优化重型片段着色器的性能

我需要帮助优化以下一组着色器：

顶点：

    precision mediump float;

uniform vec2 rubyTextureSize;

attribute vec4 vPosition;
attribute vec2 a_TexCoordinate;

varying vec2 tc;

void main() {
    gl_Position = vPosition;

    tc = a_TexCoordinate;
}

分段：

precision mediump float;

/*
 Uniforms
 - rubyTexture: texture sampler
 - rubyTextureSize: size of the texture before rendering
 */

uniform sampler2D rubyTexture;
uniform vec2 rubyTextureSize;
uniform vec2 rubyTextureFract;

/*
 Varying attributes
 - tc: coordinate of the texel being processed
 - xyp_[]_[]_[]: a packed coordinate for 3 areas within the texture
 */

varying vec2 tc;

/*
 Constants
 */
/*
 Inequation coefficients for interpolation
 Equations are in the form: Ay + Bx = C
 45, 30, and 60 denote the angle from x each line the cooeficient variable set builds
 */
const vec4 Ai = vec4(1.0, -1.0, -1.0, 1.0);
const vec4 B45 = vec4(1.0, 1.0, -1.0, -1.0);
const vec4 C45 = vec4(1.5, 0.5, -0.5, 0.5);
const vec4 B30 = vec4(0.5, 2.0, -0.5, -2.0);
const vec4 C30 = vec4(1.0, 1.0, -0.5, 0.0);
const vec4 B60 = vec4(2.0, 0.5, -2.0, -0.5);
const vec4 C60 = vec4(2.0, 0.0, -1.0, 0.5);

const vec4 M45 = vec4(0.4, 0.4, 0.4, 0.4);
const vec4 M30 = vec4(0.2, 0.4, 0.2, 0.4);
const vec4 M60 = M30.yxwz;
const vec4 Mshift = vec4(0.2);

// Coefficient for weighted edge detection
const float coef = 2.0;
// Threshold for if luminance values are "equal"
const vec4 threshold = vec4(0.32);

// Conversion from RGB to Luminance (from GIMP)
const vec3 lum = vec3(0.21, 0.72, 0.07);

// Performs same logic operation as && for vectors
bvec4 _and_(bvec4 A, bvec4 B) {
    return bvec4(A.x && B.x, A.y && B.y, A.z && B.z, A.w && B.w);
}

// Performs same logic operation as || for vectors
bvec4 _or_(bvec4 A, bvec4 B) {
    return bvec4(A.x || B.x, A.y || B.y, A.z || B.z, A.w || B.w);
}

// Converts 4 3-color vectors into 1 4-value luminance vector
vec4 lum_to(vec3 v0, vec3 v1, vec3 v2, vec3 v3) {
    //    return vec4(dot(lum, v0), dot(lum, v1), dot(lum, v2), dot(lum, v3));

    return mat4(v0.x, v1.x, v2.x, v3.x, v0.y, v1.y, v2.y, v3.y, v0.z, v1.z,
            v2.z, v3.z, 0.0, 0.0, 0.0, 0.0) * vec4(lum, 0.0);
}

// Gets the difference between 2 4-value luminance vectors
vec4 lum_df(vec4 A, vec4 B) {
    return abs(A - B);
}

// Determines if 2 4-value luminance vectors are "equal" based on threshold
bvec4 lum_eq(vec4 A, vec4 B) {
    return lessThan(lum_df(A, B), threshold);
}

vec4 lum_wd(vec4 a, vec4 b, vec4 c, vec4 d, vec4 e, vec4 f, vec4 g, vec4 h) {
    return lum_df(a, b) + lum_df(a, c) + lum_df(d, e) + lum_df(d, f)
            + 4.0 * lum_df(g, h);
}

// Gets the difference between 2 3-value rgb colors
float c_df(vec3 c1, vec3 c2) {
    vec3 df = abs(c1 - c2);
    return df.r + df.g + df.b;
}

void main() {

    /*
     Mask for algorhithm
     +-----+-----+-----+-----+-----+
     |     |  1  |  2  |  3  |     |
     +-----+-----+-----+-----+-----+
     |  5  |  6  |  7  |  8  |  9  |
     +-----+-----+-----+-----+-----+
     | 10  | 11  | 12  | 13  | 14  |
     +-----+-----+-----+-----+-----+
     | 15  | 16  | 17  | 18  | 19  |
     +-----+-----+-----+-----+-----+
     |     | 21  | 22  | 23  |     |
     +-----+-----+-----+-----+-----+
     */

    float x = rubyTextureFract.x;
    float y = rubyTextureFract.y;

    vec4 xyp_1_2_3 = tc.xxxy + vec4(-x, 0.0, x, -2.0 * y);
    vec4 xyp_6_7_8 = tc.xxxy + vec4(-x, 0.0, x, -y);
    vec4 xyp_11_12_13 = tc.xxxy + vec4(-x, 0.0, x, 0.0);
    vec4 xyp_16_17_18 = tc.xxxy + vec4(-x, 0.0, x, y);
    vec4 xyp_21_22_23 = tc.xxxy + vec4(-x, 0.0, x, 2.0 * y);
    vec4 xyp_5_10_15 = tc.xyyy + vec4(-2.0 * x, -y, 0.0, y);
    vec4 xyp_9_14_9 = tc.xyyy + vec4(2.0 * x, -y, 0.0, y);

    // Get mask values by performing texture lookup with the uniform sampler
    vec3 P1 = texture2D(rubyTexture, xyp_1_2_3.xw).rgb;
    vec3 P2 = texture2D(rubyTexture, xyp_1_2_3.yw).rgb;
    vec3 P3 = texture2D(rubyTexture, xyp_1_2_3.zw).rgb;

    vec3 P6 = texture2D(rubyTexture, xyp_6_7_8.xw).rgb;
    vec3 P7 = texture2D(rubyTexture, xyp_6_7_8.yw).rgb;
    vec3 P8 = texture2D(rubyTexture, xyp_6_7_8.zw).rgb;

    vec3 P11 = texture2D(rubyTexture, xyp_11_12_13.xw).rgb;
    vec3 P12 = texture2D(rubyTexture, xyp_11_12_13.yw).rgb;
    vec3 P13 = texture2D(rubyTexture, xyp_11_12_13.zw).rgb;

    vec3 P16 = texture2D(rubyTexture, xyp_16_17_18.xw).rgb;
    vec3 P17 = texture2D(rubyTexture, xyp_16_17_18.yw).rgb;
    vec3 P18 = texture2D(rubyTexture, xyp_16_17_18.zw).rgb;

    vec3 P21 = texture2D(rubyTexture, xyp_21_22_23.xw).rgb;
    vec3 P22 = texture2D(rubyTexture, xyp_21_22_23.yw).rgb;
    vec3 P23 = texture2D(rubyTexture, xyp_21_22_23.zw).rgb;

    vec3 P5 = texture2D(rubyTexture, xyp_5_10_15.xy).rgb;
    vec3 P10 = texture2D(rubyTexture, xyp_5_10_15.xz).rgb;
    vec3 P15 = texture2D(rubyTexture, xyp_5_10_15.xw).rgb;

    vec3 P9 = texture2D(rubyTexture, xyp_9_14_9.xy).rgb;
    vec3 P14 = texture2D(rubyTexture, xyp_9_14_9.xz).rgb;
    vec3 P19 = texture2D(rubyTexture, xyp_9_14_9.xw).rgb;

    // Store luminance values of each point in groups of 4
    // so that we may operate on all four corners at once
    vec4 p7 = lum_to(P7, P11, P17, P13);
    vec4 p8 = lum_to(P8, P6, P16, P18);
    vec4 p11 = p7.yzwx; // P11, P17, P13, P7
    vec4 p12 = lum_to(P12, P12, P12, P12);
    vec4 p13 = p7.wxyz; // P13, P7,  P11, P17
    vec4 p14 = lum_to(P14, P2, P10, P22);
    vec4 p16 = p8.zwxy; // P16, P18, P8,  P6
    vec4 p17 = p7.zwxy; // P17, P13, P7,  P11
    vec4 p18 = p8.wxyz; // P18, P8,  P6,  P16
    vec4 p19 = lum_to(P19, P3, P5, P21);
    vec4 p22 = p14.wxyz; // P22, P14, P2,  P10
    vec4 p23 = lum_to(P23, P9, P1, P15);

    // Scale current texel coordinate to [0..1]
    vec2 fp = fract(tc * rubyTextureSize);

    // Determine amount of "smoothing" or mixing that could be done on texel corners
    vec4 AiMulFpy = Ai * fp.y;
    vec4 B45MulFpx = B45 * fp.x;
    vec4 ma45 = smoothstep(C45 - M45, C45 + M45, AiMulFpy + B45MulFpx);
    vec4 ma30 = smoothstep(C30 - M30, C30 + M30, AiMulFpy + B30 * fp.x);
    vec4 ma60 = smoothstep(C60 - M60, C60 + M60, AiMulFpy + B60 * fp.x);
    vec4 marn = smoothstep(C45 - M45 + Mshift, C45 + M45 + Mshift,
            AiMulFpy + B45MulFpx);

    // Perform edge weight calculations
    vec4 e45 = lum_wd(p12, p8, p16, p18, p22, p14, p17, p13);
    vec4 econt = lum_wd(p17, p11, p23, p13, p7, p19, p12, p18);
    vec4 e30 = lum_df(p13, p16);
    vec4 e60 = lum_df(p8, p17);

    // Calculate rule results for interpolation
    bvec4 r45_1 = _and_(notEqual(p12, p13), notEqual(p12, p17));
    bvec4 r45_2 = _and_(not (lum_eq(p13, p7)), not (lum_eq(p13, p8)));
    bvec4 r45_3 = _and_(not (lum_eq(p17, p11)), not (lum_eq(p17, p16)));
    bvec4 r45_4_1 = _and_(not (lum_eq(p13, p14)), not (lum_eq(p13, p19)));
    bvec4 r45_4_2 = _and_(not (lum_eq(p17, p22)), not (lum_eq(p17, p23)));
    bvec4 r45_4 = _and_(lum_eq(p12, p18), _or_(r45_4_1, r45_4_2));
    bvec4 r45_5 = _or_(lum_eq(p12, p16), lum_eq(p12, p8));
    bvec4 r45 = _and_(r45_1, _or_(_or_(_or_(r45_2, r45_3), r45_4), r45_5));
    bvec4 r30 = _and_(notEqual(p12, p16), notEqual(p11, p16));
    bvec4 r60 = _and_(notEqual(p12, p8), notEqual(p7, p8));

    // Combine rules with edge weights
    bvec4 edr45 = _and_(lessThan(e45, econt), r45);
    bvec4 edrrn = lessThanEqual(e45, econt);
    bvec4 edr30 = _and_(lessThanEqual(coef * e30, e60), r30);
    bvec4 edr60 = _and_(lessThanEqual(coef * e60, e30), r60);

    // Finalize interpolation rules and cast to float (0.0 for false, 1.0 for true)
    vec4 final45 = vec4(_and_(_and_(not (edr30), not (edr60)), edr45));
    vec4 final30 = vec4(_and_(_and_(edr45, not (edr60)), edr30));
    vec4 final60 = vec4(_and_(_and_(edr45, not (edr30)), edr60));
    vec4 final36 = vec4(_and_(_and_(edr60, edr30), edr45));
    vec4 finalrn = vec4(_and_(not (edr45), edrrn));

    // Determine the color to mix with for each corner
    vec4 px = step(lum_df(p12, p17), lum_df(p12, p13));

    // Determine the mix amounts by combining the final rule result and corresponding
    // mix amount for the rule in each corner
    vec4 mac = final36 * max(ma30, ma60) + final30 * ma30 + final60 * ma60
            + final45 * ma45 + finalrn * marn;

    /*
     Calculate the resulting color by traversing clockwise and counter-clockwise around
     the corners of the texel

     Finally choose the result that has the largest difference from the texel's original
     color
     */
    vec3 res1 = P12;
    res1 = mix(res1, mix(P13, P17, px.x), mac.x);
    res1 = mix(res1, mix(P7, P13, px.y), mac.y);
    res1 = mix(res1, mix(P11, P7, px.z), mac.z);
    res1 = mix(res1, mix(P17, P11, px.w), mac.w);

    vec3 res2 = P12;
    res2 = mix(res2, mix(P17, P11, px.w), mac.w);
    res2 = mix(res2, mix(P11, P7, px.z), mac.z);
    res2 = mix(res2, mix(P7, P13, px.y), mac.y);
    res2 = mix(res2, mix(P13, P17, px.x), mac.x);

    gl_FragColor = vec4(mix(res1, res2, step(c_df(P12, res1), c_df(P12, res2))),
            1.0);
}

着色器接收2D纹理，并打算在高分辨率2D表面（设备屏幕）上漂亮地缩放它。如果重要的话，它是对SABR缩放算法的优化。

它已经可以工作，并且在非常高端的Android设备（例如LG Nexus 4）上可以正常运行，但是在功能较弱的设备上确实很慢。

对我而言真正重要的Android设备是带有Mali 400MP GPU的Samsung Galaxy S 2 \ 3-在此着色器上的表现非常糟糕。

到目前为止，我已经尝试过：

1. 消除差异（ARM的Mali指南的建议）-进行了较小的改进。
2. 用我自己的mix（）函数覆盖-不好。
3. 将浮点精度降低到lowp-没有任何改变。

我通过计算渲染时间（在eglSwapBuffers之前和之后）来衡量性能-这为我提供了非常线性和一致的性能衡量标准。

除此之外，我真的不知道在哪里看或可以在这里优化什么...

我知道这是一个繁重的算法，并且我不要求使用哪种替代缩放方法方面的建议-我尝试了很多，并且该算法可提供最佳的视觉效果。我希望以优化的方式使用完全相同的算法。

更新

1. 我发现，如果我使用恒定向量而不是从属向量来进行所有纹理提取，则性能将得到重大改善，因此这显然是一个很大的瓶颈-可能是由于缓存。但是，我仍然需要执行这些操作。我玩过至少一些带有vec2变化的抓取（没有任何混乱），但它并没有任何改善。我想知道哪种方法可以有效地轮询21个纹理像素。

2. 我发现，计算的主要部分是使用完全相同的纹理像素进行了多次-因为输出至少缩放了x2，并且我使用GL_NEAREST进行了轮询。至少有4个片段落在完全相同的纹理像素上。如果在高分辨率设备上缩放比例为x4，则有16个片段落在同一纹理像素上-这是一个很大的浪费。有什么方法可以执行额外的着色器遍历，以计算所有在多个片段之间不变的值？我曾考虑过渲染到额外的屏幕外纹理，但是我需要为每个纹理像素存储多个值，而不仅仅是一个。

更新

1. 我还注意到，CPU几乎是未使用的，而GPU是很大的瓶颈。关于在这种情况下如何利用某些CPU功能以及将逻辑从GPU传输到CPU的任何建议？

我想知道哪种方法可以有效地轮询21个纹理像素。

答案是有效的方法是不轮询21个纹理像素的方法。很显然，很抱歉，但移动设备可能没有必要的总线宽度来支持此类内核。您需要通过减少采样器中插入的纹理的大小来进行优化，以便缓存将覆盖更大的内核半径。

另外，您可能会忘记磁盘内核，而是使用垂直内核的两次通过算法，而另一种使用纯水平内核的算法，可以这样说，从“ 2D”传递到“ 1D”，从而大大减少了采样以及通过线性访问提高缓存性能。

由于应该在GPU内存中安排Z存储纹理，因此垂直读取不会影响缓存性能。cf http://en.wikipedia.org/wiki/Z-order_curve