Files
NX2navi/tools/xlate/patches/0041-cas-loop-attempt-limit.patch
chrg6 4750e1ffa5 Bound the translator's storage buffer compare-and-swap loops at 4,096 attempts.
The loops the translator writes for narrow storage buffer stores and float atomic adds end only when their
compare-and-swap succeeds. With robust buffer access on, an atomic on a word outside its buffer may return anything,
so the compare can fail for ever and the GPU hangs until the device is lost, as reported right after the first
cutscene on an Intel Arc GPU (#610). New translator patch 0041 counts the attempts in a private word (a local cannot
count round a loop in that pass, which runs on SSA form) and leaves the loop after 4,096; contention on a word needs
about as many attempts as invocations race for it. Shared memory loops stay as they are: a shared word in range is
always settled, and the mesh programs run that loop for every primitive index byte.

Checked: all 4,238 NVN2 compute programs an installed shader cache stores translate and pass spirv-val (vulkan1.3); 45
contain a compare-and-swap, 11 of them in a storage loop, which grew by 256 bytes each and now compare the count with
4096; the other 34 are unchanged. tools/xlate_smoke.py passes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-11 19:38:18 -05:00

90 lines
5.2 KiB
Diff

--- a/src/Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs
+++ b/src/Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs
@@ -1118,7 +1118,7 @@
case Instruction.AtomicAdd | Instruction.FP32:
// NX2NAVI: the word is read, the float sum swapped in if the word is still the one read,
// and the loop repeats otherwise (InstEmitMemory EmitAtomicOp marks these)
- resultValue = GenerateAtomicCasLoop(context, wordOffset, binding, (memValue) =>
+ resultValue = GenerateAtomicCasLoop(context, resourceManager, wordOffset, binding, (memValue) =>
{
return context.FPAdd(memValue, value);
});
@@ -1137,7 +1137,7 @@
}
else
{
- resultValue = GenerateAtomicCasLoop(context, wordOffset, binding, (memValue) =>
+ resultValue = GenerateAtomicCasLoop(context, resourceManager, wordOffset, binding, (memValue) =>
{
return context.IMaximumS32(memValue, value);
});
@@ -1154,7 +1154,7 @@
}
else
{
- resultValue = GenerateAtomicCasLoop(context, wordOffset, binding, (memValue) =>
+ resultValue = GenerateAtomicCasLoop(context, resourceManager, wordOffset, binding, (memValue) =>
{
return context.IMinimumS32(memValue, value);
});
@@ -1190,7 +1190,7 @@
{
Operand bitOffset = HelperFunctionManager.GetBitOffset(context, offset);
- GenerateAtomicCasLoop(context, wordOffset, binding, (memValue) =>
+ GenerateAtomicCasLoop(context, resourceManager, wordOffset, binding, (memValue) =>
{
return context.BitfieldInsert(memValue, value, bitOffset, Const(bitSize));
});
@@ -1235,8 +1235,19 @@
return true;
}
- private static Operand GenerateAtomicCasLoop(EmitterContext context, Operand wordOffset, int binding, Func<Operand, Operand> opCallback)
+ // NX2NAVI: the most attempts a storage buffer compare-and-swap loop makes before it gives up. The loop otherwise
+ // ends only when its compare-and-swap succeeds; on a word outside its buffer robust buffer access lets the
+ // atomics return anything, so the compare may never match, and the GPU hangs (device lost). Contention on a
+ // word takes about as many attempts as invocations race for it, far below this. Shared memory loops
+ // (HelperFunctionManager) stay unbounded: a shared word in range is always settled, and the mesh programs'
+ // byte stores run that loop for every primitive index.
+ private const int CasAttemptLimit = 4096;
+
+ private static Operand GenerateAtomicCasLoop(EmitterContext context, ResourceManager resourceManager, Operand wordOffset, int binding, Func<Operand, Operand> opCallback)
{
+ int attempts = resourceManager.CasAttemptsMemoryId;
+ context.Store(StorageKind.LocalMemory, attempts, Const(0));
+
// NX2NAVI: each attempt reads the word with an atomic OR of 0, which returns its current value; a plain load
// could keep returning a stale cached copy (seen on AMD), so an invocation that lost the race never matched
// again and the loop did not end. The read stays inside the loop: this pass runs on SSA form, and the
@@ -1258,7 +1269,10 @@
Operand casFail = context.ICompareNotEqual(casResult, oldValue);
- context.BranchIfTrue(lblLoopHead, casFail);
+ Operand made = context.IAdd(context.Load(StorageKind.LocalMemory, attempts), Const(1));
+ context.Store(StorageKind.LocalMemory, attempts, made);
+
+ context.BranchIfTrue(lblLoopHead, context.BitwiseAnd(casFail, context.ICompareLessUnsigned(made, Const(CasAttemptLimit))));
return oldValue;
}
--- a/src/Ryujinx.Graphics.Shader/Translation/ResourceManager.cs
+++ b/src/Ryujinx.Graphics.Shader/Translation/ResourceManager.cs
@@ -143,6 +143,15 @@
}
}
+ // NX2NAVI: a per-invocation count of compare-and-swap attempts, for the storage buffer loops of
+ // GlobalToStorage (GenerateAtomicCasLoop). Private memory, not a local: that pass runs on SSA form, where copy
+ // propagation takes a local as assigned once, so no local can count round a loop. Made on first use.
+ private int _casAttemptsMemoryId = -1;
+
+ public int CasAttemptsMemoryId => _casAttemptsMemoryId >= 0
+ ? _casAttemptsMemoryId
+ : _casAttemptsMemoryId = AddMemoryDefinition("cas_attempts", AggregateType.U32);
+
private int AddMemoryDefinition(string name, AggregateType type, int arrayLength = 1)
{
return Properties.AddLocalMemory(new MemoryDefinition(name, type, arrayLength));