1. Add a new API -- sbgemv to support bfloat16 based gemv
2. Implement a generic kernel for sbgemv
3. Implement an avx512-bf16 based kernel for sbgemv
Signed-off-by: Chen, Guobing <guobing.chen@intel.com>
This initializes the per-thread memory buffers which get
cleared/released on a fork via pthread_at_fork. Not doing so leads to
each thread calling blas_memory_alloc on almost every execution which
slows down the code significantly as the threads race for the memory
allocation using locks to serialize that.
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>
... 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>
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>
In current code, no matter what number of threads specified, all
available CPU count is used when invoking OMP, which leads to very bad
performance if the workload is small while all available CPUs are big.
Lots of time are wasted on inter-thread sync. Fix this issue by really
using the number specified by the variable 'num' from calling API.
Signed-off-by: Chen, Guobing <guobing.chen@intel.com>
Enable new build target platform -- COOPERLAKE. This target platform
supports all the SKYLAKEX supported ISAs + avx512bf16. So all the
SKYLAKEX specific kernels/drivers and related code are now extended
to be also active on COOPERLAKE. Besides, new BF16 related kernels
are active under this target.
Extend and simplify the run-time detection for dynamic architecture support for z
to check HW_CAP and only use SIMD features if advertised by the OS.
While at it, also honor the env variable LD_HWCAP_MASK and do not use
the CPU features masked there.
Note that we can only use the SIMD features on z13 or newer (i.e.,
Vector Facility or Vector-Enhancements Facilities) when the operating
system supports properly context-switching the vector registers. The OS
advertises that support as a bit in the HW_CAP value in the auxiliary
vector. While all recent Linux kernels have that support, we should
maintain compatibility with older versions that may still be in use.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
When building OpenBLAS with DYNAMIC_ARCH=1 on s390x (aka zarch), make
sure to include support for systems without the facilities introduced
with z13 (i.e., zarch_generic). Adjust runtime detection to fallback to
that generic code when running on a unknown platform other than Z13
through Z15.
When detecting a Z13 or newer system, add a check for gcc support for
the architecture-specific features before selecting the respective
kernel. Fallback to Z13 or generic code, in case.
Signed-off-by: Marius Hillenbrand <mhillen@linux.ibm.com>
* make building the bfloat16 BLAS functions conditional on BUILD_HALF
* pass the BUILD_HALF option to gensymbol
* Pass BUILD_HALF as a compiler define for dynamic_arch builds
This patch adds support for bfloat16 data type matrix multiplication kernel.
For architectures that don't support bfloat16, it is defined as unsigned short
(2 bytes). Default unroll sizes can be changed as per architecture as done for
SGEMM and for now 8 and 4 are used for M and N. Size of ncopy/tcopy can be
changed as per architecture requirement and for now, size 2 is used.
Added shgemm in kernel/power/KERNEL.POWER9 and tested in powerpc64le and
powerpc64. For reference, added a small test compare_sgemm_shgemm.c to compare
sgemm and shgemm output.
This patch does not cover OpenBLAS test, benchmark and lapack tests for shgemm.
Complex type implementation can be discussed and added once this is approved.
* Add another memory barrier in memory.c to prevent races in memory slot allocation
* Add an all-core test on Drone.io's ThunderX platform and modify dgemm_tester to use all 96 cores
* Add an ARMV7 iOS build on Travis
* thread_local appears to be unavailable on ARMV7 iOS
* Add no-thumb option for ARMV7 IOS build to get it to accept DMB ISH
* Make local labels in macros of nrm2_vfpv3.S compatible with the xcode assembler
Setting DYNAMIC_ARCH=1 on POWER9 does not build POWER9 files due to some
compiler version checks. This patch fixes some of the macros that are used
to check compiler version. On fixing those checks, there are some new make
failures related to icamin, icamax, isamin, isamax and caxpy files on POWER9.
This patch fixes those failures as well.
Heavy-weight locking isn't required to pass the work queue
pointer between threads and simple atomic acquire/release
semantics can be used instead. This is especially important as
pthread_mutex_lock() isn't fair.
We've observed substantial variation in runtime because of the
the unfairness of these locks which complety goes away with
this implementation.
The locks themselves are left to provide a portable way for
idling threads to sleep/wakeup after many unsuccessful iterations
waiting.
It used to be 5ms, which might not be long enough in some cases for the
thread to exit well, but then when set to 5000 (5s), it would slow down
any program depending on OpenBlas.
Let's just set it to 50ms, which is at least 10 times longer than
originally, but still reasonable in case of failed thread termination.
* Update DYNAMIC_ARCH list of ARM64 targets for gmake
* Update arm64 cpu list for runtime detection
* Update DYNAMIC_ARCH list of ARM64 targets for cmake and add POWERPC targets
This patch was submitted to the GIMP project by a publisher wishing to
keep confidentiality (hence anonymously). I just pass along the patch.
Here is the patch explanation which came with:
First they remind us what Microsoft documentation says about
TerminateThread:
> TerminateThread is a dangerous function that should only be used in
> the most extreme cases. You should call TerminateThread only if you
> know exactly what the target thread is doing, and you control all of
> the code that the target thread could possibly be running at the time
> of the termination.
(https://docs.microsoft.com/en-us/windows/win32/api/processthreadsapi/nf-processthreadsapi-terminatethread)
Then they say that 5 milliseconds time-out might not be long enough for
the thread to exit gracefully. They propose to set it to a much higher
value (for instance here 5 seconds).
And finally you should always check the return value of
WaitForSingleObject(). In particular you want to run TerminateThread()
only if WaitForSingleObject() failed, not on success case.
1. qalloc() appears to have been a special routine written for the PPC440-based QCDOC supercomputer(s) from around 2005, its source does not seem to be readily available. So switch the #if 1 in the code to rely on standard malloc() by default.
2. Utility functions like get_num_procs, get_num_threads that were added to the "normally" used memory.c in the meantime were still missing here.
* Add automatic fixups for OSX/IOS cross-compilation
* Add OSX/IOS cross-compilation test to Travis CI
* Handle platforms that lack hwcap.h by falling back to ARMV8
* Fix PROLOGUE for OSX/IOS
* Add gcc7-generated assembly files for POWER8/9 isa/ica-min/max and POWER9 caxpy
To work around internal compiler errors encountered when compiling the original C source with gcc 4 and 5, and wrong code generated by gcc 8.3.0
* Use gcc-generated assembly instead of original C sources
to work around internal compiler errors encountered with gcc 4.8/5.4 and wrong code generation by gcc 8.3
* Use gcc-generated assembly instead of the original C source
to work around internal compiler errors encountered with gcc 4.8 and 5.4, and wrong code generation by gcc 8.3
* Add gcc7-generated assembler version of caxpy for power8
to work around wrong code generated by gcc 8.3
* Handle CONJ define for caxpyc
* Handle CONJ define for caxpyc
* Add gcc7-generated assembly cdot for POWER9
* Use prebuilt assembly for POWER9 cdot
created with gcc 7.3.1 to work around ICE in older gcc versions
* Exclude POWER9 from DYNAMIC_ARCH when gcc versions is lower than 6
* Update Makefile.system
* Use PROLOGUE macro to ensure correct function name for DYNAMIC_ARCH
* Disable POWER9 with old gcc versions
* Fix potential memory leak in cpu enumeration with glibc
An early return after a failed call to sched_getaffinity would leak the previously allocated cpu_set_t. Wrong calculation of the size argument in that call increased the likelyhood of that failure. Fixes#2003
* Restore dropped patches in the non-TLS branch of memory.c
As discovered in #2002, the reintroduction of the "original" non-TLS version of memory.c as an alternate branch had inadvertently used ba1f91f rather than a8002e2 , thereby dropping the commits for #1450, #1468, #1501, #1504 and #1520.
The current gemm threading code can make very unfortunate choices, for
example on my 10 core system a 1024x1024x1024 matrix multiply ends up
chunking into blocks of 102... which is not a vector friendly size
and performance ends up horrible.
this patch adds a helper define where an architecture can specify
a preference for size multiples.
This is different from existing defines that are minimum sizes and such.
The performance increase with this patch for the 1024x1024x1024 sgemm
is 2.3x (!!)
Enable DYNAMIC_ARCH feature on ARM64. This patch uses the cpuid
feature in linux kernel to detect the core type at runtime
(https://www.kernel.org/doc/Documentation/arm64/cpu-feature-registers.txt).
If this feature is missing in kernel, then the user should use the
OPENBLAS_CORETYPE env variable to select the desired core type.
Currently the generic ARMV8 target uses C implementations
for many routines. Replace these with the neon implementations
written for THUNDERX2T99 target which are upto 6x faster for
certain routines.
Since we now use an allocation size that isn't a multiple of PAGESIZE, finding
the pages for run_bench wasn't terminating properly. Now we detect if we've
found enough pages for the allocation and terminate the loop.
- Compiler TLS is now used only used when the compiler supports it
- If compiler TLS is unsupported, we use platform-specific TLS
- Only one variable (an index) is now in TLS
- We only access TLS once per alloc, and never when freeing
- Allocation / release info is now stored within the allocation itself, by
over-allocating; this saves having external structures do the bookkeeping, and
reduces some of the redundant data that was being stored (such as addresses)
- We never hit the alloc lock when not using SMP or when using OpenMP (that was
my fault)
- Now that there are fewer tracking structures I think this is a bit easier to
read than before
a few places in the gemm scheduler code were missing barriers;
the code likely worked OK due to heavy use of volatile / _Atomic
but there's no reason to get this incorrect
The use of _Atomic leads to really bad code generation in the compiler
(on x86, you get 2 "mfence" memory barriers around each access with gcc8, despite
x86 being ordered and cache coherent). But there's a fallback in the code that
just uses volatile which is more than plenty in practice.
If we're nervous about cross thread synchronization for these variables, we should
make the YIELD function be a compiler/memory barrier instead.
performance before (after last commit)
Matrix SGEMM cycles MPC DGEMM cycles MPC
48 x 48 10630.0 10.6 0.7% 18112.8 6.2 -0.7%
64 x 64 20374.8 13.0 1.9% 40487.0 6.5 0.4%
65 x 65 141955.2 1.9 -428.3% 146708.8 1.9 -179.2%
80 x 80 178921.1 2.9 -369.6% 186032.7 2.8 -156.6%
96 x 96 205436.2 4.3 -233.4% 224513.1 3.9 -97.0%
112 x 112 244408.2 5.8 -162.7% 262158.7 5.4 -47.1%
128 x 128 321334.5 6.5 -141.3% 333829.0 6.3 -29.2%
Performance with this patch (roughly a 2x improvement):
Matrix SGEMM cycles MPC DGEMM cycles MPC
48 x 48 10756.0 10.5 -0.5% 18296.7 6.1 -1.7%
64 x 64 20490.0 12.9 1.4% 40615.0 6.5 0.0%
65 x 65 83528.3 3.3 -210.9% 96319.0 2.9 -83.3%
80 x 80 101453.5 5.1 -166.3% 128021.7 4.0 -76.6%
96 x 96 149795.1 5.9 -143.1% 168059.4 5.3 -47.4%
112 x 112 191481.2 7.3 -105.8% 204165.0 6.9 -14.6%
128 x 128 265019.2 7.9 -99.0% 272006.4 7.7 -5.3%
Martin Kroeker
Gengxin Xie
Guillaume Horel
Chen, Guobing
Rajalakshmi Srinivasaraghavan
User User-User
Alexander Grund
y00512012
fossum
Marius Hillenbrand
Ashwin Sekhar T K
Matthew Treinish
Rajalakshmi Srinivasaraghavan
Sharvil Nanavati
Ali Saidi
Susi Lehtola
wjc404
int_13h
Jehan
Sebastian Berg
Andrew
luz.paz
Jeff Baylor
Rashmica Gupta
Erik M. Bray
caiyu
Arjan van de Ven
Ashwin Sekhar T K
Zoltán Mizsei
Craig Donner
oon3m0oo