Reduce local memory usage in int3c2e CUDA kernels

[wxj6000]

The issue (see #75)

Current int3c2e kernels allocate a large amount of local memory, which may lead to out of memory when cupy memory pool occupies 90% of global memory. This PR tries to resolve this issue.

TODOs

• Reduce local memory in int3c2e gradient and hessian
  • Reduce 2x gsize -> 1x gsize in gradient
  • Reduce 4x gsize -> 1x gsize in Hessian
• Map each task to one block for high angular momentum

Additionally

In DF/grad/JK kernels, the intermediates (3,nao) are eliminated. The derivatives w.r.t atomic positions are directly calculated in CUDA kernels.

[Copilot]

Copilot reviewed 41 out of 45 changed files in this pull request and generated 1 comment.

Files not reviewed (4)

• gpu4pyscf/lib/gint/CMakeLists.txt: Language not supported
• gpu4pyscf/lib/gint/bpcache.cu: Language not supported
• gpu4pyscf/lib/gint/g2e.cu: Language not supported
• gpu4pyscf/lib/gint/g2e.h: Language not supported

Comments suppressed due to low confidence (1)

gpu4pyscf/df/grad/uhf.py:174

• [nitpick] The new variable names 'ej' and 'ek' are not self-explanatory. Consider using more descriptive names (e.g., 'grad_j' and 'grad_k') to clearly indicate their purpose.

ej = -ej

[Copilot]

The use of 'exit()' here forces an early termination of the example, which may be unintended. Consider removing it to allow the script to complete its execution.

Suggested change

exit()

[Copilot]

Copilot reviewed 39 out of 44 changed files in this pull request and generated no comments.

Files not reviewed (5)

• gpu4pyscf/lib/gint/CMakeLists.txt: Language not supported
• gpu4pyscf/lib/gint/bpcache.cu: Language not supported
• gpu4pyscf/lib/gint/g2e.cu: Language not supported
• gpu4pyscf/lib/gint/g2e.h: Language not supported
• gpu4pyscf/lib/gint/g3c2e.cu: Language not supported

Comments suppressed due to low confidence (2)

gpu4pyscf/df/int3c2e.py:446

• [nitpick] The variable naming has been updated from 'vj' to 'ej' (and similarly for 'vk' to 'ek'). Please update the associated inline comments to reflect this change for clarity.

            ej = cupy.zeros([natm,3], order='C')

gpu4pyscf/df/grad/uhf.py:174

• [nitpick] The inline comment about variables remaining in Cartesian coordinates appears to reference the old 'vj/vk' names. Consider updating or removing the comment to avoid confusion given the new 'ej/ek' naming.

        ej = -ej

[sunqm]

Please check the size of the output binary file. If the int3c2e code leads to a very big sized wheel, adjust the unrolling level.