[CSV 4/9] Word-at-a-time field seek and warp-aggregated validity and type counts in the CSV reader by karthikeyann · Pull Request #19 · karthikeyann/cudf

karthikeyann · 2026-09-24T10:48:46Z

Summary
The CSV type-detection and decode kernels process one row per thread. This PR cuts their memory transactions and atomic contention:

Field ends are found with aligned 8-byte word loads and a SWAR search, with an exact re-check.
Validity bits are set with one atomicOr per group of lanes on the same column. Null counts come from the finished masks via cudf::batch_null_count, replacing a same-address atomicAdd per field.
Type-detection counts are warp-aggregated. For tables of up to 32 inferred columns they are also accumulated per block in shared memory.
cudf::batch_null_count now supports more than 65,535 bitmasks. This is a libcudf-wide fix; see Notes.

Motivation
With one thread per row, a warp's loads are spread over 32 rows, so every character costs a separate memory transaction. The per-column counters (valid counts and type histograms) are a few addresses that every thread updates, so their atomics serialize.

Design

seek_field_end_by_words: returns exactly what seek_field_end returns without escape characters.
- Unquoted fields: bytes that may be a delimiter, terminator or '\r' are flagged with a SWAR zero-byte test that flags every occurrence. Flags are checked in increasing order with the exact test, so the first one that passes is the field end.
- Quoted fields: the same quote state machine runs over the loaded word. delim_whitespace keeps the old seek.
- Only words entirely within [data begin, row end) are loaded, whatever the alignment. Bytes outside the row are never used, which is what in-place unescaping in the decode kernel requires.
Validity: a warp's 32 rows share one mask word (bit = lane), so lanes on the same column are grouped with __match_any_sync and set their bits with one atomicOr. Null counts are computed from the zeroed masks over exactly [0, num_records), which adds no synchronization.
Type detection: fields are classified into a field_class that indexes the counters, and increments are warp-aggregated. classify_integral_field holds the integer logic and is shared with the JSON type inference. Up to 32 columns use per-block shared-memory counters (a separate template variant, which compiles to shared-memory atomics ATOMS; a runtime-flag build compiled the increment to a generic ATOM.E.ADD); wider tables use global counters. The cutoff was measured on one GPU and only affects speed.

Testing

New gtests: 1–17 byte inputs from exact-size device buffers at every alignment; field ends at every word position with delimiters/terminators \x01, \xFE, CR-only, and CRLF inside quotes; CRLF input including with delim_whitespace; validity at mask-word edges with short rows and use_cols; types decided by single fields in later blocks (8, 32, 33 and 40 columns); wide-table inference; more than 65,535 columns end to end plus a 70,000-mask batch_null_count test.
All CSV, JSON, type-inference and bitmask tests pass. memcheck (--rmm_mode=cuda) and racecheck are clean; the 70,000-column test allocates from its own pool so it stays fast under compute-sanitizer. Mutation checks on the counting paths are caught by the new tests.
Output is bit-identical to the parent on 8 oracle programs and on more than 20,000 fuzzed reads (host buffer and device buffers at offsets 0 and 3).

Performance (local, RTX PRO 6000, 256 MB, minimum of 3 interleaved rounds; official best-of-2 numbers for all sources are in the benchmark comment below)

Device buffer: TAXI −20%, LOGS −11%, ANALYTICS −15%, QUOTE_0 −37%, QUOTE_25 −23%; reads with type inference −26% to −60%, without −4% to −40%.
Host buffer: TAXI −11%, QUOTE_25 −6%. Per-kernel numbers are in the commit messages.

Notes for reviewers

Commit 2 changes cudf::batch_null_count (src/bitmask/null_mask.cu) for all of libcudf. Behaviour is unchanged for 65,535 masks or fewer. Owners of that utility should review it; it could also land separately first.
Commit 4 alone makes the type-detection kernel slower on 8-column float tables (5.38 → 5.73 ms); commit 5 fixes this (2.8 ms). Commit 5 costs 2–3% end to end on 8-column integer and mixed tables.
The shared-memory threshold is empirical (RTX PRO 6000, sm_120) and only affects speed.

This is PR 4 of a stack of draft PRs extracted from an automated (CORAL) optimization run of the CSV reader, reviewed by a separate principal-engineer review pass; it builds on #18.

🤖 Generated with Claude Code

The type detection and decode kernels process one row per thread, and `seek_field_end` reads each row one character at a time. The loads of a warp are therefore scattered over 32 rows, and each character costs a separate memory transaction. Add `seek_field_end_by_words`, which returns the same position as `seek_field_end` (without escape characters) but loads the row in aligned 8-byte words, and use it in both kernels: - In a field that does not start with a quote, only a delimiter, a terminator or a '\r' can end the field. The bytes of the word that may hold one of them are flagged with the SWAR zero-byte test on the word XOR the broadcast character. That test flags every occurrence (it may also flag other bytes), and the flags are checked in increasing order with the exact test of `seek_field_end` ('\r' only ends a field before '\n'), so the first flag that passes is the position `seek_field_end` returns, and false flags fail the test. - The characters of a quoted field are run through the same quote state machine as in `seek_field_end`, from the loaded word. - Multi-character delimiters (delim_whitespace) keep using `seek_field_end`. Only words that lie entirely within [data begin, row end) are loaded, whatever the alignment of the data, so no memory outside of the data is read even for unaligned or unpadded buffers; the characters of the row outside of such words are read one at a time. (The reader parses its own aligned copy of the data for now, so the lower bound matters for sources that would be parsed in place.) A word may start in the previous row, whose bytes the decode kernel may be rewriting (in place unescaping); those bytes are never used, as the documented invariant of the decode kernel requires. Median kernel times (nsys, RTX PRO 6000, 256MB benchmarks): benchmark kernel before after REALISTIC TAXI decode 5.64 ms 3.90 ms REALISTIC LOGS decode 1.62 ms 1.25 ms REALISTIC ANALYTICS decode 5.57 ms 4.12 ms QUOTING QUOTE_0_PCT decode 9.38 ms 6.53 ms QUOTING QUOTE_25_PCT decode 4.72 ms 3.02 ms QUOTING QUOTE_100_PCT decode 1.91 ms 1.22 ms INFERENCE ALL_INTEGRAL, 64 detection 11.36 ms 8.89 ms INFERENCE ALL_FLOAT, 8 detection 6.61 ms 5.38 ms Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

`batch_count_set_bits`, which backs `cudf::batch_null_count`, indexes the bitmasks with the second dimension of the grid, whose size is limited to 65535, so counting more bitmasks failed with cudaErrorInvalidValue. Launch the kernel once per batch of at most 65535 bitmasks. Nothing changes for fewer bitmasks. The CSV reader is about to count the nulls of its columns with this function, and CSV files can have more columns than that. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

The decode kernel set the validity bit of each valid field with its own atomicOr, and counted the valid fields of each column with an atomicAdd per field. The 32 lanes of a warp decode 32 consecutive rows that start at a multiple of 32, so their bits of a column all lie in the same mask word and their atomics serialize on it; the counter of a column is a single address that every warp of the grid updates for every field. Group the lanes that mark the same column with __match_any_sync and let the lowest of them set all of their bits with one atomicOr (the bit of a row is its lane index). The lanes that are active together normally mark the same column, but that is not guaranteed after divergence, so they are grouped by column. The atomic is still required because different groups of the same warp can update the word at different times. Stop counting in the kernel: after decoding, count the nulls of the non-string columns from their masks with cudf::batch_null_count. This is exact because the masks are zero-initialized, bits are only set for the valid fields of rows in [0, num_records), and exactly that range is counted. String columns get their null counts when they are built from the decoded string pairs, as before, so their masks are not counted. The count replaces the copy of the kernel's counters to the host, so the number of synchronizations does not change. A test reads a table with more columns than the batched count can index in one grid (65535), which the previous commit enables. Median time of the decode kernel (nsys, RTX PRO 6000, 256MB benchmarks): benchmark before after INFERENCE ALL_INTEGRAL, 8 4.24 ms 2.26 ms INFERENCE ALL_INTEGRAL, 64 7.51 ms 4.36 ms INFERENCE MIXED, 8 3.83 ms 2.26 ms INFERENCE MIXED, 64 3.72 ms 2.97 ms QUOTING QUOTE_0_PCT 6.53 ms 4.33 ms QUOTING QUOTE_25_PCT 3.02 ms 2.31 ms REALISTIC TAXI 3.90 ms 3.39 ms REALISTIC LOGS 1.25 ms 1.10 ms Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

The type detection kernel incremented a member of the column's column_type_histogram with a global atomicAdd in each branch of its classification. The counters of a column are a few addresses that every thread of the grid updates for each row, so these atomics serialize. Classify each field into a `field_class`, which indexes an array of `num_field_classes` counters per column, and increment the counter in one place. The lanes of a warp that increment the same counter at the same time are grouped with __match_any_sync, and the lowest of them adds their number with a single atomic. The counts are unchanged. Integers are classified by the new classify_integral_field, which holds the logic of infer_integral_field_counter; the latter is now a thin wrapper around it for the JSON type inference. The histograms are assembled on the host from the counters, so the device code no longer depends on the layout of column_type_histogram. A new test decides column types with single fields in later blocks of rows and with the number of NA fields, so that a lost or miscounted increment changes an inferred type. Median time of the type detection kernel (nsys, RTX PRO 6000, 256MB CSV_READER_TYPE_INFERENCE_NVBENCH with inference): profile, columns before after ALL_INTEGRAL, 8 2.77 ms 1.85 ms ALL_INTEGRAL, 64 8.00 ms 5.67 ms MIXED, 64 11.48 ms 5.92 ms ALL_STRING, 8 8.68 ms 1.15 ms ALL_STRING, 64 6.08 ms 2.03 ms ALL_FLOAT, 8 5.38 ms 5.73 ms (see the next commit) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Narrow tables have few type detection counters, and every warp of the grid updates them, so even warp-aggregated global atomics contend on them. This is worst when the lanes of a warp reach their increments at different times, as with variable-length floating point fields. For tables of up to 32 inferred columns, each block counts in shared memory, zeroed on entry, and adds its non-zero counts to the global counters at the end; threads without a row now stay until then instead of returning early. The counts are exact either way. The kernel is a template on where the counters are, rather than taking a runtime flag. With the counters known to be in shared memory, the compiler emits shared memory atomics (ATOMS in the SASS of this version); in a build that selected the counters with a runtime flag, the increment compiled to a generic atomic (ATOM.E.ADD), since the counter pointer can be in either address space. The global memory instance also keeps its early return and has no block barriers. The threshold only affects speed, and was measured on one GPU (RTX PRO 6000 Blackwell). Larger shared memory allocations per block made the kernel slower than counting in global memory (64 columns: 8.9 ms instead of 5.7 ms; 32 columns with the allocation padded to 2KB: 8.3 ms instead of 3.6 ms), likely because the kernel, which reads rows through the L1 cache, then gets a larger shared memory carveout. Wider tables spread their updates over enough counters that global memory atomics do not contend much on them. Median time of the type detection kernel (nsys, 256MB CSV_READER_TYPE_INFERENCE_NVBENCH with inference, num_cols overridden for 16 and 32): profile, columns global shared ALL_FLOAT, 8 5.6 ms 2.8 ms ALL_FLOAT, 32 4.9 ms 3.5 ms ALL_INTEGRAL, 16 6.4 ms 3.3 ms ALL_INTEGRAL, 32 5.5 ms 3.6 ms MIXED, 16 4.9 ms 3.2 ms ALL_INTEGRAL, 8 1.76 ms 1.88 ms Tables with 8 integer columns get slightly slower (ALL_INTEGRAL and MIXED: 2-3% end to end), which the gains above outweigh. The tests that decide column types with single fields in later blocks now also cover 32 and 33 columns, on both sides of the threshold, and fail if a block's shared counts are not added to the global counters. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

karthikeyann · 2026-09-24T11:12:08Z

Note

Re-measured cleanly (2026-09-25); these numbers replace the earlier ones in this comment. The earlier numbers were taken while other validation jobs (tests, sanitizers, fuzzers) ran on the same host, against a baseline measured hours earlier, so the pinned-host, pageable-host and file results in particular could be skewed.

Methodology (clean re-measurement). All ten builds of the stack (base and PRs 1–9) were measured in one session on the same GPU (host GPU 1), in two interleaved rounds (base, PR 1, …, PR 9, then again), so drift affects every build equally. Each benchmark run started only when no other process was using the GPU and no build/test/sanitizer/fuzzer was running on the host in our environment, and would have been re-run if anything appeared during it; all 160 runs were clean on the first attempt. Each number below is the best of the two rounds per side.

CSV reader benchmarks: PR 4/9 (word-at-a-time seek, warp-aggregated validity and type counts)

This PR 00ec4cd593 vs base coral/csv-base 63712844b0 and vs parent PR c7d51a913d. Measured on one NVIDIA RTX PRO 6000 Blackwell Server Edition, CUDA 13.3, Release build, with no builds running during the measurements. Throughput is NVBench bytes_per_second (uncompressed table bytes / mean GPU time); each side is the best of 2 run(s) (this PR) and 2 run(s) (base) per config. Run-to-run noise is a few percent, so treat changes under 3% as noise. Bold means at least 3% faster.

Speedup vs base (cumulative), by input source

Config	Device buffer	Pinned host	Pageable host	File (kvikio)
realistic TAXI 256MB	2.42× _{45.7 GiB/s}	1.51× _{16.9 GiB/s}	1.14× _{5.4 GiB/s}	1.53× _{9.7 GiB/s}
realistic LOGS 256MB	2.53× _{68.8 GiB/s}	1.58× _{26.9 GiB/s}	1.23× _{8.8 GiB/s}	1.57× _{15.5 GiB/s}
realistic ANALYTICS 256MB	1.73× _{43.8 GiB/s}	1.28× _{16.1 GiB/s}	1.02× _{4.9 GiB/s}	1.35× _{9.0 GiB/s}
inference MIXED 8 cols	1.89× _{39.9 GiB/s}	1.34× _{14.7 GiB/s}	1.13× _{4.6 GiB/s}	1.44× _{8.3 GiB/s}
inference MIXED 64 cols	2.61× _{17.2 GiB/s}	1.88× _{9.2 GiB/s}	1.38× _{3.6 GiB/s}	1.84× _{5.8 GiB/s}
no inference MIXED 8 cols	2.00× _{58.3 GiB/s}	1.28× _{16.5 GiB/s}	1.03× _{4.6 GiB/s}	1.25× _{8.0 GiB/s}
25% quoted 256MB	3.30× _{42.0 GiB/s}	2.08× _{20.3 GiB/s}	1.37× _{6.8 GiB/s}	1.76× _{11.4 GiB/s}
100% quoted 256MB	5.33× _{26.0 GiB/s}	3.60× _{19.5 GiB/s}	2.22× _{9.3 GiB/s}	2.84× _{13.3 GiB/s}
Weighted geomean	2.53×	1.68×	1.26×	1.62×

Speedup vs parent PR (this PR's own contribution)

Config	Device buffer	Pinned host	Pageable host	File (kvikio)
realistic TAXI 256MB	1.31× _{45.7 GiB/s}	1.11× _{16.9 GiB/s}	1.07× _{5.4 GiB/s}	1.14× _{9.7 GiB/s}
realistic LOGS 256MB	1.15× _{68.8 GiB/s}	1.05× _{26.9 GiB/s}	1.10× _{8.8 GiB/s}	1.09× _{15.5 GiB/s}
realistic ANALYTICS 256MB	1.18× _{43.8 GiB/s}	1.08× _{16.1 GiB/s}	1.03× _{4.9 GiB/s}	1.09× _{9.0 GiB/s}
inference MIXED 8 cols	1.41× _{39.9 GiB/s}	1.17× _{14.7 GiB/s}	1.10× _{4.6 GiB/s}	1.12× _{8.3 GiB/s}
inference MIXED 64 cols	1.75× _{17.2 GiB/s}	1.44× _{9.2 GiB/s}	1.20× _{3.6 GiB/s}	1.28× _{5.8 GiB/s}
no inference MIXED 8 cols	1.39× _{58.3 GiB/s}	1.11× _{16.5 GiB/s}	1.00× _{4.6 GiB/s}	0.96× _{8.0 GiB/s}
25% quoted 256MB	1.28× _{42.0 GiB/s}	1.13× _{20.3 GiB/s}	1.00× _{6.8 GiB/s}	1.03× _{11.4 GiB/s}
100% quoted 256MB	1.06× _{26.0 GiB/s}	1.05× _{19.5 GiB/s}	1.02× _{9.3 GiB/s}	0.98× _{13.3 GiB/s}
Weighted geomean	1.31×	1.14×	1.06×	1.09×

_{Sources: the *_IO_NVBENCH benchmarks (same data as the primary benchmarks, 256MB, with the input in device memory, pinned host memory, pageable host memory, or a file read via kvikio; file reads hit the OS page cache). Weighted geomean uses weight 2 for TAXI, MIXED-8 with inference and 25% quoted, weight 1 otherwise.}

Full results: every CSV reader benchmark config

`CSV_READER_NVBENCH` (18 configs)

Benchmark	Config	Base	Parent	This PR	vs base	vs parent
csv_read_column_selection	column_selection=ALL, num_chunks=1, row_selection=ALL	2.63 GiB/s	3.08 GiB/s	3.44 GiB/s	1.31×	1.12×
csv_read_column_selection	column_selection=ALTERNATE, num_chunks=1, row_selection=ALL	1.45 GiB/s	1.63 GiB/s	1.80 GiB/s	1.24×	1.11×
csv_read_column_selection	column_selection=FIRST_HALF, num_chunks=1, row_selection=ALL	1.49 GiB/s	1.69 GiB/s	1.85 GiB/s	1.24×	1.10×
csv_read_column_selection	column_selection=SECOND_HALF, num_chunks=1, row_selection=ALL	1.49 GiB/s	1.69 GiB/s	1.86 GiB/s	1.24×	1.10×
csv_read_data_type	data_type=DECIMAL, io=DEVICE_BUFFER	15.33 GiB/s	23.72 GiB/s	34.56 GiB/s	2.25×	1.46×
csv_read_data_type	data_type=DURATION, io=DEVICE_BUFFER	12.78 GiB/s	18.73 GiB/s	32.63 GiB/s	2.55×	1.74×
csv_read_data_type	data_type=FLOAT, io=DEVICE_BUFFER	18.66 GiB/s	23.27 GiB/s	26.45 GiB/s	1.42×	1.14×
csv_read_data_type	data_type=INTEGRAL, io=DEVICE_BUFFER	17.43 GiB/s	23.30 GiB/s	37.60 GiB/s	2.16×	1.61×
csv_read_data_type	data_type=STRING, io=DEVICE_BUFFER	4.68 GiB/s	25.79 GiB/s	26.88 GiB/s	5.74×	1.04×
csv_read_data_type	data_type=TIMESTAMP, io=DEVICE_BUFFER	12.37 GiB/s	17.60 GiB/s	26.20 GiB/s	2.12×	1.49×
csv_read_io	io=FILEPATH	4.55 GiB/s	6.47 GiB/s	7.09 GiB/s	1.56×	1.10×
csv_read_io	io=HOST_BUFFER	3.38 GiB/s	3.87 GiB/s	3.85 GiB/s	1.14×	0.99×
csv_read_row_selection	column_selection=ALL, num_chunks=1, row_selection=BYTE_RANGE	2.56 GiB/s	2.86 GiB/s	3.19 GiB/s	1.25×	1.11×
csv_read_row_selection	column_selection=ALL, num_chunks=1, row_selection=NROWS	2.54 GiB/s	2.87 GiB/s	3.32 GiB/s	1.30×	1.16×
csv_read_row_selection	column_selection=ALL, num_chunks=1, row_selection=SKIPFOOTER	2.67 GiB/s	3.10 GiB/s	3.61 GiB/s	1.35×	1.17×
csv_read_row_selection	column_selection=ALL, num_chunks=8, row_selection=BYTE_RANGE	1.93 GiB/s	2.48 GiB/s	2.66 GiB/s	1.38×	1.07×
csv_read_row_selection	column_selection=ALL, num_chunks=8, row_selection=NROWS	0.71 GiB/s	0.76 GiB/s	0.78 GiB/s	1.09×	1.02×
csv_read_row_selection	column_selection=ALL, num_chunks=8, row_selection=SKIPFOOTER	0.45 GiB/s	0.47 GiB/s	0.47 GiB/s	1.05×	1.00×

`CSV_READER_QUOTING_IO_NVBENCH` (12 configs)

Benchmark	Config	Base	Parent	This PR	vs base	vs parent
csv_read_quoting_io	data_size_mb=256, io=DEVICE_BUFFER, quote_pct=QUOTE_0_PCT	17.11 GiB/s	23.33 GiB/s	37.03 GiB/s	2.16×	1.59×
csv_read_quoting_io	data_size_mb=256, io=DEVICE_BUFFER, quote_pct=QUOTE_100_PCT	4.87 GiB/s	24.59 GiB/s	25.96 GiB/s	5.33×	1.06×
csv_read_quoting_io	data_size_mb=256, io=DEVICE_BUFFER, quote_pct=QUOTE_25_PCT	12.75 GiB/s	32.74 GiB/s	42.03 GiB/s	3.30×	1.28×
csv_read_quoting_io	data_size_mb=256, io=FILEPATH, quote_pct=QUOTE_0_PCT	5.25 GiB/s	6.93 GiB/s	7.38 GiB/s	1.41×	1.06×
csv_read_quoting_io	data_size_mb=256, io=FILEPATH, quote_pct=QUOTE_100_PCT	4.69 GiB/s	13.53 GiB/s	13.32 GiB/s	2.84×	0.98×
csv_read_quoting_io	data_size_mb=256, io=FILEPATH, quote_pct=QUOTE_25_PCT	6.47 GiB/s	11.03 GiB/s	11.40 GiB/s	1.76×	1.03×
csv_read_quoting_io	data_size_mb=256, io=HOST_BUFFER, quote_pct=QUOTE_0_PCT	3.80 GiB/s	4.03 GiB/s	4.27 GiB/s	1.12×	1.06×
csv_read_quoting_io	data_size_mb=256, io=HOST_BUFFER, quote_pct=QUOTE_100_PCT	4.18 GiB/s	9.08 GiB/s	9.28 GiB/s	2.22×	1.02×
csv_read_quoting_io	data_size_mb=256, io=HOST_BUFFER, quote_pct=QUOTE_25_PCT	4.99 GiB/s	6.82 GiB/s	6.85 GiB/s	1.37×	1.00×
csv_read_quoting_io	data_size_mb=256, io=PINNED_BUFFER, quote_pct=QUOTE_0_PCT	9.57 GiB/s	11.45 GiB/s	13.90 GiB/s	1.45×	1.21×
csv_read_quoting_io	data_size_mb=256, io=PINNED_BUFFER, quote_pct=QUOTE_100_PCT	5.40 GiB/s	18.59 GiB/s	19.45 GiB/s	3.60×	1.05×
csv_read_quoting_io	data_size_mb=256, io=PINNED_BUFFER, quote_pct=QUOTE_25_PCT	9.77 GiB/s	17.88 GiB/s	20.29 GiB/s	2.08×	1.13×

`CSV_READER_QUOTING_NVBENCH` (6 configs)

Benchmark	Config	Base	Parent	This PR	vs base	vs parent
csv_read_quoting	data_size_mb=256, quote_pct=QUOTE_0_PCT	17.51 GiB/s	24.55 GiB/s	38.28 GiB/s	2.19×	1.56×
csv_read_quoting	data_size_mb=256, quote_pct=QUOTE_100_PCT	4.80 GiB/s	25.28 GiB/s	27.23 GiB/s	5.68×	1.08×
csv_read_quoting	data_size_mb=256, quote_pct=QUOTE_25_PCT	13.88 GiB/s	33.07 GiB/s	42.54 GiB/s	3.06×	1.29×
csv_read_quoting	data_size_mb=64, quote_pct=QUOTE_0_PCT	14.10 GiB/s	18.42 GiB/s	26.77 GiB/s	1.90×	1.45×
csv_read_quoting	data_size_mb=64, quote_pct=QUOTE_100_PCT	1.37 GiB/s	8.34 GiB/s	8.61 GiB/s	6.30×	1.03×
csv_read_quoting	data_size_mb=64, quote_pct=QUOTE_25_PCT	4.57 GiB/s	20.47 GiB/s	21.62 GiB/s	4.73×	1.06×

`CSV_READER_REALISTIC_IO_NVBENCH` (12 configs)

Benchmark	Config	Base	Parent	This PR	vs base	vs parent
csv_read_realistic_io	data_size_mb=256, io=DEVICE_BUFFER, profile=ANALYTICS	25.28 GiB/s	37.00 GiB/s	43.83 GiB/s	1.73×	1.18×
csv_read_realistic_io	data_size_mb=256, io=DEVICE_BUFFER, profile=LOGS	27.19 GiB/s	60.08 GiB/s	68.82 GiB/s	2.53×	1.15×
csv_read_realistic_io	data_size_mb=256, io=DEVICE_BUFFER, profile=TAXI	18.91 GiB/s	34.85 GiB/s	45.69 GiB/s	2.42×	1.31×
csv_read_realistic_io	data_size_mb=256, io=FILEPATH, profile=ANALYTICS	6.63 GiB/s	8.25 GiB/s	8.96 GiB/s	1.35×	1.09×
csv_read_realistic_io	data_size_mb=256, io=FILEPATH, profile=LOGS	9.85 GiB/s	14.16 GiB/s	15.48 GiB/s	1.57×	1.09×
csv_read_realistic_io	data_size_mb=256, io=FILEPATH, profile=TAXI	6.37 GiB/s	8.55 GiB/s	9.73 GiB/s	1.53×	1.14×
csv_read_realistic_io	data_size_mb=256, io=HOST_BUFFER, profile=ANALYTICS	4.83 GiB/s	4.79 GiB/s	4.92 GiB/s	1.02×	1.03×
csv_read_realistic_io	data_size_mb=256, io=HOST_BUFFER, profile=LOGS	7.16 GiB/s	7.98 GiB/s	8.78 GiB/s	1.23×	1.10×
csv_read_realistic_io	data_size_mb=256, io=HOST_BUFFER, profile=TAXI	4.77 GiB/s	5.07 GiB/s	5.43 GiB/s	1.14×	1.07×
csv_read_realistic_io	data_size_mb=256, io=PINNED_BUFFER, profile=ANALYTICS	12.58 GiB/s	14.93 GiB/s	16.09 GiB/s	1.28×	1.08×
csv_read_realistic_io	data_size_mb=256, io=PINNED_BUFFER, profile=LOGS	17.05 GiB/s	25.59 GiB/s	26.92 GiB/s	1.58×	1.05×
csv_read_realistic_io	data_size_mb=256, io=PINNED_BUFFER, profile=TAXI	11.20 GiB/s	15.19 GiB/s	16.91 GiB/s	1.51×	1.11×

`CSV_READER_REALISTIC_NVBENCH` (9 configs)

Benchmark	Config	Base	Parent	This PR	vs base	vs parent
csv_read_realistic	data_size_mb=1024, profile=ANALYTICS	25.11 GiB/s	40.80 GiB/s	49.03 GiB/s	1.95×	1.20×
csv_read_realistic	data_size_mb=1024, profile=LOGS	31.88 GiB/s	69.84 GiB/s	81.58 GiB/s	2.56×	1.17×
csv_read_realistic	data_size_mb=1024, profile=TAXI	23.43 GiB/s	39.91 GiB/s	52.73 GiB/s	2.25×	1.32×
csv_read_realistic	data_size_mb=256, profile=ANALYTICS	25.24 GiB/s	37.14 GiB/s	43.96 GiB/s	1.74×	1.18×
csv_read_realistic	data_size_mb=256, profile=LOGS	27.37 GiB/s	60.37 GiB/s	69.43 GiB/s	2.54×	1.15×
csv_read_realistic	data_size_mb=256, profile=TAXI	19.01 GiB/s	35.43 GiB/s	45.63 GiB/s	2.40×	1.29×
csv_read_realistic	data_size_mb=512, profile=ANALYTICS	25.64 GiB/s	39.44 GiB/s	47.23 GiB/s	1.84×	1.20×
csv_read_realistic	data_size_mb=512, profile=LOGS	31.05 GiB/s	67.20 GiB/s	77.72 GiB/s	2.50×	1.16×
csv_read_realistic	data_size_mb=512, profile=TAXI	22.88 GiB/s	38.22 GiB/s	50.45 GiB/s	2.20×	1.32×

`CSV_READER_SCALE_NVBENCH` (5 configs)

Benchmark	Config	Base	Parent	This PR	vs base	vs parent
csv_read_scale	data_size_mb=1024	3.54 GiB/s	3.66 GiB/s	4.05 GiB/s	1.14×	1.11×
csv_read_scale	data_size_mb=2048	3.52 GiB/s	3.85 GiB/s	4.06 GiB/s	1.15×	1.05×
csv_read_scale	data_size_mb=256	3.59 GiB/s	3.62 GiB/s	3.99 GiB/s	1.11×	1.10×
csv_read_scale	data_size_mb=4096	3.38 GiB/s	3.71 GiB/s	3.95 GiB/s	1.17×	1.06×
csv_read_scale	data_size_mb=512	3.60 GiB/s	3.65 GiB/s	4.03 GiB/s	1.12×	1.11×

`CSV_READER_TYPE_INFERENCE_IO_NVBENCH` (64 configs)

Benchmark	Config	Base	Parent	This PR	vs base	vs parent
csv_read_with_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=64, type_profile=ALL_FLOAT	10.11 GiB/s	13.29 GiB/s	17.94 GiB/s	1.77×	1.35×
csv_read_with_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=64, type_profile=ALL_INTEGRAL	9.02 GiB/s	11.63 GiB/s	21.21 GiB/s	2.35×	1.82×
csv_read_with_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=64, type_profile=ALL_STRING	4.40 GiB/s	15.06 GiB/s	22.05 GiB/s	5.01×	1.46×
csv_read_with_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=64, type_profile=MIXED	6.59 GiB/s	9.82 GiB/s	17.18 GiB/s	2.61×	1.75×
csv_read_with_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=8, type_profile=ALL_FLOAT	12.31 GiB/s	16.49 GiB/s	22.54 GiB/s	1.83×	1.37×
csv_read_with_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=8, type_profile=ALL_INTEGRAL	21.14 GiB/s	28.36 GiB/s	39.96 GiB/s	1.89×	1.41×
csv_read_with_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=8, type_profile=ALL_STRING	11.91 GiB/s	18.97 GiB/s	50.17 GiB/s	4.21×	2.65×
csv_read_with_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=8, type_profile=MIXED	21.15 GiB/s	28.38 GiB/s	39.92 GiB/s	1.89×	1.41×
csv_read_with_inference_io	data_size_mb=256, io=FILEPATH, num_cols=64, type_profile=ALL_FLOAT	4.03 GiB/s	5.20 GiB/s	5.80 GiB/s	1.44×	1.11×
csv_read_with_inference_io	data_size_mb=256, io=FILEPATH, num_cols=64, type_profile=ALL_INTEGRAL	4.12 GiB/s	5.01 GiB/s	6.37 GiB/s	1.55×	1.27×
csv_read_with_inference_io	data_size_mb=256, io=FILEPATH, num_cols=64, type_profile=ALL_STRING	3.94 GiB/s	9.98 GiB/s	12.28 GiB/s	3.12×	1.23×
csv_read_with_inference_io	data_size_mb=256, io=FILEPATH, num_cols=64, type_profile=MIXED	3.16 GiB/s	4.53 GiB/s	5.80 GiB/s	1.84×	1.28×
csv_read_with_inference_io	data_size_mb=256, io=FILEPATH, num_cols=8, type_profile=ALL_FLOAT	4.37 GiB/s	5.73 GiB/s	6.14 GiB/s	1.41×	1.07×
csv_read_with_inference_io	data_size_mb=256, io=FILEPATH, num_cols=8, type_profile=ALL_INTEGRAL	5.98 GiB/s	7.10 GiB/s	7.96 GiB/s	1.33×	1.12×
csv_read_with_inference_io	data_size_mb=256, io=FILEPATH, num_cols=8, type_profile=ALL_STRING	7.49 GiB/s	11.38 GiB/s	17.26 GiB/s	2.31×	1.52×
csv_read_with_inference_io	data_size_mb=256, io=FILEPATH, num_cols=8, type_profile=MIXED	5.79 GiB/s	7.47 GiB/s	8.34 GiB/s	1.44×	1.12×
csv_read_with_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=64, type_profile=ALL_FLOAT	3.07 GiB/s	3.31 GiB/s	3.53 GiB/s	1.15×	1.07×
csv_read_with_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=64, type_profile=ALL_INTEGRAL	3.21 GiB/s	3.26 GiB/s	3.90 GiB/s	1.22×	1.20×
csv_read_with_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=64, type_profile=ALL_STRING	3.50 GiB/s	7.23 GiB/s	8.63 GiB/s	2.47×	1.19×
csv_read_with_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=64, type_profile=MIXED	2.60 GiB/s	3.00 GiB/s	3.59 GiB/s	1.38×	1.20×
csv_read_with_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=8, type_profile=ALL_FLOAT	3.30 GiB/s	3.39 GiB/s	3.54 GiB/s	1.07×	1.05×
csv_read_with_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=8, type_profile=ALL_INTEGRAL	4.04 GiB/s	4.25 GiB/s	4.38 GiB/s	1.08×	1.03×
csv_read_with_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=8, type_profile=ALL_STRING	6.36 GiB/s	7.83 GiB/s	10.91 GiB/s	1.72×	1.39×
csv_read_with_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=8, type_profile=MIXED	4.06 GiB/s	4.17 GiB/s	4.58 GiB/s	1.13×	1.10×
csv_read_with_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=64, type_profile=ALL_FLOAT	6.80 GiB/s	8.01 GiB/s	9.64 GiB/s	1.42×	1.20×
csv_read_with_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=64, type_profile=ALL_INTEGRAL	6.43 GiB/s	7.37 GiB/s	10.86 GiB/s	1.69×	1.47×
csv_read_with_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=64, type_profile=ALL_STRING	4.36 GiB/s	12.45 GiB/s	16.85 GiB/s	3.86×	1.35×
csv_read_with_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=64, type_profile=MIXED	4.90 GiB/s	6.37 GiB/s	9.20 GiB/s	1.88×	1.44×
csv_read_with_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=8, type_profile=ALL_FLOAT	7.73 GiB/s	9.12 GiB/s	10.79 GiB/s	1.39×	1.18×
csv_read_with_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=8, type_profile=ALL_INTEGRAL	10.97 GiB/s	12.56 GiB/s	14.63 GiB/s	1.33×	1.16×
csv_read_with_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=8, type_profile=ALL_STRING	10.30 GiB/s	14.79 GiB/s	28.79 GiB/s	2.79×	1.95×
csv_read_with_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=8, type_profile=MIXED	10.97 GiB/s	12.61 GiB/s	14.70 GiB/s	1.34×	1.17×
csv_read_without_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=64, type_profile=ALL_FLOAT	18.66 GiB/s	23.22 GiB/s	26.36 GiB/s	1.41×	1.14×
csv_read_without_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=64, type_profile=ALL_INTEGRAL	17.43 GiB/s	24.48 GiB/s	38.11 GiB/s	2.19×	1.56×
csv_read_without_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=64, type_profile=ALL_STRING	4.91 GiB/s	24.50 GiB/s	25.98 GiB/s	5.29×	1.06×
csv_read_without_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=64, type_profile=MIXED	14.50 GiB/s	23.39 GiB/s	34.86 GiB/s	2.40×	1.49×
csv_read_without_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=8, type_profile=ALL_FLOAT	21.45 GiB/s	26.91 GiB/s	27.76 GiB/s	1.29×	1.03×
csv_read_without_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=8, type_profile=ALL_INTEGRAL	28.99 GiB/s	41.64 GiB/s	58.82 GiB/s	2.03×	1.41×
csv_read_without_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=8, type_profile=ALL_STRING	20.79 GiB/s	59.74 GiB/s	65.06 GiB/s	3.13×	1.09×
csv_read_without_inference_io	data_size_mb=256, io=DEVICE_BUFFER, num_cols=8, type_profile=MIXED	29.16 GiB/s	42.00 GiB/s	58.35 GiB/s	2.00×	1.39×
csv_read_without_inference_io	data_size_mb=256, io=FILEPATH, num_cols=64, type_profile=ALL_FLOAT	5.12 GiB/s	6.62 GiB/s	6.37 GiB/s	1.24×	0.96×
csv_read_without_inference_io	data_size_mb=256, io=FILEPATH, num_cols=64, type_profile=ALL_INTEGRAL	5.02 GiB/s	7.09 GiB/s	7.68 GiB/s	1.53×	1.08×
csv_read_without_inference_io	data_size_mb=256, io=FILEPATH, num_cols=64, type_profile=ALL_STRING	4.78 GiB/s	13.46 GiB/s	13.38 GiB/s	2.80×	0.99×
csv_read_without_inference_io	data_size_mb=256, io=FILEPATH, num_cols=64, type_profile=MIXED	4.60 GiB/s	6.50 GiB/s	6.61 GiB/s	1.44×	1.02×
csv_read_without_inference_io	data_size_mb=256, io=FILEPATH, num_cols=8, type_profile=ALL_FLOAT	5.29 GiB/s	6.94 GiB/s	6.49 GiB/s	1.23×	0.94×
csv_read_without_inference_io	data_size_mb=256, io=FILEPATH, num_cols=8, type_profile=ALL_INTEGRAL	6.01 GiB/s	8.31 GiB/s	8.65 GiB/s	1.44×	1.04×
csv_read_without_inference_io	data_size_mb=256, io=FILEPATH, num_cols=8, type_profile=ALL_STRING	11.19 GiB/s	19.09 GiB/s	18.49 GiB/s	1.65×	0.97×
csv_read_without_inference_io	data_size_mb=256, io=FILEPATH, num_cols=8, type_profile=MIXED	6.39 GiB/s	8.32 GiB/s	7.96 GiB/s	1.25×	0.96×
csv_read_without_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=64, type_profile=ALL_FLOAT	3.75 GiB/s	3.90 GiB/s	3.80 GiB/s	1.01×	0.97×
csv_read_without_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=64, type_profile=ALL_INTEGRAL	3.79 GiB/s	4.21 GiB/s	4.45 GiB/s	1.17×	1.06×
csv_read_without_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=64, type_profile=ALL_STRING	4.25 GiB/s	9.12 GiB/s	9.14 GiB/s	2.15×	1.00×
csv_read_without_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=64, type_profile=MIXED	3.33 GiB/s	3.68 GiB/s	3.89 GiB/s	1.17×	1.05×
csv_read_without_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=8, type_profile=ALL_FLOAT	3.90 GiB/s	4.01 GiB/s	3.90 GiB/s	1.00×	0.97×
csv_read_without_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=8, type_profile=ALL_INTEGRAL	4.30 GiB/s	4.64 GiB/s	4.75 GiB/s	1.10×	1.02×
csv_read_without_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=8, type_profile=ALL_STRING	8.50 GiB/s	11.04 GiB/s	11.32 GiB/s	1.33×	1.03×
csv_read_without_inference_io	data_size_mb=256, io=HOST_BUFFER, num_cols=8, type_profile=MIXED	4.49 GiB/s	4.66 GiB/s	4.64 GiB/s	1.03×	1.00×
csv_read_without_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=64, type_profile=ALL_FLOAT	9.60 GiB/s	10.89 GiB/s	11.44 GiB/s	1.19×	1.05×
csv_read_without_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=64, type_profile=ALL_INTEGRAL	9.67 GiB/s	11.47 GiB/s	14.00 GiB/s	1.45×	1.22×
csv_read_without_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=64, type_profile=ALL_STRING	5.47 GiB/s	18.53 GiB/s	19.51 GiB/s	3.57×	1.05×
csv_read_without_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=64, type_profile=MIXED	7.96 GiB/s	10.71 GiB/s	12.72 GiB/s	1.60×	1.19×
csv_read_without_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=8, type_profile=ALL_FLOAT	10.51 GiB/s	11.67 GiB/s	11.73 GiB/s	1.12×	1.01×
csv_read_without_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=8, type_profile=ALL_INTEGRAL	12.94 GiB/s	14.77 GiB/s	16.66 GiB/s	1.29×	1.13×
csv_read_without_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=8, type_profile=ALL_STRING	17.06 GiB/s	31.61 GiB/s	33.26 GiB/s	1.95×	1.05×
csv_read_without_inference_io	data_size_mb=256, io=PINNED_BUFFER, num_cols=8, type_profile=MIXED	12.86 GiB/s	14.78 GiB/s	16.47 GiB/s	1.28×	1.11×

`CSV_READER_TYPE_INFERENCE_NVBENCH` (32 configs)

Benchmark	Config	Base	Parent	This PR	vs base	vs parent
csv_read_with_inference	data_size_mb=256, num_cols=64, type_profile=ALL_FLOAT	10.18 GiB/s	13.44 GiB/s	18.06 GiB/s	1.77×	1.34×
csv_read_with_inference	data_size_mb=256, num_cols=64, type_profile=ALL_INTEGRAL	9.10 GiB/s	12.01 GiB/s	21.53 GiB/s	2.37×	1.79×
csv_read_with_inference	data_size_mb=256, num_cols=64, type_profile=ALL_STRING	4.40 GiB/s	16.23 GiB/s	22.80 GiB/s	5.18×	1.40×
csv_read_with_inference	data_size_mb=256, num_cols=64, type_profile=MIXED	6.80 GiB/s	9.82 GiB/s	17.33 GiB/s	2.55×	1.76×
csv_read_with_inference	data_size_mb=256, num_cols=8, type_profile=ALL_FLOAT	12.32 GiB/s	17.01 GiB/s	22.61 GiB/s	1.84×	1.33×
csv_read_with_inference	data_size_mb=256, num_cols=8, type_profile=ALL_INTEGRAL	21.29 GiB/s	28.43 GiB/s	40.00 GiB/s	1.88×	1.41×
csv_read_with_inference	data_size_mb=256, num_cols=8, type_profile=ALL_STRING	12.00 GiB/s	20.59 GiB/s	50.31 GiB/s	4.19×	2.44×
csv_read_with_inference	data_size_mb=256, num_cols=8, type_profile=MIXED	21.64 GiB/s	28.95 GiB/s	40.71 GiB/s	1.88×	1.41×
csv_read_with_inference	data_size_mb=64, num_cols=64, type_profile=ALL_FLOAT	10.15 GiB/s	10.41 GiB/s	13.54 GiB/s	1.33×	1.30×
csv_read_with_inference	data_size_mb=64, num_cols=64, type_profile=ALL_INTEGRAL	8.33 GiB/s	10.47 GiB/s	15.68 GiB/s	1.88×	1.50×
csv_read_with_inference	data_size_mb=64, num_cols=64, type_profile=ALL_STRING	1.42 GiB/s	7.10 GiB/s	7.84 GiB/s	5.51×	1.10×
csv_read_with_inference	data_size_mb=64, num_cols=64, type_profile=MIXED	4.05 GiB/s	7.30 GiB/s	10.58 GiB/s	2.61×	1.45×
csv_read_with_inference	data_size_mb=64, num_cols=8, type_profile=ALL_FLOAT	11.10 GiB/s	14.46 GiB/s	18.60 GiB/s	1.68×	1.29×
csv_read_with_inference	data_size_mb=64, num_cols=8, type_profile=ALL_INTEGRAL	18.41 GiB/s	23.81 GiB/s	30.59 GiB/s	1.66×	1.28×
csv_read_with_inference	data_size_mb=64, num_cols=8, type_profile=ALL_STRING	5.52 GiB/s	16.80 GiB/s	31.08 GiB/s	5.63×	1.85×
csv_read_with_inference	data_size_mb=64, num_cols=8, type_profile=MIXED	19.86 GiB/s	25.03 GiB/s	32.60 GiB/s	1.64×	1.30×
csv_read_without_inference	data_size_mb=256, num_cols=64, type_profile=ALL_FLOAT	18.10 GiB/s	23.49 GiB/s	26.40 GiB/s	1.46×	1.12×
csv_read_without_inference	data_size_mb=256, num_cols=64, type_profile=ALL_INTEGRAL	17.51 GiB/s	24.68 GiB/s	38.37 GiB/s	2.19×	1.55×
csv_read_without_inference	data_size_mb=256, num_cols=64, type_profile=ALL_STRING	4.70 GiB/s	25.21 GiB/s	27.15 GiB/s	5.78×	1.08×
csv_read_without_inference	data_size_mb=256, num_cols=64, type_profile=MIXED	14.57 GiB/s	23.41 GiB/s	35.16 GiB/s	2.41×	1.50×
csv_read_without_inference	data_size_mb=256, num_cols=8, type_profile=ALL_FLOAT	21.55 GiB/s	26.93 GiB/s	27.92 GiB/s	1.30×	1.04×
csv_read_without_inference	data_size_mb=256, num_cols=8, type_profile=ALL_INTEGRAL	29.30 GiB/s	42.93 GiB/s	58.48 GiB/s	2.00×	1.36×
csv_read_without_inference	data_size_mb=256, num_cols=8, type_profile=ALL_STRING	21.92 GiB/s	63.01 GiB/s	64.91 GiB/s	2.96×	1.03×
csv_read_without_inference	data_size_mb=256, num_cols=8, type_profile=MIXED	29.24 GiB/s	42.16 GiB/s	59.39 GiB/s	2.03×	1.41×
csv_read_without_inference	data_size_mb=64, num_cols=64, type_profile=ALL_FLOAT	15.91 GiB/s	16.00 GiB/s	18.68 GiB/s	1.17×	1.17×
csv_read_without_inference	data_size_mb=64, num_cols=64, type_profile=ALL_INTEGRAL	14.41 GiB/s	18.52 GiB/s	26.43 GiB/s	1.83×	1.43×
csv_read_without_inference	data_size_mb=64, num_cols=64, type_profile=ALL_STRING	1.44 GiB/s	8.51 GiB/s	8.69 GiB/s	6.05×	1.02×
csv_read_without_inference	data_size_mb=64, num_cols=64, type_profile=MIXED	9.73 GiB/s	17.03 GiB/s	23.28 GiB/s	2.39×	1.37×
csv_read_without_inference	data_size_mb=64, num_cols=8, type_profile=ALL_FLOAT	18.58 GiB/s	21.68 GiB/s	23.24 GiB/s	1.25×	1.07×
csv_read_without_inference	data_size_mb=64, num_cols=8, type_profile=ALL_INTEGRAL	25.71 GiB/s	35.53 GiB/s	43.52 GiB/s	1.69×	1.23×
csv_read_without_inference	data_size_mb=64, num_cols=8, type_profile=ALL_STRING	7.32 GiB/s	40.94 GiB/s	37.78 GiB/s	5.16×	0.92×
csv_read_without_inference	data_size_mb=64, num_cols=8, type_profile=MIXED	25.23 GiB/s	35.45 GiB/s	43.45 GiB/s	1.72×	1.23×

karthikeyann and others added 5 commits September 24, 2026 09:07

github-actions Bot assigned karthikeyann Sep 24, 2026

github-actions Bot added the libcudf label Sep 24, 2026

This was referenced Sep 24, 2026

[CSV 5/9] Stage the rows of CSV decode and type detection blocks in shared memory #20

Draft

[CSV all] Upstreamable CSV reader optimizations (combined #16–#24) #25

Draft

Navigation Menu

Repository navigation

[CSV 4/9] Word-at-a-time field seek and warp-aggregated validity and type counts in the CSV reader - #19

[CSV 4/9] Word-at-a-time field seek and warp-aggregated validity and type counts in the CSV reader#19
karthikeyann wants to merge 5 commits into
coral/csv-pr03-single-pass-rowsfrom
coral/csv-pr04-word-scan-validity

Uh oh!

`CSV_READER_NVBENCH` (18 configs)

`CSV_READER_QUOTING_IO_NVBENCH` (12 configs)

`CSV_READER_QUOTING_NVBENCH` (6 configs)

`CSV_READER_REALISTIC_IO_NVBENCH` (12 configs)

`CSV_READER_REALISTIC_NVBENCH` (9 configs)

`CSV_READER_SCALE_NVBENCH` (5 configs)

`CSV_READER_TYPE_INFERENCE_IO_NVBENCH` (64 configs)

`CSV_READER_TYPE_INFERENCE_NVBENCH` (32 configs)

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Uh oh!

Uh oh!

CSV reader benchmarks: PR 4/9 (word-at-a-time seek, warp-aggregated validity and type counts)

Speedup vs base (cumulative), by input source

Speedup vs parent PR (this PR's own contribution)

CSV_READER_NVBENCH (18 configs)

CSV_READER_QUOTING_IO_NVBENCH (12 configs)

CSV_READER_QUOTING_NVBENCH (6 configs)

CSV_READER_REALISTIC_IO_NVBENCH (12 configs)

CSV_READER_REALISTIC_NVBENCH (9 configs)

CSV_READER_SCALE_NVBENCH (5 configs)

CSV_READER_TYPE_INFERENCE_IO_NVBENCH (64 configs)

CSV_READER_TYPE_INFERENCE_NVBENCH (32 configs)

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

`CSV_READER_NVBENCH` (18 configs)

`CSV_READER_QUOTING_IO_NVBENCH` (12 configs)

`CSV_READER_QUOTING_NVBENCH` (6 configs)

`CSV_READER_REALISTIC_IO_NVBENCH` (12 configs)

`CSV_READER_REALISTIC_NVBENCH` (9 configs)

`CSV_READER_SCALE_NVBENCH` (5 configs)

`CSV_READER_TYPE_INFERENCE_IO_NVBENCH` (64 configs)

`CSV_READER_TYPE_INFERENCE_NVBENCH` (32 configs)