siz_t bli_thread_get_range_ndim
(
dir_t direct,
thrinfo_t* thr,
obj_t* a,
obj_t* b,
obj_t* c,
cntl_t* cntl,
cntx_t* cntx,
dim_t* start,
dim_t* end
)
{
bszid_t bszid = bli_cntl_bszid( cntl );
opid_t family = bli_cntx_get_family( cntx );
// This is part of trsm's current implementation, whereby right side
// cases are implemented in left-side micro-kernels, which requires
// we swap the usage of the register blocksizes for the purposes of
// packing A and B.
if ( family == BLIS_TRSM )
{
if ( bli_obj_root_is_triangular( *b ) ) bszid = BLIS_MR;
else bszid = BLIS_NR;
}
blksz_t* bmult = bli_cntx_get_bmult( bszid, cntx );
obj_t* x;
bool_t use_weighted;
// Use the operation family to choose the one of the two matrices
// being partitioned that potentially has structure, and also to
// decide whether or not we need to use weighted range partitioning.
// NOTE: It's important that we use non-weighted range partitioning
// for hemm and symm (ie: the gemm family) because the weighted
// function will mistakenly skip over unstored regions of the
// structured matrix, even though they represent part of that matrix
// that will be dense and full (after packing).
if ( family == BLIS_GEMM ) { x = b; use_weighted = FALSE; }
else if ( family == BLIS_HERK ) { x = c; use_weighted = TRUE; }
else if ( family == BLIS_TRMM ) { x = b; use_weighted = TRUE; }
else /*family == BLIS_TRSM*/ { x = b; use_weighted = FALSE; }
if ( use_weighted )
{
if ( direct == BLIS_FWD )
return bli_thread_get_range_weighted_l2r( thr, x, bmult, start, end );
else
return bli_thread_get_range_weighted_r2l( thr, x, bmult, start, end );
}
else
{
if ( direct == BLIS_FWD )
return bli_thread_get_range_l2r( thr, x, bmult, start, end );
else
return bli_thread_get_range_r2l( thr, x, bmult, start, end );
}
}
void bli_trmm_blk_var2b( obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
gemm_t* cntl,
thrinfo_t* thread )
{
obj_t a_pack_s;
obj_t b1_pack_s, c1_pack_s;
obj_t b1, c1;
obj_t* a_pack = NULL;
obj_t* b1_pack = NULL;
obj_t* c1_pack = NULL;
dim_t i;
dim_t b_alg;
// Prune any zero region that exists along the partitioning dimension.
bli_trmm_prune_unref_mparts_n( a, b, c );
if( bli_thread_am_ochief( thread ) ) {
// Initialize object for packing A
bli_obj_init_pack( &a_pack_s );
bli_packm_init( a, &a_pack_s,
cntx, bli_cntl_sub_packm_a( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntx, bli_cntl_sub_scalm( cntl ) );
}
a_pack = bli_thread_obroadcast( thread, &a_pack_s );
// Initialize pack objects for B and C that are passed into packm_init().
if( bli_thread_am_ichief( thread ) ) {
bli_obj_init_pack( &b1_pack_s );
bli_obj_init_pack( &c1_pack_s );
}
b1_pack = bli_thread_ibroadcast( thread, &b1_pack_s );
c1_pack = bli_thread_ibroadcast( thread, &c1_pack_s );
// Pack A (if instructed).
bli_packm_int( a, a_pack,
cntx, bli_cntl_sub_packm_a( cntl ),
bli_thrinfo_sub_opackm( thread ) );
dim_t my_start, my_end;
bli_thread_get_range_weighted_r2l( thread, b,
bli_cntx_get_bmult( bli_cntl_bszid( cntl ), cntx ),
&my_start, &my_end );
// Partition along the n dimension.
for ( i = my_start; i < my_end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_b( i, my_end, b,
bli_cntl_bszid( cntl ), cntx );
// Acquire partitions for B1 and C1.
bli_acquire_mpart_r2l( BLIS_SUBPART1,
i, b_alg, b, &b1 );
bli_acquire_mpart_r2l( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize objects for packing A1 and B1.
if( bli_thread_am_ichief( thread ) ) {
bli_packm_init( &b1, b1_pack,
cntx, bli_cntl_sub_packm_b( cntl ) );
bli_packm_init( &c1, c1_pack,
cntx, bli_cntl_sub_packm_c( cntl ) );
}
bli_thread_ibarrier( thread );
// Pack B1 (if instructed).
bli_packm_int( &b1, b1_pack,
cntx, bli_cntl_sub_packm_b( cntl ),
bli_thrinfo_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntx, bli_cntl_sub_packm_c( cntl ),
bli_thrinfo_sub_ipackm( thread ) );
// Perform trmm subproblem.
bli_trmm_int( &BLIS_ONE,
a_pack,
b1_pack,
&BLIS_ONE,
c1_pack,
cntx,
bli_cntl_sub_gemm( cntl ),
bli_thrinfo_sub_self( thread ) );
bli_thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
cntx, bli_cntl_sub_unpackm_c( cntl ),
bli_thrinfo_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_thread_obarrier( thread );
if( bli_thread_am_ochief( thread ) )
bli_packm_release( a_pack, bli_cntl_sub_packm_a( cntl ) );
if( bli_thread_am_ichief( thread ) ) {
bli_packm_release( b1_pack, bli_cntl_sub_packm_b( cntl ) );
bli_packm_release( c1_pack, bli_cntl_sub_packm_c( cntl ) );
}
}
int main( int argc, char** argv )
{
//bli_init();
#if 0
obj_t a, b, c;
obj_t aa, bb, cc;
dim_t m, n, k;
num_t dt;
uplo_t uploa, uplob, uploc;
{
dt = BLIS_DOUBLE;
m = 6;
k = 6;
n = 6;
bli_obj_create( dt, m, k, 0, 0, &a );
bli_obj_create( dt, k, n, 0, 0, &b );
bli_obj_create( dt, m, n, 0, 0, &c );
uploa = BLIS_UPPER;
uploa = BLIS_LOWER;
bli_obj_set_struc( BLIS_TRIANGULAR, &a );
bli_obj_set_uplo( uploa, &a );
bli_obj_set_diag_offset( -2, &a );
uplob = BLIS_UPPER;
uplob = BLIS_LOWER;
bli_obj_set_struc( BLIS_TRIANGULAR, &b );
bli_obj_set_uplo( uplob, &b );
bli_obj_set_diag_offset( -2, &b );
uploc = BLIS_UPPER;
//uploc = BLIS_LOWER;
//uploc = BLIS_ZEROS;
//uploc = BLIS_DENSE;
bli_obj_set_struc( BLIS_HERMITIAN, &c );
//bli_obj_set_struc( BLIS_TRIANGULAR, &c );
bli_obj_set_uplo( uploc, &c );
bli_obj_set_diag_offset( 1, &c );
bli_obj_alias_to( &a, &aa ); (void)aa;
bli_obj_alias_to( &b, &bb ); (void)bb;
bli_obj_alias_to( &c, &cc ); (void)cc;
bli_randm( &a );
bli_randm( &b );
bli_randm( &c );
//bli_mkherm( &a );
//bli_mktrim( &a );
bli_prune_unref_mparts( &cc, BLIS_M,
&aa, BLIS_N );
bli_printm( "c orig", &c, "%4.1f", "" );
bli_printm( "c alias", &cc, "%4.1f", "" );
bli_printm( "a orig", &a, "%4.1f", "" );
bli_printm( "a alias", &aa, "%4.1f", "" );
//bli_obj_print( "a struct", &a );
}
#endif
dim_t p_begin, p_max, p_inc;
gint_t m_input, n_input;
char uploa_ch;
doff_t diagoffa;
dim_t bf;
dim_t n_way;
char part_dim_ch;
bool_t go_fwd;
char out_ch;
obj_t a;
blksz_t bfs;
thrinfo_t thrinfo;
dim_t m, n;
uplo_t uploa;
bool_t part_m_dim, part_n_dim;
bool_t go_bwd;
dim_t p;
num_t dt;
dim_t start, end;
dim_t width;
siz_t area;
gint_t t_begin, t_stop, t_inc;
dim_t t;
if ( argc == 13 )
{
sscanf( argv[1], "%u", &p_begin );
sscanf( argv[2], "%u", &p_max );
sscanf( argv[3], "%u", &p_inc );
sscanf( argv[4], "%d", &m_input );
sscanf( argv[5], "%d", &n_input );
sscanf( argv[6], "%c", &uploa_ch );
sscanf( argv[7], "%d", &diagoffa );
sscanf( argv[8], "%u", &bf );
sscanf( argv[9], "%u", &n_way );
sscanf( argv[10], "%c", &part_dim_ch );
sscanf( argv[11], "%u", &go_fwd );
sscanf( argv[12], "%c", &out_ch );
}
else
{
printf( "\n" );
printf( " %s\n", argv[0] );
printf( "\n" );
printf( " Simulate the dimension ranges assigned to threads when\n" );
printf( " partitioning a matrix for parallelism in BLIS.\n" );
printf( "\n" );
printf( " Usage:\n" );
printf( "\n" );
printf( " %s p_beg p_max p_inc m n uplo doff bf n_way part_dim go_fwd out\n", argv[0] );
printf( "\n" );
printf( " p_beg: the first problem size p to test.\n" );
printf( " p_max: the maximum problem size p to test.\n" );
printf( " p_inc: the increase in problem size p between tests.\n" );
printf( " m: the m dimension:\n" );
printf( " n: the n dimension:\n" );
printf( " if m,n = -1: bind m,n to problem size p.\n" );
printf( " if m,n = 0: bind m,n to p_max.\n" );
printf( " if m,n > 0: hold m,n = c constant for all p.\n" );
printf( " uplo: the uplo field of the matrix being partitioned:\n" );
printf( " 'l': lower-stored (BLIS_LOWER)\n" );
printf( " 'u': upper-stored (BLIS_UPPER)\n" );
printf( " 'd': densely-stored (BLIS_DENSE)\n" );
printf( " doff: the diagonal offset of the matrix being partitioned.\n" );
printf( " bf: the simulated blocking factor. all thread ranges must\n" );
printf( " be a multiple of bf, except for the range that contains\n" );
printf( " the edge case (if one exists). the blocking factor\n" );
printf( " would typically correspond to a register blocksize.\n" );
printf( " n_way: the number of ways of parallelism for which we are\n" );
printf( " partitioning (i.e.: the number of threads, or thread\n" );
printf( " groups).\n" );
printf( " part_dim: the dimension to partition:\n" );
printf( " 'm': partition the m dimension.\n" );
printf( " 'n': partition the n dimension.\n" );
printf( " go_fwd: the direction to partition:\n" );
printf( " '1': forward, e.g. left-to-right (part_dim = 'm') or\n" );
printf( " top-to-bottom (part_dim = 'n')\n" );
printf( " '0': backward, e.g. right-to-left (part_dim = 'm') or\n" );
printf( " bottom-to-top (part_dim = 'n')\n" );
printf( " NOTE: reversing the direction does not change the\n" );
printf( " subpartitions' widths, but it does change which end of\n" );
printf( " the index range receives the edge case, if it exists.\n" );
printf( " out: the type of output per thread-column:\n" );
printf( " 'w': the width (and area) of the thread's subpartition\n" );
printf( " 'r': the actual ranges of the thread's subpartition\n" );
printf( " where the start and end points of each range are\n" );
printf( " inclusive and exclusive, respectively.\n" );
printf( "\n" );
exit(1);
}
if ( m_input == 0 ) m_input = p_max;
if ( n_input == 0 ) n_input = p_max;
if ( part_dim_ch == 'm' ) { part_m_dim = TRUE; part_n_dim = FALSE; }
else { part_m_dim = FALSE; part_n_dim = TRUE; }
go_bwd = !go_fwd;
if ( uploa_ch == 'l' ) uploa = BLIS_LOWER;
else if ( uploa_ch == 'u' ) uploa = BLIS_UPPER;
else uploa = BLIS_DENSE;
if ( part_n_dim )
{
if ( bli_is_upper( uploa ) ) { t_begin = n_way-1; t_stop = -1; t_inc = -1; }
else /* if lower or dense */ { t_begin = 0; t_stop = n_way; t_inc = 1; }
}
else // if ( part_m_dim )
{
if ( bli_is_lower( uploa ) ) { t_begin = n_way-1; t_stop = -1; t_inc = -1; }
else /* if upper or dense */ { t_begin = 0; t_stop = n_way; t_inc = 1; }
}
printf( "\n" );
printf( " part: %3s doff: %3d bf: %3d output: %s\n",
( part_n_dim ? ( go_fwd ? "l2r" : "r2l" )
: ( go_fwd ? "t2b" : "b2t" ) ),
( int )diagoffa, ( int )bf,
( out_ch == 'w' ? "width(area)" : "ranges" ) );
printf( " uplo: %3c nt: %3u\n", uploa_ch, ( unsigned )n_way );
printf( "\n" );
printf( " " );
for ( t = t_begin; t != t_stop; t += t_inc )
{
if ( part_n_dim )
{
if ( t == t_begin ) printf( "left... " );
else if ( t == t_stop-t_inc ) printf( " ...right" );
else printf( " " );
}
else // if ( part_m_dim )
{
if ( t == t_begin ) printf( "top... " );
else if ( t == t_stop-t_inc ) printf( " ...bottom" );
else printf( " " );
}
}
printf( "\n" );
printf( "%4c x %4c ", 'm', 'n' );
for ( t = t_begin; t != t_stop; t += t_inc )
{
printf( "%9s %u ", "thread", ( unsigned )t );
}
printf( "\n" );
printf( "-------------" );
for ( t = t_begin; t != t_stop; t += t_inc )
{
printf( "-------------" );
}
printf( "\n" );
for ( p = p_begin; p <= p_max; p += p_inc )
{
if ( m_input < 0 ) m = ( dim_t )p;
else m = ( dim_t )m_input;
if ( n_input < 0 ) n = ( dim_t )p;
else n = ( dim_t )n_input;
dt = BLIS_DOUBLE;
bli_obj_create( dt, m, n, 0, 0, &a );
bli_obj_set_struc( BLIS_TRIANGULAR, &a );
bli_obj_set_uplo( uploa, &a );
bli_obj_set_diag_offset( diagoffa, &a );
bli_randm( &a );
bli_blksz_init_easy( &bfs, bf, bf, bf, bf );
printf( "%4u x %4u ", ( unsigned )m, ( unsigned )n );
for ( t = t_begin; t != t_stop; t += t_inc )
{
thrinfo.n_way = n_way;
thrinfo.work_id = t;
if ( part_n_dim && go_fwd )
area = bli_thread_get_range_weighted_l2r( &thrinfo, &a, &bfs, &start, &end );
else if ( part_n_dim && go_bwd )
area = bli_thread_get_range_weighted_r2l( &thrinfo, &a, &bfs, &start, &end );
else if ( part_m_dim && go_fwd )
area = bli_thread_get_range_weighted_t2b( &thrinfo, &a, &bfs, &start, &end );
else // ( part_m_dim && go_bwd )
area = bli_thread_get_range_weighted_b2t( &thrinfo, &a, &bfs, &start, &end );
width = end - start;
if ( out_ch == 'w' ) printf( "%4u(%6u) ", ( unsigned )width,
( unsigned )area );
else printf( "[%4u,%4u) ", ( unsigned )start,
( unsigned )end );
}
printf( "\n" );
bli_obj_free( &a );
}
//bli_finalize();
return 0;
}
