[AGC] Implement RDNA2 buffer/image/DS 32-bit atomic instructions (#222)

Adds decode and SPIR-V translation for the missing MUBUF, MIMG and DS
atomic instructions in the Gen5 shader translator, generalizing the
existing BufferAtomicAdd path. Covers swap, cmpswap, add, sub,
smin/smax, umin/umax, and/or/xor, inc and dec, plus the DS RTN
variants. Image atomics go through OpImageTexelPointer on the storage
image binding.

Notable: DS_CMPST operand order (DATA0 = comparator, DATA1 = new
value) is reversed relative to buffer/image cmpswap, which a dedicated
test locks in. ATOMIC_INC/DEC are approximated with
OpAtomicIIncrement/IDecrement, exact for the common 0xFFFFFFFF clamp.

Verified with 9 new synthetic decoder and end-to-end SPIR-V tests
(part of the #36 test corpus effort); full suite passes 35/35.

Signed-off-by: missatjuhvdk1 <177474143+missatjuhvdk1@users.noreply.github.com>
Co-authored-by: missatjuhvdk1 <177474143+missatjuhvdk1@users.noreply.github.com>
This commit is contained in:
Mees van den Kieboom
2026-07-16 16:28:39 +02:00
committed by GitHub
parent f2d9051358
commit 9883a9445d
5 changed files with 551 additions and 33 deletions
@@ -1802,24 +1802,6 @@ public static partial class Gen5SpirvTranslator
switch (instruction.Opcode)
{
case "DsAddU32":
{
if (instruction.Sources.Count < 2)
{
error = "missing LDS atomic-add source";
return false;
}
var address = GetRawSource(instruction, 0);
_module.AddInstruction(
SpirvOp.AtomicIAdd,
_uintType,
LdsPointer(address, control.Offset0),
UInt(2), // Workgroup scope.
UInt(0x108), // AcquireRelease | WorkgroupMemory.
GetRawSource(instruction, 1));
return true;
}
case "DsWriteB32":
{
if (instruction.Sources.Count < 2)
@@ -1964,6 +1946,11 @@ public static partial class Gen5SpirvTranslator
return true;
}
default:
if (Gen5ShaderTranslator.IsDataShareAtomic(instruction.Opcode))
{
return TryEmitDataShareAtomic(instruction, control, out error);
}
error = $"unsupported LDS opcode {instruction.Opcode}";
return false;
}
@@ -2004,6 +1991,126 @@ public static partial class Gen5SpirvTranslator
Store(pointer, selected);
}
private bool TryEmitDataShareAtomic(
Gen5ShaderInstruction instruction,
Gen5DataShareControl control,
out string error)
{
error = string.Empty;
var atomicOp = instruction.Opcode switch
{
"DsAddU32" or "DsAddRtnU32" => SpirvOp.AtomicIAdd,
"DsSubU32" or "DsSubRtnU32" => SpirvOp.AtomicISub,
"DsIncU32" or "DsIncRtnU32" => SpirvOp.AtomicIIncrement,
"DsDecU32" or "DsDecRtnU32" => SpirvOp.AtomicIDecrement,
"DsMinI32" or "DsMinRtnI32" => SpirvOp.AtomicSMin,
"DsMaxI32" or "DsMaxRtnI32" => SpirvOp.AtomicSMax,
"DsMinU32" or "DsMinRtnU32" => SpirvOp.AtomicUMin,
"DsMaxU32" or "DsMaxRtnU32" => SpirvOp.AtomicUMax,
"DsAndB32" or "DsAndRtnB32" => SpirvOp.AtomicAnd,
"DsOrB32" or "DsOrRtnB32" => SpirvOp.AtomicOr,
"DsXorB32" or "DsXorRtnB32" => SpirvOp.AtomicXor,
"DsWrxchgRtnB32" => SpirvOp.AtomicExchange,
"DsCmpstB32" or "DsCmpstRtnB32" => SpirvOp.AtomicCompareExchange,
_ => SpirvOp.Nop,
};
if (atomicOp == SpirvOp.Nop)
{
error = $"unsupported LDS opcode {instruction.Opcode}";
return false;
}
var address = GetRawSource(instruction, 0);
var pointer = LdsPointer(address, control.Offset0);
EmitExecConditional(() =>
{
var original = EmitAtomic(
atomicOp,
_uintType,
pointer,
scope: 2,
semantics: 0x108,
// DS_CMPST sources: DATA0 is the comparator, DATA1 the new value.
value: () => GetRawSource(
instruction,
atomicOp == SpirvOp.AtomicCompareExchange ? 2 : 1),
comparator: () => GetRawSource(instruction, 1));
if (instruction.Destinations.Count > 0)
{
StoreV(instruction.Destinations[0].Value, original);
}
});
return true;
}
// Maps the AMD atomic-op name suffix shared by buffer/image atomics to a SPIR-V opcode.
// Inc/Dec approximate the AMD wrap-clamp semantics (MEM = tmp >= DATA ? 0 : tmp + 1),
// which is exact for the common 0xFFFFFFFF clamp operand.
private static bool TryGetAtomicOp(string name, out SpirvOp op)
{
op = name switch
{
"Swap" => SpirvOp.AtomicExchange,
"Cmpswap" => SpirvOp.AtomicCompareExchange,
"Add" => SpirvOp.AtomicIAdd,
"Sub" => SpirvOp.AtomicISub,
"Smin" => SpirvOp.AtomicSMin,
"Umin" => SpirvOp.AtomicUMin,
"Smax" => SpirvOp.AtomicSMax,
"Umax" => SpirvOp.AtomicUMax,
"And" => SpirvOp.AtomicAnd,
"Or" => SpirvOp.AtomicOr,
"Xor" => SpirvOp.AtomicXor,
"Inc" => SpirvOp.AtomicIIncrement,
"Dec" => SpirvOp.AtomicIDecrement,
_ => SpirvOp.Nop,
};
return op != SpirvOp.Nop;
}
private uint EmitAtomic(
SpirvOp op,
uint type,
uint pointer,
uint scope,
uint semantics,
Func<uint> value,
Func<uint> comparator)
{
if (op is SpirvOp.AtomicIIncrement or SpirvOp.AtomicIDecrement)
{
return _module.AddInstruction(
op,
type,
pointer,
UInt(scope),
UInt(semantics));
}
if (op == SpirvOp.AtomicCompareExchange)
{
// The unequal semantics must not contain Release; downgrade it to Acquire.
return _module.AddInstruction(
op,
type,
pointer,
UInt(scope),
UInt(semantics),
UInt((semantics & ~0x8u) | 0x2u),
value(),
comparator());
}
return _module.AddInstruction(
op,
type,
pointer,
UInt(scope),
UInt(semantics),
value());
}
private bool TryEmitInterpolation(
Gen5ShaderInstruction instruction,
Gen5InterpolationControl interpolation,
@@ -2239,22 +2346,27 @@ public static partial class Gen5SpirvTranslator
byteAddress = ApplyGuestBufferByteBias(bindingIndex, byteAddress);
var dwordAddress = ShiftRightLogical(byteAddress, UInt(2));
if (instruction.Opcode is "BufferAtomicAdd" or "BufferAtomicUMax")
if (instruction.Opcode.StartsWith("BufferAtomic", StringComparison.Ordinal))
{
if (!TryGetAtomicOp(instruction.Opcode["BufferAtomic".Length..], out var atomicOp))
{
error = $"unsupported buffer opcode {instruction.Opcode}";
return false;
}
EmitExecConditional(() =>
{
var inRange = IsBufferWordInRange(bindingIndex, dwordAddress);
EmitConditional(inRange, () =>
{
var original = _module.AddInstruction(
instruction.Opcode == "BufferAtomicAdd"
? SpirvOp.AtomicIAdd
: SpirvOp.AtomicUMax,
var original = EmitAtomic(
atomicOp,
_uintType,
BufferWordPointer(bindingIndex, dwordAddress),
UInt(1),
UInt(0x48),
LoadV(control.VectorData));
scope: 1,
semantics: 0x48,
value: () => LoadV(control.VectorData),
comparator: () => LoadV(control.VectorData + 1));
if (control.Glc)
{
StoreV(control.VectorData, original);
@@ -3196,6 +3308,60 @@ public static partial class Gen5SpirvTranslator
return true;
}
if (instruction.Opcode.StartsWith("ImageAtomic", StringComparison.Ordinal))
{
if (!resource.IsStorage)
{
error = "image atomic is not bound as storage";
return false;
}
if (resource.ComponentKind == ImageComponentKind.Float ||
!TryGetAtomicOp(instruction.Opcode["ImageAtomic".Length..], out var atomicOp))
{
error = $"unsupported storage image opcode {instruction.Opcode}";
return false;
}
var signed = resource.ComponentKind == ImageComponentKind.Sint;
var atomicImageSize = _module.AddInstruction(
SpirvOp.ImageQuerySize,
_module.TypeVector(_intType, 2),
imageObject);
var coordinates = BuildClampedIntegerCoordinates(
image,
0,
atomicImageSize);
EmitExecConditional(() =>
{
var pointer = _module.AddInstruction(
SpirvOp.ImageTexelPointer,
_module.TypePointer(SpirvStorageClass.Image, resource.ComponentType),
resource.Variable,
coordinates,
UInt(0));
uint LoadData(uint register) => signed
? Bitcast(_intType, LoadV(register))
: LoadV(register);
var original = EmitAtomic(
atomicOp,
resource.ComponentType,
pointer,
scope: 1,
semantics: 0x808,
value: () => LoadData(image.VectorData),
comparator: () => LoadData(image.VectorData + 1));
if (image.Glc)
{
StoreV(
image.VectorData,
signed ? Bitcast(_uintType, original) : original);
}
});
return true;
}
if (resource.IsStorage &&
instruction.Opcode is not ("ImageLoad" or "ImageLoadMip"))
{