Skip to content

Commit

Permalink
Android app + doc cleanup
Browse files Browse the repository at this point in the history
  • Loading branch information
sevagh committed Apr 6, 2024
1 parent 2483139 commit fedd352
Show file tree
Hide file tree
Showing 5 changed files with 68 additions and 58 deletions.
9 changes: 0 additions & 9 deletions .github/BLAS_benchmarks.md

This file was deleted.

56 changes: 56 additions & 0 deletions .github/PERFORMANCE.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,56 @@
### Multi-core, OpenMP, BLAS, etc.

:warning: `demucs.cpp` library code in `./src` **should not use any threading (e.g. pthread or OpenMP) except through the BLAS interface.** This is because demucs.cpp is compiled to a single-threaded WebAssembly module in <https://freemusicdemixer.com>.

If you have OpenMP and OpenBLAS installed, OpenBLAS might automatically use all of the threads on your machine, which doesn't always run the fastest. Use the `OMP_NUM_THREADS` environment variable to limit this. On my 16c/32t machine, I found `OMP_NUM_THREADS=16` to be the fastest. This matches the [Eigen recommendation](https://eigen.tuxfamily.org/dox/TopicMultiThreading.html) to use the same number of threads as physical cores:
>On most OS it is very important to limit the number of threads to the number of physical cores, otherwise significant slowdowns are expected, especially for operations involving dense matrices.
### BLAS benchmarks

The benchmark plots below show the performance of different BLAS libraries (OpenBLAS, Intel MKL, AMD AOCL BLIS) with different numbers of threads on my Ryzen Zen3 5950X (16c/32t). In my case, 16 threads with OpenBLAS is a good blend of performance and memory usage.

<img alt="bench-wall-time" src="./wall_time_comparison.png" width="500"/>
<img alt="bench-cpu-time" src="./cpu_time_comparison.png" width="500"/>
<img alt="bench-memory" src="./memory_usage_comparison.png" width="500"/>

I didn't include any GPU BLAS libraries (NVBLAS, cuBLAS, etc.) because the I'm limiting the scope of demucs.cpp to use only the CPU. The real PyTorch version of Demucs is suitable for GPU acceleration.

### GPUs, cuBLAS, NVBLAS

There is a [branch](https://github.com/sevagh/demucs.cpp/tree/nvblas) where I explored NVBLAS (a cuBLAS wrapper with automatic host-GPU memory transfers). It's not very useful, but it's what I expect. Demucs.cpp is heavy on the for-loops and small matrix-vector or matrix-matrix multiplications. This is to run on Android phones (typically with small amounts of memory, 6-8 GB on flagships) and in WebAssembly (which has a 4 GB memory limit per module).

If I wrote it to use large matrix broadcasts, it would probably be faster (while consuming more memory and breaking the intended usecase), and accelerate much better on GPUs.

### Multi-threading

There are two new programs, `demucs_mt.cpp.main` and `demucs_ft_mt.cpp.main` that use C++11 [std::threads](https://en.cppreference.com/w/cpp/thread/thread).

In the single-threaded programs:

* User supplies a waveform of length N seconds
* Waveform is split into 7.8-second segments for Demucs inference
* Segments are processed sequentially, where each segment inference can use >1 core with `OMP_NUM_THREADS`

In the multi-threaded programs:
* User supplies a waveform of length N seconds and a `num_threads` argument
* Waveform is split into `num_threads` sub-waveforms (of length M < N) to process in parallel with a 0.75-second overlap
* We always need overlapping segments in audio applications to eliminate [boundary artifacts](https://freemusicdemixer.com/under-the-hood/2024/02/23/Demucs-segmentation#boundary-artifacts-and-the-overlap-add-method)
* `num_threads` threads are launched to perform Demucs inference on the sub-waveforms in parallel
* Within each thread, the sub-waveform is split into 7.8-second segments
* Segments within a thread are still processed sequentially, where each segment inference can use >1 core with `OMP_NUM_THREADS`

For the single-threaded `demucs.cpp.main`, my suggestion is `OMP_NUM_THREADS=$num_physical_cores`. On my 5950X system with 16 cores, execution time for a 4-minute song:
```
real 10m23.201s
user 29m42.190s
sys 4m17.248s
```

For the multi-threaded `demucs_mt.cpp.main`, using 4 `std::thread` and OMP threads = 4 (4x4 = 16 physical cores):
```
real 4m9.331s
user 18m59.731s
sys 3m28.465s
```

More than 2x faster for 4 threads. This is inspired by the parallelism strategy used in <https://freemusicdemixer.com>.
Binary file added .github/android-screenshot.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added .github/google-play-badge.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
61 changes: 12 additions & 49 deletions README.md
Original file line number Diff line number Diff line change
@@ -1,6 +1,16 @@
# demucs.cpp

C++17 implementation of the [Demucs v4 hybrid transformer](https://github.com/facebookresearch/demucs), a PyTorch neural network for music demixing. Similar project to [umx.cpp](https://github.com/sevagh/umx.cpp). This code powers my site <https://freemusicdemixer.com>.
C++17 library that implements the inference of the [Demucs v4 hybrid transformer model](https://github.com/facebookresearch/demucs), a PyTorch neural network for music demixing.

It uses only the standard library and the header-only library [Eigen](https://eigen.tuxfamily.org/index.php?title=Main_Page) as dependencies, making it suitable to compile and run on many platforms. It was designed for low-memory environments by sacrificing the speed of the Torch implementation.

Demucs.cpp powers my websites (<https://freemusicdemixer.com>, <https://pro.freemusicdemixer.com>) and now my new Android app [Music Demixer](https://play.google.com/store/apps/details?id=com.freemusicdemixer.pro) to bring Demucs to your pocket!

<a href="https://play.google.com/store/apps/details?id=com.freemusicdemixer.pro"><img src=".github/android-screenshot.png" width="128px" alt="music-demixer-android"/></a> <a href="https://play.google.com/store/apps/details?id=com.freemusicdemixer.pro"><img alt="google-play-badge" width="150px" src=".github/google-play-badge.png"/></a>

See my other project [umx.cpp](https://github.com/sevagh/umx.cpp) for a similar library for Open-Unmix.

### Library design

It uses [libnyquist](https://github.com/ddiakopoulos/libnyquist) to load audio files, the [ggml](https://github.com/ggerganov/ggml) file format to serialize the PyTorch weights of `htdemucs`, `htdemucs_6s`, and `htdemucs_ft` (4-source, 6-source, fine-tuned) to a binary file format, and [Eigen](https://eigen.tuxfamily.org/index.php?title=Main_Page) (+ OpenMP) to implement the inference. There are also programs for multi-threaded Demucs inference using C++11's `std::thread`.

Expand All @@ -14,48 +24,7 @@ It uses [libnyquist](https://github.com/ddiakopoulos/libnyquist) to load audio f
1. `demucs_mt.cpp.main`: run a single model, multi-threaded
1. `demucs_ft_mt.cpp.main`: run all four fine-tuned models, multi-threaded

### Multi-core, OpenMP, BLAS, etc.

:warning: `demucs.cpp` library code in `./src` **should not use any threading (e.g. pthread or OpenMP) except through the BLAS interface.** This is because demucs.cpp is compiled to a single-threaded WebAssembly module in <https://freemusicdemixer.com>.

If you have OpenMP and OpenBLAS installed, OpenBLAS might automatically use all of the threads on your machine, which doesn't always run the fastest. Use the `OMP_NUM_THREADS` environment variable to limit this. On my 16c/32t machine, I found `OMP_NUM_THREADS=16` to be the fastest. This matches the [Eigen recommendation](https://eigen.tuxfamily.org/dox/TopicMultiThreading.html) to use the same number of threads as physical cores:
>On most OS it is very important to limit the number of threads to the number of physical cores, otherwise significant slowdowns are expected, especially for operations involving dense matrices.
See the [BLAS benchmarks doc](./.github/BLAS_benchmarks.md) for more details.

### Multi-threading

There are two new programs, `demucs_mt.cpp.main` and `demucs_ft_mt.cpp.main` that use C++11 [std::threads](https://en.cppreference.com/w/cpp/thread/thread).

In the single-threaded programs:

* User supplies a waveform of length N seconds
* Waveform is split into 7.8-second segments for Demucs inference
* Segments are processed sequentially, where each segment inference can use >1 core with `OMP_NUM_THREADS`

In the multi-threaded programs:
* User supplies a waveform of length N seconds and a `num_threads` argument
* Waveform is split into `num_threads` sub-waveforms (of length M < N) to process in parallel with a 0.75-second overlap
* We always need overlapping segments in audio applications to eliminate [boundary artifacts](https://freemusicdemixer.com/under-the-hood/2024/02/23/Demucs-segmentation#boundary-artifacts-and-the-overlap-add-method)
* `num_threads` threads are launched to perform Demucs inference on the sub-waveforms in parallel
* Within each thread, the sub-waveform is split into 7.8-second segments
* Segments within a thread are still processed sequentially, where each segment inference can use >1 core with `OMP_NUM_THREADS`

For the single-threaded `demucs.cpp.main`, my suggestion is `OMP_NUM_THREADS=$num_physical_cores`. On my 5950X system with 16 cores, execution time for a 4-minute song:
```
real 10m23.201s
user 29m42.190s
sys 4m17.248s
```

For the multi-threaded `demucs_mt.cpp.main`, using 4 `std::thread` and OMP threads = 4 (4x4 = 16 physical cores):
```
real 4m9.331s
user 18m59.731s
sys 3m28.465s
```

More than 2x faster for 4 threads. This is inspired by the parallelism strategy used in <https://freemusicdemixer.com>.
See the [PERFORMANCE doc](./.github/PERFORMANCE.md) for details on multi-threading, external BLAS libraries, etc..

## Instructions

Expand Down Expand Up @@ -149,9 +118,3 @@ Encoder Status: 0
```

For the 6-source model, additional targets 4 and 5 correspond to guitar and piano.

## Dev tips

* make lint
* Valgrind memory error test: `valgrind --leak-check=full --show-leak-kinds=all --track-origins=yes --verbose ./demucs.cpp.main ../ggml-demucs/ggml-model-htdemucs-f16.bin ../test/data/gspi_stereo.wav ./demucs-out-cpp/`
* Callgrind + KCachegrind: `valgrind --tool=callgrind ./demucs.cpp.test --gtest_filter='*FreqDec*'`

0 comments on commit fedd352

Please sign in to comment.