Vulkan: Allow custom texture upscaling shaders.

Let's not just hardcode.
This commit is contained in:
Unknown W. Brackets committed 2020-08-01 22:00:04 -07:00
1 parent 5ff75f4ca8
commit d99e67a061
5 files changed
+342 -297

No files matched your search

+1 -1
View File
@@ -198,7 +198,7 @@ public:
// FramebufferManager keeps TextureCache updated about what regions of memory are being rendered to.
void NotifyFramebuffer(u32 address, VirtualFramebuffer *framebuffer, FramebufferNotification msg);
void NotifyConfigChanged();
virtual void NotifyConfigChanged();
void NotifyVideoUpload(u32 addr, int size, int width, GEBufferFormat fmt);
int AttachedDrawingHeight();
+52 -295
View File
@@ -19,6 +19,7 @@
#include <cstring>
#include "ext/xxhash.h"
#include "file/vfs.h"
#include "i18n/i18n.h"
#include "math/math_util.h"
#include "profiler/profiler.h"
@@ -38,6 +39,7 @@
#include "GPU/ge_constants.h"
#include "GPU/GPUState.h"
#include "GPU/Common/PostShader.h"
#include "GPU/Common/TextureCacheCommon.h"
#include "GPU/Common/TextureDecoder.h"
#include "GPU/Vulkan/TextureCacheVulkan.h"
@@ -62,288 +64,6 @@ static const VkComponentMapping VULKAN_1555_SWIZZLE = { VK_COMPONENT_SWIZZLE_B,
static const VkComponentMapping VULKAN_565_SWIZZLE = { VK_COMPONENT_SWIZZLE_R, VK_COMPONENT_SWIZZLE_G, VK_COMPONENT_SWIZZLE_B, VK_COMPONENT_SWIZZLE_A };
static const VkComponentMapping VULKAN_8888_SWIZZLE = { VK_COMPONENT_SWIZZLE_R, VK_COMPONENT_SWIZZLE_G, VK_COMPONENT_SWIZZLE_B, VK_COMPONENT_SWIZZLE_A };
// 4xBRZ shader - Copyright (C) 2014-2016 DeSmuME team (GPL2+)
// Hyllian's xBR-vertex code and texel mapping
// Copyright (C) 2011/2016 Hyllian - [email protected]
const char *shader4xbrz = R"(
#define BLEND_ALPHA 1
#define BLEND_NONE 0
#define BLEND_NORMAL 1
#define BLEND_DOMINANT 2
#define LUMINANCE_WEIGHT 1.0
#define EQUAL_COLOR_TOLERANCE 30.0/255.0
#define STEEP_DIRECTION_THRESHOLD 2.2
#define DOMINANT_DIRECTION_THRESHOLD 3.6
float reduce(vec4 color) {
return dot(color.rgb, vec3(65536.0, 256.0, 1.0));
}
float DistYCbCr(vec4 pixA, vec4 pixB) {
// https://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.2020_conversion
const vec3 K = vec3(0.2627, 0.6780, 0.0593);
const mat3 MATRIX = mat3(K,
-.5 * K.r / (1.0 - K.b), -.5 * K.g / (1.0 - K.b), .5,
.5, -.5 * K.g / (1.0 - K.r), -.5 * K.b / (1.0 - K.r));
vec4 diff = pixA - pixB;
vec3 YCbCr = diff.rgb * MATRIX;
YCbCr.x *= LUMINANCE_WEIGHT;
float d = length(YCbCr);
return sqrt(pixA.a * pixB.a * d * d + diff.a * diff.a);
}
bool IsPixEqual(const vec4 pixA, const vec4 pixB) {
return (DistYCbCr(pixA, pixB) < EQUAL_COLOR_TOLERANCE);
}
bool IsBlendingNeeded(const ivec4 blend) {
ivec4 diff = blend - ivec4(BLEND_NONE);
return diff.x != 0 || diff.y != 0 || diff.z != 0 || diff.w != 0;
}
vec4 applyScalingf(uvec2 origxy, uvec2 xy) {
float dx = 1.0 / params.width;
float dy = 1.0 / params.height;
// A1 B1 C1
// A0 A B C C4
// D0 D E F F4
// G0 G H I I4
// G5 H5 I5
uvec4 t1 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y - 2); // A1 B1 C1
uvec4 t2 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y - 1); // A B C
uvec4 t3 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y + 0); // D E F
uvec4 t4 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y + 1); // G H I
uvec4 t5 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y + 2); // G5 H5 I5
uvec4 t6 = uvec4(origxy.x - 2, origxy.y - 1, origxy.y, origxy.y + 1); // A0 D0 G0
uvec4 t7 = uvec4(origxy.x + 2, origxy.y - 1, origxy.y, origxy.y + 1); // C4 F4 I4
vec2 f = fract(vec2(float(xy.x) / float(params.scale), float(xy.y) / float(params.scale)));
//---------------------------------------
// Input Pixel Mapping: |21|22|23|
// 19|06|07|08|09
// 18|05|00|01|10
// 17|04|03|02|11
// |15|14|13|
vec4 src[25];
src[21] = readColorf(t1.xw);
src[22] = readColorf(t1.yw);
src[23] = readColorf(t1.zw);
src[ 6] = readColorf(t2.xw);
src[ 7] = readColorf(t2.yw);
src[ 8] = readColorf(t2.zw);
src[ 5] = readColorf(t3.xw);
src[ 0] = readColorf(t3.yw);
src[ 1] = readColorf(t3.zw);
src[ 4] = readColorf(t4.xw);
src[ 3] = readColorf(t4.yw);
src[ 2] = readColorf(t4.zw);
src[15] = readColorf(t5.xw);
src[14] = readColorf(t5.yw);
src[13] = readColorf(t5.zw);
src[19] = readColorf(t6.xy);
src[18] = readColorf(t6.xz);
src[17] = readColorf(t6.xw);
src[ 9] = readColorf(t7.xy);
src[10] = readColorf(t7.xz);
src[11] = readColorf(t7.xw);
float v[9];
v[0] = reduce(src[0]);
v[1] = reduce(src[1]);
v[2] = reduce(src[2]);
v[3] = reduce(src[3]);
v[4] = reduce(src[4]);
v[5] = reduce(src[5]);
v[6] = reduce(src[6]);
v[7] = reduce(src[7]);
v[8] = reduce(src[8]);
ivec4 blendResult = ivec4(BLEND_NONE);
// Preprocess corners
// Pixel Tap Mapping: --|--|--|--|--
// --|--|07|08|--
// --|05|00|01|10
// --|04|03|02|11
// --|--|14|13|--
// Corner (1, 1)
if ( ((v[0] == v[1] && v[3] == v[2]) || (v[0] == v[3] && v[1] == v[2])) == false) {
float dist_03_01 = DistYCbCr(src[ 4], src[ 0]) + DistYCbCr(src[ 0], src[ 8]) + DistYCbCr(src[14], src[ 2]) + DistYCbCr(src[ 2], src[10]) + (4.0 * DistYCbCr(src[ 3], src[ 1]));
float dist_00_02 = DistYCbCr(src[ 5], src[ 3]) + DistYCbCr(src[ 3], src[13]) + DistYCbCr(src[ 7], src[ 1]) + DistYCbCr(src[ 1], src[11]) + (4.0 * DistYCbCr(src[ 0], src[ 2]));
bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_03_01) < dist_00_02;
blendResult[2] = ((dist_03_01 < dist_00_02) && (v[0] != v[1]) && (v[0] != v[3])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
}
// Pixel Tap Mapping: --|--|--|--|--
// --|06|07|--|--
// 18|05|00|01|--
// 17|04|03|02|--
// --|15|14|--|--
// Corner (0, 1)
if ( ((v[5] == v[0] && v[4] == v[3]) || (v[5] == v[4] && v[0] == v[3])) == false) {
float dist_04_00 = DistYCbCr(src[17], src[ 5]) + DistYCbCr(src[ 5], src[ 7]) + DistYCbCr(src[15], src[ 3]) + DistYCbCr(src[ 3], src[ 1]) + (4.0 * DistYCbCr(src[ 4], src[ 0]));
float dist_05_03 = DistYCbCr(src[18], src[ 4]) + DistYCbCr(src[ 4], src[14]) + DistYCbCr(src[ 6], src[ 0]) + DistYCbCr(src[ 0], src[ 2]) + (4.0 * DistYCbCr(src[ 5], src[ 3]));
bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_05_03) < dist_04_00;
blendResult[3] = ((dist_04_00 > dist_05_03) && (v[0] != v[5]) && (v[0] != v[3])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
}
// Pixel Tap Mapping: --|--|22|23|--
// --|06|07|08|09
// --|05|00|01|10
// --|--|03|02|--
// --|--|--|--|--
// Corner (1, 0)
if ( ((v[7] == v[8] && v[0] == v[1]) || (v[7] == v[0] && v[8] == v[1])) == false) {
float dist_00_08 = DistYCbCr(src[ 5], src[ 7]) + DistYCbCr(src[ 7], src[23]) + DistYCbCr(src[ 3], src[ 1]) + DistYCbCr(src[ 1], src[ 9]) + (4.0 * DistYCbCr(src[ 0], src[ 8]));
float dist_07_01 = DistYCbCr(src[ 6], src[ 0]) + DistYCbCr(src[ 0], src[ 2]) + DistYCbCr(src[22], src[ 8]) + DistYCbCr(src[ 8], src[10]) + (4.0 * DistYCbCr(src[ 7], src[ 1]));
bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_07_01) < dist_00_08;
blendResult[1] = ((dist_00_08 > dist_07_01) && (v[0] != v[7]) && (v[0] != v[1])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
}
// Pixel Tap Mapping: --|21|22|--|--
// 19|06|07|08|--
// 18|05|00|01|--
// --|04|03|--|--
// --|--|--|--|--
// Corner (0, 0)
if ( ((v[6] == v[7] && v[5] == v[0]) || (v[6] == v[5] && v[7] == v[0])) == false) {
float dist_05_07 = DistYCbCr(src[18], src[ 6]) + DistYCbCr(src[ 6], src[22]) + DistYCbCr(src[ 4], src[ 0]) + DistYCbCr(src[ 0], src[ 8]) + (4.0 * DistYCbCr(src[ 5], src[ 7]));
float dist_06_00 = DistYCbCr(src[19], src[ 5]) + DistYCbCr(src[ 5], src[ 3]) + DistYCbCr(src[21], src[ 7]) + DistYCbCr(src[ 7], src[ 1]) + (4.0 * DistYCbCr(src[ 6], src[ 0]));
bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_05_07) < dist_06_00;
blendResult[0] = ((dist_05_07 < dist_06_00) && (v[0] != v[5]) && (v[0] != v[7])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
}
vec4 dst[16];
dst[ 0] = src[0];
dst[ 1] = src[0];
dst[ 2] = src[0];
dst[ 3] = src[0];
dst[ 4] = src[0];
dst[ 5] = src[0];
dst[ 6] = src[0];
dst[ 7] = src[0];
dst[ 8] = src[0];
dst[ 9] = src[0];
dst[10] = src[0];
dst[11] = src[0];
dst[12] = src[0];
dst[13] = src[0];
dst[14] = src[0];
dst[15] = src[0];
// Scale pixel
if (IsBlendingNeeded(blendResult) == true) {
float dist_01_04 = DistYCbCr(src[1], src[4]);
float dist_03_08 = DistYCbCr(src[3], src[8]);
bool haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[4]) && (v[5] != v[4]);
bool haveSteepLine = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[8]) && (v[7] != v[8]);
bool needBlend = (blendResult[2] != BLEND_NONE);
bool doLineBlend = ( blendResult[2] >= BLEND_DOMINANT ||
((blendResult[1] != BLEND_NONE && !IsPixEqual(src[0], src[4])) ||
(blendResult[3] != BLEND_NONE && !IsPixEqual(src[0], src[8])) ||
(IsPixEqual(src[4], src[3]) && IsPixEqual(src[3], src[2]) && IsPixEqual(src[2], src[1]) && IsPixEqual(src[1], src[8]) && IsPixEqual(src[0], src[2]) == false) ) == false );
vec4 blendPix = ( DistYCbCr(src[0], src[1]) <= DistYCbCr(src[0], src[3]) ) ? src[1] : src[3];
dst[ 2] = mix(dst[ 2], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
dst[ 9] = mix(dst[ 9], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
dst[10] = mix(dst[10], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
dst[11] = mix(dst[11], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[12] = mix(dst[12], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
dst[13] = mix(dst[13], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[14] = mix(dst[14], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
dst[15] = mix(dst[15], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
dist_01_04 = DistYCbCr(src[7], src[2]);
dist_03_08 = DistYCbCr(src[1], src[6]);
haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[2]) && (v[3] != v[2]);
haveSteepLine = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[6]) && (v[5] != v[6]);
needBlend = (blendResult[1] != BLEND_NONE);
doLineBlend = ( blendResult[1] >= BLEND_DOMINANT ||
!((blendResult[0] != BLEND_NONE && !IsPixEqual(src[0], src[2])) ||
(blendResult[2] != BLEND_NONE && !IsPixEqual(src[0], src[6])) ||
(IsPixEqual(src[2], src[1]) && IsPixEqual(src[1], src[8]) && IsPixEqual(src[8], src[7]) && IsPixEqual(src[7], src[6]) && !IsPixEqual(src[0], src[8])) ) );
blendPix = ( DistYCbCr(src[0], src[7]) <= DistYCbCr(src[0], src[1]) ) ? src[7] : src[1];
dst[ 1] = mix(dst[ 1], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
dst[ 6] = mix(dst[ 6], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
dst[ 7] = mix(dst[ 7], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
dst[ 8] = mix(dst[ 8], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[ 9] = mix(dst[ 9], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
dst[10] = mix(dst[10], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[11] = mix(dst[11], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
dst[12] = mix(dst[12], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
dist_01_04 = DistYCbCr(src[5], src[8]);
dist_03_08 = DistYCbCr(src[7], src[4]);
haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[8]) && (v[1] != v[8]);
haveSteepLine = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[4]) && (v[3] != v[4]);
needBlend = (blendResult[0] != BLEND_NONE);
doLineBlend = ( blendResult[0] >= BLEND_DOMINANT ||
!((blendResult[3] != BLEND_NONE && !IsPixEqual(src[0], src[8])) ||
(blendResult[1] != BLEND_NONE && !IsPixEqual(src[0], src[4])) ||
(IsPixEqual(src[8], src[7]) && IsPixEqual(src[7], src[6]) && IsPixEqual(src[6], src[5]) && IsPixEqual(src[5], src[4]) && !IsPixEqual(src[0], src[6])) ) );
blendPix = ( DistYCbCr(src[0], src[5]) <= DistYCbCr(src[0], src[7]) ) ? src[5] : src[7];
dst[ 0] = mix(dst[ 0], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
dst[15] = mix(dst[15], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
dst[ 4] = mix(dst[ 4], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
dst[ 5] = mix(dst[ 5], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[ 6] = mix(dst[ 6], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
dst[ 7] = mix(dst[ 7], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[ 8] = mix(dst[ 8], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
dst[ 9] = mix(dst[ 9], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
dist_01_04 = DistYCbCr(src[3], src[6]);
dist_03_08 = DistYCbCr(src[5], src[2]);
haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[6]) && (v[7] != v[6]);
haveSteepLine = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[2]) && (v[1] != v[2]);
needBlend = (blendResult[3] != BLEND_NONE);
doLineBlend = ( blendResult[3] >= BLEND_DOMINANT ||
!((blendResult[2] != BLEND_NONE && !IsPixEqual(src[0], src[6])) ||
(blendResult[0] != BLEND_NONE && !IsPixEqual(src[0], src[2])) ||
(IsPixEqual(src[6], src[5]) && IsPixEqual(src[5], src[4]) && IsPixEqual(src[4], src[3]) && IsPixEqual(src[3], src[2]) && !IsPixEqual(src[0], src[4])) ) );
blendPix = ( DistYCbCr(src[0], src[3]) <= DistYCbCr(src[0], src[5]) ) ? src[3] : src[5];
dst[ 3] = mix(dst[ 3], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
dst[12] = mix(dst[12], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
dst[13] = mix(dst[13], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
dst[14] = mix(dst[14], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[15] = mix(dst[15], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
dst[ 4] = mix(dst[ 4], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[ 5] = mix(dst[ 5], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
dst[ 6] = mix(dst[ 6], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
}
// select output pixel
vec4 res = mix(mix(mix(mix(dst[ 6], dst[ 7], step(0.25, f.x)),
mix(dst[ 8], dst[ 9], step(0.75, f.x)),
step(0.50, f.x)),
mix(mix(dst[ 5], dst[ 0], step(0.25, f.x)),
mix(dst[ 1], dst[10], step(0.75, f.x)),
step(0.50, f.x)),
step(0.25, f.y)),
mix(mix(mix(dst[ 4], dst[ 3], step(0.25, f.x)),
mix(dst[ 2], dst[11], step(0.75, f.x)),
step(0.50, f.x)),
mix(mix(dst[15], dst[14], step(0.25, f.x)),
mix(dst[13], dst[12], step(0.75, f.x)),
step(0.50, f.x)),
step(0.75, f.y)),
step(0.50, f.y));
return res;
}
uint applyScalingu(uvec2 origxy, uvec2 xy) {
return packUnorm4x8(applyScalingf(origxy, xy));
}
)";
const char *copyShader = R"(
#version 450
#extension GL_ARB_separate_shader_objects : enable
@@ -645,20 +365,55 @@ void TextureCacheVulkan::DeviceRestore(VulkanContext *vulkan, Draw::DrawContext
samp.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST;
vkCreateSampler(vulkan_->GetDevice(), &samp, nullptr, &samplerNearest_);
std::string error;
std::string fullUploadShader = StringFromFormat(uploadShader, shader4xbrz);
std::string fullCopyShader = StringFromFormat(copyShader, shader4xbrz);
if (g_Config.bTexHardwareScaling) {
uploadCS_ = CompileShaderModule(vulkan_, VK_SHADER_STAGE_COMPUTE_BIT, fullUploadShader.c_str(), &error);
_dbg_assert_msg_(uploadCS_ != VK_NULL_HANDLE, "failed to compile upload shader");
copyCS_ = CompileShaderModule(vulkan_, VK_SHADER_STAGE_COMPUTE_BIT, fullCopyShader.c_str(), &error);
_dbg_assert_msg_(copyCS_!= VK_NULL_HANDLE, "failed to compile copy shader");
}
CompileScalingShader();
computeShaderManager_.DeviceRestore(vulkan);
}
void TextureCacheVulkan::NotifyConfigChanged() {
TextureCacheCommon::NotifyConfigChanged();
CompileScalingShader();
}
static std::string ReadShaderSrc(const std::string &filename) {
size_t sz = 0;
char *data = (char *)VFSReadFile(filename.c_str(), &sz);
if (!data)
return "";
std::string src(data, sz);
free(data);
return src;
}
void TextureCacheVulkan::CompileScalingShader() {
if (!g_Config.bTexHardwareScaling || g_Config.sTextureShaderName != textureShader_) {
if (uploadCS_ != VK_NULL_HANDLE)
vulkan_->Delete().QueueDeleteShaderModule(uploadCS_);
if (copyCS_ != VK_NULL_HANDLE)
vulkan_->Delete().QueueDeleteShaderModule(copyCS_);
textureShader_.clear();
}
if (!g_Config.bTexHardwareScaling)
return;
const TextureShaderInfo *shaderInfo = GetTextureShaderInfo(g_Config.sTextureShaderName);
if (!shaderInfo || shaderInfo->computeShaderFile.empty())
return;
std::string shaderSource = ReadShaderSrc(shaderInfo->computeShaderFile);
std::string fullUploadShader = StringFromFormat(uploadShader, shaderSource.c_str());
std::string fullCopyShader = StringFromFormat(copyShader, shaderSource.c_str());
std::string error;
uploadCS_ = CompileShaderModule(vulkan_, VK_SHADER_STAGE_COMPUTE_BIT, fullUploadShader.c_str(), &error);
_dbg_assert_msg_(uploadCS_ != VK_NULL_HANDLE, "failed to compile upload shader");
copyCS_ = CompileShaderModule(vulkan_, VK_SHADER_STAGE_COMPUTE_BIT, fullCopyShader.c_str(), &error);
_dbg_assert_msg_(copyCS_ != VK_NULL_HANDLE, "failed to compile copy shader");
textureShader_ = g_Config.sTextureShaderName;
}
void TextureCacheVulkan::ReleaseTexture(TexCacheEntry *entry, bool delete_them) {
DEBUG_LOG(G3D, "Deleting texture %p", entry->vkTex);
delete entry->vkTex;
@@ -1026,17 +781,19 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
badMipSizes = false;
}
bool hardwareScaling = g_Config.bTexHardwareScaling && (uploadCS_ != VK_NULL_HANDLE || copyCS_ != VK_NULL_HANDLE);
// Don't scale the PPGe texture.
if (entry->addr > 0x05000000 && entry->addr < PSP_GetKernelMemoryEnd())
scaleFactor = 1;
if ((entry->status & TexCacheEntry::STATUS_CHANGE_FREQUENT) != 0 && scaleFactor != 1 && !g_Config.bTexHardwareScaling) {
if ((entry->status & TexCacheEntry::STATUS_CHANGE_FREQUENT) != 0 && scaleFactor != 1 && !hardwareScaling) {
// Remember for later that we /wanted/ to scale this texture.
entry->status |= TexCacheEntry::STATUS_TO_SCALE;
scaleFactor = 1;
}
if (scaleFactor != 1) {
if (texelsScaledThisFrame_ >= TEXCACHE_MAX_TEXELS_SCALED && !g_Config.bTexHardwareScaling) {
if (texelsScaledThisFrame_ >= TEXCACHE_MAX_TEXELS_SCALED && !hardwareScaling) {
entry->status |= TexCacheEntry::STATUS_TO_SCALE;
scaleFactor = 1;
} else {
@@ -1088,7 +845,7 @@ void TextureCacheVulkan::BuildTexture(TexCacheEntry *const entry) {
// If we want to use the GE debugger, we should add VK_IMAGE_USAGE_TRANSFER_SRC_BIT too...
// Compute experiment
if (actualFmt == VULKAN_8888_FORMAT && scaleFactor > 1 && g_Config.bTexHardwareScaling) {
if (actualFmt == VULKAN_8888_FORMAT && scaleFactor > 1 && hardwareScaling) {
// Enable the experiment you want.
if (uploadCS_ != VK_NULL_HANDLE)
computeUpload = true;
+5
View File
@@ -94,6 +94,8 @@ public:
}
}
void NotifyConfigChanged() override;
void GetVulkanHandles(VkImageView &imageView, VkSampler &sampler) {
imageView = imageView_;
sampler = curSampler_;
@@ -123,6 +125,8 @@ private:
void ApplyTextureFramebuffer(TexCacheEntry *entry, VirtualFramebuffer *framebuffer) override;
void BuildTexture(TexCacheEntry *const entry) override;
void CompileScalingShader();
VulkanContext *vulkan_ = nullptr;
VulkanDeviceAllocator *allocator_ = nullptr;
VulkanPushBuffer *push_ = nullptr;
@@ -145,6 +149,7 @@ private:
DrawEngineVulkan *drawEngine_;
Vulkan2D *vulkan2D_;
std::string textureShader_;
VkShaderModule uploadCS_ = VK_NULL_HANDLE;
VkShaderModule copyCS_ = VK_NULL_HANDLE;
+5 -1
View File
@@ -144,4 +144,8 @@ SettingDefaultValue4=1.0
SettingMaxValue4=2.0
SettingMinValue4=0.1
SettingStep4=0.05
[Tex4xBRZ]
Type=Texture
Name=4xBRZ
Author=Hyllian
Compute=tex_4xbrz.csh
+279
View File
@@ -0,0 +1,279 @@
// 4xBRZ shader - Copyright (C) 2014-2016 DeSmuME team (GPL2+)
// Hyllian's xBR-vertex code and texel mapping
// Copyright (C) 2011/2016 Hyllian - [email protected]
#define BLEND_ALPHA 1
#define BLEND_NONE 0
#define BLEND_NORMAL 1
#define BLEND_DOMINANT 2
#define LUMINANCE_WEIGHT 1.0
#define EQUAL_COLOR_TOLERANCE 30.0/255.0
#define STEEP_DIRECTION_THRESHOLD 2.2
#define DOMINANT_DIRECTION_THRESHOLD 3.6
float reduce(vec4 color) {
return dot(color.rgb, vec3(65536.0, 256.0, 1.0));
}
float DistYCbCr(vec4 pixA, vec4 pixB) {
// https://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.2020_conversion
const vec3 K = vec3(0.2627, 0.6780, 0.0593);
const mat3 MATRIX = mat3(K,
-.5 * K.r / (1.0 - K.b), -.5 * K.g / (1.0 - K.b), .5,
.5, -.5 * K.g / (1.0 - K.r), -.5 * K.b / (1.0 - K.r));
vec4 diff = pixA - pixB;
vec3 YCbCr = diff.rgb * MATRIX;
YCbCr.x *= LUMINANCE_WEIGHT;
float d = length(YCbCr);
return sqrt(pixA.a * pixB.a * d * d + diff.a * diff.a);
}
bool IsPixEqual(const vec4 pixA, const vec4 pixB) {
return (DistYCbCr(pixA, pixB) < EQUAL_COLOR_TOLERANCE);
}
bool IsBlendingNeeded(const ivec4 blend) {
ivec4 diff = blend - ivec4(BLEND_NONE);
return diff.x != 0 || diff.y != 0 || diff.z != 0 || diff.w != 0;
}
vec4 applyScalingf(uvec2 origxy, uvec2 xy) {
float dx = 1.0 / params.width;
float dy = 1.0 / params.height;
// A1 B1 C1
// A0 A B C C4
// D0 D E F F4
// G0 G H I I4
// G5 H5 I5
uvec4 t1 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y - 2); // A1 B1 C1
uvec4 t2 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y - 1); // A B C
uvec4 t3 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y + 0); // D E F
uvec4 t4 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y + 1); // G H I
uvec4 t5 = uvec4(origxy.x - 1, origxy.x, origxy.x + 1, origxy.y + 2); // G5 H5 I5
uvec4 t6 = uvec4(origxy.x - 2, origxy.y - 1, origxy.y, origxy.y + 1); // A0 D0 G0
uvec4 t7 = uvec4(origxy.x + 2, origxy.y - 1, origxy.y, origxy.y + 1); // C4 F4 I4
vec2 f = fract(vec2(float(xy.x) / float(params.scale), float(xy.y) / float(params.scale)));
//---------------------------------------
// Input Pixel Mapping: |21|22|23|
// 19|06|07|08|09
// 18|05|00|01|10
// 17|04|03|02|11
// |15|14|13|
vec4 src[25];
src[21] = readColorf(t1.xw);
src[22] = readColorf(t1.yw);
src[23] = readColorf(t1.zw);
src[ 6] = readColorf(t2.xw);
src[ 7] = readColorf(t2.yw);
src[ 8] = readColorf(t2.zw);
src[ 5] = readColorf(t3.xw);
src[ 0] = readColorf(t3.yw);
src[ 1] = readColorf(t3.zw);
src[ 4] = readColorf(t4.xw);
src[ 3] = readColorf(t4.yw);
src[ 2] = readColorf(t4.zw);
src[15] = readColorf(t5.xw);
src[14] = readColorf(t5.yw);
src[13] = readColorf(t5.zw);
src[19] = readColorf(t6.xy);
src[18] = readColorf(t6.xz);
src[17] = readColorf(t6.xw);
src[ 9] = readColorf(t7.xy);
src[10] = readColorf(t7.xz);
src[11] = readColorf(t7.xw);
float v[9];
v[0] = reduce(src[0]);
v[1] = reduce(src[1]);
v[2] = reduce(src[2]);
v[3] = reduce(src[3]);
v[4] = reduce(src[4]);
v[5] = reduce(src[5]);
v[6] = reduce(src[6]);
v[7] = reduce(src[7]);
v[8] = reduce(src[8]);
ivec4 blendResult = ivec4(BLEND_NONE);
// Preprocess corners
// Pixel Tap Mapping: --|--|--|--|--
// --|--|07|08|--
// --|05|00|01|10
// --|04|03|02|11
// --|--|14|13|--
// Corner (1, 1)
if ( ((v[0] == v[1] && v[3] == v[2]) || (v[0] == v[3] && v[1] == v[2])) == false) {
float dist_03_01 = DistYCbCr(src[ 4], src[ 0]) + DistYCbCr(src[ 0], src[ 8]) + DistYCbCr(src[14], src[ 2]) + DistYCbCr(src[ 2], src[10]) + (4.0 * DistYCbCr(src[ 3], src[ 1]));
float dist_00_02 = DistYCbCr(src[ 5], src[ 3]) + DistYCbCr(src[ 3], src[13]) + DistYCbCr(src[ 7], src[ 1]) + DistYCbCr(src[ 1], src[11]) + (4.0 * DistYCbCr(src[ 0], src[ 2]));
bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_03_01) < dist_00_02;
blendResult[2] = ((dist_03_01 < dist_00_02) && (v[0] != v[1]) && (v[0] != v[3])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
}
// Pixel Tap Mapping: --|--|--|--|--
// --|06|07|--|--
// 18|05|00|01|--
// 17|04|03|02|--
// --|15|14|--|--
// Corner (0, 1)
if ( ((v[5] == v[0] && v[4] == v[3]) || (v[5] == v[4] && v[0] == v[3])) == false) {
float dist_04_00 = DistYCbCr(src[17], src[ 5]) + DistYCbCr(src[ 5], src[ 7]) + DistYCbCr(src[15], src[ 3]) + DistYCbCr(src[ 3], src[ 1]) + (4.0 * DistYCbCr(src[ 4], src[ 0]));
float dist_05_03 = DistYCbCr(src[18], src[ 4]) + DistYCbCr(src[ 4], src[14]) + DistYCbCr(src[ 6], src[ 0]) + DistYCbCr(src[ 0], src[ 2]) + (4.0 * DistYCbCr(src[ 5], src[ 3]));
bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_05_03) < dist_04_00;
blendResult[3] = ((dist_04_00 > dist_05_03) && (v[0] != v[5]) && (v[0] != v[3])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
}
// Pixel Tap Mapping: --|--|22|23|--
// --|06|07|08|09
// --|05|00|01|10
// --|--|03|02|--
// --|--|--|--|--
// Corner (1, 0)
if ( ((v[7] == v[8] && v[0] == v[1]) || (v[7] == v[0] && v[8] == v[1])) == false) {
float dist_00_08 = DistYCbCr(src[ 5], src[ 7]) + DistYCbCr(src[ 7], src[23]) + DistYCbCr(src[ 3], src[ 1]) + DistYCbCr(src[ 1], src[ 9]) + (4.0 * DistYCbCr(src[ 0], src[ 8]));
float dist_07_01 = DistYCbCr(src[ 6], src[ 0]) + DistYCbCr(src[ 0], src[ 2]) + DistYCbCr(src[22], src[ 8]) + DistYCbCr(src[ 8], src[10]) + (4.0 * DistYCbCr(src[ 7], src[ 1]));
bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_07_01) < dist_00_08;
blendResult[1] = ((dist_00_08 > dist_07_01) && (v[0] != v[7]) && (v[0] != v[1])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
}
// Pixel Tap Mapping: --|21|22|--|--
// 19|06|07|08|--
// 18|05|00|01|--
// --|04|03|--|--
// --|--|--|--|--
// Corner (0, 0)
if ( ((v[6] == v[7] && v[5] == v[0]) || (v[6] == v[5] && v[7] == v[0])) == false) {
float dist_05_07 = DistYCbCr(src[18], src[ 6]) + DistYCbCr(src[ 6], src[22]) + DistYCbCr(src[ 4], src[ 0]) + DistYCbCr(src[ 0], src[ 8]) + (4.0 * DistYCbCr(src[ 5], src[ 7]));
float dist_06_00 = DistYCbCr(src[19], src[ 5]) + DistYCbCr(src[ 5], src[ 3]) + DistYCbCr(src[21], src[ 7]) + DistYCbCr(src[ 7], src[ 1]) + (4.0 * DistYCbCr(src[ 6], src[ 0]));
bool dominantGradient = (DOMINANT_DIRECTION_THRESHOLD * dist_05_07) < dist_06_00;
blendResult[0] = ((dist_05_07 < dist_06_00) && (v[0] != v[5]) && (v[0] != v[7])) ? ((dominantGradient) ? BLEND_DOMINANT : BLEND_NORMAL) : BLEND_NONE;
}
vec4 dst[16];
dst[ 0] = src[0];
dst[ 1] = src[0];
dst[ 2] = src[0];
dst[ 3] = src[0];
dst[ 4] = src[0];
dst[ 5] = src[0];
dst[ 6] = src[0];
dst[ 7] = src[0];
dst[ 8] = src[0];
dst[ 9] = src[0];
dst[10] = src[0];
dst[11] = src[0];
dst[12] = src[0];
dst[13] = src[0];
dst[14] = src[0];
dst[15] = src[0];
// Scale pixel
if (IsBlendingNeeded(blendResult) == true) {
float dist_01_04 = DistYCbCr(src[1], src[4]);
float dist_03_08 = DistYCbCr(src[3], src[8]);
bool haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[4]) && (v[5] != v[4]);
bool haveSteepLine = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[8]) && (v[7] != v[8]);
bool needBlend = (blendResult[2] != BLEND_NONE);
bool doLineBlend = ( blendResult[2] >= BLEND_DOMINANT ||
((blendResult[1] != BLEND_NONE && !IsPixEqual(src[0], src[4])) ||
(blendResult[3] != BLEND_NONE && !IsPixEqual(src[0], src[8])) ||
(IsPixEqual(src[4], src[3]) && IsPixEqual(src[3], src[2]) && IsPixEqual(src[2], src[1]) && IsPixEqual(src[1], src[8]) && IsPixEqual(src[0], src[2]) == false) ) == false );
vec4 blendPix = ( DistYCbCr(src[0], src[1]) <= DistYCbCr(src[0], src[3]) ) ? src[1] : src[3];
dst[ 2] = mix(dst[ 2], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
dst[ 9] = mix(dst[ 9], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
dst[10] = mix(dst[10], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
dst[11] = mix(dst[11], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[12] = mix(dst[12], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
dst[13] = mix(dst[13], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[14] = mix(dst[14], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
dst[15] = mix(dst[15], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
dist_01_04 = DistYCbCr(src[7], src[2]);
dist_03_08 = DistYCbCr(src[1], src[6]);
haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[2]) && (v[3] != v[2]);
haveSteepLine = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[6]) && (v[5] != v[6]);
needBlend = (blendResult[1] != BLEND_NONE);
doLineBlend = ( blendResult[1] >= BLEND_DOMINANT ||
!((blendResult[0] != BLEND_NONE && !IsPixEqual(src[0], src[2])) ||
(blendResult[2] != BLEND_NONE && !IsPixEqual(src[0], src[6])) ||
(IsPixEqual(src[2], src[1]) && IsPixEqual(src[1], src[8]) && IsPixEqual(src[8], src[7]) && IsPixEqual(src[7], src[6]) && !IsPixEqual(src[0], src[8])) ) );
blendPix = ( DistYCbCr(src[0], src[7]) <= DistYCbCr(src[0], src[1]) ) ? src[7] : src[1];
dst[ 1] = mix(dst[ 1], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
dst[ 6] = mix(dst[ 6], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
dst[ 7] = mix(dst[ 7], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
dst[ 8] = mix(dst[ 8], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[ 9] = mix(dst[ 9], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
dst[10] = mix(dst[10], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[11] = mix(dst[11], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
dst[12] = mix(dst[12], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
dist_01_04 = DistYCbCr(src[5], src[8]);
dist_03_08 = DistYCbCr(src[7], src[4]);
haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[8]) && (v[1] != v[8]);
haveSteepLine = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[4]) && (v[3] != v[4]);
needBlend = (blendResult[0] != BLEND_NONE);
doLineBlend = ( blendResult[0] >= BLEND_DOMINANT ||
!((blendResult[3] != BLEND_NONE && !IsPixEqual(src[0], src[8])) ||
(blendResult[1] != BLEND_NONE && !IsPixEqual(src[0], src[4])) ||
(IsPixEqual(src[8], src[7]) && IsPixEqual(src[7], src[6]) && IsPixEqual(src[6], src[5]) && IsPixEqual(src[5], src[4]) && !IsPixEqual(src[0], src[6])) ) );
blendPix = ( DistYCbCr(src[0], src[5]) <= DistYCbCr(src[0], src[7]) ) ? src[5] : src[7];
dst[ 0] = mix(dst[ 0], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
dst[15] = mix(dst[15], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
dst[ 4] = mix(dst[ 4], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
dst[ 5] = mix(dst[ 5], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[ 6] = mix(dst[ 6], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
dst[ 7] = mix(dst[ 7], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[ 8] = mix(dst[ 8], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
dst[ 9] = mix(dst[ 9], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
dist_01_04 = DistYCbCr(src[3], src[6]);
dist_03_08 = DistYCbCr(src[5], src[2]);
haveShallowLine = (STEEP_DIRECTION_THRESHOLD * dist_01_04 <= dist_03_08) && (v[0] != v[6]) && (v[7] != v[6]);
haveSteepLine = (STEEP_DIRECTION_THRESHOLD * dist_03_08 <= dist_01_04) && (v[0] != v[2]) && (v[1] != v[2]);
needBlend = (blendResult[3] != BLEND_NONE);
doLineBlend = ( blendResult[3] >= BLEND_DOMINANT ||
!((blendResult[2] != BLEND_NONE && !IsPixEqual(src[0], src[6])) ||
(blendResult[0] != BLEND_NONE && !IsPixEqual(src[0], src[2])) ||
(IsPixEqual(src[6], src[5]) && IsPixEqual(src[5], src[4]) && IsPixEqual(src[4], src[3]) && IsPixEqual(src[3], src[2]) && !IsPixEqual(src[0], src[4])) ) );
blendPix = ( DistYCbCr(src[0], src[3]) <= DistYCbCr(src[0], src[5]) ) ? src[3] : src[5];
dst[ 3] = mix(dst[ 3], blendPix, (needBlend && doLineBlend) ? ((haveShallowLine) ? ((haveSteepLine) ? 1.0/3.0 : 0.25) : ((haveSteepLine) ? 0.25 : 0.00)) : 0.00);
dst[12] = mix(dst[12], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.25 : 0.00);
dst[13] = mix(dst[13], blendPix, (needBlend && doLineBlend && haveSteepLine) ? 0.75 : 0.00);
dst[14] = mix(dst[14], blendPix, (needBlend) ? ((doLineBlend) ? ((haveSteepLine) ? 1.00 : ((haveShallowLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[15] = mix(dst[15], blendPix, (needBlend) ? ((doLineBlend) ? 1.00 : 0.6848532563) : 0.00);
dst[ 4] = mix(dst[ 4], blendPix, (needBlend) ? ((doLineBlend) ? ((haveShallowLine) ? 1.00 : ((haveSteepLine) ? 0.75 : 0.50)) : 0.08677704501) : 0.00);
dst[ 5] = mix(dst[ 5], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.75 : 0.00);
dst[ 6] = mix(dst[ 6], blendPix, (needBlend && doLineBlend && haveShallowLine) ? 0.25 : 0.00);
}
// select output pixel
vec4 res = mix(mix(mix(mix(dst[ 6], dst[ 7], step(0.25, f.x)),
mix(dst[ 8], dst[ 9], step(0.75, f.x)),
step(0.50, f.x)),
mix(mix(dst[ 5], dst[ 0], step(0.25, f.x)),
mix(dst[ 1], dst[10], step(0.75, f.x)),
step(0.50, f.x)),
step(0.25, f.y)),
mix(mix(mix(dst[ 4], dst[ 3], step(0.25, f.x)),
mix(dst[ 2], dst[11], step(0.75, f.x)),
step(0.50, f.x)),
mix(mix(dst[15], dst[14], step(0.25, f.x)),
mix(dst[13], dst[12], step(0.75, f.x)),
step(0.50, f.x)),
step(0.75, f.y)),
step(0.50, f.y));
return res;
}
uint applyScalingu(uvec2 origxy, uvec2 xy) {
return packUnorm4x8(applyScalingf(origxy, xy));
}