111 lines · plain
1.. SPDX-License-Identifier: GPL-2.02 3============4Introduction5============6 7The Linux compute accelerators subsystem is designed to expose compute8accelerators in a common way to user-space and provide a common set of9functionality.10 11These devices can be either stand-alone ASICs or IP blocks inside an SoC/GPU.12Although these devices are typically designed to accelerate13Machine-Learning (ML) and/or Deep-Learning (DL) computations, the accel layer14is not limited to handling these types of accelerators.15 16Typically, a compute accelerator will belong to one of the following17categories:18 19- Edge AI - doing inference at an edge device. It can be an embedded ASIC/FPGA,20 or an IP inside a SoC (e.g. laptop web camera). These devices21 are typically configured using registers and can work with or without DMA.22 23- Inference data-center - single/multi user devices in a large server. This24 type of device can be stand-alone or an IP inside a SoC or a GPU. It will25 have on-board DRAM (to hold the DL topology), DMA engines and26 command submission queues (either kernel or user-space queues).27 It might also have an MMU to manage multiple users and might also enable28 virtualization (SR-IOV) to support multiple VMs on the same device. In29 addition, these devices will usually have some tools, such as profiler and30 debugger.31 32- Training data-center - Similar to Inference data-center cards, but typically33 have more computational power and memory b/w (e.g. HBM) and will likely have34 a method of scaling-up/out, i.e. connecting to other training cards inside35 the server or in other servers, respectively.36 37All these devices typically have different runtime user-space software stacks,38that are tailored-made to their h/w. In addition, they will also probably39include a compiler to generate programs to their custom-made computational40engines. Typically, the common layer in user-space will be the DL frameworks,41such as PyTorch and TensorFlow.42 43Sharing code with DRM44=====================45 46Because this type of devices can be an IP inside GPUs or have similar47characteristics as those of GPUs, the accel subsystem will use the48DRM subsystem's code and functionality. i.e. the accel core code will49be part of the DRM subsystem and an accel device will be a new type of DRM50device.51 52This will allow us to leverage the extensive DRM code-base and53collaborate with DRM developers that have experience with this type of54devices. In addition, new features that will be added for the accelerator55drivers can be of use to GPU drivers as well.56 57Differentiation from GPUs58=========================59 60Because we want to prevent the extensive user-space graphic software stack61from trying to use an accelerator as a GPU, the compute accelerators will be62differentiated from GPUs by using a new major number and new device char files.63 64Furthermore, the drivers will be located in a separate place in the kernel65tree - drivers/accel/.66 67The accelerator devices will be exposed to the user space with the dedicated68261 major number and will have the following convention:69 70- device char files - /dev/accel/accel\*71- sysfs - /sys/class/accel/accel\*/72- debugfs - /sys/kernel/debug/accel/\*/73 74Getting Started75===============76 77First, read the DRM documentation at Documentation/gpu/index.rst.78Not only it will explain how to write a new DRM driver but it will also79contain all the information on how to contribute, the Code Of Conduct and80what is the coding style/documentation. All of that is the same for the81accel subsystem.82 83Second, make sure the kernel is configured with CONFIG_DRM_ACCEL.84 85To expose your device as an accelerator, two changes are needed to86be done in your driver (as opposed to a standard DRM driver):87 88- Add the DRIVER_COMPUTE_ACCEL feature flag in your drm_driver's89 driver_features field. It is important to note that this driver feature is90 mutually exclusive with DRIVER_RENDER and DRIVER_MODESET. Devices that want91 to expose both graphics and compute device char files should be handled by92 two drivers that are connected using the auxiliary bus framework.93 94- Change the open callback in your driver fops structure to accel_open().95 Alternatively, your driver can use DEFINE_DRM_ACCEL_FOPS macro to easily96 set the correct function operations pointers structure.97 98External References99===================100 101email threads102-------------103 104* `Initial discussion on the New subsystem for acceleration devices <https://lore.kernel.org/lkml/CAFCwf11=9qpNAepL7NL+YAV_QO=Wv6pnWPhKHKAepK3fNn+2Dg@mail.gmail.com/>`_ - Oded Gabbay (2022)105* `patch-set to add the new subsystem <https://lore.kernel.org/lkml/20221022214622.18042-1-ogabbay@kernel.org/>`_ - Oded Gabbay (2022)106 107Conference talks108----------------109 110* `LPC 2022 Accelerators BOF outcomes summary <https://airlied.blogspot.com/2022/09/accelerators-bof-outcomes-summary.html>`_ - Dave Airlie (2022)111