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 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
25 
26 #include <algorithm>
27 #include <arm_neon.h>
28 #include <cstddef>
29 
30 namespace arm_conv {
31 namespace winograd {
32 namespace output_transform {
33 
a64_fp16_4x4_3x3(unsigned int n_channels,const __fp16 * inptr,const size_t matrix_stride,const __fp16 * bptr,__fp16 * const output,const size_t output_row_stride,const size_t output_col_stride,const __fp16 output_min,const __fp16 output_max)34 void a64_fp16_4x4_3x3(
35     unsigned int n_channels,
36     const __fp16* inptr,
37     const size_t matrix_stride,
38     const __fp16* bptr,
39     __fp16* const output,
40     const size_t output_row_stride,
41     const size_t output_col_stride,
42     const __fp16 output_min,
43     const __fp16 output_max
44 )
45 {
46     constexpr int output_tile_rows = 4, output_tile_cols = 4;
47 
48     // Construct a map to the output cells
49     __fp16 *outptrs[output_tile_rows][output_tile_cols];
50     for (int i = 0; i < output_tile_rows; i++)
51     {
52         for (int j = 0; j < output_tile_cols; j++)
53         {
54             outptrs[i][j] = output + i*output_row_stride + j*output_col_stride;
55         }
56     }
57 
58     // For each channel of the output
59     int channels_remaining = n_channels;
60 
61 #ifdef __aarch64__
62     for (; channels_remaining >= 8; channels_remaining -= 8)
63   {
64     // Matrices used and computed during this transform
65     float16x8_t F[6][6], FZ[6][4], f[4][4], b;
66 
67     // Read a 6x6 tile in the Winograd domain
68     for (int i = 0, m = 0; i < 6; i++)
69     {
70       for (int j = 0; j < 6; j++, m++)
71       {
72         F[i][j] = vld1q_f16(inptr + m*matrix_stride);
73       }
74     }
75     inptr += 8;
76 
77     // Compute the matrix F Z
78     for (int i = 0; i < 6; i++)
79     {
80       // FZ[i][0] =  1*F[i][0] +  1*F[i][1] +  1*F[i][2] +  1*F[i][3] +  1*F[i][4];
81       FZ[i][0] = vaddq_f16(vaddq_f16(vaddq_f16(F[i][0], F[i][1]), vaddq_f16(F[i][2], F[i][3])), F[i][4]);
82 
83       // FZ[i][1] =  1*F[i][1] + -1*F[i][2] +  2*F[i][3] + -2*F[i][4];
84       FZ[i][1] = vaddq_f16(vsubq_f16(F[i][1], F[i][2]), vmulq_f16(vsubq_f16(F[i][3], F[i][4]), vdupq_n_f16(2.0f)));
85 
86       // FZ[i][2] =  1*F[i][1] +  1*F[i][2] +  4*F[i][3] +  4*F[i][4];
87       FZ[i][2] = vaddq_f16(vaddq_f16(F[i][1], F[i][2]), vmulq_f16(vaddq_f16(F[i][3], F[i][4]), vdupq_n_f16(4.0f)));
88 
89       // FZ[i][3] =  1*F[i][1] + -1*F[i][2] +  8*F[i][3] + -8*F[i][4] +  1*F[i][5];
90       FZ[i][3] = vaddq_f16(vaddq_f16(vsubq_f16(F[i][1], F[i][2]), vmulq_f16(vsubq_f16(F[i][3], F[i][4]), vdupq_n_f16(8.0f))), F[i][5]);
91     }
92 
93     // Compute the output tile f = ZT F Z
94     for (int j = 0; j < 4; j++)
95     {
96       // f[0][j] =  1*FZ[0][j] +  1*FZ[1][j] +  1*FZ[2][j] +  1*FZ[3][j] +  1*FZ[4][j];
97       f[0][j] = vaddq_f16(vaddq_f16(vaddq_f16(FZ[0][j], FZ[1][j]), vaddq_f16(FZ[2][j], FZ[3][j])), FZ[4][j]);
98 
99       // f[1][j] =  1*FZ[1][j] + -1*FZ[2][j] +  2*FZ[3][j] + -2*FZ[4][j];
100       f[1][j] = vaddq_f16(vsubq_f16(FZ[1][j], FZ[2][j]), vmulq_f16(vsubq_f16(FZ[3][j], FZ[4][j]), vdupq_n_f16(2.0f)));
101 
102       // f[2][j] =  1*FZ[1][j] +  1*FZ[2][j] +  4*FZ[3][j] +  4*FZ[4][j];
103       f[2][j] = vaddq_f16(vaddq_f16(FZ[1][j], FZ[2][j]), vmulq_f16(vaddq_f16(FZ[3][j], FZ[4][j]), vdupq_n_f16(4.0f)));
104 
105       // f[3][j] =  1*FZ[1][j] + -1*FZ[2][j] +  8*FZ[3][j] + -8*FZ[4][j] +  1*FZ[5][j];
106       f[3][j] = vaddq_f16(vaddq_f16(vsubq_f16(FZ[1][j], FZ[2][j]), vmulq_f16(vsubq_f16(FZ[3][j], FZ[4][j]), vdupq_n_f16(8.0f))), FZ[5][j]);
107     }
108 
109     // Write out the output tile
110     if (bptr != nullptr)
111     {
112       b = vld1q_f16(bptr);
113       bptr += 8;
114     }
115     else
116     {
117       b = vdupq_n_f16(0.0f);
118     }
119     for (int i = 0; i < output_tile_rows; i++)
120     {
121       for (int j = 0; j < output_tile_cols; j++)
122       {
123         const auto y =
124             vmaxq_f16(vminq_f16(vaddq_f16(f[i][j], b), vdupq_n_f16(output_max)),
125                      vdupq_n_f16(output_min));
126         vst1q_f16(outptrs[i][j], y);
127         outptrs[i][j] += 8;
128       }
129     }
130   }
131 #endif  // __aarch64__
132 #ifdef __arm_any__
133     for (; channels_remaining >= 4; channels_remaining -= 4)
134   {
135     // Matrices used and computed during this transform
136     float16x4_t F[6][6], FZ[6][4], f[4][4], b;
137 
138     // Read a 6x6 tile in the Winograd domain
139     for (int i = 0, m = 0; i < 6; i++)
140     {
141       for (int j = 0; j < 6; j++, m++)
142       {
143         F[i][j] = vld1_f16(inptr + m*matrix_stride);
144       }
145     }
146     inptr += 4;
147 
148     // Compute the matrix F Z
149     for (int i = 0; i < 6; i++)
150     {
151       // FZ[i][0] =  1*F[i][0] +  1*F[i][1] +  1*F[i][2] +  1*F[i][3] +  1*F[i][4];
152       FZ[i][0] = vadd_f16(vadd_f16(vadd_f16(F[i][0], F[i][1]), vadd_f16(F[i][2], F[i][3])), F[i][4]);
153 
154       // FZ[i][1] =  1*F[i][1] + -1*F[i][2] +  2*F[i][3] + -2*F[i][4];
155       FZ[i][1] = vadd_f16(vsub_f16(F[i][1], F[i][2]), vmul_f16(vsub_f16(F[i][3], F[i][4]), vdup_n_f16(2.0f)));
156 
157       // FZ[i][2] =  1*F[i][1] +  1*F[i][2] +  4*F[i][3] +  4*F[i][4];
158       FZ[i][2] = vadd_f16(vadd_f16(F[i][1], F[i][2]), vmul_f16(vadd_f16(F[i][3], F[i][4]), vdup_n_f16(4.0f)));
159 
160       // FZ[i][3] =  1*F[i][1] + -1*F[i][2] +  8*F[i][3] + -8*F[i][4] +  1*F[i][5];
161       FZ[i][3] = vadd_f16(vadd_f16(vsub_f16(F[i][1], F[i][2]), vmul_f16(vsub_f16(F[i][3], F[i][4]), vdup_n_f16(8.0f))), F[i][5]);
162     }
163 
164     // Compute the output tile f = ZT F Z
165     for (int j = 0; j < 4; j++)
166     {
167       // f[0][j] =  1*FZ[0][j] +  1*FZ[1][j] +  1*FZ[2][j] +  1*FZ[3][j] +  1*FZ[4][j];
168       f[0][j] = vadd_f16(vadd_f16(vadd_f16(FZ[0][j], FZ[1][j]), vadd_f16(FZ[2][j], FZ[3][j])), FZ[4][j]);
169 
170       // f[1][j] =  1*FZ[1][j] + -1*FZ[2][j] +  2*FZ[3][j] + -2*FZ[4][j];
171       f[1][j] = vadd_f16(vsub_f16(FZ[1][j], FZ[2][j]), vmul_f16(vsub_f16(FZ[3][j], FZ[4][j]), vdup_n_f16(2.0f)));
172 
173       // f[2][j] =  1*FZ[1][j] +  1*FZ[2][j] +  4*FZ[3][j] +  4*FZ[4][j];
174       f[2][j] = vadd_f16(vadd_f16(FZ[1][j], FZ[2][j]), vmul_f16(vadd_f16(FZ[3][j], FZ[4][j]), vdup_n_f16(4.0f)));
175 
176       // f[3][j] =  1*FZ[1][j] + -1*FZ[2][j] +  8*FZ[3][j] + -8*FZ[4][j] +  1*FZ[5][j];
177       f[3][j] = vadd_f16(vadd_f16(vsub_f16(FZ[1][j], FZ[2][j]), vmul_f16(vsub_f16(FZ[3][j], FZ[4][j]), vdup_n_f16(8.0f))), FZ[5][j]);
178     }
179 
180     // Write out the output tile
181     if (bptr != nullptr)
182     {
183       b = vld1_f16(bptr);
184       bptr += 4;
185     }
186     else
187     {
188       b = vdup_n_f16(0.0f);
189     }
190     for (int i = 0; i < output_tile_rows; i++)
191     {
192       for (int j = 0; j < output_tile_cols; j++)
193       {
194         const auto y =
195             vmax_f16(vmin_f16(vadd_f16(f[i][j], b), vdup_n_f16(output_max)),
196                      vdup_n_f16(output_min));
197         vst1_f16(outptrs[i][j], y);
198         outptrs[i][j] += 4;
199       }
200     }
201   }
202 #endif  // __arm_any__
203     for (; channels_remaining; channels_remaining--)
204     {
205         // Matrices used and computed during this transform
206         __fp16 F[6][6], FZ[6][4], f[4][4], b;
207 
208         // Read a 6x6 tile in the Winograd domain
209         for (int i = 0, m = 0; i < 6; i++)
210         {
211             for (int j = 0; j < 6; j++, m++)
212             {
213                 F[i][j] = *(inptr + m*matrix_stride);
214             }
215         }
216         inptr++;
217 
218         // Compute the matrix F Z
219         for (int i = 0; i < 6; i++)
220         {
221             FZ[i][0] =  1*F[i][0] +  1*F[i][1] +  1*F[i][2] +  1*F[i][3] +  1*F[i][4];
222             FZ[i][1] =  1*F[i][1] + -1*F[i][2] +  2*F[i][3] + -2*F[i][4];
223             FZ[i][2] =  1*F[i][1] +  1*F[i][2] +  4*F[i][3] +  4*F[i][4];
224             FZ[i][3] =  1*F[i][1] + -1*F[i][2] +  8*F[i][3] + -8*F[i][4] +  1*F[i][5];
225         }
226 
227         // Compute the output tile f = ZT F Z
228         for (int j = 0; j < 4; j++)
229         {
230             f[0][j] =  1*FZ[0][j] +  1*FZ[1][j] +  1*FZ[2][j] +  1*FZ[3][j] +  1*FZ[4][j];
231             f[1][j] =  1*FZ[1][j] + -1*FZ[2][j] +  2*FZ[3][j] + -2*FZ[4][j];
232             f[2][j] =  1*FZ[1][j] +  1*FZ[2][j] +  4*FZ[3][j] +  4*FZ[4][j];
233             f[3][j] =  1*FZ[1][j] + -1*FZ[2][j] +  8*FZ[3][j] + -8*FZ[4][j] +  1*FZ[5][j];
234         }
235 
236         // Write out the output tile
237         if (bptr != nullptr)
238         {
239             b = *(bptr++);
240         }
241         else
242         {
243             b = 0.0f;
244         }
245         for (int i = 0; i < output_tile_rows; i++)
246         {
247             for (int j = 0; j < output_tile_cols; j++)
248             {
249                 const auto y = std::max(std::min<__fp16>(f[i][j] + b, output_max), output_min);
250                 *(outptrs[i][j]++) = y;
251             }
252         }
253     }
254 }
255 
256 } // namespace output_transform
257 } // namespace winograd
258 } // namespace arm_conv
259 
260 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
261