mirror of
https://github.com/tangerineturtle1/NX2navi.git
synced 2026-11-20 12:29:57 -04:00
Keep the ASTC decoder's scratch across upload batches and stage its blocks in the staging arena by default.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
@@ -444,7 +444,7 @@ static void mem_barrier(VkCommandBuffer cb, VkPipelineStageFlags2 src, VkAccessF
|
||||
static struct { VkBuffer buf; VmaAllocation alloc; VkDeviceSize size; VkCommandBuffer cb; VkDevice dev; uint64_t big_ns; } S;
|
||||
static uint64_t n_scratch_new, n_scratch_reuse;
|
||||
|
||||
// NX2NAVI_ASTC_SCRATCH_KEEP=1 (default 0 until a play test measured it): the scratch outlives its upload batch. Every
|
||||
// NX2NAVI_ASTC_SCRATCH_KEEP=1 (the default): the scratch outlives its upload batch. Every
|
||||
// batch goes to gfx_queue, so the reuse barrier at the start of the next batch's first chunk (vk_astc_record) orders it
|
||||
// after the earlier batches' copies and dispatches, as it does between two textures of one batch; it then also waits
|
||||
// for their writes (WAW). With
|
||||
@@ -452,6 +452,8 @@ static uint64_t n_scratch_new, n_scratch_reuse;
|
||||
// old one (324 in one Kokiri Forest to Hyrule Field run), and one that found no room in a VMA block allocated a new
|
||||
// block on the replay thread during an area load (8-19 ms per block on NVIDIA, vk_headroom.c). A scratch grown past
|
||||
// SCRATCH_MIN for a large texture goes back to SCRATCH_MIN once no texture needed more for SCRATCH_SHRINK_NS.
|
||||
// On a Kakariko walk at 1440p (2026-11-11) the scratch was made 3 times instead of 60, the walk allocated device memory
|
||||
// 265 times instead of 342, and NX2NAVI_ASTC=check found 0 of 1920 textures differing from the CPU decode.
|
||||
// 0: a scratch per batch.
|
||||
#define SCRATCH_SHRINK_NS 5000000000ull
|
||||
static int scratch_keep(void) {
|
||||
|
||||
Reference in New Issue
Block a user