Devices and external libraries options, make install, Makefile fix #4
Conversation
nicolaspi
commented
Oct 9, 2017
nicolaspi
commented
- New features that allow to generate code according to operator's devices compatibility (cpu and/or gpu), and to specify external libraries to be linked.
- The Makefile has been fixed to work with the latest tensorflow version (tested on the version 1.3 built from source).
- Added 'make install' that copies the built library to the tensorflow's library directory.
and external library links. (see "get_slide_patch.yml" example) - Fix Makefile - Added an install option in the Makefile that copies the user's lib in the tensorflow's lib directory - Fix unit test template.
|
Looks like a version issue with tensorflow, the linker does not find |
|
Thanks for the PR. It looks like the new |
| @@ -8,7 +8,7 @@ | |||
|
|
|||
| def make_template_kwargs(op_name, py_op_name, | |||
| project, library, op_inputs, op_outputs, | |||
| op_type_attrs, op_other_attrs, op_doc): | |||
| op_type_attrs, op_other_attrs, op_doc, op_libs, op_devices_comps, op_abi): | |||
There was a problem hiding this comment.
Move op_libs, op_devices_comps, op_abi to a newline.
| - "gpu" | ||
| extra_libraries_deps: | ||
| - "m" | ||
| use_old_abi: yes | ||
|
|
There was a problem hiding this comment.
I think if we start mixing the Makefile options with the operator options, the yaml will start to be cluttered.
How about a makefile section to distinguish it from the rest of the yaml?
makefile:
use_old_abi: yes
libraries:
- "m"| @@ -73,6 +75,12 @@ For example, we can define the outline for a ``ComplexPhase`` operator in the `` | |||
| (2) of (L, M) sky coordinates with shape (nsrc, 2) | |||
| (3) of frequencies, | |||
| compute the complex phase with shape (nsrc, ntime, nbl, nchan) | |||
| devices_compatibilities: | |||
There was a problem hiding this comment.
Rename devices_compatibilities to devices
| @@ -49,6 +49,8 @@ The user should provide a YAML configuration file defining the operator: | |||
| - polymorphic type attributes. | |||
| - other attributes. | |||
| - documentation. | |||
| - devices compatibilities. | |||
| - optional external libraries. | |||
There was a problem hiding this comment.
device compatibilities to devices
optional external libraries to makefile options
| @@ -147,6 +155,33 @@ generator code as the range of attribute behaviour is complex. | |||
| op_other_attrs: | |||
| - "iterations: int32 >= 2", | |||
|
|
|||
| Sources codes will be generated according to the specified devices. | |||
There was a problem hiding this comment.
Source code will be generated for the requested devices
| devices_compatibilities: | ||
| - "cpu" | ||
| - "gpu" | ||
| extra_libraries_deps: |
There was a problem hiding this comment.
extra_library_deps to libraries
| extra_libraries_deps: | ||
| - "openslide" | ||
| devices_compatibilities: | ||
| - "cpu" |
| @@ -24,7 +24,7 @@ class Test{{op_name}}(unittest.TestCase): | |||
| {% set comma = last_joiner(",") -%} | |||
| type_permutations = [ | |||
| {%- for perm in op_np_type_perms %} | |||
| {{ perm | replace("'", "") | indent(4, True) }}{{comma(loop.last)}} | |||
| {{ perm[0] | replace("'", "") | indent(4, True) }}{{comma(loop.last)}} | |||
| The binary pip packages available on the TensorFlow website are built | ||
| with gcc4 that uses the older ABI. Set the option ``use_old_abi`` with the value ``yes`` | ||
| to allow the library to be compatible with the older abi. | ||
|
|
There was a problem hiding this comment.
Change to:
Note that gcc version 5 and above uses a new C++ ABI.
However, the official tensorflow binary pip packages are built with gcc 4
which uses an older ABI. Set the option `use_old_abi: yes` to make
the library compatible with the older ABI.
| CPPFLAGS=-std=c++11 $(TF_FLAGS) $(INCLUDES) -fPIC -fopenmp -O2 -march=native -mtune=native | ||
| NVCCFLAGS=-std=c++11 -D GOOGLE_CUDA=$(TF_CUDA) $(TF_FLAGS) $(INCLUDES) \ | ||
| -x cu --compiler-options "-fPIC" --gpu-architecture=sm_30 -lineinfo | ||
|
|
||
| LDFLAGS = -fPIC -fopenmp | ||
| LDFLAGS = -fPIC -fopenmp -L$(TF_LIB) -ltensorflow_framework{{extra_libs}} | ||
|
|
There was a problem hiding this comment.
One possible way of handling the the presence or absence of the -ltensorflow_framework library would be to add tensorflow_version to the YAML.
