- Source file extension:
- CUDA program names must end in .cu
- Compiling and running
CUDA C programs:
Single program file compilation: nvcc -o prog prog.cu
Multiple program file compilation: nvcc -c prog1.cu nvcc -c prog2.cu nvcc -o prog prog1.o prog2.o
- CUDA servers available
for gerenal usage:
lab1a lab1b ... lab1h
The CUDA servers are only accessible via lab0z.mathcs.emory.edu (or a lab machine in the CS lab)
- How to logon
to a CUDA server:
Login to the MathCS gateway lab0z.mathcs.emory.edu Then on lab0z.mathcs.emory.edu, execute: ssh -X lab1a or ssh -X lab1b etc.You must logon to a CUDA server before you can compile/run your CUDA C program(s) !!!
- CUDA header file name:
#include <cuda.h> (Path: /usr/local/cuda-9.0/include/cuda.h)The CUDA header file is automatically included by the CUDA C compiler nvcc !!!
- A failed
Hello World
CUDA C program:
/* --------------------------------------------------- My Hello world for CUDA programming --------------------------------------------------- */ #include <stdio.h> // C programming header file #include <unistd.h> // C programming header file // cude.h is automatically included by nvcc... /* ------------------------------------ Your first kernel (= GPU function) ------------------------------------ */ __global__ void hello( ) { printf("Hello World !\n"); } int main() { /* ------------------------------------ Call the hello( ) kernel function ------------------------------------ */ hello<<< 1, 4 >>>( ); printf("I am the CPU: Hello World ! \n"); return 0; } - Example Program:
(Demo above code)
/home/cs355001/demo/CUDA/1-intro/hello1 Output: I am the CPU: Hello World !
- A successful
Hello World
CUDA C program:
/* --------------------------------------------------- My Hello world for CUDA programming --------------------------------------------------- */ #include <stdio.h> // C programming header file #include <unistd.h> // C programming header file // cude.h is automatically included by nvcc... /* ------------------------------------ Your first kernel (= GPU function) ------------------------------------ */ __global__ void hello( ) { printf("Hello World !\n"); } int main() { /* ------------------------------------ Call the hello( ) kernel function ------------------------------------ */ hello<<< 1, 4 >>>( ); printf("I am the CPU: Hello World ! \n"); sleep(1); // Necessary to give time to let GPU threads run !!! return 0; } - Example Program:
(Demo above code)
/home/cs355001/demo/CUDA/1-intro/hello1 Output: I am the CPU: Hello World ! Hello World ! Hello World ! Hello World ! Hello World !
- Explanation:
- Kernel function:
- Kernel function = a function that is (stored and) run on the Graphics Processing Unit (GPU)
How to define a kernel function:
__global__ C-function-definitionExample:
__global__ void hello( ) // This function is run by the GPU !!! { printf("Hello World !\n"); }
- Program starts
its execution in the
main( ) C function
- Executing a
kernel function
(a.k.a. a launching a kernel):
/* ------------------------------------------------ Call the kernel function "hello( )" using a "1,4 grid" ------------------------------------------------ */ hello<<< 1, 4 >>>( );A grid of GPU threads will start to execute the code in the hello( ) function
The kernel launch will return immediately and the CPU will continue its program execution (while the GPU runs the kernel !!!)
- Therefore:
- In hello1:
- The CPU will
print
the message
"I am the CPU: Hello World !"
and exit
(so you do not see the Hello World messages printed by the kernel)
- The CPU will
print
the message
"I am the CPU: Hello World !"
and exit
- In hello1:
- The CPU will
print
the message
"I am the CPU: Hello World !"
and waits 1 sec
- There was enough delay before program exit for the kernel to print the Hello World messages
- The CPU will
print
the message
"I am the CPU: Hello World !"
and waits 1 sec
Why were there 4 Hello messages:
- The call:
hello<<< 1, 4 >>>( );
launches a grid that contains 4 "threads" of executions
Each thread will print one Hello Wolrd message
- In hello1:
- Kernel function: