Trim swapchain buffer #1727
Trim swapchain buffer #1727
Conversation
|
CI gfxreconstruct build queued with queue ID 250630. |
|
CI gfxreconstruct build # 4787 running. |
|
CI gfxreconstruct build # 4787 passed. |
| swapchain = reinterpret_cast<IDXGISwapChain3*>(swapchain_info->object); | ||
| while (extra_info->buffer_index != swapchain->GetCurrentBackBufferIndex()) | ||
| { | ||
| swapchain->Present(0, 0); |
There was a problem hiding this comment.
Making the Present() calls here will inject extra frames when a trimmed capture is recaptured. Dx12ReplayConsumerBase::ProcessSetSwapchainImageStateCommand can also inject Present() frames. Those get reported in the output of gfxrecon-info as Blank frames. Dx12StatsConsumer tracks those here:
gfxreconstruct/framework/decode/dx12_stats_consumer.h
Lines 298 to 304 in 9c295b4
The value of Blank frames is used by our CI to offset screenshot frame indices after recapture. However this PR passed the CI tests. So some follow up questions are:
- Why didn't this change affect CI success?
- Do we need to account for these Present calls in
Blank frames, and if so, how can it be done?
One thought I have as a solution for 2 (if it is needed) is that we add a new meta data command that is like InitSubresourceCommand but is specific for swapchain images, something like InitDxgiSwapchainBufferCommand and use that for restoring swapchain buffer data instead.
There was a problem hiding this comment.
For question 1:
The test needs to trimming capture, and then capture the trimmed file again, and then replay the file, it will happen the Blank frames issue. Our ci didn't do that kind of test.
Plus, although, some original files were trimmed captured files, those files were captured by old code. The old code didn't write swapchain buffer, so even it captured the trim captured files, it didn't include Blank frames.
Dx12ReplayConsumerBase::ProcessSetSwapchainImageStateQueueSubmit that initialize swapchain buffer index also runs Present to match the correct buffer index if in need. So Blank frames issue exists for long.
locke-lunarg
force-pushed
the
trim-swapchain
branch
from
5b3dde3
to
36eba5d
Compare
|
CI gfxreconstruct build queued with queue ID 251211. |
|
CI gfxreconstruct build # 4792 running. |
|
CI gfxreconstruct build # 4792 passed. |
For copy swapchain buffers: 1. The queue has to been swapchain's queue. 2. One ExecuteCommandLists could work for only one swapchain buffer. 3. The current back buffer index has to match the swapchain buffer. 4. After ExecuteCommandLists, the current back buffer index has to back init.
locke-lunarg
force-pushed
the
trim-swapchain
branch
from
36eba5d
to
a545e40
Compare
|
CI gfxreconstruct build queued with queue ID 251278. |
|
CI gfxreconstruct build # 4793 running. |
|
CI gfxreconstruct build # 4793 passed. |
No description provided.