purpose computation. Then the architecture of a GPU is illustrated. There are
many fundamental differences between a GPU and a tradition processor. In order
to benefit from the power of a GPU, a CUDA programmer needs to use threads. They
-have some particularities which enables the CUDA model to be efficient and
+have some particularities which enable the CUDA model to be efficient and
scalable when some constraints are addressed.
\section{Brief history of Video Card}
-Video card or Graphics card have been introduced in personnal computers to
-produce high quality graphics faster than classical Central Processing Unit
+Video cards or Graphics cards have been introduced in personal computers to
+produce high quality graphics faster than classical Central Processing Units
(CPU) and to alleviate CPU from this task. In general, display tasks are very
repetitive and very specific. Hence, some manufacturers have produced more and
more sofisticated video cards, providing 2D accelerations then 3D accelerations,
then some light transforms. Video cards own their own memory to perform their
-computation. From at least two dedaces, every personnal computer has a video
-card which is simple for desktop computers or which provides many accelerations
-for game and/or graphic oriented computers. In the latter case, graphic cards
+computation. For at least two dedaces, every personnal computer has had a video
+card which is simple for desktop computers or which provides many accelerations
+for game and/or graphic oriented computers. In the latter case, graphic cards
may be more expensive than a CPU.
-After 2000, video cards allowed to apply arithmetics operations simulatenously
-on a sequence of pixels, also later called stream processing. In this case,
-information of the pixels (color, location and other information) are combined
-in order to produce a pixel color that can be displayed on a
-screen. Simultaneous computations are provided by shaders which calculate
-rendering effects on graphics hardware with a high degree of flexibility. These
-shaders handles the stream data with pipelines.
+Since 2000, video cards have allowed users to apply arithmetics operations
+simultaneously on a sequence of pixels, also later called stream processing. In
+this case, the information of the pixels (color, location and other information) are
+combined in order to produce a pixel color that can be displayed on a screen.
+Simultaneous computations are provided by shaders which calculate rendering
+effects on graphics hardware with a high degree of flexibility. These shaders
+handles the stream data with pipelines.
-Some reasearchers tried to apply those operations on other data, representing
+Some researchers tried to apply those operations on other data, representing
something different from pixels, and consequently this resulted in the first
uses of video cards for performing general purpose computation. The programming
model was not easy to use at all and was very dependent of the hardware
-constraints. More precisely it consisted in using either DirectX of OpenGL
+constraints. More precisely it consisted in using either DirectX of OpenGL
functions providing an interface to some classical operations for videos
-operations (memory transfers, texture manipulation, ...). Floating point
-operations were most of the time unimaginable. Obviously when something bad
-happened, programmers had no way (and tools) to detect it.
+operations (memory transfers, texture manipulation, ...). Floating point
+operations were most of the time unimaginable. Obviously when something went
+wrong, programmers had no way (and neither the tools) to detect it.
\section{GPGPU}
was first proposed in 2007 by NVidia. It unifies the programming model for some
of their most performant video cards. Cuda~\cite{ch1:cuda} has quickly been
considered by the scientific community as a great advance for general purpose
-graphics processing unit (GPGPU) computing. Of course other programming model
+graphics processing unit (GPGPU) computing. Of course other programming models
have been proposed. The other well-known alternative is OpenCL which aims at
proposing an alternative to Cuda and which is multi-platform and portable. This
is a great advantage since it is even possible to execute OpenCL programs on
-traditionnal CPUs. The main drawbacks is that it is less tight with the
-hardware and consequently provides sometimes less efficient programs. Moreover,
-Cuda benefits from more mature compilation and optimization procedures. Other
-less known environment have been proposed, but most of them have been stopped,
-for example we can cited: FireStream by ATI which is not maintened anymore and
-replaced by OpenCL, BrookGPU by Standford University~\cite{ch1:Buck:2004:BGS}.
+traditionnal CPUs. The main drawback is that it is less tight with the hardware
+and consequently sometimes provides less efficient programs. Moreover, Cuda
+benefits from more mature compilation and optimization procedures. Other less
+known environments have been proposed, but most of them have been stopped, for
+example we can cite: FireStream by ATI which is not maintened anymore and
+replaced by OpenCL, BrookGPU by Standford University~\cite{ch1:Buck:2004:BGS}.
Another environment based on pragma (insertion of pragma directives inside the
-code to help the compiler to generate efficient code) is call OpenACC. For a
+code to help the compiler to generate efficient code) is call OpenACC. For a
comparison with OpenCL, interested readers may refer to~\cite{ch1:CMR:12}.
\section{Architecture of current GPUs}
-Architecture \index{architecture of a GPU} of current GPUs is constantly
-evolving. Nevertheless some trends remains true through this
-evolution. Processing units composing a GPU are far more simpler than a
-traditional CPU but it is much easier to integrate many computing units inside a
-GPU card than many cores inside a CPU. This is due to the fact that cores of a
-GPU are simpler than cores of a CPU. In 2012, the most powerful GPUs own more
-than 500 cores and the most powerful CPUs have 8
+The architecture \index{architecture of a GPU} of current GPUs is constantly
+evolving. Nevertheless some trends remain constant throughout this evolution.
+Processing units composing a GPU are far more simple than a traditional CPU but
+it is much easier to integrate many computing units inside a GPU card than to do
+so with many cores inside a CPU. This is due to the fact that the cores of a GPU are
+simpler than the cores of a CPU. In 2012, the most powerful GPUs own more than 500
+cores and the most powerful CPUs have 8
cores. Figure~\ref{ch1:fig:comparison_cpu_gpu} shows the number of cores inside
a CPU and inside a GPU. In fact, in a current NVidia GPU, there are
multiprocessors which have 32 cores (for example on Fermi cards). The core clock
-of CPU is generally around 3GHz and the one of GPU is about 1.5GHz. Although the
-core clock of GPU cores is slower, the amount of cores inside a GPU provides
-more computational power. This measure is commonly represented by the number of
-floating point operation per seconds. Nowadays most powerful GPUs provide more
-than 1TFlops, i.e. $10^{12}$ floating point operations per second. Nevertheless
-GPUs are very efficient to perform some operations but not all kinds of
-operations. They are very efficient to execute repetitive work in which only the
-data change. It is important to keep in mind that multiprocessors inside a GPU
-have 32 cores. Later we will see that these 32 cores need to do the same work to
-get maximum performance.
+of CPU is generally around 3GHz and the one of GPU is about 1.5GHz. Although
+the core clock of GPU cores is slower, the amount of cores inside a GPU provides
+more computational power. This measure is commonly represented by the number of
+floating point operation per seconds. Nowadays the most powerful GPUs provide more
+than 1TFlops, i.e. $10^{12}$ floating point operations per second.
+Nevertheless GPUs are very efficient to perform some operations but not all
+kinds of operations. They are very efficient to execute repetitive work in which
+only the data change. It is important to keep in mind that multiprocessors
+inside a GPU have 32 cores. Later we will see that these 32 cores need to do the
+same work to get maximum performance.
\begin{figure}[b!]
\centerline{\includegraphics[]{Chapters/chapter1/figures/nb_cores_CPU_GPU.pdf}}
-\caption[Comparison of number of cores in a CPU and in a GPU]{Comparison of number of cores in a CPU and in a GPU.}
+\caption{Comparison of number of cores in a CPU and in a GPU.}
+%[Comparison of number of cores in a CPU and in a GPU]
\label{ch1:fig:comparison_cpu_gpu}
\end{figure}
-On most powerful GPU cards, called Fermi, multiprocessors are called streaming
+On the most powerful GPU cards, called Fermi, multiprocessors are called streaming
multiprocessors (SM). Each SM contains 32 cores and is able to perform 32
-floating point or integer operations on 32bits numbers per clock or 16 floating
-point on 64bits number per clock. SMs have their own registers, execution
+floating points or integer operations on 32 bits numbers per clock or 16 floating
+points on 64 bits number per clock. SMs have their own registers, execution
pipelines and caches. On Fermi architecture, there are 64Kb shared memory + L1
cache and 32,536 32bits registers per SM. More precisely the programmer can
-decide what amount of shared memory and L1 cache SM can use. The constaint is
-that the sum of both amounts is less or equal to 64Kb.
+decide what amount of shared memory and L1 cache SM can use. The constraint is
+that the sum of both amounts should be less or equal to 64Kb.
Threads are used to benefit from the important number of cores of a GPU. Those
threads are different from traditional threads for CPU. In
chapter~\ref{chapter2}, some examples of GPU programming will explicit the
details of the GPU threads. However, threads are gathered into blocks of 32
-threads, called ``warp''. Those warps are important when designing an algorithm
+threads, called ``warps''. Those warps are important when designing an algorithm
for GPU.
through the use of cache memories. Moreover, nowadays CPUs perform many
performance optimizations such as speculative execution which roughly speaking
consists in executing a small part of code in advance even if later this work
-reveals to be useless. In opposite, GPUs do not have low latency memory. In
-comparison GPUs have ridiculous cache memories. Nevertheless the architecture of
-GPUs is optimized for throughtput computation and it takes into account the
-memory latency.
+reveals itself to be useless. On the contrary, GPUs do not have low latency
+memory. In comparison GPUs have small cache memories. Nevertheless the
+architecture of GPUs is optimized for throughtput computation and it takes into
+account the memory latency.
\begin{figure}[b!]
\centerline{\includegraphics[scale=0.7]{Chapters/chapter1/figures/low_latency_vs_high_throughput.pdf}}
-\caption[Comparison of low latency of CPU and highthroughput of GPU]{Comparison of low latency of CPU and highthroughput of GPU.}
+\caption{Comparison of low latency of CPU and high throughput of GPU.}
\label{ch1:fig:latency_throughput}
\end{figure}
the principle to obtain a high throughput is to have many tasks to
compute. Later we will see that those tasks are called threads with CUDA. With
this principle, as soon as a task is finished the next one is ready to be
-executed while the waiting for data for the previous task is overlapped by
+executed while the wait for data for the previous task is overlapped by
computation of other tasks.
\section{Kinds of parallelism}
Many kinds of parallelism are avaible according to the type of hardware.
-Roughtly speaking, there are three classes of parallelism: instruction-level
+Roughly speaking, there are three classes of parallelism: instruction-level
parallelism, data parallelism and task parallelism.
Instruction-level parallelism consists in re-ordering some instructions in order
to execute some of them in parallel without changing the result of the code.
-In modern CPUs, instruction pipelines allow processor to execute instruction
+In modern CPUs, instruction pipelines allow processor to execute instructions
faster. With a pipeline a processor can execute multiple instructions
simultaneously due to the fact that the output of a task is the input of the
next one.
Data parallelism consists in executing the same program with different data on
-different computing units. Of course, no depency should exist between the the
+different computing units. Of course, no dependency should exist between the
data. For example, it is easy to parallelize loops without dependency using the
data parallelism paradigm. This paradigm is linked with the Single Instructions
-Multiple Data (SIMD) architecture. This is the kind of parallism providing by
+Multiple Data (SIMD) architecture. This is the kind of parallelism provided by
GPUs.
Taks parallelism is the common parallism achieved out on clusters and grids and
Multiple Thread (SIMT) model. This is due to the fact that a programmer accesses
to the cores by the intermediate of threads. In the CUDA model, all cores
execute the same set of instructions but with different data. This model has
-similarities with vector programming model proposed for vector machines through
+similarities with the vector programming model proposed for vector machines through
the 1970s into the 90s, notably the various Cray platforms. On the CUDA
architecture, the performance is led by the use of a huge number of threads
-(from thousand upto to millions). The particularity of the model is that there
+(from thousands up to to millions). The particularity of the model is that there
is no context switching as in CPUs and each thread has its own registers. In
practice, threads are executed by SM and are gathered into groups of 32
-threads. Those groups are call ``warps''. Each SM alternatively executes
-``active warps'' and warps becoming temporaly inactive due to waiting of data
+threads. Those groups are called ``warps''. Each SM alternatively executes
+``active warps'' and warps becoming temporarilly inactive due to waiting of data
(as shown in Figure~\ref{ch1:fig:latency_throughput}).
The key to scalability in the CUDA model is the use of a huge number of threads.
In practice, threads are not only gathered in warps but also in thread blocks. A
-thread block is executed by only one SM and it cannot migrate. Typical size of
-thread block is a number power of two (for example: 64, 128, 256 or 512).
+thread block is executed by only one SM and it cannot migrate. The typical size of
+a thread block is a number power of two (for example: 64, 128, 256 or 512).
In this case, without changing anything inside a CUDA code, it is possible to
-run your code with a small CUDA device or most performant Tesla CUDA cards.
+run your code with a small CUDA device or the most performing Tesla CUDA cards.
Blocks are executed in any order depending on the number of SMs available. So
the programmer must conceive its code having this issue in mind. This
-independence between threads blocks provides the scalability of CUDA codes.
+independence between thread blocks provides the scalability of CUDA codes.
\begin{figure}[b!]
\centerline{\includegraphics[scale=0.65]{Chapters/chapter1/figures/scalability.pdf}}
-\caption[Scalability of GPU]{Scalability of GPU.}
+\caption{Scalability of GPU.}
\label{ch1:fig:scalability}
\end{figure}
A kernel is a function which contains a block of instructions that are executed
-by the threads of a GPU. When the problem considered is a 2 dimensions or 3
-dimensions problem, it is possible to group thread blocks into grid. In
-practice, the number of thread blocks and the size of thread block is given in
-parameter to each kernel. Figure~\ref{ch1:fig:scalability} illustrates an
+by the threads of a GPU. When the problem considered is a two dimensional or three
+dimensional problem, it is possible to group thread blocks into a grid. In
+practice, the number of thread blocks and the size of thread blocks is given as
+parameters to each kernel. Figure~\ref{ch1:fig:scalability} illustrates an
example of a kernel composed of 8 thread blocks. Then this kernel is executed on
-a small device containing only 2 SMs. So in in this case, blocks are executed 2
+a small device containing only 2 SMs. So in this case, blocks are executed 2
by 2 in any order. If the kernel is executed on a larger CUDA device containing
4 SMs, blocks are executed 4 by 4 simultaneously. The execution times should be
approximately twice faster in the latter case. Of course, that depends on other
parameters that will be described later.
-Thread blocks provide a way to cooperation in the sens that threads of the same
+Thread blocks provide a way to cooperation in the sense that threads of the same
block cooperatively load and store blocks of memory they all
use. Synchronizations of threads in the same block are possible (but not between
threads of different blocks). Threads of the same block can also share results
As previously mentioned each thread can access its own registers. It is
important to keep in mind that the number of registers per block is limited. On
recent cards, this number is limited to 64Kb per SM. Access to registers is
-very fast, so when possible it is a good idea to use them.
+very fast, so it is a good idea to use them whenever possible.
Likewise each thread can access local memory which, in practice, is much slower
than registers. Local memory is automatically used by the compiler when all the
\begin{figure}[hbtp!]
\centerline{\includegraphics[scale=0.60]{Chapters/chapter1/figures/memory_hierarchy.pdf}}
-\caption[Memory hierarchy of a GPU]{Memory hierarchy of a GPU.}
+\caption{Memory hierarchy of a GPU.}
\label{ch1:fig:memory_hierarchy}
\end{figure}
\section{Conclusion}
-In this chapter, a brief presentation of the video card, which later have been
+In this chapter, a brief presentation of the video card, which has later been
used to perform computation, has been given. The architecture of a GPU has been
illustrated focusing on the particularity of GPUs in term of parallelism, memory
latency and threads. In order to design an efficient algorithm for GPU, it is
