View Code? Open in Web Editor
NEW
This project forked from chenhui1016/hb-pytorch-apps
PyTorch Applications
Python 98.98%
Shell 0.11%
Makefile 0.91%
hb-pytorch-apps's Introduction
- Multilayer Perceptron for MNIST (mlp_mnist)
- Autoencoder Based Recommendation (Recsys)
- identify a kernel to port
- (possibly) register the kernel with PyTorch
- add host code, tests, kernel device code
- test using emulation
- test using cosim
- optimize
- identify a workload to port
- develop workload in pytorch-apps following coding conventions
- test serial native version (determine kernels not ported to HB)
- if any kernels not ported yet, goto kernel developer flow
- test on HB emulation (possibly full workload, maybe one batch)
- profile parallel native version
- identify key kernels using native profiling data
- develop workload kernel file with reduced chunks
- test workload kernels on emulation
- test workload kernels on cosim
- profile workload kernels natively for baseline kernel results
- profile workload kernels on cosim
- combine full workload profile data + workload kernel data to
estimate speedup of full workload on HB vs baseline
hb-pytorch-apps's People
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent