At http://go/skiareview/c/skia/+/400622/55/bench/SkSLBench.cpp#84,
during the initial DSL parser land, Nanobench was updated to construct
a DSLParser directly instead of calling SkSLCompiler::convertProgram.
This skipped over some important compiler setup--specifically, applying
optimization flags and resetting the inliner.
Unfortunately this means that our optimized benchmark results for the
past ~9 months are not trustworthy at all. The numbers are wildly off:
http://screen/YfZHDvhs9V9Ka6x
"Unoptimized" benchmarks seem to be correct in practice, due to
redundant flag checks. "Optimized" benchmarks, however, are reporting
much better values than they should. The benchmark would stop inlining
after a few iterations of the test, because we'd quickly hit our
internal inlining limit
(https://osscs.corp.google.com/skia/skia/+/main:src/sksl/SkSLInliner.cpp;l=70)
at which point the inliner would turn itself off.
To fix the root cause, Nanobench uses SkSLCompiler::convertProgram
again. Additionally, Compiler::optimize now defensively calls
inliner.reset() before inlining. A second reset shouldn't be necessary,
but it's harmless and free. (This just sets two int fields to zero.)
Change-Id: I690ead14137e4bf74d0f7aa32812bd5b3c21bef5
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/534560
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
This will allow perf to give us insights on SkVM code generation
performance, so that any regressions are automatically noticed and
trends over time can be observed.
This CL includes three separate benchmarks:
- skvm: Compile SkSL to SkVM with no optimization pass
- skvm_opt: Compile to SkVM and perform optimization/DCE
- skvm_jit: Compile, optimize/DCE, and generate native assembly via JIT
Change-Id: I54ad20d7adb9c5e66d33a2fb761303a269341ff2
Bug: skia:13259
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/533957
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
SkVM does not support nonconstant array indexing because this is an ES3
feature, but sksl_large was using the ES3-specific looping binary
colorizer for gradients. This prevented SkVM from compiling the
benchmark. sksl_large has been regenerated with
`nonconstantArrayIndexSupport` set to false in the GrShaderCaps.
Change-Id: If706d9544d1bc0d9c49c59c925a517f677262c85
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/533956
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Previously, we used the Imgui paint shader. However, it uses a custom
mesh now. This increases its starting size substantially. The bitmap
text shader is a better example of a compact but still functionally-
useful shader.
(I also checked the medium and large shaders; these have not changed at
all since our last update, since Ganesh rarely changes now.)
Change-Id: I8e41aeaf4241d64910bd8188155ba428a6b0dbaa
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/533879
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Graphite code snippets can be moved into these files; they will be
compiled once at build time and dehydrated for later use. This allows us
to avoid synthesizing and compiling them in each shader where they are
referenced. (Unfortunately, the GPU driver will still need to compile
them each time.)
Change-Id: I5cdc5881d71d7b81a02c91a84d52804f2909b483
Bug: skia:13110
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/532259
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Change-Id: I238d29ba0250224fa593845ae65192653f58faff
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/528156
Reviewed-by: Kevin Lubick <kjlubick@google.com>
Reviewed-by: Jim Van Verth <jvanverth@google.com>
Commit-Queue: Greg Daniel <egdaniel@google.com>
The previous CLs have removed the last significant differences between
SkSL::String and std::string. This CL removes SkSL::String entirely and
replaces it with std::string throughout the code.
Apologies for the very long CL, but I have done my best to make it as
simple and reviewable as possible. The vast majority of changes are
simple replacement of `SkSL::String` with `std::string`. In the rare
spots where code is moved from one place to another, it is logically
unchanged.
Change-Id: I39563d2db45da229f17f4504dfd63e00bde7a96e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/503339
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Things have evolved since these were last updated. In particular, we no
longer emit binding and set decorations for each individual uniform.
Backends that rely on that (SPIRV, Metal), generate SkSL that has all
the uniforms grouped in a buffer. GLSL emits the uniforms with no layout
qualifiers. SPIRV supported this already (and automatically applied the
default binding/set to the implicitly created uniform buffer). Metal
*almost* supported this - it handled the uniform buffer correctly, but
didn't apply the default binding to any un-decorated samplers.
Bug: skia:12670
Change-Id: I7b193b89f86441344cf7ad5b7ed5e07451a1cc6d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/475157
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Bug: skia:12559
Change-Id: I91e0b46702ca361737459d64d1cdb57f8dd2f64e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/461697
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
This completes the long process of moving all of IRGenerator's code and
data into better homes and finally kills the class altogether. There is
a lot of #include churn here due to poor header hygiene in the past;
IRGenerator.h included a ton of stuff and almost everything included
IRGenerator, allowing us to inadvertently be sloppy with our includes.
Change-Id: I70d854e57dec7bd9a311b72f9f72d978d354da98
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/459936
Reviewed-by: John Stiles <johnstiles@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Now that DSLParser is used for everything we can get rid of this
unused code.
Change-Id: I7b586a7d7a51ac9df7d2c52baa1f525233f489b9
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/443076
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This will allow us to use AutoSymbolTable in more places (and helps cut
down IRGenerator, as we transition away from it).
Change-Id: Ie88c149e0f7eee080a8a4c8cbbc34c8a2a5401d1
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/449756
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This eliminates the SkSL ErrorReporter class and funnels everything
through the DSL ErrorHandler. Since the DSL error handler can be
changed, this required a number of updates to ensure that things work
properly in the face of custom error handlers. There is probably more
work to be done in that area, but this at least passes all existing
tests.
Change-Id: Iaee27b79fc4ed426c484ccab257c09d28619ead5
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/438116
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
The DSLParser is an alternative to the existing SkSLParser which goes
directly to IR code using the SkSL DSL. It is substantially faster and
simpler than the existing parser->IRGenerator pipeline, but not yet
feature complete nor fully tested.
Change-Id: Iee45e9b527a3b88faa2ea74fc512051c8a38c5d4
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/400622
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Like SkColorFilter, SkShader, etc., this has a public-facing component
(SkBlender) and a private subclass (SkBlenderBase) which can be
obtained via a helper function (as_BB). At present there are no public-
facing methods, but the type needs to be exposed to be usable by the
outside world.
These classes exist for SkRuntimeEffect to subclass. The blender base
provides a `program` method with the parameters that blending will use.
Change-Id: I75c772fd4108a9c21fbda84201a8b23d3750a0df
Bug: skia:12080
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/416916
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Change-Id: I17c624e9145d7152f65695ca3ce4592a11bb09a2
Bug: skia:12080
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/418637
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Change-Id: I9f4378b622699d80bef449f9857117e9cc437956
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/418140
Reviewed-by: Brian Osman <brianosman@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Runtime Blend effects always take two input colors--source and
destination--instead of one. This CL adds a new ProgramKind for blend
effects, a new program module (empty for now), and adds a test to
confirm that the signature for blend functions is checked. Currently
these are only accessible via skslc; there's no Runtime Effect API to
create one and the dest color isn't hooked up to anything.
Change-Id: I5272a811d2d76b878cfdf3429efa78c9c8b3fd97
Bug: skia:12080
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/416798
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
These enforce stricter rules about the signature of main, and each one
uses a separate pre-include module. That prevents color filters from
being able to reference sk_FragCoord (or coords passed to main) at all.
It also limits the versions of sample() that are exposed.
In the new world, an effect created for a specific stage of the Skia
pipeline can only be used to create instances of that stage (SkShader or
SkColorFilter). For now, SkRuntimeEffect::Make uses kRuntimeEffect,
which continues to be more lenient and allow creation of either shaders
or color filters from a single effect. After we migrate all clients, we
can deprecate and then delete that mode.
Bug: skia:11813
Change-Id: I0afd79a72beeec84da42c86146e8fcd8d0e4c09f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/395716
Reviewed-by: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Our benchmarks had drifted away from actual codegen in several ways, and
were no longer representative of actual output. In particular, every
function in the existing benchmarks used a scratch `_output` variable
which would make it ineligible for return-statement-only inlining.
Updated the tests to match modern codegen. `small` and `large` use the
same sources as before. The code from `medium` is older and had diverged
quite far from today's output; I could not find a representative code
snippet at all. It has been replaced with SkSL of roughly similar
complexity.
Change-Id: Iadd32cbb6725554355ddfe384ce538f3dc49aca3
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/385160
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
This change will allow these types to be forward-declared; C++ doesn't
allow forward declaration of types declared inside a struct. Moving
these types out of Programs resulted in a large diff.
The Settings::Value helper class has been moved inside of the
IRGenerator. In practice, it was actually just an implementation detail
of how IRGenerator looks up caps-values by name. It seems very unlikely
that this will be necessary elsewhere going forward.
Change-Id: I6119417fae608f1c492a27de746d2b550ef8ca20
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/370836
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This maps to usage better, and makes some code simpler to understand.
Note that there is still a PipelineStage *back-end*, which is specific
to the runtime-effect FP. A kRuntimeEffect_Kind program can be used to
generate a PipelineStage (for the GPU backend), or an skvm program (for
the CPU backend).
Change-Id: Id3f535db93a239726c595225aafe9467f0d19817
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/344969
Reviewed-by: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
We were always pre-loading the fragment and vertex modules, but deferred
loading all others. Those two take up about 300 KB of heap. Now, all
modules are deferred, so compiler instances that don't need them (like
the one used for runtime effects) are much smaller.
Now that we can get better fine-grained numbers, added two more
benchmarks, to track actual baseline usage, plus the usage in the two
most likely configurations.
Change-Id: Idfbcd52c8afee566ac42ab827c80c940f91c4ad7
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/337176
Commit-Queue: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
Previously, any builtin functions would be optimized as a side-effect of
optimizing programs that used them. Now that shared elements aren't
being optimized in that way, we explicitly optimize any shared modules
when they are first created. We don't remove dead elements, but we
we do substitute settings, simplify, and inline.
Bug: skia:10905
Change-Id: I701b5e9f52fb880ef3e6f4c67694d08602f47e95
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/336440
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
This reverts commit 1277971939.
Change-Id: I7985ef22ddd19adcab468acc684b330ce6978c8d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/332738
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
This reverts commit a2d6b31f66.
Reason for revert: breaking bots
Original change's description:
> Additional SkSL benches
>
> This adds additional benchmarks to give us better insight into how long
> the various compilation phases take. The four benches per size now
> cover just parsing, parsing + converting to IR, parsing + converting to
> IR + optimizing, and finally parsing + converting to IR + optimizing +
> generating GLSL.
>
> Change-Id: I7099a5a9f40ae5031e330dc4e1bb08c2a20ada63
> Bug: skia:10805
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/332262
> Reviewed-by: Brian Osman <brianosman@google.com>
> Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
TBR=brianosman@google.com,ethannicholas@google.com,johnstiles@google.com
Change-Id: Idb3e3082d11f72978131068b2229ce2578d645c3
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: skia:10805
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/332601
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
This adds additional benchmarks to give us better insight into how long
the various compilation phases take. The four benches per size now
cover just parsing, parsing + converting to IR, parsing + converting to
IR + optimizing, and finally parsing + converting to IR + optimizing +
generating GLSL.
Change-Id: I7099a5a9f40ae5031e330dc4e1bb08c2a20ada63
Bug: skia:10805
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/332262
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
This ties the caps to the compiler instance, paving the way for
pre-optimizing the shared code. Most of the time, the compiler is
created and owned the GPU instance, so this is fine. For runtime
effects, we now use the shared (device-agnostic) compiler instance
for the first compile, even on GPU. It's configured with caps that
apply no workarounds. We pass the user's SkSL to the backend as
cleanly as possible, and then apply any workarounds once it's part
of the full program.
Bug: skia:10905
Bug: skia:10868
Change-Id: Ifcf8d7ebda5d43ad8e180f06700a261811da83de
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/331493
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Salomon <bsalomon@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
This is the standard no-op shader that Viewer uses to blit the software
rasterizer's back-buffer to the screen. It modulates against white and
samples a texture using an identity matrix.
This gives us a realistic best-case-scenario test. `sksl_tiny` is too
small to be a realistic test (although it's a good data point to show
our ultimate speed-of-light).
Change-Id: Ic697cb1301752574ab63a9a0d7b07a0ff81c3a88
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/329476
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
The "sksl_huge" test has been renamed to "sksl_medium."
Added a new "sksl_large" test which contains a dithered gradient shader.
Change-Id: Ia8488460022008ae4efecb5593f01db06c1c283f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/320827
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This is basically timing the rehydrator, but is helpful in conjunction
with heap usage if a client wants to decide to throw away a compiler
that they don't anticipate using for some amount of time.
Change-Id: Ie9b27d9a3b7667b5492d870f242cdafd00784708
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/320617
Commit-Queue: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: Brian Osman <brianosman@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
separate step.
This ended up uncovering an optimization bug. SkSLNonConstantCase
started failing; it turns out that the optimizer was never being run on
this test, and so we hadn't noticed that it didn't actually work in the
presence of optimization.
Change-Id: Iff1d330be7534113b86f86b00c39f91282903ae3
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/316568
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Mechanically updated via Xcode "Replace Regular Expression":
typedef (.*) INHERITED;
-->
using INHERITED = $1;
The ClangTidy approach generated an even larger CL which would have
required a significant amount of hand-tweaking to be usable.
Change-Id: I671dc9d9efdf6d60151325c8d4d13fad7e10a15b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/314999
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Mike Klein <mtklein@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Early results from measuring RSS (via /proc or similar) shows results
that aren't very accurate or stable. From previous detailed testing, I
know that mallinfo gives a good answer. We only really need results
from ~one machine, so limiting to UNIX bots isn't a big deal.
Change-Id: I73af043720b1204e1da436e2e63b19766a97a9a2
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/301445
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
This will let us track a new metric that measures the heap usage of
constructing a default compiler object. I anticipate adding similar
stats about the heap usage from compiling simple vs. complex SkSL.
Change-Id: Idb814c0b5d210d00a06ce5dc7147437aabcba1bb
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/301359
Reviewed-by: Mike Klein <mtklein@google.com>
Reviewed-by: Joe Gregorio <jcgregorio@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Bug: skia:
Change-Id: I2ae0caf08f8434302cae8151ae1ea0fda8d56928
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/230397
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Bug: skia:
Change-Id: I83e3a094d26085fc4d586e5d2581e0d61c55634e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/145080
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Bug: skia:
Change-Id: I8497ddbc5c5fef4e0fa87f3b9828034ba34284cc
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/217636
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
