Send atomics through the device address table to guest memory too, so the trees no longer flicker or stretch into columns.

The culling passes count each frame's visible foliage with atomics through the device address table into small records (an indirect draw's count, then 1, 1). Stores through the table already reached both the VRAM mirror and its guest copy, but atomics reached the mirror only, so guest memory held the CPU's template of zeros. The CPU writes other words of the same pages every frame, and the sync of such a page copied those zeros over the counts: in Hyrule Field the trees flickered, went missing or stretched into columns reaching the sky. NX2NAVI_MEMORY=verify found 5 to 10 such pages a minute, all in ten pools of 20 and 44 KiB.

Translator patch 0036 repeats every table atomic on the twin (a compare-exchange with the same operands), leaving its result unused so the driver need not wait on guest memory. A buffer fill into a mirror now fills its guest copy at once (vk_mem_guest_twin), so a counter the GPU resets that way counts on from the same value in both; that mismatch is why table atomics stayed out of the twin before. The memory check's mismatch lines now also give the pool's last written binding, its last use, the CPU's write stamp for the page and the first bytes of both copies.

With the change the trees stayed steady at the spot where they flickered, all 37,848 shaders translated without a failure, and the GPU time per frame stayed at about 9 ms (75 fps uncapped).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
chrg6
2026-11-08 08:59:26 -05:00
parent a8f34e48c8
commit 82257eec3d
5 changed files with 117 additions and 3 deletions

View File

@@ -253,6 +253,7 @@ void vk_mem_dirty_submit(void); // vk_sync_submit, under its l
extern int vk_mirror_on;
void vk_mem_use(VkMemSpan *sp);
void vk_mem_pick(VkMemSpan *sp);
VkBuffer vk_mem_guest_twin(const VkMemSpan *sp); // a VRAM mirror span: the pool's imported buffer (guest memory), same offset; else none
#define VK_MEM_USE(sp) do { if (vk_mirror_on && (sp).pool) vk_mem_use(&(sp)); } while (0)
#define VK_MEM_PICK(sp) do { if (vk_mirror_on && (sp).pool) vk_mem_pick(&(sp)); } while (0)
void vk_mem_gpu_wrote(const VkMemSpan *sp, uint64_t size);

View File

@@ -700,6 +700,10 @@ static void clear_buf(VkExec *ex, const p_clear_buf *p) {
if (!started) vk_exec_pass_end(ex), vk_exec_barrier(ex), started = 1;
VK_MEM_PICK(d);
vkCmdFillBuffer(ex->cb, d.buffer, d.offset, n, p->value);
// A VRAM mirror's guest copy is filled too, at once: atomics through the device address table count in both
// copies (translator patch 0036), and a counter reset here must start from the same value in both.
VkBuffer twin = vk_mem_guest_twin(&d);
if (twin) vkCmdFillBuffer(ex->cb, twin, d.offset, n, p->value), ex->recorded++;
vk_mem_gpu_wrote(&d, n);
ex->recorded++;
vk_mem_count(d.cls, VK_MUSE_GPU_WRITE, n);

View File

@@ -63,6 +63,8 @@ typedef struct PoolRec {
VmaAllocation malloc_;
uint32_t acc_sub; // replay: the last submission that synced the mirror (mirror_sync)
uint32_t wb_sub; // replay: the last submission that wrote mirror bytes back to guest memory
uint32_t vd_wrote_sub, vd_use_sub; // verify mode: the last submission that bound it written / used its mirror
uint64_t vd_wrote_n; // verify mode: written bindings of it so far
uint32_t mgen; // gen_lock: every page write stamped up to this generation is in the mirror
// (or in a copy queued before its next use)
FlushRange fl[FLUSH_RANGES + 1]; // flush_lock: nvnMemoryPoolFlushMappedRange pages not copied yet (flush_add)
@@ -1213,7 +1215,7 @@ static void dirty_frame(uint64_t frame) {
// another pool or the allocation fails.
#define VERIFY_BYTES (4u << 20)
#define VERIFY_SAMPLES 64
#define VERIFY_LOG 16
#define VERIFY_LOG 200
enum { MR_BUDGET, MR_UNTRACKED, MR_OVERLAP, MR_ALLOC, MR_N };
int vk_mirror_on;
@@ -1502,6 +1504,7 @@ void vk_mem_use(VkMemSpan *sp) {
if (r->mstate == MS_NONE) mirror_create(r);
if (r->mstate != MS_ON) return;
if (r->acc_sub != msub) mirror_sync(r);
r->vd_use_sub = msub;
sp->buffer = r->mbuf;
}
@@ -1512,6 +1515,11 @@ void vk_mem_pick(VkMemSpan *sp) {
sp->buffer = r->mbuf;
}
VkBuffer vk_mem_guest_twin(const VkMemSpan *sp) {
const PoolRec *r = sp->pool;
return r && r->mstate == MS_ON && sp->buffer == r->mbuf ? r->buf : VK_NULL_HANDLE;
}
int vk_mem_mirror_probe(uint64_t addr) {
VkMemSpan sp;
if (!vk_mirror_on || !vk_mem_resolve(addr, 1, &sp) || !sp.pool) return 0;
@@ -1755,6 +1763,7 @@ static void snap_free(void *arg) { // completion thread: the submission is don
static void gpu_wrote(const VkMemSpan *sp, uint64_t size, int shader) {
PoolRec *r = sp->pool;
if (!r || r->mstate != MS_ON || sp->buffer != r->mbuf || !size) return;
r->vd_wrote_sub = msub, r->vd_wrote_n++;
// A binding to the end of its pool (no size) can span hundreds of MiB: those stay copies.
if (shader && !(sp->offset & 3) && !(size & 3) && size <= MERGE_MAX && merging() && snap_add(r, sp->offset, size))
return;
@@ -1948,6 +1957,13 @@ static void verify_cb(void *arg) { // completion thread
rt_log("mem: verify mismatch: pool %#llx+%#llx (%s) offset %#llx: bytes %#x..%#x of the page differ, "
"mirror %02x guest %02x at the first (submission %u)", (unsigned long long)r->lo,
(unsigned long long)r->size, f, (unsigned long long)(a - r->lo), k, e, m[k], g[k], j->sub);
const uint32_t *gs = gen_slot(a & ~((1ull << GEN_PAGE_SHIFT) - 1), 0);
char mh[100], gh[100];
for (int b = 0; b < 16; b++) snprintf(mh + 3 * b, 4, "%02x ", m[b]), snprintf(gh + 3 * b, 4, "%02x ", g[b]);
rt_log("mem: verify mismatch detail: written binds %llu, last %u; used %u; written back %u; gpa record %u; "
"page cpu stamp %u (synced to %u, now %u); mirror %s| guest %s",
(unsigned long long)r->vd_wrote_n, r->vd_wrote_sub, r->vd_use_sub, wb, r->gpa_rec, gs ? *gs : 0,
s->mgen, gen_now, mh, gh);
}
}
}

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) != 34:
sys.exit(f"expected 34 patches, found {len(patches)}")
if len(patches) != 35:
sys.exit(f"expected 35 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 +1,93 @@
--- a/src/Ryujinx.Graphics.Shader/CodeGen/Spirv/Instructions.cs
+++ b/src/Ryujinx.Graphics.Shader/CodeGen/Spirv/Instructions.cs
@@ -1795,19 +1795,25 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
AstOperation operation,
Func<SpvInstruction, SpvInstruction, SpvInstruction, SpvInstruction, SpvInstruction, SpvInstruction> emitU)
{
- SpvInstruction elemPointer = GetStoragePointer(context, operation, out AggregateType varType);
+ SpvInstruction elemPointer = GetStoragePointer(context, operation, out AggregateType varType, out SpvInstruction twin);
SpvInstruction value = context.Get(varType, operation.GetSource(operation.SourcesCount - 1));
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 (twin != null)
+ {
+ emitU(context.GetType(varType), twin, one, zero, value); // NX2NAVI: guest memory's copy (GetDeviceAddressPointer)
+ }
+
+ return new OperationResult(varType, result);
}
private static OperationResult GenerateAtomicMemoryCas(CodeGenContext context, AstOperation operation)
{
- SpvInstruction elemPointer = GetStoragePointer(context, operation, out AggregateType varType);
+ SpvInstruction elemPointer = GetStoragePointer(context, operation, out AggregateType varType, out SpvInstruction twin);
SpvInstruction value0 = context.Get(varType, operation.GetSource(operation.SourcesCount - 2));
SpvInstruction value1 = context.Get(varType, operation.GetSource(operation.SourcesCount - 1));
@@ -1815,7 +1821,14 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
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 (twin != null)
+ {
+ // NX2NAVI: the same exchange on guest memory's copy, which holds the same word while both are in step
+ context.AtomicCompareExchange(context.GetType(varType), twin, one, zero, zero, value1, value0);
+ }
+
+ return new OperationResult(varType, result);
}
private static OperationResult GenerateLoadOrStore(CodeGenContext context, AstOperation operation, bool isStore)
@@ -1864,9 +1877,12 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
// The pointer addresses the 32-bit word at the address's 4-byte-aligned word. Stores also go to the twin
// (GetDeviceAddressPointers): 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). Atomics do not: the twin's word may be stale (a counter the GPU reset with a buffer fill
- // went on counting from its old value in guest memory), and that page's next sync copied the wrong count into
- // the mirror, so culled instance lists drew garbage for a frame (flickering foliage).
+ // (runtime/vk_gpa.c). So do atomics (GenerateAtomicMemoryBinary, GenerateAtomicMemoryCas): the culling passes
+ // count instances with them into records the CPU rewrites other words of each frame, and with the counts in the
+ // mirror alone the sync of such a page copied the guest's zeros over them (foliage that flickered, went missing or
+ // stretched into columns). The twin's atomic leaves its result unused, so the driver can issue it without a round
+ // trip to guest memory. Buffer fills reach the twin at once (runtime/vk_exec_xfer.c clear_buf), so a counter the
+ // GPU resets with one counts on from the same value in both.
private static SpvInstruction GetDeviceAddressPointer(CodeGenContext context, int binding, SpvInstruction low, SpvInstruction high)
{
return GetDeviceAddressPointers(context, binding, low, high, out _);
@@ -1983,19 +1999,27 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Spirv
}
private static SpvInstruction GetStoragePointer(CodeGenContext context, AstOperation operation, out AggregateType varType)
+ {
+ return GetStoragePointer(context, operation, out varType, out _);
+ }
+
+ // twin: for a device address table pointer, the twin GetDeviceAddressPointers gives; null for any other storage.
+ private static SpvInstruction GetStoragePointer(CodeGenContext context, AstOperation operation, out AggregateType varType, out SpvInstruction twin)
{
StorageKind storageKind = operation.StorageKind;
+ twin = null;
// NX2NAVI: atomics on the device address table's pointers (GenerateDeviceAddressLoadOrStore covers loads and
// stores): sources are the table binding and the address's two words.
if (storageKind == StorageKind.GlobalMemory)
{
varType = AggregateType.U32;
- return GetDeviceAddressPointer(
+ return GetDeviceAddressPointers(
context,
((AstOperand)operation.GetSource(0)).Value,
context.Get(AggregateType.U32, operation.GetSource(1)),
- context.Get(AggregateType.U32, operation.GetSource(2)));
+ context.Get(AggregateType.U32, operation.GetSource(2)),
+ out twin);
}
StorageClass storageClass;