(This mirrors an optimization performed in the constant folder.)
Expressions like `OpIEqual %20 %20` or `OpFUnordNotEqual %15 %15` can be
replaced by `true` or `false` on sight. The GLSL spec makes it clear
that checking for NaN is optional:
4.7.1 Range and Precision
"... NaNs are not required to be generated. Support for signaling NaNs
is not required and exceptions are never raised. Operations and built-in
functions that operate on a NaN are not required to return a NaN as the
result."
Change-Id: I2e29b659a73582e9ade0eb61f70f7d362a007c50
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/531550
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>
We now have two functions `writeOpLoad` and `writeOpStore` which are
in charge of writing SpvOpLoad and SpvOpStore instructions.
`writeOpStore` also keeps track of pointer stores in a "store cache."
Subsequent loads from that same pointer will be found in the cache and
will return the value stored in that pointer instead.
Such a cache definitely cannot work in the face of control flow, so we
make the following concessions:
- `pruneReachableOps` is now `pruneConditionalOps`. Any pointers that
are altered inside a potentially-unreachable block are cleared from
the cache entirely.
- The entire store cache is cleared at all OpLabels within a loop.
The cache also cannot work in the presence of swizzled stores, so we
make another significant concession:
- The entire store cache is cleared whenever we store into a non-memory
pointer (e.g., assigning into a swizzled LValue, such as `foo.xz`).
Despite these significant limitations, this manages to dramatically
shrink many real-world examples.
Change-Id: I0981a0cf7b45b064e153e9ada271494c8e00cad5
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/530054
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Previously, we stringized the types and put them into fTypeMap. Using
the op cache is a simpler mechanism that should work equally well.
Output diffs are almost all ID reorderings. In a few cases we
managed to deduplicate function types that stringize differently but
come out the same in SPIR-V (e.g. no float/half distinction).
Change-Id: If7de5b2dafa12d05c3c2c497a243e9e3908dfee7
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/529805
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Previously, we didn't usually generate OpConstantComposite ops for
matrices. Now, a matrix assembled from constants should come out as a
constant.
Change-Id: I458718901686dffb84e4079a81017d61195420d3
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/529338
Reviewed-by: Arman Uguray <armansito@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
The output changes here are almost entirely a wash, because we already
had support for caching scalars and vectors. Almost all changes are just
inconsequential reorderings of IDs, and the removal of RelaxedPrecision
decorators on constants (which were not meaningful).
Change-Id: I45340c4a240cb504b7c4a934b3db178d2f39ec99
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/528709
Reviewed-by: Arman Uguray <armansito@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Some GPUs (Adrenos in particular) perform noticeably better when we
use OpConstantComposite instead of OpCompositeConstruct. This also gives
us some deduplication of redundant ops.
Change-Id: I53b7a3e1cf61e51647a661a08ff4c7b53ee60f10
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/528636
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Compiling a program with "allow narrowing conversions" actually fixes up
narrowing casts in the program by inserting casts wherever they would be
needed for type-correctness. For instance, compiling the statement
`half h = myFloat;`
inserts an appropriate narrowing cast:
`half h = half(myFloat);`.
The Pipeline stage code generator relies on this behavior, as when it
re-emits a runtime effect into a complete SkSL program, the narrowing-
conversions flag will no longer be set, but that is okay, because the
emitted code now contains typecasts anywhere they would be necessary.
Logically, this implies that anything which supports narrowing
conversions must be castable between high and low precision. In GLSL and
SPIR-V, such a cast is trivial, because the types are the same and the
precision qualifiers are treated as individual hints on each variable.
In Metal, we dodge the issue by only emitting full-precision types. But
we also need to emit raw SkSL from an SkSL program (that is what the
Pipeline stage generator does).
SkSL already supported every typical cast, but GLSL lacked any syntax
for casting an array to a different type. This meant SkSL had no array
casting syntax as well. SkSL now has array-cast syntax, but it is only
allowed for casting low/high-precision arrays to the same base type.
(You can't cast an int array to float, or a signed array to unsigned.)
Change-Id: Ia20933541c3bd4a946c1ea38209f93008acdb9cb
Bug: skia:12248
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/437687
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
