1 | #ifdef __OPENCL_VERSION__ |
---|
2 | # define USE_OPENCL |
---|
3 | #elif defined(_OPENMP) |
---|
4 | # define USE_OPENMP |
---|
5 | #endif |
---|
6 | |
---|
7 | // If opencl is not available, then we are compiling a C function |
---|
8 | // Note: if using a C++ compiler, then define kernel as extern "C" |
---|
9 | #ifdef USE_OPENCL |
---|
10 | typedef int int32_t; |
---|
11 | # if defined(USE_SINCOS) |
---|
12 | # define SINCOS(angle,svar,cvar) svar=sincos(angle,&cvar) |
---|
13 | # else |
---|
14 | # define SINCOS(angle,svar,cvar) do {const double _t_=angle; svar=sin(_t_);cvar=cos(_t_);} while (0) |
---|
15 | # endif |
---|
16 | // Intel CPU on Mac gives strange values for erf(); on the verified |
---|
17 | // platforms (intel, nvidia, amd), the cephes erf() is significantly |
---|
18 | // faster than that available in the native OpenCL. |
---|
19 | #define NEED_ERF |
---|
20 | // OpenCL only has type generic math |
---|
21 | #define expf exp |
---|
22 | #ifndef NEED_ERF |
---|
23 | # define erff erf |
---|
24 | # define erfcf erfc |
---|
25 | #endif |
---|
26 | #else // !USE_OPENCL |
---|
27 | // Use SAS_DOUBLE to force the use of double even for float kernels |
---|
28 | # define SAS_DOUBLE dou ## ble |
---|
29 | # ifdef __cplusplus |
---|
30 | #include <cstdio> |
---|
31 | #include <cmath> |
---|
32 | using namespace std; |
---|
33 | #if defined(_MSC_VER) |
---|
34 | #include <limits> |
---|
35 | #include <float.h> |
---|
36 | #define kernel extern "C" __declspec( dllexport ) |
---|
37 | inline double trunc(double x) { return x>=0?floor(x):-floor(-x); } |
---|
38 | inline double fmin(double x, double y) { return x>y ? y : x; } |
---|
39 | inline double fmax(double x, double y) { return x<y ? y : x; } |
---|
40 | #define isnan(x) _isnan(x) |
---|
41 | #define isinf(x) (!_finite(x)) |
---|
42 | #define isfinite(x) _finite(x) |
---|
43 | #define NAN (std::numeric_limits<double>::quiet_NaN()) // non-signalling NaN |
---|
44 | #define INFINITY (std::numeric_limits<double>::infinity()) |
---|
45 | #define NEED_ERF |
---|
46 | #define NEED_EXPM1 |
---|
47 | #define NEED_TGAMMA |
---|
48 | #else |
---|
49 | #define kernel extern "C" |
---|
50 | #include <cstdint> |
---|
51 | #endif |
---|
52 | inline void SINCOS(double angle, double &svar, double &cvar) { svar=sin(angle); cvar=cos(angle); } |
---|
53 | # else // !__cplusplus |
---|
54 | #include <inttypes.h> // C99 guarantees that int32_t types is here |
---|
55 | #include <stdio.h> |
---|
56 | #if defined(__TINYC__) |
---|
57 | typedef int int32_t; |
---|
58 | #include <math.h> |
---|
59 | // TODO: check isnan is correct |
---|
60 | inline double _isnan(double x) { return x != x; } // hope this doesn't optimize away! |
---|
61 | #undef isnan |
---|
62 | #define isnan(x) _isnan(x) |
---|
63 | // Defeat the double->float conversion since we don't have tgmath |
---|
64 | inline SAS_DOUBLE trunc(SAS_DOUBLE x) { return x>=0?floor(x):-floor(-x); } |
---|
65 | inline SAS_DOUBLE fmin(SAS_DOUBLE x, SAS_DOUBLE y) { return x>y ? y : x; } |
---|
66 | inline SAS_DOUBLE fmax(SAS_DOUBLE x, SAS_DOUBLE y) { return x<y ? y : x; } |
---|
67 | #define NEED_ERF |
---|
68 | #define NEED_EXPM1 |
---|
69 | #define NEED_TGAMMA |
---|
70 | // expf missing from windows? |
---|
71 | #define expf exp |
---|
72 | #else |
---|
73 | #include <tgmath.h> // C99 type-generic math, so sin(float) => sinf |
---|
74 | #endif |
---|
75 | // MSVC doesn't support C99, so no need for dllexport on C99 branch |
---|
76 | #define kernel |
---|
77 | #define SINCOS(angle,svar,cvar) do {const double _t_=angle; svar=sin(_t_);cvar=cos(_t_);} while (0) |
---|
78 | # endif // !__cplusplus |
---|
79 | # define global |
---|
80 | # define local |
---|
81 | # define constant const |
---|
82 | // OpenCL powr(a,b) = C99 pow(a,b), b >= 0 |
---|
83 | // OpenCL pown(a,b) = C99 pow(a,b), b integer |
---|
84 | # define powr(a,b) pow(a,b) |
---|
85 | # define pown(a,b) pow(a,b) |
---|
86 | #endif // !USE_OPENCL |
---|
87 | |
---|
88 | #if defined(NEED_EXPM1) |
---|
89 | // TODO: precision is a half digit lower than numpy on mac in [1e-7, 0.5] |
---|
90 | // Run "explore/precision.py sas_expm1" to see this (may have to fiddle |
---|
91 | // the xrange for log to see the complete range). |
---|
92 | static SAS_DOUBLE expm1(SAS_DOUBLE x_in) { |
---|
93 | double x = (double)x_in; // go back to float for single precision kernels |
---|
94 | // Adapted from the cephes math library. |
---|
95 | // Copyright 1984 - 1992 by Stephen L. Moshier |
---|
96 | if (x != x || x == 0.0) { |
---|
97 | return x; // NaN and +/- 0 |
---|
98 | } else if (x < -0.5 || x > 0.5) { |
---|
99 | return exp(x) - 1.0; |
---|
100 | } else { |
---|
101 | const double xsq = x*x; |
---|
102 | const double p = ((( |
---|
103 | +1.2617719307481059087798E-4)*xsq |
---|
104 | +3.0299440770744196129956E-2)*xsq |
---|
105 | +9.9999999999999999991025E-1); |
---|
106 | const double q = (((( |
---|
107 | +3.0019850513866445504159E-6)*xsq |
---|
108 | +2.5244834034968410419224E-3)*xsq |
---|
109 | +2.2726554820815502876593E-1)*xsq |
---|
110 | +2.0000000000000000000897E0); |
---|
111 | double r = x * p; |
---|
112 | r = r / (q - r); |
---|
113 | return r+r; |
---|
114 | } |
---|
115 | } |
---|
116 | #endif |
---|
117 | |
---|
118 | // Standard mathematical constants: |
---|
119 | // M_E, M_LOG2E, M_LOG10E, M_LN2, M_LN10, M_PI, M_PI_2=pi/2, M_PI_4=pi/4, |
---|
120 | // M_1_PI=1/pi, M_2_PI=2/pi, M_2_SQRTPI=2/sqrt(pi), SQRT2, SQRT1_2=sqrt(1/2) |
---|
121 | // OpenCL defines M_constant_F for float constants, and nothing if double |
---|
122 | // is not enabled on the card, which is why these constants may be missing |
---|
123 | #ifndef M_PI |
---|
124 | # define M_PI 3.141592653589793 |
---|
125 | #endif |
---|
126 | #ifndef M_PI_2 |
---|
127 | # define M_PI_2 1.570796326794897 |
---|
128 | #endif |
---|
129 | #ifndef M_PI_4 |
---|
130 | # define M_PI_4 0.7853981633974483 |
---|
131 | #endif |
---|
132 | #ifndef M_E |
---|
133 | # define M_E 2.718281828459045091 |
---|
134 | #endif |
---|
135 | #ifndef M_SQRT1_2 |
---|
136 | # define M_SQRT1_2 0.70710678118654746 |
---|
137 | #endif |
---|
138 | |
---|
139 | // Non-standard function library |
---|
140 | // pi/180, used for converting between degrees and radians |
---|
141 | // 4/3 pi for computing sphere volumes |
---|
142 | // square and cube for computing squares and cubes |
---|
143 | #ifndef M_PI_180 |
---|
144 | # define M_PI_180 0.017453292519943295 |
---|
145 | #endif |
---|
146 | #ifndef M_4PI_3 |
---|
147 | # define M_4PI_3 4.18879020478639 |
---|
148 | #endif |
---|
149 | inline double square(double x) { return x*x; } |
---|
150 | inline double cube(double x) { return x*x*x; } |
---|
151 | inline double sas_sinx_x(double x) { return x==0 ? 1.0 : sin(x)/x; } |
---|