Marek Working On 32-bit GPU Pointers For RadeonSI

[phoronix]

Phoronix: Marek Working On 32-bit GPU Pointers For RadeonSI

Well known open-source AMD 3D driver developer Marek Olšák has published a set of new patches featuring his latest optimization work: 32-bit GPU pointers...

http://www.phoronix.com/scan.php?page=news_item&px=RadeonSI-32-bit-GPU-Pointers

[cb88]

That sounds like it should have significant performance impact especially for compute... perhaps they are looking at updating clover or better integrating with ROCm

[rene]

this sounds like x32 of amd64 or n32 of mips64, aka 32 bit pointers for 64 bit CPUs, to use shorter pointers to free registers, and cache utilization?

[mannerov]

This looks like a good idea, and it seems to reduce spilled VGPR too, thus it helps the llvm scheduler.

[boxie]

this sounds like x32 of amd64 or n32 of mips64, aka 32 bit pointers for 64 bit CPUs, to use shorter pointers to free registers, and cache utilization?

it does seem like that, it will be interesting to see what the performance impact is, and if there are any regressions

[Jahimself]

Marek is everywhere?

[davidbepo]

Marek is everywhere?

marek is like santa he gives christmas presents(https://www.phoronix.com/scan.php?pa...-Glxgears-Xmas) and can be everywhere or nowhere

[wizard69]

it does seem like that, it will be interesting to see what the performance impact is, and if there are any regressions

It may very well help some apps running on GPUs. I just find it a bit funny when you can buy GPU cards with far more addressable RAM than a 32 bit pointer allows for. Of course these days buying a GPU card with more RAM than most of the computers ive ever owned is a possibility.

While i can see a positive performance impact im not convinced there will be many regressions. I would imagine this feature will be exposed as new datatypes for the programmer to exploit. One thing for sure, if there is a problem Micheals testing will show it to us.

[boxie]

It may very well help some apps running on GPUs. I just find it a bit funny when you can buy GPU cards with far more addressable RAM than a 32 bit pointer allows for. Of course these days buying a GPU card with more RAM than most of the computers ive ever owned is a possibility.

While i can see a positive performance impact im not convinced there will be many regressions. I would imagine this feature will be exposed as new datatypes for the programmer to exploit. One thing for sure, if there is a problem Micheals testing will show it to us.

i wonder if this gets enabled for only 4GiB < cards (as you would need the 64bit pointers for accessing the large memory areas).

[M@yeulC]

i wonder if this gets enabled for only 4GiB < cards (as you would need the 64bit pointers for accessing the large memory areas).

Most likely you can use two addressing modes... Then I am just speculating, I must admit that I didn't bother researching. Another possibility *could* be addressing larger chunks of memory.

Edit: it does look like there are some new heap flags, so you can most certainly use it along with 64bit addressing for other parts of the program. Mesa code certainly looks like a treasure trove to anyone interested in GPUs, I'll have to have a deep look at it one of these days