vk: Resource management improvements #5988
Conversation
kd-11
commented
May 17, 2019
•
kd-11
commented
- Bumps max allocated draw call resources up to 16k from 4k to avoid easily running out of resources mid-frame in heavy scenes.
- Use a simple FIFO queue for frame data handling. In future, this can easily be expanded to use a present scheduler thread for frame-pacing support.
- The frame FIFO allows better management of descriptor pools; allowing them to be reset on acquisition of a frame context object that is guaranteed to have finished referencing it. This avoids a use-after-free situation that causes visual flicker on AMD hardware on some games like Infamous.
- Also fixes a broken optimization check (inFamous [NPUA80480] crashes on Vulkan after #5565 #5600)
|
Fixed exploding vertices on AMD gpu's: |
|
Perhaps it's unrelated to this PR, but found a little improvement lately |
|
I see slight fps improvement in P5 in Shibuya streets (master 8143 = 30-35FPS, PR = 38-44 FPS) and bigger one in inFamous at the safehouse (master 8143 = 40-47FPS, PR = 55-60FPS). No other improvements or regressions. |
|
performance in draw call limited starting scene in R&C ToD decreased from 12.5 to 10.5 fps |
|
400% scaling works again in NHL 16 👍 |
|
@isJuhn Can't have both; either we get flickering or we get a lower performance if we exceed allocation. I'm surprised ToD is exhausing 16k draw calls. |
|
@kd-11 after retesting with everything else closed the difference isn't as large as in my first test, 13-13.5 -> ~12 fps. what exactly do you mean by exhausting 16k draw calls? in the debug overlay it says there is only ~7700 draw calls. |
|
Still can't play at 300% (4K) with NCAA Football 10, previously worked prior to PR #5977 |
|
@isJuhn That number is the number of batches, not individual draw invocations. Its kinda my fault really, I need to rewrite the dispatcher to not rely on descriptors too much. A streaming approach should work, I can use push constants to set up the draw uid. |
- Use a simple queue to avoid redundant checks over all the contexts - Poll queue if RSX pipe is idle - Only check the queue when the frame context is dirty (after a queue operation) - Reset descriptors at the start of the frame context to avoid having to synchronize mid-frame - Fully synchronize if a descriptor reset is required mid-frame (spec compliance; also fixes flickering verts on some hardware)
…iagnostic message for OOM
- The 'max' index should take the first assigned ID; fixes problems with divisors
|
This commit on my system (Win10, i7 7700, gtx1060 3gb) causes a lot of stuttering with NiNoKuni (Fps still show 30 but the game stutters as hell), any reason why? |
|
Framepacing probably; especially if your "RSX load" shows constant 99% and the game never checks for pending submissions. |
|
Do I have to open a new issue about this, or were you already aware and working on it? |
|
Just open an issue as I already started on another task. |
|
the last commit that was merged into master fixes my issues with NCAA Football 10. Fully playable at 4K again :D |
|
less fps with this build |
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
|
Open a proper issue tracker if you have issues. Don't spam closed threads. |
