Tags

follow bug29_1, can not find the bugs!!!!!!!!!!!!!!!!, really set ccdptzyx, psctzyx as rule. I am really sosososososososo tired. *************************

follow dev68, the p...xyzt version of qcu (besides clover.....) test pass, but when give " if (1) {
    if_b_x = 0;
    if_b_y = 0;
    if_b_z = 0;
    if_b_t = 0;
    if_f_x = 0;
    if_f_y = 0;
    if_f_z = 0;
    if_f_t = 0;
    // if_b_x_b_y = 0;
    // if_b_x_b_z = 0;
    // if_b_x_b_t = 0;
    // if_b_y_b_z = 0;
    // if_b_y_b_t = 0;
    // if_b_z_b_t = 0;
    // if_f_x_b_y = 0;
    // if_f_x_b_z = 0;
    // if_f_x_b_t = 0;
    // if_f_y_b_z = 0;
    // if_f_y_b_t = 0;
    // if_f_z_b_t = 0;
    // if_b_x_f_y = 0;
    // if_b_x_f_z = 0;
    // if_b_x_f_t = 0;
    // if_b_y_f_z = 0;
    // if_b_y_f_t = 0;
    // if_b_z_f_t = 0;
  }" to PyQCU/cpp/cuda/qcu/src/clover_dslash_multi.cu, test pass, focus on this!!!

follow bug29, a new bug found in bistabcg, the target of dev67 cost too much, now get back to dev67_1, set ccdptzyx, psctzyx as rule. I am so tired. ***

follow dev67_1, the p...xyzt version of qcu (besides clover.....) test pass, but when give " if (1) {
    if_b_x = 0;
    if_b_y = 0;
    if_b_z = 0;
    if_b_t = 0;
    if_f_x = 0;
    if_f_y = 0;
    if_f_z = 0;
    if_f_t = 0;
    if_b_x_b_y = 0;
    // if_b_x_b_z = 0;
    // if_b_x_b_t = 0;
    // if_b_y_b_z = 0;
    // if_b_y_b_t = 0;
    // if_b_z_b_t = 0;
    if_f_x_b_y = 0;
    // if_f_x_b_z = 0;
    // if_f_x_b_t = 0;
    // if_f_y_b_z = 0;
    // if_f_y_b_t = 0;
    // if_f_z_b_t = 0;
    if_b_x_f_y = 0;
    // if_b_x_f_z = 0;
    // if_b_x_f_t = 0;
    // if_b_y_f_z = 0;
    // if_b_y_f_t = 0;
    // if_b_z_f_t = 0;
  }" to PyQCU/cpp/cuda/qcu/src/clover_dslash_multi.cu, test pass, focus on this!!!

follow dev67, referring to PyQCU/examples/qcu/conftest-v20260413.ipynb, the single-threaded version of qcu (ccdptzyx, psctzyx) test pass.

follow dev66, changing tzyx to xyzt in PyQCU/src/cuda/qcu...... this is a copy of dev66.

follow test10, CUDA version of QCU restarts.

follow dev65, the newly added PyQCU functions, except for TileLang (due to software version incompatibility in the environment), have been debugged and tested successfully on the NPU (including cross-node)......

follow dev64, referring to PyQCU/pyqcu/dslash/_clover.py, the multi-threaded version of make-clover done and pass the test. :)

follow stab21, 1. referring to PyQCU/pyqcu/smear/_stout.py, the multi-threaded version of stout-smear done and pass the test; 2. NEVER NEVER USE THE SHIT LIKE THAT "[[torch.zeros([]), torch.zeros([]), torch.zeros([]), torch.zeros([])]]*4"; 3. NEVER NEVER USE THE SHIT LIKE THAT "[[torch.zeros([]), torch.zeros([]), torch.zeros([]), torch.zeros([])]]*4"; 4. NEVER NEVER USE THE SHIT LIKE THAT "[[torch.zeros([]), torch.zeros([]), torch.zeros([]), torch.zeros([])]]*4". ^_^

follow dev63, yht is coming...

follow dev62, 1. referring to PyQCU/examples/profiler/conftest.py, using ui.perfetto.dev and torch.profiler, some detailed performance optimizations were performed on the operator; 2. referring to PyQCU/pyqcu/tools/_einsum.py, the TileLang version of Eexyzt_exyzt2Exyzt was completed, but its performance was mediocre........

follow dev61, referring to PyQCU/examples/init/log-v20260225.txt, the multi-threaded version of parity multigrid solver (pyqcu::testing::solver::parity) pass, next, we will focus on operator optimization based on TileLang. ^_^

follow bug28, 1. the bug with slover_parity is cleared, the matvec_eo (or matvec_oe) is equal to -kappa*D_eo (or D_oe), so the right matvec_parity is little different to 1-166 (PhD thesis of Jiang Xiangyu); 2. referring to PyQCU/examples/init/log-v20260222.txt, the multi-threaded version of parity solver (pyqcu::testing::solver::parity) pass. :)

follow stab20, 1. pyqcu::dslash::operator::matvec_all multi-threaded test pass; 2. but pyqcu::testing::solver::parity runs wrong......

follow bug27, 1. "U_dag_minus = _torch.roll(U_dag_mu, shifts=1, dims=ward)", the "ward" is changed to 0123 from -1-2-3-4 in bug27, now I add "ward = lattice.wards[ward_key]" below "ward_key = lattice.ward_keys[ward]"; 2. pyqcu::dslash::operator::matvec_* test pass. orz

follow stab19, a serious bug appear to pyqcu:dslash:wilson, all test run wrong!!!

follow dev60_1, 1. removed the no-longer-maintained Cupy interface and its corresponding CUDA C++ source code, the Pytorch and Tilelang projects have been merged, future focus will be solely on supplementing and optimizing this part; 2. referring to PyQCU/examples/init/log-v20260212.txt, changed the last four dimensions of the array from tzyx to xyzt, and made numerous corresponding code modifications, some infrequently used code was also removed or merged, single-threaded and multi-threaded tests were then passed; 3. referring to PyQCU/examples/data or P:\Aliyun Disk\Workings\LQCD\data\data-v20260206.rar, based on the previous QUDA data, a new test set prepared for the new version of PyQCU is provided; 4. large-scale performance optimizations for Tilelang will follow; hopefully, everything will go smoothly. :)

follow dev60, referring to PyQCU/test/dev60_1, the tilelang seems to be very powerfull, so I prepare to so focus on tilelang that I will remove other support......

follow bug26, prepare for new version of pyqcu for tilelang, this is a copy of v20260203.