wr

warp-rnnt

CUDA-Warp RNN-Transducer

Showing:

Popularity

Downloads/wk

0

GitHub Stars

146

Maintenance

Last Commit

1mo ago

Contributors

4

Package

Dependencies

0

License

MIT

Categories

Readme

PyPI Downloads

CUDA-Warp RNN-Transducer

A GPU implementation of RNN Transducer (Graves 2012, 2013). This code is ported from the reference implementation (by Awni Hannun) and fully utilizes the CUDA warp mechanism.

The main bottleneck in the loss is a forward/backward pass, which based on the dynamic programming algorithm. In particular, there is a nested loop to populate a lattice with shape (T, U), and each value in this lattice depend on the two previous cells from each dimension (e.g. forward pass).

CUDA executes threads in groups of 32 parallel threads called warps. Full efficiency is realized when all 32 threads of a warp agree on their execution path. This is exactly what is used to optimize the RNN Transducer. The lattice is split into warps in the T dimension. In each warp, variables between threads exchanged using a fast operations. As soon as the current warp fills the last value, the next two warps (t+32, u) and (t, u+1) are start running. A schematic procedure for the forward pass is shown in the figure below, where T - number of frames, U - number of labels, W - warp size. The similar procedure for the backward pass runs in parallel.

Performance

NVIDIA Profiler shows advantage of the warp implementation over the non-warp implementation.

This warp implementation:

Non-warp implementation warp-transducer:

Unfortunately, in practice this advantage disappears because the memory operations takes much longer. Especially if you synchronize memory on each iteration.

warp_rnnt (gather=False)warp_rnnt (gather=True)warprnnt_pytorchtransducer (CPU)
T=150, U=40, V=28
N=10.50 ms0.54 ms0.63 ms1.28 ms
N=161.79 ms1.72 ms1.85 ms6.15 ms
N=323.09 ms2.94 ms2.97 ms12.72 ms
N=645.83 ms5.54 ms5.23 ms23.73 ms
N=12811.30 ms10.74 ms9.99 ms47.93 ms
T=150, U=20, V=5000
N=10.95 ms0.80 ms1.74 ms21.18 ms
N=168.74 ms6.24 ms16.20 ms240.11 ms
N=3217.26 ms12.35 ms31.64 ms490.66 ms
N=64out-of-memoryout-of-memoryout-of-memory944.73 ms
N=128out-of-memoryout-of-memoryout-of-memory1894.93 ms
T=1500, U=300, V=50
N=15.89 ms4.99 ms10.02 ms121.82 ms
N=1695.46 ms78.88 ms76.66 ms732.50 ms
N=32out-of-memory157.86 ms165.38 ms1448.54 ms
N=64out-of-memoryout-of-memoryout-of-memory2767.59 ms

Benchmarked on a GeForce RTX 2070 Super GPU, Intel i7-10875H CPU @ 2.30GHz.

Note

  • This implementation assumes that the input is log_softmax.

  • In addition to alphas/betas arrays, counts array is allocated with shape (N, U * 2), which is used as a scheduling mechanism.

  • core_gather.cu is a memory-efficient version that expects log_probs with the shape (N, T, U, 2) only for blank and labels values. It shows excellent performance with a large vocabulary.

  • Do not expect that this implementation will greatly reduce the training time of RNN Transducer model. Probably, the main bottleneck will be a trainable joint network with an output (N, T, U, V).

  • Also, there is a restricted version, called Recurrent Neural Aligner, with assumption that the length of input sequence is equal to or greater than the length of target sequence.

Install

There are two bindings for the core algorithm:

Reference

Rate & Review

Great Documentation0
Easy to Use0
Performant0
Highly Customizable0
Bleeding Edge0
Responsive Maintainers0
Poor Documentation0
Hard to Use0
Slow0
Buggy0
Abandoned0
Unwelcoming Community0
100