Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
* Remove .DS_Store * Initial code for loading MFA * Fix error messages * Attempt to compile C++ code * Remove unused dependency * [PATCH] Fix cpp compilation errors. * Fix error output * Refactor the types * Dispatch C++ bindings * Add hasher * Refactor MFA errors * Skeleton implementation of 'encode_gemm' * Calculate threadgroup memory and grid XY * Finish draft of encoding code * MTL::CommandBatch * Detect compatible MFA GEMMs * Add bias restriction * Fix up MTL::CommandBatch * Fix typo * 48x48 is the most appropriate default for FP32 * It is running * Introduce max stream count when schedule a graph. * Avoid stack allocation if it is too large. * Don't allocate gradients if it doesn't compute. * No gradients are allocated if these are not trainable. * Fix up the map * Document the block sizes * Block size selection heuristic * Preparation for batching * Add code for encoding the matrix offsets in a batch * Support a subset of batching * Caught a bug in the dispatching code * Move files around to make compilation happy. * Regenerate configure file. * Add unordered map import for ccv_nnc_mfa.hpp * Style updates for ccv_nnc_gemm_mps. * Revert an update to lib/ccv.h * Move 2-space to tab. * Minor change to force init mps. * Commit the lib file and pass in the path. * Gate with OSX, MAC is 1 on both iOS and Mac. * call useResource on the buffers. * Update lib/nnc/mfa/ccv_nnc_mfa_gemm.cpp Co-authored-by: Philip Turner <[email protected]> --------- Co-authored-by: Philip Turner <[email protected]>
- Loading branch information