Returning the layouts from an export operation and then using
them in a subsequent import operation allows the import to preserve
the texture contents.
This fixes Vukan image wrapping on some AMD/NVIDIA devices.
Bug: dawn:200
Change-Id: Icbb6e759856d410bb69724b9f439bc3088756d19
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/28380
Commit-Queue: Austin Eng <enga@chromium.org>
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
More code is now shared across backends.
Bug: dawn:483
Change-Id: I7ca1b8cbc2f12e408c94fbe5bca9fd29e47e0004
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/27021
Commit-Queue: Tomek Ponitka <tommek@google.com>
Reviewed-by: Austin Eng <enga@chromium.org>
- Implement QuerySet on Vulkan backend.
- Enable end2end tests on Vulkan. The timestamp tests will be enabled
in following CL which implement timestamp query.
Bug: dawn:434
Change-Id: I7ee04380c5f6b5af561cc23e28637dcae70bc7b9
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/26360
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Hao Li <hao.x.li@intel.com>
This enables depth-stencil textures to track per aspect state
independently. It lifts the restriction that depth and stencil
store ops must be the same as they now have independent clear
states. It will also enable correct barriers on Vulkan and D3D12.
Bug: dawn:439
Change-Id: I8a73187df57a1d7eee6790cb4395bdecf42b63aa
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/26127
Reviewed-by: Jiawei Shao <jiawei.shao@intel.com>
Commit-Queue: Austin Eng <enga@chromium.org>
Added the alphaToCoverage functionality with some
end2end tests. There is no validation for disabling alphaToCoverage
mode if SV_Coverage is statically used by the shader yet.
Bug: dawn:494
Change-Id: I9df15b35697ea05a064b092edae9d5d20f73c4d8
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/25761
Commit-Queue: Tomek Ponitka <tommek@google.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
This should not be necessary as the command pool should own the
memory if it was created without the creation flag
VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT. However, some drivers
leak memory if vkFreeCommandBuffers is not called.
Bug: chromium:1082181
Change-Id: Ia437fc17b2a304a248b9227b43cfff1868f9f10e
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/25062
Commit-Queue: Austin Eng <enga@chromium.org>
Reviewed-by: Stephen White <senorblanco@chromium.org>
This patch implements buffer lazy initialization before
CopyBufferToBuffer() behind the toggle LazyClearBufferOnFirstUse.
- If the source buffer is not initialized, it will be cleared to 0
before CopyBufferToBuffer().
- If the destination buffer is not initialized and the copy doesn't
overwrite the whole buffer, it will be cleared to 0 before
CopyBufferToBuffer(), otherwise the buffer shouldn't be cleared.
BUG=dawn:414
TEST=dawn_end2end_tests
Change-Id: I3d0512c6376a1ed8928e86f8e56fefebc16910fa
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/24360
Commit-Queue: Jiawei Shao <jiawei.shao@intel.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
This required changing DeviceBase::CreateBufferImpl to return
ResultOrError<Ref<BufferBase>>
Bug: chromium:1103154
Change-Id: I1a5811d293333b6ef29c988a08f2f1f84ac65702
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/24500
Commit-Queue: Kai Ninomiya <kainino@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
GetMappedRange never produces errors and instead returns nullptr when it
is disallowed. When in a correct state, should return a valid pointer as
much as possible, even if the buffer is an error or if the device is
lost.
Adds tests for error buffers and device loss, and modify existing tests
to not expect a device error.
Also removes some dead code in the Vulkan backend and adds a fix for
missing deallocation of VkMemory on device shutdown.
Bug: dawn:445
Change-Id: Ia844ee3493cdaf75083424743dd194fa94faf591
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/24160
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Stephen White <senorblanco@chromium.org>
This patch implements buffer lazy initialization before writeBuffer():
if the buffer is not initialized and writeBuffer() doesn't cover the
whole buffer, the buffer will be cleared to 0, otherwise the buffer
shouldn't be cleared.
This patch also introduces a toggle LazyClearBufferOnFirstUse for the
development of buffer lazy initialization: before buffer lazy
initialization being completely supported, all the related code will
only be enabled behind this toggle to prevent the buffers with valid
content being unexpectedly cleared.
BUG=dawn:414
TEST=dawn_end2end_tests
Change-Id: I99a2aa98ca4b9b21d69c6b32080afb525e2c4ad3
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/24041
Commit-Queue: Jiawei Shao <jiawei.shao@intel.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
All the querying of extensions was correctly handled for the WebGPU
ShaderFloat16 extensions, except that we forgot to request it when
creating the VkDevice.
In practice this doesn't crash as most drivers won't check if the
extension is enabled, and just compile the shader code correctly.
However the Vulkan Validation Layers complain. This CL fixes it by
requesting the features on device creation.
Bug: dawn:426
Change-Id: Id7307bb4f1d08a10bec9fc7bf0500c5761f96ff7
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/23203
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
Reviewed-by: Stephen White <senorblanco@chromium.org>
Reland after a refactor of how the extension handling work in the Vulkan
backend.
The original author is David Turner <david.turner.dev@gmail.com>.
This CL ensures that, on architectures with a varying subgroup size,
compute shaders are always compiled with a fixed subgroup size to
avoid consistency issues when one shader writes data in a subgroup-size
dependent layout to GPU memory, to be read by another shader in a
future dispatch.
At the moment, only Intel ICDs are known to implement this [1],
and the code uses a heuristics to chose the size of 16, which seems to
be the sweet spot according to Intel engineers.
+ Update the PNextChainBuilder class to deal with the fact that
VkComputePipelineCreateInfo::pNext is defined as a const void*,
which created compiler errors in the previous implementation.
[1] https://bugs.freedesktop.org/show_bug.cgi?id=108875
Bug: dawn:464
Change-Id: I035ee06084fcc964742f0bff4c54cff257c742ae
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/23202
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
Reviewed-by: Stephen White <senorblanco@chromium.org>
Reland after a refactor of how the extension handling work in the Vulkan
backend.
The original author is David Turner <david.turner.dev@gmail.com>.
Certain Vulkan ICDs (Intel ones notably) will compile SPIR-V
shaders with an liberal, compiler-selected, subgroup size (i.e.
either 8, 16 or 32). For more context, see [1].
This can be a problem for compute, when one shader stores data
in device memory using a subgroup-size dependent layout, to be
consumed by a another shader. Problems arise when the compiler
decides to compile both shaders with different subgroup sizes.
To work-around this, the VK_EXT_subgroup_size_control device
extension was introduced recently: it allows the device to
report the min/max subgroup sizes it provides, and allows
the Vulkan program to control the subgroup size precisely
if it wants to.
This patch adds support to the Vulkan backend to report and
enable the extension if it is available. Note that:
- This changes the definition of VulkanDeviceKnobs to
make room for the required pNext-linked chains of
extensions.
- A helper class, PNextChainBuilder is also provided in
UtilsVulkan.h to make it easy to build pNext-linked
extension struct chains at runtime, as required when
probing device propertires/features, or when
creating a new VkDevice handle.
- This modifies VulkanDeviceInfo::GatherDeviceInfo() to
use PNextChainBuilder to query extension features and properties
in a single call to vkGetPhysicalDevice{Properties,Features}2.
Apart from that, there is no change in behaviour in this CL.
I.e. a later CL might force a specific subgroup size for
consistency, or introduce a new API to let Dawn clients
select a fixed subgroup size.
[1] https://bugs.freedesktop.org/show_bug.cgi?id=108875
Bug: dawn:464
Change-Id: I01e5c28e7dac66f0a57bf35532eb192912b254fa
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/23201
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
This updates CL:
- Adds a deprecation warning to use size.depth instead of
arrayLayerCount.
- Changes all tests and samples to use size.depth.
- Adds deprecation tests for the change.
In particular the state tracking in TextureBase isn't changed yet
because it requires non-trivial changes in the backends. It will be done
in a follow-up CL.
Bug:dawn:22
Change-Id: Ic02dfb5baaba8d5b06cd339ce988e9b1d16cb5e9
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/23101
Reviewed-by: Jiawei Shao <jiawei.shao@intel.com>
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Similarly to the CL for Instance extensions, it makes each device
extension linked to an enum class and a bitset. Logic surrounding device
extensions is changed to take advantage of this to be more programmatic
and less error prone when adding support for a new extension.
Bug: dawn:457
Change-Id: Iecf623c40b890b7e00ba972d5eac0712866692b5
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/22941
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
This reverts commit 0357eed7de15e901d14f6149e91bfe03d86958de
and reland commit bdc05c3d5fef780382c3fd5e5ebcb14bd929819a.
The Vulkan-Loader has a bug where if the instance is created
with Vulkan 1.1 and not the promoted extensions, it will skip
emulation and if the ICD doesn't support Vulkan 1.1 nor the
extensions. Enable the promoted extensions, even when creating
a Vulkan 1.1 instance.
Original change's description:
> Check FP16 support on vulkan backend
>
> This patch check FP16 support on vulkan backend, and introduces
> the shader_float16 extension.
>
> BUG=dawn:426
> TEST=dawn_end2end_tests
>
> Change-Id: Ie09568a416ce9eb2c11afeede3e7da520550d5fb
> Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/21901
> Commit-Queue: Xinghua Cao <xinghua.cao@intel.com>
> Reviewed-by: Corentin Wallez <cwallez@chromium.org>
> Reviewed-by: Jiawei Shao <jiawei.shao@intel.com>
> Reviewed-by: Kai Ninomiya <kainino@chromium.org>
> Reviewed-by: Austin Eng <enga@chromium.org>
Bug: chromium:1087896, dawn:426
Change-Id: I2c4465fb2fe957966b44d3e5840112219481c639
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/22781
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Zero-sized copies are invalid in a couple backends, and in follow up
CLs CreateBufferMapped will be change to handle zero-sized buffers
correctly.
Bug: chromium:1069076
Change-Id: Ieef62a13182bbe1e939a3847980c91339e42aa8f
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/22460
Reviewed-by: Stephen White <senorblanco@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
To avoid overly ticking, we only want to tick when:
1. the last submitted serial has moved beyond the completed serial
2. or the completed serial has not reached the future command serial added
by the trackers (MapRequestTracker, FenceSignalTracker, ErrorScopeTracker).
Bug: dawn:400
Change-Id: Ie7c65acc332846ac1a27f9a18f230149d96d2189
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/19062
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Natasha Lee <natlee@microsoft.com>
This patch check FP16 support on vulkan backend, and introduces
the shader_float16 extension.
BUG=dawn:426
TEST=dawn_end2end_tests
Change-Id: Ie09568a416ce9eb2c11afeede3e7da520550d5fb
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/21901
Commit-Queue: Xinghua Cao <xinghua.cao@intel.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Jiawei Shao <jiawei.shao@intel.com>
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
All the buffer backend files had basically the same implemenations
of MapRequestTracker and the tracker was owned by device backends.
This refactor puts MapRequestTracker into its own file
and has the tracker be owned by DeviceBase and BufferBase.
Bug: dawn:400
Change-Id: Id28422b575e9c04d4435d5f119e0ffe08c2d1ce8
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/21760
Commit-Queue: Natasha Lee <natlee@microsoft.com>
Reviewed-by: Austin Eng <enga@chromium.org>
Move mCompletedSerial and mLastSubmittedSerial to Device frontend and
add getters and setters for the device backend to access such.
This is to aid the Device in taking more ownership of Serials and Ticking.
Bug: dawn:400
Change-Id: Ifa53ac294a871e484716842a3d212373b57847c4
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/20480
Commit-Queue: Natasha Lee <natlee@microsoft.com>
Reviewed-by: Austin Eng <enga@chromium.org>
This is a reland of 96c4019214e0b2d7c0843eacf96398a6a1198d1f
It includes a fix to add a dummy descriptor count if the
VkDescriptorPool would be empty, and adds a test that a bind group
with an empty bind group layout may be created and used.
Original change's description:
> Slab-allocate VkDescriptorSets
>
> This introduces a slab allocator for VkDescriptorSets which creates
> a VkDescriptorPool pre-allocated with multiple VkDescriptorSets per
> BindGroupLayout. In the future, we can deduplicate pools that have
> the same, or roughly the same, descriptor counts.
>
> This CL also removes the old DescriptorSetService and moves most of
> the functionality onto the DescriptorSetAllocator itself to keep
> the tracking logic in one place.
>
> Bug: dawn:340
> Change-Id: I785b17f4353fb3d40c9ccc33746600d6794efe7c
> Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/19320
> Reviewed-by: Austin Eng <enga@chromium.org>
> Commit-Queue: Austin Eng <enga@chromium.org>
Bug: dawn:340
Change-Id: Iabb744f110d0cab442bb857b31c87ba46bf0ad7a
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/20321
Commit-Queue: Austin Eng <enga@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
To avoid accidental memory leaks on account of using raw pointers,
use Ref<TextureBase> as method return type except at Dawn interface
boundaries.
Change-Id: I6459062ee28984de2cb1d5a2059bc70cf82b2faf
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/19580
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Rafael Cintron <rafael.cintron@microsoft.com>
This reverts commit 96c4019214e0b2d7c0843eacf96398a6a1198d1f.
Reason for revert: Breaks the roll, WebGPU CTS hits the
ASSERT(totalDescriptorCount > 0)
Original change's description:
> Slab-allocate VkDescriptorSets
>
> This introduces a slab allocator for VkDescriptorSets which creates
> a VkDescriptorPool pre-allocated with multiple VkDescriptorSets per
> BindGroupLayout. In the future, we can deduplicate pools that have
> the same, or roughly the same, descriptor counts.
>
> This CL also removes the old DescriptorSetService and moves most of
> the functionality onto the DescriptorSetAllocator itself to keep
> the tracking logic in one place.
>
> Bug: dawn:340
> Change-Id: I785b17f4353fb3d40c9ccc33746600d6794efe7c
> Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/19320
> Reviewed-by: Austin Eng <enga@chromium.org>
> Commit-Queue: Austin Eng <enga@chromium.org>
TBR=cwallez@chromium.org,kainino@chromium.org,enga@chromium.org,bryan.bernhart@intel.com
Change-Id: Icb58485f1080eab79b24fbcd834a89fc6206b80e
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: dawn:340
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/20280
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
This introduces a slab allocator for VkDescriptorSets which creates
a VkDescriptorPool pre-allocated with multiple VkDescriptorSets per
BindGroupLayout. In the future, we can deduplicate pools that have
the same, or roughly the same, descriptor counts.
This CL also removes the old DescriptorSetService and moves most of
the functionality onto the DescriptorSetAllocator itself to keep
the tracking logic in one place.
Bug: dawn:340
Change-Id: I785b17f4353fb3d40c9ccc33746600d6794efe7c
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/19320
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Austin Eng <enga@chromium.org>
This makes all backends register the default queue at device
initialization time, so that the same queue is returned by
each call to GetDefaultQueue.
All usages of CreateQueue are replaced by GetDefaultQueue
except a couple ones that could use the queue initialized by
DawnTest::SetUp.
A deprecation warning mechanism is added so that users of Dawn
can now that they should upgrade their usage of the API. It also
comes with a backdoor so we can test that they are emitted.
New DeprecatedAPITests are added that will contain tests for
deprecated APIs, and will also check that deprecation warnings
are produced.
The special casing of GetDefaultQueue in the wire will be done
in a follow-up CL to ease the review. It happens to work through
the regular wire mechanisms at the moment but returns a different
object on each GetDefaultQueue call.
Bug: dawn:22
Change-Id: I78dc1fa474769674278d30040e8d05c658b88360
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/19724
Reviewed-by: Austin Eng <enga@chromium.org>
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
DynamicUploader/RingBuffer were incorrectly assuming that they could not
get empty allocation requests. Fix this and add a test.
The test also surfaced a bug in the Metal backend where the command
recording context could be left with a blit encoder open that was not
properly handled on device shutdown.
Bug: chromium:1069076
Change-Id: I9793b37142bd509254ce2894fa9f6208e9a68048
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/19291
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
What was previously the Device's loss status is now a state that also
contains the "being created" state. Its transitions are entirely
handled in the frontend which enforces somewhat uniform lifecycles
between backends.
The backend devices' ShutDownImpl() function is now guaranteed to be
called only during the destructor, which leads to further simplification.
Previously Destroy() could also be called when the device was first
lost. This require complications because, for example, a WGPUBuffer
could still exist, and would want to call some resource allocator
service after the call to Destroy(). Now destruction of the device's
backing API objects is deferred to the destructor. (that's ok as long
as the application can't submit any more work).
WaitForCompletion is now guaranteed to be called before ShutDownImpl() iff
the call to DeviceBase::Initialize was succesful and the backing device
not lost. The idea is that after DeviceBase::Initialize, the GPU can
have some work enqueued and we need to wait for it to complete before
deleting backing API objects. In the future we might also have backend
be reentrant, using WebGPU itself to implement parts of the backend.
Reentrant calls would only be allowed after DeviceBase::Initialize.
Also the DynamicUploader that was special-cased in all backends is
now handled entirely by the frontend.
Bug: dawn:373
Change-Id: I985417d67727ea3bc11849c999c5ef0e02403223
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/18801
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
RefCounted (and derived) destructors should be protected on the class
to ensure the objects can ONLY be destructed by calling Release. This
avoids errors cause by destroying objects out from under code which
has an active reference count.
Unfortunately, many of the 'base' classes must continue having public
destructors because they are used as "blueprint" objects created on
the stack.
Added final on most-derived classes.
Ideas for future improvement:
- Change "base" objects to have protected destructors but create new
blueprint objects that privately derive from base objects. This
limits the blueprint object's usefulness to only be a blueprint.
- Modify createX methods to return Ref<Object> instead of Object*
Change-Id: I6f3b3b178118d135c4342cb912e982a3873d71af
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/18780
Commit-Queue: Rafael Cintron <rafael.cintron@microsoft.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
The effect to the user is the same, the Dawn device gets lost. However
we need to make the difference internally because when the backend
device is lost we can clean up immediately. On the contrary on internal
errors, the backend device is still alive and processing commands so we
need to gracefully shut it down.
Bug: dawn:269
Change-Id: Ie13b33a4f9ac2e1f5f98b3723d83cf1c6205c988
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/17965
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
Previously the code ASSERTed that the semaphores didn't exist on
destroy, but that's not necessarily the case. Handle destruction more
correctly.
Bug: dawn:269
Change-Id: If123e0e20b4ee157c70a1b8cc2f3b20a9473f55e
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/17963
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
This reverts commit 4ae315b0d11882f341c22147672e1661bcb3b7d7.
Reason for revert: crbug.com/1059205
Bug: chromium:1059205
Original change's description:
> Vulkan: Report and enable subgroup size control device extension.
>
> Certain Vulkan ICDs (Intel ones notably) will compile SPIR-V
> shaders with an liberal, compiler-selected, subgroup size (i.e.
> either 8, 16 or 32). For more context, see [1].
>
> This can be a problem for compute, when one shader stores data
> in device memory using a subgroup-size dependent layout, to be
> consumed by a another shader. Problems arise when the compiler
> decides to compile both shaders with different subgroup sizes.
>
> To work-around this, the VK_EXT_subgroup_size_control device
> extension was introduced recently: it allows the device to
> report the min/max subgroup sizes it provides, and allows
> the Vulkan program to control the subgroup size precisely
> if it wants to.
>
> This patch adds support to the Vulkan backend to report and
> enable the extension if it is available. Note that:
>
> - The corresponding VkStructureType enum values and
> struct types are not rolled to the third-party Vulkan
> headers used by Dawn yet, so vulkan_platform.h has been
> modified to define them if necessary. This can be
> removed in the future when the Vulkan-Headers are
> updated in a different patch.
>
> - This modifies VulkanDeviceInfo::GatherDeviceInfo() to
> use VkGetPhysicalDevice{Properties2,Features2} if the
> VK_KHR_get_device_properties2 instance extension is
> available. Otherwise, the Vulkan 1.0 APIs
> VkGetPhysicalDevice{Properties,Features} are used instead
> (and it is assumed that no subgroup size control is
> possible).
>
> - This changes the definition of VulkanDeviceKnobs to
> make room for the required pNext-linked chains of
> extensions.
>
> - A helper class, PNextChainBuilder is also provided in
> UtilsVulkan.h to make it easy to build pNext-linked
> extension struct chains at runtime, as required when
> probing device propertires/features, or when
> creating a new VkDevice handle.
>
> Apart from that, there is no change in behaviour in this CL.
> I.e. a later CL might force a specific subgroup size for
> consistency, or introduce a new API to let Dawn clients
> select a fixed subgroup size.
>
> [1] https://bugs.freedesktop.org/show_bug.cgi?id=108875
>
> Change-Id: I524af6ff3479f25b0a8bb139a062fe632c826893
> Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/16020
> Reviewed-by: Austin Eng <enga@chromium.org>
> Reviewed-by: Corentin Wallez <cwallez@chromium.org>
> Commit-Queue: Corentin Wallez <cwallez@chromium.org>
TBR=cwallez@google.com,cwallez@chromium.org,enga@chromium.org,enga@google.com,david.turner.dev@gmail.com
# Not skipping CQ checks because original CL landed > 1 day ago.
Change-Id: I893d771d7effdf83685dda3edac8a08f98d2f6e5
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/16522
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
This reverts commit 56f1678437ba88107630c97033615eaf22eb996c.
Reason for revert: crbug.com/1059205
Bug: chromium:1059205
Original change's description:
> Vulkan: Enforce fixed subgroup size for compute shaders.
>
> This CL ensures that, on architectures with a varying subgroup size,
> compute shaders are always compiled with a fixed subgroup size to
> avoid consistency issues when one shader writes data in a subgroup-size
> dependent layout to GPU memory, to be read by another shader in a
> future dispatch.
>
> At the moment, only Intel ICDs are known to implement this [1],
> and the code uses a heuristics to chose the size of 16, which seems to
> be the sweet spot according to Intel engineers.
>
> + Update the PNextChainBuilder class to deal with the fact that
> VkComputePipelineCreateInfo::pNext is defined as a const void*,
> which created compiler errors in the previous implementation.
>
> [1] https://bugs.freedesktop.org/show_bug.cgi?id=108875
>
> Change-Id: I332faa53b9f854a8abe43a7271f30d8c5deb2142
> Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/16021
> Commit-Queue: Corentin Wallez <cwallez@chromium.org>
> Reviewed-by: Corentin Wallez <cwallez@chromium.org>
TBR=cwallez@google.com,cwallez@chromium.org,enga@chromium.org,enga@google.com,david.turner.dev@gmail.com
# Not skipping CQ checks because original CL landed > 1 day ago.
Change-Id: I922eccc310505da4b4a9fc853335733ca4900fc8
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/16521
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
This CL ensures that, on architectures with a varying subgroup size,
compute shaders are always compiled with a fixed subgroup size to
avoid consistency issues when one shader writes data in a subgroup-size
dependent layout to GPU memory, to be read by another shader in a
future dispatch.
At the moment, only Intel ICDs are known to implement this [1],
and the code uses a heuristics to chose the size of 16, which seems to
be the sweet spot according to Intel engineers.
+ Update the PNextChainBuilder class to deal with the fact that
VkComputePipelineCreateInfo::pNext is defined as a const void*,
which created compiler errors in the previous implementation.
[1] https://bugs.freedesktop.org/show_bug.cgi?id=108875
Change-Id: I332faa53b9f854a8abe43a7271f30d8c5deb2142
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/16021
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Certain Vulkan ICDs (Intel ones notably) will compile SPIR-V
shaders with an liberal, compiler-selected, subgroup size (i.e.
either 8, 16 or 32). For more context, see [1].
This can be a problem for compute, when one shader stores data
in device memory using a subgroup-size dependent layout, to be
consumed by a another shader. Problems arise when the compiler
decides to compile both shaders with different subgroup sizes.
To work-around this, the VK_EXT_subgroup_size_control device
extension was introduced recently: it allows the device to
report the min/max subgroup sizes it provides, and allows
the Vulkan program to control the subgroup size precisely
if it wants to.
This patch adds support to the Vulkan backend to report and
enable the extension if it is available. Note that:
- The corresponding VkStructureType enum values and
struct types are not rolled to the third-party Vulkan
headers used by Dawn yet, so vulkan_platform.h has been
modified to define them if necessary. This can be
removed in the future when the Vulkan-Headers are
updated in a different patch.
- This modifies VulkanDeviceInfo::GatherDeviceInfo() to
use VkGetPhysicalDevice{Properties2,Features2} if the
VK_KHR_get_device_properties2 instance extension is
available. Otherwise, the Vulkan 1.0 APIs
VkGetPhysicalDevice{Properties,Features} are used instead
(and it is assumed that no subgroup size control is
possible).
- This changes the definition of VulkanDeviceKnobs to
make room for the required pNext-linked chains of
extensions.
- A helper class, PNextChainBuilder is also provided in
UtilsVulkan.h to make it easy to build pNext-linked
extension struct chains at runtime, as required when
probing device propertires/features, or when
creating a new VkDevice handle.
Apart from that, there is no change in behaviour in this CL.
I.e. a later CL might force a specific subgroup size for
consistency, or introduce a new API to let Dawn clients
select a fixed subgroup size.
[1] https://bugs.freedesktop.org/show_bug.cgi?id=108875
Change-Id: I524af6ff3479f25b0a8bb139a062fe632c826893
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/16020
Reviewed-by: Austin Eng <enga@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Overriding VK_DEFINE_NON_DISPATCHABLE_HANDLE changes the function
signatures of Vulkan functions, changing their ABI and making us
incompatible with real drivers. This removes that magic, and replaces it
with an explicit wrapper, VkHandle, which has much of the same
functionality as the original VkNonDispatchableHandle.
It adds definitions for dawn_native::vulkan::VkBuffer et al, which
shadow the native ::VkBuffer et al. This retains type safety throughout
the Vulkan backend without changing every single usage.
Notably, the following things had to change:
- An explicit conversion from VkBuffer* to ::VkBuffer* is needed for
arrays. This is implemented as a reinterpret_cast, which is still
safe as the new VkHandle still has the same memory layout properties
as VkNonDispatchableHandle did.
- When pointing to a VkHandle as an output pointer, it's now necessary
to explicitly get the native ::VkBuffer (via operator*) and point to
it.
Previously reviewed on:
https://dawn-review.googlesource.com/c/dawn/+/15580
Bug: chromium:1046362
Change-Id: I7d34ec38a805025f92165ea9a7ee07ae5c182076
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/15641
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
Commit-Queue: Kai Ninomiya <kainino@chromium.org>
This reverts commit 4e17d5c2483b63d4863162d692a1a961d1dcb958.
Reason for revert: broken on chromeos
Original change's description:
> Remove VK_DEFINE_NON_DISPATCHABLE_HANDLE magic, use explicit VkHandle wrapper
>
> Overriding VK_DEFINE_NON_DISPATCHABLE_HANDLE changes the function
> signatures of Vulkan functions, changing their ABI and making us
> incompatible with real drivers. This removes that magic, and replaces it
> with an explicit wrapper, VkHandle, which has much of the same
> functionality as the original VkNonDispatchableHandle.
>
> It adds definitions for dawn_native::vulkan::VkBuffer et al, which
> shadow the native ::VkBuffer et al. This retains type safety throughout
> the Vulkan backend without changing every single usage.
>
> Notably, the following things had to change:
> - An explicit conversion from VkBuffer* to ::VkBuffer* is needed for
> arrays. This is implemented as a reinterpret_cast, which is still
> safe as the new VkHandle still has the same memory layout properties
> as VkNonDispatchableHandle did.
> - When pointing to a VkHandle as an output pointer, it's now necessary
> to explicitly get the native ::VkBuffer (via operator*) and point to it.
>
> Bug: chromium:1046362
> Change-Id: I9c5691b6e295aca1b46d4e3d0203956e4d570285
> Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/15580
> Reviewed-by: Austin Eng <enga@chromium.org>
> Reviewed-by: Kai Ninomiya <kainino@chromium.org>
> Commit-Queue: Kai Ninomiya <kainino@chromium.org>
TBR=cwallez@chromium.org,kainino@chromium.org,enga@chromium.org
Change-Id: I500df2e34fd0f245ad04c517ff028ddd7bb5a2bf
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: chromium:1046362
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/15620
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
Commit-Queue: Kai Ninomiya <kainino@chromium.org>
Overriding VK_DEFINE_NON_DISPATCHABLE_HANDLE changes the function
signatures of Vulkan functions, changing their ABI and making us
incompatible with real drivers. This removes that magic, and replaces it
with an explicit wrapper, VkHandle, which has much of the same
functionality as the original VkNonDispatchableHandle.
It adds definitions for dawn_native::vulkan::VkBuffer et al, which
shadow the native ::VkBuffer et al. This retains type safety throughout
the Vulkan backend without changing every single usage.
Notably, the following things had to change:
- An explicit conversion from VkBuffer* to ::VkBuffer* is needed for
arrays. This is implemented as a reinterpret_cast, which is still
safe as the new VkHandle still has the same memory layout properties
as VkNonDispatchableHandle did.
- When pointing to a VkHandle as an output pointer, it's now necessary
to explicitly get the native ::VkBuffer (via operator*) and point to it.
Bug: chromium:1046362
Change-Id: I9c5691b6e295aca1b46d4e3d0203956e4d570285
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/15580
Reviewed-by: Austin Eng <enga@chromium.org>
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
Commit-Queue: Kai Ninomiya <kainino@chromium.org>
The state-tracking of the webgpu.h swapchain is a bit complicated
because contrary to implementation-based swapchains, they have more
guarantees and a "replacing mechanism". For example instead of hoping
the implementation-based swapchain resize automatically, the
surface-based swapchain needs to be replaced by a new swapchain and
invalidated.
This mechanism of invalidation also needs to be triggered when the last
reference to the surface is lost because we don't want to risk the
application destroying the window from under us.
Adds tests for all the cases of invalidation I could think of apart from
device loss.
Bug: dawn:269
Change-Id: Id515dbb640e13c6e30bb1f1e93b8e54f1e2bba4b
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/15400
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
If Device creation fails, several things are just partially initialized
and the destroy sequence crashes dereferencing null data.
This commit marks the Vulkan device as lost until after it is created.
This avoids parts of the destroy sequence which are unecessary since
the Device was never successfully created and no commands are in flight.
Bug: chromium:1043095
Change-Id: I8e121709fa19b215e118a615b639380d1db1f3f2
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/15460
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
