dmeoli / ADS / 40981254551

97%

main: 98%

Build Type

push

github

Committed by dmeoli

Commit Message

the backends of GMP and of a CBLAS on by default where found

CMake takes ADS_WITH_GMP and the new ADS_WITH_BLAS as AUTO (on when a
program that calls the library compiles and links, ON an error if it does
not, OFF the code of ADS alone), lib/makefile-c does the same with the
probes of extlib/makefile-libGMP and the new extlib/makefile-libBLAS (make
backends tells what was found), ADSConfig.cmake looks for them first where
they were found, and the recipes of packaging/ build with both off. The
ranges of BigInteger on GMP were measured again against mpz_class on every
case of bench/bignum.cpp: GMP also takes the square root, the quotients by
divisors of 4 to 2047 digits straight into the result, the conversions of
1 to 3 digits and the gcd of 2 digits, while the long quotients and
conversions go back to the code of ADS, now the faster there. Rational
goes to mpq_add(), mpq_sub(), mpq_mul(), mpq_div() and mpq_cmp() on terms
of 2 and 3 digits, read in place. gemm() of double and float goes to
cblas_dgemm() and cblas_sgemm() on the small and the thin products, where
OpenBLAS was the faster, the parallel gemm() a range of rows per thread;
on macOS the CBLAS of Accelerate comes from vecLib/cblas.h alone, since
the whole Accelerate brings the old types of macOS (Point) into the global
namespace. The tests check the edges of all the ranges with and without
the backends; the CI runs the tests on the backends found and with both
off on every system and compiler of its matrix, and the job backends with
both required and the parallel gemm() on OpenBLAS under ThreadSanitizer;
the theory of the numeric area gains the section on the backends.

Coverage Stats

88 of 91 new or added lines in 2 files covered. (96.7%)

195 existing lines in 2 files now uncovered.

47630 of 48978 relevant lines covered (97.25%)

9717428.32 hits per line