embedded3 R6-3: RTX_E3_FIELD_P_ONLY=1 (default off) — the rebuild's field transfer limited to p when the classification and imposition ran on the device

The download waits until the step's rebuild route is known; with the device
mask and device imposition the host rebuild reads and writes only the
pressure (the fresh cells' refill), so u, v, w and the old velocities stay on
the device (their upload wrote back the values just downloaded). Not under
RTX_E3_BAND_CHECK (its references read the whole mirror).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/step/device.rs

@@ -292,6 +292,24 @@ impl DeviceStep {
         rt.stream.synchronize().expect("sync");
     }
 
+    /// R6-3 (`RTX_E3_FIELD_P_ONLY=1`): the pressure alone down (the
+    /// device-imposed rebuild's only host read).
+    pub fn download_p(&self, field: &mut Field) {
+        let rt = runtime();
+        assert_eq!(field.grid, self.grid);
+        rt.stream.memcpy_dtoh(&self.p, &mut field.p).expect("p");
+        rt.stream.synchronize().expect("sync");
+    }
+
+    /// R6-3: the pressure alone up (the fresh cells' refill; u, v, w were
+    /// not touched on the host).
+    pub fn upload_p(&mut self, field: &Field) {
+        let rt = runtime();
+        assert_eq!(field.grid, self.grid);
+        rt.stream.memcpy_htod(&field.p, &mut self.p).expect("p");
+        rt.stream.synchronize().expect("sync");
+    }
+
     /// The arrays a moving body's mask rebuild reads (P1-3 (c)): the
     /// corrected velocities, the pressure and the previous velocities.
     pub fn download_for_rebuild(&self, field: &mut Field) {

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/step/device/cut.rs

@@ -909,7 +909,16 @@ impl DeviceStep {
             // w, p (fresh cells take their neighbours' pressure) and the
             // previous velocities (the imposition's source).
             let mut field = self.host_field.take().unwrap_or_else(|| Field::new(g));
+            // R6-3 (`RTX_E3_FIELD_P_ONLY=1`, default off; not under the band
+            // check, whose references read the whole mirror): the download
+            // waits until the step's rebuild route is known — with the device
+            // classification and imposition the host rebuild reads and writes
+            // the pressure alone (the fresh cells' refill), so only p moves.
+            let p_only = super::poisson_setup::field_p_only()
+                && std::env::var("RTX_E3_BAND_CHECK").is_err();
+            if !p_only {
                 self.download_for_rebuild(&mut field);
+            }
             let l_down = lap.elapsed();
             // R6-1 (default ON, `RTX_E3_GEOM_DEVICE=0` off): the end-of-step
             // cut geometry on the device, its face tables written into the
@@ -985,10 +994,30 @@ impl DeviceStep {
                     );
                 }
             }
+            let mut l_down = l_down;
+            let partial = p_only
+                && self
+                    .solver
+                    .pending_mask
+                    .as_ref()
+                    .is_some_and(|u| u.impose_on_device);
+            if p_only {
+                let ld = Instant::now();
+                if partial {
+                    self.download_p(&mut field);
+                } else {
+                    self.download_for_rebuild(&mut field);
+                }
+                l_down += ld.elapsed();
+            }
             let l_geom = lap.elapsed();
             fresh_cells = self.solver.rebuild_moving_mask(&mut field, dt, t_new);
             let l_mask = lap.elapsed();
+            if partial {
+                self.upload_p(&field);
+            } else {
                 self.upload_after_rebuild(&field);
+            }
             self.host_field = Some(field);
             let l_up = lap.elapsed();
             // P1-5: the persistent tables updated in place when the mask has a

crates/specialized/rtx-cfd/src/solvers/incompressible/embedded3/step/device/poisson_setup.rs

@@ -52,3 +52,10 @@ pub(super) fn device_grid(solver: &Solver, g: Grid, dt: f64) -> Option<OperatorG
         dt,
     })
 }
+
+/// `RTX_E3_FIELD_P_ONLY=1` (default off): the rebuild's field transfer
+/// limited to the pressure when the classification and the imposition ran on
+/// the device.
+pub(super) fn field_p_only() -> bool {
+    std::env::var("RTX_E3_FIELD_P_ONLY").is_ok_and(|v| v == "1")
+}