Bring the 0.10.0 line into the VRAM pass, keeping the line's own releases where the two overlap.

The line freed the upscaler stand-in's scratch images, FSR 1's and SMAA's images and the DLSS sizes on its own while
this pass did: its versions stay. Of this pass's versions, what they do not cover stays: the stand-in's views of a game
image it read as it is now go with that image (vk_tonemap_forget from vk_exec_image_gone, under a lock), and the
upscaler release frees only what the line's idle release leaves, the reactive mask's snapshots, the FSR upscale
contexts and the DLSS sizes after a failure, at the same 600 presents and with vram_use = high keeping them as well.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
chrg6
2026-11-13 00:50:21 -05:00
36 changed files with 2196 additions and 280 deletions

View File

@@ -32,6 +32,7 @@
#include "vk_gpa.h"
#include "mesh_fb.h"
#include "vk_mesh_fb.h"
#include "vk_scratch_lru.h"
#include "platform/platform.h"
#include "vk_dlaa.h"
#include "mesh_emu_spirv.h"
@@ -920,8 +921,8 @@ VkImageView vk_exec_target_view(const VkTexImage *img, VkFormat format, VkImageV
}
int vk_exec_image_gone(VkImage image) {
vk_tonemap_image_gone(image); // its views in the tonemap pass, keyed by the handle a new image may get
vk_exec_bc_gone(image);
vk_tonemap_forget(image); // the stand-in's views of a game image it read as it is
vk_hdr_image_gone(image);
if (!views_by_image.klen) return 0;
int purged = 0;
@@ -6828,37 +6829,23 @@ static int fmt_rgba8(VkFormat f) {
f == VK_FORMAT_B8G8R8A8_SRGB || f == VK_FORMAT_A8B8G8R8_UNORM_PACK32 || f == VK_FORMAT_A8B8G8R8_SRGB_PACK32;
}
// Scratch images for those blits, one per (size, format). Up to FB_SCRATCH are kept; a new size past them takes the
// place of the one used longest ago, which goes once the GPU is past the submission being recorded (vk_defer_image),
// with the tonemap pass's view of it. Before, the first eight sizes were kept for the process and a ninth got no
// scratch, so its blit was left out: the sizes come from the scene's resolution, which DLSS modes, the dynamic
// resolution and a window or picture size change move, at up to 32 MiB each at 2160 lines. One used in the
// submission being recorded is never the one replaced: the blit may have recorded it already.
#define FB_SCRATCH 8
// Scratch images for those blits, one per (size, format), at most SCRATCH_N; a new size takes the place of the one used
// longest ago (vk_scratch_lru.h), whose image goes through the deferred destruction.
#define SCRATCH_N 8
static VkImage fallback_scratch(VkExec *ex, uint32_t w, uint32_t h, VkFormat f) {
static struct { uint32_t w, h; VkFormat f; VkImage img; VmaAllocation a; uint64_t used; } cache[FB_SCRATCH];
static int n;
static PlatMutex lock = PLAT_MUTEX_INIT; // replay workers record blits side by side
plat_mutex_lock(&lock);
for (int i = 0; i < n; i++)
if (cache[i].w == w && cache[i].h == h && cache[i].f == f) {
cache[i].used = ex->fence;
VkImage img = cache[i].img;
plat_mutex_unlock(&lock);
return img;
}
int slot = n;
if (n == FB_SCRATCH) { // the one used longest ago, unless the submission being recorded uses it
slot = -1;
for (int i = 0; i < n; i++)
if (cache[i].used < ex->fence && (slot < 0 || cache[i].used < cache[slot].used)) slot = i;
if (slot < 0) {
plat_mutex_unlock(&lock);
return VK_NULL_HANDLE;
}
vk_tonemap_image_gone(cache[slot].img);
vk_defer_image(cache[slot].img, cache[slot].a);
cache[slot].img = VK_NULL_HANDLE;
static ScrEntry ent[SCRATCH_N];
static struct { VkImage img; VmaAllocation a; } cache[SCRATCH_N];
static int evictions;
int hit, evict, i = scr_pick(ent, SCRATCH_N, w, h, (uint32_t)f,
__atomic_load_n(&gfx_counters.frames, __ATOMIC_RELAXED), &hit, &evict);
if (hit) return cache[i].img;
if (evict) {
if (evictions++ < 8)
rt_log("vk: a stand-in scratch image of %ux%u (format %d) makes room for one of %ux%u (format %d)",
ent[i].w, ent[i].h, (int)ent[i].f, w, h, (int)f);
vk_tonemap_forget(cache[i].img); // the stand-in's views of it, before a new image can take its handle
vk_defer_image(cache[i].img, cache[i].a);
ent[i].live = 0, cache[i].img = VK_NULL_HANDLE, cache[i].a = NULL;
}
VkCtx *c = vk_ctx();
VkImageCreateInfo ii = {.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO, .imageType = VK_IMAGE_TYPE_2D, .format = f,
@@ -6869,11 +6856,7 @@ static VkImage fallback_scratch(VkExec *ex, uint32_t w, uint32_t h, VkFormat f)
.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED};
VmaAllocationCreateInfo ia = {.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE};
VkImage img; VmaAllocation a;
if (vmaCreateImage(c->vma, &ii, &ia, &img, &a, NULL) != VK_SUCCESS) {
if (slot < n) cache[slot] = cache[--n]; // the replaced entry's place stays empty
plat_mutex_unlock(&lock);
return VK_NULL_HANDLE;
}
if (vmaCreateImage(c->vma, &ii, &ia, &img, &a, NULL) != VK_SUCCESS) return VK_NULL_HANDLE;
VkImageMemoryBarrier2 bar = {.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER_2,
.dstStageMask = VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
.dstAccessMask = VK_ACCESS_2_MEMORY_READ_BIT | VK_ACCESS_2_MEMORY_WRITE_BIT,
@@ -6882,10 +6865,8 @@ static VkImage fallback_scratch(VkExec *ex, uint32_t w, uint32_t h, VkFormat f)
.image = img, .subresourceRange = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 1}};
VkDependencyInfo di = {.sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO, .imageMemoryBarrierCount = 1, .pImageMemoryBarriers = &bar};
vkCmdPipelineBarrier2(ex->cb, &di);
cache[slot].w = w, cache[slot].h = h, cache[slot].f = f, cache[slot].img = img, cache[slot].a = a;
cache[slot].used = ex->fence;
if (slot == n) n++;
plat_mutex_unlock(&lock);
ent[i] = (ScrEntry){w, h, (uint32_t)f, __atomic_load_n(&gfx_counters.frames, __ATOMIC_RELAXED), 1};
cache[i].img = img, cache[i].a = a;
return img;
}
@@ -7391,7 +7372,14 @@ static void op_update_ub(VkExec *ex, const p_update_ub *u) {
VK_MEM_PICK(sp);
VkBufferCopy cp = {o, sp.offset, u->size};
vkCmdCopyBuffer(ex->cb, b, sp.buffer, 1, &cp);
vk_mem_gpu_wrote(&sp, u->size);
// In a submission's tail (vk_exec.c "the early tail") an update of a mirrored block reaches guest memory with the
// results the game reads, from the bytes here, and takes no write-back: one after the tail would land when the
// game is already writing the next frame's data. Not with the lazy write-back, where the block's pages may be
// armed and a store into them would wait for the GPU (memory.c lazy_fault): the update is noted as ever there,
// and the tail goes after the write-backs as before.
if (vk_exec_tail() && !vk_mem_wb_lazy() && vk_mem_mirrored(u->dst, u->size))
vk_exec_tail_ub(ex->fence, u->dst, u->data, u->size);
else vk_mem_gpu_wrote(&sp, u->size);
vk_usnap_drop(u->dst, u->size);
ex->recorded++;
vk_mem_count(sp.cls, VK_MUSE_GPU_WRITE, u->size);