Marius Hillenbrand [Mon, 14 Sep 2020 16:36:31 +0000 (18:36 +0200)]
s390x: fix cscal and zscal implementations
The implementation of complex scalar * vector multiplication for Z14
makes some LAPACK tests fail because the numerical differences to the
reference implementation exceed the threshold (as can be seen by running
make lapack-test and replacing kernel/zarch/cscal.c with a generic
implementation for comparison).
The complex multiplication uses terms of the form a * b + c * d for both
real and imaginary parts. The assembly code (and compiler-emitted code
as well) uses fused multiply add operations for the second product and
sum. The results can be "surprising", for example when both terms in the
imaginary part nearly cancel each other out. In that case, the second
product contributes more digits to the sum than the first product that
has been rounded before.
One option is to use separate multiplications (which then round the same
way) and a distinct add. Change the code to pursue that path, by (1)
requesting the compiler not to contract the operations into FMAs and (2)
replacing the assembly kernel with corresponding vectorized C code
(where change 1 also applies).
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Marius Hillenbrand [Wed, 16 Sep 2020 13:55:38 +0000 (15:55 +0200)]
s390x: for clang use fp-contract=on instead of fast
Make clang slightly more cautious when contracting floating-point
operations (e.g., when applying fused multiply add) by setting
-ffp-contract=on (instead of fast).
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Marius Hillenbrand [Tue, 15 Sep 2020 08:54:37 +0000 (10:54 +0200)]
s390x: move common vector definitions and utils into header
... to facilitate reuse beyond gemm_vec.c and avoid code duplication.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Martin Kroeker [Fri, 18 Sep 2020 21:18:41 +0000 (23:18 +0200)]
Merge pull request #2845 from martin-frbg/lapack443
Fix workspace query in LAPACK xGELQ (Reference-LAPACK 443)
Martin Kroeker [Fri, 18 Sep 2020 07:19:46 +0000 (09:19 +0200)]
Fix workspace query in xGELQ (Reference-LAPACK PR443)
Martin Kroeker [Fri, 18 Sep 2020 07:17:49 +0000 (09:17 +0200)]
Merge pull request #86 from xianyi/develop
rebase
Martin Kroeker [Thu, 17 Sep 2020 21:46:32 +0000 (23:46 +0200)]
Merge pull request #2844 from RajalakshmiSR/daxpy_p10
Optimize daxpy/zaxpy for POWER10
Rajalakshmi Srinivasaraghavan [Thu, 17 Sep 2020 17:56:28 +0000 (12:56 -0500)]
Optimize daxpy/zaxpy for POWER10
This patch makes use of new POWER10 vector pair instructions for
loads and stores. Tested in simulator and no new failures.
Martin Kroeker [Thu, 17 Sep 2020 15:29:56 +0000 (17:29 +0200)]
Merge pull request #2841 from martin-frbg/cpp_gemvtest
Make thread safety tests available to CMAKE and support running only the GEMV version
Martin Kroeker [Thu, 17 Sep 2020 15:28:43 +0000 (17:28 +0200)]
Merge pull request #2843 from mhillenibm/fixup_merge_dynamic_zarch
s390x/DYNAMIC_ARCH: fixup broken merge and reapply simplification
Marius Hillenbrand [Thu, 17 Sep 2020 14:45:07 +0000 (16:45 +0200)]
s390x/DYNAMIC_ARCH: fixup broken merge and reapply simplification
An unrelated commit and merge inadvertently reverted our recent two
changes for simplifying DYNAMIC_ARCH on s390x. Simply reapply the
changes.
Simplify detection of which kernels we can compile on s390x. Instead of
decoding the gcc version in a complicated manner, just check if CC
supports a given -march=archXY flag. Together with the next patch, we
thereby gain support for builds with LLVM/clang with DYNAMIC_ARCH=1.
To enable builds with DYNAMIC_ARCH with older compiler releases, the
Makefile and drivers/other/dynamic_arch.c need a common view of the
architecture support built into the library.
We follow the notation from x86 when used with DYNAMIC_LIST, where
defines DYN_<ARCH NAME> denote support for a given generation to be
built in. Since there are far fewer architecture generations in OpenBLAS
for s390x, that does not bloat command lines too much.
Closes: #2842
Fixes:
ba644378dce7 ("Copy BUILD_ options available to the compiler flags"
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Martin Kroeker [Thu, 17 Sep 2020 11:49:24 +0000 (13:49 +0200)]
Add option for running only the less demanding GEMV version of the thread safety tests
Martin Kroeker [Thu, 17 Sep 2020 11:46:41 +0000 (13:46 +0200)]
Support running just the GEMV version of the thread safety test
Martin Kroeker [Thu, 17 Sep 2020 11:45:40 +0000 (13:45 +0200)]
Add cpp_thread_test options
Martin Kroeker [Thu, 17 Sep 2020 11:43:55 +0000 (13:43 +0200)]
Add CMakeLists.txt
Martin Kroeker [Thu, 17 Sep 2020 11:42:47 +0000 (13:42 +0200)]
Merge pull request #85 from xianyi/develop
rebase
Martin Kroeker [Wed, 16 Sep 2020 16:55:50 +0000 (18:55 +0200)]
Merge pull request #2840 from martin-frbg/fixup2833
Fix for cmake BUILD_ settings PR 2833
Martin Kroeker [Tue, 15 Sep 2020 21:15:34 +0000 (23:15 +0200)]
Activate all BUILD_ options if none was specified
Martin Kroeker [Tue, 15 Sep 2020 21:13:30 +0000 (23:13 +0200)]
Merge pull request #84 from xianyi/develop
rebase
Martin Kroeker [Tue, 15 Sep 2020 19:17:48 +0000 (21:17 +0200)]
Merge pull request #2838 from austinpagan/gordon_trmm
Adding performance patch for trmm, just like trsm (#2836)
fossum [Tue, 15 Sep 2020 13:59:50 +0000 (08:59 -0500)]
Adding performance patch for trmm, just like #2836
Martin Kroeker [Tue, 15 Sep 2020 09:26:37 +0000 (11:26 +0200)]
Merge pull request #2836 from austinpagan/gordon_trsm
Fixing a performance bug in trsm_[LR].c.
fossum [Mon, 14 Sep 2020 18:10:48 +0000 (13:10 -0500)]
Fixing a performance bug in trsm_[LR].c.
Martin Kroeker [Mon, 14 Sep 2020 13:00:19 +0000 (15:00 +0200)]
Merge pull request #2796 from Guobing-Chen/BF16_dot_coversion_apis
Add bfloat16 based dot and conversion with single/double
Martin Kroeker [Mon, 14 Sep 2020 05:24:23 +0000 (07:24 +0200)]
Merge pull request #2833 from martin-frbg/issue2830
Make building the tests for individual data types conditional on the respective BUILD option
Martin Kroeker [Sun, 13 Sep 2020 22:03:33 +0000 (00:03 +0200)]
Copy BUILD_ options available to the compiler flags
Martin Kroeker [Sun, 13 Sep 2020 21:55:11 +0000 (23:55 +0200)]
Add BUILD_SINGLE etc
Martin Kroeker [Sun, 13 Sep 2020 21:29:01 +0000 (23:29 +0200)]
Rearrange ifdefs
Martin Kroeker [Sun, 13 Sep 2020 20:20:41 +0000 (22:20 +0200)]
Remove spurious tests for complex ASUM and NRM2
Martin Kroeker [Sun, 13 Sep 2020 20:17:46 +0000 (22:17 +0200)]
Make tests conditional on BUILD_DOUBLE
Martin Kroeker [Sun, 13 Sep 2020 19:52:18 +0000 (21:52 +0200)]
Make tests for individual variable types conditional on the respective BUILD_ option
Martin Kroeker [Sun, 13 Sep 2020 19:50:12 +0000 (21:50 +0200)]
Make building individual tests depend on BUILD_SINGLE etc defines
Martin Kroeker [Sun, 13 Sep 2020 19:49:01 +0000 (21:49 +0200)]
Remove spurious complex16 tests
Martin Kroeker [Sun, 13 Sep 2020 19:47:55 +0000 (21:47 +0200)]
Copy BUILD_* directives to the compiler options to allow ifdef in tests
Martin Kroeker [Sun, 13 Sep 2020 19:20:30 +0000 (21:20 +0200)]
Merge pull request #2832 from martin-frbg/issue2831
Fix gfortran detection by vendor matching
Martin Kroeker [Sun, 13 Sep 2020 16:40:59 +0000 (18:40 +0200)]
Fix vendor match for GCC gfortran
Martin Kroeker [Sun, 13 Sep 2020 16:30:11 +0000 (18:30 +0200)]
Merge pull request #83 from xianyi/develop
rebase
Martin Kroeker [Tue, 8 Sep 2020 21:36:41 +0000 (23:36 +0200)]
Merge pull request #2829 from mhillenibm/clang_s390x
Fix DYNAMIC_ARCH=1 with clang s390x
Marius Hillenbrand [Tue, 8 Sep 2020 17:30:37 +0000 (19:30 +0200)]
Add an s390 build with clang to the Travis configuration
Since clang builds have been fixed on s390x, including support for
DYNAMIC_ARCH, cover that build type in Travis.
Explicitly request Ubuntu 20.04 (codename focal) to get a recent
LLVM/clang version 10.x and thereby cover all s390x architecture
generations supported in OpenBLAS. Ubuntu 18.10's LLVM/clang 6.x cannot
build the inline assembly in some of the Z13 and Z14 kernels.
LLVM/clang currently does not support OpenMP on s390x, so disable that
in the build.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Marius Hillenbrand [Tue, 8 Sep 2020 13:15:15 +0000 (15:15 +0200)]
Update CONTRIBUTERS.md - clang build fixes for IBM z
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Marius Hillenbrand [Mon, 7 Sep 2020 15:13:03 +0000 (17:13 +0200)]
s390x/DYNAMIC_ARCH: define a HW_CAP flag to support slightly older glibc versions
Enable building DYNAMIC_ARCH support with older versions of glibc that
do not know about the hwcap flag HWCAP_S390_VXE yet.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Marius Hillenbrand [Mon, 7 Sep 2020 15:04:03 +0000 (17:04 +0200)]
s390x/DYNAMIC_ARCH: pass supported arch levels from Makefile to run-time code
... instead of duplicating the (old) mechanism from the Makefile that
aimed to derive supported architecture generations from the gcc
version.
To enable builds with DYNAMIC_ARCH with older compiler releases, the
Makefile and drivers/other/dynamic_arch.c need a common view of the
architecture support built into the library.
We follow the notation from x86 when used with DYNAMIC_LIST, where
defines DYN_<ARCH NAME> denote support for a given generation to be
built in. Since there are far fewer architecture generations in OpenBLAS
for s390x, that does not bloat command lines too much.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Marius Hillenbrand [Fri, 4 Sep 2020 14:32:45 +0000 (16:32 +0200)]
s390x/DYNAMIC_ARCH: generalize detecting supported archs for clang
Simplify detection of which kernels we can compile on s390x. Instead of
decoding the gcc version in a complicated manner, just check if CC
supports a given -march=archXY flag. Together with the next patch, we
thereby gain support for builds with LLVM/clang with DYNAMIC_ARCH=1.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Martin Kroeker [Tue, 8 Sep 2020 08:25:19 +0000 (10:25 +0200)]
Merge pull request #2828 from martin-frbg/lapack438
Correct xLASET arguments in LAPACK EIG tests
Martin Kroeker [Mon, 7 Sep 2020 20:03:46 +0000 (22:03 +0200)]
Correct dimension argument to xLASET
from Reference-LAPACK PR 438
Martin Kroeker [Mon, 7 Sep 2020 19:59:13 +0000 (21:59 +0200)]
Merge pull request #82 from xianyi/develop
rebase
Martin Kroeker [Sun, 6 Sep 2020 16:32:15 +0000 (18:32 +0200)]
Merge pull request #2803 from xiegengxin/AVX2-asum
Implementaion of dasum, sasum with AVX2 & AVX512 intrinsic
Martin Kroeker [Sun, 6 Sep 2020 08:05:47 +0000 (10:05 +0200)]
Merge pull request #2824 from martin-frbg/asumbench
Use POSIX2001 clock.gettime in asum benchmark if available
Martin Kroeker [Sat, 5 Sep 2020 17:44:01 +0000 (19:44 +0200)]
Use POSIX2001 clock.gettime for higher resolution
Martin Kroeker [Sat, 5 Sep 2020 17:17:59 +0000 (19:17 +0200)]
Merge pull request #2816 from martin-frbg/silicon
Add basic support for Apple Vortex (ARM64) cpu
Martin Kroeker [Sat, 5 Sep 2020 15:29:38 +0000 (17:29 +0200)]
Merge pull request #2823 from martin-frbg/fix2778
Improve fix for lapack-test EIG/cchkhb2stg from PR 2778
Martin Kroeker [Sat, 5 Sep 2020 11:06:31 +0000 (13:06 +0200)]
Correct argument to SLASET (Improves fix from PR2778)
as explained by serguei-patchkovskii in Reference-LAPACK/lapack#438 (comment) , passing in an index of 1 instead of N leads to a standards violation accessing matrix A in SLASET, i.e. undefined behavior
Martin Kroeker [Sat, 5 Sep 2020 10:47:03 +0000 (12:47 +0200)]
Merge pull request #81 from xianyi/develop
rebase
Martin Kroeker [Sat, 5 Sep 2020 10:39:32 +0000 (12:39 +0200)]
Merge pull request #2822 from martin-frbg/issue2821
Fix potential domain error in sqrt
Martin Kroeker [Sat, 5 Sep 2020 07:44:33 +0000 (09:44 +0200)]
Fix potentiol domain error in sqrt
Martin Kroeker [Fri, 4 Sep 2020 21:50:43 +0000 (23:50 +0200)]
Merge pull request #2819 from h-vetinari/carry_lapack_437
Carry lapack#437
Martin Kroeker [Fri, 4 Sep 2020 21:09:31 +0000 (23:09 +0200)]
Merge pull request #2820 from RajalakshmiSR/clang
POWER9: Fix mcpu option with clang
Rajalakshmi Srinivasaraghavan [Fri, 4 Sep 2020 15:36:19 +0000 (10:36 -0500)]
POWER9: Fix mcpu option with clang
Adding check for compiler type before checking GCC version in Makefile.
This allows clang to use power9 instead of power8 when CORE is POWER9.
H. Vetinari [Wed, 2 Sep 2020 20:46:47 +0000 (22:46 +0200)]
adapt ?ggsv?-functions to ambient code style in LAPACKE/include/lapack.h
H. Vetinari [Wed, 2 Sep 2020 20:41:50 +0000 (22:41 +0200)]
Follow-up to lapack#434 & lapack#409: add missing 'const' in signatures
Based on how the surrounding functions in lapack.h are handling the
parameters, particularly the ?ggsv?3-variants of the affected functions
H. Vetinari [Wed, 2 Sep 2020 20:38:56 +0000 (22:38 +0200)]
Follow-up to lapack#434 & lapack#409: fix signature mismatches
Martin Kroeker [Fri, 4 Sep 2020 08:06:02 +0000 (10:06 +0200)]
Merge pull request #2778 from martin-frbg/lapackeig
Fix various wrong calls to SLASET/DLASET in the EIG part of the LAPACK testsuite
Chen, Guobing [Wed, 26 Aug 2020 22:42:28 +0000 (06:42 +0800)]
Add bfloat16 based dot and conversion with single/double
1. Added bfloat16 based dot as new API: shdot
2. Implemented generic kernel and cooperlake-specific (AVX512-BF16) kernel for shdot
3. Added 4 conversion APIs for bfloat16 data type <=> single/double: shstobf16 shdtobf16 sbf16tos dbf16tod
shstobf16 -- convert single float array to bfloat16 array
shdtobf16 -- convert double float array to bfloat16 array
sbf16tos -- convert bfloat16 array to single float array
dbf16tod -- convert bfloat16 array to double float array
4. Implemented generic kernels for all 4 conversion APIs, and cooperlake-specific kernel for shstobf16 and shdtobf16
5. Update level1 thread facilitate functions and macros to support multi-threading for these new APIs
6. Fix Cooperlake platform detection/specify issue when under dynamic-arch building
7. Change the typedef of bfloat16 from unsigned short to more strict uint16_t
Signed-off-by: Chen, Guobing <guobing.chen@intel.com>
Martin Kroeker [Thu, 3 Sep 2020 15:10:23 +0000 (17:10 +0200)]
Merge pull request #2817 from martin-frbg/lapack436
LAPACKE: fix declaration of work arrays in [cz]gesvdq
Martin Kroeker [Thu, 3 Sep 2020 06:44:20 +0000 (08:44 +0200)]
Rename KERNEL.SILICON to KERNEL.VORTEX
Martin Kroeker [Thu, 3 Sep 2020 06:43:26 +0000 (08:43 +0200)]
Rename SILICON to VORTEX and fix duplicate numbering
Martin Kroeker [Thu, 3 Sep 2020 06:38:53 +0000 (08:38 +0200)]
Rename SILICON to VORTEX
Martin Kroeker [Thu, 3 Sep 2020 06:38:08 +0000 (08:38 +0200)]
rename SILICON to VORTEX
Gengxin Xie [Tue, 1 Sep 2020 07:41:48 +0000 (15:41 +0800)]
align to 64, using SSE when input size is small
Martin Kroeker [Wed, 2 Sep 2020 21:44:44 +0000 (23:44 +0200)]
Fix data type of work array in zgesvdq prototype
Martin Kroeker [Wed, 2 Sep 2020 21:41:51 +0000 (23:41 +0200)]
Fix data type of rwork array
Martin Kroeker [Wed, 2 Sep 2020 20:56:58 +0000 (22:56 +0200)]
Create KERNEL.SILICON
Martin Kroeker [Wed, 2 Sep 2020 20:52:12 +0000 (22:52 +0200)]
Add AppleSIlicon cpu
Martin Kroeker [Wed, 2 Sep 2020 20:48:49 +0000 (22:48 +0200)]
Add Apple Silicon
Martin Kroeker [Wed, 2 Sep 2020 20:47:38 +0000 (22:47 +0200)]
Detect AppleSilicon cpu on OSX
Martin Kroeker [Wed, 2 Sep 2020 20:16:41 +0000 (22:16 +0200)]
Merge pull request #80 from xianyi/develop
rebase
Martin Kroeker [Wed, 2 Sep 2020 14:56:01 +0000 (16:56 +0200)]
Merge pull request #2815 from mhillenibm/clang_s390x
Fix build with clang on s390x
Marius Hillenbrand [Tue, 1 Sep 2020 14:16:53 +0000 (16:16 +0200)]
s390x: enable S/DGEMM block with explicit loop unrolling + interleaving with clang
The code for SGEMM 16x4 and DGEMM 8x4 blocks on z14 and z15 uses
explicit unrolling and interleaving to improve performance. The code
employs an empty inline asm statement with operands that constrain the
compiler's instruction scheduling and thereby enforce proper overlapping
of load and compute phases. Fix an ifdef to apply that for clang builds,
as well.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Marius Hillenbrand [Tue, 1 Sep 2020 13:09:32 +0000 (15:09 +0200)]
s390x: allow clang to emit fused multiply-adds (replicates gcc's default behavior)
gcc's default setting for floating-point expression contraction is
"fast", which allows the compiler to emit fused multiply adds instead of
separate multiplies and adds (amongst others). Fused multiply-adds,
which assembly kernels typically apply, also bring a significant
performance advantage to the C implementation for matrix-matrix
multiplication on s390x. To enable that performance advantage for builds
with clang, add -ffp-contract=fast to the compiler options.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Marius Hillenbrand [Tue, 1 Sep 2020 10:08:05 +0000 (12:08 +0200)]
s390x: avoid variable-length arrays in struct for asm operands
... since it is not required and clang does not support that gcc
extension. Instead, use a variable-length array directly for these
operands.
Note that, while the actual inline assembly code does not directly use
these memory operands, they serve to inform the compiler that it cannot
reorder reads or writes to/from the input and output data across the
inline asm statements.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Marius Hillenbrand [Tue, 1 Sep 2020 10:04:28 +0000 (12:04 +0200)]
s390x: avoid inline assembly for vector loads for clang
... since clang does not support the instruction format for inline
assembly and also it is not required for current versions of clang.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Marius Hillenbrand [Tue, 1 Sep 2020 09:58:48 +0000 (11:58 +0200)]
s390x: replace nop with "nop 0" in inline assembly
... as a bandaid for building with clang until LLVM's internal assembler
supports nops without operand.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Marius Hillenbrand [Tue, 1 Sep 2020 11:59:06 +0000 (13:59 +0200)]
s390x: use "lghi" for immediate values to fix build with clang
Some of the kernels written in assembly utilize a "load address"
instruction for loading an immediate value into a register. That is
both unnecessarily complex and LLVM's assembler does not understand that
specific syntax. Thus, replace with the appropriate "load immediate"
instruction, which is also clearer to read.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
Martin Kroeker [Tue, 1 Sep 2020 21:39:46 +0000 (23:39 +0200)]
Merge pull request #2813 from martin-frbg/issue2804-2
Fix for c_check misinterpreting arm64 in uname -m output as armv7
Martin Kroeker [Tue, 1 Sep 2020 17:54:08 +0000 (19:54 +0200)]
Fix c_check misinterpreting arm64 in uname output to mean armv7
additionla fix for upcoming OSX on ARM64 related to #2804, as suggested by fxcoudert in #2805
Martin Kroeker [Tue, 1 Sep 2020 15:19:14 +0000 (17:19 +0200)]
Merge pull request #2811 from martin-frbg/issue2806
Make NO_AVX512 option override the AVX512 compile test in CMAKE builds as well
Martin Kroeker [Tue, 1 Sep 2020 14:04:03 +0000 (16:04 +0200)]
Merge pull request #2797 from martin-frbg/relafixes1
ReLAPACK fixes
Martin Kroeker [Tue, 1 Sep 2020 10:03:53 +0000 (12:03 +0200)]
Merge pull request #79 from xianyi/develop
rebase
Martin Kroeker [Tue, 1 Sep 2020 08:44:48 +0000 (10:44 +0200)]
Fix misnaming of LAPACK_?ggsvp function prototypes as LAPACKE_ (#2808)
* Fix misnaming of LAPACK_?ggsvp and ?ggsvd function prototypes as LAPACKE_
* Drop the LAPACKE matrix_layout parameter from the argument lists, change ints to pointers and add missing work arguments.
Martin Kroeker [Mon, 31 Aug 2020 21:44:56 +0000 (23:44 +0200)]
Merge pull request #2807 from martin-frbg/issue2804
Work around ARMV8 build-time cpu detection problems on non-Linux systems
Martin Kroeker [Mon, 31 Aug 2020 18:03:21 +0000 (20:03 +0200)]
Report cpu as ARMV8 instead of just giving up on non-Linux hosts
Martin Kroeker [Mon, 31 Aug 2020 18:02:08 +0000 (20:02 +0200)]
Handle Apple labeling armv8 as arm64 rather than aarch64
Gengxin Xie [Mon, 31 Aug 2020 06:39:08 +0000 (14:39 +0800)]
define __AVX2__ to ensure the haswell code compiled with avx2
Gengxin Xie [Fri, 21 Aug 2020 06:44:36 +0000 (14:44 +0800)]
Implementaion of dasum, sasum with AVX2 & AVX512 intrinsic
Martin Kroeker [Fri, 28 Aug 2020 20:52:11 +0000 (22:52 +0200)]
Merge pull request #2799 from RajalakshmiSR/p10_ger
POWER10: Avoid setting accumulators to zero in gemm kernels
Rajalakshmi Srinivasaraghavan [Fri, 28 Aug 2020 15:42:54 +0000 (10:42 -0500)]
POWER10: Avoid setting accumulators to zero in gemm kernels
For the first iteration, it is better to use xvf*ger instead of xvf*gerpp
builtins which helps to avoid setting accumulators to zero. This helps
to reduce few instructions.
Martin Kroeker [Fri, 28 Aug 2020 06:30:59 +0000 (08:30 +0200)]
Merge pull request #2798 from kadler/aix-cpuid
Fix compile error on AIX cpuid detection
Kevin Adler [Fri, 28 Aug 2020 04:08:33 +0000 (23:08 -0500)]
Fix compile error on AIX cpuid detection
In 589c74a the cpuid detection was changed to use systemcfg, but a copy
and paste error was introduced during some refactoring that caused
POWER7 detection to reference CPUTYPE_POWER7 (which doesn't exist)
instead of CPUTYPE_POWER6.
Martin Kroeker [Thu, 27 Aug 2020 09:25:18 +0000 (11:25 +0200)]
Add early returns and fix sign errors in workspace calculations
Martin Kroeker [Thu, 27 Aug 2020 09:22:50 +0000 (11:22 +0200)]
Add early returns