Tags

follow test3, 1. 整合 QCU 的 CUDA/MPI/NCCL 源码与头文件，调整格点参数和构建配置; 2. 集成 PyQcu/PyQuda 接口、PyQuda 版本代码及依赖元数据; 3. 补充单卡/多卡、Wilson/Clover/求解器测试脚本、示例和运行数据; [codex].

follow test2, qcu has been merged into pyqcu/extern, this library will be archived and the future will be better!

follow test1, perpare for pyqcu :)

follow test1, first finish arnoldi function, almostly pass the random test, but still a serious bug appears (lattice_multiplier operator* with Eigen vector)......

follow test0, add ./test/README.md, wait for multi-gpu test support......

follow dev26, rm all files that go with pyquda, just remain a test inface for qcu (./test/test-np1.sh), wish success. orz

follow dev25, finish single-gpu float version of qcu that is able to run without pyquda (refer to ./test/*-np1.py & test.cu), wait for multi-gpu test again......

follow dev24, 1. updata ./doc; 2. find bug (bug10) caused by io between pyquda and qcu, that has been cleared (refer to ./test/test.dslash-np1.py:float test); 3.float version qcu pass single-gpu test in computer-4060, wait for multi-gpu test......

follow dev23, save data from quda (pyquda) in bin and load data to qcu in bin. prepare for the test of huawei, data files is in computer-4060-docker:/root/qcu/test. :) orz

follow bug10, 1. gmres-ir function done, then will be truely tested after multi-precision done; 2. optimize cg and bistabcg after nccl been removed (dev22), bistabcg has been faster about 20% !!!

follow dev22, change qcu and quda precision to single (float), but qcu output nan......

follow dev21, template define all done: 1. remove all nccl support for efficiency; 2. rename some source files; 3. redefine some cublas and mpi functions. <^_^>

follow bug9, template define part done, wait mpi,nccl,cublas template done :)

follow dev20, unkown bug appear after template add, debug or not......

follow stab16, 1. finish dot (MPI version); 2. rename part of test files; 3. add switch function bewteen mpi and nccl, and be able to compile without nccl library. all the test run in x99 with 2 NV-P100.

follow dev19, finish all the test in snsc with 8 NV-A100, just after done dot (MPI version), we will go ahead next step. :)

follow dev18, find a foolish bug in clover dslash caused by writing wrong (replace move_wards[_FX_FY_] by move_wards[_FX_FT_]), after fix jod done in lattice_set.h, finally run right in snsc with 8 NV-A100.

follow bug8, clover dslash dead-lock in snsc with 8 NV-A100......

follow dev17, 1. find a small bug in clover dslash when lat size cute into <=2; 2. change grid cut in tzyx way instead of xyzt way (follow to the cut way of pyquda); so, finally run right in snsc with 8 NV-A100!!!!!!

follow bug7, ready debug in x99 with 2 NV-P100.