Global load

[karloballa]

I want to optimize memory loads and tried several different instructions to see how it goes. Up to now, however, I got success only with flat_load.

I would like to compile a kernel for RX 4xx/5xx, but if I use global_load_dword** I'm getting "unknown instruction". I suppose that there are some flags for that...

Also, I have a question about "buffer resources". How to access them, and how to use them? I tried several kernel setups but didn't have luck. Is there any table which explains what is where? Who decides what info goes to which register (driver?)? Infos on the net are sparse and most of them are contradictional...

Thanks in advance

[matszpk]

The GLOBAL_* and SCRATCH_* instructions has been introduced in the RX Vega GPU and they are not available in Fiji/Polaris GPU's. The resource buffer was used in old OpenCL implementation and in the first GCN GPU generation (Tahiti, Pitcairn, HD 7xxx). For new OpenCL drivers and newer GPU's the FLAT_* instructions are recommended to access memory.

[karloballa]

Thanks for the explanation,

It seems that I misinterpreted GCN generations. I thought that Polaris is GCN 1.4 and Vega GCN 1.5.

As for buffer_load it looked to me that I could spare some instructions. Now it is understandable why the compiler is forcing flat_load.