blob: 308a31e3384ff0bcd89d9e04cc79a4586f84cc6c [file] [log] [blame]
XNNPACK Teamb455b122019-09-27 18:10:33 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
Frank Barchard21be34f2019-10-09 19:32:19 -07005- name: xnn_f32_gemminc_ukernel_1x8__aarch64_neonfma_cortex_a53
6 k-block: 8
7 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -07008 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -07009- name: xnn_f32_gemminc_ukernel_1x8__aarch64_neonfma_cortex_a57
10 k-block: 8
11 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070012 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070013- name: xnn_f32_gemminc_ukernel_1x8__aarch64_neonfma_cortex_a75
14 k-block: 8
15 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070016 assembly: true
Frank Barchard46fb8072019-10-25 12:54:22 -070017- name: xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_cortex_a53
Frank Barchard0ecc2ab2019-11-14 10:57:48 -080018 k-block: 4
19 pipelined: true
Frank Barchard46fb8072019-10-25 12:54:22 -070020 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070021- name: xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_cortex_a57
22 k-block: 8
23 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070024 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070025- name: xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_cortex_a75
26 k-block: 8
27 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070028 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070029- name: xnn_f32_gemminc_ukernel_5x8__aarch64_neonfma_cortex_a75
30 k-block: 8
31 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070032 assembly: true
Frank Barcharda7fb8552019-10-23 17:14:17 -070033- name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a53
Frank Barchard00bf68e2019-10-27 03:00:09 -070034 k-block: 4
Frank Barcharde64f91a2019-11-11 13:18:00 -080035 pipelined: true
Frank Barcharda7fb8552019-10-23 17:14:17 -070036 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070037- name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a57
Frank Barchardbd419712019-10-31 14:15:36 -070038 k-block: 4
Frank Barchard7e955972019-10-11 10:34:25 -070039 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070040- name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a73
41 k-block: 8
42 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070043 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070044- name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_cortex_a75
45 k-block: 8
46 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070047 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070048- name: xnn_f32_gemminc_ukernel_1x12__aarch64_neonfma_cortex_a53
49 k-block: 4
50 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070051 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070052- name: xnn_f32_gemminc_ukernel_4x12__aarch64_neonfma_cortex_a53
53 k-block: 4
54 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070055 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070056- name: xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_ld64
57 k-block: 2
Frank Barchard7e955972019-10-11 10:34:25 -070058 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070059- name: xnn_f32_gemminc_ukernel_4x8__aarch64_neonfma_ld128
60 k-block: 4
Frank Barchard7e955972019-10-11 10:34:25 -070061 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070062- name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_ld64
63 k-block: 2
Frank Barchard7e955972019-10-11 10:34:25 -070064 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070065- name: xnn_f32_gemminc_ukernel_6x8__aarch64_neonfma_ld128
66 k-block: 4
Frank Barchard7e955972019-10-11 10:34:25 -070067 assembly: true
Frank Barchard91317c52019-11-22 10:54:35 -080068- name: xnn_f32_gemminc_ukernel_1x8__neon_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070069 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080070- name: xnn_f32_gemminc_ukernel_4x8__neon_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070071 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080072- name: xnn_f32_gemminc_ukernel_4x8__neon_lane_ld128
XNNPACK Teamb455b122019-09-27 18:10:33 -070073 k-block: 4
Frank Barchard91317c52019-11-22 10:54:35 -080074- name: xnn_f32_gemminc_ukernel_5x8__neon_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070075 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080076- name: xnn_f32_gemminc_ukernel_6x8__neon_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070077 k-block: 2
Frank Barchard69172d92019-11-26 16:22:39 -080078- name: xnn_f32_gemminc_ukernel_6x8__neon_lane_ld128
79 k-block: 4
Frank Barchard91317c52019-11-22 10:54:35 -080080- name: xnn_f32_gemminc_ukernel_1x8__neonfma_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070081 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080082 arch:
83 - aarch64
84- name: xnn_f32_gemminc_ukernel_4x8__neonfma_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070085 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080086 arch:
87 - aarch64
88- name: xnn_f32_gemminc_ukernel_4x8__neonfma_lane_ld128
XNNPACK Teamb455b122019-09-27 18:10:33 -070089 k-block: 4
Frank Barchard91317c52019-11-22 10:54:35 -080090 arch:
91 - aarch64
92- name: xnn_f32_gemminc_ukernel_5x8__neonfma_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070093 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080094 arch:
95 - aarch64
96- name: xnn_f32_gemminc_ukernel_6x8__neonfma_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070097 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080098 arch:
99 - aarch64
Frank Barchard69172d92019-11-26 16:22:39 -0800100- name: xnn_f32_gemminc_ukernel_6x8__neonfma_lane_ld128
101 k-block: 4
102 arch:
103 - aarch64
Frank Barchard5243bb02019-11-22 16:37:50 -0800104- name: xnn_f32_gemminc_ukernel_1x8__neon_dup_ld64
105 k-block: 2
106- name: xnn_f32_gemminc_ukernel_4x8__neon_dup_ld64
107 k-block: 2
108- name: xnn_f32_gemminc_ukernel_4x8__neon_dup_ld128
109 k-block: 4
110- name: xnn_f32_gemminc_ukernel_6x8__neon_dup_ld64
111 k-block: 2
Frank Barchard69172d92019-11-26 16:22:39 -0800112- name: xnn_f32_gemminc_ukernel_6x8__neon_dup_ld128
113 k-block: 4
Frank Barchard5243bb02019-11-22 16:37:50 -0800114- name: xnn_f32_gemminc_ukernel_1x8__neonfma_dup_ld64
115 k-block: 2
116- name: xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld64
117 k-block: 2
118- name: xnn_f32_gemminc_ukernel_4x8__neonfma_dup_ld128
119 k-block: 4
120- name: xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld64
121 k-block: 2
Frank Barchard69172d92019-11-26 16:22:39 -0800122- name: xnn_f32_gemminc_ukernel_6x8__neonfma_dup_ld128
123 k-block: 4
Frank Barchard5243bb02019-11-22 16:37:50 -0800124- name: xnn_f32_gemminc_ukernel_1x8s4__neon
125 k-block: 4
126- name: xnn_f32_gemminc_ukernel_4x8s4__neon
127 k-block: 4
128- name: xnn_f32_gemminc_ukernel_6x8s4__neon
129 k-block: 4
130- name: xnn_f32_gemminc_ukernel_8x8s4__neon
131 k-block: 4
Frank Barcharddf06d802019-11-20 15:53:46 -0800132- name: xnn_f32_gemminc_ukernel_1x8s4__neonfma
133 k-block: 4
134- name: xnn_f32_gemminc_ukernel_4x8s4__neonfma
135 k-block: 4
136- name: xnn_f32_gemminc_ukernel_6x8s4__neonfma
137 k-block: 4
138- name: xnn_f32_gemminc_ukernel_8x8s4__neonfma
139 k-block: 4
XNNPACK Teamb455b122019-09-27 18:10:33 -0700140- name: xnn_f32_gemminc_ukernel_1x8__sse_load1
141 k-block: 1
142- name: xnn_f32_gemminc_ukernel_4x8__sse_load1
143 k-block: 1
144- name: xnn_f32_gemminc_ukernel_1x8__sse_dup
145 k-block: 4
146- name: xnn_f32_gemminc_ukernel_4x8__sse_dup
147 k-block: 4
148- name: xnn_f32_gemminc_ukernel_1x8s4__sse
149 k-block: 4
150- name: xnn_f32_gemminc_ukernel_4x8s4__sse
151 k-block: 4
Marat Dukhanfda12b82019-11-21 12:27:59 -0800152- name: xnn_f32_gemminc_ukernel_1x8__avx_broadcast
153 k-block: 1
154- name: xnn_f32_gemminc_ukernel_4x8__avx_broadcast
155 k-block: 1
156- name: xnn_f32_gemminc_ukernel_5x8__avx_broadcast
157 k-block: 1
158- name: xnn_f32_gemminc_ukernel_6x8__avx_broadcast
159 k-block: 1
160- name: xnn_f32_gemminc_ukernel_7x8__avx_broadcast
161 k-block: 1
162- name: xnn_f32_gemminc_ukernel_1x8__fma3_broadcast
163 k-block: 1
164- name: xnn_f32_gemminc_ukernel_4x8__fma3_broadcast
165 k-block: 1
166- name: xnn_f32_gemminc_ukernel_5x8__fma3_broadcast
167 k-block: 1
168- name: xnn_f32_gemminc_ukernel_6x8__fma3_broadcast
169 k-block: 1
170- name: xnn_f32_gemminc_ukernel_7x8__fma3_broadcast
171 k-block: 1
172- name: xnn_f32_gemminc_ukernel_8x8__fma3_broadcast
173 k-block: 1
Marat Dukhan0f349c42019-11-27 11:58:54 -0800174- name: xnn_f32_gemminc_ukernel_1x16__avx512f_broadcast
175 k-block: 1
176- name: xnn_f32_gemminc_ukernel_4x16__avx512f_broadcast
177 k-block: 1
178- name: xnn_f32_gemminc_ukernel_5x16__avx512f_broadcast
179 k-block: 1
180- name: xnn_f32_gemminc_ukernel_6x16__avx512f_broadcast
181 k-block: 1
182- name: xnn_f32_gemminc_ukernel_7x16__avx512f_broadcast
183 k-block: 1
184- name: xnn_f32_gemminc_ukernel_8x16__avx512f_broadcast
185 k-block: 1
XNNPACK Teamb455b122019-09-27 18:10:33 -0700186- name: xnn_f32_gemminc_ukernel_1x8__psimd_loadsplat
187 k-block: 1
188- name: xnn_f32_gemminc_ukernel_4x8__psimd_loadsplat
189 k-block: 1
190- name: xnn_f32_gemminc_ukernel_6x8__psimd_loadsplat
191 k-block: 1
192- name: xnn_f32_gemminc_ukernel_1x8__psimd_splat
193 k-block: 4
194- name: xnn_f32_gemminc_ukernel_4x8__psimd_splat
195 k-block: 4
196- name: xnn_f32_gemminc_ukernel_6x8__psimd_splat
197 k-block: 4
198- name: xnn_f32_gemminc_ukernel_1x8s4__psimd
199 k-block: 4
200- name: xnn_f32_gemminc_ukernel_4x8s4__psimd
201 k-block: 4
202- name: xnn_f32_gemminc_ukernel_6x8s4__psimd
203 k-block: 4
204- name: xnn_f32_gemminc_ukernel_1x4__scalar
205 k-block: 1
206- name: xnn_f32_gemminc_ukernel_2x4__scalar
207 k-block: 1
208- name: xnn_f32_gemminc_ukernel_4x4__scalar
209 k-block: 1