1) Append "Shadow" to samplers representing depth textures.
2) Sampling a depth texture returns f32, not vec4<f32>
3) Sampling a depth texture requires a Dref parameter, so we must
generate one if none is provided.
4) GLSL requires Dref to be appended to the texture coordinates vector,
*unless* it's a samplerCubeArrayShadow, since this would require vec5.
In that case, it's passed as a separate parameter.
5) GLSL's textureGather() with a depth sampler always requires a refZ
parameter, so provide zero to emulate WGSL's compare-less textureGather().
6) texelFetch() does not support depth textures, so this will have to be
validated out.
7) textureOffset() does not support sampler2DArrayShadow in GLES, so this will
have to be validated out.
Bug: tint:1298
Change-Id: Idaebe89cac6c1ec97c50a361b1d3aa3b84fb6c12
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/78760
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Stephen White <senorblanco@chromium.org>
Split IsAnyOf() into log(n) stages, where each stage performs a hashcode
check.
Previously there was a single hash test across the bitwise-or of all the
types being considered. If this passed, then each type would be tested
with Is<T>() individually. With this change, the list of types will be
recursively split into two, which each block hash-code checked. This is
repeated until we reach fewer than 4 types to check, where the test
decays to using Is<T>() for each type.
Also renamed `combined_hashcode` to `full_hashcode`, and used the term
CombinedHash for new helpers that bitwise-or the hashes from a number
of types.
Bug: tint:1383
Change-Id: Id056b9f7a9792430bd75ce554cb5fe73221ca4c7
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/78580
Reviewed-by: Antonio Maiorano <amaiorano@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Kokoro: Ben Clayton <bclayton@google.com>
When converting a for-loop to a loop, we were not cloning the for-loop's
body, but rather the statements within it. This worked fine, except if
we also hoisted a variable to a let within that body, which requires the
body to be cloned for the 'insert before' to work. This change clones
the for-loop body, which fixes the problem, but introduces a block in
the destination AST, which is ugly, but not incorrect.
Bug: tint:1300
Change-Id: I478244d87f8cf58837102004242ba1c835e21710
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/78821
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
It's too easy to copy diagnostics around and lose track of Source::File
ownership. Ideally we'd place the shared_ptr on the Source, but Sources
are copied _very_ frequently, and we'd lose a huge amount of
performance. Typically, Source::Files are owned externally. The only
time we really need to hold a shared_ptr to these is when a Source::File
is generated by an ICE, as the File points to the C++ source file that
raised the error.
Bug: chromium:1292829
Bug: tint:1383
Change-Id: I2706de8775bc3366115865b5a94785c0b2fefaae
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/78782
Reviewed-by: Antonio Maiorano <amaiorano@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Don't terminate on first error. Sleep a bit and try again.
Post a message to a gerrit change if it cannot be built. The fact the PS has a message from perfmon will prevent it from retrying the same change.
Remove trailing newlines from log.Printf() messages, they're automatically added.
Bug: tint:1383
Change-Id: I78a627c53c492e7da33a74470d5a064e90a7a753
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/78783
Auto-Submit: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ryan Harrison <rharrison@chromium.org>
Commit-Queue: Ryan Harrison <rharrison@chromium.org>
The current directory and namespace collides quite spectacularly with
the google benchmark include directory and namespace.
This becomes very hard to work around when there's a .cc file in src/
that wants to include "benchmark/benchmark.h", as MSVC appears to
resolve this to the relative path, while GCC and Clang resolve to the
compiler specified include directory.
Bug: tint:1383
Change-Id: Icc8891718d1d8a1b55c2ac4b2bb1487e8d09e629
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/78740
Kokoro: Kokoro <noreply+kokoro@google.com>
Auto-Submit: Ben Clayton <bclayton@google.com>
Reviewed-by: Antonio Maiorano <amaiorano@google.com>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
More line() and less std::endl.
More automated indents and less manual spacing.
Put a single newline after every struct and function declaration.
Note that this does touch every test result, but only affects whitespace.
Change-Id: I7506b9029b79b91fb335911dba44369b36f09bbe
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/78300
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: Stephen White <senorblanco@chromium.org>
OpenGL clips against [-w, w], while WebGPU uses [0, w]. The fix is to
modify gl_Position.z on output to match GLSL semantics. (This is the
same code used by SPIRV-Cross under the fixup_clipspace option.)
Bug: tint:1401
Change-Id: I1b1511352eee11cd9b095cd809aa8e1263c6cf4c
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/78261
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Stephen White <senorblanco@chromium.org>
This change adds an override for Transform::ShouldRun() for many of the transforms that can trivially detect whether running would be a no-op or not. Most programs do not require all the transforms to be run, and by skipping those that are not needed, significant performance wins can be had.
This change also removes Transform::Requires() and Program::HasTransformApplied(). This makes little sense now that transforms can be skipped, and the usefulness of this information has been severely reduced since the introduction of transforms that need to be run more than once.
Instread, just document on the transform class what the expectations are.
Issue: tint:1383
Change-Id: I1a6f27cc4ba61ca1475a4ba912c465db619f76c7
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/77121
Reviewed-by: Antonio Maiorano <amaiorano@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Transforms are supposed to be immutable, operating on the DataMaps provided for input and output, so make the methods const.
Add a ShouldRun() method which the Manager can use to skip over transforms that do not need to be run.
Change-Id: I320ac964577e94ac988748d8aca85bd43ee8d3b5
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/77120
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Antonio Maiorano <amaiorano@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
This transform converts all separate texture/sampler references
in a program into combined texture/samplers. This is required for GLSL,
which does not support separate texture/samplers.
As input, the transform requires a map from the unique sampler/texture
pairs previously gathered by the Resolver to strings, which will be
used as the names of the newly-generated combined samplers. Note that
binding points are unused by GLSL, and so are set to (0, 0) with
collision detection disabled.
All function signatures containing textures or samplers are rewritten,
as well as function calls and texture intrinsic calls. For texture
intrinsic calls, a placeholder sampler is used to satisfy the subsequent
Resolver pass (GLSL texture intrinsics do not require a separate sampler,
but WGSL intrinsics do). The placeholder is also used if the shader
contains only texture references (e.g., textureLoad).
Bug: tint:1366
Change-Id: Iff8407d28fdc2a8adac5cb655707a08c8553c389
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/77080
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Stephen White <senorblanco@chromium.org>
Fold both PromoteInitializersToConstVar and VarForDynamicIndex into this
new transform, with a config to enable either type of transformation.
This is almost a no-op change, except that VarForDynamicIndex required
ForLoopToLoop, while this tranformation in PromoteInitializersToConstVar
converts for-loops only as needed, so it doesn't do so when the
expression is in the for-loop initializer.
This transform will be extended to handle ensuring order of execution of
expressions.
Bug: tint:1300
Change-Id: I4d00984346a2c92b2d8563b459898f8f737589fd
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/77140
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
Kindly donated by Brandon Jones and Austin Eng.
Tint has been used to convert all the deprecated attributes to the new style. In doing so, comments have been stripped. These are not massively important for the benchmarking.
Bindings have also been adjusted to be sequential and unique so that the MSL backend doesn't have to deal with binding remapping.
Existing benchmark files that used an underscore '_' have been renamed to use a dash '-' instead, to match the new files.
Change-Id: If5fb507b981f107ed570f6eedb55b232448f67aa
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/77443
Reviewed-by: Brandon Jones <bajones@google.com>
Reviewed-by: Austin Eng <enga@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
For the directories "/test/unittest/" and "/test/vk-gl-cts/", continue to prevent the emission of PASS expectations, but now:
* Generate SKIP expectations in these directories if `--generate-skip` is passed.
* Delete expectation files in these directories if the tests now pass and `--generate-expected` is passed.
There's no change in behaviour for other test directories.
Change-Id: Ibd9e84a51029715e7c3d9e22e76a65770b66f1c7
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/77442
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: James Price <jrprice@google.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
If a WGSL test contains no entry points, add an empty one.
In that case, do not pass its name to the generator, so we
generate code for all functions, so they aren't culled for
reachability.
Add new test results for formerly empty tests.
Bug: tint:1376
Change-Id: Ibf371b943fb273d44712dfcc9dc1b7bb4ab071db
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/76540
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: Stephen White <senorblanco@chromium.org>
