patch_ib Memory Copy Reduction

[danieljvickers]

Description

The patch_ib array was being updated on the CPU. This had the downstream effect of more host-device memory copies that strictly required. This PR modified the code to use the GPU copy of patch_ib everywhere and only copy out during file IO. This should remove memory copies and enable parallelism that was not present before.

Fixes #1453

Type of change

• Refactor

[github-actions]

Claude Code Review

Head SHA: 0072f07

Files changed:

• 4
• src/simulation/m_data_output.fpp
• src/simulation/m_ib_patches.fpp
• src/simulation/m_ibm.fpp
• src/simulation/m_time_steppers.fpp

Findings

1. Uninitialized length in rectangle bounding-box computation (m_ib_patches.fpp)

The diff removes the assignments that initialised length before the corner_distance calculation in the 2D rectangle subroutine:

-        length(1) = patch_ib(patch_id)%length_x
-        length(2) = patch_ib(patch_id)%length_y
-        inverse_rotation(:,:) = patch_ib(patch_id)%rotation_matrix_inverse(:,:)

But the declaration real(wp), dimension(1:2) :: length, center is retained (context line 452) and length is still consumed immediately after (context line 468):

corner_distance = sqrt(dot_product(length, length))/2._wp

length is now uninitialized at that point. corner_distance will be garbage, so get_bounding_indices will produce wrong il/ir/jl/jr bounds. The GPU parallel loop then iterates over those wrong bounds, potentially missing IB cells or walking off the array.

Fix: Replace with corner_distance = 0.5_wp*sqrt(patch_ib(patch_id)%length_x**2 + patch_ib(patch_id)%length_y**2) (consistent with the cuboid fix in the same PR), and drop the now-unused length and inverse_rotation declarations.

---

2. Uninitialized length in cylinder bounding-box computation (m_ib_patches.fpp)

Same pattern in the 3D cylinder subroutine. The diff removes:

-        length(1) = patch_ib(patch_id)%length_x
-        length(2) = patch_ib(patch_id)%length_y
-        length(3) = patch_ib(patch_id)%length_z
-        radius = patch_ib(patch_id)%radius

The declaration real(wp), dimension(1:3) :: xyz_local, center, length remains (context line 616), and length is still read (context line 635):

corner_distance = sqrt(patch_ib(patch_id)%radius**2 + maxval(length)**2)

maxval(length) is undefined, so the cylinder's bounding box is wrong. Only radius was migrated to the direct patch_ib access; length was not.

Fix: Replace maxval(length) with max(patch_ib(patch_id)%length_x, patch_ib(patch_id)%length_y, patch_ib(patch_id)%length_z) (or the axis-appropriate subset), and remove the length declaration.

---

3. Missing GPU_UPDATE(host=...) before s_update_mib calls (m_time_steppers.fpp, m_ibm.fpp)

The PR converts two host-side do i = 1, num_ibs loops into GPU_PARALLEL_LOOPs, so patch_ib centroids, velocities, angles, and rotation matrices are now updated exclusively on the device. Both sites then immediately call s_update_mib (which performs MPI communication and broadcasts IB positions) without first pulling the device-modified data back to the host.

m_time_steppers.fpp (context lines 763–765):

$:END_GPU_PARALLEL_LOOP()   ! patch_ib updated on device only
call s_update_mib(num_ibs)  ! reads stale host copy -> broadcasts wrong positions

The old flow was host→GPU_UPDATE(device)→s_update_mib; the new GPU loop writes to device but the host copy is never refreshed.

m_ibm.fpp (rotation-matrix recalculation, diff hunk ~line 871):

+$:GPU_PARALLEL_LOOP(private='[i]')
 do i = 1, num_ibs
     if (patch_ib(i)%moving_ibm /= 0) then
         call s_update_ib_rotation_matrix(i)
     end if
 end do
-$:GPU_UPDATE(device='[patch_ib]')
+$:END_GPU_PARALLEL_LOOP()
 ! recompute the new ib_patch locations and broadcast them.

The old GPU_UPDATE(device=...) pushed host→device; removing it without adding GPU_UPDATE(host=...) means s_update_mib (called just below via the APPLY-IB-PATCHES block) operates on the pre-loop host state.

Fix: Insert $:GPU_UPDATE(host='[patch_ib(1:num_ibs)]') immediately after each END_GPU_PARALLEL_LOOP and before s_update_mib.

[codecov]

Codecov Report

❌ Patch coverage is 78.94737% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 60.78%. Comparing base (68dbb6b) to head (a6ab163).

Files with missing lines	Patch %	Lines
src/simulation/m_ib_patches.fpp	78.94%	3 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1531      +/-   ##
==========================================
- Coverage   60.80%   60.78%   -0.02%     
==========================================
  Files          73       73              
  Lines       20199    20182      -17     
  Branches     2932     2932              
==========================================
- Hits        12281    12267      -14     
+ Misses       5932     5929       -3     
  Partials     1986     1986              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[danieljvickers]

There was one $:GPU_UPDATE(host=...) that I simply could not remove unless I wanted to break another more-important optimization. I have a solution for this, but it requires the changes that will be made from the projection optimization first. Choosing to merge this in its current state, as I removed several copies that should improve performance. Will address the final memory copy in the m_ib_patch.fpp refactor coming from issue #1454

[danieljvickers]

All of the AI comments were correct this time and have not been merged.