sharedsort.comp

#version 430

#define N_FRAGS 32

#define ROWS 8

layout(local_size_x = N_FRAGS, local_size_y = ROWS, local_size_z = 1) in;

#include "lfb.glsl"
LFB_DEC(lfb);

/*
in uvec3 gl_NumWorkGroups;
in uvec3 gl_WorkGroupID;
in uvec3 gl_LocalInvocationID;
in uvec3 gl_GlobalInvocationID;
in uint  gl_LocalInvocationIndex;
*/


//gl_LocalInvocationIndex

//memoryBarrierShared
//groupMemoryBarrier
//barrier

shared LFB_FRAG_TYPE frags[N_FRAGS * ROWS];

uniform int totalPixels;

void main()
{
	LFB_FRAG_TYPE tmp;
	#define SWAP_BITONIC_SORT(a, b) \
		if (LFB_FRAG_DEPTH(frags[so+a]) > LFB_FRAG_DEPTH(frags[so+b])) \
		{ \
			tmp = frags[so+a]; frags[so+a] = frags[so+b]; frags[so+b] = tmp; \
		}
	
	int pixelOffset = int(gl_LocalInvocationID.y);
	int pixel = int(gl_WorkGroupID.x * ROWS) + pixelOffset;
	
	if (pixel >= totalPixels)
		return;
	
	int i = int(gl_LocalInvocationID.x);
	int so = pixelOffset * N_FRAGS;
	
	LFB_INIT(lfb, pixel);
	//LFB_ITER_BEGIN(lfb);
	
	int fragCount = LFB_COUNT(lfb);
	
	if (i < fragCount)
		frags[so+i] = LFB_LOAD(lfb, fragCount - i);
	else
		LFB_FRAG_DEPTH(frags[so+i]) = 9999.0;
	
	//memoryBarrierShared();
	
	#if 1
	
	#if N_FRAGS == 8
	#define LOG2_NFRAGS 3
	#elif N_FRAGS == 16
	#define LOG2_NFRAGS 4
	#elif N_FRAGS == 32
	#define LOG2_NFRAGS 5
	#elif N_FRAGS == 64
	#define LOG2_NFRAGS 6
	#elif N_FRAGS == 128
	#define LOG2_NFRAGS 7
	#endif

	#pragma optionNV(unroll all)
	for (int lk = 1; lk <= LOG2_NFRAGS; ++lk)
	{
		int k = 1<<lk;
		for (int lj = lk-1; lj >= 0; --lj)
		{
			int j = 1<<lj;
			int ixj = i ^ j;
			if (ixj > i)
			{
				if ((i & k) == 0) SWAP_BITONIC_SORT(i,ixj)
				if ((i & k) != 0) SWAP_BITONIC_SORT(ixj,i)
				//memoryBarrierShared();
			}
		}
	}
	
	#endif
	
	if (i < fragCount)
	{
		//LFB_STORE(lfb, fragCount - i, frags[so+i]);
		imageStore(datalfb, lfbTmplfb.fragOffset + (fragCount - i), vec4(frags[so+i] LFB_FRAG_PAD));
	}
}