Skip to content

Commit c9209e5

Browse files
TzugimaaBlinue
Tzugimaa
and
Blinue
authored
Update LCAS.hlsl (Blinue#324)
* Update LCAS.hlsl Added masking and reduced sharpness so it would not over sharpen by default. * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Further optimize parallelism and port for current mega/compute * Update LCAS.hlsl * Changed to 16 16 block The result seems to be relatively equal to 8 8 block. May be I have done something wrong. * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Update LCAS.hlsl * Changed LCAS sampling pattern Co-authored-by: 刘旭 <blinue@outlook.com>
1 parent ba95aec commit c9209e5

File tree

1 file changed

+68
-46
lines changed

1 file changed

+68
-46
lines changed

‎Effects/LCAS.hlsl‎

Lines changed: 68 additions & 46 deletions
Original file line numberDiff line numberDiff line change
@@ -1,15 +1,13 @@
11
// This is a combination of linear interpolation and light version of cas
22

33
//!MAGPIE EFFECT
4-
//!VERSION 1
4+
//!VERSION 2
55

6-
//!CONSTANT
7-
//!VALUE INPUT_PT_X
8-
float inputPtX;
9-
10-
//!CONSTANT
11-
//!VALUE INPUT_PT_Y
12-
float inputPtY;
6+
//!PARAMETER
7+
//!DEFAULT 0.5
8+
//!MIN 0
9+
//!MAX 1
10+
float sharpness;
1311

1412
//!TEXTURE
1513
Texture2D INPUT;
@@ -18,55 +16,79 @@ Texture2D INPUT;
1816
//!FILTER LINEAR
1917
SamplerState sam;
2018

21-
//!CONSTANT
22-
//!DEFAULT 0.33
23-
//!MIN 0
24-
//!MAX 1
25-
float sharpness;
19+
//!PASS 1
20+
//!IN INPUT
21+
//!BLOCK_SIZE 16
22+
//!NUM_THREADS 64
2623

27-
//!CONSTANT
28-
//!DEFAULT 0.2
29-
//!MIN 0
30-
//!MAX 1
31-
float threshold;
24+
float3 LCAS(uint2 ip, float peak) {
3225

33-
//!PASS 1
34-
//!BIND INPUT
26+
float2 pos = (ip + 0.5f) * GetOutputPt();
27+
float2 inputPt = GetInputPt();
3528

36-
float4 Pass1(float2 pos) {
37-
// fetch a 4 neighborhood pixels around the pixel 'e',
38-
// b
29+
// fetch a 3x3 neighborhood around the pixel 'e',
30+
// a b c
3931
// d(e)f
40-
// h
41-
float3 b = INPUT.Sample(sam, pos + float2(0, -inputPtY)).rgb;
42-
float3 d = INPUT.Sample(sam, pos + float2(-inputPtX, 0)).rgb;
43-
float3 e = INPUT.Sample(sam, pos).rgb;
44-
float3 f = INPUT.Sample(sam, pos + float2(inputPtX, 0)).rgb;
45-
float3 h = INPUT.Sample(sam, pos + float2(0, inputPtY)).rgb;
46-
47-
// Edge checker
48-
float edge = length(abs(d - f) + abs(b - h));
32+
// g h i
33+
float3 a = INPUT.SampleLevel(sam, pos + float2(-inputPt.x, -inputPt.y), 0).rgb;
34+
float3 b = INPUT.SampleLevel(sam, pos + float2(0, -inputPt.y), 0).rgb;
35+
float3 c = INPUT.SampleLevel(sam, pos + float2(inputPt.x, -inputPt.y), 0).rgb;
36+
float3 d = INPUT.SampleLevel(sam, pos + float2(-inputPt.x, 0), 0).rgb;
37+
float3 f = INPUT.SampleLevel(sam, pos + float2(inputPt.x, 0), 0).rgb;
38+
float3 g = INPUT.SampleLevel(sam, pos + float2(-inputPt.x, inputPt.y), 0).rgb;
39+
float3 h = INPUT.SampleLevel(sam, pos + float2(0, inputPt.y), 0).rgb;
40+
float3 i = INPUT.SampleLevel(sam, pos + float2(inputPt.x, inputPt.y), 0).rgb;
41+
42+
float3 x = a + c + g + i;
43+
float3 y = b + d + f + h;
4944

45+
float3 e = INPUT.SampleLevel(sam, pos + float2(inputPt.x * 0.25, inputPt.y * 0.5), 0).rgb;
46+
e += INPUT.SampleLevel(sam, pos + float2(-inputPt.x * 0.25, -inputPt.y * 0.5), 0).rgb;
47+
e += INPUT.SampleLevel(sam, pos + float2(inputPt.x * 0.5, -inputPt.y * 0.25), 0).rgb;
48+
e += INPUT.SampleLevel(sam, pos + float2(-inputPt.x * 0.5, inputPt.y * 0.25), 0).rgb;
49+
e /= 4;
50+
5051
// Soft min and max.
51-
// b
52+
// a b c
5253
// d e f
53-
// h
54-
55-
float3 mnRGB = min(min(min(d, e), min(f, b)), h);
54+
// g h i
55+
float3 mnRGB = min(min(min(min(d, e), min(f, b)), h), min(min(a, i), min(c, g)));
5656

57-
float3 mxRGB = max(max(max(d, e), max(f, b)), h);
57+
float3 mxRGB = max(max(max(max(d, e), max(f, b)), h), max(max(a, i), max(c, g)));
5858

5959
// Shaping amount of sharpening.
60-
float3 wRGB = sqrt(min(mnRGB, 1.0 - mxRGB) / mxRGB) * lerp(- 0.125, - 0.2, sharpness);
60+
float3 wRGB = sqrt(min(mnRGB, 1.0 - mxRGB) / mxRGB) * peak;
6161

6262
// Filter shape.
63-
// w
63+
// w w w
6464
// w 1 w
65-
// w
66-
// If is edge
67-
if(edge >= threshold)
68-
return float4((((b + d + f + h) * wRGB + (e * 2 - (b + d + f + h) * 0.25)) / (1.0 + 4.0 * wRGB)).rgb, 1);
69-
else
70-
return float4((((b + d + f + h) * wRGB + e) / (1.0 + 4.0 * wRGB)).rgb, 1);
71-
// If is not edge
65+
// w w w
66+
float3 color = ((x + y) * wRGB + (e * 5.0 - (x + y * 2.0 + e * 4.0) / 4.0)) / (1.0 + 8.0 * wRGB);
67+
return (color + clamp(color, mnRGB, mxRGB) * 4.0) / 5.0;
68+
}
69+
70+
void Pass1(uint2 blockStart, uint3 threadId) {
71+
uint2 gxy = blockStart + Rmp8x8(threadId.x);
72+
if (!CheckViewport(gxy)) {
73+
return;
74+
}
75+
76+
const float peak = lerp(0, -0.1111111111111111, sharpness);
77+
78+
WriteToOutput(gxy, LCAS(gxy, peak));
79+
80+
gxy.x += 8u;
81+
if (CheckViewport(gxy)) {
82+
WriteToOutput(gxy, LCAS(gxy, peak));
83+
}
84+
85+
gxy.y += 8u;
86+
if (CheckViewport(gxy)) {
87+
WriteToOutput(gxy, LCAS(gxy, peak));
88+
}
89+
90+
gxy.x -= 8u;
91+
if (CheckViewport(gxy)) {
92+
WriteToOutput(gxy, LCAS(gxy, peak));
93+
}
7294
}

0 commit comments

Comments
 (0)