Cusolver: Internal review version by haxushu · Pull Request #3 · haxushu/abacus-develop

haxushu · 2022-03-10T12:03:30Z

Internal review version

darelbeida · 2022-03-17T13:18:32Z

comment one

darelbeida

test

darelbeida · 2022-03-14T10:49:51Z

            GlobalC::wf.npwx));
 	}
-	if(m == 1)
+	if(m != 1)


可以单独提一个PR修这个错误

deepmodeling#736

darelbeida · 2022-03-14T10:50:27Z

 	}
+	else if(GlobalV::KS_SOLVER=="cusolver")
+	{
+        label = "CU";


上面的缩进两种风格都有😂

darelbeida · 2022-03-14T10:51:02Z

+#endif
+
+
+


空行数-2

darelbeida · 2022-03-14T10:54:48Z

+	int maxncol;
+
+	MPI_Allreduce(&ncol, &maxncol, 1, MPI_INT, MPI_MAX, comm_2D);
+	// MPI_Reduce(&ncol, &maxncol, 1, MPI_LONG, MPI_MAX, 0, comm_2D);


注释清理一下

darelbeida · 2022-03-14T10:55:58Z

 #include "../module_base/memory.h"
 #include "../module_base/timer.h"
-
+#include <chrono>


关于chrono这部分应该都是调试的，可以在适当时候清理掉

Have been removed.

darelbeida · 2022-03-14T14:34:47Z

+			{
+
+#ifdef __MPI
+				GlobalV::ofs_warning << "Cusolver is now only available for LCAO." << std::endl;


全局的cusolver可以统一一下大小写，感觉和后面的写法不太一致，找一个统一的写法

我不是说和ks_solver里面的字符串内容，我是说后面又看到了cuSolver写法，与这里的Cusolver不一致

后面都是全小写的，这里因为是一个句子，首字母大写了一下。

除了这个我后面都check统一一下

darelbeida · 2022-03-14T14:38:25Z

                            int iic;

-                            if(GlobalV::KS_SOLVER=="genelpa" || GlobalV::KS_SOLVER=="scalapack_gvx")  // save the matrix as column major format
+                            if(GlobalV::KS_SOLVER=="genelpa" || GlobalV::KS_SOLVER=="scalapack_gvx" || GlobalV::KS_SOLVER=="cusolver")  // save the matrix as column major format


这个判断后面用了非常多次，能不能新构造一个GlobalV的变量记录一下这个布尔值？

变量名可能和列优先相关

abacus对genelpa和scalapack_gvx都没有这样做，感觉判断的代价不大。

darelbeida · 2022-03-17T11:02:08Z

+int Diag_cuSolver_gvd::Dngvd_double(int N, int M, double *A, double *B, double *W, double *V){
+
+        copy_double(N, M, A, B);
+        buffer_double();


这步有办法只做第一次吗，后面只需要判断形状是否与之前一致，一致就不重复做，否则销毁d_work并重新做

补充一下，这个我评论错了，需要改在source/src_pdiag/diag_cusolver.cu里，而不是这个为了测试而存在的文件里，需要特别注意

cusolverDnDsygvd_bufferSize不清楚内部机制，不能确定是否一致。

darelbeida · 2022-03-17T11:04:01Z

+}
+int Diag_cuSolver_gvd::Dngvd_double(int N, int M, double *A, double *B, double *W, double *V){
+
+        copy_double(N, M, A, B);


我建议copy和recopy用上CHECK_CUDA宏之后，可以直接把代码写在这个函数里，加上一行注释即可，不需要单独的类方法。

compute也是，核心只有一个cusolverDnDsygvd调用和一个sync，足够简短可以写到这个函数里面来

darelbeida · 2022-03-17T11:06:52Z

+}
+
+
+void Diag_cuSolver_gvd::finalize(){


和析构函数合并起来吧，看起来并没有除了析构以外的逻辑会调用它

Have accepted.

darelbeida · 2022-03-18T11:06:55Z

+	double eig[dim];				// reference
+
+	// construct the matrix A and B w.r.t the problem Ax = \lambda Bx
+	for (int i = 0; i < dim*dim; i++) a[i] = {1, 0};


这里能不能构造出一个带虚部的例子？

我再加一组

Have added to test complex_2.

darelbeida · 2022-03-18T11:10:04Z

+
+# compile the test executable
+file(GLOB TEST_SRC 
+    ../../source/module_cusolver/test/cusolver_test.cpp


我在测试的cpp文件内以及这里都没有看到对CUDA_LCAO或者MPI宏的判断，我觉得至少应该加在某处？

考虑到单元测试的独立性，如果要分开测，可以在file(GLOB TEST_SRC单列cusolver_test.cpp或mpi.cpp

fix : UT for diago_cg.cpp and diago_david.cpp should not depend on __CUDA and __ROCM

… initial

1. change HONG in Makefile

* fix : when input structure doesn't have most symmetry rotation matrix, ABACUS will print which structure is better and then using the input structure for calculating. * perf : add warning output for symmetry searching part; modified tests/integral/207_NO_DP_OD/result.ref to correct version. * fix : UT for diago_cg.cpp and diago_david.cpp should not depend on __CUDA and __ROCM * fix : support both element magnetization and atomic magnetization for initial * add 304_NO_GO_AF_atommag case for testing initial atomic magnetization

fix : support both element magnetization and atomic magnetization for initial setting in STRU file

…eepmodeling#120) * fix : when input structure doesn't have most symmetry rotation matrix, ABACUS will print which structure is better and then using the input structure for calculating. * perf : add warning output for symmetry searching part; modified tests/integral/207_NO_DP_OD/result.ref to correct version. * fix : UT for diago_cg.cpp and diago_david.cpp should not depend on __CUDA and __ROCM * fix : support both element magnetization and atomic magnetization for initial * add 304_NO_GO_AF_atommag case for testing initial atomic magnetization Co-authored-by: Chun Cai <amoycaic@gmail.com>

synchronization between abacusmodeling and deepmodeling abacus-develop/develop branch

merge

range: source/src_pw/test/diago_david_test.cpp

… into develop

1. update Makefile.system and Makefile.vars

… new

… develop

fix uninited rcutmax

test(Parallel): add UT of `split_diag_world` and `split_grid_world`

… develop

added test case for DeePKS multi-k band gap label

gint : removed obsolete code; implement new gamma force & stress

1. Fix the bug that multiple line heading of blocks <PP_R>, <PP_RAB>, <PP_DIJ>, and <PP_RHOATOM> in "PROJECT-Eu.UPF-1.txt" cannot be recognized by Pseudopot_upf::read_pseudo_upf201. 2. Release the restriction that off-diagonal elements of DIJ (Pseudopot_upf::dion) must be zero.

Fix: not using `DM[is]` when no atom found on grid

into debug

fix: Add the support to the multiple line heading in pseudopotential.

into LCAO-refactor

test(parallel): UT of methods in parallel_kpoints

Merge LCAO-refactor into develop

haxushu assigned darelbeida Mar 17, 2022

haxushu requested a review from darelbeida March 17, 2022 13:19

darelbeida requested changes Mar 17, 2022

View reviewed changes

darelbeida reviewed Mar 18, 2022

View reviewed changes

haxushu force-pushed the cusolver branch 2 times, most recently from d83df2b to a19a372 Compare March 25, 2022 12:04

darelbeida force-pushed the cusolver branch from c218aa6 to e70c9d4 Compare April 1, 2022 10:28

caic99 and others added 5 commits April 4, 2022 15:35

Merge pull request deepmodeling#817 from dyzheng/new

99eb436

fix : UT for diago_cg.cpp and diago_david.cpp should not depend on __CUDA and __ROCM

fix : support both element magnetization and atomic magnetization for…

e02c9c0

… initial

Merge pull request deepmodeling#118 from PeizeLin/develop

e8fcf75

1. change HONG in Makefile

add 304_NO_GO_AF_atommag case for testing initial atomic magnetization

7791c4f

darelbeida force-pushed the cusolver branch from e70c9d4 to 9471065 Compare April 6, 2022 02:25

dyzheng and others added 10 commits April 6, 2022 16:25

Merge pull request deepmodeling#818 from dyzheng/new

0e74934

fix : support both element magnetization and atomic magnetization for initial setting in STRU file

Merge pull request deepmodeling#819 from abacusmodeling/develop

41386d0

synchronization between abacusmodeling and deepmodeling abacus-develop/develop branch

Merge pull request deepmodeling#121 from deepmodeling/develop

b4f8dd5

merge

fix bug: the opended file 'data-H' is not closed

1789577

range: source/src_pw/test/diago_david_test.cpp

1. update Makefile.system and Makefile.vars

96212f0

Merge branch 'develop' of https://gitee.com/deepmodeling/abacus-develop…

2887736

… into develop

Merge pull request deepmodeling#122 from PeizeLin/develop

13d0a5e

1. update Makefile.system and Makefile.vars

Fix : occupied band should be number of electrons instead of half

9fc9671

Merge branch 'develop' of github.com:deepmodeling/abacus-develop into…

41e19ec

… new

darelbeida force-pushed the cusolver branch from 9471065 to b5a51c1 Compare April 11, 2022 05:33

maki49 and others added 4 commits April 12, 2022 00:21

fix uninited rcutmax

7c2f518

deepks : allow reading embedding model

81e13e7

Merge branch 'develop' of github.com:deepmodeling/abacus-develop into…

e2fc3ed

… develop

Merge pull request deepmodeling#821 from maki49/develop

3e76325

fix uninited rcutmax

wenfei-li and others added 4 commits May 10, 2022 15:04

gint : update makefile

d0adbc2

Merge pull request deepmodeling#906 from hongriTianqi/develop

e34725f

test(Parallel): add UT of `split_diag_world` and `split_grid_world`

Merge branch 'develop' of github.com:deepmodeling/abacus-develop into…

53a6361

… develop

Merge pull request deepmodeling#907 from ouqi0711/develop

51200fe

added test case for DeePKS multi-k band gap label

darelbeida force-pushed the cusolver branch from 3252c0c to 7192d33 Compare May 10, 2022 11:01

Merge pull request deepmodeling#897 from wenfei-li/develop

823c5a2

gint : removed obsolete code; implement new gamma force & stress

darelbeida force-pushed the cusolver branch from 7192d33 to f57f9b4 Compare May 10, 2022 11:16

maki49 and others added 20 commits May 10, 2022 20:44

revert private

23b76a7

move LOOP_elec.cpp to module_esolver/esolver_ks_lcao_elec.cpp

c8a0558

add a bool have_force in ESolver_LCAO

44e89e6

reformat some files

dbb0ded

reformat all files in module_esolver

5e55e09

reformat some files

104b138

add comments for public variables

5f0113d

pull develop

e6513bf

fix: set have_force=false after every md step

04857ac

<fix>

e3e57df

1. Fix the bug that multiple line heading of blocks <PP_R>, <PP_RAB>, <PP_DIJ>, and <PP_RHOATOM> in "PROJECT-Eu.UPF-1.txt" cannot be recognized by Pseudopot_upf::read_pseudo_upf201. 2. Release the restriction that off-diagonal elements of DIJ (Pseudopot_upf::dion) must be zero.

gint : fix bug in gamma force

56c26b1

Merge pull request deepmodeling#911 from wenfei-li/develop

8f6f4c7

Fix: not using `DM[is]` when no atom found on grid

Merge branch 'develop' of https://github.com/deepmodeling/abacus-develop

16a5348

into debug

Merge pull request deepmodeling#910 from sunliang98/debug

8f70b45

fix: Add the support to the multiple line heading in pseudopotential.

Merge branch 'develop' of https://github.com/deepmodeling/abacus-develop

577a927

into LCAO-refactor

test(parallel): UT of methods in parallel_kpoints

b7acbbf

modify default number of processes to 4

3ecd932

Merge pull request deepmodeling#912 from hongriTianqi/develop

d9055d8

test(parallel): UT of methods in parallel_kpoints

add new deepks-bandgap-code into esolver

b625243

Merge pull request deepmodeling#909 from maki49/LCAO-refactor

2db26f6

Merge LCAO-refactor into develop

darelbeida force-pushed the cusolver branch from f57f9b4 to 7479b3b Compare May 12, 2022 23:08

implemented a new ks_solver called "cusolver" in LCAO

de1305f

darelbeida force-pushed the cusolver branch from 7479b3b to de1305f Compare May 13, 2022 11:38

		#endif

		}


		void Diag_cuSolver_gvd::finalize(){

Conversation

haxushu commented Mar 10, 2022

Uh oh!

darelbeida commented Mar 17, 2022

Uh oh!

darelbeida left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

16 participants