Mesorasi: Architecture Support for Point Cloud Analytics via Delayed-Aggregation

MICRO 2020 talk slides by Yu Feng and Tiancheng Xu.

Transcript

1. 1 Mesorasi: Architecture Support for Point Cloud Analytics via Delayed-Aggregation Yu Feng, Boyuan Tian, Tiancheng Xu with Paul Whatmough (Arm Research) and Yuhao Zhu Department of Computer Science University of Rochester http://horizon-lab.org https://github.com/horizon-research/Efficient-Deep-Learning-for-Point-Clouds
3. 2 ✘
6. 4 Autonomous Driving Robotics Mixed Reality Drone Navigation
7. Classification Segmentation Detection SLAM Deep Learning on Point Clouds 5
8. Classification Segmentation Detection SLAM Deep Learning on Point Clouds 5
9. Classification Segmentation Detection SLAM Deep Learning on Point Clouds 5 
 Performance
 Energy-efficiency

10. Classification Segmentation Detection SLAM Deep Learning on Point Clouds 5 Can we use existing neural network accelerators on point cloud workloads?
11. 6 Aggregation P3 P4 P12 P13 Neighbor Search P0 P2 P1 P6 P5 P7 P8 P9 P10 P11 Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU Key Operators P1: {P2, P3, ….} P3: {P8, P7, ….} P6: {P8, P5, ….} … … P6: … P1:
12. P1: {P2, P3, ….} P3: {P8, P7, ….} P6: {P8, P5, ….} … … P6: … P1: 7 Aggregation P3 P4 P12 P13 Neighbor Search P0 P2 P1 P6 P5 P7 P8 P9 P10 P11 Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU Key Operators
13. 8 Neighbor Search in Point Clouds
14. 8 Neighbor Search in Point Clouds P1 P8 P0 P2 P3 P6 P5 P7 P9 P10 P12 P13 P11 P4
15. 8 Neighbor Search in Point Clouds P1 P8 P0 P2 P3 P6 P5 P7 P9 P10 P12 P13 P11 P4
16. 8 Neighbor Search in Point Clouds N P1 P8 P0 P2 P3 P6 P5 P7 P9 P10 P12 P13 P11 P4
17. 8 Neighbor Search in Point Clouds N P1: { P0, P2, P3 P4, P5, P6 } P8: { P5, P6, P7 P9, P10, P11 } P1 P8 P0 P2 P3 P6 P5 P7 P9 P10 P12 P13 P11 P4
18. 8 Neighbor Search in Point Clouds N P1: { P0, P2, P3 P4, P5, P6 } P8: { P5, P6, P7 P9, P10, P11 } P1 P8 P0 P2 P3 P6 P5 P7 P9 P10 P12 P13 P11 P4 Sorting, KD-Tree Irregular Memory Access
19. Neighbor Search in conventional DNN 9 P0,0 P0,1 P0,2 P1,0 P1,1 P1,2 P2,0 P2,1 P2,2 P1,1 P1,2 P1,3 P2,1 P2,2 P2,3 P3,1 P3,2 P3,3 N P0,0 P0,1 P0,2 P0,3 P0,4 P1,0 P1,1 P1,2 P1,3 P1,4 P2,0 P2,1 P2,2 P2,3 P2,4 P3,0 P3,1 P3,2 P3,3 P3,4 Pixels: inherently regular 
 Points: irregularly scattered Regular DNNs don’t need explicit neighbor search
20. 10 Aggregation P3 P4 P12 P13 Neighbor Search P0 P2 P1 P6 P5 P7 P8 P9 P10 P11 Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU Key Operators P1: {P2, P3, ….} P3: {P8, P7, ….} P6: {P8, P5, ….} … … P6: … P1:
21. 11 Aggregation P0 P2 P3 P1 P6 P5 P7 P8 P9 P10 P12 P13 P11 N P1: { P0, P2, P3, P4, P5, P6 } P8: { P5, P6, P7, P9, P10, P11 } P4 N A Points: feature vectors Neighbor Index Table
22. 11 Aggregation P0 P2 P3 P1 P6 P5 P7 P8 P9 P10 P12 P13 P11 N P1: { P0, P2, P3, P4, P5, P6 } P8: { P5, P6, P7, P9, P10, P11 } P4 N A P1: Feature Matrix Points: feature vectors Neighbor Index Table
23. 11 Aggregation P0 P2 P3 P1 P6 P5 P7 P8 P9 P10 P12 P13 P11 N P1: { P0, P2, P3, P4, P5, P6 } P8: { P5, P6, P7, P9, P10, P11 } P4 N A P0 P1: Feature Matrix Points: feature vectors Neighbor Index Table
24. 11 Aggregation P0 P2 P3 P1 P6 P5 P7 P8 P9 P10 P12 P13 P11 N P1: { P0, P2, P3, P4, P5, P6 } P8: { P5, P6, P7, P9, P10, P11 } P4 N A P0 - P1 P1: Feature Matrix Points: feature vectors Neighbor Index Table
25. 11 Aggregation P0 P2 P3 P1 P6 P5 P7 P8 P9 P10 P12 P13 P11 N P1: { P0, P2, P3, P4, P5, P6 } P8: { P5, P6, P7, P9, P10, P11 } P4 N A P0 - P1 P2 - P1 P3 - P1 P4 - P1 P5 - P1 P6 - P1 P1: Feature Matrix Points: feature vectors Neighbor Index Table
26. 11 Aggregation P0 P2 P3 P1 P6 P5 P7 P8 P9 P10 P12 P13 P11 N P1: { P0, P2, P3, P4, P5, P6 } P8: { P5, P6, P7, P9, P10, P11 } P4 N A P0 - P1 P2 - P1 P3 - P1 P4 - P1 P5 - P1 P6 - P1 P1: Feature Matrix P5 - P8 P6 - P8 P7 - P8 P9 - P8 P10 - P8 P11 - P8 P8: Feature Matrix Points: feature vectors Neighbor Index Table
27. P1: {P2, P3, ….} P3: {P8, P7, ….} P6: {P8, P5, ….} … … P6: … P1: 12 Aggregation P3 P4 P12 P13 Neighbor Search P0 P2 P1 P6 P5 P7 P8 P9 P10 P11 Key Operators Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU
28. 13 Feature Computation MatMul ReLU MatMul ReLU MLP layers P2 - P1 P3 - P1 P4 - P1 P5 - P1 P6 - P1 P1: Feature Matrix P5 - P8 P6 - P8 P7 - P8 P9 - P8 P10 - P8 P11 - P8 P8: Feature Matrix P0 - P1 P0 - P1
29. 13 Feature Computation MatMul ReLU MatMul ReLU MLP layers P2 - P1 P3 - P1 P4 - P1 P5 - P1 P6 - P1 P1: Feature Matrix P5 - P8 P6 - P8 P7 - P8 P9 - P8 P10 - P8 P11 - P8 P8: Feature Matrix P0’ P0 - P1
30. 14 Feature Computation MatMul ReLU MatMul ReLU P1: Feature Matrix P5 - P8 P6 - P8 P7 - P8 P9 - P8 P10 - P8 P11 - P8 P8: Feature Matrix P2 - P1 P3 - P1 P4 - P1 P5 - P1 P6 - P1 P0 - P1 P0’ P2’ P3’ P4’ P5’ P6’ P1’: New Feature Matrix P5’ P6’ P7’ P9’ P10’ P11’ P8’: New Feature Matrix
31. 14 Feature Computation MatMul ReLU MatMul ReLU P1: Feature Matrix P5 - P8 P6 - P8 P7 - P8 P9 - P8 P10 - P8 P11 - P8 P8: Feature Matrix P2 - P1 P3 - P1 P4 - P1 P5 - P1 P6 - P1 P0 - P1 P0’ P2’ P3’ P4’ P5’ P6’ P1’: New Feature Matrix P5’ P6’ P7’ P9’ P10’ P11’ P8’: New Feature Matrix Reduction P1’ P1’: New Feature Vector P8’ P8’: New Feature Vector
32. Point Cloud Network Layer 15 P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool Aggregation Neighbor Feature Matrix (NFM) Each centroid point has a NFM N x M P1: P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 N neighbors (N == 2) Neighbor Search Neighbor Index Table P1: {P2, P3} P5: {P3, P6} P1, P5: centroid points M-dimension feature vectors P1 P4 P6 P2 P3 P5 Points
33. Can We Use Existing DNN Accelerators? 16
34. Can We Use Existing DNN Accelerators? 16 ▸ No. They are not enough.
35. Can We Use Existing DNN Accelerators? 16 ▸ No. They are not enough. ▹ Neighbor Search
36. Can We Use Existing DNN Accelerators? 16 ▸ No. They are not enough. ▹ Neighbor Search ▹ Aggregation
37. Can We Use Existing DNN Accelerators? 16 ▸ No. They are not enough. ▹ Neighbor Search ▹ Aggregation 0 25 50 75 100 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et Neighbor Search Aggregation Feature Computation Others Characterization of 
 Point Cloud Networks % Execution Time
38. Can We Use Existing DNN Accelerators? 16 ▸ No. They are not enough. ▹ Neighbor Search ▹ Aggregation 0 25 50 75 100 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et Neighbor Search Aggregation Feature Computation Others Characterization of 
 Point Cloud Networks % Execution Time
39. Can We Use Existing DNN Accelerators? 16 ▸ No. They are not enough. ▹ Neighbor Search ▹ Aggregation 0 25 50 75 100 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et Neighbor Search Aggregation Feature Computation Others Characterization of 
 Point Cloud Networks % Execution Time ▹ Slow on high-end mobile GPUs: PointNet++: 132 ms FPointNet: 141 ms DGCNN: 5200 ms
40. Point Cloud Network Layer 17 P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool Aggregation Neighbor Feature Matrix (NFM) Each centroid point has a NFM N x M P1: P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 N neighbors (N == 2) Neighbor Search Neighbor Index Table P1: {P2, P3} P5: {P3, P6} P1, P5: centroid points M-dimension feature vectors P1 P4 P6 P2 P3 P5 Points
41. Neighbor Feature Matrix (NFM) N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1: Each centroid point has a NFM P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool Optimization 18
42. Neighbor Feature Matrix (NFM) N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1: Each centroid point has a NFM P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool Optimization 18 N x M P3 - P1 Mat Mul ( Kernel weights , )
43. Neighbor Feature Matrix (NFM) N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1: Each centroid point has a NFM P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool Optimization 18 N x M P3 - P1 Mat Mul ( Kernel weights , )
44. Neighbor Feature Matrix (NFM) N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1: Each centroid point has a NFM P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool = Optimization 18 N x M P3 - P1 Mat Mul ( Kernel weights , ) Mat Mul ( Kernel weights , ) N x M P3 Mat Mul ( Kernel weights , ) N x M P1 -
45. Neighbor Feature Matrix (NFM) N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1: Each centroid point has a NFM P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool = Optimization 18 N x M P3 - P1 Mat Mul ( Kernel weights , ) Mat Mul ( Kernel weights , ) N x M P3 Mat Mul ( Kernel weights , ) N x M P1 -
46. Neighbor Feature Matrix (NFM) N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1: Each centroid point has a NFM P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool = Optimization 18 N x M P3 - P1 Mat Mul ( Kernel weights , ) Mat Mul ( Kernel weights , ) N x M P3 Mat Mul ( Kernel weights , ) N x M P1 - Benefit:
47. Neighbor Feature Matrix (NFM) N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1: Each centroid point has a NFM P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool = Optimization 18 N x M P3 - P1 Mat Mul ( Kernel weights , ) Mat Mul ( Kernel weights , ) N x M P3 Mat Mul ( Kernel weights , ) N x M P1 - ▸ Effectively introduces reuse opportunities
 Benefit:
48. Neighbor Feature Matrix (NFM) N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1: Each centroid point has a NFM P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool = Optimization 18 N x M P3 - P1 Mat Mul ( Kernel weights , ) Mat Mul ( Kernel weights , ) N x M P3 Mat Mul ( Kernel weights , ) N x M P1 - ▸ Effectively introduces reuse opportunities
 Reduces up to 90% 
 MAC (Multiply-Accumulate) operations Benefit: Each point is used ~30 times
49. Neighbor Feature Matrix (NFM) N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1: Each centroid point has a NFM P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool = Optimization 18 N x M P3 - P1 Mat Mul ( Kernel weights , ) Mat Mul ( Kernel weights , ) N x M P3 Mat Mul ( Kernel weights , ) N x M P1 - ▸ Effectively introduces reuse opportunities
 Reduces up to 90% 
 MAC (Multiply-Accumulate) operations ▸ Dependency elimination Benefit: Each point is used ~30 times
50. Neighbor Feature Matrix (NFM) N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1: Each centroid point has a NFM P1’ P5’ M’-dimension features Feature Computation Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool Optimization 18 N x M P3 - P1 Mat Mul ( Kernel weights , ) Mat Mul ( Kernel weights , ) N x M P3 Mat Mul ( Kernel weights , ) N x M P1 - ReLU ( ) ReLU ( ) ReLU ( ) ≠ ▸ Effectively introduces reuse opportunities
 Reduces up to 90% 
 MAC (Multiply-Accumulate) operations ▸ Dependency elimination Benefit: Each point is used ~30 times
51. Approximation 19 Neighbor Feature Matrix (NFM) MLP + Reduction Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool N x M P3 - P1 Mat Mul ( Kernel weights , ) Mat Mul ( Kernel weights , ) N x M P3 Mat Mul ( Kernel weights , ) N x M P1 - ReLU ( ) ReLU ( ) ReLU ( ) ≈ P1’ P5’ M’-dimension features Each centroid point has a NFM N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1:
52. Approximation 19 Neighbor Feature Matrix (NFM) MLP + Reduction Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP + Reduction Max Pool N x M P3 - P1 Mat Mul ( Kernel weights , ) Mat Mul ( Kernel weights , ) N x M P3 Mat Mul ( Kernel weights , ) N x M P1 - ReLU ( ) ReLU ( ) ReLU ( ) ≈ P1’ P5’ M’-dimension features Each centroid point has a NFM N x M P3 - P1 P2 - P1 N x M P5: P6 - P5 P3 - P5 P1: The accuracy change 
 ranges from -0.9% to +1.2%
53. Delayed Aggregation 20 Neighbor Search Neighbor Index Table Feature Compute Point Feature Table Aggregation Neighbor Feature Matrix (NFM) } Reduction P1’ P5’ M’-dimension features Mat Mul 1 ReLU Mat Mul 2 ReLU Mat Mul 3 ReLU MLP M-dimension feature vectors P1 P4 P6 P2 P3 P5 Points
54. Bottleneck Shift 21 Time (msec.) 0 6 12 18 24 30 Original Delayed Aggr. Feature Computation Aggregation Neighbor Search Execution Time Distribution of PointNet++
55. Bottleneck Shift 21 Time (msec.) 0 6 12 18 24 30 Original Delayed Aggr. Feature Computation Aggregation Neighbor Search Execution Time Distribution of PointNet++
56. Bottleneck Shift 21 Time (msec.) 0 6 12 18 24 30 Original Delayed Aggr. Feature Computation Aggregation Neighbor Search Execution Time Distribution of PointNet++
57. Mesorasi: Point Cloud Acceleration Frame 22 Algorithm Hardware Delayed Aggregation Aggregation Acceleration
58. Aggregation Operation 23 Neighbor Feature Matrix Point Feature Table N neighbors P1: {P2, P3, P11, P23, P31, P39, …} P3: {P5, P7, P16, P19, P38, P41, …} … Neighbor Index Table P99:{P16, P71, P96, P119, P128, P142, …} P11 P31 P23 P3 P39 P2 … …
59. Aggregation Operation 23 Neighbor Feature Matrix Point Feature Table N neighbors P1: {P2, P3, P11, P23, P31, P39, …} P3: {P5, P7, P16, P19, P38, P41, …} … Neighbor Index Table P99:{P16, P71, P96, P119, P128, P142, …} P11 P31 P23 P3 P39 P2 … … P1: {P2, P3, P11, P23, P31, P39, …}
60. Aggregation Operation 23 Neighbor Feature Matrix Point Feature Table N neighbors P1: {P2, P3, P11, P23, P31, P39, …} P3: {P5, P7, P16, P19, P38, P41, …} … Neighbor Index Table P99:{P16, P71, P96, P119, P128, P142, …} P11 P31 P23 P3 P39 P2 … … P1: {P2, P3, P11, P23, P31, P39, …}
61. Aggregation Operation 23 Neighbor Feature Matrix Point Feature Table N neighbors P1: {P2, P3, P11, P23, P31, P39, …} P3: {P5, P7, P16, P19, P38, P41, …} … Neighbor Index Table P99:{P16, P71, P96, P119, P128, P142, …} P11 P31 P23 P3 P39 P2 … … P1: {P2, P3, P11, P23, P31, P39, …}
62. Aggregation Operation 23 Neighbor Feature Matrix Point Feature Table N neighbors P1: {P2, P3, P11, P23, P31, P39, …} P3: {P5, P7, P16, P19, P38, P41, …} … Neighbor Index Table P99:{P16, P71, P96, P119, P128, P142, …} P11 P31 P23 P3 P39 P2 … … P1: {P2, P3, P11, P23, P31, P39, …}
63. Aggregation Unit 24 1: {2, 17, 9, …} 3: {8, 17, 6, …} 6: {18, 25, 34, …} …. 900: {832, 987, …} Point Feature Table (PFT) Neighbor Index Table
64. Aggregation Unit 24 1: {2, 17, 9, …} 3: {8, 17, 6, …} 6: {18, 25, 34, …} …. 900: {832, 987, …} Address Generation Point Feature Table (PFT) Neighbor Index Table
65. Aggregation Unit 24 1: {2, 17, 9, …} 3: {8, 17, 6, …} 6: {18, 25, 34, …} …. 900: {832, 987, …} Address Generation Bank 1 Bank 2 Bank B … Bank 3 … Point Feature Table (PFT) Neighbor Index Table
66. Aggregation Unit 24 1: {2, 17, 9, …} 3: {8, 17, 6, …} 6: {18, 25, 34, …} …. 900: {832, 987, …} Address Generation Bank 1 Bank 2 Bank B … Bank 3 … B-ported, B-banked; No Crossbar Point Feature Table (PFT) Neighbor Index Table
67. Aggregation Unit 25 1: {2, 17, 9, …} 3: {8, 17, 6, …} 6: {18, 25, 34, …} …. 900: {832, 987, …} Address Generation Bank 1 Bank 2 Bank B … Bank 3 … Neighbor Index Table Point Feature Table (PFT) Reduction (Max) … Shift Registers B-ported, B-banked; No Crossbar
68. Aggregation Unit 25 1: {2, 17, 9, …} 3: {8, 17, 6, …} 6: {18, 25, 34, …} …. 900: {832, 987, …} Address Generation Bank 1 Bank 2 Bank B … Bank 3 … Neighbor Index Table Point Feature Table (PFT) Reduction (Max) … Shift Registers B-ported, B-banked; No Crossbar
69. Aggregation Unit 25 1: {2, 17, 9, …} 3: {8, 17, 6, …} 6: {18, 25, 34, …} …. 900: {832, 987, …} Address Generation Bank 1 Bank 2 Bank B … Bank 3 … Neighbor Index Table Point Feature Table (PFT) Reduction (Max) … Shift Registers MUX … (Store centroid’s feature vector) B-ported, B-banked; No Crossbar
70. Aggregation Unit 25 1: {2, 17, 9, …} 3: {8, 17, 6, …} 6: {18, 25, 34, …} …. 900: {832, 987, …} Address Generation Bank 1 Bank 2 Bank B … Bank 3 … Neighbor Index Table Point Feature Table (PFT) Reduction (Max) Sub Global Buffer … Shift Registers MUX … (Store centroid’s feature vector) B-ported, B-banked; No Crossbar
71. DNN Accelerator (NPU) DRAM CPU GPU Overall Hardware Design 26
72. DNN Accelerator (NPU) DRAM CPU GPU Overall Hardware Design 26 Input Point Cloud MLP Kernel Weights MLP Intermediate Activations Neighbor Index Table
73. DNN Accelerator (NPU) DRAM CPU GPU Overall Hardware Design 26 Input Point Cloud MLP Kernel Weights MLP Intermediate Activations Neighbor Index Table GPU (Neighbor Search)
74. DNN Accelerator (NPU) DRAM CPU GPU Overall Hardware Design 26 Input Point Cloud MLP Kernel Weights MLP Intermediate Activations Neighbor Index Table GPU (Neighbor Search) Feature Extraction + Aggregation
75. DNN Accelerator (NPU) DRAM CPU GPU Overall Hardware Design 26 Systolic MAC Unit Array BN/ReLU/ MaxPool Input Point Cloud MLP Kernel Weights MLP Intermediate Activations Neighbor Index Table GPU (Neighbor Search)
76. DNN Accelerator (NPU) DRAM CPU GPU Overall Hardware Design 26 Global Buffer (Weights /Fmaps) Global Buffer (Weights /FMaps) MCU MCU Systolic MAC Unit Array BN/ReLU/ MaxPool Input Point Cloud MLP Kernel Weights MLP Intermediate Activations Neighbor Index Table GPU (Neighbor Search)
77. DNN Accelerator (NPU) DRAM CPU GPU Overall Hardware Design 26 Global Buffer (Weights /Fmaps) Global Buffer (Weights /FMaps) MCU MCU Systolic MAC Unit Array BN/ReLU/ MaxPool Input Point Cloud MLP Kernel Weights MLP Intermediate Activations Neighbor Index Table Aggregation Logic Neighbor Index Table Point Feature Buffer Reduction (Max) Neighbor Index Table Neighbor Index Buffer GPU (Neighbor Search)
78. DNN Accelerator (NPU) DRAM CPU GPU Overall Hardware Design 26 Global Buffer (Weights /Fmaps) Global Buffer (Weights /FMaps) MCU MCU Systolic MAC Unit Array BN/ReLU/ MaxPool Input Point Cloud MLP Kernel Weights MLP Intermediate Activations Neighbor Index Table Aggregation Logic Neighbor Index Table Point Feature Buffer Reduction (Max) Neighbor Index Table Neighbor Index Buffer GPU (Neighbor Search) 
 With 3.8% area overhead to the NPU 

79. Experimental Setup 27
80. Experimental Setup 27 Three Point Cloud Applications: ▹ Object Classification, Object Segmentation, and Object Detection
81. Experimental Setup 27 Three Point Cloud Applications: ▹ Object Classification, Object Segmentation, and Object Detection Datasets: ▹ ModelNet40, ShapeNet, and KITTI dataset
82. Experimental Setup 27 Three Point Cloud Applications: ▹ Object Classification, Object Segmentation, and Object Detection Datasets: ▹ ModelNet40, ShapeNet, and KITTI dataset Models: ▹ Classification: PointNet++ (c), DGCNN (c), LDGCNN, DensePoint ▹ Segmentation: PointNet++ (s), DGCNN (s) ▹ Detection: F-PointNet
83. Experimental Setup 27 Three Point Cloud Applications: ▹ Object Classification, Object Segmentation, and Object Detection Datasets: ▹ ModelNet40, ShapeNet, and KITTI dataset Models: ▹ Classification: PointNet++ (c), DGCNN (c), LDGCNN, DensePoint ▹ Segmentation: PointNet++ (s), DGCNN (s) ▹ Detection: F-PointNet https://github.com/horizon-research/Efficient-Deep-Learning-for-Point-Clouds Github:
84. Accuracy Comparison 28 Accuracy (%) 0 20 40 60 80 100 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint Orignal Delayed Aggr.
85. Accuracy Comparison 28 Accuracy (%) 0 20 40 60 80 100 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint Orignal Delayed Aggr.
86. Accuracy Comparison 28 Accuracy (%) 0 20 40 60 80 100 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint Orignal Delayed Aggr.
87. Speedup and Energy Saving on GPU 29
88. Speedup and Energy Saving on GPU 29 Speedup 0 0.5 1 1.5 2 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint AVG .
89. Speedup and Energy Saving on GPU 29 Speedup 0 0.5 1 1.5 2 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint AVG . 1.6
90. Speedup and Energy Saving on GPU 29 Speedup 0 0.5 1 1.5 2 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint AVG . Saving % 0 25 50 75 100 1.6
91. Speedup and Energy Saving on GPU 29 Speedup 0 0.5 1 1.5 2 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint AVG . Saving % 0 25 50 75 100 1.6 51.1%
92. Hardware Performance Evaluation 30
93. Hardware Performance Evaluation 30 Hardware Baseline: ▹ A generic NPU + GPU SoC.
94. Hardware Performance Evaluation 30 Variants: ▹ Mesorasi-SW: delayed-aggregation without AU support. ▹ Mesorasi-HW: delayed-aggregation with AU support. Hardware Baseline: ▹ A generic NPU + GPU SoC.
95. Hardware Performance Evaluation 30 Variants: ▹ Mesorasi-SW: delayed-aggregation without AU support. ▹ Mesorasi-HW: delayed-aggregation with AU support. Hardware Baseline: ▹ A generic NPU + GPU SoC. Implementation: ▹ 16x16 Systolic Array ▹ Synposys synthesis, TSMC 16nm FinFET technology ▹
96. Speedup 31 Speedup 0 0.5 1 1.5 2 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint AVG . Mesorasi-SW Mesorasi-HW
97. Speedup 31 Speedup 0 0.5 1 1.5 2 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint AVG . Mesorasi-SW Mesorasi-HW 1.3
98. Speedup 31 Speedup 0 0.5 1 1.5 2 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint AVG . Mesorasi-SW Mesorasi-HW 2.2 3.6 1.3 1.9
99. Energy Savings 32 Saving (%) 0 25 50 75 100 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint AVG . Mesorasi-SW Mesorasi-HW
100. Energy Savings 32 Saving (%) 0 25 50 75 100 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint AVG . Mesorasi-SW Mesorasi-HW 22 %
101. Energy Savings 32 Saving (%) 0 25 50 75 100 PointN et++ (c) PointN et++ (s) DG C N N (c) DG C N N (s) F-PointN et LDG C N N DensePoint AVG . Mesorasi-SW Mesorasi-HW 22 % 38%
102. Conclusion 33 https://github.com/horizon-research/Efficient-Deep-Learning-for-Point-Clouds
103. Conclusion 33 ‣Delayed-aggregation decouples neighbor search with feature computation and significantly reduces the overall workload. https://github.com/horizon-research/Efficient-Deep-Learning-for-Point-Clouds
104. Conclusion 33 ‣Delayed-aggregation decouples neighbor search with feature computation and significantly reduces the overall workload. ‣Hardware support further maximizes the effectiveness of delayed-aggregation. https://github.com/horizon-research/Efficient-Deep-Learning-for-Point-Clouds