cuda_loop, a FORTRAN90 code which demonstrates how the user's choice of CUDA blocks and threads determines how the user's tasks will be distributed across the GPU.
A CUDA kernel "kernel()" is invoked by a command of the form
kernel << blocks, threads >> ( args )
where blocks and threads are each vectors of up to 3 values,
listing the number of blocks and number of threads to be used.
If a problem involves N tasks, then tasks are allotted to specific CUDA processes in an organized fashion. Some processes may get no tasks, one task, or multiple tasks.
Each process is given variables that can be used to determine the tasks to be performed:
Essentially, a process can determine its linear index K by:
K = threadIdx.x
+ blockdim.x * threadIdx.y
+ blockDim.x * blockDim.y * threadIdx.z
+ blockDim.x * blockDim.y * blockDim.z * blockIdx.x
+ blockDim.x * blockDim.y * blockDim.z * gridDim.x * blockIdx.y
+ blockDim.x * blockDim.y * blockDim.z * gridDim.x * gridDim.y * blockIdx.z
It should use this index as follow:
Set task T = K.
while ( T < N )
carry out task T;
T = T + blockDim.x * blockDim.y * blockDim.z * gridDim.x * gridDim.y * gridDim.z.
The CUDA_LOOP program suggests how a specific set of block and thread parameters would determine the assignment of individual tasks to CUDA processes.
The computer code and data files made available on this web page are distributed under the MIT license
cuda_loop is available in a C version and a C++ version and a FORTRAN90 version and a MATLAB version and a Python version.