nengel@2: /*
nengel@2:  * Copyright (c) 2009 TUDelft 
nengel@2:  * 
nengel@2:  * Cell Parallel SPU - 2DWave Macroblock Decoding. 
nengel@2:  */
nengel@2: 
nengel@2: /**
nengel@2:  * @file libavcodec/cell/spu/h264_main_spu.c
nengel@2:  * Cell Parallel SPU - 2DWave Macroblock Decoding
nengel@2:  * @author C C Chi <c.c.chi@student.tudelft.nl>
nengel@2:  * 
nengel@2:  * SIMD kernels 
nengel@2:  * H.264/AVC motion compensation
nengel@2:  * @author Mauricio Alvarez <alvarez@ac.upc.edu>
nengel@2:  * @author Albert Paradis <apar7632@hotmail.com>
nengel@2:  */ 
nengel@2: 
nengel@2: 
nengel@2: /* Enable this lines to enable simulator statistic or generate traces */
nengel@2: 
nengel@2: //#define ENABLE_SIMULATOR
nengel@2: //#define ENABLE_PARAVER_TRACING_CELL
nengel@2: 
nengel@2: #ifdef ENABLE_SIMULATOR
nengel@2: 	#include "/opt/ibm/systemsim-cell/include/callthru/spu/profile.h"
nengel@2: #endif
nengel@2: 
nengel@2: #ifdef ENABLE_TRACES
nengel@2: 	#include "spu_trace.h"
nengel@2: #endif
nengel@2: #include <unistd.h>
nengel@2: #include <stdio.h>
nengel@2: #include <spu_intrinsics.h>
nengel@2: #include <spu_mfcio.h>
nengel@2: #include <libsync.h>
nengel@2: #include <sys/time.h>
nengel@2: #include <assert.h>
nengel@2: 
nengel@2: //#include "dsputil_cell.h"
nengel@2: #include "types_spu.h"
nengel@2: #include "h264_intra_spu.h"
nengel@2: #include "h264_decode_mb_spu.h"
nengel@2: #include "h264_mc_spu.h"
nengel@2: #include "h264_tables.h"
nengel@2: #include "h264_dma.h"
nengel@2: 
nengel@2: 
nengel@2: /** functions for supporting tracing with paraver for the SPU 
nengel@2:  *
nengel@2:  */
nengel@2: inline void trace_init_SPU(){
nengel@2: #ifdef ENABLE_PARAVER_TRACING_CELL
nengel@2: 	SPUtrace_init ();
nengel@2: #endif
nengel@2: }
nengel@2: 
nengel@2: inline void trace_fini_SPU(){
nengel@2: #ifdef ENABLE_PARAVER_TRACING_CELL
nengel@2: 	SPUtrace_fini ();
nengel@2: #endif
nengel@2: }
nengel@2: 
nengel@2: inline void trace_event_SPU(int event, int id){
nengel@2: #ifdef ENABLE_PARAVER_TRACING_CELL
nengel@2: 	SPUtrace_event (event, id);
nengel@2: #else
nengel@2: 	(void) event;
nengel@2: 	(void) id;
nengel@2: #endif
nengel@2: }
nengel@2: 
nengel@2: // for simulator statistic
nengel@2: inline void clear_statistic(){
nengel@2: #ifdef ENABLE_SIMULATOR
nengel@2: 	prof_clear();
nengel@2: #endif
nengel@2: }
nengel@2: 
nengel@2: inline void start_statistic(){
nengel@2: #ifdef ENABLE_SIMULATOR
nengel@2: 	prof_start();
nengel@2: #endif
nengel@2: }
nengel@2: 
nengel@2: inline void stop_statistic(){
nengel@2: #ifdef ENABLE_SIMULATOR
nengel@2: 	prof_stop();
nengel@2: #endif
nengel@2: }
nengel@2: 
nengel@2: H264Context_spu h_context;  // struct that contain all the params to decode a macroblock
nengel@2: 
nengel@2: DECLARE_ALIGNED_16(spe_pos, dma_temp); //dma temp for sending
nengel@2: //mb position of neighbouring spes
nengel@2: DECLARE_ALIGNED_16(volatile spe_pos, src_spe); //written by SPE_ID -1
nengel@2: //DECLARE_ALIGNED_16(spe_pos, tgt_spe); //written by SPE_ID +1
nengel@2: 
nengel@2: /**	
nengel@2: *	Initializes the buffering of the mb data and associated mc data. The init_mb_buffer needs to 
nengel@2: *	be called before any get_next_mb and only once at the beginning of the slice.
nengel@2: *
nengel@2: *	Note: init_mc_buffer and get_next_mb expect the width of the picture to be more than 2 mb's
nengel@2: */
nengel@2: #define TAG_OFFSET_MB MBD_buf1
nengel@2: #define TAG_OFFSET_MC MBD_mc_buf1
nengel@2: static void init_mb_buffer(H264Context_spu* h){
nengel@2: 	H264slice *s = h->s;
nengel@2: 	H264Mb *next_mb;
nengel@2: 	int mb_height = s->mb_height;
nengel@2: 	int mb_width = s->mb_width;
nengel@2: 
nengel@2: 	h->mc_idx =0;
nengel@2: 	
nengel@2: 	h->mb_dec = 0;
nengel@2: 	h->mb_mc = 0;
nengel@2: 	h->mb_dma = 0;
nengel@2: 		
nengel@2: 	h->curr_line %= mb_height;
nengel@2: 	h->next_mb_idx = h->curr_line * mb_width;
nengel@2: 	h->mb_id = h->curr_line * mb_width;
nengel@2: 	h->n_mc= h->curr_line * mb_width;
nengel@2: 	
nengel@2: 	next_mb = s->blocks + h->mb_id;
nengel@2: 	spu_dma_get(&h->mb_buf[h->mb_dma], (unsigned) next_mb, sizeof(H264Mb), h->mb_dma + TAG_OFFSET_MB);
nengel@2: 	h->mb_dma++;
nengel@2: 	h->mb_id++;
nengel@2: 	
nengel@2: 	next_mb = s->blocks + h->mb_id;
nengel@2: 	spu_dma_get(&h->mb_buf[h->mb_dma], (unsigned) next_mb, sizeof(H264Mb), h->mb_dma + TAG_OFFSET_MB);
nengel@2: 	h->mb_dma++;
nengel@2: 	h->mb_id++;
nengel@2: 	wait_dma_id(0 + TAG_OFFSET_MB);	
nengel@2: 	
nengel@2: 	H264Mb *mb = &h->mb_buf[0];
nengel@2: 	H264mc *mc = &h->mc_buf[0];
nengel@2: 	if(!IS_INTRA(mb->mb_type)){
nengel@2: 		calc_mc_params(mb, mc);
nengel@2: 		fill_ref_buf(h, mb, mc);
nengel@2: 	}
nengel@2: 	h->n_mc++;
nengel@2: 	h->mb_mc++;
nengel@2: }
nengel@2: 
nengel@2: static void *get_next_mb(H264Context_spu *h){
nengel@2: 	H264slice *s = h->s;
nengel@2: 	H264spe *spe = &h->spe;
nengel@2: 	H264Mb *mb_buf = h->mb_buf;	
nengel@2: 	H264mc *mc_buf = h->mc_buf;
nengel@2: 	H264Mb *next_mb;
nengel@2: 	H264Mb *next_dma_mb;
nengel@2: 	
nengel@2: 	if (h->curr_line >= s->mb_height)
nengel@2: 		return NULL;
nengel@2: 	
nengel@2: 	if (h->mb_id < h->mb_total){
nengel@2: 		next_dma_mb = s->blocks + h->mb_id;
nengel@2: 		spu_dma_get(&mb_buf[h->mb_dma], (unsigned) next_dma_mb, sizeof(H264Mb), h->mb_dma + TAG_OFFSET_MB);
nengel@2: 		h->mb_dma = (h->mb_dma+1)%3;
nengel@2: 		h->mb_id++;
nengel@2: 		if (h->mb_id%s->mb_width ==0){
nengel@2: 			h->mb_id+=(spe->spe_total-1)*s->mb_width;			
nengel@2: 		}
nengel@2: 	}
nengel@2: 	
nengel@2: 	h->mc = &mc_buf[h->mc_idx];
nengel@2: 	wait_dma_id(h->mc_idx + TAG_OFFSET_MC);
nengel@2: 	h->mc_idx = (h->mc_idx+1)%2;
nengel@2: 	if (h->n_mc < h->mb_total){
nengel@2: 		wait_dma_id(h->mb_mc + TAG_OFFSET_MB);
nengel@2: 		H264Mb *mb = &mb_buf[h->mb_mc];
nengel@2: 		H264mc *mc = &mc_buf[h->mc_idx];
nengel@2: 		if(!IS_INTRA(mb->mb_type)){
nengel@2: 			calc_mc_params(mb, mc);
nengel@2: 			fill_ref_buf(h, mb, mc);
nengel@2: 		}
nengel@2: 		h->n_mc++;
nengel@2: 		if (h->n_mc%s->mb_width ==0){
nengel@2: 			h->n_mc+=(spe->spe_total-1)*s->mb_width;			
nengel@2: 		}
nengel@2: 	}
nengel@2: 	h->next_mb_idx++;
nengel@2: 	if (h->next_mb_idx % s->mb_width ==0){
nengel@2: 		h->next_mb_idx+=(spe->spe_total-1)*s->mb_width;
nengel@2: 		h->curr_line+=spe->spe_total;		
nengel@2: 	}
nengel@2: 	
nengel@2: 	h->mb_mc = (h->mb_mc+1)%3;	
nengel@2: 	next_mb = &mb_buf[h->mb_dec];
nengel@2: 	h->mb_dec = (h->mb_dec+1)%3;
nengel@2: 	return next_mb;
nengel@2: }
nengel@2: 
nengel@2: static void *get_next_mb_blocking(H264Context_spu *h){
nengel@2: 	H264slice *s = h->s;
nengel@2: 	H264spe *spe = &h->spe;
nengel@2: 	H264Mb *mb_buf = h->mb_buf;
nengel@2: 	H264mc *mc_buf = h->mc_buf;
nengel@2: 	H264Mb *next_mb;
nengel@2: 	H264Mb *next_dma_mb;
nengel@2: 
nengel@2: 	if (h->mb_id >= h->mb_total)
nengel@2: 		return NULL;
nengel@2: 
nengel@2: 	//printf("%d\n", h->mb_id);
nengel@2: 	next_dma_mb = s->blocks + h->mb_id;
nengel@2: 	spu_dma_get(&mb_buf[0], (unsigned) next_dma_mb, sizeof(H264Mb), MBD_buf1);
nengel@2: 	//h->mb_dma = (h->mb_dma+1)%3;
nengel@2: 	h->mb_id++;
nengel@2: 	if (h->mb_id%s->mb_width ==0){
nengel@2: 		h->mb_id+=(spe->spe_total-1)*s->mb_width;
nengel@2: 	}
nengel@2: 	wait_dma_id(MBD_buf1);
nengel@2: 
nengel@2: 	h->mc = &mc_buf[0];	
nengel@2: 	//h->mc_idx = (h->mc_idx+1)%2;
nengel@2: 	//if (h->n_mc < h->mb_total){
nengel@2: 	H264Mb *mb = &mb_buf[0];
nengel@2: 	H264mc *mc = &mc_buf[0];
nengel@2: 	if(!IS_INTRA(mb->mb_type)){
nengel@2: 		calc_mc_params(mb, mc);
nengel@2: 		fill_ref_buf(h, mb, mc);
nengel@2: 	}
nengel@2: 	//h->n_mc++;
nengel@2: 	/*if (h->n_mc%s->mb_width ==0){
nengel@2: 		h->n_mc+=(spe->spe_total-1)*s->mb_width;
nengel@2: 	}*/	
nengel@2: //	wait_dma_id(MBD_mc_buf1);
nengel@2: 
nengel@2: // 	h->next_mb_idx++;
nengel@2: // 	if (h->next_mb_idx % s->mb_width ==0){
nengel@2: // 		h->next_mb_idx+=(spe->spe_total-1)*s->mb_width;
nengel@2: // 		h->curr_line+=spe->spe_total;
nengel@2: // 	}
nengel@2: 
nengel@2: // 	h->mb_mc = (h->mb_mc+1)%3;
nengel@2: 	next_mb = &mb_buf[0];
nengel@2: // 	h->mb_dec = (h->mb_dec+1)%3;
nengel@2: 	return next_mb;
nengel@2: }
nengel@2: 
nengel@2: 
nengel@2: #undef TAG_OFFSET_MB
nengel@2: #undef TAG_OFFSET_MC
nengel@2: static inline int dep_resolved(H264Context_spu *h){
nengel@2: 	H264slice *s = h->s;
nengel@2: 	int spe_id = h->spe.spe_id;
nengel@2: 	volatile int mb_proc_dep = src_spe.count;
nengel@2: 	if (spe_id==0)
nengel@2: 		return (h->mb_proc < mb_proc_dep-1 +s->mb_width)? 1:0;
nengel@2: 	else
nengel@2: 		return (h->mb_proc < mb_proc_dep-1)? 1:0;
nengel@2: }
nengel@2: 
nengel@2: void update_tgt_spe_dep(H264Context_spu *h, int end){
nengel@2: 	H264Mb *mb = h->mb;
nengel@2: 	H264slice *s = h->s;
nengel@2: 	H264spe *spe = &h->spe;
nengel@2: 	int mb_x = mb->mb_x;
nengel@2: 	
nengel@2: 	if (end || (mb_x%2==0 && mb_x!=0) || mb_x==s->mb_width-1){
nengel@2: 		spe_pos* dma_spe = &dma_temp;
nengel@2: 		spe_pos* tgt_spe = (spe_pos*) ((unsigned) spe->tgt_spe + (unsigned) &src_spe); //located in target spe local store
nengel@2: 		dma_spe->count = end? h->mb_proc+1: h->mb_proc;
nengel@2: 		spu_dma_barrier_put(dma_spe, (unsigned) tgt_spe, sizeof(dma_temp), MBD_put);
nengel@2: 	}
nengel@2: 	h->mb_proc++;
nengel@2: }
nengel@2: 
nengel@2: 
nengel@2: int main(unsigned long long id, unsigned long long argp)
nengel@2: {
nengel@2: 	(void) id;
nengel@2: 	H264Context_spu* h = &h_context;
nengel@2: 	H264spe *spe_params = (H264spe *) (unsigned) argp;    
nengel@2: 	
nengel@2: 	spu_dma_get(&h->spe, (unsigned) spe_params, sizeof(H264spe), MBD_slice); //ID_slice is used out of convienience
nengel@2: 	wait_dma_id(MBD_slice);
nengel@2: 
nengel@2:     //clear_statistic();
nengel@2:     dsputil_h264_init_cell(&h->dsp);
nengel@2:     ff_cropTbl_init();
nengel@2:     init_pred_ptrs(&h->hpc);
nengel@2: 
nengel@2: 	//send slice_buf to ppe
nengel@2: 	spu_write_out_mbox((unsigned) h->slice_buf);
nengel@2: 	h->sl_idx=0;
nengel@2: 	// initialize tracing with paraver
nengel@2:     //trace_init_SPU();
nengel@2: 	h->frames =0;	
nengel@2: 	src_spe.count =0;
nengel@2: 	h->mb_proc = 0;
nengel@2: 
nengel@2: 	h->mb_id=0;
nengel@2: 	h->mc_idx=0;
nengel@2: 	h->mb_dec=0;
nengel@2: 	h->mb_mc=0;
nengel@2: 	h->mb_dma=0;
nengel@2: 	h->next_mb_idx=0;
nengel@2: 
nengel@2: 	h->blocking=0;
nengel@2: 
nengel@2: 
nengel@2: 	H264spe* p = &h->spe;
nengel@2: 	h->curr_line =p->spe_id;
nengel@2: 	h->mb_total = p->mb_height*p->mb_width;
nengel@2: 	int stride_y = 32;
nengel@2: 	int stride_c = 16;
nengel@2: 	//init block_offset array
nengel@2: 	init_block_offset(stride_y, stride_c);
nengel@2: 	for(;;){
nengel@2: 		spu_read_in_mbox();
nengel@2: 
nengel@2: 		h->s = &h->slice_buf[h->sl_idx];
nengel@2: 		h->sl_idx++; h->sl_idx%=2;
nengel@2: 
nengel@2: 		if (h->s->state< 0){			
nengel@2: 			break;
nengel@2: 		}
nengel@2: 
nengel@2: 		{
nengel@2: 			if(!h->blocking){
nengel@2: 				init_mb_buffer(h);
nengel@2: 				while((h->mb=(H264Mb *)get_next_mb(h))){
nengel@2: 					while(!dep_resolved(h));
nengel@2: 					//printf("frame %d mbx %d\t mby %d id %d\n", h->frames, h->mb->mb_x, h->mb->mb_y, p-	>spe_id);
nengel@2: 					hl_decode_mb_internal(h, stride_y, stride_c);
nengel@2: 				}
nengel@2: 				update_tgt_spe_dep(h, 1);
nengel@2: 			}else{
nengel@2: 				h->mb_id=0;
nengel@2: 				while((h->mb=(H264Mb *)get_next_mb_blocking(h))){
nengel@2: 					while(!dep_resolved(h));
nengel@2: 					//printf("frame %d mbx %d\t mby %d id %d\n", h->frames, h->mb->mb_x, h->mb->mb_y, p-	>spe_id);
nengel@2: 					hl_decode_mb_internal(h, stride_y, stride_c);
nengel@2: 				}
nengel@2: 				update_tgt_spe_dep(h, 1);
nengel@2: 			}
nengel@2: 			
nengel@2: 		}
nengel@2: 
nengel@2: 		h->frames++;
nengel@2: 		
nengel@2: 		if (p->spe_id == ((h->frames*p->mb_height -1)%p->spe_total)){
nengel@2: 			//printf("spe %d, %d\n", atomic_read(p->rl_cnt), h->frames);
nengel@2: 			//MBSlice is copied beforehand.
nengel@2: 			//only inc cnt.
nengel@2: 			atomic_inc(p->rl_cnt);		
nengel@2: 		}
nengel@2: 		{
nengel@2: 			atomic_dec(p->cnt);
nengel@2: 		}
nengel@2: 	}
nengel@2: 	
nengel@2: 	return 0;
nengel@2: }
nengel@2: