I followed NVIDIA's tutorial (CSM and PSSM pdf) and sample project but I'm confused how to compare depth with split distances. 

My fragment shader looks like:

uniform sampler2DArrayShadow uShadMap;
uniform mat4                 uShadMVP[SHAD_SPLIT];
uniform float                uShadDist[SHAD_SPLIT];

float shadowCoef() {
  vec4 shadCoord;
  int  i;

  for (i = 0; i < SHAD_SPLIT; i++) {
    if (gl_FragCoord.z < uShadDist[i])
      break;
  }

  if (i >= SHAD_SPLIT)
    return 1.0;

  shadCoord   = uShadMVP[i] * vPos;
  shadCoord.w = shadCoord.z;
  shadCoord.z = float(i);

  return texture(uShadMap, shadCoord);
}

uShadDist contains distances in view space e.g. [ 12.818679, 26.606153, 46.403923, 100.000061 ] but I guess I need to convert them to clip space coord: [ 0 : 1 ] before comparing with gl_FragCoord.z, right?

Here NVIDIA's code:
 

// f[i].fard is originally in eye space - tell's us how far we can see.
// Here we compute it in camera homogeneous coordinates. Basically, we calculate
// cam_proj * (0, 0, f[i].fard, 1)^t and then normalize to [0; 1]
far_bound[i] = 0.5f*(-f[i].fard*cam_proj[10]+cam_proj[14])/f[i].fard + 0.5f;

I don't understand what (or how) it does  I tried to multiply split dist with camera's perspective proj but I always get >= 1, Can someone explain what value should I compare with gl_FragCoord.z?

One another question is that I'm using Practical Split Scheme but I don't understand what lambda value should I use, I used 0.5.  C  = (Clog + Cuni) * 0.5f . But I'm not sure if it is best value or not. 

Ok I think the I found the error about first question, I was used wrong index:

/* wrong */
0.5f * (-C * camProj[2][3] + camProj[3][3]) / C + 0.5f

/* fixed */
0.5f * (-C * camProj[2][2] + camProj[3][2]) / C + 0.5f

now I got colorful scene, I will try to understand the simple math used here 

What about lambda value ?