source: icGREP/icgrep-devel/icgrep/kernels/deletion.cpp @ 5071

Last change on this file since 5071 was 5071, checked in by cameron, 3 years ago

Update u8u16 to use new kernel infrastructure

File size: 5.1 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 <kernels/kernel.h>
7#include <kernels/deletion.h>
8#include <IDISA/idisa_builder.h>
9#include <llvm/IR/Value.h>
10
11std::vector<Value *> parallel_prefix_deletion_masks(IDISA::IDISA_Builder * iBuilder, unsigned fw, Value * del_mask) {
12    Value * m = iBuilder->simd_not(del_mask);
13    Value * mk = iBuilder->simd_slli(fw, del_mask, 1);
14    std::vector<Value *> move_masks;
15    for (unsigned shift = 1; shift < fw; shift *= 2) {
16        Value * mp = mk;
17        for (unsigned lookright = 1; lookright < fw; lookright *= 2) {
18            mp = iBuilder->simd_xor(mp, iBuilder->simd_slli(fw, mp, lookright));
19        }
20        Value * mv = iBuilder->simd_and(mp, m);
21        m = iBuilder->simd_or(iBuilder->simd_xor(m, mv), iBuilder->simd_srli(fw, mv, shift));
22        mk = iBuilder->simd_and(mk, iBuilder->simd_not(mp));
23        move_masks.push_back(mv);
24    }
25    return move_masks;
26}
27
28Value * apply_parallel_prefix_deletion(IDISA::IDISA_Builder * iBuilder, unsigned fw, Value * del_mask, std::vector<Value *> mv, Value * strm) {
29    Value * s = iBuilder->simd_and(strm, iBuilder->simd_not(del_mask));
30    for (unsigned i = 0; i < mv.size(); i++) {
31        unsigned shift = 1 << i;
32        Value * t = iBuilder->simd_and(s, mv[i]);
33        s = iBuilder->simd_or(iBuilder->simd_xor(s, t), iBuilder->simd_srli(fw, t, shift));
34    }
35    return s;
36}
37
38Value * partial_sum_popcount(IDISA::IDISA_Builder * iBuilder, unsigned fw, Value * mask) {
39    Value * per_field = iBuilder->simd_popcount(fw, mask);
40    for (unsigned move = 1; move < iBuilder->getBitBlockWidth()/fw; move *= 2) {
41        per_field = iBuilder->simd_add(fw, per_field, iBuilder->mvmd_slli(fw, per_field, move));
42    }
43    return per_field;
44}
45
46// Apply deletion to a set of stream_count input streams to produce a set of output streams.
47// Kernel inputs: stream_count data streams plus one del_mask stream
48// Outputs: the deleted streams, plus a partial sum popcount
49
50
51void deletionKernel::generateKernel() {
52    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
53    if (mKernelStateType == nullptr) finalizeKernelStateType();
54    KernelBuilder::generateKernel();
55   
56    Module * m = iBuilder->getModule();
57    unsigned blockSize = iBuilder->getBitBlockWidth();
58    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
59    Function * finalBlockFunction = m->getFunction(mKernelName + finalBlock_suffix);
60   
61    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", doBlockFunction, 0));
62   
63    Value * inputStreamBlock = getParameter(doBlockFunction, "inputStreamSet");
64    Value * outputStreamBlock = getParameter(doBlockFunction, "outputStreamSet");
65    Value * delCountBlock = getParameter(doBlockFunction, "deletionCounts");
66   
67    Value * del_mask = iBuilder->CreateBlockAlignedLoad(inputStreamBlock, {iBuilder->getInt32(0), iBuilder->getInt32(mStreamCount)});
68   
69    std::vector<Value *> move_masks = parallel_prefix_deletion_masks(iBuilder, mDeletionFieldWidth, del_mask);
70       
71    for (unsigned j = 0; j < mStreamCount; ++j) {
72        Value * input = iBuilder->CreateBlockAlignedLoad(inputStreamBlock, {iBuilder->getInt32(0), iBuilder->getInt32(j)});
73        Value * output = apply_parallel_prefix_deletion(iBuilder, mDeletionFieldWidth, del_mask, move_masks, input);
74        iBuilder->CreateBlockAlignedStore(output, outputStreamBlock, {iBuilder->getInt32(0), iBuilder->getInt32(j)});
75    }
76    Value * counts = partial_sum_popcount(iBuilder, mDeletionFieldWidth, iBuilder->simd_not(del_mask));
77    iBuilder->CreateBlockAlignedStore(iBuilder->bitCast(counts), delCountBlock, {iBuilder->getInt32(0), iBuilder->getInt32(0)});
78                                         
79    iBuilder->CreateRetVoid();
80   
81    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", finalBlockFunction, 0));
82    Value * remainingBytes = getParameter(finalBlockFunction, "remainingBytes");
83    inputStreamBlock = getParameter(finalBlockFunction, "inputStreamSet");
84    outputStreamBlock = getParameter(finalBlockFunction, "outputStreamSet");
85    delCountBlock = getParameter(finalBlockFunction, "deletionCounts");
86    Value * remaining = iBuilder->CreateZExt(remainingBytes, iBuilder->getIntNTy(blockSize));
87    Value * EOF_del = iBuilder->bitCast(iBuilder->CreateShl(Constant::getAllOnesValue(iBuilder->getIntNTy(blockSize)), remaining));
88    Value * const delmaskPtr = iBuilder->CreateGEP(inputStreamBlock, {iBuilder->getInt32(0), iBuilder->getInt32(16)});
89    Value * const delmaskVal = iBuilder->CreateBlockAlignedLoad(delmaskPtr);
90    iBuilder->CreateBlockAlignedStore(iBuilder->CreateOr(EOF_del, delmaskVal), delmaskPtr);
91    Function::arg_iterator args = finalBlockFunction->arg_begin();
92    Value * self = &*(args++);
93    /* Skip "remaining" arg */ args++;
94    std::vector<Value *> doBlockArgs = {self};
95    while (args != finalBlockFunction->arg_end()){
96        doBlockArgs.push_back(&*args++);
97    }
98    iBuilder->CreateCall(doBlockFunction, doBlockArgs);   
99    iBuilder->CreateRetVoid();
100    iBuilder->restoreIP(savePoint);
101}
102
103
Note: See TracBrowser for help on using the repository browser.