to_abs = And(to_as_int, ScalarLike(to_as_int, ~sign_mask));
1380 
1381     // When both "from" and "to" are equal, the result is "to".
1382     // N.B. It would not make a difference if we chose the result to be "from".
1383     auto from_and_to_are_equal = Eq(from_as_int, to_as_int);
1384     auto result_for_equal = to_as_int;
1385 
1386     // When both "from" and "to" are both 0, the result is "to". This ensures we
1387     // get a zero signed like "to".
1388     auto from_is_zero = Eq(from_abs, ZerosLike(from_abs));
1389     auto to_is_zero = Eq(to_abs, ZerosLike(to_abs));
1390     auto result_for_both_zero = to_as_int;
1391 
1392     auto from_sign = And(from_as_int, ScalarLike(from_as_int, sign_mask));
1393     auto to_sign = And(to_as_int, ScalarLike(to_as_int, sign_mask));
1394 
1395     // If from == 0 && to != 0, we need to return the smallest subnormal number
1396     // signed like "to".
1397     auto result_for_from_zero_to_non_zero =
1398         Or(to_sign, ScalarLike(from_as_int, 1));
1399 
1400     // If the sign of "from" and "to" disagree:
1401     // - we need to make the magnitude of "from" smaller so that it is closer to
1402     //   zero.
1403     //
1404     // Otherwise the signs agree:
1405     // - "from" with a magnitude larger than "to" means we need to make the
1406     //   magnitude smaller.
1407     // - "from" with a magnitude smaller than "to" means we need to make the
1408     //   magnitude larger.
1409     // - "from" with the same magnitude and sign as "to" has already been
1410     //   handled.
1411     auto signs_disagree = Ne(from_sign, to_sign);
1412     auto from_magnitude_larger_than_to = Gt(from_abs, to_abs);
1413     auto result_has_smaller_magnitude =
1414         Or(from_magnitude_larger_than_to, signs_disagree);
1415     auto magnitude_adjustment =
1416         Select(result_has_smaller_magnitude,
1417                Broadcast(ScalarLike(from_as_int, -1), shape.dimensions()),
1418                Broadcast(ScalarLike(from_as_int, 1), shape.dimensions()));
1419     auto result = Add(from_as_int, magnitude_adjustment);
1420     // Handle from == ±0.
1421     result = Select(from_is_zero,
1422                     Select(to_is_zero, result_for_both_zero,
1423                            result_for_from_zero_to_non_zero),
1424                     result);
1425     // Handle from == to.
1426     result = Select(from_and_to_are_equal, result_for_equal, result);
1427     // Handle isnan(from) || isnan(to).
1428     result = Select(nan_input, result_for_nan, result);
1429 
1430     // Cast back to the original type.
1431     return BitcastConvertType(result, shape.element_type());
1432   });
1433 }
1434 
1435 // Computes an approximation to the modified Bessel function of the first kind,
1436 // zeroth order.
1437 // The following implementation follows Cephes' F32 and F64 implementation of
1438 // i0e.
I0eImpl32(XlaOp x)1439 static XlaOp I0eImpl32(XlaOp x) {
1440   static const std::array<float, 18> kI0eCoeffsA{
1441       -1.30002500998624804212E-8f, 6.04699502254191894932E-8f,
1442       -2.67079385394061173391E-7f, 1.11738753912010371815E-6f,
1443       -4.41673835845875056359E-6f, 1.64484480707288970893E-5f,
1444       -5.75419501008210370398E-5f, 1.88502885095841655729E-4f,
1445       -5.76375574538582365885E-4f, 1.63947561694133579842E-3f,
1446       -4.32430999505057594430E-3f, 1.05464603945949983183E-2f,
1447       -2.37374148058994688156E-2f, 4.93052842396707084878E-2f,
1448       -9.49010970480476444210E-2f, 1.71620901522208775349E-1f,
1449       -3.04682672343198398683E-1f, 6.76795274409476084995E-1f};
1450 
1451   static const std::array<float, 7> kI0eCoeffsB{
1452       3.39623202570838634515E-9f, 2.26666899049817806459E-8f,
1453       2.04891858946906374183E-7f, 2.89137052083475648297E-6f,
1454       6.88975834691682398426E-5f, 3.36911647825569408990E-3f,
1455       8.04490411014108831608E-1f};
1456 
1457   x = Abs(x);
1458   auto half = xla::ScalarLike(x, 0.5);
1459   auto two = xla::ScalarLike(x, 2.0);
1460   auto thirty_two = xla::ScalarLike(x, 32.0);
1461   auto result_le_8 =
1462       EvaluateChebyshevPolynomial<float>(half * x - two, kI0eCoeffsA);
1463   auto result_gt_8 =
1464       EvaluateChebyshevPolynomial<float>(thirty_two / x - two, kI0eCoeffsB) /
1465       Sqrt(x);
1466   return Select(Le(x, xla::ScalarLike(x, 8.0)), result_le_8, result_gt_8);
1467 }
1468 
I0eImpl64(XlaOp x)1469 static XlaOp I0eImpl64(XlaOp x) {
1470   static const std::array<double, 30> kI0eCoeffsA{
1471       -4.41534164647933937950E-18, 3.33079451882223809783E-17,
1472       -2.43127984654795469359E-16, 1.71539128555513303061E-15,
1473       -1.16853328779934516808E-14, 7.67618549860493561688E-14,
1474       -4.85644678311192946090E-13, 2.95505266312963983461E-12,
1475       -1.72682629144155570723E-11, 9.67580903537323691224E-11,
1476       -5.18979560163526290666E-10, 2.65982372468238665035E-9,
1477       -1.30002500998624804212E-8,  6.04699502254191894932E-8,
1478       -2.67079385394061173391E-7,  1.11738753912010371815E-6,
1479       -4.41673835845875056359E-6,  1.64484480707288970893E-5,
1480       -5.75419501008210370398E-5,  1.88502885095841655729E-4,
1481       -5.76375574538582365885E-4,  1.63947561694133579842E-3,
1482       -4.32430999505057594430E-3,  1.05464603945949983183E-2,
1483       -2.37374148058994688156E-2,  4.93052842396707084878E-2,
1484       -9.49010970480476444210E-2,  1.71620901522208775349E-1,
1485       -3.04682672343198398683E-1,  6.76795274409476084995E-1};
1486 
1487   static const std::array<double, 25> kI0eCoeffsB{
1488       -7.23318048787475395456E-18, -4.83050448594418207126E-18,
1489       4.46562142029675999901E-17,  3.46122286769746109310E-17,
1490       -2.82762398051658348494E-16, -3.42548561967721913462E-16,
1491       1.77256013305652638360E-15,  3.81168066935262242075E-15,
1492       -9.55484669882830764870E-15, -4.15056934728722208663E-14,
1493       1.54008621752140982691E-14,  3.85277838274214270114E-13,
1494       7.18012445138366623367E-13,  -1.79417853150680611778E-12,
1495       -1.32158118404477131188E-11, -3.14991652796324136454E-11,
1496       1.18891471078464383424E-11,  4.94060238822496958910E-10,
1497       3.39623202570838634515E-9,   2.26666899049817806459E-8,
1498       2.04891858946906374183E-7,   2.89137052083475648297E-6,
1499       6.88975834691682398426E-5,   3.36911647825569408990E-3,
1500       8.04490411014108831608E-1};
1501 
1502   x = Abs(x);
1503   auto half = xla::ScalarLike(x, 0.5);
1504   auto two = xla::ScalarLike(x, 2.0);
1505   auto thirty_two = xla::ScalarLike(x, 32.0);
1506   auto result_le_8 =
1507       EvaluateChebyshevPolynomial<double>(half * x - two, kI0eCoeffsA);
1508   auto result_gt_8 =
1509       EvaluateChebyshevPolynomial<double>(thirty_two / x - two, kI0eCoeffsB) /
1510       Sqrt(x);
1511   return Select(Le(x, xla::ScalarLike(x, 8.0)), result_le_8, result_gt_8);
1512 }
1513 
BesselI0e(XlaOp x)1514 XlaOp BesselI0e(XlaOp x) {
1515   auto& b = *x.builder();
1516   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1517     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("BesselI0e", x));
1518     TF_ASSIGN_OR_RETURN(auto shape, b.GetShape(x));
1519     if (shape.element_type() == F64) {
1520       return I0eImpl64(x);
1521     }
1522     // I0eF32Impl don't have enough precision when run with bf16 intermediates
1523     // (not surprising!), so upcast to f32 in this case.
1524     return DoWithUpcastToF32(x, {BF16, F16},
1525                              [](XlaOp x) { return I0eImpl32(x); });
1526   });
1527 }
1528 
1529 // Computes an approximation to the modified Bessel function of the first kind,
1530 // first order.
1531 // The following implementation follows Cephes' F32 and F64 implementation of
1532 // i1e.
1533 
I1eImpl32(XlaOp x)1534 static XlaOp I1eImpl32(XlaOp x) {
1535   static const std::array<float, 17> kI1eCoeffsA{
1536       9.38153738649577178388E-9f, -4.44505912879632808065E-8f,
1537       2.00329475355213526229E-7f, -8.56872026469545474066E-7f,
1538       3.47025130813767847674E-6f, -1.32731636560394358279E-5f,
1539       4.78156510755005422638E-5f, -1.61760815825896745588E-4f,
1540       5.12285956168575772895E-4f, -1.51357245063125314899E-3f,
1541       4.15642294431288815669E-3f, -1.05640848946261981558E-2f,
1542       2.47264490306265168283E-2f, -5.29459812080949914269E-2f,
1543       1.02643658689847095384E-1f, -1.76416518357834055153E-1f,
1544       2.52587186443633654823E-1f};
1545 
1546   static const std::array<float, 7> kI1eCoeffsB{
1547       -3.83538038596423702205E-9f, -2.63146884688951950684E-8f,
1548       -2.51223623787020892529E-7f, -3.88256480887769039346E-6f,
1549       -1.10588938762623716291E-4f, -9.76109749136146840777E-3f,
1550       7.78576235018280120474E-1f};
1551   XlaOp z = Abs(x);
1552   auto half = xla::ScalarLike(x, 0.5);
1553   auto two = xla::ScalarLike(x, 2.0);
1554   auto thirty_two = xla::ScalarLike(x, 32.0);
1555   auto result_le_8 =
1556       z * EvaluateChebyshevPolynomial<float>(half * z - two, kI1eCoeffsA);
1557   auto result_gt_8 =
1558       EvaluateChebyshevPolynomial<float>(thirty_two / z - two, kI1eCoeffsB) /
1559       Sqrt(z);
1560   return Sign(x) *
1561          Select(Le(z, xla::ScalarLike(x, 8.0)), result_le_8, result_gt_8);
1562 }
1563 
I1eImpl64(XlaOp x)1564 static XlaOp I1eImpl64(XlaOp x) {
1565   static const std::array<double, 29> kI1eCoeffsA{
1566       2.77791411276104639959E-18, -2.11142121435816608115E-17,
1567       1.55363195773620046921E-16, -1.10559694773538630805E-15,
1568       7.60068429473540693410E-15, -5.04218550472791168711E-14,
1569       3.22379336594557470981E-13, -1.98397439776494371520E-12,
1570       1.17361862988909016308E-11, -6.66348972350202774223E-11,
1571       3.62559028155211703701E-10, -1.88724975172282928790E-9,
1572       9.38153738649577178388E-9,  -4.44505912879632808065E-8,
1573       2.00329475355213526229E-7,  -8.56872026469545474066E-7,
1574       3.47025130813767847674E-6,  -1.32731636560394358279E-5,
1575       4.78156510755005422638E-5,  -1.61760815825896745588E-4,
1576       5.12285956168575772895E-4,  -1.51357245063125314899E-3,
1577       4.15642294431288815669E-3,  -1.05640848946261981558E-2,
1578       2.47264490306265168283E-2,  -5.29459812080949914269E-2,
1579       1.02643658689847095384E-1,  -1.76416518357834055153E-1,
1580       2.52587186443633654823E-1};
1581 
1582   static const std::array<double, 25> kI1eCoeffsB{
1583       7.51729631084210481353E-18,  4.41434832307170791151E-18,
1584       -4.65030536848935832153E-17, -3.20952592199342395980E-17,
1585       2.96262899764595013876E-16,  3.30820231092092828324E-16,
1586       -1.88035477551078244854E-15, -3.81440307243700780478E-15,
1587       1.04202769841288027642E-14,  4.27244001671195135429E-14,
1588       -2.10154184277266431302E-14, -4.08355111109219731823E-13,
1589       -7.19855177624590851209E-13, 2.03562854414708950722E-12,
1590       1.41258074366137813316E-11,  3.25260358301548823856E-11,
1591       -1.89749581235054123450E-11, -5.58974346219658380687E-10,
1592       -3.83538038596423702205E-9,  -2.63146884688951950684E-8,
1593       -2.51223623787020892529E-7,  -3.88256480887769039346E-6,
1594       -1.10588938762623716291E-4,  -9.76109749136146840777E-3,
1595       7.78576235018280120474E-1};
1596 
1597   XlaOp z = Abs(x);
1598   auto half = xla::ScalarLike(x, 0.5);
1599   auto two = xla::ScalarLike(x, 2.0);
1600   auto thirty_two = xla::ScalarLike(x, 32.0);
1601   auto result_le_8 =
1602       z * EvaluateChebyshevPolynomial<double>(half * z - two, kI1eCoeffsA);
1603   auto result_gt_8 =
1604       EvaluateChebyshevPolynomial<double>(thirty_two / z - two, kI1eCoeffsB) /
1605       Sqrt(z);
1606   return Sign(x) *
1607          Select(Le(z, xla::ScalarLike(x, 8.0)), result_le_8, result_gt_8);
1608 }
1609 
BesselI1e(XlaOp x)1610 XlaOp BesselI1e(XlaOp x) {
1611   auto& b = *x.builder();
1612   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1613     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("BesselI1e", x));
1614     TF_ASSIGN_OR_RETURN(auto shape, b.GetShape(x));
1615     if (shape.element_type() == F64) {
1616       return I1eImpl64(x);
1617     }
1618     // I1eF32Impl don't have enough precision when run with bf16 intermediates
1619     // (not surprising!), so upcast to f32 in this case.
1620     return DoWithUpcastToF32(x, {BF16, F16},
1621                              [](XlaOp x) { return I1eImpl32(x); });
1622   });
1623 }
1624 
1625 // I J Thompson and A R Barnett. 1986. Coulomb and Bessel functions of complex
1626 // arguments and order. J. Comput. Phys. 64, 2 (June 1986), 490-509.
1627 // DOI=http://dx.doi.org/10.1016/0021-9991(86)90046-X
LentzThompsonBarnettAlgorithm(int64 num_iterations,double small,double threshold,const ForEachIndexBodyFunction & nth_partial_numerator,const ForEachIndexBodyFunction & nth_partial_denominator,absl::Span<const XlaOp> inputs,absl::string_view name)1628 static XlaOp LentzThompsonBarnettAlgorithm(
1629     int64 num_iterations, double small, double threshold,
1630     const ForEachIndexBodyFunction& nth_partial_numerator,
1631     const ForEachIndexBodyFunction& nth_partial_denominator,
1632     absl::Span<const XlaOp> inputs, absl::string_view name) {
1633   auto& b = *inputs.front().builder();
1634   return b.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1635     TF_RET_CHECK(num_iterations < INT32_MAX);
1636 
1637     enum {
1638       // Position in the evaluation.
1639       kIterationIdx,
1640       // Whether or not we have reached the desired tolerance.
1641       kValuesUnconvergedIdx,
1642       // Ratio between nth canonical numerator and the nth-1 canonical
1643       // numerator.
1644       kCIdx,
1645       // Ratio between nth-1 canonical denominator and the nth canonical
1646       // denominator.
1647       kDIdx,
1648       // Computed approximant in the evaluation.
1649       kHIdx,
1650       // Inputs follow all of the other state.
1651       kFirstInputIdx,
1652     };
1653     auto while_cond_fn = [num_iterations](
1654                              absl::Span<const XlaOp> values,
1655                              XlaBuilder* cond_builder) -> StatusOr<XlaOp> {
1656       auto iteration = values[kIterationIdx];
1657       auto iterations_remain_cond =
1658           Lt(iteration, ScalarLike(iteration, num_iterations));
1659       auto values_unconverged_cond = values[kValuesUnconvergedIdx];
1660       return And(iterations_remain_cond, values_unconverged_cond);
1661     };
1662 
1663     auto while_body_fn =
1664         [small, threshold, &nth_partial_numerator, &nth_partial_denominator](
1665             absl::Span<const XlaOp> values,
1666             XlaBuilder* body_builder) -> StatusOr<std::vector<XlaOp>> {
1667       XlaOp iteration = values[kIterationIdx];
1668 
1669       TF_ASSIGN_OR_RETURN(
1670           std::vector<XlaOp> partial_numerator,
1671           nth_partial_numerator(iteration, values.subspan(kFirstInputIdx),
1672                                 body_builder));
1673       TF_RET_CHECK(partial_numerator.size() == 1);
1674 
1675       TF_ASSIGN_OR_RETURN(
1676           std::vector<XlaOp> partial_denominator,
1677           nth_partial_denominator(iteration, values.subspan(kFirstInputIdx),
1678                                   body_builder));
1679       TF_RET_CHECK(partial_denominator.size() == 1);
1680 
1681       auto c = partial_denominator[0] + partial_numerator[0] / values[kCIdx];
1682       auto small_constant = FullLike(c, small);
1683       c = Select(Lt(Abs(c), small_constant), small_constant, c);
1684 
1685       auto d = partial_denominator[0] + partial_numerator[0] * values[kDIdx];
1686       d = Select(Lt(Abs(d), small_constant), small_constant, d);
1687 
1688       d = Reciprocal(d);
1689 
1690       auto delta = c * d;
1691       auto h = values[kHIdx] * delta;
1692 
1693       std::vector<XlaOp> updated_values(values.size());
1694       updated_values[kIterationIdx] = Add(iteration, ScalarLike(iteration, 1));
1695       updated_values[kCIdx] = c;
1696       updated_values[kDIdx] = d;
1697       updated_values[kHIdx] = h;
1698       std::copy(values.begin() + kFirstInputIdx, values.end(),
1699                 updated_values.begin() + kFirstInputIdx);
1700 
1701       // If any values are greater than the tolerance, we have not converged.
1702       auto tolerance_comparison =
1703           Ge(Abs(Sub(delta, FullLike(delta, 1.0))), FullLike(delta, threshold));
1704       updated_values[kValuesUnconvergedIdx] =
1705           ReduceAll(tolerance_comparison, ConstantR0<bool>(body_builder, false),
1706                     CreateScalarOrComputation(PRED, body_builder));
1707       return updated_values;
1708     };
1709 
1710     TF_ASSIGN_OR_RETURN(std::vector<XlaOp> partial_denominator,
1711                         nth_partial_denominator(Zero(&b, U32), inputs, &b));
1712     TF_RET_CHECK(partial_denominator.size() == 1);
1713     auto h = partial_denominator[0];
1714     auto small_constant = FullLike(h, small);
1715     h = Select(Lt(Abs(h), small_constant), small_constant, h);
1716 
1717     std::vector<XlaOp> values(kFirstInputIdx + inputs.size());
1718     values[kIterationIdx] = One(&b, U32);
1719     values[kValuesUnconvergedIdx] = ConstantR0<bool>(&b, true);
1720     values[kCIdx] = h;
1721     values[kDIdx] = FullLike(h, 0.0);
1722     values[kHIdx] = h;
1723     std::copy(inputs.begin(), inputs.end(), values.begin() + kFirstInputIdx);
1724     TF_ASSIGN_OR_RETURN(values, WhileLoopHelper(while_cond_fn, while_body_fn,
1725                                                 values, name, &b));
1726     return values[kHIdx];
1727   });
1728 }
1729 
RegularizedIncompleteBeta(XlaOp a,XlaOp b,XlaOp x)1730 XlaOp RegularizedIncompleteBeta(XlaOp a, XlaOp b, XlaOp x) {
1731   auto& builder = *x.builder();
1732   return builder.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1733     TF_ASSIGN_OR_RETURN(Shape shape, builder.GetShape(a));
1734     TF_ASSIGN_OR_RETURN(Shape b_shape, builder.GetShape(b));
1735     TF_ASSIGN_OR_RETURN(Shape x_shape, builder.GetShape(x));
1736     if (b_shape.element_type() != shape.element_type() ||
1737         x_shape.element_type() != shape.element_type()) {
1738       return InvalidArgument(
1739           "Operands to RegularizedIncompleteBeta must have identical types, "
1740           "got shapes %s, %s, and %s",
1741           shape.ToString(), b_shape.ToString(), x_shape.ToString());
1742     }
1743     if (!primitive_util::IsFloatingPointType(shape.element_type())) {
1744       return InvalidArgument(
1745           "Operands to RegularizedIncompleteBeta must be real-valued "
1746           "floating-point, but got %s",
1747           PrimitiveType_Name(shape.element_type()));
1748     }
1749     PrimitiveType element_type = shape.element_type();
1750     if (element_type == F16 || element_type == BF16) {
1751       element_type = F32;
1752       a = ConvertElementType(a, F32);
1753       b = ConvertElementType(b, F32);
1754       x = ConvertElementType(x, F32);
1755     }
1756 
1757     // The partial numerator for the incomplete beta function is given
1758     // here: http://dlmf.nist.gov/8.17.E23 Note that there is a special
1759     // case: the partial numerator for the first iteration is one.
1760     auto NthPartialBetaincNumerator =
1761         [&](XlaOp iteration, absl::Span<const XlaOp> inputs,
1762             XlaBuilder* builder) -> StatusOr<std::vector<XlaOp>> {
1763       auto a = inputs[0];
1764       auto b = inputs[1];
1765       auto x = inputs[2];
1766       auto iteration_bcast = Broadcast(iteration, shape.dimensions());
1767       auto iteration_is_even =
1768           Eq(iteration_bcast % FullLike(iteration_bcast, 2),
1769              FullLike(iteration_bcast, 0));
1770       auto iteration_is_one = Eq(iteration_bcast, FullLike(iteration_bcast, 1));
1771       auto iteration_minus_one = iteration_bcast - FullLike(iteration_bcast, 1);
1772       auto m = iteration_minus_one / FullLike(iteration_minus_one, 2);
1773       m = ConvertElementType(m, element_type);
1774       auto one = FullLike(a, 1.0);
1775       auto two = FullLike(a, 2.0);
1776       // Partial numerator terms.
1777       auto even_numerator =
1778           -(a + m) * (a + b + m) * x / ((a + two * m) * (a + two * m + one));
1779       auto odd_numerator =
1780           m * (b - m) * x / ((a + two * m - one) * (a + two * m));
1781       auto one_numerator = ScalarLike(x, 1.0);
1782       auto numerator = Select(iteration_is_even, even_numerator, odd_numerator);
1783       return std::vector<XlaOp>{
1784           Select(iteration_is_one, one_numerator, numerator)};
1785     };
1786 
1787     auto NthPartialBetaincDenominator =
1788         [&shape](XlaOp iteration, absl::Span<const XlaOp> inputs,
1789                  XlaBuilder* builder) -> StatusOr<std::vector<XlaOp>> {
1790       auto x = inputs[2];
1791       auto iteration_bcast = Broadcast(iteration, shape.dimensions());
1792       return std::vector<XlaOp>{
1793           Select(Eq(iteration_bcast, ScalarLike(iteration_bcast, 0)),
1794                  ScalarLike(x, 0.0), ScalarLike(x, 1.0))};
1795     };
1796 
1797     // Determine if the inputs are out of range.
1798     auto result_is_nan =
1799         Or(Or(Or(Le(a, ScalarLike(a, 0.0)), Le(b, ScalarLike(b, 0.0))),
1800               Lt(x, ScalarLike(x, 0.0))),
1801            Gt(x, ScalarLike(x, 1.0)));
1802 
1803     // The continued fraction will converge rapidly when x < (a+1)/(a+b+2)
1804     // as per: http://dlmf.nist.gov/8.17.E23
1805     //
1806     // Otherwise, we can rewrite using the symmetry relation as per:
1807     // http://dlmf.nist.gov/8.17.E4
1808     auto converges_rapidly =
1809         Lt(x, (a + FullLike(a, 1.0)) / (a + b + FullLike(b, 2.0)));
1810     auto a_orig = a;
1811     a = Select(converges_rapidly, a, b);
1812     b = Select(converges_rapidly, b, a_orig);
1813     x = Select(converges_rapidly, x, Sub(FullLike(x, 1.0), x));
1814 
1815     XlaOp continued_fraction;
1816 
1817     // Thresholds and iteration counts taken from Cephes.
1818     if (element_type == F32) {
1819       continued_fraction = LentzThompsonBarnettAlgorithm(
1820           /*num_iterations=*/200,
1821           /*small=*/std::numeric_limits<float>::epsilon() / 2.0f,
1822           /*threshold=*/std::numeric_limits<float>::epsilon() / 2.0f,
1823           /*nth_partial_numerator=*/NthPartialBetaincNumerator,
1824           /*nth_partial_denominator=*/NthPartialBetaincDenominator, {a, b, x},
1825           "Betainc");
1826     } else {
1827       TF_RET_CHECK(element_type == F64);
1828       continued_fraction = LentzThompsonBarnettAlgorithm(
1829           /*num_iterations=*/600,
1830           /*small=*/std::numeric_limits<double>::epsilon() / 2.0f,
1831           /*threshold=*/std::numeric_limits<double>::epsilon() / 2.0f,
1832           /*nth_partial_numerator=*/NthPartialBetaincNumerator,
1833           /*nth_partial_denominator=*/NthPartialBetaincDenominator, {a, b, x},
1834           "Betainc");
1835     }
1836 
1837     // We want to compute the regularized complete beta function so we need to
1838     // combine the continued fraction with a few more terms as well as dividing
1839     // it by Beta(a, b). To avoid overflow, we compute in the log domain.
1840     // See http://dlmf.nist.gov/8.17.E22 for an easier to read version of this
1841     // formula.
1842     auto lbeta = Lbeta(a, b);
1843     auto result =
1844         continued_fraction * Exp(Log(x) * a + Log1p(-x) * b - lbeta) / a;
1845     result = Select(result_is_nan, NanValue(&builder, element_type), result);
1846 
1847     // We have an additional fixup to do if we are taking advantage of the
1848     // symmetry relation.
1849     auto out =
1850         Select(converges_rapidly, result, Sub(FullLike(result, 1.0), result));
1851     return shape.element_type() == element_type
1852                ? out
1853                : ConvertElementType(out, shape.element_type());
1854   });
1855 }
1856 
Polygamma(XlaOp n,XlaOp x)1857 XlaOp Polygamma(XlaOp n, XlaOp x) {
1858   auto& builder = *x.builder();
1859   auto doit = [](XlaOp n, XlaOp x, PrimitiveType type) -> XlaOp {
1860     XlaOp n_plus_one = n + ScalarLike(n, 1.);
1861     XlaOp sign =
1862         (ScalarLike(n, 2.) * Rem(n, ScalarLike(n, 2.)) - ScalarLike(n, 1.));
1863 
1864     const double nan = std::numeric_limits<double>::quiet_NaN();
1865 
1866     XlaOp output = Select(Eq(n, ScalarLike(n, 0.)), Digamma(x),
1867                           sign * Exp(Lgamma(n_plus_one)) * Zeta(n_plus_one, x));
1868     // Check that n is a natural number.
1869     output = Select(Or(Ne(n, Floor(n)), Lt(n, ScalarLike(n, 0.))),
1870                     ScalarLike(n, nan), output);
1871     return output;
1872   };
1873   return builder.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1874     TF_ASSIGN_OR_RETURN(auto n_shape, builder.GetShape(n));
1875     TF_ASSIGN_OR_RETURN(auto x_shape, builder.GetShape(x));
1876     if (n_shape != x_shape) {
1877       return InvalidArgument(
1878           "Arguments to Polygamma must have equal shapes and types; "
1879           "got %s and %s",
1880           n_shape.ToString(), x_shape.ToString());
1881     }
1882     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("Zeta", x));
1883     bool needs_upcast =
1884         n_shape.element_type() == F16 || x_shape.element_type() == BF16;
1885 
1886     if (needs_upcast) {
1887       n = ConvertElementType(n, F32);
1888       x = ConvertElementType(x, F32);
1889     }
1890     XlaOp result = doit(n, x, n_shape.element_type());
1891     if (needs_upcast) {
1892       result = ConvertElementType(result, n_shape.element_type());
1893     }
1894     return result;
1895   });
1896 }
1897 
Zeta(XlaOp x,XlaOp q)1898 XlaOp Zeta(XlaOp x, XlaOp q) {
1899   auto& builder = *x.builder();
1900   auto doit = [&builder](XlaOp x, XlaOp q, PrimitiveType type) -> XlaOp {
1901     // (2k) ! / B_{2k}, where B_{2k} are the Bernoulli numbers.
1902     // These are ordered in reverse.
1903     static const std::array<double, 12> kZetaCoeffs{
1904         -7.1661652561756670113e18,
1905         1.8152105401943546773e17,
1906         -4.5979787224074726105e15,
1907         1.1646782814350067249e14,
1908         -2.950130727918164224e12,
1909         7.47242496e10,
1910         -1.8924375803183791606e9,
1911         47900160.0,
1912         -1209600.0,
1913         30240.0,
1914         -720.0,
1915         12.0,
1916     };
1917 
1918     // For speed we'll always use 9 iterations for the initial series estimate,
1919     // and a 12 term expansion for the Euler-Maclaurin formula.
1920 
1921     XlaOp a = q;
1922     XlaOp neg_power = ScalarLike(a, 0.);
1923     XlaOp initial_sum = Pow(q, Neg(x));
1924     for (int i = 0; i < 9; ++i) {
1925       a = a + ScalarLike(a, 1.);
1926       neg_power = Pow(a, Neg(x));
1927       initial_sum = initial_sum + neg_power;
1928     }
1929     a = a + ScalarLike(a, 1.);
1930     neg_power = Pow(a, Neg(x));
1931     XlaOp s = initial_sum + neg_power * a / (x - ScalarLike(a, 1.));
1932     XlaOp a_inverse_square = Reciprocal(Square(a));
1933     XlaOp horner_sum = ScalarLike(a, 0.);
1934     XlaOp factor = ScalarLike(a, 1.);
1935     // Use Horner's rule for this.
1936     // Note this differs from Cephes which does a 'naive' polynomial evaluation.
1937     // Using Horner's rule allows to avoid some NaN's and Infs from happening,
1938     // resulting in more numerically stable code.
1939     for (int i = 0; i < 11; ++i) {
1940       factor =
1941           (x - ScalarLike(x, 22 - 2 * i)) * (x - ScalarLike(x, 21 - 2 * i));
1942       horner_sum = factor * a_inverse_square *
1943                    (horner_sum + ScalarLike(a, 1. / kZetaCoeffs[i]));
1944     }
1945     s = s + neg_power *
1946                 (ScalarLike(neg_power, 0.5) +
1947                  x / a * (ScalarLike(a, 1. / kZetaCoeffs[11]) + horner_sum));
1948 
1949     const double nan = std::numeric_limits<double>::quiet_NaN();
1950     const double inf = std::numeric_limits<double>::infinity();
1951     // Use the initial zeta sum without the correction term coming
1952     // from Euler-Maclaurin if it is accurate enough.
1953     XlaOp output =
1954         Select(Lt(Abs(neg_power), Abs(initial_sum) * Epsilon(&builder, type)),
1955                initial_sum, s);
1956     // This is the harmonic series.
1957     output = Select(Eq(x, ScalarLike(x, 1.)), ScalarLike(x, inf), output);
1958     // Function is not defined for x < 1.
1959     output = Select(Lt(x, ScalarLike(x, 1.)), ScalarLike(x, nan), output);
1960     // If q <= 0, then when q is an integer or x is not an integer, this is
1961     // NaN.
1962     XlaOp domain_error = And(Le(q, ScalarLike(x, 0.)), Ne(x, Floor(x)));
1963     XlaOp negative_integer_q = And(Le(q, ScalarLike(x, 0.)), Eq(q, Floor(q)));
1964     output = Select(negative_integer_q, ScalarLike(x, inf), output);
1965     output = Select(domain_error, ScalarLike(x, nan), output);
1966     return output;
1967   };
1968   return builder.ReportErrorOrReturn([&]() -> StatusOr<XlaOp> {
1969     TF_ASSIGN_OR_RETURN(auto x_shape, builder.GetShape(x));
1970     TF_ASSIGN_OR_RETURN(auto q_shape, builder.GetShape(q));
1971     if (x_shape != q_shape) {
1972       return InvalidArgument(
1973           "Arguments to Zeta must have equal shapes and types; got %s and %s",
1974           x_shape.ToString(), q_shape.ToString());
1975     }
1976     TF_RETURN_IF_ERROR(EnsureOperandIsRealFp("Zeta", x));
1977     bool needs_upcast =
1978         x_shape.element_type() == F16 || x_shape.element_type() == BF16;
1979 
1980     if (needs_upcast) {
1981       x = ConvertElementType(x, F32);
1982       q = ConvertElementType(q, F32);
1983     }
1984     XlaOp result = doit(x, q, x_shape.element_type());
1985     if (needs_upcast) {
1986       result = ConvertElementType(result, x_shape.element_type());
1987     }
1988     return result;
1989   });
1990 }
1991 
1992 }  // namespace xla
1993