XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #pragma once |
| 10 | |
| 11 | #include <stddef.h> |
| 12 | #include <stdint.h> |
| 13 | |
| 14 | #include <xnnpack/params.h> |
| 15 | #include <xnnpack/common.h> |
| 16 | |
| 17 | #ifdef __cplusplus |
| 18 | extern "C" { |
| 19 | #endif |
| 20 | |
| 21 | |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 22 | #define DECLARE_F32_GEMM_UKERNEL_FUNCTION(fn_name) \ |
| 23 | XNN_INTERNAL void fn_name( \ |
| 24 | size_t mr, \ |
| 25 | size_t nr, \ |
| 26 | size_t k, \ |
| 27 | const float* a, \ |
| 28 | size_t a_stride, \ |
| 29 | const float* w, \ |
| 30 | float* c, \ |
| 31 | size_t cm_stride, \ |
| 32 | size_t cn_stride, \ |
| 33 | const union xnn_f32_default_params* params); |
| 34 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 35 | #define DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(fn_name) \ |
| 36 | XNN_INTERNAL void fn_name( \ |
| 37 | size_t mr, \ |
| 38 | size_t nr, \ |
| 39 | size_t k, \ |
| 40 | const float* a, \ |
| 41 | size_t a_stride, \ |
| 42 | const float* w, \ |
| 43 | float* c, \ |
| 44 | size_t cm_stride, \ |
| 45 | size_t cn_stride, \ |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 46 | const union xnn_f32_minmax_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 47 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 48 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x2__neon_lane_ld64) |
| 49 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64) |
| 50 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64) |
| 51 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld128) |
| 52 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64) |
| 53 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64) |
| 54 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld128) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 55 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 56 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x2__neonfma_lane_ld64) |
| 57 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__neonfma_lane_ld64) |
| 58 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64) |
| 59 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld128) |
| 60 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64) |
| 61 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64) |
| 62 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld128) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 63 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 64 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__neon_dup_ld64) |
| 65 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neon_dup_ld64) |
| 66 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neon_dup_ld128) |
| 67 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64) |
| 68 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld128) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 69 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 70 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__neonfma_dup_ld64) |
| 71 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64) |
| 72 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld128) |
| 73 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64) |
| 74 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld128) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 75 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 76 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8s4__neon) |
| 77 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8s4__neon) |
| 78 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8s4__neon) |
| 79 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_8x8s4__neon) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 80 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 81 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8s4__neonfma) |
| 82 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8s4__neonfma) |
| 83 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8s4__neonfma) |
| 84 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_8x8s4__neonfma) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 85 | |
Marat Dukhan | 32f9381 | 2020-05-17 20:31:21 -0700 | [diff] [blame] | 86 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x4__aarch32_vfp_ld64) |
| 87 | |
Marat Dukhan | 3b98f6b | 2020-05-17 10:09:22 -0700 | [diff] [blame] | 88 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x4__aarch32_vfp_ld64) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 89 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_ld64) |
| 90 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53) |
| 91 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55) |
| 92 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a75) |
| 93 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_pld_cortex_a75) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 94 | |
Frank Barchard | 3cb54f9 | 2020-04-10 10:46:08 -0700 | [diff] [blame] | 95 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_ld64) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 96 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x12__aarch64_neonfma_cortex_a53) |
| 97 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53) |
| 98 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a57) |
| 99 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75) |
| 100 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x12__aarch64_neonfma_cortex_a53) |
| 101 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53) |
| 102 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55) |
| 103 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a57) |
| 104 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75) |
| 105 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_ld128) |
| 106 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_ld64) |
| 107 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__aarch64_neonfma_cortex_a57) |
| 108 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__aarch64_neonfma_cortex_a75) |
| 109 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53) |
| 110 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55) |
| 111 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73) |
| 112 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a57) |
| 113 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75) |
| 114 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_ios) |
| 115 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_ld128) |
| 116 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_ld64) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 117 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 118 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__sse_load1) |
| 119 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__sse_load1) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 120 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 121 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__sse_dup) |
| 122 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__sse_dup) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 123 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 124 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8s4__sse) |
| 125 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8s4__sse) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 126 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 127 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x2c4__sse) |
Marat Dukhan | b00004d | 2020-02-13 09:28:35 -0800 | [diff] [blame] | 128 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 129 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__avx_broadcast) |
| 130 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__avx_broadcast) |
| 131 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__avx_broadcast) |
| 132 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__avx_broadcast) |
| 133 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_7x8__avx_broadcast) |
| 134 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x16__avx_broadcast) |
| 135 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x16__avx_broadcast) |
| 136 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x16__avx_broadcast) |
| 137 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x16__avx_broadcast) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 138 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 139 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__fma3_broadcast) |
| 140 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__fma3_broadcast) |
| 141 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__fma3_broadcast) |
| 142 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__fma3_broadcast) |
| 143 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_7x8__fma3_broadcast) |
| 144 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_8x8__fma3_broadcast) |
| 145 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x16__fma3_broadcast) |
| 146 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x16__fma3_broadcast) |
| 147 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x16__fma3_broadcast) |
| 148 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x16__fma3_broadcast) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 149 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 150 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast) |
| 151 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast) |
| 152 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast) |
| 153 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x16s4__fma3_broadcast) |
Marat Dukhan | 2712132 | 2019-12-09 14:57:40 -0800 | [diff] [blame] | 154 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 155 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x16__avx512f_broadcast) |
| 156 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast) |
| 157 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast) |
| 158 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast) |
| 159 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast) |
| 160 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 161 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 162 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__psimd_loadsplat) |
| 163 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__psimd_loadsplat) |
| 164 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__psimd_loadsplat) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 165 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 166 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__psimd_splat) |
| 167 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__psimd_splat) |
| 168 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__psimd_splat) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 169 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 170 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8s4__psimd) |
| 171 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8s4__psimd) |
| 172 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8s4__psimd) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 173 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 174 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x2c4__psimd) |
Marat Dukhan | b00004d | 2020-02-13 09:28:35 -0800 | [diff] [blame] | 175 | |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 176 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x2__wasm) |
| 177 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x4__wasm) |
| 178 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_2x4__wasm) |
| 179 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x4__wasm) |
| 180 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 181 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x2__wasm) |
| 182 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x4__wasm) |
| 183 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_2x4__wasm) |
| 184 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x4__wasm) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 185 | |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 186 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x2__scalar) |
| 187 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x4__scalar) |
| 188 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_2x4__scalar) |
| 189 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x4__scalar) |
| 190 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 191 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x2__scalar) |
| 192 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x4__scalar) |
| 193 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_2x4__scalar) |
| 194 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x4__scalar) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 195 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 196 | #define DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(fn_name) \ |
| 197 | XNN_INTERNAL void fn_name( \ |
| 198 | size_t mr, \ |
| 199 | size_t nr, \ |
| 200 | size_t k, \ |
| 201 | const float* a, \ |
| 202 | size_t a_stride, \ |
| 203 | const float* w, \ |
| 204 | float* c, \ |
| 205 | size_t cm_stride, \ |
| 206 | size_t cn_stride, \ |
| 207 | const float* acc, \ |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 208 | const union xnn_f32_minmax_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 209 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 210 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__neon_lane_ld64) |
| 211 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neon_lane_ld64) |
| 212 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neon_lane_ld128) |
| 213 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64) |
| 214 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64) |
| 215 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld128) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 216 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 217 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__neonfma_lane_ld64) |
| 218 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64) |
| 219 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld128) |
| 220 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64) |
| 221 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64) |
| 222 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld128) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 223 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 224 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__neon_dup_ld64) |
| 225 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64) |
| 226 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld128) |
| 227 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64) |
| 228 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld128) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 229 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 230 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__neonfma_dup_ld64) |
| 231 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_dup_ld64) |
| 232 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_dup_ld128) |
| 233 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64) |
| 234 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld128) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 235 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 236 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8s4__neon) |
| 237 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8s4__neon) |
| 238 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8s4__neon) |
| 239 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_8x8s4__neon) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 240 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 241 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8s4__neonfma) |
| 242 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8s4__neonfma) |
| 243 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8s4__neonfma) |
| 244 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_8x8s4__neonfma) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 245 | |
Frank Barchard | 3cb54f9 | 2020-04-10 10:46:08 -0700 | [diff] [blame] | 246 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__aarch64_neonfma_ld64) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 247 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x12__aarch64_neonfma_cortex_a53) |
| 248 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53) |
| 249 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__aarch64_neonfma_cortex_a57) |
| 250 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75) |
| 251 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x12__aarch64_neonfma_cortex_a53) |
| 252 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53) |
| 253 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55) |
| 254 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_cortex_a57) |
| 255 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75) |
| 256 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_ld128) |
| 257 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_ld64) |
| 258 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__aarch64_neonfma_cortex_a57) |
| 259 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__aarch64_neonfma_cortex_a75) |
| 260 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53) |
| 261 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55) |
| 262 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73) |
| 263 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_cortex_a57) |
| 264 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75) |
| 265 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_ios) |
| 266 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_ld128) |
| 267 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_ld64) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 268 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 269 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__sse_load1) |
| 270 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__sse_load1) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 271 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 272 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__sse_dup) |
| 273 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 274 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 275 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8s4__sse) |
| 276 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8s4__sse) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 277 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 278 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__avx_broadcast) |
| 279 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__avx_broadcast) |
| 280 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__avx_broadcast) |
| 281 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__avx_broadcast) |
| 282 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_7x8__avx_broadcast) |
| 283 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x16__avx_broadcast) |
| 284 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x16__avx_broadcast) |
| 285 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x16__avx_broadcast) |
| 286 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x16__avx_broadcast) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 287 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 288 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__fma3_broadcast) |
| 289 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__fma3_broadcast) |
| 290 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__fma3_broadcast) |
| 291 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__fma3_broadcast) |
| 292 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_7x8__fma3_broadcast) |
| 293 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_8x8__fma3_broadcast) |
| 294 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x16__fma3_broadcast) |
| 295 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x16__fma3_broadcast) |
| 296 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x16__fma3_broadcast) |
| 297 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x16__fma3_broadcast) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 298 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 299 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast) |
| 300 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast) |
| 301 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast) |
| 302 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x16s4__fma3_broadcast) |
Marat Dukhan | 2712132 | 2019-12-09 14:57:40 -0800 | [diff] [blame] | 303 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 304 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x16__avx512f_broadcast) |
| 305 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast) |
| 306 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast) |
| 307 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast) |
| 308 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast) |
| 309 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 310 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 311 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__psimd_loadsplat) |
| 312 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__psimd_loadsplat) |
| 313 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__psimd_loadsplat) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 314 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 315 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__psimd_splat) |
| 316 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__psimd_splat) |
| 317 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__psimd_splat) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 318 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 319 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8s4__psimd) |
| 320 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8s4__psimd) |
| 321 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8s4__psimd) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 322 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 323 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x4__wasm) |
| 324 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_2x4__wasm) |
| 325 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x2__wasm) |
| 326 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x4__wasm) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 327 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 328 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x4__scalar) |
| 329 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_2x4__scalar) |
| 330 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x4__scalar) |
Marat Dukhan | 436ebe6 | 2019-12-04 15:10:12 -0800 | [diff] [blame] | 331 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 332 | #define DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(fn_name) \ |
| 333 | void fn_name( \ |
| 334 | size_t mr, \ |
| 335 | size_t nr, \ |
| 336 | size_t k, \ |
| 337 | const void* a, \ |
| 338 | size_t a_stride, \ |
| 339 | const void* w, \ |
| 340 | void* c, \ |
| 341 | size_t cm_stride, \ |
| 342 | size_t cn_stride, \ |
Frank Barchard | 875be77 | 2020-04-16 09:50:09 -0700 | [diff] [blame] | 343 | const struct xnn_f16_scaleminmax_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 344 | |
Frank Barchard | 36b76b6 | 2020-04-10 12:39:17 -0700 | [diff] [blame] | 345 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld32) |
| 346 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32) |
| 347 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32) |
Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 348 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_1x8__aarch64_neonfp16arith_ld64) |
| 349 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_4x8__aarch64_neonfp16arith_ld64) |
| 350 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_6x8__aarch64_neonfp16arith_ld64) |
Frank Barchard | 3b8e566 | 2020-04-20 12:12:53 -0700 | [diff] [blame] | 351 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_8x8__aarch64_neonfp16arith_ld64) |
Frank Barchard | 1f4e461 | 2020-04-13 18:24:54 -0700 | [diff] [blame] | 352 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_1x8__neonfp16arith_ld64) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 353 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64) |
| 354 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64) |
| 355 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64) |
Frank Barchard | 3f9f99f | 2020-05-06 01:12:04 -0700 | [diff] [blame] | 356 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64) |
| 357 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64) |
| 358 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64) |
| 359 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 360 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 361 | #define DECLARE_Q8_GEMM_MINMAX_UKERNEL_FUNCTION(fn_name) \ |
| 362 | XNN_INTERNAL void fn_name( \ |
| 363 | size_t mr, \ |
| 364 | size_t nr, \ |
| 365 | size_t k, \ |
| 366 | const uint8_t* a, \ |
| 367 | size_t a_stride, \ |
| 368 | const void* w, \ |
| 369 | uint8_t* c, \ |
| 370 | size_t cm_stride, \ |
| 371 | size_t cn_stride, \ |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 372 | const union xnn_q8_gemm_params* params); |
| 373 | |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 374 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 375 | DECLARE_Q8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_q8_gemm_minmax_ukernel_4x8__neon) |
| 376 | DECLARE_Q8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_q8_gemm_minmax_ukernel_8x8__neon) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 377 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 378 | DECLARE_Q8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_q8_gemm_minmax_ukernel_4x8__aarch32_neon) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 379 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 380 | DECLARE_Q8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_q8_gemm_minmax_ukernel_8x8__aarch64_neon) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 381 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 382 | DECLARE_Q8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_q8_gemm_minmax_ukernel_2x4c8__sse2) |
| 383 | DECLARE_Q8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_q8_gemm_minmax_ukernel_4x4c2__sse2) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 384 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 385 | DECLARE_Q8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_q8_gemm_minmax_ukernel_2x2__scalar) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 386 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 387 | #ifdef __cplusplus |
Marat Dukhan | 80fc932 | 2019-09-29 21:06:36 -0700 | [diff] [blame] | 388 | } // extern "C" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 389 | #endif |