diff options
author | Devtools Arcadia <arcadia-devtools@yandex-team.ru> | 2022-02-07 18:08:42 +0300 |
---|---|---|
committer | Devtools Arcadia <arcadia-devtools@mous.vla.yp-c.yandex.net> | 2022-02-07 18:08:42 +0300 |
commit | 1110808a9d39d4b808aef724c861a2e1a38d2a69 (patch) | |
tree | e26c9fed0de5d9873cce7e00bc214573dc2195b7 /library/cpp/codecs/float_huffman.cpp | |
download | ydb-1110808a9d39d4b808aef724c861a2e1a38d2a69.tar.gz |
intermediate changes
ref:cde9a383711a11544ce7e107a78147fb96cc4029
Diffstat (limited to 'library/cpp/codecs/float_huffman.cpp')
-rw-r--r-- | library/cpp/codecs/float_huffman.cpp | 333 |
1 files changed, 333 insertions, 0 deletions
diff --git a/library/cpp/codecs/float_huffman.cpp b/library/cpp/codecs/float_huffman.cpp new file mode 100644 index 0000000000..c4a8bd228f --- /dev/null +++ b/library/cpp/codecs/float_huffman.cpp @@ -0,0 +1,333 @@ +#include "float_huffman.h" + +#include <util/generic/array_ref.h> +#include <util/generic/bitops.h> +#include <util/generic/cast.h> +#include <util/generic/yexception.h> +#include <util/system/unaligned_mem.h> +#include <util/system/yassert.h> +#include <util/stream/format.h> + +namespace NCodecs::NFloatHuff { + namespace { + struct THuffEntry { + ui32 CodeBase; + ui16 Prefix; + int PrefLength; + int CodeLength; + int TotalLength; + ui32 Mask; + ui64 Offset; + + THuffEntry() = default; + + constexpr THuffEntry(ui32 codeBase, ui16 prefix, int prefLength, int codeLength) + : CodeBase(codeBase) + , Prefix(prefix) + , PrefLength(prefLength) + , CodeLength(codeLength) + , TotalLength(prefLength + codeLength) + , Mask(Mask64(codeLength)) + , Offset(-(ui64(codeBase) << prefLength) + prefix) + {} + + bool Fit(ui32 code) const { + return code >= CodeBase && code < CodeBase + (1ULL << CodeLength); + } + }; + + // NB. There is a typo in the penultimate line (34 instead of 24). It was there from the very + // first commit and we cannot fix it without breaking all the clients. + constexpr THuffEntry entries[16] = { + {0x00000000, 0x01, 1, 0}, // Only +0.0f, 1 bit, prefix [1] + {0x3f800000, 0x0e, 4, 0}, // Only +1.0f, 4 bits, prefix [0111] + {0x3f700000, 0x08, 5, 20}, // [0.9375, 1.0), 25 bits, prefix [00010] + {0x3f000000, 0x00, 5, 20}, // [0.5, 0.5625), 25 bits, prefx [00000] + {0x3f400000, 0x06, 6, 20}, // [0.75, 0.8125), 26 bits, prefix [011000] + {0x3f500000, 0x22, 6, 20}, // [0.8125, 0.875), 26 bits, prefix [010001] + {0x3f200000, 0x02, 6, 20}, // [0.625, 0.6875), 26 bits, prefix [010000] + {0x3f100000, 0x38, 6, 20}, // [0.5625, 0.625), 26 bits, prefix [000111] + {0x3f600000, 0x18, 6, 20}, // [0.875, 0.9375), 26 bits, prefix [000110] + {0x3f300000, 0x30, 6, 20}, // [0.6875, 0.75), 26 bits, prefix [000011] + {0x3e800000, 0x10, 6, 20}, // [0.25, 0.28125), 26 bits, prefix [000010] + {0x3e000000, 0x04, 3, 24}, // [0.125, 0.5), 27 bits, prefix [001] + {0x3d000000, 0x0a, 4, 24}, // [0.03125, 0.125), 28 bits, prefix [0101] + {0x3c000000, 0x12, 5, 24}, // [0.0078125, 0.03125), 29 bits, prefix [01001] + {0x3b000000, 0x26, 6, 34}, // [0.001953125, end of range), 40 bits, prefix [011001] + {0x00000000, 0x16, 5, 32}, // whole range, 37 bits, prefix [01101] + }; + + [[noreturn]] Y_NO_INLINE void ThrowInvalidOffset(size_t size, size_t byteOffset) { + ythrow yexception() << + "Decompression error: requested decoding 8 bytes past end of input buffer of " << size << " bytes size at position " << byteOffset << ". "; + } + + struct THuffInfo { + constexpr THuffInfo() { + for (size_t i = 0; i < 64; ++i) { + bool init = false; + for (size_t j = 0; j != 16; ++j) { + ui16 prefix = i & Mask64(entries[j].PrefLength); + if (entries[j].Prefix == prefix) { + init = true; + DecodeLookup[i] = entries[j]; + break; + } + } + Y_ASSERT(init); + } + + for (ui32 i = 0; i < (1 << 12); ++i) { + // First two entries (+0.0f and +1.0f) are not present in the lookup, they are handled separately + for (int value = 2; value < 16; ++value) { + if (entries[value].Fit(i << 20)) { + EncodeLookup[i] = value; + break; + } + } + } + } + + std::pair<ui64, int> GetCode(ui32 value) const { + // Zeros are handled separately in the main loop + Y_ASSERT(value != 0); + + if (value == 0x3f800000) { + return {0x0e, 4}; + } + + const auto& entry = entries[EncodeLookup[value >> 20]]; + + return { + (ui64(value) << entry.PrefLength) + entry.Offset, + entry.TotalLength + }; + } + + THuffEntry DecodeLookup[64]; + ui8 EncodeLookup[1 << 12]; + }; + + const THuffInfo huffInfo; + /// End Of Stream + const ui32 EOS = ui32(-1); + } + + TString Encode(TArrayRef<const float> factors) { + TString result; + result.resize((factors.size() + 1) * 40 / 8 + 8, 0); // Max code length is 40 bits + int usedBits = 0; + ui64 buffer = 0; + char* writePtr = result.begin(); + + auto writeBits = [&](ui64 code, int size) { + const auto bitsToTransfer = Min(size, 64 - usedBits); + buffer |= (code << usedBits); + usedBits += bitsToTransfer; + if (usedBits == 64) { + memcpy(writePtr, &buffer, 8); + usedBits = size - bitsToTransfer; + if (bitsToTransfer != 64) { + buffer = code >> bitsToTransfer; + } else { + buffer = 0; + } + writePtr += 8; + } + }; + + for (size_t i = 0; i != factors.size();) { + if (BitCast<ui32>(factors[i]) == 0) { + int zeroCount = 1; + for (;;) { + ++i; + if (i == factors.size() || BitCast<ui32>(factors[i]) != 0) { + break; + } + ++zeroCount; + } + for (; zeroCount >= 64; zeroCount -= 64) { + writeBits(ui64(-1), 64); + } + writeBits(Mask64(zeroCount), zeroCount); + } else { + const auto [code, codeSize] = huffInfo.GetCode(BitCast<ui32>(factors[i])); + writeBits(code, codeSize); + ++i; + } + } + // Write EOS. + // We use precomputed constants instead of the following: + // auto [code, codeSize] = huffInfo.GetCode(EOS); + // writeBits(code, codeSize); + writeBits(211527139302, 40); + memcpy(writePtr, &buffer, 8); + result.resize(writePtr - result.begin() + usedBits / 8 + 8); + + return result; + } + + TDecoder::TDecoder(TStringBuf data) + : State{ + .Workspace = data.size() < 8 ? ThrowInvalidOffset(data.size(), 0), 0 : ReadUnaligned<ui64>(data.data()), + .WorkspaceSize = 64, + .Position = 8, + .Data = data + } + { + FillDecodeBuffer(); + } + + TVector<float> TDecoder::DecodeAll(size_t sizeHint) { + TVector<float> result; + result.reserve(sizeHint); + + while (Begin != End) { + result.insert(result.end(), Begin, End); + FillDecodeBuffer(); + } + return result; + } + + size_t TDecoder::Decode(TArrayRef<float> dest) { + size_t count = 0; + while (count < dest.size()) { + if (dest.size() - count < size_t(End - Begin)) { + const auto size = dest.size() - count; + std::copy(Begin, Begin + size, dest.data() + count); + Begin += size; + return dest.size(); + } else { + std::copy(Begin, End, dest.data() + count); + count += End - Begin; + FillDecodeBuffer(); + if (Begin == End) { + break; + } + } + } + return count; + } + + size_t TDecoder::Skip(size_t count) { + size_t skippedCount = 0; + while (skippedCount < count) { + if (count - skippedCount < size_t(End - Begin)) { + const auto size = count - skippedCount; + Begin += size; + return count; + } else { + skippedCount += End - Begin; + FillDecodeBuffer(); + if (Begin == End) { + break; + } + } + } + return skippedCount; + } + + bool TDecoder::HasMore() const { + return Begin != End; + } + + void TDecoder::FillDecodeBuffer() { + Begin = DecodeBuffer.data(); + End = DecodeBuffer.data(); + + if (HitEos) { + return; + } + + // This helps to keep most of the variables in the registers. + float* end = End; + TState state = State; + + // It is faster to just zero all the memory here in one go + // and then avoid inner loop when writing zeros. There we + // can just increment end pointer. + std::fill(DecodeBuffer.begin(), DecodeBuffer.end(), 0.0f); + + // Make sure that inside the loop we always have space to put 64 zeros and one other + // value. + float* cap = DecodeBuffer.data() + DecodeBuffer.size() - 64 - 1; + + while (end < cap) { + if (state.Workspace % 2 == 1) { + // Decode zeros + // There we can just scan whole state.Workspace for ones because it contains + // zeros outside of the WorkspaceSize bits. + const auto negWorkspace = ~state.Workspace; + const int zeroCount = negWorkspace ? CountTrailingZeroBits(negWorkspace) : 64; + end += zeroCount; + state.SkipBits(zeroCount); + continue; + } + if (state.PeekBits(4) == 0x0e) { + *end++ = 1.0f; + state.SkipBits(4); + continue; + } + const auto& entry = huffInfo.DecodeLookup[state.PeekBits(6)]; + const auto code = ui32((state.NextBitsUnmasked(entry.TotalLength) >> entry.PrefLength) & entry.Mask) + entry.CodeBase; + if (Y_UNLIKELY(code == EOS)) { + HitEos = true; + break; + } + *end++ = BitCast<float>(code); + } + + End = end; + State = state; + } + + ui64 TDecoder::TState::PeekBits(int count) { + if (WorkspaceSize > count) { + return Workspace & Mask64(count); + } else { + if (Y_UNLIKELY(Position + 8 > Data.size())) { + ThrowInvalidOffset(Data.size(), Position); + } + return (Workspace | (ReadUnaligned<ui64>(Data.data() + Position) << WorkspaceSize)) & Mask64(count); + } + } + + ui64 TDecoder::TState::NextBitsUnmasked(int count) { + if (WorkspaceSize > count) { + const auto result = Workspace; + Workspace >>= count; + WorkspaceSize -= count; + return result; + } else { + if (Y_UNLIKELY(Position + 8 > Data.size())) { + ThrowInvalidOffset(Data.size(), Position); + } + ui64 result = Workspace; + Workspace = ReadUnaligned<ui64>(Data.data() + Position); + Position += 8; + result |= Workspace << WorkspaceSize; + Workspace >>= count - WorkspaceSize; + WorkspaceSize += 64 - count; + return result; + } + } + + void TDecoder::TState::SkipBits(int count) { + if (WorkspaceSize > count) { + Workspace >>= count; + WorkspaceSize -= count; + } else { + if (Y_UNLIKELY(Position + 8 > Data.size())) { + ThrowInvalidOffset(Data.size(), Position); + } + Workspace = ReadUnaligned<ui64>(Data.data() + Position); + Position += 8; + Workspace >>= count - WorkspaceSize; + WorkspaceSize += 64 - count; + } + } + + TVector<float> Decode(TStringBuf data, size_t sizeHint) { + return TDecoder(data).DecodeAll(sizeHint); + } +} // namespace NCodecs::NFloatHuff |