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
24 #include "helpers_asymm.h"
28 #if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
33 #define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr)
35 #define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr)
36 #else /* STRIDE_X not equals 1 or 2 */
37 #error "STRIDE_X larger than 2 is not supported"
40 #define CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr) \
42 int4 weights_values0 = convert_int4(vload4(0, weights_row_ptr)); \
43 int weights_value1 = convert_int(*(weights_row_ptr + 4)); \
44 int8 src0 = convert_int8(vload8(0, src_row_ptr)); \
45 int4 src1 = convert_int4(vload4(0, src_row_ptr + 8)); \
46 acc += (src0 + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
47 acc += ((int8)(src0.s1234, src0.s567, src1.s0) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
48 acc += ((int8)(src0.s234, src0.s567, src1.s01) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
49 acc += ((int8)(src0.s345, src0.s67, src1.s012) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \
50 acc += ((int8)(src0.s45, src0.s67, src1.s0123) + input_offset) * ((int8)weights_value1 + weight_offset); \
53 #define CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr) \
55 int4 weights_values0 = convert_int4(vload4(0, weights_row_ptr)); \
56 int weights_value1 = convert_int(*(weights_row_ptr + 4)); \
57 int16 src0 = convert_int16(vload16(0, src_row_ptr)); \
58 int4 src1 = convert_int4(vload4(0, src_row_ptr + 16)); \
59 acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
60 acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
61 acc += ((int8)(src0.s2468, src0.sACE, src1.s0) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
62 acc += ((int8)(src0.s3579, src0.sBDF, src1.s1) + input_offset) * ((int8)weights_values0.s3 + weight_offset); \
63 acc += ((int8)(src0.s468a, src0.sCE, src1.s02) + input_offset) * ((int8)weights_value1 + weight_offset); \
66 #elif KERNEL_SIZE == 3
69 #define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr)
71 #define CONVOLUTION1x3(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr)
72 #else /* STRIDE_X not equals 1 or 2 */
73 #error "STRIDE_X larger than 2 is not supported"
76 #define CONVOLUTION1x3_STRIDE1(acc, src_row_ptr, weights_row_ptr) \
78 int3 weights_values0 = convert_int3(vload3(0, weights_row_ptr)); \
79 int8 src0 = convert_int8(vload8(0, src_row_ptr)); \
80 int2 src1 = convert_int2(vload2(0, src_row_ptr + 8)); \
81 acc += (src0 + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
82 acc += ((int8)(src0.s1234, src0.s567, src1.s0) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
83 acc += ((int8)(src0.s234, src0.s567, src1.s01) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
86 #define CONVOLUTION1x3_STRIDE2(acc, src_row_ptr, weights_row_ptr) \
88 int3 weights_values0 = convert_int3(vload3(0, weights_row_ptr)); \
89 int16 src0 = convert_int16(vload16(0, src_row_ptr)); \
90 int src1 = convert_int(*(src_row_ptr + 16)); \
91 acc += (src0.even + input_offset) * ((int8)weights_values0.s0 + weight_offset); \
92 acc += ((int8)(src0.s1357, src0.s9BDF) + input_offset) * ((int8)weights_values0.s1 + weight_offset); \
93 acc += ((int8)(src0.s2468, src0.sACE, src1) + input_offset) * ((int8)weights_values0.s2 + weight_offset); \
96 #elif KERNEL_SIZE == 1
99 #define INPUT_PIXEL extract_input_stride3
101 #define INPUT_PIXEL extract_input_stride2
103 #define INPUT_PIXEL extract_input_stride1
105 #else /* STRIDE_X not equals 1, 2 or 3 */
106 #error "Only support strides 1, 2 and 3"
107 #endif /* STRIDE_X */
109 /** Extracts a 1D horizontal vector from the input tensor with stride as 1.
111 * @param[in] input_pixel Pointer to the first pixel.
113 * @return extracted input pixels.
115 inline uchar8 extract_input_stride1(__global const uchar *input_pixel)
117 return vload8(0, input_pixel);
120 /** Extracts a 1D horizontal vector from the input tensor with stride as 2.
122 * @param[in] input_pixel Pointer to the first pixel.
124 * @return extracted input pixels.
126 inline uchar8 extract_input_stride2(__global const uchar *input_pixel)
128 uchar16 temp = vload16(0, input_pixel);
129 return temp.s02468ace;
132 /** Extracts a 1D horizontal vector from the input tensor with stride as 3 and 8-bit data size.
134 * @param[in] input_pixel Pointer to the first pixel.
136 * @return extracted input pixels.
138 inline uchar8 extract_input_stride3(__global const uchar *input_pixel)
140 uchar16 temp1 = vload16(0, input_pixel);
141 uchar16 temp2 = vload16(0, input_pixel + 12);
142 return (uchar8)(temp1.s0369, temp2.s0369);
145 #else /* KERNEL_SIZE not equals 1, 3 or 5 */
146 #error "Only kernel sizes 1, 3 and 5 are supported"
147 #endif /* KERNEL_SIZE */
149 /** This kernel performs a direct convolution to convolve the low three dimensions.
151 * @note The convolution stride x must be passed at compile time using -DSTRIDE_X e.g. -DSTRIDE_X=1
152 * @note The third dimensions of the weights tensors must be passed at compile time using -DWEIGHTS_DEPTH
153 * @note If biases are used then -DHAS_BIAS has to be passed at compile time
155 * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8
156 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
157 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
158 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
159 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
160 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
161 * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
162 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
163 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
164 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
165 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
166 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
167 * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes)
168 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
169 * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
170 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
171 * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p weights_ptr
172 * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
173 * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
174 * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
175 * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes)
176 * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
177 * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes)
178 * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
179 * @param[in] biases_ptr Pointer to the biases tensor. Supported data types: S32
180 * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes)
181 * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes)
182 * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor
183 * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension
184 * @param[in] input_offset Input offset quantization parameter
185 * @param[in] weight_offset Weights offset quantization parameter
186 * @param[in] output_offset Output offset quantization parameter
187 * @param[in] output_multiplier Output integer multiplier quantization parameter
188 * @param[in] output_shift Output integer shift quantization parameter
190 __kernel void direct_convolution_1x1_3x3_5x5_quantized(
191 TENSOR3D_DECLARATION(src),
192 TENSOR3D_DECLARATION(dst),
193 TENSOR3D_DECLARATION(weights),
195 VECTOR_DECLARATION(biases),
196 #endif /* defined(HAS_BIAS) */
197 unsigned int weights_stride_w,
201 int output_multiplier,
204 Image src = CONVERT_TO_IMAGE_STRUCT(src);
205 Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
206 Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
210 __global uchar *weights_addr = (__global uchar *)tensor3D_offset(&weights, 0, 0, 0);
211 __global uchar *src_addr = (__global uchar *)offset(&src, 0, 0);
213 const int kernel_index = get_global_id(2);
214 weights_addr += kernel_index * weights_stride_w;
216 for(volatile int d = 0; d < WEIGHTS_DEPTH; ++d)
219 CONVOLUTION1x5(pixels0, (__global uchar *)src_addr, (__global uchar *)weights_addr);
220 CONVOLUTION1x5(pixels0, (__global uchar *)(src_addr + 1 * src_stride_y), (__global uchar *)(weights_addr + 1 * weights_stride_y));
221 CONVOLUTION1x5(pixels0, (__global uchar *)(src_addr + 2 * src_stride_y), (__global uchar *)(weights_addr + 2 * weights_stride_y));
222 CONVOLUTION1x5(pixels0, (__global uchar *)(src_addr + 3 * src_stride_y), (__global uchar *)(weights_addr + 3 * weights_stride_y));
223 CONVOLUTION1x5(pixels0, (__global uchar *)(src_addr + 4 * src_stride_y), (__global uchar *)(weights_addr + 4 * weights_stride_y));
224 #elif KERNEL_SIZE == 3
225 CONVOLUTION1x3(pixels0, (__global uchar *)(src_addr + 0 * src_stride_y), (__global uchar *)(weights_addr + 0 * weights_stride_y));
226 CONVOLUTION1x3(pixels0, (__global uchar *)(src_addr + 1 * src_stride_y), (__global uchar *)(weights_addr + 1 * weights_stride_y));
227 CONVOLUTION1x3(pixels0, (__global uchar *)(src_addr + 2 * src_stride_y), (__global uchar *)(weights_addr + 2 * weights_stride_y));
228 #elif KERNEL_SIZE == 1
229 int weight = convert_int(*(__global uchar *)weights_addr);
230 int8 input_pixel = convert_int8(INPUT_PIXEL((__global uchar *)src_addr));
231 pixels0 += (input_pixel + input_offset) * ((int8)weight + weight_offset);
232 #endif /* (KERNEL_SIZE == 1) || (KERNEL_SIZE == 3) || (KERNEL_SIZE == 5) */
234 src_addr += src_stride_z;
235 weights_addr += weights_stride_z;
239 Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
240 __global int *bias_addr = ((__global int *)(vector_offset(&biases, kernel_index)));
241 pixels0 += (int8)(*bias_addr);
242 #endif /* defined(HAS_BIAS) */
244 pixels0 = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(pixels0, output_multiplier, output_shift, 8);
245 pixels0 = pixels0 + output_offset;
247 vstore8(convert_uchar8_sat(pixels0), 0, (__global uchar *)dst.ptr);
249 #endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
251 /** This function computes the output stage of a depthwise convolution.
253 * @param[in] src_ptr Pointer to the source image. Supported data types: QASYMM8
254 * @param[in] src_stride_x Stride of the source image in X dimension (in bytes)
255 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
256 * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes)
257 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
258 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image
259 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
260 * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes)
261 * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: QASYMM8
262 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
263 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
264 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
265 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
266 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
267 * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes)
268 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
269 * @param[in] bias_ptr (Optional) Pointer to the biases vector. Supported data types: S32
270 * @param[in] bias_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
271 * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
272 * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
273 * @param[in] output_offset Quantized offset of zero point of the output tensor data range
274 * @param[in] output_multiplier Output scale multiplier
275 * @param[in] output_shift Output scale divisor exponent
278 __kernel void output_stage_quantized(
279 TENSOR3D_DECLARATION(src),
280 TENSOR3D_DECLARATION(dst),
281 #if defined(HAS_BIAS)
282 VECTOR_DECLARATION(bias),
283 #endif //defined(HAS_BIAS)
285 int output_multiplier,
288 Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
289 Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
290 #if defined(HAS_BIAS)
291 Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
292 #endif //defined(HAS_BIAS)
295 int16 vals = vload16(0, (__global int *)(src.ptr));
297 #if defined(HAS_BIAS)
299 int bias_value = *((__global int *)(vector_offset(&bias, get_global_id(2))));
300 vals += (int16)(bias_value);
301 #endif //defined(HAS_BIAS)
303 vals = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(vals, output_multiplier, output_shift, 16);
304 vals = vals + output_offset;
306 // Store result in dst
307 vstore16(convert_uchar16_sat(vals), 0, (__global uchar *)dst.ptr);