HIP: record direct atomic PSF timings

This commit is contained in:
wyj committed 2026-09-05 22:13:14 -04:00
1 parent 7a80086f63
commit a18ebfbf1a
9 files changed
+279 -10

No files matched your search

+25 -2
View File
@@ -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))
stats.failed = 1;
if (owns_sink) {
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;
}