summary |
shortlog |
log |
commit | commitdiff |
tree
raw |
patch |
inline | side by side (from parent 1:
beef8c4)
By holding on to them only for as long as we need to, we'll save some
peak texture RAM, and _might_ in some situations get a bit less cache
trashing.
set<Phase *> generated_mipmaps;
set<Phase *> generated_mipmaps;
- // We choose the simplest option of having one texture per output,
- // since otherwise this turns into an (albeit simple) register allocation problem.
+ // We keep one texture per output, but only for as long as we actually have any
+ // phases that need it as an input. (We don't make any effort to reorder phases
+ // to minimize the number of textures in play, as register allocation can be
+ // complicated and we rarely have much to gain, since our graphs are typically
+ // pretty linear.)
map<Phase *, GLuint> output_textures;
map<Phase *, GLuint> output_textures;
+ map<Phase *, int> ref_counts;
+ for (Phase *phase : phases) {
+ for (Phase *input : phase->inputs) {
+ ++ref_counts[input];
+ }
+ }
size_t num_phases = phases.size();
if (destinations.empty()) {
size_t num_phases = phases.size();
if (destinations.empty()) {
if (do_phase_timing) {
glEndQuery(GL_TIME_ELAPSED);
}
if (do_phase_timing) {
glEndQuery(GL_TIME_ELAPSED);
}
+
+ // Drop any input textures we don't need anymore.
+ for (Phase *input : phase->inputs) {
+ assert(ref_counts[input] > 0);
+ if (--ref_counts[input] == 0) {
+ resource_pool->release_2d_texture(output_textures[input]);
+ output_textures.erase(input);
+ }
+ }
}
for (const auto &phase_and_texnum : output_textures) {
}
for (const auto &phase_and_texnum : output_textures) {