Libcint #113

AlexanderMath · 2023-10-01T09:28:26Z

Got all integrals (except grad of kinetic) passing test case. I'd like to get this on main branch so we can start

discuss how to interface nanoDFT with different integral implementations (THO, libcint, popcint, ...)
discuss how to refactor nanoDFT for molecular dynamics (some subtleties wrt state and different combinations of arguments which may make sense)

Done compiling. Calling C code from python. 
[N=2]

[Nuclear Integral]
CPU:     2.763163926555734e-07
Compiling module jit_ipu_intor1e.0:
[##################################################] 100% Compilation Finished [Elapsed: 00:00:14.1]
IPU:     2.763163926555734e-07

[Kinetic Integral]
CPU:     -1.8022852765753328e-08
Compiling module jit_ipu_intor1e.1:
[##################################################] 100% Compilation Finished [Elapsed: 00:00:13.0]
IPU:     -1.05722721688295e-08

[Overlap Integral]
CPU:     -1.2445099606406274e-07
Compiling module jit_ipu_intor1e.2:
[##################################################] 100% Compilation Finished [Elapsed: 00:00:13.0]
IPU:     -6.484635128867211e-08

[Grad Nuclear]
CPU:     7.246001532124069e-08
Compiling module jit_ipu_intor1e.3:
[##################################################] 100% Compilation Finished [Elapsed: 00:00:13.1]
IPU:     7.246001532124069e-08

[Grad Kinetic]
CPU:     2.983345692708639e-08
Compiling module jit_ipu_intor1e.4:
[##################################################] 100% Compilation Finished [Elapsed: 00:00:13.3]
IPU:     2.983345692708639e-08

[Grad Overlap]
CPU:     6.077975783780332e-08
Compiling module jit_ipu_intor1e.5:
[##################################################] 100% Compilation Finished [Elapsed: 00:00:12.9]
IPU:     6.077975783780332e-08

[Electron Repulsion Integral]
CPU:     -4.443460513425812e-08
Compiling module jit_ipu_getints4c.6:
[##################################################] 100% Compilation Finished [Elapsed: 00:00:12.9]
IPU:     -2.953344391265489e-08

[Grad of Electron Repulsion Integral]
CPU:     1.341920186359591e-07
Compiling module jit_ipu_getints4c.7:
[##################################################] 100% Compilation Finished [Elapsed: 00:00:12.8]
IPU:     1.1929085744211143e-07

…in STO3G. Errors are ~ 1e-7.

…ch of suff: 1) float/double implciit slowdown (2) parallelize over tiles (3) controll number of rys_roots/c2s code we store on tiles

…checking if introduced any bugs, comparing different basis sets, molecules, etc.

pyscf_ipu/electron_repulsion/direct.sh

pyscf_ipu/electron_repulsion/popcint/libcint.cpp

awf · 2023-10-02T18:05:37Z

pyscf_ipu/electron_repulsion/popcint/libcint.cpp

+  Input<Vector<float>> mat;
+  Input<Vector<int>> shls_slice;
+  Input<Vector<int>> ao_loc;
+  Input<Vector<int>> atm;
+  Input<Vector<int>> bas;
+  Input<Vector<float>> env;
+  Input<Vector<int>> natm;
+  Input<Vector<int>> nbas;
+  Input<Vector<int>> which_integral;
+  Input<Vector<int>> comp;


Similarly sizes here

awf · 2023-10-02T18:10:32Z

pyscf_ipu/electron_repulsion/popcint/libcint.py

+    cintopt = None 
+
+    # type 
+    float32 = "#define dtype float" in open("libcint.c", "r").read()


I assume this is not going to be called a lot? It will be expensive, and after ~1000 calls will fail due to running out of file handles.

Suggest a @memoized function called e.g. libcint_is_float32(), which closes the file.

Or just compute it up at line 19 where you are already interacting with the filesyste,

I assume this is not going to be called a lot?

Once per integral type during tracing (i.e. at most 8 times). I'd think of this as increasing compile time from 2min to 2min 100ms. We can refactor to do once and make it 2min and 10ms.

awf · 2023-10-02T18:12:08Z

pyscf_ipu/electron_repulsion/popcint/libcint.py

+    from tessellate_ipu import create_ipu_tile_primitive, ipu_cycle_count, tile_map, tile_put_sharded, tile_put_replicated
+    vertex_filename  = osp.join(osp.dirname(__file__), "libcint.cpp")
+    #mat, shls_slice, ao_loc, atm, bas, env
+    grad = create_ipu_tile_primitive(


Again, probably want to do this only once?

awf · 2023-10-02T18:13:10Z

pyscf_ipu/electron_repulsion/popcint/libcint.py

+
+    cintopt = None 
+    prescreen = lib.c_null_ptr()
+    ic(intor_name, 'GTOnr2e_fill_'+aosym, "GTOnr2e_fill_drv")


ic = icrecream? Probably remove?

awf · 2023-10-02T18:14:36Z

pyscf_ipu/electron_repulsion/popcint/libcint.py

+    # type 
+    float32 = "#define dtype float" in open("libcint.c", "r").read()
+    if float32: 
+        out = out.astype(np.float32)
+        env = env.astype(np.float32)


There's a lot of duplicated code here - and a lot of it is slow - so it feels as if there might be an opportunity to dedup.

There's a lot of duplicated code here - and a lot of it is slow - so it feels as if there might be an opportunity to dedup.

Sure, let's make those refactors. The heavy compute part is the 20k lines of C. Most of the above is the easy 1% optimizations I was hoping to off-load.

Co-authored-by: Andrew Fitzgibbon <[email protected]>

…o libcint

…aining long term solution.

…ons).

alexandermath and others added 24 commits September 20, 2023 15:27

Initial refactor on grad.py

fb4e143

simplified python grad code calling libcint.

cb8306e

Adding draft for forces.

323ad04

work

610d64f

adding all integrals

93cc525

Work on libcint.c

549e9e8

work on integrals.

a5fc8f5

compile script for cpp compile; both c++ and c compiles.

9fc56f0

Adding support for float32 through #define dtype float/double.

8b8b931

All one electron integrals passing for C++, IPU_model and IPU for H2 …

91f5ee2

…in STO3G. Errors are ~ 1e-7.

Got ERI to pass. Next is eri_grad. After this, will have to fix a bun…

4538635

…ch of suff: 1) float/double implciit slowdown (2) parallelize over tiles (3) controll number of rys_roots/c2s code we store on tiles

Got derivative of ERI.

77175ed

Cleaning up folder and test cases.

bc71a2d

Cleaned a few things.

575dedf

Prettified test cases and put output in readme.

0571e79

Changing name to popcint.

53c07ca

changing folder name

9cf216f

Update readme.MD

6c18443

Update readme.MD

b92cbf2

Adding python suppor/test to help with tilemapping.

8ce74ac

adding single ipu profiling to direct.sh script.

20f62ea

Fixed bug in script to run test cases.

46a26f7

Adding quick test case with all cpu using -skipipu. Neat for quickly …

187e308

…checking if introduced any bugs, comparing different basis sets, molecules, etc.

Merge branch 'main' into libcint

c94b49d

AlexanderMath requested review from awf, balancap, akrzgc, blazejba and hatemhelal October 1, 2023 09:39

awf requested changes Oct 2, 2023

View reviewed changes

Update pyscf_ipu/electron_repulsion/direct.sh

a833f50

Co-authored-by: Andrew Fitzgibbon <[email protected]>

awf approved these changes Oct 4, 2023

View reviewed changes

alexandermath added 4 commits October 12, 2023 09:34

Merge branch 'libcint' of github.com:graphcore-research/pyscf-ipu int…

6adaf39

…o libcint

Removing fixed sizes by using InOut instead of Input. Added TODO expl…

6909bb6

…aining long term solution.

Fixed kinetic grad and added a todo.

2fb81a0

Changed so re-use same graph for nuc/kin/ovlp (similar for grad versi…

083f762

…ons).

AlexanderMath merged commit c21b92b into main Oct 13, 2023
4 checks passed

AlexanderMath deleted the libcint branch October 13, 2023 09:57

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Libcint #113

Libcint #113

AlexanderMath commented Oct 1, 2023 •

edited

Loading

awf Oct 2, 2023

awf Oct 2, 2023

awf Oct 2, 2023

AlexanderMath Oct 4, 2023 •

edited

Loading

awf Oct 2, 2023

awf Oct 2, 2023

AlexanderMath Oct 12, 2023

awf Oct 2, 2023

AlexanderMath Oct 4, 2023

Libcint #113

Libcint #113

Conversation

AlexanderMath commented Oct 1, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AlexanderMath Oct 4, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AlexanderMath commented Oct 1, 2023 •

edited

Loading

AlexanderMath Oct 4, 2023 •

edited

Loading