forked from ShuriZma/suyu
1
0
Fork 0

Merge pull request #3631 from ReinUsesLisp/more-astc

texture/astc: More small ASTC optimizations
This commit is contained in:
Mat M 2020-04-13 10:17:32 -04:00 committed by GitHub
commit c4001225f6
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 159 additions and 82 deletions

View File

@ -20,6 +20,8 @@
#include <cstring> #include <cstring>
#include <vector> #include <vector>
#include <boost/container/static_vector.hpp>
#include "common/common_types.h" #include "common/common_types.h"
#include "video_core/textures/astc.h" #include "video_core/textures/astc.h"
@ -39,25 +41,25 @@ constexpr u32 Popcnt(u32 n) {
class InputBitStream { class InputBitStream {
public: public:
explicit InputBitStream(const u8* ptr, std::size_t start_offset = 0) constexpr explicit InputBitStream(const u8* ptr, std::size_t start_offset = 0)
: m_CurByte(ptr), m_NextBit(start_offset % 8) {} : cur_byte{ptr}, next_bit{start_offset % 8} {}
std::size_t GetBitsRead() const { constexpr std::size_t GetBitsRead() const {
return m_BitsRead; return bits_read;
} }
u32 ReadBit() { constexpr bool ReadBit() {
u32 bit = *m_CurByte >> m_NextBit++; const bool bit = (*cur_byte >> next_bit++) & 1;
while (m_NextBit >= 8) { while (next_bit >= 8) {
m_NextBit -= 8; next_bit -= 8;
m_CurByte++; cur_byte++;
} }
m_BitsRead++; bits_read++;
return bit & 1; return bit;
} }
u32 ReadBits(std::size_t nBits) { constexpr u32 ReadBits(std::size_t nBits) {
u32 ret = 0; u32 ret = 0;
for (std::size_t i = 0; i < nBits; ++i) { for (std::size_t i = 0; i < nBits; ++i) {
ret |= (ReadBit() & 1) << i; ret |= (ReadBit() & 1) << i;
@ -66,7 +68,7 @@ public:
} }
template <std::size_t nBits> template <std::size_t nBits>
u32 ReadBits() { constexpr u32 ReadBits() {
u32 ret = 0; u32 ret = 0;
for (std::size_t i = 0; i < nBits; ++i) { for (std::size_t i = 0; i < nBits; ++i) {
ret |= (ReadBit() & 1) << i; ret |= (ReadBit() & 1) << i;
@ -75,64 +77,58 @@ public:
} }
private: private:
const u8* m_CurByte; const u8* cur_byte;
std::size_t m_NextBit = 0; std::size_t next_bit = 0;
std::size_t m_BitsRead = 0; std::size_t bits_read = 0;
}; };
class OutputBitStream { class OutputBitStream {
public: public:
explicit OutputBitStream(u8* ptr, s32 nBits = 0, s32 start_offset = 0) constexpr explicit OutputBitStream(u8* ptr, std::size_t bits = 0, std::size_t start_offset = 0)
: m_NumBits(nBits), m_CurByte(ptr), m_NextBit(start_offset % 8) {} : cur_byte{ptr}, num_bits{bits}, next_bit{start_offset % 8} {}
~OutputBitStream() = default; constexpr std::size_t GetBitsWritten() const {
return bits_written;
s32 GetBitsWritten() const {
return m_BitsWritten;
} }
void WriteBitsR(u32 val, u32 nBits) { constexpr void WriteBitsR(u32 val, u32 nBits) {
for (u32 i = 0; i < nBits; i++) { for (u32 i = 0; i < nBits; i++) {
WriteBit((val >> (nBits - i - 1)) & 1); WriteBit((val >> (nBits - i - 1)) & 1);
} }
} }
void WriteBits(u32 val, u32 nBits) { constexpr void WriteBits(u32 val, u32 nBits) {
for (u32 i = 0; i < nBits; i++) { for (u32 i = 0; i < nBits; i++) {
WriteBit((val >> i) & 1); WriteBit((val >> i) & 1);
} }
} }
private: private:
void WriteBit(s32 b) { constexpr void WriteBit(bool b) {
if (bits_written >= num_bits) {
if (done)
return; return;
}
const u32 mask = 1 << m_NextBit++; const u32 mask = 1 << next_bit++;
// clear the bit // clear the bit
*m_CurByte &= static_cast<u8>(~mask); *cur_byte &= static_cast<u8>(~mask);
// Write the bit, if necessary // Write the bit, if necessary
if (b) if (b)
*m_CurByte |= static_cast<u8>(mask); *cur_byte |= static_cast<u8>(mask);
// Next byte? // Next byte?
if (m_NextBit >= 8) { if (next_bit >= 8) {
m_CurByte += 1; cur_byte += 1;
m_NextBit = 0; next_bit = 0;
} }
done = done || ++m_BitsWritten >= m_NumBits;
} }
s32 m_BitsWritten = 0; u8* cur_byte;
const s32 m_NumBits; std::size_t num_bits;
u8* m_CurByte; std::size_t bits_written = 0;
s32 m_NextBit = 0; std::size_t next_bit = 0;
bool done = false;
}; };
template <typename IntType> template <typename IntType>
@ -195,9 +191,13 @@ struct IntegerEncodedValue {
u32 trit_value; u32 trit_value;
}; };
}; };
using IntegerEncodedVector = boost::container::static_vector<
IntegerEncodedValue, 64,
boost::container::static_vector_options<
boost::container::inplace_alignment<alignof(IntegerEncodedValue)>,
boost::container::throw_on_overflow<false>>::type>;
static void DecodeTritBlock(InputBitStream& bits, std::vector<IntegerEncodedValue>& result, static void DecodeTritBlock(InputBitStream& bits, IntegerEncodedVector& result, u32 nBitsPerValue) {
u32 nBitsPerValue) {
// Implement the algorithm in section C.2.12 // Implement the algorithm in section C.2.12
u32 m[5]; u32 m[5];
u32 t[5]; u32 t[5];
@ -255,7 +255,7 @@ static void DecodeTritBlock(InputBitStream& bits, std::vector<IntegerEncodedValu
} }
} }
static void DecodeQus32Block(InputBitStream& bits, std::vector<IntegerEncodedValue>& result, static void DecodeQus32Block(InputBitStream& bits, IntegerEncodedVector& result,
u32 nBitsPerValue) { u32 nBitsPerValue) {
// Implement the algorithm in section C.2.12 // Implement the algorithm in section C.2.12
u32 m[3]; u32 m[3];
@ -343,8 +343,8 @@ static constexpr std::array EncodingsValues = MakeEncodedValues();
// Fills result with the values that are encoded in the given // Fills result with the values that are encoded in the given
// bitstream. We must know beforehand what the maximum possible // bitstream. We must know beforehand what the maximum possible
// value is, and how many values we're decoding. // value is, and how many values we're decoding.
static void DecodeIntegerSequence(std::vector<IntegerEncodedValue>& result, InputBitStream& bits, static void DecodeIntegerSequence(IntegerEncodedVector& result, InputBitStream& bits, u32 maxRange,
u32 maxRange, u32 nValues) { u32 nValues) {
// Determine encoding parameters // Determine encoding parameters
IntegerEncodedValue val = EncodingsValues[maxRange]; IntegerEncodedValue val = EncodingsValues[maxRange];
@ -634,12 +634,14 @@ static void FillError(u32* outBuf, u32 blockWidth, u32 blockHeight) {
// Replicates low numBits such that [(toBit - 1):(toBit - 1 - fromBit)] // Replicates low numBits such that [(toBit - 1):(toBit - 1 - fromBit)]
// is the same as [(numBits - 1):0] and repeats all the way down. // is the same as [(numBits - 1):0] and repeats all the way down.
template <typename IntType> template <typename IntType>
static IntType Replicate(IntType val, u32 numBits, u32 toBit) { static constexpr IntType Replicate(IntType val, u32 numBits, u32 toBit) {
if (numBits == 0) if (numBits == 0) {
return 0; return 0;
if (toBit == 0) }
if (toBit == 0) {
return 0; return 0;
IntType v = val & static_cast<IntType>((1 << numBits) - 1); }
const IntType v = val & static_cast<IntType>((1 << numBits) - 1);
IntType res = v; IntType res = v;
u32 reslen = numBits; u32 reslen = numBits;
while (reslen < toBit) { while (reslen < toBit) {
@ -656,6 +658,89 @@ static IntType Replicate(IntType val, u32 numBits, u32 toBit) {
return res; return res;
} }
static constexpr std::size_t NumReplicateEntries(u32 num_bits) {
return std::size_t(1) << num_bits;
}
template <typename IntType, u32 num_bits, u32 to_bit>
static constexpr auto MakeReplicateTable() {
std::array<IntType, NumReplicateEntries(num_bits)> table{};
for (IntType value = 0; value < static_cast<IntType>(std::size(table)); ++value) {
table[value] = Replicate(value, num_bits, to_bit);
}
return table;
}
static constexpr auto REPLICATE_BYTE_TO_16_TABLE = MakeReplicateTable<u32, 8, 16>();
static constexpr u32 ReplicateByteTo16(std::size_t value) {
return REPLICATE_BYTE_TO_16_TABLE[value];
}
static constexpr auto REPLICATE_BIT_TO_7_TABLE = MakeReplicateTable<u32, 1, 7>();
static constexpr u32 ReplicateBitTo7(std::size_t value) {
return REPLICATE_BIT_TO_7_TABLE[value];
}
static constexpr auto REPLICATE_BIT_TO_9_TABLE = MakeReplicateTable<u32, 1, 9>();
static constexpr u32 ReplicateBitTo9(std::size_t value) {
return REPLICATE_BIT_TO_9_TABLE[value];
}
static constexpr auto REPLICATE_1_BIT_TO_8_TABLE = MakeReplicateTable<u32, 1, 8>();
static constexpr auto REPLICATE_2_BIT_TO_8_TABLE = MakeReplicateTable<u32, 2, 8>();
static constexpr auto REPLICATE_3_BIT_TO_8_TABLE = MakeReplicateTable<u32, 3, 8>();
static constexpr auto REPLICATE_4_BIT_TO_8_TABLE = MakeReplicateTable<u32, 4, 8>();
static constexpr auto REPLICATE_5_BIT_TO_8_TABLE = MakeReplicateTable<u32, 5, 8>();
static constexpr auto REPLICATE_6_BIT_TO_8_TABLE = MakeReplicateTable<u32, 6, 8>();
static constexpr auto REPLICATE_7_BIT_TO_8_TABLE = MakeReplicateTable<u32, 7, 8>();
static constexpr auto REPLICATE_8_BIT_TO_8_TABLE = MakeReplicateTable<u32, 8, 8>();
/// Use a precompiled table with the most common usages, if it's not in the expected range, fallback
/// to the runtime implementation
static constexpr u32 FastReplicateTo8(u32 value, u32 num_bits) {
switch (num_bits) {
case 1:
return REPLICATE_1_BIT_TO_8_TABLE[value];
case 2:
return REPLICATE_2_BIT_TO_8_TABLE[value];
case 3:
return REPLICATE_3_BIT_TO_8_TABLE[value];
case 4:
return REPLICATE_4_BIT_TO_8_TABLE[value];
case 5:
return REPLICATE_5_BIT_TO_8_TABLE[value];
case 6:
return REPLICATE_6_BIT_TO_8_TABLE[value];
case 7:
return REPLICATE_7_BIT_TO_8_TABLE[value];
case 8:
return REPLICATE_8_BIT_TO_8_TABLE[value];
default:
return Replicate(value, num_bits, 8);
}
}
static constexpr auto REPLICATE_1_BIT_TO_6_TABLE = MakeReplicateTable<u32, 1, 6>();
static constexpr auto REPLICATE_2_BIT_TO_6_TABLE = MakeReplicateTable<u32, 2, 6>();
static constexpr auto REPLICATE_3_BIT_TO_6_TABLE = MakeReplicateTable<u32, 3, 6>();
static constexpr auto REPLICATE_4_BIT_TO_6_TABLE = MakeReplicateTable<u32, 4, 6>();
static constexpr auto REPLICATE_5_BIT_TO_6_TABLE = MakeReplicateTable<u32, 5, 6>();
static constexpr u32 FastReplicateTo6(u32 value, u32 num_bits) {
switch (num_bits) {
case 1:
return REPLICATE_1_BIT_TO_6_TABLE[value];
case 2:
return REPLICATE_2_BIT_TO_6_TABLE[value];
case 3:
return REPLICATE_3_BIT_TO_6_TABLE[value];
case 4:
return REPLICATE_4_BIT_TO_6_TABLE[value];
case 5:
return REPLICATE_5_BIT_TO_6_TABLE[value];
default:
return Replicate(value, num_bits, 6);
}
}
class Pixel { class Pixel {
protected: protected:
using ChannelType = s16; using ChannelType = s16;
@ -674,10 +759,10 @@ public:
// significant bits when going from larger to smaller bit depth // significant bits when going from larger to smaller bit depth
// or by repeating the most significant bits when going from // or by repeating the most significant bits when going from
// smaller to larger bit depths. // smaller to larger bit depths.
void ChangeBitDepth(const u8 (&depth)[4]) { void ChangeBitDepth() {
for (u32 i = 0; i < 4; i++) { for (u32 i = 0; i < 4; i++) {
Component(i) = ChangeBitDepth(Component(i), m_BitDepth[i], depth[i]); Component(i) = ChangeBitDepth(Component(i), m_BitDepth[i]);
m_BitDepth[i] = depth[i]; m_BitDepth[i] = 8;
} }
} }
@ -689,28 +774,23 @@ public:
// Changes the bit depth of a single component. See the comment // Changes the bit depth of a single component. See the comment
// above for how we do this. // above for how we do this.
static ChannelType ChangeBitDepth(Pixel::ChannelType val, u8 oldDepth, u8 newDepth) { static ChannelType ChangeBitDepth(Pixel::ChannelType val, u8 oldDepth) {
assert(newDepth <= 8);
assert(oldDepth <= 8); assert(oldDepth <= 8);
if (oldDepth == newDepth) { if (oldDepth == 8) {
// Do nothing // Do nothing
return val; return val;
} else if (oldDepth == 0 && newDepth != 0) { } else if (oldDepth == 0) {
return static_cast<ChannelType>((1 << newDepth) - 1); return static_cast<ChannelType>((1 << 8) - 1);
} else if (newDepth > oldDepth) { } else if (8 > oldDepth) {
return Replicate(val, oldDepth, newDepth); return static_cast<ChannelType>(FastReplicateTo8(static_cast<u32>(val), oldDepth));
} else { } else {
// oldDepth > newDepth // oldDepth > newDepth
if (newDepth == 0) { const u8 bitsWasted = static_cast<u8>(oldDepth - 8);
return 0xFF; u16 v = static_cast<u16>(val);
} else { v = static_cast<u16>((v + (1 << (bitsWasted - 1))) >> bitsWasted);
u8 bitsWasted = static_cast<u8>(oldDepth - newDepth); v = ::std::min<u16>(::std::max<u16>(0, v), static_cast<u16>((1 << 8) - 1));
u16 v = static_cast<u16>(val); return static_cast<u8>(v);
v = static_cast<u16>((v + (1 << (bitsWasted - 1))) >> bitsWasted);
v = ::std::min<u16>(::std::max<u16>(0, v), static_cast<u16>((1 << newDepth) - 1));
return static_cast<u8>(v);
}
} }
assert(false && "We shouldn't get here."); assert(false && "We shouldn't get here.");
@ -760,8 +840,7 @@ public:
// up in the most-significant byte. // up in the most-significant byte.
u32 Pack() const { u32 Pack() const {
Pixel eightBit(*this); Pixel eightBit(*this);
const u8 eightBitDepth[4] = {8, 8, 8, 8}; eightBit.ChangeBitDepth();
eightBit.ChangeBitDepth(eightBitDepth);
u32 r = 0; u32 r = 0;
r |= eightBit.A(); r |= eightBit.A();
@ -816,8 +895,7 @@ static void DecodeColorValues(u32* out, u8* data, const u32* modes, const u32 nP
} }
// We now have enough to decode our integer sequence. // We now have enough to decode our integer sequence.
std::vector<IntegerEncodedValue> decodedColorValues; IntegerEncodedVector decodedColorValues;
decodedColorValues.reserve(32);
InputBitStream colorStream(data); InputBitStream colorStream(data);
DecodeIntegerSequence(decodedColorValues, colorStream, range, nValues); DecodeIntegerSequence(decodedColorValues, colorStream, range, nValues);
@ -839,12 +917,12 @@ static void DecodeColorValues(u32* out, u8* data, const u32* modes, const u32 nP
u32 A = 0, B = 0, C = 0, D = 0; u32 A = 0, B = 0, C = 0, D = 0;
// A is just the lsb replicated 9 times. // A is just the lsb replicated 9 times.
A = Replicate(bitval & 1, 1, 9); A = ReplicateBitTo9(bitval & 1);
switch (val.encoding) { switch (val.encoding) {
// Replicate bits // Replicate bits
case IntegerEncoding::JustBits: case IntegerEncoding::JustBits:
out[outIdx++] = Replicate(bitval, bitlen, 8); out[outIdx++] = FastReplicateTo8(bitval, bitlen);
break; break;
// Use algorithm in C.2.13 // Use algorithm in C.2.13
@ -962,13 +1040,13 @@ static u32 UnquantizeTexelWeight(const IntegerEncodedValue& val) {
u32 bitval = val.bit_value; u32 bitval = val.bit_value;
u32 bitlen = val.num_bits; u32 bitlen = val.num_bits;
u32 A = Replicate(bitval & 1, 1, 7); u32 A = ReplicateBitTo7(bitval & 1);
u32 B = 0, C = 0, D = 0; u32 B = 0, C = 0, D = 0;
u32 result = 0; u32 result = 0;
switch (val.encoding) { switch (val.encoding) {
case IntegerEncoding::JustBits: case IntegerEncoding::JustBits:
result = Replicate(bitval, bitlen, 6); result = FastReplicateTo6(bitval, bitlen);
break; break;
case IntegerEncoding::Trit: { case IntegerEncoding::Trit: {
@ -1047,7 +1125,7 @@ static u32 UnquantizeTexelWeight(const IntegerEncodedValue& val) {
return result; return result;
} }
static void UnquantizeTexelWeights(u32 out[2][144], const std::vector<IntegerEncodedValue>& weights, static void UnquantizeTexelWeights(u32 out[2][144], const IntegerEncodedVector& weights,
const TexelWeightParams& params, const u32 blockWidth, const TexelWeightParams& params, const u32 blockWidth,
const u32 blockHeight) { const u32 blockHeight) {
u32 weightIdx = 0; u32 weightIdx = 0;
@ -1545,8 +1623,7 @@ static void DecompressBlock(const u8 inBuf[16], const u32 blockWidth, const u32
static_cast<u8>((1 << (weightParams.GetPackedBitSize() % 8)) - 1); static_cast<u8>((1 << (weightParams.GetPackedBitSize() % 8)) - 1);
memset(texelWeightData + clearByteStart, 0, 16 - clearByteStart); memset(texelWeightData + clearByteStart, 0, 16 - clearByteStart);
std::vector<IntegerEncodedValue> texelWeightValues; IntegerEncodedVector texelWeightValues;
texelWeightValues.reserve(64);
InputBitStream weightStream(texelWeightData); InputBitStream weightStream(texelWeightData);
@ -1568,9 +1645,9 @@ static void DecompressBlock(const u8 inBuf[16], const u32 blockWidth, const u32
Pixel p; Pixel p;
for (u32 c = 0; c < 4; c++) { for (u32 c = 0; c < 4; c++) {
u32 C0 = endpos32s[partition][0].Component(c); u32 C0 = endpos32s[partition][0].Component(c);
C0 = Replicate(C0, 8, 16); C0 = ReplicateByteTo16(C0);
u32 C1 = endpos32s[partition][1].Component(c); u32 C1 = endpos32s[partition][1].Component(c);
C1 = Replicate(C1, 8, 16); C1 = ReplicateByteTo16(C1);
u32 plane = 0; u32 plane = 0;
if (weightParams.m_bDualPlane && (((planeIdx + 1) & 3) == c)) { if (weightParams.m_bDualPlane && (((planeIdx + 1) & 3) == c)) {