9 replies to this topic

[derKai]

Hi,
i heard about some optimization that is done by the driver depending on the calls you make.. this can change the branch prediction for example..

in my master thesis i measured some timings of different steps in my pipeline. 2 of 6 steps depend on the geometry of the input model.. the rest is fixed.. (e.g. rendering a background scene, composition of targets to final output, ...). on my GTX 560 the fixed parts take constant time like expected and the 2 other steps scale with model complexity.. fine ;).. . on another GPU (AMD..something).. everything is changing (more or less slightly)..

are there some articles or technical reports that can explain this behavior? (something to cite maybe ;) )
i can't find anything.. maybe you know where to start.. or wrote it already in a book ;)

thanks a lot!

[osmanb]

It depends on what those "fixed" parts are doing. I find it much more likely that those passes are not actually constant time, than that the driver is doing something inconsistent. The kinds of full-screen passes used in modern techniques can often have variable timing: Variable width blurs can be affected by texture cache, branches or loops in complex shaders can be sensitive to the input images, etc...

[Kaptein]

maybe unrelated, but is it _always_ better to use if(something) color += texture2D(...)
or is it better to avoid the if()'s where you can as a general rule?

does the GLSL specific functions work just like if's? (clamp, min, max, normalize)
i realize that drivers likely will execute clamp, min and max fast (and that normalize is always slow)

Edited by Kaptein, 01 December 2012 - 06:47 PM.

[Hodgman]

in my master thesis i measured some timings of different steps in my pipeline. 2 of 6 steps depend on the geometry of the input model.. the rest is fixed.. (e.g. rendering a background scene, composition of targets to final output, ...)

Sorry, I don't have many links to documents to help you understand the drivers/hardware behaviour, but it would be interesting to hear your approach to isolating these steps.
Typically it's hard to measure any single operation -- e.g. composing targets executes a cheap shader (low chance of ALU being the bottleneck), but puts a lot of pressure on texture fetching and an equal amount on ROP, making it hard to tell which of the two is the bottleneck.
Also, even within a single generation of GPUs, the difference in bandwidth between low and high-end models can differ by ~20x, meaning an operation that's ALU-bound on one model may be fetch-bound on another model.

There's a careful balance between fetch latencies (and the GPUs bandwidth), the number of temp registers required by a shader (and the size of the GPU's register file) and the amount of non-fetch (ALU) work (and the GPU's speed at ALU work).
e.g. if a shader is fetch-bound, then decreasing temp-reg usage (or increasing the GPU's reg file size) might remove that bottleneck. In the same situation, you may be able to add more ALU work "for free" (assuming you don't add more temp-registers as well), or you might be able to assume that when moving to a GPU with lower ALU-speed, then performance won't be affected.

on another GPU (AMD..something).. everything is changing (more or less slightly)..

What do you mean exactly? The timings change run-to-run, as in they're random? Or they've changed from the previous GPU? Or they change based on some API state that's set?

Edited by Hodgman, 01 December 2012 - 11:10 PM.

[Steve_Segreto]

I've heard about certain drivers re-compiling the current shader program(s) if the user software changes constant register values.

[derKai]

Thanks for your replies!
first of all i'll try to explain what i'm doing ;)
in general.. . it's an augmented reality application.. so there are some steps to perform..

• rendering background layer (at the moment the model of a room from the inside.. to simulate the camera).. the camera does not move and so the afford should be constant
• generate shadow maps of the room above and the model below (see next step)
• render virtual objects..they are changing for performance measures.. (i restart the application to change the model) ... these models also don't move
• render virtual shadows.. therefor the room geometry and the shadow maps are used.. should be fixed time
• blend results of 1. 3. and 4. together

these steps are always the same.. (so i always have the same calls from cpu side)

i don't use branches.. all if's are replaced by step() and lerp().. (forcing the gpu to always calculate both paths?)
the number of texture fetches per pixel is constant

to measure timings i use queries:
http://mynameismjp.wordpress.com/2011/10/13/profiling-in-dx11-with-queries/
so here can be a problem because this timer is not really high precision
but the fast GPU works as expected... the slow doesn't

so.. now to measured values:
i compare timings of the 5 steps from runs with three different models: (primitive counts: 5000, 30000 and 50000)

on the geforce the three timings of 1 4 and 5 are equal and the timings of 2 and 3 are growing with primitive counts.
over time the measurements changing very little.. e.g. 1/100ms

on the amd-chip 2 and 3 are growing with primitive count, too. but steps 1 4 and 5 are changed by -+ 1ms

e.g.:
Rendering Backgorund:
model-a model-b model-c
0.23ms 0.23ms 0.23ms - Geforce
1,82ms 1,77ms 1,74ms - AMD

Composition:
model-a model-b model-c
0.17ms 0.17ms 0.17ms - Geforce
8,17ms 7,04ms 6,72ms - AMD

i think hodgeman's explaination points in the right direction..
the AMD-Chip is a mobile chip without vram .. so texture fetches are very expensive (for cache misses)..
is it possible that the driver uses different schemes to predict what texture data blocks are needed.. so that this leads in different hit-miss-ratio ?

Edited by derKai, 02 December 2012 - 05:33 AM.

[derKai]

Update:
i posted a link to MJPs blog post and said i used this method to measure my timings..
i forgot that this is actually not totally true^^..

there are two ways..
first: wait in a while loop until the query is finished (this will sync cpu and gpu)
second: do not wait.. instead collect the results in the next frame (this will not sync)

for my measuring i took the second one.. . now i tried to sync, what was slowing down the frame rate but the timing are much closer now.. what is quite interesting (the fact that the timings got closes)

[MJP]

The problem with queries for performance measurements is that they don't always measure what you want them to measure. Looking at the difference in GPU timestamps gives you the amount of time that it takes the GPU to process all commands in the command buffer that were between the timestamps, which doesn't necessarily tell you the total time required to fully execute those commands. This is because a GPU is pipelined, and can be working on multiple commands (even draw commands) simultaneously. Therefore they're not so great for fine-grained measurements. They'll work better if you can place them around natural sync points for the GPU, such as render target switches.

[derKai]

would a context.flush help? before setting up the query and before waiting for the results?

[Steve_Segreto]

You know what I've heard people do is they have business arrangements with the top 3 graphic card manufacturers and use backdoor API calls to cause the drivers to insert sync points so they can conduct accurate timing. Not sure if any of the undocumented APIs made it out to the web?