[FEA] Use libcudf's boolean mask scattering for setitem #8667
Labels
feature request
New feature or request
Performance
Performance related issue
Python
Affects Python cuDF API.
Milestone
Comments
vyasr
added
feature request
vyasr
added
Python
|
This issue has been labeled |
This was referenced Jan 28, 2022
rapids-bot bot
pushed a commit
that referenced
this issue
Feb 23, 2022
closes #8667 This PR rewrites `column.__setitem__` and calls `boolean_mask_scatter` if keys and values meet some criteria. Benchmark shows in low-order problem size (10K ish), there are 30% speed up for aligned values and 10% for unaligned values. Note standard deviation of the unaligned case is quite high after refactor. For larger problem sizes performance is rather unaffected. Benchmarks: <details> <summary>10K</summary> ``` ---------------------------------------------------------------- benchmark 'boolean_mask_col_aligned': 2 tests ----------------------------------------------------------------- Name (time in ms) Min Max Mean StdDev Median IQR Outliers Rounds -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- column_setitem[boolean_mask_col_aligned] (afte) 1.2809 (1.0) 1.6781 (1.0) 1.3064 (1.0) 0.0364 (1.0) 1.2996 (1.0) 0.0137 (1.0) 22;34 761 column_setitem[boolean_mask_col_aligned] (befo) 1.7024 (1.33) 2.3863 (1.42) 1.7270 (1.32) 0.0523 (1.43) 1.7187 (1.32) 0.0138 (1.01) 20;31 563 -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------- benchmark 'boolean_mask_col_unaligned': 2 tests -------------------------------------------------------------------------- Name (time in us) Min Max Mean StdDev Median IQR Outliers Rounds ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- column_setitem[boolean_mask_col_unaligned] (afte) 972.3390 (1.0) 1,520.7559 (1.29) 1,008.6033 (1.0) 77.0920 (9.45) 984.0354 (1.0) 13.2429 (1.45) 83;132 984 column_setitem[boolean_mask_col_unaligned] (befo) 1,106.3821 (1.14) 1,179.5689 (1.0) 1,118.7759 (1.11) 8.1539 (1.0) 1,116.2200 (1.13) 9.1530 (1.0) 149;30 874 ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------- benchmark 'boolean_mask_scalar': 2 tests --------------------------------------------------------------------- Name (time in us) Min Max Mean StdDev Median IQR Outliers Rounds ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ column_setitem[boolean_mask_scalar] (afte) 532.2532 (1.0) 605.4689 (1.0) 542.2607 (1.0) 11.1111 (1.10) 537.9058 (1.0) 5.2921 (1.0) 178;213 1461 column_setitem[boolean_mask_scalar] (befo) 770.1530 (1.45) 863.1549 (1.43) 781.4038 (1.44) 10.0834 (1.0) 778.3370 (1.45) 7.0044 (1.32) 114;90 1219 ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ----------------------------------------------------------------- benchmark 'integer_scatter_map_col': 2 tests ---------------------------------------------------------------- Name (time in ms) Min Max Mean StdDev Median IQR Outliers Rounds ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- column_setitem[integer_scatter_map_col] (afte) 1.4785 (1.0) 1.9170 (1.21) 1.5176 (1.01) 0.0438 (3.91) 1.5098 (1.00) 0.0171 (1.37) 18;26 644 column_setitem[integer_scatter_map_col] (befo) 1.4882 (1.01) 1.5802 (1.0) 1.5084 (1.0) 0.0112 (1.0) 1.5068 (1.0) 0.0124 (1.0) 140;22 650 ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------- benchmark 'integer_scatter_map_scalar': 2 tests ---------------------------------------------------------------------- Name (time in us) Min Max Mean StdDev Median IQR Outliers Rounds --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- column_setitem[integer_scatter_map_scalar] (afte) 878.2479 (1.0) 1,343.0519 (1.39) 897.1496 (1.01) 27.5109 (3.23) 892.9770 (1.00) 7.8208 (1.0) 29;60 1074 column_setitem[integer_scatter_map_scalar] (befo) 879.3280 (1.00) 966.9410 (1.0) 890.8573 (1.0) 8.5287 (1.0) 888.7504 (1.0) 8.4981 (1.09) 171;50 1086 --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------- benchmark 'stride-1_slice_col': 2 tests ----------------------------------------------------------------------- Name (time in us) Min Max Mean StdDev Median IQR Outliers Rounds -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- column_setitem[stride-1_slice_col] (afte) 752.6411 (1.0) 852.3620 (1.0) 775.1790 (1.0) 10.9726 (1.0) 772.6796 (1.0) 14.1604 (1.32) 245;23 1152 column_setitem[stride-1_slice_col] (befo) 974.8179 (1.30) 1,307.2360 (1.53) 991.1287 (1.28) 25.8696 (2.36) 985.5330 (1.28) 10.6919 (1.0) 30;51 763 -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------- benchmark 'stride-1_slice_scalar': 2 tests ------------------------------------------------------------------- Name (time in us) Min Max Mean StdDev Median IQR Outliers Rounds ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- column_setitem[stride-1_slice_scalar] (afte) 87.3711 (1.16) 134.7861 (1.0) 89.7566 (1.15) 2.4861 (1.0) 89.5601 (1.16) 1.6061 (1.13) 95;87 2106 column_setitem[stride-1_slice_scalar] (befo) 75.3789 (1.0) 136.7659 (1.01) 78.0297 (1.0) 4.8186 (1.94) 77.0842 (1.0) 1.4230 (1.0) 122;184 2403 ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------- benchmark 'stride-2_slice_col': 2 tests ---------------------------------------------------------------------- Name (time in us) Min Max Mean StdDev Median IQR Outliers Rounds ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ column_setitem[stride-2_slice_col] (afte) 684.7882 (1.02) 972.5131 (1.01) 712.5983 (1.04) 54.4972 (3.04) 693.8996 (1.02) 10.3808 (1.18) 109;163 1338 column_setitem[stride-2_slice_col] (befo) 672.3758 (1.0) 964.4001 (1.0) 684.2917 (1.0) 17.9163 (1.0) 679.5955 (1.0) 8.7875 (1.0) 85;106 1368 ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ --------------------------------------------------------------------- benchmark 'stride-2_slice_scalar': 2 tests -------------------------------------------------------------------- Name (time in us) Min Max Mean StdDev Median IQR Outliers Rounds ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- column_setitem[stride-2_slice_scalar] (afte) 302.0421 (1.04) 374.3470 (1.0) 307.5677 (1.04) 4.3768 (1.0) 306.4690 (1.04) 2.4854 (1.17) 258;253 2532 column_setitem[stride-2_slice_scalar] (befo) 290.4800 (1.0) 378.1999 (1.01) 295.3950 (1.0) 4.6778 (1.07) 294.1729 (1.0) 2.1292 (1.0) 273;324 2977 ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ``` </details> <details> <summary> 1M </summary> ``` ------------------------------ benchmark 'boolean_mask_col_aligned': 2 tests ------------------------------ Name (time in ms) Min Max Mean ----------------------------------------------------------------------------------------------------------- column_setitem[boolean_mask_col_aligned] (afte) 75.3847 (1.0) 79.0559 (1.0) 76.2878 (1.0) column_setitem[boolean_mask_col_aligned] (befo) 76.3708 (1.01) 79.7394 (1.01) 77.0892 (1.01) ----------------------------------------------------------------------------------------------------------- ------------------------------ benchmark 'boolean_mask_col_unaligned': 2 tests ------------------------------ Name (time in ms) Min Max Mean ------------------------------------------------------------------------------------------------------------- column_setitem[boolean_mask_col_unaligned] (afte) 46.5199 (1.0) 48.3434 (1.0) 46.9222 (1.0) column_setitem[boolean_mask_col_unaligned] (befo) 46.6314 (1.00) 48.5938 (1.01) 47.1492 (1.00) ------------------------------------------------------------------------------------------------------------- ------------------------------ benchmark 'boolean_mask_scalar': 2 tests ------------------------------ Name (time in ms) Min Max Mean ------------------------------------------------------------------------------------------------------ column_setitem[boolean_mask_scalar] (afte) 17.0548 (1.0) 17.8006 (1.0) 17.5430 (1.0) column_setitem[boolean_mask_scalar] (befo) 18.4329 (1.08) 18.6918 (1.05) 18.5073 (1.05) ------------------------------------------------------------------------------------------------------ -------------------------------- benchmark 'integer_scatter_map_col': 2 tests ------------------------------- Name (time in ms) Min Max Mean ------------------------------------------------------------------------------------------------------------- column_setitem[integer_scatter_map_col] (afte) 115.7189 (1.01) 120.0585 (1.0) 116.6452 (1.0) column_setitem[integer_scatter_map_col] (befo) 114.7481 (1.0) 122.7263 (1.02) 117.5000 (1.01) ------------------------------------------------------------------------------------------------------------- ------------------------------ benchmark 'integer_scatter_map_scalar': 2 tests ------------------------------ Name (time in ms) Min Max Mean ------------------------------------------------------------------------------------------------------------- column_setitem[integer_scatter_map_scalar] (afte) 57.9951 (1.0) 62.2284 (1.0) 59.8864 (1.0) column_setitem[integer_scatter_map_scalar] (befo) 60.9071 (1.05) 62.2952 (1.00) 61.6422 (1.03) ------------------------------------------------------------------------------------------------------------- ------------------------------ benchmark 'stride-1_slice_col': 2 tests ------------------------------ Name (time in ms) Min Max Mean ----------------------------------------------------------------------------------------------------- column_setitem[stride-1_slice_col] (afte) 56.9203 (1.0) 58.0924 (1.0) 57.4940 (1.0) column_setitem[stride-1_slice_col] (befo) 58.1888 (1.02) 59.2996 (1.02) 58.5722 (1.02) ----------------------------------------------------------------------------------------------------- -------------------------------- benchmark 'stride-1_slice_scalar': 2 tests ------------------------------- Name (time in us) Min Max Mean ----------------------------------------------------------------------------------------------------------- column_setitem[stride-1_slice_scalar] (afte) 287.1200 (1.08) 515.4130 (1.24) 298.1191 (1.09) column_setitem[stride-1_slice_scalar] (befo) 265.6982 (1.0) 415.9641 (1.0) 273.5206 (1.0) ----------------------------------------------------------------------------------------------------------- ------------------------------ benchmark 'stride-2_slice_col': 2 tests ------------------------------ Name (time in ms) Min Max Mean ----------------------------------------------------------------------------------------------------- column_setitem[stride-2_slice_col] (afte) 29.1543 (1.01) 31.0217 (1.00) 29.6341 (1.00) column_setitem[stride-2_slice_col] (befo) 28.9718 (1.0) 30.8824 (1.0) 29.5045 (1.0) ----------------------------------------------------------------------------------------------------- --------------------------------- benchmark 'stride-2_slice_scalar': 2 tests -------------------------------- Name (time in us) Min Max Mean ------------------------------------------------------------------------------------------------------------- column_setitem[stride-2_slice_scalar] (afte) 780.7089 (1.00) 1,407.5749 (1.0) 817.0583 (1.0) column_setitem[stride-2_slice_scalar] (befo) 777.0571 (1.0) 2,036.4628 (1.45) 832.0608 (1.02) ------------------------------------------------------------------------------------------------------------- ``` </details> Authors: - Michael Wang (https://github.com/isVoid) Approvers: - Bradley Dice (https://github.com/bdice) - Vyas Ramasubramani (https://github.com/vyasr) URL: #10202
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Is your feature request related to a problem? Please describe.
#3681 introduced a directly boolean scattering API that could be leveraged for setting items on columns using a boolean mask in a single pass. The libcudf feature has been exposed in Cython (in
_lib/copying.pyx) but cuDF currently still implements that feature manually.Describe the solution you'd like
We should change cuDF to take advantage of the direct boolean scattering approach.
Additional context
@rgsl888prabhu may have more insight into how this should be accomplished.
The text was updated successfully, but these errors were encountered: