International Conference on Advances in Computing, Communication Control and Networking (ICACCCN2018)

GPU Computing Revolution: CUDA*

Ramandeep Singh Dehal Chirag Munjal Arquish Ali Ansari Anup Singh Kushwaha
Student, DoCST, Student, DoCST, Student, DoCST, Assistant Professor, DoCST,
MRU Faridabad, MRU Faridabad, MRU Faridabad, MRU Faridabad,
raman19930@gmail.com 123.munjal@gmail.com aliarqish@gmail.com anup.usit@gmail.com

Abstract—In this research paper, we have studied Compute

Unified Device Architecture which was developed by NVIDIA for OpenGL and CUDA
its GPUs. It is a graphic library that provides a set of APIs’ OpenGL which is a graphic library provides a set of APIs’
which lets us take advantage of the GPU to render graphics on which lets you take advantage of the GPU to render graphics
the computer screen. We have listed the advantages of CUDA
over OpenGL & discussed the CUDA process flow & model
on the computer screen. OpenGL API contains functions that
overview leading to the reviews on its Applications i.e. MRI, compile to be run on the GPUs and the code to be run is
Ultrasound, Tomography, X-Rays, CT Scan, DWT-PCA Based copied into the GPU and then executed. The program host part
Image Fusion, Morphological Image Processing, Parallelizing does this copying in order to take advantage of the GPU. By
Multi-Sensor, Pyramidal image blending, parallel wavelet the use of standard compilers, the code will be executed only
algorithm on GPU using CUDA. on the CPU. API’s present in the library helps you take
advantage of this.
Keywords—CUDA, GPU, Thread, Block, Nvidia, Kernel, CPU, X- Earlier it was thought that GPU was useful only for
ray, OpenGL, CBCT, MRI, Unified, Radiation, Histogram, graphics processing and was used for this purpose only. But in
Tomography, GeForce, Morphological, Molecular & Vertebra.
the recent research, it has been discovered that the power of
parallel execution of code can also be used for many other
I. INTRODUCTION problems which include image processing, Bioinformatics,
Molecular Dynamics and much more. The main goal of
CUDA otherwise called Compute Unified Device CUDA is to provide the ability for doing general purpose
Architecture is a parallel computing platform based on a programming on GPU, and its mainly targeted for the
programming model called CUDA C. NVIDIA family of GPUs. But CUDA doesn’t give API’s
It was introduced in 2007 in the 8th series of the Nvidia specifically for the rendering of graphics and its processing, so
Graphic Card which had TESLA Micro-architecture. After you could re-implement this stuff in CUDA and create your
that, it is also being used in FERMI, KEPLER, MAXWELL & own graphics library with a smarter way and that is OpenGL.
now PASCAL Micro-Architecture Graphic Processing Units CUDA like, we already mentioned is targeted at NVIDIA
(GPUs). Nvidia develops a different API than OpenGL graphics cards, while OpenGL will drivers compile the code at
because OpenGL was very focused on the solving specific runtime and target it for many other graphics cards as well.
rendering problems in GPU that include 2D and 3D rendering.
Nvidia wanted an API that can focus on other problems like Parallel Computing Experience With CUDA
Image Processing Algorithm (such as edge detection, Garland et.al discussed the experience of parallel
removing clouds, histogram equalization, and Discrete Cosine computing with CUDA. Increase in the parallel process is
Transform(DCT) encode and decode etc), Bioinformatics, becoming a primary goal in the CPUs rather than increasing
Molecular Dynamics etc. With the help of parallel computing the base clock speed. The old NVIDIA GPUs had multicore
API it was easy for a graphic card to perform a task processor chips, scaling from 8 to 240 cores. With the arrival
efficiently. of first GPU in view of NVIDIA's Tesla bound unified
It was designed and used by Nvidia Corporations. It engineering "NVIDIA GeForce 8800"— it has turned out to
empowers an expansion in the processing execution of the be conceivable to program GPU processors specifically and
Graphic Card by harnessing its energy and is very effective at enormously as a parallel processor instead of basically as
programming multi-threaded many-cores GPUs. In industry illustrations API accelerators.
and the scholarly community, it is utilized to accomplish The CUDA parallel programming model comprises of 2 key
dramatic speedups on creation and research codes. CUDA C is plan models. In the principal show, it plans to expand a
the Application Programming Interface which supports the C standard consecutive programming dialect, i.e. C/C++ with a
and the Fortran programming language. The CUDA C base arrangement of reflections for communicating
compiler is built on C++. parallelism. In the second model, it is outlined in a route for
The advantages of CUDA is that it provides shared composing very extensible parallel code that can keep running
memory, it is cost effective & it helps Nvidia to meet the crosswise over a huge number of simultaneous strings and
demand of gaming industries in the improvement of GPU. many processor centers. The CUDA display normally directs
Although it has disadvantages as well. It is not effective for the software engineer to compose parallel projects that
PC & despite having 100 of “cores” CUDA is not as flexible straightforwardly and productively scale over the diverse
as CPUs.

ISBN: 978-1-5386-4119-4/18/$31.00 ©2018 IEEE 197

ed licensed use limited to: CENTRE FOR DEVELOPMENT OF ADVANCED COMPUTING - CDAC - KOLKATA. Downloaded on February 24,2022 at 11:41:00 UTC from IEEE Xplore. Restriction
 International Conference on Advances in Computing, Communication Control and Networking (ICACCCN2018)

levels of parallelism. The parallel programming uses hybrid CUDA, Kernels, and Threads :- The CUDA programming
CUDA, OpenMP and MPI (Message Passing Interface) model is a model in which both CPU and GPU are used.
programming [1]. The DEVICE = GPU which execute many threads in parallel.
A CUDA program is executed into a host program, having at The HOST = CPU which executes serial portion of an
least one thread running on the host Central Processing Unit application. KERNELS are the functions that run on the
and at least one parallel kernels that are required for the device. The explanation below is illustrated by Figure 2.
execution on Graphical Processing Unit which is a parallel In CUDA programming model, the GPU is considered as
preparing device. A kernel at that point executes a scalar a device. It is a co-processor to CPU and comprise of its own
successive program on a set of parallel threads. The developer DRAM (GPU memory), which runs numerous threads in
needs to arrange these threads into a matrix of thread blocks. parallel. A kernel executes a framework which comprises of
blocks. A block comprises of a cluster of threads which
cooperate with each other. And the cooperation is done by
synchronizing their execution i.e. (shared memory) and by
sharing data. Two threads from 2 different blocks cannot
cooperate.
Threads and blocks have their own IDs. Using this ID each
thread can decide on what data to work on.
This model has solved the memory addressing problems. [2]

Fig. 1. Process flow in CUDA

The threads of solitary blocks are allowed to synchronize with Fig. 2. CUDA Model Overview
each other by means of barriers and approach a fast on-chip
memory for between threads correspondence. Threads from
different blocks in a comparative structure can facilitate just
II. APPLICATIONS OF CUDA-
by methods for shared global memory space, which is visible
to all of the thread. CUDA requires that thread blocks to be a) MRI:
free, inferring that a kernel should execute effectively Magnetic Resonance Imaging frames pictures with
regardless of the request in which blocks are asked for, high spatial-determination, The thing which has to be checked
regardless of whether all blocks are executed progressively in is set in an exceptionally strong field, that makes the turn of its
the arbitrary order. The repression on the conditions between cores adjust either parallel or restricting parallel to the circle.
blocks of a kernel gives adaptability. It similarly deduces that Radiofrequency beats are then connected to energizing the
the prerequisite for synchronization among threads is the cores, that radiate vitality once they then return to their unique
fundamental thought in separating parallel work into isolated state. The first normal core utilized in imaging is the
kernels. [3] Hydrogen. To broaden the rate parallel imaging and packed
detecting is connected. In X-ray, imaging Gridding is utilized
Process Flow on CUDA for picture remaking. To dissect the difficulties and chances of
In Figure 1 we illustrate the process flow in CUDA. The using general GPU process, the non-parallel Quick Fourier
Main Memory copies the data to the Memory of GPU. Then rebuild algorithmic program was executed by Gregerson that
the CPU sends the process instructions to the GPU where it is was called as Gridding, on a GeForce 8800 GPU Nvidia's
executed parallelly in each core. After the execution, the data CUDA system. Investigation changes over a manifestation
is copied to the Main Memory from the GPU memory. from its unique area (frequently time or space) to the
recurrence area and contrariwise. It lessens the nature of
CUDA model overview processing from O(n^2) to O(n log n). This increase GPU
execution by four-hundredth.[4]

198

ed licensed use limited to: CENTRE FOR DEVELOPMENT OF ADVANCED COMPUTING - CDAC - KOLKATA. Downloaded on February 24,2022 at 11:41:00 UTC from IEEE Xplore. Restriction
 International Conference on Advances in Computing, Communication Control and Networking (ICACCCN2018)

CUDA-based GPU usage that in result expanded the

b) ULTRASOUND: execution. CUDA uses a bar outline which in turn extends
Ultrasound guideline is generally utilized for work for X-beam image.To improve the refinement of
surgical techniques, as a result of its low value and time period computerized X-beam CUDA program on a GPU, the bar
visual response. Several ultrasounds target-hunting strategies diagram is frequently acclimated and gets association of the
need giant coaching and simulations over patients. Procedure qualification in the image process. To broaden the
resources (problems) are typically restricted by time period perceivability of the picture continuously, we tend to utilize
necessities. In contrast to earlier approaches, they simulate the bar graph to extend the work. It's difficult to actualize the
original ultrasound pictures from CT knowledge on the task on a GPU because of the propelled technique to exchange
Graphics process Unit. A life scientist “Wein” recommended a the provided information.And in this way the outcome
technique for calculative properties of tissue and adjust it to a between the memory of GPU is handled. Thus, we tend to
lot of economic forms(computationally). Addition to previous show the control on bar outline which extends the work
approach, he calculated ultrasound properties from CT rapidly on GPU by the CUDA program. [7]
knowledge. Exploitation NVIDIA's CUDA a physically
adequate reenactment of ultrasound reflection, shadowing e) CT-SCAN:
(structure like tissue), noticed noise and blurring were made. Computed Tomography (CT) is utilized inside the
This will be utilized for mimicking either direct imaging or restorative imaging area, that was started by Wilhelm Konrad
outspread imaging, and each one determination of the Rontgen with the original work on x-beam imaging. It's
reenacted investigation are intuitively configurable at runtime, predominantly a couple of gifts with the end goal that:-
as well as ultrasound frequency. With current hardware, we 1. It's high spatial determination.
are able to attain a picture breadth of 1023 pixels from raw CT 2. Short examining time.
knowledge in the time period, with none reprocessing and loss Detriments is the "ionizing radiation", in order to keep up the
of knowledge from the CT image. [5] standard of the picture. The picture denoising strategies are
also utilized. Predominantly for two reasons multidimensional
c) CBCT: CT is extra computationally demanding than 3D CT, for
CBCT a therapeutic imaging system comprise of X- example,
beam registered pivotal tomography. CBCT find its The recreation must be connected to numerous volumes (e.g.
application in many field like oral surgery and orthodontia. 10-20). There are low-measurements strategies because of the
CBCT is developed inside the clinical space on account of its amount of radiation is much more than for 3D CT.
capacity to create 3D information all through mediations, its As per the Beister et al in 2012, GPU implementation is best
high indicative quality, and its short examining circumstances suited in CT,The main obstruction is the "radiologists".The
i.e 60sec. Late improvement inside the space of GPUs heps to work of parallel register provide answer for many issues as it
possess access to unrivaled figuring answer at an espresso allows the utilization of assets. [8]
value (shabby to utilize),and allowing their utilization in a few
logic issues. Noel upheld Relate Nursing recipe for 3D f) DWT-PCA Based Image Fusion Using GPU and CUDA:
remaking of CBCT by the use of CUDA, that was executed on In their work, Wencheng Wan (2011) considered the
a NVIDIA GeForce GTX 280. His usage enhanced CUDA innovation for image fusion as takes after. The source
reproduction times from minutes and hours (efficient) to the code provided contains both Kernel-GPU and host-CPU code.
seconds. He moreover assessed execution on ten clinical The host code exchanges information to and from the GPU's
information sets and one apparition learning set to take a worldwide memory and starts the portion (Kernel) code by
glance among processor and GPU. By exploitation, their calling a function. The portion (Kernel) runs various blocks of
approach, the calculation time for 2563 is lessened from 25 threads where each thread performs a singular calculation.
min to 3.2 s on the GPU. The GPU remaking time for 5123 Threads are sorted out into a chain of importance of matrices
volumes is 8.5s. [6] of thread block and a network comprises of various blocks that
execute a similar kernel. The block comprises of strings that
d) X-RAY: entrance information from the common memory and executes
X-radiation (made out of X-beams) is a sort of directions in parallel. There is a private neighborhood memory
electromagnetic wave. The X-Ray is a picture inside which the for each thread. Amid the execution, threads may get to
value of each constituent could be a solitary example and information from different memory spaces. [9]
assemble the division to be performed. Mahmoudi arranged a
division technique which bolstered the Active frame Model. g) Morphological Image Processing on GPU using CUDA:
The Active frame Model aides in particular as edges. ASM In their work, D. B. Kirk and W. mei W. Hwu (2010)
characterizes an accumulation of structures that portray the considered the CUDA innovation for Morphology as takes
frame to be known. This application is postponed because of after. Morphology is clarified as far as set hypothesis. Sets
its calculation time. Mahmoudi explains that the extraction
speak to objects in a picture. The Morphological handling is
will be performed quickly by using the huge procedure energy
worked by performing tasks on sets of pixels. Normally it is
of the Graphics procedure Units. Mahmoudi arranged a
developed for binary pictures. This set involving all the white

199

ed licensed use limited to: CENTRE FOR DEVELOPMENT OF ADVANCED COMPUTING - CDAC - KOLKATA. Downloaded on February 24,2022 at 11:41:00 UTC from IEEE Xplore. Restriction
 International Conference on Advances in Computing, Communication Control and Networking (ICACCCN2018)

pixels in a parallel picture is a total morphological depiction of Unified Device Architecture) is proposed which is utilized to
a picture. In the paired (binary) pictures, the sets are given quicken Multi-sensor picture combination, in which, every last
individuals from the 2D whole number space Z2, where every pixel is alloted to a solitary CUDA thread with the goal that
component of a set is a tuple (2D vector) whose directions are pixel-level picture combination would be parallelized by
the (x, y) directions of a white (or dark) pixel in the picture. numerous CUDA threads all the while. The execution for
The same as above can be stretched out for greyscale pictures. picture combination is additionally enhanced by making
utilization of the texture memory in CUDA. What's more, the
While the Gray-scale pictures can be spoken to as sets, whose
exploratory outcomes have obviously demonstrated that the
parts are in Z3: two segments are directions of a pixel and the
proposed strategies would enhance execution amazingly than
third one is its discrete power esteem. A CUDA program is upgraded CPU partner accessible. [12]
considered as a serial program with parallel segments. This
code of CPU and GPU is bound together, it begins its j) Pyramidal image blending using CUDA framework:
execution on the CPU having serial code and at whatever In their work, NVIDIA Corporation (2010) considered the
point a parallel code happens it is executed onto GPU and CUDA innovation as takes after. The method of pyramidal
control is offered back to CPU when parallel code completes picture mixing is one of the vital picture mixing system which
execution. The NVIDIA CUDA compiler (NVCC) isolates is utilized as an application in the all panoramic image
two unique codes (device and host) and after that incorporates stitching. The intriguing system of pyramidal picture
them independently from each other. The host is clarified as a procedure utilizes various picture pyramids that are required to
basic C code which is gathered with ordinary C compiler. The perform at various resolution levels of pictures. The beginning
code of device is not quite the same as the host code and is of the calculation happens with building Laplacian pyramids
for the two given (input) pictures and the Gaussian pyramid
likewise customized in a programming dialect with some other
for the mask picture. A short time later these picture pyramids
little augmentations to have programming dialect. It is are joined to frame the Laplacian pyramid, which is
assembled with "nvcc". The capacities that execute on GPU subsequently fell to get a seamless panaroma as the yield. The
and information parallel are called as "kernels". [10] full procedure as specified above isn't just computationally
costly tedious also and subsequently it should be speedup, and
h) CUDA based parallel wavelet algorithm in medical image
in this manner therefore, GPU based pyramidal picture mixing
fusion: calculation is acknowledged on NVIDIA's Compute Unified
In their work IMSNA (2013) considered the Device Architecture (CUDA). [13]
CUDA innovation for medicinal picture combination (image
fusion) utilizing wavelet calculation as takes after. In the III. CONCLUSION
present date essentially, CPU and PC group are utilized by
the restorative picture combination. At that point the parallel CUDA is an advance technology that brought revolution in the
wavelet picture combination calculation is given in view of processing power of the GPUs. The parallel execution of code
the model which is CUDA stream figuring model. can also be used to solve many other problems which include
Furthermore, a while later to the definite examinations of image processing, Bioinformatics, Molecular Dynamics.
wavelet picture combination, CUDA multi-layer stockpiling CUDA C platform is a software layer that has granted
structure and SIMD design, the information gathering and direct access to the virtual instruction set of GPUs. CUDA
the thread task is being deteriorated. The clinic image integration is always fantastic as NVIDIA provides high
investigations is additionally done by the assistance of quality support to the app developers who opt for CUDA
CUDA based wavelet picture combination calculation. In acceleration. Sometimes CUDA is not as that easy for the
this manner the test demonstrates this calculation can various apps to adopt as OpenCL as CUDA is a closed Nvidia
enhance the speed of CPU based one with 8-10 times and it framework and not open source like OpenCL but on the other
likewise has a linear speed up capacity. [11] hand, CUDA is supported by a wide variety of apps which is a
major advantage of using it. CUDA cross-develops with
i) Parallelizing Multi-Sensor Image Fusion Using CUDA: confidence and ease and thus helps in maintaining and using
In their work Journal of Computational and highly customized environments. It has more mature tools
Theoretical Nanoscience 7(1):408-411 · March 2012 has which comprise of a debugger and a profiler which are
considered the CUDA innovation as given takes after. The CUBLAS and CUFFT. Wherever CUDA is integrated it
picture combination (image fusion) in view of Pixel-level ensures unparalleled performance due to the high-quality
would achieve fantastic picture combination and in this way NVIDIA support. It provides the user with problem free
help enhance the execution of picture combination for the environment for running jobs.
multi-sensors. Howsoever there is a vast volume of
information that will be prepared yet the continuous IV. REFERENCES
combination handling is costly and furthermore un-adaptable.
[1] Yang, C. T., Huang, C. L., & Lin, C. F. (2011). Hybrid CUDA,
In this way considering the over, a fine-grained pixel-level OpenMP, and MPI parallel programming on multicore GPU
picture combination execution in view of CUDA (Compute clusters. Computer Physics Communications, 182(1), 266-269.

200

ed licensed use limited to: CENTRE FOR DEVELOPMENT OF ADVANCED COMPUTING - CDAC - KOLKATA. Downloaded on February 24,2022 at 11:41:00 UTC from IEEE Xplore. Restriction
 International Conference on Advances in Computing, Communication Control and Networking (ICACCCN2018)

[2] Yang, Z., Zhu, Y., & Pu, Y. (2008, December). Parallel image
processing based on CUDA. In Computer Science and Software
Engineering, 2008 International Conference on (Vol. 3, pp. 198-201).
IEEE.

[3] Garland, M., Le Grand, S., Nickolls, J., Anderson, J., Hardwick, J.,
Morton, S., ... & Volkov, V. (2008). Parallel computing experiences
with CUDA. IEEE micro, 28(4).

[4] Gregerson, A. (2008). Implementing fast MRI gridding on GPUs via

CUDA. Nvidia Tech. Report on Medical Imaging using CUDA.

[5] Reichl, T., Passenger, J., Acosta, O., & Salvado, O. (2009, February).
Ultrasound goes GPU: real-time simulation using CUDA. In SPIE
Medical Imaging (pp. 726116-726116). International Society for Optics
and Photonics.

[6] Noël, P. B., Walczak, A. M., Xu, J., Corso, J. J., Hoffmann, K. R., &
Schafer, S. (2010). GPU-based cone beam computed tomography.
Computer methods and programs in biomedicine, 98(3), 271-277.

[7] Mahmoudi, S. A., Lecron, F., Manneback, P., Benjelloun, M., &
Mahmoudi, S. (2010, September). GPU-based segmentation of cervical
vertebra in X-ray images. In cluster Computing Workshops and Posters,
2010 IEEE International Conference on (pp. 1-8). IEEE.

[8] Beister, M., Kolditz, D., & Kalender, W. A. (2012). Iterative

reconstruction methods in X-ray CT. Physica medica, 28(2), 94-108.

[9] Jie-Lun Chiang, “Knowledge based Principal Component Analysis for

Image Fusion” International Journal of Applied Mathematics and
Information Sciences 8, No. 1L, 223-230 (2014)

[10] Pierre Sollie, Edmond J. Breen, and Ronald Jones “Recursive

implementation of erosions and dilations along discrete lines at arbitrary
angles”, Pattern analysis and machine intelligence, IEEE Transactions
on, vol 18, no 5, pp. 562-567, may 1996

[11] Jia Liu, Dustin Feld, Yong Xue, Senior Member, IEEE, Jochen Garcke,
and Thomas Soddemann, “Multicore processors and graphics
Processing Unit Accelerators forParallel Retrieval of Aerosol Optical
Depth From Satellite Data: Implementation, Performance, and Energy
Efficiency”, IEEE Journal Of Selected Topics In Applied Earth
Observations And Remote Sensing, No. 5, May 201

[12] Gagandeep Kaur and Sharad P Singh, “Image Decomposition Using

Wavelet Transform”, International Journal Of Engineering and
Computer Science ISSN: 2319-7242 Issue 12 Dec 2013 Page No.3477-
3480.

[13] Peter J. Burt & Edward H. Adelson, “A Multi-resolution Spline with

Application to Image Mosaics”, ACM Transactions on Graphics,
October 1983.

Acknowledgment: We would like to express our sincere gratitude to

Accendere Knowledge Management Services Pvt. Ltd., for his valuable
comments that led to substantial improvements on an earlier
version of this manuscript.

201

ed licensed use limited to: CENTRE FOR DEVELOPMENT OF ADVANCED COMPUTING - CDAC - KOLKATA. Downloaded on February 24,2022 at 11:41:00 UTC from IEEE Xplore. Restriction