Audio offloading with RDMA Principles - Aimed at extending offloading capacities to SiMD, Coral EdgeTPU, NPU & GPU 2025 (c)RS

Part of the JIT Compiler Dongle works, Latency is an issue, Using the GPU & CPU & Dongles such as Coral.AI EdgeTPU & Intel Movidius for Mice, Camera's, Displays through HDMI & VESA DP,..

The future of device offloading is arranged here.:

The primary goal is to Enable Low Latency Hardware Offloading to enable software stacks like Bluetooth, Codecs & Device Drivers & Firmware, Dolby Audio, DTS, Realtek, Creative Logic & so on..

The point of the GPU / Processor offloading for Audio & Video & 3D Content such as Chrome & OS is that latency should be low,..

Primary compatibility with OS Stack { DirectX, Khronos, OpenCL, Direct Compute & Storage API }

P2P must not consume more than 5% of the total resource usage of the package,..

IO & DMA must be fast,..

The package should load & unload & process fast!

The core Module Kernel shall therefore be fast & small

Package core groups

Offloaded Processing 'KC'

Processing with endline automated returns from DirectCompute/OpenCL Kernel Package

The package looks as follows

Diagram 'KC'

(Kernel Module) Static Code Object Device work package = KSCOD = 'KC'

<<< Data flow >>>
Device + RAM : DMA : 'KC' CPU / Motherboard-IP Core 'KC' : DMA : Device + RAM 'KC' (Kernel) Static Code Object Device work package

If you are familiar with RDMA the package is:

Automated DMA to DMA with direct read & write from RAM Arrays,

Signalling is IO & DMA in cache pointers with P2P Pointer tables & Ethernet addressing & Direct unloading & loading on ETA or time of arrival..

Internal device ALU / DMA Cache frame & RAM usage/Storage buffer frame cache.

Rupert S

*

Offloading, How you run it:

Auto Device configurations should be easy!..

Automate common BitDepth & Buffer Length for quality &or the number of Hardware Accelerators

Auto should pick quality first, I mean 16Bit is not everybody's tea! & Dolby Atmos & DTS Default to 16Bit!,..

But if they did, device selection will be easy!

The scenario is that CPU/NPU/GPU/USB are not the only options for offloading,

Offloading latency & error statistics have to be cached, So if the option is auto,..

Automatic Device Selection & Multiplication

Selection of the BitDepth {64Bit, 32Bit, 24Bit, 16Bi} lead to a device flow path with advice in Control Panels,..

Example Offload options (there are more)

https://www.w3.org/TR/webcodecs/#definitions
https://www.w3.org/TR/webgpu/#intro
https://www.w3.org/TR/webnn/#intro

https://is.gd/JIT_RDMA

https://www.phoronix.com/news/Device-Memory-TCP-TX-Linux-6.16
https://www.phoronix.com/news/IO_uring-ZCRX-DMA-BUF

Rupert S