source: icGREP/icgrep-devel/icgrep/kernels/p2s_kernel.cpp @ 5109

Last change on this file since 5109 was 5109, checked in by cameron, 3 years ago

u8u16 pipeline

File size: 13.9 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    setDoBlockReturnType(iBuilder->getInt32Ty());
72    KernelBuilder::prepareKernel();
73}
74
75void p2sKernel_withCompressedOutput::generateDoBlockMethod() {
76    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
77    Module * m = iBuilder->getModule();
78    Type * i8PtrTy = iBuilder->getInt8PtrTy(); 
79    Type * i32 = iBuilder->getIntNTy(32); 
80    Type * bitBlockPtrTy = llvm::PointerType::get(iBuilder->getBitBlockType(), 0); 
81   
82    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
83   
84    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", doBlockFunction, 0));
85    Value * self = getParameter(doBlockFunction, "self");
86    Value * blockNo = getScalarField(self, blockNoScalar);
87    Value * basisBitsBlock_ptr = getStreamSetBlockPtr(self, "basisBits", blockNo);
88    Value * delCountBlock_ptr = getStreamSetBlockPtr(self, "deletionCounts", blockNo);
89    Value * byteStreamBlock_ptr = getStreamSetBlockPtr(self, "byteStream", blockNo);
90   
91    Value * p_bitblock[8];
92    for (unsigned i = 0; i < 8; i++) {
93        p_bitblock[i] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(i)});
94    }
95    Value * s_bytepack[8];
96    p2s(iBuilder, p_bitblock, s_bytepack);
97   
98    unsigned units_per_register = iBuilder->getBitBlockWidth()/8;
99   
100    Value * unit_counts = iBuilder->fwCast(units_per_register, iBuilder->CreateBlockAlignedLoad(delCountBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0)}));
101   
102    Value * output_ptr = iBuilder->CreateBitCast(byteStreamBlock_ptr, i8PtrTy);
103    Value * offset = ConstantInt::get(i32, 0);
104   
105    for (unsigned j = 0; j < 8; ++j) {
106        iBuilder->CreateAlignedStore(s_bytepack[j], iBuilder->CreateBitCast(iBuilder->CreateGEP(output_ptr, offset), bitBlockPtrTy), 1);
107        offset = iBuilder->CreateZExt(iBuilder->CreateExtractElement(unit_counts, iBuilder->getInt32(j)), i32);
108    }
109    iBuilder->CreateRet(offset);
110    iBuilder->restoreIP(savePoint);
111}
112   
113   
114void p2s_16Kernel::generateDoBlockMethod() {
115    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
116    Module * m = iBuilder->getModule();
117   
118    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
119   
120    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", doBlockFunction, 0));
121    Value * self = getParameter(doBlockFunction, "self");
122    Value * blockNo = getScalarField(self, blockNoScalar);
123    Value * basisBitsBlock_ptr = getStreamSetBlockPtr(self, "basisBits", blockNo);
124    Value * i16StreamBlock_ptr = getStreamSetBlockPtr(self, "i16Stream", blockNo);
125   
126    Value * hi_input[8];
127    for (unsigned j = 0; j < 8; ++j) {
128        hi_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j)});
129    }
130    Value * hi_bytes[8];
131    p2s(iBuilder, hi_input, hi_bytes);
132   
133    Value * lo_input[8];
134    for (unsigned j = 0; j < 8; ++j) {
135        lo_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j+8)});
136    }
137    Value * lo_bytes[8];
138    p2s(iBuilder, lo_input, lo_bytes);
139   
140    for (unsigned j = 0; j < 8; ++j) {
141        Value * merge0 = iBuilder->bitCast(iBuilder->esimd_mergel(8, hi_bytes[j], lo_bytes[j]));
142        Value * merge1 = iBuilder->bitCast(iBuilder->esimd_mergeh(8, hi_bytes[j], lo_bytes[j]));
143        iBuilder->CreateBlockAlignedStore(merge0, i16StreamBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0), iBuilder->getInt32(2*j)});
144        iBuilder->CreateBlockAlignedStore(merge1, i16StreamBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0), iBuilder->getInt32(2*j+1)});
145    }
146    iBuilder->CreateRetVoid();
147    iBuilder->restoreIP(savePoint);
148}
149       
150
151void p2s_16Kernel_withCompressedOutput::prepareKernel() {
152    KernelBuilder::prepareKernel();
153}
154   
155    static Function * create_write(Module * const mod) {
156        Function * write = mod->getFunction("write");
157        if (write == nullptr) {
158            FunctionType *write_type =
159            TypeBuilder<long(int, char *, long), false>::get(mod->getContext());
160            write = cast<Function>(mod->getOrInsertFunction("write", write_type,
161                                                            AttributeSet().addAttribute(mod->getContext(), 2U, Attribute::NoAlias)));
162        }
163        return write;
164    }
165   
166   
167   
168void p2s_16Kernel_withCompressedOutput::generateDoBlockMethod() {
169    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
170    Module * m = iBuilder->getModule();
171    Type * i32 = iBuilder->getIntNTy(32); 
172    Type * bitBlockPtrTy = llvm::PointerType::get(iBuilder->getBitBlockType(), 0); 
173    Type * i8PtrTy = iBuilder->getInt8PtrTy();
174
175    Function * writefn = create_write(m);
176
177    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
178   
179    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", doBlockFunction, 0));
180    Value * self = getParameter(doBlockFunction, "self");
181    Value * blockNo = getScalarField(self, blockNoScalar);
182    Value * basisBitsBlock_ptr = getStreamSetBlockPtr(self, "basisBits", blockNo);
183    Value * delCountBlock_ptr = getStreamSetBlockPtr(self, "deletionCounts", blockNo);
184    Value * i16UnitsGenerated = getScalarField(self, "unitsGenerated");  // units generated to buffer
185    Value * i16UnitsWritten = getScalarField(self, "unitsWritten");  // units written to stdout
186   
187    Value * i16StreamBase_ptr = iBuilder->CreateBitCast(getStreamSetBasePtr(self, "i16Stream"), PointerType::get(iBuilder->getInt16Ty(), 0));
188    Value * u16_output_ptr = iBuilder->CreateGEP(i16StreamBase_ptr, iBuilder->CreateSub(i16UnitsGenerated, i16UnitsWritten));
189
190   
191    Value * hi_input[8];
192    for (unsigned j = 0; j < 8; ++j) {
193        hi_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j)});
194    }
195    Value * hi_bytes[8];
196    p2s(iBuilder, hi_input, hi_bytes);
197   
198    Value * lo_input[8];
199    for (unsigned j = 0; j < 8; ++j) {
200        lo_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j+8)});
201    }
202    Value * lo_bytes[8];
203    p2s(iBuilder, lo_input, lo_bytes);
204   
205    unsigned UTF_16_units_per_register = iBuilder->getBitBlockWidth()/16;
206   
207    Value * unit_counts = iBuilder->fwCast(UTF_16_units_per_register, iBuilder->CreateBlockAlignedLoad(delCountBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0)}));
208   
209    Value * offset = ConstantInt::get(i32, 0);
210   
211    for (unsigned j = 0; j < 8; ++j) {
212        Value * merge0 = iBuilder->bitCast(iBuilder->esimd_mergel(8, hi_bytes[j], lo_bytes[j]));
213        Value * merge1 = iBuilder->bitCast(iBuilder->esimd_mergeh(8, hi_bytes[j], lo_bytes[j]));
214        //iBuilder->CallPrintRegister("merge0", merge0);
215        iBuilder->CreateAlignedStore(merge0, iBuilder->CreateBitCast(iBuilder->CreateGEP(u16_output_ptr, offset), bitBlockPtrTy), 1);
216        offset = iBuilder->CreateZExt(iBuilder->CreateExtractElement(unit_counts, iBuilder->getInt32(2*j)), i32);
217        //iBuilder->CallPrintInt("offset", offset);
218        iBuilder->CreateAlignedStore(merge1, iBuilder->CreateBitCast(iBuilder->CreateGEP(u16_output_ptr, offset), bitBlockPtrTy), 1);
219        //iBuilder->CallPrintRegister("merge1", merge1);
220        offset = iBuilder->CreateZExt(iBuilder->CreateExtractElement(unit_counts, iBuilder->getInt32(2*j+1)), i32);
221        //iBuilder->CallPrintInt("offset", offset);
222    }
223   
224    i16UnitsGenerated = iBuilder->CreateAdd(i16UnitsGenerated, iBuilder->CreateZExt(offset, iBuilder->getSizeTy()));
225    setScalarField(self, "unitsGenerated", i16UnitsGenerated);
226   
227    Value * unitsInBuffer = iBuilder->CreateSub(i16UnitsGenerated, i16UnitsWritten);
228    Value * lessThanABlockRemaining = iBuilder->CreateICmpUGT(unitsInBuffer, ConstantInt::get(iBuilder->getSizeTy(), (getStreamSetBufferSize(self, "i16Stream") - 1) * iBuilder->getBitBlockWidth()));
229    BasicBlock * flushStmts = BasicBlock::Create(iBuilder->getContext(), "flush", doBlockFunction, 0);
230    BasicBlock * exitStmts = BasicBlock::Create(iBuilder->getContext(), "exit", doBlockFunction, 0);
231    iBuilder->CreateCondBr(lessThanABlockRemaining, flushStmts, exitStmts);
232   
233    iBuilder->SetInsertPoint(flushStmts);
234    iBuilder->CreateCall(writefn, std::vector<Value *>({iBuilder->getInt32(1), iBuilder->CreateBitCast(i16StreamBase_ptr, i8PtrTy), iBuilder->CreateAdd(unitsInBuffer, unitsInBuffer)}));
235    setScalarField(self, "unitsWritten", i16UnitsGenerated); // Everything generated has now been written.
236    iBuilder->CreateBr(exitStmts);
237   
238    iBuilder->SetInsertPoint(exitStmts);
239    iBuilder->CreateRetVoid();
240    iBuilder->restoreIP(savePoint);
241}
242
243void p2s_16Kernel_withCompressedOutput::generateFinalBlockMethod() {
244    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
245    Module * m = iBuilder->getModule();
246    Type * i32 = iBuilder->getIntNTy(32); 
247    Type * bitBlockPtrTy = llvm::PointerType::get(iBuilder->getBitBlockType(), 0); 
248    Type * i8PtrTy = iBuilder->getInt8PtrTy();
249    Function * writefn = create_write(m);
250    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
251    Function * finalBlockFunction = m->getFunction(mKernelName + finalBlock_suffix);
252    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "fb_entry", finalBlockFunction, 0));
253    // Final Block arguments: self, remaining, then the standard DoBlock args.
254    Function::arg_iterator args = finalBlockFunction->arg_begin();
255    Value * self = &*(args++);
256    /* Skip "remaining" arg */ args++;
257    std::vector<Value *> doBlockArgs = {self};
258    while (args != finalBlockFunction->arg_end()){
259        doBlockArgs.push_back(&*args++);
260    }
261    iBuilder->CreateCall(doBlockFunction, doBlockArgs);
262    Value * i16UnitsGenerated = getScalarField(self, "unitsGenerated");  // units generated to buffer
263    Value * i16UnitsWritten = getScalarField(self, "unitsWritten");  // units written to stdout
264    Value * unitsInBuffer = iBuilder->CreateSub(i16UnitsGenerated, i16UnitsWritten);
265    Value * mustFlush = iBuilder->CreateICmpUGT(unitsInBuffer, ConstantInt::get(iBuilder->getSizeTy(), 0));
266   
267    BasicBlock * flushStmts = BasicBlock::Create(iBuilder->getContext(), "flush", finalBlockFunction, 0);
268    BasicBlock * exitStmts = BasicBlock::Create(iBuilder->getContext(), "exit", finalBlockFunction, 0);
269    iBuilder->CreateCondBr(mustFlush, flushStmts, exitStmts);
270   
271    iBuilder->SetInsertPoint(flushStmts);
272    Value * i16StreamBase_ptr = iBuilder->CreateBitCast(getStreamSetBasePtr(self, "i16Stream"), PointerType::get(iBuilder->getInt16Ty(), 0));
273    iBuilder->CreateCall(writefn, std::vector<Value *>({iBuilder->getInt32(1), iBuilder->CreateBitCast(i16StreamBase_ptr, i8PtrTy), iBuilder->CreateAdd(unitsInBuffer, unitsInBuffer)}));
274    setScalarField(self, "unitsWritten", i16UnitsGenerated); // Everything generated has now been written.
275    iBuilder->CreateBr(exitStmts);
276   
277    iBuilder->SetInsertPoint(exitStmts);
278    iBuilder->CreateRetVoid();
279    iBuilder->restoreIP(savePoint);
280}
281   
282   
283}
Note: See TracBrowser for help on using the repository browser.