Optimize ebml::reader::vuint_at() #11498
Conversation
Since reader::vuint_at() returns a result of type reader::Res it makes sense to make it public. Due to rust's current behavior of externally referenced private structures, rust-lang#10573, you could still use the result and assign it to a variable if you let the compiler do the type assignment, but you could not explicitly annotate a variable to hold a reader::Res.
| @@ -130,32 +130,32 @@ pub mod reader { | |||
| return vuint_at_slow(data, start); | |||
| } | |||
|
|
|||
| static shift_table: [uint, ..16] = [ | |||
There was a problem hiding this comment.
Conventionally these would be SHIFT_TABLE and MASK_TABLE.
|
New measurement storing mask and shift in a single table: And for academic interest without the bound check at the top of the function: |
|
This looks pretty awesome, thanks! Could you add some comments to the lookup table as to why it exists and what the values/rows signify? |
Use a lookup table, SHIFT_MASK_TABLE, that for every possible four bit prefix holds the number of times the value should be right shifted and what the right shifted value should be masked with. This way we can get rid of the branches which in my testing gives approximately a 2x speedup.
|
Very cool. What kind of impact does this have on rustc compiles? |
|
You may also be interested in #9303 |
Use a lookup table, SHIFT_MASK_TABLE, that for every possible four bit prefix holds the number of times the value should be right shifted and what the right shifted value should be masked with. This way we can get rid of the branches which in my testing gives approximately a 2x speedup. Timings on Intel(R) Core(TM) i5-3570K CPU @ 3.40GHz -- Before -- running 5 tests test ebml::tests::test_vuint_at ... ok test ebml::bench::vuint_at_A_aligned ... bench: 494 ns/iter (+/- 3) test ebml::bench::vuint_at_A_unaligned ... bench: 494 ns/iter (+/- 4) test ebml::bench::vuint_at_D_aligned ... bench: 467 ns/iter (+/- 5) test ebml::bench::vuint_at_D_unaligned ... bench: 467 ns/iter (+/- 5) -- After -- running 5 tests test ebml::tests::test_vuint_at ... ok test ebml::bench::vuint_at_A_aligned ... bench: 181 ns/iter (+/- 2) test ebml::bench::vuint_at_A_unaligned ... bench: 192 ns/iter (+/- 1) test ebml::bench::vuint_at_D_aligned ... bench: 181 ns/iter (+/- 3) test ebml::bench::vuint_at_D_unaligned ... bench: 197 ns/iter (+/- 6)
…affects_lint, r=Centri3 fix enum_variant_names depending lint depending on order changelog: [`enum_variant_names`]: fix single word variants preventing lint of later variant pre/postfixed with the enum name fixes rust-lang#11494 Single word variants prevented checking the `check_enum_start` and `check_enum_end` for being run on later variants
Use a lookup table, SHIFT_MASK_TABLE, that for every possible four
bit prefix holds the number of times the value should be right shifted and what
the right shifted value should be masked with. This way we can get rid of the
branches which in my testing gives approximately a 2x speedup.
Timings on Intel(R) Core(TM) i5-3570K CPU @ 3.40GHz
-- Before --
running 5 tests
test ebml::tests::test_vuint_at ... ok
test ebml::bench::vuint_at_A_aligned ... bench: 494 ns/iter (+/- 3)
test ebml::bench::vuint_at_A_unaligned ... bench: 494 ns/iter (+/- 4)
test ebml::bench::vuint_at_D_aligned ... bench: 467 ns/iter (+/- 5)
test ebml::bench::vuint_at_D_unaligned ... bench: 467 ns/iter (+/- 5)
-- After --
running 5 tests
test ebml::tests::test_vuint_at ... ok
test ebml::bench::vuint_at_A_aligned ... bench: 181 ns/iter (+/- 2)
test ebml::bench::vuint_at_A_unaligned ... bench: 192 ns/iter (+/- 1)
test ebml::bench::vuint_at_D_aligned ... bench: 181 ns/iter (+/- 3)
test ebml::bench::vuint_at_D_unaligned ... bench: 197 ns/iter (+/- 6)