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 #if defined(__aarch64__) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
25
26 #include <arm_neon.h>
27 #include <cstddef>
28
29 namespace arm_conv {
30 namespace winograd {
31 namespace input_transform {
32
a64_fp16_6x6(const unsigned int n_channels,const __fp16 * const input_base,const size_t input_row_stride,const size_t input_col_stride,__fp16 * outptr,const size_t matrix_stride)33 void a64_fp16_6x6(
34 const unsigned int n_channels,
35 const __fp16* const input_base,
36 const size_t input_row_stride,
37 const size_t input_col_stride,
38 __fp16* outptr,
39 const size_t matrix_stride
40 )
41 {
42 constexpr int inner_tile_rows = 6;
43 constexpr int inner_tile_cols = 6;
44
45 // Get pointers into the input tile
46 const __fp16 *x_ptrs[inner_tile_rows][inner_tile_cols];
47 for (int i = 0, xi = 0; i < inner_tile_rows; i++, xi++)
48 {
49 // Get a pointer into the row
50 const __fp16* const row_ptr = input_base + xi*input_row_stride;
51
52 for (int j = 0, xj = 0; j < inner_tile_cols; j++, xj++)
53 {
54 x_ptrs[i][j] = row_ptr + xj*input_col_stride;
55 }
56 }
57
58 // Matrices used/computed in this kernel.
59 __fp16 x[inner_tile_rows][inner_tile_cols];
60 __fp16 XTx[inner_tile_rows][inner_tile_cols];
61 __fp16 U[inner_tile_rows][inner_tile_cols];
62 for (int i = 0; i < inner_tile_rows; i++)
63 {
64 for (int j = 0; j < inner_tile_cols; j++)
65 {
66 x[i][j] = XTx[i][j] = 0.0f;
67 }
68 }
69
70 // Perform the Winograd input transformation for each channel in the input
71 // tensor.
72 int channels_remaining = n_channels;
73 for (; channels_remaining >= 8; channels_remaining -= 8)
74 {
75 // Matrices used/computed in this kernel
76 float16x8_t x[inner_tile_rows][inner_tile_cols];
77 float16x8_t XTx[inner_tile_rows][inner_tile_cols];
78 float16x8_t U[inner_tile_rows][inner_tile_cols];
79 for (int i = 0; i < inner_tile_rows; i++)
80 {
81 for (int j = 0; j < inner_tile_cols; j++)
82 {
83 x[i][j] = vdupq_n_f16(0.0f);
84 XTx[i][j] = vdupq_n_f16(0.0f);
85 }
86 }
87
88 // Read a 6x6 tile in the Winograd domain
89 for (int i = 0; i < inner_tile_rows; i++)
90 {
91 for (int j = 0; j < inner_tile_cols; j++)
92 {
93 x[i][j] = vld1q_f16(x_ptrs[i][j]);
94 x_ptrs[i][j] += 8;
95 }
96 }
97
98 // Compute XT . x
99 for (int j = 0; j < inner_tile_cols; j++)
100 {
101 // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j];
102 XTx[0][j] = vsubq_f16(vaddq_f16(x[4][j], vmulq_f16(x[0][j], vdupq_n_f16(4.0f))), vmulq_f16(x[2][j], vdupq_n_f16(5.0f)));
103
104 // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j];
105 XTx[1][j] = vsubq_f16(vaddq_f16(x[3][j], x[4][j]), vmulq_f16(vaddq_f16(x[1][j], x[2][j]), vdupq_n_f16(4.0f)));
106
107 // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j];
108 XTx[2][j] = vaddq_f16(vsubq_f16(x[4][j], x[3][j]), vmulq_f16(vsubq_f16(x[1][j], x[2][j]), vdupq_n_f16(4.0f)));
109
110 // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j];
111 XTx[3][j] = vaddq_f16(vsubq_f16(x[4][j], x[2][j]), vmulq_f16(vsubq_f16(x[3][j], x[1][j]), vdupq_n_f16(2.0f)));
112
113 // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j];
114 XTx[4][j] = vaddq_f16(vsubq_f16(x[4][j], x[2][j]), vmulq_f16(vsubq_f16(x[1][j], x[3][j]), vdupq_n_f16(2.0f)));
115
116 // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j];
117 XTx[5][j] = vsubq_f16(vaddq_f16(x[5][j], vmulq_f16(x[1][j], vdupq_n_f16(4.0f))), vmulq_f16(x[3][j], vdupq_n_f16(5.0f)));
118 }
119
120 // Compute U = XT . x . X
121 for (int i = 0; i < inner_tile_rows; i++)
122 {
123 // U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4];
124 U[i][0] = vsubq_f16(vaddq_f16(XTx[i][4], vmulq_f16(XTx[i][0], vdupq_n_f16(4.0f))), vmulq_f16(XTx[i][2], vdupq_n_f16(5.0f)));
125
126 // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4];
127 U[i][1] = vsubq_f16(vaddq_f16(XTx[i][3], XTx[i][4]), vmulq_f16(vaddq_f16(XTx[i][1], XTx[i][2]), vdupq_n_f16(4.0f)));
128
129 // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4];
130 U[i][2] = vaddq_f16(vsubq_f16(XTx[i][4], XTx[i][3]), vmulq_f16(vsubq_f16(XTx[i][1], XTx[i][2]), vdupq_n_f16(4.0f)));
131
132 // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4];
133 U[i][3] = vaddq_f16(vsubq_f16(XTx[i][4], XTx[i][2]), vmulq_f16(vsubq_f16(XTx[i][3], XTx[i][1]), vdupq_n_f16(2.0f)));
134
135 // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4];
136 U[i][4] = vaddq_f16(vsubq_f16(XTx[i][4], XTx[i][2]), vmulq_f16(vsubq_f16(XTx[i][1], XTx[i][3]), vdupq_n_f16(2.0f)));
137
138 // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5];
139 U[i][5] = vsubq_f16(vaddq_f16(XTx[i][5], vmulq_f16(XTx[i][1], vdupq_n_f16(4.0f))), vmulq_f16(XTx[i][3], vdupq_n_f16(5.0f)));
140 }
141
142 // Store the transformed matrix
143 for (int i = 0, m = 0; i < inner_tile_rows; i++)
144 {
145 for (int j = 0; j < inner_tile_cols; j++, m++)
146 {
147 vst1q_f16(outptr + m*matrix_stride, U[i][j]);
148 }
149 }
150 outptr += 8;
151 }
152 for (; channels_remaining >= 4; channels_remaining -= 4)
153 {
154 // Matrices used/computed in this kernel
155 float16x4_t x[inner_tile_rows][inner_tile_cols];
156 float16x4_t XTx[inner_tile_rows][inner_tile_cols];
157 float16x4_t U[inner_tile_rows][inner_tile_cols];
158 for (int i = 0; i < inner_tile_rows; i++)
159 {
160 for (int j = 0; j < inner_tile_cols; j++)
161 {
162 x[i][j] = vdup_n_f16(0.0f);
163 XTx[i][j] = vdup_n_f16(0.0f);
164 }
165 }
166
167 // Read a 6x6 tile in the Winograd domain
168 for (int i = 0; i < inner_tile_rows; i++)
169 {
170 for (int j = 0; j < inner_tile_cols; j++)
171 {
172 x[i][j] = vld1_f16(x_ptrs[i][j]);
173 x_ptrs[i][j] += 4;
174 }
175 }
176
177 // Compute XT . x
178 for (int j = 0; j < inner_tile_cols; j++)
179 {
180 // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j];
181 XTx[0][j] = vsub_f16(vadd_f16(x[4][j], vmul_f16(x[0][j], vdup_n_f16(4.0f))), vmul_f16(x[2][j], vdup_n_f16(5.0f)));
182
183 // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j];
184 XTx[1][j] = vsub_f16(vadd_f16(x[3][j], x[4][j]), vmul_f16(vadd_f16(x[1][j], x[2][j]), vdup_n_f16(4.0f)));
185
186 // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j];
187 XTx[2][j] = vadd_f16(vsub_f16(x[4][j], x[3][j]), vmul_f16(vsub_f16(x[1][j], x[2][j]), vdup_n_f16(4.0f)));
188
189 // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j];
190 XTx[3][j] = vadd_f16(vsub_f16(x[4][j], x[2][j]), vmul_f16(vsub_f16(x[3][j], x[1][j]), vdup_n_f16(2.0f)));
191
192 // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j];
193 XTx[4][j] = vadd_f16(vsub_f16(x[4][j], x[2][j]), vmul_f16(vsub_f16(x[1][j], x[3][j]), vdup_n_f16(2.0f)));
194
195 // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j];
196 XTx[5][j] = vsub_f16(vadd_f16(x[5][j], vmul_f16(x[1][j], vdup_n_f16(4.0f))), vmul_f16(x[3][j], vdup_n_f16(5.0f)));
197 }
198
199 // Compute U = XT . x . X
200 for (int i = 0; i < inner_tile_rows; i++)
201 {
202 // U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4];
203 U[i][0] = vsub_f16(vadd_f16(XTx[i][4], vmul_f16(XTx[i][0], vdup_n_f16(4.0f))), vmul_f16(XTx[i][2], vdup_n_f16(5.0f)));
204
205 // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4];
206 U[i][1] = vsub_f16(vadd_f16(XTx[i][3], XTx[i][4]), vmul_f16(vadd_f16(XTx[i][1], XTx[i][2]), vdup_n_f16(4.0f)));
207
208 // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4];
209 U[i][2] = vadd_f16(vsub_f16(XTx[i][4], XTx[i][3]), vmul_f16(vsub_f16(XTx[i][1], XTx[i][2]), vdup_n_f16(4.0f)));
210
211 // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4];
212 U[i][3] = vadd_f16(vsub_f16(XTx[i][4], XTx[i][2]), vmul_f16(vsub_f16(XTx[i][3], XTx[i][1]), vdup_n_f16(2.0f)));
213
214 // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4];
215 U[i][4] = vadd_f16(vsub_f16(XTx[i][4], XTx[i][2]), vmul_f16(vsub_f16(XTx[i][1], XTx[i][3]), vdup_n_f16(2.0f)));
216
217 // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5];
218 U[i][5] = vsub_f16(vadd_f16(XTx[i][5], vmul_f16(XTx[i][1], vdup_n_f16(4.0f))), vmul_f16(XTx[i][3], vdup_n_f16(5.0f)));
219 }
220
221 // Store the transformed matrix
222 for (int i = 0, m = 0; i < inner_tile_rows; i++)
223 {
224 for (int j = 0; j < inner_tile_cols; j++, m++)
225 {
226 vst1_f16(outptr + m*matrix_stride, U[i][j]);
227 }
228 }
229 outptr += 4;
230 }
231 for (; channels_remaining; channels_remaining--)
232 {
233 // Load x
234 for (int i = 0; i < inner_tile_rows; i++)
235 {
236 for (int j = 0; j < inner_tile_cols; j++)
237 {
238 x[i][j] = *(x_ptrs[i][j]++);
239 }
240 }
241
242 // Compute XT . x
243 for (int j = 0; j < inner_tile_cols; j++)
244 {
245 XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j];
246 XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j];
247 XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j];
248 XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j];
249 XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j];
250 XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j];
251 }
252
253 // Compute U = XT . x . X
254 for (int i = 0; i < inner_tile_rows; i++)
255 {
256 U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4];
257 U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4];
258 U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4];
259 U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4];
260 U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4];
261 U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5];
262 }
263
264 // Store the transformed matrix
265 for (int i = 0, m = 0; i < inner_tile_rows; i++)
266 {
267 for (int j = 0; j < inner_tile_cols; j++, m++)
268 {
269 *(outptr + m*matrix_stride) = U[i][j];
270 }
271 }
272 outptr++;
273 }
274 }
275
276 } // namespace input_transform
277 } // namespace winograd
278 } // namespace arm_conv
279
280 #endif // defined(__aarch64__) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
281