Skip to content

Commit

Permalink
Bump mlir-aie to 1b059e4996656d00499598c8c1ce8da6baa71262 (#855)
Browse files Browse the repository at this point in the history 
* Change from 'applyPatternsAndFoldGreedily'  to 'applyPatternsGreedily'

* Update llvm and mlir-aie

* Disable -Werror=sign-compare due to issue with nanobind 2.4

* Add convert-arith-to-llvm pass to pipeline

* Switch air python bindings from pybind11 to nanobind

* Disable -Werror=sign-compare due to issue with nanobind 2.4

* Fixup typo nb::module_ (missing underscore)

* Replace return_policy with rv_policy

* Replace applyOpPatternsAndFold with applyOpPatternsGreedily

* Add a workaround to resolve the domain conflict introduced by nanobind 2.4

* Add missing condition in _process_target
  • Loading branch information
@erwei-xilinx
erwei-xilinx authored Jan 14, 2025
1 parent 2feb775 commit 06249fd
Show file tree
Hide file tree
Showing 18 changed files with 227 additions and 107 deletions.
2 changes: 1 addition & 1 deletion CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -108,7 +108,7 @@ include_directories(${AIE_INCLUDE_DIRS})
add_definitions(${LLVM_DEFINITIONS})

## flags duplicated from mlir-aie
add_flag_if_supported("-Werror=sign-compare" WERROR_SIGN_COMPARE)
# add_flag_if_supported("-Werror=sign-compare" WERROR_SIGN_COMPARE)
add_flag_if_supported("-Werror=unused" WERROR_USED)
# What happens when you have a non-void function with no return?
# No `ret` instruction is generated and so execution of that function just
Expand Down
12 changes: 6 additions & 6 deletions mlir/lib/Conversion/AIRRtToNpuPass.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -947,19 +947,19 @@ struct AIRRtToNpuPass : public impl::AIRRtToNpuBase<AIRRtToNpuPass> {
auto ctx = &getContext();
RewritePatternSet canoPatterns_0(ctx);
xilinx::air::populateAIRLoopIndexCanonicalizationPatterns(canoPatterns_0);
(void)applyPatternsAndFoldGreedily(module, std::move(canoPatterns_0));
(void)applyPatternsGreedily(module, std::move(canoPatterns_0));

// Specialize affine for loop nest into wraps and strides
RewritePatternSet loopFoldPattern(ctx);
loopFoldPattern.add<AIRSpecializeAIRRtDmaWrapAndStrideInAffineFor>(ctx);
air::populateAIRLoopIndexCanonicalizationPatterns(loopFoldPattern);
(void)applyPatternsAndFoldGreedily(module, std::move(loopFoldPattern));
(void)applyPatternsGreedily(module, std::move(loopFoldPattern));
unrollAffineFors(module);

// Simplify arith ops (from airrt)
RewritePatternSet canoPatterns_1(ctx);
arith::IndexCastOp::getCanonicalizationPatterns(canoPatterns_1, ctx);
(void)applyPatternsAndFoldGreedily(module, std::move(canoPatterns_1));
(void)applyPatternsGreedily(module, std::move(canoPatterns_1));

// Purge all wait ops again after unroll, in case there were loop carried
// events which couldn't be purged before
Expand All @@ -974,7 +974,7 @@ struct AIRRtToNpuPass : public impl::AIRRtToNpuBase<AIRRtToNpuPass> {
// Simplify arith ops (from airrt)
RewritePatternSet canoPatterns_3(ctx);
arith::IndexCastOp::getCanonicalizationPatterns(canoPatterns_3, ctx);
(void)applyPatternsAndFoldGreedily(module, std::move(canoPatterns_3));
(void)applyPatternsGreedily(module, std::move(canoPatterns_3));

ConversionTarget target(getContext());
target.addIllegalDialect<AIRRtDialect>();
Expand Down Expand Up @@ -1020,7 +1020,7 @@ struct AIRRtToNpuPass : public impl::AIRRtToNpuBase<AIRRtToNpuPass> {
RewritePatternSet canoPatterns_2(ctx);
canoPatterns_2.insert<RelocateAssumeAlignmentOp>(ctx);
arith::IndexCastOp::getCanonicalizationPatterns(canoPatterns_2, ctx);
(void)applyPatternsAndFoldGreedily(module, std::move(canoPatterns_2));
(void)applyPatternsGreedily(module, std::move(canoPatterns_2));

// Unroll any affine for loops
unrollAffineFors(module);
Expand All @@ -1030,7 +1030,7 @@ struct AIRRtToNpuPass : public impl::AIRRtToNpuBase<AIRRtToNpuPass> {
RewritePatternSet castPattern(ctx);
air::populateBufferMemrefToFuncArgsPattern(castPattern);
castPattern.add(CastFunctionArgs);
(void)applyPatternsAndFoldGreedily(module, std::move(castPattern));
(void)applyPatternsGreedily(module, std::move(castPattern));

// Insert sync op after copying data out to host
insertNpuSyncOpForResults(module);
Expand Down
24 changes: 12 additions & 12 deletions mlir/lib/Conversion/AIRToAIEPass.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -673,7 +673,7 @@ void specializeHerdAffineIf(AIE::DeviceOp m) {
auto ctx = m->getContext();
RewritePatternSet patterns(ctx);
patterns.insert<SpecializeAffineIfPattern>(ctx);
(void)applyPatternsAndFoldGreedily(m, std::move(patterns));
(void)applyPatternsGreedily(m, std::move(patterns));
}

struct LowerAIRExecutePattern : public OpRewritePattern<air::ExecuteOp> {
Expand Down Expand Up @@ -720,7 +720,7 @@ void lowerAirExecute(AIE::DeviceOp d) {
patterns.insert<LowerAIRExecutePattern,
CanonicalizeChannelPutWrapsAndStridesPattern,
CanonicalizeChannelGetWrapsAndStridesPattern>(ctx);
(void)applyPatternsAndFoldGreedily(d, std::move(patterns));
(void)applyPatternsGreedily(d, std::move(patterns));
}

struct LowerScfTokenPattern : public OpRewritePattern<scf::ForOp> {
Expand Down Expand Up @@ -813,7 +813,7 @@ void lowerScfAirTokens(AIE::DeviceOp m) {
auto ctx = m->getContext();
RewritePatternSet patterns(ctx);
patterns.insert<LowerScfTokenPattern>(ctx);
(void)applyPatternsAndFoldGreedily(m, std::move(patterns));
(void)applyPatternsGreedily(m, std::move(patterns));
}

struct AllocL1BuffersPattern : public OpRewritePattern<memref::AllocOp> {
Expand Down Expand Up @@ -935,7 +935,7 @@ void allocL1Buffers(AIE::DeviceOp m,
RewritePatternSet patterns(ctx);
patterns.insert<AllocL1BuffersPattern>(ctx, tileToHerdMap, BufferId);
// AllocL1TensorsPattern
(void)applyPatternsAndFoldGreedily(m, std::move(patterns));
(void)applyPatternsGreedily(m, std::move(patterns));
}

bool areReferencedByTheSameAIRChannel(Value memref_a, Value memref_b) {
Expand Down Expand Up @@ -1027,7 +1027,7 @@ void allocL2Buffers(AIE::DeviceOp m,
L2MemrefToMemTileMap(m, memrefToTileMap);
patterns.insert<AllocL2BuffersPattern>(ctx, memrefToTileMap,
bufferToMemtileMap, BufferId);
(void)applyPatternsAndFoldGreedily(m, std::move(patterns));
(void)applyPatternsGreedily(m, std::move(patterns));
}

// Remove L2 temporary buffer allocs now that
Expand Down Expand Up @@ -1359,7 +1359,7 @@ void lowerAIRChannels(
std::map<Operation *, AIE::ObjectFifoCreateOp> linksToComplete;
patterns.insert<LowerAIRChannelsPattern>(ctx, s, bufferToMemtileMap,
linksToComplete);
(void)applyPatternsAndFoldGreedily(d, std::move(patterns));
(void)applyPatternsGreedily(d, std::move(patterns));
}

struct SpecializeChannelBundlePattern
Expand Down Expand Up @@ -1561,7 +1561,7 @@ void specializeChannelBundle(
auto ctx = d->getContext();
RewritePatternSet patterns(ctx);
patterns.insert<SpecializeChannelBundlePattern>(ctx, chan_to_chan_map);
(void)applyPatternsAndFoldGreedily(d, std::move(patterns));
(void)applyPatternsGreedily(d, std::move(patterns));
}

struct LowerAIRPingPongPattern : public OpRewritePattern<scf::ForOp> {
Expand Down Expand Up @@ -1618,7 +1618,7 @@ void LowerAIRPingPong(AIE::DeviceOp &d) {
auto ctx = d->getContext();
RewritePatternSet patterns(ctx);
patterns.insert<LowerAIRPingPongPattern>(ctx);
(void)applyPatternsAndFoldGreedily(d, std::move(patterns));
(void)applyPatternsGreedily(d, std::move(patterns));
}

template <typename OpT>
Expand Down Expand Up @@ -1879,7 +1879,7 @@ class AIRToAIEPass : public air::impl::AIRToAIEBase<AIRToAIEPass> {
RewritePatternSet patterns(ctx);
xilinx::air::populateAIRunrollAIRChannelPutGetInScfParallelPatterns(
patterns);
(void)applyPatternsAndFoldGreedily(aie_device, std::move(patterns));
(void)applyPatternsGreedily(aie_device, std::move(patterns));

// Substituting index operands, such as strides and offsets, to constant
// zero for convenience. TODO: generalize this
Expand Down Expand Up @@ -3251,7 +3251,7 @@ class AIRToAIEPass : public air::impl::AIRToAIEBase<AIRToAIEPass> {
}

if (patterns.getNativePatterns().size())
(void)applyPatternsAndFoldGreedily(m, std::move(patterns));
(void)applyPatternsGreedily(m, std::move(patterns));
}

void runOnOperation() override {
Expand Down Expand Up @@ -3460,7 +3460,7 @@ class AIRToAIEPass : public air::impl::AIRToAIEBase<AIRToAIEPass> {

RewritePatternSet patterns(ctx);
air::WaitAllOp::getCanonicalizationPatterns(patterns, ctx);
(void)applyPatternsAndFoldGreedily(device, std::move(patterns));
(void)applyPatternsGreedily(device, std::move(patterns));

// Remove ops via rewrite patterns.
RewritePatternSet removepatterns(ctx);
Expand Down Expand Up @@ -3702,7 +3702,7 @@ FailureOr<ModuleOp> convertAIRToAIE(mlir::RewriterBase &rewriter,
patterns.insert<LowerAIRExecutePattern>(ctx);
patterns.insert<AllocL1BuffersPattern>(ctx, tileToHerdMap, BufferId);
air::WaitAllOp::getCanonicalizationPatterns(patterns, ctx);
(void)applyPatternsAndFoldGreedily(aie_module, std::move(patterns));
(void)applyPatternsGreedily(aie_module, std::move(patterns));
}

return aie_module;
Expand Down
8 changes: 4 additions & 4 deletions mlir/lib/Conversion/ConvertToAIRPass.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1141,7 +1141,7 @@ struct CopyToDmaPass : public air::impl::CopyToDmaBase<CopyToDmaPass> {
linalg::getLinalgTilingCanonicalizationPatterns(context);
memref::AllocOp::getCanonicalizationPatterns(stage1Patterns, context);
memref::populateComposeSubViewPatterns(stage1Patterns, context);
(void)applyPatternsAndFoldGreedily(module, std::move(stage1Patterns));
(void)applyPatternsGreedily(module, std::move(stage1Patterns));

RewritePatternSet stage2Patterns(context);
stage2Patterns
Expand Down Expand Up @@ -1174,7 +1174,7 @@ struct CopyToDmaPass : public air::impl::CopyToDmaBase<CopyToDmaPass> {

RewritePatternSet pattern(context);
air::DmaMemcpyNdOp::getCanonicalizationPatterns(pattern, context);
(void)applyPatternsAndFoldGreedily(module, std::move(pattern));
(void)applyPatternsGreedily(module, std::move(pattern));
}
};

Expand Down Expand Up @@ -1613,7 +1613,7 @@ transform::ParToHerdOp::applyToOne(transform::TransformRewriter &rewriter,
getFirstDim());
patterns.add<ScfForallToHerdConversion>(ctx, filteredOps, herdOps,
getFirstDim());
(void)applyPatternsAndFoldGreedily(
(void)applyPatternsGreedily(
target->getParentWithTrait<OpTrait::IsIsolatedFromAbove>(),
std::move(patterns));
for (auto h : herdOps) {
Expand Down Expand Up @@ -1641,7 +1641,7 @@ transform::ParToLaunchOp::applyToOne(transform::TransformRewriter &rewriter,
getHasAirSegment());
patterns.add<ScfForallToLaunchConversion>(ctx, filteredOps, launchOps,
getHasAirSegment());
(void)applyPatternsAndFoldGreedily(
(void)applyPatternsGreedily(
target->getParentWithTrait<OpTrait::IsIsolatedFromAbove>(),
std::move(patterns));
for (auto l : launchOps)
Expand Down
Loading

0 comments on commit 06249fd

Please sign in to comment.