Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 3558c5840e7c973e2b1a86ae3a9335b44cad59d4 / . / src / core / CL / cl_kernels / tile_helpers.h
blob: 861ea63eca0ebd4a4362f02afc4666e49d3e96f3 [file] [log] [blame]
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1/*
Gian Marco Iodice10e88a72021-11-29 12:49:19 +0000[diff] [blame]2 * Copyright (c) 2021-2022 Arm Limited.
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[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
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]25// *INDENT-OFF*
26// clang-format off
SiCong Lica364df2022-04-13 15:48:19 +0100[diff] [blame]27#ifndef ARM_COMPUTE_TILE_HELPERS_H
28#define ARM_COMPUTE_TILE_HELPERS_H
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]29
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]30#define TILE_VECTOR_SIZE1 1
31#define TILE_VECTOR_SIZE2 2
32#define TILE_VECTOR_SIZE3 3
33#define TILE_VECTOR_SIZE4 4
34#define TILE_VECTOR_SIZE5 8
35#define TILE_VECTOR_SIZE6 8
36#define TILE_VECTOR_SIZE7 8
37#define TILE_VECTOR_SIZE8 8
38#define TILE_VECTOR_SIZE9 16
39#define TILE_VECTOR_SIZE10 16
40#define TILE_VECTOR_SIZE11 16
41#define TILE_VECTOR_SIZE12 16
42#define TILE_VECTOR_SIZE13 16
43#define TILE_VECTOR_SIZE14 16
44#define TILE_VECTOR_SIZE15 16
45#define TILE_VECTOR_SIZE16 16
46
47#define TILE_VECTOR_TYPE1(DATA_TYPE) DATA_TYPE##1
48#define TILE_VECTOR_TYPE2(DATA_TYPE) DATA_TYPE##2
49#define TILE_VECTOR_TYPE3(DATA_TYPE) DATA_TYPE##3
50#define TILE_VECTOR_TYPE4(DATA_TYPE) DATA_TYPE##4
51#define TILE_VECTOR_TYPE5(DATA_TYPE) DATA_TYPE##8
52#define TILE_VECTOR_TYPE6(DATA_TYPE) DATA_TYPE##8
53#define TILE_VECTOR_TYPE7(DATA_TYPE) DATA_TYPE##8
54#define TILE_VECTOR_TYPE8(DATA_TYPE) DATA_TYPE##8
55#define TILE_VECTOR_TYPE9(DATA_TYPE) DATA_TYPE##16
56#define TILE_VECTOR_TYPE10(DATA_TYPE) DATA_TYPE##16
57#define TILE_VECTOR_TYPE11(DATA_TYPE) DATA_TYPE##16
58#define TILE_VECTOR_TYPE12(DATA_TYPE) DATA_TYPE##16
59#define TILE_VECTOR_TYPE13(DATA_TYPE) DATA_TYPE##16
60#define TILE_VECTOR_TYPE14(DATA_TYPE) DATA_TYPE##16
61#define TILE_VECTOR_TYPE15(DATA_TYPE) DATA_TYPE##16
62#define TILE_VECTOR_TYPE16(DATA_TYPE) DATA_TYPE##16
63
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]64/** Tile object
65 * A tile object is a 2D memory block and can be accessed using the following syntax:
66 * -# a[m0].v = access the the vector at row "m0" (OpenCL vector)
Ramy Elgammal404462a2022-11-08 02:14:46 +0000[diff] [blame]67 * -# dst[m0].s[n0] = access the scalar element at row "m0" and column "n0" (scalar access)
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]68 *
69 * @param[in] DATA_TYPE Data type of the tile
70 * @param[in] H Number of tile rows
71 * @param[in] W Number of tile colums
72 * @param[in] BASENAME Tile's name
73 */
74#define TILE(DATA_TYPE, H, W, BASENAME) TILE_STR(DATA_TYPE, H, W, BASENAME)
75#define TILE_STR(DATA_TYPE, H, W, BASENAME) \
76 union { \
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]77 DATA_TYPE s[TILE_VECTOR_SIZE##W]; \
78 TILE_VECTOR_TYPE##W(DATA_TYPE) v; \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]79 } BASENAME[H]
80
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]81#define TENSOR4D_IMAGE(name) \
82 __read_only image2d_t name##_img, \
83 __global uchar *name##_ptr, \
84 uint name##_stride_x, \
85 uint name##_step_x, \
86 uint name##_stride_y, \
87 uint name##_step_y, \
88 uint name##_stride_z, \
89 uint name##_step_z, \
90 uint name##_stride_w, \
91 uint name##_step_w, \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]92 uint name##_offset_first_element_in_bytes
93
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]94#define TENSOR4D_BUFFER(name) \
95 __global uchar *name##_ptr, \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]96 uint name##_stride_x, \
97 uint name##_step_x, \
98 uint name##_stride_y, \
99 uint name##_step_y, \
100 uint name##_stride_z, \
101 uint name##_step_z, \
102 uint name##_stride_w, \
103 uint name##_step_w, \
104 uint name##_offset_first_element_in_bytes
105
106#define TENSOR4D_STR(name, type) TENSOR4D_##type(name)
107#define TENSOR4D(name, type) TENSOR4D_STR(name, type)
108
Adnan AlSinan17975a62021-11-08 17:46:39 +0000[diff] [blame]109#define TENSOR4D_T_IMAGE(name) \
110 __read_only image2d_t name##_img, \
111 __global uchar *name##_ptr, \
112 uint name##_stride_y, \
113 uint name##_stride_z, \
114 uint name##_stride_w, \
115 uint name##_c, \
116 uint name##_w, \
117 uint name##_h, \
118 uint name##_n, \
119 uint name##_offset_first_element_in_bytes
120
121#define TENSOR4D_T_BUFFER(name) \
122 __global uchar *name##_ptr, \
123 uint name##_stride_y, \
124 uint name##_stride_z, \
125 uint name##_stride_w, \
126 uint name##_c, \
127 uint name##_w, \
128 uint name##_h, \
129 uint name##_n, \
130 uint name##_offset_first_element_in_bytes
131
132#define TENSOR4D_T_STR(name, type) TENSOR4D_T_##type(name)
133#define TENSOR4D_T(name, type) TENSOR4D_T_STR(name, type)
134
Gian Marco Iodice4fb56702021-11-10 11:18:50 +0000[diff] [blame]135#define TENSOR3D_T_IMAGE(name) \
136 __read_only image2d_t name##_img, \
137 __global uchar *name##_ptr, \
138 uint name##_stride_y, \
139 uint name##_stride_z, \
140 uint name##_w, \
141 uint name##_h, \
142 uint name##_n, \
143 uint name##_offset_first_element_in_bytes
144
145#define TENSOR3D_T_BUFFER(name) \
146 __global uchar *name##_ptr, \
147 uint name##_stride_y, \
148 uint name##_stride_z, \
149 uint name##_w, \
150 uint name##_h, \
151 uint name##_n, \
152 uint name##_offset_first_element_in_bytes
153
154#define TENSOR3D_T_STR(name, type) TENSOR3D_T_##type(name)
155#define TENSOR3D_T(name, type) TENSOR3D_T_STR(name, type)
156
Giorgio Arenaea8d2662021-05-20 11:36:56 +0100[diff] [blame]157#if !defined(UNROLL_WITH_PRAGMA)
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]158#define UNROLL_INCR(idx, step, macro) idx += (step); (macro)
159
160#define LOOP_UNROLLING_1(idx, step, macro) (macro)
161#define LOOP_UNROLLING_2(idx, step, macro) LOOP_UNROLLING_1(idx, step, macro); UNROLL_INCR(idx, step, macro)
162#define LOOP_UNROLLING_3(idx, step, macro) LOOP_UNROLLING_2(idx, step, macro); UNROLL_INCR(idx, step, macro)
163#define LOOP_UNROLLING_4(idx, step, macro) LOOP_UNROLLING_3(idx, step, macro); UNROLL_INCR(idx, step, macro)
164#define LOOP_UNROLLING_5(idx, step, macro) LOOP_UNROLLING_4(idx, step, macro); UNROLL_INCR(idx, step, macro)
165#define LOOP_UNROLLING_6(idx, step, macro) LOOP_UNROLLING_5(idx, step, macro); UNROLL_INCR(idx, step, macro)
166#define LOOP_UNROLLING_7(idx, step, macro) LOOP_UNROLLING_6(idx, step, macro); UNROLL_INCR(idx, step, macro)
167#define LOOP_UNROLLING_8(idx, step, macro) LOOP_UNROLLING_7(idx, step, macro); UNROLL_INCR(idx, step, macro)
168#define LOOP_UNROLLING_9(idx, step, macro) LOOP_UNROLLING_8(idx, step, macro); UNROLL_INCR(idx, step, macro)
169#define LOOP_UNROLLING_10(idx, step, macro) LOOP_UNROLLING_9(idx, step, macro); UNROLL_INCR(idx, step, macro)
170#define LOOP_UNROLLING_11(idx, step, macro) LOOP_UNROLLING_10(idx, step, macro); UNROLL_INCR(idx, step, macro)
171#define LOOP_UNROLLING_12(idx, step, macro) LOOP_UNROLLING_11(idx, step, macro); UNROLL_INCR(idx, step, macro)
172#define LOOP_UNROLLING_13(idx, step, macro) LOOP_UNROLLING_12(idx, step, macro); UNROLL_INCR(idx, step, macro)
173#define LOOP_UNROLLING_14(idx, step, macro) LOOP_UNROLLING_13(idx, step, macro); UNROLL_INCR(idx, step, macro)
174#define LOOP_UNROLLING_15(idx, step, macro) LOOP_UNROLLING_14(idx, step, macro); UNROLL_INCR(idx, step, macro)
175#define LOOP_UNROLLING_16(idx, step, macro) LOOP_UNROLLING_15(idx, step, macro); UNROLL_INCR(idx, step, macro)
176#define LOOP_UNROLLING_17(idx, step, macro) LOOP_UNROLLING_16(idx, step, macro); UNROLL_INCR(idx, step, macro)
177#define LOOP_UNROLLING_18(idx, step, macro) LOOP_UNROLLING_17(idx, step, macro); UNROLL_INCR(idx, step, macro)
178#define LOOP_UNROLLING_19(idx, step, macro) LOOP_UNROLLING_18(idx, step, macro); UNROLL_INCR(idx, step, macro)
179#define LOOP_UNROLLING_20(idx, step, macro) LOOP_UNROLLING_19(idx, step, macro); UNROLL_INCR(idx, step, macro)
180#define LOOP_UNROLLING_21(idx, step, macro) LOOP_UNROLLING_20(idx, step, macro); UNROLL_INCR(idx, step, macro)
181#define LOOP_UNROLLING_22(idx, step, macro) LOOP_UNROLLING_21(idx, step, macro); UNROLL_INCR(idx, step, macro)
182#define LOOP_UNROLLING_23(idx, step, macro) LOOP_UNROLLING_22(idx, step, macro); UNROLL_INCR(idx, step, macro)
183#define LOOP_UNROLLING_24(idx, step, macro) LOOP_UNROLLING_23(idx, step, macro); UNROLL_INCR(idx, step, macro)
184#define LOOP_UNROLLING_25(idx, step, macro) LOOP_UNROLLING_24(idx, step, macro); UNROLL_INCR(idx, step, macro)
185#define LOOP_UNROLLING_26(idx, step, macro) LOOP_UNROLLING_25(idx, step, macro); UNROLL_INCR(idx, step, macro)
186#define LOOP_UNROLLING_27(idx, step, macro) LOOP_UNROLLING_26(idx, step, macro); UNROLL_INCR(idx, step, macro)
187#define LOOP_UNROLLING_28(idx, step, macro) LOOP_UNROLLING_27(idx, step, macro); UNROLL_INCR(idx, step, macro)
188#define LOOP_UNROLLING_29(idx, step, macro) LOOP_UNROLLING_28(idx, step, macro); UNROLL_INCR(idx, step, macro)
189#define LOOP_UNROLLING_30(idx, step, macro) LOOP_UNROLLING_29(idx, step, macro); UNROLL_INCR(idx, step, macro)
190#define LOOP_UNROLLING_31(idx, step, macro) LOOP_UNROLLING_30(idx, step, macro); UNROLL_INCR(idx, step, macro)
191#define LOOP_UNROLLING_32(idx, step, macro) LOOP_UNROLLING_31(idx, step, macro); UNROLL_INCR(idx, step, macro)
192#define LOOP_UNROLLING_33(idx, step, macro) LOOP_UNROLLING_32(idx, step, macro); UNROLL_INCR(idx, step, macro)
193#define LOOP_UNROLLING_34(idx, step, macro) LOOP_UNROLLING_33(idx, step, macro); UNROLL_INCR(idx, step, macro)
194#define LOOP_UNROLLING_35(idx, step, macro) LOOP_UNROLLING_34(idx, step, macro); UNROLL_INCR(idx, step, macro)
195#define LOOP_UNROLLING_36(idx, step, macro) LOOP_UNROLLING_35(idx, step, macro); UNROLL_INCR(idx, step, macro)
196#define LOOP_UNROLLING_37(idx, step, macro) LOOP_UNROLLING_36(idx, step, macro); UNROLL_INCR(idx, step, macro)
197#define LOOP_UNROLLING_38(idx, step, macro) LOOP_UNROLLING_37(idx, step, macro); UNROLL_INCR(idx, step, macro)
198#define LOOP_UNROLLING_39(idx, step, macro) LOOP_UNROLLING_38(idx, step, macro); UNROLL_INCR(idx, step, macro)
199#define LOOP_UNROLLING_40(idx, step, macro) LOOP_UNROLLING_39(idx, step, macro); UNROLL_INCR(idx, step, macro)
200#define LOOP_UNROLLING_41(idx, step, macro) LOOP_UNROLLING_40(idx, step, macro); UNROLL_INCR(idx, step, macro)
201#define LOOP_UNROLLING_42(idx, step, macro) LOOP_UNROLLING_41(idx, step, macro); UNROLL_INCR(idx, step, macro)
202#define LOOP_UNROLLING_43(idx, step, macro) LOOP_UNROLLING_42(idx, step, macro); UNROLL_INCR(idx, step, macro)
203#define LOOP_UNROLLING_44(idx, step, macro) LOOP_UNROLLING_43(idx, step, macro); UNROLL_INCR(idx, step, macro)
204#define LOOP_UNROLLING_45(idx, step, macro) LOOP_UNROLLING_44(idx, step, macro); UNROLL_INCR(idx, step, macro)
205#define LOOP_UNROLLING_46(idx, step, macro) LOOP_UNROLLING_45(idx, step, macro); UNROLL_INCR(idx, step, macro)
206#define LOOP_UNROLLING_47(idx, step, macro) LOOP_UNROLLING_46(idx, step, macro); UNROLL_INCR(idx, step, macro)
207#define LOOP_UNROLLING_48(idx, step, macro) LOOP_UNROLLING_47(idx, step, macro); UNROLL_INCR(idx, step, macro)
208#define LOOP_UNROLLING_49(idx, step, macro) LOOP_UNROLLING_48(idx, step, macro); UNROLL_INCR(idx, step, macro)
209#define LOOP_UNROLLING_50(idx, step, macro) LOOP_UNROLLING_49(idx, step, macro); UNROLL_INCR(idx, step, macro)
210#define LOOP_UNROLLING_51(idx, step, macro) LOOP_UNROLLING_50(idx, step, macro); UNROLL_INCR(idx, step, macro)
211#define LOOP_UNROLLING_52(idx, step, macro) LOOP_UNROLLING_51(idx, step, macro); UNROLL_INCR(idx, step, macro)
212#define LOOP_UNROLLING_53(idx, step, macro) LOOP_UNROLLING_52(idx, step, macro); UNROLL_INCR(idx, step, macro)
213#define LOOP_UNROLLING_54(idx, step, macro) LOOP_UNROLLING_53(idx, step, macro); UNROLL_INCR(idx, step, macro)
214#define LOOP_UNROLLING_55(idx, step, macro) LOOP_UNROLLING_54(idx, step, macro); UNROLL_INCR(idx, step, macro)
215#define LOOP_UNROLLING_56(idx, step, macro) LOOP_UNROLLING_55(idx, step, macro); UNROLL_INCR(idx, step, macro)
216#define LOOP_UNROLLING_57(idx, step, macro) LOOP_UNROLLING_56(idx, step, macro); UNROLL_INCR(idx, step, macro)
217#define LOOP_UNROLLING_58(idx, step, macro) LOOP_UNROLLING_57(idx, step, macro); UNROLL_INCR(idx, step, macro)
218#define LOOP_UNROLLING_59(idx, step, macro) LOOP_UNROLLING_58(idx, step, macro); UNROLL_INCR(idx, step, macro)
219#define LOOP_UNROLLING_60(idx, step, macro) LOOP_UNROLLING_59(idx, step, macro); UNROLL_INCR(idx, step, macro)
220#define LOOP_UNROLLING_61(idx, step, macro) LOOP_UNROLLING_60(idx, step, macro); UNROLL_INCR(idx, step, macro)
221#define LOOP_UNROLLING_62(idx, step, macro) LOOP_UNROLLING_61(idx, step, macro); UNROLL_INCR(idx, step, macro)
222#define LOOP_UNROLLING_63(idx, step, macro) LOOP_UNROLLING_62(idx, step, macro); UNROLL_INCR(idx, step, macro)
223#define LOOP_UNROLLING_64(idx, step, macro) LOOP_UNROLLING_63(idx, step, macro); UNROLL_INCR(idx, step, macro)
224#define LOOP_UNROLLING_65(idx, step, macro) LOOP_UNROLLING_64(idx, step, macro); UNROLL_INCR(idx, step, macro)
225#define LOOP_UNROLLING_66(idx, step, macro) LOOP_UNROLLING_65(idx, step, macro); UNROLL_INCR(idx, step, macro)
226#define LOOP_UNROLLING_67(idx, step, macro) LOOP_UNROLLING_66(idx, step, macro); UNROLL_INCR(idx, step, macro)
227#define LOOP_UNROLLING_68(idx, step, macro) LOOP_UNROLLING_67(idx, step, macro); UNROLL_INCR(idx, step, macro)
228#define LOOP_UNROLLING_69(idx, step, macro) LOOP_UNROLLING_68(idx, step, macro); UNROLL_INCR(idx, step, macro)
229#define LOOP_UNROLLING_70(idx, step, macro) LOOP_UNROLLING_69(idx, step, macro); UNROLL_INCR(idx, step, macro)
230#define LOOP_UNROLLING_71(idx, step, macro) LOOP_UNROLLING_70(idx, step, macro); UNROLL_INCR(idx, step, macro)
231#define LOOP_UNROLLING_72(idx, step, macro) LOOP_UNROLLING_71(idx, step, macro); UNROLL_INCR(idx, step, macro)
232#define LOOP_UNROLLING_73(idx, step, macro) LOOP_UNROLLING_72(idx, step, macro); UNROLL_INCR(idx, step, macro)
233#define LOOP_UNROLLING_74(idx, step, macro) LOOP_UNROLLING_73(idx, step, macro); UNROLL_INCR(idx, step, macro)
234#define LOOP_UNROLLING_75(idx, step, macro) LOOP_UNROLLING_74(idx, step, macro); UNROLL_INCR(idx, step, macro)
235#define LOOP_UNROLLING_76(idx, step, macro) LOOP_UNROLLING_75(idx, step, macro); UNROLL_INCR(idx, step, macro)
236#define LOOP_UNROLLING_77(idx, step, macro) LOOP_UNROLLING_76(idx, step, macro); UNROLL_INCR(idx, step, macro)
237#define LOOP_UNROLLING_78(idx, step, macro) LOOP_UNROLLING_77(idx, step, macro); UNROLL_INCR(idx, step, macro)
238#define LOOP_UNROLLING_79(idx, step, macro) LOOP_UNROLLING_78(idx, step, macro); UNROLL_INCR(idx, step, macro)
239#define LOOP_UNROLLING_80(idx, step, macro) LOOP_UNROLLING_79(idx, step, macro); UNROLL_INCR(idx, step, macro)
240#define LOOP_UNROLLING_81(idx, step, macro) LOOP_UNROLLING_80(idx, step, macro); UNROLL_INCR(idx, step, macro)
241#define LOOP_UNROLLING_82(idx, step, macro) LOOP_UNROLLING_81(idx, step, macro); UNROLL_INCR(idx, step, macro)
242#define LOOP_UNROLLING_83(idx, step, macro) LOOP_UNROLLING_82(idx, step, macro); UNROLL_INCR(idx, step, macro)
243#define LOOP_UNROLLING_84(idx, step, macro) LOOP_UNROLLING_83(idx, step, macro); UNROLL_INCR(idx, step, macro)
244#define LOOP_UNROLLING_85(idx, step, macro) LOOP_UNROLLING_84(idx, step, macro); UNROLL_INCR(idx, step, macro)
245#define LOOP_UNROLLING_86(idx, step, macro) LOOP_UNROLLING_85(idx, step, macro); UNROLL_INCR(idx, step, macro)
246#define LOOP_UNROLLING_87(idx, step, macro) LOOP_UNROLLING_86(idx, step, macro); UNROLL_INCR(idx, step, macro)
247#define LOOP_UNROLLING_88(idx, step, macro) LOOP_UNROLLING_87(idx, step, macro); UNROLL_INCR(idx, step, macro)
248#define LOOP_UNROLLING_89(idx, step, macro) LOOP_UNROLLING_88(idx, step, macro); UNROLL_INCR(idx, step, macro)
249#define LOOP_UNROLLING_90(idx, step, macro) LOOP_UNROLLING_89(idx, step, macro); UNROLL_INCR(idx, step, macro)
250#define LOOP_UNROLLING_91(idx, step, macro) LOOP_UNROLLING_90(idx, step, macro); UNROLL_INCR(idx, step, macro)
251#define LOOP_UNROLLING_92(idx, step, macro) LOOP_UNROLLING_91(idx, step, macro); UNROLL_INCR(idx, step, macro)
252#define LOOP_UNROLLING_93(idx, step, macro) LOOP_UNROLLING_92(idx, step, macro); UNROLL_INCR(idx, step, macro)
253#define LOOP_UNROLLING_94(idx, step, macro) LOOP_UNROLLING_93(idx, step, macro); UNROLL_INCR(idx, step, macro)
254#define LOOP_UNROLLING_95(idx, step, macro) LOOP_UNROLLING_94(idx, step, macro); UNROLL_INCR(idx, step, macro)
255#define LOOP_UNROLLING_96(idx, step, macro) LOOP_UNROLLING_95(idx, step, macro); UNROLL_INCR(idx, step, macro)
256#define LOOP_UNROLLING_97(idx, step, macro) LOOP_UNROLLING_96(idx, step, macro); UNROLL_INCR(idx, step, macro)
257#define LOOP_UNROLLING_98(idx, step, macro) LOOP_UNROLLING_97(idx, step, macro); UNROLL_INCR(idx, step, macro)
258#define LOOP_UNROLLING_99(idx, step, macro) LOOP_UNROLLING_98(idx, step, macro); UNROLL_INCR(idx, step, macro)
259#define LOOP_UNROLLING_100(idx, step, macro) LOOP_UNROLLING_99(idx, step, macro); UNROLL_INCR(idx, step, macro)
260#define LOOP_UNROLLING_101(idx, step, macro) LOOP_UNROLLING_100(idx, step, macro); UNROLL_INCR(idx, step, macro)
261#define LOOP_UNROLLING_102(idx, step, macro) LOOP_UNROLLING_101(idx, step, macro); UNROLL_INCR(idx, step, macro)
262#define LOOP_UNROLLING_103(idx, step, macro) LOOP_UNROLLING_102(idx, step, macro); UNROLL_INCR(idx, step, macro)
263#define LOOP_UNROLLING_104(idx, step, macro) LOOP_UNROLLING_103(idx, step, macro); UNROLL_INCR(idx, step, macro)
264#define LOOP_UNROLLING_105(idx, step, macro) LOOP_UNROLLING_104(idx, step, macro); UNROLL_INCR(idx, step, macro)
265#define LOOP_UNROLLING_106(idx, step, macro) LOOP_UNROLLING_105(idx, step, macro); UNROLL_INCR(idx, step, macro)
266#define LOOP_UNROLLING_107(idx, step, macro) LOOP_UNROLLING_106(idx, step, macro); UNROLL_INCR(idx, step, macro)
267#define LOOP_UNROLLING_108(idx, step, macro) LOOP_UNROLLING_107(idx, step, macro); UNROLL_INCR(idx, step, macro)
268#define LOOP_UNROLLING_109(idx, step, macro) LOOP_UNROLLING_108(idx, step, macro); UNROLL_INCR(idx, step, macro)
269#define LOOP_UNROLLING_110(idx, step, macro) LOOP_UNROLLING_109(idx, step, macro); UNROLL_INCR(idx, step, macro)
270#define LOOP_UNROLLING_111(idx, step, macro) LOOP_UNROLLING_110(idx, step, macro); UNROLL_INCR(idx, step, macro)
271#define LOOP_UNROLLING_112(idx, step, macro) LOOP_UNROLLING_111(idx, step, macro); UNROLL_INCR(idx, step, macro)
272#define LOOP_UNROLLING_113(idx, step, macro) LOOP_UNROLLING_112(idx, step, macro); UNROLL_INCR(idx, step, macro)
273#define LOOP_UNROLLING_114(idx, step, macro) LOOP_UNROLLING_113(idx, step, macro); UNROLL_INCR(idx, step, macro)
274#define LOOP_UNROLLING_115(idx, step, macro) LOOP_UNROLLING_114(idx, step, macro); UNROLL_INCR(idx, step, macro)
275#define LOOP_UNROLLING_116(idx, step, macro) LOOP_UNROLLING_115(idx, step, macro); UNROLL_INCR(idx, step, macro)
276#define LOOP_UNROLLING_117(idx, step, macro) LOOP_UNROLLING_116(idx, step, macro); UNROLL_INCR(idx, step, macro)
277#define LOOP_UNROLLING_118(idx, step, macro) LOOP_UNROLLING_117(idx, step, macro); UNROLL_INCR(idx, step, macro)
278#define LOOP_UNROLLING_119(idx, step, macro) LOOP_UNROLLING_118(idx, step, macro); UNROLL_INCR(idx, step, macro)
279#define LOOP_UNROLLING_120(idx, step, macro) LOOP_UNROLLING_119(idx, step, macro); UNROLL_INCR(idx, step, macro)
280#define LOOP_UNROLLING_121(idx, step, macro) LOOP_UNROLLING_120(idx, step, macro); UNROLL_INCR(idx, step, macro)
281#define LOOP_UNROLLING_122(idx, step, macro) LOOP_UNROLLING_121(idx, step, macro); UNROLL_INCR(idx, step, macro)
282#define LOOP_UNROLLING_123(idx, step, macro) LOOP_UNROLLING_122(idx, step, macro); UNROLL_INCR(idx, step, macro)
283#define LOOP_UNROLLING_124(idx, step, macro) LOOP_UNROLLING_123(idx, step, macro); UNROLL_INCR(idx, step, macro)
284#define LOOP_UNROLLING_125(idx, step, macro) LOOP_UNROLLING_124(idx, step, macro); UNROLL_INCR(idx, step, macro)
285#define LOOP_UNROLLING_126(idx, step, macro) LOOP_UNROLLING_125(idx, step, macro); UNROLL_INCR(idx, step, macro)
286#define LOOP_UNROLLING_127(idx, step, macro) LOOP_UNROLLING_126(idx, step, macro); UNROLL_INCR(idx, step, macro)
287#define LOOP_UNROLLING_128(idx, step, macro) LOOP_UNROLLING_127(idx, step, macro); UNROLL_INCR(idx, step, macro)
288
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]289#define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \
290 { \
291 type idx = start; \
292 LOOP_UNROLLING_##num(idx, step, macro); \
293 }
Giorgio Arenaea8d2662021-05-20 11:36:56 +0100[diff] [blame]294#else // !defined(UNROLL_WITH_PRAGMA)
295#define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \
296 { \
297 _Pragma("unroll") \
298 for(type idx = start; idx < (num * step); idx += step) \
299 { \
300 (macro); \
301 } \
302 }
303#endif // !defined(UNROLL_WITH_PRAGMA)
304#define LOOP_UNROLLING(type, idx, start, step, num, macro) LOOP_UNROLLING_STR(type, idx, start, step, num, macro)
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]305
306/** Get the get_global_id with partial N0. This function is useful when the dimension is not multiple of N0 and we need to use a partial N0
307 * to avoid out-of-bound read/write
308 *
309 * @note PARTIAL_N0 is used for get_global_id(n) = 0.
310 *
311 * @param[in] IDX get_global_id index (0,1 and 2 only)
312 * @param[in] N0 Number of elements read/written on the IDX direction
313 * @param[in] PARTIAL_N0 Number of elements read/written on the IDX direction for get_global_id(IDX) = 0. If zero,
314 * the Number of elements read/written on the IDX direction for get_global_id(IDX) = 0 is N0
315 */
316#define GET_SPATIAL_IDX(IDX, N0, PARTIAL_N0) (max((int)(get_global_id(IDX) * N0 - (N0 - PARTIAL_N0) % N0), 0))
317
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]318/** Dot product integet 8bit function
319 *
320 * @note Performs: c += dot(a, b)
321 *
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]322 * @param[in] A_DATA_TYPE A (lhs) data type
323 * @param[in] B_DATA_TYPE B (rhs) data type
324 * @param[in] C_DATA_TYPE C (accumulator) data type
325 * @param[in] K0 Number of accumulations
326 * @param[in] a OpenCL vector a
327 * @param[in] b OpenCL vector b
328 * @param[in] c Scalar variable c
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]329 */
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]330#define DOT_PRODUCT_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c) DOT_PRODUCT_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c)
331#define DOT_PRODUCT_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c) DOT_PRODUCT##K0##_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c)
332#define DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]333 ({ \
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]334 c += (C_DATA_TYPE)(a) * (C_DATA_TYPE)(b); \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]335 })
Viet-Hoa Do82169b32022-05-26 16:50:21 +0100[diff] [blame]336#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_khr_integer_dot_product)
337#define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0)));
338#define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0));
339#define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((a), (b));
340#elif defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_khr_integer_dot_product)
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]341#define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0)), (c));
342#define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0), (c));
343#define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((a), (b), (c));
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]344#elif defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
Michalis Spyrouc38ca382021-07-14 13:30:28 +0100[diff] [blame]345#define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0)));
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]346#define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0));
347#define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((a), (b));
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]348#else // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]349#define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
350 ({ \
351 c += (C_DATA_TYPE)(a).s0 * (C_DATA_TYPE)(b).s0; \
352 c += (C_DATA_TYPE)(a).s1 * (C_DATA_TYPE)(b).s1; \
353 })
354#define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
355 ({ \
356 DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c); \
357 c += (C_DATA_TYPE)(a).s2 * (C_DATA_TYPE)(b).s2; \
358 })
359#define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, x, y, val) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]360 ({ \
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]361 val += (C_DATA_TYPE)(x).s0 * (C_DATA_TYPE)(y).s0; \
362 val += (C_DATA_TYPE)(x).s1 * (C_DATA_TYPE)(y).s1; \
363 val += (C_DATA_TYPE)(x).s2 * (C_DATA_TYPE)(y).s2; \
364 val += (C_DATA_TYPE)(x).s3 * (C_DATA_TYPE)(y).s3; \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]365 })
366#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]367#define DOT_PRODUCT5_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
368 ({ \
369 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \
370 DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s4), ((b).s4), c); \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]371 })
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]372#define DOT_PRODUCT6_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
373 ({ \
374 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \
375 DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s45), ((b).s45), c); \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]376 })
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]377#define DOT_PRODUCT7_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
378 ({ \
379 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \
380 DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s456), ((b).s456), c); \
381 })
382#define DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
383 ({ \
384 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).lo), ((b).lo), c); \
385 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).hi), ((b).hi), c); \
386 })
387#define DOT_PRODUCT9_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
388 ({ \
389 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \
390 DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s8), ((b).s8), c); \
391 })
392#define DOT_PRODUCT10_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
393 ({ \
394 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \
395 DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89), ((b).s89), c); \
396 })
397#define DOT_PRODUCT11_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
398 ({ \
399 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \
400 DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89A), ((b).s89A), c); \
401 })
402#define DOT_PRODUCT12_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
403 ({ \
404 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \
405 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89AB), ((b).s89AB), c); \
406 })
407#define DOT_PRODUCT13_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
408 ({ \
409 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \
410 DOT_PRODUCT5_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABC), ((b).s89ABC), c); \
411 })
412#define DOT_PRODUCT14_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
413 ({ \
414 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \
415 DOT_PRODUCT6_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABCD), ((b).s89ABCD), c); \
416 })
417#define DOT_PRODUCT15_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
418 ({ \
419 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \
420 DOT_PRODUCT7_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABCDE), ((b).s89ABCDE), c); \
421 })
422#define DOT_PRODUCT16_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
423 ({ \
424 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).lo), ((b).lo), c); \
425 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).hi), ((b).hi), c); \
426 })
427
428/** Dot product integet 8bit function
429 *
430 * @note Performs: c += dot(a, b)
431 *
432 * @param[in] A_DATA_TYPE A (lhs) data type
433 * @param[in] B_DATA_TYPE B (rhs) data type
434 * @param[in] C_DATA_TYPE C (accumulator) data type
435 * @param[in] K0 Number of accumulations
436 * @param[in] a OpenCL vector a
437 * @param[in] c Scalar variable c
438 */
439#define REDUCE_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c) REDUCE_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c)
440#define REDUCE_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c) DOT_PRODUCT_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, (TILE_VECTOR_TYPE##K0(B_DATA_TYPE))1, c)
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]441
442/** Load a vector from global memory (tensor)
443 *
444 * @param[in] DATA_TYPE Data type
445 * @param[in] WIDTH Number of dst columns
446 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image).
447 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
448 * @param[in] TENSOR Tensor basename
449 * @param[in] X Starting X position
450 * @param[in] Y Starting Y position
451 * @param[in] STRIDE_Y Stride Y (in bytes)
452 */
453#define V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) V_LOAD_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y)
454#define V_LOAD_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) V_LOAD_##TENSOR_TYPE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y)
455#define V_LOAD_BUFFER(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) \
456 VLOAD(WIDTH) \
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]457 (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (Y) * (STRIDE_Y)))
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]458#define V_LOAD_IMAGE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) READ_IMAGE2D(DATA_TYPE, CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(WIDTH), TENSOR##_img, (X) / 4, (Y))
459
460/** Load a tile from global memory (tensor)
461 *
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]462 * @param[in] DATA_TYPE Data type
463 * @param[in] HEIGHT Number of dst rows
464 * @param[in] WIDTH Number of dst columns
465 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image).
466 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
467 * @param[in] TENSOR Tensor basename
468 * @param[in] X Starting X position
469 * @param[in] Y Starting Y position
470 * @param[in] YI_MULTIPLIER Parameter used to multiply the internal row increment (_i).
471 * In common cases should be 1 but it becomes useful when we want to load rows which are multiple of STRIDE_Y. (e.g. loading the weights of convolution layer).
472 * In this case the address calculation is performed as: (Y + _i * Y_MULTIPLIER) * STRIDE_Y
473 * @param[in] STRIDE_Y Stride Y (in bytes) used to load each row.
474 * @param[out] dst Output tile
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]475 */
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]476#define T_LOAD(DATA_TYPE, HEIGHT, WIDTH, TENSOR_TYPE, TENSOR, X, Y, YI_MULTIPLIER, STRIDE_Y, dst) \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]477 ({ \
478 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \
479 { \
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]480 dst[_i].v = V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, ((Y) + _i * (int)(YI_MULTIPLIER)), STRIDE_Y); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]481 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]482 })
483
484/** Load a tile from global memory (tensor) using an indirect Y index tile
485 *
486 * @param[in] DATA_TYPE Data type
487 * @param[in] HEIGHT Number of dst rows
488 * @param[in] WIDTH Number of dst columns
489 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported
490 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
491 * @param[in] TENSOR Tensor basename
492 * @param[in] X Starting X position
493 * @param[in] STRIDE_Y Stride Y (in bytes)
494 * @param[in] indirect_y Indirect Y index tile
495 * @param[out] dst Output tile
496 */
497#define T_LOAD_INDIRECT(DATA_TYPE, HEIGHT, WIDTH, TENSOR_TYPE, TENSOR, X, STRIDE_Y, indirect_y, dst) \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]498 ({ \
499 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \
500 { \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]501 dst[_i].v = V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, (indirect_y[_i].v), STRIDE_Y); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]502 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]503 })
504
Adnan AlSinan3e155a52021-12-10 12:34:02 +0000[diff] [blame]505/** Load a tile from global memory (tensor) using an indirect Y index tile and conditionally use a different length for the load
506 *
507 * @note If WIDTH1_CONDITION is true, the load will use the WIDTH1 length for the store
508 * @note The vectors are stored in reverse order so the invalid rows are overwritten by the valid ones
509 *
510 * @param[in] DATA_TYPE Data type
511 * @param[in] HEIGHT Number of dst rows
512 * @param[in] WIDTH0 Store width to use if WIDTH1_CONDITION = false
513 * @param[in] WIDTH1 Store width to use if WIDTH1_CONDITION = true
514 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image).
515 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
516 * @param[in] TENSOR Tensor basename
517 * @param[in] X Starting X position
518 * @param[in] STRIDE_Y Stride Y (in bytes) used to load each row.
519 * @param[in] WIDTH1_CONDITION Condition to select the WIDTH1 store
520 * @param[out] dst Output tile
521 * @param[out] indirect_y Indirect Y index tile
522 */
523#define T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, STRIDE_Y, WIDTH1_CONDITION, dst, indirect_y) \
524 ({ \
525 if(WIDTH1_CONDITION) \
526 { \
527 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \
528 { \
529 VLOAD_PARTIAL(WIDTH0, WIDTH1) \
530 (dst[HEIGHT - 1 - _i].v, 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \
531 }) \
532 } \
533 else \
534 { \
535 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \
536 { \
537 dst[HEIGHT - 1 - _i].v = V_LOAD(DATA_TYPE, WIDTH0, TENSOR_TYPE, TENSOR, X, (indirect_y[HEIGHT - 1 - _i].v), STRIDE_Y); \
538 }) \
539 } \
540 })
Gian Marco Iodice534b8892021-04-01 16:17:16 +0100[diff] [blame]541/** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout
542 *
543 * @param[in] DATA_TYPE Data type
544 * @param[in] TILE_HEIGHT Number of elements to load from Y (height) dimension
545 * @param[in] TILE_WIDTH Number of elements to load from X (width) dimension
546 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension
547 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported
548 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16)
549 * @param[in] TENSOR Tensor basename
550 * @param[in] B Starting batch index
551 * @param[in] Y Starting Y index
552 * @param[in] X Starting X index
553 * @param[in] C Starting C index
554 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension
555 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension
556 * @param[in] STRIDE_Y Stride Y (in bytes)
557 * @param[out] dst Output tile
558 */
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]559#define T_LOAD_NHWC(DATA_TYPE, TILE_HEIGHT, TILE_WIDTH, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, STRIDE_Y, dst) \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]560 ({ \
561 LOOP_UNROLLING(int, _yk, 0, 1, TILE_HEIGHT, \
562 { \
563 LOOP_UNROLLING(int, _xk, 0, 1, TILE_WIDTH, \
564 { \
565 int _src_y = (X) + _xk + ((Y) + _yk) * (TENSOR_WIDTH); \
566 _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT); \
Gian Marco Iodice534b8892021-04-01 16:17:16 +0100[diff] [blame]567 int _src_valid_y = (((X) + _xk) >= 0 && ((X) + _xk) < (int)(TENSOR_WIDTH) && ((Y) + _yk) >= 0 && ((Y) + _yk) < (int)(TENSOR_HEIGHT)); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]568 if(_src_valid_y != 0) \
569 { \
570 dst[_xk + _yk * (TILE_WIDTH)].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]571 } \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]572 }) \
573 }) \
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]574 })
575
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]576/** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout with dilation for the X and Y increments
577 *
578 * @param[in] DATA_TYPE Data type
579 * @param[in] TILE_HEIGHT Number of elements to load from Y (height) dimension
580 * @param[in] TILE_WIDTH Number of elements to load from X (width) dimension
581 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension
582 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported
583 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16)
584 * @param[in] TENSOR Tensor basename
585 * @param[in] B Starting batch index
586 * @param[in] Y Starting Y index
587 * @param[in] X Starting X index
588 * @param[in] C Starting C index
589 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension
590 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension
591 * @param[in] DILATION_X Dilation for the X increment
592 * @param[in] DILATION_Y Dilation for the Y increment
593 * @param[in] BOUNDARY_CHECK Boundary check flag. If true, it checks for any out-of-bound reads
594 * @param[out] dst Output tile
595 */
596#define T_LOAD_NHWC_WITH_DILATION(DATA_TYPE, TILE_HEIGHT, TILE_WIDTH, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, DILATION_X, DILATION_Y, BOUNDARY_CHECK, dst) \
597 ({ \
598 LOOP_UNROLLING(int, _yk, 0, 1, TILE_HEIGHT, \
599 { \
600 LOOP_UNROLLING(int, _xk, 0, 1, TILE_WIDTH, \
601 { \
602 int _src_y = (X) + _xk * (DILATION_X); \
603 int _src_z = ((Y) + _yk * (DILATION_Y)); \
604 int _src_w = (B); \
605 bool _src_valid_y = (((X) + _xk * (DILATION_X)) >= 0) && (((X) + _xk * (DILATION_X)) < (int)(TENSOR_WIDTH)) && (((Y) + _yk * (DILATION_Y)) >= 0) && (((Y) + _yk * (DILATION_Y)) < (int)(TENSOR_HEIGHT)); \
606 if(!(BOUNDARY_CHECK)) \
607 { \
608 dst[_xk + _yk * (TILE_WIDTH)].v = VLOAD(TILE_CHANNELS) \
609 (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (C) * sizeof(DATA_TYPE) + (_src_y) * (TENSOR##_stride_y) + (_src_z) * (TENSOR##_stride_z) + (_src_w) * (TENSOR##_stride_w))); \
610 } \
611 else \
612 { \
613 if(_src_valid_y) \
614 { \
615 dst[_xk + _yk * (TILE_WIDTH)].v = VLOAD(TILE_CHANNELS) \
616 (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (C) * sizeof(DATA_TYPE) + (_src_y) * (TENSOR##_stride_y) + (_src_z) * (TENSOR##_stride_z) + (_src_w) * (TENSOR##_stride_w))); \
617 } \
618 } \
619 }) \
620 }) \
621 })
622
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]623/** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout using indirect X and Y coordinates
624 *
625 * @param[in] DATA_TYPE Data type
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]626 * @param[in] TILE_AREA Number of elements to load from Y (height) dimension * Number of elements to load from X (width) dimension
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]627 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension
628 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported
629 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16)
630 * @param[in] TENSOR Tensor basename
631 * @param[in] B Starting batch index
632 * @param[in] Y Starting Y index
633 * @param[in] X Starting X index
634 * @param[in] C Starting C index
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]635 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension
Giorgio Arena945ae9e2021-10-13 11:13:04 +0100[diff] [blame]636 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]637 * @param[in] STRIDE_Y Stride Y (in bytes)
638 * @param[out] xi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect X coordinate
639 * @param[out] yi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect Y coordinate
640 * @param[out] dst Output tile
641 */
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]642#define T_LOAD_NHWC_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, STRIDE_Y, xi, yi, dst) \
643 ({ \
644 LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \
645 { \
646 int _src_y = (X) + xi[_i].v + ((Y) + yi[_i].v) * (TENSOR_WIDTH); \
647 _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT); \
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]648 int _src_valid_y = (((X) + xi[_i].v) >= 0 && ((X) + xi[_i].v) < (int)(TENSOR_WIDTH) && ((Y) + yi[_i].v) >= 0 && ((Y) + yi[_i].v) < (int)(TENSOR_HEIGHT)); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]649 if(_src_valid_y != 0) \
650 { \
651 dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]652 } \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]653 }) \
Gian Marco Iodice534b8892021-04-01 16:17:16 +0100[diff] [blame]654 })
655
Gian Marco Iodice76335eb2022-11-17 11:03:39 +0000[diff] [blame]656/** Load a tile from global memory (tensor) using an indirect buffer for the Y coordinates
657 *
658 * @param[in] DATA_TYPE Data type
659 * @param[in] TILE_AREA Number of elements to load from Y (height) dimension * Number of elements to load from X (width) dimension
660 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension
661 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported
662 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16)
663 * @param[in] TENSOR Tensor basename
664 * @param[in] C Starting C index
665 * @param[in] STRIDE_Y Stride Y (in bytes)
666 * @param[out] yi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect Y coordinate
667 * 16 is the maximum indirect buffer size.
668 * @param[out] dst Output tile
669 */
670#define T_LOAD2D_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) \
Gian Marco Iodice3394f3e2022-09-16 14:14:21 +0100[diff] [blame]671 ({ \
672 LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \
673 { \
Gian Marco Iodice76335eb2022-11-17 11:03:39 +0000[diff] [blame]674 if(yi[0].s[_i] >= 0) \
Gian Marco Iodice3394f3e2022-09-16 14:14:21 +0100[diff] [blame]675 { \
Gian Marco Iodice76335eb2022-11-17 11:03:39 +0000[diff] [blame]676 dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, yi[0].s[_i], STRIDE_Y); \
Gian Marco Iodice3394f3e2022-09-16 14:14:21 +0100[diff] [blame]677 } \
678 }) \
679 })
680
Giorgio Arena945ae9e2021-10-13 11:13:04 +0100[diff] [blame]681/** Load a tile from global memory (tensor) when the tensor is stored using a NDHWC layout using indirect X, Y and Z coordinates
682 *
683 * @param[in] DATA_TYPE Data type
684 * @param[in] TILE_AREA Number of elements to load from Y (height) dimension * Number of elements to load from X (width) dimension
685 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension
686 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported
687 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16)
688 * @param[in] TENSOR Tensor basename
689 * @param[in] B Starting batch index
690 * @param[in] Z Starting Z index
691 * @param[in] Y Starting Y index
692 * @param[in] X Starting X index
693 * @param[in] C Starting C index
694 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension
695 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension
696 * @param[in] TENSOR_DEPTH Number of elements to load from Z (depth) dimension
697 * @param[in] STRIDE_Y Stride Y (in bytes)
698 * @param[out] xi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect X coordinate
699 * @param[out] yi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect Y coordinate
700 * @param[out] zi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect Z coordinate
701 * @param[out] dst Output tile
702 */
703#define T_LOAD_NDHWC_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Z, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, TENSOR_DEPTH, STRIDE_Y, xi, yi, zi, dst) \
704 ({ \
705 LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \
706 { \
707 int _src_y = (X) + xi[_i].v + ((Y) + yi[_i].v) * (TENSOR_WIDTH) + ((Z) + zi[_i].v) * (TENSOR_WIDTH * TENSOR_HEIGHT); \
708 _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT) * (int)(TENSOR_DEPTH); \
709 int _src_valid_y = (((X) + xi[_i].v) >= 0 && ((X) + xi[_i].v) < (int)(TENSOR_WIDTH) && ((Y) + yi[_i].v) >= 0 && ((Y) + yi[_i].v) < (int)(TENSOR_HEIGHT) \
710 && ((Z) + zi[_i].v) >= 0 && ((Z) + zi[_i].v) < (int)(TENSOR_DEPTH)); \
711 if(_src_valid_y != 0) \
712 { \
713 dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \
714 } \
715 }) \
716 })
717
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]718/** Store a tile to global memory (tensor) using an indirect Y index tile and conditionally use a different length for the store
719 *
720 * @note If WIDTH1_CONDITION is true, the store will use the WIDTH1 length for the store
721 * @note The vectors are stored in reverse order so the invalid rows are overwritten by the valid ones
722 *
723 * @param[in] DATA_TYPE Data type
724 * @param[in] HEIGHT Number of src rows
725 * @param[in] WIDTH0 Store width to use if WIDTH1_CONDITION = false
726 * @param[in] WIDTH1 Store width to use if WIDTH1_CONDITION = true
727 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported
728 * cl_image is not supported.
729 * @param[in] TENSOR Tensor basename
730 * @param[in] X Starting X position
731 * @param[in] STRIDE_Y Stride Y (in bytes)
732 * @param[in] WIDTH1_CONDITION Condition to select the WIDTH1 store
733 * @param[in] src Input tile
734 * @param[in] indirect_y Indirect Y index tile
735 */
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]736#define T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, STRIDE_Y, WIDTH1_CONDITION, src, indirect_y) \
737 ({ \
738 if(WIDTH1_CONDITION) \
739 { \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]740 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]741 { \
742 VSTORE_PARTIAL(WIDTH0, WIDTH1) \
Giorgio Arena945ae9e2021-10-13 11:13:04 +0100[diff] [blame]743 (CONVERT(src[HEIGHT - 1 - _i].v, VEC_DATA_TYPE(DATA_TYPE, WIDTH0)), 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]744 }) \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]745 } \
746 else \
747 { \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]748 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]749 { \
750 VSTORE(WIDTH0) \
Giorgio Arena945ae9e2021-10-13 11:13:04 +0100[diff] [blame]751 (CONVERT(src[HEIGHT - 1 - _i].v, VEC_DATA_TYPE(DATA_TYPE, WIDTH0)), 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]752 }) \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]753 } \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]754 })
755
756/** Offset correction for the QASYMM8 computation
757 *
758 * @param[in] ACC_DATA_TYPE Accumulator data type
759 * @param[in] M0 Number of src/dst rows
760 * @param[in] N0 Number of src/dst columns
761 * @param[in] K0 Number of src columns
762 * @param[in] SRC_OFFSET Source quantization offset
763 * @param[in] WEI_OFFSET Weights quantization shift
764 * @param[in] lhs LHS tile
765 * @param[in] rhs RHS tile
766 * @param[out] dst DST tile
767 */
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]768#define T_OFFSET_CORRECTION(ACC_DATA_TYPE, M0, N0, K0, SRC_OFFSET, WEI_OFFSET, lhs, rhs, dst) \
769 ({ \
770 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
771 { \
772 ACC_DATA_TYPE _tm = 0; \
773 LOOP_UNROLLING(int, _k0, 0, 1, K0, \
774 { \
775 _tm += ((ACC_DATA_TYPE)lhs[_m0].s[_k0] * (ACC_DATA_TYPE)WEI_OFFSET); \
776 }) \
777 LOOP_UNROLLING(int, _n0, 0, 1, N0, \
778 { \
779 dst[_m0].s[_n0] += _tm; \
780 LOOP_UNROLLING(int, _k0, 0, 1, K0, \
781 { \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]782 dst[_m0].s[_n0] += ((ACC_DATA_TYPE)rhs[_n0].s[_k0] * (ACC_DATA_TYPE)SRC_OFFSET); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]783 }) \
784 }) \
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]785 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]786 })
787
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]788/** 8-bit quantization with fixed-point scale
789 *
790 * @param[in] SRC_DATA_TYPE SRC data type
791 * @param[in] DST_DATA_TYPE DST data type
792 * @param[in] QUANTIZATION_TYPE Quantization type (PER_TENSOR or PER_CHANNEL)
793 * @param[in] M0 Number of src/dst rows
794 * @param[in] N0 Number of src/dst columns
795 * @param[in] DST_OFFSET Quantization offset used for both the per-tensor and per-channel quantization
796 * @param[in] DST_SHIFT Quantization shift for the per-tensor quantization
797 * @param[in] DST_MULTIPLIER Quantization multiplier for the per-tensor quantization
798 * @param[in] src Input tile
799 * @param[in] dst_multipliers Output multipliers tile for the per-channel quantization
800 * @param[in] dst_shifts Output shift tile for the per-channel quantization
801 * @param[out] dst Output tile
802 */
803#define T_QUANTIZE8(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) T_QUANTIZE8_STR(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst)
804#define T_QUANTIZE8_STR(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) T_QUANTIZE8_##QUANTIZATION_TYPE(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst)
805
806/** 8-bit per-tensor quantization with fixed-point scale
807 *
808 * @param[in] SRC_DATA_TYPE SRC data type
809 * @param[in] DST_DATA_TYPE DST data type
810 * @param[in] M0 Number of src/dst rows
811 * @param[in] N0 Number of src/dst columns
812 * @param[in] DST_OFFSET Quantization offset
813 * @param[in] DST_SHIFT Quantization shift for the per-tensor quantization
814 * @param[in] DST_MULTIPLIER Quantization multiplier for the per-tensor quantization
815 * @param[in] src Input tile
816 * @param[in] dst_multipliers (unused)
817 * @param[in] dst_shifts (unused)
818 * @param[out] dst Output tile
819 */
820#define T_QUANTIZE8_PER_TENSOR(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) \
821 ({ \
822 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
823 { \
824 LOOP_UNROLLING(int, _n0, 0, 1, N0, \
825 { \
826 SRC_DATA_TYPE _tmp = 0; \
827 SRC_DATA_TYPE _src = src[_m0].s[_n0]; \
828 _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-DST_SHIFT)), ((SRC_DATA_TYPE)DST_SHIFT < (SRC_DATA_TYPE)0)); \
829 SRC_DATA_TYPE overflow = _src == DST_MULTIPLIER && _src == INT_MIN; \
830 long a_64 = (long)(_src); \
831 long b_64 = (long)(DST_MULTIPLIER); \
832 long ab_64 = a_64 * b_64; \
833 long mask1 = 1 << 30; \
834 long mask2 = 1 - (1 << 30); \
835 long is_positive_or_zero = ab_64 >= 0; \
836 long nudge = select(mask2, mask1, is_positive_or_zero); \
837 SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \
838 _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \
839 if(DST_SHIFT >= 0) \
840 { \
Freddie Liardet767dbf92021-07-21 16:20:41 +0100[diff] [blame]841 long mask = ((((int)1) << DST_SHIFT) - (long)1); \
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]842 long threshold = _tmp < (int)0 ? (mask >> 1) + (long)1 : (mask >> 1) + 0; \
843 _tmp = (_tmp & mask) > threshold ? (_tmp >> DST_SHIFT) + (int)1 : (_tmp >> DST_SHIFT); \
844 } \
845 _tmp += DST_OFFSET; \
846 dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \
847 }) \
848 }) \
849 })
850
851/** 8-bit per-channel quantization with fixed-point scale
852 *
853 * @param[in] SRC_DATA_TYPE SRC data type
854 * @param[in] DST_DATA_TYPE DST data type
855 * @param[in] M0 Number of src/dst rows
856 * @param[in] N0 Number of src/dst columns
857 * @param[in] DST_OFFSET Quantization offset
858 * @param[in] DST_SHIFT (unused)
859 * @param[in] DST_MULTIPLIER (unused)
860 * @param[in] src Input tile
861 * @param[in] dst_multipliers Output multipliers tile for the per-channel quantization
862 * @param[in] dst_shifts Output shift tile for the per-channel quantization
863 * @param[out] dst Output tile
864 */
865#define T_QUANTIZE8_PER_CHANNEL(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) \
866 ({ \
867 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
868 { \
869 LOOP_UNROLLING(int, _n0, 0, 1, N0, \
870 { \
871 SRC_DATA_TYPE _tmp = 0; \
Pablo Marquez Tello96fb1942022-11-01 09:32:20 +0000[diff] [blame]872 SRC_DATA_TYPE _tmp2 = 0; \
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]873 SRC_DATA_TYPE _src = src[_m0].s[_n0]; \
874 SRC_DATA_TYPE _dst_multiplier = dst_multipliers[0].s[_n0]; \
875 SRC_DATA_TYPE _dst_shift = dst_shifts[0].s[_n0]; \
876 _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-_dst_shift)), ((SRC_DATA_TYPE)_dst_shift < (SRC_DATA_TYPE)0)); \
877 SRC_DATA_TYPE overflow = _src == _dst_multiplier && _src == INT_MIN; \
878 long a_64 = (long)(_src); \
879 long b_64 = (long)(_dst_multiplier); \
880 long ab_64 = a_64 * b_64; \
881 long mask1 = 1 << 30; \
882 long mask2 = 1 - (1 << 30); \
883 long is_positive_or_zero = ab_64 >= 0; \
884 long nudge = select(mask2, mask1, is_positive_or_zero); \
885 SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \
886 _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \
Pablo Marquez Tello96fb1942022-11-01 09:32:20 +0000[diff] [blame]887 long mask = ((((int)1) << _dst_shift) - (int)1); \
888 long threshold = (mask >> 1) + any(_tmp); \
889 _tmp2 = _tmp >> _dst_shift; \
890 _tmp2 += select(0, 1, (_tmp & mask) > threshold); \
891 _tmp = select(_tmp, _tmp2, _dst_shift >= 0); \
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]892 _tmp += DST_OFFSET; \
893 dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \
894 }) \
895 }) \
896 })
897
898/** Quantized the 8-bit tile with fixed-point scale for asymmetric
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]899 *
900 * @param[in] SRC_DATA_TYPE SRC data type
901 * @param[in] DST_DATA_TYPE DST data type
902 * @param[in] M0 Number of src/dst rows
903 * @param[in] N0 Number of src/dst columns
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]904 * @param[in] DST_OFFSET Quantization offset used for both the per-tensor and per-channel quantization
905 * @param[in] DST_SHIFT Quantization shift for the per-tensor quantization
906 * @param[in] DST_MULTIPLIER Quantization multiplier for the per-tensor quantization
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]907 * @param[in] src Input tile
908 * @param[out] dst Output tile
909 */
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]910#define T_QUANTIZE8_ASYMMETRIC(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst) \
911 ({ \
912 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
913 { \
914 LOOP_UNROLLING(int, _n0, 0, 1, N0, \
915 { \
916 SRC_DATA_TYPE _tmp = 0; \
917 SRC_DATA_TYPE _src = src[_m0].s[_n0]; \
918 _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-DST_SHIFT)), ((SRC_DATA_TYPE)DST_SHIFT < (SRC_DATA_TYPE)0)); \
919 SRC_DATA_TYPE overflow = _src == DST_MULTIPLIER && _src == INT_MIN; \
920 long a_64 = (long)(_src); \
921 long b_64 = (long)(DST_MULTIPLIER); \
922 long ab_64 = a_64 * b_64; \
923 long mask1 = 1 << 30; \
924 long mask2 = 1 - (1 << 30); \
925 long is_positive_or_zero = ab_64 >= 0; \
926 long nudge = select(mask2, mask1, is_positive_or_zero); \
927 SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \
928 _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \
929 if(DST_SHIFT >= 0) \
930 { \
931 long mask = ((((int)1) << DST_SHIFT) - (int)1); \
932 long threshold = _tmp < (int)0 ? (mask >> 1) + (long)1 : (mask >> 1) + 0; \
933 _tmp = (_tmp & mask) > threshold ? (_tmp >> DST_SHIFT) + (int)1 : (_tmp >> DST_SHIFT); \
934 } \
935 _tmp += DST_OFFSET; \
936 dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \
937 }) \
938 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]939 })
940
941/** Conditional rowset (memset by row)
942 *
943 * @note Set the row to VALUE_TO_SET if the corresponding mask == 0
944 *
945 * @param[in] DATA_TYPE Data type
946 * @param[in] M0 Number of LHS rows
947 * @param[in] N0 Number of LHS columns
948 * @param[in] VALUE_TO_SET Value to set the row
949 * @param[in, out] a Input/output tile
950 * @param[out] mask Mask to check for setting the row to VALUE_TO_SET
951 */
952#define T_ROWSET_MASK(DATA_TYPE, M0, N0, VALUE_TO_SET, a, mask) \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]953 ({ \
954 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
955 { \
956 LOOP_UNROLLING(int, _n0, 0, 1, N0, \
957 { \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]958 a[_m0].s[_n0] = select((DATA_TYPE)(a[_m0].s[_n0]), (DATA_TYPE)(VALUE_TO_SET), (SELECT_DATA_TYPE(DATA_TYPE))(mask[_m0].v == (DATA_TYPE)0)); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]959 }) \
960 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]961 })
962
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]963/** Element-wise activation for floating point types
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]964 *
965 * @note Performs: activation(LHS) = DST
966 *
967 * @param[in] DATA_TYPE SRC/DST data type
968 * @param[in] M0 Number of SRC/DST rows
969 * @param[in] N0 Number of SRC/DST columns
970 * @param[in] ACTIVATION_TYPE Activation type
971 * @param[in] A_VAL A value used for the activation (e.g. tanh_op, brelu,..)
972 * @param[in] B_VAL B value used for the activation (e.g. tanh_op, brelu,..)
973 * @param[out] src SRC tile
974 * @param[out] dst DST tile
975 */
976#define T_ACTIVATION(DATA_TYPE, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, src, dst) \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]977 ({ \
978 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
979 { \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]980 dst[_m0].v = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, N0, src[_m0].v, A_VAL, B_VAL); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]981 }) \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]982 })
983
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]984// RELU Activation
985#define relu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (max((DATA_TYPE)ZERO_VALUE, x))
986// Bounded RELU Activation
987#define brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (min((DATA_TYPE)A_VAL, max((DATA_TYPE)ZERO_VALUE, x)))
988// Lower Upper Bounded RELU Activation
989#define lu_brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (min(max(x, (DATA_TYPE)B_VAL), (DATA_TYPE)A_VAL))
990// Hard Swish Activation
991#define hard_swish_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (x * ((min(max((DATA_TYPE)(x + (DATA_TYPE)3.f), (DATA_TYPE)0.f), (DATA_TYPE)6.f)) * (DATA_TYPE)0.166666667f))
992// Identity Activation
993#define identity_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (x)
994
995#define ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) op##_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x)
996#define ACTIVATION_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x)
997
Gunes Bayir4bfc70e2021-12-10 16:17:56 +0000[diff] [blame]998#define V_ADD(A_VAL, B_VAL) ((A_VAL) + (B_VAL))
999#define V_DIV(A_VAL, B_VAL) ((A_VAL) / (B_VAL))
Michalis Spyroub1fcefd2022-06-15 19:02:28 +0100[diff] [blame]1000
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]1001/** Element-wise activation for quantized types
1002 *
1003 * @note Performs: activation(LHS) = DST
1004 *
1005 * @param[in] DATA_TYPE SRC/DST data type
1006 * @param[in] M0 Number of SRC/DST rows
1007 * @param[in] N0 Number of SRC/DST columns
1008 * @param[in] ACTIVATION_TYPE Activation type
1009 * @param[in] ZERO_VALUE The zero value to consider in the computation
1010 * @param[in] A_VAL A value used for the activation (e.g. tanh_op, brelu,..)
1011 * @param[in] B_VAL B value used for the activation (e.g. tanh_op, brelu,..)
1012 * @param[out] src SRC tile
1013 * @param[out] dst DST tile
1014 */
1015#define T_ACTIVATION_QUANTIZED(DATA_TYPE, M0, N0, ACTIVATION_TYPE, ZERO_VALUE, A_VAL, B_VAL, src, dst) \
1016 ({ \
1017 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
1018 { \
1019 dst[_m0].v = ACTIVATION_QUANTIZED(ACTIVATION_TYPE, DATA_TYPE, N0, ZERO_VALUE, A_VAL, B_VAL, src[_m0].v); \
1020 }) \
1021 })
1022
Gunes Bayir4bfc70e2021-12-10 16:17:56 +0000[diff] [blame]1023/** Element-wise addition between two tiles
1024 *
1025 * @note Performs: LHS + RHS = DST
1026 *
1027 * @param[in] DATA_TYPE LHS/RHS/DST data type
1028 * @param[in] M0 Number of LHS rows
1029 * @param[in] N0 Number of LHS columns
1030 * @param[in] lhs LHS tile
1031 * @param[in] rhs Constant RHS tile
1032 * @param[out] dst DST tile
1033 */
1034#define T_ADD(DATA_TYPE, M0, N0, lhs, rhs, dst) \
1035 ({ \
1036 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
1037 { \
1038 dst[_m0].v = lhs[_m0].v + rhs[_m0].v; \
1039 }) \
1040 })
1041
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1042/** Element-wise addition with a constant value
1043 *
1044 * @note Performs: LHS + constant = DST
1045 *
1046 * @param[in] DATA_TYPE LHS/RHS/DST data type
1047 * @param[in] M0 Number of LHS rows
1048 * @param[in] N0 Number of LHS columns
1049 * @param[in] lhs LHS tile
1050 * @param[in] rhs_constant Constant value
1051 * @param[out] dst DST tile
1052 */
1053#define T_ADD_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]1054 ({ \
1055 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
1056 { \
Gunes Bayir4bfc70e2021-12-10 16:17:56 +0000[diff] [blame]1057 dst[_m0].v = lhs[_m0].v + (DATA_TYPE)rhs_constant; \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]1058 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1059 })
1060
Gunes Bayir4bfc70e2021-12-10 16:17:56 +0000[diff] [blame]1061#define T_ELTWISE_BROADCAST_ADD_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)
1062#define T_ELTWISE_BROADCAST_DIV_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)
1063
1064/** Element-wise scale with a constant value
1065 *
1066 * @note Performs: LHS * constant = DST
1067 *
1068 * @param[in] DATA_TYPE LHS/RHS/DST data type
1069 * @param[in] M0 Number of LHS rows
1070 * @param[in] N0 Number of LHS columns
1071 * @param[in] lhs LHS tile
1072 * @param[in] rhs_constant Constant value
1073 * @param[out] dst DST tile
1074 */
1075#define T_SCALE_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \
1076 ({ \
1077 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
1078 { \
1079 dst[_m0].v = lhs[_m0].v * (DATA_TYPE)rhs_constant; \
1080 }) \
1081 })
Michalis Spyroub1fcefd2022-06-15 19:02:28 +0100[diff] [blame]1082
1083/** Element-wise operation with RHS broadcasted (RHS has the X dimension only)
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1084 *
Michalis Spyroub1fcefd2022-06-15 19:02:28 +0100[diff] [blame]1085 * @note Performs: LHS OP RHS[broadcasted] = DST
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1086 * @note Both tiles must have same data type
1087 *
Michalis Spyroub1fcefd2022-06-15 19:02:28 +0100[diff] [blame]1088 * @param[in] T_ELWISE_OP Elementwise operator to perform
Giorgio Arena945ae9e2021-10-13 11:13:04 +0100[diff] [blame]1089 * @param[in] DST_DATA_TYPE DST data type
1090 * @param[in] M0 Number of LHS rows
1091 * @param[in] N0 Number of LHS columns
1092 * @param[in] lhs LHS tile
1093 * @param[in] rhs RHS tile
1094 * @param[out] dst DST tile
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1095 */
Michalis Spyroub1fcefd2022-06-15 19:02:28 +0100[diff] [blame]1096#define T_ELTWISE_BROADCAST_X(T_ELWISE_OP, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]1097 ({ \
1098 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
1099 { \
Michalis Spyroub1fcefd2022-06-15 19:02:28 +0100[diff] [blame]1100 dst[_m0].v = T_ELWISE_OP(CONVERT(lhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)), CONVERT(rhs[0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0))); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]1101 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1102 })
1103
Gunes Bayir4bfc70e2021-12-10 16:17:56 +0000[diff] [blame]1104#define T_ELTWISE_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)
1105#define T_ELTWISE_DIV(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst)
Michalis Spyroub1fcefd2022-06-15 19:02:28 +0100[diff] [blame]1106
1107/** Element-wise operation between two tiles (LHS and RHS)
Michalis Spyrou06adbc52022-05-06 17:06:21 +0100[diff] [blame]1108 *
Michalis Spyroub1fcefd2022-06-15 19:02:28 +0100[diff] [blame]1109 * @note Performs: LHS OP RHS = DST
Michalis Spyrou06adbc52022-05-06 17:06:21 +0100[diff] [blame]1110 * @note Both tiles must have same data type
1111 *
Michalis Spyroub1fcefd2022-06-15 19:02:28 +0100[diff] [blame]1112 * @param[in] T_ELWISE_OP Elementwise operator to perform
Michalis Spyrou06adbc52022-05-06 17:06:21 +0100[diff] [blame]1113 * @param[in] DST_DATA_TYPE DST data type
1114 * @param[in] M0 Number of LHS rows
1115 * @param[in] N0 Number of LHS columns
1116 * @param[in] lhs LHS tile
1117 * @param[in] rhs RHS tile
1118 * @param[out] dst DST tile
1119 */
Michalis Spyroub1fcefd2022-06-15 19:02:28 +0100[diff] [blame]1120#define T_ELTWISE(T_ELWISE_OP, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) \
Michalis Spyrou06adbc52022-05-06 17:06:21 +0100[diff] [blame]1121 ({ \
1122 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
1123 { \
Michalis Spyroub1fcefd2022-06-15 19:02:28 +0100[diff] [blame]1124 dst[_m0].v = T_ELWISE_OP(CONVERT(lhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)), CONVERT(rhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0))); \
1125 }) \
1126 })
1127
1128/** Floor operation on a tile
1129 *
1130 * @note Performs: floor(SRC) = DST
1131 * @note Both tiles must have same data type
1132 *
1133 * @param[in] DST_DATA_TYPE DST data type
1134 * @param[in] M0 Number of SRC rows
1135 * @param[in] N0 Number of SRC columns
1136 * @param[in] src LHS tile
1137 * @param[out] dst DST tile
1138 */
1139#define T_FLOOR(DST_DATA_TYPE, M0, N0, src, dst) \
1140 ({ \
1141 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
1142 { \
1143 dst[_m0].v = floor(CONVERT(src[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0))); \
Michalis Spyrou06adbc52022-05-06 17:06:21 +0100[diff] [blame]1144 }) \
1145 })
1146
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1147/** Matrix multiplication
1148 *
1149 * @note Performs: LHS X RHS + DST = DST
1150 *
1151 * @param[in] LHS_DATA_TYPE LHS tile data type
1152 * @param[in] RHS_DATA_TYPE RHS tile data type
1153 * @param[in] DST_DATA_TYPE RHS tile data type
1154 * @param[in] M0 Number of LHS rows
1155 * @param[in] N0 Number of RHS columns
1156 * @param[in] K0 Number of LHS columns
1157 * @param[in] LHS_LAYOUT LHS layout (T= transposed, NT= not transposed)
1158 * @param[in] RHS_LAYOUT RHS layout (T= transposed, NT= not transposed)
1159 * @param[in] lhs LHS tile
1160 * @param[in] rhs RHS tile
1161 * @param[in, out] dst DST tile
1162 */
1163#define T_MMUL(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, LHS_LAYOUT, RHS_LAYOUT, lhs, rhs, dst) T_MMUL_##LHS_LAYOUT##_##RHS_LAYOUT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]1164#define T_MMUL_NT_T(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_##LHS_DATA_TYPE##_##RHS_DATA_TYPE##_##DST_DATA_TYPE(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
1165#define T_MMUL_NT_T_float_float_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
Giorgio Arena945ae9e2021-10-13 11:13:04 +0100[diff] [blame]1166#define T_MMUL_NT_T_half_half_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]1167#define T_MMUL_NT_T_half_half_half(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
1168#define T_MMUL_NT_T_char_char_int(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
1169#define T_MMUL_NT_T_uchar_uchar_uint(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
1170#define T_MMUL_NT_T_uchar_uchar_int(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
1171#define T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1172 { \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]1173 LOOP_UNROLLING(int, _m, 0, 1, M0, \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1174 { \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]1175 LOOP_UNROLLING(int, _n, 0, 1, N0, \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1176 { \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]1177 LOOP_UNROLLING(int, _k, 0, 1, K0, \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1178 { \
Giorgio Arena945ae9e2021-10-13 11:13:04 +0100[diff] [blame]1179 dst[_m].s[_n] = fma((DST_DATA_TYPE)(lhs[_m].s[_k]), (DST_DATA_TYPE)(rhs[_n].s[_k]), dst[_m].s[_n]); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]1180 }) \
1181 }) \
1182 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1183 }
Gian Marco Iodice8155c022021-04-16 15:08:59 +0100[diff] [blame]1184
1185#define T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \
1186 ({ \
1187 LOOP_UNROLLING(int, _m, 0, 1, M0, \
1188 { \
1189 LOOP_UNROLLING(int, _n, 0, 1, N0, \
1190 { \
1191 DOT_PRODUCT_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, K0, (lhs[_m].v), (rhs[_n].v), dst[_m].s[_n]); \
1192 }) \
1193 }) \
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]1194 })
SiCong Lica364df2022-04-13 15:48:19 +0100[diff] [blame]1195
Pablo Marquez Tello96fb1942022-11-01 09:32:20 +0000[diff] [blame]1196#endif // ARM_COMPUTE_TILE_HELPERS_H