1/*
2 * Copyright (C) 2018-2021 Intel Corporation
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 */
7
8#include "shared/source/helpers/ptr_math.h"
9
10#include <array>
11
12namespace NEO {
13
14bool isCompatibleWithLayoutForImages(const std::array<uint16_t, 3> &localWorkgroupSize, const std::array<uint8_t, 3> &dimensionsOrder, uint16_t simd) {
15    uint8_t xMask = simd == 8u ? 0b1 : 0b11;
16    uint8_t yMask = 0b11;
17    return dimensionsOrder.at(0) == 0 &&
18           dimensionsOrder.at(1) == 1 &&
19           (localWorkgroupSize.at(0) & xMask) == 0 &&
20           (localWorkgroupSize.at(1) & yMask) == 0 &&
21           localWorkgroupSize.at(2) == 1u;
22}
23
24inline void generateLocalIDsWithLayoutForImages(void *b, const std::array<uint16_t, 3> &localWorkgroupSize, uint16_t simd) {
25    uint8_t rowWidth = simd == 32u ? 32u : 16u;
26    uint8_t xDelta = simd == 8u ? 2u : 4u;                                                    // difference between corresponding values in consecutive X rows
27    uint8_t yDelta = (simd == 8u || localWorkgroupSize.at(1) == 4u) ? 4u : rowWidth / xDelta; // difference between corresponding values in consecutive Y rows
28
29    auto buffer = reinterpret_cast<uint16_t *>(b);
30    uint16_t offset = 0u;
31    auto numGrfs = (localWorkgroupSize.at(0) * localWorkgroupSize.at(1) * localWorkgroupSize.at(2) + (simd - 1)) / simd;
32    uint8_t xMask = simd == 8u ? 0b1 : 0b11;
33    uint16_t x = 0u;
34    uint16_t y = 0u;
35    for (auto grfId = 0; grfId < numGrfs; grfId++) {
36        auto rowX = buffer + offset;
37        auto rowY = buffer + offset + rowWidth;
38        auto rowZ = buffer + offset + 2 * rowWidth;
39        uint16_t extraX = 0u;
40        uint16_t extraY = 0u;
41
42        for (uint8_t i = 0u; i < simd; i++) {
43            if (i > 0) {
44                extraX++;
45                if (extraX == xDelta) {
46                    extraX = 0u;
47                }
48                if ((i & xMask) == 0) {
49                    extraY++;
50                    if (y + extraY == localWorkgroupSize.at(1)) {
51                        extraY = 0;
52                        x += xDelta;
53                    }
54                }
55            }
56            if (x == localWorkgroupSize.at(0)) {
57                x = 0u;
58                y += yDelta;
59                if (y >= localWorkgroupSize.at(1)) {
60                    y = 0u;
61                }
62            }
63            rowX[i] = x + extraX;
64            rowY[i] = y + extraY;
65            rowZ[i] = 0u;
66        }
67        x += xDelta;
68        offset += 3 * rowWidth;
69    }
70}
71
72void generateLocalIDsForSimdOne(void *b, const std::array<uint16_t, 3> &localWorkgroupSize,
73                                const std::array<uint8_t, 3> &dimensionsOrder, uint32_t grfSize) {
74    uint32_t xDimNum = dimensionsOrder[0];
75    uint32_t yDimNum = dimensionsOrder[1];
76    uint32_t zDimNum = dimensionsOrder[2];
77
78    for (int i = 0; i < localWorkgroupSize[zDimNum]; i++) {
79        for (int j = 0; j < localWorkgroupSize[yDimNum]; j++) {
80            for (int k = 0; k < localWorkgroupSize[xDimNum]; k++) {
81                static_cast<uint16_t *>(b)[0] = k;
82                static_cast<uint16_t *>(b)[1] = j;
83                static_cast<uint16_t *>(b)[2] = i;
84                b = ptrOffset(b, grfSize);
85            }
86        }
87    }
88}
89
90} // namespace NEO
91