Tags

follow dev40, finish bistabcg on all coarse grid in ./test/test-dev41.ipynb, however, the cg version does not converge, and a preliminary judgment is that r_matvec_p destroys the positive definiteness of the original matvec, moreover, the tolerance degree of the result error on the coarse grid needs further discussion...... orz

follow dev39, finish r_vec and p_vec in ./test/test-dev40.ipynb(refer to https://github.com/Gustavroot/MGMLMC_for_mat_with_displ/blob/main/src/build_2lvl_method_g5_comp.m). :)

follow bug17, after hard work, it was determined that the bug was caused by missing cuda synchronization of the pyqcu interface in qcu, cudaDeviceSynchronize() is added to each ./extern/qcu/src/apply.*.cu, result is shown in ./test/test-dev39.ipynb(a copy file of ./test/test-bug17.ipynb), nice one!

follow dev38, 1. optimize host_params init in ./extern/qcu/include/lattice_set.h; 2. optimize pyqcu init in ./test/test-bug17.ipynb; 3. find a very very strange bug in ./test/test-bug17.ipynb (in this file, cg_dslash should be same to matvec, but it does not happen, and more strange, this cg in python version runs right in sometime???). o.0

follow dev35, give a test format for the following naive single multigrid test in python (refer to ./test/test-dev38.ipynb), hope to be successful! o.0

follow bug16, update ./extern/qcu/env.sh (arch_sm has been automatic identified) & ./extern/qcu/include/lattice_set.h (there move_wards & host_params[_LAT_*_] init with a stupid mistake!!!), bug has been cleared (refer to ./test/log.clover-dev37.txt). -_- :)

follow bug16, update ./extern/qcu/env.sh (arch_sm has been automatic identified) & ./extern/qcu/include/lattice_set.h (there move_wards & host_params[_LAT_*_] init with a stupid mistake!!!), bug has been cleared (refer to ./test/log.clover-dev37.txt). -_- :)

follow dev36, when running ./test/test.clover-test5.py, bug appear (refer to ./test/log.clover-bug16.txt), same bug also appear when when running ./test/test.wilson-test5.py......

follow dev36, when running ./test/test.clover-test5.py, bug appear (refer to ./test/log.clover-bug16.txt), same bug also appear when when running ./test/test.wilson-test5.py......

follow bug15, h5py is successfully installed in x99 (refer to ./test/test-dev36.sh)! :)

follow bug15, h5py is successfully installed in x99 (refer to ./test/test-dev36.sh)! :)

follow dev35, prepare for test5, but bug appear when h5py compiling in x99 (refer to ./test/log-bug15.txt)......

follow dev35, prepare for test5, but bug appear when h5py compiling in x99 (refer to ./test/log-bug15.txt)......

follow bug14, debug done with bug14 (the init of rank_index is not right, define._X_ is 0 not 3......), finish hdf5 swap with grid array (refer to ./pyqcu/io.py:grid_xxxtzyx2hdf5_xxxtzyx);hdf5_xxxtzyx2grid_xxxtzyx). :)

follow dev34, finish xxxtzyx2grid_xxxtzyx & grid_xxxtzyx2xxxtzyx function (refer to ./test/test-bug14.py), but there is a strange bug ("Dest Shape: (3, 3, 4, 2, 0, 32, 32, 4)", refer to local:./test/.log.txt)......

follow dev33, the cupy implementation of odd-even partitioning and restoration has been completed (refer to ./pyqcu/io.py:xxxtzyx2pxxxtzyx;pxxxtzyx2xxxtzyx). o.0

follow dev32, the shifted_power_iteration of chebyshev_filter is used to solve the smaller eigenvalue of the cg version of dslash (refer./test/test.small.eigen-dev33.ipynb), which takes a long time and has a large error in general. but it almost meets the requirements as a test vector. ^_^

follow bug13, fix the bug that bug13 motioned, ./extern/qcu/include/define.h:_give_u_laplacian can be chosed to use, preformance test refer to ./test/log-dev32.txt, wait for multi-gpu test! :)

follow dev31, dev31 bug for boundary send-recv function construction error, for laplacian function, forward and backward transmission construction is not symmetrical, U and src mapping there is an offset, is being urgently fixed!

follow dev30, found in the dev30 Laplacian function results are not consistent with the result cupy, follow-up found that randomly generated when U dont meet./extern/qcu/include/define.h:give_u_laplacian optimization requirements, give up the optimized results with cupy close, The cause of minor difference is not found. The laplacian function of multiple Gpus is temporarily declared complete, which has obvious advantages over cupy performance (refer to./test/log-dev31.txt), and wait for subsequent testing and debugging.....