在GLSL中将采样的纹理值相乘然后相除会在某些Android设备上产生不正确的结果吗？

Question

Some background: This is my first OpenGL ES 2 program, so I'm inexperienced. I've been working on a shader to display color-indexed textures. It samples a point from one texture, multiplies it by 256.0 to scale it up to 8-bit integer range, samples the corresponding RGBA value from that offset in a palette texture, and sets the gl_FragColor.

I've had it working perfectly on my Samsung Galaxy S2 for a month now, but it never worked on my Galaxy S1, displaying only a black texture. After countless attempts of banging my head against the wall trying to find the problem through trial and error in all parts of my code, I finally narrowed it down to a piece of shader code that can effectively be reduced to the following:

precision highp float;

varying vec2 v_texCoord;
uniform sampler2D texture;

void main()
{
    float multiplied = texture2D(texture, v_texCoord).s * 256.0;

    float divided = multiplied / 256.0;

    gl_FragColor = vec4(divided, divided, divided, 1.0);
}

For that example, texture is a 256x256 pixel texture with every value set to 0xFF. If I run that code on my Galaxy S2, it works fine: I get a completely white screen, exactly 0xFFFFFF. If I run it on my original Galaxy S1, the screen is exactly 0x000000.

Now, I found that if I decrease that 256.0 value then run it, the output on my Galaxy S1 starts to increase from all black to various shades of gray. For example, 8.0 gave me 0x424142, and 4.0 gave me 0x848284. By the time I multiply / divide by 2.0, it starts outputting 0xFFFFFF just like my Galaxy S2!

Any idea what could possibly cause an issue like this? I'll take your wildest guesses, because I'm absolutely defeated at this point. I don't see anything that could be causing this.

Answer 1

Is the answer floating point truncation?

A fragment shader is allowed to use "lowp" float precision, which means 16-bit floating point, which 256 * 256 would overflow (see http://en.wikipedia.org/wiki/Half-precision_floating-point_format ). At this point the math falls apart and results are undefined.

However, as far as I can tell (see glblenchmark.com data ) the Galaxy S1 should support highp floats in the fragment shader as it has the "OES_fragment_precision_high" extension in the glGetString(GL_EXTENSIONS) output. (Though I may have the wrong model there...)

Maybe there is some other reason you're not getting high precision floats though. What #version are you using in your shader?