--- encoder.c	2002/12/08 06:43:34	1.76.2.25
+++ encoder.c	2003/01/26 05:09:00	1.76.2.42
@@ -39,11 +39,10 @@
  *             MinChen <chenm001@163.com>
  *  14.04.2002 added FrameCodeB()
  *
- *  $Id: encoder.c,v 1.76.2.25 2002/12/08 06:43:34 suxen_drol Exp $
+ *  $Id: encoder.c,v 1.76.2.42 2003/01/26 05:09:00 suxen_drol Exp $
  *
  ****************************************************************************/
 
-
 #include <stdlib.h>
 #include <stdio.h>
 #include <math.h>
@@ -73,7 +72,7 @@
  ****************************************************************************/
 
 #define ENC_CHECK(X) if(!(X)) return XVID_ERR_FORMAT
-#define SWAP(A,B)    { void * tmp = A; A = B; B = tmp; }
+#define SWAP(_T_,A,B)    { _T_ tmp = A; A = B; B = tmp; }
 
 /*****************************************************************************
  * Local function prototypes
@@ -230,8 +229,8 @@
 
 	pEnc->bitrate = pParam->rc_bitrate;
 
-	pEnc->iFrameNum = 0;
-	pEnc->iMaxKeyInterval = pParam->max_key_interval;
+	pEnc->iFrameNum = -1;
+	pEnc->mbParam.iMaxKeyInterval = pParam->max_key_interval;
 
 	/* try to allocate frame memory */
 
@@ -255,9 +254,8 @@
 
 	/* try to allocate image memory */
 
-#ifdef _DEBUG_PSNR
-	image_null(&pEnc->sOriginal);
-#endif
+	if (pParam->global & XVID_GLOBAL_EXTRASTATS)
+		image_null(&pEnc->sOriginal);
 
 	image_null(&pEnc->f_refh);
 	image_null(&pEnc->f_refv);
@@ -271,12 +269,12 @@
 	image_null(&pEnc->vInterHV);
 	image_null(&pEnc->vInterHVf);
 
-#ifdef _DEBUG_PSNR
-	if (image_create
-		(&pEnc->sOriginal, pEnc->mbParam.edged_width,
-		 pEnc->mbParam.edged_height) < 0)
-		goto xvid_err_memory3;
-#endif
+	if (pParam->global & XVID_GLOBAL_EXTRASTATS)
+	{	if (image_create
+			(&pEnc->sOriginal, pEnc->mbParam.edged_width,
+			 pEnc->mbParam.edged_height) < 0)
+			goto xvid_err_memory3;
+	}
 
 	if (image_create
 		(&pEnc->f_refh, pEnc->mbParam.edged_width,
@@ -320,15 +318,21 @@
 		 pEnc->mbParam.edged_height) < 0)
 		goto xvid_err_memory3;
 
+/* Create full bitplane for GMC, this might be wasteful */
+	if (image_create
+		(&pEnc->vGMC, pEnc->mbParam.edged_width,
+		 pEnc->mbParam.edged_height) < 0)
+		goto xvid_err_memory3;
 
 
-	/* B Frames specific init */
 
-	pEnc->global = pParam->global;
+	pEnc->mbParam.global = pParam->global;
+
+	/* B Frames specific init */
 	pEnc->mbParam.max_bframes = pParam->max_bframes;
-	pEnc->bquant_ratio = pParam->bquant_ratio;
-	pEnc->bquant_offset = pParam->bquant_offset;
-	pEnc->frame_drop_ratio = pParam->frame_drop_ratio;
+	pEnc->mbParam.bquant_ratio = pParam->bquant_ratio;
+	pEnc->mbParam.bquant_offset = pParam->bquant_offset;
+	pEnc->mbParam.frame_drop_ratio = pParam->frame_drop_ratio;
 	pEnc->bframes = NULL;
 
 	if (pEnc->mbParam.max_bframes > 0) {
@@ -460,10 +464,11 @@
 	}
 
   xvid_err_memory3:
-#ifdef _DEBUG_PSNR
-	image_destroy(&pEnc->sOriginal, pEnc->mbParam.edged_width,
-				  pEnc->mbParam.edged_height);
-#endif
+
+	if (pEnc->mbParam.global & XVID_GLOBAL_EXTRASTATS)
+	{	image_destroy(&pEnc->sOriginal, pEnc->mbParam.edged_width,
+					  pEnc->mbParam.edged_height);
+	}
 
 	image_destroy(&pEnc->f_refh, pEnc->mbParam.edged_width,
 				  pEnc->mbParam.edged_height);
@@ -487,6 +492,11 @@
 	image_destroy(&pEnc->vInterHVf, pEnc->mbParam.edged_width,
 				  pEnc->mbParam.edged_height);
 
+/* destroy GMC image */
+	image_destroy(&pEnc->vGMC, pEnc->mbParam.edged_width,
+				  pEnc->mbParam.edged_height);
+
+
   xvid_err_memory2:
 	xvid_free(pEnc->current->mbs);
 	xvid_free(pEnc->reference->mbs);
@@ -574,10 +584,10 @@
 	image_destroy(&pEnc->f_refhv, pEnc->mbParam.edged_width,
 				  pEnc->mbParam.edged_height);
 
-#ifdef _DEBUG_PSNR
-	image_destroy(&pEnc->sOriginal, pEnc->mbParam.edged_width,
-				  pEnc->mbParam.edged_height);
-#endif
+	if (pEnc->mbParam.global & XVID_GLOBAL_EXTRASTATS)
+	{	image_destroy(&pEnc->sOriginal, pEnc->mbParam.edged_width,
+					  pEnc->mbParam.edged_height);
+	}
 
 	/* Encoder structure */
 
@@ -646,6 +656,16 @@
 
 
 
+/* convert pFrame->intra to coding_type */
+static int intra2coding_type(int intra)
+{
+	if (intra < 0)  return -1;
+	if (intra == 1) return I_VOP;
+	if (intra == 2) return B_VOP;
+
+	return P_VOP;
+}
+
 
 
 /*****************************************************************************
@@ -664,16 +684,12 @@
 {
 	uint16_t x, y;
 	Bitstream bs;
-	uint32_t bits, mode;
+	uint32_t bits;
+	int mode;
 
 	int input_valid = 1;
 	int bframes_count = 0;
 
-#ifdef _DEBUG_PSNR
-	float psnr;
-	char temp[128];
-#endif
-
 	ENC_CHECK(pEnc);
 	ENC_CHECK(pFrame);
 	ENC_CHECK(pFrame->image);
@@ -703,19 +719,28 @@
 				pEnc->queue_head, pEnc->queue_tail, pEnc->queue_size);
 
 			pEnc->bframenum_tail--;
-			SWAP(pEnc->current, pEnc->reference);
+			SWAP(FRAMEINFO *, pEnc->current, pEnc->reference);
 
-			SWAP(pEnc->current, pEnc->bframes[pEnc->bframenum_tail]);
+			SWAP(FRAMEINFO *, pEnc->current, pEnc->bframes[pEnc->bframenum_tail]);
 
 			FrameCodeP(pEnc, &bs, &bits, 1, 0);
 			bframes_count = 0;
 
-			BitstreamPad(&bs);
+			BitstreamPadAlways(&bs);
 			pFrame->length = BitstreamLength(&bs);
 			pFrame->intra = 0;
 
+
 			emms();
 
+			if (pResult) {
+				pResult->quant = pEnc->current->quant;
+				pResult->hlength = pFrame->length - (pEnc->current->sStat.iTextBits / 8);
+				pResult->kblks = pEnc->current->sStat.kblks;
+				pResult->mblks = pEnc->current->sStat.mblks;
+				pResult->ublks = pEnc->current->sStat.ublks;
+			}
+
 			return XVID_ERR_OK;
 		}
 
@@ -727,10 +752,18 @@
 		FrameCodeB(pEnc, pEnc->bframes[pEnc->bframenum_head], &bs, &bits);
 		pEnc->bframenum_head++;
 
-		BitstreamPad(&bs);
+		BitstreamPadAlways(&bs);
 		pFrame->length = BitstreamLength(&bs);
 		pFrame->intra = 2;
 
+		if (pResult) {
+			pResult->quant = pEnc->current->quant;
+			pResult->hlength = pFrame->length - (pEnc->current->sStat.iTextBits / 8);
+			pResult->kblks = pEnc->current->sStat.kblks;
+			pResult->mblks = pEnc->current->sStat.mblks;
+			pResult->ublks = pEnc->current->sStat.ublks;
+		}
+
 		if (input_valid)
 			queue_image(pEnc, pFrame);
 
@@ -749,23 +782,32 @@
 		   indentical to the future-referece frame.
 		*/
 
-		if ((pEnc->global & XVID_GLOBAL_PACKED)) {
+		if ((pEnc->mbParam.global & XVID_GLOBAL_PACKED)) {
 			int tmp;
 			
 			DPRINTF(DPRINTF_DEBUG,"*** EMPTY bf: head=%i tail=%i   queue: head=%i tail=%i size=%i",
 				pEnc->bframenum_head, pEnc->bframenum_tail,
 				pEnc->queue_head, pEnc->queue_tail, pEnc->queue_size);
 
-			BitstreamPad(&bs);
 
 			tmp = pEnc->current->seconds;
 			pEnc->current->seconds = 0; /* force time_base = 0 */
+
 			BitstreamWriteVopHeader(&bs, &pEnc->mbParam, pEnc->current, 0);
 			pEnc->current->seconds = tmp;
 
+			BitstreamPadAlways(&bs);
 			pFrame->length = BitstreamLength(&bs);
 			pFrame->intra = 4;
 
+			if (pResult) {
+				pResult->quant = pEnc->current->quant;
+				pResult->hlength = pFrame->length - (pEnc->current->sStat.iTextBits / 8);
+				pResult->kblks = pEnc->current->sStat.kblks;
+				pResult->mblks = pEnc->current->sStat.mblks;
+				pResult->ublks = pEnc->current->sStat.ublks;
+			}
+
 			if (input_valid)
 				queue_image(pEnc, pFrame);
 
@@ -781,10 +823,10 @@
 	if (pEnc->bframenum_dx50bvop != -1)
 	{
 
-		SWAP(pEnc->current, pEnc->reference);
-		SWAP(pEnc->current, pEnc->bframes[pEnc->bframenum_dx50bvop]);		
+		SWAP(FRAMEINFO *, pEnc->current, pEnc->reference);
+		SWAP(FRAMEINFO *, pEnc->current, pEnc->bframes[pEnc->bframenum_dx50bvop]);		
 
-		if ((pEnc->global & XVID_GLOBAL_DEBUG)) {
+		if ((pEnc->mbParam.global & XVID_GLOBAL_DEBUG)) {
 			image_printf(&pEnc->current->image, pEnc->mbParam.edged_width, pEnc->mbParam.height, 5, 100, "DX50 IVOP");
 		}
 
@@ -796,7 +838,7 @@
 
 	} else if (input_valid) {
 
-		SWAP(pEnc->current, pEnc->reference);
+		SWAP(FRAMEINFO *, pEnc->current, pEnc->reference);
 
 		start_timer();
 		if (image_input
@@ -822,7 +864,7 @@
 
 	} else if (pEnc->queue_size > 0) {
 		
-		SWAP(pEnc->current, pEnc->reference);
+		SWAP(FRAMEINFO *, pEnc->current, pEnc->reference);
 
 		image_swap(&pEnc->current->image, &pEnc->queue[pEnc->queue_head]);
 		pEnc->queue_head =  (pEnc->queue_head + 1) % pEnc->mbParam.max_bframes;
@@ -839,12 +881,45 @@
 				pEnc->bframenum_head, pEnc->bframenum_tail,
 				pEnc->queue_head, pEnc->queue_tail, pEnc->queue_size);
 
-			BitstreamPutBits(&bs, 0x7f, 8);
+		//	BitstreamPutBits(&bs, 0x7f, 8);
 			pFrame->intra = 5;
+
+			if (pResult) {
+				/*
+				 * We must decide what to put there because i know some apps
+				 * are storing statistics about quantizers and just do
+				 * stats[quant]++ or stats[quant-1]++
+				 * transcode is one of these app with its 2pass module
+				 */
+
+				/*
+				 * For now i prefer 31 than 0 that could lead to a segfault
+				 * in transcode
+				 */
+				pResult->quant = 31;
+
+				pResult->hlength = 0;
+				pResult->kblks = 0;
+				pResult->mblks = 0;
+				pResult->ublks = 0;
+			}
+
+		} else {
+
+			if (pResult) {
+				pResult->quant = pEnc->current->quant;
+				pResult->hlength = pFrame->length - (pEnc->current->sStat.iTextBits / 8);
+				pResult->kblks = pEnc->current->sStat.kblks;
+				pResult->mblks = pEnc->current->sStat.mblks;
+				pResult->ublks = pEnc->current->sStat.ublks;
+			}
+
 		}
 
 		pFrame->length = BitstreamLength(&bs);
+
 		emms();
+
 		return XVID_ERR_OK;
 	}
 
@@ -855,6 +930,7 @@
 	// only inc frame num, adapt quant, etc. if we havent seen it before
 	if (pEnc->bframenum_dx50bvop < 0 )
 	{
+		mode = intra2coding_type(pFrame->intra);
 		if (pFrame->quant == 0)
 			pEnc->current->quant = RateControlGetQ(&pEnc->rate_control, 0);
 		else
@@ -875,14 +951,14 @@
 
 		inc_frame_num(pEnc);
 
-#ifdef _DEBUG_PSNR
-		image_copy(&pEnc->sOriginal, &pEnc->current->image,
-			   pEnc->mbParam.edged_width, pEnc->mbParam.height);
-#endif
+		if (pFrame->general & XVID_EXTRASTATS)
+		{	image_copy(&pEnc->sOriginal, &pEnc->current->image,
+				   pEnc->mbParam.edged_width, pEnc->mbParam.height);
+		}
 
 		emms();
 
-		if ((pEnc->global & XVID_GLOBAL_DEBUG)) {
+		if ((pEnc->mbParam.global & XVID_GLOBAL_DEBUG)) {
 			image_printf(&pEnc->current->image, pEnc->mbParam.edged_width, pEnc->mbParam.height, 5, 5, 
 				"%i  if:%i  st:%i", pEnc->m_framenum++, pEnc->iFrameNum, pEnc->current->stamp);
 		}
@@ -928,14 +1004,19 @@
 	 * %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% */
 	pEnc->iFrameNum++;
 
-	if (pEnc->iFrameNum == 0 || pFrame->intra == 1 || pEnc->bframenum_dx50bvop >= 0 ||
-		(pFrame->intra < 0 && pEnc->iMaxKeyInterval > 0 &&
-		 pEnc->iFrameNum >= pEnc->iMaxKeyInterval)
-		|| 2 == (mode = MEanalysis(&pEnc->reference->image, pEnc->current,
-									&pEnc->mbParam, pEnc->iMaxKeyInterval,
-									(pFrame->intra < 0) ? pEnc->iFrameNum : 0,
-									bframes_count++))) {
+	if (pEnc->iFrameNum == 0 || pEnc->bframenum_dx50bvop >= 0 ||
+		(mode < 0 && pEnc->mbParam.iMaxKeyInterval > 0 && 
+			pEnc->iFrameNum >= pEnc->mbParam.iMaxKeyInterval))
+	{
+		mode = I_VOP;
+	}else{
+		mode = MEanalysis(&pEnc->reference->image, pEnc->current,
+					&pEnc->mbParam, pEnc->mbParam.iMaxKeyInterval,
+					(mode < 0) ? pEnc->iFrameNum : 0,
+					bframes_count++);
+	}
 
+	if (mode == I_VOP) {
 		/*
 		 * This will be coded as an Intra Frame
 		 */
@@ -958,26 +1039,25 @@
 				pEnc->bframenum_head, pEnc->bframenum_tail,
 				pEnc->queue_head, pEnc->queue_tail, pEnc->queue_size);
 		
-		if ((pEnc->global & XVID_GLOBAL_DEBUG)) {
+		if ((pEnc->mbParam.global & XVID_GLOBAL_DEBUG)) {
 			image_printf(&pEnc->current->image, pEnc->mbParam.edged_width, pEnc->mbParam.height, 5, 200, "IVOP");
 		}
 
 		// when we reach an iframe in DX50BVOP mode, encode the last bframe as a pframe
 
-		if ((pEnc->global & XVID_GLOBAL_DX50BVOP) && pEnc->bframenum_tail > 0) {
+		if ((pEnc->mbParam.global & XVID_GLOBAL_DX50BVOP) && pEnc->bframenum_tail > 0) {
 
 			pEnc->bframenum_tail--;
 			pEnc->bframenum_dx50bvop = pEnc->bframenum_tail;
 
-			SWAP(pEnc->current, pEnc->bframes[pEnc->bframenum_dx50bvop]);
-			if ((pEnc->global & XVID_GLOBAL_DEBUG)) {
+			SWAP(FRAMEINFO *, pEnc->current, pEnc->bframes[pEnc->bframenum_dx50bvop]);
+			if ((pEnc->mbParam.global & XVID_GLOBAL_DEBUG)) {
 				image_printf(&pEnc->current->image, pEnc->mbParam.edged_width, pEnc->mbParam.height, 5, 100, "DX50 BVOP->PVOP");
 			}
 			FrameCodeP(pEnc, &bs, &bits, 1, 0);
 			bframes_count = 0;
-
 			pFrame->intra = 0;
-			
+
 		} else {
 
 			FrameCodeI(pEnc, &bs, &bits);
@@ -989,7 +1069,7 @@
 
 		pEnc->flush_bframes = 1;
 
-		if ((pEnc->global & XVID_GLOBAL_PACKED) && pEnc->bframenum_tail > 0) {
+		if ((pEnc->mbParam.global & XVID_GLOBAL_PACKED) && pEnc->bframenum_tail > 0) {
 			BitstreamPadAlways(&bs);
 			input_valid = 0;
 			goto ipvop_loop;
@@ -999,8 +1079,7 @@
 		 * NB : sequences like "IIBB" decode fine with msfdam but,
 		 *      go screwy with divx 5.00
 		 */
-	} else if (pEnc->bframenum_tail >= pEnc->mbParam.max_bframes || mode != 0) {
-// 	} else if (pFrame->intra == 0 || pEnc->bframenum_tail >= pEnc->mbParam.max_bframes || mode != 0) { 
+	} else if (mode == P_VOP || mode == S_VOP || pEnc->bframenum_tail >= pEnc->mbParam.max_bframes) {
 		/*
 		 * This will be coded as a Predicted Frame
 		 */
@@ -1009,7 +1088,7 @@
 				pEnc->bframenum_head, pEnc->bframenum_tail,
 				pEnc->queue_head, pEnc->queue_tail, pEnc->queue_size);
 		
-		if ((pEnc->global & XVID_GLOBAL_DEBUG)) {
+		if ((pEnc->mbParam.global & XVID_GLOBAL_DEBUG)) {
 			image_printf(&pEnc->current->image, pEnc->mbParam.edged_width, pEnc->mbParam.height, 5, 200, "PVOP");
 		}
 
@@ -1018,24 +1097,24 @@
 		pFrame->intra = 0;
 		pEnc->flush_bframes = 1;
 
-		if ((pEnc->global & XVID_GLOBAL_PACKED) && (pEnc->bframenum_tail > 0)) {
+		if ((pEnc->mbParam.global & XVID_GLOBAL_PACKED) && (pEnc->bframenum_tail > 0)) {
 			BitstreamPadAlways(&bs);
 			input_valid = 0;
 			goto ipvop_loop;
 		}
 
-	} else {
+	} else {	/* mode == B_VOP */
 		/*
 		 * This will be coded as a Bidirectional Frame
 		 */
 
-		if ((pEnc->global & XVID_GLOBAL_DEBUG)) {
+		if ((pEnc->mbParam.global & XVID_GLOBAL_DEBUG)) {
 			image_printf(&pEnc->current->image, pEnc->mbParam.edged_width, pEnc->mbParam.height, 5, 200, "BVOP");
 		}
 
 		if (pFrame->bquant < 1) {
 			pEnc->current->quant = ((((pEnc->reference->quant + pEnc->current->quant) * 
-				pEnc->bquant_ratio) / 2) + pEnc->bquant_offset)/100;
+				pEnc->mbParam.bquant_ratio) / 2) + pEnc->mbParam.bquant_offset)/100;
 
 		} else {
 			pEnc->current->quant = pFrame->bquant;
@@ -1051,8 +1130,8 @@
 				pEnc->queue_head, pEnc->queue_tail, pEnc->queue_size,pEnc->current->quant);
 
 		/* store frame into bframe buffer & swap ref back to current */
-		SWAP(pEnc->current, pEnc->bframes[pEnc->bframenum_tail]);
-		SWAP(pEnc->current, pEnc->reference);
+		SWAP(FRAMEINFO *, pEnc->current, pEnc->bframes[pEnc->bframenum_tail]);
+		SWAP(FRAMEINFO *, pEnc->current, pEnc->reference);
 
 		pEnc->bframenum_tail++;
 
@@ -1064,7 +1143,7 @@
 		goto bvop_loop;
 	}
 
-	BitstreamPad(&bs);
+	BitstreamPadAlways(&bs);
 	pFrame->length = BitstreamLength(&bs);
 
 	if (pResult) {
@@ -1073,20 +1152,27 @@
 		pResult->kblks = pEnc->current->sStat.kblks;
 		pResult->mblks = pEnc->current->sStat.mblks;
 		pResult->ublks = pEnc->current->sStat.ublks;
+
+		if (pFrame->general & XVID_EXTRASTATS)
+		{	pResult->sse_y =
+				plane_sse( pEnc->sOriginal.y, pEnc->current->image.y,
+						   pEnc->mbParam.edged_width, pEnc->mbParam.width,
+						   pEnc->mbParam.height);
+
+			pResult->sse_u =
+				plane_sse( pEnc->sOriginal.u, pEnc->current->image.u,
+						   pEnc->mbParam.edged_width/2, pEnc->mbParam.width/2,
+						   pEnc->mbParam.height/2);
+
+			pResult->sse_v =
+				plane_sse( pEnc->sOriginal.v, pEnc->current->image.v,
+						   pEnc->mbParam.edged_width/2, pEnc->mbParam.width/2,
+						   pEnc->mbParam.height/2);		
+		}
 	}
 
 	emms();
 
-#ifdef _DEBUG_PSNR
-	psnr =
-		image_psnr(&pEnc->sOriginal, &pEnc->current->image,
-				   pEnc->mbParam.edged_width, pEnc->mbParam.width,
-				   pEnc->mbParam.height);
-
-	snprintf(temp, 127, "PSNR: %f\n", psnr);
-	DEBUG(temp);
-#endif
-
 	if (pFrame->quant == 0) {
 		RateControlUpdate(&pEnc->rate_control, pEnc->current->quant,
 						  pFrame->length, pFrame->intra);
@@ -1120,10 +1206,8 @@
 	uint32_t bits;
 	uint16_t write_vol_header = 0;
 
-#ifdef _DEBUG_PSNR
 	float psnr;
 	uint8_t temp[128];
-#endif
 
 	start_global_timer();
 
@@ -1132,7 +1216,7 @@
 	ENC_CHECK(pFrame->bitstream);
 	ENC_CHECK(pFrame->image);
 
-	SWAP(pEnc->current, pEnc->reference);
+	SWAP(FRAMEINFO *, pEnc->current, pEnc->reference);
 
 	pEnc->current->global_flags = pFrame->general;
 	pEnc->current->motion_flags = pFrame->motion;
@@ -1154,10 +1238,10 @@
 		return XVID_ERR_FORMAT;
 	stop_conv_timer();
 
-#ifdef _DEBUG_PSNR
-	image_copy(&pEnc->sOriginal, &pEnc->current->image,
-			   pEnc->mbParam.edged_width, pEnc->mbParam.height);
-#endif
+	if (pFrame->general & XVID_EXTRASTATS)
+	{	image_copy(&pEnc->sOriginal, &pEnc->current->image,
+				   pEnc->mbParam.edged_width, pEnc->mbParam.height);
+	}
 
 	emms();
 
@@ -1235,8 +1319,8 @@
 
 	if (pFrame->intra < 0) {
 		if ((pEnc->iFrameNum == 0)
-			|| ((pEnc->iMaxKeyInterval > 0)
-				&& (pEnc->iFrameNum >= pEnc->iMaxKeyInterval))) {
+			|| ((pEnc->mbParam.iMaxKeyInterval > 0)
+				&& (pEnc->iFrameNum >= pEnc->mbParam.iMaxKeyInterval))) {
 			pFrame->intra = FrameCodeI(pEnc, &bs, &bits);
 		} else {
 			pFrame->intra = FrameCodeP(pEnc, &bs, &bits, 0, write_vol_header);
@@ -1250,9 +1334,9 @@
 
 	}
 
-	BitstreamPutBits(&bs, 0xFFFF, 16);
-	BitstreamPutBits(&bs, 0xFFFF, 16);
-	BitstreamPad(&bs);
+//	BitstreamPutBits(&bs, 0xFFFF, 16);
+//	BitstreamPutBits(&bs, 0xFFFF, 16);
+	BitstreamPadAlways(&bs);
 	pFrame->length = BitstreamLength(&bs);
 
 	if (pResult) {
@@ -1269,15 +1353,15 @@
 		RateControlUpdate(&pEnc->rate_control, pEnc->current->quant,
 						  pFrame->length, pFrame->intra);
 	}
-#ifdef _DEBUG_PSNR
-	psnr =
-		image_psnr(&pEnc->sOriginal, &pEnc->current->image,
-				   pEnc->mbParam.edged_width, pEnc->mbParam.width,
-				   pEnc->mbParam.height);
-
-	snprintf(temp, 127, "PSNR: %f\n", psnr);
-	DEBUG(temp);
-#endif
+	if (pFrame->general & XVID_EXTRASTATS)
+	{
+		psnr =
+			image_psnr(&pEnc->sOriginal, &pEnc->current->image,
+					   pEnc->mbParam.edged_width, pEnc->mbParam.width,
+					   pEnc->mbParam.height);
+	
+		snprintf(temp, 127, "PSNR: %f\n", psnr);
+	}
 
 	pEnc->iFrameNum++;
 
@@ -1443,7 +1527,7 @@
 
 	if (intra) {
 		if (!hint->rawhints) {
-			BitstreamPad(&bs);
+			BitstreamPadAlways(&bs);
 			hint->hintlength = BitstreamLength(&bs);
 		}
 		return;
@@ -1518,12 +1602,28 @@
 		   Bitstream * bs,
 		   uint32_t * pBits)
 {
+	int mb_width = pEnc->mbParam.mb_width;
+	int mb_height = pEnc->mbParam.mb_height;
 
 	DECLARE_ALIGNED_MATRIX(dct_codes, 6, 64, int16_t, CACHE_LINE);
 	DECLARE_ALIGNED_MATRIX(qcoeff, 6, 64, int16_t, CACHE_LINE);
 
 	uint16_t x, y;
 
+	if ((pEnc->current->global_flags & XVID_REDUCED))
+	{
+		mb_width = (pEnc->mbParam.width + 31) / 32;
+		mb_height = (pEnc->mbParam.height + 31) / 32;
+
+		/* 16x16->8x8 downsample requires 1 additional edge pixel*/
+		/* XXX: setedges is overkill */
+		start_timer();
+		image_setedges(&pEnc->current->image, 
+			pEnc->mbParam.edged_width, pEnc->mbParam.edged_height,
+			pEnc->mbParam.width, pEnc->mbParam.height);
+		stop_edges_timer();
+	}
+
 	pEnc->iFrameNum = 0;
 	pEnc->mbParam.m_rounding_type = 1;
 	pEnc->current->rounding_type = pEnc->mbParam.m_rounding_type;
@@ -1532,25 +1632,19 @@
 
 	BitstreamWriteVolHeader(bs, &pEnc->mbParam, pEnc->current);
 
-#define DIVX501B481P "DivX501b481p"
-	if ((pEnc->global & XVID_GLOBAL_PACKED)) {
-		BitstreamWriteUserData(bs, DIVX501B481P, strlen(DIVX501B481P));
-	}
-
-#define XVID_ID "XviD" XVID_BS_VERSION
-	BitstreamWriteUserData(bs, XVID_ID, strlen(XVID_ID));
-
 	set_timecodes(pEnc->current,pEnc->reference,pEnc->mbParam.fbase);
+
+	BitstreamPadAlways(bs);
 	BitstreamWriteVopHeader(bs, &pEnc->mbParam, pEnc->current, 1);
 
 	*pBits = BitstreamPos(bs);
 
 	pEnc->current->sStat.iTextBits = 0;
-	pEnc->current->sStat.kblks = pEnc->mbParam.mb_width * pEnc->mbParam.mb_height;
+	pEnc->current->sStat.kblks = mb_width * mb_height;
 	pEnc->current->sStat.mblks = pEnc->current->sStat.ublks = 0;
 
-	for (y = 0; y < pEnc->mbParam.mb_height; y++)
-		for (x = 0; x < pEnc->mbParam.mb_width; x++) {
+	for (y = 0; y < mb_height; y++)
+		for (x = 0; x < mb_width; x++) {
 			MACROBLOCK *pMB =
 				&pEnc->current->mbs[x + y * pEnc->mbParam.mb_width];
 
@@ -1573,6 +1667,12 @@
 			stop_coding_timer();
 		}
 
+	if ((pEnc->current->global_flags & XVID_REDUCED))
+	{
+		image_deblock_rrv(&pEnc->current->image, pEnc->mbParam.edged_width, 
+			pEnc->current->mbs, mb_width, mb_height, pEnc->mbParam.mb_width,
+			16, XVID_DEC_DEBLOCKY|XVID_DEC_DEBLOCKUV);
+	}
 	emms();
 
 	*pBits = BitstreamPos(bs) - *pBits;
@@ -1590,6 +1690,8 @@
 #define INTRA_THRESHOLD 0.5
 #define BFRAME_SKIP_THRESHHOLD 30
 
+
+/* FrameCodeP also handles S(GMC)-VOPs */
 static int
 FrameCodeP(Encoder * pEnc,
 		   Bitstream * bs,
@@ -1602,6 +1704,9 @@
 	DECLARE_ALIGNED_MATRIX(dct_codes, 6, 64, int16_t, CACHE_LINE);
 	DECLARE_ALIGNED_MATRIX(qcoeff, 6, 64, int16_t, CACHE_LINE);
 
+	int mb_width = pEnc->mbParam.mb_width;
+	int mb_height = pEnc->mbParam.mb_height;
+
 	int iLimit;
 	int x, y, k;
 	int iSearchRange;
@@ -1610,6 +1715,13 @@
 	/* IMAGE *pCurrent = &pEnc->current->image; */
 	IMAGE *pRef = &pEnc->reference->image;
 
+	if ((pEnc->current->global_flags & XVID_REDUCED))
+	{
+		mb_width = (pEnc->mbParam.width + 31) / 32;
+		mb_height = (pEnc->mbParam.height + 31) / 32;
+	}
+
+
 	start_timer();
 	image_setedges(pRef, pEnc->mbParam.edged_width, pEnc->mbParam.edged_height,
 				   pEnc->mbParam.width, pEnc->mbParam.height);
@@ -1621,11 +1733,9 @@
 	pEnc->current->fcode = pEnc->mbParam.m_fcode;
 
 	if (!force_inter)
-		iLimit =
-			(int) (pEnc->mbParam.mb_width * pEnc->mbParam.mb_height *
-				   INTRA_THRESHOLD);
+		iLimit = (int)(mb_width * mb_height *  INTRA_THRESHOLD);
 	else
-		iLimit = pEnc->mbParam.mb_width * pEnc->mbParam.mb_height + 1;
+		iLimit = mb_width * mb_height + 1;
 
 	if ((pEnc->current->global_flags & XVID_HALFPEL)) {
 		start_timer();
@@ -1637,42 +1747,44 @@
 		stop_inter_timer();
 	}
 
-	if (pEnc->current->global_flags & XVID_GMC) {
-//		printf("Global Motion = %d %d quarterpel=%d\n", pEnc->current->GMC_MV.x, pEnc->current->GMC_MV.y,pEnc->current->quarterpel);
-		DPRINTF(DPRINTF_HEADER, "Global Motion = %d %d quarterpel=%d\n", pEnc->current->GMC_MV.x, pEnc->current->GMC_MV.y,pEnc->current->quarterpel);
-		pEnc->current->coding_type = S_VOP;
-	} else
-		pEnc->current->coding_type = P_VOP;
+	pEnc->current->coding_type = P_VOP;
 	
 	start_timer();
-	if (pEnc->current->global_flags & XVID_HINTEDME_SET) {
+	if (pEnc->current->global_flags & XVID_HINTEDME_SET)
 		HintedMESet(pEnc, &bIntra);
-	if (bIntra == 0) {
-			pEnc->current->fcode = FindFcode(&pEnc->mbParam, pEnc->current);
-			MotionEstimationHinted(&pEnc->mbParam, pEnc->current, pEnc->reference,
-											&pEnc->vInterH, &pEnc->vInterV, &pEnc->vInterHV);
-		}
-
-	} else {
-
+	else
 		bIntra =
 			MotionEstimation(&pEnc->mbParam, pEnc->current, pEnc->reference,
                          &pEnc->vInterH, &pEnc->vInterV, &pEnc->vInterHV,
                          iLimit);
-	}
+
 	stop_motion_timer();
 
 	if (bIntra == 1) return FrameCodeI(pEnc, bs, pBits);
 
-	if ( (pEnc->current->GMC_MV.x == 0) && (pEnc->current->GMC_MV.y == 0) )
-			pEnc->current->coding_type = P_VOP;		/* no global motion -> no GMC */
+	if ( ( pEnc->current->global_flags & XVID_GMC ) 
+		&& ( (pEnc->current->warp.duv[1].x != 0) || (pEnc->current->warp.duv[1].y != 0) ) )
+	{
+		pEnc->current->coding_type = S_VOP;
 
+		generate_GMCparameters(	2, 16, &pEnc->current->warp, 
+					pEnc->mbParam.width, pEnc->mbParam.height, 
+					&pEnc->current->gmc_data);
+
+		generate_GMCimage(&pEnc->current->gmc_data, &pEnc->reference->image, 
+				pEnc->mbParam.mb_width, pEnc->mbParam.mb_height,
+				pEnc->mbParam.edged_width, pEnc->mbParam.edged_width/2, 
+				pEnc->mbParam.m_fcode, pEnc->mbParam.m_quarterpel, 0, 
+				pEnc->current->rounding_type, pEnc->current->mbs, &pEnc->vGMC);
 
-	if (vol_header)
-		BitstreamWriteVolHeader(bs, &pEnc->mbParam, pEnc->current);
-		
+	}
 
 	set_timecodes(pEnc->current,pEnc->reference,pEnc->mbParam.fbase);
+	if (vol_header)
+	{	BitstreamWriteVolHeader(bs, &pEnc->mbParam, pEnc->current);	
+		BitstreamPadAlways(bs);
+	}
+
 	BitstreamWriteVopHeader(bs, &pEnc->mbParam, pEnc->current, 1);
 
 	*pBits = BitstreamPos(bs);
@@ -1680,83 +1792,130 @@
 	pEnc->current->sStat.iTextBits = pEnc->current->sStat.iMvSum = pEnc->current->sStat.iMvCount = 
 		pEnc->current->sStat.kblks = pEnc->current->sStat.mblks = pEnc->current->sStat.ublks = 0;
 
-	for (y = 0; y < pEnc->mbParam.mb_height; y++) {
-		for (x = 0; x < pEnc->mbParam.mb_width; x++) {
+
+	for (y = 0; y < mb_height; y++) {
+		for (x = 0; x < mb_width; x++) {
 			MACROBLOCK *pMB =
 				&pEnc->current->mbs[x + y * pEnc->mbParam.mb_width];
 
+/* Mode decision: Check, if the block should be INTRA / INTER or GMC-coded */
+/* For a start, leave INTRA decision as is, only choose only between INTER/GMC  - gruel, 9.1.2002 */
+
 			bIntra = (pMB->mode == MODE_INTRA) || (pMB->mode == MODE_INTRA_Q);
 
-			if (!bIntra) {
+			if (bIntra) {
+				CodeIntraMB(pEnc, pMB);
+				MBTransQuantIntra(&pEnc->mbParam, pEnc->current, pMB, x, y,
+								  dct_codes, qcoeff);
+
 				start_timer();
-				MBMotionCompensation(pMB, x, y, &pEnc->reference->image,
-									 &pEnc->vInterH, &pEnc->vInterV,
-									 &pEnc->vInterHV, &pEnc->current->image,
-									 dct_codes, pEnc->mbParam.width,
-									 pEnc->mbParam.height,
-									 pEnc->mbParam.edged_width,
-									 pEnc->mbParam.m_quarterpel,
-									 pEnc->current->rounding_type);
-				stop_comp_timer();
-
-				if ((pEnc->current->global_flags & XVID_LUMIMASKING)) {
-					if (pMB->dquant != NO_CHANGE) {
-						pMB->mode = MODE_INTER_Q;
-						pEnc->current->quant += DQtab[pMB->dquant];
-						if (pEnc->current->quant > 31)
-							pEnc->current->quant = 31;
-						else if (pEnc->current->quant < 1)
-							pEnc->current->quant = 1;
-					}
+				MBPrediction(pEnc->current, x, y, pEnc->mbParam.mb_width, qcoeff);
+				stop_prediction_timer();
+
+				pEnc->current->sStat.kblks++;
+
+				MBCoding(pEnc->current, pMB, qcoeff, bs, &pEnc->current->sStat);
+				stop_coding_timer();
+				continue;
+			}
+				
+			if (pEnc->current->coding_type == S_VOP) {
+
+				int32_t iSAD = sad16(pEnc->current->image.y + 16*y*pEnc->mbParam.edged_width + 16*x,
+					pEnc->vGMC.y + 16*y*pEnc->mbParam.edged_width + 16*x, 
+					pEnc->mbParam.edged_width, 65536);
+				
+				if (pEnc->current->motion_flags & PMV_CHROMA16) {
+					iSAD += sad8(pEnc->current->image.u + 8*y*(pEnc->mbParam.edged_width/2) + 8*x,
+					pEnc->vGMC.u + 8*y*(pEnc->mbParam.edged_width/2) + 8*x, pEnc->mbParam.edged_width/2);
+
+					iSAD += sad8(pEnc->current->image.v + 8*y*(pEnc->mbParam.edged_width/2) + 8*x,
+					pEnc->vGMC.v + 8*y*(pEnc->mbParam.edged_width/2) + 8*x, pEnc->mbParam.edged_width/2);
 				}
-				pMB->quant = pEnc->current->quant;
 
-				pMB->field_pred = 0;
+				if (iSAD <= pMB->sad16) {		/* mode decision GMC */
 
-				if (pMB->mode != MODE_NOT_CODED)
-					pMB->cbp =
-						MBTransQuantInter(&pEnc->mbParam, pEnc->current, pMB, x, y,
-										  dct_codes, qcoeff);
+					if (pEnc->mbParam.m_quarterpel)
+						pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv;
+					else
+						pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv;
+
+					pMB->mode = MODE_INTER;
+					pMB->mcsel = 1;
+					pMB->sad16 = iSAD;
+				} else {
+					pMB->mcsel = 0;
+				}
 			} else {
-				CodeIntraMB(pEnc, pMB);
-				MBTransQuantIntra(&pEnc->mbParam, pEnc->current, pMB, x, y,
-								  dct_codes, qcoeff);
+				pMB->mcsel = 0;	/* just a precaution */
 			}
 
 			start_timer();
-			MBPrediction(pEnc->current, x, y, pEnc->mbParam.mb_width, qcoeff);
-			stop_prediction_timer();
+			MBMotionCompensation(pMB, x, y, &pEnc->reference->image,
+								 &pEnc->vInterH, &pEnc->vInterV,
+								 &pEnc->vInterHV, &pEnc->vGMC, 
+								 &pEnc->current->image,
+								 dct_codes, pEnc->mbParam.width,
+								 pEnc->mbParam.height,
+								 pEnc->mbParam.edged_width,
+								 pEnc->mbParam.m_quarterpel,
+								 (pEnc->current->global_flags & XVID_REDUCED),
+								 pEnc->current->rounding_type);
+
+			stop_comp_timer();
+
+			if ((pEnc->current->global_flags & XVID_LUMIMASKING)) {
+				if (pMB->dquant != NO_CHANGE) {
+					pMB->mode = MODE_INTER_Q;
+					pEnc->current->quant += DQtab[pMB->dquant];
+					if (pEnc->current->quant > 31)
+						pEnc->current->quant = 31;
+					else if (pEnc->current->quant < 1)
+						pEnc->current->quant = 1;
+				}
+			}
+			pMB->quant = pEnc->current->quant;
 
-			if (pMB->mode == MODE_INTRA || pMB->mode == MODE_INTRA_Q) {
-				pEnc->current->sStat.kblks++;
-			} else if (pMB->cbp || pMB->mvs[0].x || pMB->mvs[0].y ||
-					   pMB->mvs[1].x || pMB->mvs[1].y || pMB->mvs[2].x ||
-					   pMB->mvs[2].y || pMB->mvs[3].x || pMB->mvs[3].y) {
+			pMB->field_pred = 0;
+
+			if (pMB->mode != MODE_NOT_CODED)
+			{	pMB->cbp =
+					MBTransQuantInter(&pEnc->mbParam, pEnc->current, pMB, x, y,
+									  dct_codes, qcoeff);
+			}
+
+			if (pMB->cbp || pMB->mvs[0].x || pMB->mvs[0].y ||
+				   pMB->mvs[1].x || pMB->mvs[1].y || pMB->mvs[2].x ||
+				   pMB->mvs[2].y || pMB->mvs[3].x || pMB->mvs[3].y) {
 				pEnc->current->sStat.mblks++;
 			}  else {
 				pEnc->current->sStat.ublks++;
-			} 
-
+			}
+			
 			start_timer();
 
 			/* Finished processing the MB, now check if to CODE or SKIP */
 
-			skip_possible = (pMB->cbp == 0) & (pMB->mode == MODE_INTER) &
+			skip_possible = (pMB->cbp == 0) && (pMB->mode == MODE_INTER) &&
 							(pMB->dquant == NO_CHANGE);
-
-			if(pEnc->mbParam.m_quarterpel)
-			{	skip_possible &= (pMB->qmvs[0].x == pEnc->current->GMC_MV.x) & (pMB->qmvs[0].y == pEnc->current->GMC_MV.y);
-			}
-			else
-			{	skip_possible &= (pMB->mvs[0].x == pEnc->current->GMC_MV.x) & (pMB->mvs[0].y == pEnc->current->GMC_MV.y);
-			}
 			
+			if (pEnc->current->coding_type == S_VOP)
+				skip_possible &= (pMB->mcsel == 1);
+			else if (pEnc->current->coding_type == P_VOP) {
+				if (pEnc->mbParam.m_quarterpel)
+					skip_possible &= ( (pMB->qmvs[0].x == 0) && (pMB->qmvs[0].y == 0) );
+				else 
+					skip_possible &= ( (pMB->mvs[0].x == 0) && (pMB->mvs[0].y == 0) );
+			}
+
 			if ( (pMB->mode == MODE_NOT_CODED) || (skip_possible)) {
 
 /* This is a candidate for SKIPping, but for P-VOPs check intermediate B-frames first */
-				int bSkip = 1; 
 
 				if (pEnc->current->coding_type == P_VOP)	/* special rule for P-VOP's SKIP */
+				{
+					int bSkip = 1; 
+				
 					for (k=pEnc->bframenum_head; k< pEnc->bframenum_tail; k++)
 					{
 						int iSAD;
@@ -1769,42 +1928,85 @@
 						}
 					}
 					
-				if (!bSkip)
-				{
-					VECTOR predMV;
-					if(pEnc->mbParam.m_quarterpel) {
-						predMV = get_qpmv2(pEnc->current->mbs, pEnc->mbParam.mb_width, 0, x, y, 0);
-						pMB->pmvs[0].x = pMB->qmvs[0].x - predMV.x;  /* with GMC, qmvs doesn't have to be (0,0)! */
-						pMB->pmvs[0].y = pMB->qmvs[0].y - predMV.y; 
-					}
-					else {
-						predMV = get_pmv2(pEnc->current->mbs, pEnc->mbParam.mb_width, 0, x, y, 0);
-						pMB->pmvs[0].x = pMB->mvs[0].x - predMV.x; /* with GMC, mvs doesn't have to be (0,0)! */
-						pMB->pmvs[0].y = pMB->mvs[0].y - predMV.y; 
+					if (!bSkip) {	/* no SKIP, but trivial block */
+						if(pEnc->mbParam.m_quarterpel) {
+							VECTOR predMV = get_qpmv2(pEnc->current->mbs, pEnc->mbParam.mb_width, 0, x, y, 0);
+							pMB->pmvs[0].x = - predMV.x; 
+							pMB->pmvs[0].y = - predMV.y; 
+						}
+						else {
+							VECTOR predMV = get_pmv2(pEnc->current->mbs, pEnc->mbParam.mb_width, 0, x, y, 0);
+							pMB->pmvs[0].x = - predMV.x; 
+							pMB->pmvs[0].y = - predMV.y; 
+						}
+						pMB->mode = MODE_INTER;
+						pMB->cbp = 0;
+						MBCoding(pEnc->current, pMB, qcoeff, bs, &pEnc->current->sStat);
+						stop_coding_timer();
+
+						continue;	/* next MB */
 					}
-					pMB->mode = MODE_INTER;
-					pMB->cbp = 0;
-					MBCoding(pEnc->current, pMB, qcoeff, bs, &pEnc->current->sStat);
-				}
-				else 
-				{
-					pMB->mode = MODE_NOT_CODED;
-					MBSkip(bs);
 				}
-			
+				/* do SKIP */
+
+				pMB->mode = MODE_NOT_CODED;
+				MBSkip(bs);
+				stop_coding_timer();
+				continue;	/* next MB */
+			}
+			/* ordinary case: normal coded INTER/INTER4V block */
+
+			if (pEnc->current->global_flags & XVID_GREYSCALE)
+			{	pMB->cbp &= 0x3C;		/* keep only bits 5-2 */
+				qcoeff[4*64+0]=0;		/* zero, because DC for INTRA MBs DC value is saved */
+				qcoeff[5*64+0]=0;
+			}
+
+			if(pEnc->mbParam.m_quarterpel) {
+				VECTOR predMV = get_qpmv2(pEnc->current->mbs, pEnc->mbParam.mb_width, 0, x, y, 0);
+				pMB->pmvs[0].x = pMB->qmvs[0].x - predMV.x;  
+				pMB->pmvs[0].y = pMB->qmvs[0].y - predMV.y; 
+				DPRINTF(DPRINTF_MV,"mv_diff (%i,%i) pred (%i,%i) result (%i,%i)", pMB->pmvs[0].x, pMB->pmvs[0].y, predMV.x, predMV.y, pMB->mvs[0].x, pMB->mvs[0].y);
 			} else {
-				if (pEnc->current->global_flags & XVID_GREYSCALE)
-				{	pMB->cbp &= 0x3C;		/* keep only bits 5-2 */
-					qcoeff[4*64+0]=0;		/* zero, because DC for INTRA MBs DC value is saved */
-					qcoeff[5*64+0]=0;
-				}
-				MBCoding(pEnc->current, pMB, qcoeff, bs, &pEnc->current->sStat);
+				VECTOR predMV = get_pmv2(pEnc->current->mbs, pEnc->mbParam.mb_width, 0, x, y, 0);
+				pMB->pmvs[0].x = pMB->mvs[0].x - predMV.x; 
+				pMB->pmvs[0].y = pMB->mvs[0].y - predMV.y; 
+				DPRINTF(DPRINTF_MV,"mv_diff (%i,%i) pred (%i,%i) result (%i,%i)", pMB->pmvs[0].x, pMB->pmvs[0].y, predMV.x, predMV.y, pMB->mvs[0].x, pMB->mvs[0].y);
 			}
 
+
+			if (pMB->mode == MODE_INTER4V)
+			{	int k;
+				for (k=1;k<4;k++)
+				{
+					if(pEnc->mbParam.m_quarterpel) {
+						VECTOR predMV = get_qpmv2(pEnc->current->mbs, pEnc->mbParam.mb_width, 0, x, y, k);
+						pMB->pmvs[k].x = pMB->qmvs[k].x - predMV.x;  
+						pMB->pmvs[k].y = pMB->qmvs[k].y - predMV.y; 
+				DPRINTF(DPRINTF_MV,"mv_diff (%i,%i) pred (%i,%i) result (%i,%i)", pMB->pmvs[k].x, pMB->pmvs[k].y, predMV.x, predMV.y, pMB->mvs[k].x, pMB->mvs[k].y);
+					} else {
+						VECTOR predMV = get_pmv2(pEnc->current->mbs, pEnc->mbParam.mb_width, 0, x, y, k);
+						pMB->pmvs[k].x = pMB->mvs[k].x - predMV.x; 
+						pMB->pmvs[k].y = pMB->mvs[k].y - predMV.y; 
+				DPRINTF(DPRINTF_MV,"mv_diff (%i,%i) pred (%i,%i) result (%i,%i)", pMB->pmvs[k].x, pMB->pmvs[k].y, predMV.x, predMV.y, pMB->mvs[k].x, pMB->mvs[k].y);
+					}
+
+				}
+			}
+			
+			MBCoding(pEnc->current, pMB, qcoeff, bs, &pEnc->current->sStat);
 			stop_coding_timer();
+
 		}
 	}
 
+	if ((pEnc->current->global_flags & XVID_REDUCED))
+	{
+		image_deblock_rrv(&pEnc->current->image, pEnc->mbParam.edged_width, 
+			pEnc->current->mbs, mb_width, mb_height, pEnc->mbParam.mb_width,
+			16, XVID_DEC_DEBLOCKY|XVID_DEC_DEBLOCKUV);
+	}
+
 	emms();
 
 	if (pEnc->current->global_flags & XVID_HINTEDME_GET) {
@@ -1835,12 +2037,12 @@
 	pEnc->fMvPrevSigma = fSigma;
 
 	/* frame drop code */
-	// DPRINTF(DPRINTF_DEBUG, "kmu %i %i %i", pEnc->current->sStat.kblks, pEnc->current->sStat.mblks, pEnc->current->sStat.ublks);
+	DPRINTF(DPRINTF_DEBUG, "kmu %i %i %i", pEnc->current->sStat.kblks, pEnc->current->sStat.mblks, pEnc->current->sStat.ublks);
 	if (pEnc->current->sStat.kblks + pEnc->current->sStat.mblks <
-		(pEnc->frame_drop_ratio * pEnc->mbParam.mb_width * pEnc->mbParam.mb_height) / 100)
+		(pEnc->mbParam.frame_drop_ratio * mb_width * mb_height) / 100)
 	{
 		pEnc->current->sStat.kblks = pEnc->current->sStat.mblks = 0;
-		pEnc->current->sStat.ublks = pEnc->mbParam.mb_width * pEnc->mbParam.mb_height;
+		pEnc->current->sStat.ublks = mb_width * mb_height;
 
 		BitstreamReset(bs);
 
@@ -1855,10 +2057,25 @@
 		pEnc->current->fcode = pEnc->reference->fcode;
 		pEnc->current->bcode = pEnc->reference->bcode;
 		image_copy(&pEnc->current->image, &pEnc->reference->image, pEnc->mbParam.edged_width, pEnc->mbParam.height);
-		memcpy(pEnc->current->mbs, pEnc->reference->mbs, sizeof(MACROBLOCK) * pEnc->mbParam.mb_width * pEnc->mbParam.mb_height);
-
+		memcpy(pEnc->current->mbs, pEnc->reference->mbs, sizeof(MACROBLOCK) * mb_width * mb_height);
 	}
 
+	/* XXX: debug
+	{
+		char s[100];
+		sprintf(s, "\\%05i_cur.pgm", pEnc->m_framenum);
+		image_dump_yuvpgm(&pEnc->current->image, 
+			pEnc->mbParam.edged_width,
+			pEnc->mbParam.width, pEnc->mbParam.height, s);
+		
+		sprintf(s, "\\%05i_ref.pgm", pEnc->m_framenum);
+		image_dump_yuvpgm(&pEnc->reference->image, 
+			pEnc->mbParam.edged_width,
+			pEnc->mbParam.width, pEnc->mbParam.height, s);
+	} 
+	*/
+
+
 	*pBits = BitstreamPos(bs) - *pBits;
 
 	return 0;					// inter
@@ -1871,8 +2088,8 @@
 		   Bitstream * bs,
 		   uint32_t * pBits)
 {
-	int16_t dct_codes[6 * 64];
-	int16_t qcoeff[6 * 64];
+	DECLARE_ALIGNED_MATRIX(dct_codes, 6, 64, int16_t, CACHE_LINE);
+	DECLARE_ALIGNED_MATRIX(qcoeff, 6, 64, int16_t, CACHE_LINE);
 	uint32_t x, y;
 
 	IMAGE *f_ref = &pEnc->reference->image;
@@ -1885,6 +2102,8 @@
 		fprintf(fp,"Y=%3d   X=%3d   MB=%2d   CBP=%02X\n",y,x,mb->mode,mb->cbp); \
 	}
 
+	pEnc->current->global_flags &= ~XVID_REDUCED;	/* reduced resoltion not yet supported */
+
 	if (!first){
 		fp=fopen("C:\\XVIDDBGE.TXT","w");
 	}
@@ -1946,7 +2165,7 @@
 	for (y = 0; y < pEnc->mbParam.mb_height; y++) {
 		for (x = 0; x < pEnc->mbParam.mb_width; x++) {
 			MACROBLOCK * const mb = &frame->mbs[x + y * pEnc->mbParam.mb_width];
-			int direction = pEnc->global & XVID_ALTERNATESCAN ? 2 : 0;
+			int direction = pEnc->mbParam.global & XVID_ALTERNATESCAN ? 2 : 0;
 
 			// decoder ignores mb when refence block is INTER(0,0), CBP=0 
 			if (mb->mode == MODE_NOT_CODED) {
@@ -1997,18 +2216,3 @@
 	}
 #endif
 }
-
-
-/*	in case internal output is needed somewhere... */
-/*	{
-        FILE *filehandle;
-        filehandle=fopen("last-b.pgm","wb");
-        if (filehandle)
-        {
-                fprintf(filehandle,"P5\n\n");           //
-                fprintf(filehandle,"%d %d 255\n",pEnc->mbParam.edged_width,pEnc->mbParam.edged_height);
-                fwrite(frame->image.y,pEnc->mbParam.edged_width,pEnc->mbParam.edged_height,filehandle);
-                fclose(filehandle);
-		}
-	}
-*/