Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 4a578b923ed000c67fe0bc1433f945aea634ca9c / . / src / core / CL / cl_kernels / tile_helpers.h
blob: 8f5b5c4a2a5cc696ec3da7ac893190cfc6bbd813 [file] [log] [blame]
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]1/*
2 * Copyright (c) 2021 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
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]25// *INDENT-OFF*
26// clang-format off
27
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]28#define TILE_VECTOR_SIZE1 1
29#define TILE_VECTOR_SIZE2 2
30#define TILE_VECTOR_SIZE3 3
31#define TILE_VECTOR_SIZE4 4
32#define TILE_VECTOR_SIZE5 8
33#define TILE_VECTOR_SIZE6 8
34#define TILE_VECTOR_SIZE7 8
35#define TILE_VECTOR_SIZE8 8
36#define TILE_VECTOR_SIZE9 16
37#define TILE_VECTOR_SIZE10 16
38#define TILE_VECTOR_SIZE11 16
39#define TILE_VECTOR_SIZE12 16
40#define TILE_VECTOR_SIZE13 16
41#define TILE_VECTOR_SIZE14 16
42#define TILE_VECTOR_SIZE15 16
43#define TILE_VECTOR_SIZE16 16
44
45#define TILE_VECTOR_TYPE1(DATA_TYPE) DATA_TYPE##1
46#define TILE_VECTOR_TYPE2(DATA_TYPE) DATA_TYPE##2
47#define TILE_VECTOR_TYPE3(DATA_TYPE) DATA_TYPE##3
48#define TILE_VECTOR_TYPE4(DATA_TYPE) DATA_TYPE##4
49#define TILE_VECTOR_TYPE5(DATA_TYPE) DATA_TYPE##8
50#define TILE_VECTOR_TYPE6(DATA_TYPE) DATA_TYPE##8
51#define TILE_VECTOR_TYPE7(DATA_TYPE) DATA_TYPE##8
52#define TILE_VECTOR_TYPE8(DATA_TYPE) DATA_TYPE##8
53#define TILE_VECTOR_TYPE9(DATA_TYPE) DATA_TYPE##16
54#define TILE_VECTOR_TYPE10(DATA_TYPE) DATA_TYPE##16
55#define TILE_VECTOR_TYPE11(DATA_TYPE) DATA_TYPE##16
56#define TILE_VECTOR_TYPE12(DATA_TYPE) DATA_TYPE##16
57#define TILE_VECTOR_TYPE13(DATA_TYPE) DATA_TYPE##16
58#define TILE_VECTOR_TYPE14(DATA_TYPE) DATA_TYPE##16
59#define TILE_VECTOR_TYPE15(DATA_TYPE) DATA_TYPE##16
60#define TILE_VECTOR_TYPE16(DATA_TYPE) DATA_TYPE##16
61
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]62/** Tile object
63 * A tile object is a 2D memory block and can be accessed using the following syntax:
64 * -# a[m0].v = access the the vector at row "m0" (OpenCL vector)
65 * -# a[m0].s[x] = access the scalar element at row "m0" and column "n0" (scalar access)
66 *
67 * @param[in] DATA_TYPE Data type of the tile
68 * @param[in] H Number of tile rows
69 * @param[in] W Number of tile colums
70 * @param[in] BASENAME Tile's name
71 */
72#define TILE(DATA_TYPE, H, W, BASENAME) TILE_STR(DATA_TYPE, H, W, BASENAME)
73#define TILE_STR(DATA_TYPE, H, W, BASENAME) \
74 union { \
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]75 DATA_TYPE s[TILE_VECTOR_SIZE##W]; \
76 TILE_VECTOR_TYPE##W(DATA_TYPE) v; \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]77 } BASENAME[H]
78
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]79#define TENSOR4D_IMAGE(name) \
80 __read_only image2d_t name##_img, \
81 __global uchar *name##_ptr, \
82 uint name##_stride_x, \
83 uint name##_step_x, \
84 uint name##_stride_y, \
85 uint name##_step_y, \
86 uint name##_stride_z, \
87 uint name##_step_z, \
88 uint name##_stride_w, \
89 uint name##_step_w, \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]90 uint name##_offset_first_element_in_bytes
91
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]92#define TENSOR4D_BUFFER(name) \
93 __global uchar *name##_ptr, \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]94 uint name##_stride_x, \
95 uint name##_step_x, \
96 uint name##_stride_y, \
97 uint name##_step_y, \
98 uint name##_stride_z, \
99 uint name##_step_z, \
100 uint name##_stride_w, \
101 uint name##_step_w, \
102 uint name##_offset_first_element_in_bytes
103
104#define TENSOR4D_STR(name, type) TENSOR4D_##type(name)
105#define TENSOR4D(name, type) TENSOR4D_STR(name, type)
106
Giorgio Arenaea8d2662021-05-20 11:36:56 +0100[diff] [blame]107#if !defined(UNROLL_WITH_PRAGMA)
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]108#define UNROLL_INCR(idx, step, macro) idx += (step); (macro)
109
110#define LOOP_UNROLLING_1(idx, step, macro) (macro)
111#define LOOP_UNROLLING_2(idx, step, macro) LOOP_UNROLLING_1(idx, step, macro); UNROLL_INCR(idx, step, macro)
112#define LOOP_UNROLLING_3(idx, step, macro) LOOP_UNROLLING_2(idx, step, macro); UNROLL_INCR(idx, step, macro)
113#define LOOP_UNROLLING_4(idx, step, macro) LOOP_UNROLLING_3(idx, step, macro); UNROLL_INCR(idx, step, macro)
114#define LOOP_UNROLLING_5(idx, step, macro) LOOP_UNROLLING_4(idx, step, macro); UNROLL_INCR(idx, step, macro)
115#define LOOP_UNROLLING_6(idx, step, macro) LOOP_UNROLLING_5(idx, step, macro); UNROLL_INCR(idx, step, macro)
116#define LOOP_UNROLLING_7(idx, step, macro) LOOP_UNROLLING_6(idx, step, macro); UNROLL_INCR(idx, step, macro)
117#define LOOP_UNROLLING_8(idx, step, macro) LOOP_UNROLLING_7(idx, step, macro); UNROLL_INCR(idx, step, macro)
118#define LOOP_UNROLLING_9(idx, step, macro) LOOP_UNROLLING_8(idx, step, macro); UNROLL_INCR(idx, step, macro)
119#define LOOP_UNROLLING_10(idx, step, macro) LOOP_UNROLLING_9(idx, step, macro); UNROLL_INCR(idx, step, macro)
120#define LOOP_UNROLLING_11(idx, step, macro) LOOP_UNROLLING_10(idx, step, macro); UNROLL_INCR(idx, step, macro)
121#define LOOP_UNROLLING_12(idx, step, macro) LOOP_UNROLLING_11(idx, step, macro); UNROLL_INCR(idx, step, macro)
122#define LOOP_UNROLLING_13(idx, step, macro) LOOP_UNROLLING_12(idx, step, macro); UNROLL_INCR(idx, step, macro)
123#define LOOP_UNROLLING_14(idx, step, macro) LOOP_UNROLLING_13(idx, step, macro); UNROLL_INCR(idx, step, macro)
124#define LOOP_UNROLLING_15(idx, step, macro) LOOP_UNROLLING_14(idx, step, macro); UNROLL_INCR(idx, step, macro)
125#define LOOP_UNROLLING_16(idx, step, macro) LOOP_UNROLLING_15(idx, step, macro); UNROLL_INCR(idx, step, macro)
126#define LOOP_UNROLLING_17(idx, step, macro) LOOP_UNROLLING_16(idx, step, macro); UNROLL_INCR(idx, step, macro)
127#define LOOP_UNROLLING_18(idx, step, macro) LOOP_UNROLLING_17(idx, step, macro); UNROLL_INCR(idx, step, macro)
128#define LOOP_UNROLLING_19(idx, step, macro) LOOP_UNROLLING_18(idx, step, macro); UNROLL_INCR(idx, step, macro)
129#define LOOP_UNROLLING_20(idx, step, macro) LOOP_UNROLLING_19(idx, step, macro); UNROLL_INCR(idx, step, macro)
130#define LOOP_UNROLLING_21(idx, step, macro) LOOP_UNROLLING_20(idx, step, macro); UNROLL_INCR(idx, step, macro)
131#define LOOP_UNROLLING_22(idx, step, macro) LOOP_UNROLLING_21(idx, step, macro); UNROLL_INCR(idx, step, macro)
132#define LOOP_UNROLLING_23(idx, step, macro) LOOP_UNROLLING_22(idx, step, macro); UNROLL_INCR(idx, step, macro)
133#define LOOP_UNROLLING_24(idx, step, macro) LOOP_UNROLLING_23(idx, step, macro); UNROLL_INCR(idx, step, macro)
134#define LOOP_UNROLLING_25(idx, step, macro) LOOP_UNROLLING_24(idx, step, macro); UNROLL_INCR(idx, step, macro)
135#define LOOP_UNROLLING_26(idx, step, macro) LOOP_UNROLLING_25(idx, step, macro); UNROLL_INCR(idx, step, macro)
136#define LOOP_UNROLLING_27(idx, step, macro) LOOP_UNROLLING_26(idx, step, macro); UNROLL_INCR(idx, step, macro)
137#define LOOP_UNROLLING_28(idx, step, macro) LOOP_UNROLLING_27(idx, step, macro); UNROLL_INCR(idx, step, macro)
138#define LOOP_UNROLLING_29(idx, step, macro) LOOP_UNROLLING_28(idx, step, macro); UNROLL_INCR(idx, step, macro)
139#define LOOP_UNROLLING_30(idx, step, macro) LOOP_UNROLLING_29(idx, step, macro); UNROLL_INCR(idx, step, macro)
140#define LOOP_UNROLLING_31(idx, step, macro) LOOP_UNROLLING_30(idx, step, macro); UNROLL_INCR(idx, step, macro)
141#define LOOP_UNROLLING_32(idx, step, macro) LOOP_UNROLLING_31(idx, step, macro); UNROLL_INCR(idx, step, macro)
142#define LOOP_UNROLLING_33(idx, step, macro) LOOP_UNROLLING_32(idx, step, macro); UNROLL_INCR(idx, step, macro)
143#define LOOP_UNROLLING_34(idx, step, macro) LOOP_UNROLLING_33(idx, step, macro); UNROLL_INCR(idx, step, macro)
144#define LOOP_UNROLLING_35(idx, step, macro) LOOP_UNROLLING_34(idx, step, macro); UNROLL_INCR(idx, step, macro)
145#define LOOP_UNROLLING_36(idx, step, macro) LOOP_UNROLLING_35(idx, step, macro); UNROLL_INCR(idx, step, macro)
146#define LOOP_UNROLLING_37(idx, step, macro) LOOP_UNROLLING_36(idx, step, macro); UNROLL_INCR(idx, step, macro)
147#define LOOP_UNROLLING_38(idx, step, macro) LOOP_UNROLLING_37(idx, step, macro); UNROLL_INCR(idx, step, macro)
148#define LOOP_UNROLLING_39(idx, step, macro) LOOP_UNROLLING_38(idx, step, macro); UNROLL_INCR(idx, step, macro)
149#define LOOP_UNROLLING_40(idx, step, macro) LOOP_UNROLLING_39(idx, step, macro); UNROLL_INCR(idx, step, macro)
150#define LOOP_UNROLLING_41(idx, step, macro) LOOP_UNROLLING_40(idx, step, macro); UNROLL_INCR(idx, step, macro)
151#define LOOP_UNROLLING_42(idx, step, macro) LOOP_UNROLLING_41(idx, step, macro); UNROLL_INCR(idx, step, macro)
152#define LOOP_UNROLLING_43(idx, step, macro) LOOP_UNROLLING_42(idx, step, macro); UNROLL_INCR(idx, step, macro)
153#define LOOP_UNROLLING_44(idx, step, macro) LOOP_UNROLLING_43(idx, step, macro); UNROLL_INCR(idx, step, macro)
154#define LOOP_UNROLLING_45(idx, step, macro) LOOP_UNROLLING_44(idx, step, macro); UNROLL_INCR(idx, step, macro)
155#define LOOP_UNROLLING_46(idx, step, macro) LOOP_UNROLLING_45(idx, step, macro); UNROLL_INCR(idx, step, macro)
156#define LOOP_UNROLLING_47(idx, step, macro) LOOP_UNROLLING_46(idx, step, macro); UNROLL_INCR(idx, step, macro)
157#define LOOP_UNROLLING_48(idx, step, macro) LOOP_UNROLLING_47(idx, step, macro); UNROLL_INCR(idx, step, macro)
158#define LOOP_UNROLLING_49(idx, step, macro) LOOP_UNROLLING_48(idx, step, macro); UNROLL_INCR(idx, step, macro)
159#define LOOP_UNROLLING_50(idx, step, macro) LOOP_UNROLLING_49(idx, step, macro); UNROLL_INCR(idx, step, macro)
160#define LOOP_UNROLLING_51(idx, step, macro) LOOP_UNROLLING_50(idx, step, macro); UNROLL_INCR(idx, step, macro)
161#define LOOP_UNROLLING_52(idx, step, macro) LOOP_UNROLLING_51(idx, step, macro); UNROLL_INCR(idx, step, macro)
162#define LOOP_UNROLLING_53(idx, step, macro) LOOP_UNROLLING_52(idx, step, macro); UNROLL_INCR(idx, step, macro)
163#define LOOP_UNROLLING_54(idx, step, macro) LOOP_UNROLLING_53(idx, step, macro); UNROLL_INCR(idx, step, macro)
164#define LOOP_UNROLLING_55(idx, step, macro) LOOP_UNROLLING_54(idx, step, macro); UNROLL_INCR(idx, step, macro)
165#define LOOP_UNROLLING_56(idx, step, macro) LOOP_UNROLLING_55(idx, step, macro); UNROLL_INCR(idx, step, macro)
166#define LOOP_UNROLLING_57(idx, step, macro) LOOP_UNROLLING_56(idx, step, macro); UNROLL_INCR(idx, step, macro)
167#define LOOP_UNROLLING_58(idx, step, macro) LOOP_UNROLLING_57(idx, step, macro); UNROLL_INCR(idx, step, macro)
168#define LOOP_UNROLLING_59(idx, step, macro) LOOP_UNROLLING_58(idx, step, macro); UNROLL_INCR(idx, step, macro)
169#define LOOP_UNROLLING_60(idx, step, macro) LOOP_UNROLLING_59(idx, step, macro); UNROLL_INCR(idx, step, macro)
170#define LOOP_UNROLLING_61(idx, step, macro) LOOP_UNROLLING_60(idx, step, macro); UNROLL_INCR(idx, step, macro)
171#define LOOP_UNROLLING_62(idx, step, macro) LOOP_UNROLLING_61(idx, step, macro); UNROLL_INCR(idx, step, macro)
172#define LOOP_UNROLLING_63(idx, step, macro) LOOP_UNROLLING_62(idx, step, macro); UNROLL_INCR(idx, step, macro)
173#define LOOP_UNROLLING_64(idx, step, macro) LOOP_UNROLLING_63(idx, step, macro); UNROLL_INCR(idx, step, macro)
174#define LOOP_UNROLLING_65(idx, step, macro) LOOP_UNROLLING_64(idx, step, macro); UNROLL_INCR(idx, step, macro)
175#define LOOP_UNROLLING_66(idx, step, macro) LOOP_UNROLLING_65(idx, step, macro); UNROLL_INCR(idx, step, macro)
176#define LOOP_UNROLLING_67(idx, step, macro) LOOP_UNROLLING_66(idx, step, macro); UNROLL_INCR(idx, step, macro)
177#define LOOP_UNROLLING_68(idx, step, macro) LOOP_UNROLLING_67(idx, step, macro); UNROLL_INCR(idx, step, macro)
178#define LOOP_UNROLLING_69(idx, step, macro) LOOP_UNROLLING_68(idx, step, macro); UNROLL_INCR(idx, step, macro)
179#define LOOP_UNROLLING_70(idx, step, macro) LOOP_UNROLLING_69(idx, step, macro); UNROLL_INCR(idx, step, macro)
180#define LOOP_UNROLLING_71(idx, step, macro) LOOP_UNROLLING_70(idx, step, macro); UNROLL_INCR(idx, step, macro)
181#define LOOP_UNROLLING_72(idx, step, macro) LOOP_UNROLLING_71(idx, step, macro); UNROLL_INCR(idx, step, macro)
182#define LOOP_UNROLLING_73(idx, step, macro) LOOP_UNROLLING_72(idx, step, macro); UNROLL_INCR(idx, step, macro)
183#define LOOP_UNROLLING_74(idx, step, macro) LOOP_UNROLLING_73(idx, step, macro); UNROLL_INCR(idx, step, macro)
184#define LOOP_UNROLLING_75(idx, step, macro) LOOP_UNROLLING_74(idx, step, macro); UNROLL_INCR(idx, step, macro)
185#define LOOP_UNROLLING_76(idx, step, macro) LOOP_UNROLLING_75(idx, step, macro); UNROLL_INCR(idx, step, macro)
186#define LOOP_UNROLLING_77(idx, step, macro) LOOP_UNROLLING_76(idx, step, macro); UNROLL_INCR(idx, step, macro)
187#define LOOP_UNROLLING_78(idx, step, macro) LOOP_UNROLLING_77(idx, step, macro); UNROLL_INCR(idx, step, macro)
188#define LOOP_UNROLLING_79(idx, step, macro) LOOP_UNROLLING_78(idx, step, macro); UNROLL_INCR(idx, step, macro)
189#define LOOP_UNROLLING_80(idx, step, macro) LOOP_UNROLLING_79(idx, step, macro); UNROLL_INCR(idx, step, macro)
190#define LOOP_UNROLLING_81(idx, step, macro) LOOP_UNROLLING_80(idx, step, macro); UNROLL_INCR(idx, step, macro)
191#define LOOP_UNROLLING_82(idx, step, macro) LOOP_UNROLLING_81(idx, step, macro); UNROLL_INCR(idx, step, macro)
192#define LOOP_UNROLLING_83(idx, step, macro) LOOP_UNROLLING_82(idx, step, macro); UNROLL_INCR(idx, step, macro)
193#define LOOP_UNROLLING_84(idx, step, macro) LOOP_UNROLLING_83(idx, step, macro); UNROLL_INCR(idx, step, macro)
194#define LOOP_UNROLLING_85(idx, step, macro) LOOP_UNROLLING_84(idx, step, macro); UNROLL_INCR(idx, step, macro)
195#define LOOP_UNROLLING_86(idx, step, macro) LOOP_UNROLLING_85(idx, step, macro); UNROLL_INCR(idx, step, macro)
196#define LOOP_UNROLLING_87(idx, step, macro) LOOP_UNROLLING_86(idx, step, macro); UNROLL_INCR(idx, step, macro)
197#define LOOP_UNROLLING_88(idx, step, macro) LOOP_UNROLLING_87(idx, step, macro); UNROLL_INCR(idx, step, macro)
198#define LOOP_UNROLLING_89(idx, step, macro) LOOP_UNROLLING_88(idx, step, macro); UNROLL_INCR(idx, step, macro)
199#define LOOP_UNROLLING_90(idx, step, macro) LOOP_UNROLLING_89(idx, step, macro); UNROLL_INCR(idx, step, macro)
200#define LOOP_UNROLLING_91(idx, step, macro) LOOP_UNROLLING_90(idx, step, macro); UNROLL_INCR(idx, step, macro)
201#define LOOP_UNROLLING_92(idx, step, macro) LOOP_UNROLLING_91(idx, step, macro); UNROLL_INCR(idx, step, macro)
202#define LOOP_UNROLLING_93(idx, step, macro) LOOP_UNROLLING_92(idx, step, macro); UNROLL_INCR(idx, step, macro)
203#define LOOP_UNROLLING_94(idx, step, macro) LOOP_UNROLLING_93(idx, step, macro); UNROLL_INCR(idx, step, macro)
204#define LOOP_UNROLLING_95(idx, step, macro) LOOP_UNROLLING_94(idx, step, macro); UNROLL_INCR(idx, step, macro)
205#define LOOP_UNROLLING_96(idx, step, macro) LOOP_UNROLLING_95(idx, step, macro); UNROLL_INCR(idx, step, macro)
206#define LOOP_UNROLLING_97(idx, step, macro) LOOP_UNROLLING_96(idx, step, macro); UNROLL_INCR(idx, step, macro)
207#define LOOP_UNROLLING_98(idx, step, macro) LOOP_UNROLLING_97(idx, step, macro); UNROLL_INCR(idx, step, macro)
208#define LOOP_UNROLLING_99(idx, step, macro) LOOP_UNROLLING_98(idx, step, macro); UNROLL_INCR(idx, step, macro)
209#define LOOP_UNROLLING_100(idx, step, macro) LOOP_UNROLLING_99(idx, step, macro); UNROLL_INCR(idx, step, macro)
210#define LOOP_UNROLLING_101(idx, step, macro) LOOP_UNROLLING_100(idx, step, macro); UNROLL_INCR(idx, step, macro)
211#define LOOP_UNROLLING_102(idx, step, macro) LOOP_UNROLLING_101(idx, step, macro); UNROLL_INCR(idx, step, macro)
212#define LOOP_UNROLLING_103(idx, step, macro) LOOP_UNROLLING_102(idx, step, macro); UNROLL_INCR(idx, step, macro)
213#define LOOP_UNROLLING_104(idx, step, macro) LOOP_UNROLLING_103(idx, step, macro); UNROLL_INCR(idx, step, macro)
214#define LOOP_UNROLLING_105(idx, step, macro) LOOP_UNROLLING_104(idx, step, macro); UNROLL_INCR(idx, step, macro)
215#define LOOP_UNROLLING_106(idx, step, macro) LOOP_UNROLLING_105(idx, step, macro); UNROLL_INCR(idx, step, macro)
216#define LOOP_UNROLLING_107(idx, step, macro) LOOP_UNROLLING_106(idx, step, macro); UNROLL_INCR(idx, step, macro)
217#define LOOP_UNROLLING_108(idx, step, macro) LOOP_UNROLLING_107(idx, step, macro); UNROLL_INCR(idx, step, macro)
218#define LOOP_UNROLLING_109(idx, step, macro) LOOP_UNROLLING_108(idx, step, macro); UNROLL_INCR(idx, step, macro)
219#define LOOP_UNROLLING_110(idx, step, macro) LOOP_UNROLLING_109(idx, step, macro); UNROLL_INCR(idx, step, macro)
220#define LOOP_UNROLLING_111(idx, step, macro) LOOP_UNROLLING_110(idx, step, macro); UNROLL_INCR(idx, step, macro)
221#define LOOP_UNROLLING_112(idx, step, macro) LOOP_UNROLLING_111(idx, step, macro); UNROLL_INCR(idx, step, macro)
222#define LOOP_UNROLLING_113(idx, step, macro) LOOP_UNROLLING_112(idx, step, macro); UNROLL_INCR(idx, step, macro)
223#define LOOP_UNROLLING_114(idx, step, macro) LOOP_UNROLLING_113(idx, step, macro); UNROLL_INCR(idx, step, macro)
224#define LOOP_UNROLLING_115(idx, step, macro) LOOP_UNROLLING_114(idx, step, macro); UNROLL_INCR(idx, step, macro)
225#define LOOP_UNROLLING_116(idx, step, macro) LOOP_UNROLLING_115(idx, step, macro); UNROLL_INCR(idx, step, macro)
226#define LOOP_UNROLLING_117(idx, step, macro) LOOP_UNROLLING_116(idx, step, macro); UNROLL_INCR(idx, step, macro)
227#define LOOP_UNROLLING_118(idx, step, macro) LOOP_UNROLLING_117(idx, step, macro); UNROLL_INCR(idx, step, macro)
228#define LOOP_UNROLLING_119(idx, step, macro) LOOP_UNROLLING_118(idx, step, macro); UNROLL_INCR(idx, step, macro)
229#define LOOP_UNROLLING_120(idx, step, macro) LOOP_UNROLLING_119(idx, step, macro); UNROLL_INCR(idx, step, macro)
230#define LOOP_UNROLLING_121(idx, step, macro) LOOP_UNROLLING_120(idx, step, macro); UNROLL_INCR(idx, step, macro)
231#define LOOP_UNROLLING_122(idx, step, macro) LOOP_UNROLLING_121(idx, step, macro); UNROLL_INCR(idx, step, macro)
232#define LOOP_UNROLLING_123(idx, step, macro) LOOP_UNROLLING_122(idx, step, macro); UNROLL_INCR(idx, step, macro)
233#define LOOP_UNROLLING_124(idx, step, macro) LOOP_UNROLLING_123(idx, step, macro); UNROLL_INCR(idx, step, macro)
234#define LOOP_UNROLLING_125(idx, step, macro) LOOP_UNROLLING_124(idx, step, macro); UNROLL_INCR(idx, step, macro)
235#define LOOP_UNROLLING_126(idx, step, macro) LOOP_UNROLLING_125(idx, step, macro); UNROLL_INCR(idx, step, macro)
236#define LOOP_UNROLLING_127(idx, step, macro) LOOP_UNROLLING_126(idx, step, macro); UNROLL_INCR(idx, step, macro)
237#define LOOP_UNROLLING_128(idx, step, macro) LOOP_UNROLLING_127(idx, step, macro); UNROLL_INCR(idx, step, macro)
238
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]239#define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \
240 { \
241 type idx = start; \
242 LOOP_UNROLLING_##num(idx, step, macro); \
243 }
Giorgio Arenaea8d2662021-05-20 11:36:56 +0100[diff] [blame]244#else // !defined(UNROLL_WITH_PRAGMA)
245#define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \
246 { \
247 _Pragma("unroll") \
248 for(type idx = start; idx < (num * step); idx += step) \
249 { \
250 (macro); \
251 } \
252 }
253#endif // !defined(UNROLL_WITH_PRAGMA)
254#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]255
256/** 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
257 * to avoid out-of-bound read/write
258 *
259 * @note PARTIAL_N0 is used for get_global_id(n) = 0.
260 *
261 * @param[in] IDX get_global_id index (0,1 and 2 only)
262 * @param[in] N0 Number of elements read/written on the IDX direction
263 * @param[in] PARTIAL_N0 Number of elements read/written on the IDX direction for get_global_id(IDX) = 0. If zero,
264 * the Number of elements read/written on the IDX direction for get_global_id(IDX) = 0 is N0
265 */
266#define GET_SPATIAL_IDX(IDX, N0, PARTIAL_N0) (max((int)(get_global_id(IDX) * N0 - (N0 - PARTIAL_N0) % N0), 0))
267
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]268/** Dot product integet 8bit function
269 *
270 * @note Performs: c += dot(a, b)
271 *
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]272 * @param[in] A_DATA_TYPE A (lhs) data type
273 * @param[in] B_DATA_TYPE B (rhs) data type
274 * @param[in] C_DATA_TYPE C (accumulator) data type
275 * @param[in] K0 Number of accumulations
276 * @param[in] a OpenCL vector a
277 * @param[in] b OpenCL vector b
278 * @param[in] c Scalar variable c
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]279 */
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]280#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)
281#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)
282#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]283 ({ \
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]284 c += (C_DATA_TYPE)(a) * (C_DATA_TYPE)(b); \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]285 })
286#if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]287#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));
288#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));
289#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]290#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)
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]291#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));
292#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);
293#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]294#else // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]295#define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
296 ({ \
297 c += (C_DATA_TYPE)(a).s0 * (C_DATA_TYPE)(b).s0; \
298 c += (C_DATA_TYPE)(a).s1 * (C_DATA_TYPE)(b).s1; \
299 })
300#define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
301 ({ \
302 DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c); \
303 c += (C_DATA_TYPE)(a).s2 * (C_DATA_TYPE)(b).s2; \
304 })
305#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]306 ({ \
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]307 val += (C_DATA_TYPE)(x).s0 * (C_DATA_TYPE)(y).s0; \
308 val += (C_DATA_TYPE)(x).s1 * (C_DATA_TYPE)(y).s1; \
309 val += (C_DATA_TYPE)(x).s2 * (C_DATA_TYPE)(y).s2; \
310 val += (C_DATA_TYPE)(x).s3 * (C_DATA_TYPE)(y).s3; \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]311 })
312#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]313#define DOT_PRODUCT5_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
314 ({ \
315 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \
316 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]317 })
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]318#define DOT_PRODUCT6_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
319 ({ \
320 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \
321 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]322 })
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]323#define DOT_PRODUCT7_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
324 ({ \
325 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \
326 DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s456), ((b).s456), c); \
327 })
328#define DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
329 ({ \
330 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).lo), ((b).lo), c); \
331 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).hi), ((b).hi), c); \
332 })
333#define DOT_PRODUCT16_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \
334 ({ \
335 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).lo), ((b).lo), c); \
336 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).hi), ((b).hi), c); \
337 })
338
339/** Dot product integet 8bit function
340 *
341 * @note Performs: c += dot(a, b)
342 *
343 * @param[in] A_DATA_TYPE A (lhs) data type
344 * @param[in] B_DATA_TYPE B (rhs) data type
345 * @param[in] C_DATA_TYPE C (accumulator) data type
346 * @param[in] K0 Number of accumulations
347 * @param[in] a OpenCL vector a
348 * @param[in] c Scalar variable c
349 */
350#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)
351#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]352
353/** Load a vector from global memory (tensor)
354 *
355 * @param[in] DATA_TYPE Data type
356 * @param[in] WIDTH Number of dst columns
357 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image).
358 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
359 * @param[in] TENSOR Tensor basename
360 * @param[in] X Starting X position
361 * @param[in] Y Starting Y position
362 * @param[in] STRIDE_Y Stride Y (in bytes)
363 */
364#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)
365#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)
366#define V_LOAD_BUFFER(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) \
367 VLOAD(WIDTH) \
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]368 (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]369#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))
370
371/** Load a tile from global memory (tensor)
372 *
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]373 * @param[in] DATA_TYPE Data type
374 * @param[in] HEIGHT Number of dst rows
375 * @param[in] WIDTH Number of dst columns
376 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image).
377 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
378 * @param[in] TENSOR Tensor basename
379 * @param[in] X Starting X position
380 * @param[in] Y Starting Y position
381 * @param[in] YI_MULTIPLIER Parameter used to multiply the internal row increment (_i).
382 * 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).
383 * In this case the address calculation is performed as: (Y + _i * Y_MULTIPLIER) * STRIDE_Y
384 * @param[in] STRIDE_Y Stride Y (in bytes) used to load each row.
385 * @param[out] dst Output tile
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]386 */
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]387#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]388 ({ \
389 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \
390 { \
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]391 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]392 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]393 })
394
395/** Load a tile from global memory (tensor) using an indirect Y index tile
396 *
397 * @param[in] DATA_TYPE Data type
398 * @param[in] HEIGHT Number of dst rows
399 * @param[in] WIDTH Number of dst columns
400 * @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
401 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
402 * @param[in] TENSOR Tensor basename
403 * @param[in] X Starting X position
404 * @param[in] STRIDE_Y Stride Y (in bytes)
405 * @param[in] indirect_y Indirect Y index tile
406 * @param[out] dst Output tile
407 */
408#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]409 ({ \
410 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \
411 { \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]412 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]413 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]414 })
415
Gian Marco Iodice534b8892021-04-01 16:17:16 +0100[diff] [blame]416/** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout
417 *
418 * @param[in] DATA_TYPE Data type
419 * @param[in] TILE_HEIGHT Number of elements to load from Y (height) dimension
420 * @param[in] TILE_WIDTH Number of elements to load from X (width) dimension
421 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension
422 * @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
423 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16)
424 * @param[in] TENSOR Tensor basename
425 * @param[in] B Starting batch index
426 * @param[in] Y Starting Y index
427 * @param[in] X Starting X index
428 * @param[in] C Starting C index
429 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension
430 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension
431 * @param[in] STRIDE_Y Stride Y (in bytes)
432 * @param[out] dst Output tile
433 */
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]434#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]435 ({ \
436 LOOP_UNROLLING(int, _yk, 0, 1, TILE_HEIGHT, \
437 { \
438 LOOP_UNROLLING(int, _xk, 0, 1, TILE_WIDTH, \
439 { \
440 int _src_y = (X) + _xk + ((Y) + _yk) * (TENSOR_WIDTH); \
441 _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT); \
Gian Marco Iodice534b8892021-04-01 16:17:16 +0100[diff] [blame]442 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]443 if(_src_valid_y != 0) \
444 { \
445 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]446 } \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]447 }) \
448 }) \
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]449 })
450
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]451/** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout with dilation for the X and Y increments
452 *
453 * @param[in] DATA_TYPE Data type
454 * @param[in] TILE_HEIGHT Number of elements to load from Y (height) dimension
455 * @param[in] TILE_WIDTH Number of elements to load from X (width) dimension
456 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension
457 * @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
458 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16)
459 * @param[in] TENSOR Tensor basename
460 * @param[in] B Starting batch index
461 * @param[in] Y Starting Y index
462 * @param[in] X Starting X index
463 * @param[in] C Starting C index
464 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension
465 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension
466 * @param[in] DILATION_X Dilation for the X increment
467 * @param[in] DILATION_Y Dilation for the Y increment
468 * @param[in] STRIDE_Y Stride Y (in bytes)
469 * @param[in] BOUNDARY_CHECK Boundary check flag. If true, it checks for any out-of-bound reads
470 * @param[out] dst Output tile
471 */
472#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, STRIDE_Y, BOUNDARY_CHECK, dst) \
473 ({ \
474 LOOP_UNROLLING(int, _yk, 0, 1, TILE_HEIGHT, \
475 { \
476 LOOP_UNROLLING(int, _xk, 0, 1, TILE_WIDTH, \
477 { \
478 int _src_y = (X) + _xk * (DILATION_X) + ((Y) + _yk * (DILATION_Y)) * (TENSOR_WIDTH); \
479 _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT); \
480 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)); \
481 if(!(BOUNDARY_CHECK)) \
482 { \
483 dst[_xk + _yk * (TILE_WIDTH)].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \
484 } \
485 else \
486 { \
487 if(_src_valid_y) \
488 { \
489 dst[_xk + _yk * (TILE_WIDTH)].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \
490 } \
491 } \
492 }) \
493 }) \
494 })
495
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]496/** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout using indirect X and Y coordinates
497 *
498 * @param[in] DATA_TYPE Data type
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]499 * @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]500 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension
501 * @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
502 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16)
503 * @param[in] TENSOR Tensor basename
504 * @param[in] B Starting batch index
505 * @param[in] Y Starting Y index
506 * @param[in] X Starting X index
507 * @param[in] C Starting C index
508 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension
509 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension
510 * @param[in] STRIDE_Y Stride Y (in bytes)
511 * @param[out] xi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect X coordinate
512 * @param[out] yi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect Y coordinate
513 * @param[out] dst Output tile
514 */
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]515#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) \
516 ({ \
517 LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \
518 { \
519 int _src_y = (X) + xi[_i].v + ((Y) + yi[_i].v) * (TENSOR_WIDTH); \
520 _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT); \
Gian Marco Iodice0b76f7d2021-04-08 17:20:00 +0100[diff] [blame]521 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]522 if(_src_valid_y != 0) \
523 { \
524 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]525 } \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]526 }) \
Gian Marco Iodice534b8892021-04-01 16:17:16 +0100[diff] [blame]527 })
528
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]529/** Store a tile to global memory (tensor) using an indirect Y index tile and conditionally use a different length for the store
530 *
531 * @note If WIDTH1_CONDITION is true, the store will use the WIDTH1 length for the store
532 * @note The vectors are stored in reverse order so the invalid rows are overwritten by the valid ones
533 *
534 * @param[in] DATA_TYPE Data type
535 * @param[in] HEIGHT Number of src rows
536 * @param[in] WIDTH0 Store width to use if WIDTH1_CONDITION = false
537 * @param[in] WIDTH1 Store width to use if WIDTH1_CONDITION = true
538 * @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
539 * cl_image is not supported.
540 * @param[in] TENSOR Tensor basename
541 * @param[in] X Starting X position
542 * @param[in] STRIDE_Y Stride Y (in bytes)
543 * @param[in] WIDTH1_CONDITION Condition to select the WIDTH1 store
544 * @param[in] src Input tile
545 * @param[in] indirect_y Indirect Y index tile
546 */
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]547#define T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, STRIDE_Y, WIDTH1_CONDITION, src, indirect_y) \
548 ({ \
549 if(WIDTH1_CONDITION) \
550 { \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]551 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]552 { \
553 VSTORE_PARTIAL(WIDTH0, WIDTH1) \
554 (src[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)); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]555 }) \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]556 } \
557 else \
558 { \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]559 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]560 { \
561 VSTORE(WIDTH0) \
562 (src[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)); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]563 }) \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]564 } \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]565 })
566
567/** Offset correction for the QASYMM8 computation
568 *
569 * @param[in] ACC_DATA_TYPE Accumulator data type
570 * @param[in] M0 Number of src/dst rows
571 * @param[in] N0 Number of src/dst columns
572 * @param[in] K0 Number of src columns
573 * @param[in] SRC_OFFSET Source quantization offset
574 * @param[in] WEI_OFFSET Weights quantization shift
575 * @param[in] lhs LHS tile
576 * @param[in] rhs RHS tile
577 * @param[out] dst DST tile
578 */
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]579#define T_OFFSET_CORRECTION(ACC_DATA_TYPE, M0, N0, K0, SRC_OFFSET, WEI_OFFSET, lhs, rhs, dst) \
580 ({ \
581 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
582 { \
583 ACC_DATA_TYPE _tm = 0; \
584 LOOP_UNROLLING(int, _k0, 0, 1, K0, \
585 { \
586 _tm += ((ACC_DATA_TYPE)lhs[_m0].s[_k0] * (ACC_DATA_TYPE)WEI_OFFSET); \
587 }) \
588 LOOP_UNROLLING(int, _n0, 0, 1, N0, \
589 { \
590 dst[_m0].s[_n0] += _tm; \
591 LOOP_UNROLLING(int, _k0, 0, 1, K0, \
592 { \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]593 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]594 }) \
595 }) \
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]596 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]597 })
598
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]599/** 8-bit quantization with fixed-point scale
600 *
601 * @param[in] SRC_DATA_TYPE SRC data type
602 * @param[in] DST_DATA_TYPE DST data type
603 * @param[in] QUANTIZATION_TYPE Quantization type (PER_TENSOR or PER_CHANNEL)
604 * @param[in] M0 Number of src/dst rows
605 * @param[in] N0 Number of src/dst columns
606 * @param[in] DST_OFFSET Quantization offset used for both the per-tensor and per-channel quantization
607 * @param[in] DST_SHIFT Quantization shift for the per-tensor quantization
608 * @param[in] DST_MULTIPLIER Quantization multiplier for the per-tensor quantization
609 * @param[in] src Input tile
610 * @param[in] dst_multipliers Output multipliers tile for the per-channel quantization
611 * @param[in] dst_shifts Output shift tile for the per-channel quantization
612 * @param[out] dst Output tile
613 */
614#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)
615#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)
616
617/** 8-bit per-tensor quantization with fixed-point scale
618 *
619 * @param[in] SRC_DATA_TYPE SRC data type
620 * @param[in] DST_DATA_TYPE DST data type
621 * @param[in] M0 Number of src/dst rows
622 * @param[in] N0 Number of src/dst columns
623 * @param[in] DST_OFFSET Quantization offset
624 * @param[in] DST_SHIFT Quantization shift for the per-tensor quantization
625 * @param[in] DST_MULTIPLIER Quantization multiplier for the per-tensor quantization
626 * @param[in] src Input tile
627 * @param[in] dst_multipliers (unused)
628 * @param[in] dst_shifts (unused)
629 * @param[out] dst Output tile
630 */
631#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) \
632 ({ \
633 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
634 { \
635 LOOP_UNROLLING(int, _n0, 0, 1, N0, \
636 { \
637 SRC_DATA_TYPE _tmp = 0; \
638 SRC_DATA_TYPE _src = src[_m0].s[_n0]; \
639 _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-DST_SHIFT)), ((SRC_DATA_TYPE)DST_SHIFT < (SRC_DATA_TYPE)0)); \
640 SRC_DATA_TYPE overflow = _src == DST_MULTIPLIER && _src == INT_MIN; \
641 long a_64 = (long)(_src); \
642 long b_64 = (long)(DST_MULTIPLIER); \
643 long ab_64 = a_64 * b_64; \
644 long mask1 = 1 << 30; \
645 long mask2 = 1 - (1 << 30); \
646 long is_positive_or_zero = ab_64 >= 0; \
647 long nudge = select(mask2, mask1, is_positive_or_zero); \
648 SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \
649 _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \
650 if(DST_SHIFT >= 0) \
651 { \
652 long mask = ((((int)1) << DST_SHIFT) - (int)1); \
653 long threshold = _tmp < (int)0 ? (mask >> 1) + (long)1 : (mask >> 1) + 0; \
654 _tmp = (_tmp & mask) > threshold ? (_tmp >> DST_SHIFT) + (int)1 : (_tmp >> DST_SHIFT); \
655 } \
656 _tmp += DST_OFFSET; \
657 dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \
658 }) \
659 }) \
660 })
661
662/** 8-bit per-channel quantization with fixed-point scale
663 *
664 * @param[in] SRC_DATA_TYPE SRC data type
665 * @param[in] DST_DATA_TYPE DST data type
666 * @param[in] M0 Number of src/dst rows
667 * @param[in] N0 Number of src/dst columns
668 * @param[in] DST_OFFSET Quantization offset
669 * @param[in] DST_SHIFT (unused)
670 * @param[in] DST_MULTIPLIER (unused)
671 * @param[in] src Input tile
672 * @param[in] dst_multipliers Output multipliers tile for the per-channel quantization
673 * @param[in] dst_shifts Output shift tile for the per-channel quantization
674 * @param[out] dst Output tile
675 */
676#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) \
677 ({ \
678 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
679 { \
680 LOOP_UNROLLING(int, _n0, 0, 1, N0, \
681 { \
682 SRC_DATA_TYPE _tmp = 0; \
683 SRC_DATA_TYPE _src = src[_m0].s[_n0]; \
684 SRC_DATA_TYPE _dst_multiplier = dst_multipliers[0].s[_n0]; \
685 SRC_DATA_TYPE _dst_shift = dst_shifts[0].s[_n0]; \
686 _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-_dst_shift)), ((SRC_DATA_TYPE)_dst_shift < (SRC_DATA_TYPE)0)); \
687 SRC_DATA_TYPE overflow = _src == _dst_multiplier && _src == INT_MIN; \
688 long a_64 = (long)(_src); \
689 long b_64 = (long)(_dst_multiplier); \
690 long ab_64 = a_64 * b_64; \
691 long mask1 = 1 << 30; \
692 long mask2 = 1 - (1 << 30); \
693 long is_positive_or_zero = ab_64 >= 0; \
694 long nudge = select(mask2, mask1, is_positive_or_zero); \
695 SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \
696 _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \
697 if(_dst_shift >= 0) \
698 { \
699 long mask = ((((int)1) << _dst_shift) - (int)1); \
700 long threshold = _tmp < (int)0 ? (mask >> 1) + (long)1 : (mask >> 1) + 0; \
701 _tmp = (_tmp & mask) > threshold ? (_tmp >> _dst_shift) + (int)1 : (_tmp >> _dst_shift); \
702 } \
703 _tmp += DST_OFFSET; \
704 dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \
705 }) \
706 }) \
707 })
708
709/** Quantized the 8-bit tile with fixed-point scale for asymmetric
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]710 *
711 * @param[in] SRC_DATA_TYPE SRC data type
712 * @param[in] DST_DATA_TYPE DST data type
713 * @param[in] M0 Number of src/dst rows
714 * @param[in] N0 Number of src/dst columns
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]715 * @param[in] DST_OFFSET Quantization offset used for both the per-tensor and per-channel quantization
716 * @param[in] DST_SHIFT Quantization shift for the per-tensor quantization
717 * @param[in] DST_MULTIPLIER Quantization multiplier for the per-tensor quantization
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]718 * @param[in] src Input tile
719 * @param[out] dst Output tile
720 */
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]721#define T_QUANTIZE8_ASYMMETRIC(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst) \
722 ({ \
723 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
724 { \
725 LOOP_UNROLLING(int, _n0, 0, 1, N0, \
726 { \
727 SRC_DATA_TYPE _tmp = 0; \
728 SRC_DATA_TYPE _src = src[_m0].s[_n0]; \
729 _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-DST_SHIFT)), ((SRC_DATA_TYPE)DST_SHIFT < (SRC_DATA_TYPE)0)); \
730 SRC_DATA_TYPE overflow = _src == DST_MULTIPLIER && _src == INT_MIN; \
731 long a_64 = (long)(_src); \
732 long b_64 = (long)(DST_MULTIPLIER); \
733 long ab_64 = a_64 * b_64; \
734 long mask1 = 1 << 30; \
735 long mask2 = 1 - (1 << 30); \
736 long is_positive_or_zero = ab_64 >= 0; \
737 long nudge = select(mask2, mask1, is_positive_or_zero); \
738 SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \
739 _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \
740 if(DST_SHIFT >= 0) \
741 { \
742 long mask = ((((int)1) << DST_SHIFT) - (int)1); \
743 long threshold = _tmp < (int)0 ? (mask >> 1) + (long)1 : (mask >> 1) + 0; \
744 _tmp = (_tmp & mask) > threshold ? (_tmp >> DST_SHIFT) + (int)1 : (_tmp >> DST_SHIFT); \
745 } \
746 _tmp += DST_OFFSET; \
747 dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \
748 }) \
749 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]750 })
751
752/** Conditional rowset (memset by row)
753 *
754 * @note Set the row to VALUE_TO_SET if the corresponding mask == 0
755 *
756 * @param[in] DATA_TYPE Data type
757 * @param[in] M0 Number of LHS rows
758 * @param[in] N0 Number of LHS columns
759 * @param[in] VALUE_TO_SET Value to set the row
760 * @param[in, out] a Input/output tile
761 * @param[out] mask Mask to check for setting the row to VALUE_TO_SET
762 */
763#define T_ROWSET_MASK(DATA_TYPE, M0, N0, VALUE_TO_SET, a, mask) \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]764 ({ \
765 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
766 { \
767 LOOP_UNROLLING(int, _n0, 0, 1, N0, \
768 { \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]769 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]770 }) \
771 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]772 })
773
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]774/** Element-wise activation for floating point types
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]775 *
776 * @note Performs: activation(LHS) = DST
777 *
778 * @param[in] DATA_TYPE SRC/DST data type
779 * @param[in] M0 Number of SRC/DST rows
780 * @param[in] N0 Number of SRC/DST columns
781 * @param[in] ACTIVATION_TYPE Activation type
782 * @param[in] A_VAL A value used for the activation (e.g. tanh_op, brelu,..)
783 * @param[in] B_VAL B value used for the activation (e.g. tanh_op, brelu,..)
784 * @param[out] src SRC tile
785 * @param[out] dst DST tile
786 */
787#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]788 ({ \
789 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
790 { \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]791 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]792 }) \
Gian Marco Iodicea8903c82021-03-24 14:48:22 +0000[diff] [blame]793 })
794
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]795// RELU Activation
796#define relu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (max((DATA_TYPE)ZERO_VALUE, x))
797// Bounded RELU Activation
798#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)))
799// Lower Upper Bounded RELU Activation
800#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))
801// Hard Swish Activation
802#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))
803// Identity Activation
804#define identity_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (x)
805
806#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)
807#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)
808
809/** Element-wise activation for quantized types
810 *
811 * @note Performs: activation(LHS) = DST
812 *
813 * @param[in] DATA_TYPE SRC/DST data type
814 * @param[in] M0 Number of SRC/DST rows
815 * @param[in] N0 Number of SRC/DST columns
816 * @param[in] ACTIVATION_TYPE Activation type
817 * @param[in] ZERO_VALUE The zero value to consider in the computation
818 * @param[in] A_VAL A value used for the activation (e.g. tanh_op, brelu,..)
819 * @param[in] B_VAL B value used for the activation (e.g. tanh_op, brelu,..)
820 * @param[out] src SRC tile
821 * @param[out] dst DST tile
822 */
823#define T_ACTIVATION_QUANTIZED(DATA_TYPE, M0, N0, ACTIVATION_TYPE, ZERO_VALUE, A_VAL, B_VAL, src, dst) \
824 ({ \
825 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
826 { \
827 dst[_m0].v = ACTIVATION_QUANTIZED(ACTIVATION_TYPE, DATA_TYPE, N0, ZERO_VALUE, A_VAL, B_VAL, src[_m0].v); \
828 }) \
829 })
830
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]831/** Element-wise addition with a constant value
832 *
833 * @note Performs: LHS + constant = DST
834 *
835 * @param[in] DATA_TYPE LHS/RHS/DST data type
836 * @param[in] M0 Number of LHS rows
837 * @param[in] N0 Number of LHS columns
838 * @param[in] lhs LHS tile
839 * @param[in] rhs_constant Constant value
840 * @param[out] dst DST tile
841 */
842#define T_ADD_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]843 ({ \
844 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
845 { \
846 LOOP_UNROLLING(int, _n0, 0, 1, N0, \
847 { \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]848 dst[_m0].s[_n0] = lhs[_m0].s[_n0] + rhs_constant; \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]849 }) \
850 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]851 })
852
853/** Element-wise addition with RHS broadcasted (RHS has the X dimension only)
854 *
855 * @note Performs: LHS + RHS[broadcasted] = DST
856 * @note Both tiles must have same data type
857 *
858 * @param[in] DATA_TYPE LHS/RHS/DST data type
859 * @param[in] M0 Number of LHS rows
860 * @param[in] N0 Number of LHS columns
861 * @param[in] lhs LHS tile
862 * @param[in] rhs RHS tile
863 * @param[out] dst DST tile
864 */
865#define T_ADD_BROADCAST_X(DATA_TYPE, M0, N0, lhs, rhs, dst) \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]866 ({ \
867 LOOP_UNROLLING(int, _m0, 0, 1, M0, \
868 { \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]869 dst[_m0].v = lhs[_m0].v + rhs[0].v; \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]870 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]871 })
872
873/** Matrix multiplication
874 *
875 * @note Performs: LHS X RHS + DST = DST
876 *
877 * @param[in] LHS_DATA_TYPE LHS tile data type
878 * @param[in] RHS_DATA_TYPE RHS tile data type
879 * @param[in] DST_DATA_TYPE RHS tile data type
880 * @param[in] M0 Number of LHS rows
881 * @param[in] N0 Number of RHS columns
882 * @param[in] K0 Number of LHS columns
883 * @param[in] LHS_LAYOUT LHS layout (T= transposed, NT= not transposed)
884 * @param[in] RHS_LAYOUT RHS layout (T= transposed, NT= not transposed)
885 * @param[in] lhs LHS tile
886 * @param[in] rhs RHS tile
887 * @param[in, out] dst DST tile
888 */
889#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 Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]890#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)
891#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)
892#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)
893#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)
894#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)
895#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)
896#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]897 { \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]898 LOOP_UNROLLING(int, _m, 0, 1, M0, \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]899 { \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]900 LOOP_UNROLLING(int, _n, 0, 1, N0, \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]901 { \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]902 LOOP_UNROLLING(int, _k, 0, 1, K0, \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]903 { \
904 dst[_m].s[_n] = fma((lhs[_m].s[_k]), (rhs[_n].s[_k]), dst[_m].s[_n]); \
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]905 }) \
906 }) \
907 }) \
Gian Marco Iodice5c9eed82021-03-19 11:26:20 +0000[diff] [blame]908 }
Giorgio Arenabdd16d12021-05-13 16:58:51 +0100[diff] [blame]909
Gian Marco Iodice561c1762021-04-16 15:08:59 +0100[diff] [blame]910#define T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \
911 ({ \
912 LOOP_UNROLLING(int, _m, 0, 1, M0, \
913 { \
914 LOOP_UNROLLING(int, _n, 0, 1, N0, \
915 { \
916 DOT_PRODUCT_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, K0, (lhs[_m].v), (rhs[_n].v), dst[_m].s[_n]); \
917 }) \
918 }) \
919 })