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 | |
Zhi An Ng | b43b47a | 2021-12-23 16:27:22 -0800 | [diff] [blame] | 14 | #include <xnnpack/allocator.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 15 | #include <xnnpack/common.h> |
Zhi An Ng | b43b47a | 2021-12-23 16:27:22 -0800 | [diff] [blame] | 16 | #include <xnnpack/params.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 17 | |
| 18 | #ifdef __cplusplus |
| 19 | extern "C" { |
| 20 | #endif |
| 21 | |
| 22 | |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 23 | #define DECLARE_F32_GEMM_UKERNEL_FUNCTION(fn_name) \ |
| 24 | XNN_INTERNAL void fn_name( \ |
| 25 | size_t mr, \ |
| 26 | size_t nr, \ |
| 27 | size_t k, \ |
| 28 | const float* a, \ |
| 29 | size_t a_stride, \ |
| 30 | const float* w, \ |
| 31 | float* c, \ |
| 32 | size_t cm_stride, \ |
| 33 | size_t cn_stride, \ |
| 34 | const union xnn_f32_default_params* params); |
| 35 | |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 36 | #define DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(fn_name) \ |
| 37 | XNN_INTERNAL void fn_name( \ |
| 38 | size_t mr, \ |
| 39 | size_t nr, \ |
| 40 | size_t k, \ |
| 41 | const float* a, \ |
| 42 | size_t a_stride, \ |
| 43 | const float* w, \ |
| 44 | float* c, \ |
| 45 | size_t cm_stride, \ |
| 46 | size_t cn_stride, \ |
| 47 | const union xnn_f32_relu_params* params); |
| 48 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 49 | #define DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(fn_name) \ |
| 50 | XNN_INTERNAL void fn_name( \ |
| 51 | size_t mr, \ |
| 52 | size_t nr, \ |
| 53 | size_t k, \ |
| 54 | const float* a, \ |
| 55 | size_t a_stride, \ |
| 56 | const float* w, \ |
| 57 | float* c, \ |
| 58 | size_t cm_stride, \ |
| 59 | size_t cn_stride, \ |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 60 | const union xnn_f32_minmax_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 61 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 62 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x2__neon_lane_ld64) |
| 63 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64) |
| 64 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64) |
| 65 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld128) |
| 66 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64) |
| 67 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64) |
| 68 | 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] | 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_4x2__neonfma_lane_ld64) |
| 71 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__neonfma_lane_ld64) |
| 72 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64) |
| 73 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld128) |
| 74 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64) |
| 75 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64) |
| 76 | 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] | 77 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 78 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__neon_dup_ld64) |
| 79 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neon_dup_ld64) |
| 80 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neon_dup_ld128) |
| 81 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64) |
| 82 | 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] | 83 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 84 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__neonfma_dup_ld64) |
| 85 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64) |
| 86 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld128) |
| 87 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64) |
| 88 | 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] | 89 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 90 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8s4__neon) |
| 91 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8s4__neon) |
| 92 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8s4__neon) |
| 93 | 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] | 94 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 95 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8s4__neonfma) |
| 96 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8s4__neonfma) |
| 97 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8s4__neonfma) |
| 98 | 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] | 99 | |
Marat Dukhan | 32f9381 | 2020-05-17 20:31:21 -0700 | [diff] [blame] | 100 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x4__aarch32_vfp_ld64) |
| 101 | |
Marat Dukhan | 3b98f6b | 2020-05-17 10:09:22 -0700 | [diff] [blame] | 102 | 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] | 103 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_ld64) |
Frank Barchard | 490febe | 2020-07-16 18:42:17 -0700 | [diff] [blame] | 104 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a7) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 105 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53) |
| 106 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55) |
| 107 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a75) |
Frank Barchard | 7873586 | 2022-01-04 16:47:44 -0800 | [diff] [blame] | 108 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_prfm_cortex_a75) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 109 | |
Frank Barchard | 3cb54f9 | 2020-04-10 10:46:08 -0700 | [diff] [blame] | 110 | 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] | 111 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x12__aarch64_neonfma_cortex_a53) |
| 112 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 113 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 114 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 115 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x12__aarch64_neonfma_cortex_a53) |
| 116 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53) |
| 117 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 118 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 119 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_prfm_cortex_a75) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 120 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_ld128) |
| 121 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_ld64) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 122 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 123 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__aarch64_neonfma_prfm_cortex_a75) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 124 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53) |
| 125 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55) |
| 126 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 127 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 128 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a75) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 129 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_ld128) |
| 130 | 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] | 131 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 132 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__sse_load1) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 133 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x8__sse_load1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 134 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__sse_load1) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 135 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__sse_load1) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 136 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 137 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__sse_dup) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 138 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x8__sse_dup) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 139 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__sse_dup) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 140 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__sse_dup) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 141 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 142 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8s4__sse) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 143 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x8s4__sse) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 144 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8s4__sse) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 145 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8s4__sse) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 146 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 147 | 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] | 148 | |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 149 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__sse2_dup) |
| 150 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x8__sse2_dup) |
| 151 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__sse2_dup) |
| 152 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__sse2_dup) |
| 153 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 154 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__avx_broadcast) |
| 155 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__avx_broadcast) |
| 156 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__avx_broadcast) |
| 157 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__avx_broadcast) |
| 158 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_7x8__avx_broadcast) |
| 159 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x16__avx_broadcast) |
| 160 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x16__avx_broadcast) |
| 161 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x16__avx_broadcast) |
| 162 | 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] | 163 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 164 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__fma3_broadcast) |
| 165 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__fma3_broadcast) |
| 166 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__fma3_broadcast) |
| 167 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__fma3_broadcast) |
| 168 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_7x8__fma3_broadcast) |
| 169 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_8x8__fma3_broadcast) |
| 170 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x16__fma3_broadcast) |
| 171 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x16__fma3_broadcast) |
| 172 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x16__fma3_broadcast) |
| 173 | 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] | 174 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 175 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast) |
| 176 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x16s4__fma3_broadcast) |
| 177 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast) |
| 178 | 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] | 179 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 180 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x16__avx512f_broadcast) |
| 181 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast) |
| 182 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast) |
| 183 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast) |
| 184 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast) |
| 185 | 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] | 186 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 187 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_arm_loadsplat) |
| 188 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_arm_loadsplat) |
| 189 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_arm_loadsplat) |
| 190 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_arm_loadsplat) |
| 191 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_loadsplat) |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 192 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 193 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_x86_loadsplat) |
| 194 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_x86_loadsplat) |
| 195 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_x86_loadsplat) |
| 196 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_loadsplat) |
| 197 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_loadsplat) |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 198 | |
Marat Dukhan | 802808c | 2020-06-16 11:01:17 -0700 | [diff] [blame] | 199 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x8__wasmsimd_splat) |
| 200 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x8__wasmsimd_splat) |
| 201 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_5x8__wasmsimd_splat) |
| 202 | |
Marat Dukhan | 688f6d8 | 2020-07-14 17:02:11 -0700 | [diff] [blame] | 203 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_1x8__wasmsimd_splat) |
| 204 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_4x8__wasmsimd_splat) |
| 205 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_5x8__wasmsimd_splat) |
Marat Dukhan | 802808c | 2020-06-16 11:01:17 -0700 | [diff] [blame] | 206 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 207 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_arm_splat) |
| 208 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_arm_splat) |
| 209 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_arm_splat) |
| 210 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_arm_splat) |
| 211 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_arm_splat) |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 212 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 213 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_x86_splat) |
| 214 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x8__wasmsimd_x86_splat) |
| 215 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_x86_splat) |
| 216 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_x86_splat) |
| 217 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8__wasmsimd_x86_splat) |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 218 | |
| 219 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_arm) |
| 220 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x8s4__wasmsimd_arm) |
| 221 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8s4__wasmsimd_arm) |
| 222 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_arm) |
| 223 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8s4__wasmsimd_arm) |
| 224 | |
| 225 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x8s4__wasmsimd_x86) |
| 226 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_3x8s4__wasmsimd_x86) |
| 227 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x8s4__wasmsimd_x86) |
| 228 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_5x8s4__wasmsimd_x86) |
| 229 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_6x8s4__wasmsimd_x86) |
| 230 | |
Marat Dukhan | e39e646 | 2020-07-09 01:33:36 -0700 | [diff] [blame] | 231 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x2c4__wasmsimd) |
| 232 | |
Marat Dukhan | 688f6d8 | 2020-07-14 17:02:11 -0700 | [diff] [blame] | 233 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_4x2c4__wasmsimd) |
Marat Dukhan | e39e646 | 2020-07-09 01:33:36 -0700 | [diff] [blame] | 234 | |
| 235 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x2c4__wasmsimd_arm) |
| 236 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x2c4__wasmsimd_x86) |
| 237 | |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 238 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x2__wasm) |
| 239 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x4__wasm) |
| 240 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_2x4__wasm) |
| 241 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x4__wasm) |
| 242 | |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 243 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_4x2__wasm) |
| 244 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_1x4__wasm) |
| 245 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_2x4__wasm) |
| 246 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_4x4__wasm) |
| 247 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 248 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x2__wasm) |
| 249 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x4__wasm) |
| 250 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_2x4__wasm) |
| 251 | 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] | 252 | |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 253 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x2__scalar) |
| 254 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_1x4__scalar) |
| 255 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_2x4__scalar) |
| 256 | DECLARE_F32_GEMM_UKERNEL_FUNCTION(xnn_f32_gemm_ukernel_4x4__scalar) |
| 257 | |
Marat Dukhan | 467f636 | 2020-05-22 23:21:55 -0700 | [diff] [blame] | 258 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_4x2__scalar) |
| 259 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_1x4__scalar) |
| 260 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_2x4__scalar) |
| 261 | DECLARE_F32_GEMM_RELU_UKERNEL_FUNCTION(xnn_f32_gemm_relu_ukernel_4x4__scalar) |
| 262 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 263 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_4x2__scalar) |
| 264 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_1x4__scalar) |
| 265 | DECLARE_F32_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemm_minmax_ukernel_2x4__scalar) |
| 266 | 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] | 267 | |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 268 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 269 | #define DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(fn_name) \ |
| 270 | XNN_INTERNAL void fn_name( \ |
| 271 | size_t mr, \ |
| 272 | size_t nr, \ |
| 273 | size_t k, \ |
| 274 | const float* a, \ |
| 275 | size_t a_stride, \ |
| 276 | const float* w, \ |
| 277 | float* c, \ |
| 278 | size_t cm_stride, \ |
| 279 | size_t cn_stride, \ |
| 280 | const float* acc, \ |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 281 | const union xnn_f32_minmax_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 282 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 283 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__neon_lane_ld64) |
| 284 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neon_lane_ld64) |
| 285 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neon_lane_ld128) |
| 286 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64) |
| 287 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64) |
| 288 | 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] | 289 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 290 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__neonfma_lane_ld64) |
| 291 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64) |
| 292 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld128) |
| 293 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64) |
| 294 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64) |
| 295 | 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] | 296 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 297 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__neon_dup_ld64) |
| 298 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64) |
| 299 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld128) |
| 300 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64) |
| 301 | 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] | 302 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 303 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__neonfma_dup_ld64) |
| 304 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_dup_ld64) |
| 305 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_dup_ld128) |
| 306 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64) |
| 307 | 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] | 308 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 309 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8s4__neon) |
| 310 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8s4__neon) |
| 311 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8s4__neon) |
| 312 | 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] | 313 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 314 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8s4__neonfma) |
| 315 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8s4__neonfma) |
| 316 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8s4__neonfma) |
| 317 | 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] | 318 | |
Frank Barchard | 3cb54f9 | 2020-04-10 10:46:08 -0700 | [diff] [blame] | 319 | 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] | 320 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x12__aarch64_neonfma_cortex_a53) |
| 321 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 322 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 323 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 324 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x12__aarch64_neonfma_cortex_a53) |
| 325 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53) |
| 326 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 327 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 328 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_prfm_cortex_a75) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 329 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_ld128) |
| 330 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__aarch64_neonfma_ld64) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 331 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 332 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__aarch64_neonfma_prfm_cortex_a75) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 333 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53) |
| 334 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55) |
| 335 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 336 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_cortex_a75) |
Frank Barchard | 143a110 | 2021-06-15 09:15:34 -0700 | [diff] [blame] | 337 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a75) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 338 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__aarch64_neonfma_ld128) |
| 339 | 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] | 340 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 341 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__sse_load1) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 342 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x8__sse_load1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 343 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__sse_load1) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 344 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__sse_load1) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 345 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 346 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__sse_dup) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 347 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x8__sse_dup) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 348 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__sse_dup) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 349 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__sse_dup) |
| 350 | |
| 351 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__sse2_dup) |
| 352 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x8__sse2_dup) |
| 353 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__sse2_dup) |
| 354 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__sse2_dup) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 355 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 356 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8s4__sse) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 357 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x8s4__sse) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 358 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8s4__sse) |
Marat Dukhan | 802fcae | 2020-12-11 14:37:25 -0800 | [diff] [blame] | 359 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8s4__sse) |
Marat Dukhan | eccfd71 | 2019-12-08 16:49:27 -0800 | [diff] [blame] | 360 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 361 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__avx_broadcast) |
| 362 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__avx_broadcast) |
| 363 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__avx_broadcast) |
| 364 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__avx_broadcast) |
| 365 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_7x8__avx_broadcast) |
| 366 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x16__avx_broadcast) |
| 367 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x16__avx_broadcast) |
| 368 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x16__avx_broadcast) |
| 369 | 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] | 370 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 371 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__fma3_broadcast) |
| 372 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__fma3_broadcast) |
| 373 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__fma3_broadcast) |
| 374 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__fma3_broadcast) |
| 375 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_7x8__fma3_broadcast) |
| 376 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_8x8__fma3_broadcast) |
| 377 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x16__fma3_broadcast) |
| 378 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x16__fma3_broadcast) |
| 379 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x16__fma3_broadcast) |
| 380 | 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] | 381 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 382 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x16s4__fma3_broadcast) |
| 383 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x16s4__fma3_broadcast) |
| 384 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x16s4__fma3_broadcast) |
| 385 | 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] | 386 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 387 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x16__avx512f_broadcast) |
| 388 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast) |
| 389 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast) |
| 390 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast) |
| 391 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast) |
| 392 | 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] | 393 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 394 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__wasmsimd_arm_loadsplat) |
| 395 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x8__wasmsimd_arm_loadsplat) |
| 396 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__wasmsimd_arm_loadsplat) |
| 397 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_arm_loadsplat) |
| 398 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_arm_loadsplat) |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 399 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 400 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__wasmsimd_arm_splat) |
| 401 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x8__wasmsimd_arm_splat) |
| 402 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__wasmsimd_arm_splat) |
| 403 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_arm_splat) |
| 404 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_arm_splat) |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 405 | |
| 406 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8s4__wasmsimd_arm) |
| 407 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x8s4__wasmsimd_arm) |
| 408 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8s4__wasmsimd_arm) |
| 409 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_arm) |
| 410 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8s4__wasmsimd_arm) |
| 411 | |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 412 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 413 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__wasmsimd_x86_loadsplat) |
| 414 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x8__wasmsimd_x86_loadsplat) |
| 415 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__wasmsimd_x86_loadsplat) |
| 416 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_loadsplat) |
| 417 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_loadsplat) |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 418 | |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 419 | |
Frank Barchard | 0725b8d | 2020-12-07 11:07:35 -0800 | [diff] [blame] | 420 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8__wasmsimd_x86_splat) |
| 421 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x8__wasmsimd_x86_splat) |
| 422 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8__wasmsimd_x86_splat) |
| 423 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8__wasmsimd_x86_splat) |
| 424 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8__wasmsimd_x86_splat) |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 425 | |
Marat Dukhan | 1bbf96b | 2020-06-15 23:01:20 -0700 | [diff] [blame] | 426 | |
| 427 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x8s4__wasmsimd_x86) |
| 428 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_3x8s4__wasmsimd_x86) |
| 429 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x8s4__wasmsimd_x86) |
| 430 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_5x8s4__wasmsimd_x86) |
| 431 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_6x8s4__wasmsimd_x86) |
| 432 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 433 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x4__wasm) |
| 434 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_2x4__wasm) |
| 435 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_4x2__wasm) |
| 436 | 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] | 437 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 438 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_1x4__scalar) |
| 439 | DECLARE_F32_GEMMINC_MINMAX_UKERNEL_FUNCTION(xnn_f32_gemminc_minmax_ukernel_2x4__scalar) |
| 440 | 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] | 441 | |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 442 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 443 | #define DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(fn_name) \ |
| 444 | void fn_name( \ |
| 445 | size_t mr, \ |
| 446 | size_t nr, \ |
| 447 | size_t k, \ |
| 448 | const void* a, \ |
| 449 | size_t a_stride, \ |
| 450 | const void* w, \ |
| 451 | void* c, \ |
| 452 | size_t cm_stride, \ |
| 453 | size_t cn_stride, \ |
Marat Dukhan | 14dd8d0 | 2022-01-06 16:03:31 -0800 | [diff] [blame] | 454 | const union xnn_f16_scaleminmax_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 455 | |
Frank Barchard | 36b76b6 | 2020-04-10 12:39:17 -0700 | [diff] [blame] | 456 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld32) |
| 457 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32) |
| 458 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32) |
Frank Barchard | 80fc5f4 | 2021-06-07 10:43:16 -0700 | [diff] [blame] | 459 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a55) |
Frank Barchard | 9737461 | 2021-06-07 11:51:07 -0700 | [diff] [blame] | 460 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a75) |
Frank Barchard | bddfbcd | 2020-04-15 12:32:41 -0700 | [diff] [blame] | 461 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_1x8__aarch64_neonfp16arith_ld64) |
| 462 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_4x8__aarch64_neonfp16arith_ld64) |
| 463 | 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] | 464 | 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] | 465 | 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] | 466 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64) |
| 467 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64) |
| 468 | 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] | 469 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64) |
| 470 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64) |
| 471 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64) |
| 472 | 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] | 473 | |
Marat Dukhan | c4302c2 | 2022-01-06 19:27:03 -0800 | [diff] [blame] | 474 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_1x8__avx2_broadcast) |
| 475 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_4x8__avx2_broadcast) |
| 476 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_5x8__avx2_broadcast) |
| 477 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_6x8__avx2_broadcast) |
| 478 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_7x8__avx2_broadcast) |
| 479 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_1x16__avx2_broadcast) |
| 480 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_3x16__avx2_broadcast) |
| 481 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_4x16__avx2_broadcast) |
| 482 | DECLARE_F16_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_f16_gemm_minmax_ukernel_5x16__avx2_broadcast) |
| 483 | |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 484 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 485 | #define DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(fn_name) \ |
| 486 | XNN_INTERNAL void fn_name( \ |
| 487 | size_t mr, \ |
Benoit Jacob | 1065af4 | 2020-07-30 12:17:15 -0700 | [diff] [blame] | 488 | size_t nc, \ |
| 489 | size_t kc, \ |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 490 | const uint8_t* a, \ |
| 491 | size_t a_stride, \ |
| 492 | const void* w, \ |
| 493 | uint8_t* c, \ |
| 494 | size_t cm_stride, \ |
| 495 | size_t cn_stride, \ |
Marat Dukhan | e3d17bf | 2021-05-24 22:22:43 -0700 | [diff] [blame] | 496 | const union xnn_qu8_conv_minmax_params* params); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 497 | |
Marat Dukhan | 173661d | 2021-07-26 23:47:08 -0700 | [diff] [blame] | 498 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 499 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane) |
| 500 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane) |
Marat Dukhan | 173661d | 2021-07-26 23:47:08 -0700 | [diff] [blame] | 501 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__neon_mlal_lane) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 502 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane) |
Marat Dukhan | 173661d | 2021-07-26 23:47:08 -0700 | [diff] [blame] | 503 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 504 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane) |
| 505 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane) |
Marat Dukhan | 173661d | 2021-07-26 23:47:08 -0700 | [diff] [blame] | 506 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 507 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 508 | |
Digant Desai | 59d6515 | 2021-11-29 10:44:04 -0800 | [diff] [blame] | 509 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane) |
| 510 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x8__neon_mlal_lane) |
Marat Dukhan | 69c8a29 | 2021-07-14 19:34:56 -0700 | [diff] [blame] | 511 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane) |
| 512 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane) |
| 513 | |
| 514 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane) |
| 515 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x16__neonv8_mlal_lane) |
| 516 | |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 517 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53) |
| 518 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53) |
| 519 | |
Frank Barchard | 9cdc10d | 2021-11-22 19:03:54 -0800 | [diff] [blame] | 520 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_ld64) |
| 521 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64) |
| 522 | |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 523 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75) |
| 524 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75) |
| 525 | |
Frank Barchard | ca4c68e | 2021-08-25 19:06:40 -0700 | [diff] [blame] | 526 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_cortex_a55) |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 527 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128) |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 528 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128) |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 529 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55) |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 530 | |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 531 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 532 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot) |
| 533 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 534 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 535 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 536 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot) |
| 537 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot) |
| 538 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 539 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot) |
| 540 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 541 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 542 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_5x16c4__neondot) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 543 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot) |
| 544 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot) |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 545 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot) |
| 546 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot) |
| 547 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 548 | |
Digant Desai | 10f9f62 | 2021-11-23 13:33:52 -0800 | [diff] [blame] | 549 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_cortex_a55) |
Digant Desai | 2e2d179 | 2021-11-24 11:06:37 -0800 | [diff] [blame] | 550 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128) |
| 551 | |
Digant Desai | 9982ed3 | 2021-11-24 13:03:54 -0800 | [diff] [blame] | 552 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot) |
| 553 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot) |
| 554 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot) |
Digant Desai | 10f9f62 | 2021-11-23 13:33:52 -0800 | [diff] [blame] | 555 | |
Frank Barchard | 901845c | 2022-01-19 01:45:22 -0800 | [diff] [blame] | 556 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64) |
| 557 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64) |
Frank Barchard | 9e4d2aa | 2022-02-02 00:31:21 -0800 | [diff] [blame] | 558 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53) |
| 559 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53) |
Frank Barchard | 34251d8 | 2022-02-02 11:57:11 -0800 | [diff] [blame] | 560 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7) |
| 561 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7) |
Frank Barchard | 901845c | 2022-01-19 01:45:22 -0800 | [diff] [blame] | 562 | |
Marat Dukhan | ef47f8d | 2021-07-02 15:08:32 -0700 | [diff] [blame] | 563 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64) |
| 564 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld64) |
| 565 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld64) |
| 566 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld64) |
| 567 | |
| 568 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64) |
| 569 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld64) |
| 570 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld64) |
| 571 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld64) |
| 572 | |
| 573 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld64) |
| 574 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld64) |
| 575 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld64) |
| 576 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld64) |
| 577 | |
| 578 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld64) |
| 579 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64) |
| 580 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64) |
| 581 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld64) |
| 582 | |
| 583 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128) |
| 584 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld128) |
| 585 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128) |
| 586 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld128) |
| 587 | |
| 588 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128) |
| 589 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld128) |
| 590 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld128) |
| 591 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128) |
| 592 | |
| 593 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld128) |
| 594 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld128) |
| 595 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld128) |
| 596 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld128) |
| 597 | |
| 598 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128) |
| 599 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld128) |
| 600 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld128) |
| 601 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld128) |
| 602 | |
Marat Dukhan | ef47f8d | 2021-07-02 15:08:32 -0700 | [diff] [blame] | 603 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64) |
| 604 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld64) |
| 605 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64) |
| 606 | |
| 607 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64) |
| 608 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64) |
| 609 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64) |
| 610 | |
| 611 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld64) |
| 612 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld64) |
| 613 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64) |
| 614 | |
| 615 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64) |
| 616 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64) |
| 617 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64) |
| 618 | |
| 619 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld128) |
| 620 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld128) |
| 621 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128) |
| 622 | |
| 623 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128) |
| 624 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld128) |
| 625 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128) |
| 626 | |
| 627 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128) |
| 628 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128) |
| 629 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128) |
| 630 | |
| 631 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128) |
| 632 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128) |
| 633 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128) |
| 634 | |
Marat Dukhan | 902ef7f | 2021-07-02 16:11:06 -0700 | [diff] [blame] | 635 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x8c8__avx2) |
| 636 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x8c8__avx2) |
| 637 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x8c8__avx2) |
| 638 | |
Marat Dukhan | 3cf2e22 | 2021-07-08 11:38:45 -0700 | [diff] [blame] | 639 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx) |
| 640 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx) |
| 641 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx) |
| 642 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx) |
| 643 | |
Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame] | 644 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64) |
| 645 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64) |
| 646 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64) |
| 647 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64) |
| 648 | |
| 649 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128) |
| 650 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128) |
| 651 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld128) |
| 652 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld128) |
| 653 | |
Marat Dukhan | 348c377 | 2022-02-01 00:36:50 -0800 | [diff] [blame] | 654 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64) |
| 655 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld64) |
| 656 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld64) |
| 657 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld64) |
| 658 | |
| 659 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128) |
| 660 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld128) |
| 661 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld128) |
| 662 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128) |
| 663 | |
Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame] | 664 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64) |
| 665 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld64) |
| 666 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64) |
| 667 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld64) |
| 668 | |
| 669 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128) |
| 670 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128) |
| 671 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld128) |
| 672 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld128) |
| 673 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 674 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld64) |
| 675 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul32_ld64) |
| 676 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld64) |
Marat Dukhan | 43bee05 | 2021-07-14 20:57:18 -0700 | [diff] [blame] | 677 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 678 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul32_ld128) |
| 679 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul32_ld128) |
| 680 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul32_ld128) |
Marat Dukhan | 43bee05 | 2021-07-14 20:57:18 -0700 | [diff] [blame] | 681 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 682 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__wasm_fmagic) |
| 683 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__wasm_fmagic) |
| 684 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__wasm_fmagic) |
| 685 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__wasm_fmagic) |
| 686 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__wasm_fmagic) |
| 687 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__wasm_fmagic) |
| 688 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__wasm_fmagic) |
| 689 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic) |
| 690 | |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 691 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_fmagic) |
| 692 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_fmagic) |
| 693 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_fmagic) |
| 694 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_fmagic) |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 695 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_fmagic) |
| 696 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_fmagic) |
| 697 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_fmagic) |
| 698 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic) |
Marat Dukhan | 927d474 | 2021-07-15 13:42:49 -0700 | [diff] [blame] | 699 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 700 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic) |
| 701 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic) |
| 702 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic) |
| 703 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic) |
Marat Dukhan | 272d4d9 | 2022-01-04 15:07:14 -0800 | [diff] [blame] | 704 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic) |
| 705 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic) |
| 706 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic) |
| 707 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic) |
| 708 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 709 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf) |
| 710 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf) |
| 711 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf) |
| 712 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf) |
Marat Dukhan | 272d4d9 | 2022-01-04 15:07:14 -0800 | [diff] [blame] | 713 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf) |
| 714 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf) |
| 715 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf) |
| 716 | DECLARE_QU8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qu8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf) |
| 717 | |
Marat Dukhan | ef47f8d | 2021-07-02 15:08:32 -0700 | [diff] [blame] | 718 | |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 719 | #define DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(fn_name) \ |
| 720 | XNN_INTERNAL void fn_name( \ |
| 721 | size_t mr, \ |
| 722 | size_t nr, \ |
| 723 | size_t k, \ |
| 724 | const int8_t* a, \ |
| 725 | size_t a_stride, \ |
| 726 | const void* w, \ |
| 727 | int8_t* c, \ |
| 728 | size_t cm_stride, \ |
| 729 | size_t cn_stride, \ |
Marat Dukhan | e3d17bf | 2021-05-24 22:22:43 -0700 | [diff] [blame] | 730 | const union xnn_qs8_conv_minmax_params* params); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 731 | |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 732 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane) |
| 733 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane) |
| 734 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane) |
| 735 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mlal_lane) |
| 736 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane) |
| 737 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane) |
| 738 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane) |
| 739 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane) |
| 740 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane) |
| 741 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane) |
| 742 | |
Marat Dukhan | 2d3c97c | 2021-06-25 18:00:28 -0700 | [diff] [blame] | 743 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane) |
| 744 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane) |
| 745 | |
| 746 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane) |
| 747 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x16__neonv8_mlal_lane) |
| 748 | |
Frank Barchard | 27bf92c | 2021-11-24 15:47:52 -0800 | [diff] [blame] | 749 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane_prfm) |
| 750 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane_prfm) |
| 751 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane_prfm) |
| 752 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mlal_lane_prfm) |
| 753 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm) |
| 754 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane_prfm) |
| 755 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm) |
| 756 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane_prfm) |
| 757 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane_prfm) |
| 758 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane_prfm) |
| 759 | |
Frank Barchard | 510b8e0 | 2021-07-26 17:25:18 -0700 | [diff] [blame] | 760 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__neon_mull_addw_dup) |
| 761 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup) |
| 762 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8__neon_mull_addw_dup) |
| 763 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup) |
| 764 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mull_addw_dup) |
| 765 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup) |
| 766 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16__neon_mull_addw_dup) |
| 767 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mull_addw_dup) |
| 768 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 769 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mull_dup) |
| 770 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_dup) |
| 771 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup) |
| 772 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup) |
| 773 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup) |
| 774 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup) |
| 775 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup) |
| 776 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup) |
Frank Barchard | 287952a | 2021-11-03 15:26:45 -0700 | [diff] [blame] | 777 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 778 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_dup) |
| 779 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup) |
Frank Barchard | 287952a | 2021-11-03 15:26:45 -0700 | [diff] [blame] | 780 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 781 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_dup) |
| 782 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup) |
Frank Barchard | 287952a | 2021-11-03 15:26:45 -0700 | [diff] [blame] | 783 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 784 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup) |
| 785 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup) |
| 786 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup) |
| 787 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup) |
| 788 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup) |
| 789 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup) |
| 790 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup) |
| 791 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup) |
Frank Barchard | 287952a | 2021-11-03 15:26:45 -0700 | [diff] [blame] | 792 | |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 793 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r) |
| 794 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r) |
| 795 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r) |
| 796 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r) |
| 797 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r) |
| 798 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r) |
| 799 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r) |
| 800 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r) |
| 801 | |
| 802 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_ld1r) |
| 803 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r) |
| 804 | |
| 805 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld1r) |
| 806 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r) |
| 807 | |
| 808 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r) |
| 809 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r) |
| 810 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r) |
| 811 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r) |
| 812 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r) |
| 813 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r) |
| 814 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r) |
| 815 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r) |
| 816 | |
| 817 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld2r) |
| 818 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r) |
| 819 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r) |
| 820 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r) |
| 821 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r) |
| 822 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r) |
| 823 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r) |
| 824 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r) |
| 825 | |
| 826 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_ld2r) |
| 827 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r) |
| 828 | |
| 829 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld2r) |
| 830 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r) |
| 831 | |
| 832 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld2r) |
| 833 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r) |
| 834 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r) |
| 835 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r) |
| 836 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r) |
| 837 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r) |
| 838 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r) |
| 839 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r) |
| 840 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 841 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull) |
| 842 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4s2__neon_mull) |
| 843 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull) |
| 844 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull) |
| 845 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mull) |
| 846 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull) |
| 847 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mull) |
| 848 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull) |
Frank Barchard | eb704f7 | 2021-11-12 01:26:50 -0800 | [diff] [blame] | 849 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 850 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4s2__neon_mlal) |
| 851 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4s2__neon_mlal) |
Frank Barchard | eb704f7 | 2021-11-12 01:26:50 -0800 | [diff] [blame] | 852 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 853 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4s2__neonv8_mlal) |
| 854 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4s2__neonv8_mlal) |
Frank Barchard | eb704f7 | 2021-11-12 01:26:50 -0800 | [diff] [blame] | 855 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 856 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mlal) |
| 857 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal) |
| 858 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal) |
| 859 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal) |
| 860 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal) |
| 861 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal) |
| 862 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal) |
| 863 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal) |
Frank Barchard | eb704f7 | 2021-11-12 01:26:50 -0800 | [diff] [blame] | 864 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 865 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_dup) |
| 866 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_dup) |
| 867 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup) |
| 868 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup) |
| 869 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup) |
| 870 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup) |
| 871 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup) |
| 872 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup) |
Frank Barchard | 2302ffd | 2021-01-22 14:08:02 -0800 | [diff] [blame] | 873 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 874 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup) |
| 875 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup) |
| 876 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup) |
| 877 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup) |
| 878 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup) |
| 879 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup) |
| 880 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup) |
| 881 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup) |
Frank Barchard | 1d41247 | 2021-10-25 17:27:21 -0700 | [diff] [blame] | 882 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 883 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup) |
| 884 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_dup) |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 885 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 886 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup) |
| 887 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup) |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 888 | |
Frank Barchard | 15eec02 | 2021-11-17 13:26:20 -0800 | [diff] [blame] | 889 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r) |
| 890 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld1r) |
| 891 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r) |
| 892 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r) |
| 893 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r) |
| 894 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r) |
| 895 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r) |
| 896 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r) |
| 897 | |
| 898 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r) |
| 899 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld1r) |
| 900 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r) |
| 901 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r) |
| 902 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r) |
| 903 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r) |
| 904 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r) |
| 905 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r) |
| 906 | |
| 907 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r) |
| 908 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r) |
| 909 | |
| 910 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r) |
| 911 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld1r) |
| 912 | |
| 913 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld2r) |
| 914 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r) |
| 915 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r) |
| 916 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r) |
| 917 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r) |
| 918 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r) |
| 919 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r) |
| 920 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r) |
| 921 | |
| 922 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r) |
| 923 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r) |
| 924 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r) |
| 925 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r) |
| 926 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r) |
| 927 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r) |
| 928 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r) |
| 929 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r) |
| 930 | |
| 931 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r) |
| 932 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld2r) |
| 933 | |
| 934 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r) |
| 935 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld2r) |
| 936 | |
Frank Barchard | 42f5c50 | 2021-11-16 10:04:21 -0800 | [diff] [blame] | 937 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld4r) |
| 938 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld4r) |
| 939 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r) |
| 940 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r) |
| 941 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r) |
| 942 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r) |
| 943 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld4r) |
| 944 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r) |
| 945 | |
| 946 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld4r) |
| 947 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld4r) |
| 948 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r) |
| 949 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r) |
| 950 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r) |
| 951 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r) |
| 952 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r) |
| 953 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r) |
| 954 | |
| 955 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r) |
| 956 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r) |
| 957 | |
| 958 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r) |
| 959 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r) |
| 960 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 961 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull) |
| 962 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull) |
| 963 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2s4__neon_mull) |
| 964 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull) |
| 965 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull) |
| 966 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull) |
| 967 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull) |
| 968 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mull) |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 969 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 970 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal) |
| 971 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2s4__neon_mlal) |
| 972 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal) |
| 973 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal) |
| 974 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal) |
| 975 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal) |
| 976 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal) |
| 977 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal) |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 978 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 979 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2s4__neon_mlal) |
| 980 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2s4__neon_mlal) |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 981 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 982 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal) |
| 983 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal) |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 984 | |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 985 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mull) |
| 986 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull) |
| 987 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull) |
| 988 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull) |
| 989 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull) |
| 990 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mull) |
| 991 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mull) |
| 992 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mull) |
Frank Barchard | 5899012 | 2021-01-29 16:09:56 -0800 | [diff] [blame] | 993 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 994 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal) |
| 995 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal) |
Marat Dukhan | 2d3c97c | 2021-06-25 18:00:28 -0700 | [diff] [blame] | 996 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 997 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal) |
| 998 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal) |
Marat Dukhan | 2d3c97c | 2021-06-25 18:00:28 -0700 | [diff] [blame] | 999 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1000 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal) |
| 1001 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal) |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1002 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mlal) |
| 1003 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mlal) |
Marat Dukhan | e903dff | 2021-07-16 19:43:41 -0700 | [diff] [blame] | 1004 | |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1005 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal) |
| 1006 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c8__neon_mlal) |
| 1007 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c8__neon_mlal) |
| 1008 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal) |
Frank Barchard | da78da1 | 2021-03-02 14:28:00 -0800 | [diff] [blame] | 1009 | |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1010 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal) |
| 1011 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal) |
| 1012 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c16__neon_mlal) |
| 1013 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c16__neon_mlal) |
Frank Barchard | 71c4d1a | 2021-02-19 10:07:23 -0800 | [diff] [blame] | 1014 | |
Marat Dukhan | 8999190 | 2021-12-06 00:54:36 -0800 | [diff] [blame] | 1015 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal) |
| 1016 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal) |
| 1017 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c16__neon_mlal) |
| 1018 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal) |
Frank Barchard | 71c4d1a | 2021-02-19 10:07:23 -0800 | [diff] [blame] | 1019 | |
Marat Dukhan | 18630de | 2021-06-02 22:20:01 -0700 | [diff] [blame] | 1020 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c4__neondot) |
Marat Dukhan | d4c8303 | 2020-09-09 23:20:12 -0700 | [diff] [blame] | 1021 | |
Marat Dukhan | e903dff | 2021-07-16 19:43:41 -0700 | [diff] [blame] | 1022 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neondot) |
| 1023 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neondot) |
Marat Dukhan | 4486f87 | 2021-08-07 15:22:50 -0700 | [diff] [blame] | 1024 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot) |
| 1025 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot) |
Marat Dukhan | e903dff | 2021-07-16 19:43:41 -0700 | [diff] [blame] | 1026 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neondot) |
| 1027 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot) |
Marat Dukhan | 4486f87 | 2021-08-07 15:22:50 -0700 | [diff] [blame] | 1028 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot) |
| 1029 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot) |
Marat Dukhan | 18630de | 2021-06-02 22:20:01 -0700 | [diff] [blame] | 1030 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1031 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal) |
| 1032 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_prfm) |
| 1033 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53) |
| 1034 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_prfm_cortex_a53) |
| 1035 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull) |
| 1036 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal) |
| 1037 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm) |
| 1038 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53) |
| 1039 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53) |
| 1040 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__aarch64_neon_mlal) |
Frank Barchard | 13db60f | 2021-07-20 14:34:35 -0700 | [diff] [blame] | 1041 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1042 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal) |
| 1043 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm) |
| 1044 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53) |
| 1045 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm_cortex_a53) |
| 1046 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mull) |
| 1047 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal) |
| 1048 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm) |
| 1049 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53) |
| 1050 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53) |
| 1051 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal) |
Frank Barchard | 1663c0c | 2021-07-01 11:20:06 -0700 | [diff] [blame] | 1052 | |
Frank Barchard | 13db60f | 2021-07-20 14:34:35 -0700 | [diff] [blame] | 1053 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53) |
| 1054 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53) |
| 1055 | |
Frank Barchard | 98af05c | 2021-06-30 12:15:04 -0700 | [diff] [blame] | 1056 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53) |
| 1057 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53) |
| 1058 | |
Frank Barchard | 914f57b | 2021-12-13 12:31:42 -0800 | [diff] [blame] | 1059 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64) |
| 1060 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_prfm_ld64) |
| 1061 | |
Frank Barchard | 5cffb64 | 2021-11-22 13:59:43 -0800 | [diff] [blame] | 1062 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_ld64) |
| 1063 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64) |
| 1064 | |
| 1065 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_ld64) |
| 1066 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64) |
| 1067 | |
Frank Barchard | 13db60f | 2021-07-20 14:34:35 -0700 | [diff] [blame] | 1068 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__aarch64_neondot_ld32) |
| 1069 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__aarch64_neondot_ld64) |
| 1070 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld32) |
| 1071 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64) |
Frank Barchard | 60729d0 | 2021-07-20 12:25:09 -0700 | [diff] [blame] | 1072 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128) |
Frank Barchard | 13db60f | 2021-07-20 14:34:35 -0700 | [diff] [blame] | 1073 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55) |
Frank Barchard | 60729d0 | 2021-07-20 12:25:09 -0700 | [diff] [blame] | 1074 | |
Frank Barchard | 1a0b276 | 2021-06-29 18:37:59 -0700 | [diff] [blame] | 1075 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x16c4__aarch64_neondot_ld32) |
| 1076 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x16c4__aarch64_neondot_ld64) |
| 1077 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld32) |
| 1078 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld64) |
Frank Barchard | 3ebfb13 | 2021-06-29 15:08:39 -0700 | [diff] [blame] | 1079 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128) |
Frank Barchard | 1a0b276 | 2021-06-29 18:37:59 -0700 | [diff] [blame] | 1080 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_cortex_a55) |
Frank Barchard | 3ebfb13 | 2021-06-29 15:08:39 -0700 | [diff] [blame] | 1081 | |
Frank Barchard | 9f3f420 | 2021-12-16 18:13:51 -0800 | [diff] [blame] | 1082 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__aarch32_neondot_ld64) |
Frank Barchard | 0f294ad | 2022-01-24 10:48:38 -0800 | [diff] [blame] | 1083 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__aarch32_neondot_cortex_a55) |
Frank Barchard | 9f3f420 | 2021-12-16 18:13:51 -0800 | [diff] [blame] | 1084 | |
Frank Barchard | da7b2e2 | 2021-12-13 23:50:53 -0800 | [diff] [blame] | 1085 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64) |
| 1086 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64) |
Frank Barchard | 9e4d2aa | 2022-02-02 00:31:21 -0800 | [diff] [blame] | 1087 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53) |
| 1088 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53) |
Frank Barchard | 34251d8 | 2022-02-02 11:57:11 -0800 | [diff] [blame] | 1089 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7) |
| 1090 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7) |
Frank Barchard | da7b2e2 | 2021-12-13 23:50:53 -0800 | [diff] [blame] | 1091 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1092 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64) |
| 1093 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld64) |
| 1094 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld64) |
| 1095 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld64) |
| 1096 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1097 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64) |
| 1098 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld64) |
| 1099 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld64) |
| 1100 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld64) |
| 1101 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1102 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld64) |
| 1103 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld64) |
| 1104 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld64) |
| 1105 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld64) |
| 1106 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1107 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld64) |
| 1108 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64) |
| 1109 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64) |
| 1110 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld64) |
| 1111 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1112 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128) |
| 1113 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld128) |
| 1114 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128) |
| 1115 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld128) |
| 1116 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1117 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128) |
| 1118 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld128) |
| 1119 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld128) |
| 1120 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128) |
| 1121 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1122 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld128) |
| 1123 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld128) |
| 1124 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld128) |
| 1125 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld128) |
| 1126 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1127 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128) |
| 1128 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld128) |
| 1129 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld128) |
| 1130 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld128) |
| 1131 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1132 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64) |
| 1133 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld64) |
| 1134 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64) |
| 1135 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1136 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__ssse3_ld64) |
| 1137 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__ssse3_ld64) |
| 1138 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__ssse3_ld64) |
| 1139 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1140 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64) |
| 1141 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64) |
| 1142 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64) |
| 1143 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1144 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld64) |
| 1145 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld64) |
| 1146 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64) |
| 1147 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1148 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64) |
| 1149 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64) |
| 1150 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64) |
| 1151 | |
Marat Dukhan | c46e671 | 2021-06-01 19:00:16 -0700 | [diff] [blame] | 1152 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld128) |
| 1153 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld128) |
| 1154 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128) |
| 1155 | |
Marat Dukhan | ef47f8d | 2021-07-02 15:08:32 -0700 | [diff] [blame] | 1156 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__ssse3_ld128) |
| 1157 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__ssse3_ld128) |
| 1158 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__ssse3_ld128) |
| 1159 | |
| 1160 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128) |
| 1161 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld128) |
| 1162 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128) |
| 1163 | |
| 1164 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128) |
| 1165 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128) |
| 1166 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128) |
| 1167 | |
| 1168 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128) |
| 1169 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128) |
| 1170 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128) |
| 1171 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1172 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c2__sse2) |
| 1173 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__sse2) |
| 1174 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__sse2) |
| 1175 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__sse2) |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1176 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1177 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__sse2) |
| 1178 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__sse2) |
| 1179 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__sse2) |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1180 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1181 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__ssse3) |
| 1182 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__ssse3) |
| 1183 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__ssse3) |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1184 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1185 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c2__sse41) |
| 1186 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__sse41) |
| 1187 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__sse41) |
| 1188 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__sse41) |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1189 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1190 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__sse41) |
| 1191 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__sse41) |
| 1192 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__sse41) |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1193 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1194 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c2__avx) |
| 1195 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__avx) |
| 1196 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__avx) |
| 1197 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__avx) |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1198 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1199 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__avx) |
| 1200 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__avx) |
| 1201 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__avx) |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1202 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1203 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c2__xop) |
| 1204 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__xop) |
| 1205 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__xop) |
| 1206 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__xop) |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1207 | |
Marat Dukhan | 0ff7989 | 2021-08-06 16:05:06 -0700 | [diff] [blame] | 1208 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__xop) |
| 1209 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__xop) |
| 1210 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__xop) |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1211 | |
Marat Dukhan | 9b474cf | 2021-05-25 16:37:48 -0700 | [diff] [blame] | 1212 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__avx2) |
| 1213 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__avx2) |
| 1214 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x8c8__avx2) |
| 1215 | |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1216 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x8c8__avx2) |
| 1217 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x8c8__avx2) |
| 1218 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x8c8__avx2) |
| 1219 | |
Marat Dukhan | 71855ee | 2021-05-25 19:05:06 -0700 | [diff] [blame] | 1220 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx) |
| 1221 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx) |
| 1222 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx) |
| 1223 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx) |
| 1224 | |
Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame] | 1225 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64) |
| 1226 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64) |
| 1227 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64) |
| 1228 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64) |
| 1229 | |
| 1230 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128) |
| 1231 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128) |
| 1232 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld128) |
| 1233 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld128) |
| 1234 | |
| 1235 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2) |
| 1236 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2) |
| 1237 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2) |
| 1238 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2) |
| 1239 | |
Marat Dukhan | 0f1ed94 | 2021-12-08 23:25:50 -0800 | [diff] [blame] | 1240 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64) |
| 1241 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld64) |
| 1242 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld64) |
| 1243 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld64) |
| 1244 | |
| 1245 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128) |
| 1246 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld128) |
| 1247 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld128) |
| 1248 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128) |
| 1249 | |
Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame] | 1250 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64) |
| 1251 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld64) |
| 1252 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64) |
| 1253 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld64) |
| 1254 | |
| 1255 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128) |
| 1256 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128) |
| 1257 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld128) |
| 1258 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld128) |
| 1259 | |
| 1260 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2) |
| 1261 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2) |
| 1262 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2) |
| 1263 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2) |
| 1264 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1265 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_ld64) |
| 1266 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld64) |
| 1267 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld64) |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1268 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1269 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_ld128) |
| 1270 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld128) |
| 1271 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld128) |
Marat Dukhan | 4741e41 | 2021-06-30 13:38:06 -0700 | [diff] [blame] | 1272 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1273 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16) |
| 1274 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16) |
| 1275 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_xw_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16) |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1276 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1277 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__wasm_fmagic) |
| 1278 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__wasm_fmagic) |
| 1279 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__wasm_fmagic) |
| 1280 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__wasm_fmagic) |
| 1281 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__wasm_fmagic) |
| 1282 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__wasm_fmagic) |
| 1283 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__wasm_fmagic) |
| 1284 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic) |
Frank Barchard | 1a2dbe1 | 2021-07-22 20:13:58 -0700 | [diff] [blame] | 1285 | |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1286 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_fmagic) |
| 1287 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_fmagic) |
| 1288 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_fmagic) |
| 1289 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_fmagic) |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1290 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_fmagic) |
| 1291 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_fmagic) |
| 1292 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_fmagic) |
| 1293 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic) |
Marat Dukhan | 779b253 | 2021-06-29 14:14:13 -0700 | [diff] [blame] | 1294 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1295 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic) |
| 1296 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic) |
| 1297 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic) |
| 1298 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic) |
Marat Dukhan | 272d4d9 | 2022-01-04 15:07:14 -0800 | [diff] [blame] | 1299 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic) |
| 1300 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic) |
| 1301 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic) |
| 1302 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic) |
| 1303 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1304 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf) |
| 1305 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf) |
| 1306 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf) |
| 1307 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf) |
Marat Dukhan | 272d4d9 | 2022-01-04 15:07:14 -0800 | [diff] [blame] | 1308 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf) |
| 1309 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf) |
| 1310 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf) |
| 1311 | DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf) |
| 1312 | |
Benoit Jacob | a964473 | 2020-08-13 12:48:55 -0700 | [diff] [blame] | 1313 | |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1314 | #define DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(fn_name) \ |
| 1315 | XNN_INTERNAL void fn_name( \ |
| 1316 | size_t mr, \ |
| 1317 | size_t nr, \ |
| 1318 | size_t k, \ |
| 1319 | const int8_t* a, \ |
| 1320 | size_t a_stride, \ |
| 1321 | const void* w, \ |
| 1322 | int8_t* c, \ |
| 1323 | size_t cm_stride, \ |
| 1324 | size_t cn_stride, \ |
| 1325 | const union xnn_qs8_minmax_params* params); |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 1326 | |
Frank Barchard | f623740 | 2022-01-05 00:26:09 -0800 | [diff] [blame] | 1327 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane) |
| 1328 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8__neon_mlal_lane) |
| 1329 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x8__neon_mlal_lane) |
| 1330 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__neon_mlal_lane) |
| 1331 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane) |
Marat Dukhan | e76478b | 2021-06-28 16:35:40 -0700 | [diff] [blame] | 1332 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane) |
Frank Barchard | f623740 | 2022-01-05 00:26:09 -0800 | [diff] [blame] | 1333 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neon_mlal_lane) |
| 1334 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neon_mlal_lane) |
Marat Dukhan | e76478b | 2021-06-28 16:35:40 -0700 | [diff] [blame] | 1335 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane) |
Frank Barchard | f623740 | 2022-01-05 00:26:09 -0800 | [diff] [blame] | 1336 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane) |
| 1337 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neon_mlal_lane_prfm) |
| 1338 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8__neon_mlal_lane_prfm) |
| 1339 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x8__neon_mlal_lane_prfm) |
| 1340 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__neon_mlal_lane_prfm) |
| 1341 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm) |
| 1342 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neon_mlal_lane_prfm) |
| 1343 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neon_mlal_lane_prfm) |
| 1344 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neon_mlal_lane_prfm) |
| 1345 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16__neon_mlal_lane_prfm) |
| 1346 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm) |
Marat Dukhan | e76478b | 2021-06-28 16:35:40 -0700 | [diff] [blame] | 1347 | |
Frank Barchard | f623740 | 2022-01-05 00:26:09 -0800 | [diff] [blame] | 1348 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane) |
| 1349 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8__neonv8_mlal_lane) |
| 1350 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x8__neonv8_mlal_lane) |
| 1351 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__neonv8_mlal_lane) |
| 1352 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane) |
Marat Dukhan | e76478b | 2021-06-28 16:35:40 -0700 | [diff] [blame] | 1353 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane) |
Frank Barchard | f623740 | 2022-01-05 00:26:09 -0800 | [diff] [blame] | 1354 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane) |
| 1355 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane) |
Marat Dukhan | e76478b | 2021-06-28 16:35:40 -0700 | [diff] [blame] | 1356 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16__neonv8_mlal_lane) |
Frank Barchard | f623740 | 2022-01-05 00:26:09 -0800 | [diff] [blame] | 1357 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane) |
| 1358 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8__neonv8_mlal_lane_prfm) |
| 1359 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8__neonv8_mlal_lane_prfm) |
| 1360 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x8__neonv8_mlal_lane_prfm) |
| 1361 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__neonv8_mlal_lane_prfm) |
| 1362 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm) |
| 1363 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane_prfm) |
| 1364 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane_prfm) |
| 1365 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane_prfm) |
| 1366 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16__neonv8_mlal_lane_prfm) |
| 1367 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm) |
| 1368 | |
Marat Dukhan | e76478b | 2021-06-28 16:35:40 -0700 | [diff] [blame] | 1369 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1370 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_dup) |
| 1371 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup) |
Frank Barchard | 287952a | 2021-11-03 15:26:45 -0700 | [diff] [blame] | 1372 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1373 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_dup) |
| 1374 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup) |
Frank Barchard | 287952a | 2021-11-03 15:26:45 -0700 | [diff] [blame] | 1375 | |
Frank Barchard | 64ab1b7 | 2021-11-22 10:57:40 -0800 | [diff] [blame] | 1376 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_ld1r) |
| 1377 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r) |
| 1378 | |
| 1379 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld1r) |
| 1380 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r) |
| 1381 | |
| 1382 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neon_mlal_ld2r) |
| 1383 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r) |
| 1384 | |
| 1385 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld2r) |
| 1386 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r) |
| 1387 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1388 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4s2__neon_mlal) |
| 1389 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4s2__neon_mlal) |
Frank Barchard | eb704f7 | 2021-11-12 01:26:50 -0800 | [diff] [blame] | 1390 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1391 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4s2__neonv8_mlal) |
| 1392 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4s2__neonv8_mlal) |
Frank Barchard | eb704f7 | 2021-11-12 01:26:50 -0800 | [diff] [blame] | 1393 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1394 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup) |
| 1395 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_dup) |
Marat Dukhan | 14f325e | 2021-06-30 18:46:25 -0700 | [diff] [blame] | 1396 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1397 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup) |
| 1398 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_dup) |
Marat Dukhan | 14f325e | 2021-06-30 18:46:25 -0700 | [diff] [blame] | 1399 | |
Frank Barchard | 15eec02 | 2021-11-17 13:26:20 -0800 | [diff] [blame] | 1400 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r) |
| 1401 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld1r) |
| 1402 | |
| 1403 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r) |
| 1404 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld1r) |
| 1405 | |
| 1406 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r) |
| 1407 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld2r) |
| 1408 | |
| 1409 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r) |
| 1410 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld2r) |
| 1411 | |
Frank Barchard | 42f5c50 | 2021-11-16 10:04:21 -0800 | [diff] [blame] | 1412 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r) |
| 1413 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neon_mlal_ld4r) |
| 1414 | |
| 1415 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r) |
| 1416 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2__neonv8_mlal_ld4r) |
| 1417 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1418 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2s4__neon_mlal) |
| 1419 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2s4__neon_mlal) |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 1420 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1421 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal) |
| 1422 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal) |
Frank Barchard | c7a032d | 2021-11-10 12:37:49 -0800 | [diff] [blame] | 1423 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1424 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal) |
| 1425 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal) |
Marat Dukhan | e76478b | 2021-06-28 16:35:40 -0700 | [diff] [blame] | 1426 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1427 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal) |
| 1428 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal) |
Marat Dukhan | e76478b | 2021-06-28 16:35:40 -0700 | [diff] [blame] | 1429 | |
| 1430 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neondot) |
| 1431 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__neondot) |
| 1432 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot) |
| 1433 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot) |
| 1434 | |
| 1435 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__neondot) |
| 1436 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__neondot) |
| 1437 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot) |
| 1438 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot) |
| 1439 | |
Frank Barchard | e22685a | 2021-11-12 11:36:58 -0800 | [diff] [blame] | 1440 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal) |
| 1441 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm) |
| 1442 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_cortex_a53) |
| 1443 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm_cortex_a53) |
| 1444 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mull) |
| 1445 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal) |
| 1446 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm) |
| 1447 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_cortex_a53) |
| 1448 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53) |
| 1449 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal) |
Frank Barchard | 1663c0c | 2021-07-01 11:20:06 -0700 | [diff] [blame] | 1450 | |
Frank Barchard | 98af05c | 2021-06-30 12:15:04 -0700 | [diff] [blame] | 1451 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53) |
| 1452 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53) |
| 1453 | |
Frank Barchard | 5cffb64 | 2021-11-22 13:59:43 -0800 | [diff] [blame] | 1454 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_ld64) |
| 1455 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_ld64) |
| 1456 | |
Frank Barchard | f10af6c | 2021-06-30 12:42:29 -0700 | [diff] [blame] | 1457 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__aarch64_neondot_ld32) |
| 1458 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__aarch64_neondot_ld64) |
| 1459 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld32) |
| 1460 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld64) |
| 1461 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128) |
| 1462 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_cortex_a55) |
| 1463 | |
Frank Barchard | 87fe410 | 2021-12-28 14:42:23 -0800 | [diff] [blame] | 1464 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__aarch32_neondot_ld64) |
Frank Barchard | ac654f1 | 2022-01-24 23:51:04 -0800 | [diff] [blame] | 1465 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__aarch32_neondot_cortex_a55) |
Frank Barchard | 87fe410 | 2021-12-28 14:42:23 -0800 | [diff] [blame] | 1466 | |
Frank Barchard | 5e1a303 | 2022-01-14 13:12:41 -0800 | [diff] [blame] | 1467 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_ld64) |
| 1468 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64) |
Frank Barchard | 87fe410 | 2021-12-28 14:42:23 -0800 | [diff] [blame] | 1469 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_ld64) |
| 1470 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_ld64) |
Frank Barchard | 9e4d2aa | 2022-02-02 00:31:21 -0800 | [diff] [blame] | 1471 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_cortex_a53) |
| 1472 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_cortex_a53) |
Frank Barchard | 101271e | 2022-02-02 01:49:54 -0800 | [diff] [blame] | 1473 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53) |
| 1474 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53) |
Frank Barchard | 34251d8 | 2022-02-02 11:57:11 -0800 | [diff] [blame] | 1475 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7) |
| 1476 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7) |
Frank Barchard | 87fe410 | 2021-12-28 14:42:23 -0800 | [diff] [blame] | 1477 | |
Marat Dukhan | fc188ed | 2021-06-03 12:21:22 -0700 | [diff] [blame] | 1478 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64) |
| 1479 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld64) |
| 1480 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld64) |
| 1481 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld64) |
| 1482 | |
| 1483 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64) |
| 1484 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld64) |
| 1485 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld64) |
| 1486 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld64) |
| 1487 | |
| 1488 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld64) |
| 1489 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld64) |
| 1490 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld64) |
| 1491 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld64) |
| 1492 | |
| 1493 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld64) |
| 1494 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64) |
| 1495 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64) |
| 1496 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld64) |
| 1497 | |
| 1498 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128) |
| 1499 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld128) |
| 1500 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128) |
| 1501 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld128) |
| 1502 | |
| 1503 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128) |
| 1504 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld128) |
| 1505 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld128) |
| 1506 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128) |
| 1507 | |
| 1508 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld128) |
| 1509 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld128) |
| 1510 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld128) |
| 1511 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld128) |
| 1512 | |
| 1513 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128) |
| 1514 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld128) |
| 1515 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld128) |
| 1516 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld128) |
| 1517 | |
| 1518 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64) |
| 1519 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld64) |
| 1520 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64) |
| 1521 | |
| 1522 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64) |
| 1523 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64) |
| 1524 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64) |
| 1525 | |
| 1526 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld64) |
| 1527 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld64) |
| 1528 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64) |
| 1529 | |
| 1530 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64) |
| 1531 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64) |
| 1532 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64) |
| 1533 | |
| 1534 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld128) |
| 1535 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld128) |
| 1536 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128) |
| 1537 | |
| 1538 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128) |
| 1539 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld128) |
| 1540 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128) |
| 1541 | |
| 1542 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128) |
| 1543 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128) |
| 1544 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128) |
| 1545 | |
| 1546 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128) |
| 1547 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128) |
| 1548 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128) |
| 1549 | |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1550 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__avx2) |
| 1551 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__avx2) |
| 1552 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x8c8__avx2) |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 1553 | |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 1554 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_xw_minmax_fp32_ukernel_1x8c8__avx2) |
| 1555 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_xw_minmax_fp32_ukernel_2x8c8__avx2) |
| 1556 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_xw_minmax_fp32_ukernel_3x8c8__avx2) |
Marat Dukhan | 27203da | 2020-08-05 15:19:03 -0700 | [diff] [blame] | 1557 | |
Marat Dukhan | c3e3f1c | 2021-06-03 09:56:16 -0700 | [diff] [blame] | 1558 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx) |
| 1559 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x16c8__avx512skx) |
| 1560 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x16c8__avx512skx) |
| 1561 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx) |
| 1562 | |
Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame] | 1563 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld64) |
| 1564 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld64) |
| 1565 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld64) |
| 1566 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64) |
| 1567 | |
| 1568 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__wasmsimd_dot16x2_ld128) |
| 1569 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__wasmsimd_dot16x2_ld128) |
| 1570 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__wasmsimd_dot16x2_ld128) |
| 1571 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld128) |
| 1572 | |
Marat Dukhan | 0f1ed94 | 2021-12-08 23:25:50 -0800 | [diff] [blame] | 1573 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld64) |
| 1574 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld64) |
| 1575 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld64) |
| 1576 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld64) |
| 1577 | |
| 1578 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128) |
| 1579 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld128) |
| 1580 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2s4__wasmsimd_dot16x2_ld128) |
| 1581 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128) |
| 1582 | |
Marat Dukhan | 8dc106e | 2021-08-31 15:23:02 -0700 | [diff] [blame] | 1583 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld64) |
| 1584 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld64) |
| 1585 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld64) |
| 1586 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld64) |
| 1587 | |
| 1588 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_dot16x2_ld128) |
| 1589 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_dot16x2_ld128) |
| 1590 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_dot16x2_ld128) |
| 1591 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4c8__wasmsimd_dot16x2_ld128) |
| 1592 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1593 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_ld64) |
| 1594 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld64) |
| 1595 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld64) |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 1596 | |
Marat Dukhan | dfc2db0 | 2021-08-08 21:19:07 -0700 | [diff] [blame] | 1597 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__wasmsimd_mul16_ld128) |
| 1598 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__wasmsimd_mul16_ld128) |
| 1599 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__wasmsimd_mul16_ld128) |
Marat Dukhan | 47c1220 | 2021-06-30 15:09:34 -0700 | [diff] [blame] | 1600 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1601 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__wasm_fmagic) |
| 1602 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__wasm_fmagic) |
| 1603 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__wasm_fmagic) |
| 1604 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__wasm_fmagic) |
| 1605 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__wasm_fmagic) |
| 1606 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__wasm_fmagic) |
| 1607 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__wasm_fmagic) |
| 1608 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic) |
| 1609 | |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1610 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_fmagic) |
| 1611 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_fmagic) |
| 1612 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_fmagic) |
| 1613 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_fmagic) |
Marat Dukhan | 2ac722e | 2022-01-04 01:54:20 -0800 | [diff] [blame] | 1614 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_fmagic) |
| 1615 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_fmagic) |
| 1616 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_fmagic) |
| 1617 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_fmagic) |
Marat Dukhan | d602154 | 2021-06-30 09:04:20 -0700 | [diff] [blame] | 1618 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1619 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic) |
| 1620 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic) |
| 1621 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_imagic) |
| 1622 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_imagic) |
Marat Dukhan | 272d4d9 | 2022-01-04 15:07:14 -0800 | [diff] [blame] | 1623 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_imagic) |
| 1624 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_imagic) |
| 1625 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_imagic) |
| 1626 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_imagic) |
| 1627 | |
Marat Dukhan | 7c1115f | 2022-01-04 17:18:41 -0800 | [diff] [blame] | 1628 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_lrintf) |
| 1629 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_lrintf) |
| 1630 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x2__scalar_lrintf) |
| 1631 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x2__scalar_lrintf) |
Marat Dukhan | 272d4d9 | 2022-01-04 15:07:14 -0800 | [diff] [blame] | 1632 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf) |
| 1633 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x4__scalar_lrintf) |
| 1634 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf) |
| 1635 | DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x4__scalar_lrintf) |
| 1636 | |
Zhi An Ng | 9532079 | 2022-02-03 17:05:35 -0800 | [diff] [blame] | 1637 | #define DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(fn_name) \ |
| 1638 | XNN_INTERNAL enum xnn_status fn_name( \ |
| 1639 | struct xnn_code_buffer* code, \ |
| 1640 | size_t nc, \ |
| 1641 | size_t kc, \ |
| 1642 | const void* params); |
Zhi An Ng | 13b57dd | 2022-01-06 09:33:20 -0800 | [diff] [blame] | 1643 | |
Zhi An Ng | 9532079 | 2022-02-03 17:05:35 -0800 | [diff] [blame] | 1644 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a53) |
| 1645 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a55) |
| 1646 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a7) |
| 1647 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a75) |
| 1648 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_ld64) |
| 1649 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_prfm_cortex_a75) |
Zhi An Ng | b43b47a | 2021-12-23 16:27:22 -0800 | [diff] [blame] | 1650 | |
Zhi An Ng | 9532079 | 2022-02-03 17:05:35 -0800 | [diff] [blame] | 1651 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64) |
| 1652 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64) |
| 1653 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_qs8_gemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64) |
Zhi An Ng | ed73fb6 | 2022-01-06 10:19:18 -0800 | [diff] [blame] | 1654 | |
Zhi An Ng | 9532079 | 2022-02-03 17:05:35 -0800 | [diff] [blame] | 1655 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_qc8_gemm_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_ld64) |
| 1656 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_qc8_gemm_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_ld64) |
| 1657 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_qc8_gemm_fp32_ukernel_4x8c4__aarch32_neondot_ld64) |
| 1658 | |
| 1659 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_f32_gemm_ukernel_1x8__aarch64_neonfma_cortex_a75) |
| 1660 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_f32_gemm_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75) |
| 1661 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_f32_gemm_ukernel_6x8__aarch64_neonfma_cortex_a75) |
| 1662 | DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION(xnn_generate_f32_gemm_ukernel_6x8__aarch64_neonfma_prfm_cortex_a75) |
| 1663 | |
| 1664 | #undef DECLARE_GENERATE_GEMM_UKERNEL_FUNCTION |
Zhi An Ng | c2e2da8 | 2022-01-25 16:51:58 -0800 | [diff] [blame] | 1665 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1666 | #ifdef __cplusplus |
Marat Dukhan | 80fc932 | 2019-09-29 21:06:36 -0700 | [diff] [blame] | 1667 | } // extern "C" |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1668 | #endif |