x86 - How to compile Tensorflow with SSE4.2 and AVX instructions? -


this message received running script check if tensorflow working:

i tensorflow/stream_executor/dso_loader.cc:125] opened cuda library libcublas.so.8.0 locally tensorflow/stream_executor/dso_loader.cc:125] opened cuda library libcudnn.so.5 locally tensorflow/stream_executor/dso_loader.cc:125] opened cuda library libcufft.so.8.0 locally tensorflow/stream_executor/dso_loader.cc:125] opened cuda library libcuda.so.1 locally tensorflow/stream_executor/dso_loader.cc:125] opened cuda library libcurand.so.8.0 locally w tensorflow/core/platform/cpu_feature_guard.cc:95] tensorflow library wasn't compiled use sse4.2 instructions, these available on machine , speed cpu computations. w tensorflow/core/platform/cpu_feature_guard.cc:95] tensorflow library wasn't compiled use avx instructions, these available on machine , speed cpu computations. tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:910] successful numa node read sysfs had negative value (-1), there must @ least 1 numa node, returning numa node 0

i noticed has mentioned sse4.2 , avx,

1) sse4.2 , avx?

2) how these sse4.2 , avx improve cpu computations tensorflow tasks.

3) how make tensorflow compile using 2 libraries?

i ran same problem, seems yaroslav bulatov's suggestion doesn't cover sse4.2 support, adding --copt=-msse4.2 suffice. in end, built with

bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=cuda -k //tensorflow/tools/pip_package:build_pip_package

without getting warning or errors.

probably best choice system is:

bazel build -c opt --copt=-march=native --copt=-mfpmath=both --config=cuda -k //tensorflow/tools/pip_package:build_pip_package

-mfpmath=both works gcc, not clang. i'm not sure default -o2 or -o3 setting is. gcc -o3 enables auto-vectorization, won't help, , make slower.

what does: --copt bazel build passes option directly gcc compiling c , c++ files (but not linking, need different option cross-file link-time-optimization)

x86-64 gcc defaults using sse2 or older simd instructions, can run binaries on any x86-64 system. (see https://gcc.gnu.org/onlinedocs/gcc/x86-options.html). that's not want. want make binary takes advantage of instructions cpu can run, because you're running binary on system built it.

-march=native enables options cpu supports, makes -mavx512f -mavx2 -mavx -mfma -msse4.2 redundant. (also, -mavx2 enables -mavx , -msse4.2, yaroslav's command should have been fine). if you're using cpu doesn't support 1 of these options (like fma), using -mfma make binary faults illegal instructions.

tensorflow's ./configure defaults enabling -march=native, using should avoid needing specify compiler options manually.

-march=native enables -mtune=native, it optimizes cpu things sequence of avx instructions best unaligned loads.

this applies gcc, clang, or icc.


-

Comments

Post a Comment

Popular posts from this blog

python - How to insert QWidgets in the middle of a Layout? -

Image
i'm using qt framework build graphical user interface. use qgridlayout position qwidgets neatly. gui looks this: my application regularly adds new widgets gui @ runtime. these new widgets not added @ end of qlayout, somewhere in middle. the procedure bit cumbersome. applied on figure above, need take out widg_c , widg_d , ... qgridlayout. next, add widg_x and widg_y , , put other widgets again. how remove widgets qgridlayout: for in reversed(range(mygridlayout.count())): self.itemat(i).widget().setparent(none) ### as long you're dealing small amount of widgets, procedure not disaster. in application display lot of small widgets - perhaps 50 or more! application freezes second while procedure ongoing, annoying user. there way insert widgets somewhere in qlayout, without need take out other widgets? edit: apparently solution qvboxlayout simple. use function insertwidget(..) instead of addwidget(..) . docs can found link: http://doc.qt.io/qt-5/qboxlayou
Read more

python - serve multiple gunicorn django instances under nginx ubuntu -

i need serve 2 webs apps (django) on gunicorn under different domain names using nginx on ubuntu. started using tutorial: https://www.digitalocean.com/community/tutorials/how-to-set-up-django-with-postgres-nginx-and-gunicorn-on-ubuntu-16-04 it worked fine 1 using that. tried same thing second gives me 502 gateway second domain. tried following tutorial: http://michal.karzynski.pl/blog/2013/10/29/serving-multiple-django-applications-with-nginx-gunicorn-supervisor/ i'm @ part run gunicorn start script: sudo bin/gunicorn_start i response: starting webuildblack root traceback (most recent call last): file "/home/devin/webuildblack/bin/gunicorn", line 7, in <module> gunicorn.app.wsgiapp import run importerror: no module named 'gunicorn.app'; 'gunicorn' not package you have named program gunicorn.py gunicorn -package. rename file , remove gunicorn.pyc file.
Read more

module - Prestashop displayPaymentReturn hook url -

i new prestashop developer , trying create paymentmodule. have got show payment method can not proceed purchase because not know hot works. does know should redirect run hooddisplaypaymentreturn method? i happy if explain me complete navigation map make purchase. anyway, can find relation between hooks , pages? when have new module payment rely on simplest provided prestashop: bankwire. inside can find 3 hooks. hookpayment: public function hookpayment($params) { if (!$this->active) return; if (!$this->checkcurrency($params['cart'])) return; $this->smarty->assign(array( 'this_path' => $this->_path, 'this_path_bw' => $this->_path, 'this_path_ssl' => tools::getshopdomainssl(true, true).__ps_base_uri__.'modules/'.$this->name.'/' )); return $this->display(__file__, 'payment.tpl'); } hookdisplaypaymenteu: public fun
Read more