HIP: record direct atomic PSF timings

README.md

@@ -41,7 +41,10 @@ fallbacks remain CPU reference evaluations, with an ordered HDR transfer before
 and after each fallback. A HIP initialization, upload, kernel, or download
 error terminates the render; it never switches to the CPU backend silently.
 Use `make PSF_BACKEND=hip SPACETIME=minkowski hip-psf-test` for the small
-GPU-vs-CPU cache-HDR regression.
+GPU-vs-CPU cache-HDR regression. HIP renders also report event, batch, H2D,
+kernel, and HDR-download timing after each frame. To bound diagnostic-event
+storage, the report times at most the first 64 batches and labels the timed
+batch count explicitly; the event and total-batch counts are not sampled.
 
 ## Movie PNG sequence (Phase A)
 

benchmarks/hip_psf_direct_atomic_dense_tile_2026-09-05.md

@@ -0,0 +1,139 @@
+# HIP direct-atomic 高密度 PSF endpoint（2026-09-05）
+
+## 目的和边界
+
+测量 production HIP `PsfEventSink` 在一个高表面密度 2MASS tile 上的端到端表现，并将 GPU
+传输、kernel 与 HDR 下载分开记录。此记录验证 direct-global-atomic 的实际成本边界；它不是
+tile-local reduction 的性能声明，也不把 kernel 时间直接归因为 atomic contention。
+
+## 控制条件
+
+- Source baseline: `7a80086f63e8b1e5f27d9524ba0455f2d6e49be5`
+  (`Feat: add initial HIP PSF backend`)；HIP timing counters 是该基线之上的未提交诊断改动，
+  不改变 event、HDR 或输出路径。
+- Host: i7-12700K；`OMP_NUM_THREADS=16`。
+- GPU: AMD Radeon RX 9070 (`gfx1201`), 56 CUs, 15.9 GiB global memory；HIP runtime 是主机
+  实际 probe 所报告的设备。
+- Build: Release；C11/`-march=native -O2 -pipe`、OpenMP、PNG；HIP 后端由
+  `PSF_BACKEND=hip` 选择。
+- Input: `assets/2mass/processed/all_sky/tile_ra252_dec048.csv`，54,160 stars。
+- Camera: Minkowski；RA 252.5 deg、Dec -41.5 deg、FOV 0.6 deg、1920 x 1920 px、
+  `--coarse-cell-pixels 240`（128 triangles / 81 initial rays）。
+- PSF: FWHM 2.7 px、Moffat beta 3、exposure `1e15`、
+  `--max-cache-psf-flux 1e300`。cache radius 266 px；本次没有 min-Y discard 或 direct fallback。
+
+这里刻意以较大的 coarse cell 减少 triangle 枚举工作，使固定 catalog 与相同镜头映射下的
+splat 阶段成为可比较对象；它不是全天空或生产 4K 吞吐量估计。
+
+## 命令与原始输出
+
+CPU reference（同一输入与参数；在 HIP 命令确认退出后顺序运行）：
+
+```console
+export OMP_NUM_THREADS=16
+time build/Release/minkowski_sky \
+  --catalog assets/2mass/processed/all_sky/tile_ra252_dec048.csv \
+  --look-ra-deg 252.5 --look-dec-deg -41.5 --fov-deg 0.6 \
+  --width 1920 --height 1920 --coarse-cell-pixels 240 \
+  --exposure 1e15 --psf-fwhm-pixels 2.7 --psf-moffat-beta 3 \
+  --max-cache-psf-flux 1e300 --output /tmp/cpu_psf_dense.png --verbose
+PSF cache ready: 64x64 phases, radius 266 px, relative tail 1e-08, tail abs 1e-06, boundary 1e-07, build 20.738 s
+Frame 0: tracing 81 initial rays from 128 mesh triangles...
+Frame 0: initial ray trace finished in 0.009 s.
+Frame 0: traced 81 lens vertices; starting catalog render.
+Frame 0: finding and prefetching catalog tiles...
+Frame 0: catalog prefetch finished (0 candidate tiles).
+Frame 0: splatting 128 lens triangles...
+Frame 0: splat worker 14/16 started.
+Frame 0: splat worker 10/16 started.
+Frame 0: splat worker 16/16 started.
+Frame 0: splat worker 1/16 started.
+Frame 0: splat worker 12/16 started.
+Frame 0: splat worker 13/16 started.
+Frame 0: splat worker 5/16 started.
+Frame 0: splat worker 11/16 started.
+Frame 0: splat worker 7/16 started.
+Frame 0: splat worker 6/16 started.
+Frame 0: splat worker 3/16 started.
+Frame 0: splat worker 4/16 started.
+Frame 0: splat worker 2/16 started.
+Frame 0: splat worker 9/16 started.
+Frame 0: splat worker 8/16 started.
+Frame 0: splat worker 15/16 started.
+Frame 0: splat worker 14/16 reached 8 local triangles.
+Frame 0: splat worker 2/16 reached 8 local triangles.
+Frame 0: splat worker 10/16 reached 8 local triangles.
+Frame 0: splat worker 7/16 reached 8 local triangles.
+Frame 0: splat worker 1/16 reached 8 local triangles.
+Frame 0: splat worker 11/16 reached 8 local triangles.
+Frame 0: splat worker 3/16 reached 8 local triangles.
+Frame 0: splat worker 5/16 reached 8 local triangles.
+Frame 0: splat worker 4/16 reached 8 local triangles.
+Frame 0: splat worker 12/16 reached 8 local triangles.
+Frame 0: splat worker 16/16 reached 8 local triangles.
+Frame 0: splat worker 13/16 reached 8 local triangles.
+Frame 0: splat worker 9/16 finished after 4 local triangles.
+Frame 0: splat worker 6/16 finished after 5 local triangles.
+Frame 0: splat worker 15/16 finished after 5 local triangles.
+Frame 0: splat worker 8/16 finished after 5 local triangles.
+Frame 0: splat worker 11/16 finished after 9 local triangles.
+Frame 0: splat worker 1/16 finished after 9 local triangles.
+Frame 0: splat worker 5/16 finished after 9 local triangles.
+Frame 0: splat worker 7/16 finished after 9 local triangles.
+Frame 0: splat worker 10/16 finished after 9 local triangles.
+Frame 0: splat worker 16/16 finished after 8 local triangles.
+Frame 0: splat worker 13/16 finished after 8 local triangles.
+Frame 0: splat worker 4/16 finished after 8 local triangles.
+Frame 0: splat worker 3/16 finished after 9 local triangles.
+Frame 0: splat worker 12/16 finished after 8 local triangles.
+Frame 0: splat worker 2/16 finished after 10 local triangles.
+Frame 0: splat worker 14/16 finished after 13 local triangles.
+Frame 0: catalog splatting finished in 7.1 s; writing image...
+Rendered 26172 images from 54160 catalog stars to /tmp/cpu_psf_dense.png (ok)
+PSF splats: cached 26172, cached wing-clipped 0, direct fallbacks 0, discarded below min-Y 0
+build/Release/minkowski_sky --catalog  --look-ra-deg 252.5 --look-dec-deg      418.79s user 3.19s system 1472% cpu 28.654 total
+```
+
+HIP measurement（加入 timing counters 后；其余参数相同）：
+
+```console
+make -B PSF_BACKEND=hip SPACETIME=minkowski backend
+export OMP_NUM_THREADS=16
+time build/Release/minkowski_sky_hip \
+  --catalog assets/2mass/processed/all_sky/tile_ra252_dec048.csv \
+  --look-ra-deg 252.5 --look-dec-deg -41.5 --fov-deg 0.6 \
+  --width 1920 --height 1920 --coarse-cell-pixels 240 \
+  --exposure 1e15 --psf-fwhm-pixels 2.7 --psf-moffat-beta 3 \
+  --max-cache-psf-flux 1e300 --output /tmp/hip_psf_dense_timing.png --verbose
+PSF cache ready: 64x64 phases, radius 266 px, relative tail 1e-08, tail abs 1e-06, boundary 1e-07, build 20.915 s
+Frame 0: tracing 81 initial rays from 128 mesh triangles...
+Frame 0: initial ray trace finished in 0.008 s.
+Frame 0: traced 81 lens vertices; starting catalog render.
+Frame 0: finding and prefetching catalog tiles...
+Frame 0: catalog prefetch finished (0 candidate tiles).
+Frame 0: splatting 128 lens triangles...
+Frame 0: catalog splatting finished in 3.2 s; writing image...
+Rendered 26172 images from 54160 catalog stars to /tmp/hip_psf_dense_timing.png (ok)
+PSF splats: cached 26172, cached wing-clipped 0, direct fallbacks 0, discarded below min-Y 0
+HIP PSF: 26172 events in 2 batches (2 timed); upload 0.000073 s, kernel 2.950377 s, download 0.011256 s
+build/Release/minkowski_sky_hip --catalog  --look-ra-deg 252.5 --look-dec-deg  326.85s user 1.53s system 1314% cpu 24.985 total
+```
+
+两次命令都没有设置 `TIMEFMT`；末行是 zsh 默认 `%J  %U user %S system %P cpu %*E total`
+格式的原始输出。它们按顺序运行：先确认没有 renderer benchmark，再执行 HIP，确认其退出后才
+执行 CPU，期间没有第二个 renderer benchmark。
+
+`cmp` 比较 `/tmp/cpu_psf_dense.png` 与 `/tmp/hip_psf_dense_timing.png`：`PNG byte comparison:
+identical`。
+
+## 结果和决策
+
+CPU catalog splatting 为 7.1 s，HIP 为 3.2 s（约 2.22x）；完整进程的 zsh elapsed time 为
+28.654 s（CPU）与 24.985 s（HIP）。在 HIP 的 3.2 s 中，两个 batch 的 kernel event 合计
+2.950377 s；上传为 0.000073 s、最终 HDR 下载为 0.011256 s。因此这个 endpoint 的 GPU 端
+时间由 kernel 主导，而不是 PCIe。
+
+但 kernel 还包括每个 event 的圆形 PSF 支持域遍历、cache 权重读取和 HDR atomic 写入。这里只有
+一个密度／支持域组合，也没有关闭或替代 atomic 的对照，所以不能以此声称 global atomics 的热点
+竞争是主导瓶颈。结论是保留 direct atomic，并先测密度与支持域扫描；暂不引入 tile-local reduction
+或其分桶／重排复杂度。

gpu_psf_acceleration_plan.md

@@ -174,10 +174,16 @@ HIP direct-atomic 的当前实现状态：
 3. RX 9070 上的 640×360 固定 Minkowski catalog HDR 对照逐样本一致；含 2 个 cache 事件与
    1 个 direct fallback 的 `tests/data/hip_psf_mixed_catalog.csv` 场同样逐样本一致。GPU failure
    会打印阶段与 HIP 错误并令该帧返回失败，不会切换到 CPU backend。
-4. 待记录端到端基准的完整命令和原始输出：Git hash、CPU threads、GPU/runtime、块大小、上传、
+4. 已在 RX 9070 上记录一个受控的高密度 endpoint：54,160 星的 1920×1920、0.6 deg
-   kernel、download、分辨率、catalog、PSF 及所有 fallback 统计。
+   2MASS tile 产生 26,172 个 cache event、零 direct fallback。HIP 的 2 个 batch 中 H2D
-5. 只在 direct atomic profile 显示热点竞争是主要瓶颈时，比较 tile-local reduction；ATRI 的
+   为 0.073 ms、kernel 为 2.950 s、D2H 为 11.256 ms；CPU/HIP PNG 字节一致。完整命令、
-   整帧收集／重排同样必须以端到端收益证明。
+   原始输出、Git 基线、线程、设备、分辨率和 PSF 参数见
+   `benchmarks/hip_psf_direct_atomic_dense_tile_2026-09-05.md`。
+5. 该 endpoint 说明当前时间主要位于 device kernel，而非 PCIe；它**不能单独证明** kernel
+   时间主要来自热点 atomic contention（仍混有每 event 的 PSF 遍历与 cache 读取）。因此暂不
+   实现 tile-local reduction。下一步应在资源隔离条件下，以不同屏幕密度和 PSF 支持域分离竞争
+   效应；只有该证据显示竞争为主导且 tile-local 的整帧端到端测量获益时，才保留分桶／重排实现。
+   ATRI 的整帧收集／重排同样必须以端到端收益证明。
 6. HIP 路径稳定后再评估 Vulkan 的可移植后端及其他机器的实际设备矩阵。
 
 每一步独立提交，并保持 CPU reference、GPU runtime 层和 shader 资源的提交边界清晰。

src/frame.c

@@ -134,6 +134,7 @@ typedef struct {
 #ifdef PSF_BACKEND_HIP
   HipPsfSink *hip;
   char hip_message[256];
+  HipPsfTiming hip_timing;
 #endif
   int failed;
 } PsfEventSink;
@@ -227,6 +228,10 @@ static int psf_event_sink_resume_gpu(PsfEventSink *sink) {
 static int psf_event_sink_destroy(PsfEventSink *sink) {
   int result = psf_event_sink_finish_for_cpu(sink);
 #ifdef PSF_BACKEND_HIP
+  if (sink->hip != NULL && hip_psf_sink_get_timing(sink->hip, &sink->hip_timing)) {
+    psf_event_sink_mark_failed(sink, "timing collection");
+    result = -1;
+  }
   hip_psf_sink_destroy(sink->hip);
 #endif
   free(sink->events);
@@ -998,6 +1003,8 @@ typedef struct {
   size_t direct_fallbacks;
   size_t cached_wing_clipped;
   size_t discarded_below_min_y;
+  size_t gpu_event_count, gpu_batch_count, gpu_timed_batch_count;
+  double gpu_upload_seconds, gpu_kernel_seconds, gpu_download_seconds;
   int failed;
 #ifdef GR_DEBUG
   double max_raw_magnification;
@@ -1016,6 +1023,12 @@ static void copy_psf_splat_stats(PsfSplatStats *destination,
                                  .cached_wing_clipped = source.cached_wing_clipped,
                                  .direct_fallbacks = source.direct_fallbacks,
                                  .discarded_below_min_y = source.discarded_below_min_y,
+                                 .gpu_event_count = source.gpu_event_count,
+                                 .gpu_batch_count = source.gpu_batch_count,
+                                 .gpu_timed_batch_count = source.gpu_timed_batch_count,
+                                 .gpu_upload_seconds = source.gpu_upload_seconds,
+                                 .gpu_kernel_seconds = source.gpu_kernel_seconds,
+                                 .gpu_download_seconds = source.gpu_download_seconds,
 #ifdef GR_DEBUG
                                  .max_raw_magnification = source.max_raw_magnification,
                                  .magnification_clamped_triangles =
@@ -1162,8 +1175,18 @@ static CatalogSplatStats splat_catalog_triangles(
     stats.cached_wing_clipped += context.cached_wing_clipped;
     stats.discarded_below_min_y += context.discarded_below_min_y;
   }
-  if (owns_sink && psf_event_sink_destroy(&owned_sink))
+  if (owns_sink) {
-    stats.failed = 1;
+    if (psf_event_sink_destroy(&owned_sink))
+      stats.failed = 1;
+#ifdef PSF_BACKEND_HIP
+    stats.gpu_event_count = owned_sink.hip_timing.event_count;
+    stats.gpu_batch_count = owned_sink.hip_timing.batch_count;
+    stats.gpu_timed_batch_count = owned_sink.hip_timing.timed_batch_count;
+    stats.gpu_upload_seconds = owned_sink.hip_timing.upload_seconds;
+    stats.gpu_kernel_seconds = owned_sink.hip_timing.kernel_seconds;
+    stats.gpu_download_seconds = owned_sink.hip_timing.download_seconds;
+#endif
+  }
   return stats;
 }
 

src/hip_psf.h

@@ -13,6 +13,11 @@ extern "C" {
  * for PsfCachedEvent chunks, immutable cache weights, and double RGB HDR. */
 typedef struct HipPsfSink HipPsfSink;
 
+typedef struct {
+  size_t event_count, batch_count, timed_batch_count;
+  double upload_seconds, kernel_seconds, download_seconds;
+} HipPsfTiming;
+
 int hip_psf_available(char *message, size_t message_size);
 
 /* Creates a zeroed GPU HDR framebuffer. event_capacity is the reusable upload
@@ -35,6 +40,9 @@ int hip_psf_sink_submit(HipPsfSink *sink, const PsfCachedEvent *events,
 /* 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);
+/* Valid after hip_psf_sink_finish(); timings are accumulated without a
+ * per-batch host synchronization. */
+int hip_psf_sink_get_timing(const HipPsfSink *sink, HipPsfTiming *timing);
 void hip_psf_sink_destroy(HipPsfSink *sink);
 
 #ifdef __cplusplus

src/hip_psf.hip

@@ -5,6 +5,14 @@
 #include <cmath>
 #include <cstdio>
 #include <limits>
+#include <vector>
+
+struct HipPsfBatchTiming {
+  hipEvent_t upload_start = nullptr, upload_end = nullptr;
+  hipEvent_t kernel_start = nullptr, kernel_end = nullptr;
+};
+
+enum { HIP_PSF_MAX_TIMED_BATCHES = 64 };
 
 struct HipPsfSink {
   PsfCachedEvent *events = nullptr;
@@ -14,6 +22,10 @@ struct HipPsfSink {
   size_t event_capacity = 0, hdr_values = 0;
   int width = 0, height = 0, phase_resolution = 0, radius_pixels = 0;
   double max_radius_pixels = 0.0;
+  hipEvent_t download_start = nullptr, download_end = nullptr;
+  std::vector<HipPsfBatchTiming> batches;
+  size_t measured_batches = 0;
+  HipPsfTiming timing = {};
 };
 
 static int report(hipError_t status, char *message, size_t message_size) {
@@ -25,6 +37,15 @@ static void ok(char *message, size_t message_size) {
   if (message && message_size) std::snprintf(message, message_size, "ok");
 }
 
+static void destroy_batch_timing(HipPsfBatchTiming *timing) {
+  if (!timing) return;
+  (void)hipEventDestroy(timing->upload_start);
+  (void)hipEventDestroy(timing->upload_end);
+  (void)hipEventDestroy(timing->kernel_start);
+  (void)hipEventDestroy(timing->kernel_end);
+  *timing = {};
+}
+
 __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;
@@ -118,6 +139,8 @@ extern "C" int hip_psf_sink_create(HipPsfSink **out, int width, int height,
   sink->radius_pixels = cache->radius_pixels; sink->max_radius_pixels = cache->max_radius_pixels;
   const size_t weight_count = nodes * nodes * side * side;
   if (report(hipStreamCreate(&sink->stream), message, message_size) ||
+      report(hipEventCreate(&sink->download_start), message, message_size) ||
+      report(hipEventCreate(&sink->download_end), message, message_size) ||
       report(hipMalloc(&sink->events, event_capacity * sizeof *sink->events), message, message_size) ||
       report(hipMalloc(&sink->weights, weight_count * sizeof *sink->weights), message, message_size) ||
       report(hipMalloc(&sink->hdr, sink->hdr_values * sizeof *sink->hdr), message, message_size) ||
@@ -137,13 +160,35 @@ extern "C" int hip_psf_sink_submit(HipPsfSink *sink, const PsfCachedEvent *event
   if (!event_count) { ok(message, message_size); return 0; }
   const unsigned int threads = 128;
   const size_t blocks = (event_count + threads - 1) / threads;
+  HipPsfBatchTiming timing;
+  const int timed = sink->batches.size() < HIP_PSF_MAX_TIMED_BATCHES;
   if (blocks > std::numeric_limits<unsigned int>::max() ||
-      report(hipMemcpyAsync(sink->events, events, event_count * sizeof *events, hipMemcpyHostToDevice, sink->stream), message, message_size))
+      (timed && (report(hipEventCreate(&timing.upload_start), message, message_size) ||
+                 report(hipEventCreate(&timing.upload_end), message, message_size) ||
+                 report(hipEventCreate(&timing.kernel_start), message, message_size) ||
+                 report(hipEventCreate(&timing.kernel_end), message, message_size) ||
+                 report(hipEventRecord(timing.upload_start, sink->stream), message, message_size))) ||
+      report(hipMemcpyAsync(sink->events, events, event_count * sizeof *events,
+                            hipMemcpyHostToDevice, sink->stream), message, message_size) ||
+      (timed && (report(hipEventRecord(timing.upload_end, sink->stream), message, message_size) ||
+                 report(hipEventRecord(timing.kernel_start, sink->stream), message, message_size)))) {
+    destroy_batch_timing(&timing);
     return -1;
+  }
   hipLaunchKernelGGL(splat_kernel, dim3((unsigned int)blocks), dim3(threads), 0, sink->stream,
                      sink->events, event_count, sink->width, sink->height, sink->weights,
                      sink->phase_resolution, sink->radius_pixels, sink->max_radius_pixels, sink->hdr);
-  if (report(hipGetLastError(), message, message_size)) return -1;
+  if (report(hipGetLastError(), message, message_size) ||
+      (timed && report(hipEventRecord(timing.kernel_end, sink->stream), message, message_size))) {
+    destroy_batch_timing(&timing);
+    return -1;
+  }
+  if (timed) {
+    sink->batches.push_back(timing);
+    ++sink->timing.timed_batch_count;
+  }
+  ++sink->timing.batch_count;
+  sink->timing.event_count += event_count;
   ok(message, message_size); return 0;
 }
 
@@ -161,13 +206,39 @@ extern "C" int hip_psf_sink_load_hdr(HipPsfSink *sink, const double *hdr,
 
 extern "C" int hip_psf_sink_finish(HipPsfSink *sink, double *hdr, char *message, size_t message_size) {
   if (!sink || !hdr) { if (message && message_size) std::snprintf(message, message_size, "invalid HIP PSF HDR download"); return -1; }
-  if (report(hipMemcpyAsync(hdr, sink->hdr, sink->hdr_values * sizeof *hdr, hipMemcpyDeviceToHost, sink->stream), message, message_size) ||
+  if (report(hipEventRecord(sink->download_start, sink->stream), message, message_size) ||
+      report(hipMemcpyAsync(hdr, sink->hdr, sink->hdr_values * sizeof *hdr, hipMemcpyDeviceToHost, sink->stream), message, message_size) ||
+      report(hipEventRecord(sink->download_end, sink->stream), message, message_size) ||
       report(hipStreamSynchronize(sink->stream), message, message_size)) return -1;
+  float milliseconds = 0.0f;
+  for (size_t i = sink->measured_batches; i < sink->batches.size(); ++i) {
+    const HipPsfBatchTiming &batch = sink->batches[i];
+    if (report(hipEventElapsedTime(&milliseconds, batch.upload_start, batch.upload_end),
+               message, message_size)) return -1;
+    sink->timing.upload_seconds += milliseconds * 1e-3;
+    if (report(hipEventElapsedTime(&milliseconds, batch.kernel_start, batch.kernel_end),
+               message, message_size)) return -1;
+    sink->timing.kernel_seconds += milliseconds * 1e-3;
+  }
+  sink->measured_batches = sink->batches.size();
+  if (report(hipEventElapsedTime(&milliseconds, sink->download_start, sink->download_end),
+             message, message_size)) return -1;
+  sink->timing.download_seconds += milliseconds * 1e-3;
   ok(message, message_size); return 0;
 }
 
+extern "C" int hip_psf_sink_get_timing(const HipPsfSink *sink, HipPsfTiming *timing) {
+  if (!sink || !timing) return -1;
+  *timing = sink->timing;
+  return 0;
+}
+
 extern "C" void hip_psf_sink_destroy(HipPsfSink *sink) {
   if (!sink) return;
+  for (HipPsfBatchTiming &timing : sink->batches)
+    destroy_batch_timing(&timing);
+  (void)hipEventDestroy(sink->download_start);
+  (void)hipEventDestroy(sink->download_end);
   (void)hipFree(sink->events); (void)hipFree(sink->weights); (void)hipFree(sink->hdr);
   (void)hipStreamDestroy(sink->stream); delete sink;
 }

src/optics.c

@@ -301,6 +301,13 @@ void psf_kernel_cache_report(const PsfKernelCache *cache,
           stats == NULL ? 0u : stats->cached_wing_clipped,
           stats == NULL ? 0u : stats->direct_fallbacks,
           stats == NULL ? 0u : stats->discarded_below_min_y);
+  if (stats != NULL && stats->gpu_batch_count != 0)
+    fprintf(stream, "HIP PSF: %zu events in %zu batches (%zu timed); upload %.6f s, "
+                    "kernel %.6f s, download %.6f s\n",
+            stats->gpu_event_count, stats->gpu_batch_count,
+            stats->gpu_timed_batch_count,
+            stats->gpu_upload_seconds, stats->gpu_kernel_seconds,
+            stats->gpu_download_seconds);
   if (stats != NULL && stats->discarded_below_min_y != 0)
     fputs("Warning: --psf-min-y discarded one or more PSF events.\n", stream);
 #ifdef GR_DEBUG

src/optics.h

@@ -26,6 +26,8 @@ typedef struct {
   size_t cached_wing_clipped;
   size_t direct_fallbacks;
   size_t discarded_below_min_y;
+  size_t gpu_event_count, gpu_batch_count, gpu_timed_batch_count;
+  double gpu_upload_seconds, gpu_kernel_seconds, gpu_download_seconds;
 #ifdef GR_DEBUG
   double max_raw_magnification;
   size_t magnification_clamped_triangles;

tests/test_hip_psf.hip

@@ -46,6 +46,16 @@ int main() {
     std::free(cpu_hdr); std::free(gpu_hdr); psf_kernel_cache_destroy(&cache);
     return 1;
   }
+  HipPsfTiming timing = {};
+  if (hip_psf_sink_get_timing(sink, &timing) || timing.event_count != 3 ||
+      timing.batch_count != 1 || timing.timed_batch_count != 1 ||
+      timing.upload_seconds < 0.0 || timing.kernel_seconds < 0.0 ||
+      timing.download_seconds < 0.0) {
+    std::fputs("HIP PSF timing accounting failed\n", stderr);
+    hip_psf_sink_destroy(sink);
+    std::free(cpu_hdr); std::free(gpu_hdr); psf_kernel_cache_destroy(&cache);
+    return 1;
+  }
   double max_abs = 0.0, max_rel = 0.0;
   for (size_t i = 0; i < values; ++i) {
     const double absolute = std::fabs(cpu_hdr[i] - gpu_hdr[i]);