Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Single-threaded performance improvements in forward DWT for 5-3 and 9-7 (and other improvements) #1253

Merged
merged 24 commits into from
Oct 9, 2020
Merged

Single-threaded performance improvements in forward DWT for 5-3 and 9-7 (and other improvements) #1253

merged 24 commits into from
Oct 9, 2020

Conversation

rouault
Copy link
Collaborator

@rouault rouault commented May 24, 2020

Summary of improvements:

  • Switches forward 9-7 DWT to using 32-bit floating point operations instead of fixed point arithmetic
  • For 5-3 and 9-7 foward DWT, perform the vertical pass on 8 columns at a time for better cache efficiency and to be able to use SSE2 (for 5-3) and SSE (for 9-7)
  • forward DWT: more efficient 5-3 and 9-7 horizontal pass processing (removal of useless bound checking)
  • inverse 9-7 DWT: add multi-threading and operate on 8 columns at a time for vertical pass
  • bench_dwt utility extended to support irreversible wavelet
  • performance improvements in T1 encoder (~ 9%)

Combined effect of those improvements on the bench used in PR 1248 (time bin/opj_compress -threads XXXX -n 5 -c [256,256],[256,256],[256,256],[256,256],[256,256] -t 1024,1024 -PLT -i T36JTT_20160914T074612_B02.tif -o out.jp2 >/dev/null)

PR 1248 Now
Threads = 0 17.9 s 15.3 s
Threads = 4 5.7 s 5.0 s
Threads = 8 4.4 s 4.0 s

(For convenience this PR is based on top of PR 1248, so the first 2 commits "Add multithreading ..." belong to this later one)

Funded by ESA for S2-MPC project

rouault added 24 commits May 20, 2020 20:30
@rouault
Add multithreading support in the T1 (entropy phase) encoder
97eb7e0 
- API wise, opj_codec_set_threads() can be used on the encoding side
- opj_compress has a -threads switch similar to opj_uncompress
@rouault
Add multithreaded support in the DWT encoder.
07d1f77 
Update the bench_dwt utility to have a -decode/-encode switch

Measured performance gains for DWT encoder on a
Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz (4 cores, hyper threaded)

Encoding time:
$ ./bin/bench_dwt -encode -num_threads 1
time for dwt_encode: total = 8.348 s, wallclock = 8.352 s

$ ./bin/bench_dwt -encode -num_threads 2
time for dwt_encode: total = 9.776 s, wallclock = 4.904 s

$ ./bin/bench_dwt -encode -num_threads 4
time for dwt_encode: total = 13.188 s, wallclock = 3.310 s

$ ./bin/bench_dwt -encode -num_threads 8
time for dwt_encode: total = 30.024 s, wallclock = 4.064 s

Scaling is probably limited by memory access patterns causing
memory access to be the bottleneck.
The slightly worse results with threads==8 than with thread==4
is due to hyperthreading being not appropriate here.
@rouault
dwt.c: change sign of constants to match standard and compensate (no …
99107d5 
…functional change)
@rouault
Encoder: use floating-point operations for irreversible transformation
00cff6f
@rouault
tcd.c: add comment
3d35d0f
@rouault
compare_images.c: code reformatting
c2b9d09
@rouault
Testing: revise testing of lossy encoding by comparing PEAK and MSE w…
fe4c15f 
…ith original image
@rouault
opj_mct_encode_real(): add SSE optimization
c6a413a
@rouault
opj_j2k_setup_encoder(): add validation of tile width and height to a…
4ab2ed0 
…void potential division by zero
@rouault
opj_dwt_encode_1_real(): avoid many bound comparisons, similarly to d…
e46e300 
…ecoding side
@rouault
Irreversible decoding: align code more closely to the standard by avo…
f38c069 
…id messing up with stepsize (no functional change)
@rouault
Irreversible compression/decompression DWT: use 1/K constant as per s…
3cd1305 
…tandard

The previous constant opj_c13318 was mysteriously equal to 2/K , and in
the DWT, we had to divide K and opj_c13318 by 2... The issue was that the
band->stepsize computation in tcd.c didn't take into account the log2gain of
the band.

The effect of this change is expected to be mostly equivalent to the previous
situation, except some difference in rounding. But it leads to a dramatic
reduction of the mean square error and peak error in the irreversible encoding
of issue141.tif !
@rouault
Irreversible decoding: partially revert previous commit, to fix failu…
adccbc8 
…res in test suite
@rouault
bench_dwt.c: add a -I switch to test irreversible FWDT/IDWT
0c09062
@rouault
Speed-up 9x7 IDWD by ~20%
47943da 
"bench_dwt -I" time goes from 2.8s to 2.2s
@rouault
Remove useless + 5U margin in opj_dwt_decode_tile_97()
272b3e0 
Nothing in code analysis nor test suite shows that this margin is
needed.
It dates back to commit dbeebe7
where vector 9x7 decoding was introduced.
@rouault
Speed-up 9x7 IDWD by ~30% with OPJ_NUM_THREADS=2
45a3522 
"bench_dwt -I" time goes from 2.2s to 1.5s
@rouault
Forward DWT: small code refactoring to allow future improvements for …
bd5f5ee 
…the horizontal pass
@rouault
Forward DWT 5x3: performance improvements in horizontal pass, and mod…
97b384a 
…est in vertical pass
@rouault
dwt.c: remove unused typedef
33d3d0d
@rouault
Forward DWT: small code refactoring to allow future improvements for …
e69fa09 
…the vertical pass
@rouault
Forward DWT 5-3: major speed up by vectorizing vertical pass
a38e970 
`bench_dwt -encode` times goes from 7.9s to 1.7s
@rouault
Forward DWT 9-7: major speed up by vectorizing vertical pass
1e931fd 
`bench_dwt -I -encode` times goes from 8.6s to 2.1s
@rouault
T1 encoder: speed-up by aggressive inlining and more cache friendly d…
1c5627e 
…ata organization

~ 9% speed improvement seen on 10980x10980 uint16 image, T36JTT_20160914T074612_B02.tif
opj_compress time from 17.2s to 15.8s
@rouault
Copy link
Collaborator Author

rouault commented May 24, 2020

CC @savmickael This is ready for your review

@rouault
Copy link
Collaborator Author

rouault commented Oct 5, 2020

@savmickael I'll merge this soon (before end of week), unless you want to have a look before

@rouault rouault merged commit 491299e into uclouvain:master Oct 9, 2020
mtremer pushed a commit to ipfire/ipfire-2.x that referenced this pull request Apr 29, 2022
@pmu-ipf
openjpeg: Update to version 2.4.0
ca98d29 
- Update from version 2.3.1 to 2.4.0
- Update of rootfile
- Changelog
    2.4.0
	**Closed issues:**
		- OPENJPEG\_INSTALL\_DOC\_DIR does not control a destination directory where HTML docs would be installed. [\#1309](uclouvain/openjpeg#1309)
		- Heap-buffer-overflow in lib/openjp2/pi.c:312 [\#1302](uclouvain/openjpeg#1302)
		- Heap-buffer-overflow in lib/openjp2/t2.c:973 [\#1299](uclouvain/openjpeg#1299)
		- Heap-buffer-overflow in lib/openjp2/pi.c:623 [\#1293](uclouvain/openjpeg#1293)
		- Global-buffer-overflow in lib/openjp2/dwt.c:1980 [\#1286](uclouvain/openjpeg#1286)
		- Heap-buffer-overflow in lib/openjp2/tcd.c:2417 [\#1284](uclouvain/openjpeg#1284)
		- Heap-buffer-overflow in lib/openjp2/mqc.c:499 [\#1283](uclouvain/openjpeg#1283)
		- Openjpeg could not encode 32bit RGB float image [\#1281](uclouvain/openjpeg#1281)
		- Openjpeg could not encode 32bit RGB float image [\#1280](uclouvain/openjpeg#1280)
		- ISO/IEC 15444-1:2019 \(E\) compared with 'cio.h' [\#1277](uclouvain/openjpeg#1277)
		- Test-suite failure due to hash mismatch [\#1264](uclouvain/openjpeg#1264)
		- Heap use-after-free [\#1261](uclouvain/openjpeg#1261)
		- Memory leak when failing to allocate object... [\#1259](uclouvain/openjpeg#1259)
		- Memory leak of Tier 1 handle when OpenJPEG fails to set it as TLS... [\#1257](uclouvain/openjpeg#1257)
		- Any plan to build release for CVE-2020-8112/CVE-2020-6851 [\#1247](uclouvain/openjpeg#1247)
		- failing to convert 16-bit file: opj\_t2\_encode\_packet\(\): only 5251 bytes remaining in output buffer. 5621 needed. [\#1243](uclouvain/openjpeg#1243)
		- CMake+VS2017 Compile OK, thirdparty Compile OK, but thirdparty not install [\#1239](uclouvain/openjpeg#1239)
		- New release to solve CVE-2019-6988 ? [\#1238](uclouvain/openjpeg#1238)
		- Many tests fail to pass after the update of libtiff to version 4.1.0 [\#1233](uclouvain/openjpeg#1233)
		- Another heap buffer overflow in libopenjp2 [\#1231](uclouvain/openjpeg#1231)
		- Heap buffer overflow in libopenjp2 [\#1228](uclouvain/openjpeg#1228)
		- Endianness of binary volume \(JP3D\) [\#1224](uclouvain/openjpeg#1224)
		- New release to resolve CVE-2019-12973 [\#1222](uclouvain/openjpeg#1222)
		- how to set the block size,like 128,256 ? [\#1216](uclouvain/openjpeg#1216)
		- compress YUV files to motion jpeg2000 standard [\#1213](uclouvain/openjpeg#1213)
		- Repair/update Java wrapper, and include in release [\#1208](uclouvain/openjpeg#1208)
		- abc [\#1206](uclouvain/openjpeg#1206)
		- Slow decoding [\#1202](uclouvain/openjpeg#1202)
		- Installation question [\#1201](uclouvain/openjpeg#1201)
		- Typo in test\_decode\_area - \*ptilew is assigned instead of \*ptileh [\#1195](uclouvain/openjpeg#1195)
		- Creating a J2K file with one POC is broken [\#1191](uclouvain/openjpeg#1191)
		- Make fails on Arch Linux [\#1174](uclouvain/openjpeg#1174)
		- Heap buffer overflow in opj\_t1\_clbl\_decode\_processor\(\) triggered with Ghostscript [\#1158](uclouvain/openjpeg#1158)
		- opj\_stream\_get\_number\_byte\_left: Assertion `p\_stream-\>m\_byte\_offset \>= 0' failed. [\#1151](uclouvain/openjpeg#1151)
		- The fuzzer ignores too many inputs [\#1079](uclouvain/openjpeg#1079)
		- out of bounds read [\#1068](uclouvain/openjpeg#1068)
	**Merged pull requests:**
		- Change defined WIN32 [\#1310](uclouvain/openjpeg#1310) ([Jamaika1](https://github.com/Jamaika1))
		- docs: fix simple typo, producted -\> produced [\#1308](uclouvain/openjpeg#1308) ([timgates42](https://github.com/timgates42))
		- Set ${OPENJPEG\_INSTALL\_DOC\_DIR} to DESTINATION of HTMLs [\#1307](uclouvain/openjpeg#1307) ([lemniscati](https://github.com/lemniscati))
		- Use INC\_DIR for OPENJPEG\_INCLUDE\_DIRS \(fixes uclouvain\#1174\) [\#1306](uclouvain/openjpeg#1306) ([matthew-sharp](https://github.com/matthew-sharp))
		- pi.c: avoid out of bounds access with POC \(fixes \#1302\) [\#1304](uclouvain/openjpeg#1304) ([rouault](https://github.com/rouault))
		- Encoder: grow again buffer size [\#1303](uclouvain/openjpeg#1303) ([zodf0055980](https://github.com/zodf0055980))
		- opj\_j2k\_write\_sod\(\): avoid potential heap buffer overflow \(fixes \#1299\) \(probably master only\) [\#1301](uclouvain/openjpeg#1301) ([rouault](https://github.com/rouault))
		- pi.c: avoid out of bounds access with POC \(refs https://github.com/uclouvain/openjpeg/issues/1293\#issuecomment-737122836\) [\#1300](uclouvain/openjpeg#1300) ([rouault](https://github.com/rouault))
		- opj\_t2\_encode\_packet\(\): avoid out of bound access of \#1297, but likely not the proper fix [\#1298](uclouvain/openjpeg#1298) ([rouault](https://github.com/rouault))
		- opj\_t2\_encode\_packet\(\): avoid out of bound access of \#1294, but likely not the proper fix [\#1296](uclouvain/openjpeg#1296) ([rouault](https://github.com/rouault))
		- opj\_j2k\_setup\_encoder\(\): validate POC compno0 and compno1 \(fixes \#1293\) [\#1295](uclouvain/openjpeg#1295) ([rouault](https://github.com/rouault))
		- Encoder: avoid global buffer overflow on irreversible conversion when… [\#1292](uclouvain/openjpeg#1292) ([rouault](https://github.com/rouault))
		- Decoding: deal with some SPOT6 images that have tiles with a single tile-part with TPsot == 0 and TNsot == 0, and with missing EOC [\#1291](uclouvain/openjpeg#1291) ([rouault](https://github.com/rouault))
		- Free p\_tcd\_marker\_info to avoid memory leak [\#1288](uclouvain/openjpeg#1288) ([zodf0055980](https://github.com/zodf0055980))
		- Encoder: grow again buffer size [\#1287](uclouvain/openjpeg#1287) ([zodf0055980](https://github.com/zodf0055980))
		- Encoder: avoid uint32 overflow when allocating memory for codestream buffer \(fixes \#1243\) [\#1276](uclouvain/openjpeg#1276) ([rouault](https://github.com/rouault))
		- Java compatibility from 1.5 to 1.6 [\#1263](uclouvain/openjpeg#1263) ([jiapei100](https://github.com/jiapei100))
		- opj\_decompress: fix double-free on input directory with mix of valid and invalid images [\#1262](uclouvain/openjpeg#1262) ([rouault](https://github.com/rouault))
		- openjp2: Plug image leak when failing to allocate codestream index. [\#1260](uclouvain/openjpeg#1260) ([sebras](https://github.com/sebras))
		- openjp2: Plug memory leak when setting data as TLS fails. [\#1258](uclouvain/openjpeg#1258) ([sebras](https://github.com/sebras))
		- openjp2: Error out if failing to create Tier 1 handle. [\#1256](uclouvain/openjpeg#1256) ([sebras](https://github.com/sebras))
		- Testing for invalid values of width, height, numcomps [\#1254](uclouvain/openjpeg#1254) ([szukw000](https://github.com/szukw000))
		- Single-threaded performance improvements in forward DWT for 5-3 and 9-7 \(and other improvements\) [\#1253](uclouvain/openjpeg#1253) ([rouault](https://github.com/rouault))
		- Add support for multithreading in encoder [\#1248](uclouvain/openjpeg#1248) ([rouault](https://github.com/rouault))
		- Add support for generation of PLT markers in encoder [\#1246](uclouvain/openjpeg#1246) ([rouault](https://github.com/rouault))
		- Fix warnings about signed/unsigned casts in pi.c [\#1244](uclouvain/openjpeg#1244) ([rouault](https://github.com/rouault))
		- opj\_decompress: add sanity checks to avoid segfault in case of decoding error [\#1240](uclouvain/openjpeg#1240) ([rouault](https://github.com/rouault))
		- ignore wrong icc [\#1236](uclouvain/openjpeg#1236) ([szukw000](https://github.com/szukw000))
		- Implement writing of IMF profiles [\#1235](uclouvain/openjpeg#1235) ([rouault](https://github.com/rouault))
		- tests: add alternate checksums for libtiff 4.1 [\#1234](uclouvain/openjpeg#1234) ([rouault](https://github.com/rouault))
		- opj\_tcd\_init\_tile\(\): avoid integer overflow [\#1232](uclouvain/openjpeg#1232) ([rouault](https://github.com/rouault))
		- tests/fuzzers: link fuzz binaries using $LIB\_FUZZING\_ENGINE. [\#1230](uclouvain/openjpeg#1230) ([Dor1s](https://github.com/Dor1s))
		- opj\_j2k\_update\_image\_dimensions\(\): reject images whose coordinates are beyond INT\_MAX \(fixes \#1228\) [\#1229](uclouvain/openjpeg#1229) ([rouault](https://github.com/rouault))
		- Fix resource leaks [\#1226](uclouvain/openjpeg#1226) ([dodys](https://github.com/dodys))
		- abi-check.sh: fix false postive ABI error, and display output error log [\#1218](uclouvain/openjpeg#1218) ([rouault](https://github.com/rouault))
		- pi.c: avoid integer overflow, resulting in later invalid access to memory in opj\_t2\_decode\_packets\(\) [\#1217](uclouvain/openjpeg#1217) ([rouault](https://github.com/rouault))
		- Add check to validate SGcod/SPcoc/SPcod parameter values. [\#1211](uclouvain/openjpeg#1211) ([sebras](https://github.com/sebras))
		- Fix buffer overflow reading an image file less than four characters [\#1196](uclouvain/openjpeg#1196) ([robert-ancell](https://github.com/robert-ancell))
		- compression: emit POC marker when only one single POC is requested \(f… [\#1192](uclouvain/openjpeg#1192) ([rouault](https://github.com/rouault))
		- Fix several potential vulnerabilities  [\#1185](uclouvain/openjpeg#1185) ([Young-X](https://github.com/Young-X))
		- openjp2/j2k: Report error if all wanted components are not decoded. [\#1164](uclouvain/openjpeg#1164) ([sebras](https://github.com/sebras))

Signed-off-by: Adolf Belka <adolf.belka@ipfire.org>
Reviewed-by: Peter Müller <peter.mueller@ipfire.org>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
1 participant
@rouault