Seven Weapons of Cuda Optimization

To put it bluntly, I came to the company for an internship in July June as a researcher. Use CG to process previous imagesAlgorithmIt was rewritten on the GPU, and geforce8800gtx was used at that time. Later, I started to contact other jobs, learn X-ray and image acquisition. I just flipped through the manual and browsed the discussion in the Group when I was okay. In this casual day, I gradually forgot a lot of things. Now the project is restarted, and the Cuda path is also restarted.

There are not necessarily seven types of weapons.

Maybe it's just a codename. Maybe, this is imitating Gu Long ~

(1) stream

Stream

Stream stream.

There are countless streams flowing inside the invisible video card.

From main memory to global memory, stream [I] is in,

In SM, stream [I + 1] is busy with budget addition, subtraction, multiplication, and division,

At the same time, stream [I + 2] from global memory is also in.

Due to the existence of stream, the brothers on all fronts are busy at the same time. You are not fighting alone!

Everything seems to be too late. If there are too many streams, management is a problem. No, my 1024*1024 images, when 64 streams are opened, the performance is greatly reduced. Opening 4, 8, 16 seems to be okay.

This is what the manual says:

Set T1 = execution time of Kernel

T2 = memory transfer time

N = number of streams

Time (using stream) = max (T1, T2) + min (T1, T2)/n.

It seems that it is still advantageous for large-scale data throughput.

However, my concern is:

Is it difficult to divide the stream for operations related to data locations such as convolution?