perf(encoding): Dispatch loop for Varint bulk decode to handle "sporadic large values" (#594)#594
Open
srsuryadev wants to merge 8 commits intomainfrom
Open
perf(encoding): Dispatch loop for Varint bulk decode to handle "sporadic large values" (#594)#594srsuryadev wants to merge 8 commits intomainfrom
srsuryadev wants to merge 8 commits intomainfrom
Conversation
Summary: Add `decodeSingleByteRun` fast path to `bulkVarintDecode32` and `bulkVarintDecode64` that processes leading runs of single-byte varints (values 0-127) using 8-byte word reads before falling through to the BMI2 switch-based decoder. For each 8-byte word where no continuation bits are set (`word & 0x8080808080808080 == 0`), all 8 varints are decoded with simple shifts, avoiding the `_pext_u64` and 64-case switch overhead. This is placed in the caller functions rather than inside `bulkVarintDecodeBmi2` to preserve the BMI2 function's code layout and icache behavior for mixed-width data. Benchmark results (1M elements, mode/opt): | Scenario | Before | After | Speedup | |-----------------------|-----------|-----------|-----------| | 1-byte (32-bit) | 465us | 260us | 1.79x | | 5-byte (32-bit) | slower | 1.22ms | fixed | | 3-byte (32-bit) | 1.04ms | 864us | 1.20x | | 4-byte (32-bit) | 1.50ms | 1.04ms | 1.44x | | 64-bit 1-byte | 294us | 232us | 1.27x | | batch1024 | 1.96us | 1.20us | 1.63x | | Uniform/2-byte/8-byte | unchanged | unchanged | no regress| Also enhances the varint benchmark with fixed byte-width benchmarks (1-5 byte for 32-bit, 1/4/8 byte for 64-bit), skip benchmarks, and batch size benchmarks. Differential Revision: D96617939
… single-byte varints Summary: Manually loop-unroll `decodeSingleByteRun` with a 3-tier approach: 1. 32-element (4-word) unrolled loop with combined high-bit check `(w0 | w1 | w2 | w3) & kHighBits` to minimize branch overhead 2. 8-element (1-word) loop for smaller runs 3. Single-element trailing loop to pick up individual single-byte varints before multi-byte values Also extracts the byte-expansion logic into a reusable `expandWord()` helper for clarity. Differential Revision: D96619597
…gleByteRun Summary: Replace scalar byte expansion and reinterpret_cast-based uint64_t loads in decodeSingleByteRun with xsimd-based SIMD operations: - Use xsimd::batch<uint8_t>::load_unaligned for a single wide load (32 bytes on AVX2) + vptest to check all high bits at once, replacing 4 separate uint64_t loads + OR chain. - Use xsimd::batch<T> construction and store_unaligned for byte-to-element widening (compiles to vpmovzxbd on AVX2, vmovl on NEON). - Replace reinterpret_cast<const uint64_t*> with std::memcpy in the 8-byte loop to avoid strict-aliasing/alignment issues. Differential Revision: D96628007 Reviewed By: xiaoxmeng
…ncoding to make it robust Summary: Add further tests to the varint encoding to make it robust Differential Revision: D96665765
…ecode Summary: use pre-compiled lookup table for varint decode to eliminate switch case Differential Revision: D96756546
Summary: Add encoding fuzzer testing for varint encoding with randomized encode-decode-verify cycles across diverse data patterns and access patterns. Differential Revision: D97054913
Summary: Add bulkDecodeTwoByteRun() that detects runs of 2-byte varints by checking for the alternating high-bit pattern (0x0080008000800080) in 8-byte words, decoding 4 varints per word with simple scalar ops. This fixes the 2-byte regression introduced by the table-driven BMI2 decode (D96756546). The old switch perfectly predicted uniform 2-byte data, while the table-driven approach paid lookup overhead without benefiting from misprediction elimination. Benchmark results (varint_benchmark, mode/opt): - BulkDecode_2byte: 1.12ms → 506µs (2.2x faster, 11% faster than baseline) - NimbleBulkDecodeUniform: 1.49ms (preserved, 69% faster than baseline) - All other benchmarks: unchanged Differential Revision: D97189009
meta-cla
bot
added
the
CLA Signed
|
@srsuryadev has exported this pull request. If you are a Meta employee, you can view the originating Diff in D97366866. |
meta-codesync
bot
changed the title
srsuryadev
force-pushed
the
export-D97366866
branch
from
3a3822c to
81a4d26
Compare
srsuryadev
force-pushed
the
export-D97366866
branch
from
81a4d26 to
d23c0d5
Compare
srsuryadev
force-pushed
the
export-D97366866
branch
from
d23c0d5 to
8dfa177
Compare
srsuryadev
force-pushed
the
export-D97366866
branch
from
8dfa177 to
f55003f
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary:
Dispatch loop for varint bulk decode to handle sporadic large heads, i.e in a large sequence of 1-byte/2-byte values we see sporadically placed large values.
Differential Revision: D97366866