Parallel Computing in CUDA

Terminal programs applying the parallel programming model with the CUDA arquitecture

List of terminal programs

Program file	Description
show_devices	show list of available cuda devices in the system.
sum_vectors	add two vectors.
	cuda target: launch a kernel with multiple threads grouped into fix blocks.
matrix2dim	generate a matrix using a existing matrix.
	cuda target: using threads of one kernel lauching two dimensions.
pi	roughly calculate the value of pi.
	cuda target: making kernel threads work cooperatively using PR algorithm.
displacer_matrix	displacer a matrix
	cuda target: measure the execution time of a kernel.
rank_sort_algorithm	sort a vector using the rank sort algorithm.
	cuda target: take advantage of parallelism of the GPU to optimize algorithms.
performance	sort a vector in gpu/cpu to compare the performance.
	cuda target: calculate the performance of the gain of GPU relative to CPU.

source path: /root/code

Pre-requisites

1. docker
2. docker-compose
3. make (optional)
4. gpu with cuda support (optional)

NOTE

GPU with CUDA support is optional, if you don't have a gpu with cuda support, you can use the docker image of cuda to run the programs.

The docker image use an emulator (GPGPU-Sim) to emulate NVIDIA GPUs running CUDA and/or OpenCL workloads. to more informatcion, go here: https://github.com/gpgpu-sim/gpgpu-sim_distribution/

Support for CUDA Deep Neural Network (cuDNN) or Basic Linear Algebra Subprograms (cuBLAS) libraries. (OPTIONAL)

Modify the docker image and add the following steps to execute programs that use these libraries

1. Download cuDNN and cuBLAS libraries from NVIDIA website: https://developer.nvidia.com

2. export the following environment variables

$ export CUDNN_PATH=/path/to/cuddn
$ export LD_LIBRARY_PATH=$CUDA_INSTALL_PATH/lib64:$CUDA_INSTALL_PATH/lib:$CUDNN_PATH/lib64

INSTALLING, BUILDING AND RUNNING archcomputing/cuda:0.3 (Docker Image) and cuda (Docker Container)

1. Build/Rebuild image

   docker-compose build 
   

2. Create and start container

   docker-compose up -d
   

3. Enter to the container

   docker exec -it cuda /bin/bash
   

4. Now you can compile and run the programs in the container, either manually or using the Makefile

Compiling in NVIDIA CUDA COMPILER (nvcc)

Manual

nvcc --cudart shared -o out program.cu

Other commands need to be executed to clean up compiler-generated files and extent file compilation if the program has more than one file

Makefile

There is a Makefile, its execution is optional. Automates the build/compilation, execution, indentation and cleanup process

Building

make target=program_filename_without_extension

or

make all target=program_filename_without_extension

Clean

make clean target=program_filename_without_extension

Running

make run target=program_filename_without_extension

Indent

make indent target=program_filename_without_extension

Example

when a cuda device is available

License

MIT