PR925 #32
Open
PR925 #32
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
… or draw. (LeelaChessZero#822) * Also make parents Terminal if any move is a win or all moves are loss or draw. * Add --sticky-endgames param to control converting parents to terminals. * Switch back to handwritten loop instead of std::all_of. * Move MakeNotTerminal call into MakeMove from ResetToPosition. * Recompute eval in MakeNotTerminal. * Address review comments to MakeNotTerminal.
* Recalibrate centipawn formula - based on Bluefish, Leela, Komodo and Houdini evals in TCEC 15 divP - with this change, Leela's eval will shrink by a factor of ~2.6 in most cases - mate scores remain at 12800 or -12800 cp * Add switchable centipawn mode to preserve the old conversion formula.
…essZero#844) * Make --sticky-endgames on by default (still off in training) * Fix line split for the flag description.
…s to experimental RPM packages (LeelaChessZero#675) * Create openSUSE_install.md Extensive and complete installation instructions for installing lc0 on openSUSE. If accepted, recommend creating a link from README.md to this file. This Pull Request should include this file and "install_openSUSE_lc0.sh * Create install_openSUSE_lc0.sh Unattended compile script for install lc0 on any version openSUSE * Modify README.md to point to openSUSE_Install page * Numerous modifications to openSUSE_install.md include RPM and logo * Modified logo and page title placement to stacked centered * modify logo parameters * Numerous typos and a few minor formatting changes * Update openSUSE_install.md Minor typo change to "## An example setup with the Arena graphical chessboard" * Update openSUSE_install.md Minor change - Hyperlinked "RPM packages" in "RPM Packages vs Building from Source" section * Update openSUSE_install.md Attempt to fix minor formatting problem when viewed in Firefox (no problem any other browser) * Update openSUSE_install.md Test - Make the link to the downloadable script a variable path in hopes the hyperlink will work in any repo * Update openSUSE_install.md An important addition to setting up Arena, which is to re-configure permissions on ~/.configure/ This has nothing to do specifically with lc0 * Update openSUSE_install.md Fix statement changing from fixed path specifying a User to variable so will work with any User account * enabled executable permission on script by default * Modify README.md link to openSUSE_install.md hyperlink from a fixed path URL to variable path so will work in any repo * Add text instructing how to modify execute permission on script. Unknown why this may be required, has not be necessary in previous cases * Big augment RPN instructions plus minor typos * Update openSUSE_install.md Correct very minor typo * Update openSUSE_install.md Fix Arena code Engine > Engines * Update openSUSE_install.md Fix another Arena code Engine > Engines * Update openSUSE_install.md Update Text describing Arena chessboard * Modify logo URL to comply with change * Second modification to logo URL * update RPM info and instructions from preliminary to Production * remove temporary files created by text editor
…ero#849) * Support cudnn-fp16 backend on GPUs without tensor cores - try NCHW layout and winograd alogirhtm for convolutions (same as what we use for fp32). - it's slower than NHWC/fp16 on GPUs with tensor cores, but should give some speedup on GP100 and TU11x GPUs. * add check for cards with no tensor cores - GP100 (SM6.0) - GTX 16xx GPUs (unfortunately same sm 7.5 version so need a string compare) * add backend-opt to force nhwc on or off
* SE layer fix when not using fused kernel - only SE layer needs transposed weights. - need to store non-transposed weights too in case we have to fall back.
* Add 320 and 352 channel support for fused SE layer - just add template instantiations. - verified that it works and provides a (very) slight speedup. * Update fp16_kernels.cu
* Non-tangent Centipawn Conversion This conversion formula should be a better match near the asymptotes. * Added coefficient to the denominator This coefficient makes the max score top out near +128 (12826 cp, more precisely). This also fixes the divide-by-zero problem at x = +/- 1. * Exponent tweak and coefficient precision The coefficient 0.976953125 = 1 - 295/12800.
* import Eigen 3.3.7
Hmm I'm not entirely sure this is right.
…ssZero#866) * Revert "fix WinogradConvolution3 allocation size (LeelaChessZero#864)" This reverts commit 6a425b5. * better fix for WinogradConvolution3 allocation size
* Set probabilities for illegal moves in a loop Set probabilities for illegal moves in a loop instead of using std::memset. * Use std::fill instead of a for loop Use std::fill instead of a for loop during moves probabilities array initialization.
* fix maxSize for tensor buffers - no of filters can be less than no. of input planes! * fix calculation for value head allocation too - need 3x the size for wdl * take max of all layers when computing size to avoid assumptions
* change protobuf appveyor build (remove -Dprotobuf_MSVC_STATIC_RUNTIME=NO) * clear appveyor cache * set static/shared crt option for nvcc according to default_library and meson version (for fixed bug)
…Zero#821) * Hack to report average nodes per move for monitoring training. * WDLstyle resign false positive reporting * Make this less WIP and more submittable. * Cast correctly. * Fix typo in last commit... * Add comment about unsupported combination.
* Do softmax outside backend on set of legal moves. * Remove policy softmax from blas backend. * Remove policy softmax from CUDA backend. * Remove policy softmax from OpenCL backend. * Remove policy softmax from TensorFlow backend. * Use FastExp for policy softmax calculations. * Fix for negative exponentials. * Revert "Fix for negative exponentials." This reverts commit 9fb73d0. * Fuse softmax with softmax temperature. * Modify random backend policy value distribution. * Comment improvements.
* v0.22.0-dev -> v0.23.0-dev * More changes to the changelog. * Reorder changelog.
* Add git short revision to dev version string. * Address review comments. * Updated comment. * Don't let build fail if git is not installed. * Print warning if git is not installed. * Clarified warning. * Address review comments. * Updated handling of submodules + added warning. * Check for dirty working directory. * Extract revision after automatically downloading submodules to prevent 'dirty' flag. * Fix build error. * Print build identifier as message. * Revert "Fix build error." This reverts commit 37cec6f. * Properly fix build error.
Do we need to worry about fpu here?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
For testing for now