Hi. Not too long ago I found out that gpus don’t like it when execution of code is (too) branching and it made me feel like there might be some trick to make efficient conditional processing of data. I thing I made one strange snippet. ▶ code This kernel writes a value only if invocation index is …

I would not expect that a wild write to arbitrary GPU memory “just become noops!”. This behavior could be well-defined or undefined, depending on whether or not you are using robust memory accesses. I don’t know if that’s an option in OpenCL or if it’s mandatory. But regardless, wild writes are not…

[image] Alfonse_Reinheart: I would not expect that a wild write to arbitrary GPU memory “just become noops!” The reason to think they may be noops came to me from reading reference on amd gcn isa, where something like “stores to addresses not located in range of physical memory get discarded”…

Khronos Forums

Rules regarding writes above 1 << 48

OpenCL

show post in topic

Home
Categories
Guidelines
Terms of Service
Privacy Policy

Powered by Discourse, best viewed with JavaScript enabled