CODEX: Compare resident pose candidate RMS

Co-authored-by: Codex <codex@elphel.com>

src/main/java/com/elphel/imagej/cuas/rt/CuasPoseRT.java

@@ -1042,9 +1042,8 @@ public class CuasPoseRT {
 							numTiles, capturePrepared);
 			if ((result != null) && !pose_lma_step_path_reported) {
 				pose_lma_step_path_reported = true;
-				System.out.println("CuasPoseRT: resident CUDA float LMA candidate active "+
-						"(raw fx/J reused; Java-double candidate is debug/fallback only; "+
-						"Java RMS/acceptance retained)");
+				System.out.println("CuasPoseRT: resident CUDA float LMA candidate/RMS oracle active "+
+						"(separate candidate fx/J buffers; Java acceptance authoritative)");
 			}
 			return result;
 		}); // production candidate: resident raw fx/J -> parallel prepare/reduce -> fixed 3x3 tail // By Codex on 07/15/2026

src/main/java/com/elphel/imagej/gpu/jna/GpuQuadJna.java

@@ -592,7 +592,11 @@ public class GpuQuadJna extends GpuQuad {
                         packed[IntersceneLmaFloat.LMA_RESULT_FLOATS],
                         packed[IntersceneLmaFloat.LMA_RESULT_FLOATS + 1]) : null;
         return new IntersceneLmaFloat.ResidentLmaStepResult(
-                unpackPoseLmaStep(packed), prepared);
+                unpackPoseLmaStep(packed), prepared,
+                packed[IntersceneLmaFloat.LMA_RESULT_FLOATS],
+                packed[IntersceneLmaFloat.LMA_RESULT_FLOATS + 1],
+                packed[IntersceneLmaFloat.LMA_RESULT_FLOATS + 2],
+                packed[IntersceneLmaFloat.LMA_RESULT_FLOATS + 3]);
     }
 
     private static IntersceneLmaFloat.LmaStepResult unpackPoseLmaStep(final float[] packed) {

src/main/java/com/elphel/imagej/gpu/jna/Stage0.java

@@ -10,7 +10,7 @@ import com.sun.jna.Pointer;
  *        com.elphel.imagej.gpu.jna.Stage0 [kernel_src_dir] [libcudadevrt.a]
  */
 public class Stage0 {
-	private static final int EXPECTED_KERNELS = 32; // 19 base + 2 conditioning + 4 consolidate + 2 peak + pose fx/J + 3 pose LMA + products
+	private static final int EXPECTED_KERNELS = 33; // 19 base + 2 conditioning + 4 consolidate + 2 peak + pose fx/J + 4 pose LMA + products
     public static void main(String[] args) {
         String srcdir = (args.length > 0) ? args[0] : "/home/elphel/git/tile_processor_gpu/src";
         String devrt  = (args.length > 1) ? args[1] : "/usr/local/cuda/targets/x86_64-linux/lib/libcudadevrt.a";

src/main/java/com/elphel/imagej/gpu/jna/TpJna.java

@@ -12,8 +12,8 @@ public interface TpJna extends Library {
     /** NVRTC-compile the kernels in srcdir (+ getTpDefines), cuLink(libcudadevrt), load the module.
      *  Returns an opaque module handle, or null on failure (see tp_last_error()). */
     Pointer tp_create_module(String srcdir, String libcudadevrt);
-    /** Number of kernel functions resolved in the module (32 expected:
-     *  19 base + 2 conditioning + 4 consolidation + 2 peak + pose-fx/J + 3 pose-LMA + LMA-products),
+    /** Number of kernel functions resolved in the module (33 expected:
+     *  19 base + 2 conditioning + 4 consolidation + 2 peak + pose-fx/J + 4 pose-LMA + LMA-products),
      *  or -1 if handle is null. */
     int     tp_module_num_functions(Pointer module);
     /** Last error message (NVRTC/CUDA log), empty if none. */
@@ -155,7 +155,7 @@ public interface TpJna extends Library {
                                       float[] ymfxWeighted, float[] vector,
                                       int numValues, float[] result);
     /** Continue from raw fx/J left by the most recent pose kernel. Result has
-     *  the 19 candidate floats followed by current full/pure RMS. Prepared
+     *  the 19 candidate floats followed by current and candidate full/pure RMS. Prepared
      *  arrays are nullable unless capturePrepared is nonzero. */
     int     tp_proc_exec_pose_lma_resident_step(Pointer proc,
                                       float lambda, float[] weights, float[] y,

src/main/java/com/elphel/imagej/tileprocessor/IntersceneLma.java

@@ -126,6 +126,7 @@ public class IntersceneLma {
 	}
 	private PoseLmaResidentStepProvider poseLmaResidentStepProvider = null;
 	private static boolean poseLmaStepOracleReported = false;
+	private static boolean poseLmaAcceptanceOracleReported = false;
 	private IntersceneLmaFloat.Camera poseFloatReference = null;
 	private IntersceneLmaFloat.Camera poseFloatScene = null;
 	private float [] poseFloatCenters = null;
@@ -1046,6 +1047,11 @@ public class IntersceneLma {
 		if (last_ymfx == null) {
 			return null; // need to re-init/restart LMA
 		}
+		if (residentCandidateValid && (residentResult != null) && (debug_level >= 1) &&
+				!poseLmaAcceptanceOracleReported) {
+			reportPoseRmsDecision(residentResult, this.last_rms, rms, rms_diff);
+			poseLmaAcceptanceOracleReported = true;
+		}
 
 		this.good_or_bad_rms = rms.clone();
 		if (rms[0] < this.last_rms[0]) { // improved
@@ -1928,6 +1934,32 @@ public class IntersceneLma {
 				label, poseStep.valid, maxProducts, maxDelta, maxCandidate));
 	}
 
+	private static void reportPoseRmsDecision(
+			final IntersceneLmaFloat.ResidentLmaStepResult resident,
+			final double [] javaCurrent,
+			final double [] javaCandidate,
+			final double rmsDiff) {
+		final boolean finite = Float.isFinite(resident.currentRms) &&
+				Float.isFinite(resident.currentPureRms) &&
+				Float.isFinite(resident.candidateRms) &&
+				Float.isFinite(resident.candidatePureRms);
+		final boolean floatImproved = finite && (resident.candidateRms < resident.currentRms);
+		final boolean doubleImproved = javaCandidate[0] < javaCurrent[0];
+		final boolean floatStop = floatImproved &&
+				(resident.candidateRms >= resident.currentRms * (1.0f - (float) rmsDiff));
+		final boolean doubleStop = doubleImproved &&
+				(javaCandidate[0] >= javaCurrent[0] * (1.0 - rmsDiff));
+		System.out.println(String.format(
+				"IntersceneLma resident CUDA-float vs Java-double RMS/decision: finite=%s, |current RMS|=%g, |current pure RMS|=%g, |candidate RMS|=%g, |candidate pure RMS|=%g, improved-match=%s, stop-match=%s",
+				finite,
+				Math.abs(resident.currentRms - javaCurrent[0]),
+				Math.abs(resident.currentPureRms - javaCurrent[1]),
+				Math.abs(resident.candidateRms - javaCandidate[0]),
+				Math.abs(resident.candidatePureRms - javaCandidate[1]),
+				floatImproved == doubleImproved,
+				floatStop == doubleStop));
+	}
+
 	private static final class DoubleLmaStepResult {
 		final double [] products;
 		final double [] delta;

src/main/java/com/elphel/imagej/tileprocessor/IntersceneLmaFloat.java

@@ -42,7 +42,7 @@ public final class IntersceneLmaFloat {
 	public static final int NUM_COMPONENTS = 2;
 	public static final int LMA_PRODUCTS = NUM_PARAMS * NUM_PARAMS + NUM_PARAMS;
 	public static final int LMA_RESULT_FLOATS = 1 + LMA_PRODUCTS + 2 * NUM_PARAMS;
-	public static final int LMA_RESIDENT_RESULT_FLOATS = LMA_RESULT_FLOATS + 2;
+	public static final int LMA_RESIDENT_RESULT_FLOATS = LMA_RESULT_FLOATS + 4;
 	public static final float INFINITY_DISPARITY = 0.01f;
 
 	static final int ERS_XYZ = 0;
@@ -216,10 +216,24 @@ public final class IntersceneLmaFloat {
 	public static final class ResidentLmaStepResult {
 		public final LmaStepResult step;
 		public final PreparedLma prepared;
+		public final float currentRms;
+		public final float currentPureRms;
+		public final float candidateRms;
+		public final float candidatePureRms;
 
-		public ResidentLmaStepResult(final LmaStepResult step, final PreparedLma prepared) {
+		public ResidentLmaStepResult(
+				final LmaStepResult step,
+				final PreparedLma prepared,
+				final float currentRms,
+				final float currentPureRms,
+				final float candidateRms,
+				final float candidatePureRms) {
 			this.step = step;
 			this.prepared = prepared;
+			this.currentRms = currentRms;
+			this.currentPureRms = currentPureRms;
+			this.candidateRms = candidateRms;
+			this.candidatePureRms = candidatePureRms;
 		}
 	}