source: icGREP/icgrep-devel/icgrep/kernels/s2p_kernel.cpp @ 5391

Last change on this file since 5391 was 5391, checked in by cameron, 2 years ago

ParabixDriver/ObjectCache? separate compilation and linking: initial check-in with wc

File size: 6.6 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
6#include "s2p_kernel.h"
7#include <IR_Gen/idisa_builder.h>  // for IDISA_Builder
8#include <llvm/IR/Constant.h>      // for Constant
9#include <llvm/IR/Module.h>
10#include <llvm/Support/raw_ostream.h>
11namespace llvm { class BasicBlock; }
12namespace llvm { class Function; }
13namespace llvm { class Value; }
14
15using namespace llvm;
16
17namespace kernel {
18
19const int PACK_LANES = 1;
20
21void s2p_step(IDISA::IDISA_Builder * iBuilder, Value * s0, Value * s1, Value * hi_mask, unsigned shift, Value * &p0, Value * &p1) {
22    Value * t0 = nullptr;
23    Value * t1 = nullptr;
24    if ((iBuilder->getBitBlockWidth() == 256) && (PACK_LANES == 2)) {
25        Value * x0 = iBuilder->esimd_mergel(128, s0, s1);
26        Value * x1 = iBuilder->esimd_mergeh(128, s0, s1);
27        t0 = iBuilder->hsimd_packh_in_lanes(PACK_LANES, 16, x0, x1);
28        t1 = iBuilder->hsimd_packl_in_lanes(PACK_LANES, 16, x0, x1);
29    } else {
30        t0 = iBuilder->hsimd_packh(16, s0, s1);
31        t1 = iBuilder->hsimd_packl(16, s0, s1);
32    }
33    p0 = iBuilder->simd_if(1, hi_mask, t0, iBuilder->simd_srli(16, t1, shift));
34    p1 = iBuilder->simd_if(1, hi_mask, iBuilder->simd_slli(16, t0, shift), t1);
35}
36
37void s2p(IDISA::IDISA_Builder * iBuilder, Value * input[], Value * output[]) {
38    Value * bit00224466[4];
39    Value * bit11335577[4];
40
41    for (unsigned i = 0; i < 4; i++) {
42        Value * s0 = input[2 * i];
43        Value * s1 = input[2 * i + 1];
44        s2p_step(iBuilder, s0, s1, iBuilder->simd_himask(2), 1, bit00224466[i], bit11335577[i]);
45    }
46    Value * bit00004444[2];
47    Value * bit22226666[2];
48    Value * bit11115555[2];
49    Value * bit33337777[2];
50    for (unsigned j = 0; j<2; j++) {
51        s2p_step(iBuilder, bit00224466[2*j], bit00224466[2*j+1],
52                 iBuilder->simd_himask(4), 2, bit00004444[j], bit22226666[j]);
53        s2p_step(iBuilder, bit11335577[2*j], bit11335577[2*j+1],
54                 iBuilder->simd_himask(4), 2, bit11115555[j], bit33337777[j]);
55    }
56    s2p_step(iBuilder, bit00004444[0], bit00004444[1], iBuilder->simd_himask(8), 4, output[0], output[4]);
57    s2p_step(iBuilder, bit11115555[0], bit11115555[1], iBuilder->simd_himask(8), 4, output[1], output[5]);
58    s2p_step(iBuilder, bit22226666[0], bit22226666[1], iBuilder->simd_himask(8), 4, output[2], output[6]);
59    s2p_step(iBuilder, bit33337777[0], bit33337777[1], iBuilder->simd_himask(8), 4, output[3], output[7]);
60}
61
62/* Alternative transposition model, but small field width packs are problematic. */
63#if 0
64void s2p_ideal(IDISA::IDISA_Builder * iBuilder, Value * input[], Value * output[]) {
65    Value * hi_nybble[4];
66    Value * lo_nybble[4];
67    for (unsigned i = 0; i<4; i++) {
68        Value * s0 = input[2*i];
69        Value * s1 = input[2*i+1];
70        hi_nybble[i] = iBuilder->hsimd_packh(8, s0, s1);
71        lo_nybble[i] = iBuilder->hsimd_packl(8, s0, s1);
72    }
73    Value * pair01[2];
74    Value * pair23[2];
75    Value * pair45[2];
76    Value * pair67[2];
77    for (unsigned i = 0; i<2; i++) {
78        pair01[i] = iBuilder->hsimd_packh(4, hi_nybble[2*i], hi_nybble[2*i+1]);
79        pair23[i] = iBuilder->hsimd_packl(4, hi_nybble[2*i], hi_nybble[2*i+1]);
80        pair45[i] = iBuilder->hsimd_packh(4, lo_nybble[2*i], lo_nybble[2*i+1]);
81        pair67[i] = iBuilder->hsimd_packl(4, lo_nybble[2*i], lo_nybble[2*i+1]);
82    }
83    output[0] = iBuilder->hsimd_packh(2, pair01[0], pair01[1]);
84    output[1] = iBuilder->hsimd_packl(2, pair01[0], pair01[1]);
85    output[2] = iBuilder->hsimd_packh(2, pair23[0], pair23[1]);
86    output[3] = iBuilder->hsimd_packl(2, pair23[0], pair23[1]);
87    output[4] = iBuilder->hsimd_packh(2, pair45[0], pair45[1]);
88    output[5] = iBuilder->hsimd_packl(2, pair45[0], pair45[1]);
89    output[6] = iBuilder->hsimd_packh(2, pair67[0], pair67[1]);
90    output[7] = iBuilder->hsimd_packl(2, pair67[0], pair67[1]);
91}
92#endif
93   
94#if 0
95void generateS2P_16Kernel(Module *, IDISA::IDISA_Builder * iBuilder, KernelBuilder * kBuilder) {
96    kBuilder->addInputStream(16, "unit_pack");
97    for(unsigned i = 0; i < 16; i++) {
98            kBuilder->addOutputStream(1);
99    }
100    kBuilder->prepareFunction();
101
102    Value * ptr = kBuilder->getInputStream(0);
103
104    Value * lo[8];
105    Value * hi[8];
106    for (unsigned i = 0; i < 8; i++) {
107        Value * s0 = iBuilder->CreateBlockAlignedLoad(ptr, {iBuilder->getInt32(0), iBuilder->getInt32(2 * i)});
108        Value * s1 = iBuilder->CreateBlockAlignedLoad(ptr, {iBuilder->getInt32(0), iBuilder->getInt32(2 * i + 1)});
109        lo[i] = iBuilder->hsimd_packl(16, s0, s1);
110        hi[i] = iBuilder->hsimd_packh(16, s0, s1);
111    }
112
113    Value * output[16];
114    s2p(iBuilder, lo, output);
115    s2p(iBuilder, hi, output + 8);
116    for (unsigned j = 0; j < 16; j++) {
117        iBuilder->CreateBlockAlignedStore(output[j], kBuilder->getOutputStream(j));
118    }
119    kBuilder->finalize();
120}   
121#endif
122   
123void S2PKernel::generateDoBlockMethod() {
124    Value * bytepack[8];
125    for (unsigned i = 0; i < 8; i++) {
126        bytepack[i] = loadInputStreamPack("byteStream", iBuilder->getInt32(0), iBuilder->getInt32(i));
127    }
128    Value * basisbits[8];
129    s2p(iBuilder, bytepack, basisbits);
130    for (unsigned i = 0; i < 8; ++i) {
131        storeOutputStreamBlock("basisBits", iBuilder->getInt32(i), basisbits[i]);
132    }
133}
134
135void S2PKernel::generateFinalBlockMethod(Value * remainingBytes) {
136    /* Prepare the s2p final block function:
137     assumption: if remaining bytes is greater than 0, it is safe to read a full block of bytes.
138     if remaining bytes is zero, no read should be performed (e.g. for mmapped buffer).
139     */
140   
141    BasicBlock * finalPartialBlock = CreateBasicBlock("partial");
142    BasicBlock * finalEmptyBlock = CreateBasicBlock("empty");
143    BasicBlock * exitBlock = CreateBasicBlock("exit");
144   
145    Value * emptyBlockCond = iBuilder->CreateICmpEQ(remainingBytes, iBuilder->getSize(0));
146    iBuilder->CreateCondBr(emptyBlockCond, finalEmptyBlock, finalPartialBlock);
147    iBuilder->SetInsertPoint(finalPartialBlock);
148    CreateDoBlockMethodCall();
149   
150    iBuilder->CreateBr(exitBlock);
151   
152    iBuilder->SetInsertPoint(finalEmptyBlock);
153
154    for (unsigned i = 0; i < 8; ++i) {
155        storeOutputStreamBlock("basisBits", iBuilder->getInt32(i), Constant::getNullValue(iBuilder->getBitBlockType()));
156    }
157
158    iBuilder->CreateBr(exitBlock);
159   
160    iBuilder->SetInsertPoint(exitBlock);
161}
162
163S2PKernel::S2PKernel(IDISA::IDISA_Builder * builder)
164: BlockOrientedKernel(builder, "Parabix:s2p", 
165    {Binding{builder->getStreamSetTy(1, 8), "byteStream"}}, {Binding{builder->getStreamSetTy(8, 1), "basisBits"}}, {}, {}, {}) {
166    setNoTerminateAttribute(true);
167}
168
169}
Note: See TracBrowser for help on using the repository browser.