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_gemm_blk_var1f( obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
gemm_t* cntl,
gemm_thrinfo_t* thread )
{
//The s is for "lives on the stack"
obj_t b_pack_s;
obj_t a1_pack_s, c1_pack_s;
obj_t a1, c1;
obj_t* a1_pack = NULL;
obj_t* b_pack = NULL;
obj_t* c1_pack = NULL;
dim_t i;
dim_t b_alg;
if( thread_am_ochief( thread ) ) {
// Initialize object for packing B.
bli_obj_init_pack( &b_pack_s );
bli_packm_init( b, &b_pack_s,
cntx, cntl_sub_packm_b( cntl ) );
// Scale C by beta (if instructed).
// Since scalm doesn't support multithreading yet, must be done by chief thread (ew)
bli_scalm_int( &BLIS_ONE,
c,
cntx, cntl_sub_scalm( cntl ) );
}
b_pack = thread_obroadcast( thread, &b_pack_s );
// Initialize objects passed into bli_packm_init for A and C
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
bli_obj_init_pack( &c1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
c1_pack = thread_ibroadcast( thread, &c1_pack_s );
// Pack B (if instructed).
bli_packm_int( b, b_pack,
cntx, cntl_sub_packm_b( cntl ),
gemm_thread_sub_opackm( thread ) );
dim_t my_start, my_end;
bli_get_range_t2b( thread, a,
bli_cntx_get_bmult( cntl_bszid( cntl ), cntx ),
&my_start, &my_end );
// Partition along the m dimension.
for ( i = my_start; i < my_end; i += b_alg )
{
// Determine the current algorithmic blocksize.
// NOTE: Use of a (for execution datatype) is intentional!
// This causes the right blocksize to be used if c and a are
// complex and b is real.
b_alg = bli_determine_blocksize_f( i, my_end, a,
cntl_bszid( cntl ), cntx );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize objects for packing A1 and C1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntx, cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, c1_pack,
cntx, cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntx, cntl_sub_packm_a( cntl ),
gemm_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntx, cntl_sub_packm_c( cntl ),
gemm_thread_sub_ipackm( thread ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
c1_pack,
cntx,
cntl_sub_gemm( cntl ),
gemm_thread_sub_gemm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread
bli_unpackm_int( c1_pack, &c1,
cntx, cntl_sub_unpackm_c( cntl ),
gemm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( b_pack, cntl_sub_packm_b( cntl ) );
if( thread_am_ichief( thread ) ){
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
bli_packm_release( c1_pack, cntl_sub_packm_c( cntl ) );
}
}
void bli_trmm_blk_var2b( obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
gemm_t* cntl,
trmm_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( thread_am_ochief( thread ) ) {
// Initialize object for packing A
bli_obj_init_pack( &a_pack_s );
bli_packm_init( a, &a_pack_s,
cntx, cntl_sub_packm_a( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntx, cntl_sub_scalm( cntl ) );
}
a_pack = thread_obroadcast( thread, &a_pack_s );
// Initialize pack objects for B and C that are passed into packm_init().
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &b1_pack_s );
bli_obj_init_pack( &c1_pack_s );
}
b1_pack = thread_ibroadcast( thread, &b1_pack_s );
c1_pack = thread_ibroadcast( thread, &c1_pack_s );
// Pack A (if instructed).
bli_packm_int( a, a_pack,
cntx, cntl_sub_packm_a( cntl ),
trmm_thread_sub_opackm( thread ) );
dim_t my_start, my_end;
bli_get_range_weighted_r2l( thread, b,
bli_cntx_get_bmult( 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,
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( thread_am_ichief( thread ) ) {
bli_packm_init( &b1, b1_pack,
cntx, cntl_sub_packm_b( cntl ) );
bli_packm_init( &c1, c1_pack,
cntx, cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack B1 (if instructed).
bli_packm_int( &b1, b1_pack,
cntx, cntl_sub_packm_b( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntx, cntl_sub_packm_c( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Perform trmm subproblem.
bli_trmm_int( &BLIS_ONE,
a_pack,
b1_pack,
&BLIS_ONE,
c1_pack,
cntx,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
cntx, cntl_sub_unpackm_c( cntl ),
trmm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( a_pack, cntl_sub_packm_a( cntl ) );
if( thread_am_ichief( thread ) ) {
bli_packm_release( b1_pack, cntl_sub_packm_b( cntl ) );
bli_packm_release( c1_pack, cntl_sub_packm_c( cntl ) );
}
}