source: icGREP/icgrep-devel/icgrep/kernels/s2p_kernel.cpp @ 4986

Last change on this file since 4986 was 4986, checked in by nmedfort, 4 years ago

First attempt at dynamic segment size intergration.

File size: 5.0 KB
Line 
1/*
2 *  Copyright (c) 2016 International Characters.
3 *  This software is licensed to the public under the Open Software License 3.0.
4 */
5#include "s2p_kernel.h"
6#include <kernels/kernel.h>
7#include <IDISA/idisa_builder.h>
8
9namespace kernel {
10
11const int PACK_LANES = 1;
12
13void s2p_step(IDISA::IDISA_Builder * iBuilder, Value * s0, Value * s1, Value * hi_mask, unsigned shift, Value * &p0, Value * &p1) {
14    Value * t0 = nullptr;
15    Value * t1 = nullptr;
16    if ((iBuilder->getBitBlockWidth() == 256) && (PACK_LANES == 2)) {
17        Value * x0 = iBuilder->esimd_mergel(128, s0, s1);
18        Value * x1 = iBuilder->esimd_mergeh(128, s0, s1);
19        t0 = iBuilder->hsimd_packh_in_lanes(PACK_LANES, 16, x0, x1);
20        t1 = iBuilder->hsimd_packl_in_lanes(PACK_LANES, 16, x0, x1);
21    }
22    else {
23        t0 = iBuilder->hsimd_packh(16, s0, s1);
24        t1 = iBuilder->hsimd_packl(16, s0, s1);
25    }
26    p0 = iBuilder->simd_if(1, hi_mask, t0, iBuilder->simd_srli(16, t1, shift));
27    p1 = iBuilder->simd_if(1, hi_mask, iBuilder->simd_slli(16, t0, shift), t1);
28}
29
30inline void s2p(IDISA::IDISA_Builder * iBuilder, Value * input, Value * output[]) {
31    Value * bit00224466[4];
32    Value * bit11335577[4];
33
34    for (unsigned i = 0; i < 4; i++) {
35        Value * s0 = iBuilder->CreateBlockAlignedLoad(input, {iBuilder->getInt32(0), iBuilder->getInt32(2 * i)});
36        Value * s1 = iBuilder->CreateBlockAlignedLoad(input, {iBuilder->getInt32(0), iBuilder->getInt32(2 * i + 1)});
37        s2p_step(iBuilder, s0, s1, iBuilder->simd_himask(2), 1, bit00224466[i], bit11335577[i]);
38    }
39    Value * bit00004444[2];
40    Value * bit22226666[2];
41    Value * bit11115555[2];
42    Value * bit33337777[2];
43    for (unsigned j = 0; j<2; j++) {
44        s2p_step(iBuilder, bit00224466[2*j], bit00224466[2*j+1],
45                 iBuilder->simd_himask(4), 2, bit00004444[j], bit22226666[j]);
46        s2p_step(iBuilder, bit11335577[2*j], bit11335577[2*j+1],
47                 iBuilder->simd_himask(4), 2, bit11115555[j], bit33337777[j]);
48    }
49    s2p_step(iBuilder, bit00004444[0], bit00004444[1], iBuilder->simd_himask(8), 4, output[0], output[4]);
50    s2p_step(iBuilder, bit11115555[0], bit11115555[1], iBuilder->simd_himask(8), 4, output[1], output[5]);
51    s2p_step(iBuilder, bit22226666[0], bit22226666[1], iBuilder->simd_himask(8), 4, output[2], output[6]);
52    s2p_step(iBuilder, bit33337777[0], bit33337777[1], iBuilder->simd_himask(8), 4, output[3], output[7]);
53}
54
55void generateS2PKernel(Module *, IDISA::IDISA_Builder * iBuilder, KernelBuilder * kBuilder) {
56    kBuilder->addInputStream(8, "byte_pack");
57    for(unsigned i = 0; i < 8; ++i) {
58        kBuilder->addOutputStream(1);
59    }
60    kBuilder->prepareFunction();
61    Value * output[8];
62
63    Value * ptr = kBuilder->getInputStream(0);
64    //iBuilder->CallPrintInt("ptr", iBuilder->CreatePtrToInt(ptr, iBuilder->getInt64Ty()));
65    s2p(iBuilder, ptr, output);
66    for (unsigned j = 0; j < 8; ++j) {
67        //iBuilder->CallPrintRegister("bit" + std::to_string(j + 1), output[j]);
68        iBuilder->CreateBlockAlignedStore(output[j], kBuilder->getOutputStream(j));
69    }
70    kBuilder->finalize();
71}
72
73void generateS2P_idealKernel(Module *, IDISA::IDISA_Builder * iBuilder, KernelBuilder * kBuilder) {
74    kBuilder->addInputStream(8, "byte_pack");
75    for(unsigned i = 0; i < 8; ++i) {
76        kBuilder->addOutputStream(1);
77    }
78    kBuilder->prepareFunction();
79    Value * input = kBuilder->getInputStream(0);
80    Value * output[8];
81    Value * hi_nybble[4];
82    Value * lo_nybble[4];
83    for (unsigned i = 0; i<4; i++) {
84        Value * s0 = iBuilder->CreateBlockAlignedLoad(input, {iBuilder->getInt32(0), iBuilder->getInt32(2 * i)});
85        Value * s1 = iBuilder->CreateBlockAlignedLoad(input, {iBuilder->getInt32(0), iBuilder->getInt32(2 * i + 1)});
86        hi_nybble[i] = iBuilder->hsimd_packh(8, s0, s1);
87        lo_nybble[i] = iBuilder->hsimd_packl(8, s0, s1);
88    }
89    Value * pair01[2];
90    Value * pair23[2];
91    Value * pair45[2];
92    Value * pair67[2];
93    for (unsigned i = 0; i<2; i++) {
94        pair01[i] = iBuilder->hsimd_packh(4, hi_nybble[2*i], hi_nybble[2*i+1]);
95        pair23[i] = iBuilder->hsimd_packl(4, hi_nybble[2*i], hi_nybble[2*i+1]);
96        pair45[i] = iBuilder->hsimd_packh(4, lo_nybble[2*i], lo_nybble[2*i+1]);
97        pair67[i] = iBuilder->hsimd_packl(4, lo_nybble[2*i], lo_nybble[2*i+1]);
98    }
99    output[0] = iBuilder->hsimd_packh(2, pair01[0], pair01[1]);
100    output[1] = iBuilder->hsimd_packl(2, pair01[0], pair01[1]);
101    output[2] = iBuilder->hsimd_packh(2, pair23[0], pair23[1]);
102    output[3] = iBuilder->hsimd_packl(2, pair23[0], pair23[1]);
103    output[4] = iBuilder->hsimd_packh(2, pair45[0], pair45[1]);
104    output[5] = iBuilder->hsimd_packl(2, pair45[0], pair45[1]);
105    output[6] = iBuilder->hsimd_packh(2, pair67[0], pair67[1]);
106    output[7] = iBuilder->hsimd_packl(2, pair67[0], pair67[1]);
107
108    s2p(iBuilder, kBuilder->getInputStream(0), output);
109    for (unsigned j = 0; j < 8; ++j) {
110        iBuilder->CreateBlockAlignedStore(output[j], kBuilder->getOutputStream(j));
111    }
112    kBuilder->finalize();
113}
114   
115   
116   
117}
Note: See TracBrowser for help on using the repository browser.