Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #pragma once |
| 7 | |
| 8 | #include <stdint.h> |
| 9 | #include <stddef.h> |
| 10 | |
| 11 | #ifdef __cplusplus |
| 12 | extern "C" { |
| 13 | #endif |
| 14 | |
| 15 | typedef void (*xnn_f32_unary_math_function)( |
| 16 | size_t n, |
| 17 | const float* input, |
| 18 | float* output); |
| 19 | |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 20 | typedef void (*xnn_f32_ext_unary_math_function)( |
| 21 | size_t n, |
| 22 | const float* input, |
| 23 | float* output_mantissa, |
| 24 | float* output_exponent); |
| 25 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 26 | #define DECLARE_F32_UNARY_MATH_FUNCTION(fn_name) \ |
| 27 | void fn_name( \ |
| 28 | size_t n, \ |
| 29 | const float* input, \ |
| 30 | float* output); |
| 31 | |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 32 | #define DECLARE_F32_EXT_UNARY_MATH_FUNCTION(fn_name) \ |
| 33 | void fn_name( \ |
| 34 | size_t n, \ |
| 35 | const float* input, \ |
| 36 | float* output_mantissa, \ |
| 37 | float* output_exponent); |
| 38 | |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 39 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundne__neon_addsub) |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 40 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundne__neonv8) |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 41 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundne__sse_addsub) |
| 42 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundne__sse2_cvt) |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 43 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundne__sse41) |
Marat Dukhan | d3f3d87 | 2020-06-24 13:08:25 -0700 | [diff] [blame] | 44 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundne__wasmsimd_addsub) |
Marat Dukhan | 075088a | 2020-05-12 19:42:12 -0700 | [diff] [blame] | 45 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundne__scalar_addsub) |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 46 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundne__scalar_nearbyint) |
| 47 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundne__scalar_rint) |
Marat Dukhan | 8853b82 | 2020-05-07 12:19:01 -0700 | [diff] [blame] | 48 | |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 49 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundd__neon_addsub) |
| 50 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundd__neon_cvt) |
| 51 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundd__neonv8) |
| 52 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundd__sse_addsub) |
| 53 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundd__sse2_cvt) |
| 54 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundd__sse41) |
Marat Dukhan | d3f3d87 | 2020-06-24 13:08:25 -0700 | [diff] [blame] | 55 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundd__wasmsimd_addsub) |
Marat Dukhan | 88da62c | 2020-07-15 17:56:09 -0700 | [diff] [blame] | 56 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundd__wasmsimd_cvt) |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 57 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundd__scalar_addsub) |
| 58 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundd__scalar_cvt) |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 59 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundd__scalar_floor) |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 60 | |
| 61 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundu__neon_addsub) |
| 62 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundu__neon_cvt) |
| 63 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundu__neonv8) |
| 64 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundu__sse_addsub) |
| 65 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundu__sse2_cvt) |
| 66 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundu__sse41) |
Marat Dukhan | d3f3d87 | 2020-06-24 13:08:25 -0700 | [diff] [blame] | 67 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundu__wasmsimd_addsub) |
Marat Dukhan | 88da62c | 2020-07-15 17:56:09 -0700 | [diff] [blame] | 68 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundu__wasmsimd_cvt) |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 69 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundu__scalar_addsub) |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 70 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundu__scalar_ceil) |
Marat Dukhan | c9852ba | 2020-05-13 17:21:29 -0700 | [diff] [blame] | 71 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundu__scalar_cvt) |
| 72 | |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 73 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundz__neon_addsub) |
| 74 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundz__neon_cvt) |
| 75 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundz__neonv8) |
| 76 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundz__sse_addsub) |
| 77 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundz__sse2_cvt) |
| 78 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundz__sse41) |
Marat Dukhan | d3f3d87 | 2020-06-24 13:08:25 -0700 | [diff] [blame] | 79 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundz__wasmsimd_addsub) |
Marat Dukhan | 88da62c | 2020-07-15 17:56:09 -0700 | [diff] [blame] | 80 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundz__wasmsimd_cvt) |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 81 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundz__scalar_addsub) |
| 82 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundz__scalar_cvt) |
Marat Dukhan | ffbf96a | 2020-05-14 02:59:08 -0700 | [diff] [blame] | 83 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_roundz__scalar_trunc) |
Marat Dukhan | 2dbb944 | 2020-05-12 20:43:43 -0700 | [diff] [blame] | 84 | |
Marat Dukhan | 797a8fe | 2019-11-14 20:21:57 -0800 | [diff] [blame] | 85 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__neonfma_lut64_p2) |
| 86 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__neonfma_p5) |
Marat Dukhan | ffd6840 | 2019-11-15 15:19:11 -0800 | [diff] [blame] | 87 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__sse2_p5) |
Marat Dukhan | 8e17294 | 2020-09-20 23:56:23 -0700 | [diff] [blame] | 88 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__sse2_lut64_p2) |
Marat Dukhan | d08abcb | 2020-09-21 00:26:25 -0700 | [diff] [blame] | 89 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx_p5) |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 90 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx2_p5) |
Marat Dukhan | 8e17294 | 2020-09-20 23:56:23 -0700 | [diff] [blame] | 91 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx2_lut8_p3_perm) |
| 92 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx2_lut8_p4_perm) |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 93 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx512f_p5) |
| 94 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx512f_p5_scalef) |
Marat Dukhan | 8e17294 | 2020-09-20 23:56:23 -0700 | [diff] [blame] | 95 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx512f_lut16_p3_perm) |
Marat Dukhan | 80a8ac5 | 2020-09-21 17:00:27 -0700 | [diff] [blame] | 96 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx512f_lut16_p3_perm_scalef) |
Marat Dukhan | 8e17294 | 2020-09-20 23:56:23 -0700 | [diff] [blame] | 97 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx512f_lut32_p2_perm2) |
Marat Dukhan | 80a8ac5 | 2020-09-21 17:00:27 -0700 | [diff] [blame] | 98 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_exp__avx512f_lut32_p2_perm2_scalef) |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 99 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 100 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__neonfma_p5) |
Marat Dukhan | 189ae80 | 2019-11-26 11:28:44 -0800 | [diff] [blame] | 101 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__neonfma_lut64_p2) |
| 102 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__neonfma_lut2048_p1) |
Marat Dukhan | 5e9a91e | 2019-12-22 19:13:03 -0800 | [diff] [blame] | 103 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__sse2_p5) |
| 104 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__avx2_p5) |
| 105 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__scalar_p5) |
| 106 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__scalar_lut64_p2) |
| 107 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_expminus__scalar_lut2048_p1) |
Marat Dukhan | 515c977 | 2019-10-17 18:07:57 -0700 | [diff] [blame] | 108 | |
Marat Dukhan | 98ba441 | 2019-10-23 02:14:28 -0700 | [diff] [blame] | 109 | DECLARE_F32_EXT_UNARY_MATH_FUNCTION(xnn_math_f32_extexp__avx2_p5) |
| 110 | DECLARE_F32_EXT_UNARY_MATH_FUNCTION(xnn_math_f32_extexp__avx512f_p5) |
| 111 | |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 112 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neon_frac_p9_p10_nr1recps) |
Marat Dukhan | 77221d3 | 2020-01-06 10:04:39 -0800 | [diff] [blame] | 113 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neon_rr1_lut2048_p1_nr2recps) |
| 114 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_lut2048_p1_nr2recps) |
| 115 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_lut2048_p1_nr1recps1fma) |
| 116 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_lut2048_p1_nr2fma) |
| 117 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_lut2048_p1_div) |
| 118 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neon_rr2_lut2048_p1_nr2recps) |
| 119 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_lut2048_p1_nr2recps) |
| 120 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_lut2048_p1_nr1recps1fma) |
| 121 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_lut2048_p1_nr2fma) |
| 122 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_lut2048_p1_div) |
| 123 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neon_rr1_lut64_p2_nr2recps) |
| 124 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_lut64_p2_nr2recps) |
| 125 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_lut64_p2_nr1recps1fma) |
| 126 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_lut64_p2_nr2fma) |
| 127 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_lut64_p2_div) |
| 128 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neon_rr2_lut64_p2_nr2recps) |
| 129 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_lut64_p2_nr2recps) |
| 130 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_lut64_p2_nr1recps1fma) |
| 131 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_lut64_p2_nr2fma) |
| 132 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_lut64_p2_div) |
| 133 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neon_rr1_p5_nr2recps) |
| 134 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_p5_nr2recps) |
| 135 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_p5_nr1recps1fma) |
| 136 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_p5_nr2fma) |
| 137 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr1_p5_div) |
| 138 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neon_rr2_p5_nr2recps) |
| 139 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_p5_nr2recps) |
| 140 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_p5_nr1recps1fma) |
| 141 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_p5_nr2fma) |
| 142 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__neonfma_rr2_p5_div) |
Marat Dukhan | f8475d6 | 2020-09-17 15:01:43 -0700 | [diff] [blame] | 143 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__sse2_rr2_p5_div) |
Marat Dukhan | ac5f160 | 2020-10-15 21:00:45 -0700 | [diff] [blame] | 144 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__sse2_rr2_p5_nr1) |
| 145 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__sse2_rr2_p5_nr2) |
Marat Dukhan | f8475d6 | 2020-09-17 15:01:43 -0700 | [diff] [blame] | 146 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__sse2_rr2_lut64_p2_div) |
Marat Dukhan | ac5f160 | 2020-10-15 21:00:45 -0700 | [diff] [blame] | 147 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__sse2_rr2_lut64_p2_nr1) |
| 148 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__sse2_rr2_lut64_p2_nr2) |
Marat Dukhan | 2e4727b | 2020-09-24 00:48:39 -0700 | [diff] [blame] | 149 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx_rr2_p5_div) |
T.J. Alumbaugh | dc2b29c | 2020-10-14 13:56:08 -0700 | [diff] [blame] | 150 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx_rr2_p5_nr2) |
Marat Dukhan | 39ab16f | 2020-10-16 14:30:35 -0700 | [diff] [blame] | 151 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx_rr2_p5_nr1) |
| 152 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div) |
Marat Dukhan | 7241600 | 2020-01-05 21:53:19 -0800 | [diff] [blame] | 153 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr2_p5_div) |
| 154 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr2_p5_nr2fma) |
| 155 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr2_p5_nr1fma) |
| 156 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr1_p5_div) |
| 157 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr1_p5_nr2fma) |
| 158 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr1_p5_nr1fma) |
Marat Dukhan | 36173d2 | 2020-10-15 17:14:26 -0700 | [diff] [blame] | 159 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr2_lut64_p2_gather_div) |
| 160 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr2_lut64_p2_gather_nr2fma1adj) |
| 161 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr2_lut64_p2_gather_nr2fma) |
| 162 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr2_lut64_p2_gather_nr1fma) |
| 163 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr1_lut64_p2_gather_div) |
| 164 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr1_lut64_p2_gather_nr2fma1adj) |
| 165 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr1_lut64_p2_gather_nr2fma) |
| 166 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx2_rr1_lut64_p2_gather_nr1fma) |
Marat Dukhan | 47ffc41 | 2020-09-20 23:30:28 -0700 | [diff] [blame] | 167 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_div) |
Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 168 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma1adj) |
Marat Dukhan | 47ffc41 | 2020-09-20 23:30:28 -0700 | [diff] [blame] | 169 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_p5_scalef_nr1fma) |
| 170 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_div) |
Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 171 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma1adj) |
Marat Dukhan | 47ffc41 | 2020-09-20 23:30:28 -0700 | [diff] [blame] | 172 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_p5_scalef_nr1fma) |
Marat Dukhan | 6a34c5f | 2020-09-22 21:44:15 -0700 | [diff] [blame] | 173 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_div) |
| 174 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma1adj) |
| 175 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma) |
| 176 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_lut16_p3_perm_scalef_div) |
| 177 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_lut16_p3_perm_scalef_nr1fma1adj) |
| 178 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_lut16_p3_perm_scalef_nr1fma) |
| 179 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_div) |
| 180 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma1adj) |
| 181 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma) |
| 182 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_lut32_p2_perm2_scalef_div) |
| 183 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_lut32_p2_perm2_scalef_nr1fma1adj) |
| 184 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_lut32_p2_perm2_scalef_nr1fma) |
Marat Dukhan | 36173d2 | 2020-10-15 17:14:26 -0700 | [diff] [blame] | 185 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_lut64_p2_gather_scalef_div) |
| 186 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_lut64_p2_gather_scalef_nr1fma1adj) |
| 187 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr2_lut64_p2_gather_scalef_nr1fma) |
| 188 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_lut64_p2_gather_scalef_div) |
| 189 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_lut64_p2_gather_scalef_nr1fma1adj) |
| 190 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__avx512f_rr1_lut64_p2_gather_scalef_nr1fma) |
Marat Dukhan | 773da43 | 2020-07-17 18:55:36 -0700 | [diff] [blame] | 191 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__wasmsimd_rr2_lut64_p2_div) |
| 192 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__wasmsimd_rr2_p5_div) |
Marat Dukhan | f8475d6 | 2020-09-17 15:01:43 -0700 | [diff] [blame] | 193 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__scalar_rr2_p5_div) |
| 194 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__scalar_rr2_lut64_p2_div) |
| 195 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sigmoid__scalar_rr2_lut2048_p1_div) |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 196 | |
Marat Dukhan | 8400076 | 2020-06-29 18:38:43 -0700 | [diff] [blame] | 197 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__sse_nr1mac) |
| 198 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__sse_nr2mac) |
| 199 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__sse_hh1mac) |
| 200 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__fma3_nr1fma) |
| 201 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__fma3_nr2fma) |
| 202 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__fma3_nr1fma1adj) |
| 203 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__avx512f_nr1fma) |
| 204 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__avx512f_nr2fma) |
| 205 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__avx512f_nr1fma1adj) |
| 206 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__neon_nr1rsqrts) |
| 207 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__neon_nr2rsqrts) |
| 208 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__neon_nr3rsqrts) |
| 209 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__neonfma_nr1fma) |
| 210 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__neonfma_nr2fma) |
| 211 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__neonfma_nr3fma) |
| 212 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__neonfma_nr1rsqrts1fma1adj) |
| 213 | DECLARE_F32_UNARY_MATH_FUNCTION(xnn_math_f32_sqrt__neonfma_nr2fma1adj) |
| 214 | |
Marat Dukhan | 6adff4e | 2019-10-14 18:32:07 -0700 | [diff] [blame] | 215 | #ifdef __cplusplus |
| 216 | } /* extern "C" */ |
| 217 | #endif |