| 1 |
// Start of scalar_f16.h. |
| 2 |
|
| 3 |
// Half-precision is emulated if needed (e.g. in straight C) with the |
| 4 |
// native type used if possible. The emulation works by typedef'ing |
| 5 |
// 'float' to 'f16', and then implementing all operations on single |
| 6 |
// precision. To cut down on duplication, we use the same code for |
| 7 |
// those Futhark functions that require just operators or casts. The |
| 8 |
// in-memory representation for arrays will still be 16 bits even |
| 9 |
// under emulation, so the compiler will have to be careful when |
| 10 |
// generating reads or writes. |
| 11 |
|
| 12 |
#if !defined(cl_khr_fp16) && !(defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 600) && !(defined(ISPC)) |
| 13 |
#define EMULATE_F16 |
| 14 |
#endif |
| 15 |
|
| 16 |
#if !defined(EMULATE_F16) && defined(__OPENCL_VERSION__) |
| 17 |
#pragma OPENCL EXTENSION cl_khr_fp16 : enable |
| 18 |
#endif |
| 19 |
|
| 20 |
#ifdef EMULATE_F16 |
| 21 |
|
| 22 |
// Note that the half-precision storage format is still 16 bits - the |
| 23 |
// compiler will have to be real careful! |
| 24 |
typedef float f16; |
| 25 |
|
| 26 |
#elif defined(ISPC) |
| 27 |
typedef float16 f16; |
| 28 |
|
| 29 |
#else |
| 30 |
|
| 31 |
#ifdef __CUDA_ARCH__ |
| 32 |
#include <cuda_fp16.h> |
| 33 |
#endif |
| 34 |
|
| 35 |
typedef half f16; |
| 36 |
|
| 37 |
#endif |
| 38 |
|
| 39 |
// Some of these functions convert to single precision because half |
| 40 |
// precision versions are not available. |
| 41 |
SCALAR_FUN_ATTR f16 fadd16(f16 x, f16 y) { return x + y; } |
| 42 |
SCALAR_FUN_ATTR f16 fsub16(f16 x, f16 y) { return x - y; } |
| 43 |
SCALAR_FUN_ATTR f16 fmul16(f16 x, f16 y) { return x * y; } |
| 44 |
SCALAR_FUN_ATTR bool cmplt16(f16 x, f16 y) { return x < y; } |
| 45 |
SCALAR_FUN_ATTR bool cmple16(f16 x, f16 y) { return x <= y; } |
| 46 |
SCALAR_FUN_ATTR f16 sitofp_i8_f16(int8_t x) { return (f16) x; } |
| 47 |
SCALAR_FUN_ATTR f16 sitofp_i16_f16(int16_t x) { return (f16) x; } |
| 48 |
SCALAR_FUN_ATTR f16 sitofp_i32_f16(int32_t x) { return (f16) x; } |
| 49 |
SCALAR_FUN_ATTR f16 sitofp_i64_f16(int64_t x) { return (f16) x; } |
| 50 |
SCALAR_FUN_ATTR f16 uitofp_i8_f16(uint8_t x) { return (f16) x; } |
| 51 |
SCALAR_FUN_ATTR f16 uitofp_i16_f16(uint16_t x) { return (f16) x; } |
| 52 |
SCALAR_FUN_ATTR f16 uitofp_i32_f16(uint32_t x) { return (f16) x; } |
| 53 |
SCALAR_FUN_ATTR f16 uitofp_i64_f16(uint64_t x) { return (f16) x; } |
| 54 |
SCALAR_FUN_ATTR int8_t fptosi_f16_i8(f16 x) { return (int8_t) (float) x; } |
| 55 |
SCALAR_FUN_ATTR int16_t fptosi_f16_i16(f16 x) { return (int16_t) x; } |
| 56 |
SCALAR_FUN_ATTR int32_t fptosi_f16_i32(f16 x) { return (int32_t) x; } |
| 57 |
SCALAR_FUN_ATTR int64_t fptosi_f16_i64(f16 x) { return (int64_t) x; } |
| 58 |
SCALAR_FUN_ATTR uint8_t fptoui_f16_i8(f16 x) { return (uint8_t) (float) x; } |
| 59 |
SCALAR_FUN_ATTR uint16_t fptoui_f16_i16(f16 x) { return (uint16_t) x; } |
| 60 |
SCALAR_FUN_ATTR uint32_t fptoui_f16_i32(f16 x) { return (uint32_t) x; } |
| 61 |
SCALAR_FUN_ATTR uint64_t fptoui_f16_i64(f16 x) { return (uint64_t) x; } |
| 62 |
SCALAR_FUN_ATTR bool ftob_f16_bool(f16 x) { return x != (f16)0; } |
| 63 |
SCALAR_FUN_ATTR f16 btof_bool_f16(bool x) { return x ? 1 : 0; } |
| 64 |
|
| 65 |
#ifndef EMULATE_F16 |
| 66 |
|
| 67 |
SCALAR_FUN_ATTR bool futrts_isnan16(f16 x) { return isnan((float)x); } |
| 68 |
|
| 69 |
#ifdef __OPENCL_VERSION__ |
| 70 |
|
| 71 |
SCALAR_FUN_ATTR f16 fabs16(f16 x) { return fabs(x); } |
| 72 |
SCALAR_FUN_ATTR f16 fmax16(f16 x, f16 y) { return fmax(x, y); } |
| 73 |
SCALAR_FUN_ATTR f16 fmin16(f16 x, f16 y) { return fmin(x, y); } |
| 74 |
SCALAR_FUN_ATTR f16 fpow16(f16 x, f16 y) { return pow(x, y); } |
| 75 |
|
| 76 |
#elif defined(ISPC) |
| 77 |
|
| 78 |
SCALAR_FUN_ATTR f16 fabs16(f16 x) { return abs(x); } |
| 79 |
SCALAR_FUN_ATTR f16 fmax16(f16 x, f16 y) { return futrts_isnan16(x) ? y : futrts_isnan16(y) ? x : max(x, y); } |
| 80 |
SCALAR_FUN_ATTR f16 fmin16(f16 x, f16 y) { return futrts_isnan16(x) ? y : futrts_isnan16(y) ? x : min(x, y); } |
| 81 |
SCALAR_FUN_ATTR f16 fpow16(f16 x, f16 y) { return pow(x, y); } |
| 82 |
|
| 83 |
#else // Assuming CUDA. |
| 84 |
|
| 85 |
SCALAR_FUN_ATTR f16 fabs16(f16 x) { return fabsf(x); } |
| 86 |
SCALAR_FUN_ATTR f16 fmax16(f16 x, f16 y) { return fmaxf(x, y); } |
| 87 |
SCALAR_FUN_ATTR f16 fmin16(f16 x, f16 y) { return fminf(x, y); } |
| 88 |
SCALAR_FUN_ATTR f16 fpow16(f16 x, f16 y) { return powf(x, y); } |
| 89 |
|
| 90 |
#endif |
| 91 |
|
| 92 |
#if defined(ISPC) |
| 93 |
SCALAR_FUN_ATTR bool futrts_isinf16(float x) { return !futrts_isnan16(x) && futrts_isnan16(x - x); } |
| 94 |
SCALAR_FUN_ATTR bool futrts_isfinite16(float x) { return !futrts_isnan16(x) && !futrts_isinf16(x); } |
| 95 |
#else |
| 96 |
SCALAR_FUN_ATTR bool futrts_isinf16(f16 x) { return isinf((float)x); } |
| 97 |
#endif |
| 98 |
|
| 99 |
#ifdef __OPENCL_VERSION__ |
| 100 |
SCALAR_FUN_ATTR f16 futrts_log16(f16 x) { return log(x); } |
| 101 |
SCALAR_FUN_ATTR f16 futrts_log2_16(f16 x) { return log2(x); } |
| 102 |
SCALAR_FUN_ATTR f16 futrts_log10_16(f16 x) { return log10(x); } |
| 103 |
SCALAR_FUN_ATTR f16 futrts_log1p_16(f16 x) { return log1p(x); } |
| 104 |
SCALAR_FUN_ATTR f16 futrts_sqrt16(f16 x) { return sqrt(x); } |
| 105 |
SCALAR_FUN_ATTR f16 futrts_rsqrt16(f16 x) { return rsqrt(x); } |
| 106 |
SCALAR_FUN_ATTR f16 futrts_cbrt16(f16 x) { return cbrt(x); } |
| 107 |
SCALAR_FUN_ATTR f16 futrts_exp16(f16 x) { return exp(x); } |
| 108 |
SCALAR_FUN_ATTR f16 futrts_cos16(f16 x) { return cos(x); } |
| 109 |
SCALAR_FUN_ATTR f16 futrts_cospi16(f16 x) { return cospi(x); } |
| 110 |
SCALAR_FUN_ATTR f16 futrts_sin16(f16 x) { return sin(x); } |
| 111 |
SCALAR_FUN_ATTR f16 futrts_sinpi16(f16 x) { return sinpi(x); } |
| 112 |
SCALAR_FUN_ATTR f16 futrts_tan16(f16 x) { return tan(x); } |
| 113 |
SCALAR_FUN_ATTR f16 futrts_tanpi16(f16 x) { return tanpi(x); } |
| 114 |
SCALAR_FUN_ATTR f16 futrts_acos16(f16 x) { return acos(x); } |
| 115 |
SCALAR_FUN_ATTR f16 futrts_acospi16(f16 x) { return acospi(x); } |
| 116 |
SCALAR_FUN_ATTR f16 futrts_asin16(f16 x) { return asin(x); } |
| 117 |
SCALAR_FUN_ATTR f16 futrts_asinpi16(f16 x) { return asinpi(x); } |
| 118 |
SCALAR_FUN_ATTR f16 futrts_atan16(f16 x) { return atan(x); } |
| 119 |
SCALAR_FUN_ATTR f16 futrts_atanpi16(f16 x) { return atanpi(x); } |
| 120 |
SCALAR_FUN_ATTR f16 futrts_cosh16(f16 x) { return cosh(x); } |
| 121 |
SCALAR_FUN_ATTR f16 futrts_sinh16(f16 x) { return sinh(x); } |
| 122 |
SCALAR_FUN_ATTR f16 futrts_tanh16(f16 x) { return tanh(x); } |
| 123 |
SCALAR_FUN_ATTR f16 futrts_acosh16(f16 x) { return acosh(x); } |
| 124 |
SCALAR_FUN_ATTR f16 futrts_asinh16(f16 x) { return asinh(x); } |
| 125 |
SCALAR_FUN_ATTR f16 futrts_atanh16(f16 x) { return atanh(x); } |
| 126 |
SCALAR_FUN_ATTR f16 futrts_atan2_16(f16 x, f16 y) { return atan2(x, y); } |
| 127 |
SCALAR_FUN_ATTR f16 futrts_atan2pi_16(f16 x, f16 y) { return atan2pi(x, y); } |
| 128 |
SCALAR_FUN_ATTR f16 futrts_hypot16(f16 x, f16 y) { return hypot(x, y); } |
| 129 |
SCALAR_FUN_ATTR f16 futrts_gamma16(f16 x) { return tgamma(x); } |
| 130 |
SCALAR_FUN_ATTR f16 futrts_lgamma16(f16 x) { return lgamma(x); } |
| 131 |
SCALAR_FUN_ATTR f16 futrts_erf16(f16 x) { return erf(x); } |
| 132 |
SCALAR_FUN_ATTR f16 futrts_erfc16(f16 x) { return erfc(x); } |
| 133 |
SCALAR_FUN_ATTR f16 fmod16(f16 x, f16 y) { return fmod(x, y); } |
| 134 |
SCALAR_FUN_ATTR f16 futrts_round16(f16 x) { return rint(x); } |
| 135 |
SCALAR_FUN_ATTR f16 futrts_floor16(f16 x) { return floor(x); } |
| 136 |
SCALAR_FUN_ATTR f16 futrts_ceil16(f16 x) { return ceil(x); } |
| 137 |
SCALAR_FUN_ATTR f16 futrts_nextafter16(f16 x, f16 y) { return nextafter(x, y); } |
| 138 |
SCALAR_FUN_ATTR f16 futrts_lerp16(f16 v0, f16 v1, f16 t) { return mix(v0, v1, t); } |
| 139 |
SCALAR_FUN_ATTR f16 futrts_ldexp16(f16 x, int32_t y) { return ldexp(x, y); } |
| 140 |
SCALAR_FUN_ATTR f16 futrts_copysign16(f16 x, f16 y) { return copysign(x, y); } |
| 141 |
SCALAR_FUN_ATTR f16 futrts_mad16(f16 a, f16 b, f16 c) { return mad(a, b, c); } |
| 142 |
SCALAR_FUN_ATTR f16 futrts_fma16(f16 a, f16 b, f16 c) { return fma(a, b, c); } |
| 143 |
|
| 144 |
#elif defined(ISPC) |
| 145 |
|
| 146 |
SCALAR_FUN_ATTR f16 futrts_log16(f16 x) { return futrts_isfinite16(x) || (futrts_isinf16(x) && x < 0) ? log(x) : x; } |
| 147 |
SCALAR_FUN_ATTR f16 futrts_log2_16(f16 x) { return futrts_log16(x) / log(2.0f16); } |
| 148 |
SCALAR_FUN_ATTR f16 futrts_log10_16(f16 x) { return futrts_log16(x) / log(10.0f16); } |
| 149 |
SCALAR_FUN_ATTR f16 futrts_log1p_16(f16 x) { |
| 150 |
if(x == -1.0f16 || (futrts_isinf16(x) && x > 0.0f16)) return x / 0.0f16; |
| 151 |
f16 y = 1.0f16 + x; |
| 152 |
f16 z = y - 1.0f16; |
| 153 |
return log(y) - (z-x)/y; |
| 154 |
} |
| 155 |
SCALAR_FUN_ATTR f16 futrts_sqrt16(f16 x) { return (float16)sqrt((float)x); } |
| 156 |
SCALAR_FUN_ATTR f16 futrts_rsqrt16(f16 x) { return (float16)1/sqrt((float)x); } |
| 157 |
SCALAR_FUN_ATTR f16 futrts_exp16(f16 x) { return exp(x); } |
| 158 |
SCALAR_FUN_ATTR f16 futrts_cos16(f16 x) { return (float16)cos((float)x); } |
| 159 |
SCALAR_FUN_ATTR f16 futrts_cospi16(f16 x) { return (float16)cos((float)M_PI*(float)x); } |
| 160 |
SCALAR_FUN_ATTR f16 futrts_sin16(f16 x) { return (float16)sin((float)x); } |
| 161 |
SCALAR_FUN_ATTR f16 futrts_sinpi16(f16 x) { return (float16)sin((float)M_PI*(float)x); } |
| 162 |
SCALAR_FUN_ATTR f16 futrts_tan16(f16 x) { return (float16)tan((float)x); } |
| 163 |
SCALAR_FUN_ATTR f16 futrts_tanpi16(f16 x) { return (float16)(tan((float)M_PI*(float)x)); } |
| 164 |
SCALAR_FUN_ATTR f16 futrts_acos16(f16 x) { return (float16)acos((float)x); } |
| 165 |
SCALAR_FUN_ATTR f16 futrts_acospi16(f16 x) { return (float16)(acos((float)x)/(float)M_PI); } |
| 166 |
SCALAR_FUN_ATTR f16 futrts_asin16(f16 x) { return (float16)asin((float)x); } |
| 167 |
SCALAR_FUN_ATTR f16 futrts_asinpi16(f16 x) { return (float16)(asin((float)x)/(float)M_PI); } |
| 168 |
SCALAR_FUN_ATTR f16 futrts_atan16(f16 x) { return (float16)atan((float)x); } |
| 169 |
SCALAR_FUN_ATTR f16 futrts_atanpi16(f16 x) { return (float16)(atan((float)x)/(float)M_PI); } |
| 170 |
SCALAR_FUN_ATTR f16 futrts_cosh16(f16 x) { return (exp(x)+exp(-x)) / 2.0f16; } |
| 171 |
SCALAR_FUN_ATTR f16 futrts_sinh16(f16 x) { return (exp(x)-exp(-x)) / 2.0f16; } |
| 172 |
SCALAR_FUN_ATTR f16 futrts_tanh16(f16 x) { return futrts_sinh16(x)/futrts_cosh16(x); } |
| 173 |
SCALAR_FUN_ATTR f16 futrts_acosh16(f16 x) { |
| 174 |
float16 f = x+(float16)sqrt((float)(x*x-1)); |
| 175 |
if(futrts_isfinite16(f)) return log(f); |
| 176 |
return f; |
| 177 |
} |
| 178 |
SCALAR_FUN_ATTR f16 futrts_asinh16(f16 x) { |
| 179 |
float16 f = x+(float16)sqrt((float)(x*x+1)); |
| 180 |
if(futrts_isfinite16(f)) return log(f); |
| 181 |
return f; |
| 182 |
} |
| 183 |
SCALAR_FUN_ATTR f16 futrts_atanh16(f16 x) { |
| 184 |
float16 f = (1+x)/(1-x); |
| 185 |
if(futrts_isfinite16(f)) return log(f)/2.0f16; |
| 186 |
return f; |
| 187 |
} |
| 188 |
SCALAR_FUN_ATTR f16 futrts_atan2_16(f16 x, f16 y) { return (float16)atan2((float)x, (float)y); } |
| 189 |
SCALAR_FUN_ATTR f16 futrts_atan2pi_16(f16 x, f16 y) { return (float16)(atan2((float)x, (float)y)/(float)M_PI); } |
| 190 |
SCALAR_FUN_ATTR f16 futrts_hypot16(f16 x, f16 y) { return (float16)futrts_hypot32((float)x, (float)y); } |
| 191 |
|
| 192 |
extern "C" unmasked uniform float tgammaf(uniform float x); |
| 193 |
SCALAR_FUN_ATTR f16 futrts_gamma16(f16 x) { |
| 194 |
f16 res; |
| 195 |
foreach_active (i) { |
| 196 |
uniform f16 r = (f16)tgammaf(extract((float)x, i)); |
| 197 |
res = insert(res, i, r); |
| 198 |
} |
| 199 |
return res; |
| 200 |
} |
| 201 |
|
| 202 |
extern "C" unmasked uniform float lgammaf(uniform float x); |
| 203 |
SCALAR_FUN_ATTR f16 futrts_lgamma16(f16 x) { |
| 204 |
f16 res; |
| 205 |
foreach_active (i) { |
| 206 |
uniform f16 r = (f16)lgammaf(extract((float)x, i)); |
| 207 |
res = insert(res, i, r); |
| 208 |
} |
| 209 |
return res; |
| 210 |
} |
| 211 |
SCALAR_FUN_ATTR f16 futrts_cbrt16(f16 x) { return (f16)futrts_cbrt32((float)x); } |
| 212 |
SCALAR_FUN_ATTR f16 futrts_erf16(f16 x) { return (f16)futrts_erf32((float)x); } |
| 213 |
SCALAR_FUN_ATTR f16 futrts_erfc16(f16 x) { return (f16)futrts_erfc32((float)x); } |
| 214 |
SCALAR_FUN_ATTR f16 fmod16(f16 x, f16 y) { return x - y * (float16)trunc((float) (x/y)); } |
| 215 |
SCALAR_FUN_ATTR f16 futrts_round16(f16 x) { return (float16)round((float)x); } |
| 216 |
SCALAR_FUN_ATTR f16 futrts_floor16(f16 x) { return (float16)floor((float)x); } |
| 217 |
SCALAR_FUN_ATTR f16 futrts_ceil16(f16 x) { return (float16)ceil((float)x); } |
| 218 |
SCALAR_FUN_ATTR f16 futrts_nextafter16(f16 x, f16 y) { return (float16)futrts_nextafter32((float)x, (float) y); } |
| 219 |
SCALAR_FUN_ATTR f16 futrts_lerp16(f16 v0, f16 v1, f16 t) { return v0 + (v1 - v0) * t; } |
| 220 |
SCALAR_FUN_ATTR f16 futrts_ldexp16(f16 x, int32_t y) { return futrts_ldexp32((float)x, y); } |
| 221 |
SCALAR_FUN_ATTR f16 futrts_copysign16(f16 x, f16 y) { return futrts_copysign32((float)x, y); } |
| 222 |
SCALAR_FUN_ATTR f16 futrts_mad16(f16 a, f16 b, f16 c) { return a * b + c; } |
| 223 |
SCALAR_FUN_ATTR f16 futrts_fma16(f16 a, f16 b, f16 c) { return a * b + c; } |
| 224 |
|
| 225 |
#else // Assume CUDA. |
| 226 |
|
| 227 |
SCALAR_FUN_ATTR f16 futrts_log16(f16 x) { return hlog(x); } |
| 228 |
SCALAR_FUN_ATTR f16 futrts_log2_16(f16 x) { return hlog2(x); } |
| 229 |
SCALAR_FUN_ATTR f16 futrts_log10_16(f16 x) { return hlog10(x); } |
| 230 |
SCALAR_FUN_ATTR f16 futrts_log1p_16(f16 x) { return (f16)log1pf((float)x); } |
| 231 |
SCALAR_FUN_ATTR f16 futrts_sqrt16(f16 x) { return hsqrt(x); } |
| 232 |
SCALAR_FUN_ATTR f16 futrts_rsqrt16(f16 x) { return hrsqrt(x); } |
| 233 |
SCALAR_FUN_ATTR f16 futrts_cbrt16(f16 x) { return cbrtf(x); } |
| 234 |
SCALAR_FUN_ATTR f16 futrts_exp16(f16 x) { return hexp(x); } |
| 235 |
SCALAR_FUN_ATTR f16 futrts_cos16(f16 x) { return hcos(x); } |
| 236 |
SCALAR_FUN_ATTR f16 futrts_cospi16(f16 x) { return hcos((f16)M_PI*x); } |
| 237 |
SCALAR_FUN_ATTR f16 futrts_sin16(f16 x) { return hsin(x); } |
| 238 |
SCALAR_FUN_ATTR f16 futrts_sinpi16(f16 x) { return hsin((f16)M_PI*x); } |
| 239 |
SCALAR_FUN_ATTR f16 futrts_tan16(f16 x) { return tanf(x); } |
| 240 |
SCALAR_FUN_ATTR f16 futrts_tanpi16(f16 x) { return tanf((f16)M_PI*x); } |
| 241 |
SCALAR_FUN_ATTR f16 futrts_acos16(f16 x) { return acosf(x); } |
| 242 |
SCALAR_FUN_ATTR f16 futrts_acospi16(f16 x) { return (f16)acosf(x)/(f16)M_PI; } |
| 243 |
SCALAR_FUN_ATTR f16 futrts_asin16(f16 x) { return asinf(x); } |
| 244 |
SCALAR_FUN_ATTR f16 futrts_asinpi16(f16 x) { return (f16)asinf(x)/(f16)M_PI; } |
| 245 |
SCALAR_FUN_ATTR f16 futrts_atan16(f16 x) { return (f16)atanf(x); } |
| 246 |
SCALAR_FUN_ATTR f16 futrts_atanpi16(f16 x) { return (f16)atanf(x)/(f16)M_PI; } |
| 247 |
SCALAR_FUN_ATTR f16 futrts_cosh16(f16 x) { return coshf(x); } |
| 248 |
SCALAR_FUN_ATTR f16 futrts_sinh16(f16 x) { return sinhf(x); } |
| 249 |
SCALAR_FUN_ATTR f16 futrts_tanh16(f16 x) { return tanhf(x); } |
| 250 |
SCALAR_FUN_ATTR f16 futrts_acosh16(f16 x) { return acoshf(x); } |
| 251 |
SCALAR_FUN_ATTR f16 futrts_asinh16(f16 x) { return asinhf(x); } |
| 252 |
SCALAR_FUN_ATTR f16 futrts_atanh16(f16 x) { return atanhf(x); } |
| 253 |
SCALAR_FUN_ATTR f16 futrts_atan2_16(f16 x, f16 y) { return (f16)atan2f(x, y); } |
| 254 |
SCALAR_FUN_ATTR f16 futrts_atan2pi_16(f16 x, f16 y) { return (f16)atan2f(x, y)/(f16)M_PI; } |
| 255 |
SCALAR_FUN_ATTR f16 futrts_hypot16(f16 x, f16 y) { return hypotf(x, y); } |
| 256 |
SCALAR_FUN_ATTR f16 futrts_gamma16(f16 x) { return tgammaf(x); } |
| 257 |
SCALAR_FUN_ATTR f16 futrts_lgamma16(f16 x) { return lgammaf(x); } |
| 258 |
SCALAR_FUN_ATTR f16 futrts_erf16(f16 x) { return erff(x); } |
| 259 |
SCALAR_FUN_ATTR f16 futrts_erfc16(f16 x) { return erfcf(x); } |
| 260 |
SCALAR_FUN_ATTR f16 fmod16(f16 x, f16 y) { return fmodf(x, y); } |
| 261 |
SCALAR_FUN_ATTR f16 futrts_round16(f16 x) { return rintf(x); } |
| 262 |
SCALAR_FUN_ATTR f16 futrts_floor16(f16 x) { return hfloor(x); } |
| 263 |
SCALAR_FUN_ATTR f16 futrts_ceil16(f16 x) { return hceil(x); } |
| 264 |
SCALAR_FUN_ATTR f16 futrts_nextafter16(f16 x, f16 y) { return __ushort_as_half(halfbitsnextafter(__half_as_ushort(x), __half_as_ushort(y))); } |
| 265 |
SCALAR_FUN_ATTR f16 futrts_lerp16(f16 v0, f16 v1, f16 t) { return v0 + (v1 - v0) * t; } |
| 266 |
SCALAR_FUN_ATTR f16 futrts_ldexp16(f16 x, int32_t y) { return futrts_ldexp32((float)x, y); } |
| 267 |
SCALAR_FUN_ATTR f16 futrts_copysign16(f16 x, f16 y) { return futrts_copysign32((float)x, y); } |
| 268 |
SCALAR_FUN_ATTR f16 futrts_mad16(f16 a, f16 b, f16 c) { return a * b + c; } |
| 269 |
SCALAR_FUN_ATTR f16 futrts_fma16(f16 a, f16 b, f16 c) { return fmaf(a, b, c); } |
| 270 |
|
| 271 |
#endif |
| 272 |
|
| 273 |
// The CUDA __half type cannot be put in unions for some reason, so we |
| 274 |
// use bespoke conversion functions instead. |
| 275 |
#ifdef __CUDA_ARCH__ |
| 276 |
SCALAR_FUN_ATTR int16_t fptobits_f16_i16(f16 x) { return __half_as_ushort(x); } |
| 277 |
SCALAR_FUN_ATTR f16 bitstofp_i16_f16(int16_t x) { return __ushort_as_half(x); } |
| 278 |
#elif defined(ISPC) |
| 279 |
SCALAR_FUN_ATTR int16_t fptobits_f16_i16(f16 x) { varying int16_t y = *((varying int16_t * uniform)&x); return y; |
| 280 |
} |
| 281 |
SCALAR_FUN_ATTR f16 bitstofp_i16_f16(int16_t x) { varying f16 y = *((varying f16 * uniform)&x); return y; } |
| 282 |
#else |
| 283 |
SCALAR_FUN_ATTR int16_t fptobits_f16_i16(f16 x) { |
| 284 |
union { |
| 285 |
f16 f; |
| 286 |
int16_t t; |
| 287 |
} p; |
| 288 |
|
| 289 |
p.f = x; |
| 290 |
return p.t; |
| 291 |
} |
| 292 |
|
| 293 |
SCALAR_FUN_ATTR f16 bitstofp_i16_f16(int16_t x) { |
| 294 |
union { |
| 295 |
int16_t f; |
| 296 |
f16 t; |
| 297 |
} p; |
| 298 |
|
| 299 |
p.f = x; |
| 300 |
return p.t; |
| 301 |
} |
| 302 |
#endif |
| 303 |
|
| 304 |
#else // No native f16 - emulate. |
| 305 |
|
| 306 |
SCALAR_FUN_ATTR f16 fabs16(f16 x) { return fabs32(x); } |
| 307 |
SCALAR_FUN_ATTR f16 fmax16(f16 x, f16 y) { return fmax32(x, y); } |
| 308 |
SCALAR_FUN_ATTR f16 fmin16(f16 x, f16 y) { return fmin32(x, y); } |
| 309 |
SCALAR_FUN_ATTR f16 fpow16(f16 x, f16 y) { return fpow32(x, y); } |
| 310 |
SCALAR_FUN_ATTR bool futrts_isnan16(f16 x) { return futrts_isnan32(x); } |
| 311 |
SCALAR_FUN_ATTR bool futrts_isinf16(f16 x) { return futrts_isinf32(x); } |
| 312 |
SCALAR_FUN_ATTR f16 futrts_log16(f16 x) { return futrts_log32(x); } |
| 313 |
SCALAR_FUN_ATTR f16 futrts_log2_16(f16 x) { return futrts_log2_32(x); } |
| 314 |
SCALAR_FUN_ATTR f16 futrts_log10_16(f16 x) { return futrts_log10_32(x); } |
| 315 |
SCALAR_FUN_ATTR f16 futrts_log1p_16(f16 x) { return futrts_log1p_32(x); } |
| 316 |
SCALAR_FUN_ATTR f16 futrts_sqrt16(f16 x) { return futrts_sqrt32(x); } |
| 317 |
SCALAR_FUN_ATTR f16 futrts_rsqrt16(f16 x) { return futrts_rsqrt32(x); } |
| 318 |
SCALAR_FUN_ATTR f16 futrts_cbrt16(f16 x) { return futrts_cbrt32(x); } |
| 319 |
SCALAR_FUN_ATTR f16 futrts_exp16(f16 x) { return futrts_exp32(x); } |
| 320 |
SCALAR_FUN_ATTR f16 futrts_cos16(f16 x) { return futrts_cos32(x); } |
| 321 |
SCALAR_FUN_ATTR f16 futrts_cospi16(f16 x) { return futrts_cospi32(x); } |
| 322 |
SCALAR_FUN_ATTR f16 futrts_sin16(f16 x) { return futrts_sin32(x); } |
| 323 |
SCALAR_FUN_ATTR f16 futrts_sinpi16(f16 x) { return futrts_sinpi32(x); } |
| 324 |
SCALAR_FUN_ATTR f16 futrts_tan16(f16 x) { return futrts_tan32(x); } |
| 325 |
SCALAR_FUN_ATTR f16 futrts_tanpi16(f16 x) { return futrts_tanpi32(x); } |
| 326 |
SCALAR_FUN_ATTR f16 futrts_acos16(f16 x) { return futrts_acos32(x); } |
| 327 |
SCALAR_FUN_ATTR f16 futrts_acospi16(f16 x) { return futrts_acospi32(x); } |
| 328 |
SCALAR_FUN_ATTR f16 futrts_asin16(f16 x) { return futrts_asin32(x); } |
| 329 |
SCALAR_FUN_ATTR f16 futrts_asinpi16(f16 x) { return futrts_asinpi32(x); } |
| 330 |
SCALAR_FUN_ATTR f16 futrts_atan16(f16 x) { return futrts_atan32(x); } |
| 331 |
SCALAR_FUN_ATTR f16 futrts_atanpi16(f16 x) { return futrts_atanpi32(x); } |
| 332 |
SCALAR_FUN_ATTR f16 futrts_cosh16(f16 x) { return futrts_cosh32(x); } |
| 333 |
SCALAR_FUN_ATTR f16 futrts_sinh16(f16 x) { return futrts_sinh32(x); } |
| 334 |
SCALAR_FUN_ATTR f16 futrts_tanh16(f16 x) { return futrts_tanh32(x); } |
| 335 |
SCALAR_FUN_ATTR f16 futrts_acosh16(f16 x) { return futrts_acosh32(x); } |
| 336 |
SCALAR_FUN_ATTR f16 futrts_asinh16(f16 x) { return futrts_asinh32(x); } |
| 337 |
SCALAR_FUN_ATTR f16 futrts_atanh16(f16 x) { return futrts_atanh32(x); } |
| 338 |
SCALAR_FUN_ATTR f16 futrts_atan2_16(f16 x, f16 y) { return futrts_atan2_32(x, y); } |
| 339 |
SCALAR_FUN_ATTR f16 futrts_atan2pi_16(f16 x, f16 y) { return futrts_atan2pi_32(x, y); } |
| 340 |
SCALAR_FUN_ATTR f16 futrts_hypot16(f16 x, f16 y) { return futrts_hypot32(x, y); } |
| 341 |
SCALAR_FUN_ATTR f16 futrts_gamma16(f16 x) { return futrts_gamma32(x); } |
| 342 |
SCALAR_FUN_ATTR f16 futrts_lgamma16(f16 x) { return futrts_lgamma32(x); } |
| 343 |
SCALAR_FUN_ATTR f16 futrts_erf16(f16 x) { return futrts_erf32(x); } |
| 344 |
SCALAR_FUN_ATTR f16 futrts_erfc16(f16 x) { return futrts_erfc32(x); } |
| 345 |
SCALAR_FUN_ATTR f16 fmod16(f16 x, f16 y) { return fmod32(x, y); } |
| 346 |
SCALAR_FUN_ATTR f16 futrts_round16(f16 x) { return futrts_round32(x); } |
| 347 |
SCALAR_FUN_ATTR f16 futrts_floor16(f16 x) { return futrts_floor32(x); } |
| 348 |
SCALAR_FUN_ATTR f16 futrts_ceil16(f16 x) { return futrts_ceil32(x); } |
| 349 |
SCALAR_FUN_ATTR f16 futrts_nextafter16(f16 x, f16 y) { return halfbits2float(halfbitsnextafter(float2halfbits(x), float2halfbits(y))); } |
| 350 |
SCALAR_FUN_ATTR f16 futrts_lerp16(f16 v0, f16 v1, f16 t) { return futrts_lerp32(v0, v1, t); } |
| 351 |
SCALAR_FUN_ATTR f16 futrts_ldexp16(f16 x, int32_t y) { return futrts_ldexp32(x, y); } |
| 352 |
SCALAR_FUN_ATTR f16 futrts_copysign16(f16 x, f16 y) { return futrts_copysign32((float)x, y); } |
| 353 |
SCALAR_FUN_ATTR f16 futrts_mad16(f16 a, f16 b, f16 c) { return futrts_mad32(a, b, c); } |
| 354 |
SCALAR_FUN_ATTR f16 futrts_fma16(f16 a, f16 b, f16 c) { return futrts_fma32(a, b, c); } |
| 355 |
|
| 356 |
// Even when we are using an OpenCL that does not support cl_khr_fp16, |
| 357 |
// it must still support vload_half for actually creating a |
| 358 |
// half-precision number, which can then be efficiently converted to a |
| 359 |
// float. Similarly for vstore_half. |
| 360 |
#ifdef __OPENCL_VERSION__ |
| 361 |
|
| 362 |
SCALAR_FUN_ATTR int16_t fptobits_f16_i16(f16 x) { |
| 363 |
int16_t y; |
| 364 |
// Violating strict aliasing here. |
| 365 |
vstore_half((float)x, 0, (half*)&y); |
| 366 |
return y; |
| 367 |
} |
| 368 |
|
| 369 |
SCALAR_FUN_ATTR f16 bitstofp_i16_f16(int16_t x) { |
| 370 |
return (f16)vload_half(0, (half*)&x); |
| 371 |
} |
| 372 |
|
| 373 |
#else |
| 374 |
SCALAR_FUN_ATTR int16_t fptobits_f16_i16(f16 x) { return (int16_t)float2halfbits(x); } |
| 375 |
SCALAR_FUN_ATTR f16 bitstofp_i16_f16(int16_t x) { return halfbits2float((uint16_t)x); } |
| 376 |
SCALAR_FUN_ATTR f16 fsignum16(f16 x) { return futrts_isnan16(x) ? x : (x > 0 ? 1 : 0) - (x < 0 ? 1 : 0); } |
| 377 |
|
| 378 |
#endif |
| 379 |
|
| 380 |
#endif |
| 381 |
|
| 382 |
SCALAR_FUN_ATTR float fpconv_f16_f16(f16 x) { return x; } |
| 383 |
SCALAR_FUN_ATTR float fpconv_f16_f32(f16 x) { return x; } |
| 384 |
SCALAR_FUN_ATTR f16 fpconv_f32_f16(float x) { return (f16) x; } |
| 385 |
|
| 386 |
#ifdef FUTHARK_F64_ENABLED |
| 387 |
SCALAR_FUN_ATTR double fpconv_f16_f64(f16 x) { return (double) x; } |
| 388 |
#if defined(ISPC) |
| 389 |
SCALAR_FUN_ATTR f16 fpconv_f64_f16(double x) { return (f16) ((float)x); } |
| 390 |
#else |
| 391 |
SCALAR_FUN_ATTR f16 fpconv_f64_f16(double x) { return (f16) x; } |
| 392 |
#endif |
| 393 |
#endif |
| 394 |
|
| 395 |
// End of scalar_f16.h. |