Skip to content

partition_analyzer.md

Latest commit

 

History

History
198 lines (146 loc) · 4.91 KB

partition_analyzer.md

File metadata and controls

198 lines (146 loc) · 4.91 KB

How to Use: partition_analyzer.py

Overview

partition_analyzer.py analyzes and compares different partitioning strategies for FX computational graphs. It quantifies the benefits of operator fusion by comparing unfused (baseline) vs fusion strategies, showing metrics like subgraph reduction and data movement savings.

Key Capabilities:

  • Partition graphs using different strategies (unfused vs fusion)
  • Compare partitioning strategies side-by-side
  • Visualize graph structure and partitions
  • Quantify benefits of fusion (reduced subgraphs, memory traffic)

Target Users:

  • Compiler engineers evaluating fusion strategies
  • Researchers studying graph transformations
  • ML engineers understanding model structure

Quick Start

# Compare all strategies on ResNet-18
python3 cli/partition_analyzer.py --model resnet18 --strategy all --compare

# Visualize fusion partitioning
python3 cli/partition_analyzer.py --model mobilenet_v2 --strategy fusion --visualize

Command-Line Arguments

Argument Type Default Description
--model str resnet18 Model name
--strategy str all Partitioning strategy: unfused, fusion, all
--compare flag False Show side-by-side comparison
--quantify flag False Show detailed metrics
--visualize flag False Show graph visualization
--input-shape int[] 1,3,224,224 Input tensor shape

Range Selection (for visualization)

Note: Node numbers are 1-based (matching display output) and ranges are inclusive on both ends.

Argument Type Default Description
--start int None Start node (1-based, inclusive)
--end int None End node (1-based, inclusive)
--around int None Center node for context view
--context int 10 Nodes before/after center (with --around)
--max-nodes int 20 Max nodes from start (backward compatible)

Range Selection Priority:

  1. --around with --context (highest priority)
  2. --start and/or --end
  3. --max-nodes (default: first 20 nodes)

Partitioning Strategies

Unfused (Baseline)

One subgraph per operator:

  • No fusion
  • Maximum subgraph count
  • Maximum data movement
  • Useful as baseline for comparison

Fusion

Aggregate operators to minimize data movement:

  • Fuses compatible operations (conv+bn+relu)
  • Reduces subgraph count (5-10×)
  • Minimizes intermediate memory traffic
  • Production-ready strategy

Usage Examples

Example 1: Basic Comparison

python3 cli/partition_analyzer.py \
  --model resnet50 \
  --strategy all \
  --compare

Output:

================================================================================
PARTITIONING COMPARISON: ResNet-50
================================================================================

Unfused Strategy:
  Subgraphs:        177
  Total FLOPs:      8.21 G
  Total Memory:     52.4 MB
  Data Movement:    high (every operator boundary)

Fusion Strategy:
  Subgraphs:        20
  Total FLOPs:      8.21 G (same)
  Total Memory:     52.4 MB (same)
  Data Movement:    low (fused operator groups)

Fusion Benefits:
  Subgraph Reduction:  88.7% (177 → 20)
  Data Movement:       ~80% reduction
  Kernel Launch:       ~88% reduction

Example 2: Visualize Fusion

# Show first 30 nodes
python3 cli/partition_analyzer.py \
  --model mobilenet_v2 \
  --strategy fusion \
  --visualize \
  --max-nodes 30

# Show specific range (nodes 10-25, inclusive)
python3 cli/partition_analyzer.py \
  --model mobilenet_v2 \
  --strategy fusion \
  --visualize \
  --start 10 --end 25

# Investigate around node 15 (±5 nodes)
python3 cli/partition_analyzer.py \
  --model mobilenet_v2 \
  --strategy fusion \
  --visualize \
  --around 15 --context 5

Output: ASCII graph showing fused subgraphs with node-by-node details

Example 3: Quantify Fusion Benefits

python3 cli/partition_analyzer.py \
  --model efficientnet_b0 \
  --strategy fusion \
  --quantify

Output:

  • Per-subgraph FLOPs
  • Per-subgraph memory
  • Fusion decisions
  • Bottleneck analysis

Interpretation

Subgraph Reduction

High Reduction (>80%):

  • ✓ Good: Significant fusion opportunities
  • Typical for CNNs (Conv+BN+ReLU patterns)

Low Reduction (<50%):

  • May indicate complex architecture
  • Transformers have less fusion opportunity (attention blocks)

Data Movement

Fusion Benefits:

  • Eliminates intermediate tensor writes
  • Reduces DRAM traffic (40-80%)
  • Improves cache utilization

Related Tools

Tool Purpose
profile_graph.py Profile graph characteristics
analyze_graph_mapping.py Map partitioned graph to hardware

Further Reading

  • Fusion Partitioner: src/graphs/transform/partitioning/fusion_partition_analyzer.py
  • Architecture Guide: CLAUDE.md