Files
UnrealEngineUWP/Engine/Shaders/Private/BitStreamReaderImplementation.ush
rune stubbe fdf873c143 Give BitStreamReader_Create/BitStreamReader_Create_Aligned overloads unique names to fix shader compilation issues on certain compilers.
#rb lukas.hermanns
[FYI] graham.wihlidal

#ROBOMERGE-SOURCE: CL 16658432 via CL 16659805
#ROBOMERGE-BOT: STARSHIP (Main -> Release-Engine-Test) (v834-16658389)

[CL 16659878 by rune stubbe in ue5-release-engine-test branch]
2021-06-14 12:32:05 -04:00

188 lines
7.0 KiB
Plaintext

// Copyright Epic Games, Inc. All Rights Reserved.
// Bit buffer implementation:
// Maintains an internal bit buffer instead of issuing memory loads at every read operation.
// Reads extract the bits from the bottom dword of the bit buffer. Whenever the bottom dword runs out of bits,
// it is refilled by shifting the bit buffer down (v_alignbit). Only when the bit buffer also runs out of bits
// is a memory load issued that then refills the buffer using a single load4.
// If the read sizes are divergent, it is very likely that for a given read at least one thread will need to refill, so
// in the worst case the refill has to happen at every read.
// To mitigate this, all reads have to supply a compile-time constant upper bound to the size of the read.
// By keeping track of these bounds, we can conservatively determine a which reads a refill can possibly be required and only
// emit the refill code in those instances.
// Everything prefixed with CompileTime should be compile-time constant and generate no code.
// We unfortunately have no way to enforce this.
// BitStreamReader
// Helper 'class' for efficiently parsing bit streams of arbitrary length.
#define CONCAT2(A, B) A##B
#define CONCAT(A, B) CONCAT2(A, B)
#define FBitStreamReaderStateType CONCAT(FBitStreamReaderState, TYPE_SUFFIX)
struct FBitStreamReaderStateType
{
#if RO_ENABLED
ByteAddressBuffer InputBufferRO;
#endif
#if RW_ENABLED
RWByteAddressBuffer InputBufferRW;
#endif
#if RO_ENABLED && RW_ENABLED
bool bRW;
#endif
uint AlignedByteAddress;
int BitOffsetFromAddress;
uint4 BufferBits;
int BufferOffset;
int CompileTimeMinBufferBits;
int CompileTimeMinDwordBits;
int CompileTimeMaxRemainingBits;
};
FBitStreamReaderStateType CONCAT(BitStreamReader_Create_Aligned_, TYPE_SUFFIX)(
#if RO_ENABLED
ByteAddressBuffer InputBufferRO,
#endif
#if RW_ENABLED
RWByteAddressBuffer InputBufferRW,
#endif
#if RO_ENABLED && RW_ENABLED
bool bRW,
#endif
uint AlignedByteAddress, uint BitOffset, uint CompileTimeMaxRemainingBits)
{
FBitStreamReaderStateType State;
#if RO_ENABLED
State.InputBufferRO = InputBufferRO;
#endif
#if RW_ENABLED
State.InputBufferRW = InputBufferRW;
#endif
#if RO_ENABLED && RW_ENABLED
State.bRW = bRW;
#endif
State.AlignedByteAddress = AlignedByteAddress;
State.BitOffsetFromAddress = BitOffset;
State.BufferBits = 0;
State.BufferOffset = 0;
State.CompileTimeMinBufferBits = 0;
State.CompileTimeMinDwordBits = 0;
State.CompileTimeMaxRemainingBits = CompileTimeMaxRemainingBits;
return State;
}
FBitStreamReaderStateType CONCAT(BitStreamReader_Create_, TYPE_SUFFIX)(
#if RO_ENABLED
ByteAddressBuffer InputBufferRO,
#endif
#if RW_ENABLED
RWByteAddressBuffer InputBufferRW,
#endif
#if RO_ENABLED && RW_ENABLED
bool bRW,
#endif
uint ByteAddress, uint BitOffset, uint CompileTimeMaxRemainingBits)
{
uint AlignedByteAddress = ByteAddress & ~3u;
BitOffset += (ByteAddress & 3u) << 3;
return CONCAT(BitStreamReader_Create_Aligned_, TYPE_SUFFIX)(
#if RO_ENABLED
InputBufferRO,
#endif
#if RW_ENABLED
InputBufferRW,
#endif
#if RO_ENABLED && RW_ENABLED
bRW,
#endif
AlignedByteAddress, BitOffset, CompileTimeMaxRemainingBits);
}
uint BitStreamReader_Read(inout FBitStreamReaderStateType State, int NumBits, int CompileTimeMaxBits)
{
if (CompileTimeMaxBits > State.CompileTimeMinBufferBits)
{
// BitBuffer could be out of bits: Reload.
// Add cumulated offset since last refill. No need to update at every read.
State.BitOffsetFromAddress += State.BufferOffset;
uint Address = State.AlignedByteAddress + ((State.BitOffsetFromAddress >> 5) << 2);
#if RO_ENABLED && RW_ENABLED
uint4 Data = State.bRW ? State.InputBufferRW.Load4(Address) : State.InputBufferRO.Load4(Address);
#elif RO_ENABLED
uint4 Data = State.InputBufferRO.Load4(Address);
#elif RW_ENABLED
uint4 Data = State.InputBufferRW.Load4(Address);
#endif
// Shift bits down to align
State.BufferBits.x = BitAlignU32(Data.y, Data.x, State.BitOffsetFromAddress); // BitOffsetFromAddress implicitly &31
if (State.CompileTimeMaxRemainingBits > 32) State.BufferBits.y = BitAlignU32(Data.z, Data.y, State.BitOffsetFromAddress); // BitOffsetFromAddress implicitly &31
if (State.CompileTimeMaxRemainingBits > 64) State.BufferBits.z = BitAlignU32(Data.w, Data.z, State.BitOffsetFromAddress); // BitOffsetFromAddress implicitly &31
if (State.CompileTimeMaxRemainingBits > 96) State.BufferBits.w = BitAlignU32(0, Data.w, State.BitOffsetFromAddress); // BitOffsetFromAddress implicitly &31
State.BufferOffset = 0;
State.CompileTimeMinDwordBits = min(32, State.CompileTimeMaxRemainingBits);
State.CompileTimeMinBufferBits = min(97, State.CompileTimeMaxRemainingBits); // Up to 31 bits wasted to alignment
}
else if (CompileTimeMaxBits > State.CompileTimeMinDwordBits)
{
// Bottom dword could be out of bits: Shift down.
State.BitOffsetFromAddress += State.BufferOffset;
State.BufferBits.x = BitAlignU32(State.BufferBits.y, State.BufferBits.x, State.BufferOffset); // BufferOffset implicitly &31
if (State.CompileTimeMinBufferBits > 32) State.BufferBits.y = BitAlignU32(State.BufferBits.z, State.BufferBits.y, State.BufferOffset); // BufferOffset implicitly &31
if (State.CompileTimeMinBufferBits > 64) State.BufferBits.z = BitAlignU32(State.BufferBits.w, State.BufferBits.z, State.BufferOffset); // BufferOffset implicitly &31
if (State.CompileTimeMinBufferBits > 96) State.BufferBits.w = BitAlignU32(0, State.BufferBits.w, State.BufferOffset); // BufferOffset implicitly &31
State.BufferOffset = 0;
State.CompileTimeMinDwordBits = min(32, State.CompileTimeMaxRemainingBits);
}
const uint Result = BitFieldExtractU32(State.BufferBits.x, NumBits, State.BufferOffset); // BufferOffset implicitly &31
State.BufferOffset += NumBits;
State.CompileTimeMinBufferBits -= CompileTimeMaxBits;
State.CompileTimeMinDwordBits -= CompileTimeMaxBits;
State.CompileTimeMaxRemainingBits -= CompileTimeMaxBits;
return Result;
}
uint2 BitStreamReader_Read2(inout FBitStreamReaderStateType State, int2 NumBits, int2 CompileTimeMaxBits)
{
uint ResultX = BitStreamReader_Read(State, NumBits.x, CompileTimeMaxBits.x);
uint ResultY = BitStreamReader_Read(State, NumBits.y, CompileTimeMaxBits.y);
return uint2(ResultX, ResultY);
}
uint3 BitStreamReader_Read3(inout FBitStreamReaderStateType State, int3 NumBits, int3 CompileTimeMaxBits)
{
uint ResultX = BitStreamReader_Read(State, NumBits.x, CompileTimeMaxBits.x);
uint ResultY = BitStreamReader_Read(State, NumBits.y, CompileTimeMaxBits.y);
uint ResultZ = BitStreamReader_Read(State, NumBits.z, CompileTimeMaxBits.z);
return uint3(ResultX, ResultY, ResultZ);
}
uint4 BitStreamReader_Read4(inout FBitStreamReaderStateType State, int4 NumBits, int4 CompileTimeMaxBits)
{
uint ResultX = BitStreamReader_Read(State, NumBits.x, CompileTimeMaxBits.x);
uint ResultY = BitStreamReader_Read(State, NumBits.y, CompileTimeMaxBits.y);
uint ResultZ = BitStreamReader_Read(State, NumBits.z, CompileTimeMaxBits.z);
uint ResultW = BitStreamReader_Read(State, NumBits.w, CompileTimeMaxBits.w);
return uint4(ResultX, ResultY, ResultZ, ResultW);
}
#undef CONCAT
#undef CONCAT2