Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 1 | /* |
Giorgio Arena | 944d3f7 | 2018-01-16 15:38:35 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2018 ARM Limited. |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 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 "helpers_asymm.h" |
| 26 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 27 | #if defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 28 | |
Giorgio Arena | 99ac60b | 2018-02-16 15:17:23 +0000 | [diff] [blame] | 29 | #if defined(FUSED_ACTIVATION) |
| 30 | #define DATA_TYPE uchar |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 31 | #ifndef VEC_SIZE |
Giorgio Arena | 99ac60b | 2018-02-16 15:17:23 +0000 | [diff] [blame] | 32 | #define VEC_SIZE 8 |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 33 | #endif /* VEC_SIZE */ |
Giorgio Arena | 99ac60b | 2018-02-16 15:17:23 +0000 | [diff] [blame] | 34 | #include "activation_layer_qa8.cl" |
| 35 | #define ACTIVATION_FUNC(x) PERFORM_ACTIVATION_QA8(FUSED_ACTIVATION, x) |
| 36 | #else /* defined(FUSED_ACTIVATION) */ |
| 37 | #define ACTIVATION_FUNC(x) (x) |
| 38 | #endif /* defined(FUSED_ACTIVATION) */ |
| 39 | |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 40 | #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
| 41 | #if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 42 | #define ARM_DOT(x0, x1, x2, x3, y0, y1, y2, y3, val) val = arm_dot_acc((uchar4)(x0, x1, x2, x3), (uchar4)(y0, y1, y2, y3), val); |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 43 | #else // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 44 | #define ARM_DOT(x0, x1, x2, x3, y0, y1, y2, y3, val) val += arm_dot((uchar4)(x0, x1, x2, x3), (uchar4)(y0, y1, y2, y3)); |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 45 | #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) |
| 46 | #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 47 | |
| 48 | #if defined(CONV_STRIDE_Y) && defined(CONV_STRIDE_X) && defined(DEPTH_MULTIPLIER) |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 49 | |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 50 | #if CONV_STRIDE_X > 3 |
| 51 | #error "Stride X not supported" |
| 52 | #endif /* CONV_STRIDE_X > 3 */ |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 53 | |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 54 | #if !(defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 55 | |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 56 | #if CONV_STRIDE_X == 1 |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 57 | #define GET_VALUES(first_value, left, middle, right) \ |
| 58 | ({ \ |
| 59 | int8 temp0 = CONVERT(vload8(0, first_value), int8); \ |
| 60 | int2 temp1 = CONVERT(vload2(0, (first_value + 8 * sizeof(uchar))), int2); \ |
| 61 | \ |
| 62 | left = CONVERT(temp0.s01234567, int8); \ |
| 63 | middle = CONVERT((int8)(temp0.s1234, temp0.s567, temp1.s0), int8); \ |
| 64 | right = CONVERT((int8)(temp0.s2345, temp0.s67, temp1.s01), int8); \ |
| 65 | }) |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 66 | #elif CONV_STRIDE_X == 2 |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 67 | #define GET_VALUES(first_value, left, middle, right) \ |
| 68 | ({ \ |
| 69 | int16 temp0 = CONVERT(vload16(0, first_value), int16); \ |
| 70 | int temp1 = CONVERT(*(first_value + 16 * sizeof(uchar)), int); \ |
| 71 | \ |
| 72 | left = CONVERT(temp0.s02468ace, int8); \ |
| 73 | middle = CONVERT(temp0.s13579bdf, int8); \ |
| 74 | right = CONVERT((int8)(temp0.s2468, temp0.sace, temp1), int8); \ |
| 75 | }) |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 76 | #else /* CONV_STRIDE_X */ |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 77 | #define GET_VALUES(first_value, left, middle, right) \ |
| 78 | ({ \ |
| 79 | int16 temp0 = CONVERT(vload16(0, first_value), int16); \ |
| 80 | int8 temp1 = CONVERT(vload8(0, (first_value + 16 * sizeof(uchar))), int8); \ |
| 81 | \ |
| 82 | left = CONVERT((int8)(temp0.s0369, temp0.scf, temp1.s25), int8); \ |
| 83 | middle = CONVERT((int8)(temp0.s147a, temp0.sd, temp1.s036), int8); \ |
| 84 | right = CONVERT((int8)(temp0.s258b, temp0.se, temp1.s147), int8); \ |
| 85 | }) |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 86 | #endif /* CONV_STRIDE_X */ |
| 87 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 88 | /** This function computes the depthwise convolution quantized. |
Anthony Barbier | f202e50 | 2017-11-23 18:02:04 +0000 | [diff] [blame] | 89 | * |
| 90 | * @param[in] src_ptr Pointer to the source image. Supported data types: QASYMM8 |
| 91 | * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) |
| 92 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 93 | * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) |
| 94 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 95 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image |
| 96 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 97 | * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) |
| 98 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: QASYMM8 |
| 99 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 100 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 101 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 102 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 103 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 104 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 105 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 106 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: QASYMM8 |
| 107 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 108 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 109 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 110 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 111 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 112 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 113 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 114 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: QASYMM8 |
| 115 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 116 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 117 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
Anthony Barbier | f202e50 | 2017-11-23 18:02:04 +0000 | [diff] [blame] | 118 | */ |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 119 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 120 | __kernel void depthwise_convolution_3x3_quantized_nchw( |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 121 | TENSOR3D_DECLARATION(src), |
| 122 | TENSOR3D_DECLARATION(dst), |
Georgios Pinitas | 5b2191e | 2018-02-22 12:56:51 +0000 | [diff] [blame] | 123 | TENSOR3D_DECLARATION(weights) |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 124 | #if defined(HAS_BIAS) |
Georgios Pinitas | 5b2191e | 2018-02-22 12:56:51 +0000 | [diff] [blame] | 125 | , |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 126 | VECTOR_DECLARATION(biases) |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 127 | #endif //defined(HAS_BIAS) |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 128 | ) |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 129 | { |
| 130 | Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src); |
| 131 | Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); |
| 132 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT(weights); |
| 133 | #if defined(HAS_BIAS) |
| 134 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 135 | |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 136 | int bias_value = *((__global int *)(vector_offset(&biases, get_global_id(2)))); |
| 137 | #endif //defined(HAS_BIAS) |
| 138 | |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 139 | src.ptr -= (get_global_id(2) - get_global_id(2) / DEPTH_MULTIPLIER) * src_step_z; |
| 140 | |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 141 | uchar3 w0 = vload3(0, weights.ptr + 0 * weights_stride_y); |
| 142 | uchar3 w1 = vload3(0, weights.ptr + 1 * weights_stride_y); |
| 143 | uchar3 w2 = vload3(0, weights.ptr + 2 * weights_stride_y); |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 144 | |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 145 | int8 values0 = 0; |
| 146 | int8 sum0 = 0; |
| 147 | #if CONV_STRIDE_Y == 1 |
| 148 | int8 values1 = 0; |
| 149 | int8 sum1 = 0; |
| 150 | #endif /* CONV_STRIDE_Y */ |
| 151 | |
| 152 | // Row0 |
| 153 | int8 left, middle, right; |
| 154 | GET_VALUES(src.ptr + 0 * src_stride_y, left, middle, right); |
| 155 | values0 += left * (int8)(w0.s0); |
| 156 | values0 += middle * (int8)(w0.s1); |
| 157 | values0 += right * (int8)(w0.s2); |
| 158 | |
| 159 | #if WEIGHTS_OFFSET != 0 |
| 160 | sum0 += left + middle + right; |
| 161 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 162 | |
| 163 | // Row1 |
| 164 | GET_VALUES(src.ptr + 1 * src_stride_y, left, middle, right); |
| 165 | values0 += left * (int8)(w1.s0); |
| 166 | values0 += middle * (int8)(w1.s1); |
| 167 | values0 += right * (int8)(w1.s2); |
| 168 | #if CONV_STRIDE_Y == 1 |
| 169 | values1 += left * (int8)(w0.s0); |
| 170 | values1 += middle * (int8)(w0.s1); |
| 171 | values1 += right * (int8)(w0.s2); |
| 172 | #endif /* CONV_STRIDE_Y == 1 */ |
| 173 | |
| 174 | #if WEIGHTS_OFFSET != 0 |
| 175 | int8 tmp = left + middle + right; |
| 176 | sum0 += tmp; |
| 177 | #if CONV_STRIDE_Y == 1 |
| 178 | sum1 += tmp; |
| 179 | #endif /* CONV_STRIDE_Y == 1 */ |
| 180 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 181 | |
| 182 | // Row2 |
| 183 | GET_VALUES(src.ptr + 2 * src_stride_y, left, middle, right); |
| 184 | values0 += left * (int8)(w2.s0); |
| 185 | values0 += middle * (int8)(w2.s1); |
| 186 | values0 += right * (int8)(w2.s2); |
| 187 | #if CONV_STRIDE_Y == 1 |
| 188 | values1 += left * (int8)(w1.s0); |
| 189 | values1 += middle * (int8)(w1.s1); |
| 190 | values1 += right * (int8)(w1.s2); |
| 191 | #endif /* CONV_STRIDE_Y == 1 */ |
| 192 | |
| 193 | #if WEIGHTS_OFFSET != 0 |
| 194 | tmp = left + middle + right; |
| 195 | sum0 += tmp; |
| 196 | #if CONV_STRIDE_Y == 1 |
| 197 | sum1 += tmp; |
| 198 | #endif /* CONV_STRIDE_Y == 1 */ |
| 199 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 200 | |
| 201 | #if CONV_STRIDE_Y == 1 |
| 202 | // Row3 |
| 203 | GET_VALUES(src.ptr + 3 * src_stride_y, left, middle, right); |
| 204 | values1 += left * (int8)(w2.s0); |
| 205 | values1 += middle * (int8)(w2.s1); |
| 206 | values1 += right * (int8)(w2.s2); |
| 207 | |
| 208 | #if WEIGHTS_OFFSET != 0 |
| 209 | sum1 += left + middle + right; |
| 210 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 211 | #endif /* CONV_STRIDE_Y == 1 */ |
| 212 | |
| 213 | #if defined(HAS_BIAS) |
| 214 | values0 += (int8)(bias_value); |
| 215 | #if CONV_STRIDE_Y == 1 |
| 216 | values1 += (int8)(bias_value); |
| 217 | #endif /* CONV_STRIDE_Y == 1 */ |
| 218 | #endif //defined(HAS_BIAS) |
| 219 | |
| 220 | #if WEIGHTS_OFFSET != 0 |
| 221 | values0 += sum0 * (int8)(WEIGHTS_OFFSET); |
| 222 | #if CONV_STRIDE_Y == 1 |
| 223 | values1 += sum1 * (int8)(WEIGHTS_OFFSET); |
| 224 | #endif /* CONV_STRIDE_Y == 1 */ |
| 225 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 226 | |
| 227 | #if INPUT_OFFSET != 0 |
| 228 | ushort sum_weights = 0; |
| 229 | ushort3 tmp_we = convert_ushort3(w0) + convert_ushort3(w1) + convert_ushort3(w2); |
| 230 | sum_weights += tmp_we.s0 + tmp_we.s1 + tmp_we.s2; |
| 231 | values0 += sum_weights * (int8)(INPUT_OFFSET); |
| 232 | #if CONV_STRIDE_Y == 1 |
| 233 | values1 += sum_weights * (int8)(INPUT_OFFSET); |
| 234 | #endif /* CONV_STRIDE_Y == 1 */ |
| 235 | #endif /* INPUT_OFFSET != 0 */ |
| 236 | |
| 237 | #if K_OFFSET != 0 |
| 238 | values0 += (int8)(K_OFFSET); |
| 239 | #if CONV_STRIDE_Y == 1 |
| 240 | values1 += (int8)(K_OFFSET); |
| 241 | #endif /* CONV_STRIDE_Y == 1 */ |
| 242 | #endif /* K_OFFSET != 0 */ |
| 243 | |
| 244 | values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8); |
| 245 | values0 += (int8)OUTPUT_OFFSET; |
| 246 | uchar8 res0 = convert_uchar8_sat(values0); |
| 247 | res0 = max(res0, (uchar8)0); |
| 248 | res0 = min(res0, (uchar8)255); |
| 249 | |
Giorgio Arena | 99ac60b | 2018-02-16 15:17:23 +0000 | [diff] [blame] | 250 | vstore8(ACTIVATION_FUNC(res0), 0, dst.ptr); |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 251 | #if CONV_STRIDE_Y == 1 |
| 252 | |
| 253 | values1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8); |
| 254 | values1 += (int8)OUTPUT_OFFSET; |
| 255 | uchar8 res1 = convert_uchar8_sat(values1); |
| 256 | res1 = max(res1, (uchar8)0); |
| 257 | res1 = min(res1, (uchar8)255); |
| 258 | |
Giorgio Arena | 99ac60b | 2018-02-16 15:17:23 +0000 | [diff] [blame] | 259 | vstore8(ACTIVATION_FUNC(res1), 0, dst.ptr + dst_stride_y); |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 260 | #endif /* CONV_STRIDE_Y == 1 */ |
Dmitry Savenko | d7295b7 | 2017-11-20 22:00:08 +0700 | [diff] [blame] | 261 | } |
Giorgio Arena | 287b570 | 2018-02-16 11:01:04 +0000 | [diff] [blame] | 262 | |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 263 | #else // !(defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)) |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 264 | |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 265 | #if CONV_STRIDE_X == 1 |
| 266 | #define GET_VALUES(first_value, left, middle, right) \ |
| 267 | ({ \ |
| 268 | uchar8 temp0 = vload8(0, first_value); \ |
| 269 | uchar2 temp1 = vload2(0, (first_value + 8 * sizeof(uchar))); \ |
| 270 | \ |
| 271 | left = temp0.s01234567; \ |
| 272 | middle = (uchar8)(temp0.s1234, temp0.s567, temp1.s0); \ |
| 273 | right = (uchar8)(temp0.s2345, temp0.s67, temp1.s01); \ |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 274 | }) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 275 | #elif CONV_STRIDE_X == 2 |
| 276 | #define GET_VALUES(first_value, left, middle, right) \ |
| 277 | ({ \ |
| 278 | uchar16 temp0 = vload16(0, first_value); \ |
| 279 | uchar temp1 = *(first_value + 16 * sizeof(uchar)); \ |
| 280 | \ |
| 281 | left = temp0.s02468ace; \ |
| 282 | middle = temp0.s13579bdf; \ |
| 283 | right = (uchar8)(temp0.s2468, temp0.sace, temp1); \ |
| 284 | }) |
| 285 | #else /* CONV_STRIDE_X */ |
| 286 | #define GET_VALUES(first_value, left, middle, right) \ |
| 287 | ({ \ |
| 288 | uchar16 temp0 = vload16(0, first_value); \ |
| 289 | uchar8 temp1 = vload8(0, (first_value + 16 * sizeof(uchar))); \ |
| 290 | \ |
| 291 | left = (uchar8)(temp0.s0369, temp0.scf, temp1.s25); \ |
| 292 | middle = (uchar8)(temp0.s147a, temp0.sd, temp1.s036); \ |
| 293 | right = (uchar8)(temp0.s258b, temp0.se, temp1.s147); \ |
| 294 | }) |
| 295 | #endif /* CONV_STRIDE_X */ |
| 296 | /** This function computes the depthwise convolution quantized using dot product when the data layout is NCHW. |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 297 | * |
| 298 | * @param[in] src_ptr Pointer to the source image. Supported data types: QASYMM8 |
| 299 | * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) |
| 300 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 301 | * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) |
| 302 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 303 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image |
| 304 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 305 | * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) |
| 306 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: QASYMM8 |
| 307 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 308 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 309 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 310 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 311 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 312 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 313 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 314 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: QASYMM8 |
| 315 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 316 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 317 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 318 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 319 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 320 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 321 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 322 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: QASYMM8 |
| 323 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 324 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 325 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 326 | */ |
| 327 | |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 328 | __kernel void depthwise_convolution_3x3_quantized_dot8_nchw( |
| 329 | TENSOR3D_DECLARATION(src), |
| 330 | TENSOR3D_DECLARATION(dst), |
| 331 | TENSOR3D_DECLARATION(weights) |
| 332 | #if defined(HAS_BIAS) |
| 333 | , |
| 334 | VECTOR_DECLARATION(biases) |
| 335 | #endif //defined(HAS_BIAS) |
| 336 | ) |
| 337 | { |
| 338 | Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src); |
| 339 | Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); |
| 340 | Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT(weights); |
| 341 | #if defined(HAS_BIAS) |
| 342 | Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); |
| 343 | |
| 344 | const int bias_value = *((__global int *)(vector_offset(&biases, get_global_id(2)))); |
| 345 | #endif //defined(HAS_BIAS) |
| 346 | |
| 347 | src.ptr -= (get_global_id(2) - get_global_id(2) / DEPTH_MULTIPLIER) * src_step_z; |
| 348 | |
| 349 | uchar3 w0 = vload3(0, weights.ptr + 0 * weights_stride_y); |
| 350 | uchar3 w1 = vload3(0, weights.ptr + 1 * weights_stride_y); |
| 351 | uchar3 w2 = vload3(0, weights.ptr + 2 * weights_stride_y); |
| 352 | |
| 353 | uchar8 left0, middle0, right0; |
| 354 | uchar8 left1, middle1, right1; |
| 355 | uchar8 left2, middle2, right2; |
| 356 | |
| 357 | int8 values0 = 0; |
| 358 | int8 sum0 = 0; |
| 359 | |
| 360 | GET_VALUES(src.ptr + 0 * src_stride_y, left0, middle0, right0); |
| 361 | GET_VALUES(src.ptr + 1 * src_stride_y, left1, middle1, right1); |
| 362 | GET_VALUES(src.ptr + 2 * src_stride_y, left2, middle2, right2); |
| 363 | |
| 364 | #if WEIGHTS_OFFSET != 0 |
| 365 | sum0 += convert_int8(left0) + convert_int8(middle0) + convert_int8(right0); |
| 366 | sum0 += convert_int8(left1) + convert_int8(middle1) + convert_int8(right1); |
| 367 | sum0 += convert_int8(left2) + convert_int8(middle2) + convert_int8(right2); |
| 368 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 369 | |
| 370 | #if CONV_STRIDE_Y == 1 |
| 371 | // If conv_stride_y is equals to 1, we compute two output rows |
| 372 | |
| 373 | uchar8 left3, middle3, right3; |
| 374 | int8 values1 = 0; |
| 375 | int8 sum1 = 0; |
| 376 | |
| 377 | GET_VALUES(src.ptr + 3 * src_stride_y, left3, middle3, right3); |
| 378 | |
| 379 | #if WEIGHTS_OFFSET != 0 |
| 380 | sum1 += convert_int8(left1) + convert_int8(middle1) + convert_int8(right1); |
| 381 | sum1 += convert_int8(left2) + convert_int8(middle2) + convert_int8(right2); |
| 382 | sum1 += convert_int8(left3) + convert_int8(middle3) + convert_int8(right3); |
| 383 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 384 | #endif // CONV_STRIDE_Y == 1 |
| 385 | |
| 386 | ARM_DOT(left0.s0, middle0.s0, right0.s0, left1.s0, w0.s0, w0.s1, w0.s2, w1.s0, values0.s0); |
| 387 | ARM_DOT(middle1.s0, right1.s0, left2.s0, middle2.s0, w1.s1, w1.s2, w2.s0, w2.s1, values0.s0); |
| 388 | values0.s0 += right2.s0 * w2.s2; |
| 389 | |
| 390 | ARM_DOT(left0.s1, middle0.s1, right0.s1, left1.s1, w0.s0, w0.s1, w0.s2, w1.s0, values0.s1); |
| 391 | ARM_DOT(middle1.s1, right1.s1, left2.s1, middle2.s1, w1.s1, w1.s2, w2.s0, w2.s1, values0.s1); |
| 392 | values0.s1 += right2.s1 * w2.s2; |
| 393 | |
| 394 | ARM_DOT(left0.s2, middle0.s2, right0.s2, left1.s2, w0.s0, w0.s1, w0.s2, w1.s0, values0.s2); |
| 395 | ARM_DOT(middle1.s2, right1.s2, left2.s2, middle2.s2, w1.s1, w1.s2, w2.s0, w2.s1, values0.s2); |
| 396 | values0.s2 += right2.s2 * w2.s2; |
| 397 | |
| 398 | ARM_DOT(left0.s3, middle0.s3, right0.s3, left1.s3, w0.s0, w0.s1, w0.s2, w1.s0, values0.s3); |
| 399 | ARM_DOT(middle1.s3, right1.s3, left2.s3, middle2.s3, w1.s1, w1.s2, w2.s0, w2.s1, values0.s3); |
| 400 | values0.s3 += right2.s3 * w2.s2; |
| 401 | |
| 402 | ARM_DOT(left0.s4, middle0.s4, right0.s4, left1.s4, w0.s0, w0.s1, w0.s2, w1.s0, values0.s4); |
| 403 | ARM_DOT(middle1.s4, right1.s4, left2.s4, middle2.s4, w1.s1, w1.s2, w2.s0, w2.s1, values0.s4); |
| 404 | values0.s4 += right2.s4 * w2.s2; |
| 405 | |
| 406 | ARM_DOT(left0.s5, middle0.s5, right0.s5, left1.s5, w0.s0, w0.s1, w0.s2, w1.s0, values0.s5); |
| 407 | ARM_DOT(middle1.s5, right1.s5, left2.s5, middle2.s5, w1.s1, w1.s2, w2.s0, w2.s1, values0.s5); |
| 408 | values0.s5 += right2.s5 * w2.s2; |
| 409 | |
| 410 | ARM_DOT(left0.s6, middle0.s6, right0.s6, left1.s6, w0.s0, w0.s1, w0.s2, w1.s0, values0.s6); |
| 411 | ARM_DOT(middle1.s6, right1.s6, left2.s6, middle2.s6, w1.s1, w1.s2, w2.s0, w2.s1, values0.s6); |
| 412 | values0.s6 += right2.s6 * w2.s2; |
| 413 | |
| 414 | ARM_DOT(left0.s7, middle0.s7, right0.s7, left1.s7, w0.s0, w0.s1, w0.s2, w1.s0, values0.s7); |
| 415 | ARM_DOT(middle1.s7, right1.s7, left2.s7, middle2.s7, w1.s1, w1.s2, w2.s0, w2.s1, values0.s7); |
| 416 | values0.s7 += right2.s7 * w2.s2; |
| 417 | |
| 418 | #if CONV_STRIDE_Y == 1 |
| 419 | ARM_DOT(left1.s0, middle1.s0, right1.s0, left2.s0, w0.s0, w0.s1, w0.s2, w1.s0, values1.s0); |
| 420 | ARM_DOT(middle2.s0, right2.s0, left3.s0, middle3.s0, w1.s1, w1.s2, w2.s0, w2.s1, values1.s0); |
| 421 | values1.s0 += right3.s0 * w2.s2; |
| 422 | |
| 423 | ARM_DOT(left1.s1, middle1.s1, right1.s1, left2.s1, w0.s0, w0.s1, w0.s2, w1.s0, values1.s1); |
| 424 | ARM_DOT(middle2.s1, right2.s1, left3.s1, middle3.s1, w1.s1, w1.s2, w2.s0, w2.s1, values1.s1); |
| 425 | values1.s1 += right3.s1 * w2.s2; |
| 426 | |
| 427 | ARM_DOT(left1.s2, middle1.s2, right1.s2, left2.s2, w0.s0, w0.s1, w0.s2, w1.s0, values1.s2); |
| 428 | ARM_DOT(middle2.s2, right2.s2, left3.s2, middle3.s2, w1.s1, w1.s2, w2.s0, w2.s1, values1.s2); |
| 429 | values1.s2 += right3.s2 * w2.s2; |
| 430 | |
| 431 | ARM_DOT(left1.s3, middle1.s3, right1.s3, left2.s3, w0.s0, w0.s1, w0.s2, w1.s0, values1.s3); |
| 432 | ARM_DOT(middle2.s3, right2.s3, left3.s3, middle3.s3, w1.s1, w1.s2, w2.s0, w2.s1, values1.s3); |
| 433 | values1.s3 += right3.s3 * w2.s2; |
| 434 | |
| 435 | ARM_DOT(left1.s4, middle1.s4, right1.s4, left2.s4, w0.s0, w0.s1, w0.s2, w1.s0, values1.s4); |
| 436 | ARM_DOT(middle2.s4, right2.s4, left3.s4, middle3.s4, w1.s1, w1.s2, w2.s0, w2.s1, values1.s4); |
| 437 | values1.s4 += right3.s4 * w2.s2; |
| 438 | |
| 439 | ARM_DOT(left1.s5, middle1.s5, right1.s5, left2.s5, w0.s0, w0.s1, w0.s2, w1.s0, values1.s5); |
| 440 | ARM_DOT(middle2.s5, right2.s5, left3.s5, middle3.s5, w1.s1, w1.s2, w2.s0, w2.s1, values1.s5); |
| 441 | values1.s5 += right3.s5 * w2.s2; |
| 442 | |
| 443 | ARM_DOT(left1.s6, middle1.s6, right1.s6, left2.s6, w0.s0, w0.s1, w0.s2, w1.s0, values1.s6); |
| 444 | ARM_DOT(middle2.s6, right2.s6, left3.s6, middle3.s6, w1.s1, w1.s2, w2.s0, w2.s1, values1.s6); |
| 445 | values1.s6 += right3.s6 * w2.s2; |
| 446 | |
| 447 | ARM_DOT(left1.s7, middle1.s7, right1.s7, left2.s7, w0.s0, w0.s1, w0.s2, w1.s0, values1.s7); |
| 448 | ARM_DOT(middle2.s7, right2.s7, left3.s7, middle3.s7, w1.s1, w1.s2, w2.s0, w2.s1, values1.s7); |
| 449 | values1.s7 += right3.s7 * w2.s2; |
| 450 | #endif // CONV_STRIDE_Y == 1 |
| 451 | |
| 452 | #if defined(HAS_BIAS) |
| 453 | values0 += (int8)(bias_value); |
| 454 | #if CONV_STRIDE_Y == 1 |
| 455 | values1 += (int8)(bias_value); |
| 456 | #endif /* CONV_STRIDE_Y == 1 */ |
| 457 | #endif //defined(HAS_BIAS) |
| 458 | |
| 459 | #if WEIGHTS_OFFSET != 0 |
| 460 | values0 += sum0 * (int8)(WEIGHTS_OFFSET); |
| 461 | #if CONV_STRIDE_Y == 1 |
| 462 | values1 += sum1 * (int8)(WEIGHTS_OFFSET); |
| 463 | #endif /* CONV_STRIDE_Y == 1 */ |
| 464 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 465 | |
| 466 | #if INPUT_OFFSET != 0 |
| 467 | ushort sum_weights = 0; |
| 468 | ushort3 tmp_we = convert_ushort3(w0) + convert_ushort3(w1) + convert_ushort3(w2); |
| 469 | sum_weights += tmp_we.s0 + tmp_we.s1 + tmp_we.s2; |
| 470 | values0 += sum_weights * (int8)(INPUT_OFFSET); |
| 471 | #if CONV_STRIDE_Y == 1 |
| 472 | values1 += sum_weights * (int8)(INPUT_OFFSET); |
| 473 | #endif /* CONV_STRIDE_Y == 1 */ |
| 474 | #endif /* INPUT_OFFSET != 0 */ |
| 475 | |
| 476 | #if K_OFFSET != 0 |
| 477 | values0 += (int8)(K_OFFSET); |
| 478 | #if CONV_STRIDE_Y == 1 |
| 479 | values1 += (int8)(K_OFFSET); |
| 480 | #endif /* CONV_STRIDE_Y == 1 */ |
| 481 | #endif /* K_OFFSET != 0 */ |
| 482 | |
| 483 | values0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8); |
| 484 | values0 += (int8)OUTPUT_OFFSET; |
| 485 | uchar8 res0 = convert_uchar8_sat(values0); |
| 486 | res0 = max(res0, (uchar8)0); |
| 487 | res0 = min(res0, (uchar8)255); |
| 488 | |
| 489 | vstore8(ACTIVATION_FUNC(res0), 0, dst.ptr); |
| 490 | #if CONV_STRIDE_Y == 1 |
| 491 | |
| 492 | values1 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(values1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT, 8); |
| 493 | values1 += (int8)OUTPUT_OFFSET; |
| 494 | uchar8 res1 = convert_uchar8_sat(values1); |
| 495 | res1 = max(res1, (uchar8)0); |
| 496 | res1 = min(res1, (uchar8)255); |
| 497 | |
| 498 | vstore8(ACTIVATION_FUNC(res1), 0, dst.ptr + dst_stride_y); |
| 499 | #endif /* CONV_STRIDE_Y == 1 */ |
| 500 | } |
| 501 | |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 502 | #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 503 | |
| 504 | #endif /* defined(CONV_STRIDE_Y) && defined(CONV_STRIDE_X) && defined(DEPTH_MULTIPLIER) */ |
| 505 | |
| 506 | #if defined(VEC_SIZE) && defined(SRC_DIM_1) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT) |
| 507 | |
| 508 | #define asymm_mult_by_quant_multiplier_less_than_one(x, y, z) ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, y, z, VEC_SIZE) |
| 509 | |
| 510 | #define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE) |
| 511 | #define VEC_UCHAR VEC_DATA_TYPE(uchar, VEC_SIZE) |
| 512 | #define VEC_USHORT VEC_DATA_TYPE(ushort, VEC_SIZE) |
| 513 | |
| 514 | #define MULTIPLY_ADD(x, y, acc) acc += CONVERT(CONVERT(x, VEC_USHORT) * CONVERT(y, VEC_USHORT), VEC_INT) |
| 515 | |
| 516 | #if WEIGHTS_OFFSET != 0 |
| 517 | #define MULTIPLY_ADD_ACCUMULATE(x, y, acc, sum) \ |
| 518 | ({ \ |
| 519 | sum += CONVERT(x, VEC_INT); \ |
| 520 | MULTIPLY_ADD(x, y, acc); \ |
| 521 | }) |
| 522 | #else /* WEIGHTS_OFFSET != 0 */ |
| 523 | #define MULTIPLY_ADD_ACCUMULATE(x, y, acc, sum) MULTIPLY_ADD(x, y, acc) |
| 524 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 525 | |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 526 | #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 527 | #define DOT_PRODUCT(acc, val0, val1, val2, val3, val4, val5, val6, val7, val8, w0, w1, w2, w3, w4, w5, w6, w7, w8) \ |
| 528 | ({ \ |
| 529 | ARM_DOT(val0.s0, val1.s0, val2.s0, val3.s0, w0.s0, w1.s0, w2.s0, w3.s0, acc.s0); \ |
| 530 | ARM_DOT(val4.s0, val5.s0, val6.s0, val7.s0, w4.s0, w5.s0, w6.s0, w7.s0, acc.s0); \ |
| 531 | acc.s0 += val8.s0 * w8.s0; \ |
| 532 | \ |
| 533 | ARM_DOT(val0.s1, val1.s1, val2.s1, val3.s1, w0.s1, w1.s1, w2.s1, w3.s1, acc.s1); \ |
| 534 | ARM_DOT(val4.s1, val5.s1, val6.s1, val7.s1, w4.s1, w5.s1, w6.s1, w7.s1, acc.s1); \ |
| 535 | acc.s1 += val8.s1 * w8.s1; \ |
| 536 | \ |
| 537 | ARM_DOT(val0.s2, val1.s2, val2.s2, val3.s2, w0.s2, w1.s2, w2.s2, w3.s2, acc.s2); \ |
| 538 | ARM_DOT(val4.s2, val5.s2, val6.s2, val7.s2, w4.s2, w5.s2, w6.s2, w7.s2, acc.s2); \ |
| 539 | acc.s2 += val8.s2 * w8.s2; \ |
| 540 | \ |
| 541 | ARM_DOT(val0.s3, val1.s3, val2.s3, val3.s3, w0.s3, w1.s3, w2.s3, w3.s3, acc.s3); \ |
| 542 | ARM_DOT(val4.s3, val5.s3, val6.s3, val7.s3, w4.s3, w5.s3, w6.s3, w7.s3, acc.s3); \ |
| 543 | acc.s3 += val8.s3 * w8.s3; \ |
| 544 | }) |
| 545 | |
| 546 | #if WEIGHTS_OFFSET != 0 |
| 547 | #define DOT_PRODUCT_ACCUMULATE(acc, sum, val0, val1, val2, val3, val4, val5, val6, val7, val8, w0, w1, w2, w3, w4, w5, w6, w7, w8) \ |
| 548 | ({ \ |
| 549 | sum += CONVERT(val0, VEC_INT) + CONVERT(val1, VEC_INT) + CONVERT(val2, VEC_INT) + CONVERT(val3, VEC_INT) + CONVERT(val4, VEC_INT) + CONVERT(val5, VEC_INT) + CONVERT(val6, VEC_INT) + CONVERT(val7, VEC_INT) + CONVERT(val8, VEC_INT); \ |
| 550 | DOT_PRODUCT(acc, val0, val1, val2, val3, val4, val5, val6, val7, val8, w0, w1, w2, w3, w4, w5, w6, w7, w8); \ |
| 551 | }) |
| 552 | #else /* WEIGHTS_OFFSET != 0 */ |
| 553 | #define DOT_PRODUCT_ACCUMULATE(acc, sum, val0, val1, val2, val3, val4, val5, val6, val7, val8, w0, w1, w2, w3, w4, w5, w6, w7, w8) DOT_PRODUCT(acc, val0, val1, val2, val3, val4, val5, val6, val7, val8, w0, w1, w2, w3, w4, w5, w6, w7, w8) |
| 554 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 555 | |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 556 | #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 557 | |
| 558 | #if defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) |
| 559 | /** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width or height is not 1. |
| 560 | * |
| 561 | * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2) |
| 562 | * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) |
| 563 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1) |
| 564 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1) |
| 565 | * @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X) |
| 566 | * @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1) |
| 567 | * |
| 568 | * @param[in] src_ptr Pointer to the source image. Supported data types: QASYMM8 |
| 569 | * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) |
| 570 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 571 | * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) |
| 572 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 573 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image |
| 574 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 575 | * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) |
| 576 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 577 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 578 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 579 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 580 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 581 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 582 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 583 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 584 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
| 585 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 586 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 587 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 588 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 589 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 590 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 591 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 592 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: same as @p src_ptr |
| 593 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 594 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 595 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 596 | * @param[in] max_offset Max offset for the input tensor |
| 597 | */ |
| 598 | __kernel void depthwise_convolution_3x3_quantized_nhwc( |
| 599 | TENSOR3D_DECLARATION(src), |
| 600 | TENSOR3D_DECLARATION(dst), |
| 601 | TENSOR3D_DECLARATION(weights), |
| 602 | #if defined(HAS_BIAS) |
| 603 | VECTOR_DECLARATION(biases), |
| 604 | #endif /* defined(HAS_BIAS) */ |
| 605 | int max_offset) |
| 606 | { |
| 607 | const int x = get_global_id(0); // channels |
| 608 | const int y = get_global_id(1); // spatial coordinate x |
| 609 | const int z = get_global_id(2); // spatial coordinate y |
| 610 | |
| 611 | Vector weights = CONVERT_TO_VECTOR_STRUCT(weights); |
| 612 | |
| 613 | __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE; |
| 614 | |
| 615 | int z_coord = 0; |
| 616 | int4 offset = 0; |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 617 | const int4 y_offset = ((int4)(y * CONV_STRIDE_X) + (int4)(0, 1, 2, 3) - (int)CONV_PAD_LEFT) * (int4)src_stride_y; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 618 | |
| 619 | // We compute 2x1x1 [C,W,H] elements |
| 620 | VEC_INT acc = 0, sum = 0; |
| 621 | |
| 622 | // Load weights |
| 623 | VEC_UCHAR w0 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y + 0 * weights_stride_z); |
| 624 | VEC_UCHAR w1 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y + 0 * weights_stride_z); |
| 625 | VEC_UCHAR w2 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y + 0 * weights_stride_z); |
| 626 | VEC_UCHAR w3 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y + 1 * weights_stride_z); |
| 627 | VEC_UCHAR w4 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y + 1 * weights_stride_z); |
| 628 | VEC_UCHAR w5 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y + 1 * weights_stride_z); |
| 629 | VEC_UCHAR w6 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y + 2 * weights_stride_z); |
| 630 | VEC_UCHAR w7 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y + 2 * weights_stride_z); |
| 631 | VEC_UCHAR w8 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y + 2 * weights_stride_z); |
| 632 | |
| 633 | #if INPUT_OFFSET != 0 |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 634 | VEC_INT sum_we = CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT) |
| 635 | + CONVERT(w3, VEC_INT) + CONVERT(w4, VEC_INT) + CONVERT(w5, VEC_INT) |
| 636 | + CONVERT(w6, VEC_INT) + CONVERT(w7, VEC_INT) + CONVERT(w8, VEC_INT); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 637 | #endif /* INPUT_OFFSET != 0 */ |
| 638 | |
| 639 | // Load input values |
| 640 | // z == 0 |
| 641 | // Clamp z_coord as for z = 0, it can be negative |
| 642 | // z_coord is casted to unsigned int in order to use just a min() operation |
| 643 | // A "-1" 32 bit signed variable converted to unsigned gives 4294967295 |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 644 | z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 645 | z_coord = min((uint)z_coord, (uint)SRC_DIM_2); |
| 646 | offset = y_offset + (int4)(z_coord * src_stride_z); |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 647 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 648 | |
| 649 | VEC_UCHAR values0 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0); |
| 650 | VEC_UCHAR values1 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1); |
| 651 | VEC_UCHAR values2 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2); |
| 652 | |
| 653 | // z == 1 |
| 654 | // z_coord can be only negative for z = 0 so we do not need to clamp it |
| 655 | // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 656 | z_coord = z * (int)CONV_STRIDE_Y - (int)CONV_PAD_TOP + 1; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 657 | offset = y_offset + (int4)(z_coord * src_stride_z); |
| 658 | VEC_UCHAR values3 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0); |
| 659 | VEC_UCHAR values4 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1); |
| 660 | VEC_UCHAR values5 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2); |
| 661 | |
| 662 | // z == 2 |
| 663 | // After z = 1 we can simply add src_stride_z to offset without updating z_coord |
| 664 | // However offset can be out-of-bound so we need to check if it is greater than max_offset |
| 665 | offset += (int4)src_stride_z; |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 666 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 667 | VEC_UCHAR values6 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0); |
| 668 | VEC_UCHAR values7 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1); |
| 669 | VEC_UCHAR values8 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2); |
| 670 | |
| 671 | MULTIPLY_ADD_ACCUMULATE(values0, w0, acc, sum); |
| 672 | MULTIPLY_ADD_ACCUMULATE(values1, w1, acc, sum); |
| 673 | MULTIPLY_ADD_ACCUMULATE(values2, w2, acc, sum); |
| 674 | |
| 675 | MULTIPLY_ADD_ACCUMULATE(values3, w3, acc, sum); |
| 676 | MULTIPLY_ADD_ACCUMULATE(values4, w4, acc, sum); |
| 677 | MULTIPLY_ADD_ACCUMULATE(values5, w5, acc, sum); |
| 678 | |
| 679 | MULTIPLY_ADD_ACCUMULATE(values6, w6, acc, sum); |
| 680 | MULTIPLY_ADD_ACCUMULATE(values7, w7, acc, sum); |
| 681 | MULTIPLY_ADD_ACCUMULATE(values8, w8, acc, sum); |
| 682 | |
| 683 | #if defined(HAS_BIAS) |
| 684 | Vector biases = CONVERT_TO_VECTOR_STRUCT(biases); |
| 685 | VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr); |
| 686 | acc += bias_values; |
| 687 | #endif // defined(HAS_BIAS) |
| 688 | |
| 689 | #if WEIGHTS_OFFSET != 0 |
| 690 | acc += WEIGHTS_OFFSET * sum; |
| 691 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 692 | |
| 693 | #if INPUT_OFFSET != 0 |
| 694 | acc += INPUT_OFFSET * sum_we; |
| 695 | #endif /* INPUT_OFFSET != 0 */ |
| 696 | |
| 697 | #if K_OFFSET != 0 |
| 698 | acc += (VEC_INT)K_OFFSET; |
| 699 | #endif /* K_OFFSET != 0 */ |
| 700 | |
| 701 | acc = asymm_mult_by_quant_multiplier_less_than_one(acc, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); |
| 702 | acc += (VEC_INT)OUTPUT_OFFSET; |
| 703 | |
| 704 | VEC_UCHAR res = CONVERT_SAT(acc, VEC_UCHAR); |
| 705 | res = CLAMP(res, (VEC_UCHAR)0, (VEC_UCHAR)255); |
| 706 | |
| 707 | Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst); |
| 708 | VSTORE(VEC_SIZE) |
| 709 | (res, 0, dst.ptr); |
| 710 | } |
| 711 | #endif // defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) |
| 712 | |
| 713 | #if defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED) |
| 714 | /** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width and height is 1 |
| 715 | * |
| 716 | * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2) |
| 717 | * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) |
| 718 | * @note The number of rows processed per thread must be passed at compile time using -DNUM_ROWS_PROCESSED (i.e. -DNUM_ROWS_PROCESSED=2) |
| 719 | * @note The number of planes processed per thread must be passed at compile time using -DNUM_PLANES_PROCESSED (i.e. -DNUM_PLANES_PROCESSED=2) |
| 720 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1) |
| 721 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1). |
| 722 | * |
| 723 | * @param[in] src_ptr Pointer to the source image. Supported data types: QASYMM8 |
| 724 | * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) |
| 725 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 726 | * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) |
| 727 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 728 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image |
| 729 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 730 | * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) |
| 731 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr |
| 732 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 733 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 734 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 735 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 736 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 737 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 738 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 739 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr |
| 740 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 741 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 742 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 743 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 744 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 745 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 746 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 747 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: same as @p src_ptr |
| 748 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 749 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 750 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 751 | * @param[in] max_offset Max offset for the input tensor |
| 752 | */ |
| 753 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 754 | __kernel void depthwise_convolution_3x3_quantized_nhwc_stride1( |
| 755 | TENSOR3D_DECLARATION(src), |
| 756 | TENSOR3D_DECLARATION(dst), |
| 757 | TENSOR3D_DECLARATION(weights), |
| 758 | #if defined(HAS_BIAS) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 759 | VECTOR_DECLARATION(biases), |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 760 | #endif /* defined(HAS_BIAS) */ |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 761 | int max_offset) |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 762 | { |
Giorgio Arena | fa23f11 | 2018-06-19 11:27:38 +0100 | [diff] [blame] | 763 | int x = get_global_id(0); |
| 764 | int y = get_global_id(1); |
| 765 | int z = get_global_id(2); |
| 766 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 767 | Vector weights = CONVERT_TO_VECTOR_STRUCT(weights); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 768 | |
| 769 | __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE; |
| 770 | |
| 771 | int z_coord = 0; |
| 772 | int4 offset = 0; |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 773 | int4 y_offset = ((int4)(y * NUM_ROWS_PROCESSED) + (int4)(0, 1, 2, 3) - (int)CONV_PAD_LEFT) * (int4)src_stride_y; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 774 | |
| 775 | // We compute 2x2x2 [C,W,H] elements |
| 776 | VEC_INT acc0 = 0, sum0 = 0; |
| 777 | VEC_INT acc1 = 0, sum1 = 0; |
| 778 | VEC_INT acc2 = 0, sum2 = 0; |
| 779 | VEC_INT acc3 = 0, sum3 = 0; |
| 780 | |
| 781 | // Load weights |
| 782 | VEC_UCHAR w0 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y + 0 * weights_stride_z); |
| 783 | VEC_UCHAR w1 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y + 0 * weights_stride_z); |
| 784 | VEC_UCHAR w2 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y + 0 * weights_stride_z); |
| 785 | VEC_UCHAR w3 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y + 1 * weights_stride_z); |
| 786 | VEC_UCHAR w4 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y + 1 * weights_stride_z); |
| 787 | VEC_UCHAR w5 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y + 1 * weights_stride_z); |
| 788 | VEC_UCHAR w6 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y + 2 * weights_stride_z); |
| 789 | VEC_UCHAR w7 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y + 2 * weights_stride_z); |
| 790 | VEC_UCHAR w8 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y + 2 * weights_stride_z); |
| 791 | |
| 792 | #if INPUT_OFFSET != 0 |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 793 | VEC_INT sum_we = CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT) |
| 794 | + CONVERT(w3, VEC_INT) + CONVERT(w4, VEC_INT) + CONVERT(w5, VEC_INT) |
| 795 | + CONVERT(w6, VEC_INT) + CONVERT(w7, VEC_INT) + CONVERT(w8, VEC_INT); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 796 | #endif /* INPUT_OFFSET != 0 */ |
| 797 | |
| 798 | // Load input values |
| 799 | // z == 0 |
| 800 | // Clamp z_coord as for z = 0, it can be negative |
| 801 | // z_coord is casted to unsigned int in order to use just a min() operation |
| 802 | // A "-1" 32 bit signed variable converted to unsigned gives 4294967295 |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 803 | z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 804 | z_coord = min((uint)z_coord, (uint)SRC_DIM_2); |
| 805 | offset = y_offset + (int4)(z_coord * src_stride_z); |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 806 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 807 | |
| 808 | VEC_UCHAR values0 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0); |
| 809 | VEC_UCHAR values1 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1); |
| 810 | VEC_UCHAR values2 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2); |
| 811 | VEC_UCHAR values3 = VLOAD(VEC_SIZE)(0, src_addr + offset.s3); |
| 812 | |
| 813 | // z == 1 |
| 814 | // z_coord can be only negative for z = 0 so we do not need to clamp it |
| 815 | // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 816 | z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP + 1; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 817 | offset = y_offset + (int4)(z_coord * src_stride_z); |
| 818 | VEC_UCHAR values4 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0); |
| 819 | VEC_UCHAR values5 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1); |
| 820 | VEC_UCHAR values6 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2); |
| 821 | VEC_UCHAR values7 = VLOAD(VEC_SIZE)(0, src_addr + offset.s3); |
| 822 | |
| 823 | // z == 2 |
| 824 | // After z = 1 we can simply add src_stride_z to offset without updating z_coord |
| 825 | // However offset can be out-of-bound so we need to check if it is greater than max_offset |
| 826 | offset += (int4)src_stride_z; |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 827 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 828 | VEC_UCHAR values8 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0); |
| 829 | VEC_UCHAR values9 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1); |
| 830 | VEC_UCHAR values10 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2); |
| 831 | VEC_UCHAR values11 = VLOAD(VEC_SIZE)(0, src_addr + offset.s3); |
| 832 | |
| 833 | // z == 3 |
| 834 | // After z = 1 we can simply add src_stride_z to offset without updating z_coord |
| 835 | // However offset can be out-of-bound so we need to check if it is greater than max_offset |
| 836 | offset += (int4)(src_stride_z); |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 837 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 838 | VEC_UCHAR values12 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0); |
| 839 | VEC_UCHAR values13 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1); |
| 840 | VEC_UCHAR values14 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2); |
| 841 | VEC_UCHAR values15 = VLOAD(VEC_SIZE)(0, src_addr + offset.s3); |
| 842 | |
| 843 | MULTIPLY_ADD_ACCUMULATE(values0, w0, acc0, sum0); |
| 844 | MULTIPLY_ADD_ACCUMULATE(values1, w1, acc0, sum0); |
| 845 | MULTIPLY_ADD_ACCUMULATE(values2, w2, acc0, sum0); |
| 846 | MULTIPLY_ADD_ACCUMULATE(values1, w0, acc1, sum1); |
| 847 | MULTIPLY_ADD_ACCUMULATE(values2, w1, acc1, sum1); |
| 848 | MULTIPLY_ADD_ACCUMULATE(values3, w2, acc1, sum1); |
| 849 | |
| 850 | MULTIPLY_ADD_ACCUMULATE(values4, w3, acc0, sum0); |
| 851 | MULTIPLY_ADD_ACCUMULATE(values5, w4, acc0, sum0); |
| 852 | MULTIPLY_ADD_ACCUMULATE(values6, w5, acc0, sum0); |
| 853 | MULTIPLY_ADD_ACCUMULATE(values5, w3, acc1, sum1); |
| 854 | MULTIPLY_ADD_ACCUMULATE(values6, w4, acc1, sum1); |
| 855 | MULTIPLY_ADD_ACCUMULATE(values7, w5, acc1, sum1); |
| 856 | |
| 857 | MULTIPLY_ADD_ACCUMULATE(values8, w6, acc0, sum0); |
| 858 | MULTIPLY_ADD_ACCUMULATE(values9, w7, acc0, sum0); |
| 859 | MULTIPLY_ADD_ACCUMULATE(values10, w8, acc0, sum0); |
| 860 | MULTIPLY_ADD_ACCUMULATE(values9, w6, acc1, sum1); |
| 861 | MULTIPLY_ADD_ACCUMULATE(values10, w7, acc1, sum1); |
| 862 | MULTIPLY_ADD_ACCUMULATE(values11, w8, acc1, sum1); |
| 863 | |
| 864 | MULTIPLY_ADD_ACCUMULATE(values4, w0, acc2, sum2); |
| 865 | MULTIPLY_ADD_ACCUMULATE(values5, w1, acc2, sum2); |
| 866 | MULTIPLY_ADD_ACCUMULATE(values6, w2, acc2, sum2); |
| 867 | MULTIPLY_ADD_ACCUMULATE(values5, w0, acc3, sum3); |
| 868 | MULTIPLY_ADD_ACCUMULATE(values6, w1, acc3, sum3); |
| 869 | MULTIPLY_ADD_ACCUMULATE(values7, w2, acc3, sum3); |
| 870 | |
| 871 | MULTIPLY_ADD_ACCUMULATE(values8, w3, acc2, sum2); |
| 872 | MULTIPLY_ADD_ACCUMULATE(values9, w4, acc2, sum2); |
| 873 | MULTIPLY_ADD_ACCUMULATE(values10, w5, acc2, sum2); |
| 874 | MULTIPLY_ADD_ACCUMULATE(values9, w3, acc3, sum3); |
| 875 | MULTIPLY_ADD_ACCUMULATE(values10, w4, acc3, sum3); |
| 876 | MULTIPLY_ADD_ACCUMULATE(values11, w5, acc3, sum3); |
| 877 | |
| 878 | MULTIPLY_ADD_ACCUMULATE(values12, w6, acc2, sum2); |
| 879 | MULTIPLY_ADD_ACCUMULATE(values13, w7, acc2, sum2); |
| 880 | MULTIPLY_ADD_ACCUMULATE(values14, w8, acc2, sum2); |
| 881 | MULTIPLY_ADD_ACCUMULATE(values13, w6, acc3, sum3); |
| 882 | MULTIPLY_ADD_ACCUMULATE(values14, w7, acc3, sum3); |
| 883 | MULTIPLY_ADD_ACCUMULATE(values15, w8, acc3, sum3); |
| 884 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 885 | #if defined(HAS_BIAS) |
| 886 | Vector biases = CONVERT_TO_VECTOR_STRUCT(biases); |
| 887 | |
| 888 | VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr); |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 889 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 890 | acc0 += bias_values; |
| 891 | acc1 += bias_values; |
| 892 | acc2 += bias_values; |
| 893 | acc3 += bias_values; |
| 894 | #endif /* defined(HAS_BIAS) */ |
| 895 | |
| 896 | #if WEIGHTS_OFFSET != 0 |
| 897 | acc0 += WEIGHTS_OFFSET * sum0; |
| 898 | acc1 += WEIGHTS_OFFSET * sum1; |
| 899 | acc2 += WEIGHTS_OFFSET * sum2; |
| 900 | acc3 += WEIGHTS_OFFSET * sum3; |
| 901 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 902 | |
| 903 | #if INPUT_OFFSET != 0 |
| 904 | VEC_INT offs = INPUT_OFFSET * sum_we; |
| 905 | |
| 906 | acc0 += offs; |
| 907 | acc1 += offs; |
| 908 | acc2 += offs; |
| 909 | acc3 += offs; |
| 910 | #endif /* INPUT_OFFSET != 0 */ |
| 911 | |
| 912 | #if K_OFFSET != 0 |
| 913 | acc0 += (VEC_INT)K_OFFSET; |
| 914 | acc1 += (VEC_INT)K_OFFSET; |
| 915 | acc2 += (VEC_INT)K_OFFSET; |
| 916 | acc3 += (VEC_INT)K_OFFSET; |
| 917 | #endif /* K_OFFSET != 0 */ |
| 918 | |
| 919 | acc0 = asymm_mult_by_quant_multiplier_less_than_one(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); |
| 920 | acc1 = asymm_mult_by_quant_multiplier_less_than_one(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); |
| 921 | acc2 = asymm_mult_by_quant_multiplier_less_than_one(acc2, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); |
| 922 | acc3 = asymm_mult_by_quant_multiplier_less_than_one(acc3, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); |
| 923 | |
| 924 | acc0 += (VEC_INT)OUTPUT_OFFSET; |
| 925 | acc1 += (VEC_INT)OUTPUT_OFFSET; |
| 926 | acc2 += (VEC_INT)OUTPUT_OFFSET; |
| 927 | acc3 += (VEC_INT)OUTPUT_OFFSET; |
| 928 | |
| 929 | VEC_UCHAR res0 = CONVERT_SAT(acc0, VEC_UCHAR); |
| 930 | VEC_UCHAR res1 = CONVERT_SAT(acc1, VEC_UCHAR); |
| 931 | VEC_UCHAR res2 = CONVERT_SAT(acc2, VEC_UCHAR); |
| 932 | VEC_UCHAR res3 = CONVERT_SAT(acc3, VEC_UCHAR); |
| 933 | |
| 934 | res0 = CLAMP(res0, (VEC_UCHAR)0, (VEC_UCHAR)255); |
| 935 | res1 = CLAMP(res1, (VEC_UCHAR)0, (VEC_UCHAR)255); |
| 936 | res2 = CLAMP(res2, (VEC_UCHAR)0, (VEC_UCHAR)255); |
| 937 | res3 = CLAMP(res3, (VEC_UCHAR)0, (VEC_UCHAR)255); |
| 938 | |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 939 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z; |
| 940 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 941 | VSTORE(VEC_SIZE) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 942 | (res0, 0, dst_addr + 0 * dst_stride_y); |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 943 | VSTORE(VEC_SIZE) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 944 | (res1, 0, dst_addr + 1 * dst_stride_y); |
| 945 | |
| 946 | #if((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0) |
| 947 | if((z * NUM_PLANES_PROCESSED + 1) < DST_DIM_2) |
| 948 | #endif // ((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0) |
| 949 | { |
| 950 | VSTORE(VEC_SIZE) |
| 951 | (res2, 0, dst_addr + 0 * dst_stride_y + 1 * dst_stride_z); |
| 952 | VSTORE(VEC_SIZE) |
| 953 | (res3, 0, dst_addr + 1 * dst_stride_y + 1 * dst_stride_z); |
| 954 | } |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 955 | } |
| 956 | |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 957 | #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 958 | /** This function computes the depthwise convolution quantized for NHWC data layout when the stride along the width and height is 1 using dot product |
| 959 | * |
| 960 | * @note The number of elements read per thread must be passed at compile time using -DVEC_SIZE (e.g. -DVEC_SIZE=2) |
| 961 | * @note Dimension two of the input tensor (height for NHWC data layout) must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM_2=112) |
| 962 | * @note The number of rows processed per thread must be passed at compile time using -DNUM_ROWS_PROCESSED (i.e. -DNUM_ROWS_PROCESSED=2) |
| 963 | * @note The number of planes processed per thread must be passed at compile time using -DNUM_PLANES_PROCESSED (i.e. -DNUM_PLANES_PROCESSED=2) |
| 964 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1) |
| 965 | * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1). |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 966 | * |
| 967 | * @param[in] src_ptr Pointer to the source image. Supported data types: QASYMM8 |
| 968 | * @param[in] src_stride_x Stride of the source image in X dimension (in bytes) |
| 969 | * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) |
| 970 | * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes) |
| 971 | * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) |
| 972 | * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image |
| 973 | * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 974 | * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) |
| 975 | * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: QASYMM8 |
| 976 | * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 977 | * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) |
| 978 | * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 979 | * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) |
| 980 | * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 981 | * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes) |
| 982 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 983 | * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: QASYMM8 |
| 984 | * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) |
| 985 | * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) |
| 986 | * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) |
| 987 | * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes) |
| 988 | * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) |
| 989 | * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes) |
| 990 | * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor |
| 991 | * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: QASYMM8 |
| 992 | * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes) |
| 993 | * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes) |
| 994 | * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector |
| 995 | */ |
| 996 | |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 997 | __kernel void depthwise_convolution_3x3_quantized_dot8_nhwc_stride1( |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 998 | TENSOR3D_DECLARATION(src), |
| 999 | TENSOR3D_DECLARATION(dst), |
| 1000 | TENSOR3D_DECLARATION(weights), |
| 1001 | #if defined(HAS_BIAS) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1002 | VECTOR_DECLARATION(biases), |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1003 | #endif /* defined(HAS_BIAS) */ |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1004 | int max_offset) |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1005 | { |
Giorgio Arena | fa23f11 | 2018-06-19 11:27:38 +0100 | [diff] [blame] | 1006 | int x = get_global_id(0); |
| 1007 | int y = get_global_id(1); |
| 1008 | int z = get_global_id(2); |
| 1009 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1010 | Vector weights = CONVERT_TO_VECTOR_STRUCT(weights); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1011 | |
| 1012 | __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * VEC_SIZE; |
| 1013 | |
| 1014 | int z_coord = 0; |
| 1015 | int4 offset = 0; |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 1016 | int4 y_offset = ((int4)(y * NUM_ROWS_PROCESSED) + (int4)(0, 1, 2, 3) - (int)CONV_PAD_LEFT) * (int4)src_stride_y; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1017 | |
| 1018 | // We compute 2x2x2 [C,W,H] elements |
| 1019 | VEC_INT acc0 = 0, sum0 = 0; |
| 1020 | VEC_INT acc1 = 0, sum1 = 0; |
| 1021 | VEC_INT acc2 = 0, sum2 = 0; |
| 1022 | VEC_INT acc3 = 0, sum3 = 0; |
| 1023 | |
| 1024 | // Load weights |
| 1025 | VEC_UCHAR w0 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y + 0 * weights_stride_z); |
| 1026 | VEC_UCHAR w1 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y + 0 * weights_stride_z); |
| 1027 | VEC_UCHAR w2 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y + 0 * weights_stride_z); |
| 1028 | VEC_UCHAR w3 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y + 1 * weights_stride_z); |
| 1029 | VEC_UCHAR w4 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y + 1 * weights_stride_z); |
| 1030 | VEC_UCHAR w5 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y + 1 * weights_stride_z); |
| 1031 | VEC_UCHAR w6 = VLOAD(VEC_SIZE)(0, weights.ptr + 0 * weights_stride_y + 2 * weights_stride_z); |
| 1032 | VEC_UCHAR w7 = VLOAD(VEC_SIZE)(0, weights.ptr + 1 * weights_stride_y + 2 * weights_stride_z); |
| 1033 | VEC_UCHAR w8 = VLOAD(VEC_SIZE)(0, weights.ptr + 2 * weights_stride_y + 2 * weights_stride_z); |
| 1034 | |
| 1035 | #if INPUT_OFFSET != 0 |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 1036 | VEC_INT sum_we = CONVERT(w0, VEC_INT) + CONVERT(w1, VEC_INT) + CONVERT(w2, VEC_INT) |
| 1037 | + CONVERT(w3, VEC_INT) + CONVERT(w4, VEC_INT) + CONVERT(w5, VEC_INT) |
| 1038 | + CONVERT(w6, VEC_INT) + CONVERT(w7, VEC_INT) + CONVERT(w8, VEC_INT); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1039 | #endif /* INPUT_OFFSET != 0 */ |
| 1040 | |
| 1041 | // Load input values |
| 1042 | // z == 0 |
| 1043 | // Clamp z_coord as for z = 0, it can be negative |
| 1044 | // z_coord is casted to unsigned int in order to use just a min() operation |
| 1045 | // A "-1" 32 bit signed variable converted to unsigned gives 4294967295 |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 1046 | z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1047 | z_coord = min((uint)z_coord, (uint)SRC_DIM_2); |
| 1048 | offset = y_offset + (int4)(z_coord * src_stride_z); |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 1049 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1050 | |
| 1051 | VEC_UCHAR values0 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0); |
| 1052 | VEC_UCHAR values1 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1); |
| 1053 | VEC_UCHAR values2 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2); |
| 1054 | VEC_UCHAR values3 = VLOAD(VEC_SIZE)(0, src_addr + offset.s3); |
| 1055 | |
| 1056 | // z == 1 |
| 1057 | // z_coord can be only negative for z = 0 so we do not need to clamp it |
| 1058 | // Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 1059 | z_coord = z * (int)NUM_PLANES_PROCESSED - (int)CONV_PAD_TOP + 1; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1060 | offset = y_offset + (int4)(z_coord * src_stride_z); |
| 1061 | VEC_UCHAR values4 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0); |
| 1062 | VEC_UCHAR values5 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1); |
| 1063 | VEC_UCHAR values6 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2); |
| 1064 | VEC_UCHAR values7 = VLOAD(VEC_SIZE)(0, src_addr + offset.s3); |
| 1065 | |
| 1066 | // z == 2 |
| 1067 | // After z = 1 we can simply add src_stride_z to offset without updating z_coord |
| 1068 | // However offset can be out-of-bound so we need to check if it is greater than max_offset |
| 1069 | offset += (int4)src_stride_z; |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 1070 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1071 | VEC_UCHAR values8 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0); |
| 1072 | VEC_UCHAR values9 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1); |
| 1073 | VEC_UCHAR values10 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2); |
| 1074 | VEC_UCHAR values11 = VLOAD(VEC_SIZE)(0, src_addr + offset.s3); |
| 1075 | |
| 1076 | // z == 3 |
| 1077 | // After z = 1 we can simply add src_stride_z to offset without updating z_coord |
| 1078 | // However offset can be out-of-bound so we need to check if it is greater than max_offset |
| 1079 | offset += (int4)(src_stride_z); |
Georgios Pinitas | 4e0d381 | 2018-08-01 20:16:34 +0100 | [diff] [blame] | 1080 | offset = min(offset, (int4)max_offset); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1081 | VEC_UCHAR values12 = VLOAD(VEC_SIZE)(0, src_addr + offset.s0); |
| 1082 | VEC_UCHAR values13 = VLOAD(VEC_SIZE)(0, src_addr + offset.s1); |
| 1083 | VEC_UCHAR values14 = VLOAD(VEC_SIZE)(0, src_addr + offset.s2); |
| 1084 | VEC_UCHAR values15 = VLOAD(VEC_SIZE)(0, src_addr + offset.s3); |
| 1085 | |
| 1086 | DOT_PRODUCT_ACCUMULATE(acc0, sum0, values0, values1, values2, values4, values5, values6, values8, values9, values10, w0, w1, w2, w3, w4, w5, w6, w7, w8); |
| 1087 | DOT_PRODUCT_ACCUMULATE(acc1, sum1, values1, values2, values3, values5, values6, values7, values9, values10, values11, w0, w1, w2, w3, w4, w5, w6, w7, w8); |
| 1088 | DOT_PRODUCT_ACCUMULATE(acc2, sum2, values4, values5, values6, values8, values9, values10, values12, values13, values14, w0, w1, w2, w3, w4, w5, w6, w7, w8); |
| 1089 | DOT_PRODUCT_ACCUMULATE(acc3, sum3, values5, values6, values7, values9, values10, values11, values13, values14, values15, w0, w1, w2, w3, w4, w5, w6, w7, w8); |
| 1090 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1091 | #if defined(HAS_BIAS) |
| 1092 | Vector biases = CONVERT_TO_VECTOR_STRUCT(biases); |
| 1093 | |
| 1094 | VEC_INT bias_values = VLOAD(VEC_SIZE)(0, (__global int *)biases.ptr); |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1095 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1096 | acc0 += bias_values; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1097 | acc1 += bias_values; |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1098 | acc2 += bias_values; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1099 | acc3 += bias_values; |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1100 | #endif /* defined(HAS_BIAS) */ |
| 1101 | |
| 1102 | #if WEIGHTS_OFFSET != 0 |
| 1103 | acc0 += WEIGHTS_OFFSET * sum0; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1104 | acc1 += WEIGHTS_OFFSET * sum1; |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1105 | acc2 += WEIGHTS_OFFSET * sum2; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1106 | acc3 += WEIGHTS_OFFSET * sum3; |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1107 | #endif /* WEIGHTS_OFFSET != 0 */ |
| 1108 | |
| 1109 | #if INPUT_OFFSET != 0 |
| 1110 | VEC_INT offs = INPUT_OFFSET * sum_we; |
| 1111 | |
| 1112 | acc0 += offs; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1113 | acc1 += offs; |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1114 | acc2 += offs; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1115 | acc3 += offs; |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1116 | #endif /* INPUT_OFFSET != 0 */ |
| 1117 | |
| 1118 | #if K_OFFSET != 0 |
| 1119 | acc0 += (VEC_INT)K_OFFSET; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1120 | acc1 += (VEC_INT)K_OFFSET; |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1121 | acc2 += (VEC_INT)K_OFFSET; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1122 | acc3 += (VEC_INT)K_OFFSET; |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1123 | #endif /* K_OFFSET != 0 */ |
| 1124 | |
| 1125 | acc0 = asymm_mult_by_quant_multiplier_less_than_one(acc0, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1126 | acc1 = asymm_mult_by_quant_multiplier_less_than_one(acc1, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1127 | acc2 = asymm_mult_by_quant_multiplier_less_than_one(acc2, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1128 | acc3 = asymm_mult_by_quant_multiplier_less_than_one(acc3, OUTPUT_MULTIPLIER, OUTPUT_SHIFT); |
| 1129 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1130 | acc0 += (VEC_INT)OUTPUT_OFFSET; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1131 | acc1 += (VEC_INT)OUTPUT_OFFSET; |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1132 | acc2 += (VEC_INT)OUTPUT_OFFSET; |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1133 | acc3 += (VEC_INT)OUTPUT_OFFSET; |
| 1134 | |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1135 | VEC_UCHAR res0 = CONVERT_SAT(acc0, VEC_UCHAR); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1136 | VEC_UCHAR res1 = CONVERT_SAT(acc1, VEC_UCHAR); |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1137 | VEC_UCHAR res2 = CONVERT_SAT(acc2, VEC_UCHAR); |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1138 | VEC_UCHAR res3 = CONVERT_SAT(acc3, VEC_UCHAR); |
| 1139 | |
| 1140 | res0 = CLAMP(res0, (VEC_UCHAR)0, (VEC_UCHAR)255); |
| 1141 | res1 = CLAMP(res1, (VEC_UCHAR)0, (VEC_UCHAR)255); |
| 1142 | res2 = CLAMP(res2, (VEC_UCHAR)0, (VEC_UCHAR)255); |
| 1143 | res3 = CLAMP(res3, (VEC_UCHAR)0, (VEC_UCHAR)255); |
| 1144 | |
| 1145 | __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z; |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1146 | |
| 1147 | VSTORE(VEC_SIZE) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1148 | (res0, 0, dst_addr + 0 * dst_stride_y); |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1149 | VSTORE(VEC_SIZE) |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1150 | (res1, 0, dst_addr + 1 * dst_stride_y); |
| 1151 | |
| 1152 | #if((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0) |
| 1153 | if((z * NUM_PLANES_PROCESSED + 1) < DST_DIM_2) |
| 1154 | #endif // ((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0) |
| 1155 | { |
| 1156 | VSTORE(VEC_SIZE) |
| 1157 | (res2, 0, dst_addr + 0 * dst_stride_y + 1 * dst_stride_z); |
| 1158 | VSTORE(VEC_SIZE) |
| 1159 | (res3, 0, dst_addr + 1 * dst_stride_y + 1 * dst_stride_z); |
| 1160 | } |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1161 | } |
Georgios Pinitas | daa3855 | 2018-08-28 17:43:18 +0100 | [diff] [blame] | 1162 | #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1163 | |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1164 | #endif // defined(NUM_ROWS_PROCESSED) && defined(NUM_PLANES_PROCESSED) |
Giorgio Arena | dfca60b | 2018-01-31 10:30:59 +0000 | [diff] [blame] | 1165 | |
Giorgio Arena | eff8d95 | 2018-07-02 15:29:57 +0100 | [diff] [blame] | 1166 | #endif // defined(VEC_SIZE) && defined(SRC_DIM_1) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT) |
| 1167 | |
| 1168 | #endif // defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) |