mirror of
https://github.com/tangerineturtle1/NX2navi.git
synced 2026-11-19 15:09:58 -04:00
Stop the evictions from emptying VRAM, and give VRAM mirrors priority over textures.
A test run with the budget taken as 16 GiB showed two faults. The evictions measured the driver's heap use, which keeps counting a VMA block whole while one allocation lives in it, so freeing textures and mirrors in the middle of blocks did not move it: they went on until 5700 of 7232 textures and 395 of 410 mirrors were gone (9.2 GiB free inside blocks). And mirrors evicted under pressure during a menu were refused again when the world came back, so the draws read the pools over PCIe at 4 fps. The evictions now count the free space inside VMA's blocks as room (vk_mem_vram_live), mirrors are no longer evicted to make room for textures, and a mirror that does not fit after the idle mirrors made room asks idle textures for the rest (vk_texture_want_room) and is asked again 16 submissions later. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
@@ -776,13 +776,23 @@ static void drop_stale(uint64_t keep, uint64_t *count) {
|
||||
// first: those that come back intact and that no lookup reached for IDLE_ROUNDS pool walks and STALE_FRAMES presents
|
||||
// (touch: no live descriptor names them). A later lookup creates the image again and uploads it from its storage, as
|
||||
// for a texture seen the first time. Up to EVICT_STEP entries per check, until the device-local use is 2% of the budget
|
||||
// under the threshold; the next check waits until the GPU is past the submissions that may still name the images, as
|
||||
// their memory goes back to the driver only then (vk_defer_image) and the budget would show the old use.
|
||||
// under the threshold. The use counts the free space inside VMA's blocks as free (vk_mem_vram_live): a freed texture
|
||||
// leaves a hole the driver still counts, and measured that way the evictions went on until nearly every idle texture
|
||||
// was gone. The next check waits until the GPU is past the submissions that may still name the images, as their
|
||||
// memory goes back to VMA only then (vk_defer_image). A VRAM mirror that did not fit asks for room the same way, below
|
||||
// the threshold too (vk_texture_want_room): a pool without its mirror costs far more than a texture uploaded again.
|
||||
#define EVICT_STEP 64
|
||||
static int tex_evict = -1;
|
||||
static uint64_t evict_wait; // vk_sync_fence_done value the next eviction waits for
|
||||
static uint64_t n_evict, n_evict_bytes; // since the last perf line
|
||||
static int evict_logs;
|
||||
static volatile int64_t room_wanted; // bytes refused VRAM mirrors asked for (vk_texture_want_room)
|
||||
|
||||
void vk_texture_want_room(uint64_t bytes) {
|
||||
int64_t cur = __atomic_load_n(&room_wanted, __ATOMIC_RELAXED); // the largest request since the last check
|
||||
while ((int64_t)bytes > cur && !__atomic_compare_exchange_n(&room_wanted, &cur, (int64_t)bytes, 0, __ATOMIC_RELAXED,
|
||||
__ATOMIC_RELAXED)) {}
|
||||
}
|
||||
|
||||
static void evict_idle(uint64_t want, uint64_t used, uint64_t budget) { // tex_lock held
|
||||
if (vk_sync_fence_done() < evict_wait) return;
|
||||
@@ -826,9 +836,13 @@ static void sweep_pressure(void) { // tex_lock held
|
||||
if (submit_epoch & 7) return; // the budget and commit queries every 8 submits
|
||||
uint64_t used, budget;
|
||||
int exact, dropped = 0;
|
||||
if (vram_pressure && (rel_head || tex_evict) && vk_mem_vram(&used, &budget, &exact) &&
|
||||
used > budget / 100 * (uint64_t)vram_pressure) {
|
||||
uint64_t freed = 0, over = used - budget / 100 * (uint64_t)vram_pressure + budget / 50;
|
||||
uint64_t wanted = tex_evict ? (uint64_t)plat_xchg64(&room_wanted, 0) : 0;
|
||||
int known = vram_pressure && (rel_head || tex_evict || wanted) &&
|
||||
(tex_evict ? vk_mem_vram_live(&used, &budget) : vk_mem_vram(&used, &budget, &exact));
|
||||
if (known && (used > budget / 100 * (uint64_t)vram_pressure || wanted)) {
|
||||
uint64_t freed = 0, over = used > budget / 100 * (uint64_t)vram_pressure
|
||||
? used - budget / 100 * (uint64_t)vram_pressure + budget / 50 : 0;
|
||||
if (wanted > over) over = wanted;
|
||||
for (Tex *e = rel_head, *next; e && dropped < PRESSURE_STEP; e = next) {
|
||||
next = e->rel_next;
|
||||
if (!comes_back_intact(e)) continue;
|
||||
|
||||
Reference in New Issue
Block a user