1 | /* |
---|
2 | * Copyright (c) 2017 International Characters. |
---|
3 | * This software is licensed to the public under the Open Software License 3.0. |
---|
4 | */ |
---|
5 | #ifndef SWIZZLE_H |
---|
6 | #define SWIZZLE_H |
---|
7 | |
---|
8 | #include "kernel.h" |
---|
9 | namespace IDISA { class IDISA_Builder; } |
---|
10 | |
---|
11 | |
---|
12 | // The SwizzleGenerator class creates a kernel that transforms a set of bit streams into a swizzled form. |
---|
13 | // In swizzled form, one "swizzle field" each from a set of streams are grouped together to be processed |
---|
14 | // as a unit using SIMD operations. For example, for a swizzle field width of 64 and a block size of 256, |
---|
15 | // 4 streams are swizzled together to be operated on as a group. The ratio of the block size to the |
---|
16 | // swizzle field size is known as the swizzle factor, in this case 4. |
---|
17 | |
---|
18 | // Swizzled form is convenient for performing sequential operations on parallel sets of streams, |
---|
19 | // such as compression of each swizzle field by known counts, or stitching together known numbers |
---|
20 | // of bits from two different sources. |
---|
21 | |
---|
22 | // Any number of bit streams may be swizzled together. However, the outputSets are always |
---|
23 | // grouped together in multiples of the swizzle factor. If the bit stream count is not |
---|
24 | // an exact multiple of the swizzle factor, null streams (all zero bits) are added for |
---|
25 | // each swizzle group. |
---|
26 | |
---|
27 | // The input streams may come from any number of parallel input sets, each of the same size. |
---|
28 | // The number of inputSets defaults to 1. |
---|
29 | |
---|
30 | // The output streams may be separated in a number of output sets. However, output streams |
---|
31 | // sets must be a multiple of the swizzle factor size. |
---|
32 | |
---|
33 | // For example: consider the following 4 streams (32 bits each) |
---|
34 | // Stream 1: 000pomnl 000000kj 00000ihg 00fedcba |
---|
35 | // Stream 2: 00065432 0000001z 00000yxw 00vutsrq |
---|
36 | // Stream 3: 000POMNL 000000KJ 00000IHG 00FEDCBA |
---|
37 | // Stream 4: 00065432 0000001Z 00000YXW 00VUTSRQ |
---|
38 | // |
---|
39 | // The swizzled output using a field width of 8 produces the following swizzles. |
---|
40 | // |
---|
41 | // Swizzle 1: 00VUTSRQ 00FEDCBA 00vutsrq 00fedcba |
---|
42 | // Swizzle 2: 00000YXW 00000IHG 00000yxw 00000ihg |
---|
43 | // Swizzle 3: 0000001Z 000000KJ 0000001z 000000kj |
---|
44 | // Swizzle 4: 00065432 000POMNL 00065432 000pomnl |
---|
45 | // |
---|
46 | // Now it might be convenient to compress all fields of swizzle 1 by 2, all fields of swizzle 2 by 5 |
---|
47 | // and so on. |
---|
48 | // |
---|
49 | namespace kernel { |
---|
50 | |
---|
51 | class SwizzleGenerator : public BlockOrientedKernel { |
---|
52 | public: |
---|
53 | |
---|
54 | SwizzleGenerator(const std::unique_ptr<kernel::KernelBuilder> & iBuilder, unsigned bitStreamCount, unsigned outputSets = 1, unsigned inputSets = 1, unsigned fieldWidth = 64, std::string prefix = ""); |
---|
55 | |
---|
56 | protected: |
---|
57 | |
---|
58 | void generateDoBlockMethod(const std::unique_ptr<kernel::KernelBuilder> & iBuilder) override; |
---|
59 | |
---|
60 | private: |
---|
61 | const unsigned mBitStreamCount; |
---|
62 | const unsigned mFieldWidth; |
---|
63 | const unsigned mSwizzleFactor; |
---|
64 | const unsigned mInputSets; |
---|
65 | const unsigned mOutputSets; |
---|
66 | }; |
---|
67 | |
---|
68 | |
---|
69 | class SwizzleByGather : public BlockOrientedKernel { |
---|
70 | public: |
---|
71 | SwizzleByGather(const std::unique_ptr<kernel::KernelBuilder> & iBuilder); |
---|
72 | |
---|
73 | protected: |
---|
74 | void generateDoBlockMethod(const std::unique_ptr<kernel::KernelBuilder> & iBuilder) override; |
---|
75 | private: |
---|
76 | }; |
---|
77 | |
---|
78 | } |
---|
79 | |
---|
80 | #endif |
---|
81 | |
---|