| 1 |
.. SPDX-License-Identifier: GPL-2.0 |
| 2 |
|
| 3 |
============ |
| 4 |
Introduction |
| 5 |
============ |
| 6 |
|
| 7 |
The Linux compute accelerators subsystem is designed to expose compute |
| 8 |
accelerators in a common way to user-space and provide a common set of |
| 9 |
functionality. |
| 10 |
|
| 11 |
These devices can be either stand-alone ASICs or IP blocks inside an SoC/GPU. |
| 12 |
Although these devices are typically designed to accelerate |
| 13 |
Machine-Learning (ML) and/or Deep-Learning (DL) computations, the accel layer |
| 14 |
is not limited to handling these types of accelerators. |
| 15 |
|
| 16 |
Typically, a compute accelerator will belong to one of the following |
| 17 |
categories: |
| 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 devices |
| 21 |
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. This |
| 24 |
type of device can be stand-alone or an IP inside a SoC or a GPU. It will |
| 25 |
have on-board DRAM (to hold the DL topology), DMA engines and |
| 26 |
command submission queues (either kernel or user-space queues). |
| 27 |
It might also have an MMU to manage multiple users and might also enable |
| 28 |
virtualization (SR-IOV) to support multiple VMs on the same device. In |
| 29 |
addition, these devices will usually have some tools, such as profiler and |
| 30 |
debugger. |
| 31 |
|
| 32 |
- Training data-center - Similar to Inference data-center cards, but typically |
| 33 |
have more computational power and memory b/w (e.g. HBM) and will likely have |
| 34 |
a method of scaling-up/out, i.e. connecting to other training cards inside |
| 35 |
the server or in other servers, respectively. |
| 36 |
|
| 37 |
All these devices typically have different runtime user-space software stacks, |
| 38 |
that are tailored-made to their h/w. In addition, they will also probably |
| 39 |
include a compiler to generate programs to their custom-made computational |
| 40 |
engines. Typically, the common layer in user-space will be the DL frameworks, |
| 41 |
such as PyTorch and TensorFlow. |
| 42 |
|
| 43 |
Sharing code with DRM |
| 44 |
===================== |
| 45 |
|
| 46 |
Because this type of devices can be an IP inside GPUs or have similar |
| 47 |
characteristics as those of GPUs, the accel subsystem will use the |
| 48 |
DRM subsystem's code and functionality. i.e. the accel core code will |
| 49 |
be part of the DRM subsystem and an accel device will be a new type of DRM |
| 50 |
device. |
| 51 |
|
| 52 |
This will allow us to leverage the extensive DRM code-base and |
| 53 |
collaborate with DRM developers that have experience with this type of |
| 54 |
devices. In addition, new features that will be added for the accelerator |
| 55 |
drivers can be of use to GPU drivers as well. |
| 56 |
|
| 57 |
Differentiation from GPUs |
| 58 |
========================= |
| 59 |
|
| 60 |
Because we want to prevent the extensive user-space graphic software stack |
| 61 |
from trying to use an accelerator as a GPU, the compute accelerators will be |
| 62 |
differentiated from GPUs by using a new major number and new device char files. |
| 63 |
|
| 64 |
Furthermore, the drivers will be located in a separate place in the kernel |
| 65 |
tree - drivers/accel/. |
| 66 |
|
| 67 |
The accelerator devices will be exposed to the user space with the dedicated |
| 68 |
261 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 |
|
| 74 |
Getting Started |
| 75 |
=============== |
| 76 |
|
| 77 |
First, read the DRM documentation at Documentation/gpu/index.rst. |
| 78 |
Not only it will explain how to write a new DRM driver but it will also |
| 79 |
contain all the information on how to contribute, the Code Of Conduct and |
| 80 |
what is the coding style/documentation. All of that is the same for the |
| 81 |
accel subsystem. |
| 82 |
|
| 83 |
Second, make sure the kernel is configured with CONFIG_DRM_ACCEL. |
| 84 |
|
| 85 |
To expose your device as an accelerator, two changes are needed to |
| 86 |
be done in your driver (as opposed to a standard DRM driver): |
| 87 |
|
| 88 |
- Add the DRIVER_COMPUTE_ACCEL feature flag in your drm_driver's |
| 89 |
driver_features field. It is important to note that this driver feature is |
| 90 |
mutually exclusive with DRIVER_RENDER and DRIVER_MODESET. Devices that want |
| 91 |
to expose both graphics and compute device char files should be handled by |
| 92 |
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 easily |
| 96 |
set the correct function operations pointers structure. |
| 97 |
|
| 98 |
External References |
| 99 |
=================== |
| 100 |
|
| 101 |
email threads |
| 102 |
------------- |
| 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/[email protected]/>`_ - Oded Gabbay (2022) |
| 106 |
|
| 107 |
Conference talks |
| 108 |
---------------- |
| 109 |
|
| 110 |
* `LPC 2022 Accelerators BOF outcomes summary <https://airlied.blogspot.com/2022/09/accelerators-bof-outcomes-summary.html>`_ - Dave Airlie (2022) |