Could port to OpenCL? #28
Comments
|
Thanks for the feedback -- this is tracked in #22 |
benoitsteiner
pushed a commit
to benoitsteiner/tensorflow
that referenced
this issue
May 22, 2017
* Fixed AVX-512 intrinsic implementation. * OR'ed LIBXSMM_DNN_CONV_OPTION_OVERWRITE into convolution options, which folds zeroing the input buffer on first use. This removes the call to libxsmm_dnn_zero_buffer in case of LIBXSMM_DNN_COMPUTE_KIND_FWD. * Rely on libxsmm_hash rather than std::hash. Brought xsmm_conv2d.cc up-to-date with TF/master. * Code cleanup: use LIBXSMM_DNN_CONV_OPTION_WU_EXT_FILTER_REDUCE_OVERWRITE rather than assembling the option from separate flags. * Avoid to destroy the handle in case of LIBXSMM_DNN_WARN_FALLBACK since the next iteration may double-delete the same handle. One would need to update the handle-cache to allow destruction at this place. However, all handles are destructed when TF terminates (cache cleanup). * Rely on default configuration arguments, and thereby lower the dependence from LIBXSMM internals.
benoitsteiner
pushed a commit
to benoitsteiner/tensorflow
that referenced
this issue
May 22, 2017
* Fixed AVX-512 intrinsic implementation. * OR'ed LIBXSMM_DNN_CONV_OPTION_OVERWRITE into convolution options, which folds zeroing the input buffer on first use. This removes the call to libxsmm_dnn_zero_buffer in case of LIBXSMM_DNN_COMPUTE_KIND_FWD. * Rely on libxsmm_hash rather than std::hash. Brought xsmm_conv2d.cc up-to-date with TF/master. * Code cleanup: use LIBXSMM_DNN_CONV_OPTION_WU_EXT_FILTER_REDUCE_OVERWRITE rather than assembling the option from separate flags. * Avoid to destroy the handle in case of LIBXSMM_DNN_WARN_FALLBACK since the next iteration may double-delete the same handle. One would need to update the handle-cache to allow destruction at this place. However, all handles are destructed when TF terminates (cache cleanup). * Rely on default configuration arguments, and thereby lower the dependence from LIBXSMM internals.
benoitsteiner
added a commit
that referenced
this issue
May 22, 2017
* Fixed AVX-512 intrinsic layer (sparse_matmul_op.h). Incorporated LIBXSMM_DNN_CONV_OPTION_OVERWRITE. (#26) * Fixed AVX-512 intrinsic implementation. * OR'ed LIBXSMM_DNN_CONV_OPTION_OVERWRITE into convolution options, which folds zeroing the input buffer on first use. This removes the call to libxsmm_dnn_zero_buffer in case of LIBXSMM_DNN_COMPUTE_KIND_FWD. * Made xsmm_conv2d.cc up-to-date with TF/master, avoid double-free in case of LIBXSMM_DNN_WARN_FALLBACK, use libxsmm_hash instead of std::hash, code cleanup (#27) * Fixed AVX-512 intrinsic implementation. * OR'ed LIBXSMM_DNN_CONV_OPTION_OVERWRITE into convolution options, which folds zeroing the input buffer on first use. This removes the call to libxsmm_dnn_zero_buffer in case of LIBXSMM_DNN_COMPUTE_KIND_FWD. * Rely on libxsmm_hash rather than std::hash. Brought xsmm_conv2d.cc up-to-date with TF/master. * Code cleanup: use LIBXSMM_DNN_CONV_OPTION_WU_EXT_FILTER_REDUCE_OVERWRITE rather than assembling the option from separate flags. * Avoid to destroy the handle in case of LIBXSMM_DNN_WARN_FALLBACK since the next iteration may double-delete the same handle. One would need to update the handle-cache to allow destruction at this place. However, all handles are destructed when TF terminates (cache cleanup). * Configure LIBXSMM with default arguments (#28) * Fixed AVX-512 intrinsic implementation. * OR'ed LIBXSMM_DNN_CONV_OPTION_OVERWRITE into convolution options, which folds zeroing the input buffer on first use. This removes the call to libxsmm_dnn_zero_buffer in case of LIBXSMM_DNN_COMPUTE_KIND_FWD. * Rely on libxsmm_hash rather than std::hash. Brought xsmm_conv2d.cc up-to-date with TF/master. * Code cleanup: use LIBXSMM_DNN_CONV_OPTION_WU_EXT_FILTER_REDUCE_OVERWRITE rather than assembling the option from separate flags. * Avoid to destroy the handle in case of LIBXSMM_DNN_WARN_FALLBACK since the next iteration may double-delete the same handle. One would need to update the handle-cache to allow destruction at this place. However, all handles are destructed when TF terminates (cache cleanup). * Rely on default configuration arguments, and thereby lower the dependence from LIBXSMM internals.
benoitsteiner
pushed a commit
to benoitsteiner/tensorflow
that referenced
this issue
Jun 15, 2017
* Fixed AVX-512 intrinsic implementation. * OR'ed LIBXSMM_DNN_CONV_OPTION_OVERWRITE into convolution options, which folds zeroing the input buffer on first use. This removes the call to libxsmm_dnn_zero_buffer in case of LIBXSMM_DNN_COMPUTE_KIND_FWD. * Rely on libxsmm_hash rather than std::hash. Brought xsmm_conv2d.cc up-to-date with TF/master. * Code cleanup: use LIBXSMM_DNN_CONV_OPTION_WU_EXT_FILTER_REDUCE_OVERWRITE rather than assembling the option from separate flags. * Avoid to destroy the handle in case of LIBXSMM_DNN_WARN_FALLBACK since the next iteration may double-delete the same handle. One would need to update the handle-cache to allow destruction at this place. However, all handles are destructed when TF terminates (cache cleanup). * Rely on default configuration arguments, and thereby lower the dependence from LIBXSMM internals.
tarasglek
pushed a commit
to tarasglek/tensorflow
that referenced
this issue
Jun 20, 2017
Fix typo mistake, fixes tensorflow#27.
lukeiwanski
referenced
this issue
in codeplaysoftware/tensorflow
Oct 26, 2017
* [OpenCL] Registers Conv2DBackpropFilter * Aligned '\'
hfp
added a commit
to hfp/tensorflow
that referenced
this issue
Jan 4, 2019
* Fixed AVX-512 intrinsic implementation. * OR'ed LIBXSMM_DNN_CONV_OPTION_OVERWRITE into convolution options, which folds zeroing the input buffer on first use. This removes the call to libxsmm_dnn_zero_buffer in case of LIBXSMM_DNN_COMPUTE_KIND_FWD. * Rely on libxsmm_hash rather than std::hash. Brought xsmm_conv2d.cc up-to-date with TF/master. * Code cleanup: use LIBXSMM_DNN_CONV_OPTION_WU_EXT_FILTER_REDUCE_OVERWRITE rather than assembling the option from separate flags. * Avoid to destroy the handle in case of LIBXSMM_DNN_WARN_FALLBACK since the next iteration may double-delete the same handle. One would need to update the handle-cache to allow destruction at this place. However, all handles are destructed when TF terminates (cache cleanup). * Rely on default configuration arguments, and thereby lower the dependence from LIBXSMM internals.
eggonlea
pushed a commit
to eggonlea/tensorflow
that referenced
this issue
Mar 12, 2019
Revert "Add Boost.Locale as a dependency for the build"
cjolivier01
pushed a commit
to Cerebras/tensorflow
that referenced
this issue
Dec 6, 2019
…ch_to_eigen_fork Switch to using the ROCm fork for eigen.
nammbash
referenced
this issue
in Intel-tensorflow/tensorflow
May 18, 2020
Closed
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Could the project be port to OpenCL? Currently it only supports CUDA which is proprietary. OpenCL can be more wide spread and portable.
The text was updated successfully, but these errors were encountered: