source: icGREP/icgrep-devel/icgrep/kernels/s2p_kernel.cpp @ 5436

Last change on this file since 5436 was 5436, checked in by nmedfort, 2 years ago

Continued refactoring work. PabloKernel? now abstract base type with a 'generatePabloMethod' hook to generate Pablo code.

File size: 6.7 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 "s2p_kernel.h"
7#include <kernels/kernel_builder.h>
8#include <llvm/Support/raw_ostream.h>
9
10using namespace llvm;
11
12namespace kernel {
13
14const int PACK_LANES = 1;
15
16void s2p_step(IDISA::IDISA_Builder * const iBuilder, Value * s0, Value * s1, Value * hi_mask, unsigned shift, Value * &p0, Value * &p1) {
17    Value * t0 = nullptr;
18    Value * t1 = nullptr;
19    if ((iBuilder->getBitBlockWidth() == 256) && (PACK_LANES == 2)) {
20        Value * x0 = iBuilder->esimd_mergel(128, s0, s1);
21        Value * x1 = iBuilder->esimd_mergeh(128, s0, s1);
22        t0 = iBuilder->hsimd_packh_in_lanes(PACK_LANES, 16, x0, x1);
23        t1 = iBuilder->hsimd_packl_in_lanes(PACK_LANES, 16, x0, x1);
24    } else {
25        t0 = iBuilder->hsimd_packh(16, s0, s1);
26        t1 = iBuilder->hsimd_packl(16, s0, s1);
27    }
28    p0 = iBuilder->simd_if(1, hi_mask, t0, iBuilder->simd_srli(16, t1, shift));
29    p1 = iBuilder->simd_if(1, hi_mask, iBuilder->simd_slli(16, t0, shift), t1);
30}
31
32void s2p(IDISA::IDISA_Builder * const iBuilder, Value * input[], Value * output[]) {
33    Value * bit00224466[4];
34    Value * bit11335577[4];
35
36    for (unsigned i = 0; i < 4; i++) {
37        Value * s0 = input[2 * i];
38        Value * s1 = input[2 * i + 1];
39        s2p_step(iBuilder, s0, s1, iBuilder->simd_himask(2), 1, bit00224466[i], bit11335577[i]);
40    }
41    Value * bit00004444[2];
42    Value * bit22226666[2];
43    Value * bit11115555[2];
44    Value * bit33337777[2];
45    for (unsigned j = 0; j<2; j++) {
46        s2p_step(iBuilder, bit00224466[2*j], bit00224466[2*j+1],
47                 iBuilder->simd_himask(4), 2, bit00004444[j], bit22226666[j]);
48        s2p_step(iBuilder, bit11335577[2*j], bit11335577[2*j+1],
49                 iBuilder->simd_himask(4), 2, bit11115555[j], bit33337777[j]);
50    }
51    s2p_step(iBuilder, bit00004444[0], bit00004444[1], iBuilder->simd_himask(8), 4, output[0], output[4]);
52    s2p_step(iBuilder, bit11115555[0], bit11115555[1], iBuilder->simd_himask(8), 4, output[1], output[5]);
53    s2p_step(iBuilder, bit22226666[0], bit22226666[1], iBuilder->simd_himask(8), 4, output[2], output[6]);
54    s2p_step(iBuilder, bit33337777[0], bit33337777[1], iBuilder->simd_himask(8), 4, output[3], output[7]);
55}
56
57/* Alternative transposition model, but small field width packs are problematic. */
58#if 0
59void s2p_ideal(IDISA::IDISA_Builder * const iBuilder, Value * input[], Value * output[]) {
60    Value * hi_nybble[4];
61    Value * lo_nybble[4];
62    for (unsigned i = 0; i<4; i++) {
63        Value * s0 = input[2*i];
64        Value * s1 = input[2*i+1];
65        hi_nybble[i] = iBuilder->hsimd_packh(8, s0, s1);
66        lo_nybble[i] = iBuilder->hsimd_packl(8, s0, s1);
67    }
68    Value * pair01[2];
69    Value * pair23[2];
70    Value * pair45[2];
71    Value * pair67[2];
72    for (unsigned i = 0; i<2; i++) {
73        pair01[i] = iBuilder->hsimd_packh(4, hi_nybble[2*i], hi_nybble[2*i+1]);
74        pair23[i] = iBuilder->hsimd_packl(4, hi_nybble[2*i], hi_nybble[2*i+1]);
75        pair45[i] = iBuilder->hsimd_packh(4, lo_nybble[2*i], lo_nybble[2*i+1]);
76        pair67[i] = iBuilder->hsimd_packl(4, lo_nybble[2*i], lo_nybble[2*i+1]);
77    }
78    output[0] = iBuilder->hsimd_packh(2, pair01[0], pair01[1]);
79    output[1] = iBuilder->hsimd_packl(2, pair01[0], pair01[1]);
80    output[2] = iBuilder->hsimd_packh(2, pair23[0], pair23[1]);
81    output[3] = iBuilder->hsimd_packl(2, pair23[0], pair23[1]);
82    output[4] = iBuilder->hsimd_packh(2, pair45[0], pair45[1]);
83    output[5] = iBuilder->hsimd_packl(2, pair45[0], pair45[1]);
84    output[6] = iBuilder->hsimd_packh(2, pair67[0], pair67[1]);
85    output[7] = iBuilder->hsimd_packl(2, pair67[0], pair67[1]);
86}
87#endif
88   
89#if 0
90void generateS2P_16Kernel(Module *, IDISA::IDISA_Builder * const iBuilder, KernelBuilder * kBuilder) {
91    kBuilder->addInputStream(16, "unit_pack");
92    for(unsigned i = 0; i < 16; i++) {
93            kBuilder->addOutputStream(1);
94    }
95    kBuilder->prepareFunction();
96
97    Value * ptr = kBuilder->getInputStream(0);
98
99    Value * lo[8];
100    Value * hi[8];
101    for (unsigned i = 0; i < 8; i++) {
102        Value * s0 = iBuilder->CreateBlockAlignedLoad(ptr, {iBuilder->getInt32(0), iBuilder->getInt32(2 * i)});
103        Value * s1 = iBuilder->CreateBlockAlignedLoad(ptr, {iBuilder->getInt32(0), iBuilder->getInt32(2 * i + 1)});
104        lo[i] = iBuilder->hsimd_packl(16, s0, s1);
105        hi[i] = iBuilder->hsimd_packh(16, s0, s1);
106    }
107
108    Value * output[16];
109    s2p(iBuilder, lo, output);
110    s2p(iBuilder, hi, output + 8);
111    for (unsigned j = 0; j < 16; j++) {
112        iBuilder->CreateBlockAlignedStore(output[j], kBuilder->getOutputStream(j));
113    }
114    kBuilder->finalize();
115}   
116#endif
117   
118void S2PKernel::generateDoBlockMethod() {
119    Value * bytepack[8];
120    for (unsigned i = 0; i < 8; i++) {
121        if (mAligned) {
122            bytepack[i] = loadInputStreamPack("byteStream", iBuilder->getInt32(0), iBuilder->getInt32(i));
123        } else {
124            Value * ptr = getInputStreamPackPtr("byteStream", iBuilder->getInt32(0), iBuilder->getInt32(i));
125            // CreateLoad defaults to aligned here, so we need to force the alignment to 1 byte.
126            bytepack[i] = iBuilder->CreateAlignedLoad(ptr, 1);
127        }
128    }
129    Value * basisbits[8];
130    s2p(iBuilder, bytepack, basisbits);
131    for (unsigned i = 0; i < 8; ++i) {
132        storeOutputStreamBlock("basisBits", iBuilder->getInt32(i), basisbits[i]);
133    }
134}
135
136void S2PKernel::generateFinalBlockMethod(Value * remainingBytes) {
137    /* Prepare the s2p final block function:
138     assumption: if remaining bytes is greater than 0, it is safe to read a full block of bytes.
139     if remaining bytes is zero, no read should be performed (e.g. for mmapped buffer).
140     */
141   
142    BasicBlock * finalPartialBlock = CreateBasicBlock("partial");
143    BasicBlock * finalEmptyBlock = CreateBasicBlock("empty");
144    BasicBlock * exitBlock = CreateBasicBlock("exit");
145   
146    Value * emptyBlockCond = iBuilder->CreateICmpEQ(remainingBytes, iBuilder->getSize(0));
147    iBuilder->CreateCondBr(emptyBlockCond, finalEmptyBlock, finalPartialBlock);
148    iBuilder->SetInsertPoint(finalPartialBlock);
149    CreateDoBlockMethodCall();
150   
151    iBuilder->CreateBr(exitBlock);
152   
153    iBuilder->SetInsertPoint(finalEmptyBlock);
154
155    for (unsigned i = 0; i < 8; ++i) {
156        storeOutputStreamBlock("basisBits", iBuilder->getInt32(i), Constant::getNullValue(iBuilder->getBitBlockType()));
157    }
158
159    iBuilder->CreateBr(exitBlock);
160   
161    iBuilder->SetInsertPoint(exitBlock);
162}
163
164S2PKernel::S2PKernel(const std::unique_ptr<KernelBuilder> & b, bool aligned)
165: BlockOrientedKernel(aligned ? "s2p" : "s2p_unaligned",
166    {Binding{b->getStreamSetTy(1, 8), "byteStream"}}, {Binding{b->getStreamSetTy(8, 1), "basisBits"}}, {}, {}, {}),
167  mAligned(aligned) {
168    setNoTerminateAttribute(true);
169}
170
171}
Note: See TracBrowser for help on using the repository browser.