+ Advances in the Parallelization of Music and Audio Applications Eric Battenberg, David Wessel & Juan Colmenares

+ Overview Parallelism today in the popular interactive music languages Parallel Partitioned Convolution Accelerating Non-Negative Matrix Factorization (NMF) for use in audio source separation and music information retrieval and the importance of Selective, Embedded Just In Time Specialization (SEJITS) Real-time in the Tessellation OS A plea for more flexible I/0 with GPUs

+ Current Support for Parallelism is Copy-Based The widely used languages for music and audio applications are fundamentally sequential in character – this includes Max/MSP, PD, SuperCollider, and CHUCK among others. Limited multithreading One approach to exploiting multi-core processors is to run copies of the applications on separate cores. Max/MSP provides a useful multi-threading mechanism called poly~. PD provides PD~ each instance of which runs in a separate thread inside a PD patch.

+ Partitioned Convolution First real-time app in the Par Lab. Partitioned Convolution – an efficient way to do low-latency filtering with a long (> 1 sec) impulse response. Important in real-time reverb processing for environment simulation. Sound examples: Acoustic Guitar…in a giant mausoleum …convolved with a sine sweep Impulse response

+ Partitioned Convolution Convolution: a way to do linear filtering with a finite impulse response (FIR) filter. Direct convolution: For length L filter, O(L) ops per output point, zero delay. L can be greater than 100,000 samples (> 3 sec of audio) Block FFT Convolution: Only O(log(L)) ops per output point, but delay of L. How can we trade off between complexity and latency? FFT Complex Mult IFFT x x y y H H H = FFT(h)

+ Uniform Partitioned Convolution We would like the latency to be less than 10ms (512 samples) Cut an impulse response up into equal-sized blocks. Then we can use a parallel layout of Block FFT convolvers with delays to implement the filter. The latency is now N, and we still get complexity savings. L N delay(N) x x y y Block FFT Convolver

+ Frequency Delay Line Convolution We can also exploit linearity of the FFT so that only one FFT/IFFT is required. So the parallel Block FFT Convolver above becomes a Frequency Delay Line (FDL) Convolver: delay(N) x x y y FFT Complex Mult IFFT H1 Block FFT Convolver delay(N) + + x x y y Complex Mult H1 Complex Mult H2 Complex Mult H3 FFT IFFT Frequency Delay Line Convolver

+ Multiple FDL Convolution If L is big (e.g. > 100,000) and N is small (e.g. < 1000), our FDL will have 100’s of partitions to handle. We can connect multiple FDL’s in parallel to get the best of both worlds. x x delay(Nx6) delay(4Nx 4) FDL 1 FDL 2 FDL y y x x FDL y y

+ Scheduling Multiple FDLs FDLs are run in separate threads. Each is allowed to compute for a length of time corresponding to its block size. Synchronization is performed at the vertical lines.

+ Auto-Tuning for Real-Time We are not trying to only maximize throughput. We are trying to improve our ability to make real-time guarantees. For now, we estimate a Worst-Case Execution Time (WCET) for each size of FDL. Then we combine the FDLs that are most likely to meet their scheduling deadlines. In the future, we will use a notion of predictability along with more robust scheduling. We are finishing development on a Max/MSP object, Audio Unit plugin, and a portable standalone version of this.

+ Accelerating Non-Negative Matrix Factorization (NMF) NMF is widely used in audio source separation. The idea is to factor the time/frequency representation (spectogram) into source coupled spectral (W) and gain (H) matricies.

+ The Importance of SEJITS in Developing an Information Retrieval (MIR) Application Rather using a domain restricted language developers write in a full blown scripting language such as PYTHON or RUBY. Functions are selected by annotation as performance critical. If efficiency layer implementations of these functions are available appropriate code is generated and JIT compiled. If not the selected function is executed in the scripting language itself. The scripted implementation remains as the portable reference implementation.

With this simple music computer application we expect to initially show that Tessellation can provide acceptable performance and time predictability In cooperation with the OS Group 2nd-level RT scheduler A Cell A 2nd-level RT scheduler B Cell B Initial Cell Sound card Shell F OutputInput Music Program End-to-end Deadline Intermediate Deadline Audio Processing & Synthesis Engine Channe l F Most of the engine’s functionality Filter Parallel version of a partition-based convolution algorithm Audio Input Additional Cells A real-time application in Tessellation

2nd-level Scheduling Cell Tessellation Kernel (Partition Support) (*) Bottom part of the diagram was adapted from Liu and Asanovic, “Mitosys: ParLab Manycore OS Architecture,” Jan A) Cell and Space Partitioning  A Spatial Partition (or Cell) comprises a group of processors acting within a hardware boundary  Each cell receives a vector of basic resources – Some number of processors, a portion of physical memory, a portion of shared cache memory, and potentially a fraction of memory bandwidth  A cell may also receive – Exclusive access to other resources (e.g., certain hardware devices and raw storage partition) – Guaranteed fractional services (i.e., QoS guarantees) from other partitions (e.g., network service and file service) CPUCPU L1L1 L2BankL2Bank DRAMDRAM DRAM & I/O Interconnect L1 Interconnect CPUCPU L1L1 L2BankL2Bank DRAMDRAM CPUCPU L1L1 L2BankL2Bank DRAMDRAM CPUCPU L1L1 L2BankL2Bank DRAMDRAM CPUCPU L1L1 L2BankL2Bank DRAMDRAM CPUCPU L1L1 L2BankL2Bank DRAMDRAM (+) Fraction of memory bandwidth

+ Time-sensitive Network Subsystem Time-sensitive Network Subsystem Network Service (Net Partition) Network Service (Net Partition) Input device (Pinned/TT Partition) Input device (Pinned/TT Partition) Graphical Interface (GUI Partition) Graphical Interface (GUI Partition) Audio-processing / Synthesis Engine (Pinned/TT partition) Audio-processing / Synthesis Engine (Pinned/TT partition) Output device (Pinned/TT Partition) Output device (Pinned/TT Partition) GUI Subsystem Communication with other audio-processing nodes Music program Preliminary Example of Music Application

+ A plea for more flexible GPU I/O

+ Thanks for your attention.

+ Reserve Slides

+ Tessellation OSTessellation: 19 Nov emb er 12th, Tessellation in Server Environment DiskI/ODrivers OtherDevices Network QoS MonitorAndAdapt Persistent Storage & Parallel File System Large Compute-Bound Application Large I/O-Bound Application DiskI/ODrivers OtherDevices Network QoS MonitorAndAdapt Persistent Storage & Parallel File System Large Compute-Bound Application Large I/O-Bound Application DiskI/ODrivers OtherDevices Network QoS MonitorAndAdapt Persistent Storage & Parallel File System Large Compute-Bound Application Large I/O-Bound Application DiskI/ODrivers OtherDevices Network QoS MonitorAndAdapt Persistent Storage & Parallel File System Large Compute-Bound Application Large I/O-Bound Application QoSGuarantees Cloud Storage BW QoS QoSGuarantees QoSGuarantees QoSGuarantees