Figure 4-4: CNN architecture based on GPU

Download 3,22 Mb. Pdf ko'rish
bet	25/81
Sana	16.05.2024
Hajmi	3,22 Mb.
	#238917

1 ... 21 22 23 24 25 26 27 28 ... 81

Bog'liq
Alireza Fasih

Figure
4-4: CNN architecture based on GPU.
In many cases because of power consumption, size and price using computer for
processing is not a good idea; therefore we should choose another alternative platform like
embedded system, system on chip or FPGA board. In chapter 7 we have proposed a CNN
implementation on FPGA. In contrast to developing on GPU, traditional method of FPGA
designing is not very flexible and easy. There are many obstacle and reasons behind this
issue, such as low and middle level programming, hardware and software integration, low
level programming for peripherals I/O and leak of high level debugging tool. But new
technology in FPGA developers exist that called Impulse-C
TM
. This package and software
allows developers to write C-language for designing a digital system. Impulse-C directly
can optimize for XILINX
TM
FPGAs family from C-language. This tool can also integrate the
hardware and software as a mixed system. We used this system for capturing video from
camera daughter board which is connected to the FPGA board. Another main advantage of
this system is communication channels between software and hardware using
Fastest
Simple Link
(FSL) and
Processor Local Bus
(PLB). We design complete architecture of CNN
module in C in form of fixed-point. Impulse-C provided a very powerful debugger based on
GPU
Digital
Camera
Capture Video
Color Conversion
Normalization
GPU Global
Memory
2D Convolution
T
BU
= T
B
* U
2D Convolution
T
AY
= T
A
* Y
Accumulator
X(t) = T
BU
(t-1)+T
AY
(t-1) + I
Sigmoid Function
Y
=
|
| |
|
Kernel
Kernel
Kernel
Kernel
GPU

38
Microsoft Visual Studio
TM
editor and Eclipse IDE. During the design and debugging phase
we used this debugger for loading image into the CNN model and test the functionality of
different modules and complete system. After debugging system we generate the
optimized FPGA hardware and software interface and programmed on the FPGA by XILINX
EDK
TM
. Details of complete system are described in chapter 7.

Download 3,22 Mb.

1 ... 21 22 23 24 25 26 27 28 ... 81

Download 3,22 Mb.

Pdf ko'rish