// Copyright Epic Games, Inc. All Rights Reserved.

#include "chunking.h"

#include <gsl/gsl-lite.hpp>

#include <cmath>

namespace zen::detail {

static const uint32_t BuzhashTable[] = {
	0x458be752, 0xc10748cc, 0xfbbcdbb8, 0x6ded5b68, 0xb10a82b5, 0x20d75648, 0xdfc5665f, 0xa8428801, 0x7ebf5191, 0x841135c7, 0x65cc53b3,
	0x280a597c, 0x16f60255, 0xc78cbc3e, 0x294415f5, 0xb938d494, 0xec85c4e6, 0xb7d33edc, 0xe549b544, 0xfdeda5aa, 0x882bf287, 0x3116737c,
	0x05569956, 0xe8cc1f68, 0x0806ac5e, 0x22a14443, 0x15297e10, 0x50d090e7, 0x4ba60f6f, 0xefd9f1a7, 0x5c5c885c, 0x82482f93, 0x9bfd7c64,
	0x0b3e7276, 0xf2688e77, 0x8fad8abc, 0xb0509568, 0xf1ada29f, 0xa53efdfe, 0xcb2b1d00, 0xf2a9e986, 0x6463432b, 0x95094051, 0x5a223ad2,
	0x9be8401b, 0x61e579cb, 0x1a556a14, 0x5840fdc2, 0x9261ddf6, 0xcde002bb, 0x52432bb0, 0xbf17373e, 0x7b7c222f, 0x2955ed16, 0x9f10ca59,
	0xe840c4c9, 0xccabd806, 0x14543f34, 0x1462417a, 0x0d4a1f9c, 0x087ed925, 0xd7f8f24c, 0x7338c425, 0xcf86c8f5, 0xb19165cd, 0x9891c393,
	0x325384ac, 0x0308459d, 0x86141d7e, 0xc922116a, 0xe2ffa6b6, 0x53f52aed, 0x2cd86197, 0xf5b9f498, 0xbf319c8f, 0xe0411fae, 0x977eb18c,
	0xd8770976, 0x9833466a, 0xc674df7f, 0x8c297d45, 0x8ca48d26, 0xc49ed8e2, 0x7344f874, 0x556f79c7, 0x6b25eaed, 0xa03e2b42, 0xf68f66a4,
	0x8e8b09a2, 0xf2e0e62a, 0x0d3a9806, 0x9729e493, 0x8c72b0fc, 0x160b94f6, 0x450e4d3d, 0x7a320e85, 0xbef8f0e1, 0x21d73653, 0x4e3d977a,
	0x1e7b3929, 0x1cc6c719, 0xbe478d53, 0x8d752809, 0xe6d8c2c6, 0x275f0892, 0xc8acc273, 0x4cc21580, 0xecc4a617, 0xf5f7be70, 0xe795248a,
	0x375a2fe9, 0x425570b6, 0x8898dcf8, 0xdc2d97c4, 0x0106114b, 0x364dc22f, 0x1e0cad1f, 0xbe63803c, 0x5f69fac2, 0x4d5afa6f, 0x1bc0dfb5,
	0xfb273589, 0x0ea47f7b, 0x3c1c2b50, 0x21b2a932, 0x6b1223fd, 0x2fe706a8, 0xf9bd6ce2, 0xa268e64e, 0xe987f486, 0x3eacf563, 0x1ca2018c,
	0x65e18228, 0x2207360a, 0x57cf1715, 0x34c37d2b, 0x1f8f3cde, 0x93b657cf, 0x31a019fd, 0xe69eb729, 0x8bca7b9b, 0x4c9d5bed, 0x277ebeaf,
	0xe0d8f8ae, 0xd150821c, 0x31381871, 0xafc3f1b0, 0x927db328, 0xe95effac, 0x305a47bd, 0x426ba35b, 0x1233af3f, 0x686a5b83, 0x50e072e5,
	0xd9d3bb2a, 0x8befc475, 0x487f0de6, 0xc88dff89, 0xbd664d5e, 0x971b5d18, 0x63b14847, 0xd7d3c1ce, 0x7f583cf3, 0x72cbcb09, 0xc0d0a81c,
	0x7fa3429b, 0xe9158a1b, 0x225ea19a, 0xd8ca9ea3, 0xc763b282, 0xbb0c6341, 0x020b8293, 0xd4cd299d, 0x58cfa7f8, 0x91b4ee53, 0x37e4d140,
	0x95ec764c, 0x30f76b06, 0x5ee68d24, 0x679c8661, 0xa41979c2, 0xf2b61284, 0x4fac1475, 0x0adb49f9, 0x19727a23, 0x15a7e374, 0xc43a18d5,
	0x3fb1aa73, 0x342fc615, 0x924c0793, 0xbee2d7f0, 0x8a279de9, 0x4aa2d70c, 0xe24dd37f, 0xbe862c0b, 0x177c22c2, 0x5388e5ee, 0xcd8a7510,
	0xf901b4fd, 0xdbc13dbc, 0x6c0bae5b, 0x64efe8c7, 0x48b02079, 0x80331a49, 0xca3d8ae6, 0xf3546190, 0xfed7108b, 0xc49b941b, 0x32baf4a9,
	0xeb833a4a, 0x88a3f1a5, 0x3a91ce0a, 0x3cc27da1, 0x7112e684, 0x4a3096b1, 0x3794574c, 0xa3c8b6f3, 0x1d213941, 0x6e0a2e00, 0x233479f1,
	0x0f4cd82f, 0x6093edd2, 0x5d7d209e, 0x464fe319, 0xd4dcac9e, 0x0db845cb, 0xfb5e4bc3, 0xe0256ce1, 0x09fb4ed1, 0x0914be1e, 0xa5bdb2c3,
	0xc6eb57bb, 0x30320350, 0x3f397e91, 0xa67791bc, 0x86bc0e2c, 0xefa0a7e2, 0xe9ff7543, 0xe733612c, 0xd185897b, 0x329e5388, 0x91dd236b,
	0x2ecb0d93, 0xf4d82a3d, 0x35b5c03f, 0xe4e606f0, 0x05b21843, 0x37b45964, 0x5eff22f4, 0x6027f4cc, 0x77178b3c, 0xae507131, 0x7bf7cabc,
	0xf9c18d66, 0x593ade65, 0xd95ddf11,
};

// ROL operation (compiler turns this into a ROL when optimizing)
ZEN_FORCEINLINE static uint32_t
Rotate32(uint32_t Value, size_t RotateCount)
{
	RotateCount &= 31;

	return ((Value) << (RotateCount)) | ((Value) >> (32 - RotateCount));
}

}  // namespace zen::detail

namespace zen {

void
ZenChunkHelper::Reset()
{
	InternalReset();

	m_BytesScanned = 0;
}

void
ZenChunkHelper::InternalReset()
{
	m_CurrentHash	   = 0;
	m_CurrentChunkSize = 0;
	m_WindowSize	   = 0;
}

void
ZenChunkHelper::SetChunkSize(size_t MinSize, size_t MaxSize, size_t AvgSize)
{
	if (m_WindowSize)
		return;	 // Already started

	static_assert(kChunkSizeLimitMin > kWindowSize);

	if (AvgSize)
	{
		// TODO: Validate AvgSize range
	}
	else
	{
		if (MinSize && MaxSize)
		{
			AvgSize = std::lrint(std::pow(2, (std::log2(MinSize) + std::log2(MaxSize)) / 2));
		}
		else if (MinSize)
		{
			AvgSize = MinSize * 4;
		}
		else if (MaxSize)
		{
			AvgSize = MaxSize / 4;
		}
		else
		{
			AvgSize = kDefaultAverageChunkSize;
		}
	}

	if (MinSize)
	{
		// TODO: Validate MinSize range
	}
	else
	{
		MinSize = std::max(AvgSize / 4, kChunkSizeLimitMin);
	}

	if (MaxSize)
	{
		// TODO: Validate MaxSize range
	}
	else
	{
		MaxSize = std::min(AvgSize * 4, kChunkSizeLimitMax);
	}

	m_Discriminator = gsl::narrow<uint32_t>(AvgSize - MinSize);

	if (m_Discriminator < MinSize)
	{
		m_Discriminator = gsl::narrow<uint32_t>(MinSize);
	}

	if (m_Discriminator > MaxSize)
	{
		m_Discriminator = gsl::narrow<uint32_t>(MaxSize);
	}

	m_Threshold = gsl::narrow<uint32_t>((uint64_t(std::numeric_limits<uint32_t>::max()) + 1) / m_Discriminator);

	m_ChunkSizeMin = MinSize;
	m_ChunkSizeMax = MaxSize;
	m_ChunkSizeAvg = AvgSize;
}

size_t
ZenChunkHelper::ScanChunk(const void* DataBytesIn, size_t ByteCount)
{
	size_t Result = InternalScanChunk(DataBytesIn, ByteCount);

	if (Result == kNoBoundaryFound)
	{
		m_BytesScanned += ByteCount;
	}
	else
	{
		m_BytesScanned += Result;
	}

	return Result;
}

size_t
ZenChunkHelper::InternalScanChunk(const void* DataBytesIn, size_t ByteCount)
{
	size_t		   CurrentOffset = 0;
	const uint8_t* CursorPtr	 = reinterpret_cast<const uint8_t*>(DataBytesIn);

	// There's no point in updating the hash if we know we're not
	// going to have a cut point, so just skip the data. This logic currently
	// provides roughly a 20% speedup on my machine

	const size_t NeedHashOffset = m_ChunkSizeMin - kWindowSize;

	if (m_CurrentChunkSize < NeedHashOffset)
	{
		const uint32_t SkipBytes = gsl::narrow<uint32_t>(std::min<uint64_t>(ByteCount, NeedHashOffset - m_CurrentChunkSize));

		ByteCount -= SkipBytes;
		m_CurrentChunkSize += SkipBytes;
		CurrentOffset += SkipBytes;
		CursorPtr += SkipBytes;

		m_WindowSize = 0;

		if (ByteCount == 0)
		{
			return kNoBoundaryFound;
		}
	}

	// Fill window first

	if (m_WindowSize < kWindowSize)
	{
		const uint32_t FillBytes = uint32_t(std::min<size_t>(ByteCount, kWindowSize - m_WindowSize));

		memcpy(&m_Window[m_WindowSize], CursorPtr, FillBytes);

		CursorPtr += FillBytes;

		m_WindowSize += FillBytes;
		m_CurrentChunkSize += FillBytes;

		CurrentOffset += FillBytes;
		ByteCount -= FillBytes;

		if (m_WindowSize < kWindowSize)
		{
			return kNoBoundaryFound;
		}

		// We have a full window, initialize hash

		uint32_t CurrentHash = 0;

		for (int i = 1; i < kWindowSize; ++i)
		{
			CurrentHash ^= detail::Rotate32(detail::BuzhashTable[m_Window[i - 1]], kWindowSize - i);
		}

		m_CurrentHash = CurrentHash ^ detail::BuzhashTable[m_Window[kWindowSize - 1]];
	}

	// Scan for boundaries (i.e points where the hash matches the value determined by
	// the discriminator)

	uint32_t CurrentHash	  = m_CurrentHash;
	uint32_t CurrentChunkSize = m_CurrentChunkSize;

	size_t Index = CurrentChunkSize % kWindowSize;

	if (m_Threshold && m_UseThreshold)
	{
		// This is roughly 4x faster than the general modulo approach on my
		// TR 3990X (~940MB/sec) and doesn't require any special parameters to
		// achieve max performance

		while (ByteCount)
		{
			const uint8_t NewByte = *CursorPtr;
			const uint8_t OldByte = m_Window[Index];

			CurrentHash = detail::Rotate32(CurrentHash, 1) ^ detail::Rotate32(detail::BuzhashTable[OldByte], m_WindowSize) ^
						  detail::BuzhashTable[NewByte];

			CurrentChunkSize++;
			CurrentOffset++;

			if (CurrentChunkSize >= m_ChunkSizeMin)
			{
				bool FoundBoundary;

				if (CurrentChunkSize >= m_ChunkSizeMax)
				{
					FoundBoundary = true;
				}
				else
				{
					FoundBoundary = CurrentHash <= m_Threshold;
				}

				if (FoundBoundary)
				{
					// Boundary found!
					InternalReset();

					return CurrentOffset;
				}
			}

			m_Window[Index++] = *CursorPtr;

			if (Index == kWindowSize)
			{
				Index = 0;
			}

			++CursorPtr;
			--ByteCount;
		}
	}
	else if ((m_Discriminator & (m_Discriminator - 1)) == 0)
	{
		// This is quite a bit faster than the generic modulo path, but
		// requires a very specific average chunk size to be used. If you
		// pass in an even power-of-two divided by 0.75 as the average
		// chunk size you'll hit this path

		const uint32_t Mask = m_Discriminator - 1;

		while (ByteCount)
		{
			const uint8_t NewByte = *CursorPtr;
			const uint8_t OldByte = m_Window[Index];

			CurrentHash = detail::Rotate32(CurrentHash, 1) ^ detail::Rotate32(detail::BuzhashTable[OldByte], m_WindowSize) ^
						  detail::BuzhashTable[NewByte];

			CurrentChunkSize++;
			CurrentOffset++;

			if (CurrentChunkSize >= m_ChunkSizeMin)
			{
				bool FoundBoundary;

				if (CurrentChunkSize >= m_ChunkSizeMax)
				{
					FoundBoundary = true;
				}
				else
				{
					FoundBoundary = (CurrentHash & Mask) == Mask;
				}

				if (FoundBoundary)
				{
					// Boundary found!
					InternalReset();

					return CurrentOffset;
				}
			}

			m_Window[Index++] = *CursorPtr;

			if (Index == kWindowSize)
			{
				Index = 0;
			}

			++CursorPtr;
			--ByteCount;
		}
	}
	else
	{
		// This is the slowest path, which caps out around 250MB/sec for large sizes
		// on my TR3900X

		while (ByteCount)
		{
			const uint8_t NewByte = *CursorPtr;
			const uint8_t OldByte = m_Window[Index];

			CurrentHash = detail::Rotate32(CurrentHash, 1) ^ detail::Rotate32(detail::BuzhashTable[OldByte], m_WindowSize) ^
						  detail::BuzhashTable[NewByte];

			CurrentChunkSize++;
			CurrentOffset++;

			if (CurrentChunkSize >= m_ChunkSizeMin)
			{
				bool FoundBoundary;

				if (CurrentChunkSize >= m_ChunkSizeMax)
				{
					FoundBoundary = true;
				}
				else
				{
					FoundBoundary = (CurrentHash % m_Discriminator) == (m_Discriminator - 1);
				}

				if (FoundBoundary)
				{
					// Boundary found!
					InternalReset();

					return CurrentOffset;
				}
			}

			m_Window[Index++] = *CursorPtr;

			if (Index == kWindowSize)
			{
				Index = 0;
			}

			++CursorPtr;
			--ByteCount;
		}
	}

	m_CurrentChunkSize = CurrentChunkSize;
	m_CurrentHash	   = CurrentHash;

	return kNoBoundaryFound;
}

}  // namespace zen