1 /*
2 * Copyright (c) 2021-2022 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24
25 #include "arm_gemm_local.hpp"
26
27 #include "pooling_implementation.hpp"
28 #include "pooling_depthfirst.hpp"
29 #include "pooling_depthfirst_generic.hpp"
30
31 #include "kernels/cpp_nhwc_1x1_stride_any_depthfirst.hpp"
32 #if defined(__aarch64__)
33 #if defined(ARM_COMPUTE_ENABLE_SME)
34 #include "kernels/sme_u8_nhwc_avg_generic_depthfirst.hpp"
35 #include "kernels/sme_u8_nhwc_max_2x2_s1_output2x2_depthfirst.hpp"
36 #include "kernels/sme_u8_nhwc_max_generic_depthfirst.hpp"
37 #endif // defined(ARM_COMPUTE_ENABLE_SME)
38 #if defined(ARM_COMPUTE_ENABLE_SVE)
39 #include "kernels/sve_u8_nhwc_avg_generic_depthfirst.hpp"
40 #include "kernels/sve_u8_nhwc_max_2x2_s1_output2x2_depthfirst.hpp"
41 #include "kernels/sve_u8_nhwc_max_generic_depthfirst.hpp"
42 #endif // defined(ARM_COMPUTE_ENABLE_SVE)
43 #include "kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst.hpp"
44 #include "kernels/a64_u8_nhwc_avg_generic_depthfirst.hpp"
45 #include "kernels/a64_u8_nhwc_max_generic_depthfirst.hpp"
46 #endif // defined(__aarch64__)
47
48 #include <cstdint>
49
50 namespace arm_conv {
51 namespace pooling {
52
53 static const PoolingImplementation<uint8_t, uint8_t> pooling_u8_methods[] = {
54 {
55 PoolingMethod::DEPTHFIRST,
56 "cpp_u8_nhwc_1x1_stride_any_depthfirst",
__anonf8850bd90102() 57 [] (const PoolingArgs &args, const Nothing &) -> bool {
58 return args.pool_window.rows == 1 && args.pool_window.cols == 1;
59 },
60 nullptr,
__anonf8850bd90202() 61 [] (const PoolingArgs &args, const Nothing &) -> PoolingCommon<uint8_t, uint8_t> * {
62 auto strat = new cpp_nhwc_1x1_stride_any_depthfirst<uint8_t>(args.cpu_info);
63 return new PoolingDepthfirstGeneric<uint8_t>(strat, args);
64 },
65 },
66 #if defined(__aarch64__)
67 #if defined(ARM_COMPUTE_ENABLE_SME)
68 {
69 PoolingMethod::DEPTHFIRST,
70 "sme_u8_nhwc_max_2x2_s1_output2x2_depthfirst",
__anonf8850bd90302() 71 [] (const PoolingArgs &args, const Nothing &os) -> bool {
72 return args.cpu_info->has_sme() &&
73 is_supported<sme_u8_nhwc_max_2x2_s1_output2x2_depthfirst>(args, os);
74 },
75 nullptr,
__anonf8850bd90402() 76 [] (const PoolingArgs &args, const Nothing &) -> PoolingCommon<uint8_t, uint8_t> * {
77 auto strat = new sme_u8_nhwc_max_2x2_s1_output2x2_depthfirst(args.cpu_info);
78 return new PoolingDepthfirst<uint8_t>(strat, args);
79 },
80 },
81 {
82 PoolingMethod::DEPTHFIRST,
83 "sme_u8_nhwc_avg_generic_depthfirst",
__anonf8850bd90502() 84 [] (const PoolingArgs &args, const Nothing &) -> bool {
85 // This kernel can only be used when there is either no padding, or we don't care
86 // about the value of the padding. Otherwise, we would need to pass in the zero-point
87 // for the quantization regime.
88 return (args.exclude_padding ||
89 (args.padding.top == 0 && args.padding.bottom == 0 &&
90 args.padding.left == 0 && args.padding.right == 0)
91 ) && args.pool_type == PoolingType::AVERAGE &&
92 args.cpu_info->has_sme2();
93 },
94 nullptr,
__anonf8850bd90602() 95 [] (const PoolingArgs &args, const Nothing &) -> PoolingCommon<uint8_t, uint8_t> * {
96 auto strat = new sme_u8_nhwc_avg_generic_depthfirst(args.cpu_info);
97 return new PoolingDepthfirstGeneric<uint8_t>(strat, args);
98 },
99 },
100 {
101 PoolingMethod::DEPTHFIRST,
102 "sme_u8_nhwc_max_generic_depthfirst",
__anonf8850bd90702() 103 [] (const PoolingArgs &args, const Nothing &) -> bool {
104 return args.cpu_info->has_sme() && args.pool_type == PoolingType::MAX;
105 },
106 nullptr,
__anonf8850bd90802() 107 [] (const PoolingArgs &args, const Nothing &) -> PoolingCommon<uint8_t, uint8_t> * {
108 auto strat = new sme_u8_nhwc_max_generic_depthfirst(args.cpu_info);
109 return new PoolingDepthfirstGeneric<uint8_t>(strat, args);
110 },
111 },
112 #endif // defined(ARM_COMPUTE_ENABLE_SME)
113 #if defined(ARM_COMPUTE_ENABLE_SVE)
114 {
115 PoolingMethod::DEPTHFIRST,
116 "sve_u8_nhwc_max_2x2_s1_output2x2_depthfirst",
__anonf8850bd90902() 117 [] (const PoolingArgs &args, const Nothing &os) -> bool {
118 return args.cpu_info->has_sve() &&
119 is_supported<sve_u8_nhwc_max_2x2_s1_output2x2_depthfirst>(args, os);
120 },
121 nullptr,
__anonf8850bd90a02() 122 [] (const PoolingArgs &args, const Nothing &) -> PoolingCommon<uint8_t, uint8_t> * {
123 auto strat = new sve_u8_nhwc_max_2x2_s1_output2x2_depthfirst(args.cpu_info);
124 return new PoolingDepthfirst<uint8_t>(strat, args);
125 },
126 },
127 {
128 PoolingMethod::DEPTHFIRST,
129 "sve_u8_nhwc_avg_generic_depthfirst",
__anonf8850bd90b02() 130 [] (const PoolingArgs &args, const Nothing &) -> bool {
131 // This kernel can only be used when there is either no padding, or we don't care
132 // about the value of the padding. Otherwise, we would need to pass in the zero-point
133 // for the quantization regime.
134 return (args.exclude_padding ||
135 (args.padding.top == 0 && args.padding.bottom == 0 &&
136 args.padding.left == 0 && args.padding.right == 0)
137 ) && args.pool_type == PoolingType::AVERAGE &&
138 args.cpu_info->has_sve2();
139 },
140 nullptr,
__anonf8850bd90c02() 141 [] (const PoolingArgs &args, const Nothing &) -> PoolingCommon<uint8_t, uint8_t> * {
142 auto strat = new sve_u8_nhwc_avg_generic_depthfirst(args.cpu_info);
143 return new PoolingDepthfirstGeneric<uint8_t>(strat, args);
144 },
145 },
146 {
147 PoolingMethod::DEPTHFIRST,
148 "sve_u8_nhwc_max_generic_depthfirst",
__anonf8850bd90d02() 149 [] (const PoolingArgs &args, const Nothing &) -> bool {
150 return args.cpu_info->has_sve() && args.pool_type == PoolingType::MAX;
151 },
152 nullptr,
__anonf8850bd90e02() 153 [] (const PoolingArgs &args, const Nothing &) -> PoolingCommon<uint8_t, uint8_t> * {
154 auto strat = new sve_u8_nhwc_max_generic_depthfirst(args.cpu_info);
155 return new PoolingDepthfirstGeneric<uint8_t>(strat, args);
156 },
157 },
158 #endif // defined(ARM_COMPUTE_ENABLE_SVE)
159 {
160 PoolingMethod::DEPTHFIRST,
161 "a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst",
162 is_supported<a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst>,
163 nullptr,
__anonf8850bd90f02() 164 [] (const PoolingArgs &args, const Nothing &) -> PoolingCommon<uint8_t, uint8_t> * {
165 auto strat = new a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst(args.cpu_info);
166 return new PoolingDepthfirst<uint8_t>(strat, args);
167 },
168 },
169 {
170 PoolingMethod::DEPTHFIRST,
171 "a64_u8_nhwc_avg_generic_depthfirst",
__anonf8850bd91002() 172 [] (const PoolingArgs &args, const Nothing &) -> bool {
173 // This kernel can only be used when there is either no padding, or we don't care
174 // about the value of the padding. Otherwise, we would need to pass in the zero-point
175 // for the quantization regime.
176 return (args.exclude_padding ||
177 (args.padding.top == 0 && args.padding.bottom == 0 &&
178 args.padding.left == 0 && args.padding.right == 0)
179 ) && args.pool_type == PoolingType::AVERAGE;
180 },
181 nullptr,
__anonf8850bd91102() 182 [] (const PoolingArgs &args, const Nothing &) -> PoolingCommon<uint8_t, uint8_t> * {
183 auto strat = new a64_u8_nhwc_avg_generic_depthfirst(args.cpu_info);
184 return new PoolingDepthfirstGeneric<uint8_t>(strat, args);
185 },
186 },
187 {
188 PoolingMethod::DEPTHFIRST,
189 "a64_u8_nhwc_max_generic_depthfirst",
__anonf8850bd91202() 190 [] (const PoolingArgs &args, const Nothing &) -> bool { return args.pool_type == PoolingType::MAX; },
191 nullptr,
__anonf8850bd91302() 192 [] (const PoolingArgs &args, const Nothing &) -> PoolingCommon<uint8_t, uint8_t> * {
193 auto strat = new a64_u8_nhwc_max_generic_depthfirst(args.cpu_info);
194 return new PoolingDepthfirstGeneric<uint8_t>(strat, args);
195 },
196 },
197 #endif // defined(__aarch64__)
198 { PoolingMethod::DEFAULT, "", nullptr, nullptr, nullptr }, // End of list
199 };
200
201 template <>
pooling_implementation_list()202 const PoolingImplementation<uint8_t, uint8_t> *pooling_implementation_list()
203 {
204 return pooling_u8_methods;
205 }
206
207 template UniquePoolingCommon<uint8_t, uint8_t> pooling(const PoolingArgs &, const Nothing &);
208
209 } // namespace pooling
210 } // namespace arm_conv
211