CDS Introduction to CUDA

Introduction to the
CUDA (GPU) Programming

1
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

CPU - GPU
CPU vs GPU

• CPU is designed to excel at executing a sequence of operations (threads) as fast as

possible (low latency)

• CPU can execute a few tens of these threads in parallel

• GPU is designed to excel at executing thousands of threads in parallel (amortizing

the slower single-thread performance to achieve greater throughput

2
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU programming

• CUDA
• general-purpose parallel computing platform and programming model for NVIDIA GPUs
• allows developers to use C++ as a high-level programming language

• OpenCL
• General heterogenous computing framework
• https://www.khronos.org/opencl/


• OpenACC
• user-driven directive-based performance-portable parallel programming model
• supports C, C++, Fortran programming language

Note: For python usage: Check Theano, pyCUDA 3

Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA
CUDA (Compute Unified Device Architecture)
A general-purpose parallel computing platform and programming model for NVIDIA GPUs

4
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

Scalable Programming Model

GPU computing: steps involved

• Setup inputs on the host (CPU-accessible memory)

• Allocate memory for outputs on the host CPU
• Allocate memory for inputs on the GPU device
• Allocate memory for outputs on the GPU
• Copy inputs from host to GPU (slow)
• Start GPU kernel (function that executes on gpu)
• Copy output from GPU to host (slow)

5
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU: Hello World!

stardom

$> nvcc hello.cu -o hello

https://cuda-tutorial.readthedocs.io/en/latest/tutorials/tutorial01/ 6
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU: Hello World!

• __global__ specifier indicates a function that runs on device (GPU)

• the CUDA kernel cuda_hello() can be called from host
• kernel execution configuration is provided through <<<...>>>
syntax, called kernel launch
• the number of GPU threads “M" to be launched in each thread
block is indicated through kernel launch: <<<B,M>>>, where “B” is

Fm
the number of thread blocks
Bil
7
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU Computing

CK 8 1033
• Thread - distributed by the
CUDA runtime (threadIdx)
• Block - user defined group of
1 to ~T threads (blockIdx)
• Grid - a group of one or
more blocks. A grid is
created for each CUDA
kernel function call.

8
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU: Thread Hierarchy

NVIDIA Tesla V100 SXM2 Module with Volta GV100 GPU

https://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf
9
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA
GPU: Thread Hierarchy
NVIDIA Tesla V100 SXM2 Module with Volta GV100 GPU

Source: https://images.nvidia.com/content/volta-architecture/pdf/volta-architecture-whitepaper.pdf 10
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA
GPU: Thread Hierarchy
NVIDIA Tesla V100 SXM2 Module with Volta GV100 GPU

GPC - GPU Processing Clusters, TPCs -Texture Processing Clusters, SM - Streaming Multiprocessors
11
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU Computing

• Thread - distributed by the

CUDA runtime (threadIdx)
• Block - user defined group of
1 to ~T threads (blockIdx)
• Grid - a group of one or
more blocks. A grid is
created for each CUDA
kernel function call.

12
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU Computing

A f

• Thread - distributed by the CUDA

runtime (threadIdx)
• Block - user defined group of 1 to
~512 threads (blockIdx)
• Grid - a group of one or more blocks.
A grid is created for each CUDA
kernel function call.

13
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU Computing

Indexing Arrays with Blocks and Threads

• consider indexing an array with one element per thread (8 threads/block)

01 7 15 17

n
I
p
• with M threads per block, a unique index for each thread is given by:
• index = M * blockIdx.x + threadIdx.x
• use the built-in variable blockDim.x for threads per block (M)

mis index E 8 2 I i 17 14
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU Computing

Indexing Arrays with Blocks and Threads

• blockIdx.x, blockIdx.y, blockIdx.z built-in variables return the block ID in the x-axis, y-
axis, and z-axis of the block
• threadIdx.x, threadIdx.y, threadIdx.z built-in variables return the thread ID in the x-
axis, y-axis, and z-axis of the thread in a particular block
• blockDim.x, blockDim.y, blockDim.z built-in variables return the “block dimension”
(number of threads in a block in the x-axis, y-axis, and z-axis)

a fits3 t.IE
I
B
15
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA
GPU Computing: Vector addition
#define N 1618
#define T 1024
age 3 Pil
// Device code
k
__global__ void VecAdd(int* A, int* B, int* C, int N)
{
int i = blockDim.x * blockIdx.x + threadIdx.x; BCN3n
if (i < N)
ACN
n
r
C[i] = A[i] + B[i];
}
tT
SEI
// main code
W
int main() {
int a[N], b[N], c[N]; IB
int *dev_a, *dev_b, *dev_c;
Ia

Ii
// initialize a and b with int values

size = N * sizeof(int);
cudaMalloc((void**)&dev_a, size);
cudaMalloc((void**)&dev_b, size);
cudaMalloc((void**)&dev_c, size);

E
cudaMemcpy(dev_a, a, size,cudaMemcpyHostToDevice);

tac
cudaMemcpy(dev_b, b, size,cudaMemcpyHostToDevice);

vecAdd<<<(int)ceil(N/T),T>>>(dev_a,dev_b,dev_c);

cudaMemcpy(c, dev_c, size,cudaMemcpyDeviceToHost);

cudaFree(dev_a);

exit (0); }
cudaFree(dev_b); cudaFree(dev_c);
apu
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU Computing: Matrix multiplication

void matrixMult (int a[N][N], int b[N][N], int c[N][N], int width)
{
for (int i = 0; i < width; i++)
for (int j = 0; j < width; j++) {
int sum = 0;
for (int k = 0; k < width; k++) {
int m = a[i][k];
int n = b[k][j];
sum += m * n;
}

} -
c[I][j] = sum;

Can it be parallelized?

Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore

 CDS Introduction to CUDA

GPU Computing: Matrix multiplication

void matrixMult (int a[N][N], int b[N][N], int c[N][N], int width)
{
for (int i = 0; i < width; i++)
for (int j = 0; j < width; j++) {
int sum = 0;
for (int k = 0; k < width; k++) {
int m = a[i][k];
int n = b[k][j];
sum += m * n;
}
c[I][j] = sum;
}

} n Ro
Rtl
// Device code n
__global__ void matrixMult (int *a, int *b, int *c, int width) {
int k, sum = 0;
int col = threadIdx.x + blockDim.x * blockIdx.x;
int row = threadIdx.y + blockDim.y * blockIdx.y;
if(col < width && row < width) {
for (k = 0; k < width; k++)
In RI
sum += a[row * width + k] * b[k * width + col];
c[row * width + col] = sum;
}
}

Module 1
w
Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU Computing: Asynchronous

Concurrent Execution

CUDA exposes the following operations as independent tasks that can

operate concurrently with one another:
Computation on the host;
Computation on the device;
Memory transfers from the host to the device;
Memory transfers from the device to the host;
Memory transfers within the memory of a given device;
Memory transfers among devices.

More Details: https://docs.nvidia.com/cuda/cuda-c-programming-guide/

Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore

 CDS Introduction to CUDA

GPU Computing

EX: Write CUDA parallel code for

1. Matrix addition

2. Matrix vector multiplication

3. Matrix multiplication

20
Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore
 CDS Introduction to CUDA

GPU Computing (cuDNN)

in Data Science

Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore

 CDS Introduction to CUDA

GPU Computing (cuDNN)

in Data Science

Source: https://developer.nvidia.com/blog/category/artificial-intelligence/

Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore

 CDS Introduction to CUDA

GPU Computing (cuDNN)

in Data Science

Source: https://developer.nvidia.com/deep-learning-frameworks

Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore

 CDS Introduction to CUDA

GPU Computing (cuDNN)

in Data Science

Source: https://developer.nvidia.com/deep-learning-frameworks

Module 1 Sashikumaar Ganesan, CDS, IISc Bangalore

 CDS Introduction to CUDA

Numba for CUDA GPUs

https://numba.pydata.org/numba-doc/latest/cuda/index.html 25
Module 1 https://nyu-cds.github.io/python-numba/05-cuda/ Sashikumaar Ganesan, CDS, IISc Bangalore