|
b3cbfd55
|
2021-07-07T12:31:57
|
|
Fix multiple end2end crashes in direct-to-MSL compiler.
This set of changes:
- Adds a single place (CompilerMtl::useDirectToMSLCompiler) where the
direct-to-MSL compiler can be enabled, still at compile time rather
than run time. It is still disabled by default.
- Initializes MTLWaitableCompileEventImpl::mShader, without which all
shader compiles crash.
- Fixes a mismatch of the ANGLERasterizerDisabled variable name in
the generated MSL.
- Removes the transform feedback bindings as an argument to the
vertex shader's main0(). Transform feedback support needs to be
reimplemented in this compiler backend.
- Added an option to the DriverUniform class to emit as either an
interface block or a struct. The direct-to-MSL backend assumes it
is a struct.
- Disable variable reference validation when referencing either
sample mask or rasterizer discard functionality in the shader.
- Disable struct usage validation when inserting references to
ANGLE_TextureEnv. This occurs during pipeline rewriting and it's
infeasible to disable it at that point, so disable it during
pipeline rewriting - which means it's disabled for all shaders.
- For angle_end2end_tests, disable the file API hooking which
disables the Metal shader cache. This speeds up the tests
significantly - by roughly a factor of 5.
With these changes, several hundred angle_end2end_tests run without
any assertion failures, though still with many test failures. The next
crash is in:
GLSLTest.NamelessScopedStructs/ES2_Metal
and will likely require a larger bug fix, or more changes to be
upstreamed from WebKit's repository.
Bug: angleproject:5505
Change-Id: Ia1ea9a13867f00035d4aeccf907fd032255588e3
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/3010486
Commit-Queue: Kenneth Russell <kbr@chromium.org>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Jonah Ryan-Davis <jonahr@google.com>
|
|
daeac238
|
2021-05-08T22:09:38
|
|
Translator: Ensure structs and blocks are uniquely defined
A new AST validation is added to ensure that the same TStructure or
TInterfaceBlock is not redundantly defined. This helps with SPIR-V
generation by allowing the id to be used as key in a hash map that looks
up the corresponding SPIR-V type id.
A bug is fixed where the Vulkan driver uniform declaration created two
identical declarations for ANGLEDepthRangeParams.
A number of other bugs are also fixed in this change, where if a
variable declaration is eliminated (for example due to constant folding,
or inactive interface variable removal) and it contained a struct
specifier, the struct declaration was also removed. OutputGLSLBase had
a hack where structs were declared on first encounter, which was
incorrect as the scope of the declaration could change. Those bugs are
fixed and this hack is removed.
Bug: angleproject:2733
Bug: angleproject:4889
Bug: angleproject:5936
Change-Id: I8e13748c0bf552ae8b052249282769a1f0775603
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2881942
Reviewed-by: Charlie Lao <cclao@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
776c6015
|
2021-02-26T00:00:57
|
|
Vulkan: Call glslang at compile time
With this change, the ANGLE translator immediately compiles the
generated GLSL into SPIR-V with glslang and discards the source. This
is in preparation for generating SPIR-V directly, by making the frontend
and backend already able to digest it.
This change also allows the expensive glslang calls to be parallelized,
improving the following perf test by about 20%:
LinkProgramBenchmark.Run/vulkan_compile_and_link_multi_thread
Previously, the test was run as such in the Vulkan backend:
Main Thread 1 Thread 2
Compile1 --->
Compile2 --------------------->
Translator Translator
<---
<---------------------
Link
glslang
for
shader1
glslang
for
shader2
Done
With this change, it is run as such:
Main Thread 1 Thread 2
Compile1 --->
Compile2 --------------------->
Translator Translator
glslang glslang
<---
<---------------------
Link
Done
glslang_wrapper_utils no longer interacts with glslang! A rename will
follow.
Bug: angleproject:4889
Change-Id: If4303e8ba0ba43b1a2f47f8c0a9133d0bee1a19a
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2721195
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
|
|
|
2f737c25
|
2021-02-25T16:55:14
|
|
Translator: Rename GLSL to SPIR-V in Vulkan and Metal output
Preparation for actual SPIR-V output instead of GLSL.
Bug: angleproject:4889
Change-Id: Ic279b23d3a817bd5dca66a844905378207afdbac
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2721194
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
|
|
4a782eaf
|
2021-02-11T16:13:33
|
|
Remove redundant (c) from license headers.
According to the Chromium C++ style guide, there must be no `(c)` after
`Copyright`.
Bug: angleproject:5516
Change-Id: Id5b170b99a1aaf95ab7fabb35ed644b99b756ce1
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2690948
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Cody Northrop <cnorthrop@google.com>
|
|
|
b1deba2f
|
2021-01-28T16:06:51
|
|
Vulkan: Don't emit spec const declaration in compute shaders
Bug: angleproject:5070
Change-Id: I74becfeac88b4a8ff66dd36670792082d0b941cd
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2658882
Reviewed-by: Charlie Lao <cclao@google.com>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
64508f44
|
2020-12-12T18:00:10
|
|
Vulkan: Reduce driver uniform data to minimum if specConst is used
If specialization constant is used, driver uniform data structure should
be reduced to minimum to increase cache locality.
Bug: b/175479076
Change-Id: I1fc50666542c6763c60bfe011cde5bc77ccc08e1
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2588549
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
Commit-Queue: Charlie Lao <cclao@google.com>
|
|
|
c75473c2
|
2020-11-23T14:26:32
|
|
Vulkan: Generalize FlipRotationSpecConst to SpecializationConstant
Specialization constant are used not just for flip/rotation. It also
used for other things. This CL merges all specialization constant usage
(lineRasterEmulation, flip, rotation, halfRenderArea) into one class and
rename FlipRotationSpecConst to SpecConst.
Bug: b/173800146
Change-Id: I8dc3354b6caedbb183cec29855fc1c301ec8872a
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2555812
Commit-Queue: Charlie Lao <cclao@google.com>
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
995493cc
|
2020-11-18T20:33:33
|
|
Revert "Vulkan: Remove rotation related data from driver uniform"
This reverts commit fd97c3343e9ff23677124b7528ec945bca65796a.
Reason for revert: The specialization constant for rotation causes
performance regression with older qualcomm driver. We need to keep the
driver uniform code path for the older driver.
Original change's description:
> Vulkan: Remove rotation related data from driver uniform
>
> Rotation is now handled in the shader compiler with specialization
> constant, it should be removed from driver uniforms. Since Metal is
> using the flipXY, flipXY/negFlipXY are still kept in the shader side
> implementation, but have moved to TranslatorMetal in this CL.
>
> Bug: b/171750979
> Change-Id: Ie8d15ef227cb52a6e19e4319ecc9f09bda42e667
> Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2519863
> Reviewed-by: Ian Elliott <ianelliott@google.com>
> Reviewed-by: Tim Van Patten <timvp@google.com>
> Reviewed-by: Jamie Madill <jmadill@chromium.org>
> Commit-Queue: Charlie Lao <cclao@google.com>
TBR=ianelliott@google.com,timvp@google.com,jmadill@chromium.org,cclao@google.com
Bug: b/171750979
Change-Id: Iff9cffb28851ade1d9c5cd23fde73910a19867ce
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2547808
Reviewed-by: Ian Elliott <ianelliott@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Charlie Lao <cclao@google.com>
Commit-Queue: Charlie Lao <cclao@google.com>
|
|
|
fd97c334
|
2020-11-12T11:05:40
|
|
Vulkan: Remove rotation related data from driver uniform
Rotation is now handled in the shader compiler with specialization
constant, it should be removed from driver uniforms. Since Metal is
using the flipXY, flipXY/negFlipXY are still kept in the shader side
implementation, but have moved to TranslatorMetal in this CL.
Bug: b/171750979
Change-Id: Ie8d15ef227cb52a6e19e4319ecc9f09bda42e667
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2519863
Reviewed-by: Ian Elliott <ianelliott@google.com>
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Charlie Lao <cclao@google.com>
|
|
|
3f18ad09
|
2020-11-11T14:12:18
|
|
Vulkan: Factor out DriverUniform code to tree_utils/DriverUniform.cpp
The driver uniform code is used by various tree_ops functions. But all
driver uniform code are defined and implemented in the
TranslatorVulkan.cpp file. There is dependency rule that tree_ops code
can not call into vulkan specific function. Right now we are working
around this problem by always having TranslatorVulkan creates uniform
and pass it down to tree_ops. This creates inefficiency for cases that
dFdy/dFdx where we don't know if we will need the driver uniform or not,
until we walk the tree and see the dFdx node. This CL refactors driver
uniform code into its own file and class under tree_utils so that
everybody can use. Mean time we can also make it much easier for metal
to expand it to have its own uniform structure.
Bug: b/173047182
Change-Id: I06bd9a005ccd6dc0a21321a3010dda1eab9d6fdb
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2533443
Commit-Queue: Charlie Lao <cclao@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
|
|
6136cbcb
|
2020-09-23T21:31:05
|
|
Metal: Implement transform feedback
- XFB is currently emulated by writing to storage buffers.
- Metal doesn't allow vertex shader to both write to storage buffers and
to stage output (i.e clip position). So if GL_RASTERIZER_DISCARD is
NOT enabled, the draw with XFB enabled will have 2 passes:
+ First pass: vertex shader writes to XFB buffers + not write to stage
output + disable rasterizer.
+ Second pass: vertex shader writes to stage output (i.e.
[[position]]) + enable rasterizer. If GL_RASTERIZER_DISCARD is
enabled, the second pass is omitted.
+ This effectively executes the same vertex shader twice. TODO:
possible improvement is writing vertex outputs to buffer in first
pass then re-use that buffer as input for second pass which has a
passthrough vertex shader.
- If GL_RASTERIZER_DISCARD is enabled, and XFB is enabled:
+ Only first pass above will be executed, and the render pass will use
an empty 1x1 texture attachment since rasterization is not needed.
- If GL_RASTERIZER_DISCARD is enabled, but XFB is NOT enabled:
+ we still enable Metal rasterizer.
+ but vertex shader must emulate the discard by writing gl_Position =
(-3, -3, -3, 1). This effectively moves the vertex out of clip
space's visible area.
+ This is because GLSL still allows vertex shader to write to stage
output when rasterizer is disabled. However, Metal doesn't allow
that. In Metal, if rasterizer is disabled, then vertex shader must
not write to stage output.
- See src/libANGLE/renderer/metal/doc/TransformFeedback.md for more
details.
Bug: angleproject:2634
Change-Id: I6c700e031052560326b7f660ee7597202d38e6aa
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2408594
Reviewed-by: Jonah Ryan-Davis <jonahr@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Jonah Ryan-Davis <jonahr@google.com>
|
|
|
9277ee74
|
2020-07-06T12:40:58
|
|
Metal: Implement MSAA default framebuffer.
GL_SAMPLE_COVERAGE_VALUE is implemented by inserting gl_SampleMask
writing logic to fragment shader.
New test added: MultisampleTest.ContentPresevedAfterInterruption.
- Skip on D3D11 (Bug: angleproject:4609)
Bug: angleproject:2634
Change-Id: Ib44daf0baccc36ea320596d81713156047da059c
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2281783
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Jonah Ryan-Davis <jonahr@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
|
|
|
4cef8a49
|
2019-11-26T00:04:00
|
|
Metal: Support depthRange near > far
Metal doesn't natively support near > far depthRange, need to
emulate it via shader.
Bug: angleproject:2634
Change-Id: I1885cad3467478a0dcc4b25b7d41f980b9e03103
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/1919282
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Jonah Ryan-Davis <jonahr@google.com>
|
|
|
a5a04ac0
|
2019-10-29T22:57:55
|
|
Metal backend pt3: shader translator
Implementation of GLSL to MSL translator
Bug: angleproject:2634
Change-Id: I66e2374b461548fac46163ea79790a488515e6b0
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/1887251
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
|
|
|
d200a77a
|
2019-10-10T00:44:01
|
|
Metal backend skeleton implementation.
Bug: angleproject:2634
Change-Id: I34be82f4a80a6851fecb53a51e069b134d82613a
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/1849079
Commit-Queue: Le Hoang Quyen <le.hoang.q@gmail.com>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
|