Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [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 <stddef.h> |
| 9 | #include <stdint.h> |
| 10 | |
| 11 | #include <xnnpack/params.h> |
| 12 | #include <xnnpack/common.h> |
| 13 | |
| 14 | #ifdef __cplusplus |
| 15 | extern "C" { |
| 16 | #endif |
| 17 | |
| 18 | |
| 19 | #define DECLARE_F32_VUNARY_UKERNEL_FUNCTION(fn_name) \ |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 20 | XNN_INTERNAL void fn_name( \ |
| 21 | size_t n, \ |
| 22 | const float* x, \ |
| 23 | float* y, \ |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 24 | const void* params); |
| 25 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 26 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x4) |
| 27 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x8) |
| 28 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x12) |
| 29 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16) |
| 30 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20) |
| 31 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 32 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 33 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x4) |
| 34 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x8) |
| 35 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12) |
| 36 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16) |
| 37 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20) |
| 38 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 39 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 40 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x4) |
| 41 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8) |
| 42 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x12) |
| 43 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16) |
| 44 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20) |
| 45 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 46 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 47 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x4) |
| 48 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8) |
| 49 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12) |
| 50 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16) |
| 51 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20) |
| 52 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 53 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 54 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x4) |
| 55 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8) |
| 56 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12) |
| 57 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16) |
| 58 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20) |
| 59 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 60 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 61 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x4) |
| 62 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x8) |
| 63 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12) |
| 64 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16) |
| 65 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20) |
| 66 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24) |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 67 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 68 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x4) |
| 69 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8) |
| 70 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12) |
| 71 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16) |
| 72 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20) |
| 73 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24) |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 74 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 75 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x4) |
| 76 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8) |
| 77 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12) |
| 78 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16) |
| 79 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20) |
| 80 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24) |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 81 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 82 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x4) |
| 83 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8) |
| 84 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12) |
| 85 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16) |
| 86 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20) |
| 87 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24) |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 88 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 89 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x4) |
| 90 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8) |
| 91 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x12) |
| 92 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16) |
| 93 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20) |
| 94 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x24) |
Marat Dukhan | 68b3b45 | 2020-01-02 10:11:15 -0800 | [diff] [blame] | 95 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 96 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x4) |
| 97 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x8) |
| 98 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x12) |
| 99 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16) |
| 100 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20) |
| 101 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 102 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 103 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x4) |
| 104 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8) |
| 105 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12) |
| 106 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16) |
| 107 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20) |
| 108 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 109 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 110 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x4) |
| 111 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8) |
| 112 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12) |
| 113 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16) |
| 114 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20) |
| 115 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 116 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 117 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x4) |
| 118 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8) |
| 119 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12) |
| 120 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16) |
| 121 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20) |
| 122 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 123 | |
Marat Dukhan | 4a24a58 | 2020-01-06 13:30:00 -0800 | [diff] [blame] | 124 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x4) |
| 125 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8) |
| 126 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x12) |
| 127 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16) |
| 128 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20) |
| 129 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24) |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 130 | |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 131 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__neon_frac_p9_p10_nr1recps_x16) |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 132 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 133 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_p5_div_x4) |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 134 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_p5_div_x8) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 135 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_p5_div_x12) |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 136 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_p5_div_x16) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 137 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_p5_div_x20) |
| 138 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_p5_div_x24) |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 139 | |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 140 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_p5_div_x4) |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 141 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_p5_div_x8) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 142 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_p5_div_x12) |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 143 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_p5_div_x16) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 144 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_p5_div_x20) |
| 145 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_p5_div_x24) |
| 146 | |
Marat Dukhan | d243c1a | 2020-09-17 18:48:10 -0700 | [diff] [blame] | 147 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x4) |
| 148 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x8) |
| 149 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x12) |
| 150 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x16) |
| 151 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x20) |
| 152 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse2_lut64_p2_div_x24) |
| 153 | |
| 154 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x4) |
| 155 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x8) |
| 156 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12) |
| 157 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x16) |
| 158 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x20) |
| 159 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x24) |
| 160 | |
T.J. Alumbaugh | dc2b29c | 2020-10-14 13:56:08 -0700 | [diff] [blame] | 161 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x8) |
| 162 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x16) |
| 163 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x24) |
| 164 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x32) |
| 165 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x40) |
| 166 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x48) |
| 167 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56) |
| 168 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64) |
| 169 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x72) |
| 170 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80) |
| 171 | |
| 172 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x8) |
| 173 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x16) |
| 174 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x24) |
| 175 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x32) |
| 176 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x40) |
| 177 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x48) |
| 178 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x56) |
| 179 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x64) |
| 180 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x72) |
| 181 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x80) |
| 182 | |
Marat Dukhan | fa0a432 | 2020-01-06 16:14:29 -0800 | [diff] [blame] | 183 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x8) |
| 184 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x16) |
| 185 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x24) |
| 186 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x32) |
| 187 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x40) |
| 188 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x48) |
| 189 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x56) |
| 190 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x64) |
| 191 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x72) |
| 192 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x80) |
| 193 | |
| 194 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x8) |
| 195 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x16) |
| 196 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x24) |
| 197 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x32) |
| 198 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x40) |
| 199 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x48) |
| 200 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x56) |
| 201 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x64) |
| 202 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x72) |
| 203 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x80) |
| 204 | |
| 205 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x8) |
| 206 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x16) |
| 207 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x24) |
| 208 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x32) |
| 209 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x40) |
| 210 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x48) |
| 211 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x56) |
| 212 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x64) |
| 213 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x72) |
| 214 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x80) |
| 215 | |
Marat Dukhan | d9ca7e6 | 2020-09-23 23:45:29 -0700 | [diff] [blame] | 216 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x16) |
| 217 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x32) |
| 218 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x48) |
| 219 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x64) |
| 220 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x80) |
| 221 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x96) |
| 222 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112) |
| 223 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128) |
| 224 | |
| 225 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x16) |
| 226 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x32) |
| 227 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x48) |
| 228 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x64) |
| 229 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x80) |
| 230 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x96) |
| 231 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112) |
| 232 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128) |
| 233 | |
| 234 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x16) |
| 235 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x32) |
| 236 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48) |
| 237 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64) |
| 238 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80) |
| 239 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96) |
| 240 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112) |
| 241 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128) |
| 242 | |
| 243 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x16) |
| 244 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x32) |
| 245 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48) |
| 246 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64) |
| 247 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80) |
| 248 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96) |
| 249 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112) |
| 250 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128) |
| 251 | |
| 252 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x16) |
| 253 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x32) |
| 254 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x48) |
| 255 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64) |
| 256 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x80) |
| 257 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96) |
| 258 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112) |
| 259 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128) |
| 260 | |
| 261 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x16) |
| 262 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x32) |
| 263 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x48) |
| 264 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x64) |
| 265 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x80) |
| 266 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x96) |
| 267 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112) |
| 268 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128) |
| 269 | |
Marat Dukhan | b3635ed | 2020-07-16 12:36:28 -0700 | [diff] [blame] | 270 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x4) |
| 271 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x8) |
| 272 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x12) |
| 273 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x16) |
| 274 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x20) |
| 275 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_p5_div_x24) |
| 276 | |
Marat Dukhan | d187a5b | 2020-07-20 01:07:17 -0700 | [diff] [blame] | 277 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x4) |
| 278 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x8) |
| 279 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x12) |
| 280 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x16) |
| 281 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x20) |
| 282 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__wasmsimd_lut64_p2_div_x24) |
| 283 | |
Marat Dukhan | 3a77ea7 | 2019-12-23 12:10:24 -0800 | [diff] [blame] | 284 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x1) |
| 285 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x2) |
| 286 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__scalar_lut2048_p1_div_x4) |
| 287 | |
| 288 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x1) |
| 289 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x2) |
| 290 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__scalar_lut64_p2_div_x4) |
| 291 | |
| 292 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__scalar_p5_div_x1) |
| 293 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__scalar_p5_div_x2) |
| 294 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_sigmoid_ukernel__scalar_p5_div_x4) |
| 295 | |
Marat Dukhan | 2b9efd8 | 2020-06-08 01:09:31 -0700 | [diff] [blame] | 296 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__neon_x4) |
| 297 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__neon_x8) |
| 298 | |
| 299 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__sse_x4) |
| 300 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__sse_x8) |
| 301 | |
| 302 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__avx_x8) |
| 303 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__avx_x16) |
| 304 | |
| 305 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__avx512f_x16) |
| 306 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__avx512f_x32) |
| 307 | |
Marat Dukhan | 37c8351 | 2020-06-29 13:25:53 -0700 | [diff] [blame] | 308 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__wasmsimd_x4) |
| 309 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__wasmsimd_x8) |
| 310 | |
Marat Dukhan | 2b9efd8 | 2020-06-08 01:09:31 -0700 | [diff] [blame] | 311 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__scalar_x1) |
| 312 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__scalar_x2) |
| 313 | DECLARE_F32_VUNARY_UKERNEL_FUNCTION(xnn_f32_vsqr_ukernel__scalar_x4) |
| 314 | |
| 315 | |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 316 | #define DECLARE_F32_VSQRT_UKERNEL_FUNCTION(fn_name) \ |
| 317 | XNN_INTERNAL void fn_name( \ |
| 318 | size_t n, \ |
| 319 | const float* x, \ |
| 320 | float* y, \ |
| 321 | const union xnn_f32_sqrt_params* params); |
| 322 | |
| 323 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neon_sqrt_x4) |
| 324 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neon_sqrt_x8) |
| 325 | |
| 326 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x4) |
| 327 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8) |
| 328 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12) |
| 329 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16) |
| 330 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20) |
| 331 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24) |
| 332 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28) |
| 333 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32) |
| 334 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36) |
| 335 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40) |
| 336 | |
| 337 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x4) |
| 338 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8) |
| 339 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12) |
| 340 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16) |
| 341 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20) |
| 342 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24) |
| 343 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28) |
| 344 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32) |
| 345 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36) |
| 346 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40) |
| 347 | |
| 348 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__sse_sqrt_x4) |
| 349 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__sse_sqrt_x8) |
| 350 | |
| 351 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__avx_sqrt_x8) |
| 352 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__avx_sqrt_x16) |
| 353 | |
| 354 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8) |
| 355 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16) |
| 356 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24) |
| 357 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32) |
| 358 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40) |
| 359 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48) |
| 360 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56) |
| 361 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64) |
| 362 | |
| 363 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16) |
| 364 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32) |
| 365 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48) |
| 366 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64) |
| 367 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80) |
| 368 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96) |
| 369 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112) |
| 370 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128) |
| 371 | |
| 372 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x4) |
| 373 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x8) |
| 374 | |
| 375 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__scalar_sqrt_x1) |
| 376 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__scalar_sqrt_x2) |
| 377 | DECLARE_F32_VSQRT_UKERNEL_FUNCTION(xnn_f32_vsqrt_ukernel__scalar_sqrt_x4) |
| 378 | |
| 379 | |
Marat Dukhan | 2b9efd8 | 2020-06-08 01:09:31 -0700 | [diff] [blame] | 380 | #define DECLARE_F32_VABS_UKERNEL_FUNCTION(fn_name) \ |
| 381 | XNN_INTERNAL void fn_name( \ |
| 382 | size_t n, \ |
| 383 | const float* x, \ |
| 384 | float* y, \ |
| 385 | const union xnn_f32_abs_params* params); |
| 386 | |
| 387 | |
| 388 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__neon_x4) |
| 389 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__neon_x8) |
| 390 | |
| 391 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__sse_x4) |
| 392 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__sse_x8) |
| 393 | |
| 394 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__avx_x8) |
| 395 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__avx_x16) |
| 396 | |
| 397 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__avx512f_x16) |
| 398 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__avx512f_x32) |
| 399 | |
Marat Dukhan | 37c8351 | 2020-06-29 13:25:53 -0700 | [diff] [blame] | 400 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__wasmsimd_x4) |
| 401 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__wasmsimd_x8) |
| 402 | |
Marat Dukhan | 2b9efd8 | 2020-06-08 01:09:31 -0700 | [diff] [blame] | 403 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__scalar_x1) |
| 404 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__scalar_x2) |
| 405 | DECLARE_F32_VABS_UKERNEL_FUNCTION(xnn_f32_vabs_ukernel__scalar_x4) |
| 406 | |
| 407 | |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 408 | #define DECLARE_F32_VLRELU_UKERNEL_FUNCTION(fn_name) \ |
| 409 | XNN_INTERNAL void fn_name( \ |
| 410 | size_t n, \ |
| 411 | const float* x, \ |
| 412 | float* y, \ |
| 413 | const union xnn_f32_lrelu_params* params); |
| 414 | |
| 415 | |
| 416 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__neon_x4) |
| 417 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__neon_x8) |
| 418 | |
| 419 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__sse_x4) |
| 420 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__sse_x8) |
| 421 | |
Marat Dukhan | 0d3f467 | 2020-06-25 16:42:58 -0700 | [diff] [blame] | 422 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__sse2_x4) |
| 423 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__sse2_x8) |
| 424 | |
| 425 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__sse41_x4) |
| 426 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__sse41_x8) |
| 427 | |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 428 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__avx_x8) |
| 429 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__avx_x16) |
| 430 | |
| 431 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__avx512f_x16) |
| 432 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__avx512f_x32) |
| 433 | |
Marat Dukhan | e6502da | 2020-07-15 16:31:36 -0700 | [diff] [blame] | 434 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__wasmsimd_bitselect_x4) |
| 435 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__wasmsimd_bitselect_x8) |
| 436 | |
| 437 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__wasmsimd_minmax_x4) |
| 438 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__wasmsimd_minmax_x8) |
| 439 | |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 440 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__wasm_x1) |
| 441 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__wasm_x2) |
| 442 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__wasm_x4) |
| 443 | |
| 444 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__scalar_x1) |
| 445 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__scalar_x2) |
| 446 | DECLARE_F32_VLRELU_UKERNEL_FUNCTION(xnn_f32_vlrelu_ukernel__scalar_x4) |
| 447 | |
| 448 | |
Marat Dukhan | 2b9efd8 | 2020-06-08 01:09:31 -0700 | [diff] [blame] | 449 | #define DECLARE_F32_VNEG_UKERNEL_FUNCTION(fn_name) \ |
| 450 | XNN_INTERNAL void fn_name( \ |
| 451 | size_t n, \ |
| 452 | const float* x, \ |
| 453 | float* y, \ |
| 454 | const union xnn_f32_neg_params* params); |
| 455 | |
| 456 | |
| 457 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__neon_x4) |
| 458 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__neon_x8) |
| 459 | |
| 460 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__sse_x4) |
| 461 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__sse_x8) |
| 462 | |
| 463 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__avx_x8) |
| 464 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__avx_x16) |
| 465 | |
| 466 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__avx512f_x16) |
| 467 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__avx512f_x32) |
| 468 | |
Marat Dukhan | 37c8351 | 2020-06-29 13:25:53 -0700 | [diff] [blame] | 469 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__wasmsimd_x4) |
| 470 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__wasmsimd_x8) |
| 471 | |
Marat Dukhan | 2b9efd8 | 2020-06-08 01:09:31 -0700 | [diff] [blame] | 472 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__scalar_x1) |
| 473 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__scalar_x2) |
| 474 | DECLARE_F32_VNEG_UKERNEL_FUNCTION(xnn_f32_vneg_ukernel__scalar_x4) |
| 475 | |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 476 | |
| 477 | #define DECLARE_F32_VRND_UKERNEL_FUNCTION(fn_name) \ |
| 478 | XNN_INTERNAL void fn_name( \ |
| 479 | size_t n, \ |
| 480 | const float* x, \ |
| 481 | float* y, \ |
| 482 | const union xnn_f32_rnd_params* params); |
| 483 | |
| 484 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__neon_x4) |
| 485 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__neon_x8) |
| 486 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__neonv8_x4) |
| 487 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__neonv8_x8) |
| 488 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__sse2_x4) |
| 489 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__sse2_x8) |
| 490 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__sse41_x4) |
| 491 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__sse41_x8) |
| 492 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__avx_x8) |
| 493 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__avx_x16) |
| 494 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__avx512f_x16) |
| 495 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__avx512f_x32) |
Marat Dukhan | b82b2cd | 2020-07-16 02:23:42 -0700 | [diff] [blame] | 496 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__wasmsimd_addsub_x4) |
| 497 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__wasmsimd_addsub_x8) |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 498 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__scalar_libm_x1) |
| 499 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__scalar_libm_x2) |
| 500 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndne_ukernel__scalar_libm_x4) |
| 501 | |
| 502 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__neon_x4) |
| 503 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__neon_x8) |
| 504 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__neonv8_x4) |
| 505 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__neonv8_x8) |
| 506 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__sse2_x4) |
| 507 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__sse2_x8) |
| 508 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__sse41_x4) |
| 509 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__sse41_x8) |
| 510 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__avx_x8) |
| 511 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__avx_x16) |
| 512 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__avx512f_x16) |
| 513 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__avx512f_x32) |
Marat Dukhan | b82b2cd | 2020-07-16 02:23:42 -0700 | [diff] [blame] | 514 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__wasmsimd_addsub_x4) |
| 515 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__wasmsimd_addsub_x8) |
| 516 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__wasmsimd_cvt_x4) |
| 517 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__wasmsimd_cvt_x8) |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 518 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__scalar_libm_x1) |
| 519 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__scalar_libm_x2) |
| 520 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndz_ukernel__scalar_libm_x4) |
| 521 | |
| 522 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__neon_x4) |
| 523 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__neon_x8) |
| 524 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__neonv8_x4) |
| 525 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__neonv8_x8) |
| 526 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__sse2_x4) |
| 527 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__sse2_x8) |
| 528 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__sse41_x4) |
| 529 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__sse41_x8) |
| 530 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__avx_x8) |
| 531 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__avx_x16) |
| 532 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__avx512f_x16) |
| 533 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__avx512f_x32) |
Marat Dukhan | b82b2cd | 2020-07-16 02:23:42 -0700 | [diff] [blame] | 534 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__wasmsimd_addsub_x4) |
| 535 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__wasmsimd_addsub_x8) |
| 536 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__wasmsimd_cvt_x4) |
| 537 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__wasmsimd_cvt_x8) |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 538 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__scalar_libm_x1) |
| 539 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__scalar_libm_x2) |
| 540 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndu_ukernel__scalar_libm_x4) |
| 541 | |
| 542 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__neon_x4) |
| 543 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__neon_x8) |
| 544 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__neonv8_x4) |
| 545 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__neonv8_x8) |
| 546 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__sse2_x4) |
| 547 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__sse2_x8) |
| 548 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__sse41_x4) |
| 549 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__sse41_x8) |
| 550 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__avx_x8) |
| 551 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__avx_x16) |
| 552 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__avx512f_x16) |
| 553 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__avx512f_x32) |
Marat Dukhan | b82b2cd | 2020-07-16 02:23:42 -0700 | [diff] [blame] | 554 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__wasmsimd_addsub_x4) |
| 555 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__wasmsimd_addsub_x8) |
| 556 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__wasmsimd_cvt_x4) |
| 557 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__wasmsimd_cvt_x8) |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 558 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__scalar_libm_x1) |
| 559 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__scalar_libm_x2) |
| 560 | DECLARE_F32_VRND_UKERNEL_FUNCTION(xnn_f32_vrndd_ukernel__scalar_libm_x4) |
| 561 | |
Marat Dukhan | b82b2cd | 2020-07-16 02:23:42 -0700 | [diff] [blame] | 562 | #define DECLARE_F16_RELU_UKERNEL_FUNCTION(fn_name) \ |
| 563 | XNN_INTERNAL void fn_name( \ |
| 564 | size_t n, \ |
| 565 | const void* x, \ |
| 566 | void* y, \ |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 567 | const struct xnn_f16_relu_params* params); |
| 568 | |
| 569 | DECLARE_F16_RELU_UKERNEL_FUNCTION(xnn_f16_relu_ukernel__neonfp16arith_x8) |
| 570 | DECLARE_F16_RELU_UKERNEL_FUNCTION(xnn_f16_relu_ukernel__neonfp16arith_x16) |
| 571 | |
Marat Dukhan | b82b2cd | 2020-07-16 02:23:42 -0700 | [diff] [blame] | 572 | #define DECLARE_F32_RELU_UKERNEL_FUNCTION(fn_name) \ |
| 573 | XNN_INTERNAL void fn_name( \ |
| 574 | size_t n, \ |
| 575 | const float* x, \ |
| 576 | float* y, \ |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 577 | const union xnn_f32_relu_params* params); |
| 578 | |
| 579 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__neon_x4) |
| 580 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__neon_x8) |
| 581 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__sse_x4) |
| 582 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__sse_x8) |
| 583 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__avx_x8) |
| 584 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__avx_x16) |
| 585 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__avx512f_x16) |
| 586 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__avx512f_x32) |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 587 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__scalar_x1) |
| 588 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__scalar_x2) |
| 589 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__scalar_x4) |
Frank Barchard | 37297a6 | 2020-08-27 12:15:10 -0700 | [diff] [blame] | 590 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__scalar_x8) |
| 591 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__wasmsimd_x4) |
| 592 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__wasmsimd_x8) |
| 593 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__wasmsimd_x16) |
| 594 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__wasm_x1) |
| 595 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__wasm_x2) |
| 596 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__wasm_x4) |
| 597 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__wasm_x8) |
Frank Barchard | d91624e | 2020-08-12 13:45:56 -0700 | [diff] [blame] | 598 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__wasm32_shr_x1) |
Frank Barchard | 9f5ccb3 | 2020-08-17 14:00:51 -0700 | [diff] [blame] | 599 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__wasm32_shr_x2) |
| 600 | DECLARE_F32_RELU_UKERNEL_FUNCTION(xnn_f32_relu_ukernel__wasm32_shr_x4) |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 601 | |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 602 | #ifdef __cplusplus |
| 603 | } // extern "C" |
| 604 | #endif |