Migrated to portable simd #747

jorgecarleitao · 2022-01-09T20:00:46Z

This PR replaces the packed_simd dependency by the nightly-available std::simd (aka portable simd).

Closes #580

codecov · 2022-01-09T20:08:03Z

Codecov Report

Merging #747 (db5cdce) into main (1b7e8ee) will decrease coverage by 1.02%.
The diff coverage is 60.00%.

@@            Coverage Diff             @@
##             main     #747      +/-   ##
==========================================
- Coverage   71.52%   70.49%   -1.03%     
==========================================
  Files         337      343       +6     
  Lines       18443    18735     +292     
==========================================
+ Hits        13191    13207      +16     
- Misses       5252     5528     +276

Impacted Files	Coverage Δ
src/types/simd/mod.rs	`90.90% <ø> (ø)`
src/types/simd/packed.rs	`0.00% <0.00%> (ø)`
src/compute/aggregate/min_max.rs	`65.90% <100.00%> (ø)`
src/compute/aggregate/simd/native.rs	`94.11% <100.00%> (ø)`
src/compute/arithmetics/time.rs	`25.68% <0.00%> (-0.92%)`	⬇️
src/io/odbc/read/deserialize.rs	`0.00% <0.00%> (ø)`
src/io/odbc/write/serialize.rs	`0.00% <0.00%> (ø)`
src/io/odbc/read/schema.rs	`0.00% <0.00%> (ø)`
src/io/odbc/write/schema.rs	`0.00% <0.00%> (ø)`
... and 3 more

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1b7e8ee...db5cdce. Read the comment docs.

jorgecarleitao · 2022-01-09T21:05:23Z

Compiling and tests passing 🎉🎉🎉🎉🎉

Unfortunately, we do have major regressions:

git checkout main
cargo bench --bench aggregate --features compute_aggregate,benchmarks,simd -- "2\^20"
cargo bench --bench comparison_kernels --features compute_comparison,benchmarks,simd -- "2\^20"
git checkout portable_simd
cargo bench --bench aggregate --features compute_aggregate,benchmarks,simd -- "2\^20"
cargo bench --bench comparison_kernels --features compute_comparison,benchmarks,simd -- "2\^20"

sum 2^20 f32            time:   [176.60 us 176.83 us 177.07 us]                         
                        change: [+13.785% +16.097% +18.402%] (p = 0.00 < 0.05)
min 2^20 f32            time:   [301.82 us 303.29 us 304.81 us]                         
                        change: [+23.534% +24.444% +25.363%] (p = 0.00 < 0.05)
sum null 2^20 f32       time:   [212.85 us 214.04 us 215.33 us]                              
                        change: [-5.2111% -4.5145% -3.7872%] (p = 0.00 < 0.05)
min null 2^20 f32       time:   [474.66 us 477.09 us 479.59 us]                              
                        change: [+8.4680% +9.2737% +10.047%] (p = 0.00 < 0.05)

f32 2^20                time:   [2.1006 ms 2.1116 ms 2.1240 ms]                      
                        change: [+437.37% +440.05% +443.38%] (p = 0.00 < 0.0
f32 scalar 2^20         time:   [2.1254 ms 2.1360 ms 2.1466 ms]                             
                        change: [+961.65% +968.65% +974.66%] (p = 0.00 < 0.0
bool 2^20               time:   [189.05 us 190.03 us 191.03 us]                      
                        change: [-11.617% -10.306% -9.1736%] (p = 0.00 < 0.05)
bool scalar 2^20        time:   [31.371 us 31.527 us 31.687 us]                              
                        change: [-1.3276% -0.4781% +0.3641%] (p = 0.26 > 0.05)

ritchie46 · 2022-01-09T21:33:49Z

Man, those are huge regressions :(. Given that it compiles, maybe just wait a bit and do a benchmark once in a while?

jorgecarleitao · 2022-01-09T21:47:36Z

I agree that we should not merge with these numbers. I will reach the working group for feedback and advice.

src/compute/aggregate/simd/packed.rs

jorgecarleitao · 2022-01-11T07:23:16Z

Converted to draft to indicate that this is blocked by a perf. regression.

jorgecarleitao · 2022-01-13T19:56:51Z

after some iterations on zulip, the tracking issues are:

jhorstmann · 2022-02-11T14:17:08Z

FYI, I also experimented with portable_simd today and created an llvm issue about a missing/incorrect optimization with from_bitmask on non-avx512 machines.

jorgecarleitao · 2022-02-11T15:23:11Z

Wow, awesome finding and summary @jhorstmann - thank you for reporting it over there 🙇

ritchie46 · 2022-03-05T15:41:45Z

What do you think of keeping this one aside of current SIMD (until the performance is comparable)? Then there is already SIMD support for the stable compiler. And can people opt in to the nightly version for the maximal performance.

jorgecarleitao · 2022-03-05T15:49:09Z

:) portable simd released the required APIs. I am re-benching this as of now to evaluate where we at wrt to performance. :)

ritchie46 · 2022-03-05T15:51:07Z

🤞

jorgecarleitao · 2022-03-05T16:18:31Z

There are still 2 regressions, so this continues blocked.

sum 2^20 f32            time:   [164.88 us 165.03 us 165.23 us]                         
                        change: [-1.3535% -0.2982% +0.8551%] (p = 0.62 > 0.05)
min 2^20 f32            time:   [211.05 us 212.38 us 214.02 us]                         
                        change: [-0.9052% +0.2757% +1.3189%] (p = 0.64 > 0.05)
sum 2^20 i32            time:   [167.82 us 168.38 us 169.12 us]                         
min 2^20 i32            time:   [165.78 us 166.48 us 167.59 us]                         
sum null 2^20 f32       time:   [858.64 us 860.01 us 861.73 us]                              
                        change: [+352.55% +356.83% +360.89%] (p = 0.00 < 0.05)
min null 2^20 f32       time:   [874.36 us 877.98 us 883.40 us]                              
                        change: [+128.24% +131.57% +136.81%] (p = 0.00 < 0.05)
f32 2^20                time:   [349.08 us 354.17 us 364.40 us]                     
                        change: [-3.7451% -2.3785% -0.2625%] (p = 0.00 < 0.05)
f32 scalar 2^20         time:   [201.41 us 202.66 us 204.41 us]                            
                        change: [-2.4794% -1.2193% -0.0931%] (p = 0.05 < 0.05)
bool 2^20               time:   [179.35 us 180.41 us 182.26 us]                      
                        change: [+4.1593% +5.9573% +8.4415%] (p = 0.00 < 0.05)
bool scalar 2^20        time:   [26.768 us 26.920 us 27.116 us]                              
                        change: [+0.2871% +1.5636% +2.6602%] (p = 0.00 < 0.05)
utf8 2^20               time:   [6.6215 ms 6.6433 ms 6.6747 ms]                      
                        change: [-3.8776% -3.2642% -2.6263%] (p = 0.00 < 0.05)

jorgecarleitao · 2022-03-05T16:27:23Z

actually, the regression was only from the last commit, which tries to use the intrinsics for the from_bitmask. Without that, there is no regression. Just reverted that one and will post the updated benches.

jorgecarleitao · 2022-03-05T17:39:19Z

There is some noise, but the overall is that we are now equal to packed simd:

sum 2^20 f32            time:   [166.41 us 167.83 us 169.74 us]                         
                        change: [-0.0800% +1.1867% +3.0513%] (p = 0.11 > 0.05)
min 2^20 f32            time:   [210.90 us 211.92 us 213.13 us]                         
                        change: [+0.2811% +1.5970% +2.8225%] (p = 0.02 < 0.05)
sum 2^20 i32            time:   [167.38 us 167.71 us 168.16 us]                         
                        change: [-6.1677% -4.9397% -3.5768%] (p = 0.00 < 0.05)
min 2^20 i32            time:   [198.58 us 199.09 us 199.78 us]                         
                        change: [+14.587% +16.306% +18.393%] (p = 0.00 < 0.05)
sum null 2^20 f32       time:   [186.71 us 188.37 us 190.54 us]                              
                        change: [-1.6255% +0.2482% +2.1276%] (p = 0.81 > 0.05)
min null 2^20 f32       time:   [368.12 us 370.28 us 373.81 us]                              
                        change: [-4.1516% -3.1519% -2.1852%] (p = 0.00 < 0.05)
f32 2^20                time:   [351.59 us 352.98 us 355.48 us]                     
                        change: [-0.8055% +0.4256% +1.7513%] (p = 0.52 > 0.05)
f32 scalar 2^20         time:   [235.50 us 237.09 us 239.49 us]                            
                        change: [-4.2012% -1.9893% -0.1382%] (p = 0.06 > 0.05)
bool 2^20               time:   [192.72 us 193.63 us 195.08 us]                      
                        change: [+14.066% +16.238% +19.238%] (p = 0.00 < 0.05
bool scalar 2^20        time:   [26.444 us 26.557 us 26.692 us]                              
                        change: [-2.1207% -0.6513% +0.5040%] (p = 0.36 > 0.05)
utf8 2^20               time:   [6.5841 ms 6.6033 ms 6.6358 ms]                      
                        change: [-0.3873% +0.1999% +0.8762%] (p = 0.57 > 0.05)

ritchie46 · 2022-03-05T18:00:57Z

Awesome. Nice work!

jorgecarleitao force-pushed the portable_simd branch 3 times, most recently from e1e7af8 to 099d5ed Compare January 9, 2022 20:30

jorgecarleitao marked this pull request as ready for review January 9, 2022 20:33

programmerjake reviewed Jan 10, 2022

View reviewed changes

src/compute/aggregate/simd/packed.rs Outdated Show resolved Hide resolved

jorgecarleitao force-pushed the portable_simd branch 3 times, most recently from 8badb7d to 81aacde Compare January 10, 2022 05:38

jorgecarleitao marked this pull request as draft January 11, 2022 07:22

jorgecarleitao mentioned this pull request Jan 13, 2022

Perf. regression of min over packed_simd rust-lang/portable-simd#222

Closed

ritchie46 mentioned this pull request Jan 18, 2022

Add extend_from_iter into growable trait #773

Closed

alamb mentioned this pull request Jan 18, 2022

Broken CI SIMD and wasm checks on master: unknown feature llvm_asm apache/arrow-rs#1198

Closed

jorgecarleitao mentioned this pull request Feb 6, 2022

Upgrade packed_simd to 3.6.7 for arrow2 databendlabs/databend#4062

Closed

jorgecarleitao force-pushed the portable_simd branch from 81aacde to ede1db6 Compare March 5, 2022 15:03

jorgecarleitao force-pushed the portable_simd branch from fb754d7 to 46f361e Compare March 5, 2022 16:25

jorgecarleitao added 2 commits March 5, 2022 17:20

Improved bench

8e55011

Migrated to portable simd

767b9fe

jorgecarleitao added 2 commits March 5, 2022 17:20

Used intrinsic for to_bitmask

1df6e1c

Bumped nightly

de4f3da

jorgecarleitao force-pushed the portable_simd branch from 46f361e to de4f3da Compare March 5, 2022 17:39

jorgecarleitao marked this pull request as ready for review March 5, 2022 18:13

Bumped nightly to miri

db5cdce

jorgecarleitao force-pushed the portable_simd branch from c52abe0 to db5cdce Compare March 5, 2022 19:50

jorgecarleitao merged commit 25b8ef8 into main Mar 5, 2022

jorgecarleitao deleted the portable_simd branch March 5, 2022 20:26

jorgecarleitao added the feature A new feature label Mar 6, 2022

This was referenced Apr 1, 2022

packed_simd requires nightly apache/arrow-rs#54

Closed

packed_simd v.s. portable_simd, which should be used? apache/arrow-rs#1492

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Migrated to portable simd #747

Migrated to portable simd #747

jorgecarleitao commented Jan 9, 2022 •

edited

Loading

codecov bot commented Jan 9, 2022 •

edited

Loading

jorgecarleitao commented Jan 9, 2022 •

edited

Loading

ritchie46 commented Jan 9, 2022

jorgecarleitao commented Jan 9, 2022

jorgecarleitao commented Jan 11, 2022

jorgecarleitao commented Jan 13, 2022 •

edited

Loading

jhorstmann commented Feb 11, 2022

jorgecarleitao commented Feb 11, 2022

ritchie46 commented Mar 5, 2022

jorgecarleitao commented Mar 5, 2022

ritchie46 commented Mar 5, 2022

jorgecarleitao commented Mar 5, 2022

jorgecarleitao commented Mar 5, 2022

jorgecarleitao commented Mar 5, 2022

ritchie46 commented Mar 5, 2022

Migrated to portable simd #747

Migrated to portable simd #747

Conversation

jorgecarleitao commented Jan 9, 2022 • edited Loading

codecov bot commented Jan 9, 2022 • edited Loading

Codecov Report

jorgecarleitao commented Jan 9, 2022 • edited Loading

ritchie46 commented Jan 9, 2022

jorgecarleitao commented Jan 9, 2022

jorgecarleitao commented Jan 11, 2022

jorgecarleitao commented Jan 13, 2022 • edited Loading

jhorstmann commented Feb 11, 2022

jorgecarleitao commented Feb 11, 2022

ritchie46 commented Mar 5, 2022

jorgecarleitao commented Mar 5, 2022

ritchie46 commented Mar 5, 2022

jorgecarleitao commented Mar 5, 2022

jorgecarleitao commented Mar 5, 2022

jorgecarleitao commented Mar 5, 2022

ritchie46 commented Mar 5, 2022

jorgecarleitao commented Jan 9, 2022 •

edited

Loading

codecov bot commented Jan 9, 2022 •

edited

Loading

jorgecarleitao commented Jan 9, 2022 •

edited

Loading

jorgecarleitao commented Jan 13, 2022 •

edited

Loading