wyj/GR-raytracing
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Frame: buffer cached PSF events per worker

Browse Source
This commit is contained in:
wyj committed 2026-09-05 03:06:46 -04:00
1 parent 4953775cdd
commit 2a03cbf724
5 files changed
+187 -32
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/frame.c
+90 -13
View File
@@ -10,6 +10,10 @@
#include <stdlib.h>
#include <string.h>
#ifndef FRAME_PSF_EVENT_SINK
#define FRAME_PSF_EVENT_SINK 1
#endif
/* Numerical metric backends may reserve substantial memory for slabs and
* thread-local evaluators, so they retain this private-HDR allocation budget.
* Analytic backends deliberately use all OpenMP render workers instead. */
@@ -115,6 +119,48 @@ int frame_lens_mesh_trace(FrameLensMesh *mesh, const SpacetimeSource *spacetime,
return 0;
}
enum { PSF_EVENT_SINK_CAPACITY = 16384 };
typedef struct {
PsfCachedEvent *events;
size_t count;
double *hdr;
int width, height;
const PsfKernelCache *cache;
} PsfEventSink;
static void psf_event_sink_init(PsfEventSink *sink, double *hdr, int width,
int height, const PsfKernelCache *cache) {
*sink = (PsfEventSink){.hdr = hdr, .width = width, .height = height,
.cache = cache};
if (cache != NULL)
sink->events = malloc(PSF_EVENT_SINK_CAPACITY * sizeof *sink->events);
}
static void psf_event_sink_flush(PsfEventSink *sink) {
for (size_t i = 0; i < sink->count; ++i)
splat_prepared_cached_event(sink->hdr, sink->width, sink->height,
&sink->events[i], sink->cache);
sink->count = 0;
}
static void psf_event_sink_destroy(PsfEventSink *sink) {
psf_event_sink_flush(sink);
free(sink->events);
}
#if FRAME_PSF_EVENT_SINK
static void psf_event_sink_emit(PsfEventSink *sink, const PsfCachedEvent *event) {
if (sink->events == NULL) {
splat_prepared_cached_event(sink->hdr, sink->width, sink->height, event, sink->cache);
return;
}
if (sink->count == PSF_EVENT_SINK_CAPACITY)
psf_event_sink_flush(sink);
sink->events[sink->count++] = *event;
}
#endif
typedef struct {
size_t a, b, triangle;
unsigned int side;
@@ -852,6 +898,7 @@ typedef struct {
double max_cache_psf_flux;
double psf_relative_tail;
double psf_min_y;
PsfEventSink *event_sink;
size_t images;
size_t direct_fallbacks;
size_t cached_wing_clipped;
@@ -920,11 +967,26 @@ static int splat_catalog_tile(const Star *stars, size_t count,
weights[2] * context->vertex[2]->log_frequency_ratio;
const LinearRgb color = blackbody_to_linear_rgb(star->temperature_K * exp(log_g));
const double flux = context->exposure * star->amplitude * context->magnification;
const int direct_fallback = splat_moffat_cached(
context->hdr, context->width, context->height, image_x, image_y, color,
flux,
context->psf, context->psf_cache, context->max_cache_psf_flux,
context->psf_relative_tail, context->psf_min_y);
PsfCachedEvent event;
const int direct_fallback = psf_prepare_cached_event(
&event, image_x, image_y, color, flux, context->psf, context->psf_cache,
context->max_cache_psf_flux, context->psf_relative_tail, context->psf_min_y);
if (direct_fallback == 1) {
/* Preserve the old per-event accumulation order exactly while this is
* the CPU reference sink. A future asynchronous GPU sink must instead
* finish the submitted chunk before this CPU reference contribution. */
psf_event_sink_flush(context->event_sink);
splat_moffat_direct(context->hdr, context->width, context->height, image_x,
image_y, color, flux, context->psf,
context->psf_relative_tail, context->psf_min_y);
} else if (direct_fallback != 3) {
#if FRAME_PSF_EVENT_SINK
psf_event_sink_emit(context->event_sink, &event);
#else
splat_prepared_cached_event(context->hdr, context->width, context->height,
&event, context->psf_cache);
#endif
}
context->direct_fallbacks += direct_fallback == 1;
context->cached_wing_clipped += direct_fallback == 2;
context->discarded_below_min_y += direct_fallback == 3;
@@ -951,8 +1013,14 @@ static CatalogSplatStats splat_catalog_triangles(
const PsfKernelCache *psf_cache, double max_magnification,
double max_cache_psf_flux, double psf_relative_tail,
double psf_min_y,
size_t first_triangle, size_t last_triangle) {
size_t first_triangle, size_t last_triangle, PsfEventSink *event_sink) {
CatalogSplatStats stats = {0};
PsfEventSink owned_sink;
const int owns_sink = event_sink == NULL;
if (owns_sink) {
psf_event_sink_init(&owned_sink, hdr, width, height, psf_cache);
event_sink = &owned_sink;
}
for (size_t t = first_triangle; t < last_triangle; ++t) {
const LensVertex *vertex[3];
if (!usable_triangle(mesh, &mesh->triangles[t], vertex))
@@ -986,7 +1054,8 @@ static CatalogSplatStats splat_catalog_triangles(
.psf = psf, .psf_cache = psf_cache,
.max_cache_psf_flux = max_cache_psf_flux,
.psf_relative_tail = psf_relative_tail,
.psf_min_y = psf_min_y};
.psf_min_y = psf_min_y,
.event_sink = event_sink};
if (catalog_visit_source_triangle(catalog, direction, 0, splat_catalog_tile,
&context) == 0)
stats.images += context.images;
@@ -994,6 +1063,8 @@ 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);
return stats;
}
@@ -1061,7 +1132,7 @@ size_t frame_splat_catalog(const FrameLensMesh *mesh,
const CatalogSplatStats stats = splat_catalog_triangles(
mesh, catalog, hdr, width, height, exposure, psf, psf_cache,
max_magnification, max_cache_psf_flux, psf_relative_tail, psf_min_y,
0, mesh->triangle_count);
0, mesh->triangle_count, NULL);
copy_psf_splat_stats(psf_stats, stats);
return stats.images;
}
@@ -1078,7 +1149,7 @@ size_t frame_splat_catalog(const FrameLensMesh *mesh,
const CatalogSplatStats stats = splat_catalog_triangles(
mesh, catalog, hdr, width, height, exposure, psf, psf_cache,
max_magnification, max_cache_psf_flux, psf_relative_tail, psf_min_y,
0, mesh->triangle_count);
0, mesh->triangle_count, NULL);
copy_psf_splat_stats(psf_stats, stats);
return stats.images;
}
@@ -1089,7 +1160,7 @@ size_t frame_splat_catalog(const FrameLensMesh *mesh,
const CatalogSplatStats stats = splat_catalog_triangles(
mesh, catalog, hdr, width, height, exposure, psf, psf_cache,
max_magnification, max_cache_psf_flux, psf_relative_tail, psf_min_y,
0, mesh->triangle_count);
0, mesh->triangle_count, NULL);
copy_psf_splat_stats(psf_stats, stats);
return stats.images;
}
@@ -1106,7 +1177,7 @@ size_t frame_splat_catalog(const FrameLensMesh *mesh,
const CatalogSplatStats stats = splat_catalog_triangles(
mesh, catalog, hdr, width, height, exposure, psf, psf_cache,
max_magnification, max_cache_psf_flux, psf_relative_tail, psf_min_y,
0, mesh->triangle_count);
0, mesh->triangle_count, NULL);
copy_psf_splat_stats(psf_stats, stats);
return stats.images;
}
@@ -1126,6 +1197,8 @@ size_t frame_splat_catalog(const FrameLensMesh *mesh,
#endif
{
const size_t worker = (size_t)omp_get_thread_num();
PsfEventSink event_sink;
psf_event_sink_init(&event_sink, private_hdr[worker], width, height, psf_cache);
size_t local_triangles = 0, next_report = 8;
progress->worker_callback(progress->context, worker, worker_count, 0, 0);
/* Source density and lens magnification can vary by orders of magnitude
@@ -1138,7 +1211,7 @@ size_t frame_splat_catalog(const FrameLensMesh *mesh,
mesh, catalog, private_hdr[worker], width, height, exposure, psf,
psf_cache, max_magnification, max_cache_psf_flux, psf_relative_tail,
psf_min_y,
triangle, triangle + 1);
triangle, triangle + 1, &event_sink);
images += stats.images;
direct_fallbacks += stats.direct_fallbacks;
cached_wing_clipped += stats.cached_wing_clipped;
@@ -1155,6 +1228,7 @@ size_t frame_splat_catalog(const FrameLensMesh *mesh,
next_report *= 2;
}
}
psf_event_sink_destroy(&event_sink);
progress->worker_callback(progress->context, worker, worker_count,
local_triangles, 1);
}
@@ -1166,13 +1240,15 @@ size_t frame_splat_catalog(const FrameLensMesh *mesh,
#endif
{
const size_t worker = (size_t)omp_get_thread_num();
PsfEventSink event_sink;
psf_event_sink_init(&event_sink, private_hdr[worker], width, height, psf_cache);
#pragma omp for schedule(dynamic, 1)
for (size_t triangle = 0; triangle < mesh->triangle_count; ++triangle) {
const CatalogSplatStats stats = splat_catalog_triangles(
mesh, catalog, private_hdr[worker], width, height, exposure, psf,
psf_cache, max_magnification, max_cache_psf_flux, psf_relative_tail,
psf_min_y,
triangle, triangle + 1);
triangle, triangle + 1, &event_sink);
images += stats.images;
direct_fallbacks += stats.direct_fallbacks;
cached_wing_clipped += stats.cached_wing_clipped;
@@ -1182,6 +1258,7 @@ size_t frame_splat_catalog(const FrameLensMesh *mesh,
magnification_clamped_triangles += stats.magnification_clamped_triangles;
#endif
}
psf_event_sink_destroy(&event_sink);
}
}
#pragma omp parallel for schedule(static)