Unroll loop in lookup_2_lanes

The current loop in function lookup_2_lanes infais/utils/simdlib_emulated.h
goes from 0 to 31.  It has an if statement to do an assignment for j < 16
and a different assignment for j >= 16.  By unrolling the loop to do the j
< 16 and the j >= 16 iterations in parallel the if j < 16 is eliminated and
the number of loop iterations is reduced in half.

Then unroll the loop for the j < 16 and the j >=16 to a depth of 2.

This change results in approximately a 55% reduction in the execution time
for the bench_ivf_fastscan.py workload on Power 10 when compiled with
CMAKE_INSTALL_CONFIG_NAME=Release.

The removal of the if (j < 16) statement and the unrolling of the loop
removes branch cycle stall and register dependencies on instruction issue.
The result is the unrolled code is able issue instructions earlier thus
reducing the total number of cycles required to execute the function.

This patch makes a copy of faiss/utils/simdlib_emulated.h and names it
faiss/utils/simdlib_emulated_ppc64.h.  The new file has the new version
of lookup_2_lanes.  The new included file is gets included in file
faiss/utils/simdlib.h used if the define __PPC64__ is set by the GCC
compiler on Linux or the XLC clang compiler for AIX.  Otherwise, the
original fine simdlib_emulated.h is included.

Author: Carl Love

faiss/utils/simdlib.h

@@ -30,8 +30,16 @@
 #else
 
 // emulated = all operations are implemented as scalars
+#if defined(__PPC64__)
+
+#include <faiss/utils/simdlib_emulated_ppc64.h>
+
+#else
+
 #include <faiss/utils/simdlib_emulated.h>
 
+#endif
+
 // FIXME: make a SSE version
 // is this ever going to happen? We will probably rather implement AVX512