From 2985e5e94c82febcf215feb0023f4184b38bb24a Mon Sep 17 00:00:00 2001 From: ameerj <52414509+ameerj@users.noreply.github.com> Date: Sat, 13 Feb 2021 15:50:12 -0500 Subject: renderer_opengl: Accelerate ASTC texture decoding with a compute shader ASTC texture decoding is currently handled by a CPU decoder for GPU's without native ASTC decoding support (most desktop GPUs). This is the cause for noticeable performance degradation in titles which use the format extensively. This commit adds support to accelerate ASTC decoding using a compute shader on OpenGL for GPUs without native support. --- src/video_core/textures/astc.h | 190 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 190 insertions(+) (limited to 'src/video_core/textures') diff --git a/src/video_core/textures/astc.h b/src/video_core/textures/astc.h index 9105119bc..bc8bddaec 100644 --- a/src/video_core/textures/astc.h +++ b/src/video_core/textures/astc.h @@ -8,6 +8,196 @@ namespace Tegra::Texture::ASTC { +/// Count the number of bits set in a number. +constexpr u32 Popcnt(u32 n) { + u32 c = 0; + for (; n; c++) { + n &= n - 1; + } + return c; +} + +enum class IntegerEncoding { JustBits, Qus32, Trit }; + +struct IntegerEncodedValue { + constexpr IntegerEncodedValue() = default; + + constexpr IntegerEncodedValue(IntegerEncoding encoding_, u32 num_bits_) + : encoding{encoding_}, num_bits{num_bits_} {} + + constexpr bool MatchesEncoding(const IntegerEncodedValue& other) const { + return encoding == other.encoding && num_bits == other.num_bits; + } + + // Returns the number of bits required to encode nVals values. + u32 GetBitLength(u32 nVals) const { + u32 totalBits = num_bits * nVals; + if (encoding == IntegerEncoding::Trit) { + totalBits += (nVals * 8 + 4) / 5; + } else if (encoding == IntegerEncoding::Qus32) { + totalBits += (nVals * 7 + 2) / 3; + } + return totalBits; + } + + IntegerEncoding encoding{}; + u32 num_bits = 0; + u32 bit_value = 0; + union { + u32 qus32_value = 0; + u32 trit_value; + }; +}; + +// Returns a new instance of this struct that corresponds to the +// can take no more than maxval values +static constexpr IntegerEncodedValue CreateEncoding(u32 maxVal) { + while (maxVal > 0) { + u32 check = maxVal + 1; + + // Is maxVal a power of two? + if (!(check & (check - 1))) { + return IntegerEncodedValue(IntegerEncoding::JustBits, Popcnt(maxVal)); + } + + // Is maxVal of the type 3*2^n - 1? + if ((check % 3 == 0) && !((check / 3) & ((check / 3) - 1))) { + return IntegerEncodedValue(IntegerEncoding::Trit, Popcnt(check / 3 - 1)); + } + + // Is maxVal of the type 5*2^n - 1? + if ((check % 5 == 0) && !((check / 5) & ((check / 5) - 1))) { + return IntegerEncodedValue(IntegerEncoding::Qus32, Popcnt(check / 5 - 1)); + } + + // Apparently it can't be represented with a bounded integer sequence... + // just iterate. + maxVal--; + } + return IntegerEncodedValue(IntegerEncoding::JustBits, 0); +} + +static constexpr std::array MakeEncodedValues() { + std::array encodings{}; + for (std::size_t i = 0; i < encodings.size(); ++i) { + encodings[i] = CreateEncoding(static_cast(i)); + } + return encodings; +} + +static constexpr std::array EncodingsValues = MakeEncodedValues(); + +// Replicates low numBits such that [(toBit - 1):(toBit - 1 - fromBit)] +// is the same as [(numBits - 1):0] and repeats all the way down. +template +static constexpr IntType Replicate(IntType val, u32 numBits, u32 toBit) { + if (numBits == 0) { + return 0; + } + if (toBit == 0) { + return 0; + } + const IntType v = val & static_cast((1 << numBits) - 1); + IntType res = v; + u32 reslen = numBits; + while (reslen < toBit) { + u32 comp = 0; + if (numBits > toBit - reslen) { + u32 newshift = toBit - reslen; + comp = numBits - newshift; + numBits = newshift; + } + res = static_cast(res << numBits); + res = static_cast(res | (v >> comp)); + reslen += numBits; + } + return res; +} + +static constexpr std::size_t NumReplicateEntries(u32 num_bits) { + return std::size_t(1) << num_bits; +} + +template +static constexpr auto MakeReplicateTable() { + std::array table{}; + for (IntType value = 0; value < static_cast(std::size(table)); ++value) { + table[value] = Replicate(value, num_bits, to_bit); + } + return table; +} + +static constexpr auto REPLICATE_BYTE_TO_16_TABLE = MakeReplicateTable(); +static constexpr u32 ReplicateByteTo16(std::size_t value) { + return REPLICATE_BYTE_TO_16_TABLE[value]; +} + +static constexpr auto REPLICATE_BIT_TO_7_TABLE = MakeReplicateTable(); +static constexpr u32 ReplicateBitTo7(std::size_t value) { + return REPLICATE_BIT_TO_7_TABLE[value]; +} + +static constexpr auto REPLICATE_BIT_TO_9_TABLE = MakeReplicateTable(); +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(); +static constexpr auto REPLICATE_2_BIT_TO_8_TABLE = MakeReplicateTable(); +static constexpr auto REPLICATE_3_BIT_TO_8_TABLE = MakeReplicateTable(); +static constexpr auto REPLICATE_4_BIT_TO_8_TABLE = MakeReplicateTable(); +static constexpr auto REPLICATE_5_BIT_TO_8_TABLE = MakeReplicateTable(); +static constexpr auto REPLICATE_6_BIT_TO_8_TABLE = MakeReplicateTable(); +static constexpr auto REPLICATE_7_BIT_TO_8_TABLE = MakeReplicateTable(); +static constexpr auto REPLICATE_8_BIT_TO_8_TABLE = MakeReplicateTable(); +/// 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(); +static constexpr auto REPLICATE_2_BIT_TO_6_TABLE = MakeReplicateTable(); +static constexpr auto REPLICATE_3_BIT_TO_6_TABLE = MakeReplicateTable(); +static constexpr auto REPLICATE_4_BIT_TO_6_TABLE = MakeReplicateTable(); +static constexpr auto REPLICATE_5_BIT_TO_6_TABLE = MakeReplicateTable(); + +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); + } +} + void Decompress(std::span data, uint32_t width, uint32_t height, uint32_t depth, uint32_t block_width, uint32_t block_height, std::span output); -- cgit v1.2.3