Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -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 <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 | #define DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(fn_name) \ |
| 19 | XNN_INTERNAL void fn_name( \ |
| 20 | size_t n, \ |
| 21 | const float* input, \ |
| 22 | float* output, \ |
| 23 | float* sum, \ |
| 24 | float max); |
| 25 | |
Marat Dukhan | 8137e4c | 2020-01-25 12:56:58 -0800 | [diff] [blame] | 26 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x4) |
| 27 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8) |
| 28 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2) |
| 29 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12) |
| 30 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2) |
| 31 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3) |
| 32 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16) |
| 33 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2) |
| 34 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc4) |
| 35 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20) |
| 36 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc2) |
| 37 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x20_acc5) |
| 38 | |
| 39 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x4) |
| 40 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8) |
| 41 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x8_acc2) |
| 42 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12) |
| 43 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc2) |
| 44 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x12_acc3) |
| 45 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16) |
| 46 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc2) |
| 47 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x16_acc4) |
| 48 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20) |
| 49 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc2) |
| 50 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x20_acc5) |
| 51 | |
| 52 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x4) |
| 53 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8) |
| 54 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2) |
| 55 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12) |
| 56 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2) |
| 57 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3) |
| 58 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16) |
| 59 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2) |
| 60 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4) |
| 61 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20) |
| 62 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc2) |
| 63 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x20_acc5) |
| 64 | |
| 65 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x4) |
| 66 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8) |
| 67 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x8_acc2) |
| 68 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12) |
| 69 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc2) |
| 70 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x12_acc3) |
| 71 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16) |
| 72 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc2) |
| 73 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x16_acc4) |
| 74 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20) |
| 75 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc2) |
| 76 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x20_acc5) |
| 77 | |
Marat Dukhan | b39689d | 2020-01-24 13:32:20 -0800 | [diff] [blame] | 78 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x4) |
| 79 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8) |
| 80 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x8_acc2) |
| 81 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12) |
| 82 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc2) |
| 83 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x12_acc3) |
| 84 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16) |
| 85 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc2) |
| 86 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x16_acc4) |
| 87 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20) |
| 88 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc2) |
| 89 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__sse2_p5_x20_acc5) |
| 90 | |
Marat Dukhan | 4c4eb00 | 2019-12-08 21:27:49 -0800 | [diff] [blame] | 91 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64) |
| 92 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc2) |
| 93 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x64_acc4) |
| 94 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72) |
| 95 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x72_acc3) |
| 96 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80) |
| 97 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc2) |
| 98 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x80_acc5) |
| 99 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96) |
| 100 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc2) |
| 101 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc3) |
| 102 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6) |
| 103 | |
| 104 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x128) |
| 105 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x128_acc2) |
| 106 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4) |
| 107 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144) |
| 108 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144_acc3) |
| 109 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160) |
| 110 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc2) |
| 111 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5) |
| 112 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192) |
| 113 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2) |
| 114 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3) |
| 115 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6) |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 116 | |
Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 117 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x4) |
| 118 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8) |
| 119 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8_acc2) |
| 120 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x12) |
| 121 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x12_acc2) |
| 122 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x12_acc3) |
| 123 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16) |
| 124 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc2) |
| 125 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x16_acc4) |
| 126 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20) |
| 127 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc2) |
| 128 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x20_acc5) |
| 129 | |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 130 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x1) |
| 131 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x2) |
| 132 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x2_acc2) |
| 133 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x4) |
| 134 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x4_acc2) |
| 135 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x4_acc4) |
| 136 | |
| 137 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x1) |
| 138 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x2) |
| 139 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x2_acc2) |
| 140 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x4) |
| 141 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x4_acc2) |
| 142 | DECLARE_F32_RADDSTOREEXPMINUSMAX_UKERNEL_FUNCTION(xnn_f32_raddstoreexpminusmax_ukernel__scalar_lut64_p2_x4_acc4) |
| 143 | |
Marat Dukhan | 9757953 | 2019-10-18 16:40:39 -0700 | [diff] [blame] | 144 | #ifdef __cplusplus |
| 145 | } /* extern "C" */ |
| 146 | #endif |