void SERVICE_fwd(float* in, int in_size, float* out, int out_size,
Net<float>* net) {
string net_name = net->name();
STATS_INIT("service", "DjiNN service inference");
PRINT_STAT_STRING("network", net_name.c_str());
if (Caffe::mode() == Caffe::CPU)
PRINT_STAT_STRING("platform", "cpu");
else
PRINT_STAT_STRING("platform", "gpu");
float loss;
vector<Blob<float>*> in_blobs = net->input_blobs();
tic();
in_blobs[0]->set_cpu_data(in);
vector<Blob<float>*> out_blobs = net->ForwardPrefilled(&loss);
memcpy(out, out_blobs[0]->cpu_data(), sizeof(float));
PRINT_STAT_DOUBLE("inference latency", toc());
STATS_END();
if (out_size != out_blobs[0]->count())
LOG(FATAL) << "out_size =! out_blobs[0]->count())";
else
memcpy(out, out_blobs[0]->cpu_data(), out_size * sizeof(float));
}
int
main (const int argc, char *argv[])
{
parse_args (argc, argv, &initial_graph_name, &action_stream_name, &batch_size, &nbatch);
STATS_INIT();
load_graph_and_action_stream (initial_graph_name, &nv, &ne, (int64_t**)&off,
(int64_t**)&ind, (int64_t**)&weight, (int64_t**)&graphmem,
action_stream_name, &naction, (int64_t**)&action, (int64_t**)&actionmem);
print_initial_graph_stats (nv, ne, batch_size, nbatch, naction);
BATCH_SIZE_CHECK();
#if defined(_OPENMP)
OMP(omp parallel)
{
OMP(omp master)
PRINT_STAT_INT64 ("num_threads", (long int) omp_get_num_threads());
}
int
main (const int argc, char *argv[])
{
parse_args (argc, argv, &initial_graph_name, &action_stream_name,
&batch_size, &nbatch);
STATS_INIT ();
load_graph_and_action_stream (initial_graph_name, &nv, &ne,
(int64_t **) & off, (int64_t **) & ind,
(int64_t **) & weight,
(int64_t **) & graphmem, action_stream_name,
&naction, (int64_t **) & action,
(int64_t **) & actionmem);
print_initial_graph_stats (nv, ne, batch_size, nbatch, naction);
BATCH_SIZE_CHECK ();
/* Convert to STINGER */
tic ();
S = stinger_new ();
stinger_set_initial_edges (S, nv, 0, off, ind, weight, NULL, NULL, 0);
PRINT_STAT_DOUBLE ("time_stinger", toc ());
fflush (stdout);
int64_t numSteps = 3;
int64_t src_dest_pair[2] = { 124, 381 };
tic ();
int64_t size_intersect =
st_conn_stinger (S, nv, ne, src_dest_pair, 1, numSteps);
PRINT_STAT_DOUBLE ("time_st_conn_stinger", toc ());
PRINT_STAT_INT64 ("size_intersect", size_intersect);
stinger_free_all (S);
free (graphmem);
free (actionmem);
STATS_END ();
}
int
main (const int argc, char *argv[])
{
parse_args (argc, argv, &initial_graph_name, &action_stream_name,
&batch_size, &nbatch);
STATS_INIT ();
load_graph_and_action_stream (initial_graph_name, &nv, &ne,
(int64_t **) & off, (int64_t **) & ind,
(int64_t **) & weight,
(int64_t **) & graphmem, action_stream_name,
&naction, (int64_t **) & action,
(int64_t **) & actionmem);
print_initial_graph_stats (nv, ne, batch_size, nbatch, naction);
BATCH_SIZE_CHECK ();
int64_t *component_map = xmalloc (nv * sizeof (int64_t));
/* Convert to STINGER */
tic ();
S = stinger_new ();
stinger_set_initial_edges (S, nv, 0, off, ind, weight, NULL, NULL, 0);
PRINT_STAT_DOUBLE ("time_stinger", toc ());
fflush (stdout);
tic ();
int64_t num_comp_end =
connected_components_stinger (S, nv, ne, component_map, NULL, NULL, NULL,
NULL, NULL);
PRINT_STAT_DOUBLE ("time_components_tree", toc ());
PRINT_STAT_INT64 ("number_of_components", num_comp_end);
stinger_free_all (S);
free (graphmem);
free (actionmem);
STATS_END ();
}
int main(int argc, char *argv[]) {
if (argc < 4) {
fprintf(stderr, "[ERROR] Invalid arguments provided.\n\n");
fprintf(stderr, "Usage: %s [NUMBER OF THREADS] [WORDS] [INPUT FILE]\n\n", argv[0]);
exit(0);
}
/* Timing */
STATS_INIT("kernel", "pthread_porter_stemming");
PRINT_STAT_STRING("abrv", "pthread_stemmer");
NTHREADS = atoi(argv[1]);
int WORDS = atoi(argv[2]);
PRINT_STAT_INT("threads", NTHREADS);
FILE *f = fopen(argv[3], "r");
if (f == 0) {
fprintf(stderr, "File %s not found\n", argv[1]);
exit(1);
}
stem_list =
(struct stemmer **)sirius_malloc(WORDS * sizeof(struct stemmer *));
int words = load_data(WORDS, stem_list, f);
fclose(f);
if (words < 0)
goto out;
PRINT_STAT_INT("words", words);
tic();
int start, tids[NTHREADS];
pthread_t threads[NTHREADS];
pthread_attr_t attr;
iterations = words / NTHREADS;
sirius_pthread_attr_init(&attr);
sirius_pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
for (int i = 0; i < NTHREADS; i++) {
tids[i] = i;
sirius_pthread_create(&threads[i], &attr, stem_thread, (void *)&tids[i]);
}
for (int i = 0; i < NTHREADS; i++) {
sirius_pthread_join(threads[i], NULL);
}
PRINT_STAT_DOUBLE("pthread_stemmer", toc());
STATS_END();
#ifdef TESTING
f = fopen("../input/stem_porter.pthread", "w");
for (int i = 0; i < words; ++i) fprintf(f, "%s\n", stem_list[i]->b);
fclose(f);
#endif
out:
sirius_free(s);
// free up allocated data
for (int i = 0; i < words; i++) {
sirius_free(stem_list[i]->b);
sirius_free(stem_list[i]);
}
return 0;
}
static int
ssh_aes_ctr_init(EVP_CIPHER_CTX *ctx, const u_char *key, const u_char *iv,
int enc)
{
struct ssh_aes_ctr_ctx *c;
int i;
if ((c = EVP_CIPHER_CTX_get_app_data(ctx)) == NULL) {
c = xmalloc(sizeof(*c));
c->state = HAVE_NONE;
for (i = 0; i < NUMKQ; i++) {
pthread_mutex_init(&c->q[i].lock, NULL);
pthread_cond_init(&c->q[i].cond, NULL);
}
STATS_INIT(c->stats);
EVP_CIPHER_CTX_set_app_data(ctx, c);
}
if (c->state == (HAVE_KEY | HAVE_IV)) {
/* Cancel pregen threads */
for (i = 0; i < CIPHER_THREADS; i++)
pthread_cancel(c->tid[i]);
for (i = 0; i < CIPHER_THREADS; i++)
pthread_join(c->tid[i], NULL);
/* Start over getting key & iv */
c->state = HAVE_NONE;
}
if (key != NULL) {
AES_set_encrypt_key(key, EVP_CIPHER_CTX_key_length(ctx) * 8,
&c->aes_ctx);
c->state |= HAVE_KEY;
}
if (iv != NULL) {
memcpy(ctx->iv, iv, AES_BLOCK_SIZE);
c->state |= HAVE_IV;
}
if (c->state == (HAVE_KEY | HAVE_IV)) {
/* Clear queues */
memcpy(c->q[0].ctr, ctx->iv, AES_BLOCK_SIZE);
c->q[0].qstate = KQINIT;
for (i = 1; i < NUMKQ; i++) {
memcpy(c->q[i].ctr, ctx->iv, AES_BLOCK_SIZE);
ssh_ctr_add(c->q[i].ctr, i * KQLEN, AES_BLOCK_SIZE);
c->q[i].qstate = KQEMPTY;
}
c->qidx = 0;
c->ridx = 0;
/* Start threads */
for (i = 0; i < CIPHER_THREADS; i++) {
debug("spawned a thread");
pthread_create(&c->tid[i], NULL, thread_loop, c);
}
pthread_mutex_lock(&c->q[0].lock);
while (c->q[0].qstate != KQDRAINING)
pthread_cond_wait(&c->q[0].cond, &c->q[0].lock);
pthread_mutex_unlock(&c->q[0].lock);
}
return 1;
}
/*
* The life of a pregen thread:
* Find empty keystream queues and fill them using their counter.
* When done, update counter for the next fill.
*/
static void *
thread_loop(void *x)
{
AES_KEY key;
STATS_STRUCT(stats);
struct ssh_aes_ctr_ctx *c = x;
struct kq *q;
int i;
int qidx;
/* Threads stats on cancellation */
STATS_INIT(stats);
#ifdef CIPHER_THREAD_STATS
pthread_cleanup_push(thread_loop_stats, &stats);
#endif
/* Thread local copy of AES key */
memcpy(&key, &c->aes_ctx, sizeof(key));
/*
* Handle the special case of startup, one thread must fill
* the first KQ then mark it as draining. Lock held throughout.
*/
if (pthread_equal(pthread_self(), c->tid[0])) {
q = &c->q[0];
pthread_mutex_lock(&q->lock);
if (q->qstate == KQINIT) {
for (i = 0; i < KQLEN; i++) {
AES_encrypt(q->ctr, q->keys[i], &key);
ssh_ctr_inc(q->ctr, AES_BLOCK_SIZE);
}
ssh_ctr_add(q->ctr, KQLEN * (NUMKQ - 1), AES_BLOCK_SIZE);
q->qstate = KQDRAINING;
STATS_FILL(stats);
pthread_cond_broadcast(&q->cond);
}
pthread_mutex_unlock(&q->lock);
} else
STATS_SKIP(stats);
/*
* Normal case is to find empty queues and fill them, skipping over
* queues already filled by other threads and stopping to wait for
* a draining queue to become empty.
*
* Multiple threads may be waiting on a draining queue and awoken
* when empty. The first thread to wake will mark it as filling,
* others will move on to fill, skip, or wait on the next queue.
*/
for (qidx = 1;; qidx = (qidx + 1) % NUMKQ) {
/* Check if I was cancelled, also checked in cond_wait */
pthread_testcancel();
/* Lock queue and block if its draining */
q = &c->q[qidx];
pthread_mutex_lock(&q->lock);
pthread_cleanup_push(thread_loop_cleanup, &q->lock);
while (q->qstate == KQDRAINING || q->qstate == KQINIT) {
STATS_WAIT(stats);
pthread_cond_wait(&q->cond, &q->lock);
}
pthread_cleanup_pop(0);
/* If filling or full, somebody else got it, skip */
if (q->qstate != KQEMPTY) {
pthread_mutex_unlock(&q->lock);
STATS_SKIP(stats);
continue;
}
/*
* Empty, let's fill it.
* Queue lock is relinquished while we do this so others
* can see that it's being filled.
*/
q->qstate = KQFILLING;
pthread_mutex_unlock(&q->lock);
for (i = 0; i < KQLEN; i++) {
AES_encrypt(q->ctr, q->keys[i], &key);
ssh_ctr_inc(q->ctr, AES_BLOCK_SIZE);
}
/* Re-lock, mark full and signal consumer */
pthread_mutex_lock(&q->lock);
ssh_ctr_add(q->ctr, KQLEN * (NUMKQ - 1), AES_BLOCK_SIZE);
q->qstate = KQFULL;
STATS_FILL(stats);
pthread_cond_signal(&q->cond);
pthread_mutex_unlock(&q->lock);
}
#ifdef CIPHER_THREAD_STATS
/* Stats */
pthread_cleanup_pop(1);
#endif
return NULL;
}
int main (int argc, char * argv[])
{
APPROX int * frame;
APPROX int * output;
int i;
int nFilterRowsFD = 9;
int nFilterColsFD = 9;
APPROX fltPixel_t FD[] = {
1, 3, 4, 5, 6, 5, 4, 3, 1,
3, 9, 12, 15, 18, 15, 12, 9, 3,
4, 12, 16, 20, 24, 20, 16, 12, 4,
5, 15, 20, 25, 30, 25, 20, 15, 5,
6, 18, 24, 30, 36, 30, 24, 18, 6,
5, 15, 20, 25, 30, 25, 20, 15, 5,
4, 12, 16, 20, 24, 20, 16, 12, 4,
3, 9, 12, 15, 18, 15, 12, 9, 3,
1, 3, 4, 5, 6, 5, 4, 3, 1
};
for (i = 0; i < nFilterRowsFD * nFilterColsFD; i++) // ACCEPT_FORBID
{
FD[i] /= (1024.0);
}
srand (time (NULL));
STATS_INIT ();
PRINT_STAT_STRING ("kernel", "2d_convolution");
PRINT_STAT_INT ("rows", N);
PRINT_STAT_INT ("columns", M);
PRINT_STAT_INT ("num_frames", BATCH_SIZE);
frame = calloc (M * N * BATCH_SIZE, sizeof(algPixel_t));
output = calloc (M * N * BATCH_SIZE, sizeof(algPixel_t));
if (!frame || !output) {
fprintf(stderr, "ERROR: Allocation failed.\n");
exit(-1);
}
/* load image */
tic ();
read_array_from_octave (ENDORSE(frame), N, M, FILENAME);
PRINT_STAT_DOUBLE ("time_load_image", toc ());
/* Make BATCH_SIZE-1 copies */
tic ();
for (i = 1; i < BATCH_SIZE; i++) // ACCEPT_FORBID
{
memcpy (&frame[i * M * N], frame, M * N * sizeof(algPixel_t));
}
PRINT_STAT_DOUBLE ("time_copy", toc ());
/* Perform the 2D convolution */
tic ();
accept_roi_begin();
for (i = 0; i < BATCH_SIZE; i++) // ACCEPT_FORBID
{
conv2d (&frame[i * M * N], &output[i * M * N], N, M, FD, 1.0, nFilterRowsFD, nFilterColsFD);
}
accept_roi_end();
PRINT_STAT_DOUBLE ("time_2d_convolution", toc ());
/* Write the results out to disk */
for (i = 0; i < BATCH_SIZE; i++) // ACCEPT_FORBID
{
char buffer [30];
sprintf (buffer, "2dconv_output.%d.mat", i);
write_array_to_octave (ENDORSE(&output[i * M * N]), N, M, buffer, "output_" SIZE);
}
PRINT_STAT_STRING ("output_file", "2dconv_output." SIZE ".#.mat");
STATS_END ();
free (output);
free (frame);
return 0;
}
