Add strip_delimiters option to read_text

[upsj]

Description

This adds a strip_delimiters post-processing option to read_text. I needed to implement some lightweight striping because a thread-per-row parallelization of the string gather gave pretty bad performance.

For consistency, I also removed the special-case handling of delimiters at the end (previously adding an empty row), to match the read_csv behavior.

Benchmark results:

benchmarks/MULTIBYTE_SPLIT_NVBENCH --axis size_approx[pow2]=30 --axis byte_range_percent=100 --axis T=device --axis delim_size=4

[0] Tesla T4

T	strip_delimiters	delim_percent	size_approx	CPU Time	Noise	Peak Memory Usage	Encoded file size
device	0	1	2^30 = 1073741824	178.133 ms	0.36%	3.709 GiB	1014.442 MiB
device	1	1	2^30 = 1073741824	188.328 ms	0.31%	4.690 GiB	1014.442 MiB
device	0	25	2^30 = 1073741824	206.188 ms	0.03%	5.292 GiB	953.075 MiB
device	1	25	2^30 = 1073741824	242.534 ms	0.50%	5.975 GiB	953.075 MiB

Closes #11625

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[davidwendt]

... thread-per-row parallelization of the string gather gave pretty bad performance

Can you elaborate on this? I think there is a fast strings gather that may be possible to use here.

[codecov]

Codecov Report

Base: 87.40% // Head: 88.15% // Increases project coverage by +0.74% 🎉

Coverage data is based on head (ca5568a) compared to base (f72c4ce).
Patch has no changes to coverable lines.

Additional details and impacted files

@@               Coverage Diff                @@
##           branch-22.12   #11946      +/-   ##
================================================
+ Coverage         87.40%   88.15%   +0.74%     
================================================
  Files               133      133              
  Lines             21833    21995     +162     
================================================
+ Hits              19084    19389     +305     
+ Misses             2749     2606     -143     

Impacted Files	Coverage Δ
python/strings_udf/strings_udf/__init__.py	86.27% <0.00%> (-10.61%)	⬇️
python/cudf/cudf/io/text.py	91.66% <0.00%> (-8.34%)	⬇️
python/cudf/cudf/core/_base_index.py	82.20% <0.00%> (-3.35%)	⬇️
python/strings_udf/strings_udf/_typing.py	94.73% <0.00%> (-1.06%)	⬇️
python/cudf/cudf/testing/dataset_generator.py	72.83% <0.00%> (-0.42%)	⬇️
python/dask_cudf/dask_cudf/backends.py	84.90% <0.00%> (-0.37%)	⬇️
python/dask_cudf/dask_cudf/core.py	73.92% <0.00%> (-0.21%)	⬇️
python/cudf/cudf/io/orc.py	92.94% <0.00%> (-0.09%)	⬇️
python/cudf/cudf/__init__.py	90.69% <0.00%> (ø)
python/cudf/cudf/core/udf/_ops.py	100.00% <0.00%> (ø)
... and 23 more

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[upsj]

Can you elaborate on this? I think there is a fast strings gather that may be possible to use here.

When running for_each_n with one thread per row copying the elements excluding the delimiter, I got runtimes around 1s instead of 20 ms with the current solution. NSight Compute profiles showed horrible cache utilization numbers. cudf::strings::detail::gather looks almost right - I need to transform individual strings to strip off a fixed number of characters (except for the last row), not gather a subset of strings from a larger column. Maybe the word gather is a bit misleading. But the underlying kernels look pretty close to what I need.

[davidwendt]

Can you elaborate on this? I think there is a fast strings gather that may be possible to use here.

When running for_each_n with one thread per row copying the elements excluding the delimiter, I got runtimes around 1s instead of 20 ms with the current solution. NSight Compute profiles showed horrible cache utilization numbers. cudf::strings::detail::gather looks almost right - I need to transform individual strings to strip off a fixed number of characters (except for the last row), not gather a subset of strings from a larger column. Maybe the word gather is a bit misleading. But the underlying kernels look pretty close to what I need.

Ok. I was thinking more along the lines of this make_strings_column function that can take a device-span of string_view (or equivalent thrust::pair): https://docs.rapids.ai/api/libcudf/stable/group__column__factories.html#ga993941cbf14270bcea2cc95427996de1

It would just be a matter of building a device-uvector of these to call one of these factory functions which has a highly tuned gather operation for building a strings column from individual strings in device memory.

For reference, both of these factory functions (string_view and thrust_pair) call into

cudf/cpp/include/cudf/strings/detail/strings_column_factories.cuh

Lines 72 to 76 in 6ca2ceb

	template <typename IndexPairIterator>
	std::unique_ptr<column> make_strings_column(IndexPairIterator begin,
	IndexPairIterator end,
	rmm::cuda_stream_view stream,
	rmm::mr::device_memory_resource* mr)

[upsj]

@davidwendt thanks for the details, that shaved another 18ms off the runtime for the long string case (at the cost of maybe 20 ms for the short string case, but I'll take the added simplicity :) )

[upsj]

rerun tests

[upsj]

rerun tests

[upsj]

@gpucibot merge