1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
|
/*
* Copyright (C) 2018-2021 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "shared/source/helpers/local_id_gen.h"
#include <array>
namespace NEO {
template <typename Vec, int simd>
inline void generateLocalIDsSimd(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t threadsPerWorkGroup,
const std::array<uint8_t, 3> &dimensionsOrder, bool chooseMaxRowSize) {
const int passes = simd / Vec::numChannels;
int pass = 0;
uint32_t xDimNum = dimensionsOrder[0];
uint32_t yDimNum = dimensionsOrder[1];
uint32_t zDimNum = dimensionsOrder[2];
const Vec vLwsX(localWorkgroupSize[xDimNum]);
const Vec vLwsY(localWorkgroupSize[yDimNum]);
auto zero = Vec::zero();
auto one = Vec::one();
const auto threadSkipSize = ((simd == 32 || chooseMaxRowSize) ? 32 : 16) * sizeof(uint16_t);
Vec vSimdX(simd);
Vec vSimdY = zero;
Vec vSimdZ = zero;
Vec xWrap;
Vec yWrap;
// We need to convert simd into appropriate delta adders
do {
xWrap = vSimdX >= vLwsX;
// xWrap ? lwsX : 0;
auto deltaX = blend(vLwsX, zero, xWrap);
// x -= xWrap ? lwsX : 0;
vSimdX -= deltaX;
// xWrap ? 1 : 0;
auto deltaY = blend(one, zero, xWrap);
// y += xWrap ? 1 : 0;
vSimdY += deltaY;
yWrap = vSimdY >= vLwsY;
// yWrap ? lwsY : 0;
auto deltaY2 = blend(vLwsY, zero, yWrap);
// y -= yWrap ? lwsY : 0;
vSimdY -= deltaY2;
// yWrap ? 1 : 0;
auto deltaZ = blend(one, zero, yWrap);
// z += yWrap ? 1 : 0;
vSimdZ += deltaZ;
} while (xWrap || yWrap);
// Loop for each of the passes
do {
auto buffer = b;
Vec x(&initialLocalID[pass * Vec::numChannels]);
Vec y = zero;
Vec z = zero;
// Convert the initial SIMD lanes to localIDs
do {
xWrap = x >= vLwsX;
// xWrap ? lwsX : 0;
auto deltaX = blend(vLwsX, zero, xWrap);
// x -= xWrap ? lwsX : 0;
x -= deltaX;
// xWrap ? 1 : 0;
auto deltaY = blend(one, zero, xWrap);
// y += xWrap ? 1 : 0;
y += deltaY;
yWrap = y >= vLwsY;
// yWrap ? lwsY : 0;
auto deltaY2 = blend(vLwsY, zero, yWrap);
// y -= yWrap ? lwsY : 0;
y -= deltaY2;
// yWrap ? 1 : 0;
auto deltaZ = blend(one, zero, yWrap);
// z += yWrap ? 1 : 0;
z += deltaZ;
} while (xWrap);
for (size_t i = 0; i < threadsPerWorkGroup; ++i) {
x.store(ptrOffset(buffer, xDimNum * threadSkipSize));
y.store(ptrOffset(buffer, yDimNum * threadSkipSize));
z.store(ptrOffset(buffer, zDimNum * threadSkipSize));
x += vSimdX;
y += vSimdY;
z += vSimdZ;
xWrap = x >= vLwsX;
// xWrap ? lwsX : 0;
auto deltaX = blend(vLwsX, zero, xWrap);
// x -= xWrap ? lwsX : 0;
x -= deltaX;
// xWrap ? 1 : 0;
auto deltaY = blend(one, zero, xWrap);
// y += xWrap ? 1 : 0;
y += deltaY;
yWrap = y >= vLwsY;
// yWrap ? lwsY : 0;
auto deltaY2 = blend(vLwsY, zero, yWrap);
// y -= yWrap ? lwsY : 0;
y -= deltaY2;
// yWrap ? 1 : 0;
auto deltaZ = blend(one, zero, yWrap);
// z += yWrap ? 1 : 0;
z += deltaZ;
buffer = ptrOffset(buffer, 3 * threadSkipSize);
}
// Adjust buffer for next pass
b = ptrOffset(b, Vec::numChannels * sizeof(uint16_t));
} while (++pass < passes);
}
} // namespace NEO
|
