Why GEMM is at the heart of deep learning

I spend most of my time worrying about how to make deep learning with neural networks faster and more power efficient. In practice that means focusing on a function called GEMM. It’s part of the BLAS (Basic Linear Algebra Subprograms) library that was first created in 1979, and until I started trying to optimize neural networks I’d never heard of it.
继续阅读Why GEMM is at the heart of deep learning

Windows 7 系统电脑安装RNDIS驱动

本教程小编和大家分享 Windows 7 系统电脑安装RNDIS驱动的正确方法，RNDIS驱动是什么？ Windows 7 系统驱动RNDIS是远端网络驱动接口协议，设备通过USB方式同主机连接，模拟网络连接以便用于下载和调试工作。但是很多 Windows 7 系统用户安装RNDIS的设备时失败，遇到无法安装的问题，所以小编给大家介绍 Windows 7 系统电脑安装RNDIS驱动的正确方法。

继续阅读Windows 7 系统电脑安装RNDIS驱动

粗略判断Shader每条代码的成本

GPU IS a processor (graphics proccessing unit). Anywho, i remember seeing somewhere that in geforce 6 series cards its a signle cycle (maybe i was just dreaming :-p) but i have that memory

radeon x800 has it anyways
EDIT:

Quote:

ORIGINALLY AT: http://gear.ibuypower.com/GVE/Store/ProductDetails.aspx?sku=VC-POWERC-147
Smartshader HD•Support for Microsoft® DirectX® 9.0 programmable vertex and pixel shaders in hardware
• DirectX 9.0 Vertex Shaders
- Vertex programs up to 65,280 instructions with flow control
- Single cycle trigonometric operations (SIN & COS)
• Direct X 9.0 Extended Pixel Shaders
- Up to 1,536 instructions and 16 textures per rendering pass
- 32 temporary and constant registers
- Facing register for two-sided lighting
- 128-bit, 64-bit & 32-bit per pixel floating point color formats
- Multiple Render Target (MRT) support
• Complete feature set also supported in OpenGL® via extensions

继续阅读粗略判断Shader每条代码的成本

Android Gradle Plugin源码解析之externalNativeBuild

在Android Studio 2.2开始的Android Gradle Plugin版本中，Google集成了对cmake的完美支持，而原先的ndkBuild的方式支持也变得更加良好。这篇文章就来说说Android Gradle Plugin与交叉编译之间的一些事，即externalNativeBuild相关的task，主要是解读一下gradle构建系统相关的源码。

继续阅读Android Gradle Plugin源码解析之externalNativeBuild

Overriding a default option(…) value in CMake from a parent CMakeLists.txt

子 CMakeLists.txt

父 CMakeLists.txt

执行如下命令的时候：

会观察到生成的配置文件中 BUILD_FOR_ANDROID 不一定能生效。

需要如下配置才行：
父 CMakeLists.txt

参考链接

Use ccache with CMake for faster compilation

C and C++ compilers aren’t the fastest pieces of software out there and there’s no lack of programmer jokes based on tedium of waiting for their work to complete.

There are ways to fix the pain though - one of them is ccache. CCache improves compilation times by caching previously built object files in private cache and reusing them when you’re recompiling same objects with same parameters. Obviously it will not help if you’re compiling the code for the first time and it also won’t help if you often change compilation flags. Most C/C++ development however involves recompiling same object files with the same parameters and ccache helps alot.

For illustration, here’s the comparison of first and subsequent compilation times of a largish C++ project:

Original run with empty cache:

Recompilation with warm cache:

Installation

CCache is available in repositories on pretty much all distributions. On OS X use homebrew:

and on Debian-based distros use apt:

CMake configuration

After ccache is installed, you need to tell CMake to use it as a wrapper for the compiler. Add these lines to your CMakeLists.txt:

Rerun cmake and next make should use ccache for wrapper.

Usage with Android NDK

CCache can even be used on Android NDK - you just need to export NDK_CCACHE environment variable with path to ccache binary. ndk-build script will automatically use it. E.g.

(Note that on Debian/Ubuntu the path will probably be /usr/bin/ccache)

CCache statistics

To see if ccache is really working, you can use ccache -s command, which will display ccache statistics:

On second and all subsequent compilations the “cache hit” values should increase and thus show that ccache is working.

参考链接

Use ccache with CMake for faster compilation

macOS Mojave(10.14.3)系统QEMU虚拟机运行Clockwork OS


编译 u-boot

编译 Linux 内核


手工编译 qemu

可惜到这一步了，还是没办法成功运行系统。

参考链接

Using QEMU to emulate a Raspberry Pi

If you're building software for the Raspberry Pi (like I sometimes do), it can be a pain to have to constantly keep Pi hardware around and spotting Pi-specific problems can be difficult until too late.

One option (and the one I most like) is to emulate a Raspberry Pi locally before ever hitting the device. Why?

  • Works anywhere you can install QEMU
  • No hardware setup needed (no more scratching around for a power supply)
  • Faster feedback cycle compared to hardware
  • I can use Pi software (like Raspbian) in a virtual context
  • I can prep my "virtual Pi" with all the tools I need regardless of my physical Pi's use case

Given I'm next-to-useless at Python, that last one is pretty important as it allows me to install every Python debugging and testing tool known to man on my virtual Pi while my end-product hardware stays comparatively pristine.

Getting started

First, you'll need a few prerequisites:

QEMU (more specifically qemu-system-arm)

You can find all the packages for your chosen platform on the QEMU website and is installable across Linux, macOS and even Windows.

Raspbian

Simply download the copy of Raspbian you need from the official site. Personally, I used the 2018-11-13 version of Raspbian Lite, since I don't need an X server.

Kernel

Since the standard RPi kernel can't be booted out of the box on QEMU, we'll need a custom kernel. We'll cover that in the next step.

Preparing

Get your kernel

First, you'll need to download a kernel. Personally, I (along with most people) use the dhruvvyas90/qemu-rpi-kernel repository's kernels. Either clone the repo:

or download a kernel directly:

or download a snapshot from my website directly:

For the rest of these steps I'm going to be using the kernel-qemu-4.4.34-jessiekernel, so update the commands as needed if you're using another version.

Filesystem image

This step is optional, but recommended

When you download the Raspbian image it will be in the raw format, a plain disk image (generally with an .img extension).

A more efficient option is to convert this to a qcow2 image first. Use the qemu-imgcommand to do this:

Now we can also easily expand the image:

You can check on your image using the qemu-img info command

Starting

You've got everything you need now: a kernel, a disk image, and QEMU!

Actually running the virtual Pi is done using the qemu-system-arm command and it can be quite complicated. The full command is this (don't worry it's explained below):

如果需要指定上网方式的话，执行如下命令：

So, in order:

  • sudo qemu-system-arm: you need to run QEMU as root
  • -kernel: this is the path to the QEMU kernel we downloaded in the previous step
  • -append: here we are providing the boot args direct to the kernel, telling it where to find it's root filesytem and what type it is
  • -hda: here we're attaching the disk image itself
  • -cpu/-m: this sets the CPU type and RAM limit to match a Raspberry Pi
  • -M: this sets the machine we are emulating. versatilepb is the 'ARM Versatile/PB' machine
  • -no-reboot: just tells QEMU to exit rather than rebooting the machine
  • -serial: redirects the machine's virtual serial port to our host's stdio
  • -net: this configures the machine's network stack to attach a NIC, use the user-mode stack, connect the host's vnet0 TAP device to the new NIC and don't use config scripts.

If it's all gone well, you should now have a QEMU window pop up and you should see the familiar Raspberry Pi boot screen show up.

Now, go get yourself a drink to celebrate, because it might take a little while.

Networking

Now, that's all well and good, but without networking, we may as well be back on hardware. When the machine started, it will have attached a NIC and connected it to the host's vnet0 TAP device. If we configure that device with an IP and add it to a bridge on our host, you should be able to reliably access it like any other virtual machine.

(on host) Find a bridge and address

This will vary by host, but on my Fedora machine, for example, there is a pre-configured virbr0 bridge interface with an address in the 192.168.122.0/24 space:

I'm going to use this bridge and just pick a static address for my Pi: 192.168.122.200

Reusing an existing (pre-configured) bridge means you won't need to sort your own routing

(in guest) Configure interface

NOTE: I'm assuming Stretch here.

Open /etc/dhcpcd.conf in your new virtual Pi and configure the eth0 interface with a static address in your bridge's subnet. For example, for my bridge:

You may need to reboot for this to take effect

(in host) Add TAP to bridge

Finally, add the machine's TAP interface to your chosen bridge with the brctl command:

Now, on your host, you should be able to ping 192.168.122.200 (or your Pi's address).

Set up SSH

Now, in your machine, you can run sudo raspi-config and enable the SSH server (in the "Interfacing Options" menu at time of writing).

Make sure you change the password from default while you're there!

Finally, on your host, run ssh-copy-id pi@192.168.122.200 to copy your SSH key into the Pi's pi user and you can now SSH directly into your Pi without a password prompt.

参考链接

Using QEMU to emulate a Raspberry Pi

Simple ARM NEON optimized sin, cos, log and exp

This is the sequel of the single precision SSE optimized sin, cos, log and exp that I wrote some time ago. Adapted to the NEON fpu of my pandaboard. Precision and range are exactly the same than the SSE version, so I won't repeat them.

The code

The functions below are licensed under the zlib license, so you can do basically what you want with them.

  • neon_mathfun.h source code for sin_ps, cos_ps, sincos_ps, exp_ps, log_ps, as straight C.
  • neon_mathfun_test.c Validation+Bench program for those function. Do not forget to run it once.

Performance

Results on a pandaboard with a 1GHz dual-core ARM Cortex A9 (OMAP4), using gcc 4.6.1

command line: gcc -O3 -mfloat-abi=softfp -mfpu=neon -march=armv7-a -mtune=cortex-a9 -Wall -W neon_mathfun_test.c -lm

So performance is not stellar. I recommend to use gcc 4.6.1 or newer as it generates much better code than previous (gcc 4.5) versions -- almost 20% faster here. I believe rewriting these functions in assembly would improve the performance by 30%, and should not be very hard as the ARM and NEON asm is quite nice and easy to write -- maybe I'll do it. Computing two SIMD vectors at once would also help to improve a lot the performance as there are enough registers on NEON, and it would reduce the dependancies between neon instructions.

Note also that I have no idea of the performance on a Cortex A8 -- it may be extremely bad, I don't know.

Comparison with an Intel Atom

For comparison purposes, here is the performance of the SSE version on a single core Intel Atom N270 running at 1.66GHz

command line: cl.exe /arch:SSE /O2 /TP /MD sse_mathfun_test.c (this is msvc 2010)

The number of cycles is quite similar -- but the atom has a higher clock..

Last modified: 2011/05/29

参考链接

Simple ARM NEON optimized sin, cos, log and exp