xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/cl_kernels/nhwc/direct_convolution.cl (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1/*
2 * Copyright (c) 2021-2023 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24
25#include "activation_float_helpers.h"
26#include "helpers.h"
27#include "helpers_asymm.h"
28#include "tile_helpers.h"
29
30//! @cond Doxygen_Suppress
31/** OpenCL kernel to compute the direct convolution.
32 *
33 * @note Data layout supported: NHWC
34 * @note Data type supported: F32/F16/QASYMM8/QASYMM8_SIGNED
35 * @note The accumulation data type must be passed at compile time using -DACC_DATA_TYPE (e.g. -DDATA_TYPE_PROMOTED=half)
36 * @note The convolution padding (left and top) must be passed at compile time using -DPAD_LEFT and -DPAD_TOP (e.g. -DPAD_LEFT=2, -DPAD_TOP=2)
37 * @note The convolution strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y (e.g. -DSTRIDE_X=2, -DSTRIDE_Y=2)
38 * @note The spatial dimensions of the weights must be passed at compile time using -DWEI_WIDTH and -DWEI_HEIGHT (e.g. -DWEI_WIDTH=9, -DWEI_HEIGHT=9)
39 * @note The spatial dimensions of the source tensor must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT (e.g. -DSRC_WIDTH=96, -DSRC_HEIGHT=64)
40 * @note The spatial dimensions of the destination tensor must be passed at compile time using -DDST_WIDTH and -DDST_HEIGHT (e.g. -DDST_WIDTH=96, -DDST_HEIGHT=64)
41 * @note The channels of the source tensor must be passed at compile time using -DSRC_CHANNELS (e.g. -DSRC_CHANNELS=64)
42 * @note The channels of the destination tensor must be passed at compile time using -DDST_CHANNELS (e.g. -DDDST_CHANNELS=64)
43 * @note The tensor type ("BUFFER" or "IMAGE") of the source tensor must be passed at compile time using -DSRC_TENSOR_TYPE (e.g. -DSRC_TENSOR_TYPE=BUFFER)
44 * @note The tensor type ("BUFFER" or "IMAGE") of the weights tensor must be passed at compile time using -DWEI_TENSOR_TYPE (e.g. -DWEI_TENSOR_TYPE=BUFFER)
45 * @note The tensor type ("BUFFER" or "IMAGE") of the destination tensor must be passed at compile time using -DDST_TENSOR_TYPE (e.g. -DDST_TENSOR_TYPE=BUFFER)
46 * @note The data type of the source tensor must be passed at compile time using -DSRC_DATA_TYPE (e.g. -DSRC_DATA_TYPE=float)
47 * @note The data type of the weights tensor must be passed at compile time using -DWEI_DATA_TYPE (e.g. -DWEI_DATA_TYPE=float)
48 * @note The data type of the destination tensor must be passed at compile time using -DDST_DATA_TYPE (e.g. -DDST_DATA_TYPE=float)
49 * @note The data type of the accumulators must be passed at compile time using -DACC_DATA_TYPE (e.g. -DACC_DATA_TYPE=float)
50 * @note The number of M0 rows (width*height) to process must be passed at compile time using -DM0 (e.g. -DM0=2)
51 * @note The number of N0 output channels to process must be passed at compile time using -DN0 (e.g. -DN0=2)
52 * @note The number of K0 inner accumulations must be passed at compile time using -DK0 (e.g. -DK0=2)
53 * @note The size of the partial store block in x must be passed at compile time using -DPARTIAL_N0 (e.g. -DPARTIAL_N0=1)
54 * @note The zero value must be passed at compile time using -DZERO_VALUE (e.g. -DZERO_VALUE=0)
55 * @note Only the following configurations of M0, N0 and K0 are currently supported:
56 *  - M0 = 1, 2, 3, 4, 5, 6, 7, and 8
57 *  - N0 = 2, 3, 4, 8, 16
58 *  - K0 = 2, 3, 4, 8, 16 (only 4, 8 and 16 if WEI_TENSOR_TYPE=IMAGE)
59 *
60 *@note In case of QASYMM8/QASYMM8_SIGNED, the following extra information must be passed at compile time:
61 * - -DIS_QUANTIZED
62 * - The destination quantization multiplier e.g. -DDST_MULTIPLIER=1234
63 * - The destination quantization shift e.g. -DDST_SHIFT=4
64 * - The destination offset e.g. -DDST_OFFSET=4
65 * - The source offset e.g. -DSRC_OFFSET=4
66 * - The weights offset e.g. -DWEI_OFFSET=4
67 * - The quantized zero value e.g. -DZERO_VALUE=4
68 *
69 * @param[in]  src_img                           (Not supported) Read only cl_image object for the source tensor. Included when SRC_TENSOR_TYPE=IMAGE
70 * @param[in]  src_ptr                           Pointer to the source tensor. Supported data type: F16/F32/QASYMM8
71 * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
72 * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
73 * @param[in]  src_stride_w                      Stride of the source tensor in W dimension (in bytes)
74 * @param[in]  src_c                             The size of the channels dimension of the source tensor
75 * @param[in]  src_w                             The size of the width dimension of the source tensor
76 * @param[in]  src_h                             The size of the height dimension of the source tensor
77 * @param[in]  src_n                             The size of the batches dimension of the source tensor
78 * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
79 * @param[out] dst_img                           (Not supported) Write only cl_image object for the destination tensor. Included when DST_TENSOR_TYPE=IMAGE
80 * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data type: same as @p src_ptr
81 * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
82 * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
83 * @param[in]  dst_stride_w                      Stride of the destination tensor in W dimension (in bytes)
84 * @param[in]  dst_c                             The size of the channels dimension of the destination tensor
85 * @param[in]  dst_w                             The size of the width dimension of the destination tensor
86 * @param[in]  dst_h                             The size of the height dimension of the destination tensor
87 * @param[in]  dst_n                             The size of the batches dimension of the destination tensor
88 * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
89 * @param[in]  wei_img                           (Optional) Read only cl_image object for the weights tensor. Included when WEI_TENSOR_TYPE=IMAGE
90 * @param[in]  wei_ptr                           Pointer to the weights tensor. Supported data type: same as @p src_ptr
91 * @param[in]  wei_stride_y                      Stride of the weights tensor in Y dimension (in bytes)
92 * @param[in]  wei_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
93 * @param[in]  wei_stride_w                      Stride of the weights tensor in W dimension (in bytes)
94 * @param[in]  wei_c                             The size of the channels dimension of the weights tensor
95 * @param[in]  wei_w                             The size of the width dimension of the weights tensor
96 * @param[in]  wei_h                             The size of the height dimension of the weights tensor
97 * @param[in]  wei_n                             The size of the batches dimension of the weights tensor
98 * @param[in]  wei_offset_first_element_in_bytes The offset of the first element in the weights matrix
99 * @param[in]  bia_ptr                           (Optional) Pointer to the bias tensor Supported data type: same as @p src_ptr (if F32/F16) or S32 (if QASYMM8/QASYMM8_SIGNED)
100 * @param[in]  bia_stride_x                      (Optional) Stride of the bias tensor in X dimension (in bytes)
101 * @param[in]  bia_step_x                        (Optional) bia_stride_x * number of elements along X processed per workitem(in bytes)
102 * @param[in]  bia_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
103 */
104//! @endcond
105__kernel void direct_convolution_nhwc(
106    TENSOR4D_RO_T(src, SRC_TENSOR_TYPE),
107    TENSOR4D_WO_T(dst, DST_TENSOR_TYPE),
108    TENSOR4D_RO_T(wei, WEI_TENSOR_TYPE)
109#if defined(HAS_BIAS)
110    ,
111    VECTOR_DECLARATION(bia)
112#endif // defined(HAS_BIAS)
113)
114{
115    // All the tensor dimensions are passed at compile time.
116    // In case of dynamic tensor support, the following dimensions should be passed as function argument.
117#define _IWEI_WIDTH WEI_WIDTH
118#define _IWEI_HEIGHT WEI_HEIGHT
119#define _ISRC_WIDTH SRC_WIDTH
120#define _ISRC_HEIGHT SRC_HEIGHT
121#define _ISRC_CHANNELS SRC_CHANNELS
122#define _IDST_WIDTH DST_WIDTH
123#define _IDST_HEIGHT DST_HEIGHT
124#define _IDST_CHANNELS DST_CHANNELS
125#define _IY_MULTIPLIER (_IWEI_WIDTH * _IWEI_HEIGHT)
126
127    // If quantized, the output tile has to be quantized first before being stored to global memory
128#if defined(IS_QUANTIZED)
129#define _IOUTPUT_TILE cq
130#else // defined(IS_QUANTIZED)
131#define _IOUTPUT_TILE c
132#endif // defined(IS_QUANTIZED)
133
134    const int cout = GET_SPATIAL_IDX(0, N0, PARTIAL_N0); // OFM
135    const int mout = GET_SPATIAL_IDX(1, M0, 0);          // WIDTH x HEIGHT
136    const int bout = GET_SPATIAL_IDX(2, 1, 0);           // BATCH SIZE IDX
137
138    // .v    = access the whole vector (OpenCL vector)
139    // .s[x] = access the vector element at position x (scalar access)
140    TILE(int, 1, M0, xi);
141    TILE(int, 1, M0, yi);
142
143    // Convert the linear index to coordinate
144    LOOP_UNROLLING(int, i, 0, 1, M0,
145    {
146        xi[0].s[i] = ((mout + i) % _IDST_WIDTH) * STRIDE_X;
147        yi[0].s[i] = ((mout + i) / _IDST_WIDTH) * STRIDE_Y;
148        xi[0].s[i] -= PAD_LEFT;
149        yi[0].s[i] -= PAD_TOP;
150    })
151
152    // Initialize the accumulators
153    TILE(ACC_DATA_TYPE, M0, N0, c);
154
155    LOOP_UNROLLING(int, i, 0, 1, M0,
156    {
157        c[i].v = 0;
158    })
159
160    for(int i = 0; i < (_IWEI_WIDTH * _IWEI_HEIGHT); ++i)
161    {
162        int xk = i % _IWEI_WIDTH;
163        int yk = i / _IWEI_WIDTH;
164
165        TILE(int, 1, M0, my);
166
167        LOOP_UNROLLING(int, i, 0, 1, M0,
168        {
169            int x_s    = xi[0].s[i] + xk;
170            int y_s    = yi[0].s[i] + yk;
171            my[0].s[i] = x_s + y_s *_ISRC_WIDTH;
172            my[0].s[i] = my[0].s[i] + bout * (int)(_ISRC_WIDTH * _ISRC_HEIGHT);
173            my[0].s[i] = select(-1, my[0].s[i], x_s >= 0);
174            my[0].s[i] = select(-1, my[0].s[i], x_s < _ISRC_WIDTH);
175            my[0].s[i] = select(-1, my[0].s[i], y_s >= 0);
176            my[0].s[i] = select(-1, my[0].s[i], y_s < _ISRC_HEIGHT);
177        })
178
179        int ck = 0;
180        for(; ck <= (_ISRC_CHANNELS - K0); ck += K0)
181        {
182            TILE(SRC_DATA_TYPE, M0, K0, a);
183            TILE(WEI_DATA_TYPE, N0, K0, b);
184
185            // Initialize tiles
186            LOOP_UNROLLING(int, i, 0, 1, M0,
187            {
188                a[i].v = ZERO_VALUE;
189            })
190
191            LOOP_UNROLLING(int, i, 0, 1, N0,
192            {
193                b[i].v = ZERO_VALUE;
194            })
195
196            // Load tile from the src tensor
197            T_LOAD2D_INDIRECT(SRC_DATA_TYPE, M0, K0, SRC_TENSOR_TYPE, src, ck, src_stride_y, my, a);
198
199            // Load tile from the weights tensor
200            T_LOAD(WEI_DATA_TYPE, N0, K0, WEI_TENSOR_TYPE, wei, ck, cout * _IY_MULTIPLIER + i, _IY_MULTIPLIER, wei_stride_y, b);
201
202            // Compute the matrix multiplication between two tiles
203            T_MMUL(SRC_DATA_TYPE, WEI_DATA_TYPE, ACC_DATA_TYPE, M0, N0, K0, NT, T, a, b, c);
204
205            // Apply the offset correction (correction usually needed for asymmetric quantized computation)
206            // The computation is not performed if both SRC_OFFSET and WEI_OFFSET are zero
207            T_OFFSET_CORRECTION(ACC_DATA_TYPE, M0, N0, K0, SRC_OFFSET, WEI_OFFSET, a, b, c);
208        }
209
210        // This #if directive should be removed in case of dynamic tensor support
211#if defined(LEFTOVER_LOOP)
212        // Left-over accumulations
213        for(; ck < _ISRC_CHANNELS; ++ck)
214        {
215            TILE(SRC_DATA_TYPE, M0, 1, a);
216            TILE(WEI_DATA_TYPE, N0, 1, b);
217
218            // Initialize tiles
219            LOOP_UNROLLING(int, i, 0, 1, M0,
220            {
221                a[i].v = ZERO_VALUE;
222            })
223
224            LOOP_UNROLLING(int, i, 0, 1, N0,
225            {
226                b[i].v = ZERO_VALUE;
227            })
228
229            // Load tile from the src tensor
230            T_LOAD2D_INDIRECT(SRC_DATA_TYPE, M0, 1, SRC_TENSOR_TYPE, src, ck, src_stride_y, my, a);
231
232            // Load tile from the weights tensor
233            // The T_LOAD for the left-over elements can only use BUFFER because we load one element per iteration
234            T_LOAD(WEI_DATA_TYPE, N0, 1, BUFFER, wei, ck, cout * _IY_MULTIPLIER + i, _IY_MULTIPLIER, wei_stride_y, b);
235
236            // Compute the matrix multiplication between two tiles
237            T_MMUL(SRC_DATA_TYPE, WEI_DATA_TYPE, ACC_DATA_TYPE, M0, N0, 1, NT, T, a, b, c);
238
239            // Apply the offset correction (operation usually needed for asymmetric quantized computation)
240            // The computation is not performed if both SRC_OFFSET and WEI_OFFSET are zero
241            T_OFFSET_CORRECTION(ACC_DATA_TYPE, M0, N0, 1, SRC_OFFSET, WEI_OFFSET, a, b, c);
242        }
243#endif // defined(LEFTOVER_LOOP)
244    }
245
246    // Offset correction required for the quantized asymmetric computation
247    // The computation is not performed if both SRC_OFFSET and WEI_OFFSET are zero
248    T_ADD_CONSTANT(ACC_DATA_TYPE, M0, N0, c, (_IWEI_WIDTH * _IWEI_HEIGHT * _ISRC_CHANNELS * SRC_OFFSET * WEI_OFFSET), c);
249
250#if defined(HAS_BIAS)
251    TILE(BIA_DATA_TYPE, 1, N0, bias0);
252
253    T_LOAD(BIA_DATA_TYPE, 1, N0, BUFFER, bia, cout, 0, 1, 0, bias0);
254
255    // c = c + bias[broadcasted]
256    T_ELTWISE_BROADCAST_ADD_X(ACC_DATA_TYPE, M0, N0, c, bias0, c);
257
258#endif // HAS_BIAS
259
260#if defined(IS_QUANTIZED)
261
262    TILE(DST_DATA_TYPE, M0, N0, cq);
263
264    // Quantize the tile
265    T_QUANTIZE8_ASYMMETRIC(ACC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, c, cq);
266#endif // defined(IS_QUANTIZED)
267
268    // Apply activation
269    T_ACTIVATION(DST_DATA_TYPE, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, _IOUTPUT_TILE, _IOUTPUT_TILE);
270
271    TILE(uint, M0, 1, dst_indirect_y);
272
273    // Calculate the destination indirect Y
274    LOOP_UNROLLING(int, i, 0, 1, M0,
275    {
276        dst_indirect_y[i].v = (uint)min(mout + i, (int)(_IDST_WIDTH * _IDST_HEIGHT) - 1);
277        dst_indirect_y[i].v += bout * (int)(_IDST_WIDTH * _IDST_HEIGHT);
278    })
279
280    bool x_cond = PARTIAL_N0 != 0 && get_global_id(0) == 0;
281
282    // _IOUTPUT_TILE: c = fp32/fp16, cq=qasymm8
283    // Store the tile in reverse order so the invalid values are overwritten with the valid ones
284    T_STORE_INDIRECT_WIDTH_SELECT(DST_DATA_TYPE, M0, N0, PARTIAL_N0, DST_TENSOR_TYPE, dst, cout, dst_stride_y, x_cond, _IOUTPUT_TILE, dst_indirect_y);
285
286#undef _IWEI_WIDTH
287#undef _IWEI_HEIGHT
288#undef _ISRC_WIDTH
289#undef _ISRC_HEIGHT
290#undef _ISRC_CHANNELS
291#undef _IDST_WIDTH
292#undef _IDST_HEIGHT
293#undef _IDST_CHANNELS
294#undef _IY_MULTIPLIER
295}
296