diff options
Diffstat (limited to 'thirdparty/etcpak')
43 files changed, 9382 insertions, 0 deletions
diff --git a/thirdparty/etcpak/AUTHORS.txt b/thirdparty/etcpak/AUTHORS.txt new file mode 100644 index 0000000000..e7bae62c85 --- /dev/null +++ b/thirdparty/etcpak/AUTHORS.txt @@ -0,0 +1,3 @@ +Bartosz Taudul <wolf@nereid.pl> +Daniel Jungmann <el.3d.source@gmail.com> +Florian Penzkofer <fp@nullptr.de> diff --git a/thirdparty/etcpak/Bitmap.cpp b/thirdparty/etcpak/Bitmap.cpp new file mode 100644 index 0000000000..ef318318ac --- /dev/null +++ b/thirdparty/etcpak/Bitmap.cpp @@ -0,0 +1,216 @@ +#include <ctype.h> +#include <stdlib.h> +#include <string.h> +#include <assert.h> + +#include <png.h> +#include "lz4/lz4.h" + +#include "Bitmap.hpp" +#include "Debug.hpp" + +Bitmap::Bitmap( const char* fn, unsigned int lines, bool bgr ) + : m_block( nullptr ) + , m_lines( lines ) + , m_alpha( true ) + , m_sema( 0 ) +{ + FILE* f = fopen( fn, "rb" ); + assert( f ); + + char buf[4]; + fread( buf, 1, 4, f ); + if( memcmp( buf, "raw4", 4 ) == 0 ) + { + uint8_t a; + fread( &a, 1, 1, f ); + m_alpha = a == 1; + uint32_t d; + fread( &d, 1, 4, f ); + m_size.x = d; + fread( &d, 1, 4, f ); + m_size.y = d; + DBGPRINT( "Raw bitmap " << fn << " " << m_size.x << "x" << m_size.y ); + + assert( m_size.x % 4 == 0 ); + assert( m_size.y % 4 == 0 ); + + int32_t csize; + fread( &csize, 1, 4, f ); + char* cbuf = new char[csize]; + fread( cbuf, 1, csize, f ); + fclose( f ); + + m_block = m_data = new uint32_t[m_size.x*m_size.y]; + m_linesLeft = m_size.y / 4; + + LZ4_decompress_fast( cbuf, (char*)m_data, m_size.x*m_size.y*4 ); + delete[] cbuf; + + for( int i=0; i<m_size.y/4; i++ ) + { + m_sema.unlock(); + } + } + else + { + fseek( f, 0, SEEK_SET ); + + unsigned int sig_read = 0; + int bit_depth, color_type, interlace_type; + + png_structp png_ptr = png_create_read_struct( PNG_LIBPNG_VER_STRING, NULL, NULL, NULL ); + png_infop info_ptr = png_create_info_struct( png_ptr ); + setjmp( png_jmpbuf( png_ptr ) ); + + png_init_io( png_ptr, f ); + png_set_sig_bytes( png_ptr, sig_read ); + + png_uint_32 w, h; + + png_read_info( png_ptr, info_ptr ); + png_get_IHDR( png_ptr, info_ptr, &w, &h, &bit_depth, &color_type, &interlace_type, NULL, NULL ); + + m_size = v2i( w, h ); + + png_set_strip_16( png_ptr ); + if( color_type == PNG_COLOR_TYPE_PALETTE ) + { + png_set_palette_to_rgb( png_ptr ); + } + else if( color_type == PNG_COLOR_TYPE_GRAY && bit_depth < 8 ) + { + png_set_expand_gray_1_2_4_to_8( png_ptr ); + } + if( png_get_valid( png_ptr, info_ptr, PNG_INFO_tRNS ) ) + { + png_set_tRNS_to_alpha( png_ptr ); + } + if( color_type == PNG_COLOR_TYPE_GRAY_ALPHA ) + { + png_set_gray_to_rgb(png_ptr); + } + if( bgr ) + { + png_set_bgr(png_ptr); + } + + switch( color_type ) + { + case PNG_COLOR_TYPE_PALETTE: + if( !png_get_valid( png_ptr, info_ptr, PNG_INFO_tRNS ) ) + { + png_set_filler( png_ptr, 0xff, PNG_FILLER_AFTER ); + m_alpha = false; + } + break; + case PNG_COLOR_TYPE_GRAY_ALPHA: + png_set_gray_to_rgb( png_ptr ); + break; + case PNG_COLOR_TYPE_RGB: + png_set_filler( png_ptr, 0xff, PNG_FILLER_AFTER ); + m_alpha = false; + break; + default: + break; + } + + DBGPRINT( "Bitmap " << fn << " " << w << "x" << h ); + + assert( w % 4 == 0 ); + assert( h % 4 == 0 ); + + m_block = m_data = new uint32_t[w*h]; + m_linesLeft = h / 4; + + m_load = std::async( std::launch::async, [this, f, png_ptr, info_ptr]() mutable + { + auto ptr = m_data; + unsigned int lines = 0; + for( int i=0; i<m_size.y / 4; i++ ) + { + for( int j=0; j<4; j++ ) + { + png_read_rows( png_ptr, (png_bytepp)&ptr, NULL, 1 ); + ptr += m_size.x; + } + lines++; + if( lines >= m_lines ) + { + lines = 0; + m_sema.unlock(); + } + } + + if( lines != 0 ) + { + m_sema.unlock(); + } + + png_read_end( png_ptr, info_ptr ); + png_destroy_read_struct( &png_ptr, &info_ptr, NULL ); + fclose( f ); + } ); + } +} + +Bitmap::Bitmap( const v2i& size ) + : m_data( new uint32_t[size.x*size.y] ) + , m_block( nullptr ) + , m_lines( 1 ) + , m_linesLeft( size.y / 4 ) + , m_size( size ) + , m_sema( 0 ) +{ +} + +Bitmap::Bitmap( const Bitmap& src, unsigned int lines ) + : m_lines( lines ) + , m_alpha( src.Alpha() ) + , m_sema( 0 ) +{ +} + +Bitmap::~Bitmap() +{ + delete[] m_data; +} + +void Bitmap::Write( const char* fn ) +{ + FILE* f = fopen( fn, "wb" ); + assert( f ); + + png_structp png_ptr = png_create_write_struct( PNG_LIBPNG_VER_STRING, NULL, NULL, NULL ); + png_infop info_ptr = png_create_info_struct( png_ptr ); + setjmp( png_jmpbuf( png_ptr ) ); + png_init_io( png_ptr, f ); + + png_set_IHDR( png_ptr, info_ptr, m_size.x, m_size.y, 8, PNG_COLOR_TYPE_RGB_ALPHA, PNG_INTERLACE_NONE, PNG_COMPRESSION_TYPE_BASE, PNG_FILTER_TYPE_BASE ); + + png_write_info( png_ptr, info_ptr ); + + uint32_t* ptr = m_data; + for( int i=0; i<m_size.y; i++ ) + { + png_write_rows( png_ptr, (png_bytepp)(&ptr), 1 ); + ptr += m_size.x; + } + + png_write_end( png_ptr, info_ptr ); + png_destroy_write_struct( &png_ptr, &info_ptr ); + + fclose( f ); +} + +const uint32_t* Bitmap::NextBlock( unsigned int& lines, bool& done ) +{ + std::lock_guard<std::mutex> lock( m_lock ); + lines = std::min( m_lines, m_linesLeft ); + auto ret = m_block; + m_sema.lock(); + m_block += m_size.x * 4 * lines; + m_linesLeft -= lines; + done = m_linesLeft == 0; + return ret; +} diff --git a/thirdparty/etcpak/Bitmap.hpp b/thirdparty/etcpak/Bitmap.hpp new file mode 100644 index 0000000000..fae8c936ed --- /dev/null +++ b/thirdparty/etcpak/Bitmap.hpp @@ -0,0 +1,50 @@ +#ifndef __DARKRL__BITMAP_HPP__ +#define __DARKRL__BITMAP_HPP__ + +#include <future> +#include <memory> +#include <mutex> +#include <stdint.h> + +#include "Semaphore.hpp" +#include "Vector.hpp" + +enum class Channels +{ + RGB, + Alpha +}; + +class Bitmap +{ +public: + Bitmap( const char* fn, unsigned int lines, bool bgr ); + Bitmap( const v2i& size ); + virtual ~Bitmap(); + + void Write( const char* fn ); + + uint32_t* Data() { if( m_load.valid() ) m_load.wait(); return m_data; } + const uint32_t* Data() const { if( m_load.valid() ) m_load.wait(); return m_data; } + const v2i& Size() const { return m_size; } + bool Alpha() const { return m_alpha; } + + const uint32_t* NextBlock( unsigned int& lines, bool& done ); + +protected: + Bitmap( const Bitmap& src, unsigned int lines ); + + uint32_t* m_data; + uint32_t* m_block; + unsigned int m_lines; + unsigned int m_linesLeft; + v2i m_size; + bool m_alpha; + Semaphore m_sema; + std::mutex m_lock; + std::future<void> m_load; +}; + +typedef std::shared_ptr<Bitmap> BitmapPtr; + +#endif diff --git a/thirdparty/etcpak/BitmapDownsampled.cpp b/thirdparty/etcpak/BitmapDownsampled.cpp new file mode 100644 index 0000000000..0eb0d81185 --- /dev/null +++ b/thirdparty/etcpak/BitmapDownsampled.cpp @@ -0,0 +1,86 @@ +#include <string.h> +#include <utility> + +#include "BitmapDownsampled.hpp" +#include "Debug.hpp" + +BitmapDownsampled::BitmapDownsampled( const Bitmap& bmp, unsigned int lines ) + : Bitmap( bmp, lines ) +{ + m_size.x = std::max( 1, bmp.Size().x / 2 ); + m_size.y = std::max( 1, bmp.Size().y / 2 ); + + int w = std::max( m_size.x, 4 ); + int h = std::max( m_size.y, 4 ); + + DBGPRINT( "Subbitmap " << m_size.x << "x" << m_size.y ); + + m_block = m_data = new uint32_t[w*h]; + + if( m_size.x < w || m_size.y < h ) + { + memset( m_data, 0, w*h*sizeof( uint32_t ) ); + m_linesLeft = h / 4; + unsigned int lines = 0; + for( int i=0; i<h/4; i++ ) + { + for( int j=0; j<4; j++ ) + { + lines++; + if( lines > m_lines ) + { + lines = 0; + m_sema.unlock(); + } + } + } + if( lines != 0 ) + { + m_sema.unlock(); + } + } + else + { + m_linesLeft = h / 4; + m_load = std::async( std::launch::async, [this, &bmp, w, h]() mutable + { + auto ptr = m_data; + auto src1 = bmp.Data(); + auto src2 = src1 + bmp.Size().x; + unsigned int lines = 0; + for( int i=0; i<h/4; i++ ) + { + for( int j=0; j<4; j++ ) + { + for( int k=0; k<m_size.x; k++ ) + { + int r = ( ( *src1 & 0x000000FF ) + ( *(src1+1) & 0x000000FF ) + ( *src2 & 0x000000FF ) + ( *(src2+1) & 0x000000FF ) ) / 4; + int g = ( ( ( *src1 & 0x0000FF00 ) + ( *(src1+1) & 0x0000FF00 ) + ( *src2 & 0x0000FF00 ) + ( *(src2+1) & 0x0000FF00 ) ) / 4 ) & 0x0000FF00; + int b = ( ( ( *src1 & 0x00FF0000 ) + ( *(src1+1) & 0x00FF0000 ) + ( *src2 & 0x00FF0000 ) + ( *(src2+1) & 0x00FF0000 ) ) / 4 ) & 0x00FF0000; + int a = ( ( ( ( ( *src1 & 0xFF000000 ) >> 8 ) + ( ( *(src1+1) & 0xFF000000 ) >> 8 ) + ( ( *src2 & 0xFF000000 ) >> 8 ) + ( ( *(src2+1) & 0xFF000000 ) >> 8 ) ) / 4 ) & 0x00FF0000 ) << 8; + *ptr++ = r | g | b | a; + src1 += 2; + src2 += 2; + } + src1 += m_size.x * 2; + src2 += m_size.x * 2; + } + lines++; + if( lines >= m_lines ) + { + lines = 0; + m_sema.unlock(); + } + } + + if( lines != 0 ) + { + m_sema.unlock(); + } + } ); + } +} + +BitmapDownsampled::~BitmapDownsampled() +{ +} diff --git a/thirdparty/etcpak/BitmapDownsampled.hpp b/thirdparty/etcpak/BitmapDownsampled.hpp new file mode 100644 index 0000000000..b7313808df --- /dev/null +++ b/thirdparty/etcpak/BitmapDownsampled.hpp @@ -0,0 +1,13 @@ +#ifndef __DARKRL__BITMAPDOWNSAMPLED_HPP__ +#define __DARKRL__BITMAPDOWNSAMPLED_HPP__ + +#include "Bitmap.hpp" + +class BitmapDownsampled : public Bitmap +{ +public: + BitmapDownsampled( const Bitmap& bmp, unsigned int lines ); + ~BitmapDownsampled(); +}; + +#endif diff --git a/thirdparty/etcpak/BlockData.cpp b/thirdparty/etcpak/BlockData.cpp new file mode 100644 index 0000000000..4906e69492 --- /dev/null +++ b/thirdparty/etcpak/BlockData.cpp @@ -0,0 +1,1296 @@ +#include <assert.h> +#include <string.h> + +#include "BlockData.hpp" +#include "ColorSpace.hpp" +#include "Debug.hpp" +#include "MipMap.hpp" +#include "mmap.hpp" +#include "ProcessRGB.hpp" +#include "ProcessDxtc.hpp" +#include "Tables.hpp" +#include "TaskDispatch.hpp" + +#ifdef __ARM_NEON +# include <arm_neon.h> +#endif + +#if defined __SSE4_1__ || defined __AVX2__ || defined _MSC_VER +# ifdef _MSC_VER +# include <intrin.h> +# include <Windows.h> +# define _bswap(x) _byteswap_ulong(x) +# define _bswap64(x) _byteswap_uint64(x) +# else +# include <x86intrin.h> +# endif +#endif + +#ifndef _bswap +# define _bswap(x) __builtin_bswap32(x) +# define _bswap64(x) __builtin_bswap64(x) +#endif + +static uint8_t table59T58H[8] = { 3,6,11,16,23,32,41,64 }; + +BlockData::BlockData( const char* fn ) + : m_file( fopen( fn, "rb" ) ) +{ + assert( m_file ); + fseek( m_file, 0, SEEK_END ); + m_maplen = ftell( m_file ); + fseek( m_file, 0, SEEK_SET ); + m_data = (uint8_t*)mmap( nullptr, m_maplen, PROT_READ, MAP_SHARED, fileno( m_file ), 0 ); + + auto data32 = (uint32_t*)m_data; + if( *data32 == 0x03525650 ) + { + // PVR + switch( *(data32+2) ) + { + case 6: + m_type = Etc1; + break; + case 7: + m_type = Dxt1; + break; + case 11: + m_type = Dxt5; + break; + case 22: + m_type = Etc2_RGB; + break; + case 23: + m_type = Etc2_RGBA; + break; + default: + assert( false ); + break; + } + + m_size.y = *(data32+6); + m_size.x = *(data32+7); + m_dataOffset = 52 + *(data32+12); + } + else if( *data32 == 0x58544BAB ) + { + // KTX + switch( *(data32+7) ) + { + case 0x9274: + m_type = Etc2_RGB; + break; + case 0x9278: + m_type = Etc2_RGBA; + break; + default: + assert( false ); + break; + } + + m_size.x = *(data32+9); + m_size.y = *(data32+10); + m_dataOffset = sizeof( uint32_t ) * 17 + *(data32+15); + } + else + { + assert( false ); + } +} + +static uint8_t* OpenForWriting( const char* fn, size_t len, const v2i& size, FILE** f, int levels, BlockData::Type type ) +{ + *f = fopen( fn, "wb+" ); + assert( *f ); + fseek( *f, len - 1, SEEK_SET ); + const char zero = 0; + fwrite( &zero, 1, 1, *f ); + fseek( *f, 0, SEEK_SET ); + + auto ret = (uint8_t*)mmap( nullptr, len, PROT_WRITE, MAP_SHARED, fileno( *f ), 0 ); + auto dst = (uint32_t*)ret; + + *dst++ = 0x03525650; // version + *dst++ = 0; // flags + switch( type ) // pixelformat[0] + { + case BlockData::Etc1: + *dst++ = 6; + break; + case BlockData::Etc2_RGB: + *dst++ = 22; + break; + case BlockData::Etc2_RGBA: + *dst++ = 23; + break; + case BlockData::Dxt1: + *dst++ = 7; + break; + case BlockData::Dxt5: + *dst++ = 11; + break; + default: + assert( false ); + break; + } + *dst++ = 0; // pixelformat[1] + *dst++ = 0; // colourspace + *dst++ = 0; // channel type + *dst++ = size.y; // height + *dst++ = size.x; // width + *dst++ = 1; // depth + *dst++ = 1; // num surfs + *dst++ = 1; // num faces + *dst++ = levels; // mipmap count + *dst++ = 0; // metadata size + + return ret; +} + +static int AdjustSizeForMipmaps( const v2i& size, int levels ) +{ + int len = 0; + v2i current = size; + for( int i=1; i<levels; i++ ) + { + assert( current.x != 1 || current.y != 1 ); + current.x = std::max( 1, current.x / 2 ); + current.y = std::max( 1, current.y / 2 ); + len += std::max( 4, current.x ) * std::max( 4, current.y ) / 2; + } + assert( current.x == 1 && current.y == 1 ); + return len; +} + +BlockData::BlockData( const char* fn, const v2i& size, bool mipmap, Type type ) + : m_size( size ) + , m_dataOffset( 52 ) + , m_maplen( m_size.x*m_size.y/2 ) + , m_type( type ) +{ + assert( m_size.x%4 == 0 && m_size.y%4 == 0 ); + + uint32_t cnt = m_size.x * m_size.y / 16; + DBGPRINT( cnt << " blocks" ); + + int levels = 1; + + if( mipmap ) + { + levels = NumberOfMipLevels( size ); + DBGPRINT( "Number of mipmaps: " << levels ); + m_maplen += AdjustSizeForMipmaps( size, levels ); + } + + if( type == Etc2_RGBA || type == Dxt5 ) m_maplen *= 2; + + m_maplen += m_dataOffset; + m_data = OpenForWriting( fn, m_maplen, m_size, &m_file, levels, type ); +} + +BlockData::BlockData( const v2i& size, bool mipmap, Type type ) + : m_size( size ) + , m_dataOffset( 52 ) + , m_file( nullptr ) + , m_maplen( m_size.x*m_size.y/2 ) + , m_type( type ) +{ + assert( m_size.x%4 == 0 && m_size.y%4 == 0 ); + if( mipmap ) + { + const int levels = NumberOfMipLevels( size ); + m_maplen += AdjustSizeForMipmaps( size, levels ); + } + + if( type == Etc2_RGBA || type == Dxt5 ) m_maplen *= 2; + + m_maplen += m_dataOffset; + m_data = new uint8_t[m_maplen]; +} + +BlockData::~BlockData() +{ + if( m_file ) + { + munmap( m_data, m_maplen ); + fclose( m_file ); + } + else + { + delete[] m_data; + } +} + +void BlockData::Process( const uint32_t* src, uint32_t blocks, size_t offset, size_t width, Channels type, bool dither ) +{ + auto dst = ((uint64_t*)( m_data + m_dataOffset )) + offset; + + if( type == Channels::Alpha ) + { + if( m_type != Etc1 ) + { + CompressEtc2Alpha( src, dst, blocks, width ); + } + else + { + CompressEtc1Alpha( src, dst, blocks, width ); + } + } + else + { + switch( m_type ) + { + case Etc1: + if( dither ) + { + CompressEtc1RgbDither( src, dst, blocks, width ); + } + else + { + CompressEtc1Rgb( src, dst, blocks, width ); + } + break; + case Etc2_RGB: + CompressEtc2Rgb( src, dst, blocks, width ); + break; + case Dxt1: + if( dither ) + { + CompressDxt1Dither( src, dst, blocks, width ); + } + else + { + CompressDxt1( src, dst, blocks, width ); + } + break; + default: + assert( false ); + break; + } + } +} + +void BlockData::ProcessRGBA( const uint32_t* src, uint32_t blocks, size_t offset, size_t width ) +{ + auto dst = ((uint64_t*)( m_data + m_dataOffset )) + offset * 2; + + switch( m_type ) + { + case Etc2_RGBA: + CompressEtc2Rgba( src, dst, blocks, width ); + break; + case Dxt5: + CompressDxt5( src, dst, blocks, width ); + break; + default: + assert( false ); + break; + } +} + +namespace +{ + +static etcpak_force_inline int32_t expand6(uint32_t value) +{ + return (value << 2) | (value >> 4); +} + +static etcpak_force_inline int32_t expand7(uint32_t value) +{ + return (value << 1) | (value >> 6); +} + +static etcpak_force_inline void DecodeT( uint64_t block, uint32_t* dst, uint32_t w ) +{ + const auto r0 = ( block >> 24 ) & 0x1B; + const auto rh0 = ( r0 >> 3 ) & 0x3; + const auto rl0 = r0 & 0x3; + const auto g0 = ( block >> 20 ) & 0xF; + const auto b0 = ( block >> 16 ) & 0xF; + + const auto r1 = ( block >> 12 ) & 0xF; + const auto g1 = ( block >> 8 ) & 0xF; + const auto b1 = ( block >> 4 ) & 0xF; + + const auto cr0 = ( ( rh0 << 6 ) | ( rl0 << 4 ) | ( rh0 << 2 ) | rl0); + const auto cg0 = ( g0 << 4 ) | g0; + const auto cb0 = ( b0 << 4 ) | b0; + + const auto cr1 = ( r1 << 4 ) | r1; + const auto cg1 = ( g1 << 4 ) | g1; + const auto cb1 = ( b1 << 4 ) | b1; + + const auto codeword_hi = ( block >> 2 ) & 0x3; + const auto codeword_lo = block & 0x1; + const auto codeword = ( codeword_hi << 1 ) | codeword_lo; + + const auto c2r = clampu8( cr1 + table59T58H[codeword] ); + const auto c2g = clampu8( cg1 + table59T58H[codeword] ); + const auto c2b = clampu8( cb1 + table59T58H[codeword] ); + + const auto c3r = clampu8( cr1 - table59T58H[codeword] ); + const auto c3g = clampu8( cg1 - table59T58H[codeword] ); + const auto c3b = clampu8( cb1 - table59T58H[codeword] ); + + const uint32_t col_tab[4] = { + uint32_t(cr0 | ( cg0 << 8 ) | ( cb0 << 16 ) | 0xFF000000), + uint32_t(c2r | ( c2g << 8 ) | ( c2b << 16 ) | 0xFF000000), + uint32_t(cr1 | ( cg1 << 8 ) | ( cb1 << 16 ) | 0xFF000000), + uint32_t(c3r | ( c3g << 8 ) | ( c3b << 16 ) | 0xFF000000) + }; + + const uint32_t indexes = ( block >> 32 ) & 0xFFFFFFFF; + for( uint8_t j = 0; j < 4; j++ ) + { + for( uint8_t i = 0; i < 4; i++ ) + { + //2bit indices distributed on two lane 16bit numbers + const uint8_t index = ( ( ( indexes >> ( j + i * 4 + 16 ) ) & 0x1 ) << 1) | ( ( indexes >> ( j + i * 4 ) ) & 0x1); + dst[j * w + i] = col_tab[index]; + } + } +} + +static etcpak_force_inline void DecodeTAlpha( uint64_t block, uint64_t alpha, uint32_t* dst, uint32_t w ) +{ + const auto r0 = ( block >> 24 ) & 0x1B; + const auto rh0 = ( r0 >> 3 ) & 0x3; + const auto rl0 = r0 & 0x3; + const auto g0 = ( block >> 20 ) & 0xF; + const auto b0 = ( block >> 16 ) & 0xF; + + const auto r1 = ( block >> 12 ) & 0xF; + const auto g1 = ( block >> 8 ) & 0xF; + const auto b1 = ( block >> 4 ) & 0xF; + + const auto cr0 = ( ( rh0 << 6 ) | ( rl0 << 4 ) | ( rh0 << 2 ) | rl0); + const auto cg0 = ( g0 << 4 ) | g0; + const auto cb0 = ( b0 << 4 ) | b0; + + const auto cr1 = ( r1 << 4 ) | r1; + const auto cg1 = ( g1 << 4 ) | g1; + const auto cb1 = ( b1 << 4 ) | b1; + + const auto codeword_hi = ( block >> 2 ) & 0x3; + const auto codeword_lo = block & 0x1; + const auto codeword = (codeword_hi << 1) | codeword_lo; + + const int32_t base = alpha >> 56; + const int32_t mul = ( alpha >> 52 ) & 0xF; + const auto tbl = g_alpha[( alpha >> 48 ) & 0xF]; + + const auto c2r = clampu8( cr1 + table59T58H[codeword] ); + const auto c2g = clampu8( cg1 + table59T58H[codeword] ); + const auto c2b = clampu8( cb1 + table59T58H[codeword] ); + + const auto c3r = clampu8( cr1 - table59T58H[codeword] ); + const auto c3g = clampu8( cg1 - table59T58H[codeword] ); + const auto c3b = clampu8( cb1 - table59T58H[codeword] ); + + const uint32_t col_tab[4] = { + uint32_t(cr0 | ( cg0 << 8 ) | ( cb0 << 16 )), + uint32_t(c2r | ( c2g << 8 ) | ( c2b << 16 )), + uint32_t(cr1 | ( cg1 << 8 ) | ( cb1 << 16 )), + uint32_t(c3r | ( c3g << 8 ) | ( c3b << 16 )) + }; + + const uint32_t indexes = ( block >> 32 ) & 0xFFFFFFFF; + for( uint8_t j = 0; j < 4; j++ ) + { + for( uint8_t i = 0; i < 4; i++ ) + { + //2bit indices distributed on two lane 16bit numbers + const uint8_t index = ( ( ( indexes >> ( j + i * 4 + 16 ) ) & 0x1 ) << 1 ) | ( ( indexes >> ( j + i * 4 ) ) & 0x1 ); + const auto amod = tbl[( alpha >> ( 45 - j * 3 - i * 12 ) ) & 0x7]; + const uint32_t a = clampu8( base + amod * mul ); + dst[j * w + i] = col_tab[index] | ( a << 24 ); + } + } +} + +static etcpak_force_inline void DecodeH( uint64_t block, uint32_t* dst, uint32_t w ) +{ + const uint32_t indexes = ( block >> 32 ) & 0xFFFFFFFF; + + const auto r0444 = ( block >> 27 ) & 0xF; + const auto g0444 = ( ( block >> 20 ) & 0x1 ) | ( ( ( block >> 24 ) & 0x7 ) << 1 ); + const auto b0444 = ( ( block >> 15 ) & 0x7 ) | ( ( ( block >> 19 ) & 0x1 ) << 3 ); + + const auto r1444 = ( block >> 11 ) & 0xF; + const auto g1444 = ( block >> 7 ) & 0xF; + const auto b1444 = ( block >> 3 ) & 0xF; + + const auto r0 = ( r0444 << 4 ) | r0444; + const auto g0 = ( g0444 << 4 ) | g0444; + const auto b0 = ( b0444 << 4 ) | b0444; + + const auto r1 = ( r1444 << 4 ) | r1444; + const auto g1 = ( g1444 << 4 ) | g1444; + const auto b1 = ( b1444 << 4 ) | b1444; + + const auto codeword_hi = ( ( block & 0x1 ) << 1 ) | ( ( block & 0x4 ) ); + const auto c0 = ( r0444 << 8 ) | ( g0444 << 4 ) | ( b0444 << 0 ); + const auto c1 = ( block >> 3 ) & ( ( 1 << 12 ) - 1 ); + const auto codeword_lo = ( c0 >= c1 ) ? 1 : 0; + const auto codeword = codeword_hi | codeword_lo; + + const uint32_t col_tab[] = { + uint32_t(clampu8( r0 + table59T58H[codeword] ) | ( clampu8( g0 + table59T58H[codeword] ) << 8 ) | ( clampu8( b0 + table59T58H[codeword] ) << 16 )), + uint32_t(clampu8( r0 - table59T58H[codeword] ) | ( clampu8( g0 - table59T58H[codeword] ) << 8 ) | ( clampu8( b0 - table59T58H[codeword] ) << 16 )), + uint32_t(clampu8( r1 + table59T58H[codeword] ) | ( clampu8( g1 + table59T58H[codeword] ) << 8 ) | ( clampu8( b1 + table59T58H[codeword] ) << 16 )), + uint32_t(clampu8( r1 - table59T58H[codeword] ) | ( clampu8( g1 - table59T58H[codeword] ) << 8 ) | ( clampu8( b1 - table59T58H[codeword] ) << 16 )) + }; + + for( uint8_t j = 0; j < 4; j++ ) + { + for( uint8_t i = 0; i < 4; i++ ) + { + const uint8_t index = ( ( ( indexes >> ( j + i * 4 + 16 ) ) & 0x1 ) << 1 ) | ( ( indexes >> ( j + i * 4 ) ) & 0x1 ); + dst[j * w + i] = col_tab[index] | 0xFF000000; + } + } +} + +static etcpak_force_inline void DecodeHAlpha( uint64_t block, uint64_t alpha, uint32_t* dst, uint32_t w ) +{ + const uint32_t indexes = ( block >> 32 ) & 0xFFFFFFFF; + + const auto r0444 = ( block >> 27 ) & 0xF; + const auto g0444 = ( ( block >> 20 ) & 0x1 ) | ( ( ( block >> 24 ) & 0x7 ) << 1 ); + const auto b0444 = ( ( block >> 15 ) & 0x7 ) | ( ( ( block >> 19 ) & 0x1 ) << 3 ); + + const auto r1444 = ( block >> 11 ) & 0xF; + const auto g1444 = ( block >> 7 ) & 0xF; + const auto b1444 = ( block >> 3 ) & 0xF; + + const auto r0 = ( r0444 << 4 ) | r0444; + const auto g0 = ( g0444 << 4 ) | g0444; + const auto b0 = ( b0444 << 4 ) | b0444; + + const auto r1 = ( r1444 << 4 ) | r1444; + const auto g1 = ( g1444 << 4 ) | g1444; + const auto b1 = ( b1444 << 4 ) | b1444; + + const auto codeword_hi = ( ( block & 0x1 ) << 1 ) | ( ( block & 0x4 ) ); + const auto c0 = ( r0444 << 8 ) | ( g0444 << 4 ) | ( b0444 << 0 ); + const auto c1 = ( block >> 3 ) & ( ( 1 << 12 ) - 1 ); + const auto codeword_lo = ( c0 >= c1 ) ? 1 : 0; + const auto codeword = codeword_hi | codeword_lo; + + const int32_t base = alpha >> 56; + const int32_t mul = ( alpha >> 52 ) & 0xF; + const auto tbl = g_alpha[(alpha >> 48) & 0xF]; + + const uint32_t col_tab[] = { + uint32_t(clampu8( r0 + table59T58H[codeword] ) | ( clampu8( g0 + table59T58H[codeword] ) << 8 ) | ( clampu8( b0 + table59T58H[codeword] ) << 16 )), + uint32_t(clampu8( r0 - table59T58H[codeword] ) | ( clampu8( g0 - table59T58H[codeword] ) << 8 ) | ( clampu8( b0 - table59T58H[codeword] ) << 16 )), + uint32_t(clampu8( r1 + table59T58H[codeword] ) | ( clampu8( g1 + table59T58H[codeword] ) << 8 ) | ( clampu8( b1 + table59T58H[codeword] ) << 16 )), + uint32_t(clampu8( r1 - table59T58H[codeword] ) | ( clampu8( g1 - table59T58H[codeword] ) << 8 ) | ( clampu8( b1 - table59T58H[codeword] ) << 16 )) + }; + + for( uint8_t j = 0; j < 4; j++ ) + { + for( uint8_t i = 0; i < 4; i++ ) + { + const uint8_t index = ( ( ( indexes >> ( j + i * 4 + 16 ) ) & 0x1 ) << 1 ) | ( ( indexes >> ( j + i * 4 ) ) & 0x1 ); + const auto amod = tbl[( alpha >> ( 45 - j * 3 - i * 12) ) & 0x7]; + const uint32_t a = clampu8( base + amod * mul ); + dst[j * w + i] = col_tab[index] | ( a << 24 ); + } + } +} + +static etcpak_force_inline void DecodePlanar( uint64_t block, uint32_t* dst, uint32_t w ) +{ + const auto bv = expand6((block >> ( 0 + 32)) & 0x3F); + const auto gv = expand7((block >> ( 6 + 32)) & 0x7F); + const auto rv = expand6((block >> (13 + 32)) & 0x3F); + + const auto bh = expand6((block >> (19 + 32)) & 0x3F); + const auto gh = expand7((block >> (25 + 32)) & 0x7F); + + const auto rh0 = (block >> (32 - 32)) & 0x01; + const auto rh1 = ((block >> (34 - 32)) & 0x1F) << 1; + const auto rh = expand6(rh0 | rh1); + + const auto bo0 = (block >> (39 - 32)) & 0x07; + const auto bo1 = ((block >> (43 - 32)) & 0x3) << 3; + const auto bo2 = ((block >> (48 - 32)) & 0x1) << 5; + const auto bo = expand6(bo0 | bo1 | bo2); + const auto go0 = (block >> (49 - 32)) & 0x3F; + const auto go1 = ((block >> (56 - 32)) & 0x01) << 6; + const auto go = expand7(go0 | go1); + const auto ro = expand6((block >> (57 - 32)) & 0x3F); + +#ifdef __ARM_NEON + uint64_t init = uint64_t(uint16_t(rh-ro)) | ( uint64_t(uint16_t(gh-go)) << 16 ) | ( uint64_t(uint16_t(bh-bo)) << 32 ); + int16x8_t chco = vreinterpretq_s16_u64( vdupq_n_u64( init ) ); + init = uint64_t(uint16_t( (rv-ro) - 4 * (rh-ro) )) | ( uint64_t(uint16_t( (gv-go) - 4 * (gh-go) )) << 16 ) | ( uint64_t(uint16_t( (bv-bo) - 4 * (bh-bo) )) << 32 ); + int16x8_t cvco = vreinterpretq_s16_u64( vdupq_n_u64( init ) ); + init = uint64_t(4*ro+2) | ( uint64_t(4*go+2) << 16 ) | ( uint64_t(4*bo+2) << 32 ) | ( uint64_t(0xFFF) << 48 ); + int16x8_t col = vreinterpretq_s16_u64( vdupq_n_u64( init ) ); + + for( int j=0; j<4; j++ ) + { + for( int i=0; i<4; i++ ) + { + uint8x8_t c = vqshrun_n_s16( col, 2 ); + vst1_lane_u32( dst+j*w+i, vreinterpret_u32_u8( c ), 0 ); + col = vaddq_s16( col, chco ); + } + col = vaddq_s16( col, cvco ); + } +#elif defined __AVX2__ + const auto R0 = 4*ro+2; + const auto G0 = 4*go+2; + const auto B0 = 4*bo+2; + const auto RHO = rh-ro; + const auto GHO = gh-go; + const auto BHO = bh-bo; + + __m256i cvco = _mm256_setr_epi16( rv - ro, gv - go, bv - bo, 0, rv - ro, gv - go, bv - bo, 0, rv - ro, gv - go, bv - bo, 0, rv - ro, gv - go, bv - bo, 0 ); + __m256i col = _mm256_setr_epi16( R0, G0, B0, 0xFFF, R0+RHO, G0+GHO, B0+BHO, 0xFFF, R0+2*RHO, G0+2*GHO, B0+2*BHO, 0xFFF, R0+3*RHO, G0+3*GHO, B0+3*BHO, 0xFFF ); + + for( int j=0; j<4; j++ ) + { + __m256i c = _mm256_srai_epi16( col, 2 ); + __m128i s = _mm_packus_epi16( _mm256_castsi256_si128( c ), _mm256_extracti128_si256( c, 1 ) ); + _mm_storeu_si128( (__m128i*)(dst+j*w), s ); + col = _mm256_add_epi16( col, cvco ); + } +#elif defined __SSE4_1__ + __m128i chco = _mm_setr_epi16( rh - ro, gh - go, bh - bo, 0, 0, 0, 0, 0 ); + __m128i cvco = _mm_setr_epi16( (rv - ro) - 4 * (rh - ro), (gv - go) - 4 * (gh - go), (bv - bo) - 4 * (bh - bo), 0, 0, 0, 0, 0 ); + __m128i col = _mm_setr_epi16( 4*ro+2, 4*go+2, 4*bo+2, 0xFFF, 0, 0, 0, 0 ); + + for( int j=0; j<4; j++ ) + { + for( int i=0; i<4; i++ ) + { + __m128i c = _mm_srai_epi16( col, 2 ); + __m128i s = _mm_packus_epi16( c, c ); + dst[j*w+i] = _mm_cvtsi128_si32( s ); + col = _mm_add_epi16( col, chco ); + } + col = _mm_add_epi16( col, cvco ); + } +#else + for( int j=0; j<4; j++ ) + { + for( int i=0; i<4; i++ ) + { + const uint32_t r = (i * (rh - ro) + j * (rv - ro) + 4 * ro + 2) >> 2; + const uint32_t g = (i * (gh - go) + j * (gv - go) + 4 * go + 2) >> 2; + const uint32_t b = (i * (bh - bo) + j * (bv - bo) + 4 * bo + 2) >> 2; + if( ( ( r | g | b ) & ~0xFF ) == 0 ) + { + dst[j*w+i] = r | ( g << 8 ) | ( b << 16 ) | 0xFF000000; + } + else + { + const auto rc = clampu8( r ); + const auto gc = clampu8( g ); + const auto bc = clampu8( b ); + dst[j*w+i] = rc | ( gc << 8 ) | ( bc << 16 ) | 0xFF000000; + } + } + } +#endif +} + +static etcpak_force_inline void DecodePlanarAlpha( uint64_t block, uint64_t alpha, uint32_t* dst, uint32_t w ) +{ + const auto bv = expand6((block >> ( 0 + 32)) & 0x3F); + const auto gv = expand7((block >> ( 6 + 32)) & 0x7F); + const auto rv = expand6((block >> (13 + 32)) & 0x3F); + + const auto bh = expand6((block >> (19 + 32)) & 0x3F); + const auto gh = expand7((block >> (25 + 32)) & 0x7F); + + const auto rh0 = (block >> (32 - 32)) & 0x01; + const auto rh1 = ((block >> (34 - 32)) & 0x1F) << 1; + const auto rh = expand6(rh0 | rh1); + + const auto bo0 = (block >> (39 - 32)) & 0x07; + const auto bo1 = ((block >> (43 - 32)) & 0x3) << 3; + const auto bo2 = ((block >> (48 - 32)) & 0x1) << 5; + const auto bo = expand6(bo0 | bo1 | bo2); + const auto go0 = (block >> (49 - 32)) & 0x3F; + const auto go1 = ((block >> (56 - 32)) & 0x01) << 6; + const auto go = expand7(go0 | go1); + const auto ro = expand6((block >> (57 - 32)) & 0x3F); + + const int32_t base = alpha >> 56; + const int32_t mul = ( alpha >> 52 ) & 0xF; + const auto tbl = g_alpha[( alpha >> 48 ) & 0xF]; + +#ifdef __ARM_NEON + uint64_t init = uint64_t(uint16_t(rh-ro)) | ( uint64_t(uint16_t(gh-go)) << 16 ) | ( uint64_t(uint16_t(bh-bo)) << 32 ); + int16x8_t chco = vreinterpretq_s16_u64( vdupq_n_u64( init ) ); + init = uint64_t(uint16_t( (rv-ro) - 4 * (rh-ro) )) | ( uint64_t(uint16_t( (gv-go) - 4 * (gh-go) )) << 16 ) | ( uint64_t(uint16_t( (bv-bo) - 4 * (bh-bo) )) << 32 ); + int16x8_t cvco = vreinterpretq_s16_u64( vdupq_n_u64( init ) ); + init = uint64_t(4*ro+2) | ( uint64_t(4*go+2) << 16 ) | ( uint64_t(4*bo+2) << 32 ); + int16x8_t col = vreinterpretq_s16_u64( vdupq_n_u64( init ) ); + + for( int j=0; j<4; j++ ) + { + for( int i=0; i<4; i++ ) + { + const auto amod = tbl[(alpha >> ( 45 - j*3 - i*12 )) & 0x7]; + const uint32_t a = clampu8( base + amod * mul ); + uint8x8_t c = vqshrun_n_s16( col, 2 ); + dst[j*w+i] = vget_lane_u32( vreinterpret_u32_u8( c ), 0 ) | ( a << 24 ); + col = vaddq_s16( col, chco ); + } + col = vaddq_s16( col, cvco ); + } +#elif defined __SSE4_1__ + __m128i chco = _mm_setr_epi16( rh - ro, gh - go, bh - bo, 0, 0, 0, 0, 0 ); + __m128i cvco = _mm_setr_epi16( (rv - ro) - 4 * (rh - ro), (gv - go) - 4 * (gh - go), (bv - bo) - 4 * (bh - bo), 0, 0, 0, 0, 0 ); + __m128i col = _mm_setr_epi16( 4*ro+2, 4*go+2, 4*bo+2, 0, 0, 0, 0, 0 ); + + for( int j=0; j<4; j++ ) + { + for( int i=0; i<4; i++ ) + { + const auto amod = tbl[(alpha >> ( 45 - j*3 - i*12 )) & 0x7]; + const uint32_t a = clampu8( base + amod * mul ); + __m128i c = _mm_srai_epi16( col, 2 ); + __m128i s = _mm_packus_epi16( c, c ); + dst[j*w+i] = _mm_cvtsi128_si32( s ) | ( a << 24 ); + col = _mm_add_epi16( col, chco ); + } + col = _mm_add_epi16( col, cvco ); + } +#else + for (auto j = 0; j < 4; j++) + { + for (auto i = 0; i < 4; i++) + { + const uint32_t r = (i * (rh - ro) + j * (rv - ro) + 4 * ro + 2) >> 2; + const uint32_t g = (i * (gh - go) + j * (gv - go) + 4 * go + 2) >> 2; + const uint32_t b = (i * (bh - bo) + j * (bv - bo) + 4 * bo + 2) >> 2; + const auto amod = tbl[(alpha >> ( 45 - j*3 - i*12 )) & 0x7]; + const uint32_t a = clampu8( base + amod * mul ); + if( ( ( r | g | b ) & ~0xFF ) == 0 ) + { + dst[j*w+i] = r | ( g << 8 ) | ( b << 16 ) | ( a << 24 ); + } + else + { + const auto rc = clampu8( r ); + const auto gc = clampu8( g ); + const auto bc = clampu8( b ); + dst[j*w+i] = rc | ( gc << 8 ) | ( bc << 16 ) | ( a << 24 ); + } + } + } +#endif +} + +} + +BitmapPtr BlockData::Decode() +{ + switch( m_type ) + { + case Etc1: + case Etc2_RGB: + return DecodeRGB(); + case Etc2_RGBA: + return DecodeRGBA(); + case Dxt1: + return DecodeDxt1(); + case Dxt5: + return DecodeDxt5(); + default: + assert( false ); + return nullptr; + } +} + +static etcpak_force_inline uint64_t ConvertByteOrder( uint64_t d ) +{ + uint32_t word[2]; + memcpy( word, &d, 8 ); + word[0] = _bswap( word[0] ); + word[1] = _bswap( word[1] ); + memcpy( &d, word, 8 ); + return d; +} + +static etcpak_force_inline void DecodeRGBPart( uint64_t d, uint32_t* dst, uint32_t w ) +{ + d = ConvertByteOrder( d ); + + uint32_t br[2], bg[2], bb[2]; + + if( d & 0x2 ) + { + int32_t dr, dg, db; + + uint32_t r0 = ( d & 0xF8000000 ) >> 27; + uint32_t g0 = ( d & 0x00F80000 ) >> 19; + uint32_t b0 = ( d & 0x0000F800 ) >> 11; + + dr = ( int32_t(d) << 5 ) >> 29; + dg = ( int32_t(d) << 13 ) >> 29; + db = ( int32_t(d) << 21 ) >> 29; + + int32_t r1 = int32_t(r0) + dr; + int32_t g1 = int32_t(g0) + dg; + int32_t b1 = int32_t(b0) + db; + + // T mode + if ( (r1 < 0) || (r1 > 31) ) + { + DecodeT( d, dst, w ); + return; + } + + // H mode + if ((g1 < 0) || (g1 > 31)) + { + DecodeH( d, dst, w ); + return; + } + + // P mode + if( (b1 < 0) || (b1 > 31) ) + { + DecodePlanar( d, dst, w ); + return; + } + + br[0] = ( r0 << 3 ) | ( r0 >> 2 ); + br[1] = ( r1 << 3 ) | ( r1 >> 2 ); + bg[0] = ( g0 << 3 ) | ( g0 >> 2 ); + bg[1] = ( g1 << 3 ) | ( g1 >> 2 ); + bb[0] = ( b0 << 3 ) | ( b0 >> 2 ); + bb[1] = ( b1 << 3 ) | ( b1 >> 2 ); + } + else + { + br[0] = ( ( d & 0xF0000000 ) >> 24 ) | ( ( d & 0xF0000000 ) >> 28 ); + br[1] = ( ( d & 0x0F000000 ) >> 20 ) | ( ( d & 0x0F000000 ) >> 24 ); + bg[0] = ( ( d & 0x00F00000 ) >> 16 ) | ( ( d & 0x00F00000 ) >> 20 ); + bg[1] = ( ( d & 0x000F0000 ) >> 12 ) | ( ( d & 0x000F0000 ) >> 16 ); + bb[0] = ( ( d & 0x0000F000 ) >> 8 ) | ( ( d & 0x0000F000 ) >> 12 ); + bb[1] = ( ( d & 0x00000F00 ) >> 4 ) | ( ( d & 0x00000F00 ) >> 8 ); + } + + unsigned int tcw[2]; + tcw[0] = ( d & 0xE0 ) >> 5; + tcw[1] = ( d & 0x1C ) >> 2; + + uint32_t b1 = ( d >> 32 ) & 0xFFFF; + uint32_t b2 = ( d >> 48 ); + + b1 = ( b1 | ( b1 << 8 ) ) & 0x00FF00FF; + b1 = ( b1 | ( b1 << 4 ) ) & 0x0F0F0F0F; + b1 = ( b1 | ( b1 << 2 ) ) & 0x33333333; + b1 = ( b1 | ( b1 << 1 ) ) & 0x55555555; + + b2 = ( b2 | ( b2 << 8 ) ) & 0x00FF00FF; + b2 = ( b2 | ( b2 << 4 ) ) & 0x0F0F0F0F; + b2 = ( b2 | ( b2 << 2 ) ) & 0x33333333; + b2 = ( b2 | ( b2 << 1 ) ) & 0x55555555; + + uint32_t idx = b1 | ( b2 << 1 ); + + if( d & 0x1 ) + { + for( int i=0; i<4; i++ ) + { + for( int j=0; j<4; j++ ) + { + const auto mod = g_table[tcw[j/2]][idx & 0x3]; + const auto r = br[j/2] + mod; + const auto g = bg[j/2] + mod; + const auto b = bb[j/2] + mod; + if( ( ( r | g | b ) & ~0xFF ) == 0 ) + { + dst[j*w+i] = r | ( g << 8 ) | ( b << 16 ) | 0xFF000000; + } + else + { + const auto rc = clampu8( r ); + const auto gc = clampu8( g ); + const auto bc = clampu8( b ); + dst[j*w+i] = rc | ( gc << 8 ) | ( bc << 16 ) | 0xFF000000; + } + idx >>= 2; + } + } + } + else + { + for( int i=0; i<4; i++ ) + { + const auto tbl = g_table[tcw[i/2]]; + const auto cr = br[i/2]; + const auto cg = bg[i/2]; + const auto cb = bb[i/2]; + + for( int j=0; j<4; j++ ) + { + const auto mod = tbl[idx & 0x3]; + const auto r = cr + mod; + const auto g = cg + mod; + const auto b = cb + mod; + if( ( ( r | g | b ) & ~0xFF ) == 0 ) + { + dst[j*w+i] = r | ( g << 8 ) | ( b << 16 ) | 0xFF000000; + } + else + { + const auto rc = clampu8( r ); + const auto gc = clampu8( g ); + const auto bc = clampu8( b ); + dst[j*w+i] = rc | ( gc << 8 ) | ( bc << 16 ) | 0xFF000000; + } + idx >>= 2; + } + } + } +} + +static etcpak_force_inline void DecodeRGBAPart( uint64_t d, uint64_t alpha, uint32_t* dst, uint32_t w ) +{ + d = ConvertByteOrder( d ); + alpha = _bswap64( alpha ); + + uint32_t br[2], bg[2], bb[2]; + + if( d & 0x2 ) + { + int32_t dr, dg, db; + + uint32_t r0 = ( d & 0xF8000000 ) >> 27; + uint32_t g0 = ( d & 0x00F80000 ) >> 19; + uint32_t b0 = ( d & 0x0000F800 ) >> 11; + + dr = ( int32_t(d) << 5 ) >> 29; + dg = ( int32_t(d) << 13 ) >> 29; + db = ( int32_t(d) << 21 ) >> 29; + + int32_t r1 = int32_t(r0) + dr; + int32_t g1 = int32_t(g0) + dg; + int32_t b1 = int32_t(b0) + db; + + // T mode + if ( (r1 < 0) || (r1 > 31) ) + { + DecodeTAlpha( d, alpha, dst, w ); + return; + } + + // H mode + if ( (g1 < 0) || (g1 > 31) ) + { + DecodeHAlpha( d, alpha, dst, w ); + return; + } + + // P mode + if ( (b1 < 0) || (b1 > 31) ) + { + DecodePlanarAlpha( d, alpha, dst, w ); + return; + } + + br[0] = ( r0 << 3 ) | ( r0 >> 2 ); + br[1] = ( r1 << 3 ) | ( r1 >> 2 ); + bg[0] = ( g0 << 3 ) | ( g0 >> 2 ); + bg[1] = ( g1 << 3 ) | ( g1 >> 2 ); + bb[0] = ( b0 << 3 ) | ( b0 >> 2 ); + bb[1] = ( b1 << 3 ) | ( b1 >> 2 ); + } + else + { + br[0] = ( ( d & 0xF0000000 ) >> 24 ) | ( ( d & 0xF0000000 ) >> 28 ); + br[1] = ( ( d & 0x0F000000 ) >> 20 ) | ( ( d & 0x0F000000 ) >> 24 ); + bg[0] = ( ( d & 0x00F00000 ) >> 16 ) | ( ( d & 0x00F00000 ) >> 20 ); + bg[1] = ( ( d & 0x000F0000 ) >> 12 ) | ( ( d & 0x000F0000 ) >> 16 ); + bb[0] = ( ( d & 0x0000F000 ) >> 8 ) | ( ( d & 0x0000F000 ) >> 12 ); + bb[1] = ( ( d & 0x00000F00 ) >> 4 ) | ( ( d & 0x00000F00 ) >> 8 ); + } + + unsigned int tcw[2]; + tcw[0] = ( d & 0xE0 ) >> 5; + tcw[1] = ( d & 0x1C ) >> 2; + + uint32_t b1 = ( d >> 32 ) & 0xFFFF; + uint32_t b2 = ( d >> 48 ); + + b1 = ( b1 | ( b1 << 8 ) ) & 0x00FF00FF; + b1 = ( b1 | ( b1 << 4 ) ) & 0x0F0F0F0F; + b1 = ( b1 | ( b1 << 2 ) ) & 0x33333333; + b1 = ( b1 | ( b1 << 1 ) ) & 0x55555555; + + b2 = ( b2 | ( b2 << 8 ) ) & 0x00FF00FF; + b2 = ( b2 | ( b2 << 4 ) ) & 0x0F0F0F0F; + b2 = ( b2 | ( b2 << 2 ) ) & 0x33333333; + b2 = ( b2 | ( b2 << 1 ) ) & 0x55555555; + + uint32_t idx = b1 | ( b2 << 1 ); + + const int32_t base = alpha >> 56; + const int32_t mul = ( alpha >> 52 ) & 0xF; + const auto atbl = g_alpha[( alpha >> 48 ) & 0xF]; + + if( d & 0x1 ) + { + for( int i=0; i<4; i++ ) + { + for( int j=0; j<4; j++ ) + { + const auto mod = g_table[tcw[j/2]][idx & 0x3]; + const auto r = br[j/2] + mod; + const auto g = bg[j/2] + mod; + const auto b = bb[j/2] + mod; + const auto amod = atbl[(alpha >> ( 45 - j*3 - i*12 )) & 0x7]; + const uint32_t a = clampu8( base + amod * mul ); + if( ( ( r | g | b ) & ~0xFF ) == 0 ) + { + dst[j*w+i] = r | ( g << 8 ) | ( b << 16 ) | ( a << 24 ); + } + else + { + const auto rc = clampu8( r ); + const auto gc = clampu8( g ); + const auto bc = clampu8( b ); + dst[j*w+i] = rc | ( gc << 8 ) | ( bc << 16 ) | ( a << 24 ); + } + idx >>= 2; + } + } + } + else + { + for( int i=0; i<4; i++ ) + { + const auto tbl = g_table[tcw[i/2]]; + const auto cr = br[i/2]; + const auto cg = bg[i/2]; + const auto cb = bb[i/2]; + + for( int j=0; j<4; j++ ) + { + const auto mod = tbl[idx & 0x3]; + const auto r = cr + mod; + const auto g = cg + mod; + const auto b = cb + mod; + const auto amod = atbl[(alpha >> ( 45 - j*3 - i*12 )) & 0x7]; + const uint32_t a = clampu8( base + amod * mul ); + if( ( ( r | g | b ) & ~0xFF ) == 0 ) + { + dst[j*w+i] = r | ( g << 8 ) | ( b << 16 ) | ( a << 24 ); + } + else + { + const auto rc = clampu8( r ); + const auto gc = clampu8( g ); + const auto bc = clampu8( b ); + dst[j*w+i] = rc | ( gc << 8 ) | ( bc << 16 ) | ( a << 24 ); + } + idx >>= 2; + } + } + } +} + +BitmapPtr BlockData::DecodeRGB() +{ + auto ret = std::make_shared<Bitmap>( m_size ); + + const uint64_t* src = (const uint64_t*)( m_data + m_dataOffset ); + uint32_t* dst = ret->Data(); + + for( int y=0; y<m_size.y/4; y++ ) + { + for( int x=0; x<m_size.x/4; x++ ) + { + uint64_t d = *src++; + DecodeRGBPart( d, dst, m_size.x ); + dst += 4; + } + dst += m_size.x*3; + } + + return ret; +} + +BitmapPtr BlockData::DecodeRGBA() +{ + auto ret = std::make_shared<Bitmap>( m_size ); + + const uint64_t* src = (const uint64_t*)( m_data + m_dataOffset ); + uint32_t* dst = ret->Data(); + + for( int y=0; y<m_size.y/4; y++ ) + { + for( int x=0; x<m_size.x/4; x++ ) + { + uint64_t a = *src++; + uint64_t d = *src++; + DecodeRGBAPart( d, a, dst, m_size.x ); + dst += 4; + } + dst += m_size.x*3; + } + + return ret; +} + +static etcpak_force_inline void DecodeDxt1Part( uint64_t d, uint32_t* dst, uint32_t w ) +{ + uint8_t* in = (uint8_t*)&d; + uint16_t c0, c1; + uint32_t idx; + memcpy( &c0, in, 2 ); + memcpy( &c1, in+2, 2 ); + memcpy( &idx, in+4, 4 ); + + uint8_t r0 = ( ( c0 & 0xF800 ) >> 8 ) | ( ( c0 & 0xF800 ) >> 13 ); + uint8_t g0 = ( ( c0 & 0x07E0 ) >> 3 ) | ( ( c0 & 0x07E0 ) >> 9 ); + uint8_t b0 = ( ( c0 & 0x001F ) << 3 ) | ( ( c0 & 0x001F ) >> 2 ); + + uint8_t r1 = ( ( c1 & 0xF800 ) >> 8 ) | ( ( c1 & 0xF800 ) >> 13 ); + uint8_t g1 = ( ( c1 & 0x07E0 ) >> 3 ) | ( ( c1 & 0x07E0 ) >> 9 ); + uint8_t b1 = ( ( c1 & 0x001F ) << 3 ) | ( ( c1 & 0x001F ) >> 2 ); + + uint32_t dict[4]; + + dict[0] = 0xFF000000 | ( b0 << 16 ) | ( g0 << 8 ) | r0; + dict[1] = 0xFF000000 | ( b1 << 16 ) | ( g1 << 8 ) | r1; + + uint32_t r, g, b; + if( c0 > c1 ) + { + r = (2*r0+r1)/3; + g = (2*g0+g1)/3; + b = (2*b0+b1)/3; + dict[2] = 0xFF000000 | ( b << 16 ) | ( g << 8 ) | r; + r = (2*r1+r0)/3; + g = (2*g1+g0)/3; + b = (2*b1+b0)/3; + dict[3] = 0xFF000000 | ( b << 16 ) | ( g << 8 ) | r; + } + else + { + r = (int(r0)+r1)/2; + g = (int(g0)+g1)/2; + b = (int(b0)+b1)/2; + dict[2] = 0xFF000000 | ( b << 16 ) | ( g << 8 ) | r; + dict[3] = 0xFF000000; + } + + memcpy( dst+0, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+1, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+2, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+3, dict + (idx & 0x3), 4 ); + idx >>= 2; + dst += w; + + memcpy( dst+0, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+1, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+2, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+3, dict + (idx & 0x3), 4 ); + idx >>= 2; + dst += w; + + memcpy( dst+0, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+1, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+2, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+3, dict + (idx & 0x3), 4 ); + idx >>= 2; + dst += w; + + memcpy( dst+0, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+1, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+2, dict + (idx & 0x3), 4 ); + idx >>= 2; + memcpy( dst+3, dict + (idx & 0x3), 4 ); +} + +static etcpak_force_inline void DecodeDxt5Part( uint64_t a, uint64_t d, uint32_t* dst, uint32_t w ) +{ + uint8_t* ain = (uint8_t*)&a; + uint8_t a0, a1; + uint64_t aidx = 0; + memcpy( &a0, ain, 1 ); + memcpy( &a1, ain+1, 1 ); + memcpy( &aidx, ain+2, 6 ); + + uint8_t* in = (uint8_t*)&d; + uint16_t c0, c1; + uint32_t idx; + memcpy( &c0, in, 2 ); + memcpy( &c1, in+2, 2 ); + memcpy( &idx, in+4, 4 ); + + uint32_t adict[8]; + adict[0] = a0 << 24; + adict[1] = a1 << 24; + if( a0 > a1 ) + { + adict[2] = ( (6*a0+1*a1)/7 ) << 24; + adict[3] = ( (5*a0+2*a1)/7 ) << 24; + adict[4] = ( (4*a0+3*a1)/7 ) << 24; + adict[5] = ( (3*a0+4*a1)/7 ) << 24; + adict[6] = ( (2*a0+5*a1)/7 ) << 24; + adict[7] = ( (1*a0+6*a1)/7 ) << 24; + } + else + { + adict[2] = ( (4*a0+1*a1)/5 ) << 24; + adict[3] = ( (3*a0+2*a1)/5 ) << 24; + adict[4] = ( (2*a0+3*a1)/5 ) << 24; + adict[5] = ( (1*a0+4*a1)/5 ) << 24; + adict[6] = 0; + adict[7] = 0xFF000000; + } + + uint8_t r0 = ( ( c0 & 0xF800 ) >> 8 ) | ( ( c0 & 0xF800 ) >> 13 ); + uint8_t g0 = ( ( c0 & 0x07E0 ) >> 3 ) | ( ( c0 & 0x07E0 ) >> 9 ); + uint8_t b0 = ( ( c0 & 0x001F ) << 3 ) | ( ( c0 & 0x001F ) >> 2 ); + + uint8_t r1 = ( ( c1 & 0xF800 ) >> 8 ) | ( ( c1 & 0xF800 ) >> 13 ); + uint8_t g1 = ( ( c1 & 0x07E0 ) >> 3 ) | ( ( c1 & 0x07E0 ) >> 9 ); + uint8_t b1 = ( ( c1 & 0x001F ) << 3 ) | ( ( c1 & 0x001F ) >> 2 ); + + uint32_t dict[4]; + + dict[0] = ( b0 << 16 ) | ( g0 << 8 ) | r0; + dict[1] = ( b1 << 16 ) | ( g1 << 8 ) | r1; + + uint32_t r, g, b; + if( c0 > c1 ) + { + r = (2*r0+r1)/3; + g = (2*g0+g1)/3; + b = (2*b0+b1)/3; + dict[2] = ( b << 16 ) | ( g << 8 ) | r; + r = (2*r1+r0)/3; + g = (2*g1+g0)/3; + b = (2*b1+b0)/3; + dict[3] = ( b << 16 ) | ( g << 8 ) | r; + } + else + { + r = (int(r0)+r1)/2; + g = (int(g0)+g1)/2; + b = (int(b0)+b1)/2; + dict[2] = ( b << 16 ) | ( g << 8 ) | r; + dict[3] = 0; + } + + dst[0] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[1] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[2] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[3] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst += w; + + dst[0] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[1] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[2] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[3] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst += w; + + dst[0] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[1] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[2] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[3] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst += w; + + dst[0] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[1] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[2] = dict[idx & 0x3] | adict[aidx & 0x7]; + idx >>= 2; + aidx >>= 3; + dst[3] = dict[idx & 0x3] | adict[aidx & 0x7]; +} + +BitmapPtr BlockData::DecodeDxt1() +{ + auto ret = std::make_shared<Bitmap>( m_size ); + + const uint64_t* src = (const uint64_t*)( m_data + m_dataOffset ); + uint32_t* dst = ret->Data(); + + for( int y=0; y<m_size.y/4; y++ ) + { + for( int x=0; x<m_size.x/4; x++ ) + { + uint64_t d = *src++; + DecodeDxt1Part( d, dst, m_size.x ); + dst += 4; + } + dst += m_size.x*3; + } + + return ret; +} + +BitmapPtr BlockData::DecodeDxt5() +{ + auto ret = std::make_shared<Bitmap>( m_size ); + + const uint64_t* src = (const uint64_t*)( m_data + m_dataOffset ); + uint32_t* dst = ret->Data(); + + for( int y=0; y<m_size.y/4; y++ ) + { + for( int x=0; x<m_size.x/4; x++ ) + { + uint64_t a = *src++; + uint64_t d = *src++; + DecodeDxt5Part( a, d, dst, m_size.x ); + dst += 4; + } + dst += m_size.x*3; + } + + return ret; +} diff --git a/thirdparty/etcpak/BlockData.hpp b/thirdparty/etcpak/BlockData.hpp new file mode 100644 index 0000000000..209e35b4e6 --- /dev/null +++ b/thirdparty/etcpak/BlockData.hpp @@ -0,0 +1,56 @@ +#ifndef __BLOCKDATA_HPP__ +#define __BLOCKDATA_HPP__ + +#include <condition_variable> +#include <future> +#include <memory> +#include <mutex> +#include <stdint.h> +#include <stdio.h> +#include <vector> + +#include "Bitmap.hpp" +#include "ForceInline.hpp" +#include "Vector.hpp" + +class BlockData +{ +public: + enum Type + { + Etc1, + Etc2_RGB, + Etc2_RGBA, + Dxt1, + Dxt5 + }; + + BlockData( const char* fn ); + BlockData( const char* fn, const v2i& size, bool mipmap, Type type ); + BlockData( const v2i& size, bool mipmap, Type type ); + ~BlockData(); + + BitmapPtr Decode(); + + void Process( const uint32_t* src, uint32_t blocks, size_t offset, size_t width, Channels type, bool dither ); + void ProcessRGBA( const uint32_t* src, uint32_t blocks, size_t offset, size_t width ); + + const v2i& Size() const { return m_size; } + +private: + etcpak_no_inline BitmapPtr DecodeRGB(); + etcpak_no_inline BitmapPtr DecodeRGBA(); + etcpak_no_inline BitmapPtr DecodeDxt1(); + etcpak_no_inline BitmapPtr DecodeDxt5(); + + uint8_t* m_data; + v2i m_size; + size_t m_dataOffset; + FILE* m_file; + size_t m_maplen; + Type m_type; +}; + +typedef std::shared_ptr<BlockData> BlockDataPtr; + +#endif diff --git a/thirdparty/etcpak/ColorSpace.cpp b/thirdparty/etcpak/ColorSpace.cpp new file mode 100644 index 0000000000..0411541066 --- /dev/null +++ b/thirdparty/etcpak/ColorSpace.cpp @@ -0,0 +1,114 @@ +#include <math.h> +#include <stdint.h> + +#include "Math.hpp" +#include "ColorSpace.hpp" + +namespace Color +{ + + static const XYZ white( v3b( 255, 255, 255 ) ); + static const v3f rwhite( 1.f / white.x, 1.f / white.y, 1.f / white.z ); + + + XYZ::XYZ( float _x, float _y, float _z ) + : x( _x ) + , y( _y ) + , z( _z ) + { + } + + XYZ::XYZ( const v3b& rgb ) + { + const float r = rgb.x / 255.f; + const float g = rgb.y / 255.f; + const float b = rgb.z / 255.f; + + const float rl = sRGB2linear( r ); + const float gl = sRGB2linear( g ); + const float bl = sRGB2linear( b ); + + x = 0.4124f * rl + 0.3576f * gl + 0.1805f * bl; + y = 0.2126f * rl + 0.7152f * gl + 0.0722f * bl; + z = 0.0193f * rl + 0.1192f * gl + 0.9505f * bl; + } + + static float revlab( float t ) + { + const float p1 = 6.f/29.f; + const float p2 = 4.f/29.f; + + if( t > p1 ) + { + return t*t*t; + } + else + { + return 3 * sq( p1 ) * ( t - p2 ); + } + } + + XYZ::XYZ( const Lab& lab ) + { + y = white.y * revlab( 1.f/116.f * ( lab.L + 16 ) ); + x = white.x * revlab( 1.f/116.f * ( lab.L + 16 ) + 1.f/500.f * lab.a ); + z = white.z * revlab( 1.f/116.f * ( lab.L + 16 ) - 1.f/200.f * lab.b ); + } + + v3i XYZ::RGB() const + { + const float rl = 3.2406f * x - 1.5372f * y - 0.4986f * z; + const float gl = -0.9689f * x + 1.8758f * y + 0.0415f * z; + const float bl = 0.0557f * x - 0.2040f * y + 1.0570f * z; + + const float r = linear2sRGB( rl ); + const float g = linear2sRGB( gl ); + const float b = linear2sRGB( bl ); + + return v3i( clampu8( int32_t( r * 255 ) ), clampu8( int32_t( g * 255 ) ), clampu8( int32_t( b * 255 ) ) ); + } + + + Lab::Lab() + : L( 0 ) + , a( 0 ) + , b( 0 ) + { + } + + Lab::Lab( float L, float a, float b ) + : L( L ) + , a( a ) + , b( b ) + { + } + + static float labfunc( float t ) + { + const float p1 = (6.f/29.f)*(6.f/29.f)*(6.f/29.f); + const float p2 = (1.f/3.f)*(29.f/6.f)*(29.f/6.f); + const float p3 = (4.f/29.f); + + if( t > p1 ) + { + return pow( t, 1.f/3.f ); + } + else + { + return p2 * t + p3; + } + } + + Lab::Lab( const XYZ& xyz ) + { + L = 116 * labfunc( xyz.y * rwhite.y ) - 16; + a = 500 * ( labfunc( xyz.x * rwhite.x ) - labfunc( xyz.y * rwhite.y ) ); + b = 200 * ( labfunc( xyz.y * rwhite.y ) - labfunc( xyz.z * rwhite.z ) ); + } + + Lab::Lab( const v3b& rgb ) + { + new(this) Lab( XYZ( rgb ) ); + } + +} diff --git a/thirdparty/etcpak/ColorSpace.hpp b/thirdparty/etcpak/ColorSpace.hpp new file mode 100644 index 0000000000..c9d0a9cf3f --- /dev/null +++ b/thirdparty/etcpak/ColorSpace.hpp @@ -0,0 +1,36 @@ +#ifndef __DARKRL__COLORSPACE_HPP__ +#define __DARKRL__COLORSPACE_HPP__ + +#include "Vector.hpp" + +namespace Color +{ + + class Lab; + + class XYZ + { + public: + XYZ( float x, float y, float z ); + XYZ( const v3b& rgb ); + XYZ( const Lab& lab ); + + v3i RGB() const; + + float x, y, z; + }; + + class Lab + { + public: + Lab(); + Lab( float L, float a, float b ); + Lab( const XYZ& xyz ); + Lab( const v3b& rgb ); + + float L, a, b; + }; + +} + +#endif diff --git a/thirdparty/etcpak/DataProvider.cpp b/thirdparty/etcpak/DataProvider.cpp new file mode 100644 index 0000000000..6bd4b105ed --- /dev/null +++ b/thirdparty/etcpak/DataProvider.cpp @@ -0,0 +1,77 @@ +#include <assert.h> +#include <utility> + +#include "BitmapDownsampled.hpp" +#include "DataProvider.hpp" +#include "MipMap.hpp" + +DataProvider::DataProvider( const char* fn, bool mipmap, bool bgr ) + : m_offset( 0 ) + , m_mipmap( mipmap ) + , m_done( false ) + , m_lines( 32 ) +{ + m_bmp.emplace_back( new Bitmap( fn, m_lines, bgr ) ); + m_current = m_bmp[0].get(); +} + +DataProvider::~DataProvider() +{ +} + +unsigned int DataProvider::NumberOfParts() const +{ + unsigned int parts = ( ( m_bmp[0]->Size().y / 4 ) + m_lines - 1 ) / m_lines; + + if( m_mipmap ) + { + v2i current = m_bmp[0]->Size(); + int levels = NumberOfMipLevels( current ); + unsigned int lines = m_lines; + for( int i=1; i<levels; i++ ) + { + assert( current.x != 1 || current.y != 1 ); + current.x = std::max( 1, current.x / 2 ); + current.y = std::max( 1, current.y / 2 ); + lines *= 2; + parts += ( ( std::max( 4, current.y ) / 4 ) + lines - 1 ) / lines; + } + assert( current.x == 1 && current.y == 1 ); + } + + return parts; +} + +DataPart DataProvider::NextPart() +{ + assert( !m_done ); + + unsigned int lines = m_lines; + bool done; + + const auto ptr = m_current->NextBlock( lines, done ); + DataPart ret = { + ptr, + std::max<unsigned int>( 4, m_current->Size().x ), + lines, + m_offset + }; + + m_offset += m_current->Size().x / 4 * lines; + + if( done ) + { + if( m_mipmap && ( m_current->Size().x != 1 || m_current->Size().y != 1 ) ) + { + m_lines *= 2; + m_bmp.emplace_back( new BitmapDownsampled( *m_current, m_lines ) ); + m_current = m_bmp[m_bmp.size()-1].get(); + } + else + { + m_done = true; + } + } + + return ret; +} diff --git a/thirdparty/etcpak/DataProvider.hpp b/thirdparty/etcpak/DataProvider.hpp new file mode 100644 index 0000000000..e773801ed6 --- /dev/null +++ b/thirdparty/etcpak/DataProvider.hpp @@ -0,0 +1,41 @@ +#ifndef __DATAPROVIDER_HPP__ +#define __DATAPROVIDER_HPP__ + +#include <memory> +#include <stdint.h> +#include <vector> + +#include "Bitmap.hpp" + +struct DataPart +{ + const uint32_t* src; + unsigned int width; + unsigned int lines; + unsigned int offset; +}; + +class DataProvider +{ +public: + DataProvider( const char* fn, bool mipmap, bool bgr ); + ~DataProvider(); + + unsigned int NumberOfParts() const; + + DataPart NextPart(); + + bool Alpha() const { return m_bmp[0]->Alpha(); } + const v2i& Size() const { return m_bmp[0]->Size(); } + const Bitmap& ImageData() const { return *m_bmp[0]; } + +private: + std::vector<std::unique_ptr<Bitmap>> m_bmp; + Bitmap* m_current; + unsigned int m_offset; + unsigned int m_lines; + bool m_mipmap; + bool m_done; +}; + +#endif diff --git a/thirdparty/etcpak/Debug.cpp b/thirdparty/etcpak/Debug.cpp new file mode 100644 index 0000000000..72dc4e0526 --- /dev/null +++ b/thirdparty/etcpak/Debug.cpp @@ -0,0 +1,31 @@ +#include <algorithm> +#include <vector> +#include "Debug.hpp" + +static std::vector<DebugLog::Callback*> s_callbacks; + +void DebugLog::Message( const char* msg ) +{ + for( auto it = s_callbacks.begin(); it != s_callbacks.end(); ++it ) + { + (*it)->OnDebugMessage( msg ); + } +} + +void DebugLog::AddCallback( Callback* c ) +{ + const auto it = std::find( s_callbacks.begin(), s_callbacks.end(), c ); + if( it == s_callbacks.end() ) + { + s_callbacks.push_back( c ); + } +} + +void DebugLog::RemoveCallback( Callback* c ) +{ + const auto it = std::find( s_callbacks.begin(), s_callbacks.end(), c ); + if( it != s_callbacks.end() ) + { + s_callbacks.erase( it ); + } +} diff --git a/thirdparty/etcpak/Debug.hpp b/thirdparty/etcpak/Debug.hpp new file mode 100644 index 0000000000..524eaa7baf --- /dev/null +++ b/thirdparty/etcpak/Debug.hpp @@ -0,0 +1,27 @@ +#ifndef __DARKRL__DEBUG_HPP__ +#define __DARKRL__DEBUG_HPP__ + +#ifdef DEBUG +# include <sstream> +# define DBGPRINT(msg) { std::stringstream __buf; __buf << msg; DebugLog::Message( __buf.str().c_str() ); } +#else +# define DBGPRINT(msg) ((void)0) +#endif + +class DebugLog +{ +public: + struct Callback + { + virtual void OnDebugMessage( const char* msg ) = 0; + }; + + static void Message( const char* msg ); + static void AddCallback( Callback* c ); + static void RemoveCallback( Callback* c ); + +private: + DebugLog() {} +}; + +#endif diff --git a/thirdparty/etcpak/Dither.cpp b/thirdparty/etcpak/Dither.cpp new file mode 100644 index 0000000000..355686f26b --- /dev/null +++ b/thirdparty/etcpak/Dither.cpp @@ -0,0 +1,120 @@ +#include <algorithm> +#include <string.h> + +#include "Dither.hpp" +#include "Math.hpp" +#ifdef __SSE4_1__ +# ifdef _MSC_VER +# include <intrin.h> +# include <Windows.h> +# else +# include <x86intrin.h> +# endif +#endif + +#ifdef __AVX2__ +void DitherAvx2( uint8_t* data, __m128i px0, __m128i px1, __m128i px2, __m128i px3 ) +{ + static constexpr uint8_t a31[] = { 0, 0, 0, 1, 2, 0, 4, 0, 0, 2, 0, 0, 4, 0, 3, 0 }; + static constexpr uint8_t a63[] = { 0, 0, 0, 0, 1, 0, 2, 0, 0, 1, 0, 0, 2, 0, 1, 0 }; + static constexpr uint8_t s31[] = { 5, 0, 4, 0, 0, 2, 0, 1, 3, 0, 4, 0, 0, 0, 0, 2 }; + static constexpr uint8_t s63[] = { 2, 0, 2, 0, 0, 1, 0, 0, 1, 0, 2, 0, 0, 0, 0, 1 }; + + const __m256i BayerAdd0 = _mm256_setr_epi8( + a31[0], a63[0], a31[0], 0, a31[1], a63[1], a31[1], 0, a31[2], a63[2], a31[2], 0, a31[3], a63[3], a31[3], 0, + a31[4], a63[4], a31[4], 0, a31[5], a63[5], a31[5], 0, a31[6], a63[6], a31[6], 0, a31[7], a63[7], a31[7], 0 + ); + const __m256i BayerAdd1 = _mm256_setr_epi8( + a31[8], a63[8], a31[8], 0, a31[9], a63[9], a31[9], 0, a31[10], a63[10], a31[10], 0, a31[11], a63[11], a31[11], 0, + a31[12], a63[12], a31[12], 0, a31[13], a63[13], a31[13], 0, a31[14], a63[14], a31[14], 0, a31[15], a63[15], a31[15], 0 + ); + const __m256i BayerSub0 = _mm256_setr_epi8( + s31[0], s63[0], s31[0], 0, s31[1], s63[1], s31[1], 0, s31[2], s63[2], s31[2], 0, s31[3], s63[3], s31[3], 0, + s31[4], s63[4], s31[4], 0, s31[5], s63[5], s31[5], 0, s31[6], s63[6], s31[6], 0, s31[7], s63[7], s31[7], 0 + ); + const __m256i BayerSub1 = _mm256_setr_epi8( + s31[8], s63[8], s31[8], 0, s31[9], s63[9], s31[9], 0, s31[10], s63[10], s31[10], 0, s31[11], s63[11], s31[11], 0, + s31[12], s63[12], s31[12], 0, s31[13], s63[13], s31[13], 0, s31[14], s63[14], s31[14], 0, s31[15], s63[15], s31[15], 0 + ); + + __m256i l0 = _mm256_inserti128_si256( _mm256_castsi128_si256( px0 ), px1, 1 ); + __m256i l1 = _mm256_inserti128_si256( _mm256_castsi128_si256( px2 ), px3, 1 ); + + __m256i a0 = _mm256_adds_epu8( l0, BayerAdd0 ); + __m256i a1 = _mm256_adds_epu8( l1, BayerAdd1 ); + __m256i s0 = _mm256_subs_epu8( a0, BayerSub0 ); + __m256i s1 = _mm256_subs_epu8( a1, BayerSub1 ); + + _mm256_storeu_si256( (__m256i*)(data ), s0 ); + _mm256_storeu_si256( (__m256i*)(data+32), s1 ); + +} +#endif + +void Dither( uint8_t* data ) +{ +#ifdef __AVX2__ + static constexpr uint8_t a31[] = { 0, 0, 0, 1, 2, 0, 4, 0, 0, 2, 0, 0, 4, 0, 3, 0 }; + static constexpr uint8_t a63[] = { 0, 0, 0, 0, 1, 0, 2, 0, 0, 1, 0, 0, 2, 0, 1, 0 }; + static constexpr uint8_t s31[] = { 5, 0, 4, 0, 0, 2, 0, 1, 3, 0, 4, 0, 0, 0, 0, 2 }; + static constexpr uint8_t s63[] = { 2, 0, 2, 0, 0, 1, 0, 0, 1, 0, 2, 0, 0, 0, 0, 1 }; + + const __m256i BayerAdd0 = _mm256_setr_epi8( + a31[0], a63[0], a31[0], 0, a31[1], a63[1], a31[1], 0, a31[2], a63[2], a31[2], 0, a31[3], a63[3], a31[3], 0, + a31[4], a63[4], a31[4], 0, a31[5], a63[5], a31[5], 0, a31[6], a63[6], a31[6], 0, a31[7], a63[7], a31[7], 0 + ); + const __m256i BayerAdd1 = _mm256_setr_epi8( + a31[8], a63[8], a31[8], 0, a31[9], a63[9], a31[9], 0, a31[10], a63[10], a31[10], 0, a31[11], a63[11], a31[11], 0, + a31[12], a63[12], a31[12], 0, a31[13], a63[13], a31[13], 0, a31[14], a63[14], a31[14], 0, a31[15], a63[15], a31[15], 0 + ); + const __m256i BayerSub0 = _mm256_setr_epi8( + s31[0], s63[0], s31[0], 0, s31[1], s63[1], s31[1], 0, s31[2], s63[2], s31[2], 0, s31[3], s63[3], s31[3], 0, + s31[4], s63[4], s31[4], 0, s31[5], s63[5], s31[5], 0, s31[6], s63[6], s31[6], 0, s31[7], s63[7], s31[7], 0 + ); + const __m256i BayerSub1 = _mm256_setr_epi8( + s31[8], s63[8], s31[8], 0, s31[9], s63[9], s31[9], 0, s31[10], s63[10], s31[10], 0, s31[11], s63[11], s31[11], 0, + s31[12], s63[12], s31[12], 0, s31[13], s63[13], s31[13], 0, s31[14], s63[14], s31[14], 0, s31[15], s63[15], s31[15], 0 + ); + + __m256i px0 = _mm256_loadu_si256( (__m256i*)(data ) ); + __m256i px1 = _mm256_loadu_si256( (__m256i*)(data+32) ); + + __m256i a0 = _mm256_adds_epu8( px0, BayerAdd0 ); + __m256i a1 = _mm256_adds_epu8( px1, BayerAdd1 ); + __m256i s0 = _mm256_subs_epu8( a0, BayerSub0 ); + __m256i s1 = _mm256_subs_epu8( a1, BayerSub1 ); + + _mm256_storeu_si256( (__m256i*)(data ), s0 ); + _mm256_storeu_si256( (__m256i*)(data+32), s1 ); +#else + static constexpr int8_t Bayer31[16] = { + ( 0-8)*2/3, ( 8-8)*2/3, ( 2-8)*2/3, (10-8)*2/3, + (12-8)*2/3, ( 4-8)*2/3, (14-8)*2/3, ( 6-8)*2/3, + ( 3-8)*2/3, (11-8)*2/3, ( 1-8)*2/3, ( 9-8)*2/3, + (15-8)*2/3, ( 7-8)*2/3, (13-8)*2/3, ( 5-8)*2/3 + }; + static constexpr int8_t Bayer63[16] = { + ( 0-8)*2/6, ( 8-8)*2/6, ( 2-8)*2/6, (10-8)*2/6, + (12-8)*2/6, ( 4-8)*2/6, (14-8)*2/6, ( 6-8)*2/6, + ( 3-8)*2/6, (11-8)*2/6, ( 1-8)*2/6, ( 9-8)*2/6, + (15-8)*2/6, ( 7-8)*2/6, (13-8)*2/6, ( 5-8)*2/6 + }; + + for( int i=0; i<16; i++ ) + { + uint32_t col; + memcpy( &col, data, 4 ); + uint8_t r = col & 0xFF; + uint8_t g = ( col >> 8 ) & 0xFF; + uint8_t b = ( col >> 16 ) & 0xFF; + + r = clampu8( r + Bayer31[i] ); + g = clampu8( g + Bayer63[i] ); + b = clampu8( b + Bayer31[i] ); + + col = r | ( g << 8 ) | ( b << 16 ); + memcpy( data, &col, 4 ); + data += 4; + } +#endif +} diff --git a/thirdparty/etcpak/Dither.hpp b/thirdparty/etcpak/Dither.hpp new file mode 100644 index 0000000000..e43ce5676d --- /dev/null +++ b/thirdparty/etcpak/Dither.hpp @@ -0,0 +1,21 @@ +#ifndef __DITHER_HPP__ +#define __DITHER_HPP__ + +#include <stddef.h> +#include <stdint.h> + +#ifdef __AVX2__ +# ifdef _MSC_VER +# include <intrin.h> +# else +# include <x86intrin.h> +# endif +#endif + +void Dither( uint8_t* data ); + +#ifdef __AVX2__ +void DitherAvx2( uint8_t* data, __m128i px0, __m128i px1, __m128i px2, __m128i px3 ); +#endif + +#endif diff --git a/thirdparty/etcpak/Error.cpp b/thirdparty/etcpak/Error.cpp new file mode 100644 index 0000000000..014ecdab66 --- /dev/null +++ b/thirdparty/etcpak/Error.cpp @@ -0,0 +1,48 @@ +#include <stdint.h> + +#include "Error.hpp" +#include "Math.hpp" + +float CalcMSE3( const Bitmap& bmp, const Bitmap& out ) +{ + float err = 0; + + const uint32_t* p1 = bmp.Data(); + const uint32_t* p2 = out.Data(); + size_t cnt = bmp.Size().x * bmp.Size().y; + + for( size_t i=0; i<cnt; i++ ) + { + uint32_t c1 = *p1++; + uint32_t c2 = *p2++; + + err += sq( ( c1 & 0x000000FF ) - ( c2 & 0x000000FF ) ); + err += sq( ( ( c1 & 0x0000FF00 ) >> 8 ) - ( ( c2 & 0x0000FF00 ) >> 8 ) ); + err += sq( ( ( c1 & 0x00FF0000 ) >> 16 ) - ( ( c2 & 0x00FF0000 ) >> 16 ) ); + } + + err /= cnt * 3; + + return err; +} + +float CalcMSE1( const Bitmap& bmp, const Bitmap& out ) +{ + float err = 0; + + const uint32_t* p1 = bmp.Data(); + const uint32_t* p2 = out.Data(); + size_t cnt = bmp.Size().x * bmp.Size().y; + + for( size_t i=0; i<cnt; i++ ) + { + uint32_t c1 = *p1++; + uint32_t c2 = *p2++; + + err += sq( ( c1 >> 24 ) - ( c2 & 0xFF ) ); + } + + err /= cnt; + + return err; +} diff --git a/thirdparty/etcpak/Error.hpp b/thirdparty/etcpak/Error.hpp new file mode 100644 index 0000000000..9817754b74 --- /dev/null +++ b/thirdparty/etcpak/Error.hpp @@ -0,0 +1,9 @@ +#ifndef __ERROR_HPP__ +#define __ERROR_HPP__ + +#include "Bitmap.hpp" + +float CalcMSE3( const Bitmap& bmp, const Bitmap& out ); +float CalcMSE1( const Bitmap& bmp, const Bitmap& out ); + +#endif diff --git a/thirdparty/etcpak/ForceInline.hpp b/thirdparty/etcpak/ForceInline.hpp new file mode 100644 index 0000000000..b6f012841b --- /dev/null +++ b/thirdparty/etcpak/ForceInline.hpp @@ -0,0 +1,20 @@ +#ifndef __FORCEINLINE_HPP__ +#define __FORCEINLINE_HPP__ + +#if defined(__GNUC__) +# define etcpak_force_inline __attribute__((always_inline)) inline +#elif defined(_MSC_VER) +# define etcpak_force_inline __forceinline +#else +# define etcpak_force_inline inline +#endif + +#if defined(__GNUC__) +# define etcpak_no_inline __attribute__((noinline)) +#elif defined(_MSC_VER) +# define etcpak_no_inline __declspec(noinline) +#else +# define etcpak_no_inline +#endif + +#endif diff --git a/thirdparty/etcpak/LICENSE.txt b/thirdparty/etcpak/LICENSE.txt new file mode 100644 index 0000000000..59e85d6ea5 --- /dev/null +++ b/thirdparty/etcpak/LICENSE.txt @@ -0,0 +1,26 @@ +etcpak, an extremely fast ETC compression utility (https://github.com/wolfpld/etcpak) + +Copyright (c) 2013-2021, Bartosz Taudul <wolf@nereid.pl> +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + * Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright + notice, this list of conditions and the following disclaimer in the + documentation and/or other materials provided with the distribution. + * Neither the name of the <organization> nor the + names of its contributors may be used to endorse or promote products + derived from this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL <COPYRIGHT HOLDER> BE LIABLE FOR ANY +DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. diff --git a/thirdparty/etcpak/Math.hpp b/thirdparty/etcpak/Math.hpp new file mode 100644 index 0000000000..994e1ac4ea --- /dev/null +++ b/thirdparty/etcpak/Math.hpp @@ -0,0 +1,92 @@ +#ifndef __DARKRL__MATH_HPP__ +#define __DARKRL__MATH_HPP__ + +#include <algorithm> +#include <cmath> +#include <stdint.h> + +#include "ForceInline.hpp" + +template<typename T> +static etcpak_force_inline T AlignPOT( T val ) +{ + if( val == 0 ) return 1; + val--; + for( unsigned int i=1; i<sizeof( T ) * 8; i <<= 1 ) + { + val |= val >> i; + } + return val + 1; +} + +static etcpak_force_inline int CountSetBits( uint32_t val ) +{ + val -= ( val >> 1 ) & 0x55555555; + val = ( ( val >> 2 ) & 0x33333333 ) + ( val & 0x33333333 ); + val = ( ( val >> 4 ) + val ) & 0x0f0f0f0f; + val += val >> 8; + val += val >> 16; + return val & 0x0000003f; +} + +static etcpak_force_inline int CountLeadingZeros( uint32_t val ) +{ + val |= val >> 1; + val |= val >> 2; + val |= val >> 4; + val |= val >> 8; + val |= val >> 16; + return 32 - CountSetBits( val ); +} + +static etcpak_force_inline float sRGB2linear( float v ) +{ + const float a = 0.055f; + if( v <= 0.04045f ) + { + return v / 12.92f; + } + else + { + return pow( ( v + a ) / ( 1 + a ), 2.4f ); + } +} + +static etcpak_force_inline float linear2sRGB( float v ) +{ + const float a = 0.055f; + if( v <= 0.0031308f ) + { + return 12.92f * v; + } + else + { + return ( 1 + a ) * pow( v, 1/2.4f ) - a; + } +} + +template<class T> +static etcpak_force_inline T SmoothStep( T x ) +{ + return x*x*(3-2*x); +} + +static etcpak_force_inline uint8_t clampu8( int32_t val ) +{ + if( ( val & ~0xFF ) == 0 ) return val; + return ( ( ~val ) >> 31 ) & 0xFF; +} + +template<class T> +static etcpak_force_inline T sq( T val ) +{ + return val * val; +} + +static etcpak_force_inline int mul8bit( int a, int b ) +{ + int t = a*b + 128; + return ( t + ( t >> 8 ) ) >> 8; +} + +#endif diff --git a/thirdparty/etcpak/MipMap.hpp b/thirdparty/etcpak/MipMap.hpp new file mode 100644 index 0000000000..d3b4bc9e7c --- /dev/null +++ b/thirdparty/etcpak/MipMap.hpp @@ -0,0 +1,11 @@ +#ifndef __MIPMAP_HPP__ +#define __MIPMAP_HPP__ + +#include "Vector.hpp" + +inline int NumberOfMipLevels( const v2i& size ) +{ + return (int)floor( log2( std::max( size.x, size.y ) ) ) + 1; +} + +#endif diff --git a/thirdparty/etcpak/ProcessCommon.hpp b/thirdparty/etcpak/ProcessCommon.hpp new file mode 100644 index 0000000000..657d68888f --- /dev/null +++ b/thirdparty/etcpak/ProcessCommon.hpp @@ -0,0 +1,50 @@ +#ifndef __PROCESSCOMMON_HPP__ +#define __PROCESSCOMMON_HPP__ + +#include <assert.h> +#include <stddef.h> +#include <stdint.h> + +template<class T> +static size_t GetLeastError( const T* err, size_t num ) +{ + size_t idx = 0; + for( size_t i=1; i<num; i++ ) + { + if( err[i] < err[idx] ) + { + idx = i; + } + } + return idx; +} + +static uint64_t FixByteOrder( uint64_t d ) +{ + return ( ( d & 0x00000000FFFFFFFF ) ) | + ( ( d & 0xFF00000000000000 ) >> 24 ) | + ( ( d & 0x000000FF00000000 ) << 24 ) | + ( ( d & 0x00FF000000000000 ) >> 8 ) | + ( ( d & 0x0000FF0000000000 ) << 8 ); +} + +template<class T, class S> +static uint64_t EncodeSelectors( uint64_t d, const T terr[2][8], const S tsel[16][8], const uint32_t* id ) +{ + size_t tidx[2]; + tidx[0] = GetLeastError( terr[0], 8 ); + tidx[1] = GetLeastError( terr[1], 8 ); + + d |= tidx[0] << 26; + d |= tidx[1] << 29; + for( int i=0; i<16; i++ ) + { + uint64_t t = tsel[i][tidx[id[i]%2]]; + d |= ( t & 0x1 ) << ( i + 32 ); + d |= ( t & 0x2 ) << ( i + 47 ); + } + + return d; +} + +#endif diff --git a/thirdparty/etcpak/ProcessDxtc.cpp b/thirdparty/etcpak/ProcessDxtc.cpp new file mode 100644 index 0000000000..508d55fd75 --- /dev/null +++ b/thirdparty/etcpak/ProcessDxtc.cpp @@ -0,0 +1,956 @@ +#include "Dither.hpp" +#include "ForceInline.hpp" +#include "ProcessDxtc.hpp" + +#include <assert.h> +#include <stdint.h> +#include <string.h> + +#ifdef __ARM_NEON +# include <arm_neon.h> +#endif + +#if defined __AVX__ && !defined __SSE4_1__ +# define __SSE4_1__ +#endif + +#if defined __SSE4_1__ || defined __AVX2__ +# ifdef _MSC_VER +# include <intrin.h> +# else +# include <x86intrin.h> +# ifndef _mm256_cvtsi256_si32 +# define _mm256_cvtsi256_si32( v ) ( _mm_cvtsi128_si32( _mm256_castsi256_si128( v ) ) ) +# endif +# endif +#endif + + +static etcpak_force_inline uint16_t to565( uint8_t r, uint8_t g, uint8_t b ) +{ + return ( ( r & 0xF8 ) << 8 ) | ( ( g & 0xFC ) << 3 ) | ( b >> 3 ); +} + +static etcpak_force_inline uint16_t to565( uint32_t c ) +{ + return + ( ( c & 0xF80000 ) >> 19 ) | + ( ( c & 0x00FC00 ) >> 5 ) | + ( ( c & 0x0000F8 ) << 8 ); +} + +static const uint8_t DxtcIndexTable[256] = { + 85, 87, 86, 84, 93, 95, 94, 92, 89, 91, 90, 88, 81, 83, 82, 80, + 117, 119, 118, 116, 125, 127, 126, 124, 121, 123, 122, 120, 113, 115, 114, 112, + 101, 103, 102, 100, 109, 111, 110, 108, 105, 107, 106, 104, 97, 99, 98, 96, + 69, 71, 70, 68, 77, 79, 78, 76, 73, 75, 74, 72, 65, 67, 66, 64, + 213, 215, 214, 212, 221, 223, 222, 220, 217, 219, 218, 216, 209, 211, 210, 208, + 245, 247, 246, 244, 253, 255, 254, 252, 249, 251, 250, 248, 241, 243, 242, 240, + 229, 231, 230, 228, 237, 239, 238, 236, 233, 235, 234, 232, 225, 227, 226, 224, + 197, 199, 198, 196, 205, 207, 206, 204, 201, 203, 202, 200, 193, 195, 194, 192, + 149, 151, 150, 148, 157, 159, 158, 156, 153, 155, 154, 152, 145, 147, 146, 144, + 181, 183, 182, 180, 189, 191, 190, 188, 185, 187, 186, 184, 177, 179, 178, 176, + 165, 167, 166, 164, 173, 175, 174, 172, 169, 171, 170, 168, 161, 163, 162, 160, + 133, 135, 134, 132, 141, 143, 142, 140, 137, 139, 138, 136, 129, 131, 130, 128, + 21, 23, 22, 20, 29, 31, 30, 28, 25, 27, 26, 24, 17, 19, 18, 16, + 53, 55, 54, 52, 61, 63, 62, 60, 57, 59, 58, 56, 49, 51, 50, 48, + 37, 39, 38, 36, 45, 47, 46, 44, 41, 43, 42, 40, 33, 35, 34, 32, + 5, 7, 6, 4, 13, 15, 14, 12, 9, 11, 10, 8, 1, 3, 2, 0 +}; + +static const uint8_t AlphaIndexTable_SSE[64] = { + 9, 15, 14, 13, 12, 11, 10, 8, 57, 63, 62, 61, 60, 59, 58, 56, + 49, 55, 54, 53, 52, 51, 50, 48, 41, 47, 46, 45, 44, 43, 42, 40, + 33, 39, 38, 37, 36, 35, 34, 32, 25, 31, 30, 29, 28, 27, 26, 24, + 17, 23, 22, 21, 20, 19, 18, 16, 1, 7, 6, 5, 4, 3, 2, 0, +}; + +static const uint16_t DivTable[255*3+1] = { + 0xffff, 0xffff, 0xffff, 0xffff, 0xcccc, 0xaaaa, 0x9249, 0x8000, 0x71c7, 0x6666, 0x5d17, 0x5555, 0x4ec4, 0x4924, 0x4444, 0x4000, + 0x3c3c, 0x38e3, 0x35e5, 0x3333, 0x30c3, 0x2e8b, 0x2c85, 0x2aaa, 0x28f5, 0x2762, 0x25ed, 0x2492, 0x234f, 0x2222, 0x2108, 0x2000, + 0x1f07, 0x1e1e, 0x1d41, 0x1c71, 0x1bac, 0x1af2, 0x1a41, 0x1999, 0x18f9, 0x1861, 0x17d0, 0x1745, 0x16c1, 0x1642, 0x15c9, 0x1555, + 0x14e5, 0x147a, 0x1414, 0x13b1, 0x1352, 0x12f6, 0x129e, 0x1249, 0x11f7, 0x11a7, 0x115b, 0x1111, 0x10c9, 0x1084, 0x1041, 0x1000, + 0x0fc0, 0x0f83, 0x0f48, 0x0f0f, 0x0ed7, 0x0ea0, 0x0e6c, 0x0e38, 0x0e07, 0x0dd6, 0x0da7, 0x0d79, 0x0d4c, 0x0d20, 0x0cf6, 0x0ccc, + 0x0ca4, 0x0c7c, 0x0c56, 0x0c30, 0x0c0c, 0x0be8, 0x0bc5, 0x0ba2, 0x0b81, 0x0b60, 0x0b40, 0x0b21, 0x0b02, 0x0ae4, 0x0ac7, 0x0aaa, + 0x0a8e, 0x0a72, 0x0a57, 0x0a3d, 0x0a23, 0x0a0a, 0x09f1, 0x09d8, 0x09c0, 0x09a9, 0x0991, 0x097b, 0x0964, 0x094f, 0x0939, 0x0924, + 0x090f, 0x08fb, 0x08e7, 0x08d3, 0x08c0, 0x08ad, 0x089a, 0x0888, 0x0876, 0x0864, 0x0853, 0x0842, 0x0831, 0x0820, 0x0810, 0x0800, + 0x07f0, 0x07e0, 0x07d1, 0x07c1, 0x07b3, 0x07a4, 0x0795, 0x0787, 0x0779, 0x076b, 0x075d, 0x0750, 0x0743, 0x0736, 0x0729, 0x071c, + 0x070f, 0x0703, 0x06f7, 0x06eb, 0x06df, 0x06d3, 0x06c8, 0x06bc, 0x06b1, 0x06a6, 0x069b, 0x0690, 0x0685, 0x067b, 0x0670, 0x0666, + 0x065c, 0x0652, 0x0648, 0x063e, 0x0634, 0x062b, 0x0621, 0x0618, 0x060f, 0x0606, 0x05fd, 0x05f4, 0x05eb, 0x05e2, 0x05d9, 0x05d1, + 0x05c9, 0x05c0, 0x05b8, 0x05b0, 0x05a8, 0x05a0, 0x0598, 0x0590, 0x0588, 0x0581, 0x0579, 0x0572, 0x056b, 0x0563, 0x055c, 0x0555, + 0x054e, 0x0547, 0x0540, 0x0539, 0x0532, 0x052b, 0x0525, 0x051e, 0x0518, 0x0511, 0x050b, 0x0505, 0x04fe, 0x04f8, 0x04f2, 0x04ec, + 0x04e6, 0x04e0, 0x04da, 0x04d4, 0x04ce, 0x04c8, 0x04c3, 0x04bd, 0x04b8, 0x04b2, 0x04ad, 0x04a7, 0x04a2, 0x049c, 0x0497, 0x0492, + 0x048d, 0x0487, 0x0482, 0x047d, 0x0478, 0x0473, 0x046e, 0x0469, 0x0465, 0x0460, 0x045b, 0x0456, 0x0452, 0x044d, 0x0448, 0x0444, + 0x043f, 0x043b, 0x0436, 0x0432, 0x042d, 0x0429, 0x0425, 0x0421, 0x041c, 0x0418, 0x0414, 0x0410, 0x040c, 0x0408, 0x0404, 0x0400, + 0x03fc, 0x03f8, 0x03f4, 0x03f0, 0x03ec, 0x03e8, 0x03e4, 0x03e0, 0x03dd, 0x03d9, 0x03d5, 0x03d2, 0x03ce, 0x03ca, 0x03c7, 0x03c3, + 0x03c0, 0x03bc, 0x03b9, 0x03b5, 0x03b2, 0x03ae, 0x03ab, 0x03a8, 0x03a4, 0x03a1, 0x039e, 0x039b, 0x0397, 0x0394, 0x0391, 0x038e, + 0x038b, 0x0387, 0x0384, 0x0381, 0x037e, 0x037b, 0x0378, 0x0375, 0x0372, 0x036f, 0x036c, 0x0369, 0x0366, 0x0364, 0x0361, 0x035e, + 0x035b, 0x0358, 0x0355, 0x0353, 0x0350, 0x034d, 0x034a, 0x0348, 0x0345, 0x0342, 0x0340, 0x033d, 0x033a, 0x0338, 0x0335, 0x0333, + 0x0330, 0x032e, 0x032b, 0x0329, 0x0326, 0x0324, 0x0321, 0x031f, 0x031c, 0x031a, 0x0317, 0x0315, 0x0313, 0x0310, 0x030e, 0x030c, + 0x0309, 0x0307, 0x0305, 0x0303, 0x0300, 0x02fe, 0x02fc, 0x02fa, 0x02f7, 0x02f5, 0x02f3, 0x02f1, 0x02ef, 0x02ec, 0x02ea, 0x02e8, + 0x02e6, 0x02e4, 0x02e2, 0x02e0, 0x02de, 0x02dc, 0x02da, 0x02d8, 0x02d6, 0x02d4, 0x02d2, 0x02d0, 0x02ce, 0x02cc, 0x02ca, 0x02c8, + 0x02c6, 0x02c4, 0x02c2, 0x02c0, 0x02be, 0x02bc, 0x02bb, 0x02b9, 0x02b7, 0x02b5, 0x02b3, 0x02b1, 0x02b0, 0x02ae, 0x02ac, 0x02aa, + 0x02a8, 0x02a7, 0x02a5, 0x02a3, 0x02a1, 0x02a0, 0x029e, 0x029c, 0x029b, 0x0299, 0x0297, 0x0295, 0x0294, 0x0292, 0x0291, 0x028f, + 0x028d, 0x028c, 0x028a, 0x0288, 0x0287, 0x0285, 0x0284, 0x0282, 0x0280, 0x027f, 0x027d, 0x027c, 0x027a, 0x0279, 0x0277, 0x0276, + 0x0274, 0x0273, 0x0271, 0x0270, 0x026e, 0x026d, 0x026b, 0x026a, 0x0268, 0x0267, 0x0265, 0x0264, 0x0263, 0x0261, 0x0260, 0x025e, + 0x025d, 0x025c, 0x025a, 0x0259, 0x0257, 0x0256, 0x0255, 0x0253, 0x0252, 0x0251, 0x024f, 0x024e, 0x024d, 0x024b, 0x024a, 0x0249, + 0x0247, 0x0246, 0x0245, 0x0243, 0x0242, 0x0241, 0x0240, 0x023e, 0x023d, 0x023c, 0x023b, 0x0239, 0x0238, 0x0237, 0x0236, 0x0234, + 0x0233, 0x0232, 0x0231, 0x0230, 0x022e, 0x022d, 0x022c, 0x022b, 0x022a, 0x0229, 0x0227, 0x0226, 0x0225, 0x0224, 0x0223, 0x0222, + 0x0220, 0x021f, 0x021e, 0x021d, 0x021c, 0x021b, 0x021a, 0x0219, 0x0218, 0x0216, 0x0215, 0x0214, 0x0213, 0x0212, 0x0211, 0x0210, + 0x020f, 0x020e, 0x020d, 0x020c, 0x020b, 0x020a, 0x0209, 0x0208, 0x0207, 0x0206, 0x0205, 0x0204, 0x0203, 0x0202, 0x0201, 0x0200, + 0x01ff, 0x01fe, 0x01fd, 0x01fc, 0x01fb, 0x01fa, 0x01f9, 0x01f8, 0x01f7, 0x01f6, 0x01f5, 0x01f4, 0x01f3, 0x01f2, 0x01f1, 0x01f0, + 0x01ef, 0x01ee, 0x01ed, 0x01ec, 0x01eb, 0x01ea, 0x01e9, 0x01e9, 0x01e8, 0x01e7, 0x01e6, 0x01e5, 0x01e4, 0x01e3, 0x01e2, 0x01e1, + 0x01e0, 0x01e0, 0x01df, 0x01de, 0x01dd, 0x01dc, 0x01db, 0x01da, 0x01da, 0x01d9, 0x01d8, 0x01d7, 0x01d6, 0x01d5, 0x01d4, 0x01d4, + 0x01d3, 0x01d2, 0x01d1, 0x01d0, 0x01cf, 0x01cf, 0x01ce, 0x01cd, 0x01cc, 0x01cb, 0x01cb, 0x01ca, 0x01c9, 0x01c8, 0x01c7, 0x01c7, + 0x01c6, 0x01c5, 0x01c4, 0x01c3, 0x01c3, 0x01c2, 0x01c1, 0x01c0, 0x01c0, 0x01bf, 0x01be, 0x01bd, 0x01bd, 0x01bc, 0x01bb, 0x01ba, + 0x01ba, 0x01b9, 0x01b8, 0x01b7, 0x01b7, 0x01b6, 0x01b5, 0x01b4, 0x01b4, 0x01b3, 0x01b2, 0x01b2, 0x01b1, 0x01b0, 0x01af, 0x01af, + 0x01ae, 0x01ad, 0x01ad, 0x01ac, 0x01ab, 0x01aa, 0x01aa, 0x01a9, 0x01a8, 0x01a8, 0x01a7, 0x01a6, 0x01a6, 0x01a5, 0x01a4, 0x01a4, + 0x01a3, 0x01a2, 0x01a2, 0x01a1, 0x01a0, 0x01a0, 0x019f, 0x019e, 0x019e, 0x019d, 0x019c, 0x019c, 0x019b, 0x019a, 0x019a, 0x0199, + 0x0198, 0x0198, 0x0197, 0x0197, 0x0196, 0x0195, 0x0195, 0x0194, 0x0193, 0x0193, 0x0192, 0x0192, 0x0191, 0x0190, 0x0190, 0x018f, + 0x018f, 0x018e, 0x018d, 0x018d, 0x018c, 0x018b, 0x018b, 0x018a, 0x018a, 0x0189, 0x0189, 0x0188, 0x0187, 0x0187, 0x0186, 0x0186, + 0x0185, 0x0184, 0x0184, 0x0183, 0x0183, 0x0182, 0x0182, 0x0181, 0x0180, 0x0180, 0x017f, 0x017f, 0x017e, 0x017e, 0x017d, 0x017d, + 0x017c, 0x017b, 0x017b, 0x017a, 0x017a, 0x0179, 0x0179, 0x0178, 0x0178, 0x0177, 0x0177, 0x0176, 0x0175, 0x0175, 0x0174, 0x0174, + 0x0173, 0x0173, 0x0172, 0x0172, 0x0171, 0x0171, 0x0170, 0x0170, 0x016f, 0x016f, 0x016e, 0x016e, 0x016d, 0x016d, 0x016c, 0x016c, + 0x016b, 0x016b, 0x016a, 0x016a, 0x0169, 0x0169, 0x0168, 0x0168, 0x0167, 0x0167, 0x0166, 0x0166, 0x0165, 0x0165, 0x0164, 0x0164, + 0x0163, 0x0163, 0x0162, 0x0162, 0x0161, 0x0161, 0x0160, 0x0160, 0x015f, 0x015f, 0x015e, 0x015e, 0x015d, 0x015d, 0x015d, 0x015c, + 0x015c, 0x015b, 0x015b, 0x015a, 0x015a, 0x0159, 0x0159, 0x0158, 0x0158, 0x0158, 0x0157, 0x0157, 0x0156, 0x0156 +}; +static const uint16_t DivTableNEON[255*3+1] = { + 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, + 0x0000, 0x1c71, 0x1af2, 0x1999, 0x1861, 0x1745, 0x1642, 0x1555, 0x147a, 0x13b1, 0x12f6, 0x1249, 0x11a7, 0x1111, 0x1084, 0x1000, + 0x0f83, 0x0f0f, 0x0ea0, 0x0e38, 0x0dd6, 0x0d79, 0x0d20, 0x0ccc, 0x0c7c, 0x0c30, 0x0be8, 0x0ba2, 0x0b60, 0x0b21, 0x0ae4, 0x0aaa, + 0x0a72, 0x0a3d, 0x0a0a, 0x09d8, 0x09a9, 0x097b, 0x094f, 0x0924, 0x08fb, 0x08d3, 0x08ad, 0x0888, 0x0864, 0x0842, 0x0820, 0x0800, + 0x07e0, 0x07c1, 0x07a4, 0x0787, 0x076b, 0x0750, 0x0736, 0x071c, 0x0703, 0x06eb, 0x06d3, 0x06bc, 0x06a6, 0x0690, 0x067b, 0x0666, + 0x0652, 0x063e, 0x062b, 0x0618, 0x0606, 0x05f4, 0x05e2, 0x05d1, 0x05c0, 0x05b0, 0x05a0, 0x0590, 0x0581, 0x0572, 0x0563, 0x0555, + 0x0547, 0x0539, 0x052b, 0x051e, 0x0511, 0x0505, 0x04f8, 0x04ec, 0x04e0, 0x04d4, 0x04c8, 0x04bd, 0x04b2, 0x04a7, 0x049c, 0x0492, + 0x0487, 0x047d, 0x0473, 0x0469, 0x0460, 0x0456, 0x044d, 0x0444, 0x043b, 0x0432, 0x0429, 0x0421, 0x0418, 0x0410, 0x0408, 0x0400, + 0x03f8, 0x03f0, 0x03e8, 0x03e0, 0x03d9, 0x03d2, 0x03ca, 0x03c3, 0x03bc, 0x03b5, 0x03ae, 0x03a8, 0x03a1, 0x039b, 0x0394, 0x038e, + 0x0387, 0x0381, 0x037b, 0x0375, 0x036f, 0x0369, 0x0364, 0x035e, 0x0358, 0x0353, 0x034d, 0x0348, 0x0342, 0x033d, 0x0338, 0x0333, + 0x032e, 0x0329, 0x0324, 0x031f, 0x031a, 0x0315, 0x0310, 0x030c, 0x0307, 0x0303, 0x02fe, 0x02fa, 0x02f5, 0x02f1, 0x02ec, 0x02e8, + 0x02e4, 0x02e0, 0x02dc, 0x02d8, 0x02d4, 0x02d0, 0x02cc, 0x02c8, 0x02c4, 0x02c0, 0x02bc, 0x02b9, 0x02b5, 0x02b1, 0x02ae, 0x02aa, + 0x02a7, 0x02a3, 0x02a0, 0x029c, 0x0299, 0x0295, 0x0292, 0x028f, 0x028c, 0x0288, 0x0285, 0x0282, 0x027f, 0x027c, 0x0279, 0x0276, + 0x0273, 0x0270, 0x026d, 0x026a, 0x0267, 0x0264, 0x0261, 0x025e, 0x025c, 0x0259, 0x0256, 0x0253, 0x0251, 0x024e, 0x024b, 0x0249, + 0x0246, 0x0243, 0x0241, 0x023e, 0x023c, 0x0239, 0x0237, 0x0234, 0x0232, 0x0230, 0x022d, 0x022b, 0x0229, 0x0226, 0x0224, 0x0222, + 0x021f, 0x021d, 0x021b, 0x0219, 0x0216, 0x0214, 0x0212, 0x0210, 0x020e, 0x020c, 0x020a, 0x0208, 0x0206, 0x0204, 0x0202, 0x0200, + 0x01fe, 0x01fc, 0x01fa, 0x01f8, 0x01f6, 0x01f4, 0x01f2, 0x01f0, 0x01ee, 0x01ec, 0x01ea, 0x01e9, 0x01e7, 0x01e5, 0x01e3, 0x01e1, + 0x01e0, 0x01de, 0x01dc, 0x01da, 0x01d9, 0x01d7, 0x01d5, 0x01d4, 0x01d2, 0x01d0, 0x01cf, 0x01cd, 0x01cb, 0x01ca, 0x01c8, 0x01c7, + 0x01c5, 0x01c3, 0x01c2, 0x01c0, 0x01bf, 0x01bd, 0x01bc, 0x01ba, 0x01b9, 0x01b7, 0x01b6, 0x01b4, 0x01b3, 0x01b2, 0x01b0, 0x01af, + 0x01ad, 0x01ac, 0x01aa, 0x01a9, 0x01a8, 0x01a6, 0x01a5, 0x01a4, 0x01a2, 0x01a1, 0x01a0, 0x019e, 0x019d, 0x019c, 0x019a, 0x0199, + 0x0198, 0x0197, 0x0195, 0x0194, 0x0193, 0x0192, 0x0190, 0x018f, 0x018e, 0x018d, 0x018b, 0x018a, 0x0189, 0x0188, 0x0187, 0x0186, + 0x0184, 0x0183, 0x0182, 0x0181, 0x0180, 0x017f, 0x017e, 0x017d, 0x017b, 0x017a, 0x0179, 0x0178, 0x0177, 0x0176, 0x0175, 0x0174, + 0x0173, 0x0172, 0x0171, 0x0170, 0x016f, 0x016e, 0x016d, 0x016c, 0x016b, 0x016a, 0x0169, 0x0168, 0x0167, 0x0166, 0x0165, 0x0164, + 0x0163, 0x0162, 0x0161, 0x0160, 0x015f, 0x015e, 0x015d, 0x015c, 0x015b, 0x015a, 0x0159, 0x0158, 0x0158, 0x0157, 0x0156, 0x0155, + 0x0154, 0x0153, 0x0152, 0x0151, 0x0150, 0x0150, 0x014f, 0x014e, 0x014d, 0x014c, 0x014b, 0x014a, 0x014a, 0x0149, 0x0148, 0x0147, + 0x0146, 0x0146, 0x0145, 0x0144, 0x0143, 0x0142, 0x0142, 0x0141, 0x0140, 0x013f, 0x013e, 0x013e, 0x013d, 0x013c, 0x013b, 0x013b, + 0x013a, 0x0139, 0x0138, 0x0138, 0x0137, 0x0136, 0x0135, 0x0135, 0x0134, 0x0133, 0x0132, 0x0132, 0x0131, 0x0130, 0x0130, 0x012f, + 0x012e, 0x012e, 0x012d, 0x012c, 0x012b, 0x012b, 0x012a, 0x0129, 0x0129, 0x0128, 0x0127, 0x0127, 0x0126, 0x0125, 0x0125, 0x0124, + 0x0123, 0x0123, 0x0122, 0x0121, 0x0121, 0x0120, 0x0120, 0x011f, 0x011e, 0x011e, 0x011d, 0x011c, 0x011c, 0x011b, 0x011b, 0x011a, + 0x0119, 0x0119, 0x0118, 0x0118, 0x0117, 0x0116, 0x0116, 0x0115, 0x0115, 0x0114, 0x0113, 0x0113, 0x0112, 0x0112, 0x0111, 0x0111, + 0x0110, 0x010f, 0x010f, 0x010e, 0x010e, 0x010d, 0x010d, 0x010c, 0x010c, 0x010b, 0x010a, 0x010a, 0x0109, 0x0109, 0x0108, 0x0108, + 0x0107, 0x0107, 0x0106, 0x0106, 0x0105, 0x0105, 0x0104, 0x0104, 0x0103, 0x0103, 0x0102, 0x0102, 0x0101, 0x0101, 0x0100, 0x0100, + 0x00ff, 0x00ff, 0x00fe, 0x00fe, 0x00fd, 0x00fd, 0x00fc, 0x00fc, 0x00fb, 0x00fb, 0x00fa, 0x00fa, 0x00f9, 0x00f9, 0x00f8, 0x00f8, + 0x00f7, 0x00f7, 0x00f6, 0x00f6, 0x00f5, 0x00f5, 0x00f4, 0x00f4, 0x00f4, 0x00f3, 0x00f3, 0x00f2, 0x00f2, 0x00f1, 0x00f1, 0x00f0, + 0x00f0, 0x00f0, 0x00ef, 0x00ef, 0x00ee, 0x00ee, 0x00ed, 0x00ed, 0x00ed, 0x00ec, 0x00ec, 0x00eb, 0x00eb, 0x00ea, 0x00ea, 0x00ea, + 0x00e9, 0x00e9, 0x00e8, 0x00e8, 0x00e7, 0x00e7, 0x00e7, 0x00e6, 0x00e6, 0x00e5, 0x00e5, 0x00e5, 0x00e4, 0x00e4, 0x00e3, 0x00e3, + 0x00e3, 0x00e2, 0x00e2, 0x00e1, 0x00e1, 0x00e1, 0x00e0, 0x00e0, 0x00e0, 0x00df, 0x00df, 0x00de, 0x00de, 0x00de, 0x00dd, 0x00dd, + 0x00dd, 0x00dc, 0x00dc, 0x00db, 0x00db, 0x00db, 0x00da, 0x00da, 0x00da, 0x00d9, 0x00d9, 0x00d9, 0x00d8, 0x00d8, 0x00d7, 0x00d7, + 0x00d7, 0x00d6, 0x00d6, 0x00d6, 0x00d5, 0x00d5, 0x00d5, 0x00d4, 0x00d4, 0x00d4, 0x00d3, 0x00d3, 0x00d3, 0x00d2, 0x00d2, 0x00d2, + 0x00d1, 0x00d1, 0x00d1, 0x00d0, 0x00d0, 0x00d0, 0x00cf, 0x00cf, 0x00cf, 0x00ce, 0x00ce, 0x00ce, 0x00cd, 0x00cd, 0x00cd, 0x00cc, + 0x00cc, 0x00cc, 0x00cb, 0x00cb, 0x00cb, 0x00ca, 0x00ca, 0x00ca, 0x00c9, 0x00c9, 0x00c9, 0x00c9, 0x00c8, 0x00c8, 0x00c8, 0x00c7, + 0x00c7, 0x00c7, 0x00c6, 0x00c6, 0x00c6, 0x00c5, 0x00c5, 0x00c5, 0x00c5, 0x00c4, 0x00c4, 0x00c4, 0x00c3, 0x00c3, 0x00c3, 0x00c3, + 0x00c2, 0x00c2, 0x00c2, 0x00c1, 0x00c1, 0x00c1, 0x00c1, 0x00c0, 0x00c0, 0x00c0, 0x00bf, 0x00bf, 0x00bf, 0x00bf, 0x00be, 0x00be, + 0x00be, 0x00bd, 0x00bd, 0x00bd, 0x00bd, 0x00bc, 0x00bc, 0x00bc, 0x00bc, 0x00bb, 0x00bb, 0x00bb, 0x00ba, 0x00ba, 0x00ba, 0x00ba, + 0x00b9, 0x00b9, 0x00b9, 0x00b9, 0x00b8, 0x00b8, 0x00b8, 0x00b8, 0x00b7, 0x00b7, 0x00b7, 0x00b7, 0x00b6, 0x00b6, 0x00b6, 0x00b6, + 0x00b5, 0x00b5, 0x00b5, 0x00b5, 0x00b4, 0x00b4, 0x00b4, 0x00b4, 0x00b3, 0x00b3, 0x00b3, 0x00b3, 0x00b2, 0x00b2, 0x00b2, 0x00b2, + 0x00b1, 0x00b1, 0x00b1, 0x00b1, 0x00b0, 0x00b0, 0x00b0, 0x00b0, 0x00af, 0x00af, 0x00af, 0x00af, 0x00ae, 0x00ae, 0x00ae, 0x00ae, + 0x00ae, 0x00ad, 0x00ad, 0x00ad, 0x00ad, 0x00ac, 0x00ac, 0x00ac, 0x00ac, 0x00ac, 0x00ab, 0x00ab, 0x00ab, 0x00ab, +}; + +static const uint16_t DivTableAlpha[256] = { + 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xe38e, 0xcccc, 0xba2e, 0xaaaa, 0x9d89, 0x9249, 0x8888, 0x8000, + 0x7878, 0x71c7, 0x6bca, 0x6666, 0x6186, 0x5d17, 0x590b, 0x5555, 0x51eb, 0x4ec4, 0x4bda, 0x4924, 0x469e, 0x4444, 0x4210, 0x4000, + 0x3e0f, 0x3c3c, 0x3a83, 0x38e3, 0x3759, 0x35e5, 0x3483, 0x3333, 0x31f3, 0x30c3, 0x2fa0, 0x2e8b, 0x2d82, 0x2c85, 0x2b93, 0x2aaa, + 0x29cb, 0x28f5, 0x2828, 0x2762, 0x26a4, 0x25ed, 0x253c, 0x2492, 0x23ee, 0x234f, 0x22b6, 0x2222, 0x2192, 0x2108, 0x2082, 0x2000, + 0x1f81, 0x1f07, 0x1e91, 0x1e1e, 0x1dae, 0x1d41, 0x1cd8, 0x1c71, 0x1c0e, 0x1bac, 0x1b4e, 0x1af2, 0x1a98, 0x1a41, 0x19ec, 0x1999, + 0x1948, 0x18f9, 0x18ac, 0x1861, 0x1818, 0x17d0, 0x178a, 0x1745, 0x1702, 0x16c1, 0x1681, 0x1642, 0x1605, 0x15c9, 0x158e, 0x1555, + 0x151d, 0x14e5, 0x14af, 0x147a, 0x1446, 0x1414, 0x13e2, 0x13b1, 0x1381, 0x1352, 0x1323, 0x12f6, 0x12c9, 0x129e, 0x1273, 0x1249, + 0x121f, 0x11f7, 0x11cf, 0x11a7, 0x1181, 0x115b, 0x1135, 0x1111, 0x10ec, 0x10c9, 0x10a6, 0x1084, 0x1062, 0x1041, 0x1020, 0x1000, + 0x0fe0, 0x0fc0, 0x0fa2, 0x0f83, 0x0f66, 0x0f48, 0x0f2b, 0x0f0f, 0x0ef2, 0x0ed7, 0x0ebb, 0x0ea0, 0x0e86, 0x0e6c, 0x0e52, 0x0e38, + 0x0e1f, 0x0e07, 0x0dee, 0x0dd6, 0x0dbe, 0x0da7, 0x0d90, 0x0d79, 0x0d62, 0x0d4c, 0x0d36, 0x0d20, 0x0d0b, 0x0cf6, 0x0ce1, 0x0ccc, + 0x0cb8, 0x0ca4, 0x0c90, 0x0c7c, 0x0c69, 0x0c56, 0x0c43, 0x0c30, 0x0c1e, 0x0c0c, 0x0bfa, 0x0be8, 0x0bd6, 0x0bc5, 0x0bb3, 0x0ba2, + 0x0b92, 0x0b81, 0x0b70, 0x0b60, 0x0b50, 0x0b40, 0x0b30, 0x0b21, 0x0b11, 0x0b02, 0x0af3, 0x0ae4, 0x0ad6, 0x0ac7, 0x0ab8, 0x0aaa, + 0x0a9c, 0x0a8e, 0x0a80, 0x0a72, 0x0a65, 0x0a57, 0x0a4a, 0x0a3d, 0x0a30, 0x0a23, 0x0a16, 0x0a0a, 0x09fd, 0x09f1, 0x09e4, 0x09d8, + 0x09cc, 0x09c0, 0x09b4, 0x09a9, 0x099d, 0x0991, 0x0986, 0x097b, 0x0970, 0x0964, 0x095a, 0x094f, 0x0944, 0x0939, 0x092f, 0x0924, + 0x091a, 0x090f, 0x0905, 0x08fb, 0x08f1, 0x08e7, 0x08dd, 0x08d3, 0x08ca, 0x08c0, 0x08b7, 0x08ad, 0x08a4, 0x089a, 0x0891, 0x0888, + 0x087f, 0x0876, 0x086d, 0x0864, 0x085b, 0x0853, 0x084a, 0x0842, 0x0839, 0x0831, 0x0828, 0x0820, 0x0818, 0x0810, 0x0808, 0x0800, +}; + +static etcpak_force_inline uint64_t ProcessRGB( const uint8_t* src ) +{ +#ifdef __SSE4_1__ + __m128i px0 = _mm_loadu_si128(((__m128i*)src) + 0); + __m128i px1 = _mm_loadu_si128(((__m128i*)src) + 1); + __m128i px2 = _mm_loadu_si128(((__m128i*)src) + 2); + __m128i px3 = _mm_loadu_si128(((__m128i*)src) + 3); + + __m128i smask = _mm_set1_epi32( 0xF8FCF8 ); + __m128i sd0 = _mm_and_si128( px0, smask ); + __m128i sd1 = _mm_and_si128( px1, smask ); + __m128i sd2 = _mm_and_si128( px2, smask ); + __m128i sd3 = _mm_and_si128( px3, smask ); + + __m128i sc = _mm_shuffle_epi32(sd0, _MM_SHUFFLE(0, 0, 0, 0)); + + __m128i sc0 = _mm_cmpeq_epi8(sd0, sc); + __m128i sc1 = _mm_cmpeq_epi8(sd1, sc); + __m128i sc2 = _mm_cmpeq_epi8(sd2, sc); + __m128i sc3 = _mm_cmpeq_epi8(sd3, sc); + + __m128i sm0 = _mm_and_si128(sc0, sc1); + __m128i sm1 = _mm_and_si128(sc2, sc3); + __m128i sm = _mm_and_si128(sm0, sm1); + + if( _mm_testc_si128(sm, _mm_set1_epi32(-1)) ) + { + uint32_t c; + memcpy( &c, src, 4 ); + return uint64_t( to565( c ) ) << 16; + } + + __m128i min0 = _mm_min_epu8( px0, px1 ); + __m128i min1 = _mm_min_epu8( px2, px3 ); + __m128i min2 = _mm_min_epu8( min0, min1 ); + + __m128i max0 = _mm_max_epu8( px0, px1 ); + __m128i max1 = _mm_max_epu8( px2, px3 ); + __m128i max2 = _mm_max_epu8( max0, max1 ); + + __m128i min3 = _mm_shuffle_epi32( min2, _MM_SHUFFLE( 2, 3, 0, 1 ) ); + __m128i max3 = _mm_shuffle_epi32( max2, _MM_SHUFFLE( 2, 3, 0, 1 ) ); + __m128i min4 = _mm_min_epu8( min2, min3 ); + __m128i max4 = _mm_max_epu8( max2, max3 ); + + __m128i min5 = _mm_shuffle_epi32( min4, _MM_SHUFFLE( 0, 0, 2, 2 ) ); + __m128i max5 = _mm_shuffle_epi32( max4, _MM_SHUFFLE( 0, 0, 2, 2 ) ); + __m128i rmin = _mm_min_epu8( min4, min5 ); + __m128i rmax = _mm_max_epu8( max4, max5 ); + + __m128i range1 = _mm_subs_epu8( rmax, rmin ); + __m128i range2 = _mm_sad_epu8( rmax, rmin ); + + uint32_t vrange = _mm_cvtsi128_si32( range2 ) >> 1; + __m128i range = _mm_set1_epi16( DivTable[vrange] ); + + __m128i inset1 = _mm_srli_epi16( range1, 4 ); + __m128i inset = _mm_and_si128( inset1, _mm_set1_epi8( 0xF ) ); + __m128i min = _mm_adds_epu8( rmin, inset ); + __m128i max = _mm_subs_epu8( rmax, inset ); + + __m128i c0 = _mm_subs_epu8( px0, rmin ); + __m128i c1 = _mm_subs_epu8( px1, rmin ); + __m128i c2 = _mm_subs_epu8( px2, rmin ); + __m128i c3 = _mm_subs_epu8( px3, rmin ); + + __m128i is0 = _mm_maddubs_epi16( c0, _mm_set1_epi8( 1 ) ); + __m128i is1 = _mm_maddubs_epi16( c1, _mm_set1_epi8( 1 ) ); + __m128i is2 = _mm_maddubs_epi16( c2, _mm_set1_epi8( 1 ) ); + __m128i is3 = _mm_maddubs_epi16( c3, _mm_set1_epi8( 1 ) ); + + __m128i s0 = _mm_hadd_epi16( is0, is1 ); + __m128i s1 = _mm_hadd_epi16( is2, is3 ); + + __m128i m0 = _mm_mulhi_epu16( s0, range ); + __m128i m1 = _mm_mulhi_epu16( s1, range ); + + __m128i p0 = _mm_packus_epi16( m0, m1 ); + + __m128i p1 = _mm_or_si128( _mm_srai_epi32( p0, 6 ), _mm_srai_epi32( p0, 12 ) ); + __m128i p2 = _mm_or_si128( _mm_srai_epi32( p0, 18 ), p0 ); + __m128i p3 = _mm_or_si128( p1, p2 ); + __m128i p =_mm_shuffle_epi8( p3, _mm_set1_epi32( 0x0C080400 ) ); + + uint32_t vmin = _mm_cvtsi128_si32( min ); + uint32_t vmax = _mm_cvtsi128_si32( max ); + uint32_t vp = _mm_cvtsi128_si32( p ); + + return uint64_t( ( uint64_t( to565( vmin ) ) << 16 ) | to565( vmax ) | ( uint64_t( vp ) << 32 ) ); +#elif defined __ARM_NEON +# ifdef __aarch64__ + uint8x16x4_t px = vld4q_u8( src ); + + uint8x16_t lr = px.val[0]; + uint8x16_t lg = px.val[1]; + uint8x16_t lb = px.val[2]; + + uint8_t rmaxr = vmaxvq_u8( lr ); + uint8_t rmaxg = vmaxvq_u8( lg ); + uint8_t rmaxb = vmaxvq_u8( lb ); + + uint8_t rminr = vminvq_u8( lr ); + uint8_t rming = vminvq_u8( lg ); + uint8_t rminb = vminvq_u8( lb ); + + int rr = rmaxr - rminr; + int rg = rmaxg - rming; + int rb = rmaxb - rminb; + + int vrange1 = rr + rg + rb; + uint16_t vrange2 = DivTableNEON[vrange1]; + + uint8_t insetr = rr >> 4; + uint8_t insetg = rg >> 4; + uint8_t insetb = rb >> 4; + + uint8_t minr = rminr + insetr; + uint8_t ming = rming + insetg; + uint8_t minb = rminb + insetb; + + uint8_t maxr = rmaxr - insetr; + uint8_t maxg = rmaxg - insetg; + uint8_t maxb = rmaxb - insetb; + + uint8x16_t cr = vsubq_u8( lr, vdupq_n_u8( rminr ) ); + uint8x16_t cg = vsubq_u8( lg, vdupq_n_u8( rming ) ); + uint8x16_t cb = vsubq_u8( lb, vdupq_n_u8( rminb ) ); + + uint16x8_t is0l = vaddl_u8( vget_low_u8( cr ), vget_low_u8( cg ) ); + uint16x8_t is0h = vaddl_u8( vget_high_u8( cr ), vget_high_u8( cg ) ); + uint16x8_t is1l = vaddw_u8( is0l, vget_low_u8( cb ) ); + uint16x8_t is1h = vaddw_u8( is0h, vget_high_u8( cb ) ); + + int16x8_t range = vdupq_n_s16( vrange2 ); + uint16x8_t m0 = vreinterpretq_u16_s16( vqdmulhq_s16( vreinterpretq_s16_u16( is1l ), range ) ); + uint16x8_t m1 = vreinterpretq_u16_s16( vqdmulhq_s16( vreinterpretq_s16_u16( is1h ), range ) ); + + uint8x8_t p00 = vmovn_u16( m0 ); + uint8x8_t p01 = vmovn_u16( m1 ); + uint8x16_t p0 = vcombine_u8( p00, p01 ); + + uint32x4_t p1 = vaddq_u32( vshrq_n_u32( vreinterpretq_u32_u8( p0 ), 6 ), vshrq_n_u32( vreinterpretq_u32_u8( p0 ), 12 ) ); + uint32x4_t p2 = vaddq_u32( vshrq_n_u32( vreinterpretq_u32_u8( p0 ), 18 ), vreinterpretq_u32_u8( p0 ) ); + uint32x4_t p3 = vaddq_u32( p1, p2 ); + + uint16x4x2_t p4 = vuzp_u16( vget_low_u16( vreinterpretq_u16_u32( p3 ) ), vget_high_u16( vreinterpretq_u16_u32( p3 ) ) ); + uint8x8x2_t p = vuzp_u8( vreinterpret_u8_u16( p4.val[0] ), vreinterpret_u8_u16( p4.val[0] ) ); + + uint32_t vp; + vst1_lane_u32( &vp, vreinterpret_u32_u8( p.val[0] ), 0 ); + + return uint64_t( ( uint64_t( to565( minr, ming, minb ) ) << 16 ) | to565( maxr, maxg, maxb ) | ( uint64_t( vp ) << 32 ) ); +# else + uint32x4_t px0 = vld1q_u32( (uint32_t*)src ); + uint32x4_t px1 = vld1q_u32( (uint32_t*)src + 4 ); + uint32x4_t px2 = vld1q_u32( (uint32_t*)src + 8 ); + uint32x4_t px3 = vld1q_u32( (uint32_t*)src + 12 ); + + uint32x4_t smask = vdupq_n_u32( 0xF8FCF8 ); + uint32x4_t sd0 = vandq_u32( smask, px0 ); + uint32x4_t sd1 = vandq_u32( smask, px1 ); + uint32x4_t sd2 = vandq_u32( smask, px2 ); + uint32x4_t sd3 = vandq_u32( smask, px3 ); + + uint32x4_t sc = vdupq_n_u32( sd0[0] ); + + uint32x4_t sc0 = vceqq_u32( sd0, sc ); + uint32x4_t sc1 = vceqq_u32( sd1, sc ); + uint32x4_t sc2 = vceqq_u32( sd2, sc ); + uint32x4_t sc3 = vceqq_u32( sd3, sc ); + + uint32x4_t sm0 = vandq_u32( sc0, sc1 ); + uint32x4_t sm1 = vandq_u32( sc2, sc3 ); + int64x2_t sm = vreinterpretq_s64_u32( vandq_u32( sm0, sm1 ) ); + + if( sm[0] == -1 && sm[1] == -1 ) + { + return uint64_t( to565( src[0], src[1], src[2] ) ) << 16; + } + + uint32x4_t mask = vdupq_n_u32( 0xFFFFFF ); + uint8x16_t l0 = vreinterpretq_u8_u32( vandq_u32( mask, px0 ) ); + uint8x16_t l1 = vreinterpretq_u8_u32( vandq_u32( mask, px1 ) ); + uint8x16_t l2 = vreinterpretq_u8_u32( vandq_u32( mask, px2 ) ); + uint8x16_t l3 = vreinterpretq_u8_u32( vandq_u32( mask, px3 ) ); + + uint8x16_t min0 = vminq_u8( l0, l1 ); + uint8x16_t min1 = vminq_u8( l2, l3 ); + uint8x16_t min2 = vminq_u8( min0, min1 ); + + uint8x16_t max0 = vmaxq_u8( l0, l1 ); + uint8x16_t max1 = vmaxq_u8( l2, l3 ); + uint8x16_t max2 = vmaxq_u8( max0, max1 ); + + uint8x16_t min3 = vreinterpretq_u8_u32( vrev64q_u32( vreinterpretq_u32_u8( min2 ) ) ); + uint8x16_t max3 = vreinterpretq_u8_u32( vrev64q_u32( vreinterpretq_u32_u8( max2 ) ) ); + + uint8x16_t min4 = vminq_u8( min2, min3 ); + uint8x16_t max4 = vmaxq_u8( max2, max3 ); + + uint8x16_t min5 = vcombine_u8( vget_high_u8( min4 ), vget_low_u8( min4 ) ); + uint8x16_t max5 = vcombine_u8( vget_high_u8( max4 ), vget_low_u8( max4 ) ); + + uint8x16_t rmin = vminq_u8( min4, min5 ); + uint8x16_t rmax = vmaxq_u8( max4, max5 ); + + uint8x16_t range1 = vsubq_u8( rmax, rmin ); + uint8x8_t range2 = vget_low_u8( range1 ); + uint8x8x2_t range3 = vzip_u8( range2, vdup_n_u8( 0 ) ); + uint16x4_t range4 = vreinterpret_u16_u8( range3.val[0] ); + + uint16_t vrange1; + uint16x4_t range5 = vpadd_u16( range4, range4 ); + uint16x4_t range6 = vpadd_u16( range5, range5 ); + vst1_lane_u16( &vrange1, range6, 0 ); + + uint32_t vrange2 = ( 2 << 16 ) / uint32_t( vrange1 + 1 ); + uint16x8_t range = vdupq_n_u16( vrange2 ); + + uint8x16_t inset = vshrq_n_u8( range1, 4 ); + uint8x16_t min = vaddq_u8( rmin, inset ); + uint8x16_t max = vsubq_u8( rmax, inset ); + + uint8x16_t c0 = vsubq_u8( l0, rmin ); + uint8x16_t c1 = vsubq_u8( l1, rmin ); + uint8x16_t c2 = vsubq_u8( l2, rmin ); + uint8x16_t c3 = vsubq_u8( l3, rmin ); + + uint16x8_t is0 = vpaddlq_u8( c0 ); + uint16x8_t is1 = vpaddlq_u8( c1 ); + uint16x8_t is2 = vpaddlq_u8( c2 ); + uint16x8_t is3 = vpaddlq_u8( c3 ); + + uint16x4_t is4 = vpadd_u16( vget_low_u16( is0 ), vget_high_u16( is0 ) ); + uint16x4_t is5 = vpadd_u16( vget_low_u16( is1 ), vget_high_u16( is1 ) ); + uint16x4_t is6 = vpadd_u16( vget_low_u16( is2 ), vget_high_u16( is2 ) ); + uint16x4_t is7 = vpadd_u16( vget_low_u16( is3 ), vget_high_u16( is3 ) ); + + uint16x8_t s0 = vcombine_u16( is4, is5 ); + uint16x8_t s1 = vcombine_u16( is6, is7 ); + + uint16x8_t m0 = vreinterpretq_u16_s16( vqdmulhq_s16( vreinterpretq_s16_u16( s0 ), vreinterpretq_s16_u16( range ) ) ); + uint16x8_t m1 = vreinterpretq_u16_s16( vqdmulhq_s16( vreinterpretq_s16_u16( s1 ), vreinterpretq_s16_u16( range ) ) ); + + uint8x8_t p00 = vmovn_u16( m0 ); + uint8x8_t p01 = vmovn_u16( m1 ); + uint8x16_t p0 = vcombine_u8( p00, p01 ); + + uint32x4_t p1 = vaddq_u32( vshrq_n_u32( vreinterpretq_u32_u8( p0 ), 6 ), vshrq_n_u32( vreinterpretq_u32_u8( p0 ), 12 ) ); + uint32x4_t p2 = vaddq_u32( vshrq_n_u32( vreinterpretq_u32_u8( p0 ), 18 ), vreinterpretq_u32_u8( p0 ) ); + uint32x4_t p3 = vaddq_u32( p1, p2 ); + + uint16x4x2_t p4 = vuzp_u16( vget_low_u16( vreinterpretq_u16_u32( p3 ) ), vget_high_u16( vreinterpretq_u16_u32( p3 ) ) ); + uint8x8x2_t p = vuzp_u8( vreinterpret_u8_u16( p4.val[0] ), vreinterpret_u8_u16( p4.val[0] ) ); + + uint32_t vmin, vmax, vp; + vst1q_lane_u32( &vmin, vreinterpretq_u32_u8( min ), 0 ); + vst1q_lane_u32( &vmax, vreinterpretq_u32_u8( max ), 0 ); + vst1_lane_u32( &vp, vreinterpret_u32_u8( p.val[0] ), 0 ); + + return uint64_t( ( uint64_t( to565( vmin ) ) << 16 ) | to565( vmax ) | ( uint64_t( vp ) << 32 ) ); +# endif +#else + uint32_t ref; + memcpy( &ref, src, 4 ); + uint32_t refMask = ref & 0xF8FCF8; + auto stmp = src + 4; + for( int i=1; i<16; i++ ) + { + uint32_t px; + memcpy( &px, stmp, 4 ); + if( ( px & 0xF8FCF8 ) != refMask ) break; + stmp += 4; + } + if( stmp == src + 64 ) + { + return uint64_t( to565( ref ) ) << 16; + } + + uint8_t min[3] = { src[0], src[1], src[2] }; + uint8_t max[3] = { src[0], src[1], src[2] }; + auto tmp = src + 4; + for( int i=1; i<16; i++ ) + { + for( int j=0; j<3; j++ ) + { + if( tmp[j] < min[j] ) min[j] = tmp[j]; + else if( tmp[j] > max[j] ) max[j] = tmp[j]; + } + tmp += 4; + } + + const uint32_t range = DivTable[max[0] - min[0] + max[1] - min[1] + max[2] - min[2]]; + const uint32_t rmin = min[0] + min[1] + min[2]; + for( int i=0; i<3; i++ ) + { + const uint8_t inset = ( max[i] - min[i] ) >> 4; + min[i] += inset; + max[i] -= inset; + } + + uint32_t data = 0; + for( int i=0; i<16; i++ ) + { + const uint32_t c = src[0] + src[1] + src[2] - rmin; + const uint8_t idx = ( c * range ) >> 16; + data |= idx << (i*2); + src += 4; + } + + return uint64_t( ( uint64_t( to565( min[0], min[1], min[2] ) ) << 16 ) | to565( max[0], max[1], max[2] ) | ( uint64_t( data ) << 32 ) ); +#endif +} + +#ifdef __AVX2__ +static etcpak_force_inline void ProcessRGB_AVX( const uint8_t* src, char*& dst ) +{ + __m256i px0 = _mm256_loadu_si256(((__m256i*)src) + 0); + __m256i px1 = _mm256_loadu_si256(((__m256i*)src) + 1); + __m256i px2 = _mm256_loadu_si256(((__m256i*)src) + 2); + __m256i px3 = _mm256_loadu_si256(((__m256i*)src) + 3); + + __m256i smask = _mm256_set1_epi32( 0xF8FCF8 ); + __m256i sd0 = _mm256_and_si256( px0, smask ); + __m256i sd1 = _mm256_and_si256( px1, smask ); + __m256i sd2 = _mm256_and_si256( px2, smask ); + __m256i sd3 = _mm256_and_si256( px3, smask ); + + __m256i sc = _mm256_shuffle_epi32(sd0, _MM_SHUFFLE(0, 0, 0, 0)); + + __m256i sc0 = _mm256_cmpeq_epi8(sd0, sc); + __m256i sc1 = _mm256_cmpeq_epi8(sd1, sc); + __m256i sc2 = _mm256_cmpeq_epi8(sd2, sc); + __m256i sc3 = _mm256_cmpeq_epi8(sd3, sc); + + __m256i sm0 = _mm256_and_si256(sc0, sc1); + __m256i sm1 = _mm256_and_si256(sc2, sc3); + __m256i sm = _mm256_and_si256(sm0, sm1); + + const int64_t solid0 = 1 - _mm_testc_si128( _mm256_castsi256_si128( sm ), _mm_set1_epi32( -1 ) ); + const int64_t solid1 = 1 - _mm_testc_si128( _mm256_extracti128_si256( sm, 1 ), _mm_set1_epi32( -1 ) ); + + if( solid0 + solid1 == 0 ) + { + const auto c0 = uint64_t( to565( src[0], src[1], src[2] ) ); + const auto c1 = uint64_t( to565( src[16], src[17], src[18] ) ); + memcpy( dst, &c0, 8 ); + memcpy( dst+8, &c1, 8 ); + dst += 16; + return; + } + + __m256i min0 = _mm256_min_epu8( px0, px1 ); + __m256i min1 = _mm256_min_epu8( px2, px3 ); + __m256i min2 = _mm256_min_epu8( min0, min1 ); + + __m256i max0 = _mm256_max_epu8( px0, px1 ); + __m256i max1 = _mm256_max_epu8( px2, px3 ); + __m256i max2 = _mm256_max_epu8( max0, max1 ); + + __m256i min3 = _mm256_shuffle_epi32( min2, _MM_SHUFFLE( 2, 3, 0, 1 ) ); + __m256i max3 = _mm256_shuffle_epi32( max2, _MM_SHUFFLE( 2, 3, 0, 1 ) ); + __m256i min4 = _mm256_min_epu8( min2, min3 ); + __m256i max4 = _mm256_max_epu8( max2, max3 ); + + __m256i min5 = _mm256_shuffle_epi32( min4, _MM_SHUFFLE( 0, 0, 2, 2 ) ); + __m256i max5 = _mm256_shuffle_epi32( max4, _MM_SHUFFLE( 0, 0, 2, 2 ) ); + __m256i rmin = _mm256_min_epu8( min4, min5 ); + __m256i rmax = _mm256_max_epu8( max4, max5 ); + + __m256i range1 = _mm256_subs_epu8( rmax, rmin ); + __m256i range2 = _mm256_sad_epu8( rmax, rmin ); + + uint16_t vrange0 = DivTable[_mm256_cvtsi256_si32( range2 ) >> 1]; + uint16_t vrange1 = DivTable[_mm256_extract_epi16( range2, 8 ) >> 1]; + __m256i range00 = _mm256_set1_epi16( vrange0 ); + __m256i range = _mm256_inserti128_si256( range00, _mm_set1_epi16( vrange1 ), 1 ); + + __m256i inset1 = _mm256_srli_epi16( range1, 4 ); + __m256i inset = _mm256_and_si256( inset1, _mm256_set1_epi8( 0xF ) ); + __m256i min = _mm256_adds_epu8( rmin, inset ); + __m256i max = _mm256_subs_epu8( rmax, inset ); + + __m256i c0 = _mm256_subs_epu8( px0, rmin ); + __m256i c1 = _mm256_subs_epu8( px1, rmin ); + __m256i c2 = _mm256_subs_epu8( px2, rmin ); + __m256i c3 = _mm256_subs_epu8( px3, rmin ); + + __m256i is0 = _mm256_maddubs_epi16( c0, _mm256_set1_epi8( 1 ) ); + __m256i is1 = _mm256_maddubs_epi16( c1, _mm256_set1_epi8( 1 ) ); + __m256i is2 = _mm256_maddubs_epi16( c2, _mm256_set1_epi8( 1 ) ); + __m256i is3 = _mm256_maddubs_epi16( c3, _mm256_set1_epi8( 1 ) ); + + __m256i s0 = _mm256_hadd_epi16( is0, is1 ); + __m256i s1 = _mm256_hadd_epi16( is2, is3 ); + + __m256i m0 = _mm256_mulhi_epu16( s0, range ); + __m256i m1 = _mm256_mulhi_epu16( s1, range ); + + __m256i p0 = _mm256_packus_epi16( m0, m1 ); + + __m256i p1 = _mm256_or_si256( _mm256_srai_epi32( p0, 6 ), _mm256_srai_epi32( p0, 12 ) ); + __m256i p2 = _mm256_or_si256( _mm256_srai_epi32( p0, 18 ), p0 ); + __m256i p3 = _mm256_or_si256( p1, p2 ); + __m256i p =_mm256_shuffle_epi8( p3, _mm256_set1_epi32( 0x0C080400 ) ); + + __m256i mm0 = _mm256_unpacklo_epi8( _mm256_setzero_si256(), min ); + __m256i mm1 = _mm256_unpacklo_epi8( _mm256_setzero_si256(), max ); + __m256i mm2 = _mm256_unpacklo_epi64( mm1, mm0 ); + __m256i mmr = _mm256_slli_epi64( _mm256_srli_epi64( mm2, 11 ), 11 ); + __m256i mmg = _mm256_slli_epi64( _mm256_srli_epi64( mm2, 26 ), 5 ); + __m256i mmb = _mm256_srli_epi64( _mm256_slli_epi64( mm2, 16 ), 59 ); + __m256i mm3 = _mm256_or_si256( mmr, mmg ); + __m256i mm4 = _mm256_or_si256( mm3, mmb ); + __m256i mm5 = _mm256_shuffle_epi8( mm4, _mm256_set1_epi32( 0x09080100 ) ); + + __m256i d0 = _mm256_unpacklo_epi32( mm5, p ); + __m256i d1 = _mm256_permute4x64_epi64( d0, _MM_SHUFFLE( 3, 2, 2, 0 ) ); + __m128i d2 = _mm256_castsi256_si128( d1 ); + + __m128i mask = _mm_set_epi64x( 0xFFFF0000 | -solid1, 0xFFFF0000 | -solid0 ); + __m128i d3 = _mm_and_si128( d2, mask ); + _mm_storeu_si128( (__m128i*)dst, d3 ); + + for( int j=4; j<8; j++ ) dst[j] = (char)DxtcIndexTable[(uint8_t)dst[j]]; + for( int j=12; j<16; j++ ) dst[j] = (char)DxtcIndexTable[(uint8_t)dst[j]]; + + dst += 16; +} +#endif + +static const uint8_t AlphaIndexTable[8] = { 1, 7, 6, 5, 4, 3, 2, 0 }; + +static etcpak_force_inline uint64_t ProcessAlpha( const uint8_t* src ) +{ + uint8_t solid8 = *src; + uint16_t solid16 = uint16_t( solid8 ) | ( uint16_t( solid8 ) << 8 ); + uint32_t solid32 = uint32_t( solid16 ) | ( uint32_t( solid16 ) << 16 ); + uint64_t solid64 = uint64_t( solid32 ) | ( uint64_t( solid32 ) << 32 ); + if( memcmp( src, &solid64, 8 ) == 0 && memcmp( src+8, &solid64, 8 ) == 0 ) + { + return solid8; + } + + uint8_t min = src[0]; + uint8_t max = min; + for( int i=1; i<16; i++ ) + { + const auto v = src[i]; + if( v > max ) max = v; + else if( v < min ) min = v; + } + + uint32_t range = ( 8 << 13 ) / ( 1 + max - min ); + uint64_t data = 0; + for( int i=0; i<16; i++ ) + { + uint8_t a = src[i] - min; + uint64_t idx = AlphaIndexTable[( a * range ) >> 13]; + data |= idx << (i*3); + } + + return max | ( min << 8 ) | ( data << 16 ); +} + +#ifdef __SSE4_1__ +static etcpak_force_inline uint64_t ProcessRGB_SSE( __m128i px0, __m128i px1, __m128i px2, __m128i px3 ) +{ + __m128i smask = _mm_set1_epi32( 0xF8FCF8 ); + __m128i sd0 = _mm_and_si128( px0, smask ); + __m128i sd1 = _mm_and_si128( px1, smask ); + __m128i sd2 = _mm_and_si128( px2, smask ); + __m128i sd3 = _mm_and_si128( px3, smask ); + + __m128i sc = _mm_shuffle_epi32(sd0, _MM_SHUFFLE(0, 0, 0, 0)); + + __m128i sc0 = _mm_cmpeq_epi8(sd0, sc); + __m128i sc1 = _mm_cmpeq_epi8(sd1, sc); + __m128i sc2 = _mm_cmpeq_epi8(sd2, sc); + __m128i sc3 = _mm_cmpeq_epi8(sd3, sc); + + __m128i sm0 = _mm_and_si128(sc0, sc1); + __m128i sm1 = _mm_and_si128(sc2, sc3); + __m128i sm = _mm_and_si128(sm0, sm1); + + if( _mm_testc_si128(sm, _mm_set1_epi32(-1)) ) + { + return uint64_t( to565( _mm_cvtsi128_si32( px0 ) ) ) << 16; + } + + px0 = _mm_and_si128( px0, _mm_set1_epi32( 0xFFFFFF ) ); + px1 = _mm_and_si128( px1, _mm_set1_epi32( 0xFFFFFF ) ); + px2 = _mm_and_si128( px2, _mm_set1_epi32( 0xFFFFFF ) ); + px3 = _mm_and_si128( px3, _mm_set1_epi32( 0xFFFFFF ) ); + + __m128i min0 = _mm_min_epu8( px0, px1 ); + __m128i min1 = _mm_min_epu8( px2, px3 ); + __m128i min2 = _mm_min_epu8( min0, min1 ); + + __m128i max0 = _mm_max_epu8( px0, px1 ); + __m128i max1 = _mm_max_epu8( px2, px3 ); + __m128i max2 = _mm_max_epu8( max0, max1 ); + + __m128i min3 = _mm_shuffle_epi32( min2, _MM_SHUFFLE( 2, 3, 0, 1 ) ); + __m128i max3 = _mm_shuffle_epi32( max2, _MM_SHUFFLE( 2, 3, 0, 1 ) ); + __m128i min4 = _mm_min_epu8( min2, min3 ); + __m128i max4 = _mm_max_epu8( max2, max3 ); + + __m128i min5 = _mm_shuffle_epi32( min4, _MM_SHUFFLE( 0, 0, 2, 2 ) ); + __m128i max5 = _mm_shuffle_epi32( max4, _MM_SHUFFLE( 0, 0, 2, 2 ) ); + __m128i rmin = _mm_min_epu8( min4, min5 ); + __m128i rmax = _mm_max_epu8( max4, max5 ); + + __m128i range1 = _mm_subs_epu8( rmax, rmin ); + __m128i range2 = _mm_sad_epu8( rmax, rmin ); + + uint32_t vrange = _mm_cvtsi128_si32( range2 ) >> 1; + __m128i range = _mm_set1_epi16( DivTable[vrange] ); + + __m128i inset1 = _mm_srli_epi16( range1, 4 ); + __m128i inset = _mm_and_si128( inset1, _mm_set1_epi8( 0xF ) ); + __m128i min = _mm_adds_epu8( rmin, inset ); + __m128i max = _mm_subs_epu8( rmax, inset ); + + __m128i c0 = _mm_subs_epu8( px0, rmin ); + __m128i c1 = _mm_subs_epu8( px1, rmin ); + __m128i c2 = _mm_subs_epu8( px2, rmin ); + __m128i c3 = _mm_subs_epu8( px3, rmin ); + + __m128i is0 = _mm_maddubs_epi16( c0, _mm_set1_epi8( 1 ) ); + __m128i is1 = _mm_maddubs_epi16( c1, _mm_set1_epi8( 1 ) ); + __m128i is2 = _mm_maddubs_epi16( c2, _mm_set1_epi8( 1 ) ); + __m128i is3 = _mm_maddubs_epi16( c3, _mm_set1_epi8( 1 ) ); + + __m128i s0 = _mm_hadd_epi16( is0, is1 ); + __m128i s1 = _mm_hadd_epi16( is2, is3 ); + + __m128i m0 = _mm_mulhi_epu16( s0, range ); + __m128i m1 = _mm_mulhi_epu16( s1, range ); + + __m128i p0 = _mm_packus_epi16( m0, m1 ); + + __m128i p1 = _mm_or_si128( _mm_srai_epi32( p0, 6 ), _mm_srai_epi32( p0, 12 ) ); + __m128i p2 = _mm_or_si128( _mm_srai_epi32( p0, 18 ), p0 ); + __m128i p3 = _mm_or_si128( p1, p2 ); + __m128i p =_mm_shuffle_epi8( p3, _mm_set1_epi32( 0x0C080400 ) ); + + uint32_t vmin = _mm_cvtsi128_si32( min ); + uint32_t vmax = _mm_cvtsi128_si32( max ); + uint32_t vp = _mm_cvtsi128_si32( p ); + + return uint64_t( ( uint64_t( to565( vmin ) ) << 16 ) | to565( vmax ) | ( uint64_t( vp ) << 32 ) ); +} + +static etcpak_force_inline uint64_t ProcessAlpha_SSE( __m128i px0, __m128i px1, __m128i px2, __m128i px3 ) +{ + __m128i mask = _mm_setr_epi32( 0x0f0b0703, -1, -1, -1 ); + + __m128i m0 = _mm_shuffle_epi8( px0, mask ); + __m128i m1 = _mm_shuffle_epi8( px1, _mm_shuffle_epi32( mask, _MM_SHUFFLE( 3, 3, 0, 3 ) ) ); + __m128i m2 = _mm_shuffle_epi8( px2, _mm_shuffle_epi32( mask, _MM_SHUFFLE( 3, 0, 3, 3 ) ) ); + __m128i m3 = _mm_shuffle_epi8( px3, _mm_shuffle_epi32( mask, _MM_SHUFFLE( 0, 3, 3, 3 ) ) ); + __m128i m4 = _mm_or_si128( m0, m1 ); + __m128i m5 = _mm_or_si128( m2, m3 ); + __m128i a = _mm_or_si128( m4, m5 ); + + __m128i solidCmp = _mm_shuffle_epi8( a, _mm_setzero_si128() ); + __m128i cmpRes = _mm_cmpeq_epi8( a, solidCmp ); + if( _mm_testc_si128( cmpRes, _mm_set1_epi32( -1 ) ) ) + { + return _mm_cvtsi128_si32( a ) & 0xFF; + } + + __m128i a1 = _mm_shuffle_epi32( a, _MM_SHUFFLE( 2, 3, 0, 1 ) ); + __m128i max1 = _mm_max_epu8( a, a1 ); + __m128i min1 = _mm_min_epu8( a, a1 ); + __m128i amax2 = _mm_shuffle_epi32( max1, _MM_SHUFFLE( 0, 0, 2, 2 ) ); + __m128i amin2 = _mm_shuffle_epi32( min1, _MM_SHUFFLE( 0, 0, 2, 2 ) ); + __m128i max2 = _mm_max_epu8( max1, amax2 ); + __m128i min2 = _mm_min_epu8( min1, amin2 ); + __m128i amax3 = _mm_alignr_epi8( max2, max2, 2 ); + __m128i amin3 = _mm_alignr_epi8( min2, min2, 2 ); + __m128i max3 = _mm_max_epu8( max2, amax3 ); + __m128i min3 = _mm_min_epu8( min2, amin3 ); + __m128i amax4 = _mm_alignr_epi8( max3, max3, 1 ); + __m128i amin4 = _mm_alignr_epi8( min3, min3, 1 ); + __m128i max = _mm_max_epu8( max3, amax4 ); + __m128i min = _mm_min_epu8( min3, amin4 ); + __m128i minmax = _mm_unpacklo_epi8( max, min ); + + __m128i r = _mm_sub_epi8( max, min ); + int range = _mm_cvtsi128_si32( r ) & 0xFF; + __m128i rv = _mm_set1_epi16( DivTableAlpha[range] ); + + __m128i v = _mm_sub_epi8( a, min ); + + __m128i lo16 = _mm_unpacklo_epi8( v, _mm_setzero_si128() ); + __m128i hi16 = _mm_unpackhi_epi8( v, _mm_setzero_si128() ); + + __m128i lomul = _mm_mulhi_epu16( lo16, rv ); + __m128i himul = _mm_mulhi_epu16( hi16, rv ); + + __m128i p0 = _mm_packus_epi16( lomul, himul ); + __m128i p1 = _mm_or_si128( _mm_and_si128( p0, _mm_set1_epi16( 0x3F ) ), _mm_srai_epi16( _mm_and_si128( p0, _mm_set1_epi16( 0x3F00 ) ), 5 ) ); + __m128i p2 = _mm_packus_epi16( p1, p1 ); + + uint64_t pi = _mm_cvtsi128_si64( p2 ); + uint64_t data = 0; + for( int i=0; i<8; i++ ) + { + uint64_t idx = AlphaIndexTable_SSE[(pi>>(i*8)) & 0x3F]; + data |= idx << (i*6); + } + return (uint64_t)(uint16_t)_mm_cvtsi128_si32( minmax ) | ( data << 16 ); +} +#endif + +void CompressDxt1( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ) +{ +#ifdef __AVX2__ + if( width%8 == 0 ) + { + blocks /= 2; + uint32_t buf[8*4]; + int i = 0; + char* dst8 = (char*)dst; + + do + { + auto tmp = (char*)buf; + memcpy( tmp, src + width * 0, 8*4 ); + memcpy( tmp + 8*4, src + width * 1, 8*4 ); + memcpy( tmp + 16*4, src + width * 2, 8*4 ); + memcpy( tmp + 24*4, src + width * 3, 8*4 ); + src += 8; + if( ++i == width/8 ) + { + src += width * 3; + i = 0; + } + + ProcessRGB_AVX( (uint8_t*)buf, dst8 ); + } + while( --blocks ); + } + else +#endif + { + uint32_t buf[4*4]; + int i = 0; + + auto ptr = dst; + do + { + auto tmp = (char*)buf; + memcpy( tmp, src + width * 0, 4*4 ); + memcpy( tmp + 4*4, src + width * 1, 4*4 ); + memcpy( tmp + 8*4, src + width * 2, 4*4 ); + memcpy( tmp + 12*4, src + width * 3, 4*4 ); + src += 4; + if( ++i == width/4 ) + { + src += width * 3; + i = 0; + } + + const auto c = ProcessRGB( (uint8_t*)buf ); + uint8_t fix[8]; + memcpy( fix, &c, 8 ); + for( int j=4; j<8; j++ ) fix[j] = DxtcIndexTable[fix[j]]; + memcpy( ptr, fix, sizeof( uint64_t ) ); + ptr++; + } + while( --blocks ); + } +} + +void CompressDxt1Dither( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ) +{ + uint32_t buf[4*4]; + int i = 0; + + auto ptr = dst; + do + { + auto tmp = (char*)buf; + memcpy( tmp, src + width * 0, 4*4 ); + memcpy( tmp + 4*4, src + width * 1, 4*4 ); + memcpy( tmp + 8*4, src + width * 2, 4*4 ); + memcpy( tmp + 12*4, src + width * 3, 4*4 ); + src += 4; + if( ++i == width/4 ) + { + src += width * 3; + i = 0; + } + + Dither( (uint8_t*)buf ); + + const auto c = ProcessRGB( (uint8_t*)buf ); + uint8_t fix[8]; + memcpy( fix, &c, 8 ); + for( int j=4; j<8; j++ ) fix[j] = DxtcIndexTable[fix[j]]; + memcpy( ptr, fix, sizeof( uint64_t ) ); + ptr++; + } + while( --blocks ); +} + +void CompressDxt5( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ) +{ + int i = 0; + auto ptr = dst; + do + { +#ifdef __SSE4_1__ + __m128i px0 = _mm_loadu_si128( (__m128i*)( src + width * 0 ) ); + __m128i px1 = _mm_loadu_si128( (__m128i*)( src + width * 1 ) ); + __m128i px2 = _mm_loadu_si128( (__m128i*)( src + width * 2 ) ); + __m128i px3 = _mm_loadu_si128( (__m128i*)( src + width * 3 ) ); + + src += 4; + if( ++i == width/4 ) + { + src += width * 3; + i = 0; + } + + *ptr++ = ProcessAlpha_SSE( px0, px1, px2, px3 ); + + const auto c = ProcessRGB_SSE( px0, px1, px2, px3 ); + uint8_t fix[8]; + memcpy( fix, &c, 8 ); + for( int j=4; j<8; j++ ) fix[j] = DxtcIndexTable[fix[j]]; + memcpy( ptr, fix, sizeof( uint64_t ) ); + ptr++; +#else + uint32_t rgba[4*4]; + uint8_t alpha[4*4]; + + auto tmp = (char*)rgba; + memcpy( tmp, src + width * 0, 4*4 ); + memcpy( tmp + 4*4, src + width * 1, 4*4 ); + memcpy( tmp + 8*4, src + width * 2, 4*4 ); + memcpy( tmp + 12*4, src + width * 3, 4*4 ); + src += 4; + if( ++i == width/4 ) + { + src += width * 3; + i = 0; + } + + for( int i=0; i<16; i++ ) + { + alpha[i] = rgba[i] >> 24; + rgba[i] &= 0xFFFFFF; + } + *ptr++ = ProcessAlpha( alpha ); + + const auto c = ProcessRGB( (uint8_t*)rgba ); + uint8_t fix[8]; + memcpy( fix, &c, 8 ); + for( int j=4; j<8; j++ ) fix[j] = DxtcIndexTable[fix[j]]; + memcpy( ptr, fix, sizeof( uint64_t ) ); + ptr++; +#endif + } + while( --blocks ); +} diff --git a/thirdparty/etcpak/ProcessDxtc.hpp b/thirdparty/etcpak/ProcessDxtc.hpp new file mode 100644 index 0000000000..8e0b12e4bd --- /dev/null +++ b/thirdparty/etcpak/ProcessDxtc.hpp @@ -0,0 +1,11 @@ +#ifndef __PROCESSDXT1_HPP__ +#define __PROCESSDXT1_HPP__ + +#include <stddef.h> +#include <stdint.h> + +void CompressDxt1( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ); +void CompressDxt1Dither( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ); +void CompressDxt5( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ); + +#endif diff --git a/thirdparty/etcpak/ProcessRGB.cpp b/thirdparty/etcpak/ProcessRGB.cpp new file mode 100644 index 0000000000..7f4524d105 --- /dev/null +++ b/thirdparty/etcpak/ProcessRGB.cpp @@ -0,0 +1,3100 @@ +#include <array> +#include <string.h> +#include <limits> + +#ifdef __ARM_NEON +# include <arm_neon.h> +#endif + +#include "Dither.hpp" +#include "ForceInline.hpp" +#include "Math.hpp" +#include "ProcessCommon.hpp" +#include "ProcessRGB.hpp" +#include "Tables.hpp" +#include "Vector.hpp" +#if defined __SSE4_1__ || defined __AVX2__ || defined _MSC_VER +# ifdef _MSC_VER +# include <intrin.h> +# include <Windows.h> +# define _bswap(x) _byteswap_ulong(x) +# define _bswap64(x) _byteswap_uint64(x) +# else +# include <x86intrin.h> +# endif +#endif + +#ifndef _bswap +# define _bswap(x) __builtin_bswap32(x) +# define _bswap64(x) __builtin_bswap64(x) +#endif + +namespace +{ + +#if defined _MSC_VER && !defined __clang__ +static etcpak_force_inline unsigned long _bit_scan_forward( unsigned long mask ) +{ + unsigned long ret; + _BitScanForward( &ret, mask ); + return ret; +} +#endif + +typedef std::array<uint16_t, 4> v4i; + +#ifdef __AVX2__ +static etcpak_force_inline __m256i Sum4_AVX2( const uint8_t* data) noexcept +{ + __m128i d0 = _mm_loadu_si128(((__m128i*)data) + 0); + __m128i d1 = _mm_loadu_si128(((__m128i*)data) + 1); + __m128i d2 = _mm_loadu_si128(((__m128i*)data) + 2); + __m128i d3 = _mm_loadu_si128(((__m128i*)data) + 3); + + __m128i dm0 = _mm_and_si128(d0, _mm_set1_epi32(0x00FFFFFF)); + __m128i dm1 = _mm_and_si128(d1, _mm_set1_epi32(0x00FFFFFF)); + __m128i dm2 = _mm_and_si128(d2, _mm_set1_epi32(0x00FFFFFF)); + __m128i dm3 = _mm_and_si128(d3, _mm_set1_epi32(0x00FFFFFF)); + + __m256i t0 = _mm256_cvtepu8_epi16(dm0); + __m256i t1 = _mm256_cvtepu8_epi16(dm1); + __m256i t2 = _mm256_cvtepu8_epi16(dm2); + __m256i t3 = _mm256_cvtepu8_epi16(dm3); + + __m256i sum0 = _mm256_add_epi16(t0, t1); + __m256i sum1 = _mm256_add_epi16(t2, t3); + + __m256i s0 = _mm256_permute2x128_si256(sum0, sum1, (0) | (3 << 4)); // 0, 0, 3, 3 + __m256i s1 = _mm256_permute2x128_si256(sum0, sum1, (1) | (2 << 4)); // 1, 1, 2, 2 + + __m256i s2 = _mm256_permute4x64_epi64(s0, _MM_SHUFFLE(1, 3, 0, 2)); + __m256i s3 = _mm256_permute4x64_epi64(s0, _MM_SHUFFLE(0, 2, 1, 3)); + __m256i s4 = _mm256_permute4x64_epi64(s1, _MM_SHUFFLE(3, 1, 0, 2)); + __m256i s5 = _mm256_permute4x64_epi64(s1, _MM_SHUFFLE(2, 0, 1, 3)); + + __m256i sum5 = _mm256_add_epi16(s2, s3); // 3, 0, 3, 0 + __m256i sum6 = _mm256_add_epi16(s4, s5); // 2, 1, 1, 2 + return _mm256_add_epi16(sum5, sum6); // 3+2, 0+1, 3+1, 3+2 +} + +static etcpak_force_inline __m256i Average_AVX2( const __m256i data) noexcept +{ + __m256i a = _mm256_add_epi16(data, _mm256_set1_epi16(4)); + + return _mm256_srli_epi16(a, 3); +} + +static etcpak_force_inline __m128i CalcErrorBlock_AVX2( const __m256i data, const v4i a[8]) noexcept +{ + // + __m256i a0 = _mm256_load_si256((__m256i*)a[0].data()); + __m256i a1 = _mm256_load_si256((__m256i*)a[4].data()); + + // err = 8 * ( sq( average[0] ) + sq( average[1] ) + sq( average[2] ) ); + __m256i a4 = _mm256_madd_epi16(a0, a0); + __m256i a5 = _mm256_madd_epi16(a1, a1); + + __m256i a6 = _mm256_hadd_epi32(a4, a5); + __m256i a7 = _mm256_slli_epi32(a6, 3); + + __m256i a8 = _mm256_add_epi32(a7, _mm256_set1_epi32(0x3FFFFFFF)); // Big value to prevent negative values, but small enough to prevent overflow + + // average is not swapped + // err -= block[0] * 2 * average[0]; + // err -= block[1] * 2 * average[1]; + // err -= block[2] * 2 * average[2]; + __m256i a2 = _mm256_slli_epi16(a0, 1); + __m256i a3 = _mm256_slli_epi16(a1, 1); + __m256i b0 = _mm256_madd_epi16(a2, data); + __m256i b1 = _mm256_madd_epi16(a3, data); + + __m256i b2 = _mm256_hadd_epi32(b0, b1); + __m256i b3 = _mm256_sub_epi32(a8, b2); + __m256i b4 = _mm256_hadd_epi32(b3, b3); + + __m256i b5 = _mm256_permutevar8x32_epi32(b4, _mm256_set_epi32(0, 0, 0, 0, 5, 1, 4, 0)); + + return _mm256_castsi256_si128(b5); +} + +static etcpak_force_inline void ProcessAverages_AVX2(const __m256i d, v4i a[8] ) noexcept +{ + __m256i t = _mm256_add_epi16(_mm256_mullo_epi16(d, _mm256_set1_epi16(31)), _mm256_set1_epi16(128)); + + __m256i c = _mm256_srli_epi16(_mm256_add_epi16(t, _mm256_srli_epi16(t, 8)), 8); + + __m256i c1 = _mm256_shuffle_epi32(c, _MM_SHUFFLE(3, 2, 3, 2)); + __m256i diff = _mm256_sub_epi16(c, c1); + diff = _mm256_max_epi16(diff, _mm256_set1_epi16(-4)); + diff = _mm256_min_epi16(diff, _mm256_set1_epi16(3)); + + __m256i co = _mm256_add_epi16(c1, diff); + + c = _mm256_blend_epi16(co, c, 0xF0); + + __m256i a0 = _mm256_or_si256(_mm256_slli_epi16(c, 3), _mm256_srli_epi16(c, 2)); + + _mm256_store_si256((__m256i*)a[4].data(), a0); + + __m256i t0 = _mm256_add_epi16(_mm256_mullo_epi16(d, _mm256_set1_epi16(15)), _mm256_set1_epi16(128)); + __m256i t1 = _mm256_srli_epi16(_mm256_add_epi16(t0, _mm256_srli_epi16(t0, 8)), 8); + + __m256i t2 = _mm256_or_si256(t1, _mm256_slli_epi16(t1, 4)); + + _mm256_store_si256((__m256i*)a[0].data(), t2); +} + +static etcpak_force_inline uint64_t EncodeAverages_AVX2( const v4i a[8], size_t idx ) noexcept +{ + uint64_t d = ( idx << 24 ); + size_t base = idx << 1; + + __m128i a0 = _mm_load_si128((const __m128i*)a[base].data()); + + __m128i r0, r1; + + if( ( idx & 0x2 ) == 0 ) + { + r0 = _mm_srli_epi16(a0, 4); + + __m128i a1 = _mm_unpackhi_epi64(r0, r0); + r1 = _mm_slli_epi16(a1, 4); + } + else + { + __m128i a1 = _mm_and_si128(a0, _mm_set1_epi16(-8)); + + r0 = _mm_unpackhi_epi64(a1, a1); + __m128i a2 = _mm_sub_epi16(a1, r0); + __m128i a3 = _mm_srai_epi16(a2, 3); + r1 = _mm_and_si128(a3, _mm_set1_epi16(0x07)); + } + + __m128i r2 = _mm_or_si128(r0, r1); + // do missing swap for average values + __m128i r3 = _mm_shufflelo_epi16(r2, _MM_SHUFFLE(3, 0, 1, 2)); + __m128i r4 = _mm_packus_epi16(r3, _mm_setzero_si128()); + d |= _mm_cvtsi128_si32(r4); + + return d; +} + +static etcpak_force_inline uint64_t CheckSolid_AVX2( const uint8_t* src ) noexcept +{ + __m256i d0 = _mm256_loadu_si256(((__m256i*)src) + 0); + __m256i d1 = _mm256_loadu_si256(((__m256i*)src) + 1); + + __m256i c = _mm256_broadcastd_epi32(_mm256_castsi256_si128(d0)); + + __m256i c0 = _mm256_cmpeq_epi8(d0, c); + __m256i c1 = _mm256_cmpeq_epi8(d1, c); + + __m256i m = _mm256_and_si256(c0, c1); + + if (!_mm256_testc_si256(m, _mm256_set1_epi32(-1))) + { + return 0; + } + + return 0x02000000 | + ( (unsigned int)( src[0] & 0xF8 ) << 16 ) | + ( (unsigned int)( src[1] & 0xF8 ) << 8 ) | + ( (unsigned int)( src[2] & 0xF8 ) ); +} + +static etcpak_force_inline __m128i PrepareAverages_AVX2( v4i a[8], const uint8_t* src) noexcept +{ + __m256i sum4 = Sum4_AVX2( src ); + + ProcessAverages_AVX2(Average_AVX2( sum4 ), a ); + + return CalcErrorBlock_AVX2( sum4, a); +} + +static etcpak_force_inline __m128i PrepareAverages_AVX2( v4i a[8], const __m256i sum4) noexcept +{ + ProcessAverages_AVX2(Average_AVX2( sum4 ), a ); + + return CalcErrorBlock_AVX2( sum4, a); +} + +static etcpak_force_inline void FindBestFit_4x2_AVX2( uint32_t terr[2][8], uint32_t tsel[8], v4i a[8], const uint32_t offset, const uint8_t* data) noexcept +{ + __m256i sel0 = _mm256_setzero_si256(); + __m256i sel1 = _mm256_setzero_si256(); + + for (unsigned int j = 0; j < 2; ++j) + { + unsigned int bid = offset + 1 - j; + + __m256i squareErrorSum = _mm256_setzero_si256(); + + __m128i a0 = _mm_loadl_epi64((const __m128i*)a[bid].data()); + __m256i a1 = _mm256_broadcastq_epi64(a0); + + // Processing one full row each iteration + for (size_t i = 0; i < 8; i += 4) + { + __m128i rgb = _mm_loadu_si128((const __m128i*)(data + i * 4)); + + __m256i rgb16 = _mm256_cvtepu8_epi16(rgb); + __m256i d = _mm256_sub_epi16(a1, rgb16); + + // The scaling values are divided by two and rounded, to allow the differences to be in the range of signed int16 + // This produces slightly different results, but is significant faster + __m256i pixel0 = _mm256_madd_epi16(d, _mm256_set_epi16(0, 38, 76, 14, 0, 38, 76, 14, 0, 38, 76, 14, 0, 38, 76, 14)); + __m256i pixel1 = _mm256_packs_epi32(pixel0, pixel0); + __m256i pixel2 = _mm256_hadd_epi16(pixel1, pixel1); + __m128i pixel3 = _mm256_castsi256_si128(pixel2); + + __m128i pix0 = _mm_broadcastw_epi16(pixel3); + __m128i pix1 = _mm_broadcastw_epi16(_mm_srli_epi32(pixel3, 16)); + __m256i pixel = _mm256_insertf128_si256(_mm256_castsi128_si256(pix0), pix1, 1); + + // Processing first two pixels of the row + { + __m256i pix = _mm256_abs_epi16(pixel); + + // Taking the absolute value is way faster. The values are only used to sort, so the result will be the same. + // Since the selector table is symmetrical, we need to calculate the difference only for half of the entries. + __m256i error0 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[0]))); + __m256i error1 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[1]))); + + __m256i minIndex0 = _mm256_and_si256(_mm256_cmpgt_epi16(error0, error1), _mm256_set1_epi16(1)); + __m256i minError = _mm256_min_epi16(error0, error1); + + // Exploiting symmetry of the selector table and use the sign bit + // This produces slightly different results, but is significant faster + __m256i minIndex1 = _mm256_srli_epi16(pixel, 15); + + // Interleaving values so madd instruction can be used + __m256i minErrorLo = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(1, 1, 0, 0)); + __m256i minErrorHi = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(3, 3, 2, 2)); + + __m256i minError2 = _mm256_unpacklo_epi16(minErrorLo, minErrorHi); + // Squaring the minimum error to produce correct values when adding + __m256i squareError = _mm256_madd_epi16(minError2, minError2); + + squareErrorSum = _mm256_add_epi32(squareErrorSum, squareError); + + // Packing selector bits + __m256i minIndexLo2 = _mm256_sll_epi16(minIndex0, _mm_cvtsi64_si128(i + j * 8)); + __m256i minIndexHi2 = _mm256_sll_epi16(minIndex1, _mm_cvtsi64_si128(i + j * 8)); + + sel0 = _mm256_or_si256(sel0, minIndexLo2); + sel1 = _mm256_or_si256(sel1, minIndexHi2); + } + + pixel3 = _mm256_extracti128_si256(pixel2, 1); + pix0 = _mm_broadcastw_epi16(pixel3); + pix1 = _mm_broadcastw_epi16(_mm_srli_epi32(pixel3, 16)); + pixel = _mm256_insertf128_si256(_mm256_castsi128_si256(pix0), pix1, 1); + + // Processing second two pixels of the row + { + __m256i pix = _mm256_abs_epi16(pixel); + + // Taking the absolute value is way faster. The values are only used to sort, so the result will be the same. + // Since the selector table is symmetrical, we need to calculate the difference only for half of the entries. + __m256i error0 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[0]))); + __m256i error1 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[1]))); + + __m256i minIndex0 = _mm256_and_si256(_mm256_cmpgt_epi16(error0, error1), _mm256_set1_epi16(1)); + __m256i minError = _mm256_min_epi16(error0, error1); + + // Exploiting symmetry of the selector table and use the sign bit + __m256i minIndex1 = _mm256_srli_epi16(pixel, 15); + + // Interleaving values so madd instruction can be used + __m256i minErrorLo = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(1, 1, 0, 0)); + __m256i minErrorHi = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(3, 3, 2, 2)); + + __m256i minError2 = _mm256_unpacklo_epi16(minErrorLo, minErrorHi); + // Squaring the minimum error to produce correct values when adding + __m256i squareError = _mm256_madd_epi16(minError2, minError2); + + squareErrorSum = _mm256_add_epi32(squareErrorSum, squareError); + + // Packing selector bits + __m256i minIndexLo2 = _mm256_sll_epi16(minIndex0, _mm_cvtsi64_si128(i + j * 8)); + __m256i minIndexHi2 = _mm256_sll_epi16(minIndex1, _mm_cvtsi64_si128(i + j * 8)); + __m256i minIndexLo3 = _mm256_slli_epi16(minIndexLo2, 2); + __m256i minIndexHi3 = _mm256_slli_epi16(minIndexHi2, 2); + + sel0 = _mm256_or_si256(sel0, minIndexLo3); + sel1 = _mm256_or_si256(sel1, minIndexHi3); + } + } + + data += 8 * 4; + + _mm256_store_si256((__m256i*)terr[1 - j], squareErrorSum); + } + + // Interleave selector bits + __m256i minIndexLo0 = _mm256_unpacklo_epi16(sel0, sel1); + __m256i minIndexHi0 = _mm256_unpackhi_epi16(sel0, sel1); + + __m256i minIndexLo1 = _mm256_permute2x128_si256(minIndexLo0, minIndexHi0, (0) | (2 << 4)); + __m256i minIndexHi1 = _mm256_permute2x128_si256(minIndexLo0, minIndexHi0, (1) | (3 << 4)); + + __m256i minIndexHi2 = _mm256_slli_epi32(minIndexHi1, 1); + + __m256i sel = _mm256_or_si256(minIndexLo1, minIndexHi2); + + _mm256_store_si256((__m256i*)tsel, sel); +} + +static etcpak_force_inline void FindBestFit_2x4_AVX2( uint32_t terr[2][8], uint32_t tsel[8], v4i a[8], const uint32_t offset, const uint8_t* data) noexcept +{ + __m256i sel0 = _mm256_setzero_si256(); + __m256i sel1 = _mm256_setzero_si256(); + + __m256i squareErrorSum0 = _mm256_setzero_si256(); + __m256i squareErrorSum1 = _mm256_setzero_si256(); + + __m128i a0 = _mm_loadl_epi64((const __m128i*)a[offset + 1].data()); + __m128i a1 = _mm_loadl_epi64((const __m128i*)a[offset + 0].data()); + + __m128i a2 = _mm_broadcastq_epi64(a0); + __m128i a3 = _mm_broadcastq_epi64(a1); + __m256i a4 = _mm256_insertf128_si256(_mm256_castsi128_si256(a2), a3, 1); + + // Processing one full row each iteration + for (size_t i = 0; i < 16; i += 4) + { + __m128i rgb = _mm_loadu_si128((const __m128i*)(data + i * 4)); + + __m256i rgb16 = _mm256_cvtepu8_epi16(rgb); + __m256i d = _mm256_sub_epi16(a4, rgb16); + + // The scaling values are divided by two and rounded, to allow the differences to be in the range of signed int16 + // This produces slightly different results, but is significant faster + __m256i pixel0 = _mm256_madd_epi16(d, _mm256_set_epi16(0, 38, 76, 14, 0, 38, 76, 14, 0, 38, 76, 14, 0, 38, 76, 14)); + __m256i pixel1 = _mm256_packs_epi32(pixel0, pixel0); + __m256i pixel2 = _mm256_hadd_epi16(pixel1, pixel1); + __m128i pixel3 = _mm256_castsi256_si128(pixel2); + + __m128i pix0 = _mm_broadcastw_epi16(pixel3); + __m128i pix1 = _mm_broadcastw_epi16(_mm_srli_epi32(pixel3, 16)); + __m256i pixel = _mm256_insertf128_si256(_mm256_castsi128_si256(pix0), pix1, 1); + + // Processing first two pixels of the row + { + __m256i pix = _mm256_abs_epi16(pixel); + + // Taking the absolute value is way faster. The values are only used to sort, so the result will be the same. + // Since the selector table is symmetrical, we need to calculate the difference only for half of the entries. + __m256i error0 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[0]))); + __m256i error1 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[1]))); + + __m256i minIndex0 = _mm256_and_si256(_mm256_cmpgt_epi16(error0, error1), _mm256_set1_epi16(1)); + __m256i minError = _mm256_min_epi16(error0, error1); + + // Exploiting symmetry of the selector table and use the sign bit + __m256i minIndex1 = _mm256_srli_epi16(pixel, 15); + + // Interleaving values so madd instruction can be used + __m256i minErrorLo = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(1, 1, 0, 0)); + __m256i minErrorHi = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(3, 3, 2, 2)); + + __m256i minError2 = _mm256_unpacklo_epi16(minErrorLo, minErrorHi); + // Squaring the minimum error to produce correct values when adding + __m256i squareError = _mm256_madd_epi16(minError2, minError2); + + squareErrorSum0 = _mm256_add_epi32(squareErrorSum0, squareError); + + // Packing selector bits + __m256i minIndexLo2 = _mm256_sll_epi16(minIndex0, _mm_cvtsi64_si128(i)); + __m256i minIndexHi2 = _mm256_sll_epi16(minIndex1, _mm_cvtsi64_si128(i)); + + sel0 = _mm256_or_si256(sel0, minIndexLo2); + sel1 = _mm256_or_si256(sel1, minIndexHi2); + } + + pixel3 = _mm256_extracti128_si256(pixel2, 1); + pix0 = _mm_broadcastw_epi16(pixel3); + pix1 = _mm_broadcastw_epi16(_mm_srli_epi32(pixel3, 16)); + pixel = _mm256_insertf128_si256(_mm256_castsi128_si256(pix0), pix1, 1); + + // Processing second two pixels of the row + { + __m256i pix = _mm256_abs_epi16(pixel); + + // Taking the absolute value is way faster. The values are only used to sort, so the result will be the same. + // Since the selector table is symmetrical, we need to calculate the difference only for half of the entries. + __m256i error0 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[0]))); + __m256i error1 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[1]))); + + __m256i minIndex0 = _mm256_and_si256(_mm256_cmpgt_epi16(error0, error1), _mm256_set1_epi16(1)); + __m256i minError = _mm256_min_epi16(error0, error1); + + // Exploiting symmetry of the selector table and use the sign bit + __m256i minIndex1 = _mm256_srli_epi16(pixel, 15); + + // Interleaving values so madd instruction can be used + __m256i minErrorLo = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(1, 1, 0, 0)); + __m256i minErrorHi = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(3, 3, 2, 2)); + + __m256i minError2 = _mm256_unpacklo_epi16(minErrorLo, minErrorHi); + // Squaring the minimum error to produce correct values when adding + __m256i squareError = _mm256_madd_epi16(minError2, minError2); + + squareErrorSum1 = _mm256_add_epi32(squareErrorSum1, squareError); + + // Packing selector bits + __m256i minIndexLo2 = _mm256_sll_epi16(minIndex0, _mm_cvtsi64_si128(i)); + __m256i minIndexHi2 = _mm256_sll_epi16(minIndex1, _mm_cvtsi64_si128(i)); + __m256i minIndexLo3 = _mm256_slli_epi16(minIndexLo2, 2); + __m256i minIndexHi3 = _mm256_slli_epi16(minIndexHi2, 2); + + sel0 = _mm256_or_si256(sel0, minIndexLo3); + sel1 = _mm256_or_si256(sel1, minIndexHi3); + } + } + + _mm256_store_si256((__m256i*)terr[1], squareErrorSum0); + _mm256_store_si256((__m256i*)terr[0], squareErrorSum1); + + // Interleave selector bits + __m256i minIndexLo0 = _mm256_unpacklo_epi16(sel0, sel1); + __m256i minIndexHi0 = _mm256_unpackhi_epi16(sel0, sel1); + + __m256i minIndexLo1 = _mm256_permute2x128_si256(minIndexLo0, minIndexHi0, (0) | (2 << 4)); + __m256i minIndexHi1 = _mm256_permute2x128_si256(minIndexLo0, minIndexHi0, (1) | (3 << 4)); + + __m256i minIndexHi2 = _mm256_slli_epi32(minIndexHi1, 1); + + __m256i sel = _mm256_or_si256(minIndexLo1, minIndexHi2); + + _mm256_store_si256((__m256i*)tsel, sel); +} + +static etcpak_force_inline uint64_t EncodeSelectors_AVX2( uint64_t d, const uint32_t terr[2][8], const uint32_t tsel[8], const bool rotate) noexcept +{ + size_t tidx[2]; + + // Get index of minimum error (terr[0] and terr[1]) + __m256i err0 = _mm256_load_si256((const __m256i*)terr[0]); + __m256i err1 = _mm256_load_si256((const __m256i*)terr[1]); + + __m256i errLo = _mm256_permute2x128_si256(err0, err1, (0) | (2 << 4)); + __m256i errHi = _mm256_permute2x128_si256(err0, err1, (1) | (3 << 4)); + + __m256i errMin0 = _mm256_min_epu32(errLo, errHi); + + __m256i errMin1 = _mm256_shuffle_epi32(errMin0, _MM_SHUFFLE(2, 3, 0, 1)); + __m256i errMin2 = _mm256_min_epu32(errMin0, errMin1); + + __m256i errMin3 = _mm256_shuffle_epi32(errMin2, _MM_SHUFFLE(1, 0, 3, 2)); + __m256i errMin4 = _mm256_min_epu32(errMin3, errMin2); + + __m256i errMin5 = _mm256_permute2x128_si256(errMin4, errMin4, (0) | (0 << 4)); + __m256i errMin6 = _mm256_permute2x128_si256(errMin4, errMin4, (1) | (1 << 4)); + + __m256i errMask0 = _mm256_cmpeq_epi32(errMin5, err0); + __m256i errMask1 = _mm256_cmpeq_epi32(errMin6, err1); + + uint32_t mask0 = _mm256_movemask_epi8(errMask0); + uint32_t mask1 = _mm256_movemask_epi8(errMask1); + + tidx[0] = _bit_scan_forward(mask0) >> 2; + tidx[1] = _bit_scan_forward(mask1) >> 2; + + d |= tidx[0] << 26; + d |= tidx[1] << 29; + + unsigned int t0 = tsel[tidx[0]]; + unsigned int t1 = tsel[tidx[1]]; + + if (!rotate) + { + t0 &= 0xFF00FF00; + t1 &= 0x00FF00FF; + } + else + { + t0 &= 0xCCCCCCCC; + t1 &= 0x33333333; + } + + // Flip selectors from sign bit + unsigned int t2 = (t0 | t1) ^ 0xFFFF0000; + + return d | static_cast<uint64_t>(_bswap(t2)) << 32; +} + +static etcpak_force_inline __m128i r6g7b6_AVX2(__m128 cof, __m128 chf, __m128 cvf) noexcept +{ + __m128i co = _mm_cvttps_epi32(cof); + __m128i ch = _mm_cvttps_epi32(chf); + __m128i cv = _mm_cvttps_epi32(cvf); + + __m128i coh = _mm_packus_epi32(co, ch); + __m128i cv0 = _mm_packus_epi32(cv, _mm_setzero_si128()); + + __m256i cohv0 = _mm256_inserti128_si256(_mm256_castsi128_si256(coh), cv0, 1); + __m256i cohv1 = _mm256_min_epu16(cohv0, _mm256_set1_epi16(1023)); + + __m256i cohv2 = _mm256_sub_epi16(cohv1, _mm256_set1_epi16(15)); + __m256i cohv3 = _mm256_srai_epi16(cohv2, 1); + + __m256i cohvrb0 = _mm256_add_epi16(cohv3, _mm256_set1_epi16(11)); + __m256i cohvrb1 = _mm256_add_epi16(cohv3, _mm256_set1_epi16(4)); + __m256i cohvg0 = _mm256_add_epi16(cohv3, _mm256_set1_epi16(9)); + __m256i cohvg1 = _mm256_add_epi16(cohv3, _mm256_set1_epi16(6)); + + __m256i cohvrb2 = _mm256_srai_epi16(cohvrb0, 7); + __m256i cohvrb3 = _mm256_srai_epi16(cohvrb1, 7); + __m256i cohvg2 = _mm256_srai_epi16(cohvg0, 8); + __m256i cohvg3 = _mm256_srai_epi16(cohvg1, 8); + + __m256i cohvrb4 = _mm256_sub_epi16(cohvrb0, cohvrb2); + __m256i cohvrb5 = _mm256_sub_epi16(cohvrb4, cohvrb3); + __m256i cohvg4 = _mm256_sub_epi16(cohvg0, cohvg2); + __m256i cohvg5 = _mm256_sub_epi16(cohvg4, cohvg3); + + __m256i cohvrb6 = _mm256_srai_epi16(cohvrb5, 3); + __m256i cohvg6 = _mm256_srai_epi16(cohvg5, 2); + + __m256i cohv4 = _mm256_blend_epi16(cohvg6, cohvrb6, 0x55); + + __m128i cohv5 = _mm_packus_epi16(_mm256_castsi256_si128(cohv4), _mm256_extracti128_si256(cohv4, 1)); + return _mm_shuffle_epi8(cohv5, _mm_setr_epi8(6, 5, 4, -1, 2, 1, 0, -1, 10, 9, 8, -1, -1, -1, -1, -1)); +} + +struct Plane +{ + uint64_t plane; + uint64_t error; + __m256i sum4; +}; + +static etcpak_force_inline Plane Planar_AVX2(const uint8_t* src) +{ + __m128i d0 = _mm_loadu_si128(((__m128i*)src) + 0); + __m128i d1 = _mm_loadu_si128(((__m128i*)src) + 1); + __m128i d2 = _mm_loadu_si128(((__m128i*)src) + 2); + __m128i d3 = _mm_loadu_si128(((__m128i*)src) + 3); + + __m128i rgb0 = _mm_shuffle_epi8(d0, _mm_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, -1, -1, -1, -1)); + __m128i rgb1 = _mm_shuffle_epi8(d1, _mm_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, -1, -1, -1, -1)); + __m128i rgb2 = _mm_shuffle_epi8(d2, _mm_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, -1, -1, -1, -1)); + __m128i rgb3 = _mm_shuffle_epi8(d3, _mm_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, -1, -1, -1, -1)); + + __m128i rg0 = _mm_unpacklo_epi32(rgb0, rgb1); + __m128i rg1 = _mm_unpacklo_epi32(rgb2, rgb3); + __m128i b0 = _mm_unpackhi_epi32(rgb0, rgb1); + __m128i b1 = _mm_unpackhi_epi32(rgb2, rgb3); + + // swap channels + __m128i b8 = _mm_unpacklo_epi64(rg0, rg1); + __m128i g8 = _mm_unpackhi_epi64(rg0, rg1); + __m128i r8 = _mm_unpacklo_epi64(b0, b1); + + __m128i t0 = _mm_sad_epu8(r8, _mm_setzero_si128()); + __m128i t1 = _mm_sad_epu8(g8, _mm_setzero_si128()); + __m128i t2 = _mm_sad_epu8(b8, _mm_setzero_si128()); + + __m128i r8s = _mm_shuffle_epi8(r8, _mm_set_epi8(0xF, 0xE, 0xB, 0xA, 0x7, 0x6, 0x3, 0x2, 0xD, 0xC, 0x9, 0x8, 0x5, 0x4, 0x1, 0x0)); + __m128i g8s = _mm_shuffle_epi8(g8, _mm_set_epi8(0xF, 0xE, 0xB, 0xA, 0x7, 0x6, 0x3, 0x2, 0xD, 0xC, 0x9, 0x8, 0x5, 0x4, 0x1, 0x0)); + __m128i b8s = _mm_shuffle_epi8(b8, _mm_set_epi8(0xF, 0xE, 0xB, 0xA, 0x7, 0x6, 0x3, 0x2, 0xD, 0xC, 0x9, 0x8, 0x5, 0x4, 0x1, 0x0)); + + __m128i s0 = _mm_sad_epu8(r8s, _mm_setzero_si128()); + __m128i s1 = _mm_sad_epu8(g8s, _mm_setzero_si128()); + __m128i s2 = _mm_sad_epu8(b8s, _mm_setzero_si128()); + + __m256i sr0 = _mm256_insertf128_si256(_mm256_castsi128_si256(t0), s0, 1); + __m256i sg0 = _mm256_insertf128_si256(_mm256_castsi128_si256(t1), s1, 1); + __m256i sb0 = _mm256_insertf128_si256(_mm256_castsi128_si256(t2), s2, 1); + + __m256i sr1 = _mm256_slli_epi64(sr0, 32); + __m256i sg1 = _mm256_slli_epi64(sg0, 16); + + __m256i srb = _mm256_or_si256(sr1, sb0); + __m256i srgb = _mm256_or_si256(srb, sg1); + + __m128i t3 = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(t0), _mm_castsi128_ps(t1), _MM_SHUFFLE(2, 0, 2, 0))); + __m128i t4 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(3, 1, 2, 0)); + __m128i t5 = _mm_hadd_epi32(t3, t4); + __m128i t6 = _mm_shuffle_epi32(t5, _MM_SHUFFLE(1, 1, 1, 1)); + __m128i t7 = _mm_shuffle_epi32(t5, _MM_SHUFFLE(2, 2, 2, 2)); + + __m256i sr = _mm256_broadcastw_epi16(t5); + __m256i sg = _mm256_broadcastw_epi16(t6); + __m256i sb = _mm256_broadcastw_epi16(t7); + + __m256i r08 = _mm256_cvtepu8_epi16(r8); + __m256i g08 = _mm256_cvtepu8_epi16(g8); + __m256i b08 = _mm256_cvtepu8_epi16(b8); + + __m256i r16 = _mm256_slli_epi16(r08, 4); + __m256i g16 = _mm256_slli_epi16(g08, 4); + __m256i b16 = _mm256_slli_epi16(b08, 4); + + __m256i difR0 = _mm256_sub_epi16(r16, sr); + __m256i difG0 = _mm256_sub_epi16(g16, sg); + __m256i difB0 = _mm256_sub_epi16(b16, sb); + + __m256i difRyz = _mm256_madd_epi16(difR0, _mm256_set_epi16(255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255)); + __m256i difGyz = _mm256_madd_epi16(difG0, _mm256_set_epi16(255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255)); + __m256i difByz = _mm256_madd_epi16(difB0, _mm256_set_epi16(255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255)); + + __m256i difRxz = _mm256_madd_epi16(difR0, _mm256_set_epi16(255, 255, 255, 255, 85, 85, 85, 85, -85, -85, -85, -85, -255, -255, -255, -255)); + __m256i difGxz = _mm256_madd_epi16(difG0, _mm256_set_epi16(255, 255, 255, 255, 85, 85, 85, 85, -85, -85, -85, -85, -255, -255, -255, -255)); + __m256i difBxz = _mm256_madd_epi16(difB0, _mm256_set_epi16(255, 255, 255, 255, 85, 85, 85, 85, -85, -85, -85, -85, -255, -255, -255, -255)); + + __m256i difRGyz = _mm256_hadd_epi32(difRyz, difGyz); + __m256i difByzxz = _mm256_hadd_epi32(difByz, difBxz); + + __m256i difRGxz = _mm256_hadd_epi32(difRxz, difGxz); + + __m128i sumRGyz = _mm_add_epi32(_mm256_castsi256_si128(difRGyz), _mm256_extracti128_si256(difRGyz, 1)); + __m128i sumByzxz = _mm_add_epi32(_mm256_castsi256_si128(difByzxz), _mm256_extracti128_si256(difByzxz, 1)); + __m128i sumRGxz = _mm_add_epi32(_mm256_castsi256_si128(difRGxz), _mm256_extracti128_si256(difRGxz, 1)); + + __m128i sumRGByz = _mm_hadd_epi32(sumRGyz, sumByzxz); + __m128i sumRGByzxz = _mm_hadd_epi32(sumRGxz, sumByzxz); + + __m128i sumRGBxz = _mm_shuffle_epi32(sumRGByzxz, _MM_SHUFFLE(2, 3, 1, 0)); + + __m128 sumRGByzf = _mm_cvtepi32_ps(sumRGByz); + __m128 sumRGBxzf = _mm_cvtepi32_ps(sumRGBxz); + + const float value = (255 * 255 * 8.0f + 85 * 85 * 8.0f) * 16.0f; + + __m128 scale = _mm_set1_ps(-4.0f / value); + + __m128 af = _mm_mul_ps(sumRGBxzf, scale); + __m128 bf = _mm_mul_ps(sumRGByzf, scale); + + __m128 df = _mm_mul_ps(_mm_cvtepi32_ps(t5), _mm_set1_ps(4.0f / 16.0f)); + + // calculating the three colors RGBO, RGBH, and RGBV. RGB = df - af * x - bf * y; + __m128 cof0 = _mm_fnmadd_ps(af, _mm_set1_ps(-255.0f), _mm_fnmadd_ps(bf, _mm_set1_ps(-255.0f), df)); + __m128 chf0 = _mm_fnmadd_ps(af, _mm_set1_ps( 425.0f), _mm_fnmadd_ps(bf, _mm_set1_ps(-255.0f), df)); + __m128 cvf0 = _mm_fnmadd_ps(af, _mm_set1_ps(-255.0f), _mm_fnmadd_ps(bf, _mm_set1_ps( 425.0f), df)); + + // convert to r6g7b6 + __m128i cohv = r6g7b6_AVX2(cof0, chf0, cvf0); + + uint64_t rgbho = _mm_extract_epi64(cohv, 0); + uint32_t rgbv0 = _mm_extract_epi32(cohv, 2); + + // Error calculation + auto ro0 = (rgbho >> 48) & 0x3F; + auto go0 = (rgbho >> 40) & 0x7F; + auto bo0 = (rgbho >> 32) & 0x3F; + auto ro1 = (ro0 >> 4) | (ro0 << 2); + auto go1 = (go0 >> 6) | (go0 << 1); + auto bo1 = (bo0 >> 4) | (bo0 << 2); + auto ro2 = (ro1 << 2) + 2; + auto go2 = (go1 << 2) + 2; + auto bo2 = (bo1 << 2) + 2; + + __m256i ro3 = _mm256_set1_epi16(ro2); + __m256i go3 = _mm256_set1_epi16(go2); + __m256i bo3 = _mm256_set1_epi16(bo2); + + auto rh0 = (rgbho >> 16) & 0x3F; + auto gh0 = (rgbho >> 8) & 0x7F; + auto bh0 = (rgbho >> 0) & 0x3F; + auto rh1 = (rh0 >> 4) | (rh0 << 2); + auto gh1 = (gh0 >> 6) | (gh0 << 1); + auto bh1 = (bh0 >> 4) | (bh0 << 2); + + auto rh2 = rh1 - ro1; + auto gh2 = gh1 - go1; + auto bh2 = bh1 - bo1; + + __m256i rh3 = _mm256_set1_epi16(rh2); + __m256i gh3 = _mm256_set1_epi16(gh2); + __m256i bh3 = _mm256_set1_epi16(bh2); + + auto rv0 = (rgbv0 >> 16) & 0x3F; + auto gv0 = (rgbv0 >> 8) & 0x7F; + auto bv0 = (rgbv0 >> 0) & 0x3F; + auto rv1 = (rv0 >> 4) | (rv0 << 2); + auto gv1 = (gv0 >> 6) | (gv0 << 1); + auto bv1 = (bv0 >> 4) | (bv0 << 2); + + auto rv2 = rv1 - ro1; + auto gv2 = gv1 - go1; + auto bv2 = bv1 - bo1; + + __m256i rv3 = _mm256_set1_epi16(rv2); + __m256i gv3 = _mm256_set1_epi16(gv2); + __m256i bv3 = _mm256_set1_epi16(bv2); + + __m256i x = _mm256_set_epi16(3, 3, 3, 3, 2, 2, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0); + + __m256i rh4 = _mm256_mullo_epi16(rh3, x); + __m256i gh4 = _mm256_mullo_epi16(gh3, x); + __m256i bh4 = _mm256_mullo_epi16(bh3, x); + + __m256i y = _mm256_set_epi16(3, 2, 1, 0, 3, 2, 1, 0, 3, 2, 1, 0, 3, 2, 1, 0); + + __m256i rv4 = _mm256_mullo_epi16(rv3, y); + __m256i gv4 = _mm256_mullo_epi16(gv3, y); + __m256i bv4 = _mm256_mullo_epi16(bv3, y); + + __m256i rxy = _mm256_add_epi16(rh4, rv4); + __m256i gxy = _mm256_add_epi16(gh4, gv4); + __m256i bxy = _mm256_add_epi16(bh4, bv4); + + __m256i rp0 = _mm256_add_epi16(rxy, ro3); + __m256i gp0 = _mm256_add_epi16(gxy, go3); + __m256i bp0 = _mm256_add_epi16(bxy, bo3); + + __m256i rp1 = _mm256_srai_epi16(rp0, 2); + __m256i gp1 = _mm256_srai_epi16(gp0, 2); + __m256i bp1 = _mm256_srai_epi16(bp0, 2); + + __m256i rp2 = _mm256_max_epi16(_mm256_min_epi16(rp1, _mm256_set1_epi16(255)), _mm256_setzero_si256()); + __m256i gp2 = _mm256_max_epi16(_mm256_min_epi16(gp1, _mm256_set1_epi16(255)), _mm256_setzero_si256()); + __m256i bp2 = _mm256_max_epi16(_mm256_min_epi16(bp1, _mm256_set1_epi16(255)), _mm256_setzero_si256()); + + __m256i rdif = _mm256_sub_epi16(r08, rp2); + __m256i gdif = _mm256_sub_epi16(g08, gp2); + __m256i bdif = _mm256_sub_epi16(b08, bp2); + + __m256i rerr = _mm256_mullo_epi16(rdif, _mm256_set1_epi16(38)); + __m256i gerr = _mm256_mullo_epi16(gdif, _mm256_set1_epi16(76)); + __m256i berr = _mm256_mullo_epi16(bdif, _mm256_set1_epi16(14)); + + __m256i sum0 = _mm256_add_epi16(rerr, gerr); + __m256i sum1 = _mm256_add_epi16(sum0, berr); + + __m256i sum2 = _mm256_madd_epi16(sum1, sum1); + + __m128i sum3 = _mm_add_epi32(_mm256_castsi256_si128(sum2), _mm256_extracti128_si256(sum2, 1)); + + uint32_t err0 = _mm_extract_epi32(sum3, 0); + uint32_t err1 = _mm_extract_epi32(sum3, 1); + uint32_t err2 = _mm_extract_epi32(sum3, 2); + uint32_t err3 = _mm_extract_epi32(sum3, 3); + + uint64_t error = err0 + err1 + err2 + err3; + /**/ + + uint32_t rgbv = ( rgbv0 & 0x3F ) | ( ( rgbv0 >> 2 ) & 0x1FC0 ) | ( ( rgbv0 >> 3 ) & 0x7E000 ); + uint64_t rgbho0_ = ( rgbho & 0x3F0000003F ) | ( ( rgbho >> 2 ) & 0x1FC000001FC0 ) | ( ( rgbho >> 3 ) & 0x7E0000007E000 ); + uint64_t rgbho0 = ( rgbho0_ & 0x7FFFF ) | ( ( rgbho0_ >> 13 ) & 0x3FFFF80000 ); + + uint32_t hi = rgbv | ((rgbho0 & 0x1FFF) << 19); + rgbho0 >>= 13; + uint32_t lo = ( rgbho0 & 0x1 ) | ( ( rgbho0 & 0x1FE ) << 1 ) | ( ( rgbho0 & 0x600 ) << 2 ) | ( ( rgbho0 & 0x3F800 ) << 5 ) | ( ( rgbho0 & 0x1FC0000 ) << 6 ); + + uint32_t idx = ( ( rgbho >> 33 ) & 0xF ) | ( ( rgbho >> 41 ) & 0x10 ) | ( ( rgbho >> 48 ) & 0x20 ); + lo |= g_flags[idx]; + uint64_t result = static_cast<uint32_t>(_bswap(lo)); + result |= static_cast<uint64_t>(static_cast<uint32_t>(_bswap(hi))) << 32; + + Plane plane; + + plane.plane = result; + plane.error = error; + plane.sum4 = _mm256_permute4x64_epi64(srgb, _MM_SHUFFLE(2, 3, 0, 1)); + + return plane; +} + +static etcpak_force_inline uint64_t EncodeSelectors_AVX2( uint64_t d, const uint32_t terr[2][8], const uint32_t tsel[8], const bool rotate, const uint64_t value, const uint32_t error) noexcept +{ + size_t tidx[2]; + + // Get index of minimum error (terr[0] and terr[1]) + __m256i err0 = _mm256_load_si256((const __m256i*)terr[0]); + __m256i err1 = _mm256_load_si256((const __m256i*)terr[1]); + + __m256i errLo = _mm256_permute2x128_si256(err0, err1, (0) | (2 << 4)); + __m256i errHi = _mm256_permute2x128_si256(err0, err1, (1) | (3 << 4)); + + __m256i errMin0 = _mm256_min_epu32(errLo, errHi); + + __m256i errMin1 = _mm256_shuffle_epi32(errMin0, _MM_SHUFFLE(2, 3, 0, 1)); + __m256i errMin2 = _mm256_min_epu32(errMin0, errMin1); + + __m256i errMin3 = _mm256_shuffle_epi32(errMin2, _MM_SHUFFLE(1, 0, 3, 2)); + __m256i errMin4 = _mm256_min_epu32(errMin3, errMin2); + + __m256i errMin5 = _mm256_permute2x128_si256(errMin4, errMin4, (0) | (0 << 4)); + __m256i errMin6 = _mm256_permute2x128_si256(errMin4, errMin4, (1) | (1 << 4)); + + __m256i errMask0 = _mm256_cmpeq_epi32(errMin5, err0); + __m256i errMask1 = _mm256_cmpeq_epi32(errMin6, err1); + + uint32_t mask0 = _mm256_movemask_epi8(errMask0); + uint32_t mask1 = _mm256_movemask_epi8(errMask1); + + tidx[0] = _bit_scan_forward(mask0) >> 2; + tidx[1] = _bit_scan_forward(mask1) >> 2; + + if ((terr[0][tidx[0]] + terr[1][tidx[1]]) >= error) + { + return value; + } + + d |= tidx[0] << 26; + d |= tidx[1] << 29; + + unsigned int t0 = tsel[tidx[0]]; + unsigned int t1 = tsel[tidx[1]]; + + if (!rotate) + { + t0 &= 0xFF00FF00; + t1 &= 0x00FF00FF; + } + else + { + t0 &= 0xCCCCCCCC; + t1 &= 0x33333333; + } + + // Flip selectors from sign bit + unsigned int t2 = (t0 | t1) ^ 0xFFFF0000; + + return d | static_cast<uint64_t>(_bswap(t2)) << 32; +} + +#endif + +static etcpak_force_inline void Average( const uint8_t* data, v4i* a ) +{ +#ifdef __SSE4_1__ + __m128i d0 = _mm_loadu_si128(((__m128i*)data) + 0); + __m128i d1 = _mm_loadu_si128(((__m128i*)data) + 1); + __m128i d2 = _mm_loadu_si128(((__m128i*)data) + 2); + __m128i d3 = _mm_loadu_si128(((__m128i*)data) + 3); + + __m128i d0l = _mm_unpacklo_epi8(d0, _mm_setzero_si128()); + __m128i d0h = _mm_unpackhi_epi8(d0, _mm_setzero_si128()); + __m128i d1l = _mm_unpacklo_epi8(d1, _mm_setzero_si128()); + __m128i d1h = _mm_unpackhi_epi8(d1, _mm_setzero_si128()); + __m128i d2l = _mm_unpacklo_epi8(d2, _mm_setzero_si128()); + __m128i d2h = _mm_unpackhi_epi8(d2, _mm_setzero_si128()); + __m128i d3l = _mm_unpacklo_epi8(d3, _mm_setzero_si128()); + __m128i d3h = _mm_unpackhi_epi8(d3, _mm_setzero_si128()); + + __m128i sum0 = _mm_add_epi16(d0l, d1l); + __m128i sum1 = _mm_add_epi16(d0h, d1h); + __m128i sum2 = _mm_add_epi16(d2l, d3l); + __m128i sum3 = _mm_add_epi16(d2h, d3h); + + __m128i sum0l = _mm_unpacklo_epi16(sum0, _mm_setzero_si128()); + __m128i sum0h = _mm_unpackhi_epi16(sum0, _mm_setzero_si128()); + __m128i sum1l = _mm_unpacklo_epi16(sum1, _mm_setzero_si128()); + __m128i sum1h = _mm_unpackhi_epi16(sum1, _mm_setzero_si128()); + __m128i sum2l = _mm_unpacklo_epi16(sum2, _mm_setzero_si128()); + __m128i sum2h = _mm_unpackhi_epi16(sum2, _mm_setzero_si128()); + __m128i sum3l = _mm_unpacklo_epi16(sum3, _mm_setzero_si128()); + __m128i sum3h = _mm_unpackhi_epi16(sum3, _mm_setzero_si128()); + + __m128i b0 = _mm_add_epi32(sum0l, sum0h); + __m128i b1 = _mm_add_epi32(sum1l, sum1h); + __m128i b2 = _mm_add_epi32(sum2l, sum2h); + __m128i b3 = _mm_add_epi32(sum3l, sum3h); + + __m128i a0 = _mm_srli_epi32(_mm_add_epi32(_mm_add_epi32(b2, b3), _mm_set1_epi32(4)), 3); + __m128i a1 = _mm_srli_epi32(_mm_add_epi32(_mm_add_epi32(b0, b1), _mm_set1_epi32(4)), 3); + __m128i a2 = _mm_srli_epi32(_mm_add_epi32(_mm_add_epi32(b1, b3), _mm_set1_epi32(4)), 3); + __m128i a3 = _mm_srli_epi32(_mm_add_epi32(_mm_add_epi32(b0, b2), _mm_set1_epi32(4)), 3); + + _mm_storeu_si128((__m128i*)&a[0], _mm_packus_epi32(_mm_shuffle_epi32(a0, _MM_SHUFFLE(3, 0, 1, 2)), _mm_shuffle_epi32(a1, _MM_SHUFFLE(3, 0, 1, 2)))); + _mm_storeu_si128((__m128i*)&a[2], _mm_packus_epi32(_mm_shuffle_epi32(a2, _MM_SHUFFLE(3, 0, 1, 2)), _mm_shuffle_epi32(a3, _MM_SHUFFLE(3, 0, 1, 2)))); +#elif defined __ARM_NEON + uint8x16x2_t t0 = vzipq_u8(vld1q_u8(data + 0), uint8x16_t()); + uint8x16x2_t t1 = vzipq_u8(vld1q_u8(data + 16), uint8x16_t()); + uint8x16x2_t t2 = vzipq_u8(vld1q_u8(data + 32), uint8x16_t()); + uint8x16x2_t t3 = vzipq_u8(vld1q_u8(data + 48), uint8x16_t()); + + uint16x8x2_t d0 = { vreinterpretq_u16_u8(t0.val[0]), vreinterpretq_u16_u8(t0.val[1]) }; + uint16x8x2_t d1 = { vreinterpretq_u16_u8(t1.val[0]), vreinterpretq_u16_u8(t1.val[1]) }; + uint16x8x2_t d2 = { vreinterpretq_u16_u8(t2.val[0]), vreinterpretq_u16_u8(t2.val[1]) }; + uint16x8x2_t d3 = { vreinterpretq_u16_u8(t3.val[0]), vreinterpretq_u16_u8(t3.val[1]) }; + + uint16x8x2_t s0 = vzipq_u16(vreinterpretq_u16_s16( vaddq_s16(vreinterpretq_s16_u16( d0.val[0] ), vreinterpretq_s16_u16( d1.val[0] ) ) ), uint16x8_t()); + uint16x8x2_t s1 = vzipq_u16(vreinterpretq_u16_s16( vaddq_s16(vreinterpretq_s16_u16( d0.val[1] ), vreinterpretq_s16_u16( d1.val[1] ) ) ), uint16x8_t()); + uint16x8x2_t s2 = vzipq_u16(vreinterpretq_u16_s16( vaddq_s16(vreinterpretq_s16_u16( d2.val[0] ), vreinterpretq_s16_u16( d3.val[0] ) ) ), uint16x8_t()); + uint16x8x2_t s3 = vzipq_u16(vreinterpretq_u16_s16( vaddq_s16(vreinterpretq_s16_u16( d2.val[1] ), vreinterpretq_s16_u16( d3.val[1] ) ) ), uint16x8_t()); + + uint32x4x2_t sum0 = { vreinterpretq_u32_u16(s0.val[0]), vreinterpretq_u32_u16(s0.val[1]) }; + uint32x4x2_t sum1 = { vreinterpretq_u32_u16(s1.val[0]), vreinterpretq_u32_u16(s1.val[1]) }; + uint32x4x2_t sum2 = { vreinterpretq_u32_u16(s2.val[0]), vreinterpretq_u32_u16(s2.val[1]) }; + uint32x4x2_t sum3 = { vreinterpretq_u32_u16(s3.val[0]), vreinterpretq_u32_u16(s3.val[1]) }; + + uint32x4_t b0 = vaddq_u32(sum0.val[0], sum0.val[1]); + uint32x4_t b1 = vaddq_u32(sum1.val[0], sum1.val[1]); + uint32x4_t b2 = vaddq_u32(sum2.val[0], sum2.val[1]); + uint32x4_t b3 = vaddq_u32(sum3.val[0], sum3.val[1]); + + uint32x4_t a0 = vshrq_n_u32(vqaddq_u32(vqaddq_u32(b2, b3), vdupq_n_u32(4)), 3); + uint32x4_t a1 = vshrq_n_u32(vqaddq_u32(vqaddq_u32(b0, b1), vdupq_n_u32(4)), 3); + uint32x4_t a2 = vshrq_n_u32(vqaddq_u32(vqaddq_u32(b1, b3), vdupq_n_u32(4)), 3); + uint32x4_t a3 = vshrq_n_u32(vqaddq_u32(vqaddq_u32(b0, b2), vdupq_n_u32(4)), 3); + + uint16x8_t o0 = vcombine_u16(vqmovun_s32(vreinterpretq_s32_u32( a0 )), vqmovun_s32(vreinterpretq_s32_u32( a1 ))); + uint16x8_t o1 = vcombine_u16(vqmovun_s32(vreinterpretq_s32_u32( a2 )), vqmovun_s32(vreinterpretq_s32_u32( a3 ))); + + a[0] = v4i{o0[2], o0[1], o0[0], 0}; + a[1] = v4i{o0[6], o0[5], o0[4], 0}; + a[2] = v4i{o1[2], o1[1], o1[0], 0}; + a[3] = v4i{o1[6], o1[5], o1[4], 0}; +#else + uint32_t r[4]; + uint32_t g[4]; + uint32_t b[4]; + + memset(r, 0, sizeof(r)); + memset(g, 0, sizeof(g)); + memset(b, 0, sizeof(b)); + + for( int j=0; j<4; j++ ) + { + for( int i=0; i<4; i++ ) + { + int index = (j & 2) + (i >> 1); + b[index] += *data++; + g[index] += *data++; + r[index] += *data++; + data++; + } + } + + a[0] = v4i{ uint16_t( (r[2] + r[3] + 4) / 8 ), uint16_t( (g[2] + g[3] + 4) / 8 ), uint16_t( (b[2] + b[3] + 4) / 8 ), 0}; + a[1] = v4i{ uint16_t( (r[0] + r[1] + 4) / 8 ), uint16_t( (g[0] + g[1] + 4) / 8 ), uint16_t( (b[0] + b[1] + 4) / 8 ), 0}; + a[2] = v4i{ uint16_t( (r[1] + r[3] + 4) / 8 ), uint16_t( (g[1] + g[3] + 4) / 8 ), uint16_t( (b[1] + b[3] + 4) / 8 ), 0}; + a[3] = v4i{ uint16_t( (r[0] + r[2] + 4) / 8 ), uint16_t( (g[0] + g[2] + 4) / 8 ), uint16_t( (b[0] + b[2] + 4) / 8 ), 0}; +#endif +} + +static etcpak_force_inline void CalcErrorBlock( const uint8_t* data, unsigned int err[4][4] ) +{ +#ifdef __SSE4_1__ + __m128i d0 = _mm_loadu_si128(((__m128i*)data) + 0); + __m128i d1 = _mm_loadu_si128(((__m128i*)data) + 1); + __m128i d2 = _mm_loadu_si128(((__m128i*)data) + 2); + __m128i d3 = _mm_loadu_si128(((__m128i*)data) + 3); + + __m128i dm0 = _mm_and_si128(d0, _mm_set1_epi32(0x00FFFFFF)); + __m128i dm1 = _mm_and_si128(d1, _mm_set1_epi32(0x00FFFFFF)); + __m128i dm2 = _mm_and_si128(d2, _mm_set1_epi32(0x00FFFFFF)); + __m128i dm3 = _mm_and_si128(d3, _mm_set1_epi32(0x00FFFFFF)); + + __m128i d0l = _mm_unpacklo_epi8(dm0, _mm_setzero_si128()); + __m128i d0h = _mm_unpackhi_epi8(dm0, _mm_setzero_si128()); + __m128i d1l = _mm_unpacklo_epi8(dm1, _mm_setzero_si128()); + __m128i d1h = _mm_unpackhi_epi8(dm1, _mm_setzero_si128()); + __m128i d2l = _mm_unpacklo_epi8(dm2, _mm_setzero_si128()); + __m128i d2h = _mm_unpackhi_epi8(dm2, _mm_setzero_si128()); + __m128i d3l = _mm_unpacklo_epi8(dm3, _mm_setzero_si128()); + __m128i d3h = _mm_unpackhi_epi8(dm3, _mm_setzero_si128()); + + __m128i sum0 = _mm_add_epi16(d0l, d1l); + __m128i sum1 = _mm_add_epi16(d0h, d1h); + __m128i sum2 = _mm_add_epi16(d2l, d3l); + __m128i sum3 = _mm_add_epi16(d2h, d3h); + + __m128i sum0l = _mm_unpacklo_epi16(sum0, _mm_setzero_si128()); + __m128i sum0h = _mm_unpackhi_epi16(sum0, _mm_setzero_si128()); + __m128i sum1l = _mm_unpacklo_epi16(sum1, _mm_setzero_si128()); + __m128i sum1h = _mm_unpackhi_epi16(sum1, _mm_setzero_si128()); + __m128i sum2l = _mm_unpacklo_epi16(sum2, _mm_setzero_si128()); + __m128i sum2h = _mm_unpackhi_epi16(sum2, _mm_setzero_si128()); + __m128i sum3l = _mm_unpacklo_epi16(sum3, _mm_setzero_si128()); + __m128i sum3h = _mm_unpackhi_epi16(sum3, _mm_setzero_si128()); + + __m128i b0 = _mm_add_epi32(sum0l, sum0h); + __m128i b1 = _mm_add_epi32(sum1l, sum1h); + __m128i b2 = _mm_add_epi32(sum2l, sum2h); + __m128i b3 = _mm_add_epi32(sum3l, sum3h); + + __m128i a0 = _mm_add_epi32(b2, b3); + __m128i a1 = _mm_add_epi32(b0, b1); + __m128i a2 = _mm_add_epi32(b1, b3); + __m128i a3 = _mm_add_epi32(b0, b2); + + _mm_storeu_si128((__m128i*)&err[0], a0); + _mm_storeu_si128((__m128i*)&err[1], a1); + _mm_storeu_si128((__m128i*)&err[2], a2); + _mm_storeu_si128((__m128i*)&err[3], a3); +#elif defined __ARM_NEON + uint8x16x2_t t0 = vzipq_u8(vld1q_u8(data + 0), uint8x16_t()); + uint8x16x2_t t1 = vzipq_u8(vld1q_u8(data + 16), uint8x16_t()); + uint8x16x2_t t2 = vzipq_u8(vld1q_u8(data + 32), uint8x16_t()); + uint8x16x2_t t3 = vzipq_u8(vld1q_u8(data + 48), uint8x16_t()); + + uint16x8x2_t d0 = { vreinterpretq_u16_u8(t0.val[0]), vreinterpretq_u16_u8(t0.val[1]) }; + uint16x8x2_t d1 = { vreinterpretq_u16_u8(t1.val[0]), vreinterpretq_u16_u8(t1.val[1]) }; + uint16x8x2_t d2 = { vreinterpretq_u16_u8(t2.val[0]), vreinterpretq_u16_u8(t2.val[1]) }; + uint16x8x2_t d3 = { vreinterpretq_u16_u8(t3.val[0]), vreinterpretq_u16_u8(t3.val[1]) }; + + uint16x8x2_t s0 = vzipq_u16(vreinterpretq_u16_s16( vaddq_s16(vreinterpretq_s16_u16( d0.val[0] ), vreinterpretq_s16_u16( d1.val[0] ))), uint16x8_t()); + uint16x8x2_t s1 = vzipq_u16(vreinterpretq_u16_s16( vaddq_s16(vreinterpretq_s16_u16( d0.val[1] ), vreinterpretq_s16_u16( d1.val[1] ))), uint16x8_t()); + uint16x8x2_t s2 = vzipq_u16(vreinterpretq_u16_s16( vaddq_s16(vreinterpretq_s16_u16( d2.val[0] ), vreinterpretq_s16_u16( d3.val[0] ))), uint16x8_t()); + uint16x8x2_t s3 = vzipq_u16(vreinterpretq_u16_s16( vaddq_s16(vreinterpretq_s16_u16( d2.val[1] ), vreinterpretq_s16_u16( d3.val[1] ))), uint16x8_t()); + + uint32x4x2_t sum0 = { vreinterpretq_u32_u16(s0.val[0]), vreinterpretq_u32_u16(s0.val[1]) }; + uint32x4x2_t sum1 = { vreinterpretq_u32_u16(s1.val[0]), vreinterpretq_u32_u16(s1.val[1]) }; + uint32x4x2_t sum2 = { vreinterpretq_u32_u16(s2.val[0]), vreinterpretq_u32_u16(s2.val[1]) }; + uint32x4x2_t sum3 = { vreinterpretq_u32_u16(s3.val[0]), vreinterpretq_u32_u16(s3.val[1]) }; + + uint32x4_t b0 = vaddq_u32(sum0.val[0], sum0.val[1]); + uint32x4_t b1 = vaddq_u32(sum1.val[0], sum1.val[1]); + uint32x4_t b2 = vaddq_u32(sum2.val[0], sum2.val[1]); + uint32x4_t b3 = vaddq_u32(sum3.val[0], sum3.val[1]); + + uint32x4_t a0 = vreinterpretq_u32_u8( vandq_u8(vreinterpretq_u8_u32( vqaddq_u32(b2, b3) ), vreinterpretq_u8_u32( vdupq_n_u32(0x00FFFFFF)) ) ); + uint32x4_t a1 = vreinterpretq_u32_u8( vandq_u8(vreinterpretq_u8_u32( vqaddq_u32(b0, b1) ), vreinterpretq_u8_u32( vdupq_n_u32(0x00FFFFFF)) ) ); + uint32x4_t a2 = vreinterpretq_u32_u8( vandq_u8(vreinterpretq_u8_u32( vqaddq_u32(b1, b3) ), vreinterpretq_u8_u32( vdupq_n_u32(0x00FFFFFF)) ) ); + uint32x4_t a3 = vreinterpretq_u32_u8( vandq_u8(vreinterpretq_u8_u32( vqaddq_u32(b0, b2) ), vreinterpretq_u8_u32( vdupq_n_u32(0x00FFFFFF)) ) ); + + vst1q_u32(err[0], a0); + vst1q_u32(err[1], a1); + vst1q_u32(err[2], a2); + vst1q_u32(err[3], a3); +#else + unsigned int terr[4][4]; + + memset(terr, 0, 16 * sizeof(unsigned int)); + + for( int j=0; j<4; j++ ) + { + for( int i=0; i<4; i++ ) + { + int index = (j & 2) + (i >> 1); + unsigned int d = *data++; + terr[index][0] += d; + d = *data++; + terr[index][1] += d; + d = *data++; + terr[index][2] += d; + data++; + } + } + + for( int i=0; i<3; i++ ) + { + err[0][i] = terr[2][i] + terr[3][i]; + err[1][i] = terr[0][i] + terr[1][i]; + err[2][i] = terr[1][i] + terr[3][i]; + err[3][i] = terr[0][i] + terr[2][i]; + } + for( int i=0; i<4; i++ ) + { + err[i][3] = 0; + } +#endif +} + +static etcpak_force_inline unsigned int CalcError( const unsigned int block[4], const v4i& average ) +{ + unsigned int err = 0x3FFFFFFF; // Big value to prevent negative values, but small enough to prevent overflow + err -= block[0] * 2 * average[2]; + err -= block[1] * 2 * average[1]; + err -= block[2] * 2 * average[0]; + err += 8 * ( sq( average[0] ) + sq( average[1] ) + sq( average[2] ) ); + return err; +} + +static etcpak_force_inline void ProcessAverages( v4i* a ) +{ +#ifdef __SSE4_1__ + for( int i=0; i<2; i++ ) + { + __m128i d = _mm_loadu_si128((__m128i*)a[i*2].data()); + + __m128i t = _mm_add_epi16(_mm_mullo_epi16(d, _mm_set1_epi16(31)), _mm_set1_epi16(128)); + + __m128i c = _mm_srli_epi16(_mm_add_epi16(t, _mm_srli_epi16(t, 8)), 8); + + __m128i c1 = _mm_shuffle_epi32(c, _MM_SHUFFLE(3, 2, 3, 2)); + __m128i diff = _mm_sub_epi16(c, c1); + diff = _mm_max_epi16(diff, _mm_set1_epi16(-4)); + diff = _mm_min_epi16(diff, _mm_set1_epi16(3)); + + __m128i co = _mm_add_epi16(c1, diff); + + c = _mm_blend_epi16(co, c, 0xF0); + + __m128i a0 = _mm_or_si128(_mm_slli_epi16(c, 3), _mm_srli_epi16(c, 2)); + + _mm_storeu_si128((__m128i*)a[4+i*2].data(), a0); + } + + for( int i=0; i<2; i++ ) + { + __m128i d = _mm_loadu_si128((__m128i*)a[i*2].data()); + + __m128i t0 = _mm_add_epi16(_mm_mullo_epi16(d, _mm_set1_epi16(15)), _mm_set1_epi16(128)); + __m128i t1 = _mm_srli_epi16(_mm_add_epi16(t0, _mm_srli_epi16(t0, 8)), 8); + + __m128i t2 = _mm_or_si128(t1, _mm_slli_epi16(t1, 4)); + + _mm_storeu_si128((__m128i*)a[i*2].data(), t2); + } +#elif defined __ARM_NEON + for( int i=0; i<2; i++ ) + { + int16x8_t d = vld1q_s16((int16_t*)&a[i*2]); + int16x8_t t = vaddq_s16(vmulq_s16(d, vdupq_n_s16(31)), vdupq_n_s16(128)); + int16x8_t c = vshrq_n_s16(vaddq_s16(t, vshrq_n_s16(t, 8)), 8); + + int16x8_t c1 = vcombine_s16(vget_high_s16(c), vget_high_s16(c)); + int16x8_t diff = vsubq_s16(c, c1); + diff = vmaxq_s16(diff, vdupq_n_s16(-4)); + diff = vminq_s16(diff, vdupq_n_s16(3)); + + int16x8_t co = vaddq_s16(c1, diff); + + c = vcombine_s16(vget_low_s16(co), vget_high_s16(c)); + + int16x8_t a0 = vorrq_s16(vshlq_n_s16(c, 3), vshrq_n_s16(c, 2)); + + vst1q_s16((int16_t*)&a[4+i*2], a0); + } + + for( int i=0; i<2; i++ ) + { + int16x8_t d = vld1q_s16((int16_t*)&a[i*2]); + + int16x8_t t0 = vaddq_s16(vmulq_s16(d, vdupq_n_s16(15)), vdupq_n_s16(128)); + int16x8_t t1 = vshrq_n_s16(vaddq_s16(t0, vshrq_n_s16(t0, 8)), 8); + + int16x8_t t2 = vorrq_s16(t1, vshlq_n_s16(t1, 4)); + + vst1q_s16((int16_t*)&a[i*2], t2); + } +#else + for( int i=0; i<2; i++ ) + { + for( int j=0; j<3; j++ ) + { + int32_t c1 = mul8bit( a[i*2+1][j], 31 ); + int32_t c2 = mul8bit( a[i*2][j], 31 ); + + int32_t diff = c2 - c1; + if( diff > 3 ) diff = 3; + else if( diff < -4 ) diff = -4; + + int32_t co = c1 + diff; + + a[5+i*2][j] = ( c1 << 3 ) | ( c1 >> 2 ); + a[4+i*2][j] = ( co << 3 ) | ( co >> 2 ); + } + } + + for( int i=0; i<4; i++ ) + { + a[i][0] = g_avg2[mul8bit( a[i][0], 15 )]; + a[i][1] = g_avg2[mul8bit( a[i][1], 15 )]; + a[i][2] = g_avg2[mul8bit( a[i][2], 15 )]; + } +#endif +} + +static etcpak_force_inline void EncodeAverages( uint64_t& _d, const v4i* a, size_t idx ) +{ + auto d = _d; + d |= ( idx << 24 ); + size_t base = idx << 1; + + if( ( idx & 0x2 ) == 0 ) + { + for( int i=0; i<3; i++ ) + { + d |= uint64_t( a[base+0][i] >> 4 ) << ( i*8 ); + d |= uint64_t( a[base+1][i] >> 4 ) << ( i*8 + 4 ); + } + } + else + { + for( int i=0; i<3; i++ ) + { + d |= uint64_t( a[base+1][i] & 0xF8 ) << ( i*8 ); + int32_t c = ( ( a[base+0][i] & 0xF8 ) - ( a[base+1][i] & 0xF8 ) ) >> 3; + c &= ~0xFFFFFFF8; + d |= ((uint64_t)c) << ( i*8 ); + } + } + _d = d; +} + +static etcpak_force_inline uint64_t CheckSolid( const uint8_t* src ) +{ +#ifdef __SSE4_1__ + __m128i d0 = _mm_loadu_si128(((__m128i*)src) + 0); + __m128i d1 = _mm_loadu_si128(((__m128i*)src) + 1); + __m128i d2 = _mm_loadu_si128(((__m128i*)src) + 2); + __m128i d3 = _mm_loadu_si128(((__m128i*)src) + 3); + + __m128i c = _mm_shuffle_epi32(d0, _MM_SHUFFLE(0, 0, 0, 0)); + + __m128i c0 = _mm_cmpeq_epi8(d0, c); + __m128i c1 = _mm_cmpeq_epi8(d1, c); + __m128i c2 = _mm_cmpeq_epi8(d2, c); + __m128i c3 = _mm_cmpeq_epi8(d3, c); + + __m128i m0 = _mm_and_si128(c0, c1); + __m128i m1 = _mm_and_si128(c2, c3); + __m128i m = _mm_and_si128(m0, m1); + + if (!_mm_testc_si128(m, _mm_set1_epi32(-1))) + { + return 0; + } +#elif defined __ARM_NEON + int32x4_t d0 = vld1q_s32((int32_t*)src + 0); + int32x4_t d1 = vld1q_s32((int32_t*)src + 4); + int32x4_t d2 = vld1q_s32((int32_t*)src + 8); + int32x4_t d3 = vld1q_s32((int32_t*)src + 12); + + int32x4_t c = vdupq_n_s32(d0[0]); + + int32x4_t c0 = vreinterpretq_s32_u32(vceqq_s32(d0, c)); + int32x4_t c1 = vreinterpretq_s32_u32(vceqq_s32(d1, c)); + int32x4_t c2 = vreinterpretq_s32_u32(vceqq_s32(d2, c)); + int32x4_t c3 = vreinterpretq_s32_u32(vceqq_s32(d3, c)); + + int32x4_t m0 = vandq_s32(c0, c1); + int32x4_t m1 = vandq_s32(c2, c3); + int64x2_t m = vreinterpretq_s64_s32(vandq_s32(m0, m1)); + + if (m[0] != -1 || m[1] != -1) + { + return 0; + } +#else + const uint8_t* ptr = src + 4; + for( int i=1; i<16; i++ ) + { + if( memcmp( src, ptr, 4 ) != 0 ) + { + return 0; + } + ptr += 4; + } +#endif + return 0x02000000 | + ( (unsigned int)( src[0] & 0xF8 ) << 16 ) | + ( (unsigned int)( src[1] & 0xF8 ) << 8 ) | + ( (unsigned int)( src[2] & 0xF8 ) ); +} + +static etcpak_force_inline void PrepareAverages( v4i a[8], const uint8_t* src, unsigned int err[4] ) +{ + Average( src, a ); + ProcessAverages( a ); + + unsigned int errblock[4][4]; + CalcErrorBlock( src, errblock ); + + for( int i=0; i<4; i++ ) + { + err[i/2] += CalcError( errblock[i], a[i] ); + err[2+i/2] += CalcError( errblock[i], a[i+4] ); + } +} + +static etcpak_force_inline void FindBestFit( uint64_t terr[2][8], uint16_t tsel[16][8], v4i a[8], const uint32_t* id, const uint8_t* data ) +{ + for( size_t i=0; i<16; i++ ) + { + uint16_t* sel = tsel[i]; + unsigned int bid = id[i]; + uint64_t* ter = terr[bid%2]; + + uint8_t b = *data++; + uint8_t g = *data++; + uint8_t r = *data++; + data++; + + int dr = a[bid][0] - r; + int dg = a[bid][1] - g; + int db = a[bid][2] - b; + +#ifdef __SSE4_1__ + // Reference implementation + + __m128i pix = _mm_set1_epi32(dr * 77 + dg * 151 + db * 28); + // Taking the absolute value is way faster. The values are only used to sort, so the result will be the same. + __m128i error0 = _mm_abs_epi32(_mm_add_epi32(pix, g_table256_SIMD[0])); + __m128i error1 = _mm_abs_epi32(_mm_add_epi32(pix, g_table256_SIMD[1])); + __m128i error2 = _mm_abs_epi32(_mm_sub_epi32(pix, g_table256_SIMD[0])); + __m128i error3 = _mm_abs_epi32(_mm_sub_epi32(pix, g_table256_SIMD[1])); + + __m128i index0 = _mm_and_si128(_mm_cmplt_epi32(error1, error0), _mm_set1_epi32(1)); + __m128i minError0 = _mm_min_epi32(error0, error1); + + __m128i index1 = _mm_sub_epi32(_mm_set1_epi32(2), _mm_cmplt_epi32(error3, error2)); + __m128i minError1 = _mm_min_epi32(error2, error3); + + __m128i minIndex0 = _mm_blendv_epi8(index0, index1, _mm_cmplt_epi32(minError1, minError0)); + __m128i minError = _mm_min_epi32(minError0, minError1); + + // Squaring the minimum error to produce correct values when adding + __m128i minErrorLow = _mm_shuffle_epi32(minError, _MM_SHUFFLE(1, 1, 0, 0)); + __m128i squareErrorLow = _mm_mul_epi32(minErrorLow, minErrorLow); + squareErrorLow = _mm_add_epi64(squareErrorLow, _mm_loadu_si128(((__m128i*)ter) + 0)); + _mm_storeu_si128(((__m128i*)ter) + 0, squareErrorLow); + __m128i minErrorHigh = _mm_shuffle_epi32(minError, _MM_SHUFFLE(3, 3, 2, 2)); + __m128i squareErrorHigh = _mm_mul_epi32(minErrorHigh, minErrorHigh); + squareErrorHigh = _mm_add_epi64(squareErrorHigh, _mm_loadu_si128(((__m128i*)ter) + 1)); + _mm_storeu_si128(((__m128i*)ter) + 1, squareErrorHigh); + + // Taking the absolute value is way faster. The values are only used to sort, so the result will be the same. + error0 = _mm_abs_epi32(_mm_add_epi32(pix, g_table256_SIMD[2])); + error1 = _mm_abs_epi32(_mm_add_epi32(pix, g_table256_SIMD[3])); + error2 = _mm_abs_epi32(_mm_sub_epi32(pix, g_table256_SIMD[2])); + error3 = _mm_abs_epi32(_mm_sub_epi32(pix, g_table256_SIMD[3])); + + index0 = _mm_and_si128(_mm_cmplt_epi32(error1, error0), _mm_set1_epi32(1)); + minError0 = _mm_min_epi32(error0, error1); + + index1 = _mm_sub_epi32(_mm_set1_epi32(2), _mm_cmplt_epi32(error3, error2)); + minError1 = _mm_min_epi32(error2, error3); + + __m128i minIndex1 = _mm_blendv_epi8(index0, index1, _mm_cmplt_epi32(minError1, minError0)); + minError = _mm_min_epi32(minError0, minError1); + + // Squaring the minimum error to produce correct values when adding + minErrorLow = _mm_shuffle_epi32(minError, _MM_SHUFFLE(1, 1, 0, 0)); + squareErrorLow = _mm_mul_epi32(minErrorLow, minErrorLow); + squareErrorLow = _mm_add_epi64(squareErrorLow, _mm_loadu_si128(((__m128i*)ter) + 2)); + _mm_storeu_si128(((__m128i*)ter) + 2, squareErrorLow); + minErrorHigh = _mm_shuffle_epi32(minError, _MM_SHUFFLE(3, 3, 2, 2)); + squareErrorHigh = _mm_mul_epi32(minErrorHigh, minErrorHigh); + squareErrorHigh = _mm_add_epi64(squareErrorHigh, _mm_loadu_si128(((__m128i*)ter) + 3)); + _mm_storeu_si128(((__m128i*)ter) + 3, squareErrorHigh); + __m128i minIndex = _mm_packs_epi32(minIndex0, minIndex1); + _mm_storeu_si128((__m128i*)sel, minIndex); +#elif defined __ARM_NEON + int32x4_t pix = vdupq_n_s32(dr * 77 + dg * 151 + db * 28); + + // Taking the absolute value is way faster. The values are only used to sort, so the result will be the same. + uint32x4_t error0 = vreinterpretq_u32_s32(vabsq_s32(vaddq_s32(pix, g_table256_NEON[0]))); + uint32x4_t error1 = vreinterpretq_u32_s32(vabsq_s32(vaddq_s32(pix, g_table256_NEON[1]))); + uint32x4_t error2 = vreinterpretq_u32_s32(vabsq_s32(vsubq_s32(pix, g_table256_NEON[0]))); + uint32x4_t error3 = vreinterpretq_u32_s32(vabsq_s32(vsubq_s32(pix, g_table256_NEON[1]))); + + uint32x4_t index0 = vandq_u32(vcltq_u32(error1, error0), vdupq_n_u32(1)); + uint32x4_t minError0 = vminq_u32(error0, error1); + + uint32x4_t index1 = vreinterpretq_u32_s32(vsubq_s32(vdupq_n_s32(2), vreinterpretq_s32_u32(vcltq_u32(error3, error2)))); + uint32x4_t minError1 = vminq_u32(error2, error3); + + uint32x4_t blendMask = vcltq_u32(minError1, minError0); + uint32x4_t minIndex0 = vorrq_u32(vbicq_u32(index0, blendMask), vandq_u32(index1, blendMask)); + uint32x4_t minError = vminq_u32(minError0, minError1); + + // Squaring the minimum error to produce correct values when adding + uint32x4_t squareErrorLow = vmulq_u32(minError, minError); + uint32x4_t squareErrorHigh = vshrq_n_u32(vreinterpretq_u32_s32(vqdmulhq_s32(vreinterpretq_s32_u32(minError), vreinterpretq_s32_u32(minError))), 1); + uint32x4x2_t squareErrorZip = vzipq_u32(squareErrorLow, squareErrorHigh); + uint64x2x2_t squareError = { vreinterpretq_u64_u32(squareErrorZip.val[0]), vreinterpretq_u64_u32(squareErrorZip.val[1]) }; + squareError.val[0] = vaddq_u64(squareError.val[0], vld1q_u64(ter + 0)); + squareError.val[1] = vaddq_u64(squareError.val[1], vld1q_u64(ter + 2)); + vst1q_u64(ter + 0, squareError.val[0]); + vst1q_u64(ter + 2, squareError.val[1]); + + // Taking the absolute value is way faster. The values are only used to sort, so the result will be the same. + error0 = vreinterpretq_u32_s32( vabsq_s32(vaddq_s32(pix, g_table256_NEON[2]))); + error1 = vreinterpretq_u32_s32( vabsq_s32(vaddq_s32(pix, g_table256_NEON[3]))); + error2 = vreinterpretq_u32_s32( vabsq_s32(vsubq_s32(pix, g_table256_NEON[2]))); + error3 = vreinterpretq_u32_s32( vabsq_s32(vsubq_s32(pix, g_table256_NEON[3]))); + + index0 = vandq_u32(vcltq_u32(error1, error0), vdupq_n_u32(1)); + minError0 = vminq_u32(error0, error1); + + index1 = vreinterpretq_u32_s32( vsubq_s32(vdupq_n_s32(2), vreinterpretq_s32_u32(vcltq_u32(error3, error2))) ); + minError1 = vminq_u32(error2, error3); + + blendMask = vcltq_u32(minError1, minError0); + uint32x4_t minIndex1 = vorrq_u32(vbicq_u32(index0, blendMask), vandq_u32(index1, blendMask)); + minError = vminq_u32(minError0, minError1); + + // Squaring the minimum error to produce correct values when adding + squareErrorLow = vmulq_u32(minError, minError); + squareErrorHigh = vshrq_n_u32(vreinterpretq_u32_s32( vqdmulhq_s32(vreinterpretq_s32_u32(minError), vreinterpretq_s32_u32(minError)) ), 1 ); + squareErrorZip = vzipq_u32(squareErrorLow, squareErrorHigh); + squareError.val[0] = vaddq_u64(vreinterpretq_u64_u32( squareErrorZip.val[0] ), vld1q_u64(ter + 4)); + squareError.val[1] = vaddq_u64(vreinterpretq_u64_u32( squareErrorZip.val[1] ), vld1q_u64(ter + 6)); + vst1q_u64(ter + 4, squareError.val[0]); + vst1q_u64(ter + 6, squareError.val[1]); + + uint16x8_t minIndex = vcombine_u16(vqmovn_u32(minIndex0), vqmovn_u32(minIndex1)); + vst1q_u16(sel, minIndex); +#else + int pix = dr * 77 + dg * 151 + db * 28; + + for( int t=0; t<8; t++ ) + { + const int64_t* tab = g_table256[t]; + unsigned int idx = 0; + uint64_t err = sq( tab[0] + pix ); + for( int j=1; j<4; j++ ) + { + uint64_t local = sq( tab[j] + pix ); + if( local < err ) + { + err = local; + idx = j; + } + } + *sel++ = idx; + *ter++ += err; + } +#endif + } +} + +#if defined __SSE4_1__ || defined __ARM_NEON +// Non-reference implementation, but faster. Produces same results as the AVX2 version +static etcpak_force_inline void FindBestFit( uint32_t terr[2][8], uint16_t tsel[16][8], v4i a[8], const uint32_t* id, const uint8_t* data ) +{ + for( size_t i=0; i<16; i++ ) + { + uint16_t* sel = tsel[i]; + unsigned int bid = id[i]; + uint32_t* ter = terr[bid%2]; + + uint8_t b = *data++; + uint8_t g = *data++; + uint8_t r = *data++; + data++; + + int dr = a[bid][0] - r; + int dg = a[bid][1] - g; + int db = a[bid][2] - b; + +#ifdef __SSE4_1__ + // The scaling values are divided by two and rounded, to allow the differences to be in the range of signed int16 + // This produces slightly different results, but is significant faster + __m128i pixel = _mm_set1_epi16(dr * 38 + dg * 76 + db * 14); + __m128i pix = _mm_abs_epi16(pixel); + + // Taking the absolute value is way faster. The values are only used to sort, so the result will be the same. + // Since the selector table is symmetrical, we need to calculate the difference only for half of the entries. + __m128i error0 = _mm_abs_epi16(_mm_sub_epi16(pix, g_table128_SIMD[0])); + __m128i error1 = _mm_abs_epi16(_mm_sub_epi16(pix, g_table128_SIMD[1])); + + __m128i index = _mm_and_si128(_mm_cmplt_epi16(error1, error0), _mm_set1_epi16(1)); + __m128i minError = _mm_min_epi16(error0, error1); + + // Exploiting symmetry of the selector table and use the sign bit + // This produces slightly different results, but is needed to produce same results as AVX2 implementation + __m128i indexBit = _mm_andnot_si128(_mm_srli_epi16(pixel, 15), _mm_set1_epi8(-1)); + __m128i minIndex = _mm_or_si128(index, _mm_add_epi16(indexBit, indexBit)); + + // Squaring the minimum error to produce correct values when adding + __m128i squareErrorLo = _mm_mullo_epi16(minError, minError); + __m128i squareErrorHi = _mm_mulhi_epi16(minError, minError); + + __m128i squareErrorLow = _mm_unpacklo_epi16(squareErrorLo, squareErrorHi); + __m128i squareErrorHigh = _mm_unpackhi_epi16(squareErrorLo, squareErrorHi); + + squareErrorLow = _mm_add_epi32(squareErrorLow, _mm_loadu_si128(((__m128i*)ter) + 0)); + _mm_storeu_si128(((__m128i*)ter) + 0, squareErrorLow); + squareErrorHigh = _mm_add_epi32(squareErrorHigh, _mm_loadu_si128(((__m128i*)ter) + 1)); + _mm_storeu_si128(((__m128i*)ter) + 1, squareErrorHigh); + + _mm_storeu_si128((__m128i*)sel, minIndex); +#elif defined __ARM_NEON + int16x8_t pixel = vdupq_n_s16( dr * 38 + dg * 76 + db * 14 ); + int16x8_t pix = vabsq_s16( pixel ); + + int16x8_t error0 = vabsq_s16( vsubq_s16( pix, g_table128_NEON[0] ) ); + int16x8_t error1 = vabsq_s16( vsubq_s16( pix, g_table128_NEON[1] ) ); + + int16x8_t index = vandq_s16( vreinterpretq_s16_u16( vcltq_s16( error1, error0 ) ), vdupq_n_s16( 1 ) ); + int16x8_t minError = vminq_s16( error0, error1 ); + + int16x8_t indexBit = vandq_s16( vmvnq_s16( vshrq_n_s16( pixel, 15 ) ), vdupq_n_s16( -1 ) ); + int16x8_t minIndex = vorrq_s16( index, vaddq_s16( indexBit, indexBit ) ); + + int16x4_t minErrorLow = vget_low_s16( minError ); + int16x4_t minErrorHigh = vget_high_s16( minError ); + + int32x4_t squareErrorLow = vmull_s16( minErrorLow, minErrorLow ); + int32x4_t squareErrorHigh = vmull_s16( minErrorHigh, minErrorHigh ); + + int32x4_t squareErrorSumLow = vaddq_s32( squareErrorLow, vld1q_s32( (int32_t*)ter ) ); + int32x4_t squareErrorSumHigh = vaddq_s32( squareErrorHigh, vld1q_s32( (int32_t*)ter + 4 ) ); + + vst1q_s32( (int32_t*)ter, squareErrorSumLow ); + vst1q_s32( (int32_t*)ter + 4, squareErrorSumHigh ); + + vst1q_s16( (int16_t*)sel, minIndex ); +#endif + } +} +#endif + +static etcpak_force_inline uint8_t convert6(float f) +{ + int i = (std::min(std::max(static_cast<int>(f), 0), 1023) - 15) >> 1; + return (i + 11 - ((i + 11) >> 7) - ((i + 4) >> 7)) >> 3; +} + +static etcpak_force_inline uint8_t convert7(float f) +{ + int i = (std::min(std::max(static_cast<int>(f), 0), 1023) - 15) >> 1; + return (i + 9 - ((i + 9) >> 8) - ((i + 6) >> 8)) >> 2; +} + +static etcpak_force_inline std::pair<uint64_t, uint64_t> Planar(const uint8_t* src) +{ + int32_t r = 0; + int32_t g = 0; + int32_t b = 0; + + for (int i = 0; i < 16; ++i) + { + b += src[i * 4 + 0]; + g += src[i * 4 + 1]; + r += src[i * 4 + 2]; + } + + int32_t difRyz = 0; + int32_t difGyz = 0; + int32_t difByz = 0; + int32_t difRxz = 0; + int32_t difGxz = 0; + int32_t difBxz = 0; + + const int32_t scaling[] = { -255, -85, 85, 255 }; + + for (int i = 0; i < 16; ++i) + { + int32_t difB = (static_cast<int>(src[i * 4 + 0]) << 4) - b; + int32_t difG = (static_cast<int>(src[i * 4 + 1]) << 4) - g; + int32_t difR = (static_cast<int>(src[i * 4 + 2]) << 4) - r; + + difRyz += difR * scaling[i % 4]; + difGyz += difG * scaling[i % 4]; + difByz += difB * scaling[i % 4]; + + difRxz += difR * scaling[i / 4]; + difGxz += difG * scaling[i / 4]; + difBxz += difB * scaling[i / 4]; + } + + const float scale = -4.0f / ((255 * 255 * 8.0f + 85 * 85 * 8.0f) * 16.0f); + + float aR = difRxz * scale; + float aG = difGxz * scale; + float aB = difBxz * scale; + + float bR = difRyz * scale; + float bG = difGyz * scale; + float bB = difByz * scale; + + float dR = r * (4.0f / 16.0f); + float dG = g * (4.0f / 16.0f); + float dB = b * (4.0f / 16.0f); + + // calculating the three colors RGBO, RGBH, and RGBV. RGB = df - af * x - bf * y; + float cofR = std::fma(aR, 255.0f, std::fma(bR, 255.0f, dR)); + float cofG = std::fma(aG, 255.0f, std::fma(bG, 255.0f, dG)); + float cofB = std::fma(aB, 255.0f, std::fma(bB, 255.0f, dB)); + float chfR = std::fma(aR, -425.0f, std::fma(bR, 255.0f, dR)); + float chfG = std::fma(aG, -425.0f, std::fma(bG, 255.0f, dG)); + float chfB = std::fma(aB, -425.0f, std::fma(bB, 255.0f, dB)); + float cvfR = std::fma(aR, 255.0f, std::fma(bR, -425.0f, dR)); + float cvfG = std::fma(aG, 255.0f, std::fma(bG, -425.0f, dG)); + float cvfB = std::fma(aB, 255.0f, std::fma(bB, -425.0f, dB)); + + // convert to r6g7b6 + int32_t coR = convert6(cofR); + int32_t coG = convert7(cofG); + int32_t coB = convert6(cofB); + int32_t chR = convert6(chfR); + int32_t chG = convert7(chfG); + int32_t chB = convert6(chfB); + int32_t cvR = convert6(cvfR); + int32_t cvG = convert7(cvfG); + int32_t cvB = convert6(cvfB); + + // Error calculation + auto ro0 = coR; + auto go0 = coG; + auto bo0 = coB; + auto ro1 = (ro0 >> 4) | (ro0 << 2); + auto go1 = (go0 >> 6) | (go0 << 1); + auto bo1 = (bo0 >> 4) | (bo0 << 2); + auto ro2 = (ro1 << 2) + 2; + auto go2 = (go1 << 2) + 2; + auto bo2 = (bo1 << 2) + 2; + + auto rh0 = chR; + auto gh0 = chG; + auto bh0 = chB; + auto rh1 = (rh0 >> 4) | (rh0 << 2); + auto gh1 = (gh0 >> 6) | (gh0 << 1); + auto bh1 = (bh0 >> 4) | (bh0 << 2); + + auto rh2 = rh1 - ro1; + auto gh2 = gh1 - go1; + auto bh2 = bh1 - bo1; + + auto rv0 = cvR; + auto gv0 = cvG; + auto bv0 = cvB; + auto rv1 = (rv0 >> 4) | (rv0 << 2); + auto gv1 = (gv0 >> 6) | (gv0 << 1); + auto bv1 = (bv0 >> 4) | (bv0 << 2); + + auto rv2 = rv1 - ro1; + auto gv2 = gv1 - go1; + auto bv2 = bv1 - bo1; + + uint64_t error = 0; + + for (int i = 0; i < 16; ++i) + { + int32_t cR = clampu8((rh2 * (i / 4) + rv2 * (i % 4) + ro2) >> 2); + int32_t cG = clampu8((gh2 * (i / 4) + gv2 * (i % 4) + go2) >> 2); + int32_t cB = clampu8((bh2 * (i / 4) + bv2 * (i % 4) + bo2) >> 2); + + int32_t difB = static_cast<int>(src[i * 4 + 0]) - cB; + int32_t difG = static_cast<int>(src[i * 4 + 1]) - cG; + int32_t difR = static_cast<int>(src[i * 4 + 2]) - cR; + + int32_t dif = difR * 38 + difG * 76 + difB * 14; + + error += dif * dif; + } + + /**/ + uint32_t rgbv = cvB | (cvG << 6) | (cvR << 13); + uint32_t rgbh = chB | (chG << 6) | (chR << 13); + uint32_t hi = rgbv | ((rgbh & 0x1FFF) << 19); + uint32_t lo = (chR & 0x1) | 0x2 | ((chR << 1) & 0x7C); + lo |= ((coB & 0x07) << 7) | ((coB & 0x18) << 8) | ((coB & 0x20) << 11); + lo |= ((coG & 0x3F) << 17) | ((coG & 0x40) << 18); + lo |= coR << 25; + + const auto idx = (coR & 0x20) | ((coG & 0x20) >> 1) | ((coB & 0x1E) >> 1); + + lo |= g_flags[idx]; + + uint64_t result = static_cast<uint32_t>(_bswap(lo)); + result |= static_cast<uint64_t>(static_cast<uint32_t>(_bswap(hi))) << 32; + + return std::make_pair(result, error); +} + +#ifdef __ARM_NEON + +static etcpak_force_inline int32x2_t Planar_NEON_DifXZ( int16x8_t dif_lo, int16x8_t dif_hi ) +{ + int32x4_t dif0 = vmull_n_s16( vget_low_s16( dif_lo ), -255 ); + int32x4_t dif1 = vmull_n_s16( vget_high_s16( dif_lo ), -85 ); + int32x4_t dif2 = vmull_n_s16( vget_low_s16( dif_hi ), 85 ); + int32x4_t dif3 = vmull_n_s16( vget_high_s16( dif_hi ), 255 ); + int32x4_t dif4 = vaddq_s32( vaddq_s32( dif0, dif1 ), vaddq_s32( dif2, dif3 ) ); + +#ifndef __aarch64__ + int32x2_t dif5 = vpadd_s32( vget_low_s32( dif4 ), vget_high_s32( dif4 ) ); + return vpadd_s32( dif5, dif5 ); +#else + return vdup_n_s32( vaddvq_s32( dif4 ) ); +#endif +} + +static etcpak_force_inline int32x2_t Planar_NEON_DifYZ( int16x8_t dif_lo, int16x8_t dif_hi ) +{ + int16x4_t scaling = { -255, -85, 85, 255 }; + int32x4_t dif0 = vmull_s16( vget_low_s16( dif_lo ), scaling ); + int32x4_t dif1 = vmull_s16( vget_high_s16( dif_lo ), scaling ); + int32x4_t dif2 = vmull_s16( vget_low_s16( dif_hi ), scaling ); + int32x4_t dif3 = vmull_s16( vget_high_s16( dif_hi ), scaling ); + int32x4_t dif4 = vaddq_s32( vaddq_s32( dif0, dif1 ), vaddq_s32( dif2, dif3 ) ); + +#ifndef __aarch64__ + int32x2_t dif5 = vpadd_s32( vget_low_s32( dif4 ), vget_high_s32( dif4 ) ); + return vpadd_s32( dif5, dif5 ); +#else + return vdup_n_s32( vaddvq_s32( dif4 ) ); +#endif +} + +static etcpak_force_inline int16x8_t Planar_NEON_SumWide( uint8x16_t src ) +{ + uint16x8_t accu8 = vpaddlq_u8( src ); +#ifndef __aarch64__ + uint16x4_t accu4 = vpadd_u16( vget_low_u16( accu8 ), vget_high_u16( accu8 ) ); + uint16x4_t accu2 = vpadd_u16( accu4, accu4 ); + uint16x4_t accu1 = vpadd_u16( accu2, accu2 ); + return vreinterpretq_s16_u16( vcombine_u16( accu1, accu1 ) ); +#else + return vdupq_n_s16( vaddvq_u16( accu8 ) ); +#endif +} + +static etcpak_force_inline int16x8_t convert6_NEON( int32x4_t lo, int32x4_t hi ) +{ + uint16x8_t x = vcombine_u16( vqmovun_s32( lo ), vqmovun_s32( hi ) ); + int16x8_t i = vreinterpretq_s16_u16( vshrq_n_u16( vqshlq_n_u16( x, 6 ), 6) ); // clamp 0-1023 + i = vhsubq_s16( i, vdupq_n_s16( 15 ) ); + + int16x8_t ip11 = vaddq_s16( i, vdupq_n_s16( 11 ) ); + int16x8_t ip4 = vaddq_s16( i, vdupq_n_s16( 4 ) ); + + return vshrq_n_s16( vsubq_s16( vsubq_s16( ip11, vshrq_n_s16( ip11, 7 ) ), vshrq_n_s16( ip4, 7) ), 3 ); +} + +static etcpak_force_inline int16x4_t convert7_NEON( int32x4_t x ) +{ + int16x4_t i = vreinterpret_s16_u16( vshr_n_u16( vqshl_n_u16( vqmovun_s32( x ), 6 ), 6 ) ); // clamp 0-1023 + i = vhsub_s16( i, vdup_n_s16( 15 ) ); + + int16x4_t p9 = vadd_s16( i, vdup_n_s16( 9 ) ); + int16x4_t p6 = vadd_s16( i, vdup_n_s16( 6 ) ); + return vshr_n_s16( vsub_s16( vsub_s16( p9, vshr_n_s16( p9, 8 ) ), vshr_n_s16( p6, 8 ) ), 2 ); +} + +static etcpak_force_inline std::pair<uint64_t, uint64_t> Planar_NEON( const uint8_t* src ) +{ + uint8x16x4_t srcBlock = vld4q_u8( src ); + + int16x8_t bSumWide = Planar_NEON_SumWide( srcBlock.val[0] ); + int16x8_t gSumWide = Planar_NEON_SumWide( srcBlock.val[1] ); + int16x8_t rSumWide = Planar_NEON_SumWide( srcBlock.val[2] ); + + int16x8_t dif_R_lo = vsubq_s16( vreinterpretq_s16_u16( vshll_n_u8( vget_low_u8( srcBlock.val[2] ), 4) ), rSumWide ); + int16x8_t dif_R_hi = vsubq_s16( vreinterpretq_s16_u16( vshll_n_u8( vget_high_u8( srcBlock.val[2] ), 4) ), rSumWide ); + + int16x8_t dif_G_lo = vsubq_s16( vreinterpretq_s16_u16( vshll_n_u8( vget_low_u8( srcBlock.val[1] ), 4 ) ), gSumWide ); + int16x8_t dif_G_hi = vsubq_s16( vreinterpretq_s16_u16( vshll_n_u8( vget_high_u8( srcBlock.val[1] ), 4 ) ), gSumWide ); + + int16x8_t dif_B_lo = vsubq_s16( vreinterpretq_s16_u16( vshll_n_u8( vget_low_u8( srcBlock.val[0] ), 4) ), bSumWide ); + int16x8_t dif_B_hi = vsubq_s16( vreinterpretq_s16_u16( vshll_n_u8( vget_high_u8( srcBlock.val[0] ), 4) ), bSumWide ); + + int32x2x2_t dif_xz_z = vzip_s32( vzip_s32( Planar_NEON_DifXZ( dif_B_lo, dif_B_hi ), Planar_NEON_DifXZ( dif_R_lo, dif_R_hi ) ).val[0], Planar_NEON_DifXZ( dif_G_lo, dif_G_hi ) ); + int32x4_t dif_xz = vcombine_s32( dif_xz_z.val[0], dif_xz_z.val[1] ); + int32x2x2_t dif_yz_z = vzip_s32( vzip_s32( Planar_NEON_DifYZ( dif_B_lo, dif_B_hi ), Planar_NEON_DifYZ( dif_R_lo, dif_R_hi ) ).val[0], Planar_NEON_DifYZ( dif_G_lo, dif_G_hi ) ); + int32x4_t dif_yz = vcombine_s32( dif_yz_z.val[0], dif_yz_z.val[1] ); + + const float fscale = -4.0f / ( (255 * 255 * 8.0f + 85 * 85 * 8.0f ) * 16.0f ); + float32x4_t fa = vmulq_n_f32( vcvtq_f32_s32( dif_xz ), fscale ); + float32x4_t fb = vmulq_n_f32( vcvtq_f32_s32( dif_yz ), fscale ); + int16x4_t bgrgSum = vzip_s16( vzip_s16( vget_low_s16( bSumWide ), vget_low_s16( rSumWide ) ).val[0], vget_low_s16( gSumWide ) ).val[0]; + float32x4_t fd = vmulq_n_f32( vcvtq_f32_s32( vmovl_s16( bgrgSum ) ), 4.0f / 16.0f); + + float32x4_t cof = vmlaq_n_f32( vmlaq_n_f32( fd, fb, 255.0f ), fa, 255.0f ); + float32x4_t chf = vmlaq_n_f32( vmlaq_n_f32( fd, fb, 255.0f ), fa, -425.0f ); + float32x4_t cvf = vmlaq_n_f32( vmlaq_n_f32( fd, fb, -425.0f ), fa, 255.0f ); + + int32x4_t coi = vcvtq_s32_f32( cof ); + int32x4_t chi = vcvtq_s32_f32( chf ); + int32x4_t cvi = vcvtq_s32_f32( cvf ); + + int32x4x2_t tr_hv = vtrnq_s32( chi, cvi ); + int32x4x2_t tr_o = vtrnq_s32( coi, coi ); + + int16x8_t c_hvoo_br_6 = convert6_NEON( tr_hv.val[0], tr_o.val[0] ); + int16x4_t c_hvox_g_7 = convert7_NEON( vcombine_s32( vget_low_s32( tr_hv.val[1] ), vget_low_s32( tr_o.val[1] ) ) ); + int16x8_t c_hvoo_br_8 = vorrq_s16( vshrq_n_s16( c_hvoo_br_6, 4 ), vshlq_n_s16( c_hvoo_br_6, 2 ) ); + int16x4_t c_hvox_g_8 = vorr_s16( vshr_n_s16( c_hvox_g_7, 6 ), vshl_n_s16( c_hvox_g_7, 1 ) ); + + int16x4_t rec_gxbr_o = vext_s16( c_hvox_g_8, vget_high_s16( c_hvoo_br_8 ), 3 ); + + rec_gxbr_o = vadd_s16( vshl_n_s16( rec_gxbr_o, 2 ), vdup_n_s16( 2 ) ); + int16x8_t rec_ro_wide = vdupq_lane_s16( rec_gxbr_o, 3 ); + int16x8_t rec_go_wide = vdupq_lane_s16( rec_gxbr_o, 0 ); + int16x8_t rec_bo_wide = vdupq_lane_s16( rec_gxbr_o, 1 ); + + int16x4_t br_hv2 = vsub_s16( vget_low_s16( c_hvoo_br_8 ), vget_high_s16( c_hvoo_br_8 ) ); + int16x4_t gg_hv2 = vsub_s16( c_hvox_g_8, vdup_lane_s16( c_hvox_g_8, 2 ) ); + + int16x8_t scaleh_lo = { 0, 0, 0, 0, 1, 1, 1, 1 }; + int16x8_t scaleh_hi = { 2, 2, 2, 2, 3, 3, 3, 3 }; + int16x8_t scalev = { 0, 1, 2, 3, 0, 1, 2, 3 }; + + int16x8_t rec_r_1 = vmlaq_lane_s16( rec_ro_wide, scalev, br_hv2, 3 ); + int16x8_t rec_r_lo = vreinterpretq_s16_u16( vmovl_u8( vqshrun_n_s16( vmlaq_lane_s16( rec_r_1, scaleh_lo, br_hv2, 2 ), 2 ) ) ); + int16x8_t rec_r_hi = vreinterpretq_s16_u16( vmovl_u8( vqshrun_n_s16( vmlaq_lane_s16( rec_r_1, scaleh_hi, br_hv2, 2 ), 2 ) ) ); + + int16x8_t rec_b_1 = vmlaq_lane_s16( rec_bo_wide, scalev, br_hv2, 1 ); + int16x8_t rec_b_lo = vreinterpretq_s16_u16( vmovl_u8( vqshrun_n_s16( vmlaq_lane_s16( rec_b_1, scaleh_lo, br_hv2, 0 ), 2 ) ) ); + int16x8_t rec_b_hi = vreinterpretq_s16_u16( vmovl_u8( vqshrun_n_s16( vmlaq_lane_s16( rec_b_1, scaleh_hi, br_hv2, 0 ), 2 ) ) ); + + int16x8_t rec_g_1 = vmlaq_lane_s16( rec_go_wide, scalev, gg_hv2, 1 ); + int16x8_t rec_g_lo = vreinterpretq_s16_u16( vmovl_u8( vqshrun_n_s16( vmlaq_lane_s16( rec_g_1, scaleh_lo, gg_hv2, 0 ), 2 ) ) ); + int16x8_t rec_g_hi = vreinterpretq_s16_u16( vmovl_u8( vqshrun_n_s16( vmlaq_lane_s16( rec_g_1, scaleh_hi, gg_hv2, 0 ), 2 ) ) ); + + int16x8_t dif_r_lo = vsubq_s16( vreinterpretq_s16_u16( vmovl_u8( vget_low_u8( srcBlock.val[2] ) ) ), rec_r_lo ); + int16x8_t dif_r_hi = vsubq_s16( vreinterpretq_s16_u16( vmovl_u8( vget_high_u8( srcBlock.val[2] ) ) ), rec_r_hi ); + + int16x8_t dif_g_lo = vsubq_s16( vreinterpretq_s16_u16( vmovl_u8( vget_low_u8( srcBlock.val[1] ) ) ), rec_g_lo ); + int16x8_t dif_g_hi = vsubq_s16( vreinterpretq_s16_u16( vmovl_u8( vget_high_u8( srcBlock.val[1] ) ) ), rec_g_hi ); + + int16x8_t dif_b_lo = vsubq_s16( vreinterpretq_s16_u16( vmovl_u8( vget_low_u8( srcBlock.val[0] ) ) ), rec_b_lo ); + int16x8_t dif_b_hi = vsubq_s16( vreinterpretq_s16_u16( vmovl_u8( vget_high_u8( srcBlock.val[0] ) ) ), rec_b_hi ); + + int16x8_t dif_lo = vmlaq_n_s16( vmlaq_n_s16( vmulq_n_s16( dif_r_lo, 38 ), dif_g_lo, 76 ), dif_b_lo, 14 ); + int16x8_t dif_hi = vmlaq_n_s16( vmlaq_n_s16( vmulq_n_s16( dif_r_hi, 38 ), dif_g_hi, 76 ), dif_b_hi, 14 ); + + int16x4_t tmpDif = vget_low_s16( dif_lo ); + int32x4_t difsq_0 = vmull_s16( tmpDif, tmpDif ); + tmpDif = vget_high_s16( dif_lo ); + int32x4_t difsq_1 = vmull_s16( tmpDif, tmpDif ); + tmpDif = vget_low_s16( dif_hi ); + int32x4_t difsq_2 = vmull_s16( tmpDif, tmpDif ); + tmpDif = vget_high_s16( dif_hi ); + int32x4_t difsq_3 = vmull_s16( tmpDif, tmpDif ); + + uint32x4_t difsq_5 = vaddq_u32( vreinterpretq_u32_s32( difsq_0 ), vreinterpretq_u32_s32( difsq_1 ) ); + uint32x4_t difsq_6 = vaddq_u32( vreinterpretq_u32_s32( difsq_2 ), vreinterpretq_u32_s32( difsq_3) ); + + uint64x2_t difsq_7 = vaddl_u32( vget_low_u32( difsq_5 ), vget_high_u32( difsq_5 ) ); + uint64x2_t difsq_8 = vaddl_u32( vget_low_u32( difsq_6 ), vget_high_u32( difsq_6 ) ); + + uint64x2_t difsq_9 = vaddq_u64( difsq_7, difsq_8 ); + +#ifdef __aarch64__ + uint64_t error = vaddvq_u64( difsq_9 ); +#else + uint64_t error = vgetq_lane_u64( difsq_9, 0 ) + vgetq_lane_u64( difsq_9, 1 ); +#endif + + int32_t coR = c_hvoo_br_6[6]; + int32_t coG = c_hvox_g_7[2]; + int32_t coB = c_hvoo_br_6[4]; + + int32_t chR = c_hvoo_br_6[2]; + int32_t chG = c_hvox_g_7[0]; + int32_t chB = c_hvoo_br_6[0]; + + int32_t cvR = c_hvoo_br_6[3]; + int32_t cvG = c_hvox_g_7[1]; + int32_t cvB = c_hvoo_br_6[1]; + + uint32_t rgbv = cvB | ( cvG << 6 ) | ( cvR << 13 ); + uint32_t rgbh = chB | ( chG << 6 ) | ( chR << 13 ); + uint32_t hi = rgbv | ( ( rgbh & 0x1FFF ) << 19 ); + uint32_t lo = ( chR & 0x1 ) | 0x2 | ( ( chR << 1 ) & 0x7C ); + lo |= ( ( coB & 0x07 ) << 7 ) | ( ( coB & 0x18 ) << 8 ) | ( ( coB & 0x20 ) << 11 ); + lo |= ( ( coG & 0x3F) << 17) | ( (coG & 0x40 ) << 18 ); + lo |= coR << 25; + + const auto idx = ( coR & 0x20 ) | ( ( coG & 0x20 ) >> 1 ) | ( ( coB & 0x1E ) >> 1 ); + + lo |= g_flags[idx]; + + uint64_t result = static_cast<uint32_t>( _bswap(lo) ); + result |= static_cast<uint64_t>( static_cast<uint32_t>( _bswap( hi ) ) ) << 32; + + return std::make_pair( result, error ); +} + +#endif + +template<class T, class S> +static etcpak_force_inline uint64_t EncodeSelectors( uint64_t d, const T terr[2][8], const S tsel[16][8], const uint32_t* id, const uint64_t value, const uint64_t error) +{ + size_t tidx[2]; + tidx[0] = GetLeastError( terr[0], 8 ); + tidx[1] = GetLeastError( terr[1], 8 ); + + if ((terr[0][tidx[0]] + terr[1][tidx[1]]) >= error) + { + return value; + } + + d |= tidx[0] << 26; + d |= tidx[1] << 29; + for( int i=0; i<16; i++ ) + { + uint64_t t = tsel[i][tidx[id[i]%2]]; + d |= ( t & 0x1 ) << ( i + 32 ); + d |= ( t & 0x2 ) << ( i + 47 ); + } + + return FixByteOrder(d); +} + +} + +static etcpak_force_inline uint64_t ProcessRGB( const uint8_t* src ) +{ +#ifdef __AVX2__ + uint64_t d = CheckSolid_AVX2( src ); + if( d != 0 ) return d; + + alignas(32) v4i a[8]; + + __m128i err0 = PrepareAverages_AVX2( a, src ); + + // Get index of minimum error (err0) + __m128i err1 = _mm_shuffle_epi32(err0, _MM_SHUFFLE(2, 3, 0, 1)); + __m128i errMin0 = _mm_min_epu32(err0, err1); + + __m128i errMin1 = _mm_shuffle_epi32(errMin0, _MM_SHUFFLE(1, 0, 3, 2)); + __m128i errMin2 = _mm_min_epu32(errMin1, errMin0); + + __m128i errMask = _mm_cmpeq_epi32(errMin2, err0); + + uint32_t mask = _mm_movemask_epi8(errMask); + + uint32_t idx = _bit_scan_forward(mask) >> 2; + + d |= EncodeAverages_AVX2( a, idx ); + + alignas(32) uint32_t terr[2][8] = {}; + alignas(32) uint32_t tsel[8]; + + if ((idx == 0) || (idx == 2)) + { + FindBestFit_4x2_AVX2( terr, tsel, a, idx * 2, src ); + } + else + { + FindBestFit_2x4_AVX2( terr, tsel, a, idx * 2, src ); + } + + return EncodeSelectors_AVX2( d, terr, tsel, (idx % 2) == 1 ); +#else + uint64_t d = CheckSolid( src ); + if( d != 0 ) return d; + + v4i a[8]; + unsigned int err[4] = {}; + PrepareAverages( a, src, err ); + size_t idx = GetLeastError( err, 4 ); + EncodeAverages( d, a, idx ); + +#if ( defined __SSE4_1__ || defined __ARM_NEON ) && !defined REFERENCE_IMPLEMENTATION + uint32_t terr[2][8] = {}; +#else + uint64_t terr[2][8] = {}; +#endif + uint16_t tsel[16][8]; + auto id = g_id[idx]; + FindBestFit( terr, tsel, a, id, src ); + + return FixByteOrder( EncodeSelectors( d, terr, tsel, id ) ); +#endif +} + +static etcpak_force_inline uint64_t ProcessRGB_ETC2( const uint8_t* src ) +{ +#ifdef __AVX2__ + uint64_t d = CheckSolid_AVX2( src ); + if( d != 0 ) return d; + + auto plane = Planar_AVX2( src ); + + alignas(32) v4i a[8]; + + __m128i err0 = PrepareAverages_AVX2( a, plane.sum4 ); + + // Get index of minimum error (err0) + __m128i err1 = _mm_shuffle_epi32(err0, _MM_SHUFFLE(2, 3, 0, 1)); + __m128i errMin0 = _mm_min_epu32(err0, err1); + + __m128i errMin1 = _mm_shuffle_epi32(errMin0, _MM_SHUFFLE(1, 0, 3, 2)); + __m128i errMin2 = _mm_min_epu32(errMin1, errMin0); + + __m128i errMask = _mm_cmpeq_epi32(errMin2, err0); + + uint32_t mask = _mm_movemask_epi8(errMask); + + size_t idx = _bit_scan_forward(mask) >> 2; + + d = EncodeAverages_AVX2( a, idx ); + + alignas(32) uint32_t terr[2][8] = {}; + alignas(32) uint32_t tsel[8]; + + if ((idx == 0) || (idx == 2)) + { + FindBestFit_4x2_AVX2( terr, tsel, a, idx * 2, src ); + } + else + { + FindBestFit_2x4_AVX2( terr, tsel, a, idx * 2, src ); + } + + return EncodeSelectors_AVX2( d, terr, tsel, (idx % 2) == 1, plane.plane, plane.error ); +#else + uint64_t d = CheckSolid( src ); + if (d != 0) return d; + +#ifdef __ARM_NEON + auto result = Planar_NEON( src ); +#else + auto result = Planar( src ); +#endif + + v4i a[8]; + unsigned int err[4] = {}; + PrepareAverages( a, src, err ); + size_t idx = GetLeastError( err, 4 ); + EncodeAverages( d, a, idx ); + +#if ( defined __SSE4_1__ || defined __ARM_NEON ) && !defined REFERENCE_IMPLEMENTATION + uint32_t terr[2][8] = {}; +#else + uint64_t terr[2][8] = {}; +#endif + uint16_t tsel[16][8]; + auto id = g_id[idx]; + FindBestFit( terr, tsel, a, id, src ); + + return EncodeSelectors( d, terr, tsel, id, result.first, result.second ); +#endif +} + +#ifdef __SSE4_1__ +template<int K> +static etcpak_force_inline __m128i Widen( const __m128i src ) +{ + static_assert( K >= 0 && K <= 7, "Index out of range" ); + + __m128i tmp; + switch( K ) + { + case 0: + tmp = _mm_shufflelo_epi16( src, _MM_SHUFFLE( 0, 0, 0, 0 ) ); + return _mm_shuffle_epi32( tmp, _MM_SHUFFLE( 0, 0, 0, 0 ) ); + case 1: + tmp = _mm_shufflelo_epi16( src, _MM_SHUFFLE( 1, 1, 1, 1 ) ); + return _mm_shuffle_epi32( tmp, _MM_SHUFFLE( 0, 0, 0, 0 ) ); + case 2: + tmp = _mm_shufflelo_epi16( src, _MM_SHUFFLE( 2, 2, 2, 2 ) ); + return _mm_shuffle_epi32( tmp, _MM_SHUFFLE( 0, 0, 0, 0 ) ); + case 3: + tmp = _mm_shufflelo_epi16( src, _MM_SHUFFLE( 3, 3, 3, 3 ) ); + return _mm_shuffle_epi32( tmp, _MM_SHUFFLE( 0, 0, 0, 0 ) ); + case 4: + tmp = _mm_shufflehi_epi16( src, _MM_SHUFFLE( 0, 0, 0, 0 ) ); + return _mm_shuffle_epi32( tmp, _MM_SHUFFLE( 2, 2, 2, 2 ) ); + case 5: + tmp = _mm_shufflehi_epi16( src, _MM_SHUFFLE( 1, 1, 1, 1 ) ); + return _mm_shuffle_epi32( tmp, _MM_SHUFFLE( 2, 2, 2, 2 ) ); + case 6: + tmp = _mm_shufflehi_epi16( src, _MM_SHUFFLE( 2, 2, 2, 2 ) ); + return _mm_shuffle_epi32( tmp, _MM_SHUFFLE( 2, 2, 2, 2 ) ); + case 7: + tmp = _mm_shufflehi_epi16( src, _MM_SHUFFLE( 3, 3, 3, 3 ) ); + return _mm_shuffle_epi32( tmp, _MM_SHUFFLE( 2, 2, 2, 2 ) ); + } +} + +static etcpak_force_inline int GetMulSel( int sel ) +{ + switch( sel ) + { + case 0: + return 0; + case 1: + case 2: + case 3: + return 1; + case 4: + return 2; + case 5: + case 6: + case 7: + return 3; + case 8: + case 9: + case 10: + case 11: + case 12: + case 13: + return 4; + case 14: + case 15: + return 5; + } +} + +#endif + +#ifdef __ARM_NEON + +static constexpr etcpak_force_inline int GetMulSel(int sel) +{ + return ( sel < 1 ) ? 0 : ( sel < 4 ) ? 1 : ( sel < 5 ) ? 2 : ( sel < 8 ) ? 3 : ( sel < 14 ) ? 4 : 5; +} + +static constexpr int ClampConstant( int x, int min, int max ) +{ + return x < min ? min : x > max ? max : x; +} + +template <int Index> +etcpak_force_inline static uint16x8_t ErrorProbe_EAC_NEON( uint8x8_t recVal, uint8x16_t alphaBlock ) +{ + uint8x8_t srcValWide; +#ifndef __aarch64__ + if( Index < 8 ) + srcValWide = vdup_lane_u8( vget_low_u8( alphaBlock ), ClampConstant( Index, 0, 8 ) ); + else + srcValWide = vdup_lane_u8( vget_high_u8( alphaBlock ), ClampConstant( Index - 8, 0, 8 ) ); +#else + srcValWide = vdup_laneq_u8( alphaBlock, Index ); +#endif + + uint8x8_t deltaVal = vabd_u8( srcValWide, recVal ); + return vmull_u8( deltaVal, deltaVal ); +} + +etcpak_force_inline static uint16_t MinError_EAC_NEON( uint16x8_t errProbe ) +{ +#ifndef __aarch64__ + uint16x4_t tmpErr = vpmin_u16( vget_low_u16( errProbe ), vget_high_u16( errProbe ) ); + tmpErr = vpmin_u16( tmpErr, tmpErr ); + return vpmin_u16( tmpErr, tmpErr )[0]; +#else + return vminvq_u16( errProbe ); +#endif +} + +template <int Index> +etcpak_force_inline static uint64_t MinErrorIndex_EAC_NEON( uint8x8_t recVal, uint8x16_t alphaBlock ) +{ + uint16x8_t errProbe = ErrorProbe_EAC_NEON<Index>( recVal, alphaBlock ); + uint16x8_t minErrMask = vceqq_u16( errProbe, vdupq_n_u16( MinError_EAC_NEON( errProbe ) ) ); + uint64_t idx = __builtin_ctzll( vget_lane_u64( vreinterpret_u64_u8( vqmovn_u16( minErrMask ) ), 0 ) ); + idx >>= 3; + idx <<= 45 - Index * 3; + + return idx; +} + +template <int Index> +etcpak_force_inline static int16x8_t WidenMultiplier_EAC_NEON( int16x8_t multipliers ) +{ + constexpr int Lane = GetMulSel( Index ); +#ifndef __aarch64__ + if( Lane < 4 ) + return vdupq_lane_s16( vget_low_s16( multipliers ), ClampConstant( Lane, 0, 4 ) ); + else + return vdupq_lane_s16( vget_high_s16( multipliers ), ClampConstant( Lane - 4, 0, 4 ) ); +#else + return vdupq_laneq_s16( multipliers, Lane ); +#endif +} + +#endif + +static etcpak_force_inline uint64_t ProcessAlpha_ETC2( const uint8_t* src ) +{ +#if defined __SSE4_1__ + // Check solid + __m128i s = _mm_loadu_si128( (__m128i*)src ); + __m128i solidCmp = _mm_set1_epi8( src[0] ); + __m128i cmpRes = _mm_cmpeq_epi8( s, solidCmp ); + if( _mm_testc_si128( cmpRes, _mm_set1_epi32( -1 ) ) ) + { + return src[0]; + } + + // Calculate min, max + __m128i s1 = _mm_shuffle_epi32( s, _MM_SHUFFLE( 2, 3, 0, 1 ) ); + __m128i max1 = _mm_max_epu8( s, s1 ); + __m128i min1 = _mm_min_epu8( s, s1 ); + __m128i smax2 = _mm_shuffle_epi32( max1, _MM_SHUFFLE( 0, 0, 2, 2 ) ); + __m128i smin2 = _mm_shuffle_epi32( min1, _MM_SHUFFLE( 0, 0, 2, 2 ) ); + __m128i max2 = _mm_max_epu8( max1, smax2 ); + __m128i min2 = _mm_min_epu8( min1, smin2 ); + __m128i smax3 = _mm_alignr_epi8( max2, max2, 2 ); + __m128i smin3 = _mm_alignr_epi8( min2, min2, 2 ); + __m128i max3 = _mm_max_epu8( max2, smax3 ); + __m128i min3 = _mm_min_epu8( min2, smin3 ); + __m128i smax4 = _mm_alignr_epi8( max3, max3, 1 ); + __m128i smin4 = _mm_alignr_epi8( min3, min3, 1 ); + __m128i max = _mm_max_epu8( max3, smax4 ); + __m128i min = _mm_min_epu8( min3, smin4 ); + __m128i max16 = _mm_unpacklo_epi8( max, _mm_setzero_si128() ); + __m128i min16 = _mm_unpacklo_epi8( min, _mm_setzero_si128() ); + + // src range, mid + __m128i srcRange = _mm_sub_epi16( max16, min16 ); + __m128i srcRangeHalf = _mm_srli_epi16( srcRange, 1 ); + __m128i srcMid = _mm_add_epi16( min16, srcRangeHalf ); + + // multiplier + __m128i mul1 = _mm_mulhi_epi16( srcRange, g_alphaRange_SIMD ); + __m128i mul = _mm_add_epi16( mul1, _mm_set1_epi16( 1 ) ); + + // wide source + __m128i s16_1 = _mm_shuffle_epi32( s, _MM_SHUFFLE( 3, 2, 3, 2 ) ); + __m128i s16[2] = { _mm_unpacklo_epi8( s, _mm_setzero_si128() ), _mm_unpacklo_epi8( s16_1, _mm_setzero_si128() ) }; + + __m128i sr[16] = { + Widen<0>( s16[0] ), + Widen<1>( s16[0] ), + Widen<2>( s16[0] ), + Widen<3>( s16[0] ), + Widen<4>( s16[0] ), + Widen<5>( s16[0] ), + Widen<6>( s16[0] ), + Widen<7>( s16[0] ), + Widen<0>( s16[1] ), + Widen<1>( s16[1] ), + Widen<2>( s16[1] ), + Widen<3>( s16[1] ), + Widen<4>( s16[1] ), + Widen<5>( s16[1] ), + Widen<6>( s16[1] ), + Widen<7>( s16[1] ) + }; + +#ifdef __AVX2__ + __m256i srcRangeWide = _mm256_broadcastsi128_si256( srcRange ); + __m256i srcMidWide = _mm256_broadcastsi128_si256( srcMid ); + + __m256i mulWide1 = _mm256_mulhi_epi16( srcRangeWide, g_alphaRange_AVX ); + __m256i mulWide = _mm256_add_epi16( mulWide1, _mm256_set1_epi16( 1 ) ); + + __m256i modMul[8] = { + _mm256_unpacklo_epi8( _mm256_packus_epi16( _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[0] ) ), _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[0] ) ) ), _mm256_setzero_si256() ), + _mm256_unpacklo_epi8( _mm256_packus_epi16( _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[1] ) ), _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[1] ) ) ), _mm256_setzero_si256() ), + _mm256_unpacklo_epi8( _mm256_packus_epi16( _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[2] ) ), _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[2] ) ) ), _mm256_setzero_si256() ), + _mm256_unpacklo_epi8( _mm256_packus_epi16( _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[3] ) ), _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[3] ) ) ), _mm256_setzero_si256() ), + _mm256_unpacklo_epi8( _mm256_packus_epi16( _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[4] ) ), _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[4] ) ) ), _mm256_setzero_si256() ), + _mm256_unpacklo_epi8( _mm256_packus_epi16( _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[5] ) ), _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[5] ) ) ), _mm256_setzero_si256() ), + _mm256_unpacklo_epi8( _mm256_packus_epi16( _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[6] ) ), _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[6] ) ) ), _mm256_setzero_si256() ), + _mm256_unpacklo_epi8( _mm256_packus_epi16( _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[7] ) ), _mm256_add_epi16( srcMidWide, _mm256_mullo_epi16( mulWide, g_alpha_AVX[7] ) ) ), _mm256_setzero_si256() ), + }; + + // find selector + __m256i mulErr = _mm256_setzero_si256(); + for( int j=0; j<16; j++ ) + { + __m256i s16Wide = _mm256_broadcastsi128_si256( sr[j] ); + __m256i err1, err2; + + err1 = _mm256_sub_epi16( s16Wide, modMul[0] ); + __m256i localErr = _mm256_mullo_epi16( err1, err1 ); + + err1 = _mm256_sub_epi16( s16Wide, modMul[1] ); + err2 = _mm256_mullo_epi16( err1, err1 ); + localErr = _mm256_min_epu16( localErr, err2 ); + + err1 = _mm256_sub_epi16( s16Wide, modMul[2] ); + err2 = _mm256_mullo_epi16( err1, err1 ); + localErr = _mm256_min_epu16( localErr, err2 ); + + err1 = _mm256_sub_epi16( s16Wide, modMul[3] ); + err2 = _mm256_mullo_epi16( err1, err1 ); + localErr = _mm256_min_epu16( localErr, err2 ); + + err1 = _mm256_sub_epi16( s16Wide, modMul[4] ); + err2 = _mm256_mullo_epi16( err1, err1 ); + localErr = _mm256_min_epu16( localErr, err2 ); + + err1 = _mm256_sub_epi16( s16Wide, modMul[5] ); + err2 = _mm256_mullo_epi16( err1, err1 ); + localErr = _mm256_min_epu16( localErr, err2 ); + + err1 = _mm256_sub_epi16( s16Wide, modMul[6] ); + err2 = _mm256_mullo_epi16( err1, err1 ); + localErr = _mm256_min_epu16( localErr, err2 ); + + err1 = _mm256_sub_epi16( s16Wide, modMul[7] ); + err2 = _mm256_mullo_epi16( err1, err1 ); + localErr = _mm256_min_epu16( localErr, err2 ); + + // note that this can overflow, but since we're looking for the smallest error, it shouldn't matter + mulErr = _mm256_adds_epu16( mulErr, localErr ); + } + uint64_t minPos1 = _mm_cvtsi128_si64( _mm_minpos_epu16( _mm256_castsi256_si128( mulErr ) ) ); + uint64_t minPos2 = _mm_cvtsi128_si64( _mm_minpos_epu16( _mm256_extracti128_si256( mulErr, 1 ) ) ); + int sel = ( ( minPos1 & 0xFFFF ) < ( minPos2 & 0xFFFF ) ) ? ( minPos1 >> 16 ) : ( 8 + ( minPos2 >> 16 ) ); + + __m128i recVal16; + switch( sel ) + { + case 0: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<0>( mul ), g_alpha_SIMD[0] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<0>( mul ), g_alpha_SIMD[0] ) ) ), _mm_setzero_si128() ); + break; + case 1: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[1] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[1] ) ) ), _mm_setzero_si128() ); + break; + case 2: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[2] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[2] ) ) ), _mm_setzero_si128() ); + break; + case 3: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[3] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[3] ) ) ), _mm_setzero_si128() ); + break; + case 4: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<2>( mul ), g_alpha_SIMD[4] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<2>( mul ), g_alpha_SIMD[4] ) ) ), _mm_setzero_si128() ); + break; + case 5: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[5] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[5] ) ) ), _mm_setzero_si128() ); + break; + case 6: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[6] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[6] ) ) ), _mm_setzero_si128() ); + break; + case 7: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[7] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[7] ) ) ), _mm_setzero_si128() ); + break; + case 8: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[8] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[8] ) ) ), _mm_setzero_si128() ); + break; + case 9: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[9] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[9] ) ) ), _mm_setzero_si128() ); + break; + case 10: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[10] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[10] ) ) ), _mm_setzero_si128() ); + break; + case 11: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[11] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[11] ) ) ), _mm_setzero_si128() ); + break; + case 12: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[12] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[12] ) ) ), _mm_setzero_si128() ); + break; + case 13: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[13] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[13] ) ) ), _mm_setzero_si128() ); + break; + case 14: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<5>( mul ), g_alpha_SIMD[14] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<5>( mul ), g_alpha_SIMD[14] ) ) ), _mm_setzero_si128() ); + break; + case 15: + recVal16 = _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<5>( mul ), g_alpha_SIMD[15] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<5>( mul ), g_alpha_SIMD[15] ) ) ), _mm_setzero_si128() ); + break; + default: + assert( false ); + break; + } +#else + // wide multiplier + __m128i rangeMul[16] = { + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<0>( mul ), g_alpha_SIMD[0] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<0>( mul ), g_alpha_SIMD[0] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[1] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[1] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[2] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[2] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[3] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<1>( mul ), g_alpha_SIMD[3] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<2>( mul ), g_alpha_SIMD[4] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<2>( mul ), g_alpha_SIMD[4] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[5] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[5] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[6] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[6] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[7] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<3>( mul ), g_alpha_SIMD[7] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[8] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[8] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[9] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[9] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[10] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[10] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[11] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[11] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[12] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[12] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[13] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<4>( mul ), g_alpha_SIMD[13] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<5>( mul ), g_alpha_SIMD[14] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<5>( mul ), g_alpha_SIMD[14] ) ) ), _mm_setzero_si128() ), + _mm_unpacklo_epi8( _mm_packus_epi16( _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<5>( mul ), g_alpha_SIMD[15] ) ), _mm_add_epi16( srcMid, _mm_mullo_epi16( Widen<5>( mul ), g_alpha_SIMD[15] ) ) ), _mm_setzero_si128() ) + }; + + // find selector + int err = std::numeric_limits<int>::max(); + int sel; + for( int r=0; r<16; r++ ) + { + __m128i err1, err2, minerr; + __m128i recVal16 = rangeMul[r]; + int rangeErr; + + err1 = _mm_sub_epi16( sr[0], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr = _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[1], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[2], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[3], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[4], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[5], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[6], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[7], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[8], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[9], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[10], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[11], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[12], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[13], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[14], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + err1 = _mm_sub_epi16( sr[15], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + rangeErr += _mm_cvtsi128_si64( minerr ) & 0xFFFF; + + if( rangeErr < err ) + { + err = rangeErr; + sel = r; + if( err == 0 ) break; + } + } + + __m128i recVal16 = rangeMul[sel]; +#endif + + // find indices + __m128i err1, err2, minerr; + uint64_t idx = 0, tmp; + + err1 = _mm_sub_epi16( sr[0], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 15*3; + + err1 = _mm_sub_epi16( sr[1], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 14*3; + + err1 = _mm_sub_epi16( sr[2], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 13*3; + + err1 = _mm_sub_epi16( sr[3], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 12*3; + + err1 = _mm_sub_epi16( sr[4], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 11*3; + + err1 = _mm_sub_epi16( sr[5], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 10*3; + + err1 = _mm_sub_epi16( sr[6], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 9*3; + + err1 = _mm_sub_epi16( sr[7], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 8*3; + + err1 = _mm_sub_epi16( sr[8], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 7*3; + + err1 = _mm_sub_epi16( sr[9], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 6*3; + + err1 = _mm_sub_epi16( sr[10], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 5*3; + + err1 = _mm_sub_epi16( sr[11], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 4*3; + + err1 = _mm_sub_epi16( sr[12], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 3*3; + + err1 = _mm_sub_epi16( sr[13], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 2*3; + + err1 = _mm_sub_epi16( sr[14], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 1*3; + + err1 = _mm_sub_epi16( sr[15], recVal16 ); + err2 = _mm_mullo_epi16( err1, err1 ); + minerr = _mm_minpos_epu16( err2 ); + tmp = _mm_cvtsi128_si64( minerr ); + idx |= ( tmp >> 16 ) << 0*3; + + uint16_t rm[8]; + _mm_storeu_si128( (__m128i*)rm, mul ); + uint16_t sm = _mm_cvtsi128_si64( srcMid ); + + uint64_t d = ( uint64_t( sm ) << 56 ) | + ( uint64_t( rm[GetMulSel( sel )] ) << 52 ) | + ( uint64_t( sel ) << 48 ) | + idx; + + return _bswap64( d ); +#elif defined __ARM_NEON + + int16x8_t srcMidWide, multipliers; + int srcMid; + uint8x16_t srcAlphaBlock = vld1q_u8( src ); + { + uint8_t ref = src[0]; + uint8x16_t a0 = vdupq_n_u8( ref ); + uint8x16_t r = vceqq_u8( srcAlphaBlock, a0 ); + int64x2_t m = vreinterpretq_s64_u8( r ); + if( m[0] == -1 && m[1] == -1 ) + return ref; + + // srcRange +#ifdef __aarch64__ + uint8_t min = vminvq_u8( srcAlphaBlock ); + uint8_t max = vmaxvq_u8( srcAlphaBlock ); + uint8_t srcRange = max - min; + multipliers = vqaddq_s16( vshrq_n_s16( vqdmulhq_n_s16( g_alphaRange_NEON, srcRange ), 1 ), vdupq_n_s16( 1 ) ); + srcMid = min + srcRange / 2; + srcMidWide = vdupq_n_s16( srcMid ); +#else + uint8x8_t vmin = vpmin_u8( vget_low_u8( srcAlphaBlock ), vget_high_u8( srcAlphaBlock ) ); + vmin = vpmin_u8( vmin, vmin ); + vmin = vpmin_u8( vmin, vmin ); + vmin = vpmin_u8( vmin, vmin ); + uint8x8_t vmax = vpmax_u8( vget_low_u8( srcAlphaBlock ), vget_high_u8( srcAlphaBlock ) ); + vmax = vpmax_u8( vmax, vmax ); + vmax = vpmax_u8( vmax, vmax ); + vmax = vpmax_u8( vmax, vmax ); + + int16x8_t srcRangeWide = vreinterpretq_s16_u16( vsubl_u8( vmax, vmin ) ); + multipliers = vqaddq_s16( vshrq_n_s16( vqdmulhq_s16( g_alphaRange_NEON, srcRangeWide ), 1 ), vdupq_n_s16( 1 ) ); + srcMidWide = vsraq_n_s16( vreinterpretq_s16_u16(vmovl_u8(vmin)), srcRangeWide, 1); + srcMid = vgetq_lane_s16( srcMidWide, 0 ); +#endif + } + + // calculate reconstructed values +#define EAC_APPLY_16X( m ) m( 0 ) m( 1 ) m( 2 ) m( 3 ) m( 4 ) m( 5 ) m( 6 ) m( 7 ) m( 8 ) m( 9 ) m( 10 ) m( 11 ) m( 12 ) m( 13 ) m( 14 ) m( 15 ) + +#define EAC_RECONSTRUCT_VALUE( n ) vqmovun_s16( vmlaq_s16( srcMidWide, g_alpha_NEON[n], WidenMultiplier_EAC_NEON<n>( multipliers ) ) ), + uint8x8_t recVals[16] = { EAC_APPLY_16X( EAC_RECONSTRUCT_VALUE ) }; + + // find selector + int err = std::numeric_limits<int>::max(); + int sel = 0; + for( int r = 0; r < 16; r++ ) + { + uint8x8_t recVal = recVals[r]; + + int rangeErr = 0; +#define EAC_ACCUMULATE_ERROR( n ) rangeErr += MinError_EAC_NEON( ErrorProbe_EAC_NEON<n>( recVal, srcAlphaBlock ) ); + EAC_APPLY_16X( EAC_ACCUMULATE_ERROR ) + + if( rangeErr < err ) + { + err = rangeErr; + sel = r; + if ( err == 0 ) break; + } + } + + // combine results + uint64_t d = ( uint64_t( srcMid ) << 56 ) | + ( uint64_t( multipliers[GetMulSel( sel )] ) << 52 ) | + ( uint64_t( sel ) << 48); + + // generate indices + uint8x8_t recVal = recVals[sel]; +#define EAC_INSERT_INDEX(n) d |= MinErrorIndex_EAC_NEON<n>( recVal, srcAlphaBlock ); + EAC_APPLY_16X( EAC_INSERT_INDEX ) + + return _bswap64( d ); + +#undef EAC_APPLY_16X +#undef EAC_INSERT_INDEX +#undef EAC_ACCUMULATE_ERROR +#undef EAC_RECONSTRUCT_VALUE + +#else + { + bool solid = true; + const uint8_t* ptr = src + 1; + const uint8_t ref = *src; + for( int i=1; i<16; i++ ) + { + if( ref != *ptr++ ) + { + solid = false; + break; + } + } + if( solid ) + { + return ref; + } + } + + uint8_t min = src[0]; + uint8_t max = src[0]; + for( int i=1; i<16; i++ ) + { + if( min > src[i] ) min = src[i]; + else if( max < src[i] ) max = src[i]; + } + int srcRange = max - min; + int srcMid = min + srcRange / 2; + + uint8_t buf[16][16]; + int err = std::numeric_limits<int>::max(); + int sel; + int selmul; + for( int r=0; r<16; r++ ) + { + int mul = ( ( srcRange * g_alphaRange[r] ) >> 16 ) + 1; + + int rangeErr = 0; + for( int i=0; i<16; i++ ) + { + const auto srcVal = src[i]; + + int idx = 0; + const auto modVal = g_alpha[r][0] * mul; + const auto recVal = clampu8( srcMid + modVal ); + int localErr = sq( srcVal - recVal ); + + if( localErr != 0 ) + { + for( int j=1; j<8; j++ ) + { + const auto modVal = g_alpha[r][j] * mul; + const auto recVal = clampu8( srcMid + modVal ); + const auto errProbe = sq( srcVal - recVal ); + if( errProbe < localErr ) + { + localErr = errProbe; + idx = j; + } + } + } + + buf[r][i] = idx; + rangeErr += localErr; + } + + if( rangeErr < err ) + { + err = rangeErr; + sel = r; + selmul = mul; + if( err == 0 ) break; + } + } + + uint64_t d = ( uint64_t( srcMid ) << 56 ) | + ( uint64_t( selmul ) << 52 ) | + ( uint64_t( sel ) << 48 ); + + int offset = 45; + auto ptr = buf[sel]; + for( int i=0; i<16; i++ ) + { + d |= uint64_t( *ptr++ ) << offset; + offset -= 3; + } + + return _bswap64( d ); +#endif +} + + +void CompressEtc1Alpha( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ) +{ + int w = 0; + uint32_t buf[4*4]; + do + { +#ifdef __SSE4_1__ + __m128 px0 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 0 ) ) ); + __m128 px1 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 1 ) ) ); + __m128 px2 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 2 ) ) ); + __m128 px3 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 3 ) ) ); + + _MM_TRANSPOSE4_PS( px0, px1, px2, px3 ); + + __m128i c0 = _mm_castps_si128( px0 ); + __m128i c1 = _mm_castps_si128( px1 ); + __m128i c2 = _mm_castps_si128( px2 ); + __m128i c3 = _mm_castps_si128( px3 ); + + __m128i mask = _mm_setr_epi32( 0x03030303, 0x07070707, 0x0b0b0b0b, 0x0f0f0f0f ); + __m128i p0 = _mm_shuffle_epi8( c0, mask ); + __m128i p1 = _mm_shuffle_epi8( c1, mask ); + __m128i p2 = _mm_shuffle_epi8( c2, mask ); + __m128i p3 = _mm_shuffle_epi8( c3, mask ); + + _mm_store_si128( (__m128i*)(buf + 0), p0 ); + _mm_store_si128( (__m128i*)(buf + 4), p1 ); + _mm_store_si128( (__m128i*)(buf + 8), p2 ); + _mm_store_si128( (__m128i*)(buf + 12), p3 ); + + src += 4; +#else + auto ptr = buf; + for( int x=0; x<4; x++ ) + { + unsigned int a = *src >> 24; + *ptr++ = a | ( a << 8 ) | ( a << 16 ); + src += width; + a = *src >> 24; + *ptr++ = a | ( a << 8 ) | ( a << 16 ); + src += width; + a = *src >> 24; + *ptr++ = a | ( a << 8 ) | ( a << 16 ); + src += width; + a = *src >> 24; + *ptr++ = a | ( a << 8 ) | ( a << 16 ); + src -= width * 3 - 1; + } +#endif + if( ++w == width/4 ) + { + src += width * 3; + w = 0; + } + *dst++ = ProcessRGB( (uint8_t*)buf ); + } + while( --blocks ); +} + +void CompressEtc2Alpha( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ) +{ + int w = 0; + uint32_t buf[4*4]; + do + { +#ifdef __SSE4_1__ + __m128 px0 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 0 ) ) ); + __m128 px1 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 1 ) ) ); + __m128 px2 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 2 ) ) ); + __m128 px3 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 3 ) ) ); + + _MM_TRANSPOSE4_PS( px0, px1, px2, px3 ); + + __m128i c0 = _mm_castps_si128( px0 ); + __m128i c1 = _mm_castps_si128( px1 ); + __m128i c2 = _mm_castps_si128( px2 ); + __m128i c3 = _mm_castps_si128( px3 ); + + __m128i mask = _mm_setr_epi32( 0x03030303, 0x07070707, 0x0b0b0b0b, 0x0f0f0f0f ); + __m128i p0 = _mm_shuffle_epi8( c0, mask ); + __m128i p1 = _mm_shuffle_epi8( c1, mask ); + __m128i p2 = _mm_shuffle_epi8( c2, mask ); + __m128i p3 = _mm_shuffle_epi8( c3, mask ); + + _mm_store_si128( (__m128i*)(buf + 0), p0 ); + _mm_store_si128( (__m128i*)(buf + 4), p1 ); + _mm_store_si128( (__m128i*)(buf + 8), p2 ); + _mm_store_si128( (__m128i*)(buf + 12), p3 ); + + src += 4; +#else + auto ptr = buf; + for( int x=0; x<4; x++ ) + { + unsigned int a = *src >> 24; + *ptr++ = a | ( a << 8 ) | ( a << 16 ); + src += width; + a = *src >> 24; + *ptr++ = a | ( a << 8 ) | ( a << 16 ); + src += width; + a = *src >> 24; + *ptr++ = a | ( a << 8 ) | ( a << 16 ); + src += width; + a = *src >> 24; + *ptr++ = a | ( a << 8 ) | ( a << 16 ); + src -= width * 3 - 1; + } +#endif + if( ++w == width/4 ) + { + src += width * 3; + w = 0; + } + *dst++ = ProcessRGB_ETC2( (uint8_t*)buf ); + } + while( --blocks ); +} + +#include <chrono> +#include <thread> + +void CompressEtc1Rgb( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ) +{ + int w = 0; + uint32_t buf[4*4]; + do + { +#ifdef __SSE4_1__ + __m128 px0 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 0 ) ) ); + __m128 px1 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 1 ) ) ); + __m128 px2 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 2 ) ) ); + __m128 px3 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 3 ) ) ); + + _MM_TRANSPOSE4_PS( px0, px1, px2, px3 ); + + _mm_store_si128( (__m128i*)(buf + 0), _mm_castps_si128( px0 ) ); + _mm_store_si128( (__m128i*)(buf + 4), _mm_castps_si128( px1 ) ); + _mm_store_si128( (__m128i*)(buf + 8), _mm_castps_si128( px2 ) ); + _mm_store_si128( (__m128i*)(buf + 12), _mm_castps_si128( px3 ) ); + + src += 4; +#else + auto ptr = buf; + for( int x=0; x<4; x++ ) + { + *ptr++ = *src; + src += width; + *ptr++ = *src; + src += width; + *ptr++ = *src; + src += width; + *ptr++ = *src; + src -= width * 3 - 1; + } +#endif + if( ++w == width/4 ) + { + src += width * 3; + w = 0; + } + *dst++ = ProcessRGB( (uint8_t*)buf ); + } + while( --blocks ); +} + +void CompressEtc1RgbDither( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ) +{ + int w = 0; + uint32_t buf[4*4]; + do + { +#ifdef __SSE4_1__ + __m128 px0 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 0 ) ) ); + __m128 px1 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 1 ) ) ); + __m128 px2 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 2 ) ) ); + __m128 px3 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 3 ) ) ); + + _MM_TRANSPOSE4_PS( px0, px1, px2, px3 ); + +# ifdef __AVX2__ + DitherAvx2( (uint8_t*)buf, _mm_castps_si128( px0 ), _mm_castps_si128( px1 ), _mm_castps_si128( px2 ), _mm_castps_si128( px3 ) ); +# else + _mm_store_si128( (__m128i*)(buf + 0), _mm_castps_si128( px0 ) ); + _mm_store_si128( (__m128i*)(buf + 4), _mm_castps_si128( px1 ) ); + _mm_store_si128( (__m128i*)(buf + 8), _mm_castps_si128( px2 ) ); + _mm_store_si128( (__m128i*)(buf + 12), _mm_castps_si128( px3 ) ); + + Dither( (uint8_t*)buf ); +# endif + + src += 4; +#else + auto ptr = buf; + for( int x=0; x<4; x++ ) + { + *ptr++ = *src; + src += width; + *ptr++ = *src; + src += width; + *ptr++ = *src; + src += width; + *ptr++ = *src; + src -= width * 3 - 1; + } +#endif + if( ++w == width/4 ) + { + src += width * 3; + w = 0; + } + *dst++ = ProcessRGB( (uint8_t*)buf ); + } + while( --blocks ); +} + +void CompressEtc2Rgb( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ) +{ + int w = 0; + uint32_t buf[4*4]; + do + { +#ifdef __SSE4_1__ + __m128 px0 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 0 ) ) ); + __m128 px1 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 1 ) ) ); + __m128 px2 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 2 ) ) ); + __m128 px3 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 3 ) ) ); + + _MM_TRANSPOSE4_PS( px0, px1, px2, px3 ); + + _mm_store_si128( (__m128i*)(buf + 0), _mm_castps_si128( px0 ) ); + _mm_store_si128( (__m128i*)(buf + 4), _mm_castps_si128( px1 ) ); + _mm_store_si128( (__m128i*)(buf + 8), _mm_castps_si128( px2 ) ); + _mm_store_si128( (__m128i*)(buf + 12), _mm_castps_si128( px3 ) ); + + src += 4; +#else + auto ptr = buf; + for( int x=0; x<4; x++ ) + { + *ptr++ = *src; + src += width; + *ptr++ = *src; + src += width; + *ptr++ = *src; + src += width; + *ptr++ = *src; + src -= width * 3 - 1; + } +#endif + if( ++w == width/4 ) + { + src += width * 3; + w = 0; + } + *dst++ = ProcessRGB_ETC2( (uint8_t*)buf ); + } + while( --blocks ); +} + +void CompressEtc2Rgba( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ) +{ + int w = 0; + uint32_t rgba[4*4]; + uint8_t alpha[4*4]; + do + { +#ifdef __SSE4_1__ + __m128 px0 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 0 ) ) ); + __m128 px1 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 1 ) ) ); + __m128 px2 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 2 ) ) ); + __m128 px3 = _mm_castsi128_ps( _mm_loadu_si128( (__m128i*)( src + width * 3 ) ) ); + + _MM_TRANSPOSE4_PS( px0, px1, px2, px3 ); + + __m128i c0 = _mm_castps_si128( px0 ); + __m128i c1 = _mm_castps_si128( px1 ); + __m128i c2 = _mm_castps_si128( px2 ); + __m128i c3 = _mm_castps_si128( px3 ); + + _mm_store_si128( (__m128i*)(rgba + 0), c0 ); + _mm_store_si128( (__m128i*)(rgba + 4), c1 ); + _mm_store_si128( (__m128i*)(rgba + 8), c2 ); + _mm_store_si128( (__m128i*)(rgba + 12), c3 ); + + __m128i mask = _mm_setr_epi32( 0x0f0b0703, -1, -1, -1 ); + + __m128i a0 = _mm_shuffle_epi8( c0, mask ); + __m128i a1 = _mm_shuffle_epi8( c1, _mm_shuffle_epi32( mask, _MM_SHUFFLE( 3, 3, 0, 3 ) ) ); + __m128i a2 = _mm_shuffle_epi8( c2, _mm_shuffle_epi32( mask, _MM_SHUFFLE( 3, 0, 3, 3 ) ) ); + __m128i a3 = _mm_shuffle_epi8( c3, _mm_shuffle_epi32( mask, _MM_SHUFFLE( 0, 3, 3, 3 ) ) ); + + __m128i s0 = _mm_or_si128( a0, a1 ); + __m128i s1 = _mm_or_si128( a2, a3 ); + __m128i s2 = _mm_or_si128( s0, s1 ); + + _mm_store_si128( (__m128i*)alpha, s2 ); + + src += 4; +#else + auto ptr = rgba; + auto ptr8 = alpha; + for( int x=0; x<4; x++ ) + { + auto v = *src; + *ptr++ = v; + *ptr8++ = v >> 24; + src += width; + v = *src; + *ptr++ = v; + *ptr8++ = v >> 24; + src += width; + v = *src; + *ptr++ = v; + *ptr8++ = v >> 24; + src += width; + v = *src; + *ptr++ = v; + *ptr8++ = v >> 24; + src -= width * 3 - 1; + } +#endif + if( ++w == width/4 ) + { + src += width * 3; + w = 0; + } + *dst++ = ProcessAlpha_ETC2( alpha ); + *dst++ = ProcessRGB_ETC2( (uint8_t*)rgba ); + } + while( --blocks ); +} diff --git a/thirdparty/etcpak/ProcessRGB.hpp b/thirdparty/etcpak/ProcessRGB.hpp new file mode 100644 index 0000000000..c5555a5bb1 --- /dev/null +++ b/thirdparty/etcpak/ProcessRGB.hpp @@ -0,0 +1,13 @@ +#ifndef __PROCESSRGB_HPP__ +#define __PROCESSRGB_HPP__ + +#include <stdint.h> + +void CompressEtc1Alpha( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ); +void CompressEtc2Alpha( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ); +void CompressEtc1Rgb( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ); +void CompressEtc1RgbDither( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ); +void CompressEtc2Rgb( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ); +void CompressEtc2Rgba( const uint32_t* src, uint64_t* dst, uint32_t blocks, size_t width ); + +#endif diff --git a/thirdparty/etcpak/Semaphore.hpp b/thirdparty/etcpak/Semaphore.hpp new file mode 100644 index 0000000000..9e42dbb9e0 --- /dev/null +++ b/thirdparty/etcpak/Semaphore.hpp @@ -0,0 +1,46 @@ +#ifndef __DARKRL__SEMAPHORE_HPP__ +#define __DARKRL__SEMAPHORE_HPP__ + +#include <condition_variable> +#include <mutex> + +class Semaphore +{ +public: + Semaphore( int count ) : m_count( count ) {} + + void lock() + { + std::unique_lock<std::mutex> lock( m_mutex ); + m_cv.wait( lock, [this](){ return m_count != 0; } ); + m_count--; + } + + void unlock() + { + std::lock_guard<std::mutex> lock( m_mutex ); + m_count++; + m_cv.notify_one(); + } + + bool try_lock() + { + std::lock_guard<std::mutex> lock( m_mutex ); + if( m_count == 0 ) + { + return false; + } + else + { + m_count--; + return true; + } + } + +private: + std::mutex m_mutex; + std::condition_variable m_cv; + unsigned int m_count; +}; + +#endif diff --git a/thirdparty/etcpak/System.cpp b/thirdparty/etcpak/System.cpp new file mode 100644 index 0000000000..a09b289cb2 --- /dev/null +++ b/thirdparty/etcpak/System.cpp @@ -0,0 +1,68 @@ +#include <algorithm> +#ifdef _WIN32 +# include <windows.h> +#else +# include <pthread.h> +# include <unistd.h> +#endif + +#include "System.hpp" + +unsigned int System::CPUCores() +{ + static unsigned int cores = 0; + if( cores == 0 ) + { + int tmp; +#ifdef _WIN32 + SYSTEM_INFO info; + GetSystemInfo( &info ); + tmp = (int)info.dwNumberOfProcessors; +#else +# ifndef _SC_NPROCESSORS_ONLN +# ifdef _SC_NPROC_ONLN +# define _SC_NPROCESSORS_ONLN _SC_NPROC_ONLN +# elif defined _SC_CRAY_NCPU +# define _SC_NPROCESSORS_ONLN _SC_CRAY_NCPU +# endif +# endif + tmp = (int)(long)sysconf( _SC_NPROCESSORS_ONLN ); +#endif + cores = (unsigned int)std::max( tmp, 1 ); + } + return cores; +} + +void System::SetThreadName( std::thread& thread, const char* name ) +{ +#ifdef _MSC_VER + const DWORD MS_VC_EXCEPTION=0x406D1388; + +# pragma pack( push, 8 ) + struct THREADNAME_INFO + { + DWORD dwType; + LPCSTR szName; + DWORD dwThreadID; + DWORD dwFlags; + }; +# pragma pack(pop) + + DWORD ThreadId = GetThreadId( static_cast<HANDLE>( thread.native_handle() ) ); + THREADNAME_INFO info; + info.dwType = 0x1000; + info.szName = name; + info.dwThreadID = ThreadId; + info.dwFlags = 0; + + __try + { + RaiseException( MS_VC_EXCEPTION, 0, sizeof(info)/sizeof(ULONG_PTR), (ULONG_PTR*)&info ); + } + __except(EXCEPTION_EXECUTE_HANDLER) + { + } +#elif !defined(__APPLE__) + pthread_setname_np( thread.native_handle(), name ); +#endif +} diff --git a/thirdparty/etcpak/System.hpp b/thirdparty/etcpak/System.hpp new file mode 100644 index 0000000000..1a09bb15e1 --- /dev/null +++ b/thirdparty/etcpak/System.hpp @@ -0,0 +1,15 @@ +#ifndef __DARKRL__SYSTEM_HPP__ +#define __DARKRL__SYSTEM_HPP__ + +#include <thread> + +class System +{ +public: + System() = delete; + + static unsigned int CPUCores(); + static void SetThreadName( std::thread& thread, const char* name ); +}; + +#endif diff --git a/thirdparty/etcpak/Tables.cpp b/thirdparty/etcpak/Tables.cpp new file mode 100644 index 0000000000..5c7fd9cf61 --- /dev/null +++ b/thirdparty/etcpak/Tables.cpp @@ -0,0 +1,221 @@ +#include "Tables.hpp" + +const int32_t g_table[8][4] = { + { 2, 8, -2, -8 }, + { 5, 17, -5, -17 }, + { 9, 29, -9, -29 }, + { 13, 42, -13, -42 }, + { 18, 60, -18, -60 }, + { 24, 80, -24, -80 }, + { 33, 106, -33, -106 }, + { 47, 183, -47, -183 } +}; + +const int64_t g_table256[8][4] = { + { 2*256, 8*256, -2*256, -8*256 }, + { 5*256, 17*256, -5*256, -17*256 }, + { 9*256, 29*256, -9*256, -29*256 }, + { 13*256, 42*256, -13*256, -42*256 }, + { 18*256, 60*256, -18*256, -60*256 }, + { 24*256, 80*256, -24*256, -80*256 }, + { 33*256, 106*256, -33*256, -106*256 }, + { 47*256, 183*256, -47*256, -183*256 } +}; + +const uint32_t g_id[4][16] = { + { 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0 }, + { 3, 3, 2, 2, 3, 3, 2, 2, 3, 3, 2, 2, 3, 3, 2, 2 }, + { 5, 5, 5, 5, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4 }, + { 7, 7, 6, 6, 7, 7, 6, 6, 7, 7, 6, 6, 7, 7, 6, 6 } +}; + +const uint32_t g_avg2[16] = { + 0x00, + 0x11, + 0x22, + 0x33, + 0x44, + 0x55, + 0x66, + 0x77, + 0x88, + 0x99, + 0xAA, + 0xBB, + 0xCC, + 0xDD, + 0xEE, + 0xFF +}; + +const uint32_t g_flags[64] = { + 0x80800402, 0x80800402, 0x80800402, 0x80800402, + 0x80800402, 0x80800402, 0x80800402, 0x8080E002, + 0x80800402, 0x80800402, 0x8080E002, 0x8080E002, + 0x80800402, 0x8080E002, 0x8080E002, 0x8080E002, + 0x80000402, 0x80000402, 0x80000402, 0x80000402, + 0x80000402, 0x80000402, 0x80000402, 0x8000E002, + 0x80000402, 0x80000402, 0x8000E002, 0x8000E002, + 0x80000402, 0x8000E002, 0x8000E002, 0x8000E002, + 0x00800402, 0x00800402, 0x00800402, 0x00800402, + 0x00800402, 0x00800402, 0x00800402, 0x0080E002, + 0x00800402, 0x00800402, 0x0080E002, 0x0080E002, + 0x00800402, 0x0080E002, 0x0080E002, 0x0080E002, + 0x00000402, 0x00000402, 0x00000402, 0x00000402, + 0x00000402, 0x00000402, 0x00000402, 0x0000E002, + 0x00000402, 0x00000402, 0x0000E002, 0x0000E002, + 0x00000402, 0x0000E002, 0x0000E002, 0x0000E002 +}; + +const int32_t g_alpha[16][8] = { + { -3, -6, -9, -15, 2, 5, 8, 14 }, + { -3, -7, -10, -13, 2, 6, 9, 12 }, + { -2, -5, -8, -13, 1, 4, 7, 12 }, + { -2, -4, -6, -13, 1, 3, 5, 12 }, + { -3, -6, -8, -12, 2, 5, 7, 11 }, + { -3, -7, -9, -11, 2, 6, 8, 10 }, + { -4, -7, -8, -11, 3, 6, 7, 10 }, + { -3, -5, -8, -11, 2, 4, 7, 10 }, + { -2, -6, -8, -10, 1, 5, 7, 9 }, + { -2, -5, -8, -10, 1, 4, 7, 9 }, + { -2, -4, -8, -10, 1, 3, 7, 9 }, + { -2, -5, -7, -10, 1, 4, 6, 9 }, + { -3, -4, -7, -10, 2, 3, 6, 9 }, + { -1, -2, -3, -10, 0, 1, 2, 9 }, + { -4, -6, -8, -9, 3, 5, 7, 8 }, + { -3, -5, -7, -9, 2, 4, 6, 8 } +}; + +const int32_t g_alphaRange[16] = { + 0x100FF / ( 1 + g_alpha[0][7] - g_alpha[0][3] ), + 0x100FF / ( 1 + g_alpha[1][7] - g_alpha[1][3] ), + 0x100FF / ( 1 + g_alpha[2][7] - g_alpha[2][3] ), + 0x100FF / ( 1 + g_alpha[3][7] - g_alpha[3][3] ), + 0x100FF / ( 1 + g_alpha[4][7] - g_alpha[4][3] ), + 0x100FF / ( 1 + g_alpha[5][7] - g_alpha[5][3] ), + 0x100FF / ( 1 + g_alpha[6][7] - g_alpha[6][3] ), + 0x100FF / ( 1 + g_alpha[7][7] - g_alpha[7][3] ), + 0x100FF / ( 1 + g_alpha[8][7] - g_alpha[8][3] ), + 0x100FF / ( 1 + g_alpha[9][7] - g_alpha[9][3] ), + 0x100FF / ( 1 + g_alpha[10][7] - g_alpha[10][3] ), + 0x100FF / ( 1 + g_alpha[11][7] - g_alpha[11][3] ), + 0x100FF / ( 1 + g_alpha[12][7] - g_alpha[12][3] ), + 0x100FF / ( 1 + g_alpha[13][7] - g_alpha[13][3] ), + 0x100FF / ( 1 + g_alpha[14][7] - g_alpha[14][3] ), + 0x100FF / ( 1 + g_alpha[15][7] - g_alpha[15][3] ), +}; + +#ifdef __SSE4_1__ +const __m128i g_table_SIMD[2] = +{ + _mm_setr_epi16( 2, 5, 9, 13, 18, 24, 33, 47), + _mm_setr_epi16( 8, 17, 29, 42, 60, 80, 106, 183) +}; +const __m128i g_table128_SIMD[2] = +{ + _mm_setr_epi16( 2*128, 5*128, 9*128, 13*128, 18*128, 24*128, 33*128, 47*128), + _mm_setr_epi16( 8*128, 17*128, 29*128, 42*128, 60*128, 80*128, 106*128, 183*128) +}; +const __m128i g_table256_SIMD[4] = +{ + _mm_setr_epi32( 2*256, 5*256, 9*256, 13*256), + _mm_setr_epi32( 8*256, 17*256, 29*256, 42*256), + _mm_setr_epi32( 18*256, 24*256, 33*256, 47*256), + _mm_setr_epi32( 60*256, 80*256, 106*256, 183*256) +}; + +const __m128i g_alpha_SIMD[16] = { + _mm_setr_epi16( g_alpha[ 0][0], g_alpha[ 0][1], g_alpha[ 0][2], g_alpha[ 0][3], g_alpha[ 0][4], g_alpha[ 0][5], g_alpha[ 0][6], g_alpha[ 0][7] ), + _mm_setr_epi16( g_alpha[ 1][0], g_alpha[ 1][1], g_alpha[ 1][2], g_alpha[ 1][3], g_alpha[ 1][4], g_alpha[ 1][5], g_alpha[ 1][6], g_alpha[ 1][7] ), + _mm_setr_epi16( g_alpha[ 2][0], g_alpha[ 2][1], g_alpha[ 2][2], g_alpha[ 2][3], g_alpha[ 2][4], g_alpha[ 2][5], g_alpha[ 2][6], g_alpha[ 2][7] ), + _mm_setr_epi16( g_alpha[ 3][0], g_alpha[ 3][1], g_alpha[ 3][2], g_alpha[ 3][3], g_alpha[ 3][4], g_alpha[ 3][5], g_alpha[ 3][6], g_alpha[ 3][7] ), + _mm_setr_epi16( g_alpha[ 4][0], g_alpha[ 4][1], g_alpha[ 4][2], g_alpha[ 4][3], g_alpha[ 4][4], g_alpha[ 4][5], g_alpha[ 4][6], g_alpha[ 4][7] ), + _mm_setr_epi16( g_alpha[ 5][0], g_alpha[ 5][1], g_alpha[ 5][2], g_alpha[ 5][3], g_alpha[ 5][4], g_alpha[ 5][5], g_alpha[ 5][6], g_alpha[ 5][7] ), + _mm_setr_epi16( g_alpha[ 6][0], g_alpha[ 6][1], g_alpha[ 6][2], g_alpha[ 6][3], g_alpha[ 6][4], g_alpha[ 6][5], g_alpha[ 6][6], g_alpha[ 6][7] ), + _mm_setr_epi16( g_alpha[ 7][0], g_alpha[ 7][1], g_alpha[ 7][2], g_alpha[ 7][3], g_alpha[ 7][4], g_alpha[ 7][5], g_alpha[ 7][6], g_alpha[ 7][7] ), + _mm_setr_epi16( g_alpha[ 8][0], g_alpha[ 8][1], g_alpha[ 8][2], g_alpha[ 8][3], g_alpha[ 8][4], g_alpha[ 8][5], g_alpha[ 8][6], g_alpha[ 8][7] ), + _mm_setr_epi16( g_alpha[ 9][0], g_alpha[ 9][1], g_alpha[ 9][2], g_alpha[ 9][3], g_alpha[ 9][4], g_alpha[ 9][5], g_alpha[ 9][6], g_alpha[ 9][7] ), + _mm_setr_epi16( g_alpha[10][0], g_alpha[10][1], g_alpha[10][2], g_alpha[10][3], g_alpha[10][4], g_alpha[10][5], g_alpha[10][6], g_alpha[10][7] ), + _mm_setr_epi16( g_alpha[11][0], g_alpha[11][1], g_alpha[11][2], g_alpha[11][3], g_alpha[11][4], g_alpha[11][5], g_alpha[11][6], g_alpha[11][7] ), + _mm_setr_epi16( g_alpha[12][0], g_alpha[12][1], g_alpha[12][2], g_alpha[12][3], g_alpha[12][4], g_alpha[12][5], g_alpha[12][6], g_alpha[12][7] ), + _mm_setr_epi16( g_alpha[13][0], g_alpha[13][1], g_alpha[13][2], g_alpha[13][3], g_alpha[13][4], g_alpha[13][5], g_alpha[13][6], g_alpha[13][7] ), + _mm_setr_epi16( g_alpha[14][0], g_alpha[14][1], g_alpha[14][2], g_alpha[14][3], g_alpha[14][4], g_alpha[14][5], g_alpha[14][6], g_alpha[14][7] ), + _mm_setr_epi16( g_alpha[15][0], g_alpha[15][1], g_alpha[15][2], g_alpha[15][3], g_alpha[15][4], g_alpha[15][5], g_alpha[15][6], g_alpha[15][7] ), +}; + +const __m128i g_alphaRange_SIMD = _mm_setr_epi16( + g_alphaRange[0], + g_alphaRange[1], + g_alphaRange[4], + g_alphaRange[5], + g_alphaRange[8], + g_alphaRange[14], + 0, + 0 ); +#endif + +#ifdef __AVX2__ +const __m256i g_alpha_AVX[8] = { + _mm256_setr_epi16( g_alpha[ 0][0], g_alpha[ 1][0], g_alpha[ 2][0], g_alpha[ 3][0], g_alpha[ 4][0], g_alpha[ 5][0], g_alpha[ 6][0], g_alpha[ 7][0], g_alpha[ 8][0], g_alpha[ 9][0], g_alpha[10][0], g_alpha[11][0], g_alpha[12][0], g_alpha[13][0], g_alpha[14][0], g_alpha[15][0] ), + _mm256_setr_epi16( g_alpha[ 0][1], g_alpha[ 1][1], g_alpha[ 2][1], g_alpha[ 3][1], g_alpha[ 4][1], g_alpha[ 5][1], g_alpha[ 6][1], g_alpha[ 7][1], g_alpha[ 8][1], g_alpha[ 9][1], g_alpha[10][1], g_alpha[11][1], g_alpha[12][1], g_alpha[13][1], g_alpha[14][1], g_alpha[15][1] ), + _mm256_setr_epi16( g_alpha[ 0][2], g_alpha[ 1][2], g_alpha[ 2][2], g_alpha[ 3][2], g_alpha[ 4][2], g_alpha[ 5][2], g_alpha[ 6][2], g_alpha[ 7][2], g_alpha[ 8][2], g_alpha[ 9][2], g_alpha[10][2], g_alpha[11][2], g_alpha[12][2], g_alpha[13][2], g_alpha[14][2], g_alpha[15][2] ), + _mm256_setr_epi16( g_alpha[ 0][3], g_alpha[ 1][3], g_alpha[ 2][3], g_alpha[ 3][3], g_alpha[ 4][3], g_alpha[ 5][3], g_alpha[ 6][3], g_alpha[ 7][3], g_alpha[ 8][3], g_alpha[ 9][3], g_alpha[10][3], g_alpha[11][3], g_alpha[12][3], g_alpha[13][3], g_alpha[14][3], g_alpha[15][3] ), + _mm256_setr_epi16( g_alpha[ 0][4], g_alpha[ 1][4], g_alpha[ 2][4], g_alpha[ 3][4], g_alpha[ 4][4], g_alpha[ 5][4], g_alpha[ 6][4], g_alpha[ 7][4], g_alpha[ 8][4], g_alpha[ 9][4], g_alpha[10][4], g_alpha[11][4], g_alpha[12][4], g_alpha[13][4], g_alpha[14][4], g_alpha[15][4] ), + _mm256_setr_epi16( g_alpha[ 0][5], g_alpha[ 1][5], g_alpha[ 2][5], g_alpha[ 3][5], g_alpha[ 4][5], g_alpha[ 5][5], g_alpha[ 6][5], g_alpha[ 7][5], g_alpha[ 8][5], g_alpha[ 9][5], g_alpha[10][5], g_alpha[11][5], g_alpha[12][5], g_alpha[13][5], g_alpha[14][5], g_alpha[15][5] ), + _mm256_setr_epi16( g_alpha[ 0][6], g_alpha[ 1][6], g_alpha[ 2][6], g_alpha[ 3][6], g_alpha[ 4][6], g_alpha[ 5][6], g_alpha[ 6][6], g_alpha[ 7][6], g_alpha[ 8][6], g_alpha[ 9][6], g_alpha[10][6], g_alpha[11][6], g_alpha[12][6], g_alpha[13][6], g_alpha[14][6], g_alpha[15][6] ), + _mm256_setr_epi16( g_alpha[ 0][7], g_alpha[ 1][7], g_alpha[ 2][7], g_alpha[ 3][7], g_alpha[ 4][7], g_alpha[ 5][7], g_alpha[ 6][7], g_alpha[ 7][7], g_alpha[ 8][7], g_alpha[ 9][7], g_alpha[10][7], g_alpha[11][7], g_alpha[12][7], g_alpha[13][7], g_alpha[14][7], g_alpha[15][7] ), +}; + +const __m256i g_alphaRange_AVX = _mm256_setr_epi16( + g_alphaRange[ 0], g_alphaRange[ 1], g_alphaRange[ 2], g_alphaRange[ 3], g_alphaRange[ 4], g_alphaRange[ 5], g_alphaRange[ 6], g_alphaRange[ 7], + g_alphaRange[ 8], g_alphaRange[ 9], g_alphaRange[10], g_alphaRange[11], g_alphaRange[12], g_alphaRange[13], g_alphaRange[14], g_alphaRange[15] +); +#endif + +#ifdef __ARM_NEON +const int16x8_t g_table128_NEON[2] = +{ + { 2*128, 5*128, 9*128, 13*128, 18*128, 24*128, 33*128, 47*128 }, + { 8*128, 17*128, 29*128, 42*128, 60*128, 80*128, 106*128, 183*128 } +}; + +const int32x4_t g_table256_NEON[4] = +{ + { 2*256, 5*256, 9*256, 13*256 }, + { 8*256, 17*256, 29*256, 42*256 }, + { 18*256, 24*256, 33*256, 47*256 }, + { 60*256, 80*256, 106*256, 183*256 } +}; + +const int16x8_t g_alpha_NEON[16] = +{ + { -3, -6, -9, -15, 2, 5, 8, 14 }, + { -3, -7, -10, -13, 2, 6, 9, 12 }, + { -2, -5, -8, -13, 1, 4, 7, 12 }, + { -2, -4, -6, -13, 1, 3, 5, 12 }, + { -3, -6, -8, -12, 2, 5, 7, 11 }, + { -3, -7, -9, -11, 2, 6, 8, 10 }, + { -4, -7, -8, -11, 3, 6, 7, 10 }, + { -3, -5, -8, -11, 2, 4, 7, 10 }, + { -2, -6, -8, -10, 1, 5, 7, 9 }, + { -2, -5, -8, -10, 1, 4, 7, 9 }, + { -2, -4, -8, -10, 1, 3, 7, 9 }, + { -2, -5, -7, -10, 1, 4, 6, 9 }, + { -3, -4, -7, -10, 2, 3, 6, 9 }, + { -1, -2, -3, -10, 0, 1, 2, 9 }, + { -4, -6, -8, -9, 3, 5, 7, 8 }, + { -3, -5, -7, -9, 2, 4, 6, 8 } +}; + +const int16x8_t g_alphaRange_NEON = +{ + (int16_t)g_alphaRange[0], + (int16_t)g_alphaRange[1], + (int16_t)g_alphaRange[4], + (int16_t)g_alphaRange[5], + (int16_t)g_alphaRange[8], + (int16_t)g_alphaRange[14], + 0, + 0 +}; +#endif diff --git a/thirdparty/etcpak/Tables.hpp b/thirdparty/etcpak/Tables.hpp new file mode 100644 index 0000000000..69d7e8aa07 --- /dev/null +++ b/thirdparty/etcpak/Tables.hpp @@ -0,0 +1,49 @@ +#ifndef __TABLES_HPP__ +#define __TABLES_HPP__ + +#include <stdint.h> + +#ifdef __AVX2__ +# include <immintrin.h> +#endif +#ifdef __SSE4_1__ +# include <smmintrin.h> +#endif +#ifdef __ARM_NEON +# include <arm_neon.h> +#endif + +extern const int32_t g_table[8][4]; +extern const int64_t g_table256[8][4]; + +extern const uint32_t g_id[4][16]; + +extern const uint32_t g_avg2[16]; + +extern const uint32_t g_flags[64]; + +extern const int32_t g_alpha[16][8]; +extern const int32_t g_alphaRange[16]; + +#ifdef __SSE4_1__ +extern const __m128i g_table_SIMD[2]; +extern const __m128i g_table128_SIMD[2]; +extern const __m128i g_table256_SIMD[4]; + +extern const __m128i g_alpha_SIMD[16]; +extern const __m128i g_alphaRange_SIMD; +#endif + +#ifdef __AVX2__ +extern const __m256i g_alpha_AVX[8]; +extern const __m256i g_alphaRange_AVX; +#endif + +#ifdef __ARM_NEON +extern const int16x8_t g_table128_NEON[2]; +extern const int32x4_t g_table256_NEON[4]; +extern const int16x8_t g_alpha_NEON[16]; +extern const int16x8_t g_alphaRange_NEON; +#endif + +#endif diff --git a/thirdparty/etcpak/TaskDispatch.cpp b/thirdparty/etcpak/TaskDispatch.cpp new file mode 100644 index 0000000000..7287da4de2 --- /dev/null +++ b/thirdparty/etcpak/TaskDispatch.cpp @@ -0,0 +1,115 @@ +#include <assert.h> +#include <stdio.h> + +#include "Debug.hpp" +#include "System.hpp" +#include "TaskDispatch.hpp" + +static TaskDispatch* s_instance = nullptr; + +TaskDispatch::TaskDispatch( size_t workers ) + : m_exit( false ) + , m_jobs( 0 ) +{ + assert( !s_instance ); + s_instance = this; + + assert( workers >= 1 ); + workers--; + + m_workers.reserve( workers ); + for( size_t i=0; i<workers; i++ ) + { + char tmp[16]; + sprintf( tmp, "Worker %zu", i ); +#ifdef __APPLE__ + auto worker = std::thread( [this, tmp]{ + pthread_setname_np( tmp ); + Worker(); + } ); +#else + auto worker = std::thread( [this]{ Worker(); } ); +#endif + System::SetThreadName( worker, tmp ); + m_workers.emplace_back( std::move( worker ) ); + } + + DBGPRINT( "Task dispatcher with " << m_workers.size() + 1 << " workers" ); +} + +TaskDispatch::~TaskDispatch() +{ + m_exit = true; + m_queueLock.lock(); + m_cvWork.notify_all(); + m_queueLock.unlock(); + + for( auto& worker : m_workers ) + { + worker.join(); + } + + assert( s_instance ); + s_instance = nullptr; +} + +void TaskDispatch::Queue( const std::function<void(void)>& f ) +{ + std::unique_lock<std::mutex> lock( s_instance->m_queueLock ); + s_instance->m_queue.emplace_back( f ); + const auto size = s_instance->m_queue.size(); + lock.unlock(); + if( size > 1 ) + { + s_instance->m_cvWork.notify_one(); + } +} + +void TaskDispatch::Queue( std::function<void(void)>&& f ) +{ + std::unique_lock<std::mutex> lock( s_instance->m_queueLock ); + s_instance->m_queue.emplace_back( std::move( f ) ); + const auto size = s_instance->m_queue.size(); + lock.unlock(); + if( size > 1 ) + { + s_instance->m_cvWork.notify_one(); + } +} + +void TaskDispatch::Sync() +{ + std::unique_lock<std::mutex> lock( s_instance->m_queueLock ); + while( !s_instance->m_queue.empty() ) + { + auto f = s_instance->m_queue.back(); + s_instance->m_queue.pop_back(); + lock.unlock(); + f(); + lock.lock(); + } + s_instance->m_cvJobs.wait( lock, []{ return s_instance->m_jobs == 0; } ); +} + +void TaskDispatch::Worker() +{ + for(;;) + { + std::unique_lock<std::mutex> lock( m_queueLock ); + m_cvWork.wait( lock, [this]{ return !m_queue.empty() || m_exit; } ); + if( m_exit ) return; + auto f = m_queue.back(); + m_queue.pop_back(); + m_jobs++; + lock.unlock(); + f(); + lock.lock(); + m_jobs--; + bool notify = m_jobs == 0 && m_queue.empty(); + lock.unlock(); + if( notify ) + { + m_cvJobs.notify_all(); + } + } +} diff --git a/thirdparty/etcpak/TaskDispatch.hpp b/thirdparty/etcpak/TaskDispatch.hpp new file mode 100644 index 0000000000..b513de4c0c --- /dev/null +++ b/thirdparty/etcpak/TaskDispatch.hpp @@ -0,0 +1,34 @@ +#ifndef __DARKRL__TASKDISPATCH_HPP__ +#define __DARKRL__TASKDISPATCH_HPP__ + +#include <atomic> +#include <condition_variable> +#include <functional> +#include <mutex> +#include <thread> +#include <vector> + +class TaskDispatch +{ +public: + TaskDispatch( size_t workers ); + ~TaskDispatch(); + + static void Queue( const std::function<void(void)>& f ); + static void Queue( std::function<void(void)>&& f ); + + static void Sync(); + +private: + void Worker(); + + std::vector<std::function<void(void)>> m_queue; + std::mutex m_queueLock; + std::condition_variable m_cvWork, m_cvJobs; + std::atomic<bool> m_exit; + size_t m_jobs; + + std::vector<std::thread> m_workers; +}; + +#endif diff --git a/thirdparty/etcpak/Timing.cpp b/thirdparty/etcpak/Timing.cpp new file mode 100644 index 0000000000..2af851f9a9 --- /dev/null +++ b/thirdparty/etcpak/Timing.cpp @@ -0,0 +1,8 @@ +#include <chrono> + +#include "Timing.hpp" + +uint64_t GetTime() +{ + return std::chrono::time_point_cast<std::chrono::microseconds>( std::chrono::high_resolution_clock::now() ).time_since_epoch().count(); +} diff --git a/thirdparty/etcpak/Timing.hpp b/thirdparty/etcpak/Timing.hpp new file mode 100644 index 0000000000..3767e20f24 --- /dev/null +++ b/thirdparty/etcpak/Timing.hpp @@ -0,0 +1,8 @@ +#ifndef __DARKRL__TIMING_HPP__ +#define __DARKRL__TIMING_HPP__ + +#include <stdint.h> + +uint64_t GetTime(); + +#endif diff --git a/thirdparty/etcpak/Vector.hpp b/thirdparty/etcpak/Vector.hpp new file mode 100644 index 0000000000..3370a88aea --- /dev/null +++ b/thirdparty/etcpak/Vector.hpp @@ -0,0 +1,222 @@ +#ifndef __DARKRL__VECTOR_HPP__ +#define __DARKRL__VECTOR_HPP__ + +#include <assert.h> +#include <algorithm> +#include <math.h> +#include <stdint.h> + +#include "Math.hpp" + +template<class T> +struct Vector2 +{ + Vector2() : x( 0 ), y( 0 ) {} + Vector2( T v ) : x( v ), y( v ) {} + Vector2( T _x, T _y ) : x( _x ), y( _y ) {} + + bool operator==( const Vector2<T>& rhs ) const { return x == rhs.x && y == rhs.y; } + bool operator!=( const Vector2<T>& rhs ) const { return !( *this == rhs ); } + + Vector2<T>& operator+=( const Vector2<T>& rhs ) + { + x += rhs.x; + y += rhs.y; + return *this; + } + Vector2<T>& operator-=( const Vector2<T>& rhs ) + { + x -= rhs.x; + y -= rhs.y; + return *this; + } + Vector2<T>& operator*=( const Vector2<T>& rhs ) + { + x *= rhs.x; + y *= rhs.y; + return *this; + } + + T x, y; +}; + +template<class T> +Vector2<T> operator+( const Vector2<T>& lhs, const Vector2<T>& rhs ) +{ + return Vector2<T>( lhs.x + rhs.x, lhs.y + rhs.y ); +} + +template<class T> +Vector2<T> operator-( const Vector2<T>& lhs, const Vector2<T>& rhs ) +{ + return Vector2<T>( lhs.x - rhs.x, lhs.y - rhs.y ); +} + +template<class T> +Vector2<T> operator*( const Vector2<T>& lhs, const float& rhs ) +{ + return Vector2<T>( lhs.x * rhs, lhs.y * rhs ); +} + +template<class T> +Vector2<T> operator/( const Vector2<T>& lhs, const T& rhs ) +{ + return Vector2<T>( lhs.x / rhs, lhs.y / rhs ); +} + + +typedef Vector2<int32_t> v2i; +typedef Vector2<float> v2f; + + +template<class T> +struct Vector3 +{ + Vector3() : x( 0 ), y( 0 ), z( 0 ) {} + Vector3( T v ) : x( v ), y( v ), z( v ) {} + Vector3( T _x, T _y, T _z ) : x( _x ), y( _y ), z( _z ) {} + template<class Y> + Vector3( const Vector3<Y>& v ) : x( T( v.x ) ), y( T( v.y ) ), z( T( v.z ) ) {} + + T Luminance() const { return T( x * 0.3f + y * 0.59f + z * 0.11f ); } + void Clamp() + { + x = std::min( T(1), std::max( T(0), x ) ); + y = std::min( T(1), std::max( T(0), y ) ); + z = std::min( T(1), std::max( T(0), z ) ); + } + + bool operator==( const Vector3<T>& rhs ) const { return x == rhs.x && y == rhs.y && z == rhs.z; } + bool operator!=( const Vector2<T>& rhs ) const { return !( *this == rhs ); } + + T& operator[]( unsigned int idx ) { assert( idx < 3 ); return ((T*)this)[idx]; } + const T& operator[]( unsigned int idx ) const { assert( idx < 3 ); return ((T*)this)[idx]; } + + Vector3<T> operator+=( const Vector3<T>& rhs ) + { + x += rhs.x; + y += rhs.y; + z += rhs.z; + return *this; + } + + Vector3<T> operator*=( const Vector3<T>& rhs ) + { + x *= rhs.x; + y *= rhs.y; + z *= rhs.z; + return *this; + } + + Vector3<T> operator*=( const float& rhs ) + { + x *= rhs; + y *= rhs; + z *= rhs; + return *this; + } + + T x, y, z; + T padding; +}; + +template<class T> +Vector3<T> operator+( const Vector3<T>& lhs, const Vector3<T>& rhs ) +{ + return Vector3<T>( lhs.x + rhs.x, lhs.y + rhs.y, lhs.z + rhs.z ); +} + +template<class T> +Vector3<T> operator-( const Vector3<T>& lhs, const Vector3<T>& rhs ) +{ + return Vector3<T>( lhs.x - rhs.x, lhs.y - rhs.y, lhs.z - rhs.z ); +} + +template<class T> +Vector3<T> operator*( const Vector3<T>& lhs, const Vector3<T>& rhs ) +{ + return Vector3<T>( lhs.x * rhs.x, lhs.y * rhs.y, lhs.z * rhs.z ); +} + +template<class T> +Vector3<T> operator*( const Vector3<T>& lhs, const float& rhs ) +{ + return Vector3<T>( T( lhs.x * rhs ), T( lhs.y * rhs ), T( lhs.z * rhs ) ); +} + +template<class T> +Vector3<T> operator/( const Vector3<T>& lhs, const T& rhs ) +{ + return Vector3<T>( lhs.x / rhs, lhs.y / rhs, lhs.z / rhs ); +} + +template<class T> +bool operator<( const Vector3<T>& lhs, const Vector3<T>& rhs ) +{ + return lhs.Luminance() < rhs.Luminance(); +} + +typedef Vector3<int32_t> v3i; +typedef Vector3<float> v3f; +typedef Vector3<uint8_t> v3b; + + +static inline v3b v3f_to_v3b( const v3f& v ) +{ + return v3b( uint8_t( std::min( 1.f, v.x ) * 255 ), uint8_t( std::min( 1.f, v.y ) * 255 ), uint8_t( std::min( 1.f, v.z ) * 255 ) ); +} + +template<class T> +Vector3<T> Mix( const Vector3<T>& v1, const Vector3<T>& v2, float amount ) +{ + return v1 + ( v2 - v1 ) * amount; +} + +template<> +inline v3b Mix( const v3b& v1, const v3b& v2, float amount ) +{ + return v3b( v3f( v1 ) + ( v3f( v2 ) - v3f( v1 ) ) * amount ); +} + +template<class T> +Vector3<T> Desaturate( const Vector3<T>& v ) +{ + T l = v.Luminance(); + return Vector3<T>( l, l, l ); +} + +template<class T> +Vector3<T> Desaturate( const Vector3<T>& v, float mul ) +{ + T l = T( v.Luminance() * mul ); + return Vector3<T>( l, l, l ); +} + +template<class T> +Vector3<T> pow( const Vector3<T>& base, float exponent ) +{ + return Vector3<T>( + pow( base.x, exponent ), + pow( base.y, exponent ), + pow( base.z, exponent ) ); +} + +template<class T> +Vector3<T> sRGB2linear( const Vector3<T>& v ) +{ + return Vector3<T>( + sRGB2linear( v.x ), + sRGB2linear( v.y ), + sRGB2linear( v.z ) ); +} + +template<class T> +Vector3<T> linear2sRGB( const Vector3<T>& v ) +{ + return Vector3<T>( + linear2sRGB( v.x ), + linear2sRGB( v.y ), + linear2sRGB( v.z ) ); +} + +#endif diff --git a/thirdparty/etcpak/lz4/lz4.c b/thirdparty/etcpak/lz4/lz4.c new file mode 100644 index 0000000000..08cf6b5cd7 --- /dev/null +++ b/thirdparty/etcpak/lz4/lz4.c @@ -0,0 +1,1516 @@ +/* + LZ4 - Fast LZ compression algorithm + Copyright (C) 2011-2015, Yann Collet. + + BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php) + + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions are + met: + + * Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above + copyright notice, this list of conditions and the following disclaimer + in the documentation and/or other materials provided with the + distribution. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + You can contact the author at : + - LZ4 source repository : https://github.com/Cyan4973/lz4 + - LZ4 public forum : https://groups.google.com/forum/#!forum/lz4c +*/ + + +/************************************** +* Tuning parameters +**************************************/ +/* + * HEAPMODE : + * Select how default compression functions will allocate memory for their hash table, + * in memory stack (0:default, fastest), or in memory heap (1:requires malloc()). + */ +#define HEAPMODE 0 + +/* + * ACCELERATION_DEFAULT : + * Select "acceleration" for LZ4_compress_fast() when parameter value <= 0 + */ +#define ACCELERATION_DEFAULT 1 + + +/************************************** +* CPU Feature Detection +**************************************/ +/* + * LZ4_FORCE_SW_BITCOUNT + * Define this parameter if your target system or compiler does not support hardware bit count + */ +#if defined(_MSC_VER) && defined(_WIN32_WCE) /* Visual Studio for Windows CE does not support Hardware bit count */ +# define LZ4_FORCE_SW_BITCOUNT +#endif + + +/************************************** +* Includes +**************************************/ +#include "lz4.h" + + +/************************************** +* Compiler Options +**************************************/ +#ifdef _MSC_VER /* Visual Studio */ +# define FORCE_INLINE static __forceinline +# include <intrin.h> +# pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */ +# pragma warning(disable : 4293) /* disable: C4293: too large shift (32-bits) */ +#else +# if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */ +# if defined(__GNUC__) || defined(__clang__) +# define FORCE_INLINE static inline __attribute__((always_inline)) +# else +# define FORCE_INLINE static inline +# endif +# else +# define FORCE_INLINE static +# endif /* __STDC_VERSION__ */ +#endif /* _MSC_VER */ + +/* LZ4_GCC_VERSION is defined into lz4.h */ +#if (LZ4_GCC_VERSION >= 302) || (__INTEL_COMPILER >= 800) || defined(__clang__) +# define expect(expr,value) (__builtin_expect ((expr),(value)) ) +#else +# define expect(expr,value) (expr) +#endif + +#define likely(expr) expect((expr) != 0, 1) +#define unlikely(expr) expect((expr) != 0, 0) + + +/************************************** +* Memory routines +**************************************/ +#include <stdlib.h> /* malloc, calloc, free */ +#define ALLOCATOR(n,s) calloc(n,s) +#define FREEMEM free +#include <string.h> /* memset, memcpy */ +#define MEM_INIT memset + + +/************************************** +* Basic Types +**************************************/ +#if defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */ +# include <stdint.h> + typedef uint8_t BYTE; + typedef uint16_t U16; + typedef uint32_t U32; + typedef int32_t S32; + typedef uint64_t U64; +#else + typedef unsigned char BYTE; + typedef unsigned short U16; + typedef unsigned int U32; + typedef signed int S32; + typedef unsigned long long U64; +#endif + + +/************************************** +* Reading and writing into memory +**************************************/ +#define STEPSIZE sizeof(size_t) + +static unsigned LZ4_64bits(void) { return sizeof(void*)==8; } + +static unsigned LZ4_isLittleEndian(void) +{ + const union { U32 i; BYTE c[4]; } one = { 1 }; /* don't use static : performance detrimental */ + return one.c[0]; +} + + +static U16 LZ4_read16(const void* memPtr) +{ + U16 val16; + memcpy(&val16, memPtr, 2); + return val16; +} + +static U16 LZ4_readLE16(const void* memPtr) +{ + if (LZ4_isLittleEndian()) + { + return LZ4_read16(memPtr); + } + else + { + const BYTE* p = (const BYTE*)memPtr; + return (U16)((U16)p[0] + (p[1]<<8)); + } +} + +static void LZ4_writeLE16(void* memPtr, U16 value) +{ + if (LZ4_isLittleEndian()) + { + memcpy(memPtr, &value, 2); + } + else + { + BYTE* p = (BYTE*)memPtr; + p[0] = (BYTE) value; + p[1] = (BYTE)(value>>8); + } +} + +static U32 LZ4_read32(const void* memPtr) +{ + U32 val32; + memcpy(&val32, memPtr, 4); + return val32; +} + +static U64 LZ4_read64(const void* memPtr) +{ + U64 val64; + memcpy(&val64, memPtr, 8); + return val64; +} + +static size_t LZ4_read_ARCH(const void* p) +{ + if (LZ4_64bits()) + return (size_t)LZ4_read64(p); + else + return (size_t)LZ4_read32(p); +} + + +static void LZ4_copy4(void* dstPtr, const void* srcPtr) { memcpy(dstPtr, srcPtr, 4); } + +static void LZ4_copy8(void* dstPtr, const void* srcPtr) { memcpy(dstPtr, srcPtr, 8); } + +/* customized version of memcpy, which may overwrite up to 7 bytes beyond dstEnd */ +static void LZ4_wildCopy(void* dstPtr, const void* srcPtr, void* dstEnd) +{ + BYTE* d = (BYTE*)dstPtr; + const BYTE* s = (const BYTE*)srcPtr; + BYTE* e = (BYTE*)dstEnd; + do { LZ4_copy8(d,s); d+=8; s+=8; } while (d<e); +} + + +/************************************** +* Common Constants +**************************************/ +#define MINMATCH 4 + +#define COPYLENGTH 8 +#define LASTLITERALS 5 +#define MFLIMIT (COPYLENGTH+MINMATCH) +static const int LZ4_minLength = (MFLIMIT+1); + +#define KB *(1 <<10) +#define MB *(1 <<20) +#define GB *(1U<<30) + +#define MAXD_LOG 16 +#define MAX_DISTANCE ((1 << MAXD_LOG) - 1) + +#define ML_BITS 4 +#define ML_MASK ((1U<<ML_BITS)-1) +#define RUN_BITS (8-ML_BITS) +#define RUN_MASK ((1U<<RUN_BITS)-1) + + +/************************************** +* Common Utils +**************************************/ +#define LZ4_STATIC_ASSERT(c) { enum { LZ4_static_assert = 1/(int)(!!(c)) }; } /* use only *after* variable declarations */ + + +/************************************** +* Common functions +**************************************/ +static unsigned LZ4_NbCommonBytes (register size_t val) +{ + if (LZ4_isLittleEndian()) + { + if (LZ4_64bits()) + { +# if defined(_MSC_VER) && defined(_WIN64) && !defined(LZ4_FORCE_SW_BITCOUNT) + unsigned long r = 0; + _BitScanForward64( &r, (U64)val ); + return (int)(r>>3); +# elif (defined(__clang__) || (LZ4_GCC_VERSION >= 304)) && !defined(LZ4_FORCE_SW_BITCOUNT) + return (__builtin_ctzll((U64)val) >> 3); +# else + static const int DeBruijnBytePos[64] = { 0, 0, 0, 0, 0, 1, 1, 2, 0, 3, 1, 3, 1, 4, 2, 7, 0, 2, 3, 6, 1, 5, 3, 5, 1, 3, 4, 4, 2, 5, 6, 7, 7, 0, 1, 2, 3, 3, 4, 6, 2, 6, 5, 5, 3, 4, 5, 6, 7, 1, 2, 4, 6, 4, 4, 5, 7, 2, 6, 5, 7, 6, 7, 7 }; + return DeBruijnBytePos[((U64)((val & -(long long)val) * 0x0218A392CDABBD3FULL)) >> 58]; +# endif + } + else /* 32 bits */ + { +# if defined(_MSC_VER) && !defined(LZ4_FORCE_SW_BITCOUNT) + unsigned long r; + _BitScanForward( &r, (U32)val ); + return (int)(r>>3); +# elif (defined(__clang__) || (LZ4_GCC_VERSION >= 304)) && !defined(LZ4_FORCE_SW_BITCOUNT) + return (__builtin_ctz((U32)val) >> 3); +# else + static const int DeBruijnBytePos[32] = { 0, 0, 3, 0, 3, 1, 3, 0, 3, 2, 2, 1, 3, 2, 0, 1, 3, 3, 1, 2, 2, 2, 2, 0, 3, 1, 2, 0, 1, 0, 1, 1 }; + return DeBruijnBytePos[((U32)((val & -(S32)val) * 0x077CB531U)) >> 27]; +# endif + } + } + else /* Big Endian CPU */ + { + if (LZ4_64bits()) + { +# if defined(_MSC_VER) && defined(_WIN64) && !defined(LZ4_FORCE_SW_BITCOUNT) + unsigned long r = 0; + _BitScanReverse64( &r, val ); + return (unsigned)(r>>3); +# elif (defined(__clang__) || (LZ4_GCC_VERSION >= 304)) && !defined(LZ4_FORCE_SW_BITCOUNT) + return (__builtin_clzll((U64)val) >> 3); +# else + unsigned r; + if (!(val>>32)) { r=4; } else { r=0; val>>=32; } + if (!(val>>16)) { r+=2; val>>=8; } else { val>>=24; } + r += (!val); + return r; +# endif + } + else /* 32 bits */ + { +# if defined(_MSC_VER) && !defined(LZ4_FORCE_SW_BITCOUNT) + unsigned long r = 0; + _BitScanReverse( &r, (unsigned long)val ); + return (unsigned)(r>>3); +# elif (defined(__clang__) || (LZ4_GCC_VERSION >= 304)) && !defined(LZ4_FORCE_SW_BITCOUNT) + return (__builtin_clz((U32)val) >> 3); +# else + unsigned r; + if (!(val>>16)) { r=2; val>>=8; } else { r=0; val>>=24; } + r += (!val); + return r; +# endif + } + } +} + +static unsigned LZ4_count(const BYTE* pIn, const BYTE* pMatch, const BYTE* pInLimit) +{ + const BYTE* const pStart = pIn; + + while (likely(pIn<pInLimit-(STEPSIZE-1))) + { + size_t diff = LZ4_read_ARCH(pMatch) ^ LZ4_read_ARCH(pIn); + if (!diff) { pIn+=STEPSIZE; pMatch+=STEPSIZE; continue; } + pIn += LZ4_NbCommonBytes(diff); + return (unsigned)(pIn - pStart); + } + + if (LZ4_64bits()) if ((pIn<(pInLimit-3)) && (LZ4_read32(pMatch) == LZ4_read32(pIn))) { pIn+=4; pMatch+=4; } + if ((pIn<(pInLimit-1)) && (LZ4_read16(pMatch) == LZ4_read16(pIn))) { pIn+=2; pMatch+=2; } + if ((pIn<pInLimit) && (*pMatch == *pIn)) pIn++; + return (unsigned)(pIn - pStart); +} + + +#ifndef LZ4_COMMONDEFS_ONLY +/************************************** +* Local Constants +**************************************/ +#define LZ4_HASHLOG (LZ4_MEMORY_USAGE-2) +#define HASHTABLESIZE (1 << LZ4_MEMORY_USAGE) +#define HASH_SIZE_U32 (1 << LZ4_HASHLOG) /* required as macro for static allocation */ + +static const int LZ4_64Klimit = ((64 KB) + (MFLIMIT-1)); +static const U32 LZ4_skipTrigger = 6; /* Increase this value ==> compression run slower on incompressible data */ + + +/************************************** +* Local Structures and types +**************************************/ +typedef struct { + U32 hashTable[HASH_SIZE_U32]; + U32 currentOffset; + U32 initCheck; + const BYTE* dictionary; + BYTE* bufferStart; /* obsolete, used for slideInputBuffer */ + U32 dictSize; +} LZ4_stream_t_internal; + +typedef enum { notLimited = 0, limitedOutput = 1 } limitedOutput_directive; +typedef enum { byPtr, byU32, byU16 } tableType_t; + +typedef enum { noDict = 0, withPrefix64k, usingExtDict } dict_directive; +typedef enum { noDictIssue = 0, dictSmall } dictIssue_directive; + +typedef enum { endOnOutputSize = 0, endOnInputSize = 1 } endCondition_directive; +typedef enum { full = 0, partial = 1 } earlyEnd_directive; + + +/************************************** +* Local Utils +**************************************/ +int LZ4_versionNumber (void) { return LZ4_VERSION_NUMBER; } +int LZ4_compressBound(int isize) { return LZ4_COMPRESSBOUND(isize); } +int LZ4_sizeofState() { return LZ4_STREAMSIZE; } + + + +/******************************** +* Compression functions +********************************/ + +static U32 LZ4_hashSequence(U32 sequence, tableType_t const tableType) +{ + if (tableType == byU16) + return (((sequence) * 2654435761U) >> ((MINMATCH*8)-(LZ4_HASHLOG+1))); + else + return (((sequence) * 2654435761U) >> ((MINMATCH*8)-LZ4_HASHLOG)); +} + +static const U64 prime5bytes = 889523592379ULL; +static U32 LZ4_hashSequence64(size_t sequence, tableType_t const tableType) +{ + const U32 hashLog = (tableType == byU16) ? LZ4_HASHLOG+1 : LZ4_HASHLOG; + const U32 hashMask = (1<<hashLog) - 1; + return ((sequence * prime5bytes) >> (40 - hashLog)) & hashMask; +} + +static U32 LZ4_hashSequenceT(size_t sequence, tableType_t const tableType) +{ + if (LZ4_64bits()) + return LZ4_hashSequence64(sequence, tableType); + return LZ4_hashSequence((U32)sequence, tableType); +} + +static U32 LZ4_hashPosition(const void* p, tableType_t tableType) { return LZ4_hashSequenceT(LZ4_read_ARCH(p), tableType); } + +static void LZ4_putPositionOnHash(const BYTE* p, U32 h, void* tableBase, tableType_t const tableType, const BYTE* srcBase) +{ + switch (tableType) + { + case byPtr: { const BYTE** hashTable = (const BYTE**)tableBase; hashTable[h] = p; return; } + case byU32: { U32* hashTable = (U32*) tableBase; hashTable[h] = (U32)(p-srcBase); return; } + case byU16: { U16* hashTable = (U16*) tableBase; hashTable[h] = (U16)(p-srcBase); return; } + } +} + +static void LZ4_putPosition(const BYTE* p, void* tableBase, tableType_t tableType, const BYTE* srcBase) +{ + U32 h = LZ4_hashPosition(p, tableType); + LZ4_putPositionOnHash(p, h, tableBase, tableType, srcBase); +} + +static const BYTE* LZ4_getPositionOnHash(U32 h, void* tableBase, tableType_t tableType, const BYTE* srcBase) +{ + if (tableType == byPtr) { const BYTE** hashTable = (const BYTE**) tableBase; return hashTable[h]; } + if (tableType == byU32) { U32* hashTable = (U32*) tableBase; return hashTable[h] + srcBase; } + { U16* hashTable = (U16*) tableBase; return hashTable[h] + srcBase; } /* default, to ensure a return */ +} + +static const BYTE* LZ4_getPosition(const BYTE* p, void* tableBase, tableType_t tableType, const BYTE* srcBase) +{ + U32 h = LZ4_hashPosition(p, tableType); + return LZ4_getPositionOnHash(h, tableBase, tableType, srcBase); +} + +FORCE_INLINE int LZ4_compress_generic( + void* const ctx, + const char* const source, + char* const dest, + const int inputSize, + const int maxOutputSize, + const limitedOutput_directive outputLimited, + const tableType_t tableType, + const dict_directive dict, + const dictIssue_directive dictIssue, + const U32 acceleration) +{ + LZ4_stream_t_internal* const dictPtr = (LZ4_stream_t_internal*)ctx; + + const BYTE* ip = (const BYTE*) source; + const BYTE* base; + const BYTE* lowLimit; + const BYTE* const lowRefLimit = ip - dictPtr->dictSize; + const BYTE* const dictionary = dictPtr->dictionary; + const BYTE* const dictEnd = dictionary + dictPtr->dictSize; + const size_t dictDelta = dictEnd - (const BYTE*)source; + const BYTE* anchor = (const BYTE*) source; + const BYTE* const iend = ip + inputSize; + const BYTE* const mflimit = iend - MFLIMIT; + const BYTE* const matchlimit = iend - LASTLITERALS; + + BYTE* op = (BYTE*) dest; + BYTE* const olimit = op + maxOutputSize; + + U32 forwardH; + size_t refDelta=0; + + /* Init conditions */ + if ((U32)inputSize > (U32)LZ4_MAX_INPUT_SIZE) return 0; /* Unsupported input size, too large (or negative) */ + switch(dict) + { + case noDict: + default: + base = (const BYTE*)source; + lowLimit = (const BYTE*)source; + break; + case withPrefix64k: + base = (const BYTE*)source - dictPtr->currentOffset; + lowLimit = (const BYTE*)source - dictPtr->dictSize; + break; + case usingExtDict: + base = (const BYTE*)source - dictPtr->currentOffset; + lowLimit = (const BYTE*)source; + break; + } + if ((tableType == byU16) && (inputSize>=LZ4_64Klimit)) return 0; /* Size too large (not within 64K limit) */ + if (inputSize<LZ4_minLength) goto _last_literals; /* Input too small, no compression (all literals) */ + + /* First Byte */ + LZ4_putPosition(ip, ctx, tableType, base); + ip++; forwardH = LZ4_hashPosition(ip, tableType); + + /* Main Loop */ + for ( ; ; ) + { + const BYTE* match; + BYTE* token; + { + const BYTE* forwardIp = ip; + unsigned step = 1; + unsigned searchMatchNb = acceleration << LZ4_skipTrigger; + + /* Find a match */ + do { + U32 h = forwardH; + ip = forwardIp; + forwardIp += step; + step = (searchMatchNb++ >> LZ4_skipTrigger); + + if (unlikely(forwardIp > mflimit)) goto _last_literals; + + match = LZ4_getPositionOnHash(h, ctx, tableType, base); + if (dict==usingExtDict) + { + if (match<(const BYTE*)source) + { + refDelta = dictDelta; + lowLimit = dictionary; + } + else + { + refDelta = 0; + lowLimit = (const BYTE*)source; + } + } + forwardH = LZ4_hashPosition(forwardIp, tableType); + LZ4_putPositionOnHash(ip, h, ctx, tableType, base); + + } while ( ((dictIssue==dictSmall) ? (match < lowRefLimit) : 0) + || ((tableType==byU16) ? 0 : (match + MAX_DISTANCE < ip)) + || (LZ4_read32(match+refDelta) != LZ4_read32(ip)) ); + } + + /* Catch up */ + while ((ip>anchor) && (match+refDelta > lowLimit) && (unlikely(ip[-1]==match[refDelta-1]))) { ip--; match--; } + + { + /* Encode Literal length */ + unsigned litLength = (unsigned)(ip - anchor); + token = op++; + if ((outputLimited) && (unlikely(op + litLength + (2 + 1 + LASTLITERALS) + (litLength/255) > olimit))) + return 0; /* Check output limit */ + if (litLength>=RUN_MASK) + { + int len = (int)litLength-RUN_MASK; + *token=(RUN_MASK<<ML_BITS); + for(; len >= 255 ; len-=255) *op++ = 255; + *op++ = (BYTE)len; + } + else *token = (BYTE)(litLength<<ML_BITS); + + /* Copy Literals */ + LZ4_wildCopy(op, anchor, op+litLength); + op+=litLength; + } + +_next_match: + /* Encode Offset */ + LZ4_writeLE16(op, (U16)(ip-match)); op+=2; + + /* Encode MatchLength */ + { + unsigned matchLength; + + if ((dict==usingExtDict) && (lowLimit==dictionary)) + { + const BYTE* limit; + match += refDelta; + limit = ip + (dictEnd-match); + if (limit > matchlimit) limit = matchlimit; + matchLength = LZ4_count(ip+MINMATCH, match+MINMATCH, limit); + ip += MINMATCH + matchLength; + if (ip==limit) + { + unsigned more = LZ4_count(ip, (const BYTE*)source, matchlimit); + matchLength += more; + ip += more; + } + } + else + { + matchLength = LZ4_count(ip+MINMATCH, match+MINMATCH, matchlimit); + ip += MINMATCH + matchLength; + } + + if ((outputLimited) && (unlikely(op + (1 + LASTLITERALS) + (matchLength>>8) > olimit))) + return 0; /* Check output limit */ + if (matchLength>=ML_MASK) + { + *token += ML_MASK; + matchLength -= ML_MASK; + for (; matchLength >= 510 ; matchLength-=510) { *op++ = 255; *op++ = 255; } + if (matchLength >= 255) { matchLength-=255; *op++ = 255; } + *op++ = (BYTE)matchLength; + } + else *token += (BYTE)(matchLength); + } + + anchor = ip; + + /* Test end of chunk */ + if (ip > mflimit) break; + + /* Fill table */ + LZ4_putPosition(ip-2, ctx, tableType, base); + + /* Test next position */ + match = LZ4_getPosition(ip, ctx, tableType, base); + if (dict==usingExtDict) + { + if (match<(const BYTE*)source) + { + refDelta = dictDelta; + lowLimit = dictionary; + } + else + { + refDelta = 0; + lowLimit = (const BYTE*)source; + } + } + LZ4_putPosition(ip, ctx, tableType, base); + if ( ((dictIssue==dictSmall) ? (match>=lowRefLimit) : 1) + && (match+MAX_DISTANCE>=ip) + && (LZ4_read32(match+refDelta)==LZ4_read32(ip)) ) + { token=op++; *token=0; goto _next_match; } + + /* Prepare next loop */ + forwardH = LZ4_hashPosition(++ip, tableType); + } + +_last_literals: + /* Encode Last Literals */ + { + const size_t lastRun = (size_t)(iend - anchor); + if ((outputLimited) && ((op - (BYTE*)dest) + lastRun + 1 + ((lastRun+255-RUN_MASK)/255) > (U32)maxOutputSize)) + return 0; /* Check output limit */ + if (lastRun >= RUN_MASK) + { + size_t accumulator = lastRun - RUN_MASK; + *op++ = RUN_MASK << ML_BITS; + for(; accumulator >= 255 ; accumulator-=255) *op++ = 255; + *op++ = (BYTE) accumulator; + } + else + { + *op++ = (BYTE)(lastRun<<ML_BITS); + } + memcpy(op, anchor, lastRun); + op += lastRun; + } + + /* End */ + return (int) (((char*)op)-dest); +} + + +int LZ4_compress_fast_extState(void* state, const char* source, char* dest, int inputSize, int maxOutputSize, int acceleration) +{ + LZ4_resetStream((LZ4_stream_t*)state); + if (acceleration < 1) acceleration = ACCELERATION_DEFAULT; + + if (maxOutputSize >= LZ4_compressBound(inputSize)) + { + if (inputSize < LZ4_64Klimit) + return LZ4_compress_generic(state, source, dest, inputSize, 0, notLimited, byU16, noDict, noDictIssue, acceleration); + else + return LZ4_compress_generic(state, source, dest, inputSize, 0, notLimited, LZ4_64bits() ? byU32 : byPtr, noDict, noDictIssue, acceleration); + } + else + { + if (inputSize < LZ4_64Klimit) + return LZ4_compress_generic(state, source, dest, inputSize, maxOutputSize, limitedOutput, byU16, noDict, noDictIssue, acceleration); + else + return LZ4_compress_generic(state, source, dest, inputSize, maxOutputSize, limitedOutput, LZ4_64bits() ? byU32 : byPtr, noDict, noDictIssue, acceleration); + } +} + + +int LZ4_compress_fast(const char* source, char* dest, int inputSize, int maxOutputSize, int acceleration) +{ +#if (HEAPMODE) + void* ctxPtr = ALLOCATOR(1, sizeof(LZ4_stream_t)); /* malloc-calloc always properly aligned */ +#else + LZ4_stream_t ctx; + void* ctxPtr = &ctx; +#endif + + int result = LZ4_compress_fast_extState(ctxPtr, source, dest, inputSize, maxOutputSize, acceleration); + +#if (HEAPMODE) + FREEMEM(ctxPtr); +#endif + return result; +} + + +int LZ4_compress_default(const char* source, char* dest, int inputSize, int maxOutputSize) +{ + return LZ4_compress_fast(source, dest, inputSize, maxOutputSize, 1); +} + + +/* hidden debug function */ +/* strangely enough, gcc generates faster code when this function is uncommented, even if unused */ +int LZ4_compress_fast_force(const char* source, char* dest, int inputSize, int maxOutputSize, int acceleration) +{ + LZ4_stream_t ctx; + + LZ4_resetStream(&ctx); + + if (inputSize < LZ4_64Klimit) + return LZ4_compress_generic(&ctx, source, dest, inputSize, maxOutputSize, limitedOutput, byU16, noDict, noDictIssue, acceleration); + else + return LZ4_compress_generic(&ctx, source, dest, inputSize, maxOutputSize, limitedOutput, LZ4_64bits() ? byU32 : byPtr, noDict, noDictIssue, acceleration); +} + + +/******************************** +* destSize variant +********************************/ + +static int LZ4_compress_destSize_generic( + void* const ctx, + const char* const src, + char* const dst, + int* const srcSizePtr, + const int targetDstSize, + const tableType_t tableType) +{ + const BYTE* ip = (const BYTE*) src; + const BYTE* base = (const BYTE*) src; + const BYTE* lowLimit = (const BYTE*) src; + const BYTE* anchor = ip; + const BYTE* const iend = ip + *srcSizePtr; + const BYTE* const mflimit = iend - MFLIMIT; + const BYTE* const matchlimit = iend - LASTLITERALS; + + BYTE* op = (BYTE*) dst; + BYTE* const oend = op + targetDstSize; + BYTE* const oMaxLit = op + targetDstSize - 2 /* offset */ - 8 /* because 8+MINMATCH==MFLIMIT */ - 1 /* token */; + BYTE* const oMaxMatch = op + targetDstSize - (LASTLITERALS + 1 /* token */); + BYTE* const oMaxSeq = oMaxLit - 1 /* token */; + + U32 forwardH; + + + /* Init conditions */ + if (targetDstSize < 1) return 0; /* Impossible to store anything */ + if ((U32)*srcSizePtr > (U32)LZ4_MAX_INPUT_SIZE) return 0; /* Unsupported input size, too large (or negative) */ + if ((tableType == byU16) && (*srcSizePtr>=LZ4_64Klimit)) return 0; /* Size too large (not within 64K limit) */ + if (*srcSizePtr<LZ4_minLength) goto _last_literals; /* Input too small, no compression (all literals) */ + + /* First Byte */ + *srcSizePtr = 0; + LZ4_putPosition(ip, ctx, tableType, base); + ip++; forwardH = LZ4_hashPosition(ip, tableType); + + /* Main Loop */ + for ( ; ; ) + { + const BYTE* match; + BYTE* token; + { + const BYTE* forwardIp = ip; + unsigned step = 1; + unsigned searchMatchNb = 1 << LZ4_skipTrigger; + + /* Find a match */ + do { + U32 h = forwardH; + ip = forwardIp; + forwardIp += step; + step = (searchMatchNb++ >> LZ4_skipTrigger); + + if (unlikely(forwardIp > mflimit)) + goto _last_literals; + + match = LZ4_getPositionOnHash(h, ctx, tableType, base); + forwardH = LZ4_hashPosition(forwardIp, tableType); + LZ4_putPositionOnHash(ip, h, ctx, tableType, base); + + } while ( ((tableType==byU16) ? 0 : (match + MAX_DISTANCE < ip)) + || (LZ4_read32(match) != LZ4_read32(ip)) ); + } + + /* Catch up */ + while ((ip>anchor) && (match > lowLimit) && (unlikely(ip[-1]==match[-1]))) { ip--; match--; } + + { + /* Encode Literal length */ + unsigned litLength = (unsigned)(ip - anchor); + token = op++; + if (op + ((litLength+240)/255) + litLength > oMaxLit) + { + /* Not enough space for a last match */ + op--; + goto _last_literals; + } + if (litLength>=RUN_MASK) + { + unsigned len = litLength - RUN_MASK; + *token=(RUN_MASK<<ML_BITS); + for(; len >= 255 ; len-=255) *op++ = 255; + *op++ = (BYTE)len; + } + else *token = (BYTE)(litLength<<ML_BITS); + + /* Copy Literals */ + LZ4_wildCopy(op, anchor, op+litLength); + op += litLength; + } + +_next_match: + /* Encode Offset */ + LZ4_writeLE16(op, (U16)(ip-match)); op+=2; + + /* Encode MatchLength */ + { + size_t matchLength; + + matchLength = LZ4_count(ip+MINMATCH, match+MINMATCH, matchlimit); + + if (op + ((matchLength+240)/255) > oMaxMatch) + { + /* Match description too long : reduce it */ + matchLength = (15-1) + (oMaxMatch-op) * 255; + } + //printf("offset %5i, matchLength%5i \n", (int)(ip-match), matchLength + MINMATCH); + ip += MINMATCH + matchLength; + + if (matchLength>=ML_MASK) + { + *token += ML_MASK; + matchLength -= ML_MASK; + while (matchLength >= 255) { matchLength-=255; *op++ = 255; } + *op++ = (BYTE)matchLength; + } + else *token += (BYTE)(matchLength); + } + + anchor = ip; + + /* Test end of block */ + if (ip > mflimit) break; + if (op > oMaxSeq) break; + + /* Fill table */ + LZ4_putPosition(ip-2, ctx, tableType, base); + + /* Test next position */ + match = LZ4_getPosition(ip, ctx, tableType, base); + LZ4_putPosition(ip, ctx, tableType, base); + if ( (match+MAX_DISTANCE>=ip) + && (LZ4_read32(match)==LZ4_read32(ip)) ) + { token=op++; *token=0; goto _next_match; } + + /* Prepare next loop */ + forwardH = LZ4_hashPosition(++ip, tableType); + } + +_last_literals: + /* Encode Last Literals */ + { + size_t lastRunSize = (size_t)(iend - anchor); + if (op + 1 /* token */ + ((lastRunSize+240)/255) /* litLength */ + lastRunSize /* literals */ > oend) + { + /* adapt lastRunSize to fill 'dst' */ + lastRunSize = (oend-op) - 1; + lastRunSize -= (lastRunSize+240)/255; + } + ip = anchor + lastRunSize; + + if (lastRunSize >= RUN_MASK) + { + size_t accumulator = lastRunSize - RUN_MASK; + *op++ = RUN_MASK << ML_BITS; + for(; accumulator >= 255 ; accumulator-=255) *op++ = 255; + *op++ = (BYTE) accumulator; + } + else + { + *op++ = (BYTE)(lastRunSize<<ML_BITS); + } + memcpy(op, anchor, lastRunSize); + op += lastRunSize; + } + + /* End */ + *srcSizePtr = (int) (((const char*)ip)-src); + return (int) (((char*)op)-dst); +} + + +static int LZ4_compress_destSize_extState (void* state, const char* src, char* dst, int* srcSizePtr, int targetDstSize) +{ + LZ4_resetStream((LZ4_stream_t*)state); + + if (targetDstSize >= LZ4_compressBound(*srcSizePtr)) /* compression success is guaranteed */ + { + return LZ4_compress_fast_extState(state, src, dst, *srcSizePtr, targetDstSize, 1); + } + else + { + if (*srcSizePtr < LZ4_64Klimit) + return LZ4_compress_destSize_generic(state, src, dst, srcSizePtr, targetDstSize, byU16); + else + return LZ4_compress_destSize_generic(state, src, dst, srcSizePtr, targetDstSize, LZ4_64bits() ? byU32 : byPtr); + } +} + + +int LZ4_compress_destSize(const char* src, char* dst, int* srcSizePtr, int targetDstSize) +{ +#if (HEAPMODE) + void* ctx = ALLOCATOR(1, sizeof(LZ4_stream_t)); /* malloc-calloc always properly aligned */ +#else + LZ4_stream_t ctxBody; + void* ctx = &ctxBody; +#endif + + int result = LZ4_compress_destSize_extState(ctx, src, dst, srcSizePtr, targetDstSize); + +#if (HEAPMODE) + FREEMEM(ctx); +#endif + return result; +} + + + +/******************************** +* Streaming functions +********************************/ + +LZ4_stream_t* LZ4_createStream(void) +{ + LZ4_stream_t* lz4s = (LZ4_stream_t*)ALLOCATOR(8, LZ4_STREAMSIZE_U64); + LZ4_STATIC_ASSERT(LZ4_STREAMSIZE >= sizeof(LZ4_stream_t_internal)); /* A compilation error here means LZ4_STREAMSIZE is not large enough */ + LZ4_resetStream(lz4s); + return lz4s; +} + +void LZ4_resetStream (LZ4_stream_t* LZ4_stream) +{ + MEM_INIT(LZ4_stream, 0, sizeof(LZ4_stream_t)); +} + +int LZ4_freeStream (LZ4_stream_t* LZ4_stream) +{ + FREEMEM(LZ4_stream); + return (0); +} + + +#define HASH_UNIT sizeof(size_t) +int LZ4_loadDict (LZ4_stream_t* LZ4_dict, const char* dictionary, int dictSize) +{ + LZ4_stream_t_internal* dict = (LZ4_stream_t_internal*) LZ4_dict; + const BYTE* p = (const BYTE*)dictionary; + const BYTE* const dictEnd = p + dictSize; + const BYTE* base; + + if ((dict->initCheck) || (dict->currentOffset > 1 GB)) /* Uninitialized structure, or reuse overflow */ + LZ4_resetStream(LZ4_dict); + + if (dictSize < (int)HASH_UNIT) + { + dict->dictionary = NULL; + dict->dictSize = 0; + return 0; + } + + if ((dictEnd - p) > 64 KB) p = dictEnd - 64 KB; + dict->currentOffset += 64 KB; + base = p - dict->currentOffset; + dict->dictionary = p; + dict->dictSize = (U32)(dictEnd - p); + dict->currentOffset += dict->dictSize; + + while (p <= dictEnd-HASH_UNIT) + { + LZ4_putPosition(p, dict->hashTable, byU32, base); + p+=3; + } + + return dict->dictSize; +} + + +static void LZ4_renormDictT(LZ4_stream_t_internal* LZ4_dict, const BYTE* src) +{ + if ((LZ4_dict->currentOffset > 0x80000000) || + ((size_t)LZ4_dict->currentOffset > (size_t)src)) /* address space overflow */ + { + /* rescale hash table */ + U32 delta = LZ4_dict->currentOffset - 64 KB; + const BYTE* dictEnd = LZ4_dict->dictionary + LZ4_dict->dictSize; + int i; + for (i=0; i<HASH_SIZE_U32; i++) + { + if (LZ4_dict->hashTable[i] < delta) LZ4_dict->hashTable[i]=0; + else LZ4_dict->hashTable[i] -= delta; + } + LZ4_dict->currentOffset = 64 KB; + if (LZ4_dict->dictSize > 64 KB) LZ4_dict->dictSize = 64 KB; + LZ4_dict->dictionary = dictEnd - LZ4_dict->dictSize; + } +} + + +int LZ4_compress_fast_continue (LZ4_stream_t* LZ4_stream, const char* source, char* dest, int inputSize, int maxOutputSize, int acceleration) +{ + LZ4_stream_t_internal* streamPtr = (LZ4_stream_t_internal*)LZ4_stream; + const BYTE* const dictEnd = streamPtr->dictionary + streamPtr->dictSize; + + const BYTE* smallest = (const BYTE*) source; + if (streamPtr->initCheck) return 0; /* Uninitialized structure detected */ + if ((streamPtr->dictSize>0) && (smallest>dictEnd)) smallest = dictEnd; + LZ4_renormDictT(streamPtr, smallest); + if (acceleration < 1) acceleration = ACCELERATION_DEFAULT; + + /* Check overlapping input/dictionary space */ + { + const BYTE* sourceEnd = (const BYTE*) source + inputSize; + if ((sourceEnd > streamPtr->dictionary) && (sourceEnd < dictEnd)) + { + streamPtr->dictSize = (U32)(dictEnd - sourceEnd); + if (streamPtr->dictSize > 64 KB) streamPtr->dictSize = 64 KB; + if (streamPtr->dictSize < 4) streamPtr->dictSize = 0; + streamPtr->dictionary = dictEnd - streamPtr->dictSize; + } + } + + /* prefix mode : source data follows dictionary */ + if (dictEnd == (const BYTE*)source) + { + int result; + if ((streamPtr->dictSize < 64 KB) && (streamPtr->dictSize < streamPtr->currentOffset)) + result = LZ4_compress_generic(LZ4_stream, source, dest, inputSize, maxOutputSize, limitedOutput, byU32, withPrefix64k, dictSmall, acceleration); + else + result = LZ4_compress_generic(LZ4_stream, source, dest, inputSize, maxOutputSize, limitedOutput, byU32, withPrefix64k, noDictIssue, acceleration); + streamPtr->dictSize += (U32)inputSize; + streamPtr->currentOffset += (U32)inputSize; + return result; + } + + /* external dictionary mode */ + { + int result; + if ((streamPtr->dictSize < 64 KB) && (streamPtr->dictSize < streamPtr->currentOffset)) + result = LZ4_compress_generic(LZ4_stream, source, dest, inputSize, maxOutputSize, limitedOutput, byU32, usingExtDict, dictSmall, acceleration); + else + result = LZ4_compress_generic(LZ4_stream, source, dest, inputSize, maxOutputSize, limitedOutput, byU32, usingExtDict, noDictIssue, acceleration); + streamPtr->dictionary = (const BYTE*)source; + streamPtr->dictSize = (U32)inputSize; + streamPtr->currentOffset += (U32)inputSize; + return result; + } +} + + +/* Hidden debug function, to force external dictionary mode */ +int LZ4_compress_forceExtDict (LZ4_stream_t* LZ4_dict, const char* source, char* dest, int inputSize) +{ + LZ4_stream_t_internal* streamPtr = (LZ4_stream_t_internal*)LZ4_dict; + int result; + const BYTE* const dictEnd = streamPtr->dictionary + streamPtr->dictSize; + + const BYTE* smallest = dictEnd; + if (smallest > (const BYTE*) source) smallest = (const BYTE*) source; + LZ4_renormDictT((LZ4_stream_t_internal*)LZ4_dict, smallest); + + result = LZ4_compress_generic(LZ4_dict, source, dest, inputSize, 0, notLimited, byU32, usingExtDict, noDictIssue, 1); + + streamPtr->dictionary = (const BYTE*)source; + streamPtr->dictSize = (U32)inputSize; + streamPtr->currentOffset += (U32)inputSize; + + return result; +} + + +int LZ4_saveDict (LZ4_stream_t* LZ4_dict, char* safeBuffer, int dictSize) +{ + LZ4_stream_t_internal* dict = (LZ4_stream_t_internal*) LZ4_dict; + const BYTE* previousDictEnd = dict->dictionary + dict->dictSize; + + if ((U32)dictSize > 64 KB) dictSize = 64 KB; /* useless to define a dictionary > 64 KB */ + if ((U32)dictSize > dict->dictSize) dictSize = dict->dictSize; + + memmove(safeBuffer, previousDictEnd - dictSize, dictSize); + + dict->dictionary = (const BYTE*)safeBuffer; + dict->dictSize = (U32)dictSize; + + return dictSize; +} + + + +/******************************* +* Decompression functions +*******************************/ +/* + * This generic decompression function cover all use cases. + * It shall be instantiated several times, using different sets of directives + * Note that it is essential this generic function is really inlined, + * in order to remove useless branches during compilation optimization. + */ +FORCE_INLINE int LZ4_decompress_generic( + const char* const source, + char* const dest, + int inputSize, + int outputSize, /* If endOnInput==endOnInputSize, this value is the max size of Output Buffer. */ + + int endOnInput, /* endOnOutputSize, endOnInputSize */ + int partialDecoding, /* full, partial */ + int targetOutputSize, /* only used if partialDecoding==partial */ + int dict, /* noDict, withPrefix64k, usingExtDict */ + const BYTE* const lowPrefix, /* == dest if dict == noDict */ + const BYTE* const dictStart, /* only if dict==usingExtDict */ + const size_t dictSize /* note : = 0 if noDict */ + ) +{ + /* Local Variables */ + const BYTE* ip = (const BYTE*) source; + const BYTE* const iend = ip + inputSize; + + BYTE* op = (BYTE*) dest; + BYTE* const oend = op + outputSize; + BYTE* cpy; + BYTE* oexit = op + targetOutputSize; + const BYTE* const lowLimit = lowPrefix - dictSize; + + const BYTE* const dictEnd = (const BYTE*)dictStart + dictSize; + const size_t dec32table[] = {4, 1, 2, 1, 4, 4, 4, 4}; + const size_t dec64table[] = {0, 0, 0, (size_t)-1, 0, 1, 2, 3}; + + const int safeDecode = (endOnInput==endOnInputSize); + const int checkOffset = ((safeDecode) && (dictSize < (int)(64 KB))); + + + /* Special cases */ + if ((partialDecoding) && (oexit> oend-MFLIMIT)) oexit = oend-MFLIMIT; /* targetOutputSize too high => decode everything */ + if ((endOnInput) && (unlikely(outputSize==0))) return ((inputSize==1) && (*ip==0)) ? 0 : -1; /* Empty output buffer */ + if ((!endOnInput) && (unlikely(outputSize==0))) return (*ip==0?1:-1); + + + /* Main Loop */ + while (1) + { + unsigned token; + size_t length; + const BYTE* match; + + /* get literal length */ + token = *ip++; + if ((length=(token>>ML_BITS)) == RUN_MASK) + { + unsigned s; + do + { + s = *ip++; + length += s; + } + while (likely((endOnInput)?ip<iend-RUN_MASK:1) && (s==255)); + if ((safeDecode) && unlikely((size_t)(op+length)<(size_t)(op))) goto _output_error; /* overflow detection */ + if ((safeDecode) && unlikely((size_t)(ip+length)<(size_t)(ip))) goto _output_error; /* overflow detection */ + } + + /* copy literals */ + cpy = op+length; + if (((endOnInput) && ((cpy>(partialDecoding?oexit:oend-MFLIMIT)) || (ip+length>iend-(2+1+LASTLITERALS))) ) + || ((!endOnInput) && (cpy>oend-COPYLENGTH))) + { + if (partialDecoding) + { + if (cpy > oend) goto _output_error; /* Error : write attempt beyond end of output buffer */ + if ((endOnInput) && (ip+length > iend)) goto _output_error; /* Error : read attempt beyond end of input buffer */ + } + else + { + if ((!endOnInput) && (cpy != oend)) goto _output_error; /* Error : block decoding must stop exactly there */ + if ((endOnInput) && ((ip+length != iend) || (cpy > oend))) goto _output_error; /* Error : input must be consumed */ + } + memcpy(op, ip, length); + ip += length; + op += length; + break; /* Necessarily EOF, due to parsing restrictions */ + } + LZ4_wildCopy(op, ip, cpy); + ip += length; op = cpy; + + /* get offset */ + match = cpy - LZ4_readLE16(ip); ip+=2; + if ((checkOffset) && (unlikely(match < lowLimit))) goto _output_error; /* Error : offset outside destination buffer */ + + /* get matchlength */ + length = token & ML_MASK; + if (length == ML_MASK) + { + unsigned s; + do + { + if ((endOnInput) && (ip > iend-LASTLITERALS)) goto _output_error; + s = *ip++; + length += s; + } while (s==255); + if ((safeDecode) && unlikely((size_t)(op+length)<(size_t)op)) goto _output_error; /* overflow detection */ + } + length += MINMATCH; + + /* check external dictionary */ + if ((dict==usingExtDict) && (match < lowPrefix)) + { + if (unlikely(op+length > oend-LASTLITERALS)) goto _output_error; /* doesn't respect parsing restriction */ + + if (length <= (size_t)(lowPrefix-match)) + { + /* match can be copied as a single segment from external dictionary */ + match = dictEnd - (lowPrefix-match); + memmove(op, match, length); op += length; + } + else + { + /* match encompass external dictionary and current segment */ + size_t copySize = (size_t)(lowPrefix-match); + memcpy(op, dictEnd - copySize, copySize); + op += copySize; + copySize = length - copySize; + if (copySize > (size_t)(op-lowPrefix)) /* overlap within current segment */ + { + BYTE* const endOfMatch = op + copySize; + const BYTE* copyFrom = lowPrefix; + while (op < endOfMatch) *op++ = *copyFrom++; + } + else + { + memcpy(op, lowPrefix, copySize); + op += copySize; + } + } + continue; + } + + /* copy repeated sequence */ + cpy = op + length; + if (unlikely((op-match)<8)) + { + const size_t dec64 = dec64table[op-match]; + op[0] = match[0]; + op[1] = match[1]; + op[2] = match[2]; + op[3] = match[3]; + match += dec32table[op-match]; + LZ4_copy4(op+4, match); + op += 8; match -= dec64; + } else { LZ4_copy8(op, match); op+=8; match+=8; } + + if (unlikely(cpy>oend-12)) + { + if (cpy > oend-LASTLITERALS) goto _output_error; /* Error : last LASTLITERALS bytes must be literals */ + if (op < oend-8) + { + LZ4_wildCopy(op, match, oend-8); + match += (oend-8) - op; + op = oend-8; + } + while (op<cpy) *op++ = *match++; + } + else + LZ4_wildCopy(op, match, cpy); + op=cpy; /* correction */ + } + + /* end of decoding */ + if (endOnInput) + return (int) (((char*)op)-dest); /* Nb of output bytes decoded */ + else + return (int) (((const char*)ip)-source); /* Nb of input bytes read */ + + /* Overflow error detected */ +_output_error: + return (int) (-(((const char*)ip)-source))-1; +} + + +int LZ4_decompress_safe(const char* source, char* dest, int compressedSize, int maxDecompressedSize) +{ + return LZ4_decompress_generic(source, dest, compressedSize, maxDecompressedSize, endOnInputSize, full, 0, noDict, (BYTE*)dest, NULL, 0); +} + +int LZ4_decompress_safe_partial(const char* source, char* dest, int compressedSize, int targetOutputSize, int maxDecompressedSize) +{ + return LZ4_decompress_generic(source, dest, compressedSize, maxDecompressedSize, endOnInputSize, partial, targetOutputSize, noDict, (BYTE*)dest, NULL, 0); +} + +int LZ4_decompress_fast(const char* source, char* dest, int originalSize) +{ + return LZ4_decompress_generic(source, dest, 0, originalSize, endOnOutputSize, full, 0, withPrefix64k, (BYTE*)(dest - 64 KB), NULL, 64 KB); +} + + +/* streaming decompression functions */ + +typedef struct +{ + const BYTE* externalDict; + size_t extDictSize; + const BYTE* prefixEnd; + size_t prefixSize; +} LZ4_streamDecode_t_internal; + +/* + * If you prefer dynamic allocation methods, + * LZ4_createStreamDecode() + * provides a pointer (void*) towards an initialized LZ4_streamDecode_t structure. + */ +LZ4_streamDecode_t* LZ4_createStreamDecode(void) +{ + LZ4_streamDecode_t* lz4s = (LZ4_streamDecode_t*) ALLOCATOR(1, sizeof(LZ4_streamDecode_t)); + return lz4s; +} + +int LZ4_freeStreamDecode (LZ4_streamDecode_t* LZ4_stream) +{ + FREEMEM(LZ4_stream); + return 0; +} + +/* + * LZ4_setStreamDecode + * Use this function to instruct where to find the dictionary + * This function is not necessary if previous data is still available where it was decoded. + * Loading a size of 0 is allowed (same effect as no dictionary). + * Return : 1 if OK, 0 if error + */ +int LZ4_setStreamDecode (LZ4_streamDecode_t* LZ4_streamDecode, const char* dictionary, int dictSize) +{ + LZ4_streamDecode_t_internal* lz4sd = (LZ4_streamDecode_t_internal*) LZ4_streamDecode; + lz4sd->prefixSize = (size_t) dictSize; + lz4sd->prefixEnd = (const BYTE*) dictionary + dictSize; + lz4sd->externalDict = NULL; + lz4sd->extDictSize = 0; + return 1; +} + +/* +*_continue() : + These decoding functions allow decompression of multiple blocks in "streaming" mode. + Previously decoded blocks must still be available at the memory position where they were decoded. + If it's not possible, save the relevant part of decoded data into a safe buffer, + and indicate where it stands using LZ4_setStreamDecode() +*/ +int LZ4_decompress_safe_continue (LZ4_streamDecode_t* LZ4_streamDecode, const char* source, char* dest, int compressedSize, int maxOutputSize) +{ + LZ4_streamDecode_t_internal* lz4sd = (LZ4_streamDecode_t_internal*) LZ4_streamDecode; + int result; + + if (lz4sd->prefixEnd == (BYTE*)dest) + { + result = LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize, + endOnInputSize, full, 0, + usingExtDict, lz4sd->prefixEnd - lz4sd->prefixSize, lz4sd->externalDict, lz4sd->extDictSize); + if (result <= 0) return result; + lz4sd->prefixSize += result; + lz4sd->prefixEnd += result; + } + else + { + lz4sd->extDictSize = lz4sd->prefixSize; + lz4sd->externalDict = lz4sd->prefixEnd - lz4sd->extDictSize; + result = LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize, + endOnInputSize, full, 0, + usingExtDict, (BYTE*)dest, lz4sd->externalDict, lz4sd->extDictSize); + if (result <= 0) return result; + lz4sd->prefixSize = result; + lz4sd->prefixEnd = (BYTE*)dest + result; + } + + return result; +} + +int LZ4_decompress_fast_continue (LZ4_streamDecode_t* LZ4_streamDecode, const char* source, char* dest, int originalSize) +{ + LZ4_streamDecode_t_internal* lz4sd = (LZ4_streamDecode_t_internal*) LZ4_streamDecode; + int result; + + if (lz4sd->prefixEnd == (BYTE*)dest) + { + result = LZ4_decompress_generic(source, dest, 0, originalSize, + endOnOutputSize, full, 0, + usingExtDict, lz4sd->prefixEnd - lz4sd->prefixSize, lz4sd->externalDict, lz4sd->extDictSize); + if (result <= 0) return result; + lz4sd->prefixSize += originalSize; + lz4sd->prefixEnd += originalSize; + } + else + { + lz4sd->extDictSize = lz4sd->prefixSize; + lz4sd->externalDict = (BYTE*)dest - lz4sd->extDictSize; + result = LZ4_decompress_generic(source, dest, 0, originalSize, + endOnOutputSize, full, 0, + usingExtDict, (BYTE*)dest, lz4sd->externalDict, lz4sd->extDictSize); + if (result <= 0) return result; + lz4sd->prefixSize = originalSize; + lz4sd->prefixEnd = (BYTE*)dest + originalSize; + } + + return result; +} + + +/* +Advanced decoding functions : +*_usingDict() : + These decoding functions work the same as "_continue" ones, + the dictionary must be explicitly provided within parameters +*/ + +FORCE_INLINE int LZ4_decompress_usingDict_generic(const char* source, char* dest, int compressedSize, int maxOutputSize, int safe, const char* dictStart, int dictSize) +{ + if (dictSize==0) + return LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize, safe, full, 0, noDict, (BYTE*)dest, NULL, 0); + if (dictStart+dictSize == dest) + { + if (dictSize >= (int)(64 KB - 1)) + return LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize, safe, full, 0, withPrefix64k, (BYTE*)dest-64 KB, NULL, 0); + return LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize, safe, full, 0, noDict, (BYTE*)dest-dictSize, NULL, 0); + } + return LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize, safe, full, 0, usingExtDict, (BYTE*)dest, (const BYTE*)dictStart, dictSize); +} + +int LZ4_decompress_safe_usingDict(const char* source, char* dest, int compressedSize, int maxOutputSize, const char* dictStart, int dictSize) +{ + return LZ4_decompress_usingDict_generic(source, dest, compressedSize, maxOutputSize, 1, dictStart, dictSize); +} + +int LZ4_decompress_fast_usingDict(const char* source, char* dest, int originalSize, const char* dictStart, int dictSize) +{ + return LZ4_decompress_usingDict_generic(source, dest, 0, originalSize, 0, dictStart, dictSize); +} + +/* debug function */ +int LZ4_decompress_safe_forceExtDict(const char* source, char* dest, int compressedSize, int maxOutputSize, const char* dictStart, int dictSize) +{ + return LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize, endOnInputSize, full, 0, usingExtDict, (BYTE*)dest, (const BYTE*)dictStart, dictSize); +} + + +/*************************************************** +* Obsolete Functions +***************************************************/ +/* obsolete compression functions */ +int LZ4_compress_limitedOutput(const char* source, char* dest, int inputSize, int maxOutputSize) { return LZ4_compress_default(source, dest, inputSize, maxOutputSize); } +int LZ4_compress(const char* source, char* dest, int inputSize) { return LZ4_compress_default(source, dest, inputSize, LZ4_compressBound(inputSize)); } +int LZ4_compress_limitedOutput_withState (void* state, const char* src, char* dst, int srcSize, int dstSize) { return LZ4_compress_fast_extState(state, src, dst, srcSize, dstSize, 1); } +int LZ4_compress_withState (void* state, const char* src, char* dst, int srcSize) { return LZ4_compress_fast_extState(state, src, dst, srcSize, LZ4_compressBound(srcSize), 1); } +int LZ4_compress_limitedOutput_continue (LZ4_stream_t* LZ4_stream, const char* src, char* dst, int srcSize, int maxDstSize) { return LZ4_compress_fast_continue(LZ4_stream, src, dst, srcSize, maxDstSize, 1); } +int LZ4_compress_continue (LZ4_stream_t* LZ4_stream, const char* source, char* dest, int inputSize) { return LZ4_compress_fast_continue(LZ4_stream, source, dest, inputSize, LZ4_compressBound(inputSize), 1); } + +/* +These function names are deprecated and should no longer be used. +They are only provided here for compatibility with older user programs. +- LZ4_uncompress is totally equivalent to LZ4_decompress_fast +- LZ4_uncompress_unknownOutputSize is totally equivalent to LZ4_decompress_safe +*/ +int LZ4_uncompress (const char* source, char* dest, int outputSize) { return LZ4_decompress_fast(source, dest, outputSize); } +int LZ4_uncompress_unknownOutputSize (const char* source, char* dest, int isize, int maxOutputSize) { return LZ4_decompress_safe(source, dest, isize, maxOutputSize); } + + +/* Obsolete Streaming functions */ + +int LZ4_sizeofStreamState() { return LZ4_STREAMSIZE; } + +static void LZ4_init(LZ4_stream_t_internal* lz4ds, BYTE* base) +{ + MEM_INIT(lz4ds, 0, LZ4_STREAMSIZE); + lz4ds->bufferStart = base; +} + +int LZ4_resetStreamState(void* state, char* inputBuffer) +{ + if ((((size_t)state) & 3) != 0) return 1; /* Error : pointer is not aligned on 4-bytes boundary */ + LZ4_init((LZ4_stream_t_internal*)state, (BYTE*)inputBuffer); + return 0; +} + +void* LZ4_create (char* inputBuffer) +{ + void* lz4ds = ALLOCATOR(8, LZ4_STREAMSIZE_U64); + LZ4_init ((LZ4_stream_t_internal*)lz4ds, (BYTE*)inputBuffer); + return lz4ds; +} + +char* LZ4_slideInputBuffer (void* LZ4_Data) +{ + LZ4_stream_t_internal* ctx = (LZ4_stream_t_internal*)LZ4_Data; + int dictSize = LZ4_saveDict((LZ4_stream_t*)LZ4_Data, (char*)ctx->bufferStart, 64 KB); + return (char*)(ctx->bufferStart + dictSize); +} + +/* Obsolete streaming decompression functions */ + +int LZ4_decompress_safe_withPrefix64k(const char* source, char* dest, int compressedSize, int maxOutputSize) +{ + return LZ4_decompress_generic(source, dest, compressedSize, maxOutputSize, endOnInputSize, full, 0, withPrefix64k, (BYTE*)dest - 64 KB, NULL, 64 KB); +} + +int LZ4_decompress_fast_withPrefix64k(const char* source, char* dest, int originalSize) +{ + return LZ4_decompress_generic(source, dest, 0, originalSize, endOnOutputSize, full, 0, withPrefix64k, (BYTE*)dest - 64 KB, NULL, 64 KB); +} + +#endif /* LZ4_COMMONDEFS_ONLY */ + diff --git a/thirdparty/etcpak/lz4/lz4.h b/thirdparty/etcpak/lz4/lz4.h new file mode 100644 index 0000000000..3e74002256 --- /dev/null +++ b/thirdparty/etcpak/lz4/lz4.h @@ -0,0 +1,360 @@ +/* + LZ4 - Fast LZ compression algorithm + Header File + Copyright (C) 2011-2015, Yann Collet. + + BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php) + + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions are + met: + + * Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above + copyright notice, this list of conditions and the following disclaimer + in the documentation and/or other materials provided with the + distribution. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + You can contact the author at : + - LZ4 source repository : https://github.com/Cyan4973/lz4 + - LZ4 public forum : https://groups.google.com/forum/#!forum/lz4c +*/ +#pragma once + +#if defined (__cplusplus) +extern "C" { +#endif + +/* + * lz4.h provides block compression functions, and gives full buffer control to programmer. + * If you need to generate inter-operable compressed data (respecting LZ4 frame specification), + * and can let the library handle its own memory, please use lz4frame.h instead. +*/ + +/************************************** +* Version +**************************************/ +#define LZ4_VERSION_MAJOR 1 /* for breaking interface changes */ +#define LZ4_VERSION_MINOR 7 /* for new (non-breaking) interface capabilities */ +#define LZ4_VERSION_RELEASE 1 /* for tweaks, bug-fixes, or development */ +#define LZ4_VERSION_NUMBER (LZ4_VERSION_MAJOR *100*100 + LZ4_VERSION_MINOR *100 + LZ4_VERSION_RELEASE) +int LZ4_versionNumber (void); + +/************************************** +* Tuning parameter +**************************************/ +/* + * LZ4_MEMORY_USAGE : + * Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.) + * Increasing memory usage improves compression ratio + * Reduced memory usage can improve speed, due to cache effect + * Default value is 14, for 16KB, which nicely fits into Intel x86 L1 cache + */ +#define LZ4_MEMORY_USAGE 14 + + +/************************************** +* Simple Functions +**************************************/ + +int LZ4_compress_default(const char* source, char* dest, int sourceSize, int maxDestSize); +int LZ4_decompress_safe (const char* source, char* dest, int compressedSize, int maxDecompressedSize); + +/* +LZ4_compress_default() : + Compresses 'sourceSize' bytes from buffer 'source' + into already allocated 'dest' buffer of size 'maxDestSize'. + Compression is guaranteed to succeed if 'maxDestSize' >= LZ4_compressBound(sourceSize). + It also runs faster, so it's a recommended setting. + If the function cannot compress 'source' into a more limited 'dest' budget, + compression stops *immediately*, and the function result is zero. + As a consequence, 'dest' content is not valid. + This function never writes outside 'dest' buffer, nor read outside 'source' buffer. + sourceSize : Max supported value is LZ4_MAX_INPUT_VALUE + maxDestSize : full or partial size of buffer 'dest' (which must be already allocated) + return : the number of bytes written into buffer 'dest' (necessarily <= maxOutputSize) + or 0 if compression fails + +LZ4_decompress_safe() : + compressedSize : is the precise full size of the compressed block. + maxDecompressedSize : is the size of destination buffer, which must be already allocated. + return : the number of bytes decompressed into destination buffer (necessarily <= maxDecompressedSize) + If destination buffer is not large enough, decoding will stop and output an error code (<0). + If the source stream is detected malformed, the function will stop decoding and return a negative result. + This function is protected against buffer overflow exploits, including malicious data packets. + It never writes outside output buffer, nor reads outside input buffer. +*/ + + +/************************************** +* Advanced Functions +**************************************/ +#define LZ4_MAX_INPUT_SIZE 0x7E000000 /* 2 113 929 216 bytes */ +#define LZ4_COMPRESSBOUND(isize) ((unsigned)(isize) > (unsigned)LZ4_MAX_INPUT_SIZE ? 0 : (isize) + ((isize)/255) + 16) + +/* +LZ4_compressBound() : + Provides the maximum size that LZ4 compression may output in a "worst case" scenario (input data not compressible) + This function is primarily useful for memory allocation purposes (destination buffer size). + Macro LZ4_COMPRESSBOUND() is also provided for compilation-time evaluation (stack memory allocation for example). + Note that LZ4_compress_default() compress faster when dest buffer size is >= LZ4_compressBound(srcSize) + inputSize : max supported value is LZ4_MAX_INPUT_SIZE + return : maximum output size in a "worst case" scenario + or 0, if input size is too large ( > LZ4_MAX_INPUT_SIZE) +*/ +int LZ4_compressBound(int inputSize); + +/* +LZ4_compress_fast() : + Same as LZ4_compress_default(), but allows to select an "acceleration" factor. + The larger the acceleration value, the faster the algorithm, but also the lesser the compression. + It's a trade-off. It can be fine tuned, with each successive value providing roughly +~3% to speed. + An acceleration value of "1" is the same as regular LZ4_compress_default() + Values <= 0 will be replaced by ACCELERATION_DEFAULT (see lz4.c), which is 1. +*/ +int LZ4_compress_fast (const char* source, char* dest, int sourceSize, int maxDestSize, int acceleration); + + +/* +LZ4_compress_fast_extState() : + Same compression function, just using an externally allocated memory space to store compression state. + Use LZ4_sizeofState() to know how much memory must be allocated, + and allocate it on 8-bytes boundaries (using malloc() typically). + Then, provide it as 'void* state' to compression function. +*/ +int LZ4_sizeofState(void); +int LZ4_compress_fast_extState (void* state, const char* source, char* dest, int inputSize, int maxDestSize, int acceleration); + + +/* +LZ4_compress_destSize() : + Reverse the logic, by compressing as much data as possible from 'source' buffer + into already allocated buffer 'dest' of size 'targetDestSize'. + This function either compresses the entire 'source' content into 'dest' if it's large enough, + or fill 'dest' buffer completely with as much data as possible from 'source'. + *sourceSizePtr : will be modified to indicate how many bytes where read from 'source' to fill 'dest'. + New value is necessarily <= old value. + return : Nb bytes written into 'dest' (necessarily <= targetDestSize) + or 0 if compression fails +*/ +int LZ4_compress_destSize (const char* source, char* dest, int* sourceSizePtr, int targetDestSize); + + +/* +LZ4_decompress_fast() : + originalSize : is the original and therefore uncompressed size + return : the number of bytes read from the source buffer (in other words, the compressed size) + If the source stream is detected malformed, the function will stop decoding and return a negative result. + Destination buffer must be already allocated. Its size must be a minimum of 'originalSize' bytes. + note : This function fully respect memory boundaries for properly formed compressed data. + It is a bit faster than LZ4_decompress_safe(). + However, it does not provide any protection against intentionally modified data stream (malicious input). + Use this function in trusted environment only (data to decode comes from a trusted source). +*/ +int LZ4_decompress_fast (const char* source, char* dest, int originalSize); + +/* +LZ4_decompress_safe_partial() : + This function decompress a compressed block of size 'compressedSize' at position 'source' + into destination buffer 'dest' of size 'maxDecompressedSize'. + The function tries to stop decompressing operation as soon as 'targetOutputSize' has been reached, + reducing decompression time. + return : the number of bytes decoded in the destination buffer (necessarily <= maxDecompressedSize) + Note : this number can be < 'targetOutputSize' should the compressed block to decode be smaller. + Always control how many bytes were decoded. + If the source stream is detected malformed, the function will stop decoding and return a negative result. + This function never writes outside of output buffer, and never reads outside of input buffer. It is therefore protected against malicious data packets +*/ +int LZ4_decompress_safe_partial (const char* source, char* dest, int compressedSize, int targetOutputSize, int maxDecompressedSize); + + +/*********************************************** +* Streaming Compression Functions +***********************************************/ +#define LZ4_STREAMSIZE_U64 ((1 << (LZ4_MEMORY_USAGE-3)) + 4) +#define LZ4_STREAMSIZE (LZ4_STREAMSIZE_U64 * sizeof(long long)) +/* + * LZ4_stream_t + * information structure to track an LZ4 stream. + * important : init this structure content before first use ! + * note : only allocated directly the structure if you are statically linking LZ4 + * If you are using liblz4 as a DLL, please use below construction methods instead. + */ +typedef struct { long long table[LZ4_STREAMSIZE_U64]; } LZ4_stream_t; + +/* + * LZ4_resetStream + * Use this function to init an allocated LZ4_stream_t structure + */ +void LZ4_resetStream (LZ4_stream_t* streamPtr); + +/* + * LZ4_createStream will allocate and initialize an LZ4_stream_t structure + * LZ4_freeStream releases its memory. + * In the context of a DLL (liblz4), please use these methods rather than the static struct. + * They are more future proof, in case of a change of LZ4_stream_t size. + */ +LZ4_stream_t* LZ4_createStream(void); +int LZ4_freeStream (LZ4_stream_t* streamPtr); + +/* + * LZ4_loadDict + * Use this function to load a static dictionary into LZ4_stream. + * Any previous data will be forgotten, only 'dictionary' will remain in memory. + * Loading a size of 0 is allowed. + * Return : dictionary size, in bytes (necessarily <= 64 KB) + */ +int LZ4_loadDict (LZ4_stream_t* streamPtr, const char* dictionary, int dictSize); + +/* + * LZ4_compress_fast_continue + * Compress buffer content 'src', using data from previously compressed blocks as dictionary to improve compression ratio. + * Important : Previous data blocks are assumed to still be present and unmodified ! + * 'dst' buffer must be already allocated. + * If maxDstSize >= LZ4_compressBound(srcSize), compression is guaranteed to succeed, and runs faster. + * If not, and if compressed data cannot fit into 'dst' buffer size, compression stops, and function returns a zero. + */ +int LZ4_compress_fast_continue (LZ4_stream_t* streamPtr, const char* src, char* dst, int srcSize, int maxDstSize, int acceleration); + +/* + * LZ4_saveDict + * If previously compressed data block is not guaranteed to remain available at its memory location + * save it into a safer place (char* safeBuffer) + * Note : you don't need to call LZ4_loadDict() afterwards, + * dictionary is immediately usable, you can therefore call LZ4_compress_fast_continue() + * Return : saved dictionary size in bytes (necessarily <= dictSize), or 0 if error + */ +int LZ4_saveDict (LZ4_stream_t* streamPtr, char* safeBuffer, int dictSize); + + +/************************************************ +* Streaming Decompression Functions +************************************************/ + +#define LZ4_STREAMDECODESIZE_U64 4 +#define LZ4_STREAMDECODESIZE (LZ4_STREAMDECODESIZE_U64 * sizeof(unsigned long long)) +typedef struct { unsigned long long table[LZ4_STREAMDECODESIZE_U64]; } LZ4_streamDecode_t; +/* + * LZ4_streamDecode_t + * information structure to track an LZ4 stream. + * init this structure content using LZ4_setStreamDecode or memset() before first use ! + * + * In the context of a DLL (liblz4) please prefer usage of construction methods below. + * They are more future proof, in case of a change of LZ4_streamDecode_t size in the future. + * LZ4_createStreamDecode will allocate and initialize an LZ4_streamDecode_t structure + * LZ4_freeStreamDecode releases its memory. + */ +LZ4_streamDecode_t* LZ4_createStreamDecode(void); +int LZ4_freeStreamDecode (LZ4_streamDecode_t* LZ4_stream); + +/* + * LZ4_setStreamDecode + * Use this function to instruct where to find the dictionary. + * Setting a size of 0 is allowed (same effect as reset). + * Return : 1 if OK, 0 if error + */ +int LZ4_setStreamDecode (LZ4_streamDecode_t* LZ4_streamDecode, const char* dictionary, int dictSize); + +/* +*_continue() : + These decoding functions allow decompression of multiple blocks in "streaming" mode. + Previously decoded blocks *must* remain available at the memory position where they were decoded (up to 64 KB) + In the case of a ring buffers, decoding buffer must be either : + - Exactly same size as encoding buffer, with same update rule (block boundaries at same positions) + In which case, the decoding & encoding ring buffer can have any size, including very small ones ( < 64 KB). + - Larger than encoding buffer, by a minimum of maxBlockSize more bytes. + maxBlockSize is implementation dependent. It's the maximum size you intend to compress into a single block. + In which case, encoding and decoding buffers do not need to be synchronized, + and encoding ring buffer can have any size, including small ones ( < 64 KB). + - _At least_ 64 KB + 8 bytes + maxBlockSize. + In which case, encoding and decoding buffers do not need to be synchronized, + and encoding ring buffer can have any size, including larger than decoding buffer. + Whenever these conditions are not possible, save the last 64KB of decoded data into a safe buffer, + and indicate where it is saved using LZ4_setStreamDecode() +*/ +int LZ4_decompress_safe_continue (LZ4_streamDecode_t* LZ4_streamDecode, const char* source, char* dest, int compressedSize, int maxDecompressedSize); +int LZ4_decompress_fast_continue (LZ4_streamDecode_t* LZ4_streamDecode, const char* source, char* dest, int originalSize); + + +/* +Advanced decoding functions : +*_usingDict() : + These decoding functions work the same as + a combination of LZ4_setStreamDecode() followed by LZ4_decompress_x_continue() + They are stand-alone. They don't need nor update an LZ4_streamDecode_t structure. +*/ +int LZ4_decompress_safe_usingDict (const char* source, char* dest, int compressedSize, int maxDecompressedSize, const char* dictStart, int dictSize); +int LZ4_decompress_fast_usingDict (const char* source, char* dest, int originalSize, const char* dictStart, int dictSize); + + + +/************************************** +* Obsolete Functions +**************************************/ +/* Deprecate Warnings */ +/* Should these warnings messages be a problem, + it is generally possible to disable them, + with -Wno-deprecated-declarations for gcc + or _CRT_SECURE_NO_WARNINGS in Visual for example. + You can also define LZ4_DEPRECATE_WARNING_DEFBLOCK. */ +#ifndef LZ4_DEPRECATE_WARNING_DEFBLOCK +# define LZ4_DEPRECATE_WARNING_DEFBLOCK +# define LZ4_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__) +# if (LZ4_GCC_VERSION >= 405) || defined(__clang__) +# define LZ4_DEPRECATED(message) __attribute__((deprecated(message))) +# elif (LZ4_GCC_VERSION >= 301) +# define LZ4_DEPRECATED(message) __attribute__((deprecated)) +# elif defined(_MSC_VER) +# define LZ4_DEPRECATED(message) __declspec(deprecated(message)) +# else +# pragma message("WARNING: You need to implement LZ4_DEPRECATED for this compiler") +# define LZ4_DEPRECATED(message) +# endif +#endif /* LZ4_DEPRECATE_WARNING_DEFBLOCK */ + +/* Obsolete compression functions */ +/* These functions are planned to start generate warnings by r131 approximately */ +int LZ4_compress (const char* source, char* dest, int sourceSize); +int LZ4_compress_limitedOutput (const char* source, char* dest, int sourceSize, int maxOutputSize); +int LZ4_compress_withState (void* state, const char* source, char* dest, int inputSize); +int LZ4_compress_limitedOutput_withState (void* state, const char* source, char* dest, int inputSize, int maxOutputSize); +int LZ4_compress_continue (LZ4_stream_t* LZ4_streamPtr, const char* source, char* dest, int inputSize); +int LZ4_compress_limitedOutput_continue (LZ4_stream_t* LZ4_streamPtr, const char* source, char* dest, int inputSize, int maxOutputSize); + +/* Obsolete decompression functions */ +/* These function names are completely deprecated and must no longer be used. + They are only provided here for compatibility with older programs. + - LZ4_uncompress is the same as LZ4_decompress_fast + - LZ4_uncompress_unknownOutputSize is the same as LZ4_decompress_safe + These function prototypes are now disabled; uncomment them only if you really need them. + It is highly recommended to stop using these prototypes and migrate to maintained ones */ +/* int LZ4_uncompress (const char* source, char* dest, int outputSize); */ +/* int LZ4_uncompress_unknownOutputSize (const char* source, char* dest, int isize, int maxOutputSize); */ + +/* Obsolete streaming functions; use new streaming interface whenever possible */ +LZ4_DEPRECATED("use LZ4_createStream() instead") void* LZ4_create (char* inputBuffer); +LZ4_DEPRECATED("use LZ4_createStream() instead") int LZ4_sizeofStreamState(void); +LZ4_DEPRECATED("use LZ4_resetStream() instead") int LZ4_resetStreamState(void* state, char* inputBuffer); +LZ4_DEPRECATED("use LZ4_saveDict() instead") char* LZ4_slideInputBuffer (void* state); + +/* Obsolete streaming decoding functions */ +LZ4_DEPRECATED("use LZ4_decompress_safe_usingDict() instead") int LZ4_decompress_safe_withPrefix64k (const char* src, char* dst, int compressedSize, int maxDstSize); +LZ4_DEPRECATED("use LZ4_decompress_fast_usingDict() instead") int LZ4_decompress_fast_withPrefix64k (const char* src, char* dst, int originalSize); + + +#if defined (__cplusplus) +} +#endif diff --git a/thirdparty/etcpak/mmap.cpp b/thirdparty/etcpak/mmap.cpp new file mode 100644 index 0000000000..c2460ee9e4 --- /dev/null +++ b/thirdparty/etcpak/mmap.cpp @@ -0,0 +1,38 @@ +#include "mmap.hpp" + +#ifdef _WIN32 +# include <io.h> +# include <windows.h> + +void* mmap( void* addr, size_t length, int prot, int flags, int fd, off_t offset ) +{ + HANDLE hnd; + void* map = nullptr; + + switch( prot ) + { + case PROT_READ: + if( hnd = CreateFileMapping( HANDLE( _get_osfhandle( fd ) ), nullptr, PAGE_READONLY, 0, DWORD( length ), nullptr ) ) + { + map = MapViewOfFile( hnd, FILE_MAP_READ, 0, 0, length ); + CloseHandle( hnd ); + } + break; + case PROT_WRITE: + if( hnd = CreateFileMapping( HANDLE( _get_osfhandle( fd ) ), nullptr, PAGE_READWRITE, 0, DWORD( length ), nullptr ) ) + { + map = MapViewOfFile( hnd, FILE_MAP_WRITE, 0, 0, length ); + CloseHandle( hnd ); + } + break; + } + + return map ? (char*)map + offset : (void*)-1; +} + +int munmap( void* addr, size_t length ) +{ + return UnmapViewOfFile( addr ) != 0 ? 0 : -1; +} + +#endif diff --git a/thirdparty/etcpak/mmap.hpp b/thirdparty/etcpak/mmap.hpp new file mode 100644 index 0000000000..e4cfe7759c --- /dev/null +++ b/thirdparty/etcpak/mmap.hpp @@ -0,0 +1,19 @@ +#ifndef __MMAP_HPP__ +#define __MMAP_HPP__ + +#ifndef _WIN32 +# include <sys/mman.h> +#else +# include <string.h> +# include <sys/types.h> + +# define PROT_READ 1 +# define PROT_WRITE 2 +# define MAP_SHARED 0 + +void* mmap( void* addr, size_t length, int prot, int flags, int fd, off_t offset ); +int munmap( void* addr, size_t length ); + +#endif + +#endif diff --git a/thirdparty/etcpak/patches/libpng-unbundle.patch b/thirdparty/etcpak/patches/libpng-unbundle.patch new file mode 100644 index 0000000000..e3c07412c6 --- /dev/null +++ b/thirdparty/etcpak/patches/libpng-unbundle.patch @@ -0,0 +1,13 @@ +diff --git a/thirdparty/etcpak/Bitmap.cpp b/thirdparty/etcpak/Bitmap.cpp +index 6aa36f5caa..ef318318ac 100644 +--- a/thirdparty/etcpak/Bitmap.cpp ++++ b/thirdparty/etcpak/Bitmap.cpp +@@ -3,7 +3,7 @@ + #include <string.h> + #include <assert.h> + +-#include "libpng/png.h" ++#include <png.h> + #include "lz4/lz4.h" + + #include "Bitmap.hpp" diff --git a/thirdparty/etcpak/patches/llvm-c++11-narrowing-errors.patch b/thirdparty/etcpak/patches/llvm-c++11-narrowing-errors.patch new file mode 100644 index 0000000000..ab0d1e63a2 --- /dev/null +++ b/thirdparty/etcpak/patches/llvm-c++11-narrowing-errors.patch @@ -0,0 +1,64 @@ +diff --git a/thirdparty/etcpak/BlockData.cpp b/thirdparty/etcpak/BlockData.cpp +index bd738085f3..395b55246b 100644 +--- a/thirdparty/etcpak/BlockData.cpp ++++ b/thirdparty/etcpak/BlockData.cpp +@@ -334,10 +334,10 @@ static etcpak_force_inline void DecodeT( uint64_t block, uint32_t* dst, uint32_t + const auto c3b = clampu8( cb1 - table59T58H[codeword] ); + + const uint32_t col_tab[4] = { +- cr0 | ( cg0 << 8 ) | ( cb0 << 16 ) | 0xFF000000, +- c2r | ( c2g << 8 ) | ( c2b << 16 ) | 0xFF000000, +- cr1 | ( cg1 << 8 ) | ( cb1 << 16 ) | 0xFF000000, +- c3r | ( c3g << 8 ) | ( c3b << 16 ) | 0xFF000000 ++ uint32_t(cr0 | ( cg0 << 8 ) | ( cb0 << 16 ) | 0xFF000000), ++ uint32_t(c2r | ( c2g << 8 ) | ( c2b << 16 ) | 0xFF000000), ++ uint32_t(cr1 | ( cg1 << 8 ) | ( cb1 << 16 ) | 0xFF000000), ++ uint32_t(c3r | ( c3g << 8 ) | ( c3b << 16 ) | 0xFF000000) + }; + + const uint32_t indexes = ( block >> 32 ) & 0xFFFFFFFF; +@@ -389,10 +389,10 @@ static etcpak_force_inline void DecodeTAlpha( uint64_t block, uint64_t alpha, ui + const auto c3b = clampu8( cb1 - table59T58H[codeword] ); + + const uint32_t col_tab[4] = { +- cr0 | ( cg0 << 8 ) | ( cb0 << 16 ), +- c2r | ( c2g << 8 ) | ( c2b << 16 ), +- cr1 | ( cg1 << 8 ) | ( cb1 << 16 ), +- c3r | ( c3g << 8 ) | ( c3b << 16 ) ++ uint32_t(cr0 | ( cg0 << 8 ) | ( cb0 << 16 )), ++ uint32_t(c2r | ( c2g << 8 ) | ( c2b << 16 )), ++ uint32_t(cr1 | ( cg1 << 8 ) | ( cb1 << 16 )), ++ uint32_t(c3r | ( c3g << 8 ) | ( c3b << 16 )) + }; + + const uint32_t indexes = ( block >> 32 ) & 0xFFFFFFFF; +@@ -436,10 +436,10 @@ static etcpak_force_inline void DecodeH( uint64_t block, uint32_t* dst, uint32_t + const auto codeword = codeword_hi | codeword_lo; + + const uint32_t col_tab[] = { +- clampu8( r0 + table59T58H[codeword] ) | ( clampu8( g0 + table59T58H[codeword] ) << 8 ) | ( clampu8( b0 + table59T58H[codeword] ) << 16 ), +- clampu8( r0 - table59T58H[codeword] ) | ( clampu8( g0 - table59T58H[codeword] ) << 8 ) | ( clampu8( b0 - table59T58H[codeword] ) << 16 ), +- clampu8( r1 + table59T58H[codeword] ) | ( clampu8( g1 + table59T58H[codeword] ) << 8 ) | ( clampu8( b1 + table59T58H[codeword] ) << 16 ), +- clampu8( r1 - table59T58H[codeword] ) | ( clampu8( g1 - table59T58H[codeword] ) << 8 ) | ( clampu8( b1 - table59T58H[codeword] ) << 16 ) ++ uint32_t(clampu8( r0 + table59T58H[codeword] ) | ( clampu8( g0 + table59T58H[codeword] ) << 8 ) | ( clampu8( b0 + table59T58H[codeword] ) << 16 )), ++ uint32_t(clampu8( r0 - table59T58H[codeword] ) | ( clampu8( g0 - table59T58H[codeword] ) << 8 ) | ( clampu8( b0 - table59T58H[codeword] ) << 16 )), ++ uint32_t(clampu8( r1 + table59T58H[codeword] ) | ( clampu8( g1 + table59T58H[codeword] ) << 8 ) | ( clampu8( b1 + table59T58H[codeword] ) << 16 )), ++ uint32_t(clampu8( r1 - table59T58H[codeword] ) | ( clampu8( g1 - table59T58H[codeword] ) << 8 ) | ( clampu8( b1 - table59T58H[codeword] ) << 16 )) + }; + + for( uint8_t j = 0; j < 4; j++ ) +@@ -483,10 +483,10 @@ static etcpak_force_inline void DecodeHAlpha( uint64_t block, uint64_t alpha, ui + const auto tbl = g_alpha[(alpha >> 48) & 0xF]; + + const uint32_t col_tab[] = { +- clampu8( r0 + table59T58H[codeword] ) | ( clampu8( g0 + table59T58H[codeword] ) << 8 ) | ( clampu8( b0 + table59T58H[codeword] ) << 16 ), +- clampu8( r0 - table59T58H[codeword] ) | ( clampu8( g0 - table59T58H[codeword] ) << 8 ) | ( clampu8( b0 - table59T58H[codeword] ) << 16 ), +- clampu8( r1 + table59T58H[codeword] ) | ( clampu8( g1 + table59T58H[codeword] ) << 8 ) | ( clampu8( b1 + table59T58H[codeword] ) << 16 ), +- clampu8( r1 - table59T58H[codeword] ) | ( clampu8( g1 - table59T58H[codeword] ) << 8 ) | ( clampu8( b1 - table59T58H[codeword] ) << 16 ) ++ uint32_t(clampu8( r0 + table59T58H[codeword] ) | ( clampu8( g0 + table59T58H[codeword] ) << 8 ) | ( clampu8( b0 + table59T58H[codeword] ) << 16 )), ++ uint32_t(clampu8( r0 - table59T58H[codeword] ) | ( clampu8( g0 - table59T58H[codeword] ) << 8 ) | ( clampu8( b0 - table59T58H[codeword] ) << 16 )), ++ uint32_t(clampu8( r1 + table59T58H[codeword] ) | ( clampu8( g1 + table59T58H[codeword] ) << 8 ) | ( clampu8( b1 + table59T58H[codeword] ) << 16 )), ++ uint32_t(clampu8( r1 - table59T58H[codeword] ) | ( clampu8( g1 - table59T58H[codeword] ) << 8 ) | ( clampu8( b1 - table59T58H[codeword] ) << 16 )) + }; + + for( uint8_t j = 0; j < 4; j++ ) diff --git a/thirdparty/etcpak/patches/windows-mingw-fixes.patch b/thirdparty/etcpak/patches/windows-mingw-fixes.patch new file mode 100644 index 0000000000..1da60e4a4f --- /dev/null +++ b/thirdparty/etcpak/patches/windows-mingw-fixes.patch @@ -0,0 +1,63 @@ +diff --git a/thirdparty/etcpak/BlockData.cpp b/thirdparty/etcpak/BlockData.cpp +index a2cd032c5b..bd738085f3 100644 +--- a/thirdparty/etcpak/BlockData.cpp ++++ b/thirdparty/etcpak/BlockData.cpp +@@ -15,7 +15,7 @@ + # include <arm_neon.h> + #endif + +-#ifdef __SSE4_1__ ++#if defined __SSE4_1__ || defined __AVX2__ || defined _MSC_VER + # ifdef _MSC_VER + # include <intrin.h> + # include <Windows.h> +@@ -24,12 +24,6 @@ + # else + # include <x86intrin.h> + # endif +-#else +-# ifndef _MSC_VER +-# include <byteswap.h> +-# define _bswap(x) bswap_32(x) +-# define _bswap64(x) bswap_64(x) +-# endif + #endif + + #ifndef _bswap +diff --git a/thirdparty/etcpak/ProcessRGB.cpp b/thirdparty/etcpak/ProcessRGB.cpp +index 220d5c55e2..9dc5a78b67 100644 +--- a/thirdparty/etcpak/ProcessRGB.cpp ++++ b/thirdparty/etcpak/ProcessRGB.cpp +@@ -1,5 +1,6 @@ + #include <array> + #include <string.h> ++#include <limits> + + #ifdef __ARM_NEON + # include <arm_neon.h> +@@ -21,12 +22,6 @@ + # else + # include <x86intrin.h> + # endif +-#else +-# ifndef _MSC_VER +-# include <byteswap.h> +-# define _bswap(x) bswap_32(x) +-# define _bswap64(x) bswap_64(x) +-# endif + #endif + + #ifndef _bswap +diff --git a/thirdparty/etcpak/System.cpp b/thirdparty/etcpak/System.cpp +index 1383d0ecd0..a09b289cb2 100644 +--- a/thirdparty/etcpak/System.cpp ++++ b/thirdparty/etcpak/System.cpp +@@ -35,7 +35,7 @@ unsigned int System::CPUCores() + + void System::SetThreadName( std::thread& thread, const char* name ) + { +-#ifdef _WIN32 ++#ifdef _MSC_VER + const DWORD MS_VC_EXCEPTION=0x406D1388; + + # pragma pack( push, 8 ) |