CODEX: Keep pose task streams resident

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

jna/test_pose_corr_jna.cu

@@ -61,11 +61,14 @@ extern "C" {
   int   tp_proc_set_image(void*, int, const float*);
   int   tp_proc_set_tasks(void*, const float*, int, int);
   int   tp_proc_get_tasks(void*, float*, int);
+  int   tp_proc_set_tasks_slot(void*, const float*, int, int, int);
+  int   tp_proc_get_tasks_slot(void*, float*, int, int);
   int   tp_proc_set_const(void*, const char*, const float*, int);
   int   tp_proc_exec_geometry(void*, int);
   int   tp_proc_exec_convert_direct(void*, int, int, int);
   int   tp_proc_set_clt(void*, int, const float*, int);
   int   tp_proc_exec_consolidate(void*, const float*, const float*, int, int, float, float, int*);
+  int   tp_proc_exec_consolidate_slots(void*, int, int, int, int, float, float, int*);
   int   tp_proc_setup_rbg_corr(void*, int, int,int,int,int, float,float,float, int);
   int   tp_proc_exec_corr2d_inter_td(void*, int, float,float,float);
   int   tp_proc_get_corr_indices(void*, int*, int);
@@ -265,8 +268,8 @@ int main(int argc, char** argv){
       snprintf(nm,sizeof(nm),"ftasks%d_it%d",nset,k);
       const int n = (int)(data.elements(nm)/task_size);
       n_set[nset] = n;
-      if(timed(prof_tasks_h2d,[&](){ return tp_proc_set_tasks(
-          p,data.hostFloat(nm),n,n*task_size); })){
+      if(timed(prof_tasks_h2d,[&](){ return tp_proc_set_tasks_slot(
+          p,data.hostFloat(nm),n,n*task_size,nset); })){
         printf("FAIL set_tasks: %s\n",tp_last_error()); return EXIT_FAILURE; }
       // Production GpuQuadJna.execSetTilesOffsets refreshes this small vector
       // for every task set even though the current CUAS policy often repeats it.
@@ -277,8 +280,11 @@ int main(int argc, char** argv){
         printf("FAIL exec_geometry: %s\n", tp_last_error()); return EXIT_FAILURE; }
       { // checkpoint: GPU-filled tasks vs the oracle post-offsets stream (bit-exact)
         std::vector<float> got((size_t)n*task_size);
-        if(timed(prof_tasks_d2h,[&](){ return tp_proc_get_tasks(
-            p,got.data(),n*task_size); })){
+        // Production skips this readback. Keep it as the exact slot oracle, but
+        // exclude it from --profile-chain timing. By Codex on 07/15/2026.
+        const int get_rc = chain_profile ? tp_proc_get_tasks_slot(p,got.data(),n*task_size,nset) :
+            timed(prof_tasks_d2h,[&](){ return tp_proc_get_tasks_slot(p,got.data(),n*task_size,nset); });
+        if(get_rc){
           printf("FAIL get_tasks: %s\n",tp_last_error()); return EXIT_FAILURE; }
         char lbl[64]; snprintf(lbl,sizeof(lbl),"tasks%d_it%d",nset,k);
         snprintf(nm,sizeof(nm),"expected_tasks%d_it%d",nset,k);
@@ -289,15 +295,12 @@ int main(int argc, char** argv){
           p,0,(nset==0)?1:-1,0); })){
         printf("FAIL exec_convert_direct: %s\n", tp_last_error()); return EXIT_FAILURE; }
     }
-    // consolidation: the oracle post-offsets streams = the exact bytes production flattens
-    snprintf(nm,sizeof(nm),"expected_tasks0_it%d",k);
-    const float* cons0 = data.hostFloat(nm);
-    const float* cons1 = nullptr;
-    if (has_mb){ snprintf(nm,sizeof(nm),"expected_tasks1_it%d",k); cons1 = data.hostFloat(nm); }
+    // Consolidation consumes the post-offset streams directly from device slots;
+    // no task D2H -> host flatten -> H2D bridge. By Codex on 07/15/2026.
     int stats[3] = {-1,-1,-1};
-    if(timed(prof_consolidate,[&](){ return tp_proc_exec_consolidate(
-        p,cons0,cons1,n_set[0],task_size,soft_margin,hard_margin,stats); })){
-      printf("FAIL exec_consolidate: %s\n", tp_last_error()); return EXIT_FAILURE; }
+    if(timed(prof_consolidate,[&](){ return tp_proc_exec_consolidate_slots(
+        p,0,has_mb?1:-1,n_set[0],task_size,soft_margin,hard_margin,stats); })){
+      printf("FAIL exec_consolidate_slots: %s\n", tp_last_error()); return EXIT_FAILURE; }
     // the single conj-multiply vs the resident center TD; device tasks = the LAST
     // uploaded set (set 1 in MB mode), exactly the state execCorr2D_inter_TD sees
     const int num_corr = timed(prof_corr,[&](){ return tp_proc_exec_corr2d_inter_td(
@@ -475,12 +478,12 @@ int main(int argc, char** argv){
   }
   if(chain_profile){
     const double total=prof_images+prof_tasks_h2d+prof_cv_h2d+prof_geometry+
-        prof_tasks_d2h+prof_convert+prof_consolidate+prof_corr+prof_indices+
+        prof_convert+prof_consolidate+prof_corr+prof_indices+
         prof_normalize+prof_peak+prof_peaks_d2h;
     printf("production-call chain profile (1 scene, %d cycles, %d task sets):\n",num_iters,num_iters*(has_mb?2:1));
     printf("  images H2D once/scene       %9.3f ms\n",prof_images);
     printf("  tasks H2D + corr-vector H2D %9.3f ms\n",prof_tasks_h2d+prof_cv_h2d);
-    printf("  geometry + tasks D2H        %9.3f ms\n",prof_geometry+prof_tasks_d2h);
+    printf("  geometry (tasks stay GPU)   %9.3f ms\n",prof_geometry);
     printf("  convert_direct              %9.3f ms\n",prof_convert);
     printf("  consolidate                 %9.3f ms\n",prof_consolidate);
     printf("  correlate_inter             %9.3f ms\n",prof_corr);