Basic CUDA PROJECTS

Consists of 2 CUDA Projects:-

Vector Addition: Adds two simple vectors on the GPU. I used this to understand how indexing works in GPU thread blocks.
Image to Grayscale: Converts in a single image to grayscale.

Interesting Stats about the Programs (via NSight)

For the Vector Addition

We allocate 16 MB of memory for the program (indicated by the Dark Green Bar)
Kernel Execution: Represents kernel execution on the GPU, which refers to the actual code or function running on the GPU cores.
The value "2%" likely indicates that this timeline snapshot is capturing a segment where only 2% of kernel functions are active.
On the 813ms we see an increase in the memory usage. It's a gradual increase in the memory allocation which allocates each vector to the GPU. We allocated 16 MB.
Each increase indicates each vector being allocated
The GPU is using 98% (of 16MB) of its allocated memory resources.
The light green bars indicate that memory is being allocated.
The dark red bars indicate that memory is being used

For the Gray Scale Image

In this case, since we are dealing with image data, we need to allocate 64 MBs instead. (indicated by the Dark Green Bar)
Instead of using multiple vectors, we loaded all the image data. This is why we see an immediate increase instead of a gradual increadse
We are also able to see a more defined period during which the function is ran, which is indicated by the purple block in kernel.
This is because it takes a longer time to run compared to adding two vectors

The reason why we see the memory being allocated after a long period of time is because we deallocate only after the program is done.
We only free the memory after the program ends.
We could free the memory after the CUDA function is ran and the data is transferred to the CPU.
This would result in the memory usage going to 0 immediately after red.