mpc-hc Edge Sharpen mod and optimization

Edge_Sharpen.hlsl

#define Show_Edge 20  
//[0, 1, 2, 20] default 0: edge sharpen, 1: show edges in Red, 2: show luma edges, 20 luma edges on dark background 
#define Show_SplitScreen 1 //[0, 1, 2] default 0: no splitscreen, 1: horizontal split, 2 vertical split 

/*
Edge detection and Edge sharpening, using Four 3x3 Prewitt operators.

Software Licence: GPL v3
the original Edge Sharpen is distributed with MPC-HC: https://www.videohelp.com/software/Media-Player-Classic-Home-Cinema
2020: 	- modification and optimization by butterw (9 texture, 47 arithmetic) >>  (9 texture, 34 arithmetic) 
		- edge detection on Luma (Show_Edge 20, 8 texture, 30 arithmetic) 

Can be used pre or post-resize (Show_SplitScreen post-resize only really works in fullscreen). Source must be clean high-res to get nice visible edges.
*/


sampler s0: register(s0);
float2  p1: register(c1);

#define Edge_Threshold  0.2
#define Sharpen_val0    2.0
#define Sharpen_val1    0.125
//#define NbPixel     1.
#define dx    p1.x //NbPixel*p1.x
#define dy    p1.y //NbPixel*p1.y
#define CoefLuma float4(0.212656, 0.715158, 0.072186, 0) // BT.709 & sRBG luma coef (Monitors, HDTV)
#define Red float4(1, 0, 0, 0)
#define Black16  float4(16, 16, 16, 0)/255.


float4 main(float2 tex: TEXCOORD0): COLOR {
    // Get neighbor points:
    // [ 1, 2, 3 ]
    // [ 4, 0, 5 ]
    // [ 6, 7, 8 ]
    float4 c0 = tex2D(s0, tex);
	#if Show_SplitScreen == 1 
		if (tex.x<0.5) return c0;
	#elif Show_SplitScreen == 2
		if (tex.y<0.5) return c0;
	#endif
    float4 c1 = tex2D(s0, tex - p1);
    float4 c2 = tex2D(s0, tex + float2(  0, -dy));
    float4 c3 = tex2D(s0, tex + float2( dx, -dy));
    float4 c4 = tex2D(s0, tex + float2(-dx,   0));
    float4 c5 = tex2D(s0, tex + float2( dx,   0));
    float4 c6 = tex2D(s0, tex + float2(-dx,  dy));
    float4 c7 = tex2D(s0, tex + float2(  0,  dy));
    float4 c8 = tex2D(s0, tex + p1);

    /* Computation of the 3 derived vectors (hor, vert, diag1, diag2)
    // float4 delta1 = c6 + c4 + c1 - c3 - c5 - c8;
    // float4 delta2 = c1 + c2 + c4 - c5 - c8 - c7;
    // float4 delta3 = c1 + c2 + c3 - c8 - c7 - c6;
    // float4 delta4 = c2 + c3 + c5 - c7 - c6 - c4;
    Perf optimized by reusing calculated expressions: */
    float4 delta1 = -c3 + c6 + c4;
    float4 delta2 = c1 - c5 - c8;
    float4 delta4 = c2 - c7 + c5 - delta1;
    delta1+= delta2;
    delta2+= c2 - c7 + c4;
	c1 = c1 + c2 + c3;
	c6 = c6 + c7 + c8;
	float4 delta3 = c1 - c6;	
	
    // Computation of the Prewitt operator
    // float edge = length(abs(delta1) + abs(delta2) + abs(delta3) + abs(delta4))/6;
    float edge = sqrt(dot(abs(delta1) + abs(delta2) + abs(delta3) + abs(delta4), CoefLuma))*0.25; // Luma Edge detection
    if (edge > Edge_Threshold) { 
        #if Show_Edge == 0 // If we have an edge (vector length > Edge_threshold) => apply sharpen filter
			c0 = c0*Sharpen_val0 - Sharpen_val1*(c1 + c4 + c5 + c6);			
		#elif Show_Edge == 1 // Display edges in Red
			return Red;
		#elif (Show_Edge == 2 || Show_Edge == 20) // Display Luma edges
			return edge;
        #endif
	 }
	#if Show_Edge == 20 // Show Edge on Dark background
		return Black16;
	#endif 
	return c0;	
}