add support for operations: reduceMean, add, mul, div, sigmoid

src/PIM/Compiler/PimCodeGen.cpp

@@ -121,6 +121,13 @@ json::Object PimCodeGen::createEmptyOffset() {
   return offset;
 }
 
+static json::Object createRs1OnlyOffset() {
+  json::Object offset;
+  offset["offset_select"] = 1;
+  offset["offset_value"] = 0;
+  return offset;
+}
+
 void PimCodeGen::emitInstruction(json::Object instruction) const {
   coreFileStream << json::Value(std::move(instruction)) << ',';
 }
@@ -331,7 +338,8 @@ void PimCodeGen::codeGenVAvgOp(pim::PimVAvgOp vavgOp) const {
   json["op"] = "vavg";
   json["rd"] = 0;
   json["rs1"] = 1;
-  json["offset"] = createEmptyOffset();
+  json["rs2"] = 1;
+  json["offset"] = createRs1OnlyOffset();
   json["len"] = getValueSizeInBytes(vavgOp.getInput());
   emitInstruction(std::move(json));
 }

src/PIM/Conversion/ONNXToSpatial/CMakeLists.txt

@@ -4,10 +4,13 @@ add_public_tablegen_target(ONNXToSpatialIncGen)
 
 add_pim_library(OMONNXToSpatial
   Patterns/Math/Conv.cpp
+  Patterns/Math/Elementwise.cpp
   Patterns/Math/Gemm.cpp
   Patterns/Math/MatMul.cpp
+  Patterns/Math/ReduceMean.cpp
   Patterns/NN/Pool.cpp
   Patterns/NN/Relu.cpp
+  Patterns/NN/Sigmoid.cpp
   Patterns/Tensor/Concat.cpp
   Patterns/Tensor/Reshape.cpp
   ONNXToSpatialPass.cpp

src/PIM/Conversion/ONNXToSpatial/Common.hpp

@@ -14,8 +14,6 @@
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
 
-#define DEFINE_MAP_OP(opname) opname,
-
 namespace onnx_mlir {
 
 template <class ShapedType>

src/PIM/Conversion/ONNXToSpatial/ONNXToSpatialPass.cpp

@@ -72,11 +72,15 @@ void ONNXToSpatialPass::runOnOperation() {
   target.addLegalDialect<spatial::SpatialDialect, ONNXDialect, tensor::TensorDialect, arith::ArithDialect>();
   target.addDynamicallyLegalOp<ONNXMatMulOp>(
     [](ONNXMatMulOp op) { return cast<ShapedType>(op.getY().getType()).getRank() != 2; });
+  target.addIllegalOp<ONNXAddOp>();
+  target.addIllegalOp<ONNXDivOp>();
+  target.addIllegalOp<ONNXMulOp>();
   target.addIllegalOp<ONNXGemmOp>();
   target.addIllegalOp<ONNXConvOp>();
   target.addIllegalOp<ONNXMaxPoolSingleOutOp>();
   target.addIllegalOp<ONNXAveragePoolOp>();
   target.addIllegalOp<ONNXReluOp>();
+  target.addIllegalOp<ONNXSigmoidOp>();
   target.addIllegalOp<ONNXSoftmaxOp>();
   target.addIllegalOp<ONNXConcatOp>();
   target.addIllegalOp<ONNXReshapeOp>();
@@ -86,10 +90,13 @@ void ONNXToSpatialPass::runOnOperation() {
   RewritePatternSet patterns(ctx);
   patterns.add<removeLRN>(ctx);
 
+  populateElementwisePatterns(patterns, ctx);
   populateGemmPatterns(patterns, ctx);
   populateConvPatterns(patterns, ctx);
   populatePoolPatterns(patterns, ctx);
+  populateReduceMeanPatterns(patterns, ctx);
   populateReluPatterns(patterns, ctx);
+  populateSigmoidPatterns(patterns, ctx);
   populateConcatPatterns(patterns, ctx);
   populateReshapePatterns(patterns, ctx);
 

src/PIM/Conversion/ONNXToSpatial/Patterns.hpp

@@ -7,14 +7,20 @@ namespace onnx_mlir {
 
 void populateConvPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
 
+void populateElementwisePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
+
 void populateGemmPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
 
 void populateMatMulRewritePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
 
 void populatePoolPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
 
+void populateReduceMeanPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
+
 void populateReluPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
 
+void populateSigmoidPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
+
 void populateConcatPatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);
 
 void populateReshapePatterns(mlir::RewritePatternSet& patterns, mlir::MLIRContext* ctx);

src/PIM/Conversion/ONNXToSpatial/Patterns/Math/Elementwise.cpp

@@ -0,0 +1,204 @@
+#include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/IR/BuiltinAttributes.h"
+#include "mlir/IR/BuiltinTypes.h"
+#include "mlir/Transforms/DialectConversion.h"
+
+#include "llvm/ADT/SmallVector.h"
+
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
+#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
+#include "src/Dialect/ONNX/ONNXOps.hpp"
+
+using namespace mlir;
+
+namespace onnx_mlir {
+namespace {
+
+static SmallVector<int64_t> computeRowMajorStrides(ArrayRef<int64_t> shape) {
+  SmallVector<int64_t> strides(shape.size(), 1);
+  for (int64_t i = static_cast<int64_t>(shape.size()) - 2; i >= 0; --i)
+    strides[i] = strides[i + 1] * shape[i + 1];
+  return strides;
+}
+
+static DenseElementsAttr getDenseConstantAttr(Value value) {
+  if (auto constantOp = value.getDefiningOp<arith::ConstantOp>())
+    return dyn_cast<DenseElementsAttr>(constantOp.getValue());
+
+  if (auto constantOp = value.getDefiningOp<ONNXConstantOp>())
+    return dyn_cast_or_null<DenseElementsAttr>(constantOp.getValueAttr());
+
+  return nullptr;
+}
+
+static FailureOr<Value> materializeBroadcastedConstantTensor(Value value,
+                                                             RankedTensorType resultType,
+                                                             ConversionPatternRewriter& rewriter,
+                                                             Location loc) {
+  auto denseAttr = getDenseConstantAttr(value);
+  if (!denseAttr)
+    return failure();
+
+  auto sourceType = dyn_cast<RankedTensorType>(denseAttr.getType());
+  if (!sourceType || !sourceType.hasStaticShape() || !resultType.hasStaticShape())
+    return failure();
+
+  if (sourceType == resultType)
+    return value;
+
+  ArrayRef<int64_t> sourceShape = sourceType.getShape();
+  ArrayRef<int64_t> resultShape = resultType.getShape();
+  if (sourceShape.size() > resultShape.size())
+    return failure();
+
+  const int64_t rankOffset = static_cast<int64_t>(resultShape.size() - sourceShape.size());
+  for (int64_t i = 0; i < static_cast<int64_t>(resultShape.size()); ++i) {
+    const int64_t sourceIndex = i - rankOffset;
+    const int64_t sourceDim = sourceIndex < 0 ? 1 : sourceShape[sourceIndex];
+    const int64_t resultDim = resultShape[i];
+    if (sourceDim != 1 && sourceDim != resultDim)
+      return failure();
+  }
+
+  SmallVector<Attribute> sourceValues(denseAttr.getValues<Attribute>());
+  SmallVector<int64_t> sourceStrides = computeRowMajorStrides(sourceShape);
+  SmallVector<int64_t> resultStrides = computeRowMajorStrides(resultShape);
+
+  SmallVector<Attribute> resultValues;
+  resultValues.reserve(resultType.getNumElements());
+
+  for (int64_t flatIndex = 0; flatIndex < resultType.getNumElements(); ++flatIndex) {
+    int64_t remaining = flatIndex;
+    int64_t sourceFlatIndex = 0;
+
+    for (int64_t i = 0; i < static_cast<int64_t>(resultShape.size()); ++i) {
+      const int64_t resultIndex = resultStrides.empty() ? 0 : remaining / resultStrides[i];
+      remaining = resultStrides.empty() ? 0 : remaining % resultStrides[i];
+
+      const int64_t sourceIndex = i - rankOffset;
+      if (sourceIndex < 0)
+        continue;
+
+      const int64_t sourceDim = sourceShape[sourceIndex];
+      const int64_t mappedIndex = sourceDim == 1 ? 0 : resultIndex;
+      sourceFlatIndex += mappedIndex * sourceStrides[sourceIndex];
+    }
+
+    resultValues.push_back(sourceValues[sourceFlatIndex]);
+  }
+
+  auto broadcastedAttr = DenseElementsAttr::get(resultType, resultValues);
+  return arith::ConstantOp::create(rewriter, loc, resultType, broadcastedAttr).getResult();
+}
+
+static FailureOr<Value> prepareElementwiseOperand(Value value,
+                                                  RankedTensorType resultType,
+                                                  ConversionPatternRewriter& rewriter,
+                                                  Location loc) {
+  auto valueType = dyn_cast<RankedTensorType>(value.getType());
+  if (!valueType || !valueType.hasStaticShape())
+    return failure();
+
+  if (valueType == resultType)
+    return value;
+
+  return materializeBroadcastedConstantTensor(value, resultType, rewriter, loc);
+}
+
+static FailureOr<Value> materializeReciprocalTensor(Value value,
+                                                    RankedTensorType resultType,
+                                                    ConversionPatternRewriter& rewriter,
+                                                    Location loc) {
+  auto broadcastedValue = materializeBroadcastedConstantTensor(value, resultType, rewriter, loc);
+  if (failed(broadcastedValue))
+    return failure();
+
+  auto denseAttr = dyn_cast<DenseFPElementsAttr>(getDenseConstantAttr(*broadcastedValue));
+  if (!denseAttr)
+    return failure();
+
+  SmallVector<APFloat> reciprocalValues;
+  reciprocalValues.reserve(denseAttr.getNumElements());
+  for (const APFloat& valueAttr : denseAttr.getValues<APFloat>()) {
+    APFloat reciprocal(valueAttr.getSemantics(), 1);
+    auto status = reciprocal.divide(valueAttr, APFloat::rmNearestTiesToEven);
+    if (status & APFloat::opInvalidOp)
+      return failure();
+    reciprocalValues.push_back(std::move(reciprocal));
+  }
+
+  auto reciprocalAttr = DenseFPElementsAttr::get(resultType, reciprocalValues);
+  return arith::ConstantOp::create(rewriter, loc, resultType, reciprocalAttr).getResult();
+}
+
+template <typename OnnxOp, typename SpatialOp>
+struct BinaryElementwiseToSpatialCompute : OpConversionPattern<OnnxOp> {
+  using OpConversionPattern<OnnxOp>::OpConversionPattern;
+  using Adaptor = typename OnnxOp::Adaptor;
+
+  LogicalResult matchAndRewrite(OnnxOp op, Adaptor adaptor, ConversionPatternRewriter& rewriter) const override {
+    auto resultType = dyn_cast<RankedTensorType>(op->getResult(0).getType());
+    if (!resultType || !resultType.hasStaticShape())
+      return failure();
+
+    Location loc = op.getLoc();
+    auto lhs = prepareElementwiseOperand(adaptor.getOperands()[0], resultType, rewriter, loc);
+    if (failed(lhs))
+      return failure();
+
+    auto rhs = prepareElementwiseOperand(adaptor.getOperands()[1], resultType, rewriter, loc);
+    if (failed(rhs))
+      return failure();
+
+    constexpr size_t numInputs = 2;
+    auto computeOp =
+      createSpatCompute<numInputs>(rewriter, loc, resultType, {}, ValueRange {*lhs, *rhs}, [&](Value x, Value y) {
+        auto loweredOp = SpatialOp::create(rewriter, loc, resultType, x, y);
+        spatial::SpatYieldOp::create(rewriter, loc, loweredOp.getResult());
+      });
+
+    rewriter.replaceOp(op, computeOp);
+    return success();
+  }
+};
+
+struct DivToSpatialCompute : OpConversionPattern<ONNXDivOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(ONNXDivOp op, ONNXDivOpAdaptor adaptor, ConversionPatternRewriter& rewriter) const override {
+    auto resultType = dyn_cast<RankedTensorType>(op.getResult().getType());
+    if (!resultType || !resultType.hasStaticShape())
+      return failure();
+
+    Location loc = op.getLoc();
+    auto lhs = prepareElementwiseOperand(adaptor.getA(), resultType, rewriter, loc);
+    if (failed(lhs))
+      return failure();
+
+    auto reciprocalRhs = materializeReciprocalTensor(adaptor.getB(), resultType, rewriter, loc);
+    if (failed(reciprocalRhs))
+      return failure();
+
+    constexpr size_t numInputs = 2;
+    auto computeOp = createSpatCompute<numInputs>(
+      rewriter, loc, resultType, {}, ValueRange {*lhs, *reciprocalRhs}, [&](Value x, Value reciprocal) {
+        auto mulOp = spatial::SpatVMulOp::create(rewriter, loc, resultType, x, reciprocal);
+        spatial::SpatYieldOp::create(rewriter, loc, mulOp.getResult());
+      });
+
+    rewriter.replaceOp(op, computeOp);
+    return success();
+  }
+};
+
+} // namespace
+
+void populateElementwisePatterns(RewritePatternSet& patterns, MLIRContext* ctx) {
+  patterns.add<BinaryElementwiseToSpatialCompute<ONNXAddOp, spatial::SpatVAddOp>>(ctx);
+  patterns.add<BinaryElementwiseToSpatialCompute<ONNXMulOp, spatial::SpatVMulOp>>(ctx);
+  patterns.add<DivToSpatialCompute>(ctx);
+}
+
+} // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/Patterns/Math/ReduceMean.cpp

@@ -0,0 +1,163 @@
+#include "mlir/Dialect/Tensor/IR/Tensor.h"
+#include "mlir/Transforms/DialectConversion.h"
+
+#include "llvm/ADT/SmallVector.h"
+
+#include <algorithm>
+
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Patterns.hpp"
+#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
+#include "src/Dialect/ONNX/ONNXOps.hpp"
+
+using namespace mlir;
+
+namespace onnx_mlir {
+namespace {
+
+static SmallVector<int64_t> normalizeAxes(ArrayAttr axesAttr, int64_t rank) {
+  SmallVector<int64_t> normalizedAxes;
+  if (!axesAttr) {
+    normalizedAxes.reserve(rank);
+    for (int64_t axis = 0; axis < rank; axis++)
+      normalizedAxes.push_back(axis);
+    return normalizedAxes;
+  }
+
+  normalizedAxes.reserve(axesAttr.size());
+  for (Attribute attr : axesAttr) {
+    int64_t axis = cast<IntegerAttr>(attr).getInt();
+    normalizedAxes.push_back(axis >= 0 ? axis : rank + axis);
+  }
+
+  llvm::sort(normalizedAxes);
+  normalizedAxes.erase(std::unique(normalizedAxes.begin(), normalizedAxes.end()), normalizedAxes.end());
+  return normalizedAxes;
+}
+
+static SmallVector<bool> buildReducedAxesMask(ArrayRef<int64_t> axes, int64_t rank) {
+  SmallVector<bool> reducedAxes(rank, false);
+  for (int64_t axis : axes) {
+    if (axis < 0 || axis >= rank)
+      return {};
+    reducedAxes[axis] = true;
+  }
+  return reducedAxes;
+}
+
+static RankedTensorType getAllOnesType(RankedTensorType inputType, Type elementType) {
+  return RankedTensorType::get(SmallVector<int64_t>(inputType.getRank(), 1), elementType);
+}
+
+static SmallVector<ReassociationIndices> buildCollapseReassociation(ArrayRef<bool> reducedAxes) {
+  SmallVector<ReassociationIndices> reassociation;
+  ReassociationIndices currentGroup;
+
+  for (auto [axis, isReduced] : llvm::enumerate(reducedAxes)) {
+    currentGroup.push_back(axis);
+    if (!isReduced) {
+      reassociation.push_back(currentGroup);
+      currentGroup.clear();
+    }
+  }
+
+  if (!currentGroup.empty()) {
+    if (reassociation.empty())
+      reassociation.push_back(std::move(currentGroup));
+    else
+      reassociation.back().append(currentGroup.begin(), currentGroup.end());
+  }
+
+  return reassociation;
+}
+
+static Value createAverageCompute(Value input,
+                                  RankedTensorType resultType,
+                                  ConversionPatternRewriter& rewriter,
+                                  Location loc) {
+  constexpr size_t numInputs = 1;
+  auto computeOp = createSpatCompute<numInputs>(rewriter, loc, resultType, {}, ValueRange {input}, [&](Value x) {
+    auto avgOp = spatial::SpatVAvgOp::create(rewriter, loc, resultType, x);
+    spatial::SpatYieldOp::create(rewriter, loc, avgOp.getResult());
+  });
+  return computeOp.getResult(0);
+}
+
+static Value buildReduceMeanKeepdims(Value input,
+                                     ArrayRef<bool> reducedAxes,
+                                     int64_t axis,
+                                     RankedTensorType leafType,
+                                     ConversionPatternRewriter& rewriter,
+                                     Location loc) {
+  int64_t rank = cast<RankedTensorType>(input.getType()).getRank();
+  if (axis == rank)
+    return createAverageCompute(input, leafType, rewriter, loc);
+
+  if (reducedAxes[axis])
+    return buildReduceMeanKeepdims(input, reducedAxes, axis + 1, leafType, rewriter, loc);
+
+  SmallVector<Value> slices = sliceTensor(input, axis, /*sliceSize=*/1, rewriter, loc);
+  SmallVector<Value> reducedSlices;
+  reducedSlices.reserve(slices.size());
+  for (Value slice : slices)
+    reducedSlices.push_back(buildReduceMeanKeepdims(slice, reducedAxes, axis + 1, leafType, rewriter, loc));
+
+  return reducedSlices.size() == 1 ? reducedSlices.front()
+                                   : tensor::ConcatOp::create(rewriter, loc, axis, reducedSlices).getResult();
+}
+
+static Value squeezeReducedAxes(Value keepdimsValue,
+                                RankedTensorType resultType,
+                                ArrayRef<bool> reducedAxes,
+                                ConversionPatternRewriter& rewriter,
+                                Location loc) {
+  if (resultType.getRank() == 0) {
+    SmallVector<Value> indices(cast<RankedTensorType>(keepdimsValue.getType()).getRank(),
+                               arith::ConstantIndexOp::create(rewriter, loc, 0));
+    Value element = tensor::ExtractOp::create(rewriter, loc, keepdimsValue, indices);
+    return tensor::FromElementsOp::create(rewriter, loc, resultType, ValueRange {element});
+  }
+
+  return tensor::CollapseShapeOp::create(
+           rewriter, loc, resultType, keepdimsValue, buildCollapseReassociation(reducedAxes))
+    .getResult();
+}
+
+struct ReduceMeanToSpatialCompute : OpConversionPattern<ONNXReduceMeanV13Op> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult matchAndRewrite(ONNXReduceMeanV13Op reduceMeanOp,
+                                ONNXReduceMeanV13OpAdaptor adaptor,
+                                ConversionPatternRewriter& rewriter) const override {
+    auto inputType = dyn_cast<RankedTensorType>(adaptor.getData().getType());
+    auto resultType = dyn_cast<RankedTensorType>(reduceMeanOp.getReduced().getType());
+    if (!inputType || !resultType || !inputType.hasStaticShape() || !resultType.hasStaticShape())
+      return failure();
+
+    SmallVector<int64_t> axes = normalizeAxes(reduceMeanOp.getAxesAttr(), inputType.getRank());
+    SmallVector<bool> reducedAxes = buildReducedAxesMask(axes, inputType.getRank());
+    if (reducedAxes.empty() && inputType.getRank() != 0)
+      return failure();
+
+    Location loc = reduceMeanOp.getLoc();
+    RankedTensorType leafType = getAllOnesType(inputType, resultType.getElementType());
+    Value reducedKeepdims = buildReduceMeanKeepdims(adaptor.getData(), reducedAxes, /*axis=*/0, leafType, rewriter, loc);
+
+    if (reduceMeanOp.getKeepdims() != 0) {
+      rewriter.replaceOp(reduceMeanOp, reducedKeepdims);
+      return success();
+    }
+
+    Value reduced = squeezeReducedAxes(reducedKeepdims, resultType, reducedAxes, rewriter, loc);
+    rewriter.replaceOp(reduceMeanOp, reduced);
+    return success();
+  }
+};
+
+} // namespace
+
+void populateReduceMeanPatterns(RewritePatternSet& patterns, MLIRContext* ctx) {
+  patterns.add<ReduceMeanToSpatialCompute>(ctx);
+}
+
+} // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/Patterns/NN/Sigmoid.cpp

@@ -0,0 +1,36 @@
+#include "mlir/Transforms/DialectConversion.h"
+
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common.hpp"
+#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
+#include "src/Dialect/ONNX/ONNXOps.hpp"
+
+using namespace mlir;
+
+namespace onnx_mlir {
+namespace {
+
+struct SigmoidToSpatialCompute : OpConversionPattern<ONNXSigmoidOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult matchAndRewrite(ONNXSigmoidOp sigmoidOp,
+                                ONNXSigmoidOpAdaptor adaptor,
+                                ConversionPatternRewriter& rewriter) const override {
+    Location loc = sigmoidOp.getLoc();
+    Type resultType = sigmoidOp.getResult().getType();
+    constexpr size_t numInputs = 1;
+    auto computeOp = createSpatCompute<numInputs>(rewriter, loc, resultType, {}, adaptor.getX(), [&](Value x) {
+      auto spatSigmoidOp = spatial::SpatSigmoidOp::create(rewriter, loc, resultType, x);
+      spatial::SpatYieldOp::create(rewriter, loc, spatSigmoidOp.getResult());
+    });
+    rewriter.replaceOp(sigmoidOp, computeOp);
+    return success();
+  }
+};
+
+} // namespace
+
+void populateSigmoidPatterns(RewritePatternSet& patterns, MLIRContext* ctx) {
+  patterns.add<SigmoidToSpatialCompute>(ctx);
+}
+
+} // namespace onnx_mlir

src/PIM/Conversion/SpatialToPim/SpatialToPim.td

@@ -39,6 +39,12 @@ def spatToPimVVMul : Pat<
     (NativeCodeCall<"onnx_mlir::getBestOutputTensorFromOperandsOrAllocate($_builder, $0.getDefiningOp())"> $srcOpRes))
 >;
 
+def spatToPimVAvg : Pat<
+  (SpatVAvgOp:$srcOpRes $input),
+  (PimVAvgOp $input,
+    (NativeCodeCall<"onnx_mlir::getBestOutputTensorFromOperandsOrAllocate($_builder, $0.getDefiningOp())"> $srcOpRes))
+>;
+
 def spatToPimVVMax : Pat<
   (SpatVMaxOp:$srcOpRes $a, $b),
   (PimVVMaxOp $a, $b,
@@ -51,4 +57,10 @@ def spatToPimVRelu : Pat<
     (NativeCodeCall<"onnx_mlir::getBestOutputTensorFromOperandsOrAllocate($_builder, $0.getDefiningOp())"> $srcOpRes))
 >;
 
+def spatToPimVSigm : Pat<
+  (SpatSigmoidOp:$srcOpRes $input),
+  (PimVSigmOp $input,
+    (NativeCodeCall<"onnx_mlir::getBestOutputTensorFromOperandsOrAllocate($_builder, $0.getDefiningOp())"> $srcOpRes))
+>;
+
 #endif // SPATIAL_TO_PIM

src/PIM/Conversion/SpatialToPim/SpatialToPimPass.cpp

@@ -161,26 +161,41 @@ void SpatialToPimPass::runOnOperation() {
   }
 
   for (auto receiveOp : funcOp.getOps<spatial::SpatChannelReceiveOp>()) {
-    operationsToRemove.push_back(receiveOp);
+    markOpToRemove(receiveOp);
     runOnReceiveOp(receiveOp, rewriter);
   }
   for (auto computeOp : funcOp.getOps<spatial::SpatWeightedCompute>()) {
-    operationsToRemove.push_back(computeOp);
+    markOpToRemove(computeOp);
     runOnComputeOp(computeOp, rewriter);
   }
 
   enlargeVMMOutTensorsToCrossbarSize(funcOp, rewriter);
   replaceReturnOpOperands(returnOp, rewriter);
 
-  // Remove all ComputeOps
-  for (auto opToRemove : llvm::reverse(operationsToRemove)) {
-    if (!opToRemove->use_empty()) {
+  SmallVector<Operation*> pendingRemovals(operationsToRemove.begin(), operationsToRemove.end());
+  while (!pendingRemovals.empty()) {
+    bool erasedAnyOp = false;
+    for (auto it = pendingRemovals.begin(); it != pendingRemovals.end();) {
+      Operation* opToRemove = *it;
+      if (!opToRemove->use_empty()) {
+        ++it;
+        continue;
+      }
+
+      rewriter.eraseOp(opToRemove);
+      it = pendingRemovals.erase(it);
+      erasedAnyOp = true;
+    }
+
+    if (erasedAnyOp)
+      continue;
+
+    for (auto opToRemove : pendingRemovals) {
       opToRemove->dump();
       for (auto user : opToRemove->getUsers())
         user->dump();
-      assert(false && "opToRemove should be unused at this point");
     }
-    rewriter.eraseOp(opToRemove);
+    assert(false && "tracked op removal reached a cycle or missed dependency");
   }
 
   // Dump to file for debug
@@ -284,10 +299,19 @@ void SpatialToPimPass::runOnComputeOp(spatial::SpatWeightedCompute computeOp, IR
         auto concatUses = concatValue.getUses();
         auto numConcatUses = rangeLength(concatUses);
         if (numConcatUses == 1) {
-          OpOperand& concatUse = *concatUses.begin();
-          Operation* concatUser = concatUse.getOwner();
+          Value chainedValue = concatValue;
+          Operation* concatUser = concatUses.begin()->getOwner();
+
+          while (isChannelUseChainOp(concatUser)) {
+            auto chainUses = concatUser->getResult(0).getUses();
+            if (rangeLength(chainUses) != 1)
+              break;
+            chainedValue = concatUser->getResult(0);
+            concatUser = chainUses.begin()->getOwner();
+          }
+
           if (isa<func::ReturnOp>(concatUser)) {
-            size_t concatIndexInReturn = concatUse.getOperandNumber();
+            size_t concatIndexInReturn = chainedValue.getUses().begin()->getOperandNumber();
             size_t resultIndexInConcat = resultUses.begin()->getOperandNumber();
             size_t offset = 0;
             for (auto operand : concatOp->getOperands().take_front(resultIndexInConcat))
@@ -602,10 +626,22 @@ void SpatialToPimPass::replaceReturnOpOperands(func::ReturnOp& returnOp, IRRewri
     rewriter.modifyOpInPlace(returnOp,
                              [&] { returnOp.setOperand(orderWithinReturn, outputTensors[orderWithinReturn]); });
 
-    if (isa<tensor::ConcatOp>(returnOperand)) {
-      auto returnOperandUses = it.value().getUses();
-      if (rangeLength(returnOperandUses) == 0)
-        rewriter.eraseOp(returnOperand);
+    Operation* opToErase = returnOperand;
+    while (opToErase) {
+      bool isExclusivelyOwnedByReturnChain = opToErase->use_empty() || opToErase->hasOneUse();
+      if (!isExclusivelyOwnedByReturnChain)
+        break;
+
+      if (isChannelUseChainOp(opToErase)) {
+        Value source = opToErase->getOperand(0);
+        markOpToRemove(opToErase);
+        opToErase = source.getDefiningOp();
+        continue;
+      }
+
+      if (isa<tensor::ConcatOp>(opToErase))
+        markOpToRemove(opToErase);
+      break;
     }
   }
 }

src/PIM/Dialect/Spatial/Spatial.td

@@ -239,6 +239,22 @@ def SpatSumOp : SpatOp<"sum", []> {
   }];
 }
 
+def SpatVAvgOp : SpatOp<"vavg", []> {
+  let summary = "Average all elements of the input tensor to a single scalar wrapped in a tensor";
+
+  let arguments = (ins
+    SpatTensor:$input
+  );
+
+  let results = (outs
+    SpatTensor:$output
+  );
+
+  let assemblyFormat = [{
+    `(` $input `)` attr-dict `:` type($input) `->` type($output)
+  }];
+}
+
 def SpatSigmoidOp : SpatOp<"sigmoid", []> {
   let summary = "Element-wise sigmoid activation";
 

src/PIM/Dialect/Spatial/Transforms/SpatialBufferizableOpInterface.cpp

@@ -361,7 +361,7 @@ struct ChannelBroadcastReceiveOpInterface
   }
 
   /*
-   * Turn the channel receive to pim.load using by creating a new global buffer
+   * Turn the broadcast receive into a regular pim.receive from the broadcaster.
    */
   LogicalResult bufferize(Operation* op,
                           RewriterBase& rewriter,
@@ -370,8 +370,21 @@ struct ChannelBroadcastReceiveOpInterface
 
     auto outputTensor = createEmptyFromType(op->getResult(0).getType(), op->getLoc(), rewriter);
 
-    auto outputType = cast<ShapedType>(outputTensor.getType());
-    auto outputSize = outputType.getNumElements() * outputType.getElementTypeBitWidth() / 8;
+    auto numElements = cast<ShapedType>(outputTensor.getType()).getNumElements();
+    auto elementSize = cast<ShapedType>(outputTensor.getType()).getElementTypeBitWidth() / 8;
+
+    auto precomputedOtherCoreId = op->getAttr(PRECOMPUTED_OTHER_CORE_ID_ATTR_NAME);
+    if (precomputedOtherCoreId) {
+      Value newValue = pim::PimReceiveOp::create(rewriter,
+                                                 op->getLoc(),
+                                                 outputTensor.getType(),
+                                                 outputTensor,
+                                                 rewriter.getI32IntegerAttr(numElements * elementSize),
+                                                 cast<IntegerAttr>(precomputedOtherCoreId))
+                         .getOutput();
+      replaceOpWithBufferizedValues(rewriter, op, newValue);
+      return success();
+    }
 
     auto channelNewOp = op->getOperand(0).getDefiningOp<SpatChannelNewOp>();
     if (!channelNewOp) {
@@ -379,31 +392,30 @@ struct ChannelBroadcastReceiveOpInterface
       return failure();
     }
 
-    // The first 'broadcast' operation creates the buffer just after the
-    // channelNewOp, while the other 'broadcast' operation need to find this
-    // buffer allocation just after the channelNewOp
-    Value bufferAllocation;
-    if (auto allocOpAfterChannel = dyn_cast<memref::AllocOp>(channelNewOp->getNextNode())) {
-      // Buffer already allocated, load from this buffer
-      bufferAllocation = allocOpAfterChannel;
-    }
-    else {
-      // Buffer was not allocated previously, allocate it after channelNewOp
-      rewriter.setInsertionPointAfter(channelNewOp);
-      bufferAllocation = createEmptyFromType(op->getResult(0).getType(), op->getLoc(), rewriter);
-    }
+    auto srcCoreId = [&]() -> FailureOr<uint32_t> {
+      for (Operation* user : channelNewOp->getUsers()) {
+        auto sendOp = dyn_cast<SpatChannelBroadcastSendOp>(user);
+        if (!sendOp)
+          continue;
+        auto sendCoreIdAttr = cast<pim::PimCoreOp>(sendOp->getParentOp()).getCoreIdAttr();
+        op->setAttr(PRECOMPUTED_OTHER_CORE_ID_ATTR_NAME, sendCoreIdAttr);
+        return cast<pim::PimCoreOp>(sendOp->getParentOp()).getCoreId();
+      }
+      op->emitError("ChannelBroadcastReceiveOp has no matching ChannelBroadcastSendOp");
+      return failure();
+    }();
+    if (failed(srcCoreId))
+      return failure();
 
-    rewriter.setInsertionPoint(op);
-    auto memCopyHostToDevOp = pim::PimMemCopyHostToDevOp::create(rewriter,
-                                                                 op->getLoc(),
-                                                                 outputTensor.getType(),
-                                                                 outputTensor,
-                                                                 bufferAllocation,
-                                                                 rewriter.getI32IntegerAttr(0),
-                                                                 rewriter.getI32IntegerAttr(0),
-                                                                 rewriter.getI32IntegerAttr(outputSize));
+    Value newValue = pim::PimReceiveOp::create(rewriter,
+                                               op->getLoc(),
+                                               outputTensor.getType(),
+                                               outputTensor,
+                                               rewriter.getI32IntegerAttr(numElements * elementSize),
+                                               rewriter.getI32IntegerAttr(srcCoreId.value()))
+                     .getOutput();
 
-    replaceOpWithBufferizedValues(rewriter, op, memCopyHostToDevOp.getOutput());
+    replaceOpWithBufferizedValues(rewriter, op, newValue);
 
     return success();
   }
@@ -428,8 +440,7 @@ struct ChannelBroadcastSendOpInterface
   }
 
   /*
-   * Turn the channel send into a device-to-host copy into the shared
-   * broadcast buffer that receive ops load from later.
+   * Turn the broadcast send into one pim.send per broadcast receiver.
    */
   LogicalResult bufferize(Operation* op,
                           RewriterBase& rewriter,
@@ -448,32 +459,32 @@ struct ChannelBroadcastSendOpInterface
       return failure();
     }
 
-    // The first 'broadcast' operation creates the buffer just after the
-    // channelNewOp, while the other 'broadcast' operation need to find this
-    // buffer allocation just after the channelNewOp
-    Value bufferAllocation;
-    if (auto allocOpAfterChannel = dyn_cast<memref::AllocOp>(channelNewOp->getNextNode())) {
-      // Buffer already allocated, load from this buffer
-      bufferAllocation = allocOpAfterChannel;
-    }
-    else {
-      // Buffer was not allocated previously, allocate it after channelNewOp
-      rewriter.setInsertionPointAfter(channelNewOp);
-      bufferAllocation = createEmptyFromType(srcTensor.getType(), op->getLoc(), rewriter);
-    }
-
     auto srcType = cast<ShapedType>(srcTensor.getType());
     auto sizeInBytes = srcType.getNumElements() * srcType.getElementTypeBitWidth() / 8;
+    auto srcCoreIdAttr = cast<pim::PimCoreOp>(op->getParentOp()).getCoreIdAttr();
 
     rewriter.setInsertionPoint(op);
-    pim::PimMemCopyDevToHostOp::create(rewriter,
-                                       op->getLoc(),
-                                       bufferAllocation.getType(),
-                                       bufferAllocation,
-                                       srcMemRef,
-                                       rewriter.getI32IntegerAttr(0),
-                                       rewriter.getI32IntegerAttr(0),
-                                       rewriter.getI32IntegerAttr(sizeInBytes));
+    bool foundReceiver = false;
+    for (Operation* user : channelNewOp->getUsers()) {
+      auto receiveOp = dyn_cast<SpatChannelBroadcastReceiveOp>(user);
+      if (!receiveOp)
+        continue;
+
+      foundReceiver = true;
+      auto dstCoreId = cast<pim::PimCoreOp>(receiveOp->getParentOp()).getCoreId();
+      receiveOp->setAttr(PRECOMPUTED_OTHER_CORE_ID_ATTR_NAME, srcCoreIdAttr);
+      pim::PimSendOp::create(rewriter,
+                             op->getLoc(),
+                             srcMemRef,
+                             rewriter.getI32IntegerAttr(sizeInBytes),
+                             rewriter.getI32IntegerAttr(dstCoreId));
+    }
+
+    if (!foundReceiver) {
+      op->emitError("SpatChannelBroadcastSendOp has no matching ChannelBroadcastReceiveOp");
+      return failure();
+    }
+
     rewriter.eraseOp(op);
     return success();
   }