Bottle neck in fragment shader because large number of texture2D call

Question

Are there any way to optimize this flow:

1. The frame captured from camera (NV21 byte array)
2. Convert this frame to I420 format
3. Upload Y data to y_tex; U data to u_tex; V data to v_tex
4. Run the following fragment shader:

...
 float greenValue(vec2 coord)
 {
  return texture2D(y_tex, coord).r 
    - 0.344 * (texture2D(u_tex, coord).r - 0.5) 
    - 0.714 * (texture2D(v_tex, coord).r - 0.5);
 }
 // some code .... if (current pixel is skin color)
 vec2 blurCoordinates[20];
 blurCoordinates[0] = interp_tc.xy + singleStepOffset * vec2(0.0, -10.0);
 blurCoordinates[1] = interp_tc.xy + singleStepOffset * vec2(0.0, 10.0);
 blurCoordinates[2] = interp_tc.xy + singleStepOffset * vec2(-10.0, 0.0);
 blurCoordinates[3] = interp_tc.xy + singleStepOffset * vec2(10.0, 0.0);
 blurCoordinates[4] = interp_tc.xy + singleStepOffset * vec2(5.0, -8.0);
 blurCoordinates[5] = interp_tc.xy + singleStepOffset * vec2(5.0, 8.0);
 blurCoordinates[6] = interp_tc.xy + singleStepOffset * vec2(-5.0, 8.0);
 blurCoordinates[7] = interp_tc.xy + singleStepOffset * vec2(-5.0, -8.0);
 blurCoordinates[8] = interp_tc.xy + singleStepOffset * vec2(8.0, -5.0);
 blurCoordinates[9] = interp_tc.xy + singleStepOffset * vec2(8.0, 5.0);
 blurCoordinates[10] = interp_tc.xy + singleStepOffset * vec2(-8.0, 5.0);
 blurCoordinates[11] = interp_tc.xy + singleStepOffset * vec2(-8.0, -5.0);
 blurCoordinates[12] = interp_tc.xy + singleStepOffset * vec2(0.0, -6.0);
 blurCoordinates[13] = interp_tc.xy + singleStepOffset * vec2(0.0, 6.0);
 blurCoordinates[14] = interp_tc.xy + singleStepOffset * vec2(6.0, 0.0);
 blurCoordinates[15] = interp_tc.xy + singleStepOffset * vec2(-6.0, 0.0);
 blurCoordinates[16] = interp_tc.xy + singleStepOffset * vec2(-4.0, -4.0);
 blurCoordinates[17] = interp_tc.xy + singleStepOffset * vec2(-4.0, 4.0);
 blurCoordinates[18] = interp_tc.xy + singleStepOffset * vec2(4.0, -4.0);
 blurCoordinates[19] = interp_tc.xy + singleStepOffset * vec2(4.0, 4.0);
 // some code ....
 float sampleColor = centralColor.g * 20.0;
 sampleColor += greenValue(blurCoordinates[0]);
 sampleColor += greenValue(blurCoordinates[1]);
 sampleColor += greenValue(blurCoordinates[2]);
 sampleColor += greenValue(blurCoordinates[3]);
 sampleColor += greenValue(blurCoordinates[4]);
 sampleColor += greenValue(blurCoordinates[5]);
 sampleColor += greenValue(blurCoordinates[6]);
 sampleColor += greenValue(blurCoordinates[7]);
 sampleColor += greenValue(blurCoordinates[8]);
 sampleColor += greenValue(blurCoordinates[9]);
 sampleColor += greenValue(blurCoordinates[10]);
 sampleColor += greenValue(blurCoordinates[11]);
 sampleColor += greenValue(blurCoordinates[12]) * 2.0;
 sampleColor += greenValue(blurCoordinates[13]) * 2.0;
 sampleColor += greenValue(blurCoordinates[14]) * 2.0;
 sampleColor += greenValue(blurCoordinates[15]) * 2.0;
 sampleColor += greenValue(blurCoordinates[16]) * 2.0;
 sampleColor += greenValue(blurCoordinates[17]) * 2.0;
 sampleColor += greenValue(blurCoordinates[18]) * 2.0;
 sampleColor += greenValue(blurCoordinates[19]) * 2.0;

Problem:

I tested on weak device (galaxy A6) and capture with HD resolution frame (1280x720). It take ~100ms/frame. The main reason of this bottle neck is the large number of texture2D() call. Each frame make ~60,000,000 call texture2D().

Idea:

When I replace (u_tex, v_tex) by y_tex for test, it's very fast. So, I think performance maybe improved by use only one texture instead of three textures (y_tex, u_tex, v_tex).

Question:

Does my approach is feasible? Are there any way to optimize this shader or improve entire this flow?