[lkml]   [2021]   [Feb]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Patch in this message
Subject[PATCH v5 01/34] Add Vision Processing Unit (VPU) documentation.
From: mark gross <>

The Intel VPU needs a complicated SW stack to make it work. Add a
directory to hold VPU related documentation including an architectural
overview of the SW stack that the patches implement.

Cc: Jonathan Corbet <>
Signed-off-by: Mark Gross <>
Documentation/index.rst | 1 +
Documentation/vpu/index.rst | 16 ++
Documentation/vpu/vpu-stack-overview.rst | 270 +++++++++++++++++++++++
3 files changed, 287 insertions(+)
create mode 100644 Documentation/vpu/index.rst
create mode 100644 Documentation/vpu/vpu-stack-overview.rst

diff --git a/Documentation/index.rst b/Documentation/index.rst
index 5888e8a7272f..81a02f2af939 100644
--- a/Documentation/index.rst
+++ b/Documentation/index.rst
@@ -137,6 +137,7 @@ needed).
+ vpu/index

Architecture-agnostic documentation
diff --git a/Documentation/vpu/index.rst b/Documentation/vpu/index.rst
new file mode 100644
index 000000000000..7e290e048910
--- /dev/null
+++ b/Documentation/vpu/index.rst
@@ -0,0 +1,16 @@
+.. SPDX-License-Identifier: GPL-2.0-only
+Vision Processor Unit Documentation
+This documentation contains information for the Intel VPU stack.
+.. class:: toc-title
+ Table of contents
+.. toctree::
+ :maxdepth: 2
+ vpu-stack-overview
diff --git a/Documentation/vpu/vpu-stack-overview.rst b/Documentation/vpu/vpu-stack-overview.rst
new file mode 100644
index 000000000000..1fe9ce423177
--- /dev/null
+++ b/Documentation/vpu/vpu-stack-overview.rst
@@ -0,0 +1,270 @@
+.. SPDX-License-Identifier: GPL-2.0
+Intel VPU architecture
+The Intel Movidius acquisition has developed a Vision Processing Unit (VPU)
+roadmap of products starting with Keem Bay (KMB). The hardware configurations
+the VPU can support include:
+1. Standalone smart camera that does local Computer Vision (CV) processing in
+ camera
+2. Standalone appliance or signel board computer connected to a network and
+ tethered cameras doing local CV processing
+3. Embedded in a USB dongle or M.2 as an CV accelerator.
+4. Multiple VPU enabled SOC's on a PCIe card as a CV accelerator in a larger IA
+ box or server.
+Keem Bay is the first instance of this family of products. This document
+provides an architectural overview of the software stack supporting the VPU
+enabled products.
+Keem Bay (KMB) is a Computer Vision AI processing SoC based on ARM A53 CPU that
+provides Edge neural network acceleration (inference) and includes a Vision
+Processing Unit (VPU) hardware. The ARM CPU SubSystem (CPUSS) interfaces
+locally to the VPU and enables integration/interfacing with a remote host over
+PCIe or USB or Ethernet interfaces. The interface between the CPUSS and the VPU
+is implemented with hardware FIFOs (Control) and coherent memory mapping (Data)
+such that zero copy processing can happen within the VPU.
+The KMB can be used in all 4 of the above classes of designs.
+We refer to the 'local host' as being the ARM part of the SoC, while the
+'remote host' as the IA system hosting the KMB device(s). The KMB SoC boots
+from an eMMC via uBoot and ARM Linux compatible device tree interface with an
+expectation to fully boot within hundreds of milliseconds. There is also
+support for downloading the kernel and root file system image from a remote
+The eMMC can be updated with standard Mender update process.
+The VPU is started and controlled from the A53 local host. Its firmware image
+is loaded using the drive firware helper KAPI's.
+The VPU IP firware payload consists of a SPARC ISA RTEMS bootloader and/or
+application binary.
+The interface allowing (remote or local) host clients to access VPU IP
+capabilities is realized through an abstracted programming model, which
+provides Remote Proxy APIs for a host CPU application to dynamically create and
+execute CV and NN workloads on the VPU. All frameworks exposed through
+programming model’s APIs are contained in the pre-compiled standard firmware
+There is a significant software stack built up to support KMB and the use
+cases. The rest of this documentation provides an overview of the components
+of the stack.
+Keem Bay IPC
+Directly interfaces with the KMB hardware FIFOs to provide zero copy processing
+from the VPU. It implements the lowest level protocol for interacting with the
+The Keem Bay IPC mechanism is based on shared memory and hardware FIFOs.
+Specifically there are:
+* Two 128-entry hardware FIFOs, one for the CPU and one for the VPU.
+* Two shared memory regions, used as memory pool for allocating IPC buffers.
+An IPC channel is a software abstraction allowing communication multiplexing,
+so that multiple applications / users can concurrently communicate with the
+VPU. IPC channels area conceptually similar to socket ports.
+There are a total of 1024 channels, each one identified by a channel ID,
+ranging from 0 to 1023.
+Channels are divided in two categories:
+* High-Speed (HS) channels, having IDs in the 0-9 range.
+* General-Purpose (GP) channels, having IDs in the 10-1023 range.
+HS channels have higher priority over GP channels and can be used by
+applications requiring higher throughput or lower latency.
+Since all the channels share the same hardare resources (i.e., the hardware
+FIFOs and the IPC memory pools), the Keem Bay IPC driver uses software queues
+to give a higher priority to HS channels.
+The driver supports a build-time configurable number of communication channels
+defined in a so-called Channel Mapping Table.
+An IPC channel is full duplex: a pending operation from a certain channel does
+not block other operations on the same channel, regardless of their operation
+mode (blocking or non-blocking).
+Operation mode is individually selectable for each channel, per operation
+direction (read or write). All operations for that direction comply to
+Keem Bay-VPU-IPC
+This is the MMIO driver of the VPU IP block inside the SOC. It is a control
+driver mapping IPC channel communication to Xlink virtual channels.
+This driver provides the following functionality to other drivers in the
+communication stack:
+* VPU IP execution control (firmware load, start, reset)
+* VPU IP event notifications (device connected, device disconnected, WDT event)
+* VPU IP device status query (OFF, BUSY, READY, ERROR, RECOVERY)
+* Communication via the IPC protocol (wrapping the Keem Bay IPC driver and
+ exposing it to higher level Xlink layer)
+In addition to the above, the driver exposes SoC information (like stepping,
+device ID, etc.) to user-space via sysfs.
+This driver depends on the 'Keem Bay IPC' driver, which enables the Keem Bay
+IPC communication protocol.
+The driver uses the Firmware API to load the VPU firmware from user-space.
+This component implements the IPC specific Xlink protocol. It maps channel
+IDs to hardware FIFO entries, using the Keem Bay VPU IPC driver.
+Some of the main functions this driver provides:
+* establishing a connection with an IPC device
+* obtaining a list with the available devices
+* obtaining the status for a device
+* booting a device
+* resetting a device
+* opening and closing channels
+* issuing read and write operations
+This component implements an abstracted set of control and communication APIs
+based on channel identification. It is intended to support VPU technology both
+at SoC level as well as at IP level, over multiple interfaces.
+It provides symmetrical services, where the producer and the consumer have
+the same privileges.
+Xlink driver has the ability to abstract several types of communication
+channels underneath, allowing the usage of different interfaces with the same
+function calls.
+Xlink services are available to both kernel and user space clients and include:
+* interface abstract control and communication API
+* multi device support
+* concurrent communication across 4096 communication channels (from 0 to
+ 0xFFF), with customizable properties
+* full duplex channels with multiprocess and multithread support
+* channel IDs can be mapped to desired physical interface (PCIe, USB, ETH, IPC)
+ via a Channel Mapping Table
+* asynchronous fast passthrough mode: remote host data packets are directly
+ dispatched using interrupt systems running on local host to IPC calls for low
+ overhead
+* channel handshaking mechanism for peer to peer communication, without the
+ need of static channel preallocation
+* channel resource management
+* asynchronous data and device notifications to subscribers
+Xlink transports: PCIe, USB, ETH, IPC, XLink-PCIe
+This is an endpoint driver that maps Xlink channel IDs to PCIe channels.
+This component ensures (remote)host-to-(local)host communication, and VPU IP
+communication via an asynchronous passthrough mode, where PCIe data loads are
+directly dispatched to Xlink-IPC.
+The component builds and advertises Device IDs that are used by local host
+application in case of multi device scenarios.
+This is an endpoint driver that maps Xlink channel IDs to bidirectional
+USB endpoints and supports CDC USB class protocol. More than one Xlink channels
+can be mapped to a single USB endpoint.
+This component ensures host-to-host communication, and, as well, asynchronous
+passthrough communication, where USB transfer packets are directly dispatched
+to Xlink-IPC.
+The component builds and advertises Device IDs that can are used by local host
+application in case of multi device scenarios.
+This is an endpoint driver that maps Xlink channel IDs to Ethernet
+This component ensures host-to-host communication, and, as well, asynchronous
+passthrough communication, where Ethernet data loads are directly dispatched to
+The component builds and advertises Device IDs that can are used by local host
+application in case of multi device scenarios.
+Assorted drivers that depend on this stack:
+The Intel Edge.AI Computer Vision platforms have to be monitored using platform
+devices like sensors, fan controller, IO expander etc. Some of these devices
+are memory mapped and some are I2C-based. None of these devices is directly
+accessible to the host.
+The host here refers to the server to which the vision accelerators are
+connected over PCIe Interface. The Host needs to do a consolidated action based
+on the parameters of platform devices. In general, most of the standard devices
+(includes sensors, fan controller, IO expander etc) are I2C/SMBus based and are
+used to provide the status of the accelerator. Standard drivers for these
+devices are available based on I2C/SMBus APIs.
+Instead of changing the sensor drivers to adapt to PCIe interface, a generic
+I2C adapter "Xlink-SMBus" which underneath uses Xlink as physical medium is
+used. With Xlink-SMBus, the drivers for the platform devices don't need to
+undergo any interface change.
+Thermal sensor driver for exporting thermal events to the local Arm64 host as
+well as to the remote X86 host if in the PCIe add-in CV accelerator
+The driver receives the junction temperature from different heating points
+inside the SOC. The driver will receive the temperature on SMBus connection and
+forward over Xlink-smb when in a remote host configuration.
+In Keem Bay, the four thermal junction temperature points are Media Subsystem
+(mss), Neral Network subsystem (nce), Compute subsystem (cse) and SOC(maximum
+of mss, nce and cse).
+- Exports details of temperature sensor, current sensor and fan controller
+ present in Intel Edge.AI Computer Vision platforms to IA host.
+- Enable Time sync of Intel Edge.AI Computer Vision platform with IA host.
+- Handles device connect and disconnect events.
+- Receives slave address from the IA host for memory mapped thermal sensors
+ present in SoC (Documentation/hwmon/intel_tsens_sensors.rst).
+- Registers I2C slave device for slaves present in Intel Edge.AI Computer
+ Vision platform
+VPUMGR (VPU Manager)
+Bridges firmware on VPU side and applications on CPU user-space, it assists
+firmware on VPU side serving multiple user space application processes on CPU
+side concurrently while also performing necessary data buffer management on
+behalf of VPU IP.
 \ /
  Last update: 2021-02-06 05:27    [W:0.289 / U:3.952 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site