This CL refactor unittests in std140_test.cc, and add exhaustive
parameterized unittests in std140_exhaustive_test.cc. In std140_test.cc,
only test Std140 transform result for `mat2x2<f32>` for matrix used as
array element type and `mat3x2<f32>` otherwise, and keep the source and
expected programs as plain WGSL code to ensure the readablity. In
std140_exhaustive_test.cc, all matrix shape and different constant index
are tested using parameterized WGSL code, at the cost of readablity.
This CL prepares for supporting f16 in Std140 transform by allowing
testing all shape of f16 matrix as well by simply adding parameters.
Bug: tint:1473, tint:1502
Change-Id: Ib2ef5bd806ee61eab04d73a415ba62c2191e2a7e
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104282
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Zhaoming Jiang <zhaoming.jiang@intel.com>
This combination uses a suspiciously high amount of memory, triggering the fuzzer limits of 2560MB.
Tint doesn't really have any OS specific code paths, so we should still have good coverage with ASAN builds for other OSes.
Fixed: chromium:1357188
Change-Id: I4c7001f7e194ff46b2e8da635ddccdb04d60b838
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/105140
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Auto-Submit: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
There's a bug in some Qualcomm devices where using a depth/stencil
texture as a render attachment and then sampling it in a compute pass
causes a crash. This only happens, however, if the two passes occur as
part of the same Vulkan command buffer.
To work around the issue, this change splits the Vulkan command buffer
while recording any time it identifies that the problematic scenario may
occur.
Bug: dawn:1564
Change-Id: Ie137e9118ef9cc41f5908ca32c72c33f3798cd71
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104860
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Brandon Jones <bajones@chromium.org>
Reviewed-by: Kai Ninomiya <kainino@chromium.org>
Qualcomm GPUs are apparently decoding textures with a lower precision,
resulting in some of the rendered values when verifying the texture in
a test to be off by +-1 on any given channel. This change adds a
tolerance to those tests to allow a little wiggle room, since compressed
textures are inherently lossy anyway.
Allows Qualcomm GPUs to pass all compressed texture end2end tests.
Bug: dawn:1562
Change-Id: I08a21b9ce361486c247c34640080b369ae2b799d
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104622
Reviewed-by: Austin Eng <enga@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Brandon Jones <bajones@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
When converting floats to string, we now enforce the "C" locale, so
decimal points will be written as "." rather than the "," separator
used natively in some European locales.
Also, we now use operator>> to read back the number instead of
std::stof. std::stof works in the system locale, and will fail to
read back floats with the wrong decimal separator. (Also, std::stof
will throw if the number is out of range and can't fit in the
destination, which implies that the `if` check was probably never
failing.)
Skia encountered similar issues: see http://review.skia.org/587536
for the Skia implementation.
Change-Id: I5aded6acc7cfcf2ad4d5b974bc30c3b645eaec51
Bug: dawn:1686
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104680
Reviewed-by: Dan Sinclair <dsinclair@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
- Make the const eval builtin tests use the same framework as the
unary/binary op tests, allowing for Vector cases.
- No longer always use float compare, instead enable it per case.
Currently this is necessary because atan2 doesn't always return the
same constant for PI on all platforms.
- Add vector cases for atan2 and clamp.
Bug: tint:1581
Change-Id: I7eaec10b4f9685c913a9d0d17b47c413f659be7a
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104424
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
These functions don't need the device or external semaphore service
at all. Make them free functions so that a future change can allow
the handles to be closed after the device has destroyed its semaphore
service.
Bug: chromium:1359106
Change-Id: I246dd0a8f3f972c4547503d16bf8b00db14cdf58
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104542
Reviewed-by: Loko Kung <lokokung@google.com>
Commit-Queue: Austin Eng <enga@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
This moves handling the wait semaphores to the same place that the
signal semaphores are handled. It fixes a bug where the semaphores are
never waited on and never deleted if a texture is imported and then
exported without being used.
Bug: chromium:1359106
Change-Id: If226a38946d4a16598d78841e7b204ea91f8bbea
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104541
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Austin Eng <enga@chromium.org>
Reviewed-by: Loko Kung <lokokung@google.com>
The D3D12 fence share handle should be closed when the device is
deleted.
A future change will make it valid to call EndAccess after the device
is destroyed, thus the handle is closed in ~Device instead of
Device::DestroyImpl. It needs to live as long as ExternalImageDXGI
holds a reference onto the device.
Bug: chromium:1359106
Change-Id: Ib9c9aaa7fb0b5a3de035b512f8fc0316d4bd225e
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104540
Commit-Queue: Austin Eng <enga@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Sunny Sachanandani <sunnyps@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
Messages could be ignored if:
- They are not associated to any device (for example an issue around
instance or adapter operations)
- They happened between the last Tick() and device destruction.
Fix both cases to print the error to the dawn::ErrorLog and crash in
debug so that the errors are visible.
Bug: chromium:1258986
Change-Id: I9a88cd078c60b42deb2336da038902639f9a35ae
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104360
Reviewed-by: Loko Kung <lokokung@google.com>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
The SPIR-V tools roll to pull in the correct spelling of `preceded` has
landed. This CL re-enables the SPIRV-Reader tests with the correct
spelling.
Bug: tint:1406
Change-Id: I303b4b6d742f4bfcc76c6fcce66e4e1cef37b1af
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104464
Commit-Queue: Dan Sinclair <dsinclair@chromium.org>
Reviewed-by: David Neto <dneto@google.com>
This CL adds element count limits to arrays. In FXC there is a maximum
of 65536 elements in an array. This limit is not yet in WGSL, but adding
this here allows us to fix the issue with large arrays and GLSL.
Bug: chromium:1367602
Change-Id: I7df9d3e4f6c3e5107420d5f8e576d1f33e453161
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104240
Commit-Queue: Dan Sinclair <dsinclair@chromium.org>
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
When an imported texture with layout UNDEFINED was never used and then
exported with target layout UNDEFINED, Dawn would create a queue
transition barrier with dstLayout UNDEFINED which is not allowed by the
Vulkan specification. Instead detect this case and transition to
GENERAL.
Found by running dawn_end2end_tests with the VLL.
Bug: chromium:1258986
Change-Id: I5e36efda35cb27cecc0683846a314783a8a72fe6
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103025
Reviewed-by: Austin Eng <enga@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Loko Kung <lokokung@google.com>
std::stof can throw std::out_of_range if the input is not actually
representable. We had similar code in Skia which was using stof to
test that a stringized float would round-trip successfully, and it
would throw an exception on some older versions of libc++ for edge-
case inputs like FLT_MIN.
std::stof is documented as using strtof to do its conversion, so this
shouldn't change your results in practice; it just removes the part
where it could potentially throw for some inputs.
Tangentially, have you ever seen a case where the scientific-notation
path gets used? According to brucedawson@, nine digits should always
safely round-trip (in 2013, testing gcc and MSVC). See
https://randomascii.wordpress.com/2013/02/07/float-precision-revisited-nine-digit-float-portability/
Change-Id: Ie215fb8502dd8c554020c6f73432f91e3d756563
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104500
Reviewed-by: Dan Sinclair <dsinclair@chromium.org>
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Dan Sinclair <dsinclair@chromium.org>
Commit-Queue: John Stiles <johnstiles@google.com>
This is required to make importing images work on some systems. The
ideal version would be detecting whether dedicated allocations are
needed as Vulkan provides reflection for that. However this reflection
doesn't work on Nvidia, so instead Dawn requires a
NeedsDedicatedAllocation enum on import that's Yes/No/Detect so the
application can force use of a specific code path.
Support for this enum and toggling dedicated allocations on/off is added
for all external memory service implementations.
Vulkan image wrapping tests are modified to add test parameters so that
the Yes/No/Detect code paths are covered by tests.
This is technically post-V1 work, but gl_tests in Chromium fail on
Nvidia workstations without this fix, which makes it hard to debug other
issues.
Bug: dawn:1552, dawn:206, dawn:1260
Change-Id: Iee4f7bb9dbec520432ec623551221ef9e4d3d984
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103560
Reviewed-by: Austin Eng <enga@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Latest SwiftShader roll into Dawn failed because tests like
TriangleStripPrimitiveRestartTests.Uint32WithoutPrimitiveRestart relies
on robustness checks being enabled, but a recent change to SwiftShader
no longer enabled robustness by default. This change makes sure to
enable the robustness extension.
Change-Id: I7168fc440ef19ef6acac1d1ce72f4bf5a947d4dd
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104120
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Brandon Jones <bajones@chromium.org>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
The Android devices I've tested with Qualcomm GPUs (like the Pixel 4)
are exhibiting an issue where resolving timestamp queries after a
render pass is causing a crash. Until that issue can be resolved it's
safest to simply not advertise timestamp query support on these devices.
Bug: dawn:1559
Change-Id: Id76aa5095ffbb7f55579cc428388f55f4528581d
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104441
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Brandon Jones <bajones@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
