better compile times for throughput scheduling

fix failing no-plan throughput ablation runs
update results

docs/compiler/scheduling.md

@@ -63,8 +63,11 @@ lanes of an operation and computes one ordered partition with a prefix DP.
 The DP jointly assigns operation prefixes and at most sixteen units of spare
 core capacity. Each configured stage reserves one physical core. Quantization
 limits initialization only; refinement uses geometric core transfers, including one core.
-Quantization can miss a feasible allocation. Initialization failure reports a
+If the quality grid has no feasible state, a minimum-resource prefix DP checks
-bounded-search limitation, not proved hardware infeasibility.
+every physical core count for each segment. This preserves feasibility over
+the segment assignments found by the constructive packers, independently of
+resource quantization. Packing failure remains a bounded-search limitation,
+not proved hardware infeasibility.
 The seed minimizes maximum stage packing/dependency priority over the legal
 active-stage counts within the configured depth. These priorities provide
 a starting point, not a timing guarantee. Refinement can remove stage boundaries.
@@ -76,6 +79,13 @@ constructive ordered-packing curves and uses reuse-first resident-weight packing
 when order alone cannot realize a capacity bound. A feasible packing remains
 feasible with idle extra cores. All placements pass the existing topology-aware
 physical packing, operation-group, adjacency and capacity checks.
+Ordered packing returns an unavailable cost when no ordered assignment fits.
+Every finite feasibility cost has an ordered or reuse-first constructive assignment.
+Initialization retains that assignment before adding idle spare cores. If topology-aware
+greedy placement and reconstruction in physical task order fail, it uses the
+retained assignment.
+Task reordering cannot discard the initial partition's constructive capacity evidence.
+This evidence does not prove final backend legality.
 
 Refinement uses a beam of two schedules ranked by the larger queue estimate.
 Its ordered neighborhood moves operation groups across adjacent stage boundaries,
@@ -163,7 +173,9 @@ Unsupported or incomplete target timing produces an `unavailable` report.
 
 The analysis brackets the maximum cycle ratio within one nanosecond. It tests
 edge delays `w - II*d`, sweeping the distance-zero DAG between recurrence
-relaxations. At most `R+1` sweeps detect a positive cycle, where `R` counts
+relaxations. A positive predecessor cycle can terminate an infeasible probe early.
+The final witness pass retains the full relaxation procedure and deterministic
+witness selection. At most `R+1` sweeps detect a positive cycle, where `R` counts
 positive-distance edges. Bisection uses at most 64 tests. Timing outside the
 supported numeric range produces an explicit unavailable result. A predecessor
 cycle supplies a witness, including delay, iteration distance, cores, and links.
@@ -337,6 +349,11 @@ Functional simulation still executes Softmax. Estimated timing and energy exclud
 The shared queue engine models controller arrival arbitration and cross-generation dependencies.
 The compiler does not execute candidate programs or invoke a simulator.
 
+Spatial extraction permits at most one million operation visits per physical core.
+The total visit count is diagnostic; adding cores does not reduce another core's analysis allowance.
+Abstract queue construction permits two million non-work event visits per candidate.
+The complete-state recurrence limits remain unchanged.
+
 Both arbitration policies must establish complete-state recurrence within bounded analysis budgets.
 Their estimates are sensitivity checks, not formal bounds.
 The candidate with the lowest larger estimate must separate from the initialized schedule.
@@ -362,6 +379,8 @@ If every candidate fails, one diagnostic lists all attempted candidates and caus
 
 `reports/pipeline_selection.json` records Spatial costs, eligibility, selection, and backend attempts.
 It also records search parents, compact task-stage runs, core budgets, recurrence contributions, termination, attempts, queue budget, and phase times.
+Phase times separate neighborhood scoring, constructive packing, dependency-witness
+analysis, and queue execution with its complete-state proof.
 `--enable-timing` adds per-pass lowering times and IR operation/value counts.
 Final instruction generation uses the normal compiler timing report.
 Explicit Spatial exports and IR dumps reconstruct pre-lowering artifacts for the selected schedule.
@@ -424,7 +443,8 @@ fragment/consumer demands, and resource quantization `Q <= 16`.
 | Weight indexing | Up to `O(R W)` equality comparisons |
 | Weight lower bounds | `O(R + W)` / `O(W)` bits |
 | Constructive best fit | `O(T log T + P R + P W)` / `O(P W)` bits plus `O(T + P)`; bounded retries only on failure |
-| Joint initialization DP | `O(G² Q Kpack + S G² Q²)` time / `O(G² Q + S G Q)` storage; `Kpack` is one ordered-pack cost query |
+| Joint initialization DP | `O(G² Q Kpack + S G² Q²)` time / `O(G² Q + S G Q)` storage; `Kpack` is one constructive pack-cost query |
+| Feasibility completion, only if the quality grid fails | `O(G² P Kpack + S G²)` time / `O(G² + S G)` additional storage |
 | Neighborhood memory service | Expected `O(T + F)` per partition / `O(T + F)` scratch |
 | Topology clustering / cached PEFT distances | `O(P²)` time |
 | Concat cost extraction | `O(A D)` for `A` inputs of rank `D`; constant auxiliary space |

src/PIM/Compiler/PimPeriodicAnalysis.cpp

@@ -369,7 +369,7 @@ public:
   std::vector<std::pair<size_t, Cost>> memoryServices;
   std::array<Cost, pim_binary::kOpcodeCount> opcodeService{}, opcodeCounts{};
   std::string error;
-  size_t abstractBudget = 1000000;
+  size_t abstractBudget = 2000000;
 
   bool readCore(int64_t core, llvm::StringRef data, llvm::ArrayRef<int64_t> groups);
   bool readExecution(int64_t core, const std::vector<PimExecutionEvent> &program);
@@ -845,13 +845,17 @@ bool Analysis::connect() {
 
 EvaluatedAnalysis Analysis::evaluate() {
   EvaluatedAnalysis result;
+  double dependencySeconds = 0;
   if (cores.empty()) result.error = "artifact has no active repeating instruction stream";
   else if (!connect()) result.error = error;
   if (!result.error.empty()) { result.cost.reason = result.error; return result; }
   if (options.detailed || options.searchGuidance) {
+    auto start = std::chrono::steady_clock::now();
     auto recurrence = analyzePeriodicDependencies(events.size(), edges, result.error);
     if (mlir::failed(recurrence)) { result.cost.reason = result.error; return result; }
     result.recurrence = std::move(*recurrence);
+    dependencySeconds = std::chrono::duration<double>(
+      std::chrono::steady_clock::now() - start).count();
   }
   if (!options.detailed || memoryQueue) {
     auto start = std::chrono::steady_clock::now();
@@ -885,6 +889,7 @@ EvaluatedAnalysis Analysis::evaluate() {
           += static_cast<double>(edge.delayNs) / distance;
       }
   }
+  result.cost.timing.dependencySeconds = dependencySeconds;
   return result;
 }
 
@@ -894,6 +899,8 @@ llvm::json::Object Analysis::report(const EvaluatedAnalysis &evaluation) const {
     result["status"] = "unavailable"; result["reason"] = evaluation.error; return result;
   }
   result["global_memory_service_ns"] = static_cast<int64_t>(memory.serviceNs);
+  if (options.measureTime)
+    result["dependency_analysis_ms"] = evaluation.cost.timing.dependencySeconds * 1000;
   result["global_read_bytes"] = static_cast<int64_t>(memory.readBytes);
   result["global_write_bytes"] = static_cast<int64_t>(memory.writeBytes);
   result["global_read_requests"] = static_cast<int64_t>(memory.readRequests);

src/PIM/Compiler/PimPipelineSelection.cpp

@@ -113,7 +113,7 @@ struct PimPipelineSelectionPass final : PassWrapper<PimPipelineSelectionPass, Op
       PreparedSpatialSchedules &prepared, std::vector<WholeScheduleCost> &costs,
       llvm::json::Object &report) {
     llvm::json::Array summaries;
-    double realization = 0, extraction = 0, expansion = 0, queue = 0, evaluation = 0;
+    double realization = 0, extraction = 0, expansion = 0, queue = 0, evaluation = 0, dependency = 0;
     auto evaluate = [&](const MergeScheduleResult &schedule) {
       size_t index = summaries.size();
       llvm::json::Object summary{{"index", static_cast<int64_t>(index)}};
@@ -131,6 +131,7 @@ struct PimPipelineSelectionPass final : PassWrapper<PimPipelineSelectionPass, Op
       extraction += cost.timing.extractionSeconds;
       expansion += cost.timing.expansionSeconds;
       queue += cost.timing.evaluationSeconds;
+      dependency += cost.timing.dependencySeconds;
       auto convergence = cost.convergenceProved(), support = cost.loweringSupported();
       summary["cost_available"] = cost.available();
       summary["convergence_proved"] = convergence
@@ -183,6 +184,7 @@ struct PimPipelineSelectionPass final : PassWrapper<PimPipelineSelectionPass, Op
     report["event_extraction_seconds"] = extraction;
     report["event_expansion_seconds"] = expansion;
     report["queue_seconds"] = queue;
+    report["dependency_analysis_seconds"] = dependency;
     report["candidates"] = std::move(summaries);
     if (failed(search)) {
       report["status"] = "incomplete-costs";
@@ -204,6 +206,8 @@ struct PimPipelineSelectionPass final : PassWrapper<PimPipelineSelectionPass, Op
     report["single_candidate_analysis_skipped"] = single;
     report["search_rounds"] = static_cast<int64_t>(search->rounds);
     report["packing_attempts"] = static_cast<int64_t>(search->packingAttempts);
+    report["neighborhood_seconds"] = search->neighborhoodSeconds;
+    report["constructive_packing_seconds"] = search->packingSeconds;
     report["considered_assignments"] = static_cast<int64_t>(search->consideredAssignments);
     report["search_termination"] = search->termination;
     report["queue_evaluation_budget"] = static_cast<int64_t>(kPipelineQueueEvaluationBudget);

src/PIM/Compiler/PimSpatialCost.cpp

@@ -58,6 +58,10 @@ public:
         program(target.processorCount) {}
   bool build(ModuleOp module);
 private:
+  // Bound interpretation per physical core, independently of batch lane count.
+  static constexpr size_t maximumCoreVisits = 1000000;
+  size_t coreVisits = 0;
+  int64_t currentCore = -1;
   llvm::SmallDenseSet<int64_t, 16> assignedCores;
   DenseMap<Value, Residence> residence;
   DenseMap<Value, CopyAccessPlan> publicationAccess;
@@ -378,7 +382,12 @@ bool Extraction::operation(Operation &op, Sequence &sequence) {
 
 bool Extraction::block(Block &body, Sequence &sequence) {
   for (Operation &op : body) {
-    if (++visited > 1000000) { error = "abstract extraction budget exceeded"; return false; }
+    ++visited;
+    if (++coreVisits > maximumCoreVisits) {
+      error = "abstract extraction budget exceeded on physical core "
+        + std::to_string(currentCore);
+      return false;
+    }
     if (!operation(op, sequence)) return false;
     for (Value result : op.getResults())
       if (auto outputs = directOutputs.find(result); outputs != directOutputs.end())
@@ -400,6 +409,7 @@ bool Extraction::core(Operation *op, int64_t physicalCore, unsigned lane) {
     loweringSupport = StaticLoweringSupport::Unsupported;
     error = "duplicate physical core assignment: " + std::to_string(physicalCore); return false;
   }
+  coreVisits = 0; currentCore = physicalCore;
   residence.clear(); publicationAccess.clear(); directOutputs.clear(); compactStorage.clear();
   environment.bindings.clear();
   ValueRange inputs, weights;

src/PIM/Dialect/Spatial/Passes/Transforms/MergeComputeNodes/Scheduling/PeriodicDependencies.cpp

@@ -57,7 +57,29 @@ mlir::FailureOr<PeriodicDependencyEstimate> analyzePeriodicDependencies(
   const size_t absent = std::numeric_limits<size_t>::max();
   std::vector<double> potential(nodeCount);
   std::vector<size_t> predecessor(nodeCount, absent);
-  auto infeasible = [&](double interval) {
+  std::vector<size_t> cycleSeen(nodeCount);
+  auto hasPositivePredecessorCycle = [&](double interval) {
+    std::fill(cycleSeen.begin(), cycleSeen.end(), absent);
+    for (size_t start = 0; start < nodeCount; ++start) {
+      size_t node = start;
+      while (cycleSeen[node] == absent && predecessor[node] != absent) {
+        cycleSeen[node] = start;
+        node = edges[predecessor[node]].source;
+      }
+      if (cycleSeen[node] != start || predecessor[node] == absent) continue;
+      const size_t first = node;
+      Cost delay = 0, distance = 0;
+      do {
+        const auto &edge = edges[predecessor[node]];
+        delay = checkedAdd(delay, edge.delayNs);
+        distance = checkedAdd(distance, static_cast<Cost>(edge.distance));
+        node = edge.source;
+      } while (node != first);
+      if (static_cast<double>(delay) - interval * distance > 1e-7) return true;
+    }
+    return false;
+  };
+  auto infeasible = [&](double interval, bool earlyCycle = true) {
     std::fill(potential.begin(), potential.end(), 0);
     std::fill(predecessor.begin(), predecessor.end(), absent);
     auto relax = [&](size_t index) {
@@ -85,6 +107,9 @@ mlir::FailureOr<PeriodicDependencyEstimate> analyzePeriodicDependencies(
           changed |= relax(index);
       if (!changed)
         return false;
+      // A positive predecessor cycle already proves infeasibility. Keep the
+      // final witness pass unchanged so search guidance remains deterministic.
+      if (earlyCycle && hasPositivePredecessorCycle(interval)) return true;
     }
     return true;
   };
@@ -104,7 +129,7 @@ mlir::FailureOr<PeriodicDependencyEstimate> analyzePeriodicDependencies(
     error = "periodic timing exceeds one-nanosecond floating-point precision";
     return mlir::failure();
   }
-  if (!infeasible(result.recurrenceLowerNs))
+  if (!infeasible(result.recurrenceLowerNs, false))
     return result;
 
   // Recover witness cycles from the last longest-path predecessor forest.

src/PIM/Dialect/Spatial/Passes/Transforms/MergeComputeNodes/Scheduling/PipelineScheduling.cpp

@@ -5,6 +5,7 @@
 
 #include <algorithm>
 #include <cmath>
+#include <chrono>
 #include <limits>
 #include <numeric>
 #include <optional>
@@ -71,6 +72,7 @@ struct PipelineGroup {
 struct PipelineStageAssignment {
   std::vector<size_t> taskStages;
   std::vector<size_t> stageSizes;
+  std::vector<size_t> constructivePacks;
 };
 
 static bool consumesPipelineInput(const ComputeGraphNode &node) {
@@ -229,7 +231,7 @@ static FailureOr<PipelineStageAssignment> initialPipelinePartition(
     std::string &error) {
   size_t count = constraints.groups.size(), stages = layout.getStageCount();
   if (!count) return PipelineStageAssignment{{},
-    {layout.getStageSizes().begin(), layout.getStageSizes().end()}};
+    {layout.getStageSizes().begin(), layout.getStageSizes().end()}, {}};
   size_t extra = layout.getProcessorCount() - stages;
   // Resource quantization bounds seed construction only. Refinement uses
   // individual physical cores and final allocation uses every available core.
@@ -259,9 +261,9 @@ static FailureOr<PipelineStageAssignment> initialPipelinePartition(
       for (size_t amount = 0; amount <= units; ++amount) {
         size_t cores = coreCount(amount);
         if (weights.count() > checkedMultiply(cores, residentWeightCapacity)) continue;
-        costs[start][end][amount] = checkedAdd(span - largest,
+        if (auto packing = model.weights.estimatePackCost(tasks,
-          estimateOrderedPackCost(model.schedulingCosts, model.weights.taskWeightIds,
+              model.schedulingCosts, residentWeightCapacity, cores))
-            model.weights.weightCount, tasks, residentWeightCapacity, cores));
+          costs[start][end][amount] = checkedAdd(span - largest, *packing);
       }
     }
   }
@@ -290,8 +292,53 @@ static FailureOr<PipelineStageAssignment> initialPipelinePartition(
         && (!selected || best[stage][count][units].priority
           < best[selected][count][units].priority))
       selected = stage;
+  // Quantization controls quality, not feasibility. Complete the same prefix
+  // DP with minimum constructive core counts if no quantized state survives.
+  std::vector<std::vector<Cell>> feasible;
+  if (!selected) {
+    feasible.resize(stages + 1, std::vector<Cell>(count + 1));
+    feasible[0][0].valid = true;
+    std::vector<std::vector<size_t>> minimum(count + 1,
+      std::vector<size_t>(count + 1, std::numeric_limits<size_t>::max()));
+    for (size_t start = 0; start < count; ++start) {
+      TaskList tasks;
+      for (size_t end = start + 1; end <= count; ++end) {
+        llvm::append_range(tasks, constraints.groups[constraints.groupOrder[end - 1]].tasks);
+        const size_t lower = std::max<size_t>(1,
+          model.weights.getPackCountLowerBound(tasks, residentWeightCapacity));
+        const size_t maximum = layout.getProcessorCount() - stages + 1;
+        for (size_t cores = lower; cores <= maximum; ++cores)
+          if (model.weights.estimatePackCost(tasks, model.schedulingCosts,
+                residentWeightCapacity, cores)) {
+            minimum[start][end] = cores;
+            break;
+          }
+      }
+    }
+    for (size_t stage = 0; stage < stages; ++stage)
+      for (size_t start = 0; start < count; ++start) {
+        if (!feasible[stage][start].valid) continue;
+        for (size_t end = start + 1; end <= count; ++end) {
+          if (stage == 0 && end < constraints.minimumStageZeroEnd) continue;
+          if (stage && constraints.hasCrossingEdge
+              && constraints.furthestSuccessorBefore[start] >= end) continue;
+          const size_t cores = minimum[start][end];
+          if (cores > layout.getProcessorCount()) continue;
+          Cost used = checkedAdd(feasible[stage][start].priority, static_cast<Cost>(cores));
+          if (used > layout.getProcessorCount() - (stages - stage - 1)) continue;
+          auto &current = feasible[stage + 1][end];
+          if (!current.valid || used < current.priority)
+            current = {used, start, cores, true};
+        }
+      }
+    for (size_t stage = 1; stage <= stages; ++stage)
+      if (feasible[stage][count].valid
+          && (!selected || feasible[stage][count].priority < feasible[selected][count].priority))
+        selected = stage;
+  }
   if (!selected && count) {
-    error = "pipeline initialization found no dependency-adjacent operation partition within its bounded core-budget search";
+    error = "pipeline initialization found no constructively packable dependency-adjacent partition; "
+            "bounded packing failure does not prove physical infeasibility";
     return failure();
   }
   PipelineStageAssignment result;
@@ -299,13 +346,31 @@ static FailureOr<PipelineStageAssignment> initialPipelinePartition(
   result.taskStages.resize(constraints.taskToGroup.size());
   size_t end = count, remaining = units;
   for (size_t stage = selected; stage; --stage) {
-    const auto &cell = best[stage][end][remaining];
+    const auto &cell = feasible.empty() ? best[stage][end][remaining] : feasible[stage][end];
-    result.stageSizes[stage - 1] = coreCount(cell.units);
+    result.stageSizes[stage - 1] = feasible.empty() ? coreCount(cell.units) : cell.units;
     for (size_t position = cell.start; position < end; ++position)
       for (size_t task : constraints.groups[constraints.groupOrder[position]].tasks)
         result.taskStages[task] = stage - 1;
     end = cell.start;
-    remaining -= cell.units;
+    if (feasible.empty()) remaining -= cell.units;
+  }
+  // Preserve the constructive witness before adding spare cores. Reordering
+  // tasks for topology-aware placement must not lose DP feasibility evidence.
+  result.constructivePacks.resize(model.schedulingCosts.size());
+  std::vector<TaskList> stageTasks(selected);
+  for (size_t group : constraints.groupOrder)
+    for (size_t task : constraints.groups[group].tasks)
+      stageTasks[result.taskStages[task]].push_back(task);
+  for (size_t stage = 0; stage < selected; ++stage) {
+    const auto &tasks = stageTasks[stage];
+    TaskList packs;
+    if (!model.weights.estimatePackCost(tasks, model.schedulingCosts,
+          residentWeightCapacity, result.stageSizes[stage], &packs)) {
+      error = "pipeline initializer could not reconstruct its constructive packing witness";
+      return failure();
+    }
+    for (auto [index, task] : llvm::enumerate(tasks))
+      result.constructivePacks[task] = packs[index];
   }
   size_t assigned = std::accumulate(result.stageSizes.begin(), result.stageSizes.end(), size_t{0});
   for (; assigned < layout.getProcessorCount(); ++assigned) {
@@ -431,7 +496,7 @@ static LogicalResult packPipelineStages(
     std::vector<TaskList> &tasksByCpu, const PipelineCoreLayout &layout,
     size_t pipelineStages, size_t residentWeightCapacity,
     const SchedulingTarget &target, size_t &failedStage,
-    std::string &error) {
+    std::string &error, ArrayRef<size_t> constructivePacks) {
   std::vector<size_t> indegree(graph.nodes.size());
   std::vector<TaskList> successors(graph.nodes.size());
   for (size_t task = 0; task < graph.nodes.size(); ++task)
@@ -493,13 +558,11 @@ static LogicalResult packPipelineStages(
       llvm::sort(tasks, [&](size_t lhs, size_t rhs) {
         return topologicalPosition[lhs] < topologicalPosition[rhs];
       });
-      estimateOrderedPackCost(model.schedulingCosts, model.weights.taskWeightIds,
+      model.weights.estimatePackCost(tasks, model.schedulingCosts,
-        model.weights.weightCount, tasks, residentWeightCapacity, range.size, &taskPacks);
+        residentWeightCapacity, range.size, &taskPacks);
-      if (taskPacks.size() != tasks.size()) {
+      if (taskPacks.size() != tasks.size() && !constructivePacks.empty()) {
-        auto assignment = model.weights.pack(
+        taskPacks.clear();
-          tasks, model.schedulingCosts, residentWeightCapacity, range.size);
+        for (size_t task : tasks) taskPacks.push_back(constructivePacks[task]);
-        if (assignment)
-          taskPacks = std::move(*assignment);
       }
       packed = taskPacks.size() == tasks.size() && packPipelineStage(
         graph, model.schedulingCosts, tasksByCpu,
@@ -508,9 +571,9 @@ static LogicalResult packPipelineStages(
     }
     if (!packed) {
       failedStage = stage;
-      error = "pipeline resident-weight heuristic search could not pack stage "
+      error = "pipeline constructive physical-placement search could not pack stage "
             + std::to_string(stage)
-            + " within its physical core budget";
+            + " within its physical core budget; bounded search failure does not prove physical infeasibility";
       return failure();
     }
   }
@@ -633,7 +696,7 @@ static LogicalResult packPipelineAssignment(const ComputeGraph &graph,
     if (succeeded(packPipelineStages(
           graph, taskModel, tasksByPhysicalCpu, candidateLayout,
           pipelineStages, physicalTarget.residentWeightCapacity,
-          physicalTarget, failedStage, packingError))) {
+          physicalTarget, failedStage, packingError, assignment.constructivePacks))) {
       packed = true;
       break;
     }
@@ -651,6 +714,7 @@ static LogicalResult packPipelineAssignment(const ComputeGraph &graph,
         donor = stage;
     if (!donor)
       break;
+    assignment.constructivePacks.clear();
     --assignment.stageSizes[*donor];
     ++assignment.stageSizes[failedStage];
   }
@@ -748,24 +812,10 @@ static std::optional<PipelineStageAssignment> allocatePipelineStageCores(
       largest = std::max(largest, model.schedulingCosts[task]);
     }
     for (size_t cores = minimum; cores <= target.processorCount; ++cores) {
-      TaskList packs;
       Cost packedCost = std::max(largest, work[stage] / cores + (work[stage] % cores != 0));
       if (constructPacks)
-        packedCost = estimateOrderedPackCost(model.schedulingCosts,
+        packedCost = model.weights.estimatePackCost(tasks[stage], model.schedulingCosts,
-          model.weights.taskWeightIds, model.weights.weightCount, tasks[stage],
+          target.residentWeightCapacity, cores).value_or(std::numeric_limits<Cost>::max());
-          target.residentWeightCapacity, cores, &packs);
-      if (constructPacks && packs.size() != tasks[stage].size()) {
-        auto constructive = model.weights.pack(tasks[stage], model.schedulingCosts,
-          target.residentWeightCapacity, cores);
-        packedCost = std::numeric_limits<Cost>::max();
-        if (constructive) {
-          TaskCosts loads(cores);
-          for (size_t i = 0; i < tasks[stage].size(); ++i)
-            loads[(*constructive)[i]] = checkedAdd(loads[(*constructive)[i]],
-              model.schedulingCosts[tasks[stage][i]]);
-          packedCost = *llvm::max_element(loads);
-        }
-      }
       // A feasible assignment remains feasible when extra cores stay idle.
       if (cores > minimum) packedCost = std::min(packedCost, costs[stage][cores - 1]);
       costs[stage][cores] = packedCost;
@@ -1060,9 +1110,8 @@ static std::vector<PipelineProposal> pipelineNeighborhood(const ComputeGraph &gr
     auto packingCost = [&](const TaskList &tasks, size_t cores, size_t stage) {
       auto *cached = tasks == originalView.tasks[stage] ? &packCosts[stage][cores] : nullptr;
       if (cached && *cached) return **cached;
-      Cost value = estimateOrderedPackCost(model.schedulingCosts,
+      Cost value = model.weights.estimatePackCost(tasks, model.schedulingCosts,
-        model.weights.taskWeightIds, model.weights.weightCount, tasks,
+        target.residentWeightCapacity, cores).value_or(std::numeric_limits<Cost>::max());
-        target.residentWeightCapacity, cores);
       if (cached) *cached = value;
       return value;
     };
@@ -1093,6 +1142,7 @@ static std::vector<PipelineProposal> pipelineNeighborhood(const ComputeGraph &gr
         if (model.weights.getPackCountLowerBound(view.tasks[stage],
               target.residentWeightCapacity) > assignment.stageSizes[stage]) return;
         Cost work = packingCost(view.tasks[stage], assignment.stageSizes[stage], stage);
+        if (work == std::numeric_limits<Cost>::max()) return;
         maximum = std::max(maximum, work);
         // A within-generation span is latency, not an initiation-interval bound.
         total = checkedAdd(total, view.spans[stage]);
@@ -1174,7 +1224,10 @@ mlir::FailureOr<PipelineSearchResult> optimizePipelineSchedule(
   std::set<PipelineMove> visited;
   std::set<TaskList> placements{placementIdentity(graph, initial)};
   std::vector<size_t> beam{0};
+  auto neighborhoodStart = std::chrono::steady_clock::now();
   auto proposals = pipelineNeighborhood(graph, executionModel, *groups, result, beam, visited, 2 * evaluationBudget);
+  result.neighborhoodSeconds = std::chrono::duration<double>(
+    std::chrono::steady_clock::now() - neighborhoodStart).count();
   if (proposals.empty() || evaluationBudget == 1) {
     result.termination = "one neighborhood schedule; queue comparison skipped";
     return result;
@@ -1197,7 +1250,10 @@ mlir::FailureOr<PipelineSearchResult> optimizePipelineSchedule(
   if (!error.empty()) return failure();
   constexpr size_t maximumPackingAttempts = 4 * kPipelineQueueEvaluationBudget;
   while (result.costs.size() < evaluationBudget && result.packingAttempts < maximumPackingAttempts) {
+    neighborhoodStart = std::chrono::steady_clock::now();
     proposals = pipelineNeighborhood(graph, executionModel, *groups, result, beam, visited, 2 * evaluationBudget);
+    result.neighborhoodSeconds += std::chrono::duration<double>(
+      std::chrono::steady_clock::now() - neighborhoodStart).count();
     if (proposals.empty()) break;
     ++result.rounds;
     size_t evaluated = 0;
@@ -1211,17 +1267,25 @@ mlir::FailureOr<PipelineSearchResult> optimizePipelineSchedule(
       ++result.packingAttempts;
       visited.emplace(proposal.parent, proposal.assignment.taskStages,
         proposal.assignment.stageSizes, proposal.optimizeCores, proposal.adjustableStages);
+      auto packingStart = std::chrono::steady_clock::now();
       auto assignment = proposal.assignment;
       if (proposal.optimizeCores) {
         auto allocated = allocatePipelineStageCores(graph, executionModel, assignment.taskStages,
           pipelineStages, target, assignment.stageSizes, proposal.adjustableStages);
-        if (!allocated) continue;
+        if (!allocated) {
+          result.packingSeconds += std::chrono::duration<double>(
+            std::chrono::steady_clock::now() - packingStart).count();
+          continue;
+        }
         assignment = std::move(*allocated);
       }
       MergeScheduleResult schedule;
       std::string packingError;
-      if (failed(packPipelineAssignment(graph, model, assignment, schedule,
+      auto packed = packPipelineAssignment(graph, model, assignment, schedule,
-            pipelineStages, target, packingError))) continue;
+        pipelineStages, target, packingError);
+      result.packingSeconds += std::chrono::duration<double>(
+        std::chrono::steady_clock::now() - packingStart).count();
+      if (failed(packed)) continue;
       if (!placements.insert(placementIdentity(graph, schedule)).second) continue;
       auto estimate = cost(schedule);
       if (!error.empty()) return failure();
@@ -1261,7 +1325,7 @@ mlir::LogicalResult packPipelineSchedule(const ComputeGraph &graph,
   }
   PipelineTaskModel model(graph, target);
   return packPipelineAssignment(graph, model,
-    {{taskStages.begin(), taskStages.end()}, {stageSizes.begin(), stageSizes.end()}},
+    {{taskStages.begin(), taskStages.end()}, {stageSizes.begin(), stageSizes.end()}, {}},
     schedule, stageSizes.size(), target, error);
 }
 

src/PIM/Dialect/Spatial/Passes/Transforms/MergeComputeNodes/Scheduling/PipelineScheduling.hpp

@@ -114,6 +114,7 @@ struct PipelineSearchResult {
   std::vector<WholeScheduleCost> costs;
   std::vector<size_t> parents;
   size_t rounds = 0, packingAttempts = 0, consideredAssignments = 0;
+  double neighborhoodSeconds = 0, packingSeconds = 0;
   std::string termination;
 };
 

src/PIM/Dialect/Spatial/Passes/Transforms/MergeComputeNodes/Scheduling/ResidentWeightPacking.hpp

@@ -11,15 +11,18 @@
 namespace onnx_mlir::spatial {
 
 // Greedy maximal prefixes minimize ordered pack count: both constraints are
-// closed under taking subranges. The optimum cost cannot increase with packs.
+// closed under taking subranges. Feasible optimum costs cannot increase with packs.
-inline Cost estimateOrderedPackCost(
+inline std::optional<Cost> estimateOrderedPackCost(
     llvm::ArrayRef<Cost> taskCosts,
     const std::vector<std::vector<size_t>> &taskWeightIds, size_t weightCount,
     llvm::ArrayRef<size_t> tasks, size_t residentWeightCapacity,
     size_t maximumPacks, std::vector<size_t> *taskPacks = nullptr) {
+  if (taskPacks) taskPacks->clear();
+  if (!maximumPacks) return std::nullopt;
   Cost low = 0;
   Cost high = 0;
   for (size_t task : tasks) {
+    if (taskWeightIds[task].size() > residentWeightCapacity) return std::nullopt;
     low = std::max(low, taskCosts[task]);
     high = checkedAdd(high, taskCosts[task]);
   }
@@ -65,6 +68,7 @@ inline Cost estimateOrderedPackCost(
     }
     return fits;
   };
+  if (!pack(high, nullptr)) return std::nullopt;
   while (low < high) {
     Cost middle = low + (high - low) / 2;
     if (pack(middle, nullptr))
@@ -73,9 +77,7 @@ inline Cost estimateOrderedPackCost(
       low = middle + 1;
   }
   if (taskPacks) {
-    taskPacks->clear();
+    pack(low, taskPacks);
-    if (!pack(low, taskPacks))
-      taskPacks->clear();
   }
   return low;
 }
@@ -113,6 +115,22 @@ struct IndexedResidentWeights {
     return std::max<size_t>(!tasks.empty(), count / capacity + (count % capacity != 0));
   }
 
+  std::optional<Cost> estimatePackCost(llvm::ArrayRef<size_t> tasks,
+      llvm::ArrayRef<Cost> costs, size_t capacity, size_t packCount,
+      std::vector<size_t> *taskPacks = nullptr) const {
+    if (taskPacks) taskPacks->clear();
+    if (!packCount) return std::nullopt;
+    if (auto ordered = estimateOrderedPackCost(costs, taskWeightIds,
+          weightCount, tasks, capacity, packCount, taskPacks)) return ordered;
+    auto assignment = pack(tasks, costs, capacity, packCount);
+    if (!assignment) return std::nullopt;
+    std::vector<Cost> loads(packCount);
+    for (auto [index, task] : llvm::enumerate(tasks))
+      loads[(*assignment)[index]] = checkedAdd(loads[(*assignment)[index]], costs[task]);
+    if (taskPacks) *taskPacks = std::move(*assignment);
+    return *llvm::max_element(loads);
+  }
+
   // A constructive reuse-first best fit; failure is not a proof of infeasibility.
   std::optional<std::vector<size_t>> pack(
       llvm::ArrayRef<size_t> tasks, llvm::ArrayRef<Cost> costs,

src/PIM/Dialect/Spatial/Passes/Transforms/MergeComputeNodes/Scheduling/SchedulingTarget.hpp

@@ -67,6 +67,7 @@ struct WholeScheduleCost {
   std::vector<double> processorTaskWorkNs, processorCriticalDelayNs;
   struct AnalysisTiming {
     double extractionSeconds = 0, expansionSeconds = 0, evaluationSeconds = 0;
+    double dependencySeconds = 0;
   } timing;
   std::optional<bool> convergenceProved() const;
   std::optional<bool> loweringSupported() const;

test/PIM/CMakeLists.txt

@@ -69,6 +69,7 @@ add_pim_unittest(SpatialSchedulingTargetTest
 
 add_pim_unittest(PeriodicDependenciesTest
   PeriodicDependenciesTest.cpp
+  PeriodicDependenciesReference.cpp
 
   LINK_LIBS PRIVATE
   OMPimCompilerUtils

test/PIM/PeriodicDependenciesReference.cpp

@@ -0,0 +1,142 @@
+// Test-only reference from commit 3c7679594beb29ba9497aaf3e341cee74fa0be50.
+#include "src/Accelerators/PIM/Dialect/Spatial/Passes/Transforms/MergeComputeNodes/Scheduling/PeriodicDependencies.hpp"
+
+#include <limits>
+
+namespace onnx_mlir::spatial {
+
+mlir::FailureOr<PeriodicDependencyEstimate> analyzeFrozenPeriodicDependencies(
+    size_t nodeCount, llvm::ArrayRef<PeriodicDependency> edges,
+    std::string &error) {
+  std::vector<std::vector<size_t>> outgoing(nodeCount);
+  std::vector<size_t> indegree(nodeCount), recurrent;
+  Cost upper = 0;
+  for (auto [index, edge] : llvm::enumerate(edges)) {
+    if (edge.source >= nodeCount || edge.target >= nodeCount) {
+      error = "periodic dependency references an invalid event";
+      return mlir::failure();
+    }
+    if (edge.delayNs > std::numeric_limits<Cost>::max() - upper) {
+      error = "periodic delay sum exceeds the timing range";
+      return mlir::failure();
+    }
+    upper += edge.delayNs;
+    if (edge.distance) {
+      recurrent.push_back(index);
+    } else {
+      outgoing[edge.source].push_back(index);
+      ++indegree[edge.target];
+    }
+  }
+  std::vector<size_t> order;
+  for (size_t node = 0; node < nodeCount; ++node)
+    if (!indegree[node])
+      order.push_back(node);
+  for (size_t cursor = 0; cursor < order.size(); ++cursor)
+    for (size_t index : outgoing[order[cursor]])
+      if (!--indegree[edges[index].target])
+        order.push_back(edges[index].target);
+  if (order.size() != nodeCount) {
+    error = "periodic distance-zero dependencies contain a cycle";
+    return mlir::failure();
+  }
+
+  PeriodicDependencyEstimate result;
+  std::vector<Cost> span(nodeCount);
+  for (size_t node : order) {
+    result.withinGenerationSpanNs =
+      std::max(result.withinGenerationSpanNs, span[node]);
+    for (size_t index : outgoing[node]) {
+      const auto &edge = edges[index];
+      span[edge.target] = std::max(span[edge.target],
+        checkedAdd(span[node], edge.delayNs));
+    }
+  }
+  if (recurrent.empty() || !upper)
+    return result;
+
+  const size_t absent = std::numeric_limits<size_t>::max();
+  std::vector<double> potential(nodeCount);
+  std::vector<size_t> predecessor(nodeCount, absent);
+  auto infeasible = [&](double interval) {
+    std::fill(potential.begin(), potential.end(), 0);
+    std::fill(predecessor.begin(), predecessor.end(), absent);
+    auto relax = [&](size_t index) {
+      const auto &edge = edges[index];
+      double value = potential[edge.source] + edge.delayNs
+                   - interval * edge.distance;
+      if (value <= potential[edge.target] + 1e-7)
+        return false;
+      potential[edge.target] = value;
+      predecessor[edge.target] = index;
+      return true;
+    };
+    for (size_t node : order)
+      for (size_t index : outgoing[node])
+        relax(index);
+    // Eliminate distance-zero paths by a DAG sweep. A simple residual path
+    // visits at most R recurrent edges, so R+1 rounds detect a positive cycle.
+    for (size_t round = 0; round <= recurrent.size(); ++round) {
+      ++result.relaxationRounds;
+      bool changed = false;
+      for (size_t index : recurrent)
+        changed |= relax(index);
+      for (size_t node : order)
+        for (size_t index : outgoing[node])
+          changed |= relax(index);
+      if (!changed)
+        return false;
+    }
+    return true;
+  };
+  result.recurrenceUpperNs = upper;
+  for (unsigned iteration = 0;
+       iteration < 64 && result.recurrenceUpperNs - result.recurrenceLowerNs > 1;
+       ++iteration) {
+    double middle = (result.recurrenceLowerNs + result.recurrenceUpperNs) / 2;
+    if (middle == result.recurrenceLowerNs || middle == result.recurrenceUpperNs)
+      break;
+    if (infeasible(middle))
+      result.recurrenceLowerNs = middle;
+    else
+      result.recurrenceUpperNs = middle;
+  }
+  if (result.recurrenceUpperNs - result.recurrenceLowerNs > 1) {
+    error = "periodic timing exceeds one-nanosecond floating-point precision";
+    return mlir::failure();
+  }
+  if (!infeasible(result.recurrenceLowerNs))
+    return result;
+
+  // Recover witness cycles from the last longest-path predecessor forest.
+  std::vector<size_t> seen(nodeCount, absent);
+  double best = 0;
+  for (size_t start = 0; start < nodeCount; ++start) {
+    size_t node = start;
+    while (seen[node] == absent && predecessor[node] != absent) {
+      seen[node] = start;
+      node = edges[predecessor[node]].source;
+    }
+    if (seen[node] != start || predecessor[node] == absent)
+      continue;
+    size_t first = node;
+    Cost delay = 0, distance = 0;
+    std::vector<size_t> cycle;
+    do {
+      size_t index = predecessor[node];
+      cycle.push_back(index);
+      delay = checkedAdd(delay, edges[index].delayNs);
+      distance = checkedAdd(distance, static_cast<Cost>(edges[index].distance));
+      node = edges[index].source;
+    } while (node != first);
+    double ratio = distance ? static_cast<double>(delay) / distance : 0;
+    if (ratio > best) {
+      best = ratio;
+      std::reverse(cycle.begin(), cycle.end());
+      result.criticalCycle = std::move(cycle);
+    }
+  }
+  return result;
+}
+
+} // namespace onnx_mlir::spatial

test/PIM/PeriodicDependenciesTest.cpp

@@ -5,11 +5,96 @@
 #error "Periodic dependency tests require assertions"
 #endif
 #include <cmath>
+#include <chrono>
+#include <iostream>
 #include <functional>
 #include <random>
 
 using namespace onnx_mlir::spatial;
 
+namespace onnx_mlir::spatial {
+mlir::FailureOr<PeriodicDependencyEstimate> analyzeFrozenPeriodicDependencies(
+    size_t, llvm::ArrayRef<PeriodicDependency>, std::string &);
+}
+
+static size_t differentialCases = 0, acceleratedRounds = 0, frozenRounds = 0;
+static double acceleratedSeconds = 0, frozenSeconds = 0;
+
+static void differential(size_t count, const std::vector<PeriodicDependency> &edges) {
+  std::string actualError, referenceError;
+  auto start = std::chrono::steady_clock::now();
+  auto actual = analyzePeriodicDependencies(count, edges, actualError);
+  acceleratedSeconds += std::chrono::duration<double>(std::chrono::steady_clock::now() - start).count();
+  start = std::chrono::steady_clock::now();
+  auto reference = analyzeFrozenPeriodicDependencies(count, edges, referenceError);
+  frozenSeconds += std::chrono::duration<double>(std::chrono::steady_clock::now() - start).count();
+  ++differentialCases;
+  assert(mlir::succeeded(actual) == mlir::succeeded(reference));
+  assert(actualError == referenceError);
+  if (mlir::failed(actual)) return;
+  acceleratedRounds += actual->relaxationRounds;
+  frozenRounds += reference->relaxationRounds;
+  assert(actual->recurrenceLowerNs == reference->recurrenceLowerNs);
+  assert(actual->recurrenceUpperNs == reference->recurrenceUpperNs);
+  assert(actual->withinGenerationSpanNs == reference->withinGenerationSpanNs);
+  assert(actual->criticalCycle == reference->criticalCycle);
+}
+
+static void testFrozenDifferential() {
+  std::mt19937 random(171221);
+  for (size_t sample = 0; sample < 800; ++sample) {
+    const size_t count = sample < 600 ? 2 + random() % 31 : 64 + random() % 193;
+    std::vector<PeriodicDependency> edges;
+    for (size_t source = 0; source < count; ++source) {
+      const size_t component = source / 32;
+      for (size_t link = 0; link < 4; ++link) {
+        size_t target = random() % count;
+        if (sample % 3 == 0 && target / 32 != component) continue;
+        unsigned distance = target > source && link % 2 == 0 ? 0 : 1 + random() % 7;
+        edges.push_back({source, target, random() % 100001, distance});
+      }
+    }
+    differential(count, edges);
+  }
+  for (size_t sample = 0; sample < 100; ++sample) {
+    std::vector<PeriodicDependency> edges;
+    for (size_t node = 0; node < 32; ++node) {
+      edges.push_back({node, node, (Cost{1} << 40) + random() % 10001,
+        static_cast<unsigned>(1024 + random() % 7)});
+      edges.push_back({node, (node + 1) % 32, random() % 101, 1});
+    }
+    differential(32, edges);
+  }
+  differential(2, {{0,1,1,0}, {1,0,1000000000000ULL, std::numeric_limits<unsigned>::max()}});
+  for (size_t count : {2u, 64u, 1024u}) {
+    std::vector<PeriodicDependency> edges;
+    for (size_t node = 0; node < count; ++node) {
+      edges.push_back({node, node, 1000000 + node % 2, 3});
+      if (node + 1 < count) edges.push_back({node, node + 1, 1, 0});
+    }
+    differential(count, edges);
+  }
+  std::vector<PeriodicDependency> large;
+  constexpr size_t componentSize = 2048, components = 32;
+  for (size_t component = 0; component < components; ++component) {
+    size_t first = component * componentSize;
+    for (size_t node = first; node + 1 < first + componentSize; ++node)
+      large.push_back({node, node + 1, 1, 0});
+    large.push_back({first + componentSize - 1, first, 17 + component,
+      static_cast<unsigned>(1 + component % 7)});
+  }
+  differential(componentSize * components, large);
+  differential(0, {});
+  differential(2, {{0, 1, 0, 0}, {1, 0, 0, 0}});
+  differential(2, {{0, 1, 10, 0}, {1, 0, 3, 0}});
+  differential(1, {{0, 1, 1, 1}});
+  differential(1, {{0, 0, Cost{1} << 60, 1}});
+  differential(1, {{0, 0, std::numeric_limits<Cost>::max(), 1}, {0, 0, 1, 1}});
+  std::cout << "frozen_differential_cases=" << differentialCases
+    << " accelerated_rounds=" << acceleratedRounds << " frozen_rounds=" << frozenRounds
+    << " accelerated_seconds=" << acceleratedSeconds << " frozen_seconds=" << frozenSeconds << '\n';
+}
+
 static double enumerateCycles(size_t count,
                              const std::vector<PeriodicDependency> &edges) {
   double maximum = 0;
@@ -59,11 +144,22 @@ static void check(size_t count, const std::vector<PeriodicDependency> &edges,
 }
 
 int main() {
+  testFrozenDifferential();
   check(3, {{0, 1, 12, 0}, {1, 2, 7, 0}}, 0);
   check(3, {{0, 1, 12, 1}, {1, 2, 7, 1}}, 0);
   check(1, {{0, 0, 30, 2}}, 15);
   // Independent core periods remain parallel.
   check(2, {{0, 0, 10, 1}, {1, 1, 20, 1}}, 20);
+  // Positive-cycle checks must not change the optimum or the final witness.
+  std::vector<PeriodicDependency> independent;
+  for (size_t core = 0; core < 128; ++core)
+    independent.push_back({core, core, 1000 + core, 1});
+  check(128, independent, 1127);
+  std::string performanceError;
+  auto accelerated = analyzePeriodicDependencies(128, independent, performanceError);
+  assert(mlir::succeeded(accelerated));
+  assert(accelerated->criticalCycle == std::vector<size_t>{127});
+  assert(accelerated->relaxationRounds < 512);
   // Equal serial demand, different cross-core blocking recurrences.
   check(4, {{0, 1, 10, 0}, {2, 3, 10, 0},
             {1, 0, 0, 1}, {3, 2, 0, 1}}, 10);

test/PIM/PimAbstractExecutionTest.cpp

@@ -243,6 +243,17 @@ static void testSharedEngine(PimInstructionTiming target) {
   actual = analyzePimExecution(abstract, target);
   TEST_CHECK(actual.getInteger("global_read_requests") == 2);
   TEST_CHECK(actual.getInteger("global_memory_service_ns") == 160);
+  // The construction limit applies to interpreted patterns, not just traffic.
+  // These zero-trip request bodies do not allocate a large event graph.
+  abstract[0] = {{E::Work, 4}, {E::Repeat, 1500000, 0, 0, 0,
+    {{E::Repeat, 0, 0, 0, 0, {{E::Load, 1}}}}}};
+  auto largerPattern = evaluatePimExecution(abstract, target);
+  TEST_CHECK(largerPattern.cost.available());
+  TEST_CHECK(largerPattern.report.getInteger("global_read_requests") == 0);
+  abstract[0] = {{E::Repeat, 2000001, 0, 0, 0, {{E::Load, 1}}}};
+  auto aboveBudget = evaluatePimExecution(abstract, target);
+  TEST_CHECK(!aboveBudget.cost.available());
+  TEST_CHECK(aboveBudget.cost.reason == "abstract repetition budget exceeded");
   abstract[0] = {{E::Repeat, 1000000000000ull, 0, 0, 0, {{E::Load, 1}}}};
   actual = analyzePimExecution(abstract, target);
   TEST_CHECK(actual.getString("status") == "unavailable");
@@ -433,6 +444,48 @@ static void testDependencyTokens() {
   TEST_CHECK(drained.empty() && drained.remaining().empty());
 }
 
+static void testSpatialExtractionBudget(PimInstructionTiming target) {
+  MLIRContext context;
+  context.loadDialect<arith::ArithDialect, func::FuncDialect, scf::SCFDialect,
+    tensor::TensorDialect, spatial::SpatialDialect>();
+  auto module = ModuleOp::create(UnknownLoc::get(&context));
+  OpBuilder builder(module.getBodyRegion());
+  auto loc = module.getLoc();
+  auto function = func::FuncOp::create(builder, loc, "test", builder.getFunctionType({}, {}));
+  builder.setInsertionPointToStart(function.addEntryBlock());
+  auto batch = spatial::SpatScheduledComputeBatch::create(builder, loc, TypeRange{},
+    builder.getI32IntegerAttr(2), ValueRange{}, ValueRange{});
+  batch->setAttr("coreIds", builder.getDenseI32ArrayAttr({0, 1}));
+  batch.getBody().push_back(new Block);
+  batch.getBody().front().addArgument(builder.getIndexType(), loc);
+  builder.setInsertionPointToStart(&batch.getBody().front());
+  Value zero = arith::ConstantIndexOp::create(builder, loc, 0);
+  Value upper = arith::ConstantIndexOp::create(builder, loc, 100000);
+  Value one = arith::ConstantIndexOp::create(builder, loc, 1);
+  auto type = RankedTensorType::get({16}, builder.getF32Type());
+  Value local = tensor::EmptyOp::create(builder, loc, type.getShape(), type.getElementType());
+  spatial::SpatReluOp::create(builder, loc, type, local);
+  auto loop = scf::ForOp::create(builder, loc, zero, upper, one);
+  builder.setInsertionPointToStart(loop.getBody());
+  for (int64_t i = 0; i < 5; ++i) arith::ConstantIndexOp::create(builder, loc, i);
+  // A zero-count wait is runtime-free, but requires indexed interpretation.
+  spatial::SpatWaitOp::create(builder, loc, zero, zero);
+  builder.setInsertionPointAfter(loop);
+  spatial::SpatYieldOp::create(builder, loc, ValueRange{});
+  builder.setInsertionPointAfter(batch);
+  func::ReturnOp::create(builder, loc);
+  TEST_CHECK(succeeded(verify(module)));
+  auto supported = evaluatePimSpatialSchedule(module, target);
+  TEST_CHECK(supported.report.getInteger("visited_operations") > 1000000);
+  TEST_CHECK(supported.cost.available());
+  builder.setInsertionPointToStart(loop.getBody());
+  for (int64_t i = 0; i < 5; ++i) arith::ConstantIndexOp::create(builder, loc, i);
+  auto unsupported = evaluatePimSpatialSchedule(module, target);
+  TEST_CHECK(!unsupported.cost.available());
+  TEST_CHECK(unsupported.cost.reason == "abstract extraction budget exceeded on physical core 0");
+  module.erase();
+}
+
 static void testReadOnlySpatial(PimInstructionTiming target, bool batched) {
   MLIRContext context;
   context.loadDialect<arith::ArithDialect, func::FuncDialect, scf::SCFDialect,
@@ -918,6 +971,7 @@ int main() {
     testCapturedVectorWeights(hardware(period));
     testPublicationAndOutput(hardware(period));
     testUnrepresentableSpatial(hardware(period));
+    testSpatialExtractionBudget(hardware(period));
     testReadOnlySpatial(hardware(period), false);
     testReadOnlySpatial(hardware(period), true);
   }

test/PIM/PipelineOptimizerTest.cpp

@@ -3,6 +3,8 @@
 #include "src/Accelerators/PIM/Compiler/PimInstructionTiming.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/Passes/Transforms/MergeComputeNodes/Scheduling/PipelineScheduling.hpp"
 #include "src/Accelerators/PIM/Dialect/Spatial/Passes/Transforms/MergeComputeNodes/Scheduling/ScheduleEvaluation.hpp"
+#include "src/Accelerators/PIM/Dialect/Spatial/Passes/Transforms/MergeComputeNodes/Scheduling/ResidentWeightPacking.hpp"
+#include "src/Accelerators/PIM/Dialect/Spatial/SpatialOps.hpp"
 #include <chrono>
 #include <functional>
 #include <map>
@@ -14,6 +16,76 @@ using namespace onnx_mlir::spatial;
 using E = PimExecutionEvent;
 static bool sawJointMove = false, sawSuccessiveMoves = false;
 
+static void testPackingFeasibility() {
+  ComputeGraph graph;
+  for (size_t task = 0; task < 4; ++task) {
+    ResidentWeight weight;
+    weight.opaqueLane = task % 2;
+    graph.nodes.push_back({{nullptr, static_cast<uint32_t>(task), 1}, 10, {weight}, task});
+  }
+  IndexedResidentWeights weights(graph);
+  const std::vector<size_t> tasks{0, 1, 2, 3};
+  const std::vector<Cost> work{10, 10, 10, 10};
+  std::vector<size_t> assignment{99};
+  TEST_CHECK(!estimateOrderedPackCost(work, weights.taskWeightIds, weights.weightCount,
+    tasks, 1, 2, &assignment));
+  TEST_CHECK(assignment.empty());
+  TEST_CHECK(!estimateOrderedPackCost(work, weights.taskWeightIds, weights.weightCount,
+    tasks, 1, 0));
+  TEST_CHECK(weights.estimatePackCost(tasks, work, 1, 2) == 20);
+  auto placement = weights.pack(tasks, work, 1, 2);
+  TEST_CHECK(placement && (*placement)[0] == (*placement)[2]
+    && (*placement)[1] == (*placement)[3] && (*placement)[0] != (*placement)[1]);
+  TEST_CHECK(!weights.estimatePackCost(tasks, work, 1, 1));
+  for (size_t cores = 4; cores <= 8; ++cores)
+    TEST_CHECK(estimateOrderedPackCost(work, weights.taskWeightIds, weights.weightCount,
+      tasks, 1, cores) == 10);
+  weights.taskWeightIds[0] = {0, 1};
+  TEST_CHECK(!estimateOrderedPackCost(work, weights.taskWeightIds, weights.weightCount,
+    tasks, 1, 4));
+  weights.taskWeightIds = {{0, 1}, {2}, {0, 1}, {2}};
+  weights.weightCount = 3;
+  TEST_CHECK(!estimateOrderedPackCost(work, weights.taskWeightIds, weights.weightCount,
+    tasks, 2, 2));
+  TEST_CHECK(weights.estimatePackCost(tasks, work, 2, 2) == 20);
+}
+
+static void testQuantizedFeasibility(uint32_t first) {
+  mlir::MLIRContext context;
+  context.loadDialect<spatial::SpatialDialect>();
+  auto module = mlir::ModuleOp::create(mlir::UnknownLoc::get(&context));
+  mlir::OpBuilder builder(module.getBodyRegion());
+  ComputeGraph graph;
+  graph.predecessors.resize(40); graph.successors.resize(40);
+  for (uint32_t group = 0; group < 2; ++group) {
+    const uint32_t lanes = group ? 40 - first : first;
+    auto batch = spatial::SpatGraphComputeBatch::create(builder, module.getLoc(),
+      mlir::TypeRange{}, builder.getI32IntegerAttr(lanes), mlir::ValueRange{}, mlir::ValueRange{});
+    for (uint32_t lane = 0; lane < lanes; ++lane) {
+      size_t task = graph.nodes.size();
+      ResidentWeight weight; weight.opaqueLane = task;
+      graph.nodes.push_back({{batch, lane, 1}, 10, {weight}, task});
+      if (group) {
+        graph.predecessors[task].push_back({0, {0, 0}});
+        graph.successors[0].push_back({task, {0, 0}});
+      }
+    }
+  }
+  SchedulingTarget target;
+  target.processorCount = 40; target.residentWeightCapacity = 1;
+  target.interProcessorLatencyNs.assign(1600, 0);
+  MergeScheduleResult schedule;
+  std::string error;
+  TEST_CHECK(mlir::succeeded(runPipelineScheduler(graph, schedule, 2, target, error)));
+  std::vector<size_t> budgets(2);
+  for (size_t stage : schedule.processorStages) ++budgets[stage];
+  TEST_CHECK(budgets == std::vector<size_t>({first, 40 - first}));
+  for (size_t task = 0; task < graph.nodes.size(); ++task)
+    TEST_CHECK(schedule.processorStages[schedule.computeToCpuMap.lookup(graph.nodes[task].instance)]
+      == (task < first ? 0 : 1));
+  module.erase();
+}
+
 // Test-only workloads use stage-boundary host transfers and local direct links.
 // The same queue
 // engine and complete-state evidence rank reference and searched schedules.
@@ -112,6 +184,176 @@ static PimInstructionTiming hardware(size_t profile) {
   return target;
 }
 
+static void testInitializerPackingWitness() {
+  mlir::MLIRContext context;
+  context.loadDialect<spatial::SpatialDialect>();
+  auto module = mlir::ModuleOp::create(mlir::UnknownLoc::get(&context));
+  mlir::OpBuilder builder(module.getBodyRegion());
+  std::vector<mlir::Operation *> operations;
+  for (size_t i = 0; i < 3; ++i) operations.push_back(
+    spatial::SpatGraphComputeBatch::create(builder, module.getLoc(),
+      mlir::TypeRange{}, builder.getI32IntegerAttr(3), mlir::ValueRange{}, mlir::ValueRange{}));
+  const std::vector<std::vector<size_t>> resident{{2,1,5},{2,4},{0,2},{0},{4,6},{1},{2},{2},{1,4}};
+  const std::vector<Cost> costs{4,1,5,4,1,2,5,2,2};
+  ComputeGraph graph;
+  graph.predecessors.resize(9); graph.successors.resize(9);
+  for (uint32_t task = 0; task < 9; ++task) {
+    ResidentWeightSet weights;
+    for (size_t id : resident[task]) {
+      ResidentWeight w; w.opaqueLane = id; weights.push_back(w);
+    }
+    graph.nodes.push_back({{operations[task % 3], task / 3, 1}, costs[task], weights, task});
+  }
+  IndexedResidentWeights indexed(graph);
+  TEST_CHECK(!indexed.estimatePackCost({1,2,4,5,7,8}, costs, 3, 2));
+  std::vector<size_t> witness;
+  TEST_CHECK(indexed.estimatePackCost({1,4,7,2,5,8}, costs, 3, 2, &witness));
+  TEST_CHECK(witness.size() == 6);
+  auto target = hardware(1);
+  target.processorCount = 4; target.residentWeightCapacity = 3;
+  target.interProcessorLatencyNs.assign(16, 1);
+  MergeScheduleResult schedule;
+  std::string error;
+  TEST_CHECK(mlir::succeeded(runPipelineScheduler(graph, schedule, 2, target, error)));
+  std::vector<ResidentWeightSet> actual(4);
+  for (const auto &node : graph.nodes) {
+    size_t core = schedule.computeToCpuMap.lookup(node.instance);
+    TEST_CHECK(core < 4);
+    insertResidentWeights(actual[core], node.residentWeights);
+    TEST_CHECK(actual[core].size() <= 3);
+  }
+  // Exhaust all core assignments under the same stage map. The initialized
+  // assignment must belong to this physical-capacity reference set.
+  size_t feasible = 0;
+  for (size_t encoded = 0; encoded < (size_t{1} << 18); ++encoded) {
+    size_t choices = encoded;
+    std::vector<ResidentWeightSet> weights(4);
+    std::vector<size_t> groupStages(3, 2);
+    bool valid = true;
+    for (size_t task = 0; task < 9 && valid; ++task) {
+      size_t core = choices % 4; choices /= 4;
+      size_t stage = schedule.processorStages[core], group = task % 3;
+      if (groupStages[group] == 2) groupStages[group] = stage;
+      valid &= groupStages[group] == stage;
+      insertResidentWeights(weights[core], graph.nodes[task].residentWeights);
+      valid &= weights[core].size() <= 3;
+    }
+    feasible += valid;
+  }
+  TEST_CHECK(feasible > 0);
+  MergeScheduleResult alternative;
+  TEST_CHECK(mlir::succeeded(packPipelineSchedule(graph, std::vector<size_t>(9),
+    {3, 1}, target, alternative, error)));
+  std::cout << "initializer_physical_feasible_assignments=" << feasible << '\n';
+  module.erase();
+}
+
+static void testPhysicalReference() {
+  auto target = hardware(1);
+  target.processorCount = 3;
+  target.sharedMemoryRequestLatencyNs.resize(3);
+  target.sharedMemoryResponseLatencyNs.resize(3);
+  target.interProcessorLatencyNs = {0, 2, 11, 2, 0, 7, 11, 7, 0};
+  ComputeGraph graph;
+  graph.predecessors.resize(3); graph.successors.resize(3);
+  for (uint32_t task = 0; task < 3; ++task) {
+    ResidentWeight weight; weight.opaqueLane = task % 2;
+    graph.nodes.push_back({{nullptr, task, 1}, task == 1 ? 700u : 80u, {weight}, task});
+    if (task) {
+      graph.predecessors[task].push_back({task - 1, {64, 1}});
+      graph.successors[task - 1].push_back({task, {64, 1}});
+    }
+  }
+  double optimum = std::numeric_limits<double>::infinity();
+  size_t evaluations = 0;
+  for (size_t mask = 1; mask < 7; ++mask)
+    for (size_t placement = 0; placement < 27; ++placement) {
+      MergeScheduleResult schedule;
+      for (size_t core = 0; core < 3; ++core) schedule.processorStages.push_back((mask >> core) & 1);
+      size_t encoded = placement;
+      std::vector<size_t> stages, slots(3);
+      std::vector<ResidentWeightSet> resident(3);
+      bool valid = true;
+      for (const auto &node : graph.nodes) {
+        size_t core = encoded % 3; encoded /= 3;
+        schedule.computeToCpuMap[node.instance] = core;
+        schedule.computeToCpuSlotMap[node.instance] = slots[core]++;
+        stages.push_back(schedule.processorStages[core]);
+        insertResidentWeights(resident[core], node.residentWeights);
+        valid &= resident[core].size() <= target.residentWeightCapacity;
+      }
+      if (!valid || stages[0] != 0 || stages[1] < stages[0] || stages[2] < stages[1]) continue;
+      auto cost = queueCost(graph, schedule, target);
+      optimum = std::min(optimum, std::max(*cost.primary, *cost.alternative));
+      ++evaluations;
+    }
+  std::string error;
+  MergeScheduleResult initial;
+  TEST_CHECK(mlir::succeeded(runPipelineScheduler(graph, initial, 2, target, error)));
+  auto evaluate = [&](const MergeScheduleResult &schedule) { return queueCost(graph, schedule, target); };
+  auto search = optimizePipelineSchedule(graph, initial, 2, target, evaluate, error);
+  TEST_CHECK(mlir::succeeded(search));
+  size_t winner = search->costs.empty() ? 0 : selectPipelineCandidate(search->costs, error);
+  auto selected = evaluate(search->schedules[winner]);
+  double value = std::max(*selected.primary, *selected.alternative);
+  TEST_CHECK(value + 1e-6 >= optimum);
+  std::cout << "physical_reference_ns=" << optimum << " selected_ns=" << value
+    << " gap_percent=" << 100 * (value / optimum - 1)
+    << " physical_evaluations=" << evaluations
+    << " search_evaluations=" << search->costs.size()
+    << " packing_attempts=" << search->packingAttempts << '\n';
+}
+
+static void testWorseningIntermediate() {
+  // A controlled cost oracle isolates the search's non-monotone landscape.
+  // Hardware queue objectives are covered separately by runReference().
+  auto target = hardware(0);
+  target.processorCount = 8;
+  target.interProcessorLatencyNs.assign(64, 0);
+  ComputeGraph graph;
+  graph.predecessors.resize(4); graph.successors.resize(4);
+  for (uint32_t task = 0; task < 4; ++task)
+    graph.nodes.push_back({{nullptr, task, 1}, 10, {}, task});
+  auto evaluate = [&](const MergeScheduleResult &schedule) {
+    bool firstStage = llvm::all_of(graph.nodes, [&](const auto &node) {
+      return schedule.processorStages[schedule.computeToCpuMap.lookup(node.instance)] == 0;
+    });
+    size_t cores = llvm::count(schedule.processorStages, 0);
+    double value = !firstStage ? 120 : cores == 1 ? 80 : cores >= 6 ? 100 : 110;
+    WholeScheduleCost cost;
+    cost.primary = cost.alternative = value;
+    cost.loweringSupport = StaticLoweringSupport::Supported;
+    cost.selectionMargin = kPipelineSelectionMargin;
+    for (auto &proof : cost.recurrence)
+      proof = {ScheduleRecurrenceEvidence::Proved, 1, value, {}};
+    return cost;
+  };
+  std::string error;
+  MergeScheduleResult initial, intermediate;
+  TEST_CHECK(mlir::succeeded(runPipelineScheduler(graph, initial, 2, target, error)));
+  TEST_CHECK(evaluate(initial).primary == 100);
+  const std::vector<size_t> stages(4);
+  TEST_CHECK(mlir::succeeded(packPipelineSchedule(graph, stages, {3, 5}, target, intermediate, error)));
+  TEST_CHECK(evaluate(intermediate).primary == 110);
+  double reference = std::numeric_limits<double>::infinity();
+  for (size_t cores = 1; cores < 8; ++cores) {
+    MergeScheduleResult schedule;
+    TEST_CHECK(mlir::succeeded(packPipelineSchedule(graph, stages, {cores, 8 - cores}, target, schedule, error)));
+    reference = std::min(reference, *evaluate(schedule).primary);
+  }
+  TEST_CHECK(reference == 80);
+  auto search = optimizePipelineSchedule(graph, initial, 2, target, evaluate, error);
+  TEST_CHECK(mlir::succeeded(search));
+  size_t winner = selectPipelineCandidate(search->costs, error);
+  double selected = *search->costs[winner].primary;
+  TEST_CHECK(selected >= reference && selected <= 100);
+  TEST_CHECK(search->costs.size() <= kPipelineQueueEvaluationBudget);
+  std::cout << "non_monotone_oracle_reference=" << reference << " selected=" << selected
+    << " gap_percent=" << 100 * (selected / reference - 1)
+    << " evaluations=" << search->costs.size()
+    << " packing_attempts=" << search->packingAttempts << '\n';
+}
+
 static void runReference(size_t profile, size_t fixture) {
   auto target = hardware(profile);
   ComputeGraph graph;
@@ -217,8 +459,10 @@ static void runReference(size_t profile, size_t fixture) {
     << " gap_percent=" << 100 * (cost / reference - 1)
     << " best_evaluated_ns=" << bestEvaluated
     << " best_evaluated_gap_percent=" << 100 * (bestEvaluated / reference - 1)
+    << " reference_scope=logical_with_deterministic_packing"
     << " reference_evaluations=" << referenceEvaluations
-    << " search_evaluations=" << search->costs.size() << " refinement_depth=" << depth
+    << " search_evaluations=" << search->costs.size()
+    << " packing_attempts=" << search->packingAttempts << " refinement_depth=" << depth
     << " search_seconds=" << searchSeconds << " reference_seconds=" << referenceSeconds
     << " total_seconds=" << std::chrono::duration<double>(std::chrono::steady_clock::now() - begin).count()
     << '\n';
@@ -246,6 +490,11 @@ static void runReference(size_t profile, size_t fixture) {
 }
 
 int main() {
+  testInitializerPackingWitness();
+  testPackingFeasibility();
+  for (uint32_t first : {19u, 21u}) testQuantizedFeasibility(first);
+  testPhysicalReference();
+  testWorseningIntermediate();
   for (size_t profile = 0; profile < 2; ++profile)
     for (size_t fixture = 0; fixture < 6; ++fixture) runReference(profile, fixture);
   TEST_CHECK(sawJointMove && sawSuccessiveMoves);

test/PIM/SpatialSchedulingTargetTest.cpp

@@ -1642,14 +1642,14 @@ static void testOrderedPackCostCurve() {
         if (fits && packs <= cores) optimum = std::min(optimum, std::max(maximum, work));
       }
       std::vector<size_t> assignment;
-      Cost cost = estimateOrderedPackCost(costs, weights, 3, tasks, capacity, cores, &assignment);
+      auto cost = estimateOrderedPackCost(costs, weights, 3, tasks, capacity, cores, &assignment);
       if (optimum == std::numeric_limits<Cost>::max()) {
-        assert(assignment.empty());
+        assert(!cost && assignment.empty());
         continue;
       }
-      assert(cost == optimum && cost <= previous && assignment.size() == tasks.size());
+      assert(cost && *cost == optimum && *cost <= previous && assignment.size() == tasks.size());
-      plateau |= cost == previous;
+      plateau |= *cost == previous;
-      previous = cost;
+      previous = *cost;
       std::vector<Cost> loads(cores);
       std::vector<llvm::SmallBitVector> resident(cores, llvm::SmallBitVector(3));
       for (size_t task : tasks) {
@@ -1658,7 +1658,7 @@ static void testOrderedPackCostCurve() {
         for (size_t weight : weights[task]) resident[assignment[task]].set(weight);
       }
       for (size_t core = 0; core < cores; ++core) {
-        assert(loads[core] <= cost && resident[core].count() <= capacity);
+        assert(loads[core] <= *cost && resident[core].count() <= capacity);
       }
     }
     assert(plateau && previous == 11);

test/PIM/test_QueueSelection.py

@@ -78,7 +78,7 @@ def main():
     parser.add_argument("--compiler", required=True)
     args = parser.parse_args()
     compiler = str(Path(args.compiler).resolve())
-    model = REPO / "validation/networks/pimcomp_models/googlenet/googlenet-12-pimsim-nn.onnx"
+    model = REPO / "validation/networks/pimcomp_models/googlenet/googlenet-12-no-lrn.onnx"
     with tempfile.TemporaryDirectory(prefix="raptor-static-selection-") as temporary:
         root = Path(temporary)
         history = root / "validation/tools/pim/pimcomp/compare/results_comparison.csv"

validation/networks/pimcomp_models/README.md

@@ -11,11 +11,11 @@ operations, standalone commands, configuration layout, and GA/OpenMP controls.
 
 | Directory | Model / input | Source |
 |---|---|---|
-| `resnet18/` | ResNet-18 v1, `1x3x224x224` | [ONNX Model Zoo resnet18-v1-7](https://huggingface.co/onnxmodelzoo/resnet18-v1-7); symlink to `../resnet18/depth_68/resnet18_depth_68.onnx`. |
+| `resnet18/` | ResNet-18 v1, `1x3x224x224` | [ONNX Model Zoo resnet18-v1-7](https://huggingface.co/onnxmodelzoo/resnet18-v1-7); symlink to `../../resnet18/depth_68/resnet18_depth_68.onnx`. |
 | `resnet34/` | ResNet-34 v1, `1x3x224x224` | [ONNX Model Zoo resnet34-v1-7](https://huggingface.co/onnxmodelzoo/resnet34-v1-7); symbolic batch fixed to 1. |
 | `googlenet/` | GoogLeNet, `1x3x224x224` | [ONNX Model Zoo googlenet-12](https://huggingface.co/onnxmodelzoo/googlenet-12), with only its two LRN nodes bypassed. |
 | `vgg8/` | VGG-8 reconstruction, `1x1x28x28` | Six convolution and two fully connected layers with deterministic random weights. |
-| `yolo11n/` | YOLO11n detection, `1x3x640x640` | Uses `../yolo11n/depth_51/yolo11n_depth_51.onnx`; this directory retains the historical Softmax-free variant. |
+| `yolo11n/` | YOLO11n detection, `1x3x640x640` | [Ultralytics YOLO11n](https://docs.ultralytics.com/models/yolo11/); symlink to `../../yolo11n/depth_51/yolo11n_depth_51.onnx`. |
 
 The paper's checkpoints were not published. The three Model Zoo networks
 produce frontend graphs matching Pimcomp's bundled JSON graphs. VGG-8 is
@@ -29,15 +29,9 @@ Its terminal Softmax, output `prob_1`, weights, and other operations are retaine
 The native reference uses this same graph. VGG8 and both ResNet graphs have no
 removed operators; ResNet-34 fixes the symbolic batch size to one.
 
-Historical performance models change the original computation:
+The current YOLO comparison uses `yolo11n/yolo11n.onnx`,
-
-- `googlenet/googlenet-12-pimsim-nn.onnx` removes two LRN nodes and terminal
-  Softmax, which Pimcomp does not schedule.
-- `yolo11n/yolo11n-pimsim-nn.onnx` removes two internal Softmax nodes. It remains
-  available for reproducing historical measurements.
-
-The current YOLO comparison uses `../yolo11n/depth_51/yolo11n_depth_51.onnx`,
 including both Softmax operations. Its native reference uses the same full model.
+The checked-in ONNX graph uses opset 18. The original exporter version and checkpoint hash are not recorded.
 The functional simulator executes Softmax. Pimsim-nn skips `vsoftmax` before
 fetch, and Raptor's PIM queue estimator omits the same instruction cost.
 Timing and energy exclude Softmax work; they do not model hardware Softmax support.
@@ -50,9 +44,8 @@ Model SHA-256 checksums:
 c3231061d081bdd47884137b02134f85142752a39e87263c529cd14ed242b096  resnet34/resnet34-v1-7.onnx
 c99c507058eaf41de8723408fdda7db8325cb57f0a89f2ee07a716d6e963e14e  googlenet/googlenet-12.onnx
 c960c195953423388733cd0e60712cf425ca1f8917e264f5cf98e028752c61cf  googlenet/googlenet-12-no-lrn.onnx
-a26f9e33901c573e60c34a3f0abbb4744fff83e4e0f21b18fc66e20395e72982  googlenet/googlenet-12-pimsim-nn.onnx
 396cdea21e5e7d02c3f26f14d22ef20975171702493f5c5e79b8e0d896e541ef  vgg8/vgg8-mnist-reconstructed.onnx
-229f3975af8933d39aee8d9031d969bff074b69c33e304a78abb35ff0c5f445f  yolo11n/yolo11n-pimsim-nn.onnx
+5e7bd85dc60b252674f8d7e1ad24f2317d00dfb8c524826536698462de04620d  yolo11n/yolo11n.onnx
 ```
 
 ## Hardware profiles

validation/networks/pimcomp_models/yolo11n/yolo11n.onnx

@@ -0,0 +1 @@
+../../yolo11n/depth_51/yolo11n_depth_51.onnx

validation/raptor_validation/pimcomp_models.py

@@ -11,7 +11,7 @@ MODELS = {
     "resnet18": SUITE / "resnet18/resnet18-v1-7.onnx",
     "resnet34": SUITE / "resnet34/resnet34-v1-7.onnx",
     "googlenet": SUITE / "googlenet/googlenet-12-no-lrn.onnx",
-    "yolo11n": REPO / "validation/networks/yolo11n/depth_51/yolo11n_depth_51.onnx",
+    "yolo11n": SUITE / "yolo11n/yolo11n.onnx",
 }
 MODEL_NAMES = tuple(MODELS)
 DEFAULT_MODELS = MODEL_NAMES

validation/tests/test_pim_experiments.py

@@ -511,7 +511,9 @@ def test_comparison_uses_same_model_for_reference_and_execution(tmp_path, monkey
 
 
 def test_yolo_uses_full_model_without_moving_comparison_artifacts():
-    assert suite.MODELS["yolo11n"] == suite.REPO / "validation/networks/yolo11n/depth_51/yolo11n_depth_51.onnx"
+    assert suite.MODELS["yolo11n"] == suite.SUITE / "yolo11n/yolo11n.onnx"
+    assert suite.MODELS["yolo11n"].is_symlink()
+    assert suite.MODELS["yolo11n"].resolve() == suite.REPO / "validation/networks/yolo11n/depth_51/yolo11n_depth_51.onnx"
     assert suite.model_dir(None, "yolo11n") == suite.SUITE / "yolo11n"
     assert suite.common_dir(None, "yolo11n") == suite.SUITE / "yolo11n/artifacts/common-full-model"
 
@@ -756,3 +758,43 @@ def test_ablation_aggregate_keeps_pass_with_zero_performance(tmp_path):
     assert rows[0]["functional_validation"] == "PASS"
     assert rows[0]["baseline_functional_validation"] == "PASS"
     assert all(rows[0][f"{metric}_gain_percent"] == "NA" for _, metric in ablation.PERCENTAGE_FIELDS)
+
+
+def test_completed_no_sync_ablation_preserves_numerical_failure():
+    report = {"pimsim_mode": "throughput", "raptor_extra_args": ["--pim-disable-synchronization"],
+              "raptor_validation": {"status": "done", "passed": False, "max_diffs": {"output": 1}},
+              "raptor_performance": {"throughput": 100, "output_count": 8},
+              "steps": [{"name": "Functional validation Raptor", "status": "passed"}],
+              "failures": []}
+    assert suite.completed_synchronization_ablation(report)
+    assert not suite.comparison_passed(report)
+    assert suite.functional_validation_status(report["raptor_validation"]) == "FAIL"
+    for field, value in (("pimsim_mode", "latency"), ("raptor_extra_args", []),
+                         ("raptor_validation", {"status": "failed", "passed": False}),
+                         ("raptor_validation", {"status": "done", "passed": False}),
+                         ("raptor_performance", {"error": "SIGSEGV", "throughput": 100}),
+                         ("raptor_performance", {"throughput": 100, "output_count": 0}),
+                         ("steps", [{"status": "failed"}]),
+                         ("failures", [{"stage": "Compile Raptor Pim", "error": "failed"}])):
+        assert not suite.completed_synchronization_ablation({**report, field: value})
+
+
+@pytest.mark.parametrize("process_failed", [False, True])
+@pytest.mark.parametrize("variant,expected", [("no-sync", 0), ("no-plan", 1), ("baseline", 1)])
+def test_ablation_runner_distinguishes_invalid_outputs_from_failed_execution(tmp_path, monkeypatch, variant, expected, process_failed):
+    report = {"pimsim_mode": "throughput", "raptor_extra_args": ["--pim-disable-synchronization"],
+              "raptor_validation": {"status": "done", "passed": False, "max_diffs": {"output": 1}},
+              "raptor_performance": {"throughput": 100, "output_count": 8},
+              "steps": [{"status": "passed"}], "failures": []}
+    output = suite.result_dir(tmp_path, "vgg8", "arch-a", "throughput", 4, variant) / "pimcomp/comparison_report.json"
+    output.parent.mkdir(parents=True)
+    output.write_text(json.dumps(report))
+    csv = tmp_path / "results.csv"
+    csv.write_text("summary\n")
+    monkeypatch.setattr(suite, "run", lambda *a, **k: 0)
+    monkeypatch.setattr(suite, "run_comparison_jobs", lambda *a: (["arch-a/vgg8/throughput/pipeline4"] if process_failed else [], 1))
+    monkeypatch.setattr(suite, "write_results_csv", lambda *a: csv)
+    monkeypatch.setattr(sys, "argv", ["run_pimcomp_models.py", "--out-dir", str(tmp_path),
+        "--models", "vgg8", "--archs", "arch-a", "--mode", "throughput", "--pipeline", "4",
+        "--raptor-only", "--ablation-variant", variant, "--jobs", "1"])
+    assert suite.main() == (1 if process_failed else expected)

validation/tools/pim/ablation/README.md

@@ -59,8 +59,10 @@ Both programs' functional statuses are recorded as `functional_validation`
 and `baseline_functional_validation`. The latter belongs to the disabled feature.
 
 Disabling synchronization removes host-buffer lifetime and pipeline restart
-guarantees, so those variants may fail numerical validation and exit nonzero.
+guarantees, so those variants may fail numerical validation. A completed
-Baseline and `no-plan` must pass. Percentages involving failing programs remain
+`no-sync` measurement retains `FAIL` in the CSV and is reported as diagnostic
+in the runner summary. Compile, execution, missing-output, and measurement
+failures still make the run fail. Baseline and `no-plan` must pass. Percentages involving failing programs remain
 diagnostic comparisons, not valid speedups.
 
 When fast-mode convergence fails, positive fallback throughput remains available.

validation/tools/pim/ablation/results_ablation.csv

@@ -1,65 +1,25 @@
-variant,arch,model,mode,raptor_pipeline,functional_validation,latency,throughput,power,energy
+comparison,arch,model,mode,raptor_pipeline,baseline_variant,baseline_functional_validation,functional_validation,latency_gain_percent,throughput_gain_percent,power_gain_percent,energy_gain_percent
-baseline,arch-a,googlenet,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
+plan,arch-a,googlenet,latency,1,no-plan,PASS,PASS,+99.14%,NA,-226.98%,+97.19%
-no-sync,arch-a,googlenet,latency,1,PASS,+0.00%,NA,+0.00%,-0.00%
+plan,arch-a,googlenet,throughput,4,no-plan,PASS,PASS,NA,NA,-281.83%,NA
-no-plan,arch-a,googlenet,latency,1,PASS,+11513.29%,NA,-69.42%,+3452.69%
+sync,arch-a,googlenet,throughput,4,no-sync,FAIL,PASS,-5.17%,-4.96%,+12.23%,-1.65%
-no-sync-no-plan,arch-a,googlenet,latency,1,PASS,+11512.99%,NA,-69.42%,+3452.59%
+plan,arch-a,resnet18,latency,1,no-plan,PASS,PASS,+96.96%,NA,-156.23%,+92.21%
-baseline,arch-a,googlenet,throughput,4,PASS,+0.00%,+0.00%,+0.00%,+0.00%
+plan,arch-a,resnet18,throughput,4,no-plan,PASS,PASS,+95.17%,+1970.00%,-157.48%,+85.39%
-no-sync,arch-a,googlenet,throughput,4,FAIL,+26.17%,-20.74%,+9.97%,+18.50%
+sync,arch-a,resnet18,throughput,4,no-sync,FAIL,PASS,+7.15%,+7.81%,+14.70%,+7.92%
-no-plan,arch-a,googlenet,throughput,4,PASS,-100.00%,-100.00%,-62.73%,-100.00%
+plan,arch-a,resnet34,latency,1,no-plan,PASS,PASS,+95.94%,NA,-130.05%,+90.65%
-no-sync-no-plan,arch-a,googlenet,throughput,4,FAIL,+4094.63%,-97.62%,-59.86%,+1311.01%
+plan,arch-a,resnet34,throughput,4,no-plan,PASS,PASS,+96.00%,+2400.00%,-189.18%,+88.43%
-baseline,arch-a,resnet18,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
+sync,arch-a,resnet34,throughput,4,no-sync,FAIL,PASS,-9.11%,-8.42%,+7.32%,-1.02%
-no-sync,arch-a,resnet18,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
+plan,arch-a,vgg8,latency,1,no-plan,PASS,PASS,+94.75%,NA,-80.65%,+90.51%
-no-plan,arch-a,resnet18,latency,1,PASS,+3191.30%,NA,-60.97%,+1184.24%
+plan,arch-a,vgg8,throughput,4,no-plan,PASS,PASS,+92.77%,+1279.31%,-170.46%,+82.49%
-no-sync-no-plan,arch-a,resnet18,latency,1,PASS,+3191.30%,NA,-60.97%,+1184.24%
+sync,arch-a,vgg8,throughput,4,no-sync,FAIL,PASS,+19.18%,+24.03%,+21.29%,+18.62%
-baseline,arch-a,resnet18,throughput,4,PASS,+0.00%,+0.00%,+0.00%,+0.00%
+plan,arch-b,googlenet,latency,1,no-plan,PASS,PASS,+98.99%,NA,-239.10%,+96.59%
-no-sync,arch-a,resnet18,throughput,4,FAIL,+110.65%,-52.55%,+43.25%,+177.48%
+plan,arch-b,googlenet,throughput,4,no-plan,PASS,PASS,NA,NA,-306.99%,NA
-no-plan,arch-a,resnet18,throughput,4,PASS,+1598.93%,-94.12%,-51.04%,+664.87%
+sync,arch-b,googlenet,throughput,4,no-sync,FAIL,PASS,-1.28%,-1.63%,+9.99%,-4.29%
-no-sync-no-plan,arch-a,resnet18,throughput,4,FAIL,+1174.21%,-92.16%,-45.06%,+543.74%
+plan,arch-b,resnet18,latency,1,no-plan,PASS,PASS,+96.71%,NA,-171.16%,+91.07%
-baseline,arch-a,resnet34,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
+plan,arch-b,resnet18,throughput,4,no-plan,PASS,PASS,+95.46%,+2102.50%,-200.52%,+83.26%
-no-sync,arch-a,resnet34,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
+sync,arch-b,resnet18,throughput,4,no-sync,FAIL,PASS,+2.22%,+2.32%,+8.26%,-7.88%
-no-plan,arch-a,resnet34,latency,1,PASS,+2361.37%,NA,-56.53%,+969.99%
+plan,arch-b,resnet34,latency,1,no-plan,PASS,PASS,+96.23%,NA,-166.76%,+89.94%
-no-sync-no-plan,arch-a,resnet34,latency,1,PASS,+2361.37%,NA,-56.53%,+969.99%
+plan,arch-b,resnet34,throughput,4,no-plan,PASS,PASS,+97.58%,+4030.00%,-168.71%,+91.79%
-baseline,arch-a,resnet34,throughput,4,PASS,+0.00%,+0.00%,+0.00%,+0.00%
+sync,arch-b,resnet34,throughput,4,no-sync,FAIL,PASS,-0.80%,-0.96%,+10.61%,-13.51%
-no-sync,arch-a,resnet34,throughput,4,FAIL,+33.65%,-25.21%,+31.76%,+15.56%
+plan,arch-b,vgg8,latency,1,no-plan,PASS,PASS,+94.68%,NA,-98.04%,+89.47%
-no-plan,arch-a,resnet34,throughput,4,PASS,+1726.82%,-94.52%,-54.50%,+455.98%
+plan,arch-b,vgg8,throughput,4,no-plan,PASS,PASS,+93.70%,+1484.16%,-205.54%,+81.37%
-no-sync-no-plan,arch-a,resnet34,throughput,4,FAIL,+813.41%,-89.04%,-49.77%,+206.84%
+sync,arch-b,vgg8,throughput,4,no-sync,FAIL,PASS,+17.07%,+20.30%,+19.03%,+9.63%
-baseline,arch-a,vgg8,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
-no-sync,arch-a,vgg8,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
-no-plan,arch-a,vgg8,latency,1,PASS,+1800.66%,NA,-44.65%,+953.39%
-no-sync-no-plan,arch-a,vgg8,latency,1,PASS,+1800.66%,NA,-44.65%,+953.39%
-baseline,arch-a,vgg8,throughput,4,PASS,+0.00%,+0.00%,+0.00%,+0.00%
-no-sync,arch-a,vgg8,throughput,4,PASS,+2.33%,-1.72%,+18.49%,-0.98%
-no-plan,arch-a,vgg8,throughput,4,PASS,+1251.16%,-92.58%,-56.68%,+577.04%
-no-sync-no-plan,arch-a,vgg8,throughput,4,PASS,+1260.47%,-92.63%,-53.36%,+467.83%
-baseline,arch-b,googlenet,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
-no-sync,arch-b,googlenet,latency,1,PASS,+0.00%,NA,+0.00%,-0.00%
-no-plan,arch-b,googlenet,latency,1,PASS,+9849.64%,NA,-70.51%,+2833.89%
-no-sync-no-plan,arch-b,googlenet,latency,1,PASS,+9849.55%,NA,-70.51%,+2833.89%
-baseline,arch-b,googlenet,throughput,4,PASS,+0.00%,+0.00%,+0.00%,+0.00%
-no-sync,arch-b,googlenet,throughput,4,FAIL,+11.45%,-10.29%,+21.82%,+19.27%
-no-plan,arch-b,googlenet,throughput,4,PASS,-100.00%,-100.00%,-63.39%,-100.00%
-no-sync-no-plan,arch-b,googlenet,throughput,4,FAIL,+3348.28%,-97.10%,-61.58%,+1026.09%
-baseline,arch-b,resnet18,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
-no-sync,arch-b,resnet18,latency,1,PASS,+0.00%,NA,+0.00%,-0.00%
-no-plan,arch-b,resnet18,latency,1,PASS,+2935.63%,NA,-63.12%,+1019.81%
-no-sync-no-plan,arch-b,resnet18,latency,1,PASS,+2936.09%,NA,-63.12%,+1019.97%
-baseline,arch-b,resnet18,throughput,4,PASS,+0.00%,+0.00%,+0.00%,+0.00%
-no-sync,arch-b,resnet18,throughput,4,FAIL,+8.03%,-7.44%,+47.59%,+40.61%
-no-plan,arch-b,resnet18,throughput,4,PASS,+1511.85%,-93.80%,-52.93%,+560.80%
-no-sync-no-plan,arch-b,resnet18,throughput,4,FAIL,+867.12%,-89.67%,-45.41%,+359.83%
-baseline,arch-b,resnet34,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
-no-sync,arch-b,resnet34,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
-no-plan,arch-b,resnet34,latency,1,PASS,+2551.16%,NA,-62.51%,+893.70%
-no-sync-no-plan,arch-b,resnet34,latency,1,PASS,+2550.91%,NA,-62.51%,+893.62%
-baseline,arch-b,resnet34,throughput,4,PASS,+0.00%,+0.00%,+0.00%,+0.00%
-no-sync,arch-b,resnet34,throughput,4,FAIL,+28.25%,-22.13%,+49.76%,+57.30%
-no-plan,arch-b,resnet34,throughput,4,PASS,+2427.17%,-96.05%,-49.58%,+938.62%
-no-sync-no-plan,arch-b,resnet34,throughput,4,FAIL,+531.79%,-84.19%,-42.32%,+197.08%
-baseline,arch-b,vgg8,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
-no-sync,arch-b,vgg8,latency,1,PASS,+0.00%,NA,+0.00%,+0.00%
-no-plan,arch-b,vgg8,latency,1,PASS,+1784.97%,NA,-49.50%,+849.73%
-no-sync-no-plan,arch-b,vgg8,latency,1,PASS,+1784.97%,NA,-49.50%,+849.73%
-baseline,arch-b,vgg8,throughput,4,PASS,+0.00%,+0.00%,+0.00%,+0.00%
-no-sync,arch-b,vgg8,throughput,4,PASS,+23.91%,-19.82%,+25.67%,+23.93%
-no-plan,arch-b,vgg8,throughput,4,PASS,+1158.70%,-92.07%,-58.95%,+408.33%
-no-sync-no-plan,arch-b,vgg8,throughput,4,PASS,+1172.83%,-92.17%,-55.83%,+327.00%

validation/tools/pim/pimcomp/compare/results_comparison.csv

@@ -1,41 +1,41 @@
 arch,model,mode,raptor_pipeline,pimcomp_pipeline,raptor_functional_validation,pimcomp_functional_validation,raptor_latency_ms,pimcomp_latency_ms,raptor_throughput_samples_s,pimcomp_throughput_samples_s,raptor_power_mw,pimcomp_power_mw,raptor_energy_pj,pimcomp_energy_pj,better_compiler,speedup
 arch-a,vgg8,latency,1,element,PASS,PASS,1.51833,7.98507,NA,NA,321.046,200.111,4.87455e+08,1.5979e+09,raptor,5.26
-arch-a,vgg8,throughput,2,batch,PASS,PASS,0.906321,1.0364,1100,965,429.311,482.597,4.07972e+08,5.41692e+08,raptor,1.14
+arch-a,vgg8,throughput,2,batch,PASS,FAIL,0.625998,1.0364,1600,965,523.278,482.597,4.17419e+08,5.41692e+08,NA,1.66
-arch-a,vgg8,throughput,4,batch,PASS,PASS,0.778242,1.0364,1280,965,457.039,482.597,4.15762e+08,5.41692e+08,raptor,1.33
+arch-a,vgg8,throughput,4,batch,PASS,FAIL,0.625998,1.0364,1600,965,523.278,482.597,4.17419e+08,5.41692e+08,NA,1.66
-arch-a,vgg8,throughput,8,batch,PASS,PASS,0.778094,1.0364,1290,965,464.071,482.597,3.96424e+08,5.41692e+08,raptor,1.34
+arch-a,vgg8,throughput,8,batch,PASS,FAIL,0.62635,1.0364,1600,965,519.182,482.597,4.14242e+08,5.41692e+08,NA,1.66
-arch-a,resnet18,latency,1,element,PASS,FAIL,14.0332,58.1775,NA,NA,455.67,238.389,6.3945e+09,1.38689e+10,raptor,4.15
+arch-a,resnet18,latency,1,element,PASS,FAIL,14.0332,58.1775,NA,NA,455.67,238.389,6.3945e+09,1.38689e+10,NA,4.15
-arch-a,resnet18,throughput,2,batch,PASS,FAIL,11.7314,13.4585,85.2,74.3,503.111,481.604,6.62807e+09,6.56345e+09,raptor,1.15
+arch-a,resnet18,throughput,2,batch,PASS,FAIL,11.0904,13.4585,90.2,74.3,520.244,481.604,6.55167e+09,6.56345e+09,NA,1.21
-arch-a,resnet18,throughput,4,batch,PASS,FAIL,19.2072,13.4585,52.1,74.3,380.1,481.604,7.95703e+09,6.56345e+09,pimcomp,1.43
+arch-a,resnet18,throughput,4,batch,PASS,FAIL,12.0822,13.4585,82.8,74.3,487.576,481.604,6.91662e+09,6.56345e+09,NA,1.11
-arch-a,resnet18,throughput,8,batch,PASS,FAIL,14.1578,13.4585,70.6,74.3,453.416,481.604,6.77909e+09,6.56345e+09,pimcomp,1.05
+arch-a,resnet18,throughput,8,batch,PASS,FAIL,11.9279,13.4585,83.8,74.3,499.354,481.604,6.84729e+09,6.56345e+09,NA,1.13
-arch-a,resnet34,latency,1,element,PASS,FAIL,29.7186,66.7114,NA,NA,411.806,277.485,1.22383e+10,1.85114e+10,raptor,2.24
+arch-a,resnet34,latency,1,element,PASS,FAIL,29.7189,66.7114,NA,NA,411.804,277.485,1.22384e+10,1.85114e+10,NA,2.24
-arch-a,resnet34,throughput,2,batch,PASS,FAIL,23.2705,23.5817,43,42.4,471.717,515.68,1.3055e+10,1.22836e+10,raptor,1.01
+arch-a,resnet34,throughput,2,batch,PASS,FAIL,21.8196,23.5817,45.8,42.4,477.108,515.68,1.57899e+10,1.22836e+10,NA,1.08
-arch-a,resnet34,throughput,4,batch,PASS,FAIL,27.468,23.5817,36.4,42.4,422.223,515.68,1.48991e+10,1.22836e+10,pimcomp,1.16
+arch-a,resnet34,throughput,4,batch,PASS,FAIL,20,23.5817,50,42.4,548.372,515.68,1.09674e+10,1.22836e+10,NA,1.18
-arch-a,resnet34,throughput,8,batch,PASS,FAIL,39.8392,23.5817,25.1,42.4,352.28,515.68,1.70627e+10,1.22836e+10,pimcomp,1.69
+arch-a,resnet34,throughput,8,batch,PASS,FAIL,20.6363,23.5817,48.5,42.4,504.854,515.68,1.23725e+10,1.22836e+10,NA,1.14
-arch-a,googlenet,latency,1,element,PASS,PASS,9.77772,62.9248,NA,NA,563.864,231.193,5.5133e+09,1.45477e+10,raptor,6.44
+arch-a,googlenet,latency,1,element,PASS,FAIL,9.77731,62.9248,NA,NA,563.88,231.193,5.51323e+09,1.45477e+10,NA,6.44
-arch-a,googlenet,throughput,2,batch,PASS,FAIL,8.41545,18.2643,119,54.8,610.746,435.107,5.77746e+09,8.16479e+09,raptor,2.17
+arch-a,googlenet,throughput,2,batch,PASS,FAIL,7.34567,18.2643,136,54.8,679.351,435.107,5.5053e+09,8.16479e+09,NA,2.48
-arch-a,googlenet,throughput,4,batch,PASS,FAIL,10.7292,18.2643,93.2,54.8,528.026,435.107,6.59289e+09,8.16479e+09,raptor,1.70
+arch-a,googlenet,throughput,4,batch,PASS,FAIL,7.43728,18.2643,134,54.8,674.025,435.107,5.60809e+09,8.16479e+09,NA,2.45
-arch-a,googlenet,throughput,8,batch,PASS,FAIL,10.1219,18.2643,98.8,54.8,523.808,435.107,6.52174e+09,8.16479e+09,raptor,1.80
+arch-a,googlenet,throughput,8,batch,PASS,FAIL,9.01886,18.2643,111,54.8,595.582,435.107,5.71605e+09,8.16479e+09,NA,2.03
-arch-a,yolo11n,latency,1,element,PASS,FAIL,147.49,NA,NA,NA,290.174,NA,4.27977e+10,NA,NA,NA
+arch-a,yolo11n,latency,1,element,PASS,FAIL,149.46,NA,NA,NA,288.579,NA,4.31311e+10,NA,NA,NA
-arch-a,yolo11n,throughput,2,batch,SKIPPED,FAIL,102.337,NA,9.77,NA,337.882,NA,3.83081e+10,NA,NA,NA
+arch-a,yolo11n,throughput,2,batch,PASS,FAIL,94.946,NA,10.5,NA,346.466,NA,3.75539e+10,NA,NA,NA
-arch-a,yolo11n,throughput,4,batch,SKIPPED,FAIL,87.9103,NA,11.4,NA,358.201,NA,3.87298e+10,NA,NA,NA
+arch-a,yolo11n,throughput,4,batch,PASS,FAIL,80.1193,NA,12.5,NA,377.408,NA,3.8811e+10,NA,NA,NA
-arch-a,yolo11n,throughput,8,batch,SKIPPED,FAIL,82.9508,NA,12.1,NA,366.725,NA,3.85347e+10,NA,NA,NA
+arch-a,yolo11n,throughput,8,batch,PASS,FAIL,99.5488,NA,10,NA,339.331,NA,4.27113e+10,NA,NA,NA
 arch-b,vgg8,latency,1,element,PASS,PASS,1.5333,7.15212,NA,NA,292.559,173.769,4.48582e+08,1.24281e+09,raptor,4.66
-arch-b,vgg8,throughput,2,batch,PASS,PASS,0.685556,1.07282,1460,932,456.044,409.47,4.03694e+08,4.39289e+08,raptor,1.57
+arch-b,vgg8,throughput,2,batch,PASS,FAIL,0.624729,1.07282,1600,932,489.899,409.47,3.93036e+08,4.39289e+08,NA,1.72
-arch-b,vgg8,throughput,4,batch,PASS,PASS,0.673192,1.07282,1490,932,448.898,409.47,4.13527e+08,4.39289e+08,raptor,1.60
+arch-b,vgg8,throughput,4,batch,PASS,FAIL,0.62331,1.07282,1600,932,490.955,409.47,3.93418e+08,4.39289e+08,NA,1.72
-arch-b,vgg8,throughput,8,batch,PASS,PASS,0.751106,1.07282,1330,932,425.742,409.47,3.94458e+08,4.39289e+08,raptor,1.43
+arch-b,vgg8,throughput,8,batch,PASS,FAIL,0.62331,1.07282,1600,932,490.955,409.47,3.93418e+08,4.39289e+08,NA,1.72
-arch-b,resnet18,latency,1,element,PASS,FAIL,15.3758,64.3818,NA,NA,400.661,201.663,6.16048e+09,1.29834e+10,raptor,4.19
+arch-b,resnet18,latency,1,element,PASS,FAIL,15.3758,64.3818,NA,NA,400.661,201.663,6.16048e+09,1.29834e+10,NA,4.19
-arch-b,resnet18,throughput,2,batch,PASS,FAIL,12.1214,13.434,82.5,74.4,461.407,484.804,6.29532e+09,6.56966e+09,raptor,1.11
+arch-b,resnet18,throughput,2,batch,PASS,FAIL,12.9113,13.434,77.5,74.4,438.931,484.804,6.47794e+09,6.56966e+09,NA,1.04
-arch-b,resnet18,throughput,4,batch,PASS,FAIL,11.3601,13.434,88,74.4,479.333,484.804,6.98745e+09,6.56966e+09,raptor,1.18
+arch-b,resnet18,throughput,4,batch,PASS,FAIL,11.3505,13.434,88.1,74.4,478.606,484.804,6.66517e+09,6.56966e+09,NA,1.18
-arch-b,resnet18,throughput,8,batch,PASS,FAIL,13.1216,13.434,76.2,74.4,441.248,484.804,6.38271e+09,6.56966e+09,raptor,1.02
+arch-b,resnet18,throughput,8,batch,PASS,FAIL,11.3685,13.434,88,74.4,478.33,484.804,6.51513e+09,6.56966e+09,NA,1.18
-arch-b,resnet34,latency,1,element,PASS,FAIL,27.9541,107.2,NA,NA,397.206,206.476,1.11035e+10,2.21342e+10,raptor,3.83
+arch-b,resnet34,latency,1,element,PASS,FAIL,27.9541,107.2,NA,NA,397.206,206.476,1.11035e+10,2.21342e+10,NA,3.83
-arch-b,resnet34,throughput,2,batch,PASS,FAIL,22.282,22.8454,44.9,43.8,438.238,493.332,1.42187e+10,1.13922e+10,raptor,1.03
+arch-b,resnet34,throughput,2,batch,PASS,FAIL,22.4967,22.8454,44.5,43.8,440.06,493.332,1.49623e+10,1.13922e+10,NA,1.02
-arch-b,resnet34,throughput,4,batch,PASS,FAIL,29.0906,22.8454,34.4,43.8,364.956,493.332,1.63285e+10,1.13922e+10,pimcomp,1.27
+arch-b,resnet34,throughput,4,batch,PASS,FAIL,24.1838,22.8454,41.3,43.8,426.129,493.332,1.30195e+10,1.13922e+10,NA,1.06
-arch-b,resnet34,throughput,8,batch,PASS,FAIL,47.68,22.8454,21,43.8,289.821,493.332,1.63464e+10,1.13922e+10,pimcomp,2.09
+arch-b,resnet34,throughput,8,batch,PASS,FAIL,24.1516,22.8454,41.4,43.8,420.745,493.332,1.39184e+10,1.13922e+10,NA,1.06
-arch-b,googlenet,latency,1,element,PASS,PASS,11.2212,37.9367,NA,NA,483.266,242.037,5.42282e+09,9.18206e+09,raptor,3.38
+arch-b,googlenet,latency,1,element,PASS,FAIL,11.2206,37.9367,NA,NA,483.283,242.037,5.42274e+09,9.18206e+09,NA,3.38
-arch-b,googlenet,throughput,2,batch,PASS,FAIL,9.21209,21.3902,109,46.8,545.488,417.866,5.5934e+09,9.01473e+09,raptor,2.33
+arch-b,googlenet,throughput,2,batch,PASS,FAIL,9.08396,21.3902,110,46.8,550.292,417.866,5.70509e+09,9.01473e+09,NA,2.35
-arch-b,googlenet,throughput,4,batch,PASS,FAIL,8.56194,21.3902,117,46.8,555.264,417.866,5.86312e+09,9.01473e+09,raptor,2.50
+arch-b,googlenet,throughput,4,batch,PASS,FAIL,8.23423,21.3902,121,46.8,597.327,417.866,5.62683e+09,9.01473e+09,NA,2.59
-arch-b,googlenet,throughput,8,batch,PASS,FAIL,10.8652,21.3902,92,46.8,483.477,417.866,6.12683e+09,9.01473e+09,raptor,1.97
+arch-b,googlenet,throughput,8,batch,PASS,FAIL,7.78944,21.3902,128,46.8,620.65,417.866,5.47039e+09,9.01473e+09,NA,2.74
-arch-b,yolo11n,latency,1,element,PASS,FAIL,168.144,NA,NA,NA,245.314,NA,4.1248e+10,NA,NA,NA
+arch-b,yolo11n,latency,1,element,PASS,FAIL,171.598,NA,NA,NA,243.177,NA,4.17288e+10,NA,NA,NA
-arch-b,yolo11n,throughput,2,batch,SKIPPED,FAIL,104.494,NA,9.57,NA,302.121,NA,3.58015e+10,NA,NA,NA
+arch-b,yolo11n,throughput,2,batch,PASS,FAIL,99.5318,NA,10,NA,309.102,NA,3.5219e+10,NA,NA,NA
-arch-b,yolo11n,throughput,4,batch,SKIPPED,FAIL,108.247,NA,9.24,NA,297.484,NA,3.76297e+10,NA,NA,NA
+arch-b,yolo11n,throughput,4,batch,PASS,FAIL,85.3497,NA,11.7,NA,326.712,NA,3.58691e+10,NA,NA,NA
-arch-b,yolo11n,throughput,8,batch,SKIPPED,FAIL,100.129,NA,9.99,NA,308.833,NA,3.71808e+10,NA,NA,NA
+arch-b,yolo11n,throughput,8,batch,PASS,FAIL,97.9941,NA,10.2,NA,308.041,NA,3.5011e+10,NA,NA,NA

validation/tools/pim/pimcomp/compare/run_pimcomp_models.py

@@ -282,6 +282,23 @@ def comparison_passed(report: dict) -> bool:
     return result.get("status") == "done" and bool(result.get("passed"))
 
 
+def completed_synchronization_ablation(report: dict) -> bool:
+    """A completed no-sync measurement may have intentionally invalid outputs."""
+    validation = report.get("raptor_validation") or {}
+    performance = report.get("raptor_performance") or {}
+    return (
+        report.get("pimsim_mode") == "throughput"
+        and "--pim-disable-synchronization" in report.get("raptor_extra_args", [])
+        and validation.get("status") == "done" and validation.get("passed") is False
+        and bool(validation.get("max_diffs"))
+        and not report.get("failures")
+        and bool(report.get("steps"))
+        and all(step.get("status") == "passed" for step in report["steps"])
+        and not performance.get("error")
+        and performance_values(performance)["throughput"] is not None
+    )
+
+
 def functional_validation_status(result: dict | None) -> str:
     if result is None:
         return "NA"
@@ -869,6 +886,7 @@ def main() -> int:
     assert results_path is not None
     print_stage(results_path.name, STAGE_COLORS["Compare outputs"])
     print(results_path.read_text(encoding="utf-8"), end="")
+    diagnostic_failures = []
     for arch in args.archs:
         for name in args.models:
             for mode, pipeline, _ in comparisons_by_arch[arch]:
@@ -876,15 +894,19 @@ def main() -> int:
                 report_path = result_dir(
                     out_dir, name, arch, mode, pipeline, args.ablation_variant
                 ) / "pimcomp/comparison_report.json"
-                if not report_path.exists() or not comparison_passed(
+                report = json.loads(report_path.read_text(encoding="utf-8")) if report_path.exists() else {}
-                    json.loads(report_path.read_text(encoding="utf-8")),
+                if not comparison_passed(report):
-                ):
+                    if (label not in failed and args.ablation_variant == "no-sync"
-                    if label not in failed:
+                            and completed_synchronization_ablation(report)):
+                        diagnostic_failures.append(label)
+                    elif label not in failed:
                         failed.append(label)
     print(Style.BRIGHT + Fore.CYAN + "Summary" + Style.RESET_ALL)
     total_jobs = sum(len(args.models) * len(comparisons) for comparisons in comparisons_by_arch.values())
-    print(Style.BRIGHT + f"Passed: {total_jobs - len(failed)}" + Style.RESET_ALL)
+    print(Style.BRIGHT + f"Passed: {total_jobs - len(failed) - len(diagnostic_failures)}" + Style.RESET_ALL)
     print(Style.BRIGHT + f"Failed: {len(failed)}" + Style.RESET_ALL)
+    if diagnostic_failures:
+        print("No-sync numerical failures (diagnostic only): " + ", ".join(diagnostic_failures))
     print(Style.BRIGHT + f"Results: {results_path}" + Style.RESET_ALL)
     if failed:
         print(

validation/tools/pim/pimcomp/correctness/README.md

@@ -35,13 +35,15 @@ The corresponding scheduling call is disabled in
 Rust validation starts those buffers at zero, exposing the missing computation.
 
 ```bash
-.venv/bin/python validation/tools/pim/pimcomp/correctness/prefill/run_experiment.py
+.venv/bin/python validation/tools/pim/pimcomp/correctness/prefill/run_experiment.py \
+  /path/to/comparison /path/to/matching-model.onnx
 ```
 
-The default uses GoogLeNet Arch-A throughput/pipeline2. Optional positional
+The comparison directory and its matching model are required. Optional positional
-arguments are comparison directory, model, work directory, and shared reference
+arguments are work directory and shared reference directory. The default work
-directory; the default work directory is `correctness/prefill/` below the
+directory is `correctness/prefill/` below the comparison artifacts.
-comparison artifacts.
+Historical Softmax-free models can be recovered from Git for historical artifacts.
+Current GoogLeNet retains Softmax; prefill alone does not correct omitted Softmax execution.
 
 The experiment compares the original memory image with one prefilled using
 [`batch_memory.py`](prefill/batch_memory.py). In the recorded case, 20 of 39

validation/tools/pim/pimcomp/correctness/prefill/run_experiment.py

@@ -16,8 +16,6 @@ from batch_memory import prefill_batch_memory
 
 SCRIPT_DIR = Path(__file__).resolve().parent
 REPO_ROOT = SCRIPT_DIR.parents[5]
-DEFAULT_COMPARISON_DIR = REPO_ROOT / "validation/networks/pimcomp_models/googlenet/artifacts/arch-a/throughput/pipeline2"
-DEFAULT_MODEL = REPO_ROOT / "validation/networks/pimcomp_models/googlenet/googlenet-12-pimsim-nn.onnx"
 DEFAULT_COMMON_DIR = REPO_ROOT / "validation/networks/pimcomp_models/googlenet/artifacts/common"
 SIMULATOR_MANIFEST = REPO_ROOT / "backend-simulators/pim/pim-simulator/Cargo.toml"
 
@@ -78,8 +76,8 @@ def compare_outputs(
 
 def main() -> int:
     parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument("comparison_dir", nargs="?", type=Path, default=DEFAULT_COMPARISON_DIR)
+    parser.add_argument("comparison_dir", type=Path)
-    parser.add_argument("model", nargs="?", type=Path, default=DEFAULT_MODEL)
+    parser.add_argument("model", type=Path, help="ONNX model used to generate the comparison artifacts")
     parser.add_argument("work_dir", nargs="?", type=Path)
     parser.add_argument("common_dir", nargs="?", type=Path, default=DEFAULT_COMMON_DIR)
     args = parser.parse_args()