shader: port 64-bit ALU and atomic opcodes

This commit is contained in:
Senaxx
2026-10-01 18:09:22 +04:00
committed by nmzik
parent 11001c97f7
commit 0d0b215fab
20 changed files with 687 additions and 49 deletions
@@ -111,6 +111,9 @@ uint32_t EmitIEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
uint32_t EmitINotEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
uint32_t EmitULessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
uint32_t EmitSLessThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
uint32_t EmitSLessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
uint32_t EmitULessThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
uint32_t EmitUGreaterThanEqual64(EmitterState& state, uint32_t arg0, uint32_t arg1);
uint32_t EmitUGreaterThan64(EmitterState& state, uint32_t arg0, uint32_t arg1);
EMIT_NATIVE(LogicalOr, OpLogicalOr, U1, uint32_t, uint32_t)
EMIT_NATIVE(LogicalAnd, OpLogicalAnd, U1, uint32_t, uint32_t)
@@ -134,6 +137,8 @@ uint32_t EmitFloatCompare32(ValueEmitContext& ctx, const IR::Inst& inst) {
}
inline constexpr auto EmitFPOrdEqual32 = EmitFloatCompare32<spv::OpFOrdEqual>;
EMIT_NATIVE(FPOrdEqual64, OpFOrdEqual, U1, uint32_t, uint32_t)
EMIT_NATIVE(FPOrdLessThanEqual64, OpFOrdLessThanEqual, U1, uint32_t, uint32_t)
EMIT_NATIVE(FPOrdGreaterThanEqual64, OpFOrdGreaterThanEqual, U1, uint32_t, uint32_t)
inline constexpr auto EmitFPUnordEqual32 = EmitFloatCompare32<spv::OpFUnordEqual>;
inline constexpr auto EmitFPOrdNotEqual32 = EmitFloatCompare32<spv::OpFOrdNotEqual>;
inline constexpr auto EmitFPUnordNotEqual32 = EmitFloatCompare32<spv::OpFUnordNotEqual>;
@@ -150,6 +155,7 @@ inline constexpr auto EmitFPCmpClass32 = EmitClassMaskF32;
EMIT_NATIVE(FPAdd32, OpFAdd, F32, uint32_t, uint32_t)
EMIT_NATIVE(FPSub32, OpFSub, F32, uint32_t, uint32_t)
EMIT_NATIVE(FPMul32, OpFMul, F32, uint32_t, uint32_t)
EMIT_NATIVE(FPAdd64, OpFAdd, F64, uint32_t, uint32_t)
EMIT_NATIVE(FPMul64, OpFMul, F64, uint32_t, uint32_t)
inline constexpr auto EmitFPFma64 =
EmitGlsl<GLSLstd450Fma, IR::Type::F64, uint32_t, uint32_t, uint32_t>;
@@ -223,7 +229,7 @@ void EmitLoadMemory(ValueEmitContext& ctx, const IR::Inst& inst
void EmitStoreMemory(ValueEmitContext& ctx, const IR::Inst& inst);
uint32_t EmitAtomic32(ValueEmitContext& ctx, const IR::Inst& inst);
uint32_t EmitBufferAtomic64(ValueEmitContext& ctx, const IR::Inst& inst);
void EmitSharedAtomicOr64(ValueEmitContext& ctx, const IR::Inst& inst);
void EmitSharedAtomic64(ValueEmitContext& ctx, const IR::Inst& inst);
uint32_t EmitBufferFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
void EmitSharedFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
uint32_t EmitSharedIncDec(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -259,6 +265,7 @@ inline constexpr auto EmitBufferAtomicUMin32 = EmitAtomic32;
inline constexpr auto EmitBufferAtomicSMax32 = EmitAtomic32;
inline constexpr auto EmitBufferAtomicUMax32 = EmitAtomic32;
inline constexpr auto EmitBufferAtomicAnd32 = EmitAtomic32;
inline constexpr auto EmitBufferAtomicAnd64 = EmitBufferAtomic64;
inline constexpr auto EmitBufferAtomicOr32 = EmitAtomic32;
inline constexpr auto EmitBufferAtomicOr64 = EmitBufferAtomic64;
inline constexpr auto EmitBufferAtomicXor32 = EmitAtomic32;
@@ -280,6 +287,7 @@ inline constexpr auto EmitSharedAtomicFMin32 = EmitSharedFloatAtomic;
inline constexpr auto EmitSharedAtomicFMax32 = EmitSharedFloatAtomic;
inline constexpr auto EmitSharedAtomicSwap32 = EmitAtomic32;
inline constexpr auto EmitSharedAtomicIAdd32 = EmitAtomic32;
inline constexpr auto EmitSharedAtomicIAdd64 = EmitSharedAtomic64;
inline constexpr auto EmitSharedAtomicISub32 = EmitAtomic32;
inline constexpr auto EmitSharedAtomicInc32 = EmitSharedIncDec;
inline constexpr auto EmitSharedAtomicDec32 = EmitSharedIncDec;
@@ -289,6 +297,7 @@ inline constexpr auto EmitSharedAtomicSMax32 = EmitAtomic32;
inline constexpr auto EmitSharedAtomicUMax32 = EmitAtomic32;
inline constexpr auto EmitSharedAtomicAnd32 = EmitAtomic32;
inline constexpr auto EmitSharedAtomicOr32 = EmitAtomic32;
inline constexpr auto EmitSharedAtomicOr64 = EmitSharedAtomic64;
inline constexpr auto EmitSharedAtomicXor32 = EmitAtomic32;
inline constexpr auto EmitDataAppend = EmitAppendConsume;
inline constexpr auto EmitDataConsume = EmitAppendConsume;
@@ -303,7 +312,9 @@ inline constexpr auto EmitImageSampleRaw = EmitImage;
inline constexpr auto EmitImageGatherRaw = EmitImage;
inline constexpr auto EmitImageAtomicSwap32 = EmitImage;
inline constexpr auto EmitImageAtomicIAdd32 = EmitImage;
inline constexpr auto EmitImageAtomicSMin32 = EmitImage;
inline constexpr auto EmitImageAtomicUMin32 = EmitImage;
inline constexpr auto EmitImageAtomicSMax32 = EmitImage;
inline constexpr auto EmitImageAtomicUMax32 = EmitImage;
inline constexpr auto EmitImageAtomicAnd32 = EmitImage;
inline constexpr auto EmitImageAtomicOr32 = EmitImage;