fix vgg16 diamonds

updat ops validations

.agents/invariants/PERFORMANCE_OPTIMIZATION_INVARIANT.md

@@ -32,6 +32,18 @@ schedule semantics, scheduling granularity, and available parallelism.
 An optimization is acceptable only when its static runtime proxies are equal or
 better.
 
+## Asymptotic cost
+
+For each new or materially changed compiler algorithm, identify the relevant
+input size and target linear or sublinear time and space. Avoid repeated full-IR
+walks, nested scans, and per-operation recomputation when indexing, caching, or
+a single traversal can express the same behavior.
+
+When linear-or-better complexity is not possible, use the lowest justified
+complexity and report the actual time and space Big-O, the input variable, and
+why a lower bound is not practical. Include that cost in the final report; do
+not hide a superlinear path behind small current test sizes.
+
 ## Forbidden optimization trades
 
 Do not:

src/PIM/Conversion/ONNXToSpatial/Common/ComputeRegionBuilder.cpp

@@ -9,6 +9,53 @@ using namespace mlir;
 
 namespace onnx_mlir {
 
+FailureOr<Value> createFragmentAssemblyBlueprint(Value physicalBatch,
+                                                 RankedTensorType logicalType,
+                                                 ArrayRef<FragmentAssemblyEntry> entries,
+                                                 StringRef physicalLayout,
+                                                 StringRef indexMap,
+                                                 PatternRewriter& rewriter,
+                                                 Location loc) {
+  auto physicalType = dyn_cast<RankedTensorType>(physicalBatch.getType());
+  if (!physicalType || !physicalType.hasStaticShape() || !logicalType || !logicalType.hasStaticShape()
+      || physicalType.getRank() != logicalType.getRank() + 1 || entries.empty())
+    return emitError(loc, "invalid static physical batch for fragment assembly"), failure();
+
+  const int64_t rank = logicalType.getRank();
+  const int64_t laneCount = physicalType.getDimSize(0);
+  if (laneCount <= 0)
+    return emitError(loc, "fragment assembly requires at least one physical source slot"), failure();
+  const int64_t fragmentElements = physicalType.getNumElements() / laneCount;
+  SmallVector<int64_t> operandIndices(entries.size(), 0), sourceSlots, sourceOffsets, offsets, sizes,
+    strides(entries.size() * rank, 1);
+  for (const FragmentAssemblyEntry& entry : entries) {
+    if (entry.sourceSlot < 0 || entry.sourceSlot >= laneCount || entry.sourceOffset < 0
+        || entry.destinationOffsets.size() != static_cast<size_t>(rank)
+        || entry.sizes.size() != static_cast<size_t>(rank))
+      return emitError(loc, "invalid fragment assembly entry"), failure();
+    int64_t entryElements = 1;
+    for (int64_t dim = 0; dim < rank; ++dim) {
+      if (entry.destinationOffsets[dim] < 0 || entry.sizes[dim] <= 0
+          || entry.destinationOffsets[dim] + entry.sizes[dim] > logicalType.getDimSize(dim))
+        return emitError(loc, "fragment assembly entry exceeds the logical tensor"), failure();
+      entryElements *= entry.sizes[dim];
+    }
+    if (entry.sourceOffset + entryElements > fragmentElements)
+      return emitError(loc, "fragment assembly entry exceeds its physical source slot"), failure();
+    sourceSlots.push_back(entry.sourceSlot);
+    sourceOffsets.push_back(entry.sourceOffset);
+    llvm::append_range(offsets, entry.destinationOffsets);
+    llvm::append_range(sizes, entry.sizes);
+  }
+  return spatial::SpatBlueprintOp::create(rewriter, loc, logicalType, physicalBatch, ValueRange {},
+    rewriter.getStringAttr("nchw"), rewriter.getStringAttr(physicalLayout),
+    rewriter.getDenseI64ArrayAttr(offsets), rewriter.getDenseI64ArrayAttr(sizes),
+    rewriter.getStringAttr(indexMap), rewriter.getStringAttr("fragment_assembly"),
+    rewriter.getDenseI64ArrayAttr(operandIndices), rewriter.getDenseI64ArrayAttr(sourceSlots),
+    rewriter.getDenseI64ArrayAttr(sourceOffsets), rewriter.getDenseI64ArrayAttr(strides),
+    rewriter.getStringAttr("disjoint"), rewriter.getStringAttr("complete")).getOutput();
+}
+
 Value sumTensors(ArrayRef<Value> tensors, PatternRewriter& rewriter) {
   if (tensors.size() == 1)
     return tensors[0];

src/PIM/Conversion/ONNXToSpatial/Common/ComputeRegionBuilder.hpp

@@ -19,6 +19,13 @@
 
 namespace onnx_mlir {
 
+struct FragmentAssemblyEntry {
+  int64_t sourceSlot;
+  int64_t sourceOffset;
+  llvm::SmallVector<int64_t, 4> destinationOffsets;
+  llvm::SmallVector<int64_t, 4> sizes;
+};
+
 namespace detail {
 
 inline mlir::ValueRange getBlockArgs(mlir::Block* block) { return mlir::ValueRange(block->getArguments()); }
@@ -407,4 +414,12 @@ mlir::Value materializeOrComputeUnary(mlir::Value input,
 
 mlir::Value sumTensors(mlir::ArrayRef<mlir::Value> tensors, mlir::PatternRewriter& rewriter);
 
+mlir::FailureOr<mlir::Value> createFragmentAssemblyBlueprint(mlir::Value physicalBatch,
+                                                             mlir::RankedTensorType logicalType,
+                                                             llvm::ArrayRef<FragmentAssemblyEntry> entries,
+                                                             llvm::StringRef physicalLayout,
+                                                             llvm::StringRef indexMap,
+                                                             mlir::PatternRewriter& rewriter,
+                                                             mlir::Location loc);
+
 } // namespace onnx_mlir

src/PIM/Conversion/ONNXToSpatial/Common/RowStripLayoutUtils.cpp

@@ -4,6 +4,7 @@
 #include "src/Accelerators/PIM/Common/IR/ConstantUtils.hpp"
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/BiasAddUtils.hpp"
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/Common.hpp"
+#include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/ComputeRegionBuilder.hpp"
 #include "src/Accelerators/PIM/Conversion/ONNXToSpatial/Common/RowStripLayoutUtils.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include "src/Dialect/ONNX/ONNXOps.hpp"
@@ -12,6 +13,22 @@ using namespace mlir;
 
 namespace onnx_mlir {
 
+FailureOr<RowStripPhysicalValue> describeRowStripPhysicalValue(Value storage, RankedTensorType logicalType) {
+  auto storageType = dyn_cast<RankedTensorType>(storage.getType());
+  if (!storageType || !storageType.hasStaticShape() || !logicalType || !logicalType.hasStaticShape()
+      || storageType.getRank() != 5 || logicalType.getRank() != 4 || logicalType.getDimSize(0) != 1
+      || storageType.getElementType() != logicalType.getElementType()
+      || storageType.getDimSize(1) != 1 || storageType.getDimSize(2) <= 0
+      || storageType.getDimSize(3) != 1 || storageType.getDimSize(4) != logicalType.getDimSize(3))
+    return failure();
+  const int64_t tilesPerRow = ceilIntegerDivide(logicalType.getDimSize(1), storageType.getDimSize(2));
+  if (storageType.getDimSize(0) != logicalType.getDimSize(2) * tilesPerRow)
+    return failure();
+  return RowStripPhysicalValue {storage, logicalType,
+    RankedTensorType::get(storageType.getShape().drop_front(), storageType.getElementType(), storageType.getEncoding()),
+    tilesPerRow};
+}
+
 RankedTensorType getRowStripFragmentType(RankedTensorType logicalType) {
   return RankedTensorType::get({logicalType.getDimSize(0), logicalType.getDimSize(1), 1, logicalType.getDimSize(3)},
                                logicalType.getElementType(),
@@ -123,42 +140,39 @@ FailureOr<Value> createRowStripStorageFromRows(Value rows,
   return batchOp->getResult(0);
 }
 
-FailureOr<Value>
-createRowStripAssemblyBlueprint(Value storage, RankedTensorType logicalType, PatternRewriter& rewriter, Location loc) {
-  auto storageType = dyn_cast<RankedTensorType>(storage.getType());
-  if (!storageType || storageType != getRowStripStorageType(logicalType))
-    return failure();
-
-  auto [offsets, sizes] = buildRowStripMetadata(logicalType);
-  int64_t height = logicalType.getDimSize(2);
-  SmallVector<int64_t> operandIndices(height, 0), sourceSlots, sourceOffsets(height, 0), strides(height * 4, 1);
-  for (int64_t row = 0; row < height; ++row)
-    sourceSlots.push_back(row);
-  return spatial::SpatBlueprintOp::create(rewriter, loc, logicalType, storage, ValueRange {},
-    rewriter.getStringAttr("nchw"), rewriter.getStringAttr("nchw_row_strip"),
-    rewriter.getDenseI64ArrayAttr(offsets), rewriter.getDenseI64ArrayAttr(sizes),
-    rewriter.getStringAttr("nchw_row_strip_fragments"), rewriter.getStringAttr("fragment_assembly"),
-    rewriter.getDenseI64ArrayAttr(operandIndices), rewriter.getDenseI64ArrayAttr(sourceSlots),
-    rewriter.getDenseI64ArrayAttr(sourceOffsets), rewriter.getDenseI64ArrayAttr(strides),
-    rewriter.getStringAttr("disjoint"), rewriter.getStringAttr("complete")).getOutput();
+FailureOr<Value> createRowStripAssemblyBlueprint(const RowStripPhysicalValue& value,
+                                                 PatternRewriter& rewriter,
+                                                 Location loc) {
+  SmallVector<FragmentAssemblyEntry> entries;
+  const int64_t tileChannels = value.fragmentType.getDimSize(1);
+  for (int64_t row = 0; row < value.logicalType.getDimSize(2); ++row)
+    for (int64_t tile = 0; tile < value.tilesPerRow; ++tile) {
+      const int64_t channelOffset = tile * tileChannels;
+      entries.push_back({row * value.tilesPerRow + tile, 0, {0, channelOffset, row, 0},
+        {1, std::min(tileChannels, value.logicalType.getDimSize(1) - channelOffset), 1,
+         value.logicalType.getDimSize(3)}});
+    }
+  return createFragmentAssemblyBlueprint(value.storage, value.logicalType, entries, "nchw_row_strip",
+                                         kRowStripIndexMap, rewriter, loc);
 }
 
-FailureOr<Value>
-applyRowStripRelu(Value storage, RankedTensorType logicalType, PatternRewriter& rewriter, Location loc) {
-  auto fragmentType = getRowStripFragmentType(logicalType);
-  auto storageType = getRowStripStorageType(logicalType);
+FailureOr<Value> applyRowStripRelu(const RowStripPhysicalValue& value, PatternRewriter& rewriter, Location loc) {
+  auto storageType = cast<RankedTensorType>(value.storage.getType());
+  const int64_t laneCount = storageType.getDimSize(0);
   auto batchOp = createSpatComputeBatch(rewriter,
                                         loc,
                                         TypeRange {storageType},
-                                        logicalType.getDimSize(2),
+                                        laneCount,
                                         {},
-                                        ValueRange {storage},
+                                        ValueRange {value.storage},
                                         [&](detail::SpatComputeBatchBodyArgs args) {
-                                          Value fragment =
-                                            extractRowStripFragment(args.inputs.front(), logicalType, args.lane, rewriter, loc);
-                                          fragment = spatial::SpatReluOp::create(rewriter, loc, fragmentType, fragment).getResult();
-                                          insertRowStripFragment(
-                                            fragment, args.outputs.front(), logicalType, args.lane, rewriter, loc);
+                                          FailureOr<Value> fragment = extractGraphBatchPhysicalFragment(
+                                            rewriter, loc, args.inputs.front(), args.lane, value.fragmentType);
+                                          if (failed(fragment)) return failure();
+                                          Value relu = spatial::SpatReluOp::create(
+                                            rewriter, loc, value.fragmentType, *fragment).getResult();
+                                          publishGraphBatchPhysicalFragment(
+                                            rewriter, loc, relu, args.outputs.front(), args.lane);
                                           return success();
                                         });
   if (failed(batchOp))
@@ -166,41 +180,47 @@ applyRowStripRelu(Value storage, RankedTensorType logicalType, PatternRewriter&
   return batchOp->getResult(0);
 }
 
-FailureOr<Value>
-applyRowStripBiasAdd(Value storage, RankedTensorType logicalType, Value bias, PatternRewriter& rewriter, Location loc) {
+FailureOr<Value> applyRowStripBiasAdd(const RowStripPhysicalValue& value,
+                                     Value bias,
+                                     PatternRewriter& rewriter,
+                                     Location loc) {
   DenseElementsAttr denseAttr;
-  if (!isSupportedBiasAddValue(bias, logicalType, &denseAttr))
+  if (!isSupportedBiasAddValue(bias, value.logicalType, &denseAttr))
     return failure();
-  auto fragmentType = getRowStripFragmentType(logicalType);
-  auto storageType = getRowStripStorageType(logicalType);
+  FailureOr<SmallVector<Attribute>> channelValues = getBiasChannelValues(denseAttr, value.logicalType);
+  if (failed(channelValues)) return failure();
+  auto storageType = cast<RankedTensorType>(value.storage.getType());
+  auto biasStorageType = spatial::getGraphBatchPhysicalResultType(value.tilesPerRow, value.fragmentType);
+  SmallVector<Attribute> biasValues(
+    biasStorageType.getNumElements(), cast<Attribute>(rewriter.getZeroAttr(value.fragmentType.getElementType())));
+  const int64_t tileChannels = value.fragmentType.getDimSize(1);
+  const int64_t width = value.fragmentType.getDimSize(3);
+  for (int64_t channel = 0; channel < value.logicalType.getDimSize(1); ++channel)
+    for (int64_t w = 0; w < width; ++w)
+      biasValues[((channel / tileChannels) * tileChannels + channel % tileChannels) * width + w] =
+        (*channelValues)[channel];
+  Value biasStorage = getOrCreateConstant(rewriter, rewriter.getInsertionBlock()->getParentOp(),
+    DenseElementsAttr::get(biasStorageType, biasValues), biasStorageType);
+  const int64_t laneCount = storageType.getDimSize(0);
   auto batchOp = createSpatComputeBatch(rewriter,
                                         loc,
                                         TypeRange {storageType},
-                                        logicalType.getDimSize(2),
+                                        laneCount,
                                         {},
-                                        ValueRange {storage},
+                                        ValueRange {value.storage, biasStorage},
                                         [&](detail::SpatComputeBatchBodyArgs args) {
-                                          Value fragment =
-                                            extractRowStripFragment(args.inputs.front(), logicalType, args.lane, rewriter, loc);
-                                          Value constant;
-                                          if (denseAttr.isSplat()) {
-                                            constant = getOrCreateConstant(
-                                              rewriter,
-                                              rewriter.getInsertionBlock()->getParentOp(),
-                                              DenseElementsAttr::get(fragmentType, denseAttr.getSplatValue<Attribute>()),
-                                              fragmentType);
-                                          }
-                                          else {
-                                            FailureOr<Value> perChannel =
-                                              createPerChannelConstantFragment(denseAttr, fragmentType, rewriter);
-                                            if (failed(perChannel))
-                                              return failure();
-                                            constant = *perChannel;
-                                          }
-                                          fragment =
-                                            spatial::SpatVAddOp::create(rewriter, loc, fragmentType, fragment, constant).getResult();
-                                          insertRowStripFragment(
-                                            fragment, args.outputs.front(), logicalType, args.lane, rewriter, loc);
+                                          Operation* anchorOp = rewriter.getInsertionBlock()->getParentOp();
+                                          FailureOr<Value> fragment = extractGraphBatchPhysicalFragment(
+                                            rewriter, loc, args.inputs[0], args.lane, value.fragmentType);
+                                          Value tile = affineModConst(rewriter, loc, args.lane,
+                                            value.tilesPerRow, anchorOp);
+                                          FailureOr<Value> constant = extractGraphBatchPhysicalFragment(
+                                            rewriter, loc, args.inputs[1], tile, value.fragmentType);
+                                          if (failed(fragment) || failed(constant)) return failure();
+                                          Value added = spatial::SpatVAddOp::create(
+                                            rewriter, loc, value.fragmentType, *fragment, *constant).getResult();
+                                          publishGraphBatchPhysicalFragment(
+                                            rewriter, loc, added, args.outputs.front(), args.lane);
                                           return success();
                                         });
   if (failed(batchOp))

src/PIM/Conversion/ONNXToSpatial/Common/RowStripLayoutUtils.hpp

@@ -11,10 +11,13 @@ inline constexpr llvm::StringLiteral kRowStripIndexMap = "nchw_row_strip_fragmen
 struct RowStripPhysicalValue {
   mlir::Value storage;
   mlir::RankedTensorType logicalType;
-  llvm::SmallVector<int64_t, 16> fragmentOffsets;
-  llvm::SmallVector<int64_t, 16> fragmentSizes;
+  mlir::RankedTensorType fragmentType;
+  int64_t tilesPerRow;
 };
 
+mlir::FailureOr<RowStripPhysicalValue> describeRowStripPhysicalValue(mlir::Value storage,
+                                                                     mlir::RankedTensorType logicalType);
+
 std::pair<llvm::SmallVector<int64_t>, llvm::SmallVector<int64_t>>
 buildRowStripMetadata(mlir::RankedTensorType type);
 
@@ -50,18 +53,15 @@ mlir::FailureOr<mlir::Value> createRowStripStorageFromRows(mlir::Value rows,
                                                            mlir::PatternRewriter& rewriter,
                                                            mlir::Location loc);
 
-mlir::FailureOr<mlir::Value> createRowStripAssemblyBlueprint(mlir::Value storage,
-                                                               mlir::RankedTensorType logicalType,
+mlir::FailureOr<mlir::Value> createRowStripAssemblyBlueprint(const RowStripPhysicalValue& value,
                                                                mlir::PatternRewriter& rewriter,
                                                                mlir::Location loc);
 
-mlir::FailureOr<mlir::Value> applyRowStripRelu(mlir::Value storage,
-                                               mlir::RankedTensorType logicalType,
+mlir::FailureOr<mlir::Value> applyRowStripRelu(const RowStripPhysicalValue& value,
                                                mlir::PatternRewriter& rewriter,
                                                mlir::Location loc);
 
-mlir::FailureOr<mlir::Value> applyRowStripBiasAdd(mlir::Value storage,
-                                                  mlir::RankedTensorType logicalType,
+mlir::FailureOr<mlir::Value> applyRowStripBiasAdd(const RowStripPhysicalValue& value,
                                                   mlir::Value bias,
                                                   mlir::PatternRewriter& rewriter,
                                                   mlir::Location loc);

src/PIM/Conversion/ONNXToSpatial/LowerSpatialPlansPass.cpp

@@ -45,38 +45,30 @@ static FailureOr<RowStripPhysicalValue> buildRowStripValue(spatial::SpatBlueprin
   auto logicalType = dyn_cast<RankedTensorType>(blueprint.getOutput().getType());
   if (!logicalType)
     return blueprint.emitOpError("requires ranked logical output type"), failure();
-  RowStripPhysicalValue value;
-  value.storage = storage;
-  value.logicalType = logicalType;
-  value.fragmentOffsets.append(blueprint.getFragmentOffsets().begin(), blueprint.getFragmentOffsets().end());
-  value.fragmentSizes.append(blueprint.getFragmentSizes().begin(), blueprint.getFragmentSizes().end());
   if (blueprint.getIndexMap() != kRowStripIndexMap)
     return blueprint.emitOpError("requires the canonical row-strip index map"), failure();
-  auto storageType = dyn_cast<RankedTensorType>(storage.getType());
-  if (!storageType || storageType != getRowStripStorageType(logicalType))
+  FailureOr<RowStripPhysicalValue> value = describeRowStripPhysicalValue(storage, logicalType);
+  if (failed(value))
     return blueprint.emitOpError("requires physical row-strip fragment storage"), failure();
-  return value;
+  return *value;
 }
 
 static FailureOr<Value>
 lowerRowStripRelu(const RowStripPhysicalValue& input, spatial::SpatReluPlanOp planOp, PatternRewriter& rewriter) {
-  return applyRowStripRelu(input.storage, input.logicalType, rewriter, planOp.getLoc());
+  return applyRowStripRelu(input, rewriter, planOp.getLoc());
 }
 
 static FailureOr<Value> lowerRowStripBiasAdd(const RowStripPhysicalValue& input,
                                              spatial::SpatBiasAddPlanOp planOp,
                                              PatternRewriter& rewriter) {
-  return applyRowStripBiasAdd(input.storage, input.logicalType, planOp.getBias(), rewriter, planOp.getLoc());
+  return applyRowStripBiasAdd(input, planOp.getBias(), rewriter, planOp.getLoc());
 }
 
 static FailureOr<Value>
 materializeRowStripToDense(const RowStripPhysicalValue& rowStripValue, Location loc, PatternRewriter& rewriter) {
   if (rowStripValue.logicalType.getRank() != 4 || !rowStripValue.logicalType.hasStaticShape())
     return failure();
-  auto [expectedOffsets, expectedSizes] = buildRowStripMetadata(rowStripValue.logicalType);
-  if (!llvm::equal(rowStripValue.fragmentOffsets, expectedOffsets) || !llvm::equal(rowStripValue.fragmentSizes, expectedSizes))
-    return failure();
-  return createRowStripAssemblyBlueprint(rowStripValue.storage, rowStripValue.logicalType, rewriter, loc);
+  return createRowStripAssemblyBlueprint(rowStripValue, rewriter, loc);
 }
 
 static FailureOr<Value> lowerDenseBatchBiasAdd(Value input, Value bias, RankedTensorType resultType,
@@ -168,9 +160,12 @@ struct LowerSpatialPlansPass final : PassWrapper<LowerSpatialPlansPass, Operatio
         });
         if (rowStripBlueprint != planOp.getResult().getUsers().end()) {
           rewriter.setInsertionPoint(planOp);
+          std::optional<Value> physicalInput;
+          if (succeeded(rowStripInput))
+            physicalInput = rowStripInput->storage;
           FailureOr<Value> lowered = lowerSelectedConv2DPlan(
             planOp,
-            succeeded(rowStripInput) ? std::optional<Value> {rowStripInput->storage} : std::nullopt,
+            physicalInput,
             /*emitRowStripLayout=*/true,
             rewriter);
           if (failed(lowered)) {
@@ -255,8 +250,23 @@ struct LowerSpatialPlansPass final : PassWrapper<LowerSpatialPlansPass, Operatio
 
         FailureOr<RowStripPhysicalValue> input = getRowStripValue(rowStripValues, planOp.getInput());
         rewriter.setInsertionPoint(planOp);
+        std::optional<Value> physicalInput;
+        if (succeeded(input)) {
+          if (input->tilesPerRow == 1) {
+            physicalInput = input->storage;
+          }
+          else {
+            FailureOr<Value> denseInput = materializeRowStripToDense(*input, planOp.getLoc(), rewriter);
+            if (failed(denseInput)) {
+              planOp.emitOpError("failed to materialize tiled row-strip input for MaxPool");
+              signalPassFailure();
+              return;
+            }
+            planOp.getInputMutable().assign(*denseInput);
+          }
+        }
         FailureOr<Value> lowered = lowerSelectedMaxPool2DPlan(
-          planOp, succeeded(input) ? std::optional<Value> {input->storage} : std::nullopt, rewriter);
+          planOp, physicalInput, rewriter);
         if (failed(lowered)) {
           planOp.emitOpError("failed to lower selected row-strip Spatial MaxPool plan");
           signalPassFailure();

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

@@ -2326,54 +2326,25 @@ static Value maybeUnpackChunkRows(Value gemmRows,
   return unpackCompute.getResult(0);
 }
 
-static Value createChunkedConvRows(const ConvLoweringState& state,
-                                   const PreparedConvInput& preparedInput,
-                                   Value weightMatrix,
-                                   Value biasMatrix,
-                                   DenseElementsAttr wDenseAttr,
-                                   DenseElementsAttr biasDenseAttr,
-                                   int64_t forcedPackFactor,
-                                   uint64_t chunkPositions,
-                                   PatternRewriter& rewriter,
-                                   Location loc) {
-  SmallVector<Value> chunkRows;
+static Value createStreamedConvRows(const ConvLoweringState& state,
+                                    const PreparedConvInput& preparedInput,
+                                    Value weightMatrix,
+                                    Value biasMatrix,
+                                    DenseElementsAttr wDenseAttr,
+                                    DenseElementsAttr biasDenseAttr,
+                                    int64_t forcedPackFactor,
+                                    PatternRewriter& rewriter,
+                                    Location loc) {
   const int64_t totalPatches = state.batchSize * state.outHeight * state.outWidth;
-  for (int64_t chunkStart = 0; chunkStart < totalPatches; chunkStart += static_cast<int64_t>(chunkPositions)) {
-    const int64_t chunkNumPatches = std::min<int64_t>(static_cast<int64_t>(chunkPositions), totalPatches - chunkStart);
-    ConvGemmPlan chunkPlan = buildConvGemmPlan(state,
-                                               static_cast<bool>(wDenseAttr),
-                                               !state.hasBias || static_cast<bool>(biasDenseAttr),
-                                               chunkStart,
-                                               chunkNumPatches,
-                                               forcedPackFactor);
-    Value chunkInputRows = createIm2colRows(state, preparedInput, chunkPlan, rewriter, loc);
-    Value chunkB = buildPackedWeights(wDenseAttr, weightMatrix, state, chunkPlan, rewriter, loc);
-    Value gemmBias = createZeroGemmBias(chunkPlan.gemmOutputRowsType, rewriter);
-    if (state.hasBias)
-      gemmBias = state.b;
-    Value chunkC = buildPackedBias(gemmBias, biasMatrix, biasDenseAttr, state, chunkPlan, rewriter, loc);
-    Value chunkGemmRows = ONNXGemmOp::create(rewriter,
-                                             loc,
-                                             chunkPlan.gemmOutputRowsType,
-                                             chunkInputRows,
-                                             chunkB,
-                                             chunkC,
-                                             APFloat(1.0f),
-                                             APFloat(1.0f),
-                                             /*transA=*/0,
-                                             /*transB=*/0)
-                           .getY();
-    chunkRows.push_back(maybeUnpackChunkRows(chunkGemmRows, chunkPlan, rewriter, loc));
-  }
-
-  if (chunkRows.size() == 1)
-    return chunkRows.front();
-
-  auto rowType = RankedTensorType::get({totalPatches, state.numChannelsOut}, state.outType.getElementType());
-  auto collectRows = createSpatCompute(rewriter, loc, TypeRange {rowType}, {}, chunkRows, [&](ValueRange rows) {
-    spatial::SpatYieldOp::create(rewriter, loc, createSpatConcat(rewriter, loc, /*axis=*/0, rows));
-  });
-  return collectRows.getResult(0);
+  ConvGemmPlan plan = buildConvGemmPlan(state, static_cast<bool>(wDenseAttr),
+    !state.hasBias || static_cast<bool>(biasDenseAttr), 0, totalPatches, forcedPackFactor);
+  Value inputRows = createIm2colRows(state, preparedInput, plan, rewriter, loc);
+  Value packedWeights = buildPackedWeights(wDenseAttr, weightMatrix, state, plan, rewriter, loc);
+  Value gemmBias = state.hasBias ? state.b : createZeroGemmBias(plan.gemmOutputRowsType, rewriter);
+  Value packedBias = buildPackedBias(gemmBias, biasMatrix, biasDenseAttr, state, plan, rewriter, loc);
+  Value gemmRows = ONNXGemmOp::create(rewriter, loc, plan.gemmOutputRowsType, inputRows,
+    packedWeights, packedBias, APFloat(1.0f), APFloat(1.0f), 0, 0).getY();
+  return maybeUnpackChunkRows(gemmRows, plan, rewriter, loc);
 }
 
 static Value rewritePackedIm2ColConv(const ConvLoweringState& state,
@@ -2444,16 +2415,13 @@ static Value rewriteStreamedConv(const ConvLoweringState& state,
   ConvGemmPlan seedPlan = buildConvGemmPlan(
     state, static_cast<bool>(wDenseAttr), !state.hasBias || static_cast<bool>(biasDenseAttr), 0, 1, forcedPackFactor);
   Value weightMatrix = createWeightMatrix(state.w, seedPlan, rewriter, loc);
-  ConvGeometry geo = buildConvGeometry(state);
-  uint64_t chunkPositions = chooseStreamChunkPositions(geo, forcedPackFactor);
-  Value collectedRows = createChunkedConvRows(state,
+  Value collectedRows = createStreamedConvRows(state,
                                               preparedInput,
                                               weightMatrix,
                                               biasMatrix,
                                               wDenseAttr,
                                               biasDenseAttr,
                                               forcedPackFactor,
-                                              chunkPositions,
                                               rewriter,
                                               loc);
   auto gemmOutType = cast<RankedTensorType>(collectedRows.getType());
@@ -2524,6 +2492,21 @@ static bool canConsumeNchwRowStripFragments(const ConvLoweringState& state, Stri
     failureReason = "dilation_not_one";
     return false;
   }
+  ConvGeometry geometry = buildConvGeometry(state);
+  const bool pointwise = state.xHeight == 1 && state.xWidth == 1 && state.outHeight == 1 && state.outWidth == 1
+                         && state.wHeight == 1 && state.wWidth == 1 && state.padHeightBegin == 0
+                         && state.padHeightEnd == 0 && state.padWidthBegin == 0 && state.padWidthEnd == 0;
+  if (pointwise) {
+    if (!getHostConstDenseElementsAttr(state.w)) {
+      failureReason = "non_constant_weight";
+      return false;
+    }
+    if (state.hasBias && !isSupportedBiasAddValue(state.b, state.outType)) {
+      failureReason = "unsupported_bias";
+      return false;
+    }
+    return true;
+  }
   if (state.wHeight != 3 || state.wWidth != 3) {
     failureReason = "kernel_not_3x3";
     return false;
@@ -2544,7 +2527,6 @@ static bool canConsumeNchwRowStripFragments(const ConvLoweringState& state, Stri
     failureReason = "unsupported_bias";
     return false;
   }
-  ConvGeometry geometry = buildConvGeometry(state);
   if (geometry.c > geometry.xbarSize) {
     failureReason = "output_channels_exceed_crossbar";
     return false;
@@ -2575,6 +2557,25 @@ static FailureOr<Value> createPaddedBiasRowConstant(const ConvLoweringState& sta
   return getOrCreateConstant(rewriter, rewriter.getInsertionBlock()->getParentOp(), biasAttr, biasType);
 }
 
+static FailureOr<Value> createPaddedBiasTileConstant(const ConvLoweringState& state,
+                                                     int64_t tileChannels,
+                                                     PatternRewriter& rewriter) {
+  DenseElementsAttr denseAttr;
+  if (!isSupportedBiasAddValue(state.b, state.outType, &denseAttr))
+    return failure();
+  FailureOr<SmallVector<Attribute>> channelValues = getBiasChannelValues(denseAttr, state.outType);
+  if (failed(channelValues))
+    return failure();
+  const int64_t tileCount = ceilIntegerDivide(state.numChannelsOut, tileChannels);
+  auto tileType = RankedTensorType::get({tileCount, 1, tileChannels}, state.outType.getElementType());
+  SmallVector<Attribute> values(
+    tileType.getNumElements(), cast<Attribute>(rewriter.getZeroAttr(tileType.getElementType())));
+  for (int64_t channel = 0; channel < state.numChannelsOut; ++channel)
+    values[channel] = (*channelValues)[channel];
+  return getOrCreateConstant(rewriter, rewriter.getInsertionBlock()->getParentOp(),
+    DenseElementsAttr::get(tileType, values), tileType);
+}
+
 static Value createHorizontallyPaddedRowStripFragment(Value fragment,
                                                       const ConvLoweringState& state,
                                                       PatternRewriter& rewriter,
@@ -2732,8 +2733,11 @@ static FailureOr<Value> createConvInputWindow(Value input,
                         ? extractDenseConvWindowRow(input, sourceRowTable, state, outputHeight, kernelRow, rewriter, loc)
                         : extractProjectedRowStripWindowRow(
                             input, sourceRowTable, state, outputHeight, kernelRow, rewriter, loc);
-    Value mask = extractProjectedRowStripWindowMask(*maskTable, state, outputHeight, kernelRow, rewriter, loc);
-    Value semanticRow = spatial::SpatVMulOp::create(rewriter, loc, fragmentType, sourceRow, mask).getResult();
+    Value semanticRow = sourceRow;
+    if (state.padHeightBegin != 0 || state.padHeightEnd != 0) {
+      Value mask = extractProjectedRowStripWindowMask(*maskTable, state, outputHeight, kernelRow, rewriter, loc);
+      semanticRow = spatial::SpatVMulOp::create(rewriter, loc, fragmentType, sourceRow, mask).getResult();
+    }
     Value paddedRow = createHorizontallyPaddedRowStripFragment(semanticRow, state, rewriter, loc);
     window = tensor::InsertSliceOp::create(rewriter,
                                            loc,
@@ -2906,12 +2910,6 @@ static FailureOr<Value> createPaddedConvOutputRow(Value patchRow,
     .getResult();
 }
 
-static bool rowStripOutputFitsOneCore(const ConvGeometry& geometry) {
-  const int64_t inputTileCount = ceilIntegerDivide(geometry.k, geometry.xbarSize);
-  const int64_t outputTileCount = ceilIntegerDivide(geometry.c, geometry.xbarSize);
-  return inputTileCount * outputTileCount <= static_cast<int64_t>(crossbarCountInCore.getValue());
-}
-
 static bool rowStripOutputTileFitsOneCore(const ConvGeometry& geometry) {
   return ceilIntegerDivide(geometry.k, geometry.xbarSize)
          <= static_cast<int64_t>(crossbarCountInCore.getValue());
@@ -2928,38 +2926,40 @@ static FailureOr<Value> createOutputChannelTiledRowStripConvOutput(const ConvLow
   const int64_t patchSize = state.numChannelsIn * state.wHeight * state.wWidth;
   auto elementType = state.outType.getElementType();
   auto paddedPatchRowType = RankedTensorType::get({1, paddedK}, elementType);
+  auto paddedRowType = RankedTensorType::get({1, xbarDim}, elementType);
+  auto tilePixelType = RankedTensorType::get({1, xbarDim, 1, 1}, elementType);
+  auto tileFragmentType = RankedTensorType::get({1, xbarDim, 1, state.outWidth}, elementType);
   auto tileWeightsType = RankedTensorType::get({paddedK, xbarDim}, state.wType.getElementType());
-  SmallVector<Value> outputTiles;
-  outputTiles.reserve(outputTileCount);
-
-  for (int64_t outputTile = 0; outputTile < outputTileCount; ++outputTile) {
-    const int64_t channelOffset = outputTile * xbarDim;
-    const int64_t tileChannels = std::min(xbarDim, state.numChannelsOut - channelOffset);
-    auto tileRowType = RankedTensorType::get({1, tileChannels}, elementType);
-    auto tilePixelType = RankedTensorType::get({1, tileChannels, 1, 1}, elementType);
-    auto tileFragmentType = RankedTensorType::get({1, tileChannels, 1, state.outWidth}, elementType);
-    auto tileStorageType = spatial::getGraphBatchPhysicalResultType(state.outHeight, tileFragmentType);
-    SmallVector<OpFoldResult> weightOffsets {
-      rewriter.getIndexAttr(outputTile), rewriter.getIndexAttr(0), rewriter.getIndexAttr(0)};
-    SmallVector<OpFoldResult> weightSizes {
-      rewriter.getIndexAttr(1), rewriter.getIndexAttr(paddedK), rewriter.getIndexAttr(xbarDim)};
-    Value tileWeights = extractStaticSliceOrIdentity(
-      rewriter, loc, paddedWeights, tileWeightsType, weightOffsets, weightSizes, getUnitStrides(rewriter, 3));
-
-    auto tileBatch = createSpatComputeBatch(
-      rewriter,
-      loc,
-      TypeRange {tileStorageType},
-      state.outHeight,
-      ValueRange {tileWeights},
-      ValueRange {state.x},
-      [&](detail::SpatComputeBatchBodyArgs args) {
+  const int64_t laneCount = state.outHeight * outputTileCount;
+  auto tileStorageType = spatial::getGraphBatchPhysicalResultType(laneCount, tileFragmentType);
+  FailureOr<Value> paddedBias = failure();
+  if (state.hasBias)
+    paddedBias = createPaddedBiasTileConstant(state, xbarDim, rewriter);
+  if (state.hasBias && failed(paddedBias))
+    return failure();
+  auto tileBatch = createSpatComputeBatch(
+    rewriter, loc, TypeRange {tileStorageType}, laneCount, ValueRange {paddedWeights},
+    state.hasBias ? ValueRange {state.x, *paddedBias} : ValueRange {state.x},
+    [&](detail::SpatComputeBatchBodyArgs args) {
         Operation* anchorOp = rewriter.getInsertionBlock()->getParentOp();
         Value c0 = getOrCreateIndexConstant(rewriter, anchorOp, 0);
         Value c1 = getOrCreateIndexConstant(rewriter, anchorOp, 1);
         Value cOutWidth = getOrCreateIndexConstant(rewriter, anchorOp, state.outWidth);
+        Value outputRow = affineFloorDivConst(rewriter, loc, args.lane, outputTileCount, anchorOp);
+        Value outputTile = affineModConst(rewriter, loc, args.lane, outputTileCount, anchorOp);
+        SmallVector<OpFoldResult> weightOffsets {
+          outputTile, rewriter.getIndexAttr(0), rewriter.getIndexAttr(0)};
+        SmallVector<OpFoldResult> weightSizes {
+          rewriter.getIndexAttr(1), rewriter.getIndexAttr(paddedK), rewriter.getIndexAttr(xbarDim)};
+        Value tileWeights = tensor::ExtractSliceOp::create(
+          rewriter, loc, tileWeightsType, args.weights.front(), weightOffsets, weightSizes, getUnitStrides(rewriter, 3));
+        FailureOr<Value> biasTile = failure();
+        if (state.hasBias)
+          biasTile = extractGraphBatchPhysicalFragment(rewriter, loc, args.inputs[1], outputTile, paddedRowType);
+        if (state.hasBias && failed(biasTile))
+          return failure();
         FailureOr<Value> inputWindow =
-          createConvInputWindow(args.inputs.front(), state, args.lane, rewriter, loc);
+          createConvInputWindow(args.inputs.front(), state, outputRow, rewriter, loc);
         if (failed(inputWindow))
           return failure();
         Value fragmentInit = tensor::EmptyOp::create(rewriter, loc, tileFragmentType.getShape(), elementType);
@@ -2984,28 +2984,18 @@ static FailureOr<Value> createOutputChannelTiledRowStripConvOutput(const ConvLow
               paddedPatchRow = createZeroPaddedTensor(
                 paddedPatchRow, paddedPatchRowType, {0, 0}, {0, paddedK - patchSize}, rewriter, widthLoc);
             FailureOr<Value> paddedOutputRow = createPaddedConvOutputTile(
-              paddedPatchRow, args.weights.front(), numKSlices, xbarDim, rewriter, widthLoc);
+              paddedPatchRow, tileWeights, numKSlices, xbarDim, rewriter, widthLoc);
             if (failed(paddedOutputRow))
               return failure();
-            Value outputRow = *paddedOutputRow;
-            if (tileChannels != xbarDim) {
-              SmallVector<OpFoldResult> rowOffsets {rewriter.getIndexAttr(0), rewriter.getIndexAttr(0)};
-              SmallVector<OpFoldResult> rowSizes {
-                rewriter.getIndexAttr(1), rewriter.getIndexAttr(tileChannels)};
-              outputRow = tensor::ExtractSliceOp::create(rewriter,
-                                                         widthLoc,
-                                                         tileRowType,
-                                                         outputRow,
-                                                         rowOffsets,
-                                                         rowSizes,
-                                                         getUnitStrides(rewriter, 2));
-            }
+            if (state.hasBias)
+              paddedOutputRow = spatial::SpatVAddOp::create(
+                rewriter, widthLoc, paddedRowType, *paddedOutputRow, *biasTile).getResult();
             Value outputPixel = tensor::ExpandShapeOp::create(
-              rewriter, widthLoc, tilePixelType, outputRow, SmallVector<ReassociationIndices> {{0}, {1, 2, 3}});
+              rewriter, widthLoc, tilePixelType, *paddedOutputRow, SmallVector<ReassociationIndices> {{0}, {1, 2, 3}});
             SmallVector<OpFoldResult> rowOffsets {
               rewriter.getIndexAttr(0), rewriter.getIndexAttr(0), rewriter.getIndexAttr(0), widthIndex};
             SmallVector<OpFoldResult> rowSizes {rewriter.getIndexAttr(1),
-                                               rewriter.getIndexAttr(tileChannels),
+                                               rewriter.getIndexAttr(xbarDim),
                                                rewriter.getIndexAttr(1),
                                                rewriter.getIndexAttr(1)};
             Value nextFragment = tensor::InsertSliceOp::create(rewriter,
@@ -3024,58 +3014,9 @@ static FailureOr<Value> createOutputChannelTiledRowStripConvOutput(const ConvLow
           rewriter, loc, widthLoop->results.front(), args.outputs.front(), args.lane);
         return success();
       });
-    if (failed(tileBatch))
-      return failure();
-    outputTiles.push_back(tileBatch->getResult(0));
-  }
-
-  auto fragmentType = getRowStripFragmentType(state.outType);
-  auto outputStorageType = getRowStripStorageType(state.outType);
-  auto assemblyBatch = createSpatComputeBatch(rewriter,
-                                              loc,
-                                              TypeRange {outputStorageType},
-                                              state.outHeight,
-                                              {},
-                                              ValueRange(outputTiles),
-                                              [&](detail::SpatComputeBatchBodyArgs args) {
-                                                Value fragment = tensor::EmptyOp::create(
-                                                  rewriter, loc, fragmentType.getShape(), elementType);
-                                                for (int64_t outputTile = 0; outputTile < outputTileCount; ++outputTile) {
-                                                  const int64_t channelOffset = outputTile * xbarDim;
-                                                  const int64_t tileChannels =
-                                                    std::min(xbarDim, state.numChannelsOut - channelOffset);
-                                                  auto tileFragmentType = RankedTensorType::get(
-                                                    {1, tileChannels, 1, state.outWidth}, elementType);
-                                                  FailureOr<Value> tileFragment = extractGraphBatchPhysicalFragment(
-                                                    rewriter, loc, args.inputs[outputTile], args.lane, tileFragmentType);
-                                                  if (failed(tileFragment))
-                                                    return failure();
-                                                  SmallVector<OpFoldResult> offsets {rewriter.getIndexAttr(0),
-                                                                                   rewriter.getIndexAttr(channelOffset),
-                                                                                   rewriter.getIndexAttr(0),
-                                                                                   rewriter.getIndexAttr(0)};
-                                                  SmallVector<OpFoldResult> sizes {rewriter.getIndexAttr(1),
-                                                                                 rewriter.getIndexAttr(tileChannels),
-                                                                                 rewriter.getIndexAttr(1),
-                                                                                 rewriter.getIndexAttr(state.outWidth)};
-                                                  fragment = tensor::InsertSliceOp::create(rewriter,
-                                                                                          loc,
-                                                                                          *tileFragment,
-                                                                                          fragment,
-                                                                                          offsets,
-                                                                                          sizes,
-                                                                                          getUnitStrides(rewriter, 4));
-                                                }
-                                                insertRowStripFragment(
-                                                  fragment, args.outputs.front(), state.outType, args.lane, rewriter, loc);
-                                                return success();
-                                              });
-  if (failed(assemblyBatch))
+  if (failed(tileBatch))
     return failure();
-  Value output = assemblyBatch->getResult(0);
-  if (state.hasBias)
-    return applyRowStripBiasAdd(output, state.outType, state.b, rewriter, loc);
-  return output;
+  return tileBatch->getResult(0);
 }
 
 static FailureOr<Value>
@@ -3104,7 +3045,7 @@ createRowStripConvOutputFromDenseInput(const ConvLoweringState& state, PatternRe
                               weightDenseAttr, state, paddedK, xbarDim, rewriter)
                           : standard::createPaddedOutputChannelTiledWeightConstant(
                               weightDenseAttr, state, paddedK, xbarDim, rewriter);
-  if (!rowStripOutputFitsOneCore(geometry))
+  if (state.numChannelsOut > xbarDim)
     return createOutputChannelTiledRowStripConvOutput(
       state, paddedWeights, paddedK, numKSlices, xbarDim, rewriter, loc);
 
@@ -3279,11 +3220,106 @@ static FailureOr<Value> createConvOutputFromNchwRowStripFragments(Value rowStrip
   return batchOp->getResult(0);
 }
 
+static FailureOr<Value> createPointwiseOutputFromRowStripFragments(Value rowStripStorage,
+                                                                   const ConvLoweringState& state,
+                                                                   PatternRewriter& rewriter,
+                                                                   Location loc) {
+  FailureOr<RowStripPhysicalValue> input = describeRowStripPhysicalValue(rowStripStorage, state.xType);
+  if (failed(input)) return failure();
+  ConvGeometry geometry = buildConvGeometry(state);
+  const int64_t xbarDim = geometry.xbarSize;
+  const int64_t inputFragmentChannels = input->fragmentType.getDimSize(1);
+  if (inputFragmentChannels % xbarDim != 0 || state.numChannelsIn % xbarDim != 0)
+    return failure();
+  auto weightDenseAttr = getHostConstDenseElementsAttr(state.w);
+  if (!weightDenseAttr) return failure();
+
+  const int64_t outputTileCount = ceilIntegerDivide(state.numChannelsOut, xbarDim);
+  const int64_t numKSlices = state.numChannelsIn / xbarDim;
+  auto elementType = state.outType.getElementType();
+  auto paddedRowType = RankedTensorType::get({1, xbarDim}, elementType);
+  auto inputRowType = RankedTensorType::get({1, inputFragmentChannels}, elementType);
+  auto weightTileType = RankedTensorType::get({state.numChannelsIn, xbarDim}, state.wType.getElementType());
+  auto weightSliceType = RankedTensorType::get({xbarDim, xbarDim}, state.wType.getElementType());
+  auto outputFragmentType = RankedTensorType::get({1, xbarDim, 1, 1}, elementType);
+  auto outputStorageType = spatial::getGraphBatchPhysicalResultType(outputTileCount, outputFragmentType);
+  Value paddedWeights = standard::createPaddedOutputChannelTiledWeightConstant(
+    weightDenseAttr, state, state.numChannelsIn, xbarDim, rewriter);
+  FailureOr<Value> paddedBias = failure();
+  if (state.hasBias) paddedBias = createPaddedBiasTileConstant(state, xbarDim, rewriter);
+  if (state.hasBias && failed(paddedBias)) return failure();
+
+  auto batch = createSpatComputeBatch(rewriter, loc, TypeRange {outputStorageType}, outputTileCount,
+    ValueRange {paddedWeights},
+    state.hasBias ? ValueRange {rowStripStorage, *paddedBias} : ValueRange {rowStripStorage},
+    [&](detail::SpatComputeBatchBodyArgs args) {
+      Operation* anchorOp = rewriter.getInsertionBlock()->getParentOp();
+      Value c0 = getOrCreateIndexConstant(rewriter, anchorOp, 0);
+      Value c1 = getOrCreateIndexConstant(rewriter, anchorOp, 1);
+      Value cNumKSlices = getOrCreateIndexConstant(rewriter, anchorOp, numKSlices);
+      SmallVector<OpFoldResult> weightOffsets {args.lane, rewriter.getIndexAttr(0), rewriter.getIndexAttr(0)};
+      SmallVector<OpFoldResult> weightSizes {rewriter.getIndexAttr(1),
+        rewriter.getIndexAttr(state.numChannelsIn), rewriter.getIndexAttr(xbarDim)};
+      Value weightTile = tensor::ExtractSliceOp::create(
+        rewriter, loc, weightTileType, args.weights.front(), weightOffsets, weightSizes, getUnitStrides(rewriter, 3));
+      auto createPiece = [&](Value kSlice, Location pieceLoc) -> FailureOr<Value> {
+        Value channelOffset = affineMulConst(rewriter, pieceLoc, kSlice, xbarDim, anchorOp);
+        Value sourceSlot = affineFloorDivConst(
+          rewriter, pieceLoc, channelOffset, inputFragmentChannels, anchorOp);
+        Value sourceOffset = affineModConst(
+          rewriter, pieceLoc, channelOffset, inputFragmentChannels, anchorOp);
+        FailureOr<Value> fragment = extractGraphBatchPhysicalFragment(
+          rewriter, pieceLoc, args.inputs.front(), sourceSlot, input->fragmentType);
+        if (failed(fragment)) return failure();
+        Value inputRow = tensor::CollapseShapeOp::create(rewriter, pieceLoc, inputRowType, *fragment,
+          SmallVector<ReassociationIndices> {{0}, {1, 2, 3}});
+        Value inputSlice = tensor::ExtractSliceOp::create(rewriter, pieceLoc, paddedRowType, inputRow,
+          SmallVector<OpFoldResult> {rewriter.getIndexAttr(0), sourceOffset},
+          SmallVector<OpFoldResult> {rewriter.getIndexAttr(1), rewriter.getIndexAttr(xbarDim)},
+          getUnitStrides(rewriter, 2));
+        Value weightSlice = tensor::ExtractSliceOp::create(rewriter, pieceLoc, weightSliceType, weightTile,
+          SmallVector<OpFoldResult> {channelOffset, rewriter.getIndexAttr(0)},
+          SmallVector<OpFoldResult> {rewriter.getIndexAttr(xbarDim), rewriter.getIndexAttr(xbarDim)},
+          getUnitStrides(rewriter, 2));
+        return spatial::SpatVMMOp::create(rewriter, pieceLoc, paddedRowType, weightSlice, inputSlice).getResult();
+      };
+      FailureOr<Value> result = createPiece(c0, loc);
+      if (failed(result)) return failure();
+      if (numKSlices > 1) {
+        auto reduction = buildNormalizedScfFor(rewriter, loc, c1, cNumKSlices, c1, ValueRange {*result},
+          [&](OpBuilder&, Location reduceLoc, Value kSlice, ValueRange iterArgs,
+              SmallVectorImpl<Value>& yielded) {
+            FailureOr<Value> piece = createPiece(kSlice, reduceLoc);
+            if (failed(piece)) return failure();
+            yielded.push_back(spatial::SpatVAddOp::create(
+              rewriter, reduceLoc, paddedRowType, iterArgs.front(), *piece).getResult());
+            return success();
+          });
+        if (failed(reduction)) return failure();
+        result = reduction->results.front();
+      }
+      if (state.hasBias) {
+        FailureOr<Value> bias = extractGraphBatchPhysicalFragment(
+          rewriter, loc, args.inputs[1], args.lane, paddedRowType);
+        if (failed(bias)) return failure();
+        result = spatial::SpatVAddOp::create(rewriter, loc, paddedRowType, *result, *bias).getResult();
+      }
+      Value fragment = tensor::ExpandShapeOp::create(rewriter, loc, outputFragmentType, *result,
+        SmallVector<ReassociationIndices> {{0}, {1, 2, 3}});
+      publishGraphBatchPhysicalFragment(rewriter, loc, fragment, args.outputs.front(), args.lane);
+      return success();
+    });
+  if (failed(batch)) return failure();
+  return batch->getResult(0);
+}
+
 static FailureOr<Value> createConvOutputFromRowStripInput(const ConvLoweringState& state,
                                                           [[maybe_unused]] const ConvLoweringDecision& decision,
                                                           Value rowStripInput,
                                                           PatternRewriter& rewriter,
                                                           Location loc) {
+  if (state.xHeight == 1 && state.xWidth == 1 && state.wHeight == 1 && state.wWidth == 1)
+    return createPointwiseOutputFromRowStripFragments(rowStripInput, state, rewriter, loc);
   return createConvOutputFromNchwRowStripFragments(rowStripInput, state, rewriter, loc);
 }
 
@@ -4187,22 +4223,9 @@ lowerSelectedConv2DPlan(spatial::SpatConv2DPlanOp planOp,
     }
     if (failed(canLowerConvPlanToRowStrip(planOp)))
       return planOp.emitOpError("selected row-strip layout is not supported for this Conv plan"), failure();
-    ConvLoweringState rowState = *state;
-    const bool applyBiasAfterStorage = rowState.hasBias;
-    Value originalBias = rowState.b;
-    if (applyBiasAfterStorage) {
-      rowState.b = Value();
-      rowState.hasBias = false;
-    }
-
-    FailureOr<Value> rowStripStorage = createRowStripConvOutputFromDenseInput(rowState, rewriter, planOp.getLoc());
+    FailureOr<Value> rowStripStorage = createRowStripConvOutputFromDenseInput(*state, rewriter, planOp.getLoc());
     if (failed(rowStripStorage))
       return planOp.emitOpError("failed to build row-strip fragment storage for the selected Conv plan"), failure();
-    if (applyBiasAfterStorage) {
-      rowStripStorage = applyRowStripBiasAdd(*rowStripStorage, state->outType, originalBias, rewriter, planOp.getLoc());
-      if (failed(rowStripStorage))
-        return planOp.emitOpError("failed to apply row-strip Conv bias per fragment"), failure();
-    }
     return *rowStripStorage;
   }
 

src/PIM/Dialect/Spatial/Transforms/MergeComputeNodes/DeferredCommunicationRealization.cpp

@@ -77,9 +77,16 @@ static LogicalResult eraseOldGraph(func::FuncOp funcOp,
         rewriter.eraseOp(blueprint);
       continue;
     }
-    if (!op->use_empty())
-      return op->emitOpError(
-        "phase 2 cannot erase an old graph compute with live results");
+    if (!op->use_empty()) {
+      for (OpResult result : op->getResults()) {
+        if (!result.use_empty()) {
+          Operation *user = result.use_begin()->getOwner();
+          return op->emitOpError()
+                 << "phase 2 cannot erase old graph result "
+                 << result.getResultNumber() << " used by " << user->getName();
+        }
+      }
+    }
     rewriter.eraseOp(op);
   }
   return success();
@@ -100,6 +107,25 @@ static LogicalResult eraseDeferredSourceSelectors(
   return success();
 }
 
+static void eraseUnusedIdentityDeferredCommunications(
+    func::FuncOp funcOp, IRRewriter &rewriter) {
+  SmallVector<SpatDeferredCommunicationOp> unused;
+  funcOp.walk([&](SpatDeferredCommunicationOp deferred) {
+    if (!deferred.getOutput().use_empty() || !deferred.getBody().hasOneBlock())
+      return;
+    Block &body = deferred.getBody().front();
+    auto yield = dyn_cast<SpatYieldOp>(body.getTerminator());
+    auto argument = yield && yield.getOutputs().size() == 1
+                      ? dyn_cast<BlockArgument>(yield.getOutputs().front())
+                      : BlockArgument();
+    if (argument && argument.getOwner() == &body
+        && argument.getArgNumber() < deferred.getSources().size())
+      unused.push_back(deferred);
+  });
+  for (SpatDeferredCommunicationOp deferred : llvm::reverse(unused))
+    rewriter.eraseOp(deferred);
+}
+
 static LogicalResult verifyDominance(func::FuncOp funcOp) {
   DominanceInfo dominance(funcOp);
   WalkResult result = funcOp.walk([&](Operation *op) {
@@ -119,6 +145,9 @@ static LogicalResult verifyDominance(func::FuncOp funcOp) {
 LogicalResult realizeDeferredCommunication(
     func::FuncOp funcOp,
     const ScheduledComputeMaterializationResult &materialization) {
+  IRRewriter rewriter(funcOp.getContext());
+  eraseUnusedIdentityDeferredCommunications(funcOp, rewriter);
+
   auto transfers = buildDeferredTransferPlan(funcOp, materialization);
   if (failed(transfers))
     return funcOp.emitOpError(
@@ -134,7 +163,6 @@ LogicalResult realizeDeferredCommunication(
     return funcOp.emitOpError(
       "phase 2 failed to build sparse boundary programs");
 
-  IRRewriter rewriter(funcOp.getContext());
   if (failed(retargetDeferredPublications(funcOp, *transfers))
       || failed(replaceFinalGraphPublications(funcOp, *transfers)))
     return failure();

validation/operations/README.md

@@ -1,174 +1,308 @@
-# Validation Operations
+# Operation Validation Suite
 
-ONNX test models used by `validate.py` to verify the Raptor compiler + PIM simulator pipeline.
+This directory contains the ONNX models used by `validation/validate.py` to
+validate individual operations through compilation, PIM simulation, and
+comparison with the ONNX-MLIR reference runtime.
 
-Generated tests can be regenerated with:
+## Naming
 
-```
-python3 validation/operations/gen_tests.py
+Every model uses the same path convention:
+
+```text
+<category>/<case>/<category>_<case>.onnx
 ```
 
-## Conv
+The `<category>/<case>` pair is the operation ID printed by the validator.
+Use lowercase `snake_case` for both components. Keep the generator function,
+graph name, directory, and filename based on the same operation ID when adding
+a test.
 
-| Test             | Directory               | Input     | Output    | Kernel | Stride | Padding    | Bias | Notes                              |
-|------------------|-------------------------|-----------|-----------|--------|--------|------------|------|------------------------------------|
-| Simple           | `conv/simple`           | [1,3,3,3] | [1,1,2,2] | 2x2    | 1      | none       | no   | Basic conv, hand-crafted           |
-| With constant    | `conv/with_constant`    | [1,3,3,3] | [1,1,3,3] | 2x2    | 1      | SAME_UPPER | yes  | Hand-crafted, constant weight+bias |
-| Batch 2          | `conv/batch_2`          | [2,3,3,3] | [2,1,3,3] | 2x2    | 1      | SAME_UPPER | yes  | Batched input                      |
-| Kernel 3x3       | `conv/kernel_3x3`       | [1,1,5,5] | [1,1,3,3] | 3x3    | 1      | none       | no   | Larger kernel                      |
-| Stride 2         | `conv/stride_2`         | [1,1,6,6] | [1,1,2,2] | 3x3    | 2      | none       | no   | Strided convolution                |
-| Multi channel    | `conv/multi_channel`    | [1,3,5,5] | [1,4,3,3] | 3x3    | 1      | none       | no   | 3 in channels, 4 out channels      |
-| Pointwise 1x1    | `conv/pointwise_1x1`    | [1,8,4,4] | [1,4,4,4] | 1x1    | 1      | none       | no   | Channel mixing                     |
-| SAME padding 3x3 | `conv/same_padding_3x3` | [1,1,5,5] | [1,1,5,5] | 3x3    | 1      | SAME_UPPER | no   | Spatial dims preserved             |
-| Explicit padding | `conv/explicit_padding` | [1,1,4,4] | [1,1,4,4] | 3x3    | 1      | [1,1,1,1]  | no   | Symmetric explicit pads            |
-| With bias 3x3    | `conv/with_bias_3x3`    | [1,3,5,5] | [1,2,3,3] | 3x3    | 1      | none       | yes  | Multi-channel with bias            |
-| Large spatial    | `conv/large_spatial`    | [1,1,8,8] | [1,1,6,6] | 3x3    | 1      | none       | no   | Larger spatial input               |
-| Grouped two groups | `conv/grouped_two_groups` | [1,4,4,4] | [1,4,4,4] | 1x1 | 1 | none | yes | group=2 channel partitioning       |
-| Depthwise grouped | `conv/depthwise_grouped` | [1,3,4,4] | [1,3,2,2] | 3x3 | 1 | none | no | group=3, one input channel per group |
-| Dynamic          | `conv/dynamic`          | [1,1,4,4] | [1,1,2,2] | 3x3    | 1      | none       | no   | Runtime input and weight           |
+## Generate and validate
 
-## Gemm
+Regenerate all generated models from the repository root:
 
-| Test          | Directory               | A (input) | B/W tensor | Output   | transB | alpha | beta | Bias  | Notes                        |
-|---------------|-------------------------|-----------|------------|----------|--------|-------|------|-------|------------------------------|
-| Simple        | `gemm/simple`           | [10,132]  | [132,132]  | [10,132] | no     | 1     | 1    | no    | Square weights               |
-| Non-square    | `gemm/non_square`       | [4,128]   | [128,64]   | [4,64]   | no     | 1     | 1    | no    | K != N                       |
-| With bias     | `gemm/with_bias`        | [4,128]   | [128,128]  | [4,128]  | no     | 1     | 1    | [128] | Bias vector                  |
-| transB        | `gemm/transB`           | [4,128]   | [64,128]   | [4,64]   | yes    | 1     | 1    | no    | Transposed weight            |
-| Alpha/beta    | `gemm/alpha_beta`       | [4,64]    | [64,64]    | [4,64]   | no     | 0.5   | 0.25 | [64]  | Scaled matmul + bias         |
-| Small         | `gemm/small`            | [2,8]     | [8,4]      | [2,4]    | no     | 1     | 1    | no    | Tiny matrices                |
-| Large         | `gemm/large`            | [8,256]   | [256,128]  | [8,128]  | no     | 1     | 1    | no    | Larger matrices              |
-| transB + bias | `gemm/transB_with_bias` | [4,128]   | [64,128]   | [4,64]   | yes    | 1     | 1    | [64]  | Combined                     |
-| Dynamic       | `gemm/dynamic`          | [2,8]     | [8,4]      | [2,4]    | no     | 1     | 1    | no    | Runtime matrix operands      |
-| Dynamic transB | `gemm/dynamic_transB` | [2,8] | [4,8] | [2,4] | yes | 1 | 1 | no | Runtime transpose handling |
-| Dynamic bias  | `gemm/dynamic_bias`     | [2,8]     | [8,4]      | [2,4]    | no     | 1     | 1    | [4]   | Runtime bias broadcast       |
-| Dynamic alpha | `gemm/dynamic_alpha`    | [2,8]     | [8,4]      | [2,4]    | no     | 0.5   | 1    | no    | Runtime alpha scaling        |
-| Dynamic beta  | `gemm/dynamic_beta`     | [2,8]     | [8,4]      | [2,4]    | no     | 1     | 2    | [4]   | Runtime beta scaling         |
-| Dynamic bias + scale | `gemm/dynamic_bias_alpha_beta` | [2,8] | [8,4] | [2,4] | no | 0.5 | 2 | [4] | Runtime operands and bias |
+```bash
+.venv/bin/python validation/operations/gen_tests.py
+```
 
-## MatMul
+Run the complete suite with deadlock detection:
 
-| Test                | Directory                        | A input  | B tensor | Output  | Notes                                           |
-|---------------------|----------------------------------|----------|----------|---------|-------------------------------------------------|
-| Basic               | `matmul/basic`                  | [2,3]    | [3,4]    | [2,4]   | Direct 2D MatMul rewrite path                   |
-| Left constant       | `matmul/left_constant`          | [2,3]    | [3,4]    | [2,4]   | Constant LHS transpose rewrite path             |
-| Dynamic             | `matmul/dynamic`                | [2,3]    | [3,4]    | [2,4]   | Runtime matrix operands                         |
-| Batched 3D          | `matmul/batched_3d`            | [2,2,3]  | [2,3,4]  | [2,2,4] | Matching-batch direct batched lowering          |
-| Batched 3D dynamic  | `matmul/batched_3d_dynamic`    | [2,2,3]  | [2,3,4]  | [2,2,4] | Batched runtime operands                        |
-| Batched left const  | `matmul/batched_left_constant` | [2,2,3]  | [2,3,4]  | [2,2,4] | Batched constant-LHS transpose path             |
-| Batched RHS broadcast | `matmul/batched_rhs_broadcast` | [2,2,3] | [3,4] | [2,2,4] | Rank-2 RHS broadcast across batch               |
-| Batched LHS broadcast | `matmul/batched_lhs_broadcast` | [2,3] | [2,3,4] | [2,2,4] | Rank-2 LHS broadcast across batched RHS         |
+```bash
+.venv/bin/python validation/validate.py \
+  --raptor-path build_release/Release/bin/onnx-mlir \
+  --onnx-include-dir onnx-mlir/include \
+  --operations-dir validation/operations \
+  --crossbar-count 64 \
+  --crossbar-size 128 \
+  --core-count 144 \
+  --raptor-extra-arg=--pim-detect-communication-deadlock \
+  --raptor-extra-arg=--pim-export-spatial-dataflow=none
+```
 
-## Gemv
+Use `--compile-only` for compiler and deadlock checks, then `--run-only` to
+reuse those artifacts for reference execution, simulation, and comparison.
+The validator prints the complete operation results table before its summary.
 
-| Test                | Directory                          | Input    | W (weight) | Output  | Bias    | Notes                      |
-|---------------------|------------------------------------|----------|------------|---------|---------|----------------------------|
-| Simple              | `gemv/simple`                      | [1,132]  | [132,132]  | [1,132] | no      | Single-sample matmul       |
-| Constant            | `gemv/constant`                    | _(none)_ | [132,132]  | [1,132] | no      | All inputs constant        |
-| Homogeneous const   | `gemv/with_homogeneous_constant`   | [1,132]  | [132,132]  | [1,132] | [1,132] | Bias matches output shape  |
-| Heterogeneous const | `gemv/with_heterogeneous_constant` | [1,132]  | [132,132]  | [1,132] | [1,132] | Different constant pattern |
-| Scalar const        | `gemv/with_scalar_constant`        | [1,132]  | [132,132]  | [1,132] | [1,1]   | Scalar bias, broadcast     |
+## Complete inventory
 
-## Pool
+The suite contains 164 models. Tensor shapes, attributes, and constants are
+defined in `gen_tests.py` and in the checked-in ONNX models.
 
-| Test                       | Directory                       | Input     | Output    | Kernel                 | Stride | Padding    | Notes                            |
-|----------------------------|---------------------------------|-----------|-----------|------------------------|--------|------------|----------------------------------|
-| Max basic                  | `pool/max_basic`                | [1,1,4,4] | [1,1,3,3] | 2x2                    | 1      | none       | Basic max pooling                |
-| Max stride 2 multi-channel | `pool/max_stride2_multichannel` | [1,5,6,6] | [1,5,3,3] | 2x2                    | 2      | none       | Channel-preserving max pool      |
-| Max SAME_UPPER             | `pool/max_same_upper`           | [1,1,5,5] | [1,1,3,3] | 3x3                    | 2      | SAME_UPPER | Deprecated auto_pad path         |
-| Avg basic                  | `pool/avg_basic`                | [1,3,4,4] | [1,3,3,3] | 2x2                    | 1      | none       | Basic average pooling            |
-| Avg explicit padding       | `pool/avg_explicit_padding`     | [1,2,4,4] | [1,2,2,2] | 3x3                    | 2      | [1,1,1,1]  | `count_include_pad=0`            |
-| Avg include pad            | `pool/avg_include_pad`          | [1,2,4,4] | [1,2,2,2] | 3x3                    | 2      | [1,1,1,1]  | `count_include_pad=1`            |
-| Max after Conv             | `pool/max_after_conv`           | [1,3,6,6] | [1,4,2,2] | Conv 3x3 then Pool 2x2 | 2      | none       | Regression for `pool(conv(...))` |
+### Add (5)
 
-## ReduceMean
+| Case | Description |
+|---|---|
+| `after_gemm` | Gemm followed by Add with a broadcast bias vector. |
+| `basic` | Elementwise Add on two inputs with identical shapes. |
+| `broadcast_row` | Elementwise Add with row-vector broadcasting. |
+| `channel_broadcast_1024` | NCHW per-channel broadcasting over 1024 channels. |
+| `leading_dimension_broadcast` | Trailing-dimension broadcasting across leading dimensions. |
 
-| Test       | Directory                | Input     | Output    | Axes  | Keepdims | Notes                                           |
-|------------|--------------------------|-----------|-----------|-------|----------|-------------------------------------------------|
-| Basic      | `reduce_mean/basic`      | [4,8]     | [4,1]     | [1]   | 1        | Reduce feature dimension, preserving rank       |
-| Keepdims 0 | `reduce_mean/keepdims_0` | [4,8]     | [4]       | [1]   | 0        | Reduce feature dimension, dropping reduced axis |
-| 4D spatial | `reduce_mean/4d_spatial` | [1,3,4,4] | [1,3,1,1] | [2,3] | 1        | Reduce H and W on NCHW input                    |
-| After Conv | `reduce_mean/after_conv` | [1,3,5,5] | [1,2,1,1] | [2,3] | 1        | Conv 3x3 + bias, then spatial ReduceMean        |
+### Concat (3)
 
-## Relu
+| Case | Description |
+|---|---|
+| `channel_axis` | Concatenates two runtime NCHW tensors along the channel axis. |
+| `negative_axis` | Concatenates tensors using a negative axis. |
+| `three_inputs_channel_axis` | Concatenates three runtime NCHW tensors along the channel axis. |
 
-| Test       | Directory         | Input     | Output    | Notes                      |
-|------------|-------------------|-----------|-----------|----------------------------|
-| Basic      | `relu/basic`      | [4,8]     | [4,8]     | Standalone 2D Relu         |
-| 4D         | `relu/4d`         | [2,3,4,4] | [2,3,4,4] | Standalone NCHW Relu       |
-| After Conv | `relu/after_conv` | [1,3,5,5] | [1,2,3,3] | Conv 3x3 + bias, then Relu |
-| After Gemm | `relu/after_gemm` | [4,64]    | [4,32]    | Gemm + bias, then Relu     |
+### Conv (31)
 
-## Sigmoid
+| Case | Description |
+|---|---|
+| `batch_2` | Batched Conv with SAME_UPPER padding and bias. |
+| `batch_4_pointwise` | Pointwise Conv with batch size four. |
+| `depthwise_1024_channels` | Depthwise pointwise Conv with 1024 groups. |
+| `depthwise_grouped` | Depthwise-style grouped Conv with one input channel per group. |
+| `dilated_3x3` | Conv with a dilated 3x3 kernel. |
+| `dynamic` | Conv with runtime input and weight tensors. |
+| `explicit_padding` | 3x3 Conv with symmetric explicit padding. |
+| `grouped_many_groups` | Pointwise Conv with many groups and high channel counts. |
+| `grouped_two_groups` | Two-group pointwise Conv with bias. |
+| `huge_pointwise_1024` | Pointwise Conv with 1024 input and output channels. |
+| `huge_pointwise_1024_dynamic` | The 1024-channel pointwise Conv with runtime weights. |
+| `kernel_3x3` | Basic 3x3 Conv without padding. |
+| `kernel_equals_input_spatial` | Conv whose kernel covers the full spatial input. |
+| `large_input_channels_1x1` | Pointwise Conv with 1024 input channels and modest output width. |
+| `large_output_channels_1x1` | Pointwise Conv with modest input width and 1024 output channels. |
+| `large_spatial` | 3x3 Conv on a larger spatial input. |
+| `multi_channel` | 3x3 Conv with multiple input and output channels. |
+| `non_square_kernel_1x3` | Conv with a non-square 1x3 kernel. |
+| `non_square_kernel_3x1` | Conv with a non-square 3x1 kernel. |
+| `non_uniform_stride` | Conv with different height and width strides. |
+| `pointwise_1x1` | Basic pointwise channel-mixing Conv. |
+| `pointwise_tiled_chain` | Relu and chained pointwise Convs with a tiled intermediate. |
+| `real_asymmetric_padding` | Conv with asymmetric explicit padding. |
+| `relu_conv_store` | Conv, Relu, and a second Conv to validate an intermediate stored result. |
+| `same_lower_3x3` | 3x3 Conv with SAME_LOWER padding. |
+| `same_padding_3x3` | 3x3 Conv with SAME_UPPER padding. |
+| `simple` | Hand-authored basic 2x2 Conv. |
+| `stride_2` | 3x3 Conv with stride two. |
+| `with_bias_3x3` | Multi-channel 3x3 Conv with bias. |
+| `with_constant` | Hand-authored SAME_UPPER Conv with constant weight and bias. |
+| `without_kernel_shape_attr` | Conv whose kernel shape is inferred from its weight tensor. |
 
-| Test       | Directory            | Input     | Output    | Notes                     |
-|------------|----------------------|-----------|-----------|---------------------------|
-| Basic      | `sigmoid/basic`      | [4,8]     | [4,8]     | Standalone 2D Sigmoid     |
-| 4D         | `sigmoid/4d`         | [2,3,4,4] | [2,3,4,4] | Standalone NCHW Sigmoid   |
-| After Gemm | `sigmoid/after_gemm` | [4,64]    | [4,32]    | Gemm + bias, then Sigmoid |
+### Div (6)
 
-## Softmax
+| Case | Description |
+|---|---|
+| `after_gemm` | Gemm followed by Div with a broadcast divisor vector. |
+| `basic` | Elementwise Div by a same-shape constant tensor. |
+| `channel_broadcast_1024` | Div with NCHW per-channel broadcasting over 1024 channels. |
+| `leading_dimension_broadcast` | Div with trailing-dimension broadcasting. |
+| `runtime_scalar_rhs` | Div of a runtime tensor by a scalar initializer. |
+| `scalar_constant` | Div with scalar broadcasting on a 2D tensor. |
 
-| Test         | Directory                | Input       | Output      | Axis | Notes                           |
-|--------------|--------------------------|-------------|-------------|------|---------------------------------|
-| Basic        | `softmax/basic`          | [3,5]       | [3,5]       | 1    | Row-wise softmax over features  |
-| 3D last axis | `softmax/3d_last_axis`  | [2,3,4]     | [2,3,4]     | 2    | Last-dimension normalization    |
-| Channel axis | `softmax/channel_axis`  | [1,3,2,2]   | [1,3,2,2]   | 1    | NCHW channel-wise softmax       |
+### Gather (5)
 
-## Resize
+| Case | Description |
+|---|---|
+| `3d_input_axis1` | Gathers along axis one of a 3D input. |
+| `axis0_matrix_indices` | Gathers rows using a 2D indices tensor. |
+| `axis1` | Gathers selected columns from a 2D tensor. |
+| `negative_axis` | Gathers using a negative axis. |
+| `negative_indices` | Gathers with negative indices along axis zero. |
 
-| Test                | Directory               | Input     | Output    | Mode    | Notes                                   |
-|---------------------|-------------------------|-----------|-----------|---------|-----------------------------------------|
-| Nearest 2x          | `resize/nearest_2x`     | [1,1,2,3] | [1,1,4,6] | nearest | NCHW upsampling with scales [1,1,2,2]   |
-| Non-uniform scales  | `resize/non_uniform`    | [1,1,2,3] | [1,1,6,6] | nearest | Different height/width scaling factors  |
-| Explicit sizes      | `resize/with_sizes`     | [1,1,2,3] | [1,1,3,5] | nearest | Sizes input used instead of scales      |
+### Gemm (21)
 
-## Split
+| Case | Description |
+|---|---|
+| `alpha_beta` | Applies non-default alpha and beta scaling with bias. |
+| `bias_rank2_broadcast` | Broadcasts a rank-2 bias across output rows. |
+| `dynamic` | Uses both matrix operands at runtime. |
+| `dynamic_alpha` | Uses runtime operands with non-default alpha scaling. |
+| `dynamic_beta` | Uses runtime operands and bias with non-default beta scaling. |
+| `dynamic_bias` | Uses runtime matrix operands and runtime bias. |
+| `dynamic_bias_alpha_beta` | Combines runtime operands and bias with alpha and beta scaling. |
+| `dynamic_transB` | Transposes a runtime right-hand matrix. |
+| `huge_1024` | Uses 1024-wide inner and output dimensions. |
+| `large` | Exercises larger rectangular matrices. |
+| `large_k_small_n` | Uses a large reduction dimension and narrow output. |
+| `non_square` | Uses different reduction and output widths. |
+| `scalar_bias` | Broadcasts a scalar bias to the full output. |
+| `simple` | Basic Gemm with square weights. |
+| `small` | Tiny Gemm for fast focused validation. |
+| `small_k_large_n` | Uses a modest reduction dimension and wide output. |
+| `transA` | Transposes the left-hand matrix. |
+| `transA_transB` | Transposes both matrix operands. |
+| `transB` | Transposes the right-hand weight matrix. |
+| `transB_with_bias` | Combines a transposed weight matrix with bias. |
+| `with_bias` | Basic matrix product with vector bias. |
 
-| Test            | Directory                 | Input | Outputs              | Axis | Notes                               |
-|-----------------|---------------------------|-------|----------------------|------|-------------------------------------|
-| Basic           | `split/basic`             | [2,6] | [2,2], [2,4]        | 1    | Two-way split with explicit sizes   |
-| Equal three-way | `split/equal_three_way`   | [2,6] | [2,2], [2,2], [2,2] | 1    | Optional split input omitted        |
+### Gemv (5)
 
-## Gather
+| Case | Description |
+|---|---|
+| `constant` | Vector-matrix product with all inputs constant. |
+| `simple` | Basic single-row vector-matrix product. |
+| `with_heterogeneous_constant` | Adds a non-uniform constant bias pattern. |
+| `with_homogeneous_constant` | Adds a constant bias matching the output shape. |
+| `with_scalar_constant` | Adds a scalar broadcast bias. |
 
-| Test                 | Directory                      | Input | Indices | Output   | Axis | Notes                          |
-|----------------------|--------------------------------|-------|---------|----------|------|--------------------------------|
-| Axis 1               | `gather/axis1`                | [3,4] | [2]     | [3,2]    | 1    | Select two columns             |
-| Axis 0 matrix indices| `gather/axis0_matrix_indices` | [4,3] | [2,2]   | [2,2,3]  | 0    | Gather rows with 2D indices    |
+### MatMul (11)
 
-## Concat
+| Case | Description |
+|---|---|
+| `basic` | Direct 2D MatMul with constant right-hand matrix. |
+| `batched_3d` | Batched 3D MatMul with matching batch dimensions. |
+| `batched_3d_dynamic` | Batched 3D MatMul with both operands at runtime. |
+| `batched_left_constant` | Batched 3D MatMul with constant left-hand matrix. |
+| `batched_lhs_broadcast` | Broadcasts a 2D left-hand matrix across a batched right-hand tensor. |
+| `batched_rhs_broadcast` | Broadcasts a 2D right-hand matrix across a batched left-hand tensor. |
+| `dynamic` | Direct 2D MatMul with both operands at runtime. |
+| `huge_1024` | Uses 1024-wide inner and output dimensions. |
+| `left_constant` | Direct 2D MatMul with constant left-hand matrix. |
+| `matrix_vector` | Matrix-vector multiplication producing a 1D output. |
+| `vector_matrix` | Vector-matrix multiplication producing a 1D output. |
 
-| Test         | Directory             | Input(s)                  | Output    | Axis | Notes                       |
-|--------------|-----------------------|---------------------------|-----------|------|-----------------------------|
-| Channel axis | `concat/channel_axis` | A:[1,1,2,2], B:[1,2,2,2] | [1,3,2,2] | 1    | Runtime NCHW channel concat |
+### Mul (5)
 
-## Reshape
+| Case | Description |
+|---|---|
+| `after_conv` | Conv followed by per-channel scaling. |
+| `basic` | Elementwise Mul on two inputs with identical shapes. |
+| `channel_broadcast_1024` | Mul with NCHW per-channel broadcasting over 1024 channels. |
+| `leading_dimension_broadcast` | Mul with trailing-dimension broadcasting. |
+| `scalar_constant` | Mul with scalar broadcasting. |
 
-| Test      | Directory           | Input | Output | Notes                                        |
-|-----------|---------------------|-------|--------|----------------------------------------------|
-| Same rank | `reshape/same_rank` | [2,3] | [3,2]  | Runtime tensor with static shape initializer |
+### Pool (15)
 
-## Add
+| Case | Description |
+|---|---|
+| `avg_basic` | AveragePool with a 2x2 kernel and unit stride. |
+| `avg_ceil_mode` | AveragePool with ceil mode enabled. |
+| `avg_explicit_padding` | Explicitly padded AveragePool excluding pad from the divisor. |
+| `avg_include_pad` | Explicitly padded AveragePool including pad in the divisor. |
+| `avg_large_channels` | AveragePool with a large channel count and small spatial extent. |
+| `avg_non_uniform_stride` | AveragePool with different height and width strides. |
+| `avg_real_asymmetric_padding` | AveragePool with asymmetric explicit padding. |
+| `max_after_conv` | Conv followed by MaxPool. |
+| `max_basic` | MaxPool with a 2x2 kernel and unit stride. |
+| `max_ceil_mode` | MaxPool with ceil mode enabled. |
+| `max_global_style_kernel_equals_input` | MaxPool whose kernel covers the full spatial input. |
+| `max_non_square_kernel` | MaxPool with a non-square kernel. |
+| `max_real_asymmetric_padding` | MaxPool with asymmetric explicit padding. |
+| `max_same_upper` | MaxPool with SAME_UPPER padding. |
+| `max_stride2_multichannel` | Multi-channel MaxPool with stride two. |
 
-| Test          | Directory           | Input(s)         | Output | Notes                                       |
-|---------------|---------------------|------------------|--------|---------------------------------------------|
-| Basic         | `add/basic`         | A:[4,8], B:[4,8] | [4,8]  | Elementwise add, same-shape inputs          |
-| Broadcast row | `add/broadcast_row` | A:[4,8], B:[8]   | [4,8]  | Row-vector broadcasting via initializer     |
-| After Gemm    | `add/after_gemm`    | A:[4,64], D:[32] | [4,32] | Gemm + bias, then Add with broadcast vector |
+### ReduceMean (17)
 
-## Mul
+| Case | Description |
+|---|---|
+| `4d_spatial` | Reduces height and width of an NCHW tensor while preserving rank. |
+| `4d_spatial_keepdims_0` | Reduces NCHW height and width while dropping those axes. |
+| `after_conv` | Conv followed by a spatial ReduceMean. |
+| `all_axes_keepdims_0` | Reduces all axes to a scalar. |
+| `all_axes_keepdims_1` | Reduces all axes while preserving rank. |
+| `basic` | Reduces a feature dimension while preserving rank. |
+| `channel_axis_nchw` | Reduces the channel axis of an NCHW tensor. |
+| `keepdims_0` | Reduces a feature dimension and drops that axis. |
+| `large_dimension_1024` | Reduces a dimension of length 1024. |
+| `legacy_axes_1_2_keepdims_1` | Opset-18 reduction over multiple positive axes. |
+| `legacy_axis1_keepdims_0` | Opset-18 reduction over one axis while dropping it. |
+| `legacy_axis1_keepdims_1` | Opset-18 reduction over one axis while preserving rank. |
+| `legacy_empty_axes_noop` | Opset-18 empty-axes no-op followed by Relu. |
+| `legacy_nchw_spatial` | Opset-18 spatial reduction on NCHW input. |
+| `legacy_negative_axis` | Opset-18 reduction using a negative axis. |
+| `legacy_reduce_all_keepdims_1` | Opset-18 all-axis reduction with the axes input omitted. |
+| `negative_axis` | ReduceMean using a negative axis. |
 
-| Test            | Directory             | Input(s)                 | Output    | Notes                                     |
-|-----------------|-----------------------|--------------------------|-----------|-------------------------------------------|
-| Basic           | `mul/basic`           | A:[4,8], B:[4,8]         | [4,8]     | Elementwise multiply, same-shape inputs   |
-| Scalar constant | `mul/scalar_constant` | X:[4,8], S:[1]           | [4,8]     | Scalar broadcasting via initializer       |
-| After Conv      | `mul/after_conv`      | X:[1,3,5,5], S:[1,2,1,1] | [1,2,3,3] | Conv 3x3 + bias, then per-channel scaling |
+### Relu (4)
 
-## Div
+| Case | Description |
+|---|---|
+| `4d` | Standalone Relu on an NCHW tensor. |
+| `after_conv` | Conv followed by Relu. |
+| `after_gemm` | Gemm followed by Relu. |
+| `basic` | Standalone Relu on a 2D tensor. |
 
-| Test            | Directory             | Input(s)         | Output | Notes                                                |
-|-----------------|-----------------------|------------------|--------|------------------------------------------------------|
-| Basic           | `div/basic`           | X:[4,8], D:[4,8] | [4,8]  | Elementwise divide by same-shape constant tensor     |
-| Scalar constant | `div/scalar_constant` | X:[4,8], S:[1]   | [4,8]  | Scalar broadcasting via initializer                  |
-| After Gemm      | `div/after_gemm`      | A:[4,64], D:[32] | [4,32] | Gemm + bias, then Div with positive broadcast vector |
+### Reshape (4)
+
+| Case | Description |
+|---|---|
+| `4d_to_2d_flatten` | Flattens a 4D tensor into a 2D view. |
+| `infer_dim_minus_one` | Uses `-1` to infer one output dimension. |
+| `same_rank` | Changes shape without changing rank. |
+| `zero_copies_input_dim` | Uses `0` to copy an input dimension. |
+
+### Resize (6)
+
+| Case | Description |
+|---|---|
+| `height_only` | Nearest-neighbor resize of only the height dimension. |
+| `nearest_2x` | Nearest-neighbor upsampling by a factor of two. |
+| `nearest_downsample` | Nearest-neighbor downsampling. |
+| `non_uniform` | Nearest-neighbor resize with different spatial scales. |
+| `width_only` | Nearest-neighbor resize of only the width dimension. |
+| `with_sizes` | Resize using explicit output sizes instead of scales. |
+
+### Sigmoid (3)
+
+| Case | Description |
+|---|---|
+| `4d` | Standalone Sigmoid on an NCHW tensor. |
+| `after_gemm` | Gemm followed by Sigmoid. |
+| `basic` | Standalone Sigmoid on a 2D tensor. |
+
+### Slice (8)
+
+| Case | Description |
+|---|---|
+| `2d_basic` | Slices a 2D tensor with explicit axes and unit steps. |
+| `after_conv` | Conv followed by a spatial crop. |
+| `default_axes` | Omits axes and steps to use positional defaults. |
+| `large_channel_1024` | Slices a channel range from a 1024-channel tensor. |
+| `nchw_spatial_crop` | Crops the spatial axes of an NCHW tensor. |
+| `negative_axis` | Slices using a negative axis. |
+| `negative_indices` | Slices using negative indices. |
+| `step2` | Slices using a positive step greater than one. |
+
+### Softmax (5)
+
+| Case | Description |
+|---|---|
+| `3d_last_axis` | Softmax over the last axis of a 3D tensor. |
+| `basic` | Softmax over the last dimension of a 2D tensor. |
+| `channel_axis` | Softmax over the channel axis of an NCHW tensor. |
+| `large_dimension_1024` | Softmax over a last dimension of length 1024. |
+| `negative_axis` | Softmax using a negative axis. |
+
+### Split (4)
+
+| Case | Description |
+|---|---|
+| `basic` | Splits a 2D tensor into two explicit output sizes. |
+| `equal_three_way` | Splits a 2D tensor evenly into three outputs. |
+| `negative_axis` | Splits using a negative axis. |
+| `uneven_channel_axis_4d` | Splits an NCHW channel axis into uneven outputs. |
+
+### Sub (6)
+
+| Case | Description |
+|---|---|
+| `after_gemm` | Gemm followed by Sub with a broadcast constant vector. |
+| `basic` | Elementwise Sub on two runtime inputs with identical shapes. |
+| `broadcast_row` | Sub with a broadcast row-vector right-hand constant. |
+| `channel_broadcast_1024` | Sub with NCHW per-channel broadcasting over 1024 channels. |
+| `constant_lhs_broadcast` | Sub with a broadcast constant left-hand operand. |
+| `leading_dimension_broadcast` | Sub with trailing-dimension broadcasting. |

validation/operations/gen_tests.py

@@ -80,7 +80,7 @@ def conv_1x1():
                             kernel_shape=[1, 1], strides=[1, 1], pads=[0, 0, 0, 0])
     graph = helper.make_graph([node], "conv_1x1", [X], [Y], initializer=[W])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "conv/pointwise_1x1", "conv_1x1.onnx")
+    save_model(model, "conv/pointwise_1x1", "conv_pointwise_1x1.onnx")
 
 
 def conv_same_padding_3x3():
@@ -218,6 +218,33 @@ def conv_huge_pointwise_1024_dynamic():
     save_model(model, "conv/huge_pointwise_1024_dynamic", "conv_huge_pointwise_1024_dynamic.onnx")
 
 
+def conv_pointwise_tiled_chain():
+    """Chained pointwise Convs with a tiled intermediate."""
+    X = helper.make_tensor_value_info("X", TensorProto.FLOAT, [1, 1024, 1, 1])
+    Y = helper.make_tensor_value_info("Y", TensorProto.FLOAT, [1, 256, 1, 1])
+    rng = np.random.default_rng(79)
+    W1 = numpy_helper.from_array(
+        rng.uniform(-1, 1, (1024, 1024, 1, 1)).astype(np.float32), name="W1")
+    B1 = numpy_helper.from_array(
+        rng.uniform(-1, 1, 1024).astype(np.float32), name="B1")
+    W2 = numpy_helper.from_array(
+        rng.uniform(-1, 1, (256, 1024, 1, 1)).astype(np.float32), name="W2")
+    B2 = numpy_helper.from_array(
+        rng.uniform(-1, 1, 256).astype(np.float32), name="B2")
+    nodes = [
+        helper.make_node("Relu", ["X"], ["X_relu"]),
+        helper.make_node("Conv", ["X_relu", "W1", "B1"], ["hidden"],
+                         kernel_shape=[1, 1], strides=[1, 1], pads=[0, 0, 0, 0]),
+        helper.make_node("Relu", ["hidden"], ["hidden_relu"]),
+        helper.make_node("Conv", ["hidden_relu", "W2", "B2"], ["Y"],
+                         kernel_shape=[1, 1], strides=[1, 1], pads=[0, 0, 0, 0]),
+    ]
+    graph = helper.make_graph(
+        nodes, "conv_pointwise_tiled_chain", [X], [Y], initializer=[W1, B1, W2, B2])
+    model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
+    save_model(model, "conv/pointwise_tiled_chain", "conv_pointwise_tiled_chain.onnx")
+
+
 def conv_large_output_channels_1x1():
     """1x1 Conv with modest inputs and very large output channel count."""
     X = helper.make_tensor_value_info("X", TensorProto.FLOAT, [1, 64, 1, 1])
@@ -790,7 +817,7 @@ def maxpool_basic():
     node = helper.make_node("MaxPool", ["X"], ["Y"], kernel_shape=[2, 2], strides=[1, 1], pads=[0, 0, 0, 0])
     graph = helper.make_graph([node], "maxpool_basic", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/max_basic", "maxpool_basic.onnx")
+    save_model(model, "pool/max_basic", "pool_max_basic.onnx")
 
 
 def maxpool_stride2_multichannel():
@@ -800,7 +827,7 @@ def maxpool_stride2_multichannel():
     node = helper.make_node("MaxPool", ["X"], ["Y"], kernel_shape=[2, 2], strides=[2, 2], pads=[0, 0, 0, 0])
     graph = helper.make_graph([node], "maxpool_stride2_multichannel", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/max_stride2_multichannel", "maxpool_stride2_multichannel.onnx")
+    save_model(model, "pool/max_stride2_multichannel", "pool_max_stride2_multichannel.onnx")
 
 
 def maxpool_same_upper():
@@ -810,7 +837,7 @@ def maxpool_same_upper():
     node = helper.make_node("MaxPool", ["X"], ["Y"], kernel_shape=[3, 3], strides=[2, 2], auto_pad="SAME_UPPER")
     graph = helper.make_graph([node], "maxpool_same_upper", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/max_same_upper", "maxpool_same_upper.onnx")
+    save_model(model, "pool/max_same_upper", "pool_max_same_upper.onnx")
 
 
 def avgpool_basic():
@@ -820,7 +847,7 @@ def avgpool_basic():
     node = helper.make_node("AveragePool", ["X"], ["Y"], kernel_shape=[2, 2], strides=[1, 1], pads=[0, 0, 0, 0])
     graph = helper.make_graph([node], "avgpool_basic", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/avg_basic", "avgpool_basic.onnx")
+    save_model(model, "pool/avg_basic", "pool_avg_basic.onnx")
 
 
 def avgpool_explicit_padding():
@@ -831,7 +858,7 @@ def avgpool_explicit_padding():
                             kernel_shape=[3, 3], strides=[2, 2], pads=[1, 1, 1, 1], count_include_pad=0)
     graph = helper.make_graph([node], "avgpool_explicit_padding", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/avg_explicit_padding", "avgpool_explicit_padding.onnx")
+    save_model(model, "pool/avg_explicit_padding", "pool_avg_explicit_padding.onnx")
 
 
 def avgpool_include_pad():
@@ -842,7 +869,7 @@ def avgpool_include_pad():
                             kernel_shape=[3, 3], strides=[2, 2], pads=[1, 1, 1, 1], count_include_pad=1)
     graph = helper.make_graph([node], "avgpool_include_pad", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/avg_include_pad", "avgpool_include_pad.onnx")
+    save_model(model, "pool/avg_include_pad", "pool_avg_include_pad.onnx")
 
 
 def maxpool_after_conv():
@@ -855,7 +882,7 @@ def maxpool_after_conv():
     pool = helper.make_node("MaxPool", ["C"], ["Y"], kernel_shape=[2, 2], strides=[2, 2], pads=[0, 0, 0, 0])
     graph = helper.make_graph([conv, pool], "maxpool_after_conv", [X], [Y], initializer=[W])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/max_after_conv", "maxpool_after_conv.onnx")
+    save_model(model, "pool/max_after_conv", "pool_max_after_conv.onnx")
 
 
 def maxpool_ceil_mode():
@@ -866,7 +893,7 @@ def maxpool_ceil_mode():
                             kernel_shape=[2, 2], strides=[2, 2], pads=[0, 0, 0, 0], ceil_mode=1)
     graph = helper.make_graph([node], "maxpool_ceil_mode", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/max_ceil_mode", "maxpool_ceil_mode.onnx")
+    save_model(model, "pool/max_ceil_mode", "pool_max_ceil_mode.onnx")
 
 
 def avgpool_ceil_mode():
@@ -877,7 +904,7 @@ def avgpool_ceil_mode():
                             kernel_shape=[2, 2], strides=[2, 2], pads=[0, 0, 0, 0], ceil_mode=1)
     graph = helper.make_graph([node], "avgpool_ceil_mode", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/avg_ceil_mode", "avgpool_ceil_mode.onnx")
+    save_model(model, "pool/avg_ceil_mode", "pool_avg_ceil_mode.onnx")
 
 
 def maxpool_real_asymmetric_padding():
@@ -888,7 +915,7 @@ def maxpool_real_asymmetric_padding():
                             kernel_shape=[3, 3], strides=[1, 2], pads=[0, 1, 2, 1])
     graph = helper.make_graph([node], "maxpool_real_asymmetric_padding", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/max_real_asymmetric_padding", "maxpool_real_asymmetric_padding.onnx")
+    save_model(model, "pool/max_real_asymmetric_padding", "pool_max_real_asymmetric_padding.onnx")
 
 
 def avgpool_real_asymmetric_padding():
@@ -899,7 +926,7 @@ def avgpool_real_asymmetric_padding():
                             kernel_shape=[3, 3], strides=[1, 2], pads=[0, 1, 2, 1], count_include_pad=0)
     graph = helper.make_graph([node], "avgpool_real_asymmetric_padding", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/avg_real_asymmetric_padding", "avgpool_real_asymmetric_padding.onnx")
+    save_model(model, "pool/avg_real_asymmetric_padding", "pool_avg_real_asymmetric_padding.onnx")
 
 
 def maxpool_non_square_kernel():
@@ -910,7 +937,7 @@ def maxpool_non_square_kernel():
                             kernel_shape=[2, 3], strides=[1, 2], pads=[0, 0, 0, 0])
     graph = helper.make_graph([node], "maxpool_non_square_kernel", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/max_non_square_kernel", "maxpool_non_square_kernel.onnx")
+    save_model(model, "pool/max_non_square_kernel", "pool_max_non_square_kernel.onnx")
 
 
 def avgpool_non_uniform_stride():
@@ -921,7 +948,7 @@ def avgpool_non_uniform_stride():
                             kernel_shape=[2, 3], strides=[1, 2], pads=[0, 0, 0, 0])
     graph = helper.make_graph([node], "avgpool_non_uniform_stride", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/avg_non_uniform_stride", "avgpool_non_uniform_stride.onnx")
+    save_model(model, "pool/avg_non_uniform_stride", "pool_avg_non_uniform_stride.onnx")
 
 
 def maxpool_global_style_kernel_equals_input():
@@ -931,7 +958,7 @@ def maxpool_global_style_kernel_equals_input():
     node = helper.make_node("MaxPool", ["X"], ["Y"], kernel_shape=[4, 4], strides=[1, 1], pads=[0, 0, 0, 0])
     graph = helper.make_graph([node], "maxpool_global_style_kernel_equals_input", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/max_global_style_kernel_equals_input", "maxpool_global_style_kernel_equals_input.onnx")
+    save_model(model, "pool/max_global_style_kernel_equals_input", "pool_max_global_style_kernel_equals_input.onnx")
 
 
 def avgpool_large_channels():
@@ -941,7 +968,7 @@ def avgpool_large_channels():
     node = helper.make_node("AveragePool", ["X"], ["Y"], kernel_shape=[2, 2], strides=[1, 1], pads=[0, 0, 0, 0])
     graph = helper.make_graph([node], "avgpool_large_channels", [X], [Y])
     model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
-    save_model(model, "pool/avg_large_channels", "avgpool_large_channels.onnx")
+    save_model(model, "pool/avg_large_channels", "pool_avg_large_channels.onnx")
 
 
 # ---------------------------------------------------------------------------
@@ -2005,6 +2032,7 @@ if __name__ == "__main__":
     conv_dynamic()
     conv_huge_pointwise_1024()
     conv_huge_pointwise_1024_dynamic()
+    conv_pointwise_tiled_chain()
     conv_large_output_channels_1x1()
     conv_large_input_channels_1x1()
     conv_depthwise_1024_channels()

validation/validate.py

@@ -174,22 +174,18 @@ def main():
     # Summary
     n_passed = sum(1 for passed in results.values() if passed)
     n_total = len(results)
-    failing = [rel for rel, passed in results.items() if not passed]
-    if a.verbose or failing:
-        status_width = len("Result")
-        path_width = max(len("Operation"), *(len(rel) for rel in results))
-        separator = f"+-{'-' * path_width}-+-{'-' * status_width}-+"
-        print(separator)
-        print(f"| {'Operation'.ljust(path_width)} | {'Result'.ljust(status_width)} |")
-        print(separator)
-        for rel, passed in results.items():
-            if not a.verbose and passed:
-                continue
-            plain_status = "PASS" if passed else "FAIL"
-            status = Fore.GREEN + plain_status.ljust(status_width) + Style.RESET_ALL if passed else \
-                Fore.RED + plain_status.ljust(status_width) + Style.RESET_ALL
-            print(f"| {rel.ljust(path_width)} | {status} |")
-        print(separator)
+    status_width = len("Result")
+    path_width = max(len("Operation"), *(len(rel) for rel in results))
+    separator = f"+-{'-' * path_width}-+-{'-' * status_width}-+"
+    print(separator)
+    print(f"| {'Operation'.ljust(path_width)} | {'Result'.ljust(status_width)} |")
+    print(separator)
+    for rel, passed in results.items():
+        plain_status = "PASS" if passed else "FAIL"
+        status = Fore.GREEN + plain_status.ljust(status_width) + Style.RESET_ALL if passed else \
+            Fore.RED + plain_status.ljust(status_width) + Style.RESET_ALL
+        print(f"| {rel.ljust(path_width)} | {status} |")
+    print(separator)
     print("\n" + Style.BRIGHT + Fore.CYAN + "Summary" + Style.RESET_ALL)
     print(Style.BRIGHT + f"Passed: {n_passed}" + Style.RESET_ALL)
     print(Style.BRIGHT + f"Failed: {n_total - n_passed}" + Style.RESET_ALL)