mirror of
https://git.triforceproject.com/triforce/source
synced 2026-11-20 16:49:57 -04:00
Triforce v0.2.0
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
40
gpu/vk/vram_pressure.h
Normal file
40
gpu/vk/vram_pressure.h
Normal file
@@ -0,0 +1,40 @@
|
||||
/* How full VRAM really is, for the rules that give memory back near the driver's limit (vk_memory.cpp's write-back shadows).
|
||||
No Vulkan here. tests/test_vram_pressure.py */
|
||||
#pragma once
|
||||
|
||||
#include <algorithm>
|
||||
#include <cstdint>
|
||||
|
||||
namespace mkwvk::vram_pressure {
|
||||
|
||||
/// What the process really holds: the driver's count less the free space inside VMA's blocks. A destroyed image's or
|
||||
/// buffer's memory stays in its block, and the driver counts the block whole until VMA frees it
|
||||
inline uint64_t allocated(uint64_t used, uint64_t block_free) {
|
||||
return used - std::min(used, block_free);
|
||||
}
|
||||
|
||||
/// Whether the write-back's shadows are given back: what is really allocated past 95% of what the driver allows, or
|
||||
/// the driver's own count past 99%. The count alone passes 95% whenever the texture cache is full, with gigabytes of
|
||||
/// it free space inside the blocks that no eviction gives back.
|
||||
/// (A shadow over 32 MB is an allocation of its own, not a share of a block: VMA's blocks are 64 MB and anything over
|
||||
/// half of one gets its own memory. So the shadows going lowers the count by their whole size, and the lines at which
|
||||
/// they are made again, below, are 2% and 5% under these: with the count's at 98% they came back at once and went again
|
||||
/// at the next 1%, five to seven times in a five-minute run, each time read whole again)
|
||||
inline bool shadows_go(uint64_t used, uint64_t allowed, uint64_t block_free) {
|
||||
if (!allowed) return false;
|
||||
return allocated(used, block_free) > allowed / 100 * 95 || used > allowed / 100 * 99;
|
||||
}
|
||||
|
||||
/// Whether a shadow of n bytes may be made: with it, what is really allocated does not pass 90% nor the driver's own
|
||||
/// count 97%. The textures come first: their budget gives way past 90% really allocated and is squeezed past a 97%
|
||||
/// count, and a shadow adds its whole size to both (its memory is its own: the blocks' free space is of no use to it),
|
||||
/// so one made above these lines could be what starts textures being evicted, which does not lower the count. Not a
|
||||
/// step lower: the count sits at 95-98% for minutes with a full cache, and at 96% and 89% the shadows that a moment at
|
||||
/// 99% took were not back by the end of either of two five-minute runs (96 MB and none of ~400, missing in 27 of 101
|
||||
/// and 9 of 114 windows of 120 presents); at these lines 384 MB in both (missing in 7 of 105 and 10 of 120)
|
||||
inline bool shadow_fits(uint64_t n, uint64_t used, uint64_t allowed, uint64_t block_free) {
|
||||
if (!allowed) return true;
|
||||
return allocated(used, block_free) + n <= allowed / 100 * 90 && used + n <= allowed / 100 * 97;
|
||||
}
|
||||
|
||||
} // namespace mkwvk::vram_pressure
|
||||
Reference in New Issue
Block a user