source: icGREP/icgrep-devel/icgrep/kernels/s2p_kernel.cpp @ 5292

Last change on this file since 5292 was 5292, checked in by nmedfort, 3 years ago

Removed 'function' and 'self' parameters from generateXXXMethod() functions.

File size: 6.8 KB
RevLine 
[4939]1/*
[4959]2 *  Copyright (c) 2016 International Characters.
[4939]3 *  This software is licensed to the public under the Open Software License 3.0.
4 */
[5260]5
[4939]6#include "s2p_kernel.h"
[5260]7#include <IR_Gen/idisa_builder.h>  // for IDISA_Builder
8#include <llvm/IR/Constant.h>      // for Constant
9#include <llvm/IR/Module.h>
10namespace llvm { class BasicBlock; }
11namespace llvm { class Function; }
12namespace llvm { class Value; }
[4939]13
[5063]14using namespace llvm;
[4974]15
[5260]16namespace kernel {
17
[4957]18const int PACK_LANES = 1;
[4939]19
20void s2p_step(IDISA::IDISA_Builder * iBuilder, Value * s0, Value * s1, Value * hi_mask, unsigned shift, Value * &p0, Value * &p1) {
[4957]21    Value * t0 = nullptr;
22    Value * t1 = nullptr;
23    if ((iBuilder->getBitBlockWidth() == 256) && (PACK_LANES == 2)) {
24        Value * x0 = iBuilder->esimd_mergel(128, s0, s1);
25        Value * x1 = iBuilder->esimd_mergeh(128, s0, s1);
26        t0 = iBuilder->hsimd_packh_in_lanes(PACK_LANES, 16, x0, x1);
27        t1 = iBuilder->hsimd_packl_in_lanes(PACK_LANES, 16, x0, x1);
28    }
29    else {
30        t0 = iBuilder->hsimd_packh(16, s0, s1);
31        t1 = iBuilder->hsimd_packl(16, s0, s1);
32    }
[4939]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
[5045]37void s2p(IDISA::IDISA_Builder * iBuilder, Value * input[], Value * output[]) {
[4939]38    Value * bit00224466[4];
39    Value * bit11335577[4];
[4986]40
41    for (unsigned i = 0; i < 4; i++) {
[5045]42        Value * s0 = input[2 * i];
43        Value * s1 = input[2 * i + 1];
[4959]44        s2p_step(iBuilder, s0, s1, iBuilder->simd_himask(2), 1, bit00224466[i], bit11335577[i]);
[4939]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    }
[4959]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]);
[4939]60}
61
[5063]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);
[5045]72    }
[5063]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]);
[5045]91}
[5063]92#endif
93   
94#if 0
[5045]95void generateS2P_16Kernel(Module *, IDISA::IDISA_Builder * iBuilder, KernelBuilder * kBuilder) {
[5046]96    kBuilder->addInputStream(16, "unit_pack");
97    for(unsigned i = 0; i < 16; i++) {
98            kBuilder->addOutputStream(1);
99    }
100    kBuilder->prepareFunction();
[5045]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();
[5285]120}   
[5063]121#endif
[4976]122   
[5292]123void S2PKernel::generateDoBlockMethod(llvm::Value * blockNo) {
[5285]124    Value * bytepack[8];
125    for (unsigned i = 0; i < 8; i++) {
[5292]126        Value * byteStream = getStream("byteStream", blockNo, iBuilder->getInt32(0), iBuilder->getInt32(i));
[5285]127        bytepack[i] = iBuilder->CreateBlockAlignedLoad(byteStream);
128    }
129    Value * basisbits[8];
130    s2p(iBuilder, bytepack, basisbits);
131    for (unsigned i = 0; i < 8; ++i) {
[5292]132        Value * basisBits = getStream("basisBits", blockNo, iBuilder->getInt32(i));
[5285]133        iBuilder->CreateBlockAlignedStore(basisbits[i], basisBits);
134    }
135}
136
[5292]137void S2PKernel::generateFinalBlockMethod(Value * remainingBytes, llvm::Value * blockNo) {
[5074]138    /* Prepare the s2p final block function:
[5051]139     assumption: if remaining bytes is greater than 0, it is safe to read a full block of bytes.
140     if remaining bytes is zero, no read should be performed (e.g. for mmapped buffer).
141     */
[5063]142   
[5292]143    BasicBlock * finalPartialBlock = CreateBasicBlock("partial");
144    BasicBlock * finalEmptyBlock = CreateBasicBlock("empty");
145    BasicBlock * exitBlock = CreateBasicBlock("exit");
[5051]146   
[5234]147    Value * emptyBlockCond = iBuilder->CreateICmpEQ(remainingBytes, iBuilder->getSize(0));
[5051]148    iBuilder->CreateCondBr(emptyBlockCond, finalEmptyBlock, finalPartialBlock);
149    iBuilder->SetInsertPoint(finalPartialBlock);
[5292]150    CreateDoBlockMethodCall();
[5051]151   
152    iBuilder->CreateBr(exitBlock);
153   
154    iBuilder->SetInsertPoint(finalEmptyBlock);
[5292]155    Value * basisBitsPtr = getStreamView("basisBits", blockNo, iBuilder->getInt64(0));
[5260]156    iBuilder->CreateStore(Constant::getNullValue(basisBitsPtr->getType()->getPointerElementType()), basisBitsPtr);
[5051]157    iBuilder->CreateBr(exitBlock);
158   
159    iBuilder->SetInsertPoint(exitBlock);
[5063]160}
[5202]161
[5246]162S2PKernel::S2PKernel(IDISA::IDISA_Builder * builder)
[5283]163: BlockOrientedKernel(builder, "s2p", {Binding{builder->getStreamSetTy(1, 8), "byteStream"}}, {Binding{builder->getStreamSetTy(8, 1), "basisBits"}}, {}, {}, {}) {
[5252]164    setNoTerminateAttribute(true);
[5261]165    setDoBlockUpdatesProducedItemCountsAttribute(false);
[5246]166
[5051]167}
[5246]168
169}
Note: See TracBrowser for help on using the repository browser.