About:

• We implemented 3x3 convolutions.
• When using the Winograd algorithm for convolutions, we used F(2x2, 3x3), which means that the output for one tile is 2x2.

OSX setup instructions:

• install latest XCode Command Line Tools
• install brew
• install armadillo brew install armadillo
• install llvm for openmp brew install llvm

Hive setup instructions:

• Download the source for armadillo
• After unzipping the directory, cmake ., then make, then make install DESTDIR=~/lib

Generate A Problem

• Compile with make
• Create a problem file by running python3 gen_problem.py > [problem filename]

Run Naive Convolution

• Use a file of the generated format (see above) as input for the program ./naive_convolution [input filename] [output filename]

Run Winograd Convolution implented serially

• ./winograd [input filename] [output filename]

Run Winograd Convolution implemented in OpenMP

• ./winograd_openmp [input filename] [output filename]

Run Winograd Convolution implemented in OpenCL

• ./winograd_gpu [input filename] [output filename]

Flops Calculation:

• All floating point additions and multiplications are counted as separate operations.
• Manually counted, as Haswell architecture does not support hardware counters

Benchmarks

• ./bench_image_size.sh
• ./extract_perf.py BENCHMARK_FILE will extract MFlop/s and times to bench_extract_mflops.csv and bench_extract_times.csv.
• NOTE: The benchmark creates many .in input files that may take up a lot of disk space. Make sure disk quota does not fill up during benchmarking.