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
Karthik Gururaj