GPU: format and cast float

This commit is contained in:
Owain van Brakel
2020-01-19 00:34:34 +01:00
parent 952e8e8a10
commit aff45b6bbe
5 changed files with 281 additions and 304 deletions

View File

@@ -45,12 +45,13 @@ out vec4 FragColor;
void main() {
vec4 c;
if (samplingMode == SAMPLING_DEFAULT)
if (samplingMode == SAMPLING_DEFAULT) {
c = texture(tex, TexCoord);
else if (samplingMode == SAMPLING_CATROM || samplingMode == SAMPLING_MITCHELL)
} else if (samplingMode == SAMPLING_CATROM || samplingMode == SAMPLING_MITCHELL) {
c = textureCubic(tex, TexCoord, samplingMode);
else if (samplingMode == SAMPLING_XBR)
} else if (samplingMode == SAMPLING_XBR) {
c = textureXBR(tex, TexCoord, xbrTable, ceil(1.0 * targetDimensions.x / sourceDimensions.x));
}
FragColor = c;
}

View File

@@ -24,8 +24,7 @@
*/
// General case cubic filter
float cubic_custom(float x, float b, float c)
{
float cubic_custom(float x, float b, float c) {
/* A generalized cubic filter as described by Mitchell and Netravali is defined by the piecewise equation:
* if abs(x) < 1
* y = 1/6 * ( (12 - 9b - 6c) * abs(x)^3 + (-18 + 12b + 6c) * abs(x)^2 + (6 - 2b) )
@@ -40,17 +39,18 @@ float cubic_custom(float x, float b, float c)
float t2 = t * t; // t squared
float t3 = t * t * t; // t cubed
if (t < 1) // This part defines the [-1,1] region of the curve.
if (t < 1) { // This part defines the [-1,1] region of the curve.
return 1.0/6 * ((12 - 9 * b - 6 * c) * t3 + (-18 + 12 * b + 6 * c) * t2 + (6 - 2 * b));
else if (t < 2) // This part defines the [-2,-1] and [1,2] regions.
} else if (t < 2) { // This part defines the [-2,-1] and [1,2] regions.
return 1.0/6 * ((-1 * b - 6 * c) * t3 + (6 * b + 30 * c) * t2 + (-12 * b - 48 * c) * t + (8 * b + 24 * c));
else // Outside of [-2,2], the value is 0.
return 0;
}
// Outside of [-2,2], the value is 0.
return float(0);
}
// Cubic filter with Catmull-Rom parameters
float catmull_rom(float x)
{
float catmull_rom(float x) {
/*
* Generally favorable results in image upscaling are given by a cubic filter with the values b = 0 and c = 0.5.
* This is known as the Catmull-Rom filter, and it closely approximates Jinc upscaling with Lanczos input values.
@@ -63,16 +63,16 @@ float catmull_rom(float x)
float t2 = t * t;
float t3 = t * t * t;
if (t < 1)
if (t < 1) {
return 1.5 * t3 - 2.5 * t2 + 1;
else if (t < 2)
} else if (t < 2) {
return -0.5 * t3 + 2.5 * t2 - 4 * t + 2;
else
return 0;
}
float mitchell(float x)
{
return float(0);
}
float mitchell(float x) {
/*
* This is another cubic filter with less aggressive sharpening than Catmull-Rom, which some users may prefer.
* B = 1/3, C = 1/3.
@@ -82,12 +82,13 @@ float mitchell(float x)
float t2 = t * t;
float t3 = t * t * t;
if (t < 1)
if (t < 1) {
return 7.0/6 * t3 + -2 * t2 + 8.0/9;
else if (t < 2)
} else if (t < 2) {
return -7.0/18 * t3 + 2 * t2 - 10.0/3 * t + 16.0/9;
else
return 0;
}
return float(0);
}
#define CR_AR_STRENGTH 0.9
@@ -96,8 +97,7 @@ float mitchell(float x)
#define FLT_MIN 1.175494351e-38
// Calculates the distance between two points
float d(vec2 pt1, vec2 pt2)
{
float d(vec2 pt1, vec2 pt2) {
vec2 v = pt2 - pt1;
return sqrt(dot(v,v));
}
@@ -117,15 +117,12 @@ vec4 textureCubic(sampler2D sampler, vec2 texCoords, int mode){
vec4 c;
if (mode == SAMPLING_CATROM)
{
if (mode == SAMPLING_CATROM) {
// catrom benefits from anti-ringing, which requires knowledge of the minimum and maximum samples in the kernel
vec4 min_sample = vec4(FLT_MAX);
vec4 max_sample = vec4(FLT_MIN);
for (int m = -1; m <= 2; m++)
{
for (int n = -1; n <= 2; n++)
{
for (int m = -1; m <= 2; m++) {
for (int n = -1; n <= 2; n++) {
// get the raw texel, bypassing any other filters
vec4 vecData = texelFetch(sampler, texelCoords + ivec2(m, n), 0);
@@ -150,13 +147,9 @@ vec4 textureCubic(sampler2D sampler, vec2 texCoords, int mode){
c = clamp(c, min_sample, max_sample);
// mix according to anti-ringing strength
c = mix(aux, c, CR_AR_STRENGTH);
}
else if (mode == SAMPLING_MITCHELL)
{
for (int m = -1; m <= 2; m++)
{
for (int n = -1; n <= 2; n++)
{
} else if (mode == SAMPLING_MITCHELL) {
for (int m = -1; m <= 2; m++) {
for (int n = -1; n <= 2; n++) {
// get the raw texel, bypassing any other filters
vec4 vecData = texelFetch(sampler, texelCoords + ivec2(m, n), 0);

View File

@@ -24,8 +24,7 @@
Incorporates some of the ideas from SABR shader. Thanks to Joshua Street.
*/
struct XBRTable
{
struct XBRTable {
vec2 texCoord;
vec4 t1;
vec4 t2;

View File

@@ -62,53 +62,44 @@ const vec4 Ci = vec4(0.25, 0.25, 0.25, 0.25);
const vec3 Y = vec3(0.2126, 0.7152, 0.0722); // rec.709 luma weights
// Difference between vector components.
vec4 df(vec4 A, vec4 B)
{
vec4 df(vec4 A, vec4 B) {
return vec4(abs(A-B));
}
// Compare two vectors and return their components are different.
vec4 diff(vec4 A, vec4 B)
{
vec4 diff(vec4 A, vec4 B) {
return vec4(notEqual(A, B));
}
// Determine if two vector components are equal based on a threshold.
vec4 eq(vec4 A, vec4 B)
{
vec4 eq(vec4 A, vec4 B) {
return (step(df(A, B), vec4(XBR_EQ_THRESHOLD)));
}
// Determine if two vector components are NOT equal based on a threshold.
vec4 neq(vec4 A, vec4 B)
{
vec4 neq(vec4 A, vec4 B) {
return (vec4(1.0, 1.0, 1.0, 1.0) - eq(A, B));
}
// Weighted distance.
vec4 wd(vec4 a, vec4 b, vec4 c, vec4 d, vec4 e, vec4 f, vec4 g, vec4 h)
{
vec4 wd(vec4 a, vec4 b, vec4 c, vec4 d, vec4 e, vec4 f, vec4 g, vec4 h) {
return (df(a,b) + df(a,c) + df(d,e) + df(d,f) + 4.0*df(g,h));
}
vec4 weighted_distance(vec4 a, vec4 b, vec4 c, vec4 d, vec4 e, vec4 f, vec4 g, vec4 h, vec4 i, vec4 j, vec4 k, vec4 l)
{
vec4 weighted_distance(vec4 a, vec4 b, vec4 c, vec4 d, vec4 e, vec4 f, vec4 g, vec4 h, vec4 i, vec4 j, vec4 k, vec4 l) {
return (df(a,b) + df(a,c) + df(d,e) + df(d,f) + df(i,j) + df(k,l) + 2.0*df(g,h));
}
float c_df(vec3 c1, vec3 c2)
{
float c_df(vec3 c1, vec3 c2) {
vec3 df = abs(c1 - c2);
return df.r + df.g + df.b;
}
#include scale/xbr_lv2_common.glsl
// xBR-level2 upscaler. Level 2 means it detects edges in 2 directions, instead of just 1 in the most basic form of the algorithm.
// This improves quality by a good bit without adding too much complexity compared to available level-3 and level-4 algorithms.
vec4 textureXBR(sampler2D image, vec2 texCoord, XBRTable t, float scale)
{
vec4 textureXBR(sampler2D image, vec2 texCoord, XBRTable t, float scale) {
vec4 delta = vec4(1.0/scale, 1.0/scale, 1.0/scale, 1.0/scale);
vec4 delta_l = vec4(0.5/scale, 1.0/scale, 0.5/scale, 1.0/scale);
vec4 delta_u = delta_l.yxwz;
@@ -161,14 +152,11 @@ vec4 textureXBR(sampler2D image, vec2 texCoord, XBRTable t, float scale)
float y_weight = XBR_Y_WEIGHT;
if (small_details < 0.5)
{
if (small_details < 0.5) {
i4 = vec4(dot(I4.xyz,rgbw), dot(C1.xyz,rgbw), dot(A0.xyz,rgbw), dot(G5.xyz,rgbw));
i5 = vec4(dot(I5.xyz,rgbw), dot(C4.xyz,rgbw), dot(A1.xyz,rgbw), dot(G0.xyz,rgbw));
h5 = vec4(dot(H5.xyz,rgbw), dot(F4.xyz,rgbw), dot(B1.xyz,rgbw), dot(D0.xyz,rgbw));
}
else
{
} else {
i4 = mul(mat4x3(I4.xyz, C1.xyz, A0.xyz, G5.xyz), y_weight * Y);
i5 = mul(mat4x3(I5.xyz, C4.xyz, A1.xyz, G0.xyz), y_weight * Y);
h5 = mul(mat4x3(H5.xyz, F4.xyz, B1.xyz, D0.xyz), y_weight * Y);
@@ -203,13 +191,10 @@ vec4 textureXBR(sampler2D image, vec2 texCoord, XBRTable t, float scale)
vec4 fx60 = clamp((fx_u + delta_u -Cy )/(2.0*delta_u), 0.0, 1.0);
vec4 wd1, wd2;
if (small_details < 0.5)
{
if (small_details < 0.5) {
wd1 = wd( e, c, g, i, h5, f4, h, f);
wd2 = wd( h, d, i5, f, i4, b, e, i);
}
else
{
} else {
wd1 = weighted_distance( e, c, g, i, f4, h5, h, f, b, d, i4, i5);
wd2 = weighted_distance( h, d, i5, f, b, i4, e, i, g, h5, c, f4);
}

View File

@@ -26,8 +26,7 @@
#include scale/xbr_lv2_common.glsl
XBRTable xbr_vert(vec2 texCoord, ivec2 sourceDimensions)
{
XBRTable xbr_vert(vec2 texCoord, ivec2 sourceDimensions) {
float dx = (1.0/sourceDimensions.x);
float dy = (1.0/sourceDimensions.y);