source: icGREP/icgrep-devel/icgrep/kernels/p2s_kernel.cpp @ 5185

Last change on this file since 5185 was 5185, checked in by cameron, 2 years ago

Linear buffers; stdout_kernel; cache-aligned alloca, restructured u8u16, clean-ups

File size: 11.4 KB
Line 
1#include "p2s_kernel.h"
2#include "kernels/kernel.h"
3#include "IDISA/idisa_builder.h"
4#include <llvm/IR/TypeBuilder.h>
5#include <llvm/IR/Type.h>
6#include <iostream>
7#include <stdint.h>
8#include <llvm/Support/FileSystem.h>
9#include <llvm/Support/raw_ostream.h>
10
11
12
13namespace kernel{
14       
15void p2s_step(IDISA::IDISA_Builder * iBuilder, Value * p0, Value * p1, Value * hi_mask, unsigned shift, Value * &s1, Value * &s0) {
16    Value * t0 = iBuilder->simd_if(1, hi_mask, p0, iBuilder->simd_srli(16, p1, shift));
17    Value * t1 = iBuilder->simd_if(1, hi_mask, iBuilder->simd_slli(16, p0, shift), p1);
18    s1 = iBuilder->esimd_mergeh(8, t1, t0);
19    s0 = iBuilder->esimd_mergel(8, t1, t0);
20}
21
22inline void p2s(IDISA::IDISA_Builder * iBuilder, Value * p[], Value * s[]) {
23    Value * bit00004444[2];
24    Value * bit22226666[2];
25    Value * bit11115555[2];
26    Value * bit33337777[2];
27    p2s_step(iBuilder, p[0], p[4], iBuilder->simd_himask(8), 4, bit00004444[1], bit00004444[0]);
28    p2s_step(iBuilder, p[1], p[5], iBuilder->simd_himask(8), 4, bit11115555[1], bit11115555[0]);
29    p2s_step(iBuilder, p[2], p[6], iBuilder->simd_himask(8), 4, bit22226666[1], bit22226666[0]);
30    p2s_step(iBuilder, p[3], p[7], iBuilder->simd_himask(8), 4, bit33337777[1], bit33337777[0]);
31
32    Value * bit00224466[4];
33    Value * bit11335577[4];
34    for (unsigned j = 0; j<2; j++) {
35        p2s_step(iBuilder, bit00004444[j], bit22226666[j],iBuilder->simd_himask(4), 2, bit00224466[2*j+1], bit00224466[2*j]);
36        p2s_step(iBuilder, bit11115555[j], bit33337777[j],iBuilder->simd_himask(4), 2, bit11335577[2*j+1], bit11335577[2*j]);
37    }
38    for (unsigned j = 0; j<4; j++) {
39        p2s_step(iBuilder, bit00224466[j], bit11335577[j], iBuilder->simd_himask(2), 1, s[2*j+1], s[2*j]);
40    }
41}
42               
43void p2sKernel::generateDoBlockMethod() {
44    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
45    Module * m = iBuilder->getModule();
46   
47    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
48   
49    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", doBlockFunction, 0));
50   
51    Value * self = getParameter(doBlockFunction, "self");
52    Value * blockNo = getScalarField(self, blockNoScalar);
53    Value * basisBitsBlock_ptr = getStreamSetBlockPtr(self, "basisBits", blockNo);
54    Value * byteStreamBlock_ptr = getStreamSetBlockPtr(self, "byteStream", blockNo);
55
56    Value * p_bitblock[8];
57    for (unsigned i = 0; i < 8; i++) {
58        p_bitblock[i] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(i)});
59    }
60    Value * s_bytepack[8];
61    p2s(iBuilder, p_bitblock, s_bytepack);
62    for (unsigned j = 0; j < 8; ++j) {
63        iBuilder->CreateBlockAlignedStore(s_bytepack[j], byteStreamBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0), iBuilder->getInt32(j)});
64    }
65    iBuilder->CreateRetVoid();
66    iBuilder->restoreIP(savePoint);
67}
68       
69   
70void p2sKernel_withCompressedOutput::prepareKernel() {
71    KernelBuilder::prepareKernel();
72}
73
74void p2sKernel_withCompressedOutput::generateDoBlockMethod() {
75    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
76    Module * m = iBuilder->getModule();
77    Type * i8PtrTy = iBuilder->getInt8PtrTy(); 
78    Type * i32 = iBuilder->getIntNTy(32); 
79    Type * bitBlockPtrTy = llvm::PointerType::get(iBuilder->getBitBlockType(), 0); 
80   
81    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
82   
83    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", doBlockFunction, 0));
84    Value * self = getParameter(doBlockFunction, "self");
85    Value * blockNo = getScalarField(self, blockNoScalar);
86    Value * basisBitsBlock_ptr = getStreamSetBlockPtr(self, "basisBits", blockNo);
87    Value * delCountBlock_ptr = getStreamSetBlockPtr(self, "deletionCounts", blockNo);
88    Value * byteStreamBlock_ptr = getStreamSetBlockPtr(self, "byteStream", blockNo);
89   
90    Value * p_bitblock[8];
91    for (unsigned i = 0; i < 8; i++) {
92        p_bitblock[i] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(i)});
93    }
94    Value * s_bytepack[8];
95    p2s(iBuilder, p_bitblock, s_bytepack);
96   
97    unsigned units_per_register = iBuilder->getBitBlockWidth()/8;
98   
99    Value * unit_counts = iBuilder->fwCast(units_per_register, iBuilder->CreateBlockAlignedLoad(delCountBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0)}));
100   
101    Value * output_ptr = iBuilder->CreateBitCast(byteStreamBlock_ptr, i8PtrTy);
102    Value * offset = ConstantInt::get(i32, 0);
103   
104    for (unsigned j = 0; j < 8; ++j) {
105        iBuilder->CreateAlignedStore(s_bytepack[j], iBuilder->CreateBitCast(iBuilder->CreateGEP(output_ptr, offset), bitBlockPtrTy), 1);
106        offset = iBuilder->CreateZExt(iBuilder->CreateExtractElement(unit_counts, iBuilder->getInt32(j)), i32);
107    }
108    iBuilder->CreateRetVoid();
109    iBuilder->restoreIP(savePoint);
110}
111   
112   
113void p2s_16Kernel::generateDoBlockMethod() {
114    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
115    Module * m = iBuilder->getModule();
116   
117    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
118   
119    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", doBlockFunction, 0));
120    Value * self = getParameter(doBlockFunction, "self");
121    Value * blockNo = getScalarField(self, blockNoScalar);
122    Value * basisBitsBlock_ptr = getStreamSetBlockPtr(self, "basisBits", blockNo);
123    Value * i16StreamBlock_ptr = getStreamSetBlockPtr(self, "i16Stream", blockNo);
124   
125    Value * hi_input[8];
126    for (unsigned j = 0; j < 8; ++j) {
127        hi_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j)});
128    }
129    Value * hi_bytes[8];
130    p2s(iBuilder, hi_input, hi_bytes);
131   
132    Value * lo_input[8];
133    for (unsigned j = 0; j < 8; ++j) {
134        lo_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j+8)});
135    }
136    Value * lo_bytes[8];
137    p2s(iBuilder, lo_input, lo_bytes);
138   
139    for (unsigned j = 0; j < 8; ++j) {
140        Value * merge0 = iBuilder->bitCast(iBuilder->esimd_mergel(8, hi_bytes[j], lo_bytes[j]));
141        Value * merge1 = iBuilder->bitCast(iBuilder->esimd_mergeh(8, hi_bytes[j], lo_bytes[j]));
142        iBuilder->CreateBlockAlignedStore(merge0, i16StreamBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0), iBuilder->getInt32(2*j)});
143        iBuilder->CreateBlockAlignedStore(merge1, i16StreamBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0), iBuilder->getInt32(2*j+1)});
144    }
145    iBuilder->CreateRetVoid();
146    iBuilder->restoreIP(savePoint);
147}
148       
149
150void p2s_16Kernel_withCompressedOutput::prepareKernel() {
151    KernelBuilder::prepareKernel();
152}
153   
154
155void p2s_16Kernel_withCompressedOutput::generateDoBlockMethod() {
156    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
157    Module * m = iBuilder->getModule();
158    Type * i32 = iBuilder->getIntNTy(32); 
159    Type * bitBlockPtrTy = llvm::PointerType::get(iBuilder->getBitBlockType(), 0); 
160
161    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
162   
163    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", doBlockFunction, 0));
164    Constant * stride = ConstantInt::get(iBuilder->getSizeTy(), iBuilder->getStride());
165
166    Value * self = getParameter(doBlockFunction, "self");
167    Value * blockNo = getScalarField(self, blockNoScalar);
168    Value * basisBitsBlock_ptr = getStreamSetBlockPtr(self, "basisBits", blockNo);
169    Value * delCountBlock_ptr = getStreamSetBlockPtr(self, "deletionCounts", blockNo);
170    Value * i16UnitsGenerated = getProducedItemCount(self); // units generated to buffer
171    Value * i16BlockNo = iBuilder->CreateUDiv(i16UnitsGenerated, stride);
172   
173    Value * i16StreamBase_ptr = iBuilder->CreateBitCast(getStreamSetBlockPtr(self, "i16Stream", i16BlockNo), PointerType::get(iBuilder->getInt16Ty(), 0));
174   
175    Value * u16_output_ptr = iBuilder->CreateGEP(i16StreamBase_ptr, iBuilder->CreateURem(i16UnitsGenerated, stride));
176
177   
178    Value * hi_input[8];
179    for (unsigned j = 0; j < 8; ++j) {
180        hi_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j)});
181    }
182    Value * hi_bytes[8];
183    p2s(iBuilder, hi_input, hi_bytes);
184   
185    Value * lo_input[8];
186    for (unsigned j = 0; j < 8; ++j) {
187        lo_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j+8)});
188    }
189    Value * lo_bytes[8];
190    p2s(iBuilder, lo_input, lo_bytes);
191   
192    unsigned UTF_16_units_per_register = iBuilder->getBitBlockWidth()/16;
193   
194    Value * unit_counts = iBuilder->fwCast(UTF_16_units_per_register, iBuilder->CreateBlockAlignedLoad(delCountBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0)}));
195   
196    Value * offset = ConstantInt::get(i32, 0);
197   
198    for (unsigned j = 0; j < 8; ++j) {
199        Value * merge0 = iBuilder->bitCast(iBuilder->esimd_mergel(8, hi_bytes[j], lo_bytes[j]));
200        Value * merge1 = iBuilder->bitCast(iBuilder->esimd_mergeh(8, hi_bytes[j], lo_bytes[j]));
201        //iBuilder->CallPrintRegister("merge0", merge0);
202        iBuilder->CreateAlignedStore(merge0, iBuilder->CreateBitCast(iBuilder->CreateGEP(u16_output_ptr, offset), bitBlockPtrTy), 1);
203        offset = iBuilder->CreateZExt(iBuilder->CreateExtractElement(unit_counts, iBuilder->getInt32(2*j)), i32);
204        //iBuilder->CallPrintInt("offset", offset);
205        iBuilder->CreateAlignedStore(merge1, iBuilder->CreateBitCast(iBuilder->CreateGEP(u16_output_ptr, offset), bitBlockPtrTy), 1);
206        //iBuilder->CallPrintRegister("merge1", merge1);
207        offset = iBuilder->CreateZExt(iBuilder->CreateExtractElement(unit_counts, iBuilder->getInt32(2*j+1)), i32);
208        //iBuilder->CallPrintInt("offset", offset);
209    }
210   
211    i16UnitsGenerated = iBuilder->CreateAdd(i16UnitsGenerated, iBuilder->CreateZExt(offset, iBuilder->getSizeTy()));
212    setProducedItemCount(self, i16UnitsGenerated);
213    iBuilder->CreateRetVoid();
214    iBuilder->restoreIP(savePoint);
215}
216
217void p2s_16Kernel_withCompressedOutput::generateFinalBlockMethod() {
218    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
219    Module * m = iBuilder->getModule();
220    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
221    Function * finalBlockFunction = m->getFunction(mKernelName + finalBlock_suffix);
222    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "fb_entry", finalBlockFunction, 0));
223    // Final Block arguments: self, remaining, then the standard DoBlock args.
224    Function::arg_iterator args = finalBlockFunction->arg_begin();
225    Value * self = &*(args++);
226    /* Skip "remaining" arg */ args++;
227    std::vector<Value *> doBlockArgs = {self};
228    while (args != finalBlockFunction->arg_end()){
229        doBlockArgs.push_back(&*args++);
230    }
231    Value * i16UnitsGenerated = getProducedItemCount(self); // units generated to buffer
232
233    iBuilder->CreateCall(doBlockFunction, doBlockArgs);
234    i16UnitsGenerated = getProducedItemCount(self); // units generated to buffer
235    for (unsigned i = 0; i < mStreamSetOutputs.size(); i++) {
236        Value * ssStructPtr = getStreamSetStructPtr(self, mStreamSetOutputs[i].ssName);
237        Value * producerPosPtr = mStreamSetOutputBuffers[i]->getProducerPosPtr(ssStructPtr);
238        iBuilder->CreateAlignedStore(i16UnitsGenerated, producerPosPtr, sizeof(size_t))->setOrdering(AtomicOrdering::Release);
239    }
240    iBuilder->CreateRetVoid();
241    iBuilder->restoreIP(savePoint);
242}
243   
244   
245}
Note: See TracBrowser for help on using the repository browser.