void FATR
ga_antisymmetrize_(Integer *g_a) {
DoublePrecision alpha = 0.5;
int i, me = GA_Nodeid();
extern void * FATR ga_malloc(Integer nelem, int type, char *name);
extern void FATR ga_free(void *ptr);
void FATR gai_subtr(int *lo, int *hi, void *a, void *b, DoublePrecision alpha,
int type, Integer nelem, int ndim);
int alo[GA_MAX_DIM], ahi[GA_MAX_DIM], lda[GA_MAX_DIM];
int blo[GA_MAX_DIM], bhi[GA_MAX_DIM], ldb[GA_MAX_DIM];
int ndim, dims[GA_MAX_DIM], type;
Integer nelem=1;
Logical have_data;
void *a_ptr, *b_ptr;
GA_Sync();
NGA_Inquire((int)(*g_a), &type, &ndim, dims);
if (dims[0] != dims[1])
GA_Error("ga_sym: can only sym square matrix", 0L);
/* Find the local distribution */
NGA_Distribution((int)(*g_a), me, alo, ahi);
have_data = ahi[0]>=0;
for(i=1; i<ndim; i++) have_data = have_data && ahi[i]>=0;
if(have_data) {
NGA_Access((int)(*g_a), alo, ahi, &a_ptr, lda);
for(i=0; i<ndim; i++) nelem *= ahi[i]-alo[i] +1;
b_ptr = (void *) ga_malloc(nelem, MT_C_DBL, "v");
for(i=2; i<ndim; i++) {bhi[i]=ahi[i]; blo[i]=alo[i]; }
/* switch rows and cols */
blo[1]=alo[0];
bhi[1]=ahi[0];
blo[0]=alo[1];
bhi[0]=ahi[1];
for (i=0; i < ndim-1; i++)
ldb[i] = bhi[i+1] - blo[i+1] + 1;
NGA_Get((int)(*g_a), blo, bhi, b_ptr, ldb);
}
GA_Sync();
if(have_data) {
gai_subtr(alo, ahi, a_ptr, b_ptr, alpha, type, nelem, ndim);
NGA_Release_update((int)(*g_a), alo, ahi);
ga_free(b_ptr);
}
GA_Sync();
}
struct vsource_frame *
vsource_frame_init(struct vsource_frame *frame, int maxwidth, int maxheight, int maxstride) {
int i;
//
bzero(frame, sizeof(struct vsource_frame));
//
for(i = 0; i < MAX_STRIDE; i++) {
frame->linesize[i] = maxstride;
}
frame->maxstride = maxstride;
frame->imgbufsize = maxheight * maxstride;
if(ga_malloc(frame->imgbufsize, (void**) &frame->imgbuf_internal, &frame->alignment) < 0) {
return NULL;
}
frame->imgbuf = frame->imgbuf_internal + frame->alignment;
bzero(frame->imgbuf, frame->imgbufsize);
return frame;
}
void *
vencoder_threadproc(void *arg) {
// arg is pointer to source pipe
// image info
int iid;
int iwidth;
int iheight;
int rtp_id;
struct pooldata *data = NULL;
struct vsource_frame *frame = NULL;
pipeline *pipe = (pipeline*) arg;
AVCodecContext *encoder = NULL;
//
AVFrame *pic_in = NULL;
unsigned char *pic_in_buf = NULL;
int pic_in_size;
unsigned char *nalbuf = NULL, *nalbuf_a = NULL;
int nalbuf_size = 0, nalign = 0;
long long basePts = -1LL, newpts = 0LL, pts = -1LL, ptsSync = 0LL;
pthread_mutex_t condMutex = PTHREAD_MUTEX_INITIALIZER;
pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
//
int resolution[2];
int video_written = 0;
//
if(pipe == NULL) {
ga_error("video encoder: NULL pipeline specified.\n");
goto video_quit;
}
//
rtspconf = rtspconf_global();
// init variables
iid = ((struct vsource_config*) pipe->get_privdata())->id;
iwidth = video_source_maxwidth(iid);
iheight = video_source_maxheight(iid);
rtp_id = ((struct vsource_config*) pipe->get_privdata())->rtp_id;
//
outputW = iwidth; // by default, the same as max resolution
outputH = iheight;
if(ga_conf_readints("output-resolution", resolution, 2) == 2) {
outputW = resolution[0];
outputH = resolution[1];
}
//
ga_error("video encoder: image source from '%s' (%dx%d) via channel %d, resolution=%dx%d.\n",
pipe->name(), iwidth, iheight, rtp_id, outputW, outputH);
//
encoder = ga_avcodec_vencoder_init(
NULL,
rtspconf->video_encoder_codec,
outputW, outputH,
rtspconf->video_fps,
rtspconf->vso);
if(encoder == NULL) {
ga_error("video encoder: cannot initialized the encoder.\n");
goto video_quit;
}
//
nalbuf_size = 100000+12 * outputW * outputH;
if(ga_malloc(nalbuf_size, (void**) &nalbuf, &nalign) < 0) {
ga_error("video encoder: buffer allocation failed, terminated.\n");
goto video_quit;
}
nalbuf_a = nalbuf + nalign;
//
if((pic_in = avcodec_alloc_frame()) == NULL) {
ga_error("video encoder: picture allocation failed, terminated.\n");
goto video_quit;
}
pic_in_size = avpicture_get_size(PIX_FMT_YUV420P, outputW, outputH);
if((pic_in_buf = (unsigned char*) av_malloc(pic_in_size)) == NULL) {
ga_error("video encoder: picture buffer allocation failed, terminated.\n");
goto video_quit;
}
avpicture_fill((AVPicture*) pic_in, pic_in_buf,
PIX_FMT_YUV420P, outputW, outputH);
//ga_error("video encoder: linesize = %d|%d|%d\n", pic_in->linesize[0], pic_in->linesize[1], pic_in->linesize[2]);
// start encoding
ga_error("video encoding started: tid=%ld %dx%d@%dfps, nalbuf_size=%d, pic_in_size=%d.\n",
ga_gettid(),
iwidth, iheight, rtspconf->video_fps,
nalbuf_size, pic_in_size);
//
pipe->client_register(ga_gettid(), &cond);
//
while(encoder_running() > 0) {
AVPacket pkt;
int got_packet = 0;
// wait for notification
data = pipe->load_data();
if(data == NULL) {
int err;
struct timeval tv;
struct timespec to;
gettimeofday(&tv, NULL);
to.tv_sec = tv.tv_sec+1;
to.tv_nsec = tv.tv_usec * 1000;
//
if((err = pipe->timedwait(&cond, &condMutex, &to)) != 0) {
ga_error("viedo encoder: image source timed out.\n");
continue;
}
data = pipe->load_data();
if(data == NULL) {
ga_error("viedo encoder: unexpected NULL frame received (from '%s', data=%d, buf=%d).\n",
pipe->name(), pipe->data_count(), pipe->buf_count());
continue;
}
}
frame = (struct vsource_frame*) data->ptr;
// handle pts
if(basePts == -1LL) {
basePts = frame->imgpts;
ptsSync = encoder_pts_sync(rtspconf->video_fps);
newpts = ptsSync;
} else {
newpts = ptsSync + frame->imgpts - basePts;
}
// XXX: assume always YUV420P
if(pic_in->linesize[0] == frame->linesize[0]
&& pic_in->linesize[1] == frame->linesize[1]
&& pic_in->linesize[2] == frame->linesize[2]) {
bcopy(frame->imgbuf, pic_in_buf, pic_in_size);
} else {
ga_error("video encoder: YUV mode failed - mismatched linesize(s) (src:%d,%d,%d; dst:%d,%d,%d)\n",
frame->linesize[0], frame->linesize[1], frame->linesize[2],
pic_in->linesize[0], pic_in->linesize[1], pic_in->linesize[2]);
pipe->release_data(data);
goto video_quit;
}
pipe->release_data(data);
// pts must be monotonically increasing
if(newpts > pts) {
pts = newpts;
} else {
pts++;
}
// encode
pic_in->pts = pts;
av_init_packet(&pkt);
pkt.data = nalbuf_a;
pkt.size = nalbuf_size;
if(avcodec_encode_video2(encoder, &pkt, pic_in, &got_packet) < 0) {
ga_error("video encoder: encode failed, terminated.\n");
goto video_quit;
}
if(got_packet) {
if(pkt.pts == (int64_t) AV_NOPTS_VALUE) {
pkt.pts = pts;
}
pkt.stream_index = 0;
// send the packet
if(encoder_send_packet_all("video-encoder",
rtp_id/*rtspconf->video_id*/, &pkt,
pkt.pts) < 0) {
goto video_quit;
}
// free unused side-data
if(pkt.side_data_elems > 0) {
int i;
for (i = 0; i < pkt.side_data_elems; i++)
av_free(pkt.side_data[i].data);
av_freep(&pkt.side_data);
pkt.side_data_elems = 0;
}
//
if(video_written == 0) {
video_written = 1;
ga_error("first video frame written (pts=%lld)\n", pts);
}
}
}
//
video_quit:
if(pipe) {
pipe->client_unregister(ga_gettid());
pipe = NULL;
}
//
if(pic_in_buf) av_free(pic_in_buf);
if(pic_in) av_free(pic_in);
if(nalbuf) free(nalbuf);
if(encoder) ga_avcodec_close(encoder);
//
ga_error("video encoder: thread terminated (tid=%ld).\n", ga_gettid());
//
return NULL;
}
void FATR
ga_symmetrize_(Integer *g_a) {
DoublePrecision alpha = 0.5;
Integer i, me = ga_nodeid_();
Integer alo[GA_MAX_DIM], ahi[GA_MAX_DIM], lda[GA_MAX_DIM], nelem=1;
Integer blo[GA_MAX_DIM], bhi[GA_MAX_DIM], ldb[GA_MAX_DIM];
Integer ndim, dims[GA_MAX_DIM], type;
Logical have_data;
Integer g_b; /* temporary global array (b = A') */
Integer num_blocks_a;
Void *a_ptr, *b_ptr;
int local_sync_begin,local_sync_end;
char *tempB = "A_transpose";
local_sync_begin = _ga_sync_begin; local_sync_end = _ga_sync_end;
_ga_sync_begin = 1; _ga_sync_end=1; /*remove any previous masking*/
if(local_sync_begin)ga_sync_();
GA_PUSH_NAME("ga_symmetrize");
num_blocks_a = ga_total_blocks_(g_a);
nga_inquire_internal_(g_a, &type, &ndim, dims);
if (type != C_DBL)
ga_error("ga_symmetrize: only implemented for double precision",0);
if (num_blocks_a < 0) {
if (dims[ndim-1] != dims[ndim-2])
ga_error("ga_sym: can only sym square matrix", 0L);
/* Find the local distribution */
nga_distribution_(g_a, &me, alo, ahi);
have_data = ahi[0]>0;
for(i=1; i<ndim; i++) have_data = have_data && ahi[i]>0;
if(have_data) {
nga_access_ptr(g_a, alo, ahi, &a_ptr, lda);
for(i=0; i<ndim; i++) nelem *= ahi[i]-alo[i] +1;
b_ptr = (Void *) ga_malloc(nelem, MT_F_DBL, "v");
for(i=0; i<ndim-2; i++) {bhi[i]=ahi[i]; blo[i]=alo[i]; }
/* switch rows and cols */
blo[ndim-1]=alo[ndim-2];
bhi[ndim-1]=ahi[ndim-2];
blo[ndim-2]=alo[ndim-1];
bhi[ndim-2]=ahi[ndim-1];
for (i=0; i < ndim-1; i++)
ldb[i] = bhi[i] - blo[i] + 1;
nga_get_(g_a, blo, bhi, b_ptr, ldb);
}
ga_sync_();
if(have_data) {
gai_add(alo, ahi, a_ptr, b_ptr, alpha, type, nelem, ndim);
nga_release_update_(g_a, alo, ahi);
ga_free(b_ptr);
}
} else {
/* For block-cyclic data, probably most efficient solution is to
create duplicate copy, transpose it and add the results together */
DoublePrecision half = 0.5;
if (!ga_duplicate(g_a, &g_b, tempB))
ga_error("ga_symmetrize: duplicate failed", 0L);
ga_transpose_(g_a, &g_b);
ga_add_(&half, g_a, &half, &g_b, g_a);
ga_destroy_(&g_b);
}
GA_POP_NAME;
if(local_sync_end)ga_sync_();
}
