Archives Discussions

[godsic]

I have a problem with AMD Stream.

I write a program for image processing and want to use AMD Stream to accelerate it! But due to optimization for CPU all my data is represented in large 1D arrays (with size > 8192, but less then 8192x8192). Is it possible in any reasonable way copy such array into 2D stream without any conversation ? If it is inpossible can AMD implement such operations for example copy 1D to 2D, and 2D to 1D. I know it is not so hard on low level, because I use OpenGL for such purposes a lot of times. Also I notice that readStream and writeStream only works in synchronization regime. Can AMD implement async versions of such command for example readStreamAsync and writeStreamAsync?

[gaurav_garg]

You can copy 1D array into 2D stream without any problem. Brook+ data transfer methods read/write works similar to CPU memcpy and do a byte by byte copy.

You can also declare 1D streams of size > 8192. Brook+ virtualise larger 1D streams with smaller 2D GPU buffers internally, but it has some performance overhead.

[godsic]

Thanks I will try, but what size? Is ATI Stream use non_power_of_two extension for samplers? I know from my OpenGL experience that such samples GPU proceed slower than power_of_two due to memory and cache structure of AMD GPU?

[gaurav_garg]

For larger 1D streams (> 8192), Brook+ tries to use GPU buffers with width of power of 2. Otherwise, the GPU buffer is allocated of same size as of stream.

Brook+ allows non power of 2 streams, but as you said it is expected to be slower.

[godsic]

Thanks for replay. I will try!

It is a little bit hard migrate from OpenGL to ATI Stream 🙂

[ryta1203]

Would it be easier to just redo the representation into 2D, even on the CPU side? Just curious.

[godsic]

For CPU part it can reduce cache efficiency!!!!! That is why I use 1D array!

[Gipsel]

Originally posted by: godsic For CPU part it can reduce cache efficiency!!!!! That is why I use 1D array!

As you can control how you iterate through your data, it does not matter to the CPU as a 2D array is also stored in a linear fashion. You only have to access the data elements of your array in the apropriate order.