Introduction to CUDA

CUDA – AN
INTRODUCTION
Raymond Tay

CUDA - What and Why
  CUDA™ is a C/C++ SDK developed by Nvidia. Released in 2006 world-wide for
the GeForce™ 8800 graphics card. CUDA 4.0 SDK released in 2011.
  CUDA allows HPC developers, researchers to model complex problems and achieve
up to 100x performance.

CUDA
SDK

Nvidia GPUs FPS
  FPS – Floating-point per second aka flops. A measure of how many flops can a
GPU do. More is Better 

GPUs beat CPUs

Nvidia GPUs Memory Bandwidth
  With massively parallel processors in Nvidia’s GPUs, providing high memory
bandwidth plays a big role in high performance computing.

GPUs beat CPUs

GPU vs CPU

CPU GPU
"   Optimised for low-latency "   Optimised for data-parallel,
access to cached data sets throughput computation
"   Control logic for out-of-order "   Architecture tolerant of
and speculative execution memory latency
"   More transistors dedicated to
computation

I don’t know C/C++, should I leave?

  Relax, no worries. Not to fret.

Your Brain Asks:
Wait a minute, why should I learn
the C/C++ SDK?

CUDA Answers:
Efficiency!!!

I’ve heard about OpenCL. What is it?

Entry point for developers
who prefer high-level C

Entry point for developers
who want low-level API

Shared back-end compiler and
optimization technology

What do I need to begin with CUDA?

  A Nvidia CUDA enabled graphics card e.g. Fermi

How does CUDA work

PCI Bus

1.  Copy input data from CPU memory to GPU
memory
2.  Load GPU program and execute,
caching data on chip for performance
3.  Copy results from GPU memory to CPU memory

CUDA Kernels: Subdivide into Blocks

  Threads are grouped into blocks
  Blocks are grouped into a grid
  A kernel is executed as a grid of blocks of threads

Transparent Scalability – G80

1 2 3 4 5 6 7 8 9 10 11 12

9 10 11 12

1 2 3 4 5 6 7 8

As maximum blocks are executing on the GPU, blocks 9
– 12 will wait

Transparent Scalability – GT200

1 2 3 4 5 6 7 8 9 10 11 12

1 2 3 4 5 6 7 8 9 10 11 12 Idle
... Idle Idle

Arrays of Parallel Threads
  ALL threads run the same kernel code
  Each thread has an ID that’s used to compute

address & make control decisions
Block 0 Block (N -1)

0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7

… …
unsigned int tid = threadIdx.x + blockIdx.x * blockDim.x; unsigned int tid = threadIdx.x + blockIdx.x * blockDim.x;

int shifted = input_array[tid] + shift_amount; int shifted = input_array[tid] + shift_amount;
if ( shifted > alphabet_max ) if ( shifted > alphabet_max )
shifted = shifted % (alphabet_max + 1); shifted = shifted % (alphabet_max + 1);

output_array[tid] = shifted; output_array[tid] = shifted;
… …
Parallel code Parallel code

Compiling a CUDA program
C/C++ CUDA float4 me = gx[gtid];
me.x += me.y * me.z;
Application
•  Parallel Thread
eXecution (PTX)‫‏‬
–  Virtual Machine
NVCC CPU Code and ISA
–  Programming
model
Virtual PTX Code –  Execution
resources and
state
PTX to Target ld.global.v4.f32 {$f1,$f3,$f5,$f7}, [$r9+0];
mad.f32 $f1, $f5, $f3, $f1;
Compiler

G80 … GPU

Target code

Example: Block Cypher
void host_shift_cypher(unsigned int *input_array, __global__ void shift_cypher(unsigned int
unsigned int *output_array, unsigned int *input_array, unsigned int *output_array,
shift_amount, unsigned int alphabet_max, unsigned int shift_amount, unsigned int
unsigned int array_length)
alphabet_max, unsigned int array_length)
{
{
for(unsigned int i=0;i<array_length;i++)
unsigned int tid = threadIdx.x + blockIdx.x *
blockDim.x;
{
int shifted = input_array[tid] + shift_amount;
int element = input_array[i];
if ( shifted > alphabet_max )
int shifted = element + shift_amount;

shifted = shifted % (alphabet_max + 1);
if(shifted > alphabet_max)
{
output_array[tid] = shifted;
shifted = shifted % (alphabet_max + 1);
}
}
output_array[i] = shifted;
Int main() {
}
dim3 dimGrid(ceil(array_length)/block_size);
}
dim3 dimBlock(block_size);
Int main() {
shift_cypher<<<dimGrid,dimBlock>>>(input_array,
host_shift_cypher(input_array, output_array,
output_array, shift_amount, alphabet_max,
shift_amount, alphabet_max, array_length);
array_length);
}
}
CPU Program GPU Program

I see some WEIRD syntax..is it still C?

  CUDA C is an extension of C
  <<< Dg, Db, Ns, S>>> is the execution

configuration for the call to __global__ ; defines
the dimensions of the grid and blocks that’ll be used
(dynamically allocated shared memory & stream is optional)
  __global__ declares a function is a kernel which is
executed on the GPU and callable from the host
only. This call is asynchronous.
  See the CUDA C Programming Guide.

How does the CUDA Kernel get Data?

  Allocate CPU memory for n integers e.g. malloc(…)
  Allocate GPU memory for n integers e.g. cudaMalloc(…)

  Copy the CPU memory to GPU memory for n
integers e.g. cudaMemcpy(…, cudaMemcpyHostToDevice)
  Copy the GPU memory to CPU once computation is

done e.g. cudaMemcpy(…, cudaMemcpyDeviceToHost)
  Free the GPU & CPU memory e.g. cudaFree(…)

Example: Block Cypher (Host Code)
#include <stdio.h>

Int main() {
unsigned int num_bytes = sizeof(int) * (1 << 22);
unsigned int * input_array = 0;
unsigned int * output_array = 0;
…
cudaMalloc((void**)&input_array, num_bytes);
cudaMalloc((void**)&output_array, num_bytes);
cudaMemcpy(input_array, host_input_array, num_bytes, cudaMemcpyHostToDevice);
…
// gpu will compute the kernel and transfer the results out of the gpu to host.
cudaMemcpy(host_output_array, output_array, num_bytes,
cudaMemcpyDeviceToHost);
…
// free the memory
cudaFree(input_array);
cudaFree(output_array);
}

Compiling the Block Cypher GPU Code

  nvcc is the compiler and should be accessible from
your PATH variable. Set the dynamic library load
path
  UNIX: $PATH, Win: %PATH%
  UNIX: $LD_LIBRARY_PATH / $DYLD_LIBRARY_PATH

  nvcc block-cypher.cu –arch=sm_12
  Compile the GPU code for the GPU architecture sm_12
  nvcc –g –G block-cypher.cu –arch=sm_12
  Compiled
the program s.t. CPU + GPU code is in
debugged mode

Debugger
CUDA-GDB
• Based on GDB
• Linux
• Mac OS X

Parallel Nsight
• Plugin inside Visual
Studio

Visual Profiler & Memcheck
Profiler
•  Microsoft Windows
•  Linux
•  Mac OS X

•  Analyze Performance

CUDA-MEMCHECK
•  Microsoft Windows
•  Linux
•  Mac OS X

•  Detect memory access
errors

Hints
  Think about producing a serial algorithm that can
execute correctly on a CPU
  Think about producing a parallel (CUDA/OpenCL)
algorithm from that serial algorithm
  Obtain a initial run time (call it gold standard?)
  Use
the profiler to profile this initial run (Typically its quite
bad )
  Fine tune your code to take advantage of shared
memory, improving memory coalescing, reduce shared
memory conflicts etc (Consult the best practices guide &
SDK)
  Use the profiler to conduct cross comparisons

Hints (Not exhaustive!)
  Be aware of the trade offs when your kernel becomes
too complicated:
  If you noticed the kernel has a lot of local (thread) variables
e.g. int i, float j : register spilling
  If you noticed the run time is still slow EVEN AFTER you’ve
used shared memory, re-assess the memory access patterns :
shared memory conflicts
  TRY to reduce the number of conditionals e.g. Ifs : thread
divergence
  TRY to unroll ANY loops in the kernel code e.g. #pragma
unroll n
  Don’t use thread blocks that are not a multiple of warpSize.

Other cool things in the CUDA SDK 4.0
  GPUDirect
  Unified Virtual Address Space
  Multi-GPU
  P2P Memory Access/Copy (gels with the UVA)
  Concurrent Execution
  Kernel + Data
  Streams, Events
  GPU Memories
  Shared, Texture, Surface, Constant, Registers, Portable, Write-combining, Page-locked/
Pinned
  OpenGL, Direct3D interoperability
  Atomic functions, Fast Math Functions
  Dynamic Global Memory Allocation (in-kernel)
  Determine how much the device supports e.g. cudaDeviceGetLimit
  Set it before you launch the kernel e.g. cudaDeviceSetLimit
  Free it!

Additional Resources
  CUDA FAQ (http://tegradeveloper.nvidia.com/cuda-faq)
  CUDA Tools & Ecosystem (http://tegradeveloper.nvidia.com/cuda-tools-ecosystem)
  CUDA Downloads (http://tegradeveloper.nvidia.com/cuda-downloads)
  NVIDIA Forums (http://forums.nvidia.com/index.php?showforum=62)
  GPGPU (http://gpgpu.org )
  CUDA By Example (
http://tegradeveloper.nvidia.com/content/cuda-example-introduction-general-purpose-gpu-
programming-0)
  Jason Sanders & Edward Kandrot
  GPU Computing Gems Emerald Edition (
http://www.amazon.com/GPU-Computing-Gems-Emerald-Applications/dp/0123849888/ )
  Editor in Chief: Prof Hwu Wen-Mei

CUDA Libraries
  Visit this site
http://developer.nvidia.com/cuda-tools-
ecosystem#Libraries
  Thrust, CUFFT, CUBLAS, CUSP, NPP, OpenCV, GPU

AI-Tree Search, GPU AI-Path Finding
  A lot of the libraries are hosted in Google Code.

Many more gems in there too!

GPU memories: Shared

  More than 1 Tbyte/sec
aggregate memory bandwidth
  Use it
  As a cache
  To reorganize global memory accesses into
coalesced pattern
  To share data between threads

  16 kbytes per SM (Before Fermi)
  64 kbytes per SM (Fermi)

GPU memories: Texture

  Texture is an object for reading data
  Data is cached
  Host actions
  Allocate memory on GPU
  Create a texture memory reference object
  Bind the texture object to memory
  Clean up after use
  GPU actions
  Fetch using texture references
text1Dfetch(), tex1D(), tex2D(), tex3D()

GPU memories: Constant

  Write by host, read by GPU
  Data is cached

  Useful for tables of constants

  64 kbytes

Introduction to CUDA

Recommended

More Related Content

What's hot (20)

Similar to Introduction to CUDA (20)

More from Raymond Tay (7)

Recently uploaded (20)

Introduction to CUDA