1 /*
2  * Copyright (c) 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 <cstddef>
26 #include <arm_neon.h>
27 
28 namespace arm_conv {
29 namespace winograd {
30 namespace weight_transform {
31 
arm_fp32_2x2_3x3(unsigned int n_channels,const float * inptr,size_t ld_weight_row,size_t ld_weight_col,float * outptr,size_t matrix_stride)32 void arm_fp32_2x2_3x3(
33   unsigned int n_channels,
34   const float *inptr, size_t ld_weight_row, size_t ld_weight_col,
35   float *outptr, size_t matrix_stride
36 )
37 {
38   constexpr auto inner_tile_i = 4u;
39   constexpr auto inner_tile_j = 4u;
40 
41 #ifdef __aarch64__
42   // For each output channel
43   for (; n_channels >= 4u; n_channels -= 4)
44   {
45     // Matrices used and computed in this kernel
46     float32x4_t w[3][3], Ww[inner_tile_i][3], V[inner_tile_i][inner_tile_j];
47 
48     // Read weights
49     for (int i = 0; i < 3; i++)
50     {
51       for (int j = 0; j < 3; j++)
52       {
53         w[i][j] = vld1q_f32(inptr + i*ld_weight_row + j*ld_weight_col);
54       }
55     }
56 
57     // Compute the matrix W w
58     for (int j = 0; j < 3; j++)
59     {
60       Ww[0][j] = w[0][j];
61 
62       // Ww[1][j] = 0.5*(w[0][j] + w[1][j] + w[2][j]);
63       Ww[1][j] = vmulq_n_f32(vaddq_f32(vaddq_f32(w[0][j], w[1][j]), w[2][j]), 0.5f);
64 
65       // Ww[2][j] = 0.5*(w[0][j] - w[1][j] + w[2][j]);
66       Ww[2][j] = vmulq_n_f32(vaddq_f32(vsubq_f32(w[0][j], w[1][j]), w[2][j]), 0.5f);
67 
68       Ww[3][j] = w[2][j];
69     }
70 
71     // Compute V = W w WT
72     for (auto i = 0u; i < inner_tile_i; i++)
73     {
74       V[i][0] = Ww[i][0];
75 
76       // V[i][1] = 0.5*(Ww[i][0] + Ww[i][1] + Ww[i][2]);
77       V[i][1] = vmulq_n_f32(vaddq_f32(vaddq_f32(Ww[i][0], Ww[i][1]), Ww[i][2]), 0.5f);
78 
79       // V[i][2] = 0.5*(Ww[i][0] - Ww[i][1] + Ww[i][2]);
80       V[i][2] = vmulq_n_f32(vaddq_f32(vsubq_f32(Ww[i][0], Ww[i][1]), Ww[i][2]), 0.5f);
81 
82       V[i][3] = Ww[i][2];
83     }
84 
85     // Store the transformed weights
86     for (auto i = 0u, m = 0u; i < inner_tile_i; i++)
87     {
88       for (auto j = 0u; j < inner_tile_j; j++, m++)
89       {
90         vst1q_f32(outptr + m*matrix_stride, V[i][j]);
91       }
92     }
93 
94     inptr += 4;
95     outptr += 4;
96   }
97 #endif // __aarch64__
98   for (; n_channels >= 2u; n_channels -= 2)
99   {
100     // Matrices used and computed in this kernel
101     float32x2_t w[3][3], Ww[inner_tile_i][3], V[inner_tile_i][inner_tile_j];
102 
103     // Read weights
104     for (int i = 0; i < 3; i++)
105     {
106       for (int j = 0; j < 3; j++)
107       {
108         w[i][j] = vld1_f32(inptr + i*ld_weight_row + j*ld_weight_col);
109       }
110     }
111 
112     // Compute the matrix W w
113     for (int j = 0; j < 3; j++)
114     {
115       Ww[0][j] = w[0][j];
116 
117       // Ww[1][j] = 0.5*(w[0][j] + w[1][j] + w[2][j]);
118       Ww[1][j] = vmul_n_f32(vadd_f32(vadd_f32(w[0][j], w[1][j]), w[2][j]), 0.5f);
119 
120       // Ww[2][j] = 0.5*(w[0][j] - w[1][j] + w[2][j]);
121       Ww[2][j] = vmul_n_f32(vadd_f32(vsub_f32(w[0][j], w[1][j]), w[2][j]), 0.5f);
122 
123       Ww[3][j] = w[2][j];
124     }
125 
126     // Compute V = W w WT
127     for (auto i = 0u; i < inner_tile_i; i++)
128     {
129       V[i][0] = Ww[i][0];
130 
131       // V[i][1] = 0.5*(Ww[i][0] + Ww[i][1] + Ww[i][2]);
132       V[i][1] = vmul_n_f32(vadd_f32(vadd_f32(Ww[i][0], Ww[i][1]), Ww[i][2]), 0.5f);
133 
134       // V[i][2] = 0.5*(Ww[i][0] - Ww[i][1] + Ww[i][2]);
135       V[i][2] = vmul_n_f32(vadd_f32(vsub_f32(Ww[i][0], Ww[i][1]), Ww[i][2]), 0.5f);
136 
137       V[i][3] = Ww[i][2];
138     }
139 
140     // Store the transformed weights
141     for (auto i = 0u, m = 0u; i < inner_tile_i; i++)
142     {
143       for (auto j = 0u; j < inner_tile_j; j++, m++)
144       {
145         vst1_f32(outptr + m*matrix_stride, V[i][j]);
146       }
147     }
148 
149     inptr += 2;
150     outptr += 2;
151   }
152   for (; n_channels; n_channels--)
153   {
154     // Matrices used and computed in this kernel
155     float w[3][3], Ww[inner_tile_i][3], V[inner_tile_i][inner_tile_j];
156 
157     // Read weights
158     for (int i = 0; i < 3; i++)
159     {
160       for (int j = 0; j < 3; j++)
161       {
162         w[i][j] = *(inptr + i*ld_weight_row + j*ld_weight_col);
163       }
164     }
165 
166     // Compute the matrix W w
167     for (int j = 0; j < 3; j++)
168     {
169       Ww[0][j] = w[0][j];
170       Ww[1][j] = 0.5*(w[0][j] + w[1][j] + w[2][j]);
171       Ww[2][j] = 0.5*(w[0][j] - w[1][j] + w[2][j]);
172       Ww[3][j] = w[2][j];
173     }
174 
175     // Compute V = W w WT
176     for (auto i = 0u; i < inner_tile_i; i++)
177     {
178       V[i][0] = Ww[i][0];
179       V[i][1] = 0.5*(Ww[i][0] + Ww[i][1] + Ww[i][2]);
180       V[i][2] = 0.5*(Ww[i][0] - Ww[i][1] + Ww[i][2]);
181       V[i][3] = Ww[i][2];
182     }
183 
184     // Store the transformed weights
185     for (auto i = 0u, m = 0u; i < inner_tile_i; i++)
186     {
187       for (auto j = 0u; j < inner_tile_j; j++, m++)
188       {
189         *(outptr + m*matrix_stride) = V[i][j];
190       }
191     }
192 
193     inptr++;
194     outptr++;
195   }
196 }
197 
198 }  // namespace weight_transform
199 }  // namespace winograd
200 }  // namespace arm_conv
201