1 /*
2 * Copyright (c) 2020 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 #include "arm.hpp"
26 #include "input.hpp"
27
28 namespace winograd
29 {
30 template <>
transform_tile(const int n_channels,const __fp16 * const input_base,const int input_row_stride,const int input_col_stride,__fp16 * outptr,const int matrix_stride)31 void InputTransform<6, 6, __fp16, __fp16, WinogradRoots::Integers>::transform_tile(
32 const int n_channels,
33 const __fp16* const input_base,
34 const int input_row_stride,
35 const int input_col_stride,
36 __fp16* outptr,
37 const int matrix_stride
38 )
39 {
40 constexpr int inner_tile_rows = 6;
41 constexpr int inner_tile_cols = 6;
42
43 // Get pointers into the input tile
44 const __fp16 *x_ptrs[inner_tile_rows][inner_tile_cols];
45 for (int i = 0, xi = 0; i < inner_tile_rows; i++, xi++)
46 {
47 // Get a pointer into the row
48 const __fp16* const row_ptr = input_base + xi*input_row_stride;
49
50 for (int j = 0, xj = 0; j < inner_tile_cols; j++, xj++)
51 {
52 x_ptrs[i][j] = row_ptr + xj*input_col_stride;
53 }
54 }
55
56 // Matrices used/computed in this kernel.
57 __fp16 x[inner_tile_rows][inner_tile_cols];
58 __fp16 XTx[inner_tile_rows][inner_tile_cols];
59 __fp16 U[inner_tile_rows][inner_tile_cols];
60 for (int i = 0; i < inner_tile_rows; i++)
61 {
62 for (int j = 0; j < inner_tile_cols; j++)
63 {
64 x[i][j] = XTx[i][j] = 0.0f;
65 }
66 }
67
68 // Perform the Winograd input transformation for each channel in the input
69 // tensor.
70 int channels_remaining = n_channels;
71 for (; channels_remaining >= 8; channels_remaining -= 8)
72 {
73 // Matrices used/computed in this kernel
74 float16x8_t x[inner_tile_rows][inner_tile_cols];
75 float16x8_t XTx[inner_tile_rows][inner_tile_cols];
76 float16x8_t U[inner_tile_rows][inner_tile_cols];
77 for (int i = 0; i < inner_tile_rows; i++)
78 {
79 for (int j = 0; j < inner_tile_cols; j++)
80 {
81 x[i][j] = vdupq_n_f16(0.0f);
82 XTx[i][j] = vdupq_n_f16(0.0f);
83 }
84 }
85
86 // Read a 6x6 tile in the Winograd domain
87 for (int i = 0; i < inner_tile_rows; i++)
88 {
89 for (int j = 0; j < inner_tile_cols; j++)
90 {
91 x[i][j] = vld1q_f16(x_ptrs[i][j]);
92 x_ptrs[i][j] += 8;
93 }
94 }
95
96 // Compute XT . x
97 for (int j = 0; j < inner_tile_cols; j++)
98 {
99 // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j];
100 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)));
101
102 // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j];
103 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)));
104
105 // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j];
106 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)));
107
108 // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j];
109 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)));
110
111 // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j];
112 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)));
113
114 // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j];
115 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)));
116 }
117
118 // Compute U = XT . x . X
119 for (int i = 0; i < inner_tile_rows; i++)
120 {
121 // U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4];
122 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)));
123
124 // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4];
125 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)));
126
127 // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4];
128 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)));
129
130 // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4];
131 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)));
132
133 // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4];
134 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)));
135
136 // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5];
137 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)));
138 }
139
140 // Store the transformed matrix
141 for (int i = 0, m = 0; i < inner_tile_rows; i++)
142 {
143 for (int j = 0; j < inner_tile_cols; j++, m++)
144 {
145 vst1q_f16(outptr + m*matrix_stride, U[i][j]);
146 }
147 }
148 outptr += 8;
149 }
150 for (; channels_remaining >= 4; channels_remaining -= 4)
151 {
152 // Matrices used/computed in this kernel
153 float16x4_t x[inner_tile_rows][inner_tile_cols];
154 float16x4_t XTx[inner_tile_rows][inner_tile_cols];
155 float16x4_t U[inner_tile_rows][inner_tile_cols];
156 for (int i = 0; i < inner_tile_rows; i++)
157 {
158 for (int j = 0; j < inner_tile_cols; j++)
159 {
160 x[i][j] = vdup_n_f16(0.0f);
161 XTx[i][j] = vdup_n_f16(0.0f);
162 }
163 }
164
165 // Read a 6x6 tile in the Winograd domain
166 for (int i = 0; i < inner_tile_rows; i++)
167 {
168 for (int j = 0; j < inner_tile_cols; j++)
169 {
170 x[i][j] = vld1_f16(x_ptrs[i][j]);
171 x_ptrs[i][j] += 4;
172 }
173 }
174
175 // Compute XT . x
176 for (int j = 0; j < inner_tile_cols; j++)
177 {
178 // XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j];
179 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)));
180
181 // XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j];
182 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)));
183
184 // XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j];
185 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)));
186
187 // XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j];
188 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)));
189
190 // XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j];
191 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)));
192
193 // XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j];
194 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)));
195 }
196
197 // Compute U = XT . x . X
198 for (int i = 0; i < inner_tile_rows; i++)
199 {
200 // U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4];
201 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)));
202
203 // U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4];
204 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)));
205
206 // U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4];
207 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)));
208
209 // U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4];
210 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)));
211
212 // U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4];
213 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)));
214
215 // U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5];
216 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)));
217 }
218
219 // Store the transformed matrix
220 for (int i = 0, m = 0; i < inner_tile_rows; i++)
221 {
222 for (int j = 0; j < inner_tile_cols; j++, m++)
223 {
224 vst1_f16(outptr + m*matrix_stride, U[i][j]);
225 }
226 }
227 outptr += 4;
228 }
229 for (; channels_remaining; channels_remaining--)
230 {
231 // Load x
232 for (int i = 0; i < inner_tile_rows; i++)
233 {
234 for (int j = 0; j < inner_tile_cols; j++)
235 {
236 x[i][j] = *(x_ptrs[i][j]++);
237 }
238 }
239
240 // Compute XT . x
241 for (int j = 0; j < inner_tile_cols; j++)
242 {
243 XTx[0][j] = 4*x[0][j] + -5*x[2][j] + 1*x[4][j];
244 XTx[1][j] = -4*x[1][j] + -4*x[2][j] + 1*x[3][j] + 1*x[4][j];
245 XTx[2][j] = 4*x[1][j] + -4*x[2][j] + -1*x[3][j] + 1*x[4][j];
246 XTx[3][j] = -2*x[1][j] + -1*x[2][j] + 2*x[3][j] + 1*x[4][j];
247 XTx[4][j] = 2*x[1][j] + -1*x[2][j] + -2*x[3][j] + 1*x[4][j];
248 XTx[5][j] = 4*x[1][j] + -5*x[3][j] + 1*x[5][j];
249 }
250
251 // Compute U = XT . x . X
252 for (int i = 0; i < inner_tile_rows; i++)
253 {
254 U[i][0] = 4*XTx[i][0] + -5*XTx[i][2] + 1*XTx[i][4];
255 U[i][1] = -4*XTx[i][1] + -4*XTx[i][2] + 1*XTx[i][3] + 1*XTx[i][4];
256 U[i][2] = 4*XTx[i][1] + -4*XTx[i][2] + -1*XTx[i][3] + 1*XTx[i][4];
257 U[i][3] = -2*XTx[i][1] + -1*XTx[i][2] + 2*XTx[i][3] + 1*XTx[i][4];
258 U[i][4] = 2*XTx[i][1] + -1*XTx[i][2] + -2*XTx[i][3] + 1*XTx[i][4];
259 U[i][5] = 4*XTx[i][1] + -5*XTx[i][3] + 1*XTx[i][5];
260 }
261
262 // Store the transformed matrix
263 for (int i = 0, m = 0; i < inner_tile_rows; i++)
264 {
265 for (int j = 0; j < inner_tile_cols; j++, m++)
266 {
267 *(outptr + m*matrix_stride) = U[i][j];
268 }
269 }
270 outptr++;
271 }
272 }
273
274 template class InputTransform<6, 6, __fp16, __fp16, WinogradRoots::Integers>;
275
276 } // namespace winograd
277 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC