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 layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
26 #include "helpers_cs.h"
28 #if defined(DATA_TYPE_FP16)
29 precision mediump float;
30 #endif // DATA_TYPE_FP16
32 /** This kernel performs a depthwise convolution.
34 * @note The data type must be passed at compile time using "#define DATA_TYPE_NAME". e.g. "#define DATA_TYPE_FP32"
35 * @note This kernel has multiple optimized depthwise convolution options for FP16.
36 * The depthwise convolution option must be passed at compile time using "#define PROCESS_nX_nY_nZ" e.g. "#define PROCESS_8X_1Y_1Z"
37 * @note The convolution stride x must be passed at compile time using "#define STRIDE_X n" e.g. "#define STRIDE_X 1"
38 * @note In case biases will be added to the convolution "#define HAS_BIAS" has to be passed to append the final matrix with 1 in each row.
40 * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16
41 * @param[in] src_attrs The attributes of the source tensor
42 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
43 * @param[in] dst_attrs The attributes of the destination tensor
44 * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr
45 * @param[in] weights_attrs The attributes of the weights tensor
46 * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr
47 * @param[in] biases_attrs The attributes of the weights tensor
49 SHADER_PARAMS_DECLARATION
51 Tensor3DAttributes src_attrs;
52 Tensor3DAttributes dst_attrs;
53 Tensor3DAttributes weights_attrs;
55 VectorAttributes biases_attrs;
59 #if defined(DATA_TYPE_FP16)
60 #if defined(PROCESS_4X_3Y_1Z)
61 TENSOR_DECLARATION(1, srcBuffer, uvec2, src_ptr, src_shift, 3, readonly);
62 TENSOR_DECLARATION(2, dstBuffer, uvec2, dst_ptr, dst_shift, 3, writeonly);
63 TENSOR_DECLARATION(3, weightsBuffer, uvec2, weights_ptr, weights_shift, 3, readonly);
65 TENSOR_DECLARATION(4, biasesBuffer, uint, biases_ptr, biases_shift, 2, readonly);
68 #define LOAD_UNPACK_SWIZZLE(offset) load_unpack_swizzle_stride1(offset)
70 vec4 convolve1x3(vec4 s[3], vec4 w)
74 r = s[0] * w[0] + s[1] * w[1] + s[2] * w[2];
79 vec4[3] load_unpack_swizzle_stride1(uint offset)
82 s = VLOAD2_UNPACK8_HALF(src_ptr, offset);
86 r[1] = vec4(s[0].yzw, s[1].x);
87 r[2] = vec4(s[0].zw, s[1].xy);
94 Tensor3DIterator src_iter = CONVERT_TO_TENSOR3D_ITERATOR(src_attrs, src_shift);
95 Tensor3DIterator weights_iter = CONVERT_TO_TENSOR3D_ITERATOR_NO_STEP(weights_attrs, weights_shift);
96 Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift);
99 VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR_NO_STEP(biases_attrs, biases_shift);
103 for(int i = 0; i < 3; i++)
108 uint z_index = gl_GlobalInvocationID.z;
109 TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_attrs.stride_z);
111 src_iter.current_offset_in_bytes -= int((z_index - z_index / uint(DEPTH_MULTIPLIER)) * src_attrs.step_z);
114 w[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(weights_ptr, weights_iter);
115 w[1] = LOAD_UNPACK4_HALF(weights_ptr, TENSOR3D_OFFSET(weights_iter, 0, 1, 0));
116 w[2] = LOAD_UNPACK4_HALF(weights_ptr, TENSOR3D_OFFSET(weights_iter, 0, 2, 0));
121 s = LOAD_UNPACK_SWIZZLE(CURRENT_ITEM_OFFSET(src_iter));
123 r = convolve1x3(s, w[0]);
127 s = LOAD_UNPACK_SWIZZLE(TENSOR3D_OFFSET(src_iter, 0, 1, 0));
129 r = convolve1x3(s, w[1]);
131 r = convolve1x3(s, w[0]);
135 s = LOAD_UNPACK_SWIZZLE(TENSOR3D_OFFSET(src_iter, 0, 2, 0));
137 r = convolve1x3(s, w[2]);
139 r = convolve1x3(s, w[1]);
141 r = convolve1x3(s, w[0]);
145 s = LOAD_UNPACK_SWIZZLE(TENSOR3D_OFFSET(src_iter, 0, 3, 0));
147 r = convolve1x3(s, w[2]);
149 r = convolve1x3(s, w[1]);
153 s = LOAD_UNPACK_SWIZZLE(TENSOR3D_OFFSET(src_iter, 0, 4, 0));
155 r = convolve1x3(s, w[2]);
162 vec2_b = LOAD_UNPACK2_HALF(biases_ptr, VECTOR_OFFSET(biases_iter, z_index));
164 if(z_index % uint(2) == uint(0))
173 for(int i = 0; i < 3; i++)
175 pixels[i] += vec4(b);
179 STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, pixels[0]);
180 STORE_PACK4_HALF(dst_ptr, TENSOR3D_OFFSET(dst_iter, 0, 1, 0), pixels[1]);
181 STORE_PACK4_HALF(dst_ptr, TENSOR3D_OFFSET(dst_iter, 0, 2, 0), pixels[2]);
183 #elif defined(PROCESS_4X_1Y_1Z)
184 TENSOR_DECLARATION(1, srcBuffer, uvec2, src_ptr, src_shift, 3, readonly);
185 TENSOR_DECLARATION(2, dstBuffer, uvec2, dst_ptr, dst_shift, 3, writeonly);
186 TENSOR_DECLARATION(3, weightsBuffer, uvec2, weights_ptr, weights_shift, 3, readonly);
188 TENSOR_DECLARATION(4, biasesBuffer, uint, biases_ptr, biases_shift, 2, readonly);
192 #define LOAD_UNPACK_SWIZZLE(offset) load_unpack_swizzle_stride3(offset)
194 #define LOAD_UNPACK_SWIZZLE(offset) load_unpack_swizzle_stride2(offset)
195 #elif STRIDE_X == 1 /* STRIDE_X == 1 */
196 #define LOAD_UNPACK_SWIZZLE(offset) load_unpack_swizzle_stride1(offset)
197 #else /* STRIDE_X not equals 1 or 2 */
198 #error STRIDE_X larger than 2 is not supported
199 #endif /* STRIDE_X == 2 */
201 vec4 convolve1x3(vec4 s[3], vec4 w)
205 r = s[0] * w[0] + s[1] * w[1] + s[2] * w[2];
210 vec4[3] load_unpack_swizzle_stride1(uint offset)
213 s = VLOAD2_UNPACK8_HALF(src_ptr, offset);
217 r[1] = vec4(s[0].yzw, s[1].x);
218 r[2] = vec4(s[0].zw, s[1].xy);
223 vec4[3] load_unpack_swizzle_stride2(uint offset)
226 s[0] = LOAD_UNPACK4_HALF(src_ptr, offset);
227 s[1] = LOAD_UNPACK4_HALF(src_ptr, offset + uint(1));
228 s[2] = LOAD_UNPACK4_HALF(src_ptr, offset + uint(2));
231 r[0] = vec4(s[0].xz, s[1].xz);
232 r[1] = vec4(s[0].yw, s[1].yw);
233 r[2] = vec4(s[0].z, s[1].xz, s[2].x);
238 vec4[3] load_unpack_swizzle_stride3(uint offset)
241 s[0] = LOAD_UNPACK4_HALF(src_ptr, offset);
242 s[1] = LOAD_UNPACK4_HALF(src_ptr, offset + uint(1));
243 s[2] = LOAD_UNPACK4_HALF(src_ptr, offset + uint(2));
246 r[0] = vec4(s[0].xw, s[1].z, s[2].y);
247 r[1] = vec4(s[0].y, s[1].xw, s[2].z);
248 r[2] = vec4(s[0].z, s[1].y, s[2].xw);
255 Tensor3DIterator src_iter = CONVERT_TO_TENSOR3D_ITERATOR(src_attrs, src_shift);
256 Tensor3DIterator weights_iter = CONVERT_TO_TENSOR3D_ITERATOR_NO_STEP(weights_attrs, weights_shift);
257 Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift);
260 VectorIterator biases_iter = CONVERT_TO_VECTOR_ITERATOR_NO_STEP(biases_attrs, biases_shift);
263 vec4 pixels = vec4(0.f);
265 uint z_index = gl_GlobalInvocationID.z;
266 TENSOR_ITERATOR_ADVANCE_IN_BYTES(weights_iter, z_index * weights_attrs.stride_z);
268 src_iter.current_offset_in_bytes -= int((z_index - z_index / uint(DEPTH_MULTIPLIER)) * src_attrs.step_z);
271 w[0] = LOAD_UNPACK4_CURRENT_ITEM_HALF(weights_ptr, weights_iter);
272 w[1] = LOAD_UNPACK4_HALF(weights_ptr, TENSOR3D_OFFSET(weights_iter, 0, 1, 0));
273 w[2] = LOAD_UNPACK4_HALF(weights_ptr, TENSOR3D_OFFSET(weights_iter, 0, 2, 0));
278 s = LOAD_UNPACK_SWIZZLE(CURRENT_ITEM_OFFSET(src_iter));
280 r = convolve1x3(s, w[0]);
284 s = LOAD_UNPACK_SWIZZLE(TENSOR3D_OFFSET(src_iter, 0, 1, 0));
286 r = convolve1x3(s, w[1]);
290 s = LOAD_UNPACK_SWIZZLE(TENSOR3D_OFFSET(src_iter, 0, 2, 0));
292 r = convolve1x3(s, w[2]);
299 vec2_b = LOAD_UNPACK2_HALF(biases_ptr, VECTOR_OFFSET(biases_iter, z_index));
301 if(z_index % uint(2) == uint(0))
313 STORE_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, pixels);
315 #endif /* PROCESS_4X_3Y_1Z */
316 #endif /* DATA_TYPE_FP16 */