|
bcc89703
|
2021-10-07T11:26:42
|
|
Translator: Remove disableValidateFunctionCall from transformation
This validation is disabled in three transformations and re-enabled
afterwards. Two of these benefit from the simplicity it brings, and
this change removes it from the other transformation.
Bug: angleproject:6506
Change-Id: I946b547ed0e2ed39062914a7771e9635ff5e406d
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/3212289
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
|
|
|
800e82c6
|
2021-08-23T11:05:23
|
|
Translator: Validate precisions
When declaring a variable, a struct field, function parameter etc,
there's a precision necessarily applied to the entity being declared.
AST Validation is added to enforce this. Intermediate nodes derive
their precision from these entities automatically.
Consistency of intermediate nodes is not validated. This is because AST
transformations replace a node with a transformed one, and that may not
have the same precision. Take the following code:
mediump float x = ...;
mediump float y = ...;
... x + y ...
and assume is transformed as such:
highp float driver_uniform;
... (x * driver_uniform) + y ...
The addition was originally done in mediump, but would seemingly need to
be done in highp after transformation. There are a number of options
here:
- Make sure that when nodes are replaced, the precision is unaffected.
This can be intrusive, requiring temp variables.
- Bubble up the new precision
- Accept the discrepancy
ANGLE opts for the last option, which actually respects the original
shader's intended precision for operations, even if some transformation
needs to temporarily evaluate an expression at a higher precision.
Bug: angleproject:4889
Bug: angleproject:6132
Change-Id: Ibcde3a230de159157783b1c6d5ef1cd63ceb4d8f
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/3114027
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
210773db
|
2021-08-05T10:41:59
|
|
Translator: Be more explicit about precisions
GLSL ES requires that every symbol (variable, block member, function
parameter and return value) is appropriately qualified with a precision,
either individually or through the global precision specifier.
Some tree transformations however produced symbols with EbpUndefined
precision. In text GLSL output, these would produce unqualified symbols
which was often incorrect.
In this change, the transformations are made to produce explicit / more
consistent precisions. The validation (that caught these issues) is not
included in this change as there are still a few corner cases left to
address.
Bug: angleproject:4889
Bug: angleproject:6132
Change-Id: Icca8a0a5476f8646226e7243aa8f501f44acc164
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/3075127
Reviewed-by: Tim Van Patten <timvp@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
aec5e65c
|
2021-07-28T00:36:12
|
|
Get direct-to-Metal backend to run angle_end2end_tests.
Cherry-pick nameless struct fix from Apple in
https://bugs.webkit.org/show_bug.cgi?id=227482 .
Fix SeparateCompoundStructDeclarations pass to stop generating
multiple declarations; thanks syoussefi@ for advice.
Incorporate additional passes from TranslatorVulkan
(MonomorphizeUnsupportedFunctionsInVulkanGLSL,
RewriteArrayOfArrayOfOpaqueUniforms,
SeparateStructFromUniformDeclarations) needed by RewriteStructSamplers
pass in TranslatorMetalDirect. Fixes many assertion failures in GLSL
tests. Moved these passes out of tree_ops/vulkan. Thanks again to
syoussefi@ for advice and help.
Disable a validation check related to the RewritePipelines pass. Skip
two tests that were failing for other reasons.
With these changes, angle_end2end_tests runs to completion when the
direct-to-Metal backend is turned on. There are still ~1300 failures
of the ~4000 tests which will be investigated next.
Bug: angleproject:5505
Change-Id: Ibca77822543e8e8e8d2a8c862e92cdf74bfa3545
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/3058524
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Commit-Queue: Kenneth Russell <kbr@chromium.org>
|
|
|
7c6da493
|
2021-07-26T21:24:25
|
|
Translator: Make sure built-in variables are consistent
Some transformations left the tree in an inconsistent state, for example
if it used gl_SampleMask from ES3.1+OES_sample_variables and the
transformation added a usage of gl_SampleMask from ES3.2. The offending
transformations are fixed and AST validation is improved to catch such
errors.
Bug: angleproject:4889
Change-Id: I9d9ea5bb43a9408dd4c6dc3e89ec20d60dfeff73
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/3054613
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
|
|
|
0e1c90cf
|
2021-07-21T10:24:26
|
|
Translator: Validate function parameter qualifiers
This change cleans up a confusion in EvqConst and EvqConstReadOnly,
where the former was frequently used instead of the latter for function
parameters.
The change makes the following renames to make the intent of the
relevant qualifiers clearer:
EvqIn -> EvqParamIn
EvqOut -> EvqParamOut
EvqInOut -> EvqParamInOut
EvqConstReadOnly -> EvqParamConst
Bug: angleproject:4889
Change-Id: Idedd32c08a91de069b91b1657d6b783dddece04a
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/3041622
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
a3df80ea
|
2021-07-19T17:15:13
|
|
Translator: Validate AST consistency w.r.t function calls
This ensures that tree transformations don't create invalid function
calls. Additionally, it's validated that no raw function calls are
issued, as that's deprecated.
Bug: angleproject:4889
Change-Id: Iabbb493dcc0a5654bbb5f930e402bf3492c89979
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/3040119
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Jonah Ryan-Davis <jonahr@google.com>
|
|
|
0461d8d4
|
2021-06-23T23:20:33
|
|
Vulkan: SPIR-V Gen: texture and image built-ins
GLSL contains a large number of built-in texture* and image* functions,
but these map to only a handful of SPIR-V instructions. The bulk of the
work to map these is to extract the arguments from the built-ins based
on their ordinal position.
Bug: angleproject:4889
Change-Id: I760d986bd9171ddde35f9f046c549ca53252df17
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2992980
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
8b869a95
|
2021-06-13T01:09:27
|
|
Translator: Generate Ops for all built-in functions
EOpCallBuiltInFunction is removed in this change, as well as the
"op": "auto" property in builtin_function_declarations.txt. Instead,
gen_builtin_symbols.py automatically generates Ops for every built-in
function and generates the TOperator enum accordingly.
This simplifies SPIR-V code generation by allowing switches to be used
on operators instead of string comparisons.
Bug: angleproject:4589
Bug: angleproject:4889
Change-Id: Ia351524400b0e12a10a5572e27e9b88c6ec2e61c
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2958869
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Jonah Ryan-Davis <jonahr@google.com>
Reviewed-by: Tim Van Patten <timvp@google.com>
|
|
d7aa0130
|
2021-04-26T16:56:15
|
|
Upstream Apple's direct-to-Metal backend: compile translator.
This change is meant to merge the translator changes from Apple's
direct-to-Metal backend. Taken from Kyle Piddington's CL:
https://chromium-review.googlesource.com/c/angle/angle/+/2857366/
The goal of this CL is to merge the translator code in a state that
compiles, but not to switch the Metal backend over to use this
translator backend yet.
Bug: angleproject:5505
Change-Id: I68a6354604498cd5fd1eb96c13fc56f3b38f2bd0
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2897536
Reviewed-by: Jonah Ryan-Davis <jonahr@google.com>
Commit-Queue: Jonah Ryan-Davis <jonahr@google.com>
|
|
|
382bf288
|
2020-12-24T23:56:39
|
|
Organize AST transforms per backend
Most of the AST transforms are written as a workaround to an issue that
affects a single backend. This change identifies such transforms and
organizes them by backend. They are then only built if the respective
backend is.
Additionally, about half of the GL transforms are due to mac
workarounds, including the large RewriteRowMajorMatrices transform.
Mac-specific workarounds are additionally only built on said platform.
This change reduces the ANGLE binary size:
- 106KB in a Vulkan-only build on Linux
- 27KB in a GL-only build on Android (60KB on Linux)
Bug: chromium:1084580
Bug: chromium:1161513
Change-Id: I64b334332c0d4f848756c6538af0d8d96864c7e9
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/2601346
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Reviewed-by: Tim Van Patten <timvp@google.com>
|
|
155947fc
|
2019-10-24T12:55:11
|
|
Enable "-Wconditional-uninitialized".
This is a final warning used by Skia.
Bug: angleproject:4046
Change-Id: I3970e30e4bd2aef07cddadd7322ef120ac857493
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/1877481
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Jamie Madill <jmadill@chromium.org>
|
|
83a369bb
|
2019-08-14T10:39:34
|
|
Vulkan: Improve cubemap emulation seam handling
Changes seamful cubemap emulation to always compute the derivative,
emulating the bias parameter by scaling the provided derivatives.
This results in more accurate mipmap levels for seams within primitives.
There are some artifacts as a result of how derivatives are calculated,
but this matches the native driver.
Bug: angleproject:3243
Bug: angleproject:3732
Change-Id: Icb976e2a7e14cb4210645571edc037d4e607bd0d
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/1754383
Commit-Queue: James Dong <dongja@google.com>
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
472c74c6
|
2019-08-19T16:32:13
|
|
Translator: Allow tree validation in children of TCompiler
This is to be able to perform validation inside TranslatorVulkan, even
if it's through ASSERTs.
Additionally, every transformation is changed such that they do their
validation themselves. TIntermTraverser::updateTree() performs the
validation, which indirectly validates many of three tree
transformations. Some of the more ancient transformations that don't
use this function directly call TCompiler::validateAST.
Bug: angleproject:2733
Change-Id: Ie4af029d34e053c5ad1dc8c2c2568eecd625d344
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/1761149
Reviewed-by: Geoff Lang <geofflang@chromium.org>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
5a2553a7
|
2019-08-07T14:44:12
|
|
Vulkan: Emulate subgroup ops in seamful cubemap emulation
Where subgroup ops are not available, they are emulated as such:
Code with subgroup ops:
float lH = subgroupQuadSwapHorizontal(layer);
float lV = subgroupQuadSwapVertical(layer);
float lD = subgroupQuadSwapDiagonal(layer);
bool isHelperH = subgroupQuadSwapHorizontal(gl_HelperInvocation);
bool isHelperV = subgroupQuadSwapVertical(gl_HelperInvocation);
if (gl_HelperInvocation)
{
layer = !isHelperH ? lH : !isHelperV ? lV : lD;
}
Emulated code:
float nonHelperLayer = gl_HelperInvocation ? 0.0 : layer;
float lH = abs(dFdxFine(nonHelperLayer));
float lV = abs(dFdyFine(nonHelperLayer));
float lD = abs(dFdxFine(lV));
float isHelperDiffH = abs(dFdxFine(float(gl_HelperInvocation)));
bool isNonHelperH = isHelperDiffH > 0.5;
float isHelperDiffV = abs(dFdyFine(float(gl_HelperInvocation)));
bool isNonHelperV = isHelperDiffV > 0.5;
if (gl_HelperInvocation)
{
layer = isNonHelperH ? lH : isNonHelperV ? lV : lD;
}
Both paths are supported as on nvidia devices the emulated code
misbehaves. This change therefore effectively only enables seamful cube
map emulation on Android where subgroup operations are not supported.
Bug: angleproject:3243
Bug: angleproject:3732
Change-Id: I9664d9760756758748183eb121c626f176789f3a
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/1742222
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
e86a8560
|
2019-08-07T11:55:01
|
|
Vulkan: Fewer subgroup ops in seamful cubemap emulation
Two values were retrieved from quad neighbors; layer and ma. The value
of ma was retrieved as the helper invocations would otherwise have a
different value as the major axis.
This change makes the helpers rechoose ma based on the layer, which
removes a number of subgroup operations. This is also more precise as
the major axis value could be slightly different from the neighbor.
Bug: angleproject:3732
Change-Id: I3c8ca724e91c52ca2f7edc03bb0e5dca67610ff4
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/1742215
Reviewed-by: Tobin Ehlis <tobine@google.com>
Reviewed-by: Jamie Madill <jmadill@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|
|
|
93560ef5
|
2019-07-25T16:13:02
|
|
Vulkan: Seamful cube map emulation
In GLSL, a cube texture is sampled with one of textureCube* functions.
This function takes a 3D coordinate which is a vector from the center of
the cube and identifies a direction to sample from. GLES2.0 has the
following table that translates this 3D coordinate (Rx, Ry, Rz) to a
face and ST coordinates within that face. This table can be found in
Section 3.7.5 (Cube Map Texture Selection).
A compiler pass is implemented in ANGLE that replaces samplerCube
declarations with a sampler2DArray. The textureCube* functions are
replaced with the corresponding texture* functions with the translated
coordinates according to that table.
Gradients provided to textureCubeGrad are translated using the same
formulae, which is not precise but the spec specifies this projection to
be implementation dependent.
Helper invocations enabled through WQM (whole quad mode) cause a
nuisance in that the extrapolated varyings used as coordinates in a
textureCube call could have a different major axis (and therefore face)
from the non-helper invocations that lie within the geometry.
subgroupQuadSwap* operations are used in conjunction with
gl_HelperInvocation to make sure the helper threads calculate texture
UVs in the same face as the non-helper invocations.
Bug: angleproject:3300
Bug: angleproject:3240
Bug: angleproject:3243
Bug: angleproject:3732
Change-Id: I0cb6a9b1f2e1e6a392b5baca1c7118ed1c502ccf
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/1715977
Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org>
Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>
|