Multi-cgra mapper integration by guosran · Pull Request #279 · coredac/dataflow

guosran · 2026-03-06T01:12:41Z

Summary

This PR addresses issue #275. It adds multi-CGRA task placement support to MapTaskOnCgraPass and integrates it into ResourceAwareTaskOptimizationPass.

Changes

MapTaskOnCgraPass

Multi-CGRA placement in findBestPlacement: Tasks with cgra_count > 1 are now placed using a cascading shape search:
1. Try all rectangular shapes of the requested cgra_count (e.g., 1×3, 3×1 for cgra_count=3).
2. If no rectangular shape fits, try all connected non-rectangular shapes (L-shapes, T-shapes, etc.) via DFS enumeration.
3. If nothing fits, fall back to cgra_count - 1 (reject the extra CGRA, keep the previous allocation).
parseTileShapeOffsets: Parses the tile_shape attribute string (e.g., "2x2", "2x2[(0,0)(1,0)(0,1)]") into physical CGRA offset coordinates. Added assertion for empty offsets.
computeScore: Updated to handle multi-CGRA placements by computing the minimum Manhattan distance from any CGRA in the placement to the target, reflecting fast bypass paths between adjacent CGRAs.
Helper functions: Extracted tryPlaceShape, getRectShapes, and tryNonRectShapes for clarity and reuse.
Public API: Exposed runMapTaskOnCgra(FuncOp, int, int) in TaskflowPasses.h so other passes can invoke placement directly.

ResourceAwareTaskOptimizationPass

Integration: After the balance/fusion convergence loop, calls runMapTaskOnCgra() to produce task_mapping_info attributes with global grid placement that respects multi-CGRA shapes.

…ment - findBestPlacement now tries rectangular shapes first, then non-rectangular connected shapes, then falls back to k-1 CGRAs (down to 1). - Removed outdated TODO comment about MapTaskOnCgraPass not supporting multi-CGRA placement. - Added assert for empty tile_shape offsets. - Cleaned up USER COMMENT annotations.

…p comments - findBestPlacement tries rect then non-rect shapes for requested cgra_count. - If placement fails, caller falls back to cgra_count-1 (reject extra CGRA). - Normalize /// to // comment style throughout MapTaskOnCgraPass. - Remove outdated TODO comments.

- SRAM centroid now includes ALL CGRA positions of multi-CGRA tasks, not just placement[0]. - SSA proximity scoring uses min distance between two multi-CGRA placements (minDistToPlacement) instead of only comparing to the other task's primary position.

Copilot

Pull request overview

This PR adds multi-CGRA task placement support to MapTaskOnCgraPass and integrates global placement generation (task_mapping_info) into ResourceAwareTaskOptimizationPass after the balance/fusion loop converges (issue #275).

Changes:

Extend CGRA placement to support tasks spanning multiple CGRAs (rectangular and connected non-rectangular shapes), and update scoring to account for multi-tile proximity.
Expose runMapTaskOnCgra(func, rows, cols) as a callable helper and invoke it from ResourceAwareTaskOptimizationPass post-convergence.
Update multi-CGRA MLIR tests to reflect new RESOPT output formatting and/or placement results.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
lib/TaskflowDialect/Transforms/MapTaskOnCgraPass.cpp	Implements multi-CGRA placement search + scoring changes; adds `runMapTaskOnCgra` entrypoint.
lib/TaskflowDialect/Transforms/Optimizations/ResourceAwareTaskOptimizationPass.cpp	Calls `runMapTaskOnCgra` after convergence to materialize `task_mapping_info`.
include/TaskflowDialect/TaskflowPasses.h	Exposes `runMapTaskOnCgra` in the public header.
test/multi-cgra/taskflow/resource-heavy/resource-heavy.mlir	Adjusts RESOPT FileCheck expectations (line-wrapped attrs).
test/multi-cgra/taskflow/resnet/simple_resnet_tosa.mlir	Adjusts RESOPT FileCheck expectations (line-wrapped attrs).
test/multi-cgra/taskflow/parallel-nested/parallel-nested.mlir	Adjusts RESOPT FileCheck expectations (line-wrapped attrs).
test/multi-cgra/taskflow/multi-nested/multi-nested.mlir	Updates placement expectations and RESOPT FileCheck formatting.
test/multi-cgra/taskflow/irregular-loop/irregular-loop.mlir	Adjusts RESOPT FileCheck expectations (line-wrapped attrs).

Comments suppressed due to low confidence (1)

lib/TaskflowDialect/Transforms/MapTaskOnCgraPass.cpp:294

If findBestPlacement() returns an empty placement (including after the fallback), the code still commits placement.primary() which becomes (-1,-1). That invalid coordinate then propagates into task_mapping_info and into SRAM centroid computation (assignAllSRAMs sums all positions), producing incorrect placements instead of failing loudly. Handle the "no placement found" case explicitly (e.g., emit an MLIR diagnostic and signalPassFailure / return failure), and avoid pushing a sentinel position into task_node->placement.

          // Commits Placement.
          task_node->placement.push_back(placement.primary());
          for (size_t i = 1; i < placement.cgra_positions.size(); ++i) {
             task_node->placement.push_back(placement.cgra_positions[i]);
          }

          // Marks occupied.
          for (const auto &pos : placement.cgra_positions) {
            if (pos.row >= 0 && pos.row < grid_rows_ && pos.col >= 0 && pos.col < grid_cols_) {
                occupied_[pos.row][pos.col] = true;
            }
          }

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

lib/TaskflowDialect/Transforms/Optimizations/ResourceAwareTaskOptimizationPass.cpp

lib/TaskflowDialect/Transforms/AllocateCgraToTaskPass.cpp

…pper-integration

ShangkunLi · 2026-03-18T11:13:19Z

include/TaskflowDialect/TaskflowPasses.h

+
+// Runs the CGRA task placement logic directly on a function.
+// grid_rows/grid_cols default to 4x4 (kCgraGridRows/kCgraGridCols).
+void runAllocateCgraToTask(mlir::func::FuncOp func,


I think we should create two utility folders in include/TaskflowDialect/Util or lib/TaskflowDialect/Util and put these utility functions in these folders. You can refer to the folder structure in NeuraDialect.

ShangkunLi · 2026-03-18T11:16:14Z

include/TaskflowDialect/TaskflowPasses.td

@@ -70,7 +70,7 @@ def MapTaskOnCgra : Pass<"map-task-on-cgra", "func::FuncOp"> {

    Uses a default 3x3 CGRA grid.


The default configuration in runAllocateCgraToTask is 4x4. Maybe we can keep them consistent, 3x3 or 4x4.

ShangkunLi · 2026-03-18T11:35:16Z

lib/TaskflowDialect/Transforms/AllocateCgraToTaskPass.cpp


-  /// Checks if any CGRA in this task is adjacent to any in other task.
+  // Checks if any CGRA in this task is adjacent to any in other task.
  bool hasAdjacentCGRA(const TaskPlacement &other) const {


Could we rename it to hasTaskAdjacentCgra?

ShangkunLi · 2026-03-18T13:15:42Z

lib/TaskflowDialect/Transforms/AllocateCgraToTaskPass.cpp

+          // If the requested cgra_count doesn't fit, fall back to cgra_count-1
+          // (i.e. reject the extra CGRA and keep previous allocation).


If the cgra_count is calculated in resourceAwareTaskOptimization pass, we should always be able to find the best placement in the findBestPlacement function, right?

ShangkunLi · 2026-03-18T13:19:41Z

lib/TaskflowDialect/Transforms/AllocateCgraToTaskPass.cpp

          task_node->placement.push_back(placement.primary());
-          // Handles mapping one task on multi-CGRAs.
-          // TODO: Introduce explicit multi-CGRA binding logic.
          for (size_t i = 1; i < placement.cgra_positions.size(); ++i) {
             task_node->placement.push_back(placement.cgra_positions[i]);
          }


Why not use one for loop for all cgra_positions commit?

ShangkunLi · 2026-03-18T15:22:35Z

lib/TaskflowDialect/Transforms/AllocateCgraToTaskPass.cpp

+    for (auto &shape : getRectShapes(cgra_count)) {
+      TaskPlacement p = tryPlaceShape(task_node, shape, graph);
+      if (!p.cgra_positions.empty()) return p;
+    }


Why not combine getRectShapes and tryPlaceShape as a single function?

ShangkunLi · 2026-03-18T16:04:31Z

lib/TaskflowDialect/Transforms/Optimizations/ResourceAwareTaskOptimizationPass.cpp

+      // Simple hash of sorted positions.
+      int64_t hash = 0;
+      for (auto &[col_off, row_off] : sorted_positions)
+        hash = hash * 131 + col_off * 17 + row_off;


How do you choose these parameters? Maybe add a comment explaining how these parameters affect the sorting.

ShangkunLi · 2026-03-18T16:12:16Z

lib/TaskflowDialect/Transforms/Optimizations/ResourceAwareTaskOptimizationPass.cpp

+//   1. Rectangular shapes, sorted by squareness (e.g. 2×2 before 1×4),
+//      with smaller bounding-box area as tiebreaker.
+//   2. Non-rectangular shapes (L, T, etc.) in all unique rotations.
+static SmallVector<CgraShape> getAllPlacementShapes(int cgra_count) {


Can we put this function into the util folder?

ShangkunLi · 2026-03-18T16:13:15Z

lib/TaskflowDialect/Transforms/Optimizations/ResourceAwareTaskOptimizationPass.cpp

+// (including the speculatively modified one).
+//
+// Returns true if all tasks can be placed without overlap.
+static bool canAllTasksFitOnGrid(ArrayRef<int> task_cgra_counts) {


Can we put this function into the util folder? So that we can further extend the mapping to support manually assigned cgra_counts for each task.

ShangkunLi · 2026-03-18T16:17:18Z

lib/TaskflowDialect/Transforms/AllocateCgraToTaskPass.cpp

+  return std::make_unique<AllocateCgraToTaskPass>();
+}
+
+void runAllocateCgraToTask(func::FuncOp func, int grid_rows, int grid_cols) {


Put this function into the util folder.

…and optimization passes to support multi-CGRA shapes.

guosran added 4 commits March 5, 2026 12:51

feat(taskflow): integrate MapTaskOnCgraPass for multi-CGRA placements

87cb9a0

Refined fallback logic

273610a

guosran requested review from ShangkunLi and Copilot and removed request for Copilot March 6, 2026 01:12

Copilot started reviewing on behalf of guosran March 6, 2026 01:13 View session

Copilot AI review requested due to automatic review settings March 6, 2026 01:30

Copilot started reviewing on behalf of guosran March 6, 2026 01:30 View session

Copilot AI reviewed Mar 6, 2026

View reviewed changes

tancheng reviewed Mar 6, 2026

View reviewed changes

lib/TaskflowDialect/Transforms/AllocateCgraToTaskPass.cpp Outdated Show resolved Hide resolved

lib/TaskflowDialect/Transforms/AllocateCgraToTaskPass.cpp Show resolved Hide resolved

guosran added 2 commits March 7, 2026 12:06

refactor: rename MapToAcceleratorPass to MapOperationOnTilePass

c17d7fc

refactor: rename MapTaskOnCgraPass to AllocateCgraToTaskPass

2d00d5a

guosran force-pushed the feature/multi-cgra-mapper-integration branch from 07a38b2 to 2d00d5a Compare March 7, 2026 04:25

guosran added 3 commits March 7, 2026 12:28

Merge remote-tracking branch 'origin/main' into feature/multi-cgra-ma…

5b1b33b

…pper-integration

fix: resolve merge conflicts with main (FuseTaskPass, CMakeLists)

9112303

refactor: rename CGRAPosition to CgraPosition for consistency

27b7bcf

tancheng requested a review from HobbitQia March 7, 2026 06:31

guosran added 2 commits March 7, 2026 21:20

Merge branch 'main' into feature/multi-cgra-mapper-integration

f584835

resolve TODOs, add post-placement re-profiling

ea0a0b6

ShangkunLi reviewed Mar 18, 2026

View reviewed changes

guosran added 2 commits March 19, 2026 14:41

Add CGRA placement utilities and integrate them into task allocation …

a2c4da4

…and optimization passes to support multi-CGRA shapes.

Add implementation for CGRA task mapping and shared placement utilities

eabe7be

guosran marked this pull request as draft March 20, 2026 07:12

		@@ -70,7 +70,7 @@ def MapTaskOnCgra : Pass<"map-task-on-cgra", "func::FuncOp"> {

		Uses a default 3x3 CGRA grid.

		// If the requested cgra_count doesn't fit, fall back to cgra_count-1
		// (i.e. reject the extra CGRA and keep previous allocation).

Conversation

guosran commented Mar 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Changes

MapTaskOnCgraPass

ResourceAwareTaskOptimizationPass

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

guosran commented Mar 6, 2026 •

edited

Loading