I need help optimizing the following set of shaders:


    precision mediump float;

uniform vec2 rubyTextureSize;

attribute vec4 vPosition;
attribute vec2 a_TexCoordinate;

varying vec2 tc;

void main() {
    gl_Position = vPosition;

    tc = a_TexCoordinate;


precision mediump float;

 - 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;

 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
    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))),

The shaders receive a 2D texture and are meant to scale it beautifully across a high-res 2D surface (the device screen). It is an optimization of the SABR scaling algorithm in case it matters.

It already works, and performs OK on very high-end Android devices (like LG Nexus 4), but it is really slow on weaker devices.

The Android devices that really matter to me are Samsung Galaxy S 2 \ 3, with Mali 400MP GPU - which perform horribly with this shader.

So far I've tried:

  1. Eliminating varyings (advice from ARM's Mali guide) - did minor improvement.
  2. Overriding mix() functions with my own - did no good.
  3. reducing float precision to lowp - didn't change anything.

I measure performance by calculating render time (before and after eglSwapBuffers) - this gives me a very linear and consistent measurement of performance.

Beyond that, I don't really know where to look or what can be optimized here...

I know that this is a heavy algorithm, and I am not asking for advice on what alternate scaling methods to use - I've tried many and this algorithm gives the best visual result. I wish to use the exact same algorithm in an optimized way.


  1. I found that if I do all the texture fetches with a constant vector instead of dependent vectors I get a major performance improvement, so this is obviously a big bottleneck - probably because of the cache. However, I still need to do those fetches. I played with doing at least some of the fetches with vec2 varyings (without any swizzling) but it didn't improve anything. I wonder what might be a good way to efficiently poll 21 texels.

  2. I found that a major part of the calculations is being done multiple times with the exact same set of texels - because the output is scaled by at least x2, and I poll with GL_NEAREST. There at least 4 fragments that fall on exactly the same texels. If the scaling is x4 on a high-res device, there are 16 fragments that fall on the same texels - which is a big waste. Is there any way to perform an additional shader pass that will calculate all the values that don't change across multiple fragments? I thought about rendering to an additional off-screen texture, but I need to store multiple values per texel, not just one.


  1. I also noticed that the CPU is almost unused while the GPU is a big bottleneck. Any advice on how to leverage some CPU power and transfer logic from the GPU to the CPU in this situation?
  • 2
    \$\begingroup\$ You should never ever fetch the texture as a lookup. either pass the uv from the vertex so the pixelshader has the time to fetch the texture. \$\endgroup\$
    – Tordin
    May 14, 2013 at 14:02
  • \$\begingroup\$ Could you please explain? What do you mean by uv? \$\endgroup\$
    – SirKnigget
    May 14, 2013 at 14:26
  • 3
    \$\begingroup\$ Can you link to a description of the "SABR scaling algorithm"? Google doesn't find anything useful about it. By the way, a 21-texel filter (and quite math-heavy too) on a mobile GPU is just asking for trouble. I don't think you can realistically expect to make it run well without compromising quality somewhere. \$\endgroup\$ May 14, 2013 at 17:21
  • \$\begingroup\$ This gives the general idea: board.byuu.org/viewtopic.php?f=10&t=2248, although it's not the exact implementation I found. \$\endgroup\$
    – SirKnigget
    May 14, 2013 at 17:48
  • 2
    \$\begingroup\$ Regarding realistic expectations - it works great on high-end devices. I would expect to be able to fine tune what I have by about a 5x factor or similar and get it working on weaker devices. \$\endgroup\$
    – SirKnigget
    May 14, 2013 at 17:49

1 Answer 1


I wonder what might be a good way to efficiently poll 21 texels.

The answer is the efficient way is the way that is not polling 21 texels. Sorry to be obvious but mobile devices may not have the necessary bus width to support such kernels. You need to optimize by reducing the size of the texture plugged in the sampler so that caching will cover a larger kernel radius.

Also, you could forget about your disk kernel and use a two passes algorithm using a vertical kernel, and another one using a purely horizontal, this way you pass from "2D" to "1D" so to speak, and reduce drastically the number of samplings as well as improving cache performance thanks to linear access.

Vertical fetches should not affect cache performance thanks to the Z storage textures should be arranged in GPU memory. cf http://en.wikipedia.org/wiki/Z-order_curve


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .