Closed intelmark closed 4 years ago
Hi @intelmark , could you add the test result with -asm 0
, -asm 1
with avx2/avx512 here?
Please also check MD5. Thank you.
Testing shows that with -asm 0 and -asm 1, both with avx2 and avx512 encodings with the latest commits in this PR are binary identical.
Isn’t needed.
Previous check was for the Sad64x64 to be "not equal" to the best one. The fix was to also check if Sad64x64 is "equal". Hence the check now includes all cases and is no longer needed.
Hi @tszumski , could you help review the PR? Thank you.
Hi @intelmark , please rerun the CI fast type. And also check for perf. Thank you.
Hi Mark, Changes looks good but I have different proposition for AVX changes (less code): https://github.com/tszumski/SVT-HEVC/commit/80362a5b909dcce6577145b98881ecbd02118177
python SVT-HEVC_FunctionalTests.py Fast -type asm_test
Test Begin... Total Number of Tests: 108 Total Passed: 108 Percentage Passed: 100.0% Time Elapsed: 9 minute(s), 53 second(s)
Speed check: (PR 500) Bin/Release/SvtHevcEncApp -i /testfiles/bbb_1920x1080_420p.yuv -w 1920 -h 1080 -asm 0 -encMode 0 -intra-period 100 -scd 0 -rc 0 -q 22 -n 100 -b Slibbb.265 Average Speed: 2.11 fps Total Encoding Time: 47446 ms Total Execution Time: 47952 ms Average Latency: 33999 ms Max Latency: 46614 ms
(master branch) Average Speed: 2.08 fps Total Encoding Time: 48028 ms Total Execution Time: 48529 ms Average Latency: 34168 ms Max Latency: 47832 ms
Changed the check for pBestSad64x64 to also include being equal to the value, as the computed value might equal pBestSad64x64 without a correct corresponding motion vector ever being set.
We are thus changing to <= to include this corner case.
Signed-off-by: Mark Feldman mark.feldman@intel.com
Closes #481