Tags

follow bug23, after debug "applyCloverBistabCgQcu" and "void applyGaussGaugeQcu" and "PyQCU/test/test.ascend.mg-dev53.ipynb" seem to be right...... qcu yes?

follow bug22, 1. "checkCudaErrors(cudaSetDevice(host_params[_NODE_RANK_]));" lost !!!!!!, "cp.cuda.Device(rank).use()" lost !!!!!!; 2. add "void applyCloverBistabCgQcu" and "void applyGaussGaugeQcu" to PyQCU (all bug!!!), delete "void applyWilsonGmresIrQc" from PyQCU. www~www~

follow bug21, after clean a few bugs, PyQCU/pyqcu/ascend/inverse.py:mg still runs wrong. www~

follow dev52, after clean many bugs, PyQCU/pyqcu/ascend/inverse.py:mg still runs wrong. www~

follow bug20, 1. combine PyQCU/pyqcu/ascend/dslash.py and PyQCU/pyqcu/ascend/dslash_parity.py; 2. referring to PyQCU/test/test-dev52-ASCEND-RDP.*, PyQCU/pyqcu/ascend/inverse.py:restrict & prolong done. :)

follow dev51, remember that ###Schmitt orthogonalization of complex space###:"null_vecs[i] -= torch.vdot(null_vecs[j].flatten(), null_vecs[i].flatten())/torch.vdot(null_vecs[j].flatten(), null_vecs[j].flatten())*null_vecs[j]", not "null_vecs[i] -= torch.vdot(null_vecs[i].flatten(), null_vecs[j].flatten())/torch.vdot(null_vecs[j].flatten(), null_vecs[j].flatten())*null_vecs[j]"!!!, debug PyQCU/pyqcu/ascend/inverse.py:give_null_vecs done and make PyQCU/pyqcu/ascend/inverse.py:local_orthogonalize done, origin bug appear in PyQCU/test/mg_example as well as bug19......

follow bug19, 1. rename some var in PyQCU/pyqcu/ascend/dslash_parity.py:eye -> I; 2. referring to PyQCU/test/test-dev51-ASCEND-*, PyQCU/pyqcu/ascend/inverse.py:cg & bicgstab & give_null_vecs done! ^_^

follow stab18, remember that ###vdot(,)=dot(conj(),)###!!!,do not use vdot as same as dot!!!, debug PyQCU/pyqcu/ascend/inverse.py and some PyQCU/test/test-dev50-ASCEND-WILSON-* done. www~

follow dev50, referring to PyQCU/extern/cuda/qcu/include/lattice_* and PyQCU/extern/cuda/qcu/src/apply_*, add "_VERBOSE_" to the params of qcu(cuda), some test files of test5 have been added and optimized. ^_^

follow bug18, referring to PyQCU/test/test-dev50-ASCEND-* and PyQCU/pyqcu/ascend/dslash_parity.py, the bug mentioned earlier occurred due to the lack of details in parity handling: 1. The eye item should not appear in wilson_eo and wilson_oe, it should be treated as a separate diagonal item; 2. Should be classified analysis of the impact of odd-even processing in x_p dimension, for example: "parity (SRC) = (t + x + y + z) % 2 = 1, eo (SRC) = = (t + y + z) % 2 = 0, so: x_p (SRC) % 2 = 1, move_plus = 0". The implementation and debugging of dslash_parity have been completed, we will continue to make strong progress in the future!!!

follow dev49, there is a bug in PyQCU/pyqcu/ascend/dslash_parity.py(refer to PyQCU/test/test-dev50-ASCEND-WILSON-PARITY.py and /PyQCU/test/test-bug18-ASCEND-WILSON-PARITY.log)......

follow stab17, referring to PyQCU/test/test-dev49-ASCEND-* and PyQCU/pyqcu/ascend/dslash.py, I have completed the support for Wilson and Clover dslash on Ascend (PyTorch version), and the progress is quite good. :)

follow dev48, distance qcu after seven and a half months of stab16, stab17, which is actually not very stable, has been released, the directory structure of PyQCU/pyqcu has been updated, PyQCU/doc has been added, and PyQCU/pyqcu/__init__.py has been modified, ascend support, which holds the same status as PyQCU/pyqcu/cuda, has been added, more substantial updates will be available in the future, stay tuned. ^_^

follow dev46, after a long absence, I'm starting a side quest: developing LQCD software for Ascend(the NPU maked by Huawei). I've completed the single-threaded pytorch version of wilson dslash (refer to./test/ test-dev48-torch.py). I hope everything goes smoothly in the future. :)

follow dev47, same as dev47, accidentally skipped dev46......

follow dev45, rethink about the coeffi of clover(A=1+T, refer to joplin:/LQCD/some points), change some src and head files, test jods refer to ./test/test.clover-dev47.py. ^_^

follow dev44, finish save&load-able clover dslash(refer to test/test-dev45-DEMO.ipynb), add some new apply_*.cu and change some lattice_*.h, nice one~~~

follow dev43, 1. refer to the framework of GMG (geometric multimesh) and the construction methods of P matrix and R matrix of the simplest AMG (algebraic multimesh), and give an example (refer to ./test/ test-dev44-gmg.ipynb); 2. after reading the discussion of the eigen spectra of multiple operators (matvec) in <Multigrid for Chiral Lattice Fermions-Domain Wall>, in order to repeat the relevant work, a multi-lattice-sizes (the minimum was temporarily set as 4,4,4,4) & multi-precision BISTABCG solver was realized as a paving work (see./test/ test-dev44-demo.ipynb). :)

follow dev42, finish amg-2-level in test-dev43-*.ipynb, but the algorithm is wrong. I am reading the literature......

follow dev41, update eigen & cg & bistacg in pyqcu, finish some test work in test-dev42-*.ipynb, not much job done in fact. doge...