1 /*
2  * Copyright (c) 2022-2023 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 "winograd_implementations.hpp"
26 #include "weight_transform.hpp"
27 
28 namespace arm_conv {
29 namespace winograd {
30 namespace weight_transform {
31 
32 #if defined(__aarch64__)
33 #if defined(ARM_COMPUTE_ENABLE_SVE)
34 #endif  // defined(ARM_COMPUTE_ENABLE_SVE)
35 #endif  // defined(__aarch64__)
36 void arm_fp32_4x4_3x3(unsigned int, const float *, size_t, size_t, float *, size_t);
37 void arm_fp32_2x2_3x3(unsigned int, const float *, size_t, size_t, float *, size_t);
38 void arm_fp32_2x2_5x5(unsigned int, const float *, size_t, size_t, float *, size_t);
39 void cpp_fp32_1x6_1x3(unsigned int, const float *, size_t, size_t, float *, size_t);
40 void cpp_fp32_1x4_1x5(unsigned int, const float *, size_t, size_t, float *, size_t);
41 void cpp_fp32_1x2_1x7(unsigned int, const float *, size_t, size_t, float *, size_t);
42 
43 #define IMPL(KERN_ROWS, KERN_COLS, TRANS_ROWS, TRANS_COLS, KERN) \
44   new Transform<float>(#KERN, KERN_ROWS, KERN_COLS, TRANS_ROWS, TRANS_COLS, KERN)
45 
46 #define IMPL_T(KERN_ROWS, KERN_COLS, TRANS_ROWS, TRANS_COLS, KERN) \
47   new Transform<float>(#KERN, KERN_ROWS, KERN_COLS, TRANS_ROWS, TRANS_COLS, Transform<float>::get_transposed_kernel(KERN))
48 
49 static const TransformImplementation<float> transforms_fp32[] = {
50 #if defined(__aarch64__)
51 #if defined(ARM_COMPUTE_ENABLE_SVE)
52 #endif  // defined(ARM_COMPUTE_ENABLE_SVE)
53 #endif  // defined(__aarch64__)
54   { IMPL(3, 3, 6, 6, arm_fp32_4x4_3x3) },
55   { IMPL(3, 3, 4, 4, arm_fp32_2x2_3x3) },
56   { IMPL(5, 5, 6, 6, arm_fp32_2x2_5x5) },
57   { IMPL(1, 3, 1, 8, cpp_fp32_1x6_1x3) },
58   { IMPL_T(3, 1, 8, 1, cpp_fp32_1x6_1x3) },
59   { IMPL(1, 5, 1, 8, cpp_fp32_1x4_1x5) },
60   { IMPL_T(5, 1, 8, 1, cpp_fp32_1x4_1x5) },
61   { IMPL(1, 7, 1, 8, cpp_fp32_1x2_1x7) },
62   { IMPL_T(7, 1, 8, 1, cpp_fp32_1x2_1x7) },
63   { nullptr }
64 };
65 
66 template <>
implementation_list(void)67 const TransformImplementation<float> *implementation_list(void)
68 {
69   return transforms_fp32;
70 }
71 
72 }  // namespace weight_transform
73 }  // namespace winograd
74 }  // namespace arm_conv
75