Adding a simple mix() destroys the framerate

Question

I have a very simple shader and added linear fog via mix like this:

finalColor = mix(finalColor, vec3(0.5, 0.8, 0.95), vUVoutAndViewZ.z);

Note that the view Z distance is in the variable that also contains the UV coordinates, so the fog didn't even add a new interpolator.

Still, this one innocuous line brought the framerate on the OG Droid with the Power SGX chipset from 33fps to 22fps. Even the HTC Evo with the Adreno 200 GPU stays under 30fps. (The Adreno 205 is at a constant 60fps, but that thing is a beast).

The fragment shader itself is primitive (all values are hard-coded since it's a test shader):

precision mediump float;
varying mediump vec3 vUVoutAndViewZ;
varying lowp vec3 vNormalOut;

uniform lowp sampler2D diffuse;


void main() {
    lowp vec3 normal = vNormalOut;

    // Lighting
    lowp vec3 lightDir = vec3(0.5, 0.3, 0.5);
    lowp vec3 light = vec3(dot(normal, lightDir));

    lowp vec3 diffuse = texture2D(diffuse, vUVoutAndViewZ.xy).rgb;

    lowp vec3 finalColor = diffuse * light;

    // Fog
    finalColor = mix(finalColor, vec3(0.5, 0.8, 0.95), vUVoutAndViewZ.z);

    gl_FragColor = vec4(finalColor.xyz, 1.0);
}

I added the lowp/mediump declarations later (which added about 2fps), it works just as bad without them.

I refuse to believe that the PowerVR chipset is so weak that it can't handle a simple shader like this. There must be something stupid in this shader (like something implicitly swizzling a lowp register) that just completely messes up the unit.

ANSWER and EDIT:

Ellis has some fantastic information in the answer and the subsequent comments. In this particular case, it seems that mix() is flat-out broken. It brought the shader to 12 cycles (from 4) and 4 GPRs (from 2). I got back to 29fps by using this code instead:

lowp vec3 fogDiff = vec3(0.5, 0.8, 0.95) - finalColor;
fogDiff *= vUVoutAndViewZ.z;
finalColor += fogDiff;

Answer 1

Response from EboMike computed with PVRUniSCo Editor: mix() brings the shader from 4 cycles to 12 cycles (PowerVR).

The PowerVR 530/535 is very slow. Andreno 200 and PowerVR 530/535 are first GPU generation (OpenGL ES 2.x) for hdpi resolution. You can't redraw a full screen at 60FPS with a simple texture.

I wrote GPUBench to test shaders performance and it is very bad on these GPU. The second generation (Andreno 205, Powervr 540) is far better.

Today, i try to minimize Fragment shader with one or two operations. You can try to optimize your shader because GLSL compilers are bad (on mobile platform).

Note: for powervr gpu, read this document. PVRUniSCo Editor can compute cycles used in shaders.

My toughts:

• The Droid GPU is slower than Andreno 200 because of fillrate
• 30 FPS for Andreno 200 is good
• 20 FPS for Droid is good
• You can render your scene in a smaller FBO and draw it on a screen (but FBO switch is costly)
• Fragment shaders must be very short.