You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@doris.apache.org by "wumeibanfa (via GitHub)" <gi...@apache.org> on 2025/12/28 07:24:55 UTC

[PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

wumeibanfa opened a new pull request, #59426:
URL: https://github.com/apache/doris/pull/59426

   ### What problem does this PR solve?
   
   Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics
   
   ### Release note
   
   None
   
   ### Check List (For Author)
   
   - Test <!-- At least one of them must be included. -->
       - [ ] Regression test
       - [x] Unit Test
       - [ ] Manual test (add detailed scripts or steps below)
       - [ ] No need to test or manual test. Explain why:
           - [ ] This is a refactor/code format and no logic has been changed.
           - [ ] Previous test can cover this change.
           - [ ] No code files have been changed.
           - [ ] Other reason <!-- Add your reason?  -->
   
   - Behavior changed:
       - [x] No.
       - [ ] Yes. <!-- Explain the behavior change -->
   
   - Does this need documentation?
       - [x] No.
       - [ ] Yes. <!-- Add document PR link here. eg: https://github.com/apache/doris-website/pull/1214 -->
   
   ### Check List (For Reviewer who merge this PR)
   
   - [ ] Confirm the release note
   - [ ] Confirm test cases
   - [ ] Confirm document
   - [ ] Add branch pick label <!-- Add branch pick label that this PR should merge into -->
   
   


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "wumeibanfa (via GitHub)" <gi...@apache.org>.
wumeibanfa commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-3694531983

   don't use simd
   ```text
   Running ./be/output/lib/benchmark_test
   Run on (16 X 200 MHz CPU s)
   CPU Caches:
     L1 Data 64 KiB (x16)
     L1 Instruction 64 KiB (x16)
     L2 Unified 512 KiB (x16)
     L3 Unified 32768 KiB (x1)
   Load Average: 0.00, 0.00, 0.12
   ---------------------------------------------------------------------------------------------------------------
   Benchmark                                                     Time             CPU   Iterations UserCounters...
   ---------------------------------------------------------------------------------------------------------------
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_mean           3.92 ns         3.92 ns            5 bytes_per_second=7.61044G/s
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_median         3.91 ns         3.91 ns            5 bytes_per_second=7.6193G/s
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_stddev        0.011 ns        0.011 ns            5 bytes_per_second=22.2868M/s
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_cv             0.29 %          0.29 %             5 bytes_per_second=0.29%
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_mean           7.86 ns         7.86 ns            5 bytes_per_second=7.58778G/s
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_median         7.84 ns         7.84 ns            5 bytes_per_second=7.60375G/s
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_stddev        0.038 ns        0.038 ns            5 bytes_per_second=36.9531M/s
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_cv             0.48 %          0.48 %             5 bytes_per_second=0.48%
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_mean          31.4 ns         31.4 ns            5 bytes_per_second=7.60191G/s
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_median        31.3 ns         31.3 ns            5 bytes_per_second=7.61015G/s
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_stddev       0.077 ns        0.077 ns            5 bytes_per_second=19.1616M/s
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_cv            0.25 %          0.25 %             5 bytes_per_second=0.25%
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_mean          125 ns          125 ns            5 bytes_per_second=7.61962G/s
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_median        125 ns          125 ns            5 bytes_per_second=7.6198G/s
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_stddev      0.008 ns        0.008 ns            5 bytes_per_second=481.167k/s
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_cv           0.01 %          0.01 %             5 bytes_per_second=0.01%
   ```


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "wumeibanfa (via GitHub)" <gi...@apache.org>.
wumeibanfa commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-3694532404

   use simd + handle 32 byte one time
   ```
   Running ./be/output/lib/benchmark_test
   Run on (16 X 200 MHz CPU s)
   CPU Caches:
     L1 Data 64 KiB (x16)
     L1 Instruction 64 KiB (x16)
     L2 Unified 512 KiB (x16)
     L3 Unified 32768 KiB (x1)
   Load Average: 0.96, 1.81, 4.70
   ---------------------------------------------------------------------------------------------------------------
   Benchmark                                                     Time             CPU   Iterations UserCounters...
   ---------------------------------------------------------------------------------------------------------------
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_mean           1.09 ns         1.09 ns            5 bytes_per_second=27.3451G/s
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_median         1.09 ns         1.09 ns            5 bytes_per_second=27.3452G/s
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_stddev        0.000 ns        0.000 ns            5 bytes_per_second=2.48125M/s
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_cv             0.01 %          0.01 %             5 bytes_per_second=0.01%
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_mean           2.18 ns         2.18 ns            5 bytes_per_second=27.3575G/s
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_median         2.18 ns         2.18 ns            5 bytes_per_second=27.3937G/s
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_stddev        0.005 ns        0.005 ns            5 bytes_per_second=65.1549M/s
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_cv             0.23 %          0.23 %             5 bytes_per_second=0.23%
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_mean          8.67 ns         8.67 ns            5 bytes_per_second=27.5142G/s
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_median        8.67 ns         8.66 ns            5 bytes_per_second=27.5161G/s
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_stddev       0.001 ns        0.001 ns            5 bytes_per_second=4.84825M/s
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_cv            0.02 %          0.02 %             5 bytes_per_second=0.02%
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_mean         34.7 ns         34.7 ns            5 bytes_per_second=27.4944G/s
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_median       34.7 ns         34.7 ns            5 bytes_per_second=27.4981G/s
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_stddev      0.011 ns        0.011 ns            5 bytes_per_second=8.69048M/s
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_cv           0.03 %          0.03 %             5 bytes_per_second=0.03%
   ```


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "hello-stephen (via GitHub)" <gi...@apache.org>.
hello-stephen commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-3694669923

   # BE Regression && UT Coverage Report
   Increment line coverage `100% (0/0)` :tada:
   
   [Increment coverage report](http://coverage.selectdb-in.cc/coverage/59426_d860fdfa2cd9ae99881222ad8502aa83759542ae_merge/increment_report/index.html)
   [Complete coverage report](http://coverage.selectdb-in.cc/coverage/59426_d860fdfa2cd9ae99881222ad8502aa83759542ae_merge/report/index.html)
   | Category          | Coverage           |
   |-------------------|--------------------|
   | Function Coverage | 72.21% (25051/34694) |
   | Line Coverage     | 58.92% (263069/446490) |
   | Region Coverage   | 53.74% (218281/406174) |
   | Branch Coverage   | 55.36% (93831/169504) |


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "doris-robot (via GitHub)" <gi...@apache.org>.
doris-robot commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-3694580820

   
   <details>
   <summary>TPC-DS: <b>Total hot run time: 178933 ms</b></summary>
   
   ```
   machine: 'aliyun_ecs.c7a.8xlarge_32C64G'
   scripts: https://github.com/apache/doris/tree/master/tools/tpcds-tools
   TPC-DS sf100 test result on commit d860fdfa2cd9ae99881222ad8502aa83759542ae, data reload: false
   
   query5	4346	582	464	464
   query6	353	243	263	243
   query7	4232	477	279	279
   query8	309	271	253	253
   query9	8769	2553	2514	2514
   query10	515	378	349	349
   query11	15524	14909	14727	14727
   query12	175	119	118	118
   query13	1273	497	393	393
   query14	5503	3093	2744	2744
   query14_1	2664	2608	2682	2608
   query15	218	203	181	181
   query16	815	472	435	435
   query17	1161	714	609	609
   query18	2452	454	389	389
   query19	247	232	212	212
   query20	123	118	120	118
   query21	220	146	117	117
   query22	3964	4119	3841	3841
   query23	16703	16193	16167	16167
   query23_1	16239	16174	16067	16067
   query24	7390	1685	1233	1233
   query24_1	1219	1213	1226	1213
   query25	563	473	431	431
   query26	1247	274	160	160
   query27	2769	471	301	301
   query28	4498	2132	2157	2132
   query29	807	537	446	446
   query30	309	250	215	215
   query31	837	699	629	629
   query32	78	80	66	66
   query33	529	333	283	283
   query34	890	903	541	541
   query35	784	852	714	714
   query36	868	913	825	825
   query37	133	91	76	76
   query38	2933	3107	2901	2901
   query39	773	765	742	742
   query39_1	726	704	691	691
   query40	228	136	120	120
   query41	65	64	64	64
   query42	109	103	102	102
   query43	442	432	399	399
   query44	1340	735	739	735
   query45	193	191	178	178
   query46	875	970	610	610
   query47	1693	1731	1636	1636
   query48	312	326	248	248
   query49	620	434	349	349
   query50	661	300	215	215
   query51	3813	3913	3754	3754
   query52	104	106	98	98
   query53	320	361	286	286
   query54	306	262	246	246
   query55	77	77	74	74
   query56	296	304	292	292
   query57	1191	1161	1093	1093
   query58	285	253	279	253
   query59	2390	2476	2400	2400
   query60	317	313	293	293
   query61	165	165	158	158
   query62	733	763	648	648
   query63	325	293	302	293
   query64	5020	1309	1010	1010
   query65	4069	3957	3937	3937
   query66	1463	435	329	329
   query67	15169	15176	14855	14855
   query68	3500	1031	728	728
   query69	493	348	307	307
   query70	1067	967	952	952
   query71	345	298	272	272
   query72	6100	5012	5173	5012
   query73	725	674	316	316
   query74	8943	8977	8752	8752
   query75	3204	3149	2815	2815
   query76	3740	1133	741	741
   query77	519	395	298	298
   query78	9456	9396	8783	8783
   query79	1082	873	601	601
   query80	964	645	552	552
   query81	524	266	238	238
   query82	420	137	107	107
   query83	267	250	251	250
   query84	255	124	104	104
   query85	913	512	480	480
   query86	341	321	278	278
   query87	3163	3184	3066	3066
   query88	3269	2322	2314	2314
   query89	470	429	387	387
   query90	2031	184	160	160
   query91	193	184	161	161
   query92	73	70	65	65
   query93	1055	889	558	558
   query94	566	319	300	300
   query95	598	392	314	314
   query96	594	473	215	215
   query97	2293	2317	2235	2235
   query98	209	196	199	196
   query99	1362	1357	1259	1259
   Total cold run time: 254460 ms
   Total hot run time: 178933 ms
   ```
   </details>
   


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "github-actions[bot] (via GitHub)" <gi...@apache.org>.
github-actions[bot] closed pull request #59426: [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics
URL: https://github.com/apache/doris/pull/59426


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "wumeibanfa (via GitHub)" <gi...@apache.org>.
wumeibanfa commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-3694536456

   run buildall


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "Thearas (via GitHub)" <gi...@apache.org>.
Thearas commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-3694530083

   
   Thank you for your contribution to Apache Doris.
   Don't know what should be done next? See [How to process your PR](https://cwiki.apache.org/confluence/display/DORIS/How+to+process+your+PR).
   
   Please clearly describe your PR:
   1. What problem was fixed (it's best to include specific error reporting information). How it was fixed.
   2. Which behaviors were modified. What was the previous behavior, what is it now, why was it modified, and what possible impacts might there be.
   3. What features were added. Why was this function added?
   4. Which code was refactored and why was this part of the code refactored?
   5. Which functions were optimized and what is the difference before and after the optimization?
       


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "wumeibanfa (via GitHub)" <gi...@apache.org>.
wumeibanfa commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-3694532699

   use simd + handle 16 byte one time
   ```text
   Running ./be/output/lib/benchmark_test
   Run on (16 X 200 MHz CPU s)
   CPU Caches:
     L1 Data 64 KiB (x16)
     L1 Instruction 64 KiB (x16)
     L2 Unified 512 KiB (x16)
     L3 Unified 32768 KiB (x1)
   Load Average: 1.46, 1.21, 3.47
   ---------------------------------------------------------------------------------------------------------------
   Benchmark                                                     Time             CPU   Iterations UserCounters...
   ---------------------------------------------------------------------------------------------------------------
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_mean           2.77 ns         2.77 ns            5 bytes_per_second=10.7547G/s
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_median         2.77 ns         2.77 ns            5 bytes_per_second=10.7616G/s
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_stddev        0.005 ns        0.005 ns            5 bytes_per_second=18.5077M/s
   BM_bytes32_mask_to_bits32_mask/32/repeats:5_cv             0.17 %          0.17 %             5 bytes_per_second=0.17%
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_mean           5.54 ns         5.54 ns            5 bytes_per_second=10.7593G/s
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_median         5.54 ns         5.54 ns            5 bytes_per_second=10.7676G/s
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_stddev        0.010 ns        0.010 ns            5 bytes_per_second=20.5456M/s
   BM_bytes32_mask_to_bits32_mask/64/repeats:5_cv             0.19 %          0.19 %             5 bytes_per_second=0.19%
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_mean          22.2 ns         22.2 ns            5 bytes_per_second=10.7202G/s
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_median        22.2 ns         22.2 ns            5 bytes_per_second=10.7226G/s
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_stddev       0.012 ns        0.012 ns            5 bytes_per_second=5.81638M/s
   BM_bytes32_mask_to_bits32_mask/256/repeats:5_cv            0.05 %          0.05 %             5 bytes_per_second=0.05%
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_mean         88.8 ns         88.8 ns            5 bytes_per_second=10.7428G/s
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_median       88.8 ns         88.8 ns            5 bytes_per_second=10.7441G/s
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_stddev      0.050 ns        0.051 ns            5 bytes_per_second=6.29135M/s
   BM_bytes32_mask_to_bits32_mask/1024/repeats:5_cv           0.06 %          0.06 %             5 bytes_per_second=0.06%
   ```


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "doris-robot (via GitHub)" <gi...@apache.org>.
doris-robot commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-3694598397

   # BE UT Coverage Report
   Increment line coverage ` ` :tada:
   
   [Increment coverage report](http://coverage.selectdb-in.cc/coverage/d860fdfa2cd9ae99881222ad8502aa83759542ae_d860fdfa2cd9ae99881222ad8502aa83759542ae/increment_report/index.html)
   [Complete coverage report](http://coverage.selectdb-in.cc/coverage/d860fdfa2cd9ae99881222ad8502aa83759542ae_d860fdfa2cd9ae99881222ad8502aa83759542ae/report/index.html)
   | Category          | Coverage           |
   |-------------------|--------------------|
   | Function Coverage | 53.41% (18953/35489) |
   | Line Coverage     | 39.27% (175818/447682) |
   | Region Coverage   | 33.85% (136068/402009) |
   | Branch Coverage   | 34.78% (58756/168940) |


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "github-actions[bot] (via GitHub)" <gi...@apache.org>.
github-actions[bot] commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-4823418716

   We're closing this PR because it hasn't been updated in a while.
   This isn't a judgement on the merit of the PR in any way. It's just a way of keeping the PR queue manageable.
   If you'd like to revive this PR, please reopen it and feel free a maintainer to remove the Stale tag!


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "wumeibanfa (via GitHub)" <gi...@apache.org>.
wumeibanfa commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-3694536117

   # BM_bytes32_mask_to_bits32_mask Performance Evaluation
   
   ## Benchmark Results
   
   ### BM_bytes32_mask_to_bits32_mask
   
   | Size (bytes) | Origin (ns) | SVE (ns) | Improvement |
   |--------------|-------------|----------|-------------|
   | 32           | 3.92        | 1.09     | +72.2%      |
   | 64           | 7.86        | 2.18     | +72.3%      |
   | 256          | 31.4        | 8.67     | +72.4%      |
   | 1024         | 125         | 34.7     | +72.2%      |
   
   > Improvement = (Origin − SVE) / Origin
   
   ---
   
   ## Throughput Comparison
   
   - **Origin implementation**  
     ~7.6 GB/s (stable across all sizes)
   
   - **SVE implementation**  
     ~27.4–27.5 GB/s (stable across all sizes)
   
   This represents an **approximately 3.6× throughput increase** when using SVE.
   
   ## Summary
   
   SVE optimization delivers a **substantial and consistent performance improvement** for `bytes32_mask_to_bits32_mask`:
   
   - **~72% latency reduction**
   - **~3.6× throughput increase**
   ---
   
   


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "doris-robot (via GitHub)" <gi...@apache.org>.
doris-robot commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-3694583483

   
   <details>
   <summary>ClickBench: <b>Total hot run time: 27.1 s</b></summary>
   
   ```
   machine: 'aliyun_ecs.c7a.8xlarge_32C64G'
   scripts: https://github.com/apache/doris/tree/master/tools/clickbench-tools
   ClickBench test result on commit d860fdfa2cd9ae99881222ad8502aa83759542ae, data reload: false
   
   query1	0.05	0.04	0.04
   query2	0.09	0.05	0.05
   query3	0.26	0.09	0.08
   query4	1.61	0.11	0.12
   query5	0.27	0.25	0.26
   query6	1.17	0.63	0.63
   query7	0.03	0.03	0.03
   query8	0.05	0.05	0.04
   query9	0.58	0.51	0.51
   query10	0.56	0.55	0.56
   query11	0.16	0.11	0.10
   query12	0.15	0.11	0.11
   query13	0.61	0.60	0.61
   query14	1.00	0.97	0.98
   query15	0.82	0.80	0.81
   query16	0.41	0.40	0.40
   query17	1.03	1.00	1.05
   query18	0.23	0.22	0.21
   query19	1.97	1.75	1.75
   query20	0.02	0.01	0.01
   query21	15.45	0.30	0.14
   query22	4.89	0.06	0.05
   query23	16.04	0.27	0.11
   query24	1.36	0.28	0.61
   query25	0.11	0.09	0.07
   query26	0.15	0.13	0.13
   query27	0.05	0.06	0.07
   query28	4.05	1.23	1.03
   query29	12.64	3.96	3.14
   query30	0.29	0.14	0.12
   query31	2.82	0.63	0.38
   query32	3.24	0.54	0.46
   query33	2.98	3.00	2.99
   query34	16.78	5.20	4.51
   query35	4.55	4.56	4.54
   query36	0.65	0.50	0.48
   query37	0.11	0.07	0.06
   query38	0.07	0.05	0.04
   query39	0.04	0.04	0.03
   query40	0.17	0.14	0.14
   query41	0.10	0.03	0.03
   query42	0.04	0.03	0.03
   query43	0.04	0.04	0.04
   Total cold run time: 97.69 s
   Total hot run time: 27.1 s
   ```
   </details>
   


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "Copilot (via GitHub)" <gi...@apache.org>.
Copilot commented on code in PR #59426:
URL: https://github.com/apache/doris/pull/59426#discussion_r2650135951


##########
be/src/util/simd/bits.h:
##########
@@ -81,6 +85,38 @@ inline uint32_t bytes32_mask_to_bits32_mask(const uint8_t* data) {
                       _mm_loadu_si128(reinterpret_cast<const __m128i*>(data + 16)), zero16)))
               << 16) &
              0xffff0000);
+#elif defined(__ARM_FEATURE_SVE)
+    /*
+    * Assume the input data is [1 (repeated 16 times), 0 (repeated 16 times)].
+    *
+    * 1. Comparison produces a predicate mask: is_one = [1 (repeated 16 times), 0 (repeated 16 times)].
+    *
+    * 2. Construct a new vector: masked = [1, 2, 4, 8, 16, 32, 64, 128,1, 2, 4, 8, 16, 32, 64, 128, 0 (repeated 16 times)].
+    * For each position, if the corresponding element in is_one is 1, take the value from the mask array (mask_arr); otherwise, take 0.
+    *
+    * 3. Rearorder the elements: res = [1, 1, 0, 0, 2, 2, 0, 0, 4, 4, 0, 0, ...], where the index mapping is: j = (i % 4) * 8 + (i / 4).
+    *
+    * 4. Perform a horizontal sum to obtain the final result:
+    *    mask = sum_{i=0}^{7} ( res[i*4] + (res[i*4 + 1] << 8) + (res[i*4 + 2] << 16) + (res[i*4 + 3] << 24)).
+    *    This works because we use svreinterpret_u32_u8, treating every group of four u8 elements as a single u32.
+    */
+    static constexpr uint8_t mask_arr[32] = {1,  2,   4,  8,  16,  32, 64, 128, 1,  2,  4,
+                                             8,  16,  32, 64, 128, 1,  2,  4,   8,  16, 32,
+                                             64, 128, 1,  2,  4,   8,  16, 32,  64, 128};
+    svbool_t pg = svwhilelt_b8(0, 32);
+    svuint8_t mask_vec = svld1_u8(pg, mask_arr);
+    svbool_t is_one = svcmpeq_n_u8(pg, svld1_u8(pg, data), 1);
+    svuint8_t masked = svsel_u8(is_one, mask_vec, svdup_u8(0));
+
+    // j = i % 4 * 8 + i / 4

Review Comment:
   The inline comment "j = i % 4 * 8 + i / 4" lacks clarity on operator precedence. Consider adding parentheses for clarity: "j = (i % 4) * 8 + i / 4" to match the computation in the code.
   ```suggestion
       // j = (i % 4) * 8 + i / 4
   ```



##########
be/benchmark/benchmark_bit_simd.hpp:
##########
@@ -0,0 +1,114 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#include <benchmark/benchmark.h>
+#include <vector>
+#include <random>
+
+#include "util/simd/bits.h"
+
+namespace doris {
+
+static void create_test_data(std::vector<uint8_t> &data, size_t size) {
+    std::mt19937 rng(123);
+    for (int i = 0; i < size; ++i) {
+        data[i] = rng() & 1;
+    }
+}
+
+inline uint32_t bitmask_8(const uint8_t* data) {
+    static constexpr uint8_t mask_arr[16] = {1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4, 8, 16, 32, 64, 128};
+    svbool_t pg = svwhilelt_b8(0, 16);
+    svuint8_t mask_vec = svld1_u8(pg, mask_arr);
+
+    svbool_t nonzero = svcmpne_n_u8(pg, svld1_u8(pg, data), 0);
+    svuint8_t masked = svsel_u8(nonzero, mask_vec, svdup_u8(0));
+
+    svuint8_t idx = svindex_u8(0, 1);
+    svuint8_t rev_idx = svsub_u8_x(pg, svdup_u8(15), idx);
+    svuint8_t res_hi = svtbl_u8(masked, rev_idx);
+
+    uint32_t mask = static_cast<uint32_t>(svaddv_u16(pg, svreinterpret_u16_u8(svzip1_u8(masked, res_hi))));
+    return mask;
+}
+

Review Comment:
   The function bytes32_mask_to_bits32_mask in the benchmark file has a different implementation than the one in be/src/util/simd/bits.h, which may lead to confusion and inconsistent benchmarking results. The benchmark should either use the actual production function from simd::bytes32_mask_to_bits32_mask or clearly document why an alternative implementation is being compared.
   ```suggestion
   
   // Alternative benchmark-only implementation of converting 32 bytes of
   // masked data into a 32-bit mask. The production implementation is
   // simd::bytes32_mask_to_bits32_mask (see util/simd/bits.h). This version
   // is intentionally different and is used in BM_bytes32_mask_to_bits32_mask_2
   // to compare its performance against the production implementation.
   ```



##########
be/src/util/simd/bits.h:
##########
@@ -81,6 +85,38 @@ inline uint32_t bytes32_mask_to_bits32_mask(const uint8_t* data) {
                       _mm_loadu_si128(reinterpret_cast<const __m128i*>(data + 16)), zero16)))
               << 16) &
              0xffff0000);
+#elif defined(__ARM_FEATURE_SVE)
+    /*
+    * Assume the input data is [1 (repeated 16 times), 0 (repeated 16 times)].
+    *
+    * 1. Comparison produces a predicate mask: is_one = [1 (repeated 16 times), 0 (repeated 16 times)].
+    *
+    * 2. Construct a new vector: masked = [1, 2, 4, 8, 16, 32, 64, 128,1, 2, 4, 8, 16, 32, 64, 128, 0 (repeated 16 times)].
+    * For each position, if the corresponding element in is_one is 1, take the value from the mask array (mask_arr); otherwise, take 0.
+    *
+    * 3. Rearorder the elements: res = [1, 1, 0, 0, 2, 2, 0, 0, 4, 4, 0, 0, ...], where the index mapping is: j = (i % 4) * 8 + (i / 4).

Review Comment:
   The comment in step 3 contains a mathematical expression "j = (i % 4) * 8 + (i / 4)" where the operator precedence is ambiguous. It should be "j = ((i % 4) * 8) + (i / 4)" to clarify that the modulo and multiplication are evaluated before the division and addition, matching the actual code implementation.



##########
be/src/util/simd/bits.h:
##########
@@ -81,6 +85,38 @@ inline uint32_t bytes32_mask_to_bits32_mask(const uint8_t* data) {
                       _mm_loadu_si128(reinterpret_cast<const __m128i*>(data + 16)), zero16)))
               << 16) &
              0xffff0000);
+#elif defined(__ARM_FEATURE_SVE)
+    /*
+    * Assume the input data is [1 (repeated 16 times), 0 (repeated 16 times)].
+    *
+    * 1. Comparison produces a predicate mask: is_one = [1 (repeated 16 times), 0 (repeated 16 times)].
+    *
+    * 2. Construct a new vector: masked = [1, 2, 4, 8, 16, 32, 64, 128,1, 2, 4, 8, 16, 32, 64, 128, 0 (repeated 16 times)].
+    * For each position, if the corresponding element in is_one is 1, take the value from the mask array (mask_arr); otherwise, take 0.
+    *
+    * 3. Rearorder the elements: res = [1, 1, 0, 0, 2, 2, 0, 0, 4, 4, 0, 0, ...], where the index mapping is: j = (i % 4) * 8 + (i / 4).

Review Comment:
   Typo in the comment: "Rearorder" should be "Reorder".
   ```suggestion
       * 3. Reorder the elements: res = [1, 1, 0, 0, 2, 2, 0, 0, 4, 4, 0, 0, ...], where the index mapping is: j = (i % 4) * 8 + (i / 4).
   ```



##########
be/benchmark/benchmark_bit_simd.hpp:
##########
@@ -0,0 +1,114 @@
+// Licensed to the Apache Software Foundation (ASF) under one
+// or more contributor license agreements.  See the NOTICE file
+// distributed with this work for additional information
+// regarding copyright ownership.  The ASF licenses this file
+// to you under the Apache License, Version 2.0 (the
+// "License"); you may not use this file except in compliance
+// with the License.  You may obtain a copy of the License at
+//
+//   http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing,
+// software distributed under the License is distributed on an
+// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+// KIND, either express or implied.  See the License for the
+// specific language governing permissions and limitations
+// under the License.
+
+#include <benchmark/benchmark.h>
+#include <vector>
+#include <random>
+
+#include "util/simd/bits.h"
+
+namespace doris {
+
+static void create_test_data(std::vector<uint8_t> &data, size_t size) {
+    std::mt19937 rng(123);
+    for (int i = 0; i < size; ++i) {
+        data[i] = rng() & 1;
+    }
+}
+
+inline uint32_t bitmask_8(const uint8_t* data) {
+    static constexpr uint8_t mask_arr[16] = {1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4, 8, 16, 32, 64, 128};
+    svbool_t pg = svwhilelt_b8(0, 16);
+    svuint8_t mask_vec = svld1_u8(pg, mask_arr);
+
+    svbool_t nonzero = svcmpne_n_u8(pg, svld1_u8(pg, data), 0);
+    svuint8_t masked = svsel_u8(nonzero, mask_vec, svdup_u8(0));
+
+    svuint8_t idx = svindex_u8(0, 1);
+    svuint8_t rev_idx = svsub_u8_x(pg, svdup_u8(15), idx);
+    svuint8_t res_hi = svtbl_u8(masked, rev_idx);
+
+    uint32_t mask = static_cast<uint32_t>(svaddv_u16(pg, svreinterpret_u16_u8(svzip1_u8(masked, res_hi))));
+    return mask;
+}
+
+uint32_t bytes32_mask_to_bits32_mask(const uint8_t* data) {
+    return bitmask_8(data) | (bitmask_8(data + 16) << 16);
+}  

Review Comment:
   The benchmark file uses SVE intrinsics (svwhilelt_b8, svld1_u8, svcmpne_n_u8, etc.) without including the required header arm_sve.h and without guarding the code with __ARM_FEATURE_SVE preprocessor directive. This will cause compilation failures on systems without SVE support. The SVE-specific code should be wrapped in #if defined(__ARM_FEATURE_SVE) and the header should be conditionally included.



##########
be/src/util/simd/bits.h:
##########
@@ -81,6 +85,38 @@ inline uint32_t bytes32_mask_to_bits32_mask(const uint8_t* data) {
                       _mm_loadu_si128(reinterpret_cast<const __m128i*>(data + 16)), zero16)))
               << 16) &
              0xffff0000);
+#elif defined(__ARM_FEATURE_SVE)
+    /*
+    * Assume the input data is [1 (repeated 16 times), 0 (repeated 16 times)].
+    *
+    * 1. Comparison produces a predicate mask: is_one = [1 (repeated 16 times), 0 (repeated 16 times)].
+    *
+    * 2. Construct a new vector: masked = [1, 2, 4, 8, 16, 32, 64, 128,1, 2, 4, 8, 16, 32, 64, 128, 0 (repeated 16 times)].

Review Comment:
   There's a spacing issue in the comment. Missing space after "128," before "1" in the sequence. Should be "128, 1, 2, 4..." not "128,1, 2, 4..."
   ```suggestion
       * 2. Construct a new vector: masked = [1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4, 8, 16, 32, 64, 128, 0 (repeated 16 times)].
   ```



-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org


Re: [PR] [opt](arm) Improve performance of bytes32_mask_to_bits32_mask functions with SVE intrinsics [doris]

Posted by "doris-robot (via GitHub)" <gi...@apache.org>.
doris-robot commented on PR #59426:
URL: https://github.com/apache/doris/pull/59426#issuecomment-3694574461

   
   <details>
   <summary>TPC-H: <b>Total hot run time: 35087 ms</b></summary>
   
   ```
   machine: 'aliyun_ecs.c7a.8xlarge_32C64G'
   scripts: https://github.com/apache/doris/tree/master/tools/tpch-tools
   Tpch sf100 test result on commit d860fdfa2cd9ae99881222ad8502aa83759542ae, data reload: false
   
   ------ Round 1 ----------------------------------
   q1	17621	4536	4059	4059
   q2	2015	350	236	236
   q3	10181	1337	732	732
   q4	10254	936	318	318
   q5	7528	2146	1925	1925
   q6	190	167	138	138
   q7	1008	848	705	705
   q8	9375	1474	1165	1165
   q9	7165	5420	5371	5371
   q10	6846	2431	2000	2000
   q11	539	338	301	301
   q12	690	739	556	556
   q13	17777	3698	3007	3007
   q14	290	316	268	268
   q15	605	521	516	516
   q16	708	676	636	636
   q17	694	811	567	567
   q18	7758	7215	6963	6963
   q19	1152	974	600	600
   q20	396	359	257	257
   q21	4201	3905	3774	3774
   q22	1035	1009	993	993
   Total cold run time: 108028 ms
   Total hot run time: 35087 ms
   
   ----- Round 2, with runtime_filter_mode=off -----
   q1	4123	3999	4046	3999
   q2	329	393	330	330
   q3	2131	2682	2306	2306
   q4	1331	1781	1274	1274
   q5	4261	4723	4629	4629
   q6	239	189	138	138
   q7	2085	1966	1895	1895
   q8	2672	2570	2512	2512
   q9	7562	7499	7534	7499
   q10	3086	3262	2856	2856
   q11	624	551	513	513
   q12	665	898	698	698
   q13	3549	4009	3320	3320
   q14	312	325	279	279
   q15	552	494	518	494
   q16	658	693	623	623
   q17	1205	1475	1383	1383
   q18	8057	7695	7500	7500
   q19	863	830	888	830
   q20	2002	2061	1954	1954
   q21	4619	4224	4231	4224
   q22	1118	1034	981	981
   Total cold run time: 52043 ms
   Total hot run time: 50237 ms
   ```
   </details>
   


-- 
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: commits-unsubscribe@doris.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@doris.apache.org
For additional commands, e-mail: commits-help@doris.apache.org