关于具有8个纹理单元的opengl es多纹理着色器的思考

Question

Below is a fragment shader that is performing poorly. Removing the conditional branches doesnt seem to improve performance. With just 150 polys I get 10fps on Kindle Fire and 20fps on Galaxy S3. Any thoughts on the best way to optimize this, if at all? Similar shader on iPad 2 runs 30+fps.

In the code below, texture1 to texture8 are bound to the eight textures. VertexTexturesOut1 and VertexTexturesOut2 are passed from the vertex shader and have values 0.0 to 1.0 to indicate the shading amount to be blended. The idea is to blend a landscape texture, so the grass blends evenly into dirt, rock, sand, etc.

uniform sampler2D texture1;
uniform sampler2D texture2;
uniform sampler2D texture3;
uniform sampler2D texture4;
uniform sampler2D texture5;
uniform sampler2D texture6;
uniform sampler2D texture7;
uniform sampler2D texture8;

varying mediump vec2 TextureCoordOut;
varying lowp vec4 VertexTexturesOut1;
varying lowp vec4 VertexTexturesOut2;

...

    lowp vec4 Color = vec4( 0.0, 0.0, 0.0, 0.0);

    if (VertexTexturesOut1.x != 0.0) Color = Color + texture2D ( texture1, TextureCoordOut ) * VertexTexturesOut1.x;
    if (VertexTexturesOut1.y != 0.0) Color = Color + texture2D ( texture2, TextureCoordOut ) * VertexTexturesOut1.y;
    if (VertexTexturesOut1.z != 0.0) Color = Color + texture2D ( texture3, TextureCoordOut ) * VertexTexturesOut1.z;
    if (VertexTexturesOut1.w != 0.0) Color = Color + texture2D ( texture4, TextureCoordOut ) * VertexTexturesOut1.w;

    if (VertexTexturesOut2.x != 0.0) Color = Color + texture2D ( texture5, TextureCoordOut ) * VertexTexturesOut2.x;
    if (VertexTexturesOut2.y != 0.0) Color = Color + texture2D ( texture6, TextureCoordOut ) * VertexTexturesOut2.y;
    if (VertexTexturesOut2.z != 0.0) Color = Color + texture2D ( texture7, TextureCoordOut ) * VertexTexturesOut2.z;
    if (VertexTexturesOut2.w != 0.0) Color = Color + texture2D ( texture8, TextureCoordOut ) * VertexTexturesOut2.w;

    gl_FragColor = Color;    

Answer 1

You shader has many of the DONT's when designing fragment shaders. You have many conditional statements so all the threads need to wait always for all the other threads in the warp (or batch). You are also accessing 8 textures which means you need to wait for all those texture lookups (which also kills memory bandwidth). Use any optimization tools, for example the Adreno SDK, the Mali SDK or any vendor of your choice and run the shader in their optimization tools to spot where is the GPU is spending more time.

Are you using full RGB888 textures? A

Maybe if you tell us what are you trying to archive we can think on another solution without using 8 texture lookups.