Welcome to mirror list, hosted at ThFree Co, Russian Federation.

github.com/marian-nmt/marian-regression-tests.git - Unnamed repository; edit this file 'description' to name the repository.
summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat (limited to 'tests/decoder/intgemm/test_intgemm_8bit.sh')
-rw-r--r--tests/decoder/intgemm/test_intgemm_8bit.sh12
1 files changed, 9 insertions, 3 deletions
diff --git a/tests/decoder/intgemm/test_intgemm_8bit.sh b/tests/decoder/intgemm/test_intgemm_8bit.sh
index 1bdec70..d6098a0 100644
--- a/tests/decoder/intgemm/test_intgemm_8bit.sh
+++ b/tests/decoder/intgemm/test_intgemm_8bit.sh
@@ -47,10 +47,16 @@ $MRT_MARIAN/marian-decoder \
# Print current and expected BLEU for debugging
python3 $MRT_TOOLS/sacrebleu/sacrebleu.py newstest2018.ref < $prefix.out | tee $prefix.out.bleu
-cat $prefix.$suffix.expected.bleu
+# BLEU scores calculated on AVX, AVX2, AVX512 should be very similar, but does not have to be identical
+tail -n1 $prefix.*.expected.bleu || true
-# Compare with the expected output
-$MRT_TOOLS/diff.sh $prefix.out $prefix.$suffix.expected > $prefix.diff
+# Compare with the expected output, allow up to 20 different lines
+$MRT_TOOLS/diff.sh $prefix.out $prefix.$suffix.expected 20 > $prefix.diff
+
+# Compare BLEU scores
+cut -f3 -d' ' $prefix.out.bleu > $prefix.out.bleu_score
+cut -f3 -d' ' $prefix.$suffix.expected.bleu > $prefix.out.bleu_score.expected
+$MRT_TOOLS/diff-nums.py -a -p 0.9 $prefix.out.bleu_score{,.expected}
# Exit with success code