绝区零角色后处理NapBloom

完成图前置

我在UE中实现的bloom

Bloom

1.降采样与提取亮度

原始分辨率为2448×1368，降采样至344×192的同时根据阈值参数拿到发光颜色。

根据四个方向偏移一像素采样四次，求和平均（降采样），然后减去一个用户阈值

//降采样
float4 a = tex.Sample(linear, offsetA);
float4 b = tex.Sample(linear, offsetB);
float4 c = tex.Sample(linear, offsetC);
float4 d = tex.Sample(linear, offsetD);

float color = (a + b + c + d) * 0.25;

//减去用户参数-阈值
color.rgb = max(color.rgb - Threshould;, 0) * Intensity;

return color;

这里的Threshould为0.44922

2 在344×192上做横纵两次模糊

模糊shader为：MultipleGaussPassFilter

分别采样六次，并乘以一个系数后相加，六个系数相加后正好为一，应该就是卷积核：

float4 a = t.Sample(s, User * -2.2226281 + uv) * 0.1334844;
float4 b = t.Sample(s, User * -4.0952849 + uv) * 0.0057047;
float4 c = t.Sample(s, User * -0.437803 + uv) * 0.501892;
float4 d = t.Sample(s, User * 1.320767 + uv) * 0.3234969;
float4 e = t.Sample(s, User * 3.147974 + uv) * 0.0348785;
float4 f = t.Sample(s, User * 5 + uv) * 0.0005436;
return a + b + c + d + e + f;

第一次横向模糊的User值为(0.0029, 0)，这个值正好为纹理宽度的倒数(1/344)，代表一像素。在代码中的User * -2.2226281偏移量就是就是按像素偏移。

用第一次横向模糊的结果在进行第二次纵向模糊，User值为(0. 0.0052)，同样是高度(192)的一像素

3 多次降采样与模糊

将344×192降采样三次至153×161纹理的大小，应该是为了节省RT申请和切换。要先在同一张RT上算完三张图的横向模糊，在切换RT计算纵向模糊。

• Pass0
  • 降采样大图
  • 降采样中图
  • 降采样小图
• Pass1
  • 横向模糊大图
  • 横向模糊中图
  • 横向模糊小图
• Pass2
  • 纵向模糊大图
  • 纵向模糊中图
  • 纵向模糊小图

降采样算法同上一次降采样一致，采样四次后平均。

但两次模糊算法的采样次数和卷积核有所不同。并且分辨率越低，模糊采样次数越多，扩散范围越大（卷积核越大）。

大图模糊的9次采样

float a = t.Sample(s, uv + User * -5.2274981) * 0.0151298;
float b = t.Sample(s, uv + User * -7.1588202) * 0.0009649;
float c = t.Sample(s, uv + User * -3.3147621) * 0.1009583;
float d = t.Sample(s, uv + User * -1.417412) * 0.2889;
float e = t.Sample(s, uv + User * 0.4722446) * 0.3564036;
float f = t.Sample(s, uv + User * 2.364548) * 0.1897708;
float g = t.Sample(s, uv + User * 4.268898) * 0.0434656;
float h = t.Sample(s, uv + User * 6.1908078) * 0.0042536;
float i = t.Sample(s, uv + User * 8) * 0.0001532;
return a + b + c + d + e + f + g + h + i;

横向模糊User为(0.00653, 0)，纵向模糊User为(0, 0.00616)

中图的16次采样

float4 blur = 0;
blur += t.Sample(s, uv + User * -12.2933798) * 0.0009471;
blur += t.Sample(s, uv + User * -14.26509) * 0.0001463;
blur += t.Sample(s, uv + User * -10.3233604) * 0.0046463;
blur += t.Sample(s, uv + User * -8.3548632) * 0.0172796;
blur += t.Sample(s, uv + User * -6.3876772) * 0.0487266;
blur += t.Sample(s, uv + User * -4.4215422) * 0.1042022;
blur += t.Sample(s, uv + User * -2.456162) * 0.1690129;
blur += t.Sample(s, uv + User * -0.4912108) * 0.207937;
blur += t.Sample(s, uv + User * 1.473654) * 0.1940565;
blur += t.Sample(s, uv + User * 3.4387779) * 0.1373738;
blur += t.Sample(s, uv + User * 5.4044962) * 0.0737621;
blur += t.Sample(s, uv + User * 7.3711209) * 0.0300379;
blur += t.Sample(s, uv + User * 9.338933) * 0.0092757;
blur += t.Sample(s, uv + User * 11.3081703) * 0.0021717;
blur += t.Sample(s, uv + User * 13.2790203) * 0.0003854;
blur += t.Sample(s, uv + User * 15) * 3.8788519e-005;
return blur;

横向模糊User为(0.00653, 0)，纵向模糊User为(0, 0.00616)

小图的20次采样

float4 blur = 0;
blur += t.Sample(s, uv + User * -16.3224392) * 0.0003934;
blur += t.Sample(s, uv + User * -18.3031006) * 8.2805367e-005;
blur += t.Sample(s, uv + User * -14.3424101) * 0.0015638;
blur += t.Sample(s, uv + User * -12.3629599) * 0.0052015;
blur += t.Sample(s, uv + User * -10.3840103) * 0.0144784;
blur += t.Sample(s, uv + User * -8.4055147) * 0.033726;
blur += t.Sample(s, uv + User * -6.4273849) * 0.0657471;
blur += t.Sample(s, uv + User * -4.449542) * 0.1072673;
blur += t.Sample(s, uv + User * -2.4719019) * 0.1464697;
blur += t.Sample(s, uv + User * -0.4943747) * 0.1673879;
blur += t.Sample(s, uv + User * 1.48313) * 0.1601027;
blur += t.Sample(s, uv + User * 3.4607019) * 0.1281654;
blur += t.Sample(s, uv + User * 5.4384332) * 0.0858689;
blur += t.Sample(s, uv + User * 7.416409) * 0.0481489;
blur += t.Sample(s, uv + User * 9.3947134) * 0.0225949;
blur += t.Sample(s, uv + User * 11.3734198) * 0.0088735;
blur += t.Sample(s, uv + User * 13.3526201) * 0.0029162;
blur += t.Sample(s, uv + User * 15.3323498) * 0.000802;
blur += t.Sample(s, uv + User * 17.3126907) * 0.0001846;
blur += t.Sample(s, uv + User * 19) * 2.509824e-005;

横向模糊User为(0.00653, 0)，纵向模糊User为(0, 0.00616)

将以上四张图使用的卷积核做成数组，可以用循环来采样

static const float2 CORE0[6] = {
    float2(-2.2226281, 0.1334844),
    float2(-4.0952849, 0.0057047),
    float2(-0.437803, 0.501892),
    float2(1.320767, 0.3234969),
    float2(3.147974, 0.0348785),
    float2(5, 0.0005436),
};
static const float2 CORE1[9] = {
    float2(-5.2274981, 0.0151298),
    float2(-7.1588202, 0.0009649),
    float2(-3.3147621, 0.1009583),
    float2(-1.417412, 0.2889),
    float2(0.4722446, 0.3564036),
    float2(2.364548, 0.1897708),
    float2(4.268898, 0.0434656),
    float2(6.1908078, 0.0042536),
    float2(8, 0.0001532),
};
static const float2 CORE2[16] = {
    float2(-12.2933798, 0.0009471),
    float2(-14.26509, 0.0001463),
    float2(-10.3233604, 0.0046463),
    float2(-8.3548632, 0.0172796),
    float2(-6.3876772, 0.0487266),
    float2(-4.4215422, 0.1042022),
    float2(-2.456162, 0.1690129),
    float2(-0.4912108, 0.207937),
    float2(1.473654, 0.1940565),
    float2(3.4387779, 0.1373738),
    float2(5.4044962, 0.0737621),
    float2(7.3711209, 0.0300379),
    float2(9.338933, 0.0092757),
    float2(11.3081703, 0.0021717),
    float2(13.2790203, 0.0003854),
    float2(15, 3.8788519e-005),
};
static const float2 CORE3[20] = {
    float2(-16.3224392, 0.0003934),
    float2(-18.3031006, 8.2805367e-005),
    float2(-14.3424101, 0.0015638),
    float2(-12.3629599, 0.0052015),
    float2(-10.3840103, 0.0144784),
    float2(-8.4055147, 0.033726),
    float2(-6.4273849, 0.0657471),
    float2(-4.449542, 0.1072673),
    float2(-2.4719019, 0.1464697),
    float2(-0.4943747, 0.1673879),
    float2(1.48313, 0.1601027),
    float2(3.4607019, 0.1281654),
    float2(5.4384332, 0.0858689),
    float2(7.416409, 0.0481489),
    float2(9.3947134, 0.0225949),
    float2(11.3734198, 0.0088735),
    float2(13.3526201, 0.0029162),
    float2(15.3323498, 0.000802),
    float2(17.3126907, 0.0001846),
    float2(19, 2.509824e-005),
};

4 四层bloom合并

合并就是把四张图乘以四个用户的强度系数后给加起来，系数的结果正好为1：

• 第0张系数：0.29688
• 第1张系数：0.29688
• 第2张系数：0.25781
• 第3张系数：0.14844

5 合并回原图像

过亮抑制，计算出bloom部分

const float threshold = 0.3;
float3 bloomSup = pow(abs(bloomTex.rgb) / threshold, 1.0/3.0) - (1-threshold);
float3 bloom = lerp(bloomTex.rgb, bloomSup , bloomTex.rgb >= threshold);

bloom *= Intensity;

其中Intensity参数为用户亮度系数，值为1.3

然后加起来获得最终输出

out.rgb = bloom + orignal.rgb;
out.a = orignal.a + (bloom.r + bloom.g + bloom.b) * 0.3333;