RadeonSI Picks Up Primitive Culling With Async Compute For Performance Wins

[jrch2k8]

As much as I love everything Marek does, those results seem waaaay too good to be true. A 30% improvement is a hefty achievement, but the Radeon 7 nearly got a 300% improvement. That's insane, when you consider that GPU was running approximately where it was expected to run before async compute.

i would not be sure about your PoV because if you stall the geometry pipeline processing a huge amount of invisible data(from the rendering PoV) the driver may stall all or most other pipelines waiting for results and severely affect performance and FirePro(on windows at least) are know to be monsters on pro apps specially when hardcore geometry is handled.

On the other hand i don't think this will help games at all or very very few at best because simply speaking games usually dumb down geometry to the extreme

[skeevy420]

Pretty cool. I managed to find a 4GB RX 580 for $120 after shipping and it'll be here Friday. Obviously, that's a used price, but I just need something to hold me over until this time next year and the 7nm AMD GPUs have been out long enough for both the price drop and the drivers to get better. I've read enough stories on Phoronix about AMD drivers sucking the first month of a GPUs release to not want to be an early adopter. 4GB is just fine since 1080p is the highest resolution I plan on using with it.

All I know is it'll be one hell of an upgrade from my R7 260x that's in the process of dying. Hopefully, the dual 6 pin to 8 pin converter will be here by then too. It'll really suck looking at an RX 580 for a week without being able to use it. Awesome news when you buy a card and a day later it gets patches that might make it 30% more powerful and compete with a 1080.

I haven't been able to play anything more demanding than an SNES emulator since October...I'll be nice to play my games again at damn-near maximum settings.

[marek]

As much as I love everything Marek does, those results seem waaaay too good to be true. A 30% improvement is a hefty achievement, but the Radeon 7 nearly got a 300% improvement. That's insane, when you consider that GPU was running approximately where it was expected to run before async compute.

In my synthetic tests, the maximum improvement is 450% on Radeon VII when all geometry is invisible. It's just a good compute chip.

[Peter Fodrek]

Many thanks to Marek

[nuetzel]

Did some first tests on Polaris 20 (RX580 8 GB):

Maybe my LLVM git (7. Jan / last 8.0 before cut) is to old.
Or it's Timo's new 'A few NIR compile time optimisations' (have to revert)...

Blender 2.79b / UH / UV
LLVM ERROR: Cannot select: 0x7fedd4599bc8: i64 = SETCC 0x7fedd4599888, Constant:i1<0>, setne:ch
0x7fedd4599888: i1,ch = CopyFromReg 0x7fee03071b58, Register:i1 %14
0x7fedd459b208: i1 = Register %14
0x7fedd45999c0: i1 = Constant<0>
In function: prim_discard_cs

LLVM ERROR: Cannot select: 0x7f6576b9ac30: i64 = SETCC 0x7f6576b9a8f0, Constant:i1<0>, setne:ch
0x7f6576b9a8f0: i1,ch = CopyFromReg 0x7f65a1cbc558, Register:i1 %14
0x7f6576b9c270: i1 = Register %14
0x7f6576b9aa28: i1 = Constant<0>
In function: prim_discard_cs
Writing: /tmp/bh.crash.txt

FreeCAD
works mostly, but no numbers, yet.

mplayer / mpv (vdpau)
crash

mplayer: ../src/gallium/drivers/radeon/radeon_winsys.h:866: radeon_get_heap_index: Assertion `!"32BIT without WC is disallowed"' failed.
Abbruch (core dumped)

mpv: ../src/gallium/drivers/radeon/radeon_winsys.h:866: radeon_get_heap_index: Assertion `!"32BIT without WC is disallowed"' failed.
Abbruch (core dumped)

Cheers,
Dieter

[tildearrow]

The cool thing here: Although the feature is disabled by default for consumer chips, we can still play around with it:

I have a question. Is Vega FE considered a "pro" card in this case?

[marek]

Did some first tests on Polaris 20 (RX580 8 GB):

Maybe my LLVM git (7. Jan / last 8.0 before cut) is to old.
Or it's Timo's new 'A few NIR compile time optimisations' (have to revert)...

Blender 2.79b / UH / UV
LLVM ERROR: Cannot select: 0x7fedd4599bc8: i64 = SETCC 0x7fedd4599888, Constant:i1<0>, setne:ch
0x7fedd4599888: i1,ch = CopyFromReg 0x7fee03071b58, Register:i1 %14
0x7fedd459b208: i1 = Register %14
0x7fedd45999c0: i1 = Constant<0>
In function: prim_discard_cs

LLVM ERROR: Cannot select: 0x7f6576b9ac30: i64 = SETCC 0x7f6576b9a8f0, Constant:i1<0>, setne:ch
0x7f6576b9a8f0: i1,ch = CopyFromReg 0x7f65a1cbc558, Register:i1 %14
0x7f6576b9c270: i1 = Register %14
0x7f6576b9aa28: i1 = Constant<0>
In function: prim_discard_cs
Writing: /tmp/bh.crash.txt

FreeCAD
works mostly, but no numbers, yet.

mplayer / mpv (vdpau)
crash

mplayer: ../src/gallium/drivers/radeon/radeon_winsys.h:866: radeon_get_heap_index: Assertion `!"32BIT without WC is disallowed"' failed.
Abbruch (core dumped)

mpv: ../src/gallium/drivers/radeon/radeon_winsys.h:866: radeon_get_heap_index: Assertion `!"32BIT without WC is disallowed"' failed.
Abbruch (core dumped)

Cheers,
Dieter

You need LLVM 8.0 final version.

[Dukenukemx]

So does this make OpenGL faster than RADV Vulkan?

[cb88]

So does this make OpenGL faster than RADV Vulkan?

Probably it has the same effect on both if implemented for both as it is a GPU side optimization, Vulkan just has lower CPU overhead.

[clapbr]

You need LLVM 8.0 final version.

LLVM 8 isn't officially released at least in their official binaries or my distros. Kerberizer llvm-svn repo has arch binaries but they're 2 months old (his pipeline was broken by git migration). I use them and had same prob as guy above. Time to build this monster... https://github.com/llvm/llvm-project/tree/release/8.x