Context Navigation

source: soft/giet_vm/convol/main.c @ 405

Last change on this file since 405 was 402, checked in by alain, 10 years ago
supporting X_SIZE = 16 and Y_SIZE = 16
File size: 24.8 KB

Rev	Line
[334]	1	////////////////////////////////////////////////////////////////////////////////////////////
	2	// File : main.c (for convol application)
	3	// Date : june 2014
	4	// author : Alain Greiner
	5	//
	6	// The "convol" application implements a 2D convolution product.
	7	////////////////////////////////////////////////////////////////////////////////////////////
	8
	9	#include "hard_config.h"
	10	#include "stdio.h"
[353]	11	#include "stdlib.h"
[334]	12	#include "barrier.h"
[384]	13	#include "malloc.h"
[334]	14
[372]	15	#define USE_SBT_BARRIER 1
[353]	16	#define VERBOSE 0
	17	#define SUPER_VERBOSE 0
[334]	18
[353]	19	#define INITIAL_DISPLAY_ENABLE 0
[334]	20	#define FINAL_DISPLAY_ENABLE 1
	21
[353]	22	#define NB_CLUSTERS (X_SIZE * Y_SIZE)
[334]	23	#define PIXEL_SIZE 2
	24	#define NL 1024
	25	#define NP 1024
	26	#define NB_PIXELS (NP * NL)
	27	#define FRAME_SIZE (NB_PIXELS * PIXEL_SIZE)
	28
	29	#define TA(c,l,p) (A[c][((NP) * (l)) + (p)])
	30	#define TB(c,p,l) (B[c][((NL) * (p)) + (l)])
	31	#define TC(c,l,p) (C[c][((NP) * (l)) + (p)])
	32	#define TD(c,l,p) (D[c][((NP) * (l)) + (p)])
	33	#define TZ(c,l,p) (Z[c][((NP) * (l)) + (p)])
	34
	35	#define max(x,y) ((x) > (y) ? (x) : (y))
	36	#define min(x,y) ((x) < (y) ? (x) : (y))
	37
[372]	38	// global instrumentation counters (cluster_id, lpid]
	39
[353]	40	unsigned int START[NB_CLUSTERS][NB_PROCS_MAX];
	41	unsigned int H_BEG[NB_CLUSTERS][NB_PROCS_MAX];
	42	unsigned int H_END[NB_CLUSTERS][NB_PROCS_MAX];
	43	unsigned int V_BEG[NB_CLUSTERS][NB_PROCS_MAX];
	44	unsigned int V_END[NB_CLUSTERS][NB_PROCS_MAX];
	45	unsigned int D_BEG[NB_CLUSTERS][NB_PROCS_MAX];
	46	unsigned int D_END[NB_CLUSTERS][NB_PROCS_MAX];
[334]	47
[372]	48	// global synchronization barrier
[334]	49
[372]	50	#if USE_SBT_BARRIER
	51	giet_sbt_barrier_t barrier;
	52	#else
	53	giet_barrier_t barrier;
	54	#endif
[334]	55
[372]	56	volatile unsigned int barrier_init_ok = 0;
	57	volatile unsigned int load_image_ok = 0;
	58	volatile unsigned int instrumentation_ok = 0;
[334]	59
[372]	60	// global pointers on distributed buffers in all clusters
	61	unsigned short * GA[NB_CLUSTERS];
	62	int * GB[NB_CLUSTERS];
	63	int * GC[NB_CLUSTERS];
	64	int * GD[NB_CLUSTERS];
	65	unsigned char * GZ[NB_CLUSTERS];
	66
[334]	67	///////////////////////////////////////////
	68	__attribute__ ((constructor)) void main()
	69	///////////////////////////////////////////
	70	{
	71	//////////////////////////////////
	72	// convolution kernel parameters
	73	// The content of this section is
	74	// Philips proprietary information.
	75	///////////////////////////////////
	76
	77	int vnorm = 115;
	78	int vf[35] = { 1, 1, 2, 2, 2,
	79	2, 3, 3, 3, 4,
	80	4, 4, 4, 5, 5,
	81	5, 5, 5, 5, 5,
	82	5, 5, 4, 4, 4,
	83	4, 3, 3, 3, 2,
	84	2, 2, 2, 1, 1 };
	85
	86	int hrange = 100;
	87	int hnorm = 201;
	88
	89	unsigned int date = 0;
	90
	91	int c; // cluster index for loops
	92	int l; // line index for loops
	93	int p; // pixel index for loops
	94	int z; // vertical filter index for loops
	95
[353]	96	int file = 0; // file descriptor
[334]	97	unsigned int pid = giet_procid(); // processor id
	98	unsigned int nprocs = NB_PROCS_MAX; // procs per cluster
[362]	99	unsigned int nclusters = NB_CLUSTERS; // number of clusters
[334]	100	unsigned int lpid = pid % nprocs; // local task id
	101	unsigned int cluster_xy = pid / nprocs; // cluster index
	102	unsigned int x = cluster_xy >> Y_WIDTH; // x coordinate
	103	unsigned int y = cluster_xy & ((1<<Y_WIDTH)-1); // y coordinate
	104	unsigned int cluster_id = (x * Y_SIZE) + y; // continuous cluster index
	105	unsigned int task_id = (cluster_id * nprocs) + lpid; // continuous task index
	106	unsigned int ntasks = nclusters * nprocs; // number of tasks
	107	unsigned int npixels = NB_PIXELS; // pixels per frame
	108	unsigned int frame_size = FRAME_SIZE; // total size (bytes)
	109	unsigned int nblocks = frame_size / 512; // number of blocks per frame
	110
	111	unsigned int lines_per_task = NL / ntasks; // lines per task
	112	unsigned int lines_per_cluster = NL / nclusters; // lines per cluster
	113	unsigned int pixels_per_task = NP / ntasks; // columns per task
	114	unsigned int pixels_per_cluster = NP / nclusters; // columns per cluster
	115
	116	int first, last;
	117
	118	date = giet_proctime();
	119	START[cluster_id][lpid] = date;
	120
[353]	121	#if VERBOSE
	122	giet_shr_printf( "\n[CONVOL] task[%d,%d,%d] starts at cycle %d\n", x,y,lpid, date );
	123	#endif
	124
[334]	125	// parameters checking
	126
	127	if ((nprocs != 1) && (nprocs != 2) && (nprocs != 4))
	128	giet_exit( "[CONVOL ERROR] NB_PROCS_MAX must be 1, 2 or 4\n");
	129
[402]	130	if ((X_SIZE!=1) && (X_SIZE!=2) && (X_SIZE!=4) && (X_SIZE!=8) && (X_SIZE!=16))
	131	giet_exit( "[CONVOL ERROR] X_SIZE must be 1, 2, 4, 8, 16\n");
[334]	132
[402]	133	if ((Y_SIZE!=1) && (Y_SIZE!=2) && (Y_SIZE!=4) && (Y_SIZE!=8) && (Y_SIZE!=16))
	134	giet_exit( "[CONVOL ERROR] Y_SIZE must be 1, 2, 4, 8, 16\n");
[334]	135
	136	if ( NL % nclusters != 0 )
	137	giet_exit( "[CONVOL ERROR] NB_CLUSTERS must be a divider of NL");
	138
	139	if ( NP % nclusters != 0 )
	140	giet_exit( "[CONVOL ERROR] NB_CLUSTERS must be a divider of NP");
	141
[372]	142
[334]	143	///////////////////////////////////////////////////////////////////
[372]	144	// task[0][0][0] makes barrier initialisation
[334]	145	///////////////////////////////////////////////////////////////////
[372]	146
	147	if ( pid == 0 )
	148	{
	149	giet_shr_printf("\n[CONVOL] task[0,0,0] starts barrier init at cycle %d\n"
	150	"- NB_CLUSTERS = %d\n"
	151	"- NB_LOCAL_PROCS = %d\n"
	152	"- NB_TASKS = %d\n"
	153	"- NB_PIXELS = %x\n"
	154	"- FRAME_SIZE = %x\n"
	155	"- NB_BLOCKS = %x\n",
	156	giet_proctime(), nclusters, nprocs, ntasks,
	157	npixels, frame_size, nblocks );
	158	#if USE_SBT_BARRIER
	159	sbt_barrier_init( &barrier, ntasks );
	160	#else
	161	barrier_init( &barrier, ntasks );
	162	#endif
[334]	163
[372]	164	giet_shr_printf( "\n[CONVOL] task[0,0,0] completes barrier init at cycle %d\n",
	165	giet_proctime() );
	166
	167	barrier_init_ok = 1;
	168	}
	169	else
	170	{
	171	while ( barrier_init_ok == 0 );
	172	}
	173
	174	///////////////////////////////////////////////////////////////////
	175	// All task[x][y][0] allocate the global buffers in cluster(x,y)
	176	// These buffers mut be sector-aligned.
	177	///////////////////////////////////////////////////////////////////
	178	if ( lpid == 0 )
	179	{
	180
	181	#if VERBOSE
	182	giet_shr_printf( "\n[CONVOL] task[%d,%d,%d] enters malloc at cycle %d\n", x,y,lpid, date );
	183	#endif
	184
[384]	185	GA[cluster_id] = remote_malloc( (FRAME_SIZE/nclusters) , x , y );
	186	GB[cluster_id] = remote_malloc( (FRAME_SIZE/nclusters)*2 , x , y );
	187	GC[cluster_id] = remote_malloc( (FRAME_SIZE/nclusters)*2 , x , y );
	188	GD[cluster_id] = remote_malloc( (FRAME_SIZE/nclusters)*2 , x , y );
	189	GZ[cluster_id] = remote_malloc( (FRAME_SIZE/nclusters)/2 , x , y );
[372]	190
	191	#if VERBOSE
	192	giet_shr_printf( "\n[CONVOL] Shared Buffer Virtual Addresses in cluster(%d,%d)\n"
	193	"### GA = %x\n"
	194	"### GB = %x\n"
	195	"### GC = %x\n"
	196	"### GD = %x\n"
	197	"### GZ = %x\n",
	198	x, y,
	199	GA[cluster_id],
	200	GB[cluster_id],
	201	GC[cluster_id],
	202	GD[cluster_id],
	203	GZ[cluster_id] );
	204	#endif
	205	}
	206
[377]	207	///////////////////////////////
	208	#if USE_SBT_BARRIER
	209	sbt_barrier_wait( &barrier );
	210	#else
	211	barrier_wait( &barrier );
	212	#endif
[372]	213
	214	///////////////////////////////////////////////////////////////////
	215	// All tasks initialise in their private stack a copy of the
	216	// arrays of pointers on the shared, distributed buffers.
	217	///////////////////////////////////////////////////////////////////
	218
[353]	219	unsigned short * A[NB_CLUSTERS];
	220	int * B[NB_CLUSTERS];
	221	int * C[NB_CLUSTERS];
	222	int * D[NB_CLUSTERS];
	223	unsigned char * Z[NB_CLUSTERS];
[334]	224
	225	for (c = 0; c < nclusters; c++)
	226	{
[372]	227	A[c] = GA[c];
	228	B[c] = GB[c];
	229	C[c] = GC[c];
	230	D[c] = GD[c];
	231	Z[c] = GZ[c];
[334]	232	}
	233
	234	///////////////////////////////////////////////////////////////////////////
[372]	235	// task[0,0,0] open the file containing image, and load it from disk
	236	// to all A[c] buffers (nblocks / nclusters loaded in each cluster).
[334]	237	// Other tasks are waiting on the init_ok condition.
	238	//////////////////////////////////////////////////////////////////////////
	239	if ( pid == 0 )
	240	{
	241	// open file
	242	file = giet_fat_open("misc/philips_image.raw", 0 );
	243	if ( file < 0 ) giet_exit( "[CONVOL ERROR] task[0,0,0] cannot open"
	244	" file misc/philips_image.raw" );
	245
	246	giet_shr_printf( "\n[CONVOL] task[0,0,0] open file misc/philips_image.raw"
	247	" at cycle %d\n", giet_proctime() );
	248
[353]	249	for ( c = 0 ; c < NB_CLUSTERS ; c++ )
[334]	250	{
[353]	251	giet_shr_printf( "\n[CONVOL] task[0,0,0] starts load "
	252	"for cluster %d at cycle %d\n", c, giet_proctime() );
	253
[334]	254	giet_fat_read( file,
	255	A[c],
	256	nblocks/nclusters,
	257	(nblocks/nclusters)*c );
	258
[353]	259	giet_shr_printf( "\n[CONVOL] task[0,0,0] completes load "
	260	"for cluster %d at cycle %d\n", c, giet_proctime() );
[334]	261	}
[372]	262	load_image_ok = 1;
[334]	263	}
	264	else
	265	{
[372]	266	while ( load_image_ok == 0 );
[334]	267	}
	268
	269	/////////////////////////////////////////////////////////////////////////////
	270	// Optionnal parallel display of the initial image stored in A[c] buffers.
	271	// Eah task displays (NL/ntasks) lines. (one byte per pixel).
	272	/////////////////////////////////////////////////////////////////////////////
	273
	274	if ( INITIAL_DISPLAY_ENABLE )
	275	{
	276
[353]	277	#if VERBOSE
	278	giet_shr_printf( "\n[CONVOL] task[%d,%d,%d] starts initial display"
	279	" at cycle %d\n",
	280	x, y, lpid, giet_proctime() );
	281	#endif
	282
[334]	283	unsigned int line;
	284	unsigned int offset = lines_per_task * lpid;
	285
	286	for ( l = 0 ; l < lines_per_task ; l++ )
	287	{
	288	line = offset + l;
	289
	290	for ( p = 0 ; p < NP ; p++ )
	291	{
	292	TZ(cluster_id, line, p) = (unsigned char)(TA(cluster_id, line, p) >> 8);
	293	}
	294
	295	giet_fb_sync_write( NP(l + (task_id lines_per_task) ),
	296	&TZ(cluster_id, line, 0),
	297	NP);
	298	}
	299
[353]	300	#if VERBOSE
	301	giet_shr_printf( "\n[CONVOL] task[%d,%d,%d] completes initial display"
	302	" at cycle %d\n",
	303	x, y, lpid, giet_proctime() );
	304	#endif
[334]	305
[377]	306	////////////////////////////
	307	#if USE_SBT_BARRIER
[372]	308	sbt_barrier_wait( &barrier );
[377]	309	#else
[372]	310	barrier_wait( &barrier );
[377]	311	#endif
[334]	312
	313	}
	314
	315	////////////////////////////////////////////////////////
	316	// parallel horizontal filter :
	317	// B <= transpose(FH(A))
	318	// D <= A - FH(A)
	319	// Each task computes (NL/ntasks) lines
	320	// The image must be extended :
	321	// if (z<0) TA(cluster_id,l,z) == TA(cluster_id,l,0)
	322	// if (z>NP-1) TA(cluster_id,l,z) == TA(cluster_id,l,NP-1)
	323	////////////////////////////////////////////////////////
	324
	325	date = giet_proctime();
	326	H_BEG[cluster_id][lpid] = date;
	327
[353]	328	#if VERBOSE
	329	giet_shr_printf( "\n[CONVOL] task[%d,%d,%d] starts horizontal filter"
	330	" at cycle %d\n",
	331	x, y, lpid, date );
	332	#else
	333	if ( pid == 0 ) giet_shr_printf( "\n[CONVOL] task[0,0,0] starts horizontal filter"
	334	" at cycle %d\n", date );
	335	#endif
	336
[334]	337	// l = absolute line index / p = absolute pixel index
	338	// first & last define which lines are handled by a given task
	339
	340	first = task_id * lines_per_task;
	341	last = first + lines_per_task;
	342
	343	for (l = first; l < last; l++)
	344	{
	345	// src_c and src_l are the cluster index and the line index for A & D
	346	int src_c = l / lines_per_cluster;
	347	int src_l = l % lines_per_cluster;
	348
	349	// We use the specific values of the horizontal ep-filter for optimisation:
	350	// sum(p) = sum(p-1) + TA[p+hrange] - TA[p-hrange-1]
	351	// To minimize the number of tests, the loop on pixels is split in three domains
	352
	353	int sum_p = (hrange + 2) * TA(src_c, src_l, 0);
	354	for (z = 1; z < hrange; z++)
	355	{
	356	sum_p = sum_p + TA(src_c, src_l, z);
	357	}
	358
	359	// first domain : from 0 to hrange
	360	for (p = 0; p < hrange + 1; p++)
	361	{
	362	// dst_c and dst_p are the cluster index and the pixel index for B
	363	int dst_c = p / pixels_per_cluster;
	364	int dst_p = p % pixels_per_cluster;
	365	sum_p = sum_p + (int) TA(src_c, src_l, p + hrange) - (int) TA(src_c, src_l, 0);
	366	TB(dst_c, dst_p, l) = sum_p / hnorm;
	367	TD(src_c, src_l, p) = (int) TA(src_c, src_l, p) - sum_p / hnorm;
	368	}
	369	// second domain : from (hrange+1) to (NP-hrange-1)
	370	for (p = hrange + 1; p < NP - hrange; p++)
	371	{
	372	// dst_c and dst_p are the cluster index and the pixel index for B
	373	int dst_c = p / pixels_per_cluster;
	374	int dst_p = p % pixels_per_cluster;
	375	sum_p = sum_p + (int) TA(src_c, src_l, p + hrange)
	376	- (int) TA(src_c, src_l, p - hrange - 1);
	377	TB(dst_c, dst_p, l) = sum_p / hnorm;
	378	TD(src_c, src_l, p) = (int) TA(src_c, src_l, p) - sum_p / hnorm;
	379	}
	380	// third domain : from (NP-hrange) to (NP-1)
	381	for (p = NP - hrange; p < NP; p++)
	382	{
	383	// dst_c and dst_p are the cluster index and the pixel index for B
	384	int dst_c = p / pixels_per_cluster;
	385	int dst_p = p % pixels_per_cluster;
	386	sum_p = sum_p + (int) TA(src_c, src_l, NP - 1)
	387	- (int) TA(src_c, src_l, p - hrange - 1);
	388	TB(dst_c, dst_p, l) = sum_p / hnorm;
	389	TD(src_c, src_l, p) = (int) TA(src_c, src_l, p) - sum_p / hnorm;
	390	}
	391
[353]	392	#if SUPER_VERBOSE
	393	giet_shr_printf(" - line %d computed at cycle %d\n", l, giet_proctime() );
	394	#endif
	395
[334]	396	}
	397
	398	date = giet_proctime();
	399	H_END[cluster_id][lpid] = date;
	400
[353]	401	#if VERBOSE
	402	giet_shr_printf( "\n[CONVOL] task[%d,%d,%d] completes horizontal filter"
	403	" at cycle %d\n",
	404	x, y, lpid, date );
	405	#else
	406	if ( pid == 0 ) giet_shr_printf( "\n[CONVOL] task[0,0,0] completes horizontal filter"
	407	" at cycle %d\n", date );
	408	#endif
	409
[377]	410	/////////////////////////////
	411	#if USE_SBT_BARRIER
	412	sbt_barrier_wait( &barrier );
	413	#else
	414	barrier_wait( &barrier );
	415	#endif
[334]	416
[372]	417
[353]	418	///////////////////////////////////////////////////////////////
[334]	419	// parallel vertical filter :
	420	// C <= transpose(FV(B))
	421	// Each task computes (NP/ntasks) columns
	422	// The image must be extended :
	423	// if (l<0) TB(cluster_id,p,l) == TB(cluster_id,p,0)
	424	// if (l>NL-1) TB(cluster_id,p,l) == TB(cluster_id,p,NL-1)
[353]	425	///////////////////////////////////////////////////////////////
[334]	426
	427	date = giet_proctime();
	428	V_BEG[cluster_id][lpid] = date;
	429
[353]	430	#if VERBOSE
	431	giet_shr_printf( "\n[CONVOL] task[%d,%d,%d] starts vertical filter"
	432	" at cycle %d\n",
	433	x, y, lpid, date );
	434	#else
	435	if ( pid == 0 ) giet_shr_printf( "\n[CONVOL] task[0,0,0] starts vertical filter"
	436	" at cycle %d\n", date );
	437	#endif
	438
[334]	439	// l = absolute line index / p = absolute pixel index
	440	// first & last define which pixels are handled by a given task
	441
	442	first = task_id * pixels_per_task;
	443	last = first + pixels_per_task;
	444
	445	for (p = first; p < last; p++)
	446	{
	447	// src_c and src_p are the cluster index and the pixel index for B
	448	int src_c = p / pixels_per_cluster;
	449	int src_p = p % pixels_per_cluster;
	450
	451	int sum_l;
	452
	453	// We use the specific values of the vertical ep-filter
	454	// To minimize the number of tests, the NL lines are split in three domains
	455
	456	// first domain : explicit computation for the first 18 values
	457	for (l = 0; l < 18; l++)
	458	{
	459	// dst_c and dst_l are the cluster index and the line index for C
	460	int dst_c = l / lines_per_cluster;
	461	int dst_l = l % lines_per_cluster;
	462
	463	for (z = 0, sum_l = 0; z < 35; z++)
	464	{
	465	sum_l = sum_l + vf[z] * TB(src_c, src_p, max(l - 17 + z,0) );
	466	}
	467	TC(dst_c, dst_l, p) = sum_l / vnorm;
	468	}
	469	// second domain
	470	for (l = 18; l < NL - 17; l++)
	471	{
	472	// dst_c and dst_l are the cluster index and the line index for C
	473	int dst_c = l / lines_per_cluster;
	474	int dst_l = l % lines_per_cluster;
	475
	476	sum_l = sum_l + TB(src_c, src_p, l + 4)
	477	+ TB(src_c, src_p, l + 8)
	478	+ TB(src_c, src_p, l + 11)
	479	+ TB(src_c, src_p, l + 15)
	480	+ TB(src_c, src_p, l + 17)
	481	- TB(src_c, src_p, l - 5)
	482	- TB(src_c, src_p, l - 9)
	483	- TB(src_c, src_p, l - 12)
	484	- TB(src_c, src_p, l - 16)
	485	- TB(src_c, src_p, l - 18);
	486
	487	TC(dst_c, dst_l, p) = sum_l / vnorm;
	488	}
	489	// third domain
	490	for (l = NL - 17; l < NL; l++)
	491	{
	492	// dst_c and dst_l are the cluster index and the line index for C
	493	int dst_c = l / lines_per_cluster;
	494	int dst_l = l % lines_per_cluster;
	495
	496	sum_l = sum_l + TB(src_c, src_p, min(l + 4, NL - 1))
	497	+ TB(src_c, src_p, min(l + 8, NL - 1))
	498	+ TB(src_c, src_p, min(l + 11, NL - 1))
	499	+ TB(src_c, src_p, min(l + 15, NL - 1))
	500	+ TB(src_c, src_p, min(l + 17, NL - 1))
	501	- TB(src_c, src_p, l - 5)
	502	- TB(src_c, src_p, l - 9)
	503	- TB(src_c, src_p, l - 12)
	504	- TB(src_c, src_p, l - 16)
	505	- TB(src_c, src_p, l - 18);
	506
	507	TC(dst_c, dst_l, p) = sum_l / vnorm;
	508	}
	509
[353]	510	#if SUPER_VERBOSE
	511	giet_shr_printf(" - column %d computed at cycle %d\n", p, giet_proctime());
	512	#endif
	513
[334]	514	}
	515
	516	date = giet_proctime();
	517	V_END[cluster_id][lpid] = date;
	518
[353]	519	#if VERBOSE
	520	giet_shr_printf( "\n[CONVOL] task[%d,%d,%d] completes vertical filter"
	521	" at cycle %d\n",
	522	x, y, lpid, date );
	523	#else
	524	if ( pid == 0 ) giet_shr_printf( "\n[CONVOL] task[0,0,0] completes vertical filter"
	525	" at cycle %d\n", date );
	526	#endif
	527
[377]	528	////////////////////////////
	529	#if USE_SBT_BARRIER
	530	sbt_barrier_wait( &barrier );
	531	#else
	532	barrier_wait( &barrier );
	533	#endif
[334]	534
	535	////////////////////////////////////////////////////////////////
	536	// Optional parallel display of the final image Z <= D + C
	537	// Eah task displays (NL/ntasks) lines. (one byte per pixel).
	538	////////////////////////////////////////////////////////////////
	539
	540	if ( FINAL_DISPLAY_ENABLE )
	541	{
	542	date = giet_proctime();
	543	D_BEG[cluster_id][lpid] = date;
	544
[353]	545	#if VERBOSE
	546	giet_shr_printf( "\n[CONVOL] task[%d,%d,%d] starts final display"
	547	" at cycle %d\n",
	548	x, y, lpid, date);
	549	#else
	550	if ( pid == 0 ) giet_shr_printf( "\n[CONVOL] task[0,0,0] starts final display"
	551	" at cycle %d\n", date );
	552	#endif
	553
[334]	554	unsigned int line;
	555	unsigned int offset = lines_per_task * lpid;
	556
	557	for ( l = 0 ; l < lines_per_task ; l++ )
	558	{
	559	line = offset + l;
	560
	561	for ( p = 0 ; p < NP ; p++ )
	562	{
	563	TZ(cluster_id, line, p) =
	564	(unsigned char)( (TD(cluster_id, line, p) +
	565	TC(cluster_id, line, p) ) >> 8 );
	566	}
	567
	568	giet_fb_sync_write( NP(l + (task_id lines_per_task) ),
	569	&TZ(cluster_id, line, 0),
	570	NP);
	571	}
	572
	573	date = giet_proctime();
	574	D_END[cluster_id][lpid] = date;
[353]	575
	576	#if VERBOSE
	577	giet_shr_printf( "\n[CONVOL] task[%d,%d,%d] completes final display"
	578	" at cycle %d\n",
	579	x, y, lpid, date);
	580	#else
	581	if ( pid == 0 ) giet_shr_printf( "\n[CONVOL] task[0,0,0] completes final display"
	582	" at cycle %d\n", date );
	583	#endif
[334]	584
[377]	585	//////////////////////////////
	586	#if USE_SBT_BARRIER
	587	sbt_barrier_wait( &barrier );
	588	#else
	589	barrier_wait( &barrier );
	590	#endif
[372]	591
[334]	592	}
	593
	594	/////////////////////////////////////////////////////////
	595	// Task[0,0,0] makes the instrumentation
	596	/////////////////////////////////////////////////////////
	597
	598	if ( pid == 0 )
	599	{
	600	date = giet_proctime();
[353]	601	giet_shr_printf("\n[CONVOL] task[0,0,0] starts instrumentation"
	602	" at cycle %d\n\n", date );
[334]	603
	604	int cc, pp;
	605
	606	unsigned int min_start = 0xFFFFFFFF;
	607	unsigned int max_start = 0;
	608
	609	unsigned int min_h_beg = 0xFFFFFFFF;
	610	unsigned int max_h_beg = 0;
	611
	612	unsigned int min_h_end = 0xFFFFFFFF;
	613	unsigned int max_h_end = 0;
	614
	615	unsigned int min_v_beg = 0xFFFFFFFF;
	616	unsigned int max_v_beg = 0;
	617
	618	unsigned int min_v_end = 0xFFFFFFFF;
	619	unsigned int max_v_end = 0;
	620
	621	unsigned int min_d_beg = 0xFFFFFFFF;
	622	unsigned int max_d_beg = 0;
	623
	624	unsigned int min_d_end = 0xFFFFFFFF;
	625	unsigned int max_d_end = 0;
	626
	627	for (cc = 0; cc < nclusters; cc++)
	628	{
	629	for (pp = 0; pp < nprocs; pp++ )
	630	{
	631	if (START[cc][pp] < min_start) min_start = START[cc][pp];
	632	if (START[cc][pp] > max_start) max_start = START[cc][pp];
	633
	634	if (H_BEG[cc][pp] < min_h_beg) min_h_beg = H_BEG[cc][pp];
	635	if (H_BEG[cc][pp] > max_h_beg) max_h_beg = H_BEG[cc][pp];
	636
	637	if (H_END[cc][pp] < min_h_end) min_h_end = H_END[cc][pp];
	638	if (H_END[cc][pp] > max_h_end) max_h_end = H_END[cc][pp];
	639
	640	if (V_BEG[cc][pp] < min_v_beg) min_v_beg = V_BEG[cc][pp];
	641	if (V_BEG[cc][pp] > max_v_beg) max_v_beg = V_BEG[cc][pp];
	642
	643	if (V_END[cc][pp] < min_v_end) min_v_end = V_END[cc][pp];
	644	if (V_END[cc][pp] > max_v_end) max_v_end = V_END[cc][pp];
	645
	646	if (D_BEG[cc][pp] < min_d_beg) min_d_beg = D_BEG[cc][pp];
	647	if (D_BEG[cc][pp] > max_d_beg) max_d_beg = D_BEG[cc][pp];
	648
	649	if (D_END[cc][pp] < min_d_end) min_d_end = D_END[cc][pp];
	650	if (D_END[cc][pp] > max_d_end) max_d_end = D_END[cc][pp];
	651	}
	652	}
	653
	654	giet_shr_printf(" - START : min = %d / max = %d / med = %d / delta = %d\n",
	655	min_start, max_start, (min_start+max_start)/2, max_start-min_start);
	656
	657	giet_shr_printf(" - H_BEG : min = %d / max = %d / med = %d / delta = %d\n",
	658	min_h_beg, max_h_beg, (min_h_beg+max_h_beg)/2, max_h_beg-min_h_beg);
	659
[353]	660	giet_shr_printf(" - H_END : min = %d / max = %d / med = %d / delta = %d\n",
[334]	661	min_h_end, max_h_end, (min_h_end+max_h_end)/2, max_h_end-min_h_end);
	662
	663	giet_shr_printf(" - V_BEG : min = %d / max = %d / med = %d / delta = %d\n",
	664	min_v_beg, max_v_beg, (min_v_beg+max_v_beg)/2, max_v_beg-min_v_beg);
	665
[353]	666	giet_shr_printf(" - V_END : min = %d / max = %d / med = %d / delta = %d\n",
[334]	667	min_v_end, max_v_end, (min_v_end+max_v_end)/2, max_v_end-min_v_end);
	668
	669	giet_shr_printf(" - D_BEG : min = %d / max = %d / med = %d / delta = %d\n",
	670	min_d_beg, max_d_beg, (min_d_beg+max_d_beg)/2, max_d_beg-min_d_beg);
	671
[353]	672	giet_shr_printf(" - D_END : min = %d / max = %d / med = %d / delta = %d\n",
[334]	673	min_d_end, max_d_end, (min_d_end+max_d_end)/2, max_d_end-min_d_end);
	674
	675	giet_shr_printf( "\n General Scenario (Kcycles for each step)\n" );
[353]	676	giet_shr_printf( " - BOOT OS = %d\n", (min_start )/1000 );
	677	giet_shr_printf( " - LOAD IMAGE = %d\n", (min_h_beg - min_start)/1000 );
	678	giet_shr_printf( " - H_FILTER = %d\n", (max_h_end - min_h_beg)/1000 );
	679	giet_shr_printf( " - BARRIER HORI/VERT = %d\n", (min_v_beg - max_h_end)/1000 );
	680	giet_shr_printf( " - V_FILTER = %d\n", (max_v_end - min_v_beg)/1000 );
	681	giet_shr_printf( " - BARRIER VERT/DISP = %d\n", (min_d_beg - max_v_end)/1000 );
	682	giet_shr_printf( " - DISPLAY = %d\n", (max_d_end - min_d_beg)/1000 );
[334]	683
[372]	684	instrumentation_ok = 1;
[334]	685	}
	686	else
	687	{
[372]	688	while ( instrumentation_ok == 0 );
[334]	689	}
	690
	691	giet_exit( "completed");
	692
	693	} // end main()
	694
	695	// Local Variables:
	696	// tab-width: 3
	697	// c-basic-offset: 3
	698	// c-file-offsets:((innamespace . 0)(inline-open . 0))
	699	// indent-tabs-mode: nil
	700	// End:
	701
	702	// vim: filetype=cpp:expandtab:shiftwidth=3:tabstop=3:softtabstop=3
	703
	704

Note: See TracBrowser for help on using the repository browser.

Download in other formats: