Parallel-Face-Recovery-Recognition

Parallel Implementation of Nearest Neighbor (image match) algorithm and linear interpolation (occlusion recovery) algorithm; using MPI and CUDA.

How to Run (Ensure access to CUDA, MPI, and clockcycle.h); Can also use run.mk:

• mpixlc -g face.c -c -o face-mpi.o
• nvcc -g -G face.cu -c -o face-cuda.o
• mpicc -g face-mpi.o face-cuda.o -o face-exe -L/usr/local/cuda-11.2/lib64/ -lcudadevrt -lcudart -lstdc++ -lm
• mpirun --bind-to core --report-bindings -np 1 face-exe 1 1 1024

The first argument chooses which training file for input (1 indicates training file with 1*360 images)

The second argument chooses which testing file for input (1 indicates training file with 1*360 images)

The third argument chooses number of threads per block for CUDA

How to run comments also present in face-serial.c, face-mpi.c, and face.c

Test Cases:

Sequential:

• Train 1, Test 1
• Train 2, Test 2
• Train 4, Test 4
• Train 8, Test 8
• Train 16, Test 16

Strong Scaling Study 1 (Train 16, Test 16, No GPU, face-mpi.c):

• Rank 1
• Ranks 2
• Ranks 4
• Ranks 8
• Ranks 12
• Ranks 24
• Ranks 36

Strong Scaling Study 2 (Train 16, Test 16, 1 GPU):

• Ranks 1, Blocksize 1
• Ranks 2, Blocksize 8
• Ranks 4, Blocksize 16
• Ranks 8, Blocksize 32
• Ranks 12, Blocksize 128
• Ranks 24, Blocksize 512
• Ranks 36, Blocksize 1024

Weak Scaling Study(Blocksize 1024):

• 1 GPU/Rank, Train 1, Test 1
• 2 GPU/Rank, Train 2, Test 2
• 3 GPU/Rank, Train 3, Test 3
• 4 GPU/Rank, Train 4, Test 4
• 5 GPU/Rank, Train 5, Test 5
• 6 GPU/Rank, Train 6, Test 6