rtx-cfd/rtx-fsi PERF-2 P3-iii: the MarchPlane — lane-table V-cycle kernels (ml_*), the batched apply over K lanes' own operators, the plane rendezvous inside the CG's preconditioner (lane thread-local, CG guard), set_plane_lane/plane_counters; the fsi2_overset_plane driver (K lane threads, one process); G-P0 unit tests: K distinct operators batched = solo apply bit for bit, uneven lanes served exactly

Co-Authored-By: Claude Fable 5.1 <[email protected]>

crates/specialized/rtx-cfd/tests/gpu_vcycle_bench.rs

@@ -174,9 +174,21 @@ fn batched_device_vcycle_go_no_go() {
                 cells: up_u(&l.cells),
                 red: up_u(&l.red),
                 black: up_u(&l.black),
-                coarse_of: up_u(if l.coarse_of.is_empty() { &[0u32][..] } else { &l.coarse_of }),
-                children_ptr: up_u(if l.children_ptr.is_empty() { &[0u32][..] } else { &l.children_ptr }),
-                children_idx: up_u(if l.children_idx.is_empty() { &[0u32][..] } else { &l.children_idx }),
+                coarse_of: up_u(if l.coarse_of.is_empty() {
+                    &[0u32][..]
+                } else {
+                    &l.coarse_of
+                }),
+                children_ptr: up_u(if l.children_ptr.is_empty() {
+                    &[0u32][..]
+                } else {
+                    &l.children_ptr
+                }),
+                children_idx: up_u(if l.children_idx.is_empty() {
+                    &[0u32][..]
+                } else {
+                    &l.children_idx
+                }),
                 n_cells: l.cells.len(),
                 n_red: l.red.len(),
                 n_black: l.black.len(),
@@ -206,7 +218,11 @@ fn batched_device_vcycle_go_no_go() {
             // Upload the residual into level 0's b.
             stream.memcpy_htod(r_f32, &mut dev[0].b).unwrap();
             let half = |lv: &mut DeviceLevel, colour: u8| {
-                let (list, n_list) = if colour == 0 { (&lv.red, lv.n_red) } else { (&lv.black, lv.n_black) };
+                let (list, n_list) = if colour == 0 {
+                    (&lv.red, lv.n_red)
+                } else {
+                    (&lv.black, lv.n_black)
+                };
                 let n_i = lv.n as i32;
                 let n_list_i = n_list as i32;
                 unsafe {
@@ -371,7 +387,10 @@ fn batched_device_vcycle_go_no_go() {
                 let (db, sb) = cmp(&b[..n], b_ref);
                 let (dx, sx) = cmp(&x[..n], x_ref);
                 let (dr, sr) = cmp(&r[..n], r_ref);
-                println!("    level {l} ({} cells): |Δb| {db:.3e} / {sb:.3e}, |Δx| {dx:.3e} / {sx:.3e}, |Δr| {dr:.3e} / {sr:.3e}", levels[l].cells.len());
+                println!(
+                    "    level {l} ({} cells): |Δb| {db:.3e} / {sb:.3e}, |Δx| {dx:.3e} / {sx:.3e}, |Δr| {dr:.3e} / {sr:.3e}",
+                    levels[l].cells.len()
+                );
             }
         }
         let mut worst = 0.0_f64;
@@ -381,8 +400,13 @@ fn batched_device_vcycle_go_no_go() {
                 worst = worst.max(d);
             }
         }
-        println!("  K = {k}: device V-cycle vs CPU f32 reference: max |Δz| {worst:.3e} on a scale of {z_scale:.3e}");
-        assert!(worst < 1e-4 * z_scale, "device V-cycle disagrees with the CPU one");
+        println!(
+            "  K = {k}: device V-cycle vs CPU f32 reference: max |Δz| {worst:.3e} on a scale of {z_scale:.3e}"
+        );
+        assert!(
+            worst < 1e-4 * z_scale,
+            "device V-cycle disagrees with the CPU one"
+        );
 
         // Timing: 50 V-cycles, per march.
         let reps = 50;