This patch introduces new optimized version of SHGEMM kernel
using power10 Matrix-Multiply Assist (MMA) feature introduced in
POWER ISA v3.1. This patch makes use of new POWER10 compute instructions
for matrix multiplication operation.
Tested on simulator and there are no new test failures.
Comet Lake U CPUs have family: 6, model: 6, extended family: 0, and
extended model: 10 were not being correctly detected by GETARCH during
openblas builds and would show CORE=UNKNOWN and LIBCORE=unknown. This
commit adds the necessary information to cpuid_x86 to detect extended
family 10 model 6 and return the proper core information. It's
essentially just a skylake cpu, not skylake x, so I just took the used
the same return fields as skylake.
This patch introduces new optimized version of ZGEMM kernel
using power10 Matrix-Multiply Assist (MMA) feature introduced in
POWER ISA v3.1. This patch makes use of new POWER10 compute instructions
for matrix multiplication operation.
Tested on simulator and there are no new test failures.
Cycles count reduced by 30-50% compared to POWER9 version depending on
M/N/K sizes.
This patch introduces new optimized version of SGEMM, CGEMM and DGEMM
using power10 Matrix-Multiply Assist (MMA) feature introduced in
POWER ISA v3.1. This patch makes use of new POWER10 compute instructions
for matrix multiplication operation.
Tested on simulator and there are no new test failures.
Cycles count reduced by 30-50% compared to POWER9 version depending on
M/N/K sizes.
MMA GCC patch for reference:
https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=8ee2640bfdc62f835ec9740278f948034bc7d9f1
The two test cases dgemv_tester and dgemm_tester accept the degree of
concurrency as command line argument (amongst others). Fail early if
value 0 has been specified, instead of later with less-clear symptoms.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
... instead of (number of hw threads - 4) to avoid invalid numbers on
smaller systems. Currently, systems with 4 or fewer CPUs (e.g., small CI
VMs) would fail the test. Fixes one of the issues discussed in #2668
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
In gcc-7.1, the behavior of -dumpversion changed to be configured
at compile-time. On some distributions it only dumps the major version
(e.g., Ubuntu), so the current checks for the gcc minor version report
false negatives. As a replacement, gcc-7.1 introduced -dumpfullversion
which always prints the full version.
Update the gcc version detection in Makefile.system to employ
-dumpfullversion with gcc-7 and newer.
Posting this patch for discussion, since it emerged from discussions
around issue #2668 and PR #2669. It is not solving a problem right now,
but may be useful in the future.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
... to bring unified gcc version detection with common variables to the
one remaining spot in Makefile.system.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Employ common variables for gcc version detection and fix the broken
check for gcc >= 5.2.
Fixes#2668
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
