2 * Copyright (c) 2017-2018 ARM Limited.
4 * SPDX-License-Identifier: MIT
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:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
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
25 #include "helpers_asymm.h"
27 #if defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)
30 #error "Stride X not supported"
31 #endif /* CONV_STRIDE_X > 3 */
33 #if CONV_STRIDE_X == 1
34 #define GET_VALUES(first_value, left, middle, right) \
36 int8 temp0 = CONVERT(vload8(0, first_value), int8); \
37 int2 temp1 = CONVERT(vload2(0, (first_value + 8 * sizeof(uchar))), int2); \
39 left = CONVERT(temp0.s01234567, int8); \
40 middle = CONVERT((int8)(temp0.s1234, temp0.s567, temp1.s0), int8); \
41 right = CONVERT((int8)(temp0.s2345, temp0.s67, temp1.s01), int8); \
43 #elif CONV_STRIDE_X == 2
44 #define GET_VALUES(first_value, left, middle, right) \
46 int16 temp0 = CONVERT(vload16(0, first_value), int16); \
47 int temp1 = CONVERT(*(first_value + 16 * sizeof(uchar)), int); \
49 left = CONVERT(temp0.s02468ace, int8); \
50 middle = CONVERT(temp0.s13579bdf, int8); \
51 right = CONVERT((int8)(temp0.s2468, temp0.sace, temp1), int8); \
53 #else /* CONV_STRIDE_X */
54 #define GET_VALUES(first_value, left, middle, right) \
56 int16 temp0 = CONVERT(vload16(0, first_value), int16); \
57 int8 temp1 = CONVERT(vload8(0, (first_value + 16 * sizeof(uchar))), int8); \
59 left = CONVERT((int8)(temp0.s0369, temp0.scf, temp1.s25), int8); \
60 middle = CONVERT((int8)(temp0.s147a, temp0.sd, temp1.s036), int8); \
61 right = CONVERT((int8)(temp0.s258b, temp0.se, temp1.s147), int8); \
63 #endif /* CONV_STRIDE_X */
65 /** This function computes the horizontal integral of the image and adds offsets.
67 * @param[in] src_ptr Pointer to the source image. Supported data types: QASYMM8
68 * @param[in] src_stride_x Stride of the source image in X dimension (in bytes)
69 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
70 * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes)
71 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
72 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image
73 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
74 * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes)
75 * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: QASYMM8
76 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
77 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
78 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
79 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
80 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
81 * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes)
82 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
83 * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: QASYMM8
84 * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
85 * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
86 * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
87 * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes)
88 * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
89 * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes)
90 * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
91 * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: QASYMM8
92 * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
93 * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
94 * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
97 __kernel void depthwise_convolution_3x3_quantized(
98 TENSOR3D_DECLARATION(src),
99 TENSOR3D_DECLARATION(dst),
100 TENSOR3D_DECLARATION(weights)
101 #if defined(HAS_BIAS)
103 VECTOR_DECLARATION(biases)
104 #endif //defined(HAS_BIAS)
107 Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
108 Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
109 Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT(weights);
110 #if defined(HAS_BIAS)
111 Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
113 int bias_value = *((__global int *)(vector_offset(&biases, get_global_id(2))));
114 #endif //defined(HAS_BIAS)
116 uchar3 w0 = vload3(0, weights.ptr + 0 * weights_stride_y);
117 uchar3 w1 = vload3(0, weights.ptr + 1 * weights_stride_y);
118 uchar3 w2 = vload3(0, weights.ptr + 2 * weights_stride_y);
122 #if CONV_STRIDE_Y == 1
125 #endif /* CONV_STRIDE_Y */
128 int8 left, middle, right;
129 GET_VALUES(src.ptr + 0 * src_stride_y, left, middle, right);
130 values0 += left * (int8)(w0.s0);
131 values0 += middle * (int8)(w0.s1);
132 values0 += right * (int8)(w0.s2);
134 #if WEIGHTS_OFFSET != 0
135 sum0 += left + middle + right;
136 #endif /* WEIGHTS_OFFSET != 0 */
139 GET_VALUES(src.ptr + 1 * src_stride_y, left, middle, right);
140 values0 += left * (int8)(w1.s0);
141 values0 += middle * (int8)(w1.s1);
142 values0 += right * (int8)(w1.s2);
143 #if CONV_STRIDE_Y == 1
144 values1 += left * (int8)(w0.s0);
145 values1 += middle * (int8)(w0.s1);
146 values1 += right * (int8)(w0.s2);
147 #endif /* CONV_STRIDE_Y == 1 */
149 #if WEIGHTS_OFFSET != 0
150 int8 tmp = left + middle + right;
152 #if CONV_STRIDE_Y == 1
154 #endif /* CONV_STRIDE_Y == 1 */
155 #endif /* WEIGHTS_OFFSET != 0 */
158 GET_VALUES(src.ptr + 2 * src_stride_y, left, middle, right);
159 values0 += left * (int8)(w2.s0);
160 values0 += middle * (int8)(w2.s1);
161 values0 += right * (int8)(w2.s2);
162 #if CONV_STRIDE_Y == 1
163 values1 += left * (int8)(w1.s0);
164 values1 += middle * (int8)(w1.s1);
165 values1 += right * (int8)(w1.s2);
166 #endif /* CONV_STRIDE_Y == 1 */
168 #if WEIGHTS_OFFSET != 0
169 tmp = left + middle + right;
171 #if CONV_STRIDE_Y == 1
173 #endif /* CONV_STRIDE_Y == 1 */
174 #endif /* WEIGHTS_OFFSET != 0 */
176 #if CONV_STRIDE_Y == 1
178 GET_VALUES(src.ptr + 3 * src_stride_y, left, middle, right);
179 values1 += left * (int8)(w2.s0);
180 values1 += middle * (int8)(w2.s1);
181 values1 += right * (int8)(w2.s2);
183 #if WEIGHTS_OFFSET != 0
184 sum1 += left + middle + right;
185 #endif /* WEIGHTS_OFFSET != 0 */
186 #endif /* CONV_STRIDE_Y == 1 */
188 #if defined(HAS_BIAS)
189 values0 += (int8)(bias_value);
190 #if CONV_STRIDE_Y == 1
191 values1 += (int8)(bias_value);
192 #endif /* CONV_STRIDE_Y == 1 */
193 #endif //defined(HAS_BIAS)
195 #if WEIGHTS_OFFSET != 0
196 values0 += sum0 * (int8)(WEIGHTS_OFFSET);
197 #if CONV_STRIDE_Y == 1
198 values1 += sum1 * (int8)(WEIGHTS_OFFSET);
199 #endif /* CONV_STRIDE_Y == 1 */
200 #endif /* WEIGHTS_OFFSET != 0 */
202 #if INPUT_OFFSET != 0
203 ushort sum_weights = 0;
204 ushort3 tmp_we = convert_ushort3(w0) + convert_ushort3(w1) + convert_ushort3(w2);
205 sum_weights += tmp_we.s0 + tmp_we.s1 + tmp_we.s2;
206 values0 += sum_weights * (int8)(INPUT_OFFSET);
207 #if CONV_STRIDE_Y == 1
208 values1 += sum_weights * (int8)(INPUT_OFFSET);
209 #endif /* CONV_STRIDE_Y == 1 */
210 #endif /* INPUT_OFFSET != 0 */
213 values0 += (int8)(K_OFFSET);
214 #if CONV_STRIDE_Y == 1
215 values1 += (int8)(K_OFFSET);
216 #endif /* CONV_STRIDE_Y == 1 */
217 #endif /* K_OFFSET != 0 */
219 values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
220 values0 += (int8)OUTPUT_OFFSET;
221 uchar8 res0 = convert_uchar8_sat(values0);
222 res0 = max(res0, (uchar8)0);
223 res0 = min(res0, (uchar8)255);
225 vstore8(res0, 0, dst.ptr);
226 #if CONV_STRIDE_Y == 1
228 values1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8);
229 values1 += (int8)OUTPUT_OFFSET;
230 uchar8 res1 = convert_uchar8_sat(values1);
231 res1 = max(res1, (uchar8)0);
232 res1 = min(res1, (uchar8)255);
234 vstore8(res1, 0, dst.ptr + dst_stride_y);
235 #endif /* CONV_STRIDE_Y == 1 */
238 #endif /* defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) */