ColinLeeo opened a new issue, #1071:
URL: https://github.com/apache/tsfile/issues/1071
PR #1068 introduced C++ FLOAT/DOUBLE RLE (`FloatRLEDecoder` /
`DoubleRLEDecoder` in `cpp/src/encoding/float_rle_decoder.h`) and GORILLA_V1
(`FloatGorillaV1Decoder` / `DoubleGorillaV1Decoder` in
`cpp/src/encoding/gorilla_v1_decoder.h`). Both decode one value at a time and
never override the `read_batch_*` methods, so `Decoder::read_batch_double` /
`read_batch_int64` / ... (`cpp/src/encoding/decoder.h`) fall back to a
per-value loop over `read_double` / `read_int64`.
Existing codecs already received batch decode: PLAIN in #823, GORILLA and
TS_2DIFF in #898. The newly added encodings were missed.
## Impact
Profiling a wide-table read (5214 columns, RLE-encoded) shows decode
dominating the read path: `DoubleRLEDecoder::read_double` ~43% and
`Int64RleDecoder::read_int64` (time column) ~38%, essentially all inside the
per-value loop.
## Proposed changes
1. `Int64RleDecoder` / `Int32RleDecoder`
(`cpp/src/encoding/int64_rle_decoder.h` / `int32_rle_decoder.h`): add
`read_batch_int64` / `read_batch_int32`. `read_int` has two natural batch cases:
- RLE run (`is_rle_run_`): the next `current_count_` values are all the
same `rle_value_`, so fill the output range in one pass.
- bit-packing: `current_buffer_` already holds the fully unpacked group;
`memcpy` the remaining `current_count_` values.
This benefits both the time column (which uses `Int64RleDecoder`
directly) and the FLOAT/DOUBLE value columns (which wrap it).
2. `FloatRLEDecoder` / `DoubleRLEDecoder`: add `read_batch_float` /
`read_batch_double` overrides that batch-decode the inner int values and then
apply the per-value `value_at` scaling (which depends on `page_pos_`, so the
scaling itself stays per-value).
3. `GorillaV1DecoderBase`: add `read_batch_*` overrides following the
pattern already in `GorillaDecoder` (`cpp/src/encoding/gorilla_decoder.h`).
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]