#include "postprocess_mmx.h"

 
#ifdef PP_COMPONENTS_INLINE
 #ifndef INLINE
 #define INLINE static __inline
 #endif
 #include "deblock_horiz_DC_on.c"
 #include "deblock_horiz_useDC.c"
 #include "deblock_horiz_default_filter.c"
 #include "deblock_horiz_lpf9.c"
#endif

void deblock_horiz(uint8_t *image, int width, int stride, QP_STORE_T *QP_store, int QP_stride, int chromaFlag) {
	int x, y;
	int QP;
	uint8_t *v;
	int useDC, DC_on;
	#ifdef PREFETCH_AHEAD_H
	void *prefetch_addr;
	#endif

	y = 0;
		for (x=8; x<width; x+=8) {
		
			QP = chromaFlag ? QP_store[y/8*QP_stride+x/8]
			                : QP_store[y/16*QP_stride+x/16];	

			v = &(image[y*stride + x]) - 5;

			#ifdef PREFETCH_AHEAD_V
			prefetch_addr = v + PREFETCH_AHEAD_V;
			__asm {
				push eax
				push ebx
				mov eax, prefetch_addr
				mov ebx, stride
				add      eax , ebx       
				prefetcht0 [eax]           
				add      eax , ebx       
				prefetcht0 [eax]           
				add      eax , ebx       
				prefetcht0 [eax]           
				add      eax , ebx        
				prefetcht0 [eax]           
				pop ebx
				pop eax
			};
			#endif

			useDC = deblock_horiz_useDC(v, stride);

			if (useDC) { 
				
				DC_on = deblock_horiz_DC_on(v, stride, QP);

				if (DC_on) {

					deblock_horiz_lpf9(v, stride, QP); 

					#ifdef SHOWDECISIONS_H
					if (!chromaFlag) {
						v[0*stride + 4] = 
						v[1*stride + 4] = 
						v[2*stride + 4] = 
						v[3*stride + 4] = 255;  
					}
					#endif
				}

			} else {    
			
				deblock_horiz_default_filter(v, stride, QP);

				#ifdef SHOWDECISIONS_H
				if (!chromaFlag) {
					v[0*stride + 4] = 
					v[1*stride + 4] = 
					v[2*stride + 4] = 
					v[3*stride + 4] = 0;  
				}
				#endif

			}
		}
	__asm { 
		emms
	};
}