mirror of https://github.com/snes9xgit/snes9x.git
Win32: update xbrz to 1.2
This commit is contained in:
parent
47f82d7e61
commit
1d140638da
356
filter/xbrz.cpp
356
filter/xbrz.cpp
|
@ -16,12 +16,14 @@
|
|||
#include "xbrz.h"
|
||||
#include <cassert>
|
||||
#include <algorithm>
|
||||
#include <vector>
|
||||
|
||||
namespace
|
||||
{
|
||||
template <uint32_t N> inline
|
||||
unsigned char getByte(uint32_t val) { return static_cast<unsigned char>((val >> (8 * N)) & 0xff); }
|
||||
|
||||
inline unsigned char getAlpha(uint32_t val) { return getByte<3>(val); }
|
||||
inline unsigned char getRed (uint32_t val) { return getByte<2>(val); }
|
||||
inline unsigned char getGreen(uint32_t val) { return getByte<1>(val); }
|
||||
inline unsigned char getBlue (uint32_t val) { return getByte<0>(val); }
|
||||
|
@ -29,7 +31,7 @@ inline unsigned char getBlue (uint32_t val) { return getByte<0>(val); }
|
|||
template <class T> inline
|
||||
T abs(T value)
|
||||
{
|
||||
//static_assert(std::is_signed<T>::value, "");
|
||||
//static_assert(std::is_signed<T>::value, "abs() requires signed types");
|
||||
return value < 0 ? -value : value;
|
||||
}
|
||||
|
||||
|
@ -37,22 +39,28 @@ const uint32_t redMask = 0xff0000;
|
|||
const uint32_t greenMask = 0x00ff00;
|
||||
const uint32_t blueMask = 0x0000ff;
|
||||
|
||||
template <unsigned int N, unsigned int M> inline
|
||||
void alphaBlend(uint32_t& dst, uint32_t col) //blend color over destination with opacity N / M
|
||||
template <unsigned int M, unsigned int N> inline
|
||||
void alphaBlend(uint32_t& dst, uint32_t col) //blend color over destination with opacity M / N
|
||||
{
|
||||
//static_assert(N < 256, "possible overflow of (col & redMask) * N");
|
||||
//static_assert(M < 256, "possible overflow of (col & redMask ) * N + (dst & redMask ) * (M - N)");
|
||||
//static_assert(0 < N && N < M, "");
|
||||
dst = (redMask & ((col & redMask ) * N + (dst & redMask ) * (M - N)) / M) | //this works because 8 upper bits are free
|
||||
(greenMask & ((col & greenMask) * N + (dst & greenMask) * (M - N)) / M) |
|
||||
(blueMask & ((col & blueMask ) * N + (dst & blueMask ) * (M - N)) / M);
|
||||
//static_assert(0 < M && M < N && N <= 256, "possible overflow of (col & byte1Mask) * M + (dst & byte1Mask) * (N - M)");
|
||||
|
||||
const uint32_t byte1Mask = 0x000000ff;
|
||||
const uint32_t byte2Mask = 0x0000ff00;
|
||||
const uint32_t byte3Mask = 0x00ff0000;
|
||||
const uint32_t byte4Mask = 0xff000000;
|
||||
|
||||
dst = (byte1Mask & (((col & byte1Mask) * M + (dst & byte1Mask) * (N - M)) / N)) | //
|
||||
(byte2Mask & (((col & byte2Mask) * M + (dst & byte2Mask) * (N - M)) / N)) | //this works because next higher 8 bits are free
|
||||
(byte3Mask & (((col & byte3Mask) * M + (dst & byte3Mask) * (N - M)) / N)) | //
|
||||
(byte4Mask & (((((col & byte4Mask) >> 8) * M + ((dst & byte4Mask) >> 8) * (N - M)) / N) << 8)); //next 8 bits are not free, so shift
|
||||
//the last row operating on a potential alpha channel costs only ~1% perf => negligible!
|
||||
}
|
||||
|
||||
|
||||
//inline
|
||||
//double fastSqrt(double n)
|
||||
//{
|
||||
// __asm //speeds up xBRZ by about 9% compared to std::sqrt
|
||||
// __asm //speeds up xBRZ by about 9% compared to std::sqrt which internally uses the same assembler instructions but adds some "fluff"
|
||||
// {
|
||||
// fld n
|
||||
// fsqrt
|
||||
|
@ -61,17 +69,17 @@ void alphaBlend(uint32_t& dst, uint32_t col) //blend color over destination with
|
|||
//
|
||||
|
||||
|
||||
inline
|
||||
uint32_t alphaBlend2(uint32_t pix1, uint32_t pix2, double alpha)
|
||||
{
|
||||
return (redMask & static_cast<uint32_t>((pix1 & redMask ) * alpha + (pix2 & redMask ) * (1 - alpha))) |
|
||||
(greenMask & static_cast<uint32_t>((pix1 & greenMask) * alpha + (pix2 & greenMask) * (1 - alpha))) |
|
||||
(blueMask & static_cast<uint32_t>((pix1 & blueMask ) * alpha + (pix2 & blueMask ) * (1 - alpha)));
|
||||
}
|
||||
//inline
|
||||
//uint32_t alphaBlend2(uint32_t pix1, uint32_t pix2, double alpha)
|
||||
//{
|
||||
// return (redMask & static_cast<uint32_t>((pix1 & redMask ) * alpha + (pix2 & redMask ) * (1 - alpha))) |
|
||||
// (greenMask & static_cast<uint32_t>((pix1 & greenMask) * alpha + (pix2 & greenMask) * (1 - alpha))) |
|
||||
// (blueMask & static_cast<uint32_t>((pix1 & blueMask ) * alpha + (pix2 & blueMask ) * (1 - alpha)));
|
||||
//}
|
||||
|
||||
|
||||
uint32_t* byteAdvance( uint32_t* ptr, int bytes) { return reinterpret_cast< uint32_t*>(reinterpret_cast< char*>(ptr) + bytes); }
|
||||
const uint32_t* byteAdvance(const uint32_t* ptr, int bytes) { return reinterpret_cast<const uint32_t*>(reinterpret_cast<const char*>(ptr) + bytes); }
|
||||
uint32_t* byteAdvance( uint32_t* ptr, int bytes) { return reinterpret_cast< uint32_t*>(reinterpret_cast< char*>(ptr) + bytes); }
|
||||
const uint32_t* byteAdvance(const uint32_t* ptr, int bytes) { return reinterpret_cast<const uint32_t*>(reinterpret_cast<const char*>(ptr) + bytes); }
|
||||
|
||||
|
||||
//fill block with the given color
|
||||
|
@ -91,11 +99,11 @@ void fillBlock(uint32_t* trg, int pitch, uint32_t col, int n) { fillBlock(trg, p
|
|||
|
||||
|
||||
#ifdef _MSC_VER
|
||||
#define FORCE_INLINE __forceinline
|
||||
#define FORCE_INLINE __forceinline
|
||||
#elif defined __GNUC__
|
||||
#define FORCE_INLINE __attribute__((always_inline)) inline
|
||||
#define FORCE_INLINE __attribute__((always_inline)) inline
|
||||
#else
|
||||
#define FORCE_INLINE inline
|
||||
#define FORCE_INLINE inline
|
||||
#endif
|
||||
|
||||
|
||||
|
@ -182,12 +190,12 @@ void rgbtoLuv(uint32_t c, double& L, double& u, double& v)
|
|||
double y = 0.2126729 * r + 0.7151522 * g + 0.0721750 * b;
|
||||
double z = 0.0193339 * r + 0.1191920 * g + 0.9503041 * b;
|
||||
//---------------------
|
||||
double var_U = 4 * x / ( x + 15 * y + 3 * z );
|
||||
double var_V = 9 * y / ( x + 15 * y + 3 * z );
|
||||
double var_U = 4 * x / ( x + 15 * y + 3 * z );
|
||||
double var_V = 9 * y / ( x + 15 * y + 3 * z );
|
||||
double var_Y = y / 100;
|
||||
|
||||
if ( var_Y > 0.008856 ) var_Y = std::pow(var_Y , 1.0/3 );
|
||||
else var_Y = 7.787 * var_Y + 16.0 / 116;
|
||||
else var_Y = 7.787 * var_Y + 16.0 / 116;
|
||||
|
||||
const double ref_X = 95.047; //Observer= 2°, Illuminant= D65
|
||||
const double ref_Y = 100.000;
|
||||
|
@ -341,7 +349,7 @@ double distHSL(uint32_t pix1, uint32_t pix2, double lightningWeight)
|
|||
double y2 = r2 * s2 * std::sin(h2 * 2 * numeric::pi);
|
||||
double z2 = l2;
|
||||
|
||||
return 255 * std::sqrt(square(x1 - x2) + square(y1 - y2) + square(lightningWeight * (z1 - z2)));
|
||||
return 255 * std::sqrt(square(x1 - x2) + square(y1 - y2) + square(lightningWeight * (z1 - z2)));
|
||||
}
|
||||
*/
|
||||
|
||||
|
@ -380,8 +388,10 @@ double distYCbCr(uint32_t pix1, uint32_t pix2, double lumaWeight)
|
|||
const int g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2); //
|
||||
const int b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2); //substraction for int is noticeable faster than for double!
|
||||
|
||||
const double k_b = 0.0722; //ITU-R BT.709 conversion
|
||||
const double k_r = 0.2126; //
|
||||
//const double k_b = 0.0722; //ITU-R BT.709 conversion
|
||||
//const double k_r = 0.2126; //
|
||||
const double k_b = 0.0593; //ITU-R BT.2020 conversion
|
||||
const double k_r = 0.2627; //
|
||||
const double k_g = 1 - k_b - k_r;
|
||||
|
||||
const double scale_b = 0.5 / (1 - k_b);
|
||||
|
@ -392,10 +402,57 @@ double distYCbCr(uint32_t pix1, uint32_t pix2, double lumaWeight)
|
|||
const double c_r = scale_r * (r_diff - y);
|
||||
|
||||
//we skip division by 255 to have similar range like other distance functions
|
||||
return std::sqrt(square(lumaWeight * y) + square(c_b) + square(c_r));
|
||||
return std::sqrt(square(lumaWeight * y) + square(c_b) + square(c_r));
|
||||
}
|
||||
|
||||
|
||||
struct DistYCbCrBuffer //30% perf boost compared to distYCbCr()!
|
||||
{
|
||||
public:
|
||||
DistYCbCrBuffer() : buffer(256 * 256 * 256)
|
||||
{
|
||||
for (uint32_t i = 0; i < 256 * 256 * 256; ++i) //startup time: 114 ms on Intel Core i5 (four cores)
|
||||
{
|
||||
const int r_diff = getByte<2>(i) * 2 - 255;
|
||||
const int g_diff = getByte<1>(i) * 2 - 255;
|
||||
const int b_diff = getByte<0>(i) * 2 - 255;
|
||||
|
||||
const double k_b = 0.0593; //ITU-R BT.2020 conversion
|
||||
const double k_r = 0.2627; //
|
||||
const double k_g = 1 - k_b - k_r;
|
||||
|
||||
const double scale_b = 0.5 / (1 - k_b);
|
||||
const double scale_r = 0.5 / (1 - k_r);
|
||||
|
||||
const double y = k_r * r_diff + k_g * g_diff + k_b * b_diff; //[!], analog YCbCr!
|
||||
const double c_b = scale_b * (b_diff - y);
|
||||
const double c_r = scale_r * (r_diff - y);
|
||||
|
||||
buffer[i] = static_cast<float>(std::sqrt(square(y) + square(c_b) + square(c_r)));
|
||||
}
|
||||
}
|
||||
|
||||
double dist(uint32_t pix1, uint32_t pix2) const
|
||||
{
|
||||
//if (pix1 == pix2) -> 8% perf degradation!
|
||||
// return 0;
|
||||
//if (pix1 > pix2)
|
||||
// std::swap(pix1, pix2); -> 30% perf degradation!!!
|
||||
|
||||
const int r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2);
|
||||
const int g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2);
|
||||
const int b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2);
|
||||
|
||||
return buffer[(((r_diff + 255) / 2) << 16) | //slightly reduce precision (division by 2) to squeeze value into single byte
|
||||
(((g_diff + 255) / 2) << 8) |
|
||||
(( b_diff + 255) / 2)];
|
||||
}
|
||||
|
||||
private:
|
||||
std::vector<float> buffer; //consumes 64 MB memory; using double is 2% faster, but takes 128 MB
|
||||
} distYCbCrBuffer;
|
||||
|
||||
|
||||
inline
|
||||
double distYUV(uint32_t pix1, uint32_t pix2, double luminanceWeight)
|
||||
{
|
||||
|
@ -423,27 +480,11 @@ double distYUV(uint32_t pix1, uint32_t pix2, double luminanceWeight)
|
|||
#ifndef NDEBUG
|
||||
const double eps = 0.5;
|
||||
#endif
|
||||
assert(std::abs(y) <= 255 + eps);
|
||||
assert(std::abs(u) <= 255 * 2 * u_max + eps);
|
||||
assert(std::abs(v) <= 255 * 2 * v_max + eps);
|
||||
assert(abs(y) <= 255 + eps);
|
||||
assert(abs(u) <= 255 * 2 * u_max + eps);
|
||||
assert(abs(v) <= 255 * 2 * v_max + eps);
|
||||
|
||||
return std::sqrt(square(luminanceWeight * y) + square(u) + square(v));
|
||||
}
|
||||
|
||||
|
||||
inline
|
||||
double colorDist(uint32_t pix1, uint32_t pix2, double luminanceWeight)
|
||||
{
|
||||
if (pix1 == pix2) //about 8% perf boost
|
||||
return 0;
|
||||
|
||||
//return distHSL(pix1, pix2, luminanceWeight);
|
||||
//return distRGB(pix1, pix2);
|
||||
//return distLAB(pix1, pix2);
|
||||
//return distNonLinearRGB(pix1, pix2);
|
||||
//return distYUV(pix1, pix2, luminanceWeight);
|
||||
|
||||
return distYCbCr(pix1, pix2, luminanceWeight);
|
||||
return std::sqrt(square(luminanceWeight * y) + square(u) + square(v));
|
||||
}
|
||||
|
||||
|
||||
|
@ -472,19 +513,20 @@ struct Kernel_4x4 //kernel for preprocessing step
|
|||
/**/m, n, o, p;
|
||||
};
|
||||
|
||||
#define dist(col1, col2) colorDist(col1, col2, cfg.luminanceWeight_)
|
||||
#define cdist(pix1, pix2) ColorDistance::dist((pix1), (pix2), cfg.luminanceWeight_)
|
||||
/*
|
||||
input kernel area naming convention:
|
||||
-----------------
|
||||
| A | B | C | D |
|
||||
----|---|---|---|
|
||||
| E | F | G | H | //evalute the four corners between F, G, J, K
|
||||
| E | F | G | H | //evaluate the four corners between F, G, J, K
|
||||
----|---|---|---| //input pixel is at position F
|
||||
| I | J | K | L |
|
||||
----|---|---|---|
|
||||
| M | N | O | P |
|
||||
-----------------
|
||||
*/
|
||||
template <class ColorDistance>
|
||||
FORCE_INLINE //detect blend direction
|
||||
BlendResult preProcessCorners(const Kernel_4x4& ker, const xbrz::ScalerCfg& cfg) //result: F, G, J, K corners of "GradientType"
|
||||
{
|
||||
|
@ -496,11 +538,11 @@ BlendResult preProcessCorners(const Kernel_4x4& ker, const xbrz::ScalerCfg& cfg)
|
|||
ker.g == ker.k))
|
||||
return result;
|
||||
|
||||
//auto dist = [&](uint32_t col1, uint32_t col2) { return colorDist(col1, col2, cfg.luminanceWeight_); };
|
||||
//auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight_); };
|
||||
|
||||
const int weight = 4;
|
||||
double jg = dist(ker.i, ker.f) + dist(ker.f, ker.c) + dist(ker.n, ker.k) + dist(ker.k, ker.h) + weight * dist(ker.j, ker.g);
|
||||
double fk = dist(ker.e, ker.j) + dist(ker.j, ker.o) + dist(ker.b, ker.g) + dist(ker.g, ker.l) + weight * dist(ker.f, ker.k);
|
||||
double jg = cdist(ker.i, ker.f) + cdist(ker.f, ker.c) + cdist(ker.n, ker.k) + cdist(ker.k, ker.h) + weight * cdist(ker.j, ker.g);
|
||||
double fk = cdist(ker.e, ker.j) + cdist(ker.j, ker.o) + cdist(ker.b, ker.g) + cdist(ker.g, ker.l) + weight * cdist(ker.f, ker.k);
|
||||
|
||||
if (jg < fk) //test sample: 70% of values max(jg, fk) / min(jg, fk) are between 1.1 and 3.7 with median being 1.8
|
||||
{
|
||||
|
@ -578,12 +620,12 @@ template <> inline unsigned char rotateBlendInfo<ROT_270>(unsigned char b) { ret
|
|||
|
||||
|
||||
#ifndef NDEBUG
|
||||
int debugPixelX = -1;
|
||||
int debugPixelY = 84;
|
||||
bool breakIntoDebugger = false;
|
||||
int debugPixelX = -1;
|
||||
int debugPixelY = 84;
|
||||
bool breakIntoDebugger = false;
|
||||
#endif
|
||||
|
||||
#define eq(col1, col2) (colorDist(col1, col2, cfg.luminanceWeight_) < cfg.equalColorTolerance_)
|
||||
#define eq(pix1, pix2) (ColorDistance::dist((pix1), (pix2), cfg.luminanceWeight_) < cfg.equalColorTolerance_)
|
||||
|
||||
/*
|
||||
input kernel area naming convention:
|
||||
|
@ -595,9 +637,9 @@ input kernel area naming convention:
|
|||
| G | H | I |
|
||||
-------------
|
||||
*/
|
||||
template <class Scaler, RotationDegree rotDeg>
|
||||
template <class Scaler, class ColorDistance, RotationDegree rotDeg>
|
||||
FORCE_INLINE //perf: quite worth it!
|
||||
void scalePixel(const Kernel_3x3& ker,
|
||||
void blendPixel(const Kernel_3x3& ker,
|
||||
uint32_t* target, int trgWidth,
|
||||
unsigned char blendInfo, //result of preprocessing all four corners of pixel "e"
|
||||
const xbrz::ScalerCfg& cfg)
|
||||
|
@ -621,8 +663,8 @@ void scalePixel(const Kernel_3x3& ker,
|
|||
|
||||
if (getBottomR(blend) >= BLEND_NORMAL)
|
||||
{
|
||||
//auto eq = [&](uint32_t col1, uint32_t col2) { return colorDist(col1, col2, cfg.luminanceWeight_) < cfg.equalColorTolerance_; };
|
||||
//auto dist = [&](uint32_t col1, uint32_t col2) { return colorDist(col1, col2, cfg.luminanceWeight_); };
|
||||
//auto eq = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight_) < cfg.equalColorTolerance_; };
|
||||
//auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight_); };
|
||||
|
||||
bool doLineBlend = true;
|
||||
if (getBottomR(blend) >= BLEND_DOMINANT)
|
||||
|
@ -633,19 +675,19 @@ void scalePixel(const Kernel_3x3& ker,
|
|||
else if(getBottomL(blend) != BLEND_NONE && !eq(e, c))
|
||||
doLineBlend = false;
|
||||
//no full blending for L-shapes; blend corner only (handles "mario mushroom eyes")
|
||||
else if (eq(g, h) && eq(h , i) && eq(i, f) && eq(f, c) && !eq(e, i))
|
||||
else if (!eq(e, i) && eq(g, h) && eq(h , i) && eq(i, f) && eq(f, c))
|
||||
doLineBlend = false;
|
||||
else
|
||||
doLineBlend = true;
|
||||
|
||||
const uint32_t px = dist(e, f) <= dist(e, h) ? f : h; //choose most similar color
|
||||
const uint32_t px = cdist(e, f) <= cdist(e, h) ? f : h; //choose most similar color
|
||||
|
||||
OutputMatrix<Scaler::scale, rotDeg> out(target, trgWidth);
|
||||
|
||||
if (doLineBlend)
|
||||
{
|
||||
const double fg = dist(f, g); //test sample: 70% of values max(fg, hc) / min(fg, hc) are between 1.1 and 3.7 with median being 1.9
|
||||
const double hc = dist(h, c); //
|
||||
const double fg = cdist(f, g); //test sample: 70% of values max(fg, hc) / min(fg, hc) are between 1.1 and 3.7 with median being 1.9
|
||||
const double hc = cdist(h, c); //
|
||||
|
||||
const bool haveShallowLine = cfg.steepDirectionThreshold * fg <= hc && e != g && d != g;
|
||||
const bool haveSteepLine = cfg.steepDirectionThreshold * hc <= fg && e != c && b != c;
|
||||
|
@ -681,7 +723,7 @@ void scalePixel(const Kernel_3x3& ker,
|
|||
}
|
||||
|
||||
|
||||
template <class Scaler> //scaler policy: see "Scaler2x" reference implementation
|
||||
template <class Scaler, class ColorDistance> //scaler policy: see "Scaler2x" reference implementation
|
||||
void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, const xbrz::ScalerCfg& cfg, int yFirst, int yLast)
|
||||
{
|
||||
yFirst = std::max(yFirst, 0);
|
||||
|
@ -698,7 +740,7 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight,
|
|||
std::fill(preProcBuffer, preProcBuffer + bufferSize, 0);
|
||||
//static_assert(BLEND_NONE == 0, "");
|
||||
|
||||
//initialize preprocessing buffer for first row: detect upper left and right corner blending
|
||||
//initialize preprocessing buffer for first row of current stripe: detect upper left and right corner blending
|
||||
//this cannot be optimized for adjacent processing stripes; we must not allow for a memory race condition!
|
||||
if (yFirst > 0)
|
||||
{
|
||||
|
@ -715,7 +757,7 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight,
|
|||
const int x_p1 = std::min(x + 1, srcWidth - 1);
|
||||
const int x_p2 = std::min(x + 2, srcWidth - 1);
|
||||
|
||||
Kernel_4x4 ker = {}; //perf: initialization is negligable
|
||||
Kernel_4x4 ker = {}; //perf: initialization is negligible
|
||||
ker.a = s_m1[x_m1]; //read sequentially from memory as far as possible
|
||||
ker.b = s_m1[x];
|
||||
ker.c = s_m1[x_p1];
|
||||
|
@ -736,7 +778,7 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight,
|
|||
ker.o = s_p2[x_p1];
|
||||
ker.p = s_p2[x_p2];
|
||||
|
||||
const BlendResult res = preProcessCorners(ker, cfg);
|
||||
const BlendResult res = preProcessCorners<ColorDistance>(ker, cfg);
|
||||
/*
|
||||
preprocessing blend result:
|
||||
---------
|
||||
|
@ -747,7 +789,7 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight,
|
|||
*/
|
||||
setTopR(preProcBuffer[x], res.blend_j);
|
||||
|
||||
if (x + 1 < srcWidth)
|
||||
if (x + 1 < bufferSize)
|
||||
setTopL(preProcBuffer[x + 1], res.blend_k);
|
||||
}
|
||||
}
|
||||
|
@ -774,31 +816,32 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight,
|
|||
const int x_p1 = std::min(x + 1, srcWidth - 1);
|
||||
const int x_p2 = std::min(x + 2, srcWidth - 1);
|
||||
|
||||
Kernel_4x4 ker4 = {}; //perf: initialization is negligible
|
||||
|
||||
ker4.a = s_m1[x_m1]; //read sequentially from memory as far as possible
|
||||
ker4.b = s_m1[x];
|
||||
ker4.c = s_m1[x_p1];
|
||||
ker4.d = s_m1[x_p2];
|
||||
|
||||
ker4.e = s_0[x_m1];
|
||||
ker4.f = s_0[x];
|
||||
ker4.g = s_0[x_p1];
|
||||
ker4.h = s_0[x_p2];
|
||||
|
||||
ker4.i = s_p1[x_m1];
|
||||
ker4.j = s_p1[x];
|
||||
ker4.k = s_p1[x_p1];
|
||||
ker4.l = s_p1[x_p2];
|
||||
|
||||
ker4.m = s_p2[x_m1];
|
||||
ker4.n = s_p2[x];
|
||||
ker4.o = s_p2[x_p1];
|
||||
ker4.p = s_p2[x_p2];
|
||||
|
||||
//evaluate the four corners on bottom-right of current pixel
|
||||
unsigned char blend_xy = 0; //for current (x, y) position
|
||||
{
|
||||
Kernel_4x4 ker = {}; //perf: initialization is negligable
|
||||
ker.a = s_m1[x_m1]; //read sequentially from memory as far as possible
|
||||
ker.b = s_m1[x];
|
||||
ker.c = s_m1[x_p1];
|
||||
ker.d = s_m1[x_p2];
|
||||
|
||||
ker.e = s_0[x_m1];
|
||||
ker.f = s_0[x];
|
||||
ker.g = s_0[x_p1];
|
||||
ker.h = s_0[x_p2];
|
||||
|
||||
ker.i = s_p1[x_m1];
|
||||
ker.j = s_p1[x];
|
||||
ker.k = s_p1[x_p1];
|
||||
ker.l = s_p1[x_p2];
|
||||
|
||||
ker.m = s_p2[x_m1];
|
||||
ker.n = s_p2[x];
|
||||
ker.o = s_p2[x_p1];
|
||||
ker.p = s_p2[x_p2];
|
||||
|
||||
const BlendResult res = preProcessCorners(ker, cfg);
|
||||
const BlendResult res = preProcessCorners<ColorDistance>(ker4, cfg);
|
||||
/*
|
||||
preprocessing blend result:
|
||||
---------
|
||||
|
@ -816,39 +859,40 @@ void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight,
|
|||
blend_xy1 = 0;
|
||||
setTopL(blend_xy1, res.blend_k); //set 1st known corner for (x + 1, y + 1) and buffer for use on next column
|
||||
|
||||
if (x + 1 < srcWidth) //set 3rd known corner for (x + 1, y)
|
||||
if (x + 1 < bufferSize) //set 3rd known corner for (x + 1, y)
|
||||
setBottomL(preProcBuffer[x + 1], res.blend_g);
|
||||
}
|
||||
|
||||
//fill block of size scale * scale with the given color
|
||||
fillBlock(out, trgWidth * sizeof(uint32_t), s_0[x], Scaler::scale); //place *after* preprocessing step, to not overwrite the results while processing the the last pixel!
|
||||
fillBlock(out, trgWidth * sizeof(uint32_t), ker4.f, Scaler::scale); //place *after* preprocessing step, to not overwrite the results while processing the the last pixel!
|
||||
|
||||
//blend four corners of current pixel
|
||||
if (blendingNeeded(blend_xy)) //good 20% perf-improvement
|
||||
if (blendingNeeded(blend_xy)) //good 5% perf-improvement
|
||||
{
|
||||
Kernel_3x3 ker = {}; //perf: initialization is negligable
|
||||
Kernel_3x3 ker3 = {}; //perf: initialization is negligible
|
||||
|
||||
ker.a = s_m1[x_m1]; //read sequentially from memory as far as possible
|
||||
ker.b = s_m1[x];
|
||||
ker.c = s_m1[x_p1];
|
||||
ker3.a = ker4.a;
|
||||
ker3.b = ker4.b;
|
||||
ker3.c = ker4.c;
|
||||
|
||||
ker.d = s_0[x_m1];
|
||||
ker.e = s_0[x];
|
||||
ker.f = s_0[x_p1];
|
||||
ker3.d = ker4.e;
|
||||
ker3.e = ker4.f;
|
||||
ker3.f = ker4.g;
|
||||
|
||||
ker.g = s_p1[x_m1];
|
||||
ker.h = s_p1[x];
|
||||
ker.i = s_p1[x_p1];
|
||||
ker3.g = ker4.i;
|
||||
ker3.h = ker4.j;
|
||||
ker3.i = ker4.k;
|
||||
|
||||
scalePixel<Scaler, ROT_0 >(ker, out, trgWidth, blend_xy, cfg);
|
||||
scalePixel<Scaler, ROT_90 >(ker, out, trgWidth, blend_xy, cfg);
|
||||
scalePixel<Scaler, ROT_180>(ker, out, trgWidth, blend_xy, cfg);
|
||||
scalePixel<Scaler, ROT_270>(ker, out, trgWidth, blend_xy, cfg);
|
||||
blendPixel<Scaler, ColorDistance, ROT_0 >(ker3, out, trgWidth, blend_xy, cfg);
|
||||
blendPixel<Scaler, ColorDistance, ROT_90 >(ker3, out, trgWidth, blend_xy, cfg);
|
||||
blendPixel<Scaler, ColorDistance, ROT_180>(ker3, out, trgWidth, blend_xy, cfg);
|
||||
blendPixel<Scaler, ColorDistance, ROT_270>(ker3, out, trgWidth, blend_xy, cfg);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
//------------------------------------------------------------------------------------
|
||||
|
||||
struct Scaler2x
|
||||
{
|
||||
|
@ -938,7 +982,7 @@ struct Scaler3x
|
|||
{
|
||||
//model a round corner
|
||||
alphaBlend<45, 100>(out.template ref<2, 2>(), col); //exact: 0.4545939598
|
||||
//alphaBlend<14, 1000>(out.template ref<2, 1>(), col); //0.01413008627 -> negligable
|
||||
//alphaBlend<14, 1000>(out.template ref<2, 1>(), col); //0.01413008627 -> negligible
|
||||
//alphaBlend<14, 1000>(out.template ref<1, 2>(), col); //0.01413008627
|
||||
}
|
||||
};
|
||||
|
@ -1082,33 +1126,97 @@ struct Scaler5x
|
|||
alphaBlend<86, 100>(out.template ref<4, 4>(), col); //exact: 0.8631434088
|
||||
alphaBlend<23, 100>(out.template ref<4, 3>(), col); //0.2306749731
|
||||
alphaBlend<23, 100>(out.template ref<3, 4>(), col); //0.2306749731
|
||||
//alphaBlend<8, 1000>(out.template ref<4, 2>(), col); //0.008384061834 -> negligable
|
||||
//alphaBlend<8, 1000>(out.template ref<4, 2>(), col); //0.008384061834 -> negligible
|
||||
//alphaBlend<8, 1000>(out.template ref<2, 4>(), col); //0.008384061834
|
||||
}
|
||||
};
|
||||
|
||||
//------------------------------------------------------------------------------------
|
||||
|
||||
struct ColorDistanceRGB
|
||||
{
|
||||
static double dist(uint32_t pix1, uint32_t pix2, double luminanceWeight)
|
||||
{
|
||||
return distYCbCrBuffer.dist(pix1, pix2);
|
||||
|
||||
//if (pix1 == pix2) //about 4% perf boost
|
||||
// return 0;
|
||||
//return distYCbCr(pix1, pix2, luminanceWeight);
|
||||
}
|
||||
};
|
||||
|
||||
struct ColorDistanceARGB
|
||||
{
|
||||
static double dist(uint32_t pix1, uint32_t pix2, double luminanceWeight)
|
||||
{
|
||||
const double a1 = getAlpha(pix1) / 255.0 ;
|
||||
const double a2 = getAlpha(pix2) / 255.0 ;
|
||||
/*
|
||||
Requirements for a color distance handling alpha channel: with a1, a2 in [0, 1]
|
||||
|
||||
1. if a1 = a2, distance should be: a1 * distYCbCr()
|
||||
2. if a1 = 0, distance should be: a2 * distYCbCr(black, white) = a2 * 255
|
||||
3. if a1 = 1, distance should be: 255 * (1 - a2) + a2 * distYCbCr()
|
||||
*/
|
||||
|
||||
return std::min(a1, a2) * distYCbCrBuffer.dist(pix1, pix2) + 255 * abs(a1 - a2);
|
||||
|
||||
//if (pix1 == pix2)
|
||||
// return 0;
|
||||
//return std::min(a1, a2) * distYCbCr(pix1, pix2, luminanceWeight) + 255 * abs(a1 - a2);
|
||||
}
|
||||
};
|
||||
}
|
||||
|
||||
|
||||
void xbrz::scale(size_t factor, const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, const xbrz::ScalerCfg& cfg, int yFirst, int yLast)
|
||||
void xbrz::scale(size_t factor, const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, ColorFormat colFmt, const xbrz::ScalerCfg& cfg, int yFirst, int yLast)
|
||||
{
|
||||
switch (factor)
|
||||
switch (colFmt)
|
||||
{
|
||||
case 2:
|
||||
return scaleImage<Scaler2x>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
case 3:
|
||||
return scaleImage<Scaler3x>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
case 4:
|
||||
return scaleImage<Scaler4x>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
case 5:
|
||||
return scaleImage<Scaler5x>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
case ColorFormat::ARGB:
|
||||
switch (factor)
|
||||
{
|
||||
case 2:
|
||||
return scaleImage<Scaler2x, ColorDistanceARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
case 3:
|
||||
return scaleImage<Scaler3x, ColorDistanceARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
case 4:
|
||||
return scaleImage<Scaler4x, ColorDistanceARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
case 5:
|
||||
return scaleImage<Scaler5x, ColorDistanceARGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
}
|
||||
break;
|
||||
|
||||
case ColorFormat::RGB:
|
||||
switch (factor)
|
||||
{
|
||||
case 2:
|
||||
return scaleImage<Scaler2x, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
case 3:
|
||||
return scaleImage<Scaler3x, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
case 4:
|
||||
return scaleImage<Scaler4x, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
case 5:
|
||||
return scaleImage<Scaler5x, ColorDistanceRGB>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast);
|
||||
}
|
||||
break;
|
||||
}
|
||||
assert(false);
|
||||
}
|
||||
|
||||
|
||||
bool xbrz::equalColor(uint32_t col1, uint32_t col2, double luminanceWeight, double equalColorTolerance)
|
||||
bool xbrz::equalColorTest(uint32_t col1, uint32_t col2, ColorFormat colFmt, double luminanceWeight, double equalColorTolerance)
|
||||
{
|
||||
return colorDist(col1, col2, luminanceWeight) < equalColorTolerance;
|
||||
switch (colFmt)
|
||||
{
|
||||
case ColorFormat::ARGB:
|
||||
return ColorDistanceARGB::dist(col1, col2, luminanceWeight) < equalColorTolerance;
|
||||
|
||||
case ColorFormat::RGB:
|
||||
return ColorDistanceRGB::dist(col1, col2, luminanceWeight) < equalColorTolerance;
|
||||
}
|
||||
assert(false);
|
||||
return false;
|
||||
}
|
||||
|
||||
|
||||
|
|
|
@ -1,6 +1,6 @@
|
|||
// ****************************************************************************
|
||||
// * This file is part of the HqMAME project. It is distributed under *
|
||||
// * GNU General Public License: http://www.gnu.org/licenses/gpl.html *
|
||||
// * GNU General Public License: http://www.gnu.org/licenses/gpl-3.0 *
|
||||
// * Copyright (C) Zenju (zenju AT gmx DOT de) - All Rights Reserved *
|
||||
// * *
|
||||
// * Additionally and as a special exception, the author gives permission *
|
||||
|
@ -31,27 +31,33 @@ namespace xbrz
|
|||
using a modified approach of xBR:
|
||||
http://board.byuu.org/viewtopic.php?f=10&t=2248
|
||||
- new rule set preserving small image features
|
||||
- highly optimized for performance
|
||||
- support alpha channel
|
||||
- support multithreading
|
||||
- support 64 bit architectures
|
||||
- support 64-bit architectures
|
||||
- support processing image slices
|
||||
*/
|
||||
|
||||
enum ColorFormat //from high bits -> low bits, 8 bit per channel
|
||||
{
|
||||
ARGB, //including alpha channel, BGRA byte order on little-endian machines
|
||||
RGB, //8 bit for each red, green, blue, upper 8 bits unused
|
||||
};
|
||||
|
||||
/*
|
||||
-> map source (srcWidth * srcHeight) to target (scale * width x scale * height) image, optionally processing a half-open slice of rows [yFirst, yLast) only
|
||||
-> color format: ARGB (BGRA byte order), alpha channel unused
|
||||
-> support for source/target pitch in bytes!
|
||||
-> if your emulator changes only a few image slices during each cycle (e.g. Dosbox) then there's no need to run xBRZ on the complete image:
|
||||
-> if your emulator changes only a few image slices during each cycle (e.g. DOSBox) then there's no need to run xBRZ on the complete image:
|
||||
Just make sure you enlarge the source image slice by 2 rows on top and 2 on bottom (this is the additional range the xBRZ algorithm is using during analysis)
|
||||
Caveat: If there are multiple changed slices, make sure they do not overlap after adding these additional rows in order to avoid a memory race condition
|
||||
if you are using multiple threads for processing each enlarged slice!
|
||||
Caveat: If there are multiple changed slices, make sure they do not overlap after adding these additional rows in order to avoid a memory race condition
|
||||
in the target image data if you are using multiple threads for processing each enlarged slice!
|
||||
|
||||
THREAD-SAFETY: - parts of the same image may be scaled by multiple threads as long as the [yFirst, yLast) ranges do not overlap!
|
||||
- there is a minor inefficiency for the first row of a slice, so avoid processing single rows only
|
||||
|
||||
|
||||
- there is a minor inefficiency for the first row of a slice, so avoid processing single rows only; suggestion: process 6 rows at least
|
||||
*/
|
||||
void scale(size_t factor, //valid range: 2 - 5
|
||||
const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight,
|
||||
ColorFormat colFmt,
|
||||
const ScalerCfg& cfg = ScalerCfg(),
|
||||
int yFirst = 0, int yLast = std::numeric_limits<int>::max()); //slice of source image
|
||||
|
||||
|
@ -68,7 +74,7 @@ void nearestNeighborScale(const uint32_t* src, int srcWidth, int srcHeight, int
|
|||
SliceType st, int yFirst, int yLast);
|
||||
|
||||
//parameter tuning
|
||||
bool equalColor(uint32_t col1, uint32_t col2, double luminanceWeight, double equalColorTolerance);
|
||||
bool equalColorTest(uint32_t col1, uint32_t col2, ColorFormat colFmt, double luminanceWeight, double equalColorTolerance);
|
||||
|
||||
|
||||
|
||||
|
|
|
@ -183,6 +183,7 @@
|
|||
* Video output filters for the Windows port.
|
||||
*/
|
||||
|
||||
#include <algorithm>
|
||||
#include "../port.h"
|
||||
#include "wsnes9x.h"
|
||||
#include "../snes9x.h"
|
||||
|
@ -2712,7 +2713,7 @@ DWORD WINAPI ThreadProc_XBRZ(VOID * pParam)
|
|||
SetEvent(thread_data->xbrz_sync_event);
|
||||
WaitForSingleObject(thread_data->xbrz_start_event, INFINITE);
|
||||
|
||||
xbrz::scale(thread_data->scalingFactor, &renderBuffer[0], &xbrzBuffer[0], xbrz_thread_data::src->Width, xbrz_thread_data::src->Height, xbrz::ScalerCfg(), thread_data->yFirst, thread_data->yLast);
|
||||
xbrz::scale(thread_data->scalingFactor, &renderBuffer[0], &xbrzBuffer[0], xbrz_thread_data::src->Width, xbrz_thread_data::src->Height, xbrz::ColorFormat::RGB, xbrz::ScalerCfg(), thread_data->yFirst, thread_data->yLast);
|
||||
SetEvent(thread_data->xbrz_sync_event);
|
||||
WaitForSingleObject(thread_data->xbrz_start_event, INFINITE);
|
||||
|
||||
|
|
Loading…
Reference in New Issue