1 .. SPDX-License-Identifier: GPL-2.0
2
3 ============
4 Introduction
5 ============
6
7 The Linux compute accelerators subsystem is de
8 accelerators in a common way to user-space and
9 functionality.
10
11 These devices can be either stand-alone ASICs
12 Although these devices are typically designed
13 Machine-Learning (ML) and/or Deep-Learning (DL
14 is not limited to handling these types of acce
15
16 Typically, a compute accelerator will belong t
17 categories:
18
19 - Edge AI - doing inference at an edge device.
20 or an IP inside a SoC (e.g. laptop web camer
21 are typically configured using registers and
22
23 - Inference data-center - single/multi user de
24 type of device can be stand-alone or an IP i
25 have on-board DRAM (to hold the DL topology)
26 command submission queues (either kernel or
27 It might also have an MMU to manage multiple
28 virtualization (SR-IOV) to support multiple
29 addition, these devices will usually have so
30 debugger.
31
32 - Training data-center - Similar to Inference
33 have more computational power and memory b/w
34 a method of scaling-up/out, i.e. connecting
35 the server or in other servers, respectively
36
37 All these devices typically have different run
38 that are tailored-made to their h/w. In additi
39 include a compiler to generate programs to the
40 engines. Typically, the common layer in user-s
41 such as PyTorch and TensorFlow.
42
43 Sharing code with DRM
44 =====================
45
46 Because this type of devices can be an IP insi
47 characteristics as those of GPUs, the accel su
48 DRM subsystem's code and functionality. i.e. t
49 be part of the DRM subsystem and an accel devi
50 device.
51
52 This will allow us to leverage the extensive D
53 collaborate with DRM developers that have expe
54 devices. In addition, new features that will b
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-
61 from trying to use an accelerator as a GPU, th
62 differentiated from GPUs by using a new major
63
64 Furthermore, the drivers will be located in a
65 tree - drivers/accel/.
66
67 The accelerator devices will be exposed to the
68 261 major number and will have the following c
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 Documenta
78 Not only it will explain how to write a new DR
79 contain all the information on how to contribu
80 what is the coding style/documentation. All of
81 accel subsystem.
82
83 Second, make sure the kernel is configured wit
84
85 To expose your device as an accelerator, two c
86 be done in your driver (as opposed to a standa
87
88 - Add the DRIVER_COMPUTE_ACCEL feature flag in
89 driver_features field. It is important to no
90 mutually exclusive with DRIVER_RENDER and DR
91 to expose both graphics and compute device c
92 two drivers that are connected using the aux
93
94 - Change the open callback in your driver fops
95 Alternatively, your driver can use DEFINE_DR
96 set the correct function operations pointers
97
98 External References
99 ===================
100
101 email threads
102 -------------
103
104 * `Initial discussion on the New subsystem for<
105 * `patch-set to add the new subsystem <https://
106
107 Conference talks
108 ----------------
109
110 * `LPC 2022 Accelerators BOF outcomes summary
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.