Qualcomm Proposes New Linux DSP Accelerator Driver

Qualcomm has submitted a major new driver patch series to the Linux kernel that lays the groundwork for a modern DSP accelerator interface targeting the company’s Hexagon digital signal processors (DSPs), signalling a clearer path for standardised hardware offload support in open-source systems.

The new driver — called the Qualcomm DSP Accelerator (QDA) — was posted as a Request for Comments (RFC) series for inclusion in the Linux kernel’s accelerator “accel” subsystem. This marks a shift from Qualcomm’s older in-kernel FastRPC code and aims to provide a more unified, capable and maintainable architecture for developers and users working with DSPs on Qualcomm system-on-chips (SoCs).

DSPs are specialised processors built into many mobile and embedded platforms that handle real-time signal processing tasks such as audio, voice and sensor workloads. On Qualcomm SoCs, Hexagon DSPs have long powered tasks ranging from low-latency audio effects to offloading portions of modem or AI inference jobs, but until now their support in mainline Linux has been limited and fragmented. The new QDA driver seeks to change that by aligning DSP offload with the kernel’s growing ecosystem of compute accelerators.

A Standardised Interface for DSP Compute

At its core, the Qualcomm DSP Accelerator driver introduces a standard Device Resource Manager (DRM)-based accelerator interface that exposes DSP compute capabilities through devices such as /dev/accel/accelN. This approach mirrors other modern accelerator drivers that standardise interactions between user space and specialised hardware, making it easier for higher-level frameworks to submit compute jobs without bespoke, platform-specific APIs.

Key technical features of the QDA driver include:

• A GEM-based buffer management layer, which supports the import and export of DMA-BUF buffers for efficient sharing of memory between processes and hardware.

• IOMMU-based memory isolation, ensuring secure and isolated memory contexts per process, which is vital for robust multi-tenant environments.

• Implementation of the FastRPC protocol — the existing interprocessor communication method — to maintain compatibility with current user-space drivers.

• RPMsg transport support for reliable messaging between the host CPU and Hexagon DSP cores.

• Coverage for all of Qualcomm’s DSP domains — such as ADSP (Application DSP), CDSP (Compute DSP), SDSP (Sensor DSP) and GDSP (General DSP) — under a unified driver framework.
  

The initial patch series adds roughly 4,665 lines of C code to the kernel, encompassing documentation, kernel infrastructure, memory managers and IOCTL interfaces for DSP operations. An open-source user-space driver that interacts with QDA exists in a staging branch of Qualcomm’s FastRPC GitHub repository, bridging the gap between legacy tooling and the new kernel interface.

Replacing Legacy FastRPC Support

Qualcomm’s earlier strategy for DSP support in Linux centred on the FastRPC driver located under drivers/misc/, which provided primitives for communicating with DSPs but lacked integration with the kernel’s accelerator framework. The QDA driver’s emergence signals a concerted effort to migrate away from that legacy code towards a driver that various DSP use cases can build upon more consistently.

By conforming to the kernel’s accel subsystem and using common infrastructure such as GEM and DMA-BUF, QDA can offer better resource management and a clearer upgrade path for mainstream kernel inclusion. This is especially relevant as more SoCs diversify their compute resources, requiring standard interfaces that avoid fragmentation and reduce maintenance burdens.

Driver Architecture and Workflow

Under the QDA model, user-space applications will request a DSP compute context through the standard accelerator interface. Buffer objects created via GEM can then be shared securely with the hardware using DMA-BUF, and jobs are submitted to the DSP using the FastRPC protocol over an RPMsg transport layer. The IOMMU context banks ensure that each process’s memory mappings are appropriately isolated and managed for performance and safety.

This layered design positions QDA to support a wide range of workloads, including not just traditional signal processing but also AI inference, computer vision and other offloaded computations where low-power DSPs can handle tasks more efficiently than a general-purpose CPU.

Integration with the Mainline Linux Kernel

The patch submission is currently under review and discussion among kernel maintainers and subsystem experts. Laying the groundwork for QDA required not only the driver code but also documentation to guide its use, integration points and expected driver behaviour. Early patches added basic scaffolding such as Kconfig entries and Makefile support, while later patches built up the full suite of device registration, memory management and user-space interfaces.

The maintainers will likely evaluate the driver’s structure, API design and compatibility with both existing FastRPC user-space tools and the conventions of the Linux accelerator ecosystem. If accepted, QDA could become part of the kernel’s mainstream tree, improving DSP support for Linux distributions on Qualcomm-powered devices.

Broader Implications for Linux on Mobile and Embedded Platforms

Qualcomm’s QDA driver fits into a wider trend of integrating heterogeneous compute units — such as GPUs, NPUs and DSPs — under standard Linux subsystems that abstract hardware differences and enable portable software solutions. As open-source drivers for NPUs and other accelerators have matured in recent kernel releases, QDA extends that momentum into the DSP space, which has historically lagged behind in unified support.

For developers targeting audio-centric applications, real-time workloads or offload scenarios where power efficiency matters, having a standard DSP interface simplifies development and reduces reliance on proprietary drivers or platform-specific hacks. In addition, by leveraging the mainstream kernel accel infrastructure, QDA aligns Qualcomm’s DSP support with other open drivers that can benefit from shared tooling and community maintenance.

If accepted into upstream Linux, the Qualcomm DSP Accelerator driver could significantly improve the usability and performance of DSP-backed workloads on Linux systems based on Snapdragon and other Qualcomm SoCs — a noteworthy step forward for heterogeneous open-source computing on mobile and embedded platforms.