|
4a41204d
|
2020-09-02T05:04:33
|
|
Vulkan: Improve invalidate for depth/stencil
Improve state tracking when the depth and/or stencil attachments are
invalidated. Since no draw-time tracking is done, we use the number
of command-buffer commands to determine when an attachment is drawn
to. That allows all cases to be handled for store ops. Still need to
handle mContentDefined at endRP time (we have the data, just not the
plumbing).
Test: angle_white_box_tests --gtest_filter=VulkanPerformanceCounterTest.*Invalidate*/*
Test: angle_deqp_gles3_tests --gtest_filter=dEQP.GLES3/functional_fbo_invalidate_* --use-angle=vulkan
Bug: b/167276207
Change-Id: Iae10857dbb4d43b934c51ad7e400b71ae0db4f55
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2378670
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Charlie Lao <cclao@google.com>
Commit-Queue: Ian Elliott <ianelliott@google.com>
|
|
b2944fcf
|
2020-09-02T11:10:00
|
|
Vulkan: Don't break the RP on masked/scissored clears
Bug: b/166809097
Change-Id: Iedd10a6528808e859c5693a2d30c98aca1a1159c
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2390862
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Charlie Lao <cclao@google.com>
Reviewed-by: Tim Van Patten <timvp@google.com>
|
|
|
44df3fa9
|
2020-08-27T14:26:49
|
|
Vulkan: Fix whitebox test to end render pass
Since the RAII-based GLRenderbuffer goes away at the end of the test
function, we should end the render pass before ending the test.
Bug: chromium:1122621
Change-Id: I5aaf2a0b4fd63a38fc5018830ee4e9d7d93f0c81
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2380248
Reviewed-by: Cody Northrop <cnorthrop@google.com>
Reviewed-by: Ian Elliott <ianelliott@google.com>
Reviewed-by: Courtney Goeltzenleuchter <courtneygo@google.com>
Commit-Queue: Ian Elliott <ianelliott@google.com>
|
|
c5b9c49c
|
2020-08-25T18:01:29
|
|
Vulkan: Fix optimizeRenderPassForPresent regression.
gfxbench clears the depth buffer right before the swap. Even though the
last draw call that presents the frame didn't ever read or write to
depth, the tracking we added thought this depth write meant we had to
keep the LOAD_OP as CLEAR. Instead we can refine our check to treat
clears specially when enabling the depth-stencil read-only mode instead
of changing how the tracking works for clears. This way the tracking
can not affect other apps that don't use depth-stencil read-only loops.
Also adds a regression test that counts the clears after a swap.
Bug: angleproject:4959
Bug: angleproject:4979
Change-Id: I12ece6474019f7519a467f827110ad817f7d4df7
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2370364
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Charlie Lao <cclao@google.com>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
296d3bfd
|
2020-08-21T10:38:32
|
|
Vulkan: do not end render pass when invalidating
Initially, FramebufferVk::invalidateImpl() was very conservative and
always ended a render pass (if the framebuffer is part of the current
render pass). This adversely affects PUBG Mobile, which invalidates
the depth buffer every frame, causing the render pass to be split.
Test: PUBG MOBILE on Android
Test: angle_white_box_tests --gtest_filter=VulkanPerformanceCounterTest.InvalidatingAndUsingDepthDoesNotBreakRenderPass/*
Test: angle_deqp_gles3_tests --gtest_filter=dEQP.GLES3/functional_fbo_invalidate_* --use-angle=vulkan
Bug: b/163854287
Change-Id: I343dee1db3ebaf039ff92557f9ef25b24bcdcc93
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2352627
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Charlie Lao <cclao@google.com>
Commit-Queue: Ian Elliott <ianelliott@google.com>
|
|
|
d3e800e9
|
2020-08-15T17:26:04
|
|
Vulkan: Restart RenderPasses in DS read-only mode.
We can combine an initial RenderPass with a read-only RP if the first
RP never writes to depth. We can check the depth-write tracking in
CommandBufferHelper and substitute in a new Framebuffer/RP Desc in this
case as well as issue new layout barriers. We need to disable barrier
merging in this special case.
This reduces the RenderPass count in the Manhattan trace from 15->13.
The performance on the Pixel 4 benchmark goes to ~82% of native for
the on-screen version and ~88% for off-screen. There's also a ~5% bump
in speed for the desktop trace.
Bug: angleproject:4959
Change-Id: I70241824f75eaa1e11b50370f826abc36e91686e
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2358772
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Charlie Lao <cclao@google.com>
|
|
|
e689d316
|
2020-08-14T22:51:03
|
|
Vulkan: Don't end RenderPass on DS feedback loops.
Instead of always switching the Framebuffer to mask out depth/stencil
loops we now switch the RenderPass to a "read-only" depth/stencil mode.
Reduces the RenderPass count in Manhattan from 18->15. There are still
a couple extra RenderPasses inserted that we can get rid. We can merge
a few RenderPasses by retroactively changing a started RenderPass to
"read-only" when there are no prior recorded depth writes or clears.
Also adds a test to count the number of RenderPasses ANGLE generates in
DS feedback loop situaions.
Bug: angleproject:4959
Change-Id: I1855a45959655fc27ccd47a3469c1c672fc8fd9e
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2357973
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Charlie Lao <cclao@google.com>
Reviewed-by: Tim Van Patten <timvp@google.com>
|
|
dff47d5f
|
2020-07-14T19:10:12
|
|
Vulkan: Optimize MSAA using subpass resolve attachments
If a user is performing a blit to resolve a multisample color buffer
into a single attachment, ANGLE can use subpass resolve attachments to
resolve directly into the destination buffer as part of the render pass.
This allows the data to remain in tiler memory and reduce the extra
bandwidth required to write the multisampled data back to perform the
copy.
This work also requires restoring/reopening a render pass if it has been
finished already, assuming the finished render pass was started and for
the framebuffer that is the source for the blit command. Other objects
that were created when the render pass was started need to be updated as
well, such as the source FramebufferVk's resolve attachment, the
CommandBufferHelper's vk::Framebuffer and vk::RenderPassDesc, etc.
While this is better than performing vkCmdResolveImage(), there is still
another major part of optimizing MSAA using resolve attachments not
implemented here: discarding the multisampled image rather than writing
it to GMEM, which requires the user to invalidate the read FBO after the
blit.
This CL was verified with AGI to make sure there are no explicit blits
to resolve the multiple sampled image.
Bug: b/159903491
Test: FramebufferTest_ES31.*Blit*
Test: VulkanPerformanceCounterTest_ES31.MultisampleResolveWithBlit
Change-Id: I320a26088d8f614a295e7feec275d71310391806
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2298663
Commit-Queue: Tim Van Patten <timvp@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Charlie Lao <cclao@google.com>
|
|
d63bca58
|
2020-08-06T17:37:18
|
|
Vulkan: Add test for RGB texture not break renderpass
This adds a test that will use an empty RGB texture. It should not break
renderpass.
Bug: b/162603208
Change-Id: I90f5aa2fafdbd70c654c9ad0675d6def32b58313
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2338828
Commit-Queue: Charlie Lao <cclao@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
|
|
|
d201ed8b
|
2020-08-02T16:29:35
|
|
Vulkan: Track used Images in RenderPass.
Adding a simple ImageSerial tracking map in our RenderPass allows us
to know when we do or do not need to close the RenderPass on a new
Image access. This simple tracking scheme improves Manhattan
performance by up to 25% on Android. The improved perf comes from
reducing our RenderPass count (23->18 RenderPasses in our capture
scene).
Adds a FastUnorderedSet class to manage the used RP Image serials.
Updates the Query helpers to explicitly flush the RP before inserting
queries.
Bug: angleproject:4911
Change-Id: I0c34fc8e307514ebdf3e81e08d8e5aedb70ebe8f
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2334346
Reviewed-by: Charlie Lao <cclao@google.com>
Reviewed-by: Tim Van Patten <timvp@google.com>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
|
f9dd2c15
|
2020-08-02T18:03:16
|
|
Vulkan: Accumulate Buffer barriers.
Uses an unordered_map in the CommandBufferHelper to track buffer
reads and writes. Buffer barriers are tracked specially in the
CommandBufferHelper class as a barrier we execute immediately when
we execute the commands into the primary. So when we run into an
incompatible buffer access we must start a new command buffer.
The rules for an incompatible access are:
- when we are reading a buffer, any prior write in the same command
buffer is incompatible.
- when we are writing a buffer, any prior read or write in the same
command buffer is incopatible.
Also adds a regression test using a new performance counter.
Bug: angleproject:4429
Change-Id: I393a4ed87314f955eb998940b877ba76ea15a7b8
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2334091
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Charlie Lao <cclao@google.com>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
|
18dd0c28
|
2020-08-02T17:50:48
|
|
Vulkan: Add command buffer performance counter.
Adds a counter for the secondary command buffers (non-RenderPass).
We'll use this in an upcoming test that validates that ANGLE only
issues a single barrier (CB) in some buffer read/write scenarios.
Also adds a PerfCounters struct.
Bug: angleproject:4429
Change-Id: Iaf75ca89da3d02753897cb4066e2c56db497417e
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2334090
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
|
25b0de6b
|
2020-08-01T13:45:11
|
|
Vulkan: Squash Texture+ImageView Serial and improve caching.
Previously we regenerated TextureSerial on Texture state changes such
as base/max level changes. This caused ANGLE to update descriptor sets
even though it was using the same image view handles. This change
instead uses an ImageViewSubresourceSerial which includes both a
serial for the ImageView and a 32-bit packed subresource range. The CL
speeds up NBA2k because ANGLE no longer writes new descriptors
for Texture max level changes. Local testing showed up to a 40% speedup.
Also adds a regression test with a counter for the number of descriptor
set writes in a frame.
This change will also be useful in upcoming changes that track Image
serials in the RenderPass.
Bug: angleproject:4911
Change-Id: I66249634aa56288079acf2c0eb8aa3391103533c
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2333396
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
|
|
|
322cc825
|
2020-08-02T17:27:23
|
|
Rename RenderPass test to performance counter test.
This is a better naming for new tests that will check other counters
such as the number of command buffers, barriers, and descriptor set
writes.
Bug: angleproject:4911
Change-Id: I8b2c12f6e043c2833e64fa9627f781e61f8f0f3d
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2334089
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|