/src/Simd/src/Simd/SimdBaseSynetQuantizedShuffle.cpp
Line | Count | Source |
1 | | /* |
2 | | * Simd Library (http://ermig1979.github.io/Simd). |
3 | | * |
4 | | * Copyright (c) 2011-2025 Yermalayeu Ihar. |
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 deal |
8 | | * in the Software without restriction, including without limitation the rights |
9 | | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
10 | | * 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 |
14 | | * all 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 | | #include "Simd/SimdSynetQuantizeLinear.h" |
25 | | |
26 | | namespace Simd |
27 | | { |
28 | | #if defined(SIMD_SYNET_ENABLE) |
29 | | namespace Base |
30 | | { |
31 | | void SynetQuantizedShuffleLayerForwardNchw0(const uint8_t* src0, int bias0, float norm0, size_t srcC0, const uint8_t* src1, int bias1, float norm1, size_t srcC1, size_t spatial, uint8_t* dst0, uint8_t* dst1, float scale, int zero) |
32 | 0 | { |
33 | 0 | size_t dstC = (srcC0 + srcC1) / 2; |
34 | 0 | size_t cd = 0; |
35 | 0 | for (size_t cs = 0; cs < srcC0; cs += 2, cd += 1) |
36 | 0 | { |
37 | 0 | for (size_t s = 0; s < spatial; ++s) |
38 | 0 | dst0[s] = DequantizeQuantizeLinear(src0[s], bias0, norm0, scale, zero, 0, 255); |
39 | 0 | src0 += spatial; |
40 | 0 | dst0 += spatial; |
41 | 0 | for (size_t s = 0; s < spatial; ++s) |
42 | 0 | dst1[s] = DequantizeQuantizeLinear(src0[s], bias0, norm0, scale, zero, 0, 255); |
43 | 0 | src0 += spatial; |
44 | 0 | dst1 += spatial; |
45 | 0 | } |
46 | 0 | for (size_t cs = 0; cs < srcC1; cs += 2, cd += 1) |
47 | 0 | { |
48 | 0 | for (size_t s = 0; s < spatial; ++s) |
49 | 0 | dst0[s] = DequantizeQuantizeLinear(src1[s], bias1, norm1, scale, zero, 0, 255); |
50 | 0 | src1 += spatial; |
51 | 0 | dst0 += spatial; |
52 | 0 | for (size_t s = 0; s < spatial; ++s) |
53 | 0 | dst1[s] = DequantizeQuantizeLinear(src1[s], bias1, norm1, scale, zero, 0, 255); |
54 | 0 | src1 += spatial; |
55 | 0 | dst1 += spatial; |
56 | 0 | } |
57 | 0 | } |
58 | | |
59 | | void SynetQuantizedShuffleLayerForwardNhwc0(const uint8_t* src0, int bias0, float norm0, size_t srcC0, const uint8_t* src1, int bias1, float norm1, size_t srcC1, size_t spatial, uint8_t* dst0, uint8_t* dst1, float scale, int zero) |
60 | 0 | { |
61 | 0 | size_t dstC = (srcC0 + srcC1) / 2; |
62 | 0 | for (size_t s = 0; s < spatial; ++s) |
63 | 0 | { |
64 | 0 | size_t cd = 0; |
65 | 0 | for (size_t cs = 0; cs < srcC0; cs += 2, cd += 1) |
66 | 0 | { |
67 | 0 | dst0[cd] = DequantizeQuantizeLinear(src0[cs + 0], bias0, norm0, scale, zero, 0, 255); |
68 | 0 | dst1[cd] = DequantizeQuantizeLinear(src0[cs + 1], bias0, norm0, scale, zero, 0, 255); |
69 | 0 | } |
70 | 0 | for (size_t cs = 0; cs < srcC1; cs += 2, cd += 1) |
71 | 0 | { |
72 | 0 | dst0[cd] = DequantizeQuantizeLinear(src1[cs + 0], bias1, norm1, scale, zero, 0, 255); |
73 | 0 | dst1[cd] = DequantizeQuantizeLinear(src1[cs + 1], bias1, norm1, scale, zero, 0, 255); |
74 | 0 | } |
75 | 0 | src0 += srcC0; |
76 | 0 | src1 += srcC1; |
77 | 0 | dst0 += dstC; |
78 | 0 | dst1 += dstC; |
79 | 0 | } |
80 | |
|
81 | 0 | } |
82 | | |
83 | | void SynetQuantizedShuffleLayerForwardNchw1(const uint8_t* src0, int bias0, float norm0, size_t srcC0, const uint8_t* src1, int bias1, float norm1, size_t srcC1, size_t spatial, uint8_t* dst0, uint8_t* dst1, float scale, int zero) |
84 | 0 | { |
85 | 0 | size_t dstC = (srcC0 + srcC1) / 2; |
86 | 0 | size_t cs = 0; |
87 | 0 | for (size_t cd = 0; cd < srcC0; cs += 1, cd += 2) |
88 | 0 | { |
89 | 0 | for (size_t s = 0; s < spatial; ++s) |
90 | 0 | dst0[s] = DequantizeQuantizeLinear(src0[s], bias0, norm0, scale, zero, 0, 255); |
91 | 0 | src0 += spatial; |
92 | 0 | dst0 += spatial; |
93 | 0 | for (size_t s = 0; s < spatial; ++s) |
94 | 0 | dst0[s] = DequantizeQuantizeLinear(src1[s], bias1, norm1, scale, zero, 0, 255); |
95 | 0 | src1 += spatial; |
96 | 0 | dst0 += spatial; |
97 | 0 | } |
98 | 0 | for (size_t cd = 0; cd < srcC1; cs += 1, cd += 2) |
99 | 0 | { |
100 | 0 | for (size_t s = 0; s < spatial; ++s) |
101 | 0 | dst1[s] = DequantizeQuantizeLinear(src0[s], bias0, norm0, scale, zero, 0, 255); |
102 | 0 | src0 += spatial; |
103 | 0 | dst1 += spatial; |
104 | 0 | for (size_t s = 0; s < spatial; ++s) |
105 | 0 | dst1[s] = DequantizeQuantizeLinear(src1[s], bias1, norm1, scale, zero, 0, 255); |
106 | 0 | src1 += spatial; |
107 | 0 | dst1 += spatial; |
108 | 0 | } |
109 | 0 | } |
110 | | |
111 | | void SynetQuantizedShuffleLayerForwardNhwc1(const uint8_t* src0, int bias0, float norm0, size_t srcC0, const uint8_t* src1, int bias1, float norm1, size_t srcC1, size_t spatial, uint8_t* dst0, uint8_t* dst1, float scale, int zero) |
112 | 0 | { |
113 | 0 | size_t dstC = (srcC0 + srcC1) / 2; |
114 | 0 | for (size_t s = 0; s < spatial; ++s) |
115 | 0 | { |
116 | 0 | size_t cs = 0; |
117 | 0 | for (size_t cd = 0; cd < srcC0; cd += 2, cs += 1) |
118 | 0 | { |
119 | 0 | dst0[cd + 0] = DequantizeQuantizeLinear(src0[cs], bias0, norm0, scale, zero, 0, 255); |
120 | 0 | dst0[cd + 1] = DequantizeQuantizeLinear(src1[cs], bias1, norm1, scale, zero, 0, 255); |
121 | 0 | } |
122 | 0 | for (size_t cd = 0; cd < srcC1; cd += 2, cs += 1) |
123 | 0 | { |
124 | 0 | dst1[cd + 0] = DequantizeQuantizeLinear(src0[cs], bias0, norm0, scale, zero, 0, 255); |
125 | 0 | dst1[cd + 1] = DequantizeQuantizeLinear(src1[cs], bias1, norm1, scale, zero, 0, 255); |
126 | 0 | } |
127 | 0 | src0 += dstC; |
128 | 0 | src1 += dstC; |
129 | 0 | dst0 += srcC0; |
130 | 0 | dst1 += srcC1; |
131 | 0 | } |
132 | 0 | } |
133 | | |
134 | | void SynetQuantizedShuffleLayerForward(const uint8_t* src0, int bias0, const float* norm0, size_t srcC0, const uint8_t* src1, int bias1, const float* norm1, size_t srcC1, |
135 | | size_t spatial, uint8_t* dst0, uint8_t* dst1, const float* scale, int zero, SimdTensorFormatType format, int shuffleType) |
136 | 0 | { |
137 | 0 | size_t dstC = (srcC0 + srcC1) / 2; |
138 | 0 | switch (shuffleType) |
139 | 0 | { |
140 | 0 | case 0: |
141 | 0 | if (format == SimdTensorFormatNhwc) |
142 | 0 | SynetQuantizedShuffleLayerForwardNhwc0(src0, bias0, *norm0, srcC0, src1, bias1, *norm1, srcC1, spatial, dst0, dst1, *scale, zero); |
143 | 0 | else |
144 | 0 | SynetQuantizedShuffleLayerForwardNchw0(src0, bias0, *norm0, srcC0, src1, bias1, *norm1, srcC1, spatial, dst0, dst1, *scale, zero); |
145 | 0 | break; |
146 | 0 | case 1: |
147 | 0 | if (format == SimdTensorFormatNhwc) |
148 | 0 | SynetQuantizedShuffleLayerForwardNhwc1(src0, bias0, *norm0, srcC0, src1, bias1, *norm1, srcC1, spatial, dst0, dst1, *scale, zero); |
149 | 0 | else |
150 | 0 | SynetQuantizedShuffleLayerForwardNchw1(src0, bias0, *norm0, srcC0, src1, bias1, *norm1, srcC1, spatial, dst0, dst1, *scale, zero); |
151 | 0 | break; |
152 | 0 | } |
153 | 0 | } |
154 | | } |
155 | | #endif |
156 | | } |