nnicolosi/Raptor
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

add .clang-format

Browse Source 
reformat all src
This commit is contained in:
NiccoloN
2026-02-26 19:16:42 +01:00
parent a2c31836ae
commit 810e5e75f9
32 changed files with 902 additions and 953 deletions
Download Patch File Download Diff File Expand all files Collapse all files

239
src/PIM/Conversion/ONNXToSpatial/NN/Pooling.cpp
View File

@@ -6,38 +6,39 @@
#include "mlir/IR/PatternMatch.h"
#include "mlir/IR/Value.h"
#include "mlir/IR/ValueRange.h"
#include "src/Accelerators/PIM/Common/PIMCommon.hpp"
#include "src/Accelerators/PIM/Compiler/PimCompilerOptions.hpp"
#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
#include "src/Dialect/ONNX/ONNXOps.hpp"
#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ONNXToSpatialCommon.hpp"
#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Utils/SpatialReducer.hpp"
#include "llvm/ADT/SmallVector.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include <cassert>
#include <cmath>
#include <cstddef>
#include "src/Accelerators/PIM/Common/PIMCommon.hpp"
#include "src/Accelerators/PIM/Compiler/PimCompilerOptions.hpp"
#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/ONNXToSpatialCommon.hpp"
#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Utils/SpatialReducer.hpp"
#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
#include "src/Dialect/ONNX/ONNXOps.hpp"
using namespace mlir;
namespace onnx_mlir {
llvm::SmallPtrSet<Operation *, 16> oldComputeOpsReplaced;
llvm::SmallPtrSet<Operation*, 16> oldComputeOpsReplaced;
Value applyReducePatternNew(SmallVector<Value> &valuesToReduce,
ConversionPatternRewriter &rewriter,
std::function<Value(const Value &, const Value &)> reduce,
std::function<Value(const Value &)> preprocess,
std::function<Value(const Value &)> postprocess) {
Value applyReducePatternNew(SmallVector<Value>& valuesToReduce,
ConversionPatternRewriter& rewriter,
std::function<Value(const Value&, const Value&)> reduce,
std::function<Value(const Value&)> preprocess,
std::function<Value(const Value&)> postprocess) {
// Simple case: if we have only one input, just return it
if (valuesToReduce.size() == 1) {
if (valuesToReduce.size() == 1)
return valuesToReduce[0];
}
if (preprocess) {
for (auto &valToReduce : valuesToReduce) {
for (auto& valToReduce : valuesToReduce) {
rewriter.setInsertionPointAfterValue(valToReduce);
valToReduce = preprocess(valToReduce);
}
@@ -47,9 +48,9 @@ Value applyReducePatternNew(SmallVector<Value> &valuesToReduce,
// computeOp. In this case, we need to apply the reduction within-computef
// Keep a map between a computeOp and the last Value for this reduction
std::unordered_map<Operation *, Value> lastValueForCompute;
for (auto &valToReduce : valuesToReduce) {
Operation *computeOp = valToReduce.getParentBlock()->getParentOp();
std::unordered_map<Operation*, Value> lastValueForCompute;
for (auto& valToReduce : valuesToReduce) {
Operation* computeOp = valToReduce.getParentBlock()->getParentOp();
// if (valToReduce.getDefiningOp()) {
// // If the value is defined by an operation, we take the parent
// operation computeOp = valToReduce.getDefiningOp()->getParentOp();
@@ -67,12 +68,10 @@ Value applyReducePatternNew(SmallVector<Value> &valuesToReduce,
// within-compute
Value lastWithinComputeValue = it->second;
if (valToReduce.getDefiningOp()->isBeforeInBlock(
lastWithinComputeValue.getDefiningOp())) {
if (valToReduce.getDefiningOp()->isBeforeInBlock(lastWithinComputeValue.getDefiningOp()))
rewriter.setInsertionPointAfterValue(lastWithinComputeValue);
} else {
else
rewriter.setInsertionPointAfterValue(valToReduce);
}
valToReduce = reduce(lastWithinComputeValue, valToReduce);
lastValueForCompute[computeOp] = valToReduce;
}
@@ -83,9 +82,8 @@ Value applyReducePatternNew(SmallVector<Value> &valuesToReduce,
// Now, reconstruct from the map the valuesToReduce list
valuesToReduce.clear();
valuesToReduce.reserve(lastValueForCompute.size());
for (auto &entry : lastValueForCompute) {
for (auto& entry : lastValueForCompute)
valuesToReduce.push_back(entry.second);
}
Location loc = valuesToReduce[0].getLoc();
auto channelType = spatial::SpatChannelType::get(rewriter.getContext());
@@ -123,8 +121,7 @@ Value applyReducePatternNew(SmallVector<Value> &valuesToReduce,
// 3. Add a receiveOp after the second value
rewriter.setInsertionPointAfterValue(secondValue);
auto receivedValue = rewriter.create<spatial::SpatChannelReceiveOp>(
loc, secondValue.getType(), channel);
auto receivedValue = rewriter.create<spatial::SpatChannelReceiveOp>(loc, secondValue.getType(), channel);
// 4. Apply reduction between second value and received value
rewriter.setInsertionPointAfterValue(receivedValue);
@@ -135,17 +132,14 @@ Value applyReducePatternNew(SmallVector<Value> &valuesToReduce,
// If we have an odd number of inputs, we need to add the last one to the
// newInputs list.
if (valuesToReduceRef.size() % 2 == 1) {
if (valuesToReduceRef.size() % 2 == 1)
nextValuesToReduce.push_back(valuesToReduceRef.back());
}
// Replace the inputOps list with the new one.
valuesToReduceRef =
llvm::OwningArrayRef<Value>(std::move(nextValuesToReduce));
valuesToReduceRef = llvm::OwningArrayRef<Value>(std::move(nextValuesToReduce));
}
assert(valuesToReduceRef.size() == 1 &&
"Internal error: expected a single input at this point.");
assert(valuesToReduceRef.size() == 1 && "Internal error: expected a single input at this point.");
auto finalValue = valuesToReduceRef[0];
@@ -168,46 +162,49 @@ bool hasPostProcessPoolingWindow<ONNXAveragePoolOp>() {
}
template <typename PoolOp>
Value postProcessPoolingWindow(ConversionPatternRewriter &rewriter,
Location loc, PoolOp poolOp, Value valueToDivide, size_t krn_size,
size_t tilesSkippedByPadding) {
Value postProcessPoolingWindow(ConversionPatternRewriter& rewriter,
Location loc,
PoolOp poolOp,
Value valueToDivide,
size_t krn_size,
size_t tilesSkippedByPadding) {
return nullptr;
}
template <>
Value postProcessPoolingWindow<ONNXAveragePoolOp>(
ConversionPatternRewriter &rewriter, Location loc, ONNXAveragePoolOp poolOp,
Value valueToDivide, size_t krn_size, size_t tilesSkippedByPadding) {
Value postProcessPoolingWindow<ONNXAveragePoolOp>(ConversionPatternRewriter& rewriter,
Location loc,
ONNXAveragePoolOp poolOp,
Value valueToDivide,
size_t krn_size,
size_t tilesSkippedByPadding) {
bool countIncludePad = poolOp.getCountIncludePad() == 1;
size_t divisorNumber =
countIncludePad ? krn_size : krn_size - tilesSkippedByPadding;
size_t divisorNumber = countIncludePad ? krn_size : krn_size - tilesSkippedByPadding;
RankedTensorType scalarTensor =
RankedTensorType::get({1}, rewriter.getF32Type());
RankedTensorType scalarTensor = RankedTensorType::get({1}, rewriter.getF32Type());
// Put a spat.const before the computeOp, and use its value. We do this to be
// compatible with the current code generation, which assumes constant to be
// loaded in global memory, which is allocated by adding a spat.const OP
// directly under func.func (i.e. alongside ComputeOps)
auto computeOp = cast<spatial::SpatWeightedCompute>(
valueToDivide.getDefiningOp()->getParentOp());
auto computeOp = cast<spatial::SpatWeightedCompute>(valueToDivide.getDefiningOp()->getParentOp());
rewriter.setInsertionPoint(computeOp);
auto divisorValue = rewriter.create<spatial::SpatConstantOp>(loc, scalarTensor,
rewriter.getI64IntegerAttr(divisorNumber),
/* should_allocate = */ rewriter.getBoolAttr(true));
auto divisorValue = rewriter.create<spatial::SpatConstantOp>(loc,
scalarTensor,
rewriter.getI64IntegerAttr(divisorNumber),
/* should_allocate = */ rewriter.getBoolAttr(true));
rewriter.setInsertionPointAfterValue(valueToDivide);
return rewriter.create<spatial::SpatVSDivOp>(
loc, valueToDivide.getType(), valueToDivide, divisorValue);
return rewriter.create<spatial::SpatVSDivOp>(loc, valueToDivide.getType(), valueToDivide, divisorValue);
}
template <typename PoolOp, typename PoolOpAdaptor, typename ReduceOp>
struct PoolingBaseConverter : public OpConversionPattern<PoolOp> {
PoolingBaseConverter(MLIRContext *ctx) : OpConversionPattern<PoolOp>(ctx) {}
PoolingBaseConverter(MLIRContext* ctx)
: OpConversionPattern<PoolOp>(ctx) {}
LogicalResult matchAndRewrite(PoolOp poolOp, PoolOpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const final {
LogicalResult matchAndRewrite(PoolOp poolOp, PoolOpAdaptor adaptor, ConversionPatternRewriter& rewriter) const final {
Value X = adaptor.getX();
ShapedType xShape = mlir::cast<ShapedType>(X.getType());
Value Y = poolOp.getResult();
@@ -218,17 +215,13 @@ struct PoolingBaseConverter : public OpConversionPattern<PoolOp> {
unpackOptionalPairVector(adaptor.getDilations(), dilation_x, dilation_y);
unpackOptionalPairVector(adaptor.getKernelShape(), krn_w, krn_h);
if (adaptor.getAutoPad() != "NOTSET") {
return rewriter.notifyMatchFailure(
poolOp, "auto_pad != NOTSET is deprecated.");
}
if (adaptor.getAutoPad() != "NOTSET")
return rewriter.notifyMatchFailure(poolOp, "auto_pad != NOTSET is deprecated.");
size_t pad_x, pad_y;
auto padUnpackError =
unpackOptionalPadsVector(adaptor.getPads(), pad_x, pad_y);
if (padUnpackError.has_value()) {
auto padUnpackError = unpackOptionalPadsVector(adaptor.getPads(), pad_x, pad_y);
if (padUnpackError.has_value())
return rewriter.notifyMatchFailure(poolOp, padUnpackError.value());
}
Location loc = poolOp.getLoc();
@@ -236,8 +229,7 @@ struct PoolingBaseConverter : public OpConversionPattern<PoolOp> {
size_t input_w = GET_IMAGE_WIDTH(xShape);
size_t output_h = GET_IMAGE_HEIGHT(yShape);
size_t output_w = GET_IMAGE_WIDTH(yShape);
size_t channelTileCount =
ceilIntegerDivide(GET_IMAGE_CHANNEL(xShape), crossbarSize.getValue());
size_t channelTileCount = ceilIntegerDivide(GET_IMAGE_CHANNEL(xShape), crossbarSize.getValue());
size_t channelTileRest = GET_IMAGE_CHANNEL(xShape) % crossbarSize;
// 1: Tile the input tensor
@@ -249,14 +241,13 @@ struct PoolingBaseConverter : public OpConversionPattern<PoolOp> {
// Example complete input tensor: tensor<1x3x12x12xf32> (NxCxWxH)
// Suppose that the input tensor is produced by concatenating the results of
// many ComputeOps. Get the result tiles from these ComputeOps.
SmallVector<SmallVector<SmallVector<Value>>> inputTiles(channelTileCount,
SmallVector<SmallVector<Value>>(input_w, SmallVector<Value>(input_h)));
SmallVector<SmallVector<SmallVector<Value>>> inputTiles(
channelTileCount, SmallVector<SmallVector<Value>>(input_w, SmallVector<Value>(input_h)));
auto resolveErrorOpt = resolveImgInputTiles(X, inputTiles, channelTileCount,
channelTileRest, input_w, input_h, rewriter);
if (resolveErrorOpt.has_value()) {
auto resolveErrorOpt =
resolveImgInputTiles(X, inputTiles, channelTileCount, channelTileRest, input_w, input_h, rewriter);
if (resolveErrorOpt.has_value())
return rewriter.notifyMatchFailure(poolOp, *resolveErrorOpt);
}
// TODO: This requires a core for each input tile, which is not ideal. We
// can do better.
@@ -265,18 +256,17 @@ struct PoolingBaseConverter : public OpConversionPattern<PoolOp> {
for (size_t t = 0; t < channelTileCount; t++) {
for (size_t x = 0; x < input_w; x++) {
for (size_t y = 0; y < input_h; y++) {
if (auto extractSliceOp =
inputTiles[t][x][y].getDefiningOp<tensor::ExtractSliceOp>()) {
if (auto extractSliceOp = inputTiles[t][x][y].getDefiningOp<tensor::ExtractSliceOp>()) {
Location tileLoc = extractSliceOp.getLoc();
auto tempComputeOp = rewriter.create<spatial::SpatWeightedCompute>(
tileLoc, extractSliceOp.getResultType(),
/* xbarWeights =*/ValueRange(), extractSliceOp.getResult());
auto tempComputeOp = rewriter.create<spatial::SpatWeightedCompute>(tileLoc,
extractSliceOp.getResultType(),
/* xbarWeights =*/ValueRange(),
extractSliceOp.getResult());
Block *tempComputeOpBlock = new Block();
Block* tempComputeOpBlock = new Block();
tempComputeOp.getBody().push_back(tempComputeOpBlock);
auto tempComputeOpBlockArg = tempComputeOpBlock->addArgument(
extractSliceOp.getType(), tileLoc);
auto tempComputeOpBlockArg = tempComputeOpBlock->addArgument(extractSliceOp.getType(), tileLoc);
rewriter.setInsertionPointToStart(tempComputeOpBlock);
rewriter.create<spatial::SpatYieldOp>(tileLoc, tempComputeOpBlockArg);
@@ -295,8 +285,7 @@ struct PoolingBaseConverter : public OpConversionPattern<PoolOp> {
// For example: outputTiles[channelTile][x][y]
// Example complete output tensor: tensor<1x3x6x6xf32> (NxCxWxH)
SmallVector<SmallVector<SmallVector<Value>>> outputTiles(
channelTileCount, SmallVector<SmallVector<Value>>(
output_w, SmallVector<Value>(output_h, nullptr)));
channelTileCount, SmallVector<SmallVector<Value>>(output_w, SmallVector<Value>(output_h, nullptr)));
// List of values to pool for each output pixel
SmallVector<Value> valuesToPool;
@@ -312,15 +301,12 @@ struct PoolingBaseConverter : public OpConversionPattern<PoolOp> {
valuesToPool.clear();
size_t tilesSkippedByPadding = 0;
auto [start_x, end_x] = kernel_get_start_and_end(
outX, input_w, krn_w, stride_x, dilation_x, pad_x);
auto [start_y, end_y] = kernel_get_start_and_end(
outY, input_h, krn_h, stride_y, dilation_y, pad_y);
auto [start_x, end_x] = kernel_get_start_and_end(outX, input_w, krn_w, stride_x, dilation_x, pad_x);
auto [start_y, end_y] = kernel_get_start_and_end(outY, input_h, krn_h, stride_y, dilation_y, pad_y);
for (size_t inX = start_x; inX < end_x; inX += dilation_x) {
for (size_t inY = start_y; inY < end_y; inY += dilation_y) {
if (failed(verifyWithinBoundsAndPaddings(
input_w, input_h, inX, inY, pad_x, pad_y))) {
if (failed(verifyWithinBoundsAndPaddings(input_w, input_h, inX, inY, pad_x, pad_y))) {
tilesSkippedByPadding++;
continue;
}
@@ -328,78 +314,73 @@ struct PoolingBaseConverter : public OpConversionPattern<PoolOp> {
Value inputTile = inputTiles[outTile][inX][inY];
Value valueToPool;
if (auto computeProducer =
inputTile.getDefiningOp<spatial::SpatWeightedCompute>()) {
if (auto computeProducer = inputTile.getDefiningOp<spatial::SpatWeightedCompute>()) {
int resultNumber = getResultIndex(computeProducer, inputTile);
auto yieldInComputeOp = cast<spatial::SpatYieldOp>(
computeProducer.getBody().front().getTerminator());
auto yieldInComputeOp = cast<spatial::SpatYieldOp>(computeProducer.getBody().front().getTerminator());
valueToPool = yieldInComputeOp.getOperand(resultNumber);
} else if (auto receiveProducer =
inputTile
.getDefiningOp<spatial::SpatChannelReceiveOp>()) {
auto sendOpOpt =
getOtherEndOfChannel(receiveProducer, true, rewriter);
}
else if (auto receiveProducer = inputTile.getDefiningOp<spatial::SpatChannelReceiveOp>()) {
auto sendOpOpt = getOtherEndOfChannel(receiveProducer, true, rewriter);
if (failed(sendOpOpt)) {
return rewriter.notifyMatchFailure(poolOp,
"ChannelReceiveOp does not have a matching "
"ChannelSendOp.");
"ChannelReceiveOp does not have a matching "
"ChannelSendOp.");
}
auto sendOp = cast<spatial::SpatChannelSendOp>(*sendOpOpt);
valueToPool = sendOp.getData();
} else {
}
else {
return rewriter.notifyMatchFailure(poolOp,
"Input tile for Pooling is not produced by a "
"WeightedComputeOp nor a receiveOp");
"Input tile for Pooling is not produced by a "
"WeightedComputeOp nor a receiveOp");
}
valuesToPool.push_back(valueToPool);
}
}
assert(valuesToPool.size() != 0 &&
"Pooling computed on zero tiles make no sense.");
assert(valuesToPool.size() != 0 && "Pooling computed on zero tiles make no sense.");
// assert(computeOpsForPooling.size() != 1 &&
// "Pooling computed on one tiles make no sense??? Or maybe
// this " "should have been simplified earlier???");
std::function<Value(const Value &)> postProcessFn = nullptr;
std::function<Value(const Value&)> postProcessFn = nullptr;
if (hasPostProcessPoolingWindow<PoolOp>()) {
postProcessFn = [&](const Value prevFinalRes) {
return postProcessPoolingWindow(rewriter, loc, poolOp,
prevFinalRes, krn_h * krn_w, tilesSkippedByPadding);
return postProcessPoolingWindow(
rewriter, loc, poolOp, prevFinalRes, krn_h * krn_w, tilesSkippedByPadding);
};
}
Value reducedWithinCompute = applyReducePatternNew(
valuesToPool, rewriter,
[&](const Value lhs, const Value rhs) {
return rewriter.create<ReduceOp>(loc, lhs.getType(), lhs, rhs);
},
nullptr, postProcessFn);
valuesToPool,
rewriter,
[&](const Value lhs, const Value rhs) { return rewriter.create<ReduceOp>(loc, lhs.getType(), lhs, rhs); },
nullptr,
postProcessFn);
// Send this value through a channel, and receive it in the
// `func.func`. During lowering, we will need to "move it" into the
// users computeOps
auto computeOpOfReduced = cast<spatial::SpatWeightedCompute>(
reducedWithinCompute.getDefiningOp()->getParentOp());
auto computeOpOfReduced =
cast<spatial::SpatWeightedCompute>(reducedWithinCompute.getDefiningOp()->getParentOp());
// Create a new channel before the computeOp
rewriter.setInsertionPoint(computeOpOfReduced);
auto reduceChannel = rewriter.create<spatial::SpatChannelNewOp>(
loc, spatial::SpatChannelType::get(rewriter.getContext()));
auto reduceChannel =
rewriter.create<spatial::SpatChannelNewOp>(loc, spatial::SpatChannelType::get(rewriter.getContext()));
// Send value through the channel
rewriter.setInsertionPointAfterValue(reducedWithinCompute);
rewriter.create<spatial::SpatChannelSendOp>(
loc, reduceChannel, reducedWithinCompute);
rewriter.create<spatial::SpatChannelSendOp>(loc, reduceChannel, reducedWithinCompute);
// Receive after the computeOp
rewriter.setInsertionPointAfter(computeOpOfReduced);
auto receivedValue = rewriter.create<spatial::SpatChannelReceiveOp>(
loc, reducedWithinCompute.getType(), reduceChannel);
auto receivedValue =
rewriter.create<spatial::SpatChannelReceiveOp>(loc, reducedWithinCompute.getType(), reduceChannel);
outputTiles[outTile][outX][outY] = receivedValue;
}
@@ -409,9 +390,7 @@ struct PoolingBaseConverter : public OpConversionPattern<PoolOp> {
// TODO: outputTiles are not the results of the computeOps! We need to add
// them!
std::unordered_map<Operation *,
SmallVector<std::tuple<size_t, size_t, size_t, Value>>>
computeOpNeedingResults;
std::unordered_map<Operation*, SmallVector<std::tuple<size_t, size_t, size_t, Value>>> computeOpNeedingResults;
// Iterate each output tile
for (size_t outTile = 0; outTile < channelTileCount; outTile++) {
@@ -422,18 +401,16 @@ struct PoolingBaseConverter : public OpConversionPattern<PoolOp> {
auto outputTileProducer = outputTile.getDefiningOp()->getParentOp();
if (!outputTileProducer) {
return rewriter.notifyMatchFailure(poolOp,
"Output tile for Pooling is not produced by a "
"WeightedComputeOp.");
"Output tile for Pooling is not produced by a "
"WeightedComputeOp.");
}
computeOpNeedingResults[outputTileProducer].push_back(
std::make_tuple(outTile, outX, outY, outputTile));
computeOpNeedingResults[outputTileProducer].push_back(std::make_tuple(outTile, outX, outY, outputTile));
}
}
}
Value outputImage =
createImgConcatOp(outputTiles, rewriter, loc, poolOp.getType());
Value outputImage = createImgConcatOp(outputTiles, rewriter, loc, poolOp.getType());
rewriter.replaceOp(poolOp, outputImage);
@@ -441,12 +418,10 @@ struct PoolingBaseConverter : public OpConversionPattern<PoolOp> {
}
};
void populatePoolingTilingPattern(
RewritePatternSet &patterns, MLIRContext *ctx) {
patterns.insert<PoolingBaseConverter<ONNXMaxPoolSingleOutOp,
ONNXMaxPoolSingleOutOpAdaptor, spatial::SpatVMaxOp>>(ctx);
patterns.insert<PoolingBaseConverter<ONNXAveragePoolOp,
ONNXAveragePoolOpAdaptor, spatial::SpatVAddOp>>(ctx);
void populatePoolingTilingPattern(RewritePatternSet& patterns, MLIRContext* ctx) {
patterns.insert<PoolingBaseConverter<ONNXMaxPoolSingleOutOp, ONNXMaxPoolSingleOutOpAdaptor, spatial::SpatVMaxOp>>(
ctx);
patterns.insert<PoolingBaseConverter<ONNXAveragePoolOp, ONNXAveragePoolOpAdaptor, spatial::SpatVAddOp>>(ctx);
}
} // namespace onnx_mlir
} // namespace onnx_mlir