GPU vs CPU analysis

What has been made:

• Transposition
• Optimization of recurrent operations in loops
• AVX
• OpenMP
• CUDA
• Tile based

Usage

• Compile it by using "make"
• Generate a matrix with "./BINARY g L1 C1 L2 C2 s"
  • "BINARY" is the binary to be used (./cuda or ./norm)
  • "g" tells the software to generate a matrix
  • "L1 C1" are respectively matrix A lines and columns
  • "L2 C2" are respectively matrix B lines and columns
  • "s", (optional) tells the software to save matrix A and B
• Reading a matrix from file "./BINARY f L1 C1 L2 C2 FILE1 FILE2"
  • "BINARY" is the binary to be used (./cuda or ./norm)
  • "f" tells the software to read a matrix from a file
  • "L1 C1" are respectively matrix A lines and columns
  • "L2 C2" are respectively matrix B lines and columns
  • "FILE1" is the name of matrix A file
  • "FILE2" is the name of matrix B file
• Compile matrix comparator with "make comparator"
• Compare matrix multiplication results with "./comp C1.txt C2.txt"
• Rename matrices from 0....txt, 1....txt e 2....txt to a.txt, b.txt and c.txt with "make rename"
• Clean txts with "make clean"

NOTE

Source files

• matrixCuda.cu: Optimized matrix multiplication (CUDA, AVX, Transposition...)
• matrix.cu: Not optmized matrix multiplication
• comparator.c: Matrix equality file comparator

Generated binaries

• cuda: Optimized matrix multiplication (CUDA, AVX, Transposition...)
• norm: Not optimized matrix multiplication
• comp: Matrix equality file comparator

Output files saved with 's' flag:

• Matrix A: 0-rowsxcols.txt
• Matrix B: 1-rowsxcols.txt
• Matrix C: 2-rowsxcols.txt

Whether executing the software with the same matrix dimensions, the output files will be overwritten. Thus, is better to use "make rename" to avoid it.

Usage example:

./cuda g 1000 1000 1000 1000 s
make rename
./norm f 1000 1000 1000 1000 a.txt b.txt
./comp 0-1000x1000.txt c.txt