Open-DLA-Performance-Profiler
Portable (Windows) performance profiler for deep neural network. Give you the performance guideline of the model running on different openDLA architectures.
The openDLA is based on the open project NVDLA.
This tool is developed based on netron.
Usage
- Load the DLA architecture from the directory, select some architectures
- Load a model description of Caffe prototxt format
- click [view] and select summary, traffic, and detail tables
- explore the raw data
- copy raw data to your preferal spreadsheet tool to plot trends
Key Words
-
Summary Table : Summary of cycle counts, FPS of the model
- MAC Utilization: average MAC utilization under aggressive FPS
- Roofline Factor: the ratio of memory access cycles / total cycles
- Conservative FPS: consider the memory access and computation is fully overlapped
- Aggressive FPS: consider the memory access and computation is fully interleaved
-
Traffic Table : List of the memory traffic of each layer
- IFM : input feature map
- OFM : output feature map
- WM : weight matrix
-
Detail Table : List of the fused layers based on openDLA
- type : computation type
- attributes : fused layers
- ch_in : input channel number
- dim_in : input dimension
- ch_out : output channel number
- kernel_w : kernel width
- kernel_h : kernel height
- pad_w : padding size of width direction
- pad_h : padding size of height direction
- stride_w : stride size of width direction
- stride_h : stride size of height direction
Deferences between the free version here and commercial version
| Features | Free Version (here) | Commercial Version |
|---|---|---|
| Layer Fuse Info | O | O |
| Performance Table w/ DRAM | O | O |
| Performance Table w/ SRAM + DRAM | X | O |
| DRAM/SRAM access policy | X | O |
| DRAM data footprint | O | O |
| DRAM/SRAM data footprint | X | O |
| Configurable Freq./MEM BW | X | O |
| Source Codes | X | Available |