blob: 377e3cae60e934bc0856c7911487439997c53819 [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_igemm_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_igemm_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_igemm_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_igemm_ukernel_4x8__aarch64_neonfma_cortex_a53
Frank Barchard7c8e0c72019-11-17 00:02:36 -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_igemm_ukernel_4x8__aarch64_neonfma_cortex_a75
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_igemm_ukernel_5x8__aarch64_neonfma_cortex_a75
26 k-block: 8
27 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070028 assembly: true
Frank Barcharda7fb8552019-10-23 17:14:17 -070029- name: xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a53
Frank Barchard80b537a2019-11-18 10:51:33 -080030 k-block: 4
31 pipelined: true
Frank Barcharda7fb8552019-10-23 17:14:17 -070032 assembly: true
33- name: xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a57
Frank Barchardbd419712019-10-31 14:15:36 -070034 k-block: 2
Frank Barchard7e955972019-10-11 10:34:25 -070035 assembly: true
Frank Barcharda7fb8552019-10-23 17:14:17 -070036- name: xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a73
XNNPACK Teamb455b122019-09-27 18:10:33 -070037 k-block: 8
38 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070039 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070040- name: xnn_f32_igemm_ukernel_6x8__aarch64_neonfma_cortex_a75
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_igemm_ukernel_1x12__aarch64_neonfma_cortex_a53
45 k-block: 4
46 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070047 assembly: true
XNNPACK Teamb455b122019-09-27 18:10:33 -070048- name: xnn_f32_igemm_ukernel_4x12__aarch64_neonfma_cortex_a53
49 k-block: 4
50 pipelined: true
Frank Barchard7e955972019-10-11 10:34:25 -070051 assembly: true
Frank Barchard91317c52019-11-22 10:54:35 -080052- name: xnn_f32_igemm_ukernel_1x8__neon_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070053 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080054- name: xnn_f32_igemm_ukernel_4x2__neon_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070055 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080056- name: xnn_f32_igemm_ukernel_4x4__neon_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070057 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080058- name: xnn_f32_igemm_ukernel_4x8__neon_lane_ld128
XNNPACK Teamb455b122019-09-27 18:10:33 -070059 k-block: 4
Frank Barchard91317c52019-11-22 10:54:35 -080060- name: xnn_f32_igemm_ukernel_4x8__neon_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070061 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080062- name: xnn_f32_igemm_ukernel_6x8__neon_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070063 k-block: 2
Frank Barchard69172d92019-11-26 16:22:39 -080064- name: xnn_f32_igemm_ukernel_6x8__neon_lane_ld128
65 k-block: 4
Frank Barchard5243bb02019-11-22 16:37:50 -080066- name: xnn_f32_igemm_ukernel_1x8__neonfma_lane_ld64
67 k-block: 2
68 arch:
69 - aarch64
Frank Barchard91317c52019-11-22 10:54:35 -080070- name: xnn_f32_igemm_ukernel_4x2__neonfma_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070071 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080072 arch:
73 - aarch64
74- name: xnn_f32_igemm_ukernel_4x4__neonfma_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070075 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080076 arch:
77 - aarch64
78- name: xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld128
XNNPACK Teamb455b122019-09-27 18:10:33 -070079 k-block: 4
Frank Barchard91317c52019-11-22 10:54:35 -080080 arch:
81 - aarch64
82- name: xnn_f32_igemm_ukernel_4x8__neonfma_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070083 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080084 arch:
85 - aarch64
86- name: xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld64
XNNPACK Teamb455b122019-09-27 18:10:33 -070087 k-block: 2
Frank Barchard91317c52019-11-22 10:54:35 -080088 arch:
89 - aarch64
Frank Barchard69172d92019-11-26 16:22:39 -080090- name: xnn_f32_igemm_ukernel_6x8__neonfma_lane_ld128
91 k-block: 4
92 arch:
93 - aarch64
Frank Barchard5243bb02019-11-22 16:37:50 -080094- name: xnn_f32_igemm_ukernel_1x8__neon_dup_ld64
95 k-block: 2
96- name: xnn_f32_igemm_ukernel_4x8__neon_dup_ld128
97 k-block: 4
98- name: xnn_f32_igemm_ukernel_4x8__neon_dup_ld64
99 k-block: 2
100- name: xnn_f32_igemm_ukernel_6x8__neon_dup_ld64
101 k-block: 2
Frank Barchard69172d92019-11-26 16:22:39 -0800102- name: xnn_f32_igemm_ukernel_6x8__neon_dup_ld128
103 k-block: 4
Frank Barchard5243bb02019-11-22 16:37:50 -0800104- name: xnn_f32_igemm_ukernel_1x8__neonfma_dup_ld64
105 k-block: 2
106- name: xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld128
107 k-block: 4
108- name: xnn_f32_igemm_ukernel_4x8__neonfma_dup_ld64
109 k-block: 2
110- name: xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld64
111 k-block: 2
Frank Barchard69172d92019-11-26 16:22:39 -0800112- name: xnn_f32_igemm_ukernel_6x8__neonfma_dup_ld128
113 k-block: 4
Frank Barchard5243bb02019-11-22 16:37:50 -0800114- name: xnn_f32_igemm_ukernel_1x8s4__neon
115 k-block: 4
116- name: xnn_f32_igemm_ukernel_4x8s4__neon
117 k-block: 4
118- name: xnn_f32_igemm_ukernel_6x8s4__neon
119 k-block: 4
120- name: xnn_f32_igemm_ukernel_8x8s4__neon
121 k-block: 4
Frank Barcharddf06d802019-11-20 15:53:46 -0800122- name: xnn_f32_igemm_ukernel_1x8s4__neonfma
123 k-block: 4
124- name: xnn_f32_igemm_ukernel_4x8s4__neonfma
125 k-block: 4
126- name: xnn_f32_igemm_ukernel_6x8s4__neonfma
127 k-block: 4
128- name: xnn_f32_igemm_ukernel_8x8s4__neonfma
129 k-block: 4
XNNPACK Teamb455b122019-09-27 18:10:33 -0700130- name: xnn_f32_igemm_ukernel_1x8__sse_load1
131 k-block: 1
132- name: xnn_f32_igemm_ukernel_4x8__sse_load1
133 k-block: 1
134- name: xnn_f32_igemm_ukernel_1x8__sse_dup
135 k-block: 4
136- name: xnn_f32_igemm_ukernel_4x8__sse_dup
137 k-block: 4
Frank Barchardd42bdf72019-11-20 16:39:43 -0800138- name: xnn_f32_igemm_ukernel_1x8s4__sse
139 k-block: 4
140- name: xnn_f32_igemm_ukernel_4x8s4__sse
141 k-block: 4
XNNPACK Teamb455b122019-09-27 18:10:33 -0700142- name: xnn_f32_igemm_ukernel_4x2c4__sse
143 k-block: 4
144- name: xnn_f32_igemm_ukernel_4x2c4__psimd
145 k-block: 4
Marat Dukhanfda12b82019-11-21 12:27:59 -0800146- name: xnn_f32_igemm_ukernel_1x8__avx_broadcast
147 k-block: 1
148- name: xnn_f32_igemm_ukernel_4x8__avx_broadcast
149 k-block: 1
150- name: xnn_f32_igemm_ukernel_5x8__avx_broadcast
151 k-block: 1
152- name: xnn_f32_igemm_ukernel_6x8__avx_broadcast
153 k-block: 1
154- name: xnn_f32_igemm_ukernel_7x8__avx_broadcast
155 k-block: 1
156- name: xnn_f32_igemm_ukernel_1x8__fma3_broadcast
157 k-block: 1
158- name: xnn_f32_igemm_ukernel_4x8__fma3_broadcast
159 k-block: 1
160- name: xnn_f32_igemm_ukernel_5x8__fma3_broadcast
161 k-block: 1
162- name: xnn_f32_igemm_ukernel_6x8__fma3_broadcast
163 k-block: 1
164- name: xnn_f32_igemm_ukernel_7x8__fma3_broadcast
165 k-block: 1
166- name: xnn_f32_igemm_ukernel_8x8__fma3_broadcast
167 k-block: 1
XNNPACK Teamb455b122019-09-27 18:10:33 -0700168- name: xnn_f32_igemm_ukernel_1x8__psimd_loadsplat
169 k-block: 1
170- name: xnn_f32_igemm_ukernel_4x8__psimd_loadsplat
171 k-block: 1
172- name: xnn_f32_igemm_ukernel_6x8__psimd_loadsplat
173 k-block: 1
174- name: xnn_f32_igemm_ukernel_1x8__psimd_splat
175 k-block: 4
176- name: xnn_f32_igemm_ukernel_4x8__psimd_splat
177 k-block: 4
178- name: xnn_f32_igemm_ukernel_6x8__psimd_splat
179 k-block: 4
180- name: xnn_f32_igemm_ukernel_1x8s4__psimd
181 k-block: 4
182- name: xnn_f32_igemm_ukernel_4x8s4__psimd
183 k-block: 4
184- name: xnn_f32_igemm_ukernel_6x8s4__psimd
185 k-block: 4
186- name: xnn_f32_igemm_ukernel_1x4__scalar
187 k-block: 1
188- name: xnn_f32_igemm_ukernel_2x4__scalar
189 k-block: 1
190- name: xnn_f32_igemm_ukernel_4x4__scalar
191 k-block: 1
192- name: xnn_f32_igemm_ukernel_4x2__scalar
193 k-block: 1