source: icGREP/icgrep-devel/icgrep/kernels/pdep_kernel.cpp @ 6184

Last change on this file since 6184 was 6184, checked in by nmedfort, 9 months ago

Initial version of PipelineKernel? + revised StreamSet? model.

File size: 20.1 KB
Line 
1/*
2 *  Copyright (c) 2018 International Characters.
3 *  This software is licensed to the public under the Open Software License 3.0.
4 */
5#include "pdep_kernel.h"
6#include <kernels/kernel_builder.h>
7#include <llvm/Support/raw_ostream.h>
8#include <toolchain/toolchain.h>
9#include <toolchain/driver.h>
10#include <toolchain/cpudriver.h>
11#include <IR_Gen/idisa_target.h>
12#include <llvm/IR/Intrinsics.h>
13#include <llvm/IR/Module.h>
14
15
16using namespace llvm;
17
18namespace kernel {
19
20PDEPkernel::PDEPkernel(const std::unique_ptr<kernel::KernelBuilder> & b, const unsigned swizzleFactor, std::string name)
21: MultiBlockKernel(std::move(name),
22// input stream sets
23{Binding{b->getStreamSetTy(), "marker", FixedRate(), Principal()},
24Binding{b->getStreamSetTy(swizzleFactor), "source", PopcountOf("marker"), BlockSize(b->getBitBlockWidth() / swizzleFactor) }},
25// output stream set
26{Binding{b->getStreamSetTy(swizzleFactor), "output", FixedRate(), BlockSize(b->getBitBlockWidth() / swizzleFactor)}},
27{}, {}, {})
28, mSwizzleFactor(swizzleFactor) {
29
30}
31
32void PDEPkernel::generateMultiBlockLogic(const std::unique_ptr<KernelBuilder> & b, Value * const numOfBlocks) {
33    BasicBlock * const entry = b->GetInsertBlock();
34    BasicBlock * const processBlock = b->CreateBasicBlock("processBlock");
35    BasicBlock * const finishedStrides = b->CreateBasicBlock("finishedStrides");
36    const auto pdepWidth = b->getBitBlockWidth() / mSwizzleFactor;
37    ConstantInt * const BLOCK_WIDTH = b->getSize(b->getBitBlockWidth());
38    ConstantInt * const PDEP_WIDTH = b->getSize(pdepWidth);
39
40    Constant * const ZERO = b->getSize(0);
41    Value * const sourceItemCount = b->getProcessedItemCount("source");
42
43    Value * const initialSourceOffset = b->CreateURem(sourceItemCount, BLOCK_WIDTH);
44    b->CreateBr(processBlock);
45
46    b->SetInsertPoint(processBlock);
47    PHINode * const strideIndex = b->CreatePHI(b->getSizeTy(), 2);
48    strideIndex->addIncoming(ZERO, entry);
49    PHINode * const bufferPhi = b->CreatePHI(b->getBitBlockType(), 2);
50    bufferPhi->addIncoming(Constant::getNullValue(b->getBitBlockType()), entry);
51    PHINode * const sourceOffsetPhi = b->CreatePHI(b->getSizeTy(), 2);
52    sourceOffsetPhi->addIncoming(initialSourceOffset, entry);
53    PHINode * const bufferSizePhi = b->CreatePHI(b->getSizeTy(), 2);
54    bufferSizePhi->addIncoming(ZERO, entry);
55
56    // Extract the values we will use in the main processing loop
57    Value * const markerStream = b->getInputStreamBlockPtr("marker", ZERO, strideIndex);
58    Value * const markerValue = b->CreateBlockAlignedLoad(markerStream);
59    Value * const selectors = b->fwCast(pdepWidth, markerValue);
60    Value * const numOfSelectors = b->simd_popcount(pdepWidth, selectors);
61
62    // For each element of the marker block
63    Value * bufferSize = bufferSizePhi;
64    Value * sourceOffset = sourceOffsetPhi;
65    Value * buffer = bufferPhi;
66    for (unsigned i = 0; i < mSwizzleFactor; i++) {
67
68        // How many bits will we deposit?
69        Value * const required = b->CreateExtractElement(numOfSelectors, b->getSize(i));
70
71        // Aggressively enqueue any additional bits
72        BasicBlock * const entry = b->GetInsertBlock();
73        BasicBlock * const enqueueBits = b->CreateBasicBlock();
74        b->CreateBr(enqueueBits);
75
76        b->SetInsertPoint(enqueueBits);
77        PHINode * const updatedBufferSize = b->CreatePHI(bufferSize->getType(), 2);
78        updatedBufferSize->addIncoming(bufferSize, entry);
79        PHINode * const updatedSourceOffset = b->CreatePHI(sourceOffset->getType(), 2);
80        updatedSourceOffset->addIncoming(sourceOffset, entry);
81        PHINode * const updatedBuffer = b->CreatePHI(buffer->getType(), 2);
82        updatedBuffer->addIncoming(buffer, entry);
83
84        // Calculate the block and swizzle index of the current swizzle row
85        Value * const blockOffset = b->CreateUDiv(updatedSourceOffset, BLOCK_WIDTH);
86        Value * const swizzleIndex = b->CreateUDiv(b->CreateURem(updatedSourceOffset, BLOCK_WIDTH), PDEP_WIDTH);
87        Value * const swizzle = b->CreateBlockAlignedLoad(b->getInputStreamBlockPtr("source", swizzleIndex, blockOffset));
88        Value * const swizzleOffset = b->CreateURem(updatedSourceOffset, PDEP_WIDTH);
89
90        // Shift the swizzle to the right to clear off any used bits ...
91        Value * const swizzleShift = b->simd_fill(pdepWidth, swizzleOffset);
92        Value * const unreadBits = b->CreateLShr(swizzle, swizzleShift);
93
94        // ... then to the left to align the bits with the buffer and combine them.
95        Value * const bufferShift = b->simd_fill(pdepWidth, updatedBufferSize);
96        Value * const pendingBits = b->CreateShl(unreadBits, bufferShift);
97
98        buffer = b->CreateOr(updatedBuffer, pendingBits);
99        updatedBuffer->addIncoming(buffer, enqueueBits);
100
101        // Update the buffer size with the number of bits we have actually enqueued
102        Value * const maxBufferSize = b->CreateAdd(b->CreateSub(PDEP_WIDTH, swizzleOffset), updatedBufferSize);
103        bufferSize = b->CreateUMin(maxBufferSize, PDEP_WIDTH);
104        updatedBufferSize->addIncoming(bufferSize, enqueueBits);
105
106        // ... and increment the source offset by the number we actually inserted
107        Value * const inserted = b->CreateSub(bufferSize, updatedBufferSize);
108        sourceOffset = b->CreateAdd(updatedSourceOffset, inserted);
109        updatedSourceOffset->addIncoming(sourceOffset, enqueueBits);
110
111        // INVESTIGATE: we can branch at most once here. I'm not sure whether the potential
112        // branch misprediction is better or worse than always filling from two swizzles to
113        // ensure that we have enough bits to deposit.
114        BasicBlock * const depositBits = b->CreateBasicBlock();
115        b->CreateUnlikelyCondBr(b->CreateICmpULT(bufferSize, required), enqueueBits, depositBits);
116
117        b->SetInsertPoint(depositBits);
118
119        // Apply PDEP to each element of the combined swizzle using the current PDEP mask
120        Value * const mask = b->CreateExtractElement(selectors, i);
121        Value* result = b->simd_pdep(pdepWidth, buffer, b->simd_fill(pdepWidth, mask));
122
123        // Store the result
124        Value * const outputStreamPtr = b->getOutputStreamBlockPtr("output", b->getSize(i), strideIndex);
125        b->CreateBlockAlignedStore(result, outputStreamPtr);
126
127        // Shift away any used bits from the buffer and decrement our buffer size by the number we used
128        Value * const usedShift = b->simd_fill(pdepWidth, required);
129        buffer = b->CreateLShr(buffer, usedShift);
130        bufferSize = b->CreateSub(bufferSize, required);
131    }
132
133    BasicBlock * const finishedBlock = b->GetInsertBlock();
134    sourceOffsetPhi->addIncoming(sourceOffset, finishedBlock);
135    bufferSizePhi->addIncoming(bufferSize, finishedBlock);
136    bufferPhi->addIncoming(buffer, finishedBlock);
137    Value * const nextStrideIndex = b->CreateAdd(strideIndex, b->getSize(1));
138    strideIndex->addIncoming(nextStrideIndex, finishedBlock);
139    b->CreateLikelyCondBr(b->CreateICmpNE(nextStrideIndex, numOfBlocks), processBlock, finishedStrides);
140
141    b->SetInsertPoint(finishedStrides);
142}
143   
144StreamExpandKernel::StreamExpandKernel(const std::unique_ptr<kernel::KernelBuilder> &
145                                       , StreamSet * source, const unsigned base, StreamSet * mask
146                                       , StreamSet * expanded
147                                       , const unsigned FieldWidth)
148: MultiBlockKernel("streamExpand" + std::to_string(FieldWidth)
149+ "_" + std::to_string(source->getNumElements())
150+ "_" + std::to_string(base) + "_" + std::to_string(expanded->getNumElements()),
151
152{Binding{"marker", mask, FixedRate(), Principal()},
153Binding{"source", source, PopcountOf("marker")}},
154{Binding{"output", expanded, FixedRate()}},
155{}, {}, {})
156, mFieldWidth(FieldWidth)
157, mSelectedStreamBase(base)
158, mSelectedStreamCount(expanded->getNumElements()) {
159
160}
161
162void StreamExpandKernel::generateMultiBlockLogic(const std::unique_ptr<KernelBuilder> & b, llvm::Value * const numOfBlocks) {
163    Type * fieldWidthTy = b->getIntNTy(mFieldWidth);
164    Type * sizeTy = b->getSizeTy();
165    const unsigned numFields = b->getBitBlockWidth() / mFieldWidth;
166   
167    Constant * const ZERO = b->getSize(0);
168    Constant * bwConst = ConstantInt::get(sizeTy, b->getBitBlockWidth());
169    Constant * fwConst = ConstantInt::get(sizeTy, mFieldWidth);
170    Constant * fwSplat = ConstantVector::getSplat(numFields, ConstantInt::get(fieldWidthTy, mFieldWidth));
171    Constant * fw_sub1Splat = ConstantVector::getSplat(numFields, ConstantInt::get(fieldWidthTy, mFieldWidth - 1));
172   
173    BasicBlock * entry = b->GetInsertBlock();
174    BasicBlock * expandLoop = b->CreateBasicBlock("expandLoop");
175    BasicBlock * expansionDone = b->CreateBasicBlock("expansionDone");
176    Value * processedSourceItems = b->getProcessedItemCount("source");
177    Value * initialSourceOffset = b->CreateURem(processedSourceItems, bwConst);
178    Value * pendingData[mSelectedStreamCount];
179    for (unsigned i = 0; i < mSelectedStreamCount; i++) {
180        pendingData[i] = b->loadInputStreamBlock("source", b->getInt32(mSelectedStreamBase + i), ZERO);
181    }
182   
183    b->CreateBr(expandLoop);
184    // Main Loop
185    b->SetInsertPoint(expandLoop);
186    PHINode * blockNoPhi = b->CreatePHI(b->getSizeTy(), 2);
187    PHINode * pendingOffsetPhi = b->CreatePHI(b->getSizeTy(), 2);
188    PHINode * pendingDataPhi[mSelectedStreamCount];
189    blockNoPhi->addIncoming(ZERO, entry);
190    pendingOffsetPhi->addIncoming(initialSourceOffset, entry);
191    for (unsigned i = 0; i < mSelectedStreamCount; i++) {
192        pendingDataPhi[i] = b->CreatePHI(b->getBitBlockType(), 2);
193        pendingDataPhi[i]->addIncoming(pendingData[i], entry);
194    }
195    Value * deposit_mask = b->loadInputStreamBlock("marker", ZERO, blockNoPhi);
196
197    // Calculate the field values and offsets we need for assembling a
198    // a full block of source bits.  Assembly will use the following operations.
199    // A = b->simd_srlv(fw, b->mvmd_dsll(fw, source, pending, field_offset_lo), bit_offset);
200    // B = b->simd_sllv(fw, b->mvmd_dsll(fw, source, pending, field_offset_hi), shift_fwd);
201    // all_source_bits = simd_or(A, B);
202    Value * pendingOffset = b->CreateURem(pendingOffsetPhi, bwConst);
203    // Value * pendingItems = b->CreateURem(b->CreateSub(bwConst, pendingOffset), bwConst);
204    Value * pendingItems = b->CreateSub(bwConst, pendingOffset);
205
206    Value * field_offset_lo = b->CreateCeilUDiv(pendingItems, fwConst);
207    Value * bit_offset = b->simd_fill(mFieldWidth, b->CreateURem(pendingOffset, fwConst));
208    // Carefully avoid a shift by the full fieldwith (which gives a poison value).
209    // field_offset_lo + 1 unless the bit_offset is 0, in which case it is just field_offset_lo.
210    Value * field_offset_hi =  b->CreateUDiv(pendingItems, fwConst);
211    // fw - bit_offset, unless bit_offset is 0, in which case, the shift_fwd is 0.
212    Value * shift_fwd = b->CreateURem(b->CreateSub(fwSplat, bit_offset), fwSplat);
213
214    // Once all source bits are assembled, they need to be distributed to the
215    // output fields in accord with the popcounts of the deposit mask fields.
216    // The bits for each output field will typically come from (at most) two
217    // source fields, with offsets.  Calculate the field numbers and offsets.
218   
219    Value * fieldPopCounts = b->simd_popcount(mFieldWidth, deposit_mask);
220    // For each field determine the (partial) sum popcount of all fields prior to
221    // the current field.
222    Value * partialSum = fieldPopCounts;
223    for (unsigned i = 1; i < numFields; i *= 2) {
224        partialSum = b->simd_add(mFieldWidth, partialSum, b->mvmd_slli(mFieldWidth, partialSum, i));
225    }
226    Value * const blockPopCount = b->CreateZExtOrTrunc(b->CreateExtractElement(partialSum, numFields - 1), sizeTy);
227    partialSum = b->mvmd_slli(mFieldWidth, partialSum, 1);
228
229    Value * const source_field_lo = b->CreateUDiv(partialSum, fwSplat);
230    Value * const source_field_hi = b->CreateUDiv(b->CreateAdd(partialSum, fw_sub1Splat), fwSplat);
231    Value * const source_shift_lo = b->CreateAnd(partialSum, fw_sub1Splat);  // parallel URem
232    Value * const source_shift_hi = b->CreateAnd(b->CreateSub(fwSplat, source_shift_lo), fw_sub1Splat);
233
234    // The source stream may not be positioned at a block boundary.  Partial data
235    // has been saved in the kernel state, determine the next full block number
236    // for loading source streams.
237    Value * const newPendingOffset = b->CreateAdd(pendingOffsetPhi, blockPopCount);
238    Value * const srcBlockNo = b->CreateUDiv(newPendingOffset, bwConst);
239    // Now load and process source streams.
240    Value * sourceData[mSelectedStreamCount];
241    for (unsigned i = 0; i < mSelectedStreamCount; i++) {
242        sourceData[i] = b->loadInputStreamBlock("source", b->getInt32(mSelectedStreamBase + i), srcBlockNo);
243        Value * A = b->simd_srlv(mFieldWidth, b->mvmd_dsll(mFieldWidth, sourceData[i], pendingDataPhi[i], field_offset_lo), bit_offset);
244        Value * B = b->simd_sllv(mFieldWidth, b->mvmd_dsll(mFieldWidth, sourceData[i], pendingDataPhi[i], field_offset_hi), shift_fwd);
245        Value * full_source_block = b->simd_or(A, B);
246        Value * C = b->simd_srlv(mFieldWidth, b->mvmd_shuffle(mFieldWidth, full_source_block, source_field_lo), source_shift_lo);
247        Value * D = b->simd_sllv(mFieldWidth, b->mvmd_shuffle(mFieldWidth, full_source_block, source_field_hi), source_shift_hi);
248        Value * output = b->bitCast(b->simd_or(C, D));
249        b->storeOutputStreamBlock("output", b->getInt32(i), blockNoPhi, output);
250    }
251    //
252    // Update loop control Phis for the next iteration.
253    //
254    Value * nextBlk = b->CreateAdd(blockNoPhi, b->getSize(1));
255    blockNoPhi->addIncoming(nextBlk, expandLoop);
256    pendingOffsetPhi->addIncoming(newPendingOffset, expandLoop);
257    for (unsigned i = 0; i < mSelectedStreamCount; i++) {
258        pendingDataPhi[i]->addIncoming(sourceData[i], expandLoop);
259    }
260    //
261    // Now continue the loop if there are more blocks to process.
262    Value * moreToDo = b->CreateICmpNE(nextBlk, numOfBlocks);
263    b->CreateCondBr(moreToDo, expandLoop, expansionDone);
264   
265    b->SetInsertPoint(expansionDone);
266}
267
268FieldDepositKernel::FieldDepositKernel(const std::unique_ptr<kernel::KernelBuilder> &
269                                       , StreamSet * mask, StreamSet * input, StreamSet * output
270                                       , const unsigned fieldWidth)
271: MultiBlockKernel("FieldDeposit" + std::to_string(fieldWidth) + "_" + std::to_string(input->getNumElements()),
272{Binding{"depositMask", mask}
273, Binding{"inputStreamSet", input}},
274{Binding{"outputStreamSet", output}},
275{}, {}, {})
276, mFieldWidth(fieldWidth)
277, mStreamCount(input->getNumElements()) {
278
279}
280   
281void FieldDepositKernel::generateMultiBlockLogic(const std::unique_ptr<KernelBuilder> & kb, llvm::Value * const numOfBlocks) {
282    BasicBlock * entry = kb->GetInsertBlock();
283    BasicBlock * processBlock = kb->CreateBasicBlock("processBlock");
284    BasicBlock * done = kb->CreateBasicBlock("done");
285    Constant * const ZERO = kb->getSize(0);
286    kb->CreateBr(processBlock);
287    kb->SetInsertPoint(processBlock);
288    PHINode * blockOffsetPhi = kb->CreatePHI(kb->getSizeTy(), 2);
289    blockOffsetPhi->addIncoming(ZERO, entry);
290    Value * depositMask = kb->loadInputStreamBlock("depositMask", ZERO, blockOffsetPhi);
291    for (unsigned j = 0; j < mStreamCount; ++j) {
292        Value * input = kb->loadInputStreamBlock("inputStreamSet", kb->getInt32(j), blockOffsetPhi);
293        Value * output = kb->simd_pdep(mFieldWidth, input, depositMask);
294        kb->storeOutputStreamBlock("outputStreamSet", kb->getInt32(j), blockOffsetPhi, output);
295    }
296    Value * nextBlk = kb->CreateAdd(blockOffsetPhi, kb->getSize(1));
297    blockOffsetPhi->addIncoming(nextBlk, processBlock);
298    Value * moreToDo = kb->CreateICmpNE(nextBlk, numOfBlocks);
299    kb->CreateCondBr(moreToDo, processBlock, done);
300    kb->SetInsertPoint(done);
301}
302
303PDEPFieldDepositKernel::PDEPFieldDepositKernel(const std::unique_ptr<kernel::KernelBuilder> &
304                                               , StreamSet * mask, StreamSet * input, StreamSet * output
305                                               , const unsigned fieldWidth)
306: MultiBlockKernel("PDEPFieldDeposit" + std::to_string(fieldWidth) + "_" + std::to_string(input->getNumElements()) ,
307                   {Binding{"depositMask", mask},
308                    Binding{"inputStreamSet", input}},
309                   {Binding{"outputStreamSet", output}},
310                   {}, {}, {})
311, mPDEPWidth(fieldWidth)
312, mStreamCount(input->getNumElements()) {
313    if ((fieldWidth != 32) && (fieldWidth != 64))
314        llvm::report_fatal_error("Unsupported PDEP width for PDEPFieldDepositKernel");
315}
316
317void PDEPFieldDepositKernel::generateMultiBlockLogic(const std::unique_ptr<KernelBuilder> & kb, llvm::Value * const numOfBlocks) {
318    Type * fieldTy = kb->getIntNTy(mPDEPWidth);
319    Type * fieldPtrTy = PointerType::get(fieldTy, 0);
320    Constant * PDEP_func = nullptr;
321    if (mPDEPWidth == 64) {
322        PDEP_func = Intrinsic::getDeclaration(kb->getModule(), Intrinsic::x86_bmi_pdep_64);
323    } else if (mPDEPWidth == 32) {
324        PDEP_func = Intrinsic::getDeclaration(kb->getModule(), Intrinsic::x86_bmi_pdep_32);
325    }
326    BasicBlock * entry = kb->GetInsertBlock();
327    BasicBlock * processBlock = kb->CreateBasicBlock("processBlock");
328    BasicBlock * done = kb->CreateBasicBlock("done");
329    Constant * const ZERO = kb->getSize(0);
330    const unsigned fieldsPerBlock = kb->getBitBlockWidth()/mPDEPWidth;
331    kb->CreateBr(processBlock);
332    kb->SetInsertPoint(processBlock);
333    PHINode * blockOffsetPhi = kb->CreatePHI(kb->getSizeTy(), 2);
334    blockOffsetPhi->addIncoming(ZERO, entry);
335    std::vector<Value *> mask(fieldsPerBlock);
336//  When operating on fields individually, we can use vector load/store with
337//  extract/insert element operations, or we can use individual field load
338//  and stores.   Individual field operations require fewer total operations,
339//  but more memory instructions.   It may be that vector load/extract is better,
340//  while field store is better.   Vector insert then store creates long dependence
341//  chains.
342//
343#define PREFER_FIELD_STORES_OVER_INSERT_ELEMENT
344#ifdef PREFER_FIELD_LOADS_OVER_EXTRACT_ELEMENT
345    Value * depositMaskPtr = kb->getInputStreamBlockPtr("depositMask", ZERO, blockOffsetPhi);
346    depositMaskPtr = kb->CreatePointerCast(depositMaskPtr, fieldPtrTy);
347    for (unsigned i = 0; i < fieldsPerBlock; i++) {
348        mask[i] = kb->CreateLoad(kb->CreateGEP(depositMaskPtr, kb->getInt32(i)));
349    }
350#else
351    Value * depositMask = kb->fwCast(mPDEPWidth, kb->loadInputStreamBlock("depositMask", ZERO, blockOffsetPhi));
352    for (unsigned i = 0; i < fieldsPerBlock; i++) {
353        mask[i] = kb->CreateExtractElement(depositMask, kb->getInt32(i));
354    }
355#endif
356    for (unsigned j = 0; j < mStreamCount; ++j) {
357#ifdef PREFER_FIELD_LOADS_OVER_EXTRACT_ELEMENT
358        Value * inputPtr = kb->getInputStreamBlockPtr("inputStreamSet", kb->getInt32(j), blockOffsetPhi);
359        inputPtr = kb->CreatePointerCast(inputPtr, fieldPtrTy);
360#else
361        Value * inputStrm = kb->fwCast(mPDEPWidth, kb->loadInputStreamBlock("inputStreamSet", kb->getInt32(j), blockOffsetPhi));
362#endif
363#ifdef PREFER_FIELD_STORES_OVER_INSERT_ELEMENT
364        Value * outputPtr = kb->getOutputStreamBlockPtr("outputStreamSet", kb->getInt32(j), blockOffsetPhi);
365        outputPtr = kb->CreatePointerCast(outputPtr, fieldPtrTy);
366#else
367        Value * outputStrm = kb->fwCast(mPDEPWidth, kb->allZeroes());
368#endif
369        for (unsigned i = 0; i < fieldsPerBlock; i++) {
370#ifdef PREFER_FIELD_LOADS_OVER_EXTRACT_ELEMENT
371            Value * field = kb->CreateLoad(kb->CreateGEP(inputPtr, kb->getInt32(i)));
372#else
373            Value * field = kb->CreateExtractElement(inputStrm, kb->getInt32(i));
374#endif
375            Value * compressed = kb->CreateCall(PDEP_func, {field, mask[i]});
376#ifdef PREFER_FIELD_STORES_OVER_INSERT_ELEMENT
377            kb->CreateStore(compressed, kb->CreateGEP(outputPtr, kb->getInt32(i)));
378        }
379#else
380            outputStrm = kb->CreateInsertElement(outputStrm, compressed, kb->getInt32(i));
381        }
382        kb->storeOutputStreamBlock("outputStreamSet", kb->getInt32(j), blockOffsetPhi, outputStrm);
383#endif
384    }
385    Value * nextBlk = kb->CreateAdd(blockOffsetPhi, kb->getSize(1));
386    blockOffsetPhi->addIncoming(nextBlk, processBlock);
387    Value * moreToDo = kb->CreateICmpNE(nextBlk, numOfBlocks);
388    kb->CreateCondBr(moreToDo, processBlock, done);
389    kb->SetInsertPoint(done);
390}
391
392}
393
Note: See TracBrowser for help on using the repository browser.