In this lab, you will write your first OpenCL kernel from scratch to perform multiplication of matrices.
The provided file matmul.cpp implements an OpenCL host program which calls a kernel mmul.
The kernel is currently implemented as a stub in the file kernel.cl.
Starting with the sequential matrix multiplication code provided in the file matrix_lib.c, implement the mmul kernel.
Verify that your kernel produces the correct results. How many work items are created?
Measure and compare the runtime and GFLOP/s between CPU and GPU. You can choose which device is used by passing the device number as a command line argument as follows:
./matmul-c++ --device 1
Now try reducing the number of work items by assigning a whole row of A to a single work item. How does this affect the performance on each device?
On the GPU, what local work group size is used? How does this affect the scheduling of work items corresponding to individual elements of the matrix?