void square_dgemm(const int M, const double *A, const double *B, double *C)
{
if (M > 100) {
const int n_blocks = M / BLOCK_SIZE + (M%BLOCK_SIZE? 1 : 0);
//double* t = (double*)malloc(BLOCK_SIZE * BLOCK_SIZE * sizeof(double));
double aa[BLOCK_SIZE * BLOCK_SIZE] = {0};
double bb[BLOCK_SIZE * BLOCK_SIZE] = {0};
double cc[BLOCK_SIZE * BLOCK_SIZE] = {0};
int bi, bj, bk;
for (bi = 0; bi < n_blocks; ++bi) {
const int i = bi * BLOCK_SIZE;
for (bj = 0; bj < n_blocks; ++bj) {
const int j = bj * BLOCK_SIZE;
for (bk = 0; bk < n_blocks; ++bk) {
const int k = bk * BLOCK_SIZE;
do_block(M, A, B, C, aa, bb, cc, i, j, k);
}
}
}
} else {
naive_square_dgemm(M, A, B, C);
}
}
long do_sched_op_compat(int cmd, unsigned long arg)
{
long ret = 0;
switch ( cmd )
{
case SCHEDOP_yield:
{
ret = do_yield();
break;
}
case SCHEDOP_block:
{
ret = do_block();
break;
}
case SCHEDOP_shutdown:
{
TRACE_3D(TRC_SCHED_SHUTDOWN,
current->domain->domain_id, current->vcpu_id, arg);
domain_shutdown(current->domain, (u8)arg);
break;
}
default:
ret = -ENOSYS;
}
return ret;
}
/* This routine performs a dgemm operation
* C := C + A * B
* where A, B, and C are lda-by-lda matrices stored in column-major format.
* On exit, A and B maintain their input values. */
void square_dgemm(int lda, double* A, double* B, double* C) {
int block_size_row = 222;
int block_size_col = 12;
int block_size_inner = 222;
int M_even, K_even, N_even;
double new_A[50000];
double new_B[200000];
double new_C[4000];
for (int k=0; k<lda; k+=block_size_inner) {
int K = min(block_size_inner, lda-k);
copy_block(lda, K, lda, B+k, new_B);
K_even = turn_even(K);
for (int i=0; i<lda; i+=block_size_row) {
int M = min (block_size_row, lda-i);
copy_block(lda, M, K, A+i+k*lda, new_A);
M_even = turn_even(M);
for (int j=0; j<lda; j+=block_size_col) {
int N = min (block_size_col, lda-j);
N_even = turn_even(N);
copy_block(lda, M, N, C+i+j*lda, new_C);
do_block(M_even, K_even, N_even, new_A, new_B+j*K_even, new_C);
add_block(new_C, C+i+j*lda, M, N, lda, M_even);
}
}
}
}
void
output(const char *lojban, const char *trans, const char *selmao)
{
switch (ofmt) {
case OF_LATEX:
printf ("\\begin{tabular}[t]{l}"
"\\textbf{\\footnotesize %s}\\\\\n"
"\\textrm{\\footnotesize %s}\\\\\n"
"\\textit{\\footnotesize %s}\n"
"\\end{tabular}\n"
"\\rule{0in}{1.0\\baselineskip}",
lojban, selmao, trans);
break;
case OF_TEXT:
printf ("%s <%s> [%s] ", lojban, selmao, trans);
break;
case OF_TEXTBLK:
do_block(lojban, selmao, trans);
break;
#ifdef PLIST
case OF_PLIST:
dictionary = PLInsertDictionaryEntry(dictionary, PLMakeString(lojban), PLMakeString(trans));
break;
#endif //PLIST
}
}
void square_dgemm( int M, double *A, double *B, double *C )
{
for( int i = 0; i < M; i += BLOCK_SIZE )
for( int j = 0; j < M; j += BLOCK_SIZE )
for( int k = 0; k < M; k += BLOCK_SIZE )
do_block( M, A, B, C, i, j, k );
}
static
stmt_code(stream, node, brk, cont, ret) {
/* Handle the null expression. */
if ( !node ) return;
auto op = node[0];
/* For debugging purposes, put a blank line between each statement. */
fputs("\n", stream);
if ( op == 'dcls' ) declaration( stream, node );
else if ( op == 'brea' ) branch( stream, brk );
else if ( op == 'cont' ) branch( stream, cont );
else if ( op == 'retu' ) return_stmt( stream, node, ret );
else if ( op == 'goto' ) goto_stmt( stream, node );
else if ( op == 'if' ) if_stmt( stream, node, brk, cont, ret );
else if ( op == 'whil' ) while_stmt( stream, node, brk, cont, ret );
else if ( op == 'for' ) for_stmt( stream, node, brk, cont, ret );
else if ( op == 'do' ) do_stmt( stream, node, brk, cont, ret );
else if ( op == 'swit' ) switch_stmt( stream, node, brk, cont, ret );
else if ( op == 'case'
|| op == 'defa' ) case_stmt( stream, node, brk, cont, ret );
else if ( op == ':' ) label_stmt( stream, node, brk, cont, ret );
else if ( op == '{}' ) do_block( stream, node, brk, cont, ret );
else expr_code( stream, node, 0 );
}
void
output_paren(const char *text)
{
switch (ofmt) {
case OF_LATEX:
printf ("\\textrm{\\footnotesize %s}", text);
break;
case OF_TEXT:
printf ("(%s) ", text);
break;
case OF_TEXTBLK:
do_block("(", "(", "(");
do_block(text, "", "");
do_block(")", ")", ")");
break;
}
}
static
fn_decl(stream, name, decl, block, frame_sz) {
auto ret = new_label();
start_fn(decl);
prolog(stream, name, frame_sz);
do_block(stream, block, -1, -1, ret);
emit_label( stream, ret );
epilog(stream, frame_sz);
end_fn();
}
virtual void cipher(block_cipher::g_type::container_type & data)
{
int blocks = data.size() / m_blocklen;
data_type * in = data.data();
for (int i=0; i<blocks; i++)
{
do_block(in);
in += m_blocklen;
}
}
static void
input_from_stdin(size_t offset, size_t fp_len, int action, int options) {
size_t len, buf_len;
if (action == BMZ_A_LIST) {
do_list(0);
}
else {
void *data = read_from_fp(stdin, &len, &buf_len);
do_block(data, len, buf_len, offset, fp_len, action, options);
}
}
/*static*/ void *rnc_pack (void *original, long datalen, long *packlen) {
int i;
char *data = original;
long origlen = datalen;
packed = malloc(PACKED_DELTA);
if (!packed) {
perror("malloc");
exit(1);
}
packedlen = PACKED_DELTA;
packedpos = 20;
bwrite (packed+4, datalen);
bitpos = 18;
bitcount = 0;
bitbuf = 0;
write_bits (0, 2);
while (datalen > 0) {
blklen = datalen > BLOCKMAX ? BLOCKMAX : datalen;
blkstart = WINMAX - BLOCKMAX;
if (blkstart > origlen-datalen)
blkstart = origlen-datalen;
memcpy (blk, data-blkstart, blkstart+blklen);
for (i=0; i<HASHMAX; i++)
hashp[i] = -1;
ntuple = 0;
tuples[ntuple].rawlen = 0;
blklen += blkstart;
do_block();
data += bpos - blkstart;
datalen -= bpos - blkstart;
write_block();
}
if (bitcount > 0) {
write_bits (0, 17-bitcount); /* force flush */
packedpos -= 2; /* write_bits will have moved it on */
}
*packlen = packedpos;
bwrite (packed, RNC_SIGNATURE);
bwrite (packed+12, rnc_crc(packed+18, packedpos-18));
bwrite (packed+10, rnc_crc(original, origlen));
bwrite (packed+8, packedpos-18);
packed[16] = packed[17] = 0;
return packed;
}
void square_dgemm(const int M, const double *A, const double *B, double *C)
{
const int n_blocks = M / BLOCK_SIZE + (M%BLOCK_SIZE? 1 : 0);
int bi, bj, bk;
for (bi = 0; bi < n_blocks; ++bi) {
const int i = bi * BLOCK_SIZE;
for (bj = 0; bj < n_blocks; ++bj) {
const int j = bj * BLOCK_SIZE;
for (bk = 0; bk < n_blocks; ++bk) {
const int k = bk * BLOCK_SIZE;
do_block(M, A, B, C, i, j, k);
}
}
}
}
void square_dgemm (int lda, double* A, double* B, double* C)
{
double *A_block, *B_block;
posix_memalign((void **)&A_block, 64, BLOCK_L1 * BLOCK_L1 * sizeof(double));
posix_memalign((void **)&B_block, 64, BLOCK_L1 * BLOCK_L1 * sizeof(double));
// reorcder loops for cache efficiency
for (int t = 0; t < lda; t += BLOCK_L2)
{
// For each block column of B
for (int s = 0; s < lda; s += BLOCK_L2)
{
// For each block row of A
for (int r = 0; r < lda; r += BLOCK_L2)
{
// compute end index of smaller block
int end_k = t + min(BLOCK_L2, lda-t);
int end_j = s + min(BLOCK_L2, lda-s);
int end_i = r + min(BLOCK_L2, lda-r);
for (int k = t; k < end_k; k += BLOCK_L1)
{
// For each block column of B
for (int j = s; j < end_j; j += BLOCK_L1)
{
// For each block row of A
for (int i = r; i < end_i; i += BLOCK_L1)
{
// compute block size
int K = min(BLOCK_L1, end_k-k);
int N = min(BLOCK_L1, end_j-j);
int M = min(BLOCK_L1, end_i-i);
/* Performs a smaller dgemm operation
* C' := C' + A' * B'
* where C' is M-by-N, A' is M-by-K, and B' is K-by-N. */
//do_block(lda, M, N, K, A + i + k*lda, B + k + j*lda, C + i + j*lda);
do_block(lda, M, N, K, A + i + k*lda, B + k + j*lda, C + i + j*lda, A_block, B_block);
}
}
}
}
}
}
free(A_block);
free(B_block);
}
void do_block2(const int lda,
const double * A, const double * B, double * C,
const int i, const int j, const int k)
{
const int n_blocks = BLOCK2_SIZE / BLOCK3_SIZE + (BLOCK2_SIZE%BLOCK3_SIZE? 1 : 0);
int bi, bj, bk;
for (bi = 0; bi < n_blocks; ++bi) {
const int r = bi * BLOCK3_SIZE;
for (bj = 0; bj < n_blocks; ++bj) {
const int s = bj * BLOCK3_SIZE;
for (bk = 0; bk < n_blocks; ++bk) {
const int t = bk * BLOCK3_SIZE;
do_block(lda, A, B, C, i+r, j+s, k+t);
}
}
}
}
static void
input_from_file(const char *fname, size_t offset, size_t fp_len, int action,
int options) {
size_t len, buf_len;
int fd = open(fname, O_RDONLY, 0);
if (fd == -1) DIE("cannot open '%s'", fname);
if (action == BMZ_A_LIST) {
do_list(fd);
}
else {
void *data = read_from_fd(fd, &len, &buf_len);
do_block(data, len, buf_len, offset, fp_len, action, options);
}
/* close and free etc. are omitted intentionally */
}
/* This routine performs a dgemm operation
* C := C + A * B
* where A, B, and C are lda-by-lda matrices stored in column-major format.
* On exit, A and B maintain their input values. */
void square_dgemm (int lda, double* A, double* B, double* C)
{
/* For each block-row of A */
for (int i = 0; i < lda; i += BLOCK_SIZE)
/* For each block-column of B */
for (int j = 0; j < lda; j += BLOCK_SIZE)
/* Accumulate block dgemms into block of C */
for (int k = 0; k < lda; k += BLOCK_SIZE)
{
/* Correct block dimensions if block "goes off edge of" the matrix */
int M = min (BLOCK_SIZE, lda-i);
int N = min (BLOCK_SIZE, lda-j);
int K = min (BLOCK_SIZE, lda-k);
/* Perform individual block dgemm */
do_block(lda, M, N, K, A + i + k*lda, B + k + j*lda, C + i + j*lda);
}
}
void
block_sigio(void)
{
PRINT_TIME(NOFD, &tnow, &tprev, "block_sigio: entered sigio_blocked = %d",
sigio_blocked);
if (sigio_blocked == 1) {
PRINT_TIME(NOFD, &tnow, &tprev,
"block_sigio: is already blocked returning");
return;
}
DEBG(MSG_INTR, "block_sigio: blocking\n");
do_block();
num_block_sigio++;
DEBG(MSG_INTR, "block_sigio: unblocked = %15d blocked = %15d\n",
num_unblock_sigio, num_block_sigio);
DEBG(MSG_INTR, "block_sigio: setting sigio_blocked = 1\n");
sigio_blocked = 1;
DEBG(MSG_INTR, "block_sigio: done setting sigio_blocked = 1\n");
PRINT_TIME(NOFD, &tnow, &tprev, "block_sigio: returning");
}
void
block_all_signals()
{
do_block(1);
}
static void
push_buffer_impl(
XferElement *elt,
gpointer buf,
size_t len)
{
XferDestDevice *self = XFER_DEST_DEVICE(elt);
gpointer to_free = buf;
/* Handle EOF */
if (!buf) {
/* write out the partial buffer, if there's anything in it */
if (self->partial_length) {
if (!do_block(self, self->block_size, self->partial)) {
return;
}
self->partial_length = 0;
}
device_finish_file(self->device);
return;
}
/* set up the block buffer, now that we can depend on having a blocksize
* from the device */
if (!self->partial) {
self->partial = g_try_malloc(self->device->block_size);
if (self->partial == NULL) {
xfer_cancel_with_error(elt, "%s: Cannot allocate memory",
self->device->device_name);
wait_until_xfer_cancelled(elt->xfer);
return;
}
self->block_size = self->device->block_size;
self->partial_length = 0;
}
/* if we already have data in the buffer, add the new data to it */
if (self->partial_length != 0) {
gsize to_copy = min(self->block_size - self->partial_length, len);
memmove((char *)self->partial + self->partial_length, buf, to_copy);
buf = (gpointer)(to_copy + (char *)buf);
len -= to_copy;
self->partial_length += to_copy;
}
/* and if the buffer is now full, write the block */
if (self->partial_length == self->block_size) {
if (!do_block(self, self->block_size, self->partial)) {
g_free(to_free);
return;
}
self->partial_length = 0;
}
/* write any whole blocks directly from the push buffer */
while (len >= self->block_size) {
if (!do_block(self, self->block_size, buf)) {
g_free(to_free);
return;
}
buf = (gpointer)(self->block_size + (char *)buf);
len -= self->block_size;
}
/* and finally store any leftover data in the partial buffer */
if (len) {
memmove(self->partial, buf, len);
self->partial_length = len;
}
g_free(to_free);
}
void
gnupg_block_all_signals ()
{
do_block(1);
}
void
gnupg_unblock_all_signals ()
{
do_block(0);
}
void
unblock_all_signals()
{
do_block(0);
}
