Write through the device address table into guest memory too, so the GPU's writes there are no longer lost.

The table's records gain a twin: a pool with a VRAM mirror names the mirror first and its imported buffer, guest
memory, second; any other record names one buffer twice. Shaders read through the first address and store and run
atomics through both (translator patch 0034); a page without a twin sends the second write to a 4 KiB scratch buffer.
The GPU's writes through the table were not recorded as GPU writes, so they stayed in the mirror: the CPU never saw
them, and a sync of a page the CPU wrote copied the stale guest page over them. Trees drawn from records lost that
way stretched into columns reaching the sky. With NX2NAVI_MEMORY=verify the Hyrule Field walk found 16 such mirror
pages before and none or one after; GPU time per frame 5.92 ms before, 6.17 ms after.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
chrg6
2026-11-07 20:24:41 -05:00
parent 8102b6a222
commit 4cb8188491
4 changed files with 168 additions and 14 deletions

View File

@@ -8,9 +8,17 @@
// One storage buffer of 32-bit words, bound whole to every shader that declares it (res.json "table", cb_slot 0x40):
// words 0-1 the device address of a zeroed 4 KiB buffer, which an address no page maps reads and writes instead
// word 2 the word index of the records (REC_WORD)
// words 3-4 the device address of a 4 KiB scratch buffer, where a twin write goes when its page has no twin
// words 16.. L1: 65536 entries, one per 16 MiB of the 40-bit GPU address space, the index of its L2 block (0: none)
// word 65552 L2 blocks of 4096 entries, one per 4 KiB page: the record of the buffer that holds the page (0: none)
// REC_WORD records of two words: the buffer's device address minus the GPU address of its byte 0 (low word first)
// REC_WORD records of four words: the buffer's device address minus the GPU address of its byte 0, then the same
// for the page's twin (low words first)
// The twin: a pool with a VRAM mirror (vk_memory.c mirror_create) has the mirror in its record's first address and its
// imported buffer, guest memory, in the second; any other record has the same buffer in both. Shaders read through the
// first and write through both (translator patch 0034): the GPU's writes through the table are not recorded as GPU
// writes (vk_exec_draw.c ssbo_info), so with the mirror alone the CPU never saw them, and a sync of a page the CPU wrote
// copied the stale guest page over them. The game's compute passes write tree and foliage records this way, and a
// record overwritten so drew a tree stretched into a column reaching the sky (Hyrule Field, Kakariko, Death Mountain).
// A record a page entry names is never written: a new mapping takes a new record, and a record goes back to the free
// list only once the GPU finished every submission issued while a page could still name it (vk_completion_add). Page
// and L1 entries are single 32-bit stores, so a shader that reads the table while it changes sees the old mapping or
@@ -36,13 +44,14 @@
#define BLOCKS 1024u // L2 blocks: 16 GiB of mapped GPU address space; block 0 stays zero
#define REC_WORD (L2_WORD + BLOCKS * L2_N)
#define RECS 65536u // record 0 means none
#define WORDS (REC_WORD + RECS * 2u)
#define REC_WORDS 4u // words per record
#define WORDS (REC_WORD + RECS * REC_WORDS)
#define PAGE_SHIFT 12
static struct {
int state; // 0 untried, 1 on, -1 off
VkBuffer buf, dummy;
VmaAllocation alloc, dummy_alloc;
VkBuffer buf, dummy, scratch;
VmaAllocation alloc, dummy_alloc, scratch_alloc;
volatile uint32_t *gpu; // the mapped table
uint32_t *cpu; // its shadow
uint32_t nblocks; // L2 blocks handed out (block 0 is the zero block)
@@ -99,8 +108,13 @@ static int create(void) {
return 0;
}
memset(dinfo.pMappedData, 0, 4096);
uint64_t z = vk_gpa_device_address(G.dummy);
put(0, (uint32_t)z), put(1, (uint32_t)(z >> 32)), put(2, REC_WORD);
VmaAllocationCreateInfo si = {.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE};
if (vmaCreateBuffer(c->vma, &di, &si, &G.scratch, &G.scratch_alloc, NULL) != VK_SUCCESS) {
rt_log("vk_gpa: the scratch buffer could not be allocated");
return 0;
}
uint64_t z = vk_gpa_device_address(G.dummy), sc = vk_gpa_device_address(G.scratch);
put(0, (uint32_t)z), put(1, (uint32_t)(z >> 32)), put(2, REC_WORD), put(3, (uint32_t)sc), put(4, (uint32_t)(sc >> 32));
G.nblocks = 1, G.nrec = 1;
rt_log("vk_gpa: device address table on (%.1f MiB, %u blocks of 16 MiB, %u records)", WORDS * 4 / 1048576.0, BLOCKS, RECS);
return 1;
@@ -132,14 +146,16 @@ static uint32_t page_word(uint64_t page, int alloc) {
return L2_WORD + blk * L2_N + (uint32_t)(page & (L2_N - 1));
}
static uint32_t rec_alloc(uint64_t delta) {
static uint32_t rec_alloc(uint64_t delta, uint64_t twin) {
uint32_t r;
if (G.nfree) r = G.free_recs[--G.nfree];
else if (G.nrec < RECS) r = G.nrec++;
else return 0;
// The record before any page names it: the page store below is ordered after these by the release fence.
put(REC_WORD + r * 2, (uint32_t)delta);
put(REC_WORD + r * 2 + 1, (uint32_t)(delta >> 32));
put(REC_WORD + r * REC_WORDS, (uint32_t)delta);
put(REC_WORD + r * REC_WORDS + 1, (uint32_t)(delta >> 32));
put(REC_WORD + r * REC_WORDS + 2, (uint32_t)twin);
put(REC_WORD + r * REC_WORDS + 3, (uint32_t)(twin >> 32));
__atomic_thread_fence(__ATOMIC_RELEASE);
return r;
}
@@ -162,7 +178,7 @@ uint32_t vk_gpa_map(uint64_t lo, uint64_t hi, VkBuffer buf, uint64_t buf_lo) {
uint64_t p0 = (lo + 4095) >> PAGE_SHIFT, p1 = hi >> PAGE_SHIFT;
if (p1 <= p0) return 0;
plat_mutex_lock(&gpa_lock);
uint32_t r = rec_alloc(dev - buf_lo);
uint32_t r = rec_alloc(dev - buf_lo, dev - buf_lo);
if (!r) {
G.refused++;
plat_mutex_unlock(&gpa_lock);
@@ -227,7 +243,11 @@ uint32_t vk_gpa_remap(uint64_t lo, uint64_t hi, uint32_t old_rec, VkBuffer buf,
uint64_t dev = vk_gpa_device_address(buf);
uint64_t p0 = (lo + 4095) >> PAGE_SHIFT, p1 = hi >> PAGE_SHIFT;
plat_mutex_lock(&gpa_lock);
uint32_t r = dev ? rec_alloc(dev - buf_lo) : 0;
// the twin: the buffer the pages named before (the pool's imported buffer), so writes reach guest memory too
uint64_t twin = old_rec ? (uint64_t)G.cpu[REC_WORD + old_rec * REC_WORDS] |
(uint64_t)G.cpu[REC_WORD + old_rec * REC_WORDS + 1] << 32
: dev - buf_lo;
uint32_t r = dev ? rec_alloc(dev - buf_lo, twin) : 0;
if (!r) { plat_mutex_unlock(&gpa_lock); return old_rec; }
for (uint64_t p = p0; p < p1; p++) {
uint32_t w = page_word(p, 1);
@@ -253,6 +273,7 @@ void vk_gpa_shutdown(void) {
if (G.state <= 0 || !c) return;
vmaDestroyBuffer(c->vma, G.buf, G.alloc);
vmaDestroyBuffer(c->vma, G.dummy, G.dummy_alloc);
vmaDestroyBuffer(c->vma, G.scratch, G.scratch_alloc);
free(G.cpu), free(G.free_recs);
memset(&G, 0, sizeof G);
G.state = -1;

View File

@@ -18,7 +18,8 @@ uint32_t vk_gpa_map(uint64_t lo, uint64_t hi, VkBuffer buf, uint64_t buf_lo);
void vk_gpa_unmap(uint64_t lo, uint64_t hi, uint32_t rec, uint32_t (*cover)(uint64_t page_addr));
// Unmaps every page of [lo, hi), whatever record it names (pages of a virtual pool whose physical memory goes away).
void vk_gpa_clear(uint64_t lo, uint64_t hi);
// Moves the pages of [lo, hi) that name old_rec (or none) to buf (a pool's VRAM mirror); returns the new record.
// Moves the pages of [lo, hi) that name old_rec (or none) to buf (a pool's VRAM mirror), with old_rec's buffer as their
// twin, which shaders write too; returns the new record.
uint32_t vk_gpa_remap(uint64_t lo, uint64_t hi, uint32_t old_rec, VkBuffer buf, uint64_t buf_lo);
void vk_gpa_report(void);
void vk_gpa_shutdown(void);

View File

@@ -99,8 +99,8 @@ def prepare_sources() -> Path:
for junk in list(src.glob("src/*/obj")) + list(src.glob("src/*/bin")):
shutil.rmtree(junk)
patches = sorted((XLATE / "patches").glob("*.patch"))
if len(patches) != 32:
sys.exit(f"expected 32 patches, found {len(patches)}")
if len(patches) != 33:
sys.exit(f"expected 33 patches, found {len(patches)}")
for patch in patches:
# git apply works outside a repository when given a directory; --check first so a failure names the patch.
for extra in (["--check"], []):

View File

@@ -0,0 +1,132 @@
--- a/src/Ryujinx.Graphics.Shader/CodeGen/Spirv/Instructions.cs
+++ b/src/Ryujinx.Graphics.Shader/CodeGen/Spirv/Instructions.cs
@@ -1802,7 +1802,13 @@
SpvInstruction one = context.Constant(context.TypeU32(), 1);
SpvInstruction zero = context.Constant(context.TypeU32(), 0);
- return new OperationResult(varType, emitU(context.GetType(varType), elemPointer, one, zero, value));
+ SpvInstruction result = emitU(context.GetType(varType), elemPointer, one, zero, value);
+ if (operation.StorageKind == StorageKind.GlobalMemory) // NX2NAVI: the twin (GetDeviceAddressTwinPointer)
+ {
+ emitU(context.GetType(varType), GlobalTwinPointer(context, operation), one, zero, value);
+ }
+
+ return new OperationResult(varType, result);
}
private static OperationResult GenerateAtomicMemoryCas(CodeGenContext context, AstOperation operation)
@@ -1815,7 +1821,22 @@
SpvInstruction one = context.Constant(context.TypeU32(), 1);
SpvInstruction zero = context.Constant(context.TypeU32(), 0);
- return new OperationResult(varType, context.AtomicCompareExchange(context.GetType(varType), elemPointer, one, zero, zero, value1, value0));
+ SpvInstruction result = context.AtomicCompareExchange(context.GetType(varType), elemPointer, one, zero, zero, value1, value0);
+ if (operation.StorageKind == StorageKind.GlobalMemory) // NX2NAVI: the twin (GetDeviceAddressTwinPointer)
+ {
+ context.AtomicCompareExchange(context.GetType(varType), GlobalTwinPointer(context, operation), one, zero, zero, value1, value0);
+ }
+
+ return new OperationResult(varType, result);
+ }
+
+ private static SpvInstruction GlobalTwinPointer(CodeGenContext context, AstOperation operation)
+ {
+ return GetDeviceAddressTwinPointer(
+ context,
+ ((AstOperand)operation.GetSource(0)).Value,
+ context.Get(AggregateType.U32, operation.GetSource(1)),
+ context.Get(AggregateType.U32, operation.GetSource(2)));
}
private static OperationResult GenerateLoadOrStore(CodeGenContext context, AstOperation operation, bool isStore)
@@ -1856,11 +1877,29 @@
// second-level block, 0 for none (block 0 stays zero)
// word 65552 second-level blocks of 4096 entries, one per 4 KiB page: the record of the buffer that holds
// the page, 0 for none
- // records two words each, the buffer's device address minus its GPU address (low word first); the host
- // never changes a record a page names, so a page entry read at any time gives a whole one
- // The pointer addresses the 32-bit word at the address's 4-byte-aligned word.
+ // words 3-4 the device address of a 4 KiB scratch buffer, where a twin write goes when the page has no twin
+ // records four words each: the device address of the buffer the page's accesses use minus its GPU address,
+ // then the same for the page's twin, the pool's imported buffer in guest memory when the first is
+ // its VRAM mirror (low words first); the host never changes a record a page names, so a page entry
+ // read at any time gives a whole one
+ // The pointer addresses the 32-bit word at the address's 4-byte-aligned word. Stores and atomics also go to the
+ // twin (GetDeviceAddressTwinPointer): what the GPU writes into a mirror through the table is not recorded as a
+ // GPU write, so without it the CPU never saw it and a sync of a CPU-written page copied the stale guest page
+ // over it (runtime/vk_gpa.c).
private static SpvInstruction GetDeviceAddressPointer(CodeGenContext context, int binding, SpvInstruction low, SpvInstruction high)
{
+ return GetDeviceAddressPointers(context, binding, low, high, out _);
+ }
+
+ // The twin of an access's pointer: the guest copy of a mirrored page, else the scratch buffer.
+ private static SpvInstruction GetDeviceAddressTwinPointer(CodeGenContext context, int binding, SpvInstruction low, SpvInstruction high)
+ {
+ GetDeviceAddressPointers(context, binding, low, high, out SpvInstruction twin);
+ return twin;
+ }
+
+ private static SpvInstruction GetDeviceAddressPointers(CodeGenContext context, int binding, SpvInstruction low, SpvInstruction high, out SpvInstruction twinPointer)
+ {
SpvInstruction u32 = context.TypeU32();
SpvInstruction u64 = context.TypeInt(64, 0);
SpvInstruction tableVar = context.StorageBuffers[binding];
@@ -1879,16 +1918,26 @@
SpvInstruction page = context.BitwiseAnd(u32, context.ShiftRightLogical(u32, low, U(12)), U(0xfff));
SpvInstruction record = Word(context.IAdd(u32, context.IAdd(u32, context.IMul(u32, block, U(4096)), U(65552)), page));
SpvInstruction mapped = context.INotEqual(context.TypeBool(), record, U(0));
- SpvInstruction recordWord = context.IAdd(u32, Word(U(2)), context.IMul(u32, record, U(2)));
+ SpvInstruction recordWord = context.IAdd(u32, Word(U(2)), context.IMul(u32, record, U(4)));
SpvInstruction deltaLow = Word(recordWord);
SpvInstruction deltaHigh = Word(context.IAdd(u32, recordWord, U(1)));
+ SpvInstruction twinLow = Word(context.IAdd(u32, recordWord, U(2)));
+ SpvInstruction twinHigh = Word(context.IAdd(u32, recordWord, U(3)));
SpvInstruction wordLow = context.BitwiseAnd(u32, low, U(0xfffffffc));
SpvInstruction device = context.IAdd(u64, Wide(wordLow, high), Wide(deltaLow, deltaHigh));
SpvInstruction dummy = context.IAdd(u64, Wide(Word(U(0)), Word(U(1))), context.UConvert(u64, context.BitwiseAnd(u32, low, U(0xffc))));
SpvInstruction address = context.Select(u64, mapped, device, dummy);
- return context.ConvertUToPtr(context.TypePointer(StorageClass.PhysicalStorageBuffer, u32), address);
+ SpvInstruction boolType = context.TypeBool();
+ SpvInstruction sameTwin = context.LogicalAnd(boolType, context.IEqual(boolType, twinLow, deltaLow), context.IEqual(boolType, twinHigh, deltaHigh));
+ SpvInstruction hasTwin = context.LogicalAnd(boolType, mapped, context.LogicalNot(boolType, sameTwin));
+ SpvInstruction twinDevice = context.IAdd(u64, Wide(wordLow, high), Wide(twinLow, twinHigh));
+ SpvInstruction scratch = context.IAdd(u64, Wide(Word(U(3)), Word(U(4))), context.UConvert(u64, context.BitwiseAnd(u32, low, U(0xffc))));
+ SpvInstruction ptrType = context.TypePointer(StorageClass.PhysicalStorageBuffer, u32);
+ twinPointer = context.ConvertUToPtr(ptrType, context.Select(u64, hasTwin, twinDevice, scratch));
+
+ return context.ConvertUToPtr(ptrType, address);
}
private static SpvInstruction LoadAligned(CodeGenContext context, SpvInstruction type, SpvInstruction pointer)
@@ -1912,7 +1961,7 @@
SpvInstruction u32 = context.TypeU32();
SpvInstruction low = context.Get(AggregateType.U32, operation.GetSource(1));
SpvInstruction high = context.Get(AggregateType.U32, operation.GetSource(2));
- SpvInstruction pointer = GetDeviceAddressPointer(context, binding, low, high);
+ SpvInstruction pointer = GetDeviceAddressPointers(context, binding, low, high, out SpvInstruction twin);
StorageKind kind = operation.StorageKind;
int bits = kind is StorageKind.GlobalMemoryS8 or StorageKind.GlobalMemoryU8 ? 8
: kind is StorageKind.GlobalMemoryS16 or StorageKind.GlobalMemoryU16 ? 16 : 32;
@@ -1921,7 +1970,9 @@
{
if (isStore)
{
- StoreAligned(context, pointer, context.Get(AggregateType.U32, operation.GetSource(operation.SourcesCount - 1)));
+ SpvInstruction stored = context.Get(AggregateType.U32, operation.GetSource(operation.SourcesCount - 1));
+ StoreAligned(context, pointer, stored);
+ StoreAligned(context, twin, stored);
return OperationResult.Invalid;
}
@@ -1939,6 +1990,8 @@
SpvInstruction semantics = context.Constant(u32, 0);
context.AtomicAnd(u32, pointer, scope, semantics, context.Not(u32, mask));
context.AtomicOr(u32, pointer, scope, semantics, context.ShiftLeftLogical(u32, value, shift));
+ context.AtomicAnd(u32, twin, scope, semantics, context.Not(u32, mask));
+ context.AtomicOr(u32, twin, scope, semantics, context.ShiftLeftLogical(u32, value, shift));
return OperationResult.Invalid;
}