xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/cl_kernels/nhwc/dwc_native_quantized_nhwc.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 "helpers.h"
26#include "tile_helpers.h"
27
28// *INDENT-OFF*
29// clang-format off
30#define CALCULATE_WEIGHTS_OFFSET_CORRECTION(A_DATA_TYPE, B_DATA_TYPE) CALCULATE_WEIGHTS_OFFSET_CORRECTION_STR(A_DATA_TYPE, B_DATA_TYPE)
31#define CALCULATE_WEIGHTS_OFFSET_CORRECTION_STR(A_DATA_TYPE, B_DATA_TYPE) CALCULATE_WEIGHTS_OFFSET_CORRECTION_##A_DATA_TYPE##_##B_DATA_TYPE
32#define CALCULATE_WEIGHTS_OFFSET_CORRECTION_char_char (0)
33#define CALCULATE_WEIGHTS_OFFSET_CORRECTION_uchar_uchar (0)
34#define CALCULATE_WEIGHTS_OFFSET_CORRECTION_uchar_char (128)
35#define CALCULATE_WEIGHTS_OFFSET_CORRECTION_char_uchar (-128)
36
37#define T_LOAD_MULTIPLIERS_SHIFT_PER_TENSOR() \
38    ({})
39
40#define T_LOAD_MULTIPLIERS_SHIFT_PER_CHANNEL()                                                     \
41    TILE(DST_MULTIPLIERS_DATA_TYPE, 1, N0, multipliers);                                           \
42    TILE(DST_SHIFTS_DATA_TYPE, 1, N0, shifts);                                                     \
43    T_LOAD(DST_MULTIPLIERS_DATA_TYPE, 1, N0, BUFFER, dst_multipliers, cout, 0, 0, 0, multipliers); \
44    T_LOAD(DST_SHIFTS_DATA_TYPE, 1, N0, BUFFER, dst_shifts, cout, 0, 0, 0, shifts);
45
46#define T_LOAD_MULTIPLIERS_SHIFT(QUANTIZATION_TYPE) T_LOAD_MULTIPLIERS_SHIFT_STR(QUANTIZATION_TYPE)
47#define T_LOAD_MULTIPLIERS_SHIFT_STR(QUANTIZATION_TYPE) T_LOAD_MULTIPLIERS_SHIFT_##QUANTIZATION_TYPE()
48
49#if defined(WEI_WIDTH) && defined(WEI_HEIGHT) && defined(N0) && defined(M0) && defined(DILATION_X) && defined(DILATION_Y) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP)
50//! @cond Doxygen_Suppress
51/** OpenCL kernel to compute the depthwise convolution for quantized data types
52 *
53 * @note Data layout supported: NHWC
54 * @note Data type supported: QSYMM8/QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL
55 * @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)
56 * @note The convolution strides must be passed at compile time using -DSTRIDE_X and -DSTRIDE_Y (e.g. -DSTRIDE_X=2, -DSTRIDE_Y=2)
57 * @note The convolution dilations must be passed at compile time using -DDILATION_X and -DDILATION_Y (e.g. -DDILATION_X=2, -DDILATION_Y=2)
58 * @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)
59 * @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)
60 * @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)
61 * @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)
62 * @note The data type of the source tensor must be passed at compile time using -DSRC_DATA_TYPE (e.g. -DSRC_DATA_TYPE=int8)
63 * @note The data type of the weights tensor must be passed at compile time using -DWEI_DATA_TYPE (e.g. -DWEI_DATA_TYPE=int8)
64 * @note The data type of the destination tensor must be passed at compile time using -DDST_DATA_TYPE (e.g. -DDST_DATA_TYPE=int8)
65 * @note The data type of the accumulators must be passed at compile time using -DACC_DATA_TYPE (e.g. -DACC_DATA_TYPE=int)
66 * @note The number of M0 rows (width) to process must be passed at compile time using -DM0 (e.g. -DM0=2)
67 * @note The number of N0 output channels to process must be passed at compile time using -DN0 (e.g. -DN0=2)
68 * @note The size of the partial store block in the first dimension must be passed at compile time using -DPARTIAL_N0 (e.g. -DPARTIAL_N0=1)
69 * @note The activation type must be passed at compile using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu
70 * @note The A and B variables required by some activation functions must be passed at compile time using -DA_VAL= and -DB_VAL= respectively
71 * @note The quantization offset used for both the per-tensor and per-channel quantization must be passed at compile using -DDST_OFFSET (e.g., -DDST_OFFSET=3)
72 * @note The quantization shift for the per-tensor quantization must be passed at compile time using -DDST_SHIFT (e.g., -DDST_SHIFT=1)
73 * @note The quantization multiplier for the per-tensor quantization must be passed at compile using -DDST_MULTIPLIER (e.g., -DDST_MULTIPLER=121432)
74 * @note Only the following configurations of M0 and N0 are currently supported:
75 *  - M0 = 1, 2, 3, 4, 5, .... n (M0 != 1 with STRIDE_X == 1 && DILATION_X == 1 only)
76 *  - N0 = 2, 3, 4, 8, 16
77 * @note The number of rows to read from the src tensor must be passed at compile time using -DM0_A (e.g., -DM0_A=3). M0_A must be equal to WEI_WIDTH + (M0 - 1)
78 * @note The number of columns to read from the src tensor must be passed at compile time using -DN0_A. It can either be 1 (for DEPTH_MULTIPLIER > 1) or N0 (for DEPTH_MULTIPLIER == 1)
79 *
80 * @param[in]  src_img                                       (Not supported) Read only cl_image object for the source tensor. Included when SRC_TENSOR_TYPE=IMAGE
81 * @param[in]  src_ptr                                       Pointer to the source tensor. Supported data type: QSYMM8/QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL
82 * @param[in]  src_stride_y                                  Stride of the source tensor in Y dimension (in bytes)
83 * @param[in]  src_stride_z                                  Stride of the source tensor in Z dimension (in bytes)
84 * @param[in]  src_stride_w                                  Stride of the source tensor in W dimension (in bytes)
85 * @param[in]  src_c                                         The size of the channels dimension of the source tensor
86 * @param[in]  src_w                                         The size of the width dimension of the source tensor
87 * @param[in]  src_h                                         The size of the height dimension of the source tensor
88 * @param[in]  src_n                                         The size of the batches dimension of the source tensor
89 * @param[in]  src_offset_first_element_in_bytes             The offset of the first element in the source tensor
90 * @param[out] dst_img                                       (Not supported) Write only cl_image object for the destination tensor. Included when DST_TENSOR_TYPE=IMAGE
91 * @param[out] dst_ptr                                       Pointer to the destination tensor. Supported data type: same as @p src_ptr
92 * @param[in]  dst_stride_y                                  Stride of the destination tensor in Y dimension (in bytes)
93 * @param[in]  dst_stride_z                                  Stride of the destination tensor in Z dimension (in bytes)
94 * @param[in]  dst_stride_w                                  Stride of the destination tensor in W dimension (in bytes)
95 * @param[in]  dst_c                                         The size of the channels dimension of the destination tensor
96 * @param[in]  dst_w                                         The size of the width dimension of the destination tensor
97 * @param[in]  dst_h                                         The size of the height dimension of the destination tensor
98 * @param[in]  dst_n                                         The size of the batches dimension of the destination tensor
99 * @param[in]  dst_offset_first_element_in_bytes             The offset of the first element in the destination tensor
100 * @param[in]  wei_img                                       (Not supported) Read only cl_image object for the weights tensor. Included when WEI_TENSOR_TYPE=IMAGE
101 * @param[in]  wei_ptr                                       Pointer to the weights tensor. Supported data type: same as @p src_ptr
102 * @param[in]  wei_stride_y                                  Stride of the weights tensor in Y dimension (in bytes)
103 * @param[in]  wei_stride_z                                  Stride of the weights tensor in Z dimension (in bytes)
104 * @param[in]  wei_stride_w                                  Stride of the weights tensor in W dimension (in bytes)
105 * @param[in]  wei_c                                         The size of the channels dimension of the weights tensor
106 * @param[in]  wei_w                                         The size of the width dimension of the weights tensor
107 * @param[in]  wei_h                                         The size of the height dimension of the weights tensor
108 * @param[in]  wei_n                                         The size of the batches dimension of the weights tensor
109 * @param[in]  wei_step_w                                    wei_stride_w * number of elements along W processed per workitem(in bytes)
110 * @param[in]  wei_offset_first_element_in_bytes             The offset of the first element in the weights tensor
111 * @param[in]  dst_multipliers_ptr                           Pointer to the destination multipliers tensor for the per-channel quantization. Supported data type: S32
112 * @param[in]  dst_multipliers_stride_x                      Stride of the destination multipliers tensor in X dimension (in bytes)
113 * @param[in]  dst_multipliers_step_x                        dst_multipliers_stride_x * number of elements along X processed per workitem(in bytes)
114 * @param[in]  dst_multipliers_offset_first_element_in_bytes The offset of the first element in the destination multipliers tensor
115 * @param[in]  dst_shifts_ptr                                Pointer to the destination shifts tensor for the per-channel quantization. Supported data type: S32
116 * @param[in]  dst_shifts_stride_x                           Stride of the destination shifts tensor in X dimension (in bytes)
117 * @param[in]  dst_shifts_step_x                             dst_shifts_stride_x * number of elements along X processed per workitem(in bytes)
118 * @param[in]  dst_shifts_offset_first_element_in_bytes      The offset of the first element in the destination shifts tensor
119 * @param[in]  bia_ptr                                       (Optional) Pointer to the bias tensor Supported data type: S32
120 * @param[in]  bia_stride_x                                  (Optional) Stride of the bias tensor in X dimension (in bytes)
121 * @param[in]  bia_step_x                                    (Optional) bia_stride_x * number of elements along X processed per workitem(in bytes)
122 * @param[in]  bia_offset_first_element_in_bytes             (Optional) The offset of the first element in the bias tensor
123 */
124//! @endcond
125__kernel void dwc_native_quantized_nhwc(
126    TENSOR4D_RO_T(src, SRC_TENSOR_TYPE),
127    TENSOR4D_WO_T(dst, DST_TENSOR_TYPE),
128    TENSOR4D_RO_T(wei, WEI_TENSOR_TYPE),
129    VECTOR_DECLARATION(dst_multipliers),
130    VECTOR_DECLARATION(dst_shifts)
131#if defined(HAS_BIAS)
132    ,
133    VECTOR_DECLARATION(bia)
134#endif // defined(HAS_BIAS)
135)
136{
137    // Only the weight tensor dimensions are passed at compile time.
138    // In case of dynamic tensor support, the following dimensions should be passed as function argument.
139#define _IWEI_WIDTH WEI_WIDTH
140#define _IWEI_HEIGHT WEI_HEIGHT
141#define _IM0_A M0_A        // _IWEI_WIDTH + (M0 - 1) Rows tile A (If M0 != 1, the tiles overlap of 1 element on the X dimension)
142#define _IN0_A N0_A        // Cols tile A. It can be either 1 (for DEPTH_MULTIPLIER > 1) or N0 (for DEPTH_MULTIPLIER == 1)
143#define _IM0_B _IWEI_WIDTH // Rows tile B
144#define _IN0_B N0          // Cols tile B
145#define _IBOUNDARY_CHECK (!((WEI_WIDTH == 1 && WEI_HEIGHT == 1 && PAD_LEFT == 0 && PAD_TOP == 0 && M0 == 1)))
146
147    const int cout = GET_SPATIAL_IDX(0, N0, PARTIAL_N0); // OFM
148    const int xo   = GET_SPATIAL_IDX(1, M0, 0);          // WIDTH
149#if defined(BATCHED_EXECUTION)
150    const int yo   = GET_SPATIAL_IDX(2, 1, 0) % dst_h; // HEIGHT
151    const int bout = GET_SPATIAL_IDX(2, 1, 0) / dst_h; // BATCH SIZE IDX
152#else                                                  // defined(BATCHED_EXECUTION)
153    const int yo   = GET_SPATIAL_IDX(2, 1, 0); // HEIGHT
154    const int bout = 0;                        // BATCH SIZE IDX
155#endif                                                 // defined(BATCHED_EXECUTION)
156
157    int xi = xo * STRIDE_X;
158    int yi = yo * STRIDE_Y;
159    xi -= PAD_LEFT;
160    yi -= PAD_TOP;
161
162    TILE(ACC_DATA_TYPE, M0, N0, c);
163
164    // Reset accumulators
165    LOOP_UNROLLING(int, i, 0, 1, M0,
166    {
167        c[i].v = 0;
168    })
169
170#if _IWEI_HEIGHT <= 5
171    LOOP_UNROLLING(int, yk, 0, 1, _IWEI_HEIGHT,
172#else  // _IWEI_HEIGHT <= 5
173    for(int yk = 0; yk < _IWEI_HEIGHT; yk++)
174#endif // _IWEI_HEIGHT <= 5
175                   {
176                       TILE(SRC_DATA_TYPE, _IM0_A, _IN0_A, a);
177
178                       LOOP_UNROLLING(int, i, 0, 1, _IM0_A,
179    {
180        a[i].v = ZERO_VALUE;
181    })
182
183    // Load tile from the src tensor (TILE A)
184    T_LOAD_NHWC_WITH_DILATION(SRC_DATA_TYPE, 1, _IM0_A, _IN0_A, SRC_TENSOR_TYPE, src, bout, yi + yk * DILATION_Y, xi, (cout / DEPTH_MULTIPLIER), src_w, src_h, DILATION_X, 1, _IBOUNDARY_CHECK, a);
185
186    TILE(WEI_DATA_TYPE, _IM0_B, _IN0_B, b);
187
188    // Load tile from the weights tensor (TILE B)
189    T_LOAD(WEI_DATA_TYPE, _IM0_B, _IN0_B, WEI_TENSOR_TYPE, wei, cout, yk * _IM0_B, 1, wei_stride_y, b);
190
191    // Optimized path for STRIDE_X == 1
192    // If M0 != 1, we can skip the common loads between the two applied kernels on the X (WIDTH) dimension
193    LOOP_UNROLLING(int, m0, 0, 1, M0,
194    {
195        LOOP_UNROLLING(int, n0, 0, 1, N0,
196        {
197#if _IWEI_WIDTH <= 16
198#define DOT_DATA_TYPE SRC_DATA_TYPE
199#define WEI_OFFSET_CORRECTION (CALCULATE_WEIGHTS_OFFSET_CORRECTION(SRC_DATA_TYPE, WEI_DATA_TYPE))
200
201            // Optimized path for the dot instruction
202            TILE(DOT_DATA_TYPE, 1, _IWEI_WIDTH, x0);
203            TILE(DOT_DATA_TYPE, 1, _IWEI_WIDTH, y0);
204            ACC_DATA_TYPE offset_a = 0;
205            ACC_DATA_TYPE offset_b = 0;
206
207            LOOP_UNROLLING(int, xk, 0, 1, _IWEI_WIDTH,
208            {
209                x0[0].s[xk] = a[xk + m0].s[n0];
210                y0[0].s[xk] = b[xk].s[n0] + (int)WEI_OFFSET_CORRECTION;
211            })
212            DOT_PRODUCT_INTEGER8(DOT_DATA_TYPE, DOT_DATA_TYPE, ACC_DATA_TYPE, _IWEI_WIDTH, x0[0].v, y0[0].v, c[m0].s[n0]);
213            REDUCE_INTEGER8(DOT_DATA_TYPE, DOT_DATA_TYPE, ACC_DATA_TYPE, _IWEI_WIDTH, x0[0].v, offset_a);
214            REDUCE_INTEGER8(DOT_DATA_TYPE, DOT_DATA_TYPE, ACC_DATA_TYPE, _IWEI_WIDTH, y0[0].v, offset_b);
215            c[m0].s[n0] += offset_a * (ACC_DATA_TYPE)(WEI_OFFSET - (ACC_DATA_TYPE)WEI_OFFSET_CORRECTION) + offset_b * (ACC_DATA_TYPE)SRC_OFFSET;
216#else  // _IWEI_WIDTH <= 16
217            LOOP_UNROLLING(int, xk, 0, 1, _IWEI_WIDTH,
218            {
219                c[m0].s[n0] += ((ACC_DATA_TYPE)a[xk + m0].s[n0] + (ACC_DATA_TYPE)(SRC_OFFSET)) * ((ACC_DATA_TYPE)b[xk].s[n0] + (ACC_DATA_TYPE)(WEI_OFFSET));
220            })
221#endif // _IWEI_WIDTH <= 16
222        })
223    })
224                   }
225#if _IWEI_HEIGHT <= 5
226                  )
227#endif // _IWEI_HEIGHT <= 5
228
229#if _IWEI_WIDTH <= 16
230    T_ADD_CONSTANT(ACC_DATA_TYPE, M0, N0, c, (_IWEI_WIDTH * _IWEI_HEIGHT * SRC_OFFSET * (ACC_DATA_TYPE)(WEI_OFFSET - (ACC_DATA_TYPE)WEI_OFFSET_CORRECTION)), c);
231#endif // _IWEI_WIDTH <= 16
232
233#if defined(HAS_BIAS)
234    TILE(BIA_DATA_TYPE, 1, N0, bias0);
235
236    // Load bias
237    T_LOAD(BIA_DATA_TYPE, 1, N0, BUFFER, bia, cout, 0, 0, 0, bias0);
238
239    // c = c + bias[broadcasted]
240    T_ELTWISE_BROADCAST_ADD_X(ACC_DATA_TYPE, M0, N0, c, bias0, c);
241#endif // HAS_BIAS
242
243    T_LOAD_MULTIPLIERS_SHIFT(QUANTIZATION_TYPE);
244
245    // Quantize the tile
246    TILE(DST_DATA_TYPE, M0, N0, cq);
247    T_QUANTIZE8(ACC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, c, multipliers, shifts, cq);
248
249    // Perform activation
250    T_ACTIVATION_QUANTIZED(DST_DATA_TYPE, M0, N0, ACTIVATION_TYPE, DST_OFFSET, A_VAL, B_VAL, cq, cq);
251
252    bool x_cond = PARTIAL_N0 != 0 && get_global_id(0) == 0;
253
254    if(x_cond)
255    {
256        LOOP_UNROLLING(int, m0, 0, 1, M0,
257        {
258            int xi_out = min(xo + M0 - 1 - m0, (int)(dst_w) - 1);
259            VSTORE_PARTIAL(N0, PARTIAL_N0)
260            (cq[M0 - 1 - m0].v, 0, (__global DST_DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + (uint)cout * sizeof(DST_DATA_TYPE) + (uint)xi_out * dst_stride_y + (uint)yo * dst_stride_z + (uint)bout * dst_stride_w));
261        })
262    }
263    else
264    {
265        LOOP_UNROLLING(int, m0, 0, 1, M0,
266        {
267            int xi_out = min(xo + M0 - 1 - m0, (int)(dst_w) - 1);
268            VSTORE(N0)
269            (cq[M0 - 1 - m0].v, 0, (__global DST_DATA_TYPE *)(dst_ptr + dst_offset_first_element_in_bytes + (uint)cout * sizeof(DST_DATA_TYPE) + (uint)xi_out * dst_stride_y + (uint)yo * dst_stride_z + (uint)bout * dst_stride_w));
270        })
271    }
272}
273#endif // defined(WEI_WIDTH) && defined(WEI_HEIGHT) && defined(N0) && defined(M0) && defined(DILATION_X) && defined(DILATION_Y) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP)
274// *INDENT-ON*
275// clang-format on
276