source: icGREP/icgrep-devel/icgrep/kernels/p2s_kernel.cpp @ 5097

Last change on this file since 5097 was 5097, checked in by cameron, 3 years ago

Kernel/interface progress

File size: 9.7 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 = getCircularBufferBlockPointer(self, "basisBits", blockNo);
54    Value * byteStreamBlock_ptr = getCircularBufferBlockPointer(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   
86   
87    Value * self = getParameter(doBlockFunction, "self");
88    Value * blockNo = getScalarField(self, blockNoScalar);
89    Value * basisBitsBlock_ptr = getCircularBufferBlockPointer(self, "basisBits", blockNo);
90    Value * delCountBlock_ptr = getCircularBufferBlockPointer(self, "deletionCounts", blockNo);
91    Value * byteStreamBlock_ptr = getCircularBufferBlockPointer(self, "byteStream", blockNo);
92
93    Value * p_bitblock[8];
94    for (unsigned i = 0; i < 8; i++) {
95        p_bitblock[i] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(i)});
96    }
97    Value * s_bytepack[8];
98    p2s(iBuilder, p_bitblock, s_bytepack);
99   
100    unsigned units_per_register = iBuilder->getBitBlockWidth()/8;
101   
102    Value * unit_counts = iBuilder->fwCast(units_per_register, iBuilder->CreateBlockAlignedLoad(delCountBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0)}));
103   
104    Value * output_ptr = iBuilder->CreateBitCast(byteStreamBlock_ptr, i8PtrTy);
105    Value * offset = ConstantInt::get(i32, 0);
106   
107    for (unsigned j = 0; j < 8; ++j) {
108        iBuilder->CreateAlignedStore(s_bytepack[j], iBuilder->CreateBitCast(iBuilder->CreateGEP(output_ptr, offset), bitBlockPtrTy), 1);
109        offset = iBuilder->CreateZExt(iBuilder->CreateExtractElement(unit_counts, iBuilder->getInt32(j)), i32);
110    }
111    iBuilder->CreateRet(offset);
112    iBuilder->restoreIP(savePoint);
113}
114   
115   
116void p2s_16Kernel::generateDoBlockMethod() {
117    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
118    Module * m = iBuilder->getModule();
119   
120    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
121   
122    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", doBlockFunction, 0));
123   
124    Value * self = getParameter(doBlockFunction, "self");
125    Value * blockNo = getScalarField(self, blockNoScalar);
126    Value * basisBitsBlock_ptr = getCircularBufferBlockPointer(self, "basisBits", blockNo);
127    Value * i16StreamBlock_ptr = getCircularBufferBlockPointer(self, "i16Stream", blockNo);
128   
129    Value * hi_input[8];
130    for (unsigned j = 0; j < 8; ++j) {
131        hi_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j)});
132    }
133    Value * hi_bytes[8];
134    p2s(iBuilder, hi_input, hi_bytes);
135   
136    Value * lo_input[8];
137    for (unsigned j = 0; j < 8; ++j) {
138        lo_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j+8)});
139    }
140    Value * lo_bytes[8];
141    p2s(iBuilder, lo_input, lo_bytes);
142   
143    for (unsigned j = 0; j < 8; ++j) {
144        Value * merge0 = iBuilder->bitCast(iBuilder->esimd_mergel(8, hi_bytes[j], lo_bytes[j]));
145        Value * merge1 = iBuilder->bitCast(iBuilder->esimd_mergeh(8, hi_bytes[j], lo_bytes[j]));
146        iBuilder->CreateBlockAlignedStore(merge0, i16StreamBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0), iBuilder->getInt32(2*j)});
147        iBuilder->CreateBlockAlignedStore(merge1, i16StreamBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0), iBuilder->getInt32(2*j+1)});
148    }
149    iBuilder->CreateRetVoid();
150    iBuilder->restoreIP(savePoint);
151}
152       
153
154void p2s_16Kernel_withCompressedOutput::prepareKernel() {
155    setDoBlockReturnType(iBuilder->getInt32Ty());
156    KernelBuilder::prepareKernel();
157}
158   
159void p2s_16Kernel_withCompressedOutput::generateDoBlockMethod() {
160    IDISA::IDISA_Builder::InsertPoint savePoint = iBuilder->saveIP();
161    Module * m = iBuilder->getModule();
162    Type * i32 = iBuilder->getIntNTy(32); 
163    Type * bitBlockPtrTy = llvm::PointerType::get(iBuilder->getBitBlockType(), 0); 
164   
165    Function * doBlockFunction = m->getFunction(mKernelName + doBlock_suffix);
166   
167    iBuilder->SetInsertPoint(BasicBlock::Create(iBuilder->getContext(), "entry", doBlockFunction, 0));
168   
169    Value * self = getParameter(doBlockFunction, "self");
170    Value * blockNo = getScalarField(self, blockNoScalar);
171    Value * basisBitsBlock_ptr = getCircularBufferBlockPointer(self, "basisBits", blockNo);
172    Value * delCountBlock_ptr = getCircularBufferBlockPointer(self, "deletionCounts", blockNo);
173    Value * i16StreamBlock_ptr = getCircularBufferBlockPointer(self, "i16Stream", blockNo);
174
175    Value * hi_input[8];
176    for (unsigned j = 0; j < 8; ++j) {
177        hi_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j)});
178    }
179    Value * hi_bytes[8];
180    p2s(iBuilder, hi_input, hi_bytes);
181   
182    Value * lo_input[8];
183    for (unsigned j = 0; j < 8; ++j) {
184        lo_input[j] = iBuilder->CreateBlockAlignedLoad(basisBitsBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(j+8)});
185    }
186    Value * lo_bytes[8];
187    p2s(iBuilder, lo_input, lo_bytes);
188   
189    unsigned UTF_16_units_per_register = iBuilder->getBitBlockWidth()/16;
190   
191    Value * unit_counts = iBuilder->fwCast(UTF_16_units_per_register, iBuilder->CreateBlockAlignedLoad(delCountBlock_ptr, {iBuilder->getInt32(0), iBuilder->getInt32(0)}));
192   
193    Value * u16_output_ptr = iBuilder->CreateBitCast(i16StreamBlock_ptr, PointerType::get(iBuilder->getInt16Ty(), 0));
194    Value * offset = ConstantInt::get(i32, 0);
195   
196    for (unsigned j = 0; j < 8; ++j) {
197        Value * merge0 = iBuilder->bitCast(iBuilder->esimd_mergel(8, hi_bytes[j], lo_bytes[j]));
198        Value * merge1 = iBuilder->bitCast(iBuilder->esimd_mergeh(8, hi_bytes[j], lo_bytes[j]));
199        //iBuilder->CallPrintRegister("merge0", merge0);
200        iBuilder->CreateAlignedStore(merge0, iBuilder->CreateBitCast(iBuilder->CreateGEP(u16_output_ptr, offset), bitBlockPtrTy), 1);
201        offset = iBuilder->CreateZExt(iBuilder->CreateExtractElement(unit_counts, iBuilder->getInt32(2*j)), i32);
202        //iBuilder->CallPrintInt("offset", offset);
203        iBuilder->CreateAlignedStore(merge1, iBuilder->CreateBitCast(iBuilder->CreateGEP(u16_output_ptr, offset), bitBlockPtrTy), 1);
204        //iBuilder->CallPrintRegister("merge1", merge1);
205        offset = iBuilder->CreateZExt(iBuilder->CreateExtractElement(unit_counts, iBuilder->getInt32(2*j+1)), i32);
206        //iBuilder->CallPrintInt("offset", offset);
207    }
208    iBuilder->CreateRet(offset);
209    iBuilder->restoreIP(savePoint);
210}
211       
212}
Note: See TracBrowser for help on using the repository browser.