--- a/src/Ryujinx.Graphics.Shader/Instructions/InstEmitMemory.cs +++ b/src/Ryujinx.Graphics.Shader/Instructions/InstEmitMemory.cs @@ -217,6 +217,13 @@ { res = context.AtomicAdd(storageKind, e0, e1, value); } + else if (type == AtomSize.F32FtzRn && storageKind == StorageKind.GlobalMemory) + { + // NX2NAVI: a float add (RED/ATOM .F32.FTZ.RN), which accumulators of light and fog use, was + // dropped here. It is marked FP32 and becomes a compare-and-swap loop with a float add once + // GlobalToStorage knows its buffer (TryGenerateStorageOp). + res = context.Add(Instruction.AtomicAdd | Instruction.FP32, storageKind, Local(), e0, e1, value); + } else { context.TranslatorContext.GpuAccessor.Log($"Invalid reduction type: {type}."); --- a/src/Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs +++ b/src/Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs @@ -448,6 +448,13 @@ return false; } + // NX2NAVI: a float atomic add has no inline form (the SPIR-V backend only adds integers): it takes the + // helper path, whose TryGenerateStorageOp writes a compare-and-swap loop + if (operation.Inst == (Instruction.AtomicAdd | Instruction.FP32)) + { + return false; + } + return (operation.Inst != Instruction.AtomicMaxS32 && operation.Inst != Instruction.AtomicMinS32) || targetLanguage == TargetLanguage.Spirv; } @@ -928,6 +935,15 @@ case Instruction.AtomicAdd: resultValue = context.AtomicAdd(StorageKind.StorageBuffer, binding, Const(0), wordOffset, value); break; + 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) => + { + return context.FPAdd(memValue, value); + }); + + break; case Instruction.AtomicAnd: resultValue = context.AtomicAnd(StorageKind.StorageBuffer, binding, Const(0), wordOffset, value); break;