Feat: Add producer-private adaptive HIP PSF prebinning

Move adaptive tile16 selection and support-expanded CPU binning into a
per-producer HipPsfPreparedChunk computed before acquiring the shared
sink lock. The sink keeps sole ownership of the GPU HDR, stream, device
scratch and upload staging; submit_prepared runs under the lock and
preserves the existing single-stream ordering and direct fallback path.
The legacy hip_psf_sink_submit entry point prepares an internal chunk so
single-threaded replay stays compatible, and default accumulation remains
atomic.

src/frame.c

@@ -136,6 +136,8 @@ typedef struct {
   const PsfKernelCache *cache;
 #ifdef PSF_BACKEND_HIP
   HipPsfSink *hip;
+  HipPsfPreparedChunk *prepared_chunk; /* Worker-owned CPU bin. */
+  int prepared_ready;
   omp_lock_t *hip_lock; /* Shared sink lock; acquired only per chunk/fallback. */
   int borrowed_hip;
   double submit_seconds;
@@ -165,7 +167,8 @@ static int psf_event_sink_init(PsfEventSink *sink, double *hdr, int width,
   if (cache != NULL) {
 #ifdef PSF_BACKEND_DUMMY
     if (dummy_psf_sink_create(&sink->dummy, width, height,
-                              PSF_EVENT_SINK_CAPACITY)) {
+                              PSF_EVENT_SINK_CAPACITY,
+                              cache->max_radius_pixels)) {
       fputs("Dummy PSF backend initialization failed\n", stderr);
       dummy_psf_sink_destroy(sink->dummy);
       sink->dummy = NULL;
@@ -198,6 +201,22 @@ static int psf_event_sink_init(PsfEventSink *sink, double *hdr, int width,
   return 0;
 }
 
+#ifdef PSF_BACKEND_HIP
+static int psf_event_sink_prepare(PsfEventSink *sink) {
+  if (sink->prepared_chunk == NULL || sink->count == 0 ||
+      sink->prepared_ready || sink->failed)
+    return sink->failed ? -1 : 0;
+  if (hip_psf_prepared_chunk_prepare(sink->prepared_chunk, sink->events,
+                                     sink->count, sink->hip_message,
+                                     sizeof sink->hip_message)) {
+    psf_event_sink_mark_failed(sink, "parallel event preparation");
+    return -1;
+  }
+  sink->prepared_ready = 1;
+  return 0;
+}
+#endif
+
 static int psf_event_sink_flush_unlocked(PsfEventSink *sink) {
   if (sink->failed)
     return -1;
@@ -207,12 +226,18 @@ static int psf_event_sink_flush_unlocked(PsfEventSink *sink) {
 #endif
 #ifdef PSF_BACKEND_HIP
   if (sink->hip != NULL) {
-    if (hip_psf_sink_submit(sink->hip, sink->events, sink->count,
-                            sink->hip_message, sizeof sink->hip_message)) {
+    const int submit_result = sink->prepared_ready
+        ? hip_psf_sink_submit_prepared(sink->hip, sink->prepared_chunk,
+                                       sink->events, sink->count,
+                                       sink->hip_message, sizeof sink->hip_message)
+        : hip_psf_sink_submit(sink->hip, sink->events, sink->count,
+                              sink->hip_message, sizeof sink->hip_message);
+    if (submit_result) {
       psf_event_sink_mark_failed(sink, "event submission");
       return -1;
     }
     sink->count = 0;
+    sink->prepared_ready = 0;
     return 0;
   }
 #endif
@@ -226,6 +251,7 @@ static int psf_event_sink_flush_unlocked(PsfEventSink *sink) {
 static int psf_event_sink_flush(PsfEventSink *sink) {
 #ifdef PSF_BACKEND_HIP
   const double start = omp_get_wtime();
+  if (psf_event_sink_prepare(sink)) return -1;
   if (sink->hip_lock) omp_set_lock(sink->hip_lock);
 #endif
   const int result = psf_event_sink_flush_unlocked(sink);
@@ -289,6 +315,7 @@ static int psf_event_sink_destroy(PsfEventSink *sink) {
 #ifdef PSF_BACKEND_HIP
   if (sink->borrowed_hip) {
     const int result = psf_event_sink_flush(sink);
+    hip_psf_prepared_chunk_destroy(sink->prepared_chunk);
     free(sink->events);
     return result;
   }
@@ -1184,6 +1211,7 @@ static int splat_catalog_tile(const Star *stars, size_t count,
 #else
 #ifdef PSF_BACKEND_HIP
       const double submit_start = omp_get_wtime();
+      if (psf_event_sink_prepare(context->event_sink)) return -1;
       if (context->event_sink->hip_lock)
         omp_set_lock(context->event_sink->hip_lock);
 #endif
@@ -1487,6 +1515,11 @@ size_t frame_splat_catalog(const FrameLensMesh *mesh,
       snprintf(local.hip_message, sizeof local.hip_message, "worker event allocation failed");
       psf_event_sink_mark_failed(&local, "producer initialization");
     }
+    if (!local.failed && hip_psf_sink_is_adaptive(owner.hip) &&
+        hip_psf_prepared_chunk_create(&local.prepared_chunk, owner.hip,
+                                       local.hip_message,
+                                       sizeof local.hip_message))
+      psf_event_sink_mark_failed(&local, "producer preparation allocation");
     const size_t worker = (size_t)omp_get_thread_num();
     const size_t workers = (size_t)omp_get_num_threads();
     size_t completed = 0, next_report = 8;
@@ -1524,6 +1557,16 @@ size_t frame_splat_catalog(const FrameLensMesh *mesh,
   if (psf_event_sink_destroy(&owner)) hip_failed = 1;
   fprintf(stderr, "HIP producers: %d workers; summed generation %.3f s, submission/fallback %.3f s; wall %.3f s\n",
       hip_workers, hip_generate_seconds, hip_submit_seconds, omp_get_wtime() - hip_start);
+  fprintf(stderr,
+      "HIP PSF accumulation: atomic %zu batches, tile16 %zu batches; selector %.3f s, binning %.3f s\n",
+      owner.hip_timing.atomic_batch_count, owner.hip_timing.tile_batch_count,
+      owner.hip_timing.selection_seconds, owner.hip_timing.bin_seconds);
+  if (owner.hip_timing.tile_batch_count)
+    fprintf(stderr,
+        "HIP PSF tile workload: references %zu, tasks %zu, merges %zu; peak references/chunk %zu\n",
+        owner.hip_timing.tile_reference_count, owner.hip_timing.tile_task_count,
+        owner.hip_timing.tile_merge_count,
+        owner.hip_timing.peak_tile_reference_count);
   copy_psf_splat_stats(psf_stats, (CatalogSplatStats){
       .images = hip_images, .direct_fallbacks = hip_direct,
       .cached_wing_clipped = hip_clipped, .discarded_below_min_y = hip_discarded,

src/hip_psf.h

@@ -12,10 +12,15 @@ extern "C" {
 /* Opaque HIP state for one HDR framebuffer. It owns reusable device buffers
  * for PsfCachedEvent chunks, immutable cache weights, and double RGB HDR. */
 typedef struct HipPsfSink HipPsfSink;
+typedef struct HipPsfPreparedChunk HipPsfPreparedChunk;
 
 typedef struct {
   size_t event_count, batch_count, timed_batch_count;
   double upload_seconds, kernel_seconds, download_seconds;
+  size_t atomic_batch_count, tile_batch_count;
+  size_t tile_reference_count, tile_task_count, tile_merge_count;
+  size_t peak_tile_reference_count;
+  double selection_seconds, bin_seconds;
 } HipPsfTiming;
 
 int hip_psf_available(char *message, size_t message_size);
@@ -39,6 +44,25 @@ int hip_psf_sink_load_hdr(HipPsfSink *sink, const double *hdr,
 int hip_psf_sink_submit(HipPsfSink *sink, const PsfCachedEvent *events,
                         size_t event_count, char *message, size_t message_size);
 
+/* Optional adaptive path: each producer owns one preparation buffer. Prepare
+ * runs without the shared sink lock; submit_prepared runs under that lock and
+ * preserves the existing single-stream HDR ordering. Atomic sinks may keep
+ * using hip_psf_sink_submit directly. */
+int hip_psf_sink_is_adaptive(const HipPsfSink *sink);
+int hip_psf_prepared_chunk_create(HipPsfPreparedChunk **out,
+                                  const HipPsfSink *sink, char *message,
+                                  size_t message_size);
+int hip_psf_prepared_chunk_prepare(HipPsfPreparedChunk *chunk,
+                                   const PsfCachedEvent *events,
+                                   size_t event_count, char *message,
+                                   size_t message_size);
+int hip_psf_sink_submit_prepared(HipPsfSink *sink,
+                                 HipPsfPreparedChunk *chunk,
+                                 const PsfCachedEvent *events,
+                                 size_t event_count, char *message,
+                                 size_t message_size);
+void hip_psf_prepared_chunk_destroy(HipPsfPreparedChunk *chunk);
+
 /* Completes all work and overwrites hdr with double linear RGB device HDR. */
 int hip_psf_sink_finish(HipPsfSink *sink, double *hdr, char *message,
                         size_t message_size);

src/hip_psf.hip

@@ -7,8 +7,50 @@
 #include <limits>
 #include <chrono>
 #include <cstring>
+#include <vector>
+#include <algorithm>
+#include <climits>
+#include <cstdlib>
 
-enum { PSF_LANES_PER_EVENT = 32, PSF_THREADS_PER_BLOCK = 128 };
+enum { PSF_LANES_PER_EVENT = 32, PSF_THREADS_PER_BLOCK = 128,
+       TILE_SIZE = 16, TILE_TASK_EVENTS = 256 };
+
+struct ProductionTileTask { unsigned tile, first, count, partial; };
+struct ProductionTileMerge { unsigned tile, first, count; };
+struct PreparedEvent {
+  int bx, by, support;
+  size_t phase_base;
+  double tx, ty, r, g, b;
+};
+
+struct HipTileScratch {
+  PsfCachedEvent *host = nullptr, *events = nullptr;
+  PreparedEvent *prepared = nullptr;
+  int *bounds = nullptr;
+  unsigned *refs = nullptr;
+  ProductionTileTask *tasks = nullptr;
+  ProductionTileMerge *merges = nullptr;
+  double *partials = nullptr;
+  hipEvent_t upload_start = nullptr, upload_end = nullptr;
+  hipEvent_t start = nullptr, end = nullptr;
+  size_t count = 0;
+  /* Submission-owned staging remains stable until collect_tile completes. */
+  std::vector<unsigned> host_refs;
+  std::vector<ProductionTileTask> host_tasks;
+  std::vector<ProductionTileMerge> host_merges;
+};
+
+struct HipPsfPreparedChunk {
+  const HipPsfSink *owner = nullptr;
+  size_t count = 0;
+  bool use_tile = false;
+  double selection_seconds = 0.0, bin_seconds = 0.0;
+  std::vector<unsigned> counts, offsets, cursors, touched, host_refs;
+  std::vector<ProductionTileTask> host_tasks;
+  std::vector<ProductionTileMerge> host_merges;
+  std::vector<unsigned> selector_marks;
+  unsigned selector_generation = 0;
+};
 
 /* Two reusable staging slots bound host/device queue memory. A slot may be
  * overwritten only after its kernel_end event completes. All API calls for a
@@ -32,6 +74,9 @@ struct HipPsfSink {
   double max_radius_pixels = 0.0;
   hipEvent_t download_start = nullptr, download_end = nullptr;
   HipPsfTiming timing = {};
+  bool adaptive = false;
+  HipTileScratch tile;
+  HipPsfPreparedChunk *legacy_prepared = nullptr;
   std::chrono::steady_clock::time_point last_report = std::chrono::steady_clock::now();
 };
 
@@ -59,6 +104,22 @@ static int collect_batch(HipPsfSink *sink, HipPsfBatch *batch,
   return 0;
 }
 
+static int collect_tile(HipPsfSink *sink, char *message, size_t message_size) {
+  if (!sink->tile.count) return 0;
+  float upload_ms = 0.0f, kernel_ms = 0.0f;
+  if (report(hipEventSynchronize(sink->tile.end), message, message_size) ||
+      report(hipEventElapsedTime(&upload_ms, sink->tile.upload_start,
+                                 sink->tile.upload_end), message, message_size) ||
+      report(hipEventElapsedTime(&kernel_ms, sink->tile.start, sink->tile.end),
+             message, message_size)) return -1;
+  sink->timing.upload_seconds += upload_ms * 1e-3;
+  sink->timing.kernel_seconds += kernel_ms * 1e-3;
+  ++sink->timing.timed_batch_count;
+  sink->completed_events += sink->tile.count;
+  sink->tile.count = 0;
+  return 0;
+}
+
 __device__ static size_t weight_index(int phase_resolution, int radius_pixels,
                                       int phase_x, int phase_y, int offset_x, int offset_y) {
   const size_t nodes = (size_t)phase_resolution + 1;
@@ -121,6 +182,94 @@ __global__ static void splat_kernel(const PsfCachedEvent *events, size_t count,
   }
 }
 
+__global__ static void prepare_tile_events(
+    const PsfCachedEvent *events, size_t count, int phases, int radius,
+    double max_radius, PreparedEvent *prepared, int *bounds) {
+  const size_t thread = (size_t)blockIdx.x * blockDim.x + threadIdx.x;
+  const size_t index = thread / PSF_LANES_PER_EVENT;
+  const int lane = thread % PSF_LANES_PER_EVENT;
+  if (index >= count) return;
+  const PsfCachedEvent event = events[index];
+  const int bx = (int)floor(event.x), by = (int)floor(event.y);
+  const double fx = event.x - bx, fy = event.y - by;
+  const double ax = fx * phases, ay = fy * phases;
+  const int x0 = (int)floor(ax), y0 = (int)floor(ay);
+  const int support = (int)fmin(ceil(event.support_radius), max_radius);
+  if (!lane)
+    prepared[index] = {bx, by, support,
+        weight_index(phases, radius, x0, y0, 0, 0), ax - x0, ay - y0,
+        event.color.r * event.flux, event.color.g * event.flux,
+        event.color.b * event.flux};
+  const int side = 2 * radius + 1;
+  for (int row = lane; row < side; row += PSF_LANES_PER_EVENT) {
+    int first = 1, last = 0;
+    const int dy = row - radius;
+    if (dy >= -support && dy <= support)
+      row_range(event.support_radius, fx, fy, support, dy, &first, &last);
+    bounds[2 * (index * side + row)] = first;
+    bounds[2 * (index * side + row) + 1] = last;
+  }
+}
+
+__global__ static void accumulate_tiles(
+    const PreparedEvent *events, const int *bounds, const unsigned *refs,
+    const ProductionTileTask *tasks, int tiles_x, int width, int height,
+    const float *weights, int phases, int radius, double *partials,
+    double *hdr) {
+  const ProductionTileTask task = tasks[blockIdx.x];
+  const int local = threadIdx.x;
+  const int px = (task.tile % tiles_x) * TILE_SIZE + local % TILE_SIZE;
+  const int py = (task.tile / tiles_x) * TILE_SIZE + local / TILE_SIZE;
+  const int side = 2 * radius + 1;
+  const size_t plane = (size_t)side * side;
+  double r = 0.0, g = 0.0, b = 0.0;
+  if (px < width && py < height) for (unsigned k = 0; k < task.count; ++k) {
+    const unsigned id = refs[task.first + k];
+    const PreparedEvent event = events[id];
+    const int dx = px - event.bx, dy = py - event.by;
+    if (dy < -event.support || dy > event.support) continue;
+    const size_t row = 2 * ((size_t)id * side + dy + radius);
+    if (dx < bounds[row] || dx > bounds[row + 1]) continue;
+    const size_t p = event.phase_base + (ptrdiff_t)dy * side + dx;
+    const double w00 = weights[p], w10 = weights[p + plane];
+    const double w01 = weights[p + (phases + 1) * plane];
+    const double w11 = weights[p + (phases + 2) * plane];
+    const double weight = (1.0 - event.ty) *
+                              ((1.0 - event.tx) * w00 + event.tx * w10) +
+                          event.ty *
+                              ((1.0 - event.tx) * w01 + event.tx * w11);
+    r += event.r * weight; g += event.g * weight; b += event.b * weight;
+  }
+  if (task.partial == UINT_MAX) {
+    if (px < width && py < height) {
+      const size_t p = 3 * ((size_t)py * width + px);
+      hdr[p] += r; hdr[p + 1] += g; hdr[p + 2] += b;
+    }
+  } else {
+    const size_t p = 3 * ((size_t)task.partial * TILE_SIZE * TILE_SIZE + local);
+    partials[p] = r; partials[p + 1] = g; partials[p + 2] = b;
+  }
+}
+
+__global__ static void merge_tiles(const ProductionTileMerge *merges,
+                                   const ProductionTileTask *tasks, int tiles_x,
+                                   int width, int height,
+                                   const double *partials, double *hdr) {
+  const ProductionTileMerge merge = merges[blockIdx.x];
+  const int local = threadIdx.x;
+  const int px = (merge.tile % tiles_x) * TILE_SIZE + local % TILE_SIZE;
+  const int py = (merge.tile / tiles_x) * TILE_SIZE + local / TILE_SIZE;
+  if (px >= width || py >= height) return;
+  double r = 0.0, g = 0.0, b = 0.0;
+  for (unsigned k = 0; k < merge.count; ++k) {
+    const ProductionTileTask task = tasks[merge.first + k];
+    const size_t p = 3 * ((size_t)task.partial * TILE_SIZE * TILE_SIZE + local);
+    r += partials[p]; g += partials[p + 1]; b += partials[p + 2];
+  }
+  const size_t p = 3 * ((size_t)py * width + px);
+  hdr[p] += r; hdr[p + 1] += g; hdr[p + 2] += b;
+}
+
 extern "C" int hip_psf_available(char *message, size_t message_size) {
   int count = 0;
   if (report(hipGetDeviceCount(&count), message, message_size)) return -1;
@@ -157,6 +306,23 @@ extern "C" int hip_psf_sink_create(HipPsfSink **out, int width, int height,
   sink->hdr_values = (size_t)width * (size_t)height * 3;
   sink->width = width; sink->height = height; sink->phase_resolution = cache->phase_resolution;
   sink->radius_pixels = cache->radius_pixels; sink->max_radius_pixels = cache->max_radius_pixels;
+  const char *accumulation = std::getenv("HIP_PSF_ACCUMULATION");
+  if (accumulation && std::strcmp(accumulation, "atomic") &&
+      std::strcmp(accumulation, "adaptive")) {
+    if (message && message_size)
+      std::snprintf(message, message_size,
+                    "HIP_PSF_ACCUMULATION must be atomic or adaptive");
+    hip_psf_sink_destroy(sink); return -1;
+  }
+  sink->adaptive = accumulation && !std::strcmp(accumulation, "adaptive");
+  if (sink->adaptive &&
+      (event_capacity > UINT_MAX ||
+       event_capacity > std::numeric_limits<size_t>::max() / side / 2 /
+                            sizeof(int))) {
+    if (message && message_size)
+      std::snprintf(message, message_size, "adaptive HIP PSF size overflow");
+    hip_psf_sink_destroy(sink); return -1;
+  }
   const size_t weight_count = nodes * nodes * side * side;
   if (report(hipGetDevice(&sink->device), message, message_size) ||
       report(hipStreamCreate(&sink->stream), message, message_size) ||
@@ -178,6 +344,37 @@ extern "C" int hip_psf_sink_create(HipPsfSink **out, int width, int height,
       hip_psf_sink_destroy(sink); return -1;
     }
   }
+  if (sink->adaptive) {
+    const size_t tiles_x = ((size_t)width + TILE_SIZE - 1) / TILE_SIZE;
+    const size_t tiles_y = ((size_t)height + TILE_SIZE - 1) / TILE_SIZE;
+    const size_t tile_count = tiles_x * tiles_y;
+    const size_t bounds_count = event_capacity * side * 2;
+    const size_t ref_bytes = 32u * 1024u * 1024u;
+    const size_t task_bytes = 2u * 1024u * 1024u;
+    const size_t partial_bytes = 128u * 1024u * 1024u;
+    if (hip_psf_prepared_chunk_create(&sink->legacy_prepared, sink,
+                                      message, message_size)) {
+      hip_psf_sink_destroy(sink); return -1;
+    }
+    if (report(hipHostMalloc(&sink->tile.host,
+                             event_capacity * sizeof(PsfCachedEvent)), message, message_size) ||
+        report(hipMalloc(&sink->tile.events,
+                         event_capacity * sizeof(PsfCachedEvent)), message, message_size) ||
+        report(hipMalloc(&sink->tile.prepared,
+                         event_capacity * sizeof(PreparedEvent)), message, message_size) ||
+        report(hipMalloc(&sink->tile.bounds, bounds_count * sizeof(int)), message, message_size) ||
+        report(hipMalloc(&sink->tile.refs, ref_bytes), message, message_size) ||
+        report(hipMalloc(&sink->tile.tasks, task_bytes), message, message_size) ||
+        report(hipMalloc(&sink->tile.merges,
+                         tile_count * sizeof(ProductionTileMerge)), message, message_size) ||
+        report(hipMalloc(&sink->tile.partials, partial_bytes), message, message_size) ||
+        report(hipEventCreate(&sink->tile.upload_start), message, message_size) ||
+        report(hipEventCreate(&sink->tile.upload_end), message, message_size) ||
+        report(hipEventCreate(&sink->tile.start), message, message_size) ||
+        report(hipEventCreate(&sink->tile.end), message, message_size)) {
+      hip_psf_sink_destroy(sink); return -1;
+    }
+  }
   /* The caller may release the immutable host cache after creation. */
   if (report(hipStreamSynchronize(sink->stream), message, message_size)) {
     hip_psf_sink_destroy(sink); return -1;
@@ -185,14 +382,239 @@ extern "C" int hip_psf_sink_create(HipPsfSink **out, int width, int height,
   *out = sink; ok(message, message_size); return 0;
 }
 
-extern "C" int hip_psf_sink_submit(HipPsfSink *sink, const PsfCachedEvent *events,
-                                     size_t event_count, char *message, size_t message_size) {
+extern "C" int hip_psf_sink_is_adaptive(const HipPsfSink *sink) {
+  return sink && sink->adaptive;
+}
+
+extern "C" int hip_psf_prepared_chunk_create(HipPsfPreparedChunk **out,
+    const HipPsfSink *sink, char *message, size_t message_size) {
+  if (!out || !sink || !sink->adaptive) {
+    if (message && message_size) std::snprintf(message, message_size,
+                                                "invalid adaptive chunk creation");
+    return -1;
+  }
+  *out = nullptr;
+  try {
+    HipPsfPreparedChunk *chunk = new HipPsfPreparedChunk;
+    *out = chunk;
+    chunk->owner = sink;
+    const size_t tile_count = ((size_t)sink->width + TILE_SIZE - 1) / TILE_SIZE *
+                              (((size_t)sink->height + TILE_SIZE - 1) / TILE_SIZE);
+    chunk->counts.resize(tile_count);
+    chunk->offsets.resize(tile_count);
+    chunk->cursors.resize(tile_count);
+    chunk->selector_marks.resize(((size_t)sink->width + 31) / 32 *
+                                 (((size_t)sink->height + 31) / 32));
+    *out = chunk;
+  } catch (...) {
+    delete *out;
+    *out = nullptr;
+    if (message && message_size) std::snprintf(message, message_size,
+                                                "adaptive chunk host allocation failed");
+    return -1;
+  }
+  ok(message, message_size); return 0;
+}
+
+extern "C" void hip_psf_prepared_chunk_destroy(HipPsfPreparedChunk *chunk) {
+  delete chunk;
+}
+
+extern "C" int hip_psf_prepared_chunk_prepare(HipPsfPreparedChunk *chunk,
+    const PsfCachedEvent *events, size_t event_count,
+    char *message, size_t message_size) {
+  if (!chunk || !chunk->owner || (event_count && !events) ||
+      event_count > chunk->owner->event_capacity) {
+    if (message && message_size) std::snprintf(message, message_size,
+                                                "invalid adaptive chunk");
+    return -1;
+  }
+  const HipPsfSink *sink = chunk->owner;
+  chunk->count = event_count;
+  chunk->use_tile = false;
+  chunk->selection_seconds = chunk->bin_seconds = 0.0;
+  if (event_count < 8192) { ok(message, message_size); return 0; }
+  const auto select_start = std::chrono::steady_clock::now();
+  if (++chunk->selector_generation == 0) {
+    std::fill(chunk->selector_marks.begin(), chunk->selector_marks.end(), 0u);
+    chunk->selector_generation = 1;
+  }
+  const int selector_x = (sink->width + 31) / 32;
+  size_t occupied = 0;
+  for (size_t i = 0; i < event_count; ++i) {
+    const int x = (int)floor(events[i].x), y = (int)floor(events[i].y);
+    if (x < 0 || x >= sink->width || y < 0 || y >= sink->height) continue;
+    unsigned &mark = chunk->selector_marks[(size_t)(y / 32) * selector_x + x / 32];
+    if (mark != chunk->selector_generation) {
+      mark = chunk->selector_generation;
+      ++occupied;
+    }
+  }
+  bool use_tile = occupied && event_count / occupied >= 32;
+  chunk->selection_seconds = std::chrono::duration<double>(
+      std::chrono::steady_clock::now() - select_start).count();
+  if (!use_tile) { ok(message, message_size); return 0; }
+  const auto bin_start = std::chrono::steady_clock::now();
+  for (unsigned index : chunk->touched) chunk->counts[index] = 0;
+  chunk->touched.clear(); chunk->host_tasks.clear(); chunk->host_merges.clear();
+  const int tiles_x = (sink->width + TILE_SIZE - 1) / TILE_SIZE;
+  const size_t ref_capacity = 32u * 1024u * 1024u / sizeof(unsigned);
+  const size_t task_capacity = 2u * 1024u * 1024u / sizeof(ProductionTileTask);
+  const size_t partial_capacity = 128u * 1024u * 1024u /
+                                  (TILE_SIZE * TILE_SIZE * 3 * sizeof(double));
+  size_t refs_count = 0;
+  try {
+    for (size_t i = 0; i < event_count && use_tile; ++i) {
+      const int support = (int)fmin(ceil(events[i].support_radius), sink->max_radius_pixels);
+      const int bx = (int)floor(events[i].x), by = (int)floor(events[i].y);
+      const int x0 = std::max(0, bx - support), x1 = std::min(sink->width - 1, bx + support);
+      const int y0 = std::max(0, by - support), y1 = std::min(sink->height - 1, by + support);
+      if (x0 > x1 || y0 > y1) continue;
+      for (int y = y0 / TILE_SIZE; y <= y1 / TILE_SIZE && use_tile; ++y)
+        for (int x = x0 / TILE_SIZE; x <= x1 / TILE_SIZE && use_tile; ++x) {
+          const unsigned index = (unsigned)(y * tiles_x + x);
+          if (!chunk->counts[index]) chunk->touched.push_back(index);
+          ++chunk->counts[index];
+          if (++refs_count > ref_capacity) { use_tile = false; break; }
+        }
+    }
+    size_t partial_count = 0;
+    if (use_tile) {
+      chunk->host_refs.resize(refs_count);
+      size_t offset = 0;
+      for (unsigned index : chunk->touched) {
+        chunk->offsets[index] = offset;
+        chunk->cursors[index] = offset;
+        offset += chunk->counts[index];
+      }
+      for (size_t i = 0; i < event_count; ++i) {
+        const int support = (int)fmin(ceil(events[i].support_radius), sink->max_radius_pixels);
+        const int bx = (int)floor(events[i].x), by = (int)floor(events[i].y);
+        const int x0 = std::max(0, bx - support), x1 = std::min(sink->width - 1, bx + support);
+        const int y0 = std::max(0, by - support), y1 = std::min(sink->height - 1, by + support);
+        if (x0 > x1 || y0 > y1) continue;
+        for (int y = y0 / TILE_SIZE; y <= y1 / TILE_SIZE; ++y)
+          for (int x = x0 / TILE_SIZE; x <= x1 / TILE_SIZE; ++x) {
+            const unsigned index = (unsigned)(y * tiles_x + x);
+            chunk->host_refs[chunk->cursors[index]++] = (unsigned)i;
+          }
+      }
+      for (unsigned index : chunk->touched) {
+        const size_t count = chunk->counts[index];
+        const unsigned first_task = (unsigned)chunk->host_tasks.size();
+        if (chunk->host_tasks.size() + (count + TILE_TASK_EVENTS - 1) / TILE_TASK_EVENTS >
+            task_capacity) { use_tile = false; break; }
+        for (size_t k = 0; k < count; k += TILE_TASK_EVENTS) {
+          const unsigned length = (unsigned)std::min((size_t)TILE_TASK_EVENTS, count - k);
+          const unsigned partial = count <= TILE_TASK_EVENTS ? UINT_MAX : (unsigned)partial_count++;
+          if (partial_count > partial_capacity) { use_tile = false; break; }
+          chunk->host_tasks.push_back({index, (unsigned)(chunk->offsets[index] + k), length, partial});
+        }
+        if (!use_tile) break;
+        if (count > TILE_TASK_EVENTS)
+          chunk->host_merges.push_back({index, first_task,
+                                        (unsigned)(chunk->host_tasks.size() - first_task)});
+      }
+    }
+  } catch (...) {
+    if (message && message_size) std::snprintf(message, message_size,
+                                                "adaptive chunk bin allocation failed");
+    return -1;
+  }
+  chunk->bin_seconds = std::chrono::duration<double>(
+      std::chrono::steady_clock::now() - bin_start).count();
+  chunk->use_tile = use_tile && !chunk->host_tasks.empty();
+  ok(message, message_size); return 0;
+}
+
+static int submit_prepared(HipPsfSink *sink, HipPsfPreparedChunk *prepared,
+                           const PsfCachedEvent *events, size_t event_count,
+                           char *message, size_t message_size) {
   if (!sink || (event_count && !events) || event_count > sink->event_capacity) {
     if (message && message_size) std::snprintf(message, message_size, "invalid HIP PSF event chunk");
     return -1;
   }
   if (!event_count) { ok(message, message_size); return 0; }
   if (report(hipSetDevice(sink->device), message, message_size)) return -1;
+  if (sink->adaptive &&
+      (!prepared || prepared->owner != sink || prepared->count != event_count)) {
+    if (message && message_size) std::snprintf(message, message_size,
+                                                "adaptive chunk not prepared for sink/count");
+    return -1;
+  }
+  if (sink->adaptive) {
+    sink->timing.selection_seconds += prepared->selection_seconds;
+    sink->timing.bin_seconds += prepared->bin_seconds;
+  }
+  const bool use_tile = sink->adaptive && prepared->use_tile;
+  if (use_tile) {
+    HipTileScratch &tile = sink->tile;
+    if (collect_tile(sink, message, message_size)) return -1;
+    const int tiles_x = (sink->width + TILE_SIZE - 1) / TILE_SIZE;
+    try {
+      tile.host_refs = prepared->host_refs;
+      tile.host_tasks = prepared->host_tasks;
+      tile.host_merges = prepared->host_merges;
+    } catch (...) {
+      if (message && message_size) std::snprintf(message, message_size,
+                                                  "adaptive tile staging allocation failed");
+      return -1;
+    }
+    {
+      std::memcpy(tile.host, events, event_count * sizeof(PsfCachedEvent));
+      if (report(hipEventRecord(tile.upload_start, sink->stream), message, message_size) ||
+          report(hipMemcpyAsync(tile.events, tile.host,
+                                event_count * sizeof(PsfCachedEvent), hipMemcpyHostToDevice,
+                                sink->stream), message, message_size) ||
+          report(hipMemcpyAsync(tile.refs, tile.host_refs.data(),
+                                tile.host_refs.size() * sizeof(unsigned), hipMemcpyHostToDevice,
+                                sink->stream), message, message_size) ||
+          report(hipMemcpyAsync(tile.tasks, tile.host_tasks.data(),
+                                tile.host_tasks.size() * sizeof(ProductionTileTask), hipMemcpyHostToDevice,
+                                sink->stream), message, message_size) ||
+          (!tile.host_merges.empty() && report(hipMemcpyAsync(
+              tile.merges, tile.host_merges.data(),
+              tile.host_merges.size() * sizeof(ProductionTileMerge), hipMemcpyHostToDevice,
+              sink->stream), message, message_size)) ||
+          report(hipEventRecord(tile.upload_end, sink->stream), message, message_size) ||
+          report(hipEventRecord(tile.start, sink->stream), message, message_size)) return -1;
+      const size_t prepare_threads = event_count * PSF_LANES_PER_EVENT;
+      hipLaunchKernelGGL(prepare_tile_events,
+          dim3((unsigned)((prepare_threads + PSF_THREADS_PER_BLOCK - 1) /
+                          PSF_THREADS_PER_BLOCK)), dim3(PSF_THREADS_PER_BLOCK), 0,
+          sink->stream, tile.events, event_count, sink->phase_resolution,
+          sink->radius_pixels, sink->max_radius_pixels, tile.prepared, tile.bounds);
+      hipLaunchKernelGGL(accumulate_tiles, dim3((unsigned)prepared->host_tasks.size()),
+          dim3(TILE_SIZE * TILE_SIZE), 0, sink->stream, tile.prepared, tile.bounds,
+          tile.refs, tile.tasks, tiles_x, sink->width, sink->height, sink->weights,
+          sink->phase_resolution, sink->radius_pixels, tile.partials, sink->hdr);
+      if (!prepared->host_merges.empty())
+        hipLaunchKernelGGL(merge_tiles, dim3((unsigned)prepared->host_merges.size()),
+            dim3(TILE_SIZE * TILE_SIZE), 0, sink->stream, tile.merges, tile.tasks,
+            tiles_x, sink->width, sink->height, tile.partials, sink->hdr);
+      if (report(hipGetLastError(), message, message_size) ||
+          report(hipEventRecord(tile.end, sink->stream), message, message_size)) return -1;
+      tile.count = event_count;
+      sink->timing.tile_reference_count += prepared->host_refs.size();
+      sink->timing.tile_task_count += prepared->host_tasks.size();
+      sink->timing.tile_merge_count += prepared->host_merges.size();
+      sink->timing.peak_tile_reference_count = std::max(
+          sink->timing.peak_tile_reference_count, prepared->host_refs.size());
+      ++sink->timing.tile_batch_count;
+      ++sink->timing.batch_count;
+      sink->timing.event_count += event_count;
+      const auto now = std::chrono::steady_clock::now();
+      if (std::chrono::duration<double>(now - sink->last_report).count() >= 5.0) {
+        std::fprintf(stderr,
+            "HIP PSF progress: %zu submitted / %zu completed events; %zu / %zu batches timed; kernel %.3f s\n",
+            sink->timing.event_count, sink->completed_events,
+            sink->timing.timed_batch_count, sink->timing.batch_count,
+            sink->timing.kernel_seconds);
+        sink->last_report = now;
+      }
+      ok(message, message_size); return 0;
+    }
+  }
   const unsigned int threads = PSF_THREADS_PER_BLOCK;
   const size_t per_block = PSF_THREADS_PER_BLOCK / PSF_LANES_PER_EVENT;
   const size_t blocks = event_count / per_block + (event_count % per_block != 0);
@@ -214,6 +636,7 @@ extern "C" int hip_psf_sink_submit(HipPsfSink *sink, const PsfCachedEvent *event
   if (report(hipGetLastError(), message, message_size) ||
       report(hipEventRecord(batch.kernel_end, sink->stream), message, message_size)) return -1;
   batch.count = event_count;
+  ++sink->timing.atomic_batch_count;
   ++sink->timing.batch_count;
   sink->timing.event_count += event_count;
   const auto now = std::chrono::steady_clock::now();
@@ -226,6 +649,26 @@ extern "C" int hip_psf_sink_submit(HipPsfSink *sink, const PsfCachedEvent *event
   ok(message, message_size); return 0;
 }
 
+extern "C" int hip_psf_sink_submit_prepared(HipPsfSink *sink,
+    HipPsfPreparedChunk *chunk, const PsfCachedEvent *events,
+    size_t event_count, char *message, size_t message_size) {
+  return submit_prepared(sink, chunk, events, event_count, message, message_size);
+}
+
+extern "C" int hip_psf_sink_submit(HipPsfSink *sink,
+    const PsfCachedEvent *events, size_t event_count,
+    char *message, size_t message_size) {
+  if (!sink) {
+    if (message && message_size) std::snprintf(message, message_size,
+                                                "invalid HIP PSF sink");
+    return -1;
+  }
+  if (sink->adaptive && hip_psf_prepared_chunk_prepare(
+          sink->legacy_prepared, events, event_count, message, message_size)) return -1;
+  return submit_prepared(sink, sink->legacy_prepared, events, event_count,
+                         message, message_size);
+}
+
 extern "C" int hip_psf_sink_load_hdr(HipPsfSink *sink, const double *hdr,
                                       char *message, size_t message_size) {
   if (!sink || !hdr) {
@@ -233,6 +676,7 @@ extern "C" int hip_psf_sink_load_hdr(HipPsfSink *sink, const double *hdr,
     return -1;
   }
   if (report(hipSetDevice(sink->device), message, message_size) ||
+      collect_tile(sink, message, message_size) ||
       report(hipMemcpyAsync(sink->hdr, hdr, sink->hdr_values * sizeof *hdr,
                              hipMemcpyHostToDevice, sink->stream), message, message_size) ||
       report(hipStreamSynchronize(sink->stream), message, message_size))
@@ -248,6 +692,7 @@ extern "C" int hip_psf_sink_finish(HipPsfSink *sink, double *hdr, char *message,
       report(hipEventRecord(sink->download_end, sink->stream), message, message_size) ||
       report(hipStreamSynchronize(sink->stream), message, message_size)) return -1;
   float milliseconds = 0.0f;
+  if (collect_tile(sink, message, message_size)) return -1;
   for (HipPsfBatch &batch : sink->slots)
     if (collect_batch(sink, &batch, message, message_size)) return -1;
   if (report(hipEventElapsedTime(&milliseconds, sink->download_start, sink->download_end),
@@ -264,6 +709,7 @@ extern "C" int hip_psf_sink_get_timing(const HipPsfSink *sink, HipPsfTiming *tim
 
 extern "C" void hip_psf_sink_destroy(HipPsfSink *sink) {
   if (!sink) return;
+  hip_psf_prepared_chunk_destroy(sink->legacy_prepared);
   (void)hipSetDevice(sink->device);
   if (sink->stream) (void)hipStreamSynchronize(sink->stream);
   for (HipPsfBatch &batch : sink->slots) {
@@ -274,6 +720,18 @@ extern "C" void hip_psf_sink_destroy(HipPsfSink *sink) {
     if (batch.host) (void)hipHostFree(batch.host);
     if (batch.device) (void)hipFree(batch.device);
   }
+  if (sink->tile.upload_start) (void)hipEventDestroy(sink->tile.upload_start);
+  if (sink->tile.upload_end) (void)hipEventDestroy(sink->tile.upload_end);
+  if (sink->tile.start) (void)hipEventDestroy(sink->tile.start);
+  if (sink->tile.end) (void)hipEventDestroy(sink->tile.end);
+  if (sink->tile.host) (void)hipHostFree(sink->tile.host);
+  if (sink->tile.events) (void)hipFree(sink->tile.events);
+  if (sink->tile.prepared) (void)hipFree(sink->tile.prepared);
+  if (sink->tile.bounds) (void)hipFree(sink->tile.bounds);
+  if (sink->tile.refs) (void)hipFree(sink->tile.refs);
+  if (sink->tile.tasks) (void)hipFree(sink->tile.tasks);
+  if (sink->tile.merges) (void)hipFree(sink->tile.merges);
+  if (sink->tile.partials) (void)hipFree(sink->tile.partials);
   if (sink->download_start) (void)hipEventDestroy(sink->download_start);
   if (sink->download_end) (void)hipEventDestroy(sink->download_end);
   if (sink->weights) (void)hipFree(sink->weights);