Change-Id: Idb526ccbb6d552b418c569a63bd3eed83e975ab1
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540916
Reviewed-by: Kevin Lubick <kjlubick@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Since this is super performance-sensitive, I manually checked to ensure
that the generated code didn't change. Other than differences in
variable name mangling, our final output looks to be identical.
Change-Id: I1bdf012f9b15bafcd85a9e73abfb3fa444459c60
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540300
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This moves the accumulation of worst case tolerances into
VertexChunkPatchAllocator for Ganesh, and lets the tessellating
RenderSteps in Graphite calculate the exact index count for each
instance (for which DrawWriter::DynamicInstances then tracks the worst
case index count between flushes).
Bug: skia:13056, skia:13012
Change-Id: I1d7b99d3edfc5c551cbe67734405ec43fba0e2db
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539049
Reviewed-by: Robert Phillips <robertphillips@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
* Add this->'s
* Make the logic in checkForDeletedEntries more robust.
* Harden readStrikeData by checking the typefaceID before doing
the descriptor typefaceID translation.
* Rename fRemoteFontIdToTypeface to fRemoteTypefaceIdToTypeface
* Simplify some SK_DEBUG code.
Change-Id: I8498f9be7ef4874f28f595222e661b9f8330b94f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540317
Reviewed-by: Ben Wagner <bungeman@google.com>
Commit-Queue: Herb Derby <herb@google.com>
Ganesh didn't adopt DSL for its fragment processors, except for two.
Backing out those two will allow us to eliminate a lot of unused tech.
Change-Id: I705072fe589746d05e0ae54ef25cee4de368f9b5
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540318
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
With this change, make generate takes 1.8 seconds instead of
7.9 seconds.
We still have generate_force to run on everything.
Change-Id: I6d57031adbe38a7f25a59570baea89970eea024f
Bug: skia:12541
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540740
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Rather than having a monolithic third_party/BUILD.bazel, this moves
our Dawn rules to their own subdirectory and makes it callable via
@dawn instead of //third_party/dawn.
This will help with the G3 roll and make our rules more organized in
general.
This also rolls Dawn
Roll Dawn from ab9757036bd6 to e831fb61046b (22 revisions)
https://dawn.googlesource.com/dawn.git/+log/ab9757036bd6..e831fb61046b
Suggested Review Order:
- WORKSPACE.bazel, where we define @dawn and its deps
(@vulkan_headers and @vulkan_tools). I initially thought
I needed to define all of Dawn's deps in the workspace_file_content
for new_local_repository, but that WORKSPACE file is
ignored when building Skia rules.
- third_party/dawn/BUILD.bazel, the contents of which were copied
from //third_party/BUILD.bazel and modified largely via
find-and-replace to point to files relative to
//third_party/externals/dawn. One exception is the cpu_wasm
config_setting because @dawn isn't able to see Skia's
//bazel/macros.bzl.
- All other files
Change-Id: Ib2d7bc972ef00b6b68370ce5c2839ffb70ed9a2f
Bug: skia:12541, skia:13211
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/538638
Reviewed-by: Ben Wagner <bungeman@google.com>
Commit-Queue: Kevin Lubick <kjlubick@google.com>
This was our last abuse of Constructor::Convert; remaining call sites
are FunctionCall::Convert and DSLType::Construct. Both of these callers
actually require the constructor-disambiguating smarts of
Constructor::Convert.
Change-Id: I9727a63abda03613cd524be3ae9740ada6735081
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540616
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
This file was originally added in http://review.skia.org/540040.
Change-Id: I11540a8a882300e12383fac7c8fa437e1e44301b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540619
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Kevin Lubick <kjlubick@google.com>
Commit-Queue: Kevin Lubick <kjlubick@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
I keep bumping into the need for this
Change-Id: I69384f7d590c163fd6244bdc64cc5e48450fecd4
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540171
Commit-Queue: Robert Phillips <robertphillips@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This doesn't improve the quality of final code, so we can skip this pass
during normal program compilation, but there's no excuse for leaving
Nops in a dehydrated module that we always load into memory at startup.
Change-Id: I7ff94f2713cff658799c469a46b526544dfc318c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540040
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Arman Uguray <armansito@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
This describes the capabilities of a particular Skia rendering context
(GPU context, or the CPU backend). At the moment, it only contains the
supported SkSL version (with a new enum added to specify the current
value as "100" and a new ES3 value as "300".
SkCapabilities can not be retrieved from an SkCanvas - the client must
have a concrete way of knowing what their destination device that will
do the actual rendering is (GrCaps or SkSurface).
This CL doesn't make use of the SkCapabilities yet, that's coming in
follow-up CLs that alter the SkSL compiler and SkRuntimeEffect API.
Bug: skia:11209
Change-Id: I4e9fd21ff7ffd79f1926c5c2eb34e10b3af4bc9b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/537876
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Brian Salomon <bsalomon@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
This makes expressions with `initialPreLocal` eligible for constant
folding. In practice, this doesn't do much today because matrix constant
folding doesn't support no-op-arithmetic elimination, but in a followup,
we could easily simplify `initialPreLocal * foo` --> `foo` when it's the
identity matrix.
Change-Id: I4c6faf2aa550b6b9776dd8a5da9ea5ec06cd8330
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540316
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Robert Phillips <robertphillips@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
Moving it into the helper class serves a few purposes:
1. Consolidate the more complicated stroke vertex count calculations so
that they can be reused between graphite and ganesh more easily.
2. Gives us an object that the PatchWriter can pass to its
PatchAllocators that is itself not a template, so the allocators can
be easily reused across different PatchWriter configurations.
3. Gives us a good place to start experimenting with tighter tolerance
tracking (particularly around internal rotation and join rotation).
Bug: skia:13056, skia:13012
Change-Id: I6086d459a21ac7ab8e833a988cc7c403983c3dd2
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/537083
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Robert Phillips <robertphillips@google.com>
This effectively rolls back http://review.skia.org/538045.
Change-Id: I9c457a6c9c9ba98d4c2e3a6af03c5dd80c76beb9
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540082
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
At present this has no effect. There isn't any dead path in the built-in
code (as you would hope) and we don't do any sufficiently-interesting
transformations that might cause unreachable code to appear.
Change-Id: I6798ae21f8721cf9f6f61ecb1765c8d8df7a9d0d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540039
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Change-Id: Iafd9cf15ac3c42bda7fa6b6857a82dc2eae6d234
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540296
Reviewed-by: Heather Miller <hcm@google.com>
Commit-Queue: Eric Boren <borenet@google.com>
Commit-Queue: Heather Miller <hcm@google.com>
This should address the root cause of http://review.skia.org/538045.
We currently don't eliminate unreferenced names from the symbol table,
although that'd be free money if we wanted to do it.
Change-Id: I712e23e6a4f3cb48c2b5ae68f042e13031d5f632
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540038
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
In b/224677119, we found that turning on *all* of Skia tracing caused
regressions. We are separately turning it off, but we do find some value
in select tracepoints. Add back in the ones we care about. The remaining
tracepoints can still be turned on manually with the sysprop.
All of these tracepoints are already traced on other platforms (e.g.
Chrome, Flutter). Convert them to TRACE_EVENT0_ALWAYS, introduced in
I13fd77445e0d2a05e46b75629b37bab5f571b02e. (A better long term solution
will be to use categories, as tracked in b/232405757.)
For GrOps, we had an Android-specific tracepoint, added in
I18ac03676da058ba4af5bd6a0c375b3f17c3c399, as well as separate
TRACE_EVENT0 tracepoints with basically the same timing and label.
Remove the Android-specific version, and make the other one
TRACE_EVENT0_ALWAYS. This removes redundant child-parent pairs.
Bug: b/231627558
Bug: b/224677119
Bug: b/232226246
Test: perfetto
Change-Id: I03ce62062a0bde723d3b6ac026bd978b249ad9a4
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/540156
Commit-Queue: Leon Scroggins <scroggo@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This is largely a re-implementation of the logic in
src/gpu/ganesh/gradients/GrGradientShader.cpp and
src/shaders/gradients/SkTwoPointConicalGradient.cpp. For simplicity,
most everything is in the shader now, but with an eye towards moving
computations that don't use sk_FragCoord into SkKeyHelpers.cpp.
Bug: skia:13302
Change-Id: I73d2e1f6c6ac0580aa897ffff3cfa87f4ed62b11
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539569
Commit-Queue: James Godfrey-Kittle <jamesgk@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
Constructor::Convert is a very heavy hammer which does error-checking
and lots of logic to identify the right constructor type to use. It cuts
out a lot of redundant work to call the desired method
`ConstructorCompound::Make` directly instead.
Change-Id: Icf9e513dbce223c2b27ccd0d11dc08eea849ff39
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539821
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Uses the initializer_list constructor to avoid generating a bunch of
insert calls, and then moves everything right into layout().
Change-Id: I09f5538e20ea6f7d3b6117020302a41e18b66295
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539822
Reviewed-by: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Colors are premultiplied, so the specular component (which is additive)
must be modulated by alpha explicitly.
Change-Id: If5f7b456b0ce518b1806e9b1e754229c3a6d0981
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539559
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Florin Malita <fmalita@google.com>
Reviewed-by: Jorge Betancourt <jmbetancourt@google.com>
Change-Id: Iac6b8a0b74fe23959f5ee814f4a30a5458ccf21e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539936
Reviewed-by: John Stiles <johnstiles@google.com>
Auto-Submit: Michael Ludwig <michaelludwig@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
We can access shader caps bits directly from SkSL, using sk_Caps. We
don't need to pass in caps flags from fGpu->caps()->shaderCaps().
Also, you might get different caps bits depending on which
backend/device you are using, so it's a little risky to bake it into
a shared cache; SkSL can apply the right caps bits at compile time.
Change-Id: I896b3915f40fbe7f58dfe5ffb51a462b3d2096ea
Bug: skia:13302
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539572
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Robert Phillips <robertphillips@google.com>
Reviewed-by: James Godfrey-Kittle <jamesgk@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
ATRACE_ANDROID_FRAMEWORK_ALWAYS and TRACE_EVENT0_ALWAYS create a
SkAndroidFrameworkTraceUtilAlways, which formats the input, requiring a
string copy. Mimic the split in SkAndroidFrameworkTraceUtil, which has
two constructors; one for formatting, and one for static strings. This
allows skipping the copy when it's unnecessary.
Make TRACE_EVENT0_ALWAYS call the cheaper constructor, since it never
has formatted input. ATRACE_ANDROID_FRAMEWORK_ALWAYS has few callers;
only one of them forces an unnecessary copy.
Bug: b/224677119
Test: perfetto
Change-Id: I8e699ae1496c94e08e6f7fce3616254b1a627a7f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539896
Commit-Queue: Leon Scroggins <scroggo@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
ConstructorCompound::Make() will now turn `float2(myFloat2)` into the
equivalent simpler form `myFloat2` by simply returning the inner
argument as-is. In practice, this codepath is difficult to reach because
`Constructor::Convert` will interpret the above constructor as a cast,
and ConstructorCompoundCast already had a similar optimization in place
(to eliminate same-type casts).
This CL seems large but it's mostly comments. (`is_safe_to_eliminate`
could easily be a single-line return statement if the goal were to
minimize LOC!)
Change-Id: If75d1583e5fe27f7fd903175ebbd9185c67026f4
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539820
Reviewed-by: Arman Uguray <armansito@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Previously, each Expression would clone itself at its existing position
and then the position would potentially be overwritten at the end. Now,
rewriting the position happens as part of the clone operation.
This seemed simpler than the previous two-step operation, and also opens
up the possibility of passing the updated position to the clone's
children (but I didn't experiment with that here.)
Change-Id: I07d0c0221fcab460d0db358d059ceecf8a0a7dc7
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539571
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
`optimize_constructor_swizzle` already does the same transformation, but
in a far more complex way. In a followup CL, I intend to limit
`optimize_constructor_swizzle` to only operating on ConstructorCompound
objects.
Change-Id: Ieb4210a9ea9efb1a021c4b2e15ba55863000ef93
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539819
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
We added a `clone(Position pos)` overload late; this CL goes back and
uses it in all the places that needed it.
Change-Id: I6924ac9a91ebbd148531dccd25218d6f497ed0d6
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539568
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
In cases where we know the type of constructor being made, it is much
more efficient to create it directly, instead of relying on
Constructor::Convert to deduce the type. This also helps us catch
legitimate bugs that Constructor::Convert will paper over; e.g. Convert
will happily accept mismatched types and just cast them (which was
occurring here, due to calling `scalarTypeForLiteral` in the wrong
spot).
Change-Id: I972046e14cc78708a3f1b8a4a31016d91e60a431
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539508
Reviewed-by: Brian Osman <brianosman@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
In follow up changes, the PatchAllocator duck type will get a little
more involved with tessellation-specific code, so I figured it'd be good
to buffer it with this new type between GrVertexChunkBuilder (which can
be used more generally if we update ops to do so).
Bug: skia:13012
Change-Id: Ib51d7beceba98cb10678dd04b431e923518ed6d8
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/539037
Reviewed-by: Robert Phillips <robertphillips@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
