|
909ea88b
|
2020-11-20T13:07:53
|
|
Reland "Vulkan: Ignore glFlush to reduce vkQueueSubmits in Asphalt 9"
This is a reland of 5cf7472dd161bbda329dfc5e4e65bb6ce0c06fbd
The ShareGroupVk::mResourceUseLists was not being cleared each call to
RendererVk::submitFrame(), so it was growing indefinitely. Each
vk::ResourceUseList within it was cleared, so it was holding an
essentially "infinite" list of empty lists, but that caused the loop in
RendererVk::submitFrame() to take more and more time until the tests
timed out.
The fix is to do 'resourceUseLists.clear()' once the loop to release all
resources has completed, like releaseResourceUsesAndUpdateSerials() does
for each individual list. Additionally, ASSERTs are added to guarantee
that the lists are empty when the ContextVk and ShareGroupVk are
destroyed.
Original change's description:
> Vulkan: Ignore glFlush to reduce vkQueueSubmits in Asphalt 9
>
> Multithreaded apps can use the following pattern:
>
> glDrawElements()
> glFenceSync()
> glFlush()
> glWaitSync()
>
> This currently results in a vkQueueSubmit for every glFlush() to ensure
> that the work has landed in the command queue in the correct order.
> However, ANGLE can instead avoid the vkQueueSubmit during the glFlush()
> in this situation by instead flushing the ContextVk's commands and
> ending the render pass to ensure the commands are submitted in the
> correct order to the renderer. This improves performance for Asphalt 9
> by reducing frame times from 150-200msec to 35-55msec.
>
> Specifically, ANGLE will call flushCommandsAndEndRenderPass() when
> there is a sync object pending a flush or if the ContextVk is currently
> shared.
>
> Additionally, on all devices except Qualcomm, ANGLE can ignore all other
> glFlush() calls entirely and return immediately. For Qualcomm devices,
> ANGLE is still required to perform a full flush (resulting in a
> vkQueueSubmit), since ignoring the glFlush() reduces the Manhattan 3.0
> offscreen score by ~3%.
>
> Bug: angleproject:5306
> Bug: angleproject:5425
> Change-Id: I9d747caf5bf306166be0fec630a78caf41208c27
> Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2552718
> Commit-Queue: Tim Van Patten <timvp@google.com>
> Reviewed-by: Charlie Lao <cclao@google.com>
> Reviewed-by: Jamie Madill <jmadill@chromium.org>
Bug: angleproject:5306
Bug: angleproject:5425
Bug: angleproject:5470
Change-Id: I14ee424d032f22e5285d67accbec078ad1955dd0
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2595811
Commit-Queue: Tim Van Patten <timvp@google.com>
Reviewed-by: Charlie Lao <cclao@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
|
|
60641abc
|
2020-11-16T16:12:31
|
|
Vulkan: Make ImageViewHelper a Resource
Bug: angleproject:3573
Change-Id: I12e70418b3b971e802bc911409e170cbf8c61915
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2542223
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
fcd35965
|
2020-09-29T14:15:51
|
|
Vulkan: Make DescriptorPoolHelper a Resource
Descriptor pools need to live as long as the descriptor sets that are
allocated from them. Using Serials while building a command to judge a
pool's lifetime is prone to errors, since a command's Serial value
isn't known until the command is submitted, leading to deleting pools
too early relative to when the descriptor set is actually used.
This CL updates DescriptorPoolHelper to inherit from Resource, so the
descriptor pools can be retain()'ed. This allows the Resource's counter
to indicate that a pool is in use until the command's Serial is known
and can be recorded to indicate when the command completes. This
prevents descriptor pools from being destroyed before the command
completes (while the descriptor sets are still in use), or even before
the command has been submitted. Destroying a descriptor pool resets all
of the descriptors that were allocated from it, which can trigger a
variety of VVL errors depending on when it's erroneously performed.
This CL also adds the necessary retain() calls for the descriptor pools.
In particular, the pools need to be retained each time a cached
descriptor set that was allocated from it is re-used. This is relatively
simple with the current design, since we always clear the descriptor set
caches whenever a new pool is allocated, so the descriptor pool binding
is always accurate.
Bug: angleproject:5030
Test: VulkanMultithreadingTest::MultiContextDrawSmallDescriptorPools()
Change-Id: Iac9e7efef338f169a6bf8ac3b2140e03dd326641
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2504457
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Courtney Goeltzenleuchter <courtneygo@google.com>
Commit-Queue: Tim Van Patten <timvp@google.com>
|
|
9e3eec54
|
2020-10-25T15:44:09
|
|
Revert "Vulkan: Make DescriptorPoolHelper a Resource"
This reverts commit 5dcd29a6e532e4bd617af8767d488120b57f3b2c.
Reason for revert: Breaking the ANGLE -> Chromium roller:
https://chromium-review.googlesource.com/c/chromium/src/+/2496281
Original change's description:
> Vulkan: Make DescriptorPoolHelper a Resource
>
> Descriptor pools need to live as long as the descriptor sets that are
> allocated from them. Using Serials while building a command to judge a
> pool's lifetime is prone to errors, since a command's Serial value
> isn't known until the command is submitted, leading to deleting pools
> too early relative to when the descriptor set is actually used.
>
> This CL updates DescriptorPoolHelper to inherit from Resource, so the
> descriptor pools can be retain()'ed. This allows the Resource's counter
> to indicate that a pool is in use until the command's Serial is known
> and can be recorded to indicate when the command completes. This
> prevents descriptor pools from being destroyed before the command
> completes (while the descriptor sets are still in use), or even before
> the command has been submitted. Destroying a descriptor pool resets all
> of the descriptors that were allocated from it, which can trigger a
> variety of VVL errors depending on when it's erroneously performed.
>
> This CL also adds the necessary retain() calls for the descriptor pools.
> In particular, the pools need to be retained each time a cached
> descriptor set that was allocated from it is re-used. This is relatively
> simple with the current design, since we always clear the descriptor set
> caches whenever a new pool is allocated, so the descriptor pool binding
> is always accurate.
>
> Bug: angleproject:5030
> Test: MultithreadingTest::MultiContextDrawSmallDescriptorPools()
> Change-Id: I5fdeeb46159448dfd679d7169e423048348be5ab
> Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2437609
> Commit-Queue: Tim Van Patten <timvp@google.com>
> Reviewed-by: Jamie Madill <jmadill@chromium.org>
> Reviewed-by: Courtney Goeltzenleuchter <courtneygo@google.com>
TBR=courtneygo@google.com,timvp@google.com,jmadill@chromium.org
# Not skipping CQ checks because original CL landed > 1 day ago.
Bug: angleproject:5030
Change-Id: I0fd6d9a0e1b0989b22368ef98652281288699deb
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2497222
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
|
|
|
5dcd29a6
|
2020-09-29T14:15:51
|
|
Vulkan: Make DescriptorPoolHelper a Resource
Descriptor pools need to live as long as the descriptor sets that are
allocated from them. Using Serials while building a command to judge a
pool's lifetime is prone to errors, since a command's Serial value
isn't known until the command is submitted, leading to deleting pools
too early relative to when the descriptor set is actually used.
This CL updates DescriptorPoolHelper to inherit from Resource, so the
descriptor pools can be retain()'ed. This allows the Resource's counter
to indicate that a pool is in use until the command's Serial is known
and can be recorded to indicate when the command completes. This
prevents descriptor pools from being destroyed before the command
completes (while the descriptor sets are still in use), or even before
the command has been submitted. Destroying a descriptor pool resets all
of the descriptors that were allocated from it, which can trigger a
variety of VVL errors depending on when it's erroneously performed.
This CL also adds the necessary retain() calls for the descriptor pools.
In particular, the pools need to be retained each time a cached
descriptor set that was allocated from it is re-used. This is relatively
simple with the current design, since we always clear the descriptor set
caches whenever a new pool is allocated, so the descriptor pool binding
is always accurate.
Bug: angleproject:5030
Test: MultithreadingTest::MultiContextDrawSmallDescriptorPools()
Change-Id: I5fdeeb46159448dfd679d7169e423048348be5ab
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2437609
Commit-Queue: Tim Van Patten <timvp@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Courtney Goeltzenleuchter <courtneygo@google.com>
|
|
ed876984
|
2020-10-03T11:00:36
|
|
Vulkan: functionally complete worker thread
Working on enhancing worker thread to completely own primary command
buffers. This will include not only processing SCBs from main thread
into a primary, but also submitting those command buffers to the queue.
The CommandProcessor is a vk::Context so it can handle errors in the
worker thread. When the main thread submits tasks to the worker
thread it also syncs any outstanding errors from the worker.
Include asynchronousCommandProcessing feature that will control
whether the worker thread task does it's work in parallel or not. If
false, we wait for the thread to complete it's work before letting the
main thread continue. If true, the thread can execute in parallel with
the main thread.
Bug: b/154030730
Bug: b/161912801
Change-Id: I00f8f013d6cbb2af12a172c4f7927855db2f0ebf
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2328992
Commit-Queue: Courtney Goeltzenleuchter <courtneygo@google.com>
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
|
|
|
68a5baeb
|
2020-09-23T22:13:03
|
|
Revert "Vulkan: Implement a SharedResourceUse pool"
This reverts commit de335c16855f11d1f0a6f0b37bee30c8a09a6c1d.
Reason for revert: Might actually regress CPU overhead perf.
Unsure but it's possible the reported perf improvement was due
to variance.
Original change's description:
> Vulkan: Implement a SharedResourceUse pool
>
> When adding a Resource to the ResourceUseList of ContextVk
> we constructed a new SharedResourceUse object for tracking
> and update of the Resource's Serial. We would then delete
> it after releasing the resource. This incurs repeated
> memory operation costs.
>
> Instead we now allocate a pool of SharedResourceUse objects
> and acquire and release from this pool as needed.
>
> VTune profile of the Manhattan 30 offscreen benchmark
> shows the CPU occupancy of bufferRead decrease from an
> average of 0.9% -> 0.6% and imageRead decreases from
> an average of 0.4% -> 0.3%. The bottleneck for both
> these methods is the retain() method that leverages
> the new SharedResourceUse pool.
>
> Bug: angleproject:4950
> Change-Id: Ib4f67c6f101d4b2de118014546e6cc14ad108703
> Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2396597
> Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
> Reviewed-by: Jamie Madill <jmadill@chromium.org>
> Commit-Queue: Mohan Maiya <m.maiya@samsung.com>
TBR=syoussefi@chromium.org,jmadill@chromium.org,m.maiya@samsung.com
# Not skipping CQ checks because original CL landed > 1 day ago.
Bug: angleproject:4950
Change-Id: I40081551c3db67d6e55182fea40119946ed16ac3
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2426479
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
de335c16
|
2020-09-14T12:04:20
|
|
Vulkan: Implement a SharedResourceUse pool
When adding a Resource to the ResourceUseList of ContextVk
we constructed a new SharedResourceUse object for tracking
and update of the Resource's Serial. We would then delete
it after releasing the resource. This incurs repeated
memory operation costs.
Instead we now allocate a pool of SharedResourceUse objects
and acquire and release from this pool as needed.
VTune profile of the Manhattan 30 offscreen benchmark
shows the CPU occupancy of bufferRead decrease from an
average of 0.9% -> 0.6% and imageRead decreases from
an average of 0.4% -> 0.3%. The bottleneck for both
these methods is the retain() method that leverages
the new SharedResourceUse pool.
Bug: angleproject:4950
Change-Id: Ib4f67c6f101d4b2de118014546e6cc14ad108703
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2396597
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Mohan Maiya <m.maiya@samsung.com>
|
|
|
295d2ccd
|
2020-08-24T14:46:31
|
|
Vulkan: Generate perf warnings on suboptimal paths
Using KHR_debug features, this change creates a
performance-warning-generation macro and employs it in a handful of
locations to provide useful feedback to application developers.
The warnings added in this change are not exhaustive.
Bug: angleproject:3461
Bug: angleproject:4900
Change-Id: Id62435d170d90c5be9c1c5cab2d6779ccb58345e
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2372628
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
50442fac
|
2020-08-05T14:15:12
|
|
Vulkan: Fix ImageHelper's move constructor
Bug: angleproject:4913
Change-Id: Ic78a26be4c2f3fa96ef77deffc239dbb7310065e
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2339543
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
745e0712
|
2020-03-21T17:46:05
|
|
Vulkan: Enable CPU only buffers for PBOs
Add support for a CPU only buffer for PBOs that serve as
the destination for all host operations like MapBuffer*.
This removes the latency caused by waiting for the in-flight
GPU commands to be complete before handing over the buffer
to the app.
This change removes a ~6ms wait/sleep on the first call to
MapBuffer* in each frame of Manhattan
Bug: angleproject:4339
Tests: angle_end2end_tests --gtest_filter=BufferDataTest*Vulkan
Change-Id: I52016b160af8a670cc30f01c05e48f699521310f
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2116874
Commit-Queue: Mohan Maiya <m.maiya@samsung.com>
Reviewed-by: Tobin Ehlis <tobine@google.com>
|
|
449d9d76
|
2020-03-31T17:27:00
|
|
Vulkan: Refactor garbarge collection related parameter
Use RendererVk instead of VkDevice as a parameter in garbage collection
functions.
Bug: angleproject:2162
Change-Id: Ifd53e05223d6d603402c9b7fcfa82fe1f896458c
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2131882
Commit-Queue: Mohan Maiya <m.maiya@samsung.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
|
|
|
a741abb9
|
2020-02-21T16:37:37
|
|
Vulkan: Rename CommandGraphResource to Resource.
Also renames the h and cpp files to ResourceVk (to keep distinct from
other resource.h/cpp files) and renames 'onResourceAccess' to 'retain'.
Cleans up a few remaining mentions of the command graph in comments.
Bug: angleproject:4029
Change-Id: Ifc8e880c8cea3fc48a4aec4730191c88aa35a076
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2065920
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
|