JNI. See copied text from email below.
2. This commit contains all the code used in profiling, detecting FP
exceptions, dumping intermediate results. All flagged off using ifdefs,
but it's there.
--------------Text from email
As we discussed before, it's the denormal numbers that are causing the
slowdown - the core executes some microcode uops (called FP assists)
when denormal numbers are detected for FP operations (even un-vectorized
code).
The C++ compiler by default enables flush to zero (FTZ) - when set, the
hardware simply converts denormal numbers to 0. The Java binary
(executable provided by Oracle, not the native library) seems to be
compiled without FTZ (sensible choice, they want to be conservative).
Hence, the JNI invocation sees a large slowdown. Disabling FTZ in C++
slows down the C++ sandbox performance to the JNI version (fortunately,
the reverse also holds :)).
Not sure how to show the overhead for these FP assists easily - measured
a couple of counters.
FP_ASSISTS:ANY - shows number of uops executed as part of the FP
assists. When FTZ is enabled, this is 0 (both C++ and JNI), when FTZ is
disabled this value is around 203540557 (both C++ and JNI)
IDQ:MS_UOPS_CYCLES - shows the number of cycles the decoder was issuing
uops when the microcode sequencing engine was busy. When FTZ is enabled,
this is around 1.77M cycles (both C++ and JNI), when FTZ is disabled
this value is around 4.31B cycles (both C++ and JNI). This number is
still small with respect to total cycles (~40B), but it only reflects
the cycles in the decode stage. The total overhead of the microcode
assist ops could be larger.
As suggested by Mustafa, I compared intermediate values (matrices M,X,Y)
and final output of compute_full_prob. The values produced by C++ and
Java are identical to the last bit (as long as both use FTZ or no-FTZ).
Comparing the outputs of compute_full_prob for the cases no-FTZ and FTZ,
there are differences for very small values (denormal numbers).
Examples:
Diff values 1.952970E-33 1.952967E-33
Diff values 1.135071E-32 1.135070E-32
Diff values 1.135071E-32 1.135070E-32
Diff values 1.135071E-32 1.135070E-32
For this test case (low coverage NA12878), all these values would be
recomputed using the double precision version. Enabling FTZ should be
fine.
-------------------End text from email
Merge branch 'master' of /home/mghodrat/PairHMM/shared-repository into intel_pairhmm
Conflicts:
PairHMM_JNI/org_broadinstitute_sting_utils_pairhmm_VectorLoglessPairHMM.cc
2. Split into DebugJNILoglessPairHMM and VectorLoglessPairHMM with base
class JNILoglessPairHMM. DebugJNILoglessPairHMM can, in principle,
invoke any other child class of JNILoglessPairHMM.
3. Added more profiling code for Java parts of LoglessPairHMM
2. Wrapped _mm_empty() with ifdef SIMD_TYPE_SSE
3. OpenMP disabled
4. Added code for initializing PairHMM's data inside initializePairHMM -
not used yet
SSE compilation warning.
2. Added code to dynamically select between AVX, SSE4.2 and normal C++ (in
that order)
3. Created multiple files to compile with different compilation flags:
avx_function_prototypes.cc is compiled with -xAVX while
sse_function_instantiations.cc is compiled with -xSSE4.2 flag.
4. Added jniClose() and support in Java (HaplotypeCaller,
PairHMMLikelihoodCalculationEngine) to call this function at the end of
the program.
5. Removed debug code, kept assertions and profiling in C++
6. Disabled OpenMP for now.
1. Moved computeLikelihoods from PairHMM to native implementation
2. Disabled debug - debug code still left (hopefully, not part of
bytecode)
3. Added directory PairHMM_JNI in the root which holds the C++
library that contains the PairHMM AVX implementation. See
PairHMM_JNI/JNI_README first
2. Disabled debug - debug code still left (hopefully, not part of
bytecode)
3. Added directory PairHMM_JNI in the root which holds the C++ library
that contains the PairHMM AVX implementation. See PairHMM_JNI/JNI_README
first
-- New -a argument in the VQSR for specifying additional data to be used in the clustering
-- New NA12878KB walker which creates ROC curves by partitioning the data along VQSLOD and calculating how many KB TP/FP's are called.
For example, this tool can be used for processing bowtie RNA-seq data.
Each read with k N-cigar elemments is plit to k+1 reads. The split is done by hard clipping the bases rest of the bases.
In order to do it, few changes were introduced to some other clipping methods:
- make a segnificant change in ClippingOp.hardClip() that prevent the spliting of read with cigar: 1M2I1N1M3I.
- change getReadCoordinateForReferenceCoordinate in ReadUtil to recognize Ns
create unitTests for that walker:
- change ReadClipperTestUtils to be more general in order to use its code and avoid code duplication
- move some useful methods from ReadClipperTestUtils to CigarUtils
create integration test for that class
small change in a comment in FullProcessingPipeline
last commit:
Address review comments:
- move to protected under walkers/rnaseq
- change the read splitting methods to be more readable and more efficiant
- change (minor changes) some methods in ReadClipper to allow the changes in split reads
- add (minor change) one method to CigarUtils to allow the changes in split reads
- change ReadUtils.getReadCoordinateForReferenceCoordinate to include possible N in the cigar
- address the rest of the review comments (minor changes)
- fix ReadUtilsUnitTest.testReadWithNs acoording to the defult behaviour of getReadCoordinateForReferenceCoordinate (in case of refernce index that fall into deletion, return the read index of the base before the deletion).
- add another test to ReadUtilsUnitTest.testReadWithNs
- Allow the user to print the split positions (not working proparly currently)
mghodrat
Karthik Gururaj
mozdal
Eric Banks
Ryan Poplin
amilev
MauricioCarneiro