func_t* bli_gemm_cntl_ukrs( gemm_t* cntl )
{
dim_t max_depth = 10;
dim_t i;
for ( i = 0; ; ++i )
{
// If the gemm sub-tree is NULL, we are at the leaf.
if ( cntl_sub_gemm( cntl ) == NULL ) break;
// If the above branch was not taken, we can assume the gemm
// sub-tree is valid. Here, we step down into that sub-tree.
cntl = cntl_sub_gemm( cntl );
// Safeguard against infinite loops due to bad control tree
// configuration.
if ( i == max_depth ) bli_abort();
}
return cntl_gemm_ukrs( cntl );
}
void bli_trsm_l_blk_var4( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
trsm_t* cntl )
{
obj_t a1, a1_pack;
obj_t b_pack;
obj_t c1;
dim_t i;
dim_t bm_alg;
dim_t m_trans;
dim_t offB;
// Initialize all pack objects that are passed into packm_init().
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &b_pack );
// Query dimension in partitioning direction.
m_trans = bli_obj_length_after_trans( *a );
// Use the diagonal offset of A to skip over the zero region.
offB = bli_abs( bli_obj_diag_offset_after_trans( *a ) );
// Initialize object for packing B.
bli_packm_init( b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Fuse the first iteration with incremental packing and computation.
{
obj_t b_inc, b_pack_inc;
obj_t c1_inc;
dim_t j;
dim_t bn_inc;
dim_t n_trans;
// Query dimension in partitioning direction.
n_trans = bli_obj_width( b_pack );
// Determine the current algorithmic blocksize.
bm_alg = bli_determine_blocksize_f( offB, m_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
offB, bm_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
offB, bm_alg, c, &c1 );
// Initialize objects for packing A1 and C1.
bli_packm_init( &a1, &a1_pack, cntl_sub_packm_a( cntl ) );
// Pack A1 and scale by alpha (if instructed).
bli_packm_int( alpha, &a1, &a1_pack, cntl_sub_packm_a( cntl ) );
// Partition along the n dimension.
for ( j = 0; j < n_trans; j += bn_inc )
{
// Determine the current incremental packing blocksize.
bn_inc = bli_determine_blocksize_f( j, n_trans, b,
cntl_blocksize_aux( cntl ) );
// Acquire partitions.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, b, &b_inc );
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, &b_pack, &b_pack_inc );
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, &c1, &c1_inc );
// Pack B1 and scale by alpha (if instructed).
bli_packm_int( alpha, &b_inc, &b_pack_inc, cntl_sub_packm_b( cntl ) );
// Perform trsm subproblem.
bli_trsm_int( BLIS_LEFT,
alpha,
&a1_pack,
&b_pack_inc,
beta,
&c1_inc,
cntl_sub_trsm( cntl ) );
}
// Unpack B to the corresponding region of C. (Note that B and C1 are
// conformal since A1 is square.)
//bli_unpackm_int( &b_pack, &c1,
// cntl_sub_unpackm_c( cntl ) );
}
// Partition along the remaining portion of the m dimension.
for ( i = offB + bm_alg; i < m_trans; i += bm_alg )
{
// Determine the current algorithmic blocksize.
bm_alg = bli_determine_blocksize_f( i, m_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, bm_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, bm_alg, c, &c1 );
// Initialize object for packing A1.
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Pack A1 and scale by alpha (if instructed).
bli_packm_int( alpha,
&a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Perform trsm subproblem.
if ( bli_obj_intersects_diag( a1_pack ) )
bli_trsm_int( BLIS_LEFT,
alpha,
&a1_pack,
&b_pack,
beta,
&c1,
cntl_sub_trsm( cntl ) );
else
bli_gemm_int( &BLIS_MINUS_ONE,
&a1_pack,
&b_pack,
&BLIS_ONE,
&c1,
cntl_sub_gemm( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_obj_release_pack( &a1_pack );
bli_obj_release_pack( &b_pack );
}
void bli_trmm_lu_blk_var4( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
trmm_t* cntl )
{
obj_t a1, a1_pack;
obj_t b_pack;
obj_t c1, c1_pack;
dim_t i;
dim_t bm_alg;
dim_t mT_trans;
// Initialize all pack objects that are passed into packm_init().
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &b_pack );
bli_obj_init_pack( &c1_pack );
// Query dimension in partitioning direction. Use the diagonal offset
// to stop short of the zero region.
mT_trans = bli_abs( bli_obj_diag_offset_after_trans( *a ) ) +
bli_obj_width_after_trans( *a );
// Scale C by beta (if instructed).
bli_scalm_int( beta,
c,
cntl_sub_scalm( cntl ) );
// Initialize object for packing B.
bli_packm_init( b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Fuse the first iteration with incremental packing and computation.
{
obj_t b_inc, b_pack_inc;
obj_t c1_pack_inc;
dim_t j;
dim_t bn_inc;
dim_t n_trans;
// Query dimension in partitioning direction.
n_trans = bli_obj_width( b_pack );
// Determine the current algorithmic blocksize.
bm_alg = bli_determine_blocksize_f( 0, mT_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
0, bm_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
0, bm_alg, c, &c1 );
// Initialize objects for packing A1 and C1.
bli_packm_init( &a1, &a1_pack, cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, &c1_pack, cntl_sub_packm_c( cntl ) );
// Pack A1 and scale by alpha (if instructed).
bli_packm_int( alpha, &a1, &a1_pack, cntl_sub_packm_a( cntl ) );
// Pack C1 and scale by beta (if instructed).
bli_packm_int( beta, &c1, &c1_pack, cntl_sub_packm_c( cntl ) );
// Partition along the n dimension.
for ( j = 0; j < n_trans; j += bn_inc )
{
// Determine the current incremental packing blocksize.
bn_inc = bli_determine_blocksize_f( j, n_trans, b,
cntl_blocksize_aux( cntl ) );
// Acquire partitions.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, b, &b_inc );
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, &b_pack, &b_pack_inc );
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, &c1_pack, &c1_pack_inc );
// Pack B1 and scale by alpha (if instructed).
bli_packm_int( alpha, &b_inc, &b_pack_inc, cntl_sub_packm_b( cntl ) );
// Perform trmm subproblem.
bli_trmm_int( BLIS_LEFT,
alpha,
&a1_pack,
&b_pack_inc,
beta,
&c1_pack_inc,
cntl_sub_trmm( cntl ) );
}
// Unpack C1 (if C1 was packed).
bli_unpackm_int( &c1_pack, &c1, cntl_sub_unpackm_c( cntl ) );
}
// Partition along the remaining portion of the m dimension.
for ( i = bm_alg; i < mT_trans; i += bm_alg )
{
// Determine the current algorithmic blocksize.
bm_alg = bli_determine_blocksize_f( i, mT_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, bm_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, bm_alg, c, &c1 );
// Initialize objects for packing A1 and C1.
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, &c1_pack,
cntl_sub_packm_c( cntl ) );
// Pack A1 and scale by alpha (if instructed).
bli_packm_int( alpha,
&a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Pack C1 and scale by beta (if instructed).
bli_packm_int( beta,
&c1, &c1_pack,
cntl_sub_packm_c( cntl ) );
// Perform trmm subproblem.
if ( bli_obj_intersects_diag( a1_pack ) )
bli_trmm_int( BLIS_LEFT,
alpha,
&a1_pack,
&b_pack,
beta,
&c1_pack,
cntl_sub_trmm( cntl ) );
else
bli_gemm_int( alpha,
&a1_pack,
&b_pack,
&BLIS_ONE,
&c1_pack,
cntl_sub_gemm( cntl ) );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( &c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_obj_release_pack( &a1_pack );
bli_obj_release_pack( &b_pack );
bli_obj_release_pack( &c1_pack );
}
void bli_herk_blk_var1f( obj_t* a,
obj_t* ah,
obj_t* c,
gemm_t* cntl,
herk_thrinfo_t* thread )
{
obj_t ah_pack_s;
obj_t a1_pack_s, c1_pack_s;
obj_t a1, c1;
obj_t* a1_pack;
obj_t* c1_pack;
obj_t* ah_pack;
dim_t i;
dim_t b_alg;
// Prune any zero region that exists along the partitioning dimension.
bli_herk_prune_unref_mparts_m( a, ah, c );
if( thread_am_ochief( thread ) ) {
// Initialize object for packing A'.
bli_obj_init_pack( &ah_pack_s );
bli_packm_init( ah, &ah_pack_s,
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,
cntl_sub_scalm( cntl ) );
}
ah_pack = thread_obroadcast( thread, &ah_pack_s );
// Initialize pack objects that are passed into 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 A' (if instructed).
bli_packm_int( ah, ah_pack,
cntl_sub_packm_b( cntl ),
herk_thread_sub_opackm( thread ) );
dim_t my_start, my_end;
bli_get_range_weighted_t2b( thread, c,
bli_blksz_get_mult_for_obj( a, cntl_blocksize( cntl ) ),
&my_start, &my_end );
// Partition along the m dimension.
for ( i = my_start; i < my_end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, my_end, a,
cntl_blocksize( cntl ) );
// 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,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, c1_pack,
cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
herk_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntl_sub_packm_c( cntl ),
herk_thread_sub_ipackm( thread ) );
// Perform herk subproblem.
bli_herk_int( &BLIS_ONE,
a1_pack,
ah_pack,
&BLIS_ONE,
c1_pack,
cntl_sub_gemm( cntl ),
herk_thread_sub_herk( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
herk_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( ah_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_gemm_ker_var2( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl )
{
num_t dt_exec = bli_obj_execution_datatype( *c );
dim_t m = bli_obj_length( *c );
dim_t n = bli_obj_width( *c );
dim_t k = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t pd_a = bli_obj_panel_dim( *a );
inc_t cs_a = bli_obj_col_stride( *a );
inc_t ps_a = bli_obj_panel_stride( *a );
void* buf_b = bli_obj_buffer_at_off( *b );
inc_t rs_b = bli_obj_row_stride( *b );
inc_t pd_b = bli_obj_panel_dim( *b );
inc_t ps_b = bli_obj_panel_stride( *b );
void* buf_c = bli_obj_buffer_at_off( *c );
inc_t rs_c = bli_obj_row_stride( *c );
inc_t cs_c = bli_obj_col_stride( *c );
obj_t scalar_a;
obj_t scalar_b;
void* buf_alpha;
void* buf_beta;
FUNCPTR_T f;
func_t* gemm_ukrs;
void* gemm_ukr;
// Detach and multiply the scalars attached to A and B.
bli_obj_scalar_detach( a, &scalar_a );
bli_obj_scalar_detach( b, &scalar_b );
bli_mulsc( &scalar_a, &scalar_b );
// Grab the addresses of the internal scalar buffers for the scalar
// merged above and the scalar attached to C.
buf_alpha = bli_obj_internal_scalar_buffer( scalar_b );
buf_beta = bli_obj_internal_scalar_buffer( *c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Extract from the control tree node the func_t object containing
// the gemm micro-kernel function addresses, and then query the
// function address corresponding to the current datatype.
gemm_ukrs = ( func_t* )cntl_sub_gemm( cntl );
gemm_ukr = bli_func_obj_query( dt_exec, gemm_ukrs );
// Invoke the function.
f( m,
n,
k,
buf_alpha,
buf_a, cs_a, pd_a, ps_a,
buf_b, rs_b, pd_b, ps_b,
buf_beta,
buf_c, rs_c, cs_c,
gemm_ukr );
}
void bli_trsm_u_blk_var4( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
trsm_t* cntl )
{
obj_t a1, a1_pack;
obj_t b_pack;
obj_t c1;
dim_t i;
dim_t bm_alg;
dim_t m_trans;
// Initialize all pack objects that are passed into packm_init().
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &b_pack );
// Query dimension in partitioning direction.
m_trans = bli_obj_length_after_trans( *a );
// Initialize object for packing B.
bli_packm_init( b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Find the offset to the first non-zero block of A.
for ( i = 0; i < m_trans; i += bm_alg )
{
// Determine the current algorithmic blocksize.
bm_alg = bli_determine_blocksize_b( i, m_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_b2t( BLIS_SUBPART1,
i, bm_alg, a, &a1 );
if ( !bli_obj_is_zeros( a1 ) ) break;
}
// Fuse the first iteration with incremental packing and computation.
{
obj_t b_inc, b_pack_inc;
obj_t c1_inc;
dim_t j;
dim_t bn_inc;
dim_t n_trans;
// Query dimension in partitioning direction.
n_trans = bli_obj_width( b_pack );
// Determine the current algorithmic blocksize.
bm_alg = bli_determine_blocksize_b( i, m_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_b2t( BLIS_SUBPART1,
i, bm_alg, a, &a1 );
bli_acquire_mpart_b2t( BLIS_SUBPART1,
i, bm_alg, c, &c1 );
// Initialize objects for packing A1 and C1.
bli_packm_init( &a1, &a1_pack, cntl_sub_packm_a( cntl ) );
// Pack A1 and scale by alpha (if instructed).
bli_packm_int( alpha, &a1, &a1_pack, cntl_sub_packm_a( cntl ) );
// Partition along the n dimension.
for ( j = 0; j < n_trans; j += bn_inc )
{
// Determine the current incremental packing blocksize.
bn_inc = bli_determine_blocksize_f( j, n_trans, b,
cntl_blocksize_aux( cntl ) );
// Acquire partitions.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, b, &b_inc );
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, &b_pack, &b_pack_inc );
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, &c1, &c1_inc );
// Pack B1 and scale by alpha (if instructed).
bli_packm_int( alpha, &b_inc, &b_pack_inc, cntl_sub_packm_b( cntl ) );
// Perform trsm subproblem.
bli_trsm_int( BLIS_LEFT,
alpha,
&a1_pack,
&b_pack_inc,
beta,
&c1_inc,
cntl_sub_trsm( cntl ) );
}
}
// Partition along the remaining portion of the m dimension.
for ( i = i + bm_alg; i < m_trans; i += bm_alg )
{
// Determine the current algorithmic blocksize.
bm_alg = bli_determine_blocksize_b( i, m_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_b2t( BLIS_SUBPART1,
i, bm_alg, a, &a1 );
bli_acquire_mpart_b2t( BLIS_SUBPART1,
i, bm_alg, c, &c1 );
// Initialize object for packing A1.
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Pack A1 and scale by alpha (if instructed).
bli_packm_int( alpha,
&a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
if ( bli_obj_intersects_diag( a1_pack ) )
bli_trsm_int( BLIS_LEFT,
alpha,
&a1_pack,
&b_pack,
beta,
&c1,
cntl_sub_trsm( cntl ) );
else
bli_gemm_int( &BLIS_MINUS_ONE,
&a1_pack,
&b_pack,
&BLIS_ONE,
&c1,
cntl_sub_gemm( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_obj_release_pack( &a1_pack );
bli_obj_release_pack( &b_pack );
}
void bli_trmm_blk_var1f( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl,
trmm_thrinfo_t* thread )
{
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;
// Prune any zero region that exists along the partitioning dimension.
bli_trmm_prune_unref_mparts_m( a, b, c );
if( thread_am_ochief( thread ) ) {
// Initialize object for packing B.
bli_obj_init_pack( &b_pack_s );
bli_packm_init( b, &b_pack_s,
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,
cntl_sub_scalm( cntl ) );
}
b_pack = thread_obroadcast( thread, &b_pack_s );
// Initialize all pack objects that are passed into packm_init().
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,
cntl_sub_packm_b( cntl ),
trmm_thread_sub_opackm( thread ) );
// Set the default length of and offset to the non-zero part of A.
//m_trans = bli_obj_length_after_trans( *a );
//offA = 0;
// If A is lower triangular, we have to adjust where the non-zero part of
// A begins. If A is upper triangular, we have to adjust the length of
// the non-zero part. If A is general/dense, then we keep the defaults.
//if ( bli_obj_is_lower( *a ) )
// offA = bli_abs( bli_obj_diag_offset_after_trans( *a ) );
//else if ( bli_obj_is_upper( *a ) )
// m_trans = bli_abs( bli_obj_diag_offset_after_trans( *a ) ) +
// bli_obj_width_after_trans( *a );
dim_t my_start, my_end;
bli_get_range_weighted_t2b( thread, a,
bli_blksz_get_mult_for_obj( a, cntl_blocksize( cntl ) ),
&my_start, &my_end );
// Partition along the m dimension.
for ( i = my_start; i < my_end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, my_end, a,
cntl_blocksize( cntl ) );
// 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,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, c1_pack,
cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntl_sub_packm_c( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Perform trmm subproblem.
bli_trmm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
c1_pack,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
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( 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_gemm_blk_var3f( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl,
gemm_thrinfo_t* thread )
{
obj_t c_pack_s;
obj_t a1_pack_s, b1_pack_s;
obj_t a1, b1;
obj_t* a1_pack = NULL;
obj_t* b1_pack = NULL;
obj_t* c_pack = NULL;
dim_t i;
dim_t b_alg;
dim_t k_trans;
if( thread_am_ochief( thread ) ){
// Initialize object for packing C
bli_obj_init_pack( &c_pack_s );
bli_packm_init( c, &c_pack_s,
cntl_sub_packm_c( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
}
c_pack = thread_obroadcast( thread, &c_pack_s );
// Initialize pack objects for A and B that are passed into packm_init().
if( thread_am_ichief( thread ) ){
bli_obj_init_pack( &a1_pack_s );
bli_obj_init_pack( &b1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
b1_pack = thread_ibroadcast( thread, &b1_pack_s );
// Pack C (if instructed).
bli_packm_int( c, c_pack,
cntl_sub_packm_c( cntl ),
gemm_thread_sub_opackm( thread ) );
// Query dimension in partitioning direction.
k_trans = bli_obj_width_after_trans( *a );
// Partition along the k dimension.
for ( i = 0; i < k_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
// NOTE: We call a gemm/hemm/symm-specific function to determine
// the kc blocksize so that we can implement the "nudging" of kc
// to be a multiple of mr or nr, as needed.
b_alg = bli_gemm_determine_kc_f( i, k_trans, a, b,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and B1.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, b, &b1 );
// Initialize objects for packing A1 and B1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &b1, b1_pack,
cntl_sub_packm_b( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
gemm_thread_sub_ipackm( thread ) );
// Pack B1 (if instructed).
bli_packm_int( &b1, b1_pack,
cntl_sub_packm_b( cntl ),
gemm_thread_sub_ipackm( thread ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
a1_pack,
b1_pack,
&BLIS_ONE,
c_pack,
cntl_sub_gemm( cntl ),
gemm_thread_sub_gemm( thread) );
// This variant executes multiple rank-k updates. Therefore, if the
// internal beta scalar on matrix C is non-zero, we must use it
// only for the first iteration (and then BLIS_ONE for all others).
// And since c_pack is a local obj_t, we can simply overwrite the
// internal beta scalar with BLIS_ONE once it has been used in the
// first iteration.
thread_ibarrier( thread );
if ( i == 0 && thread_am_ichief( thread ) ) bli_obj_scalar_reset( c_pack );
}
thread_obarrier( thread );
// Unpack C (if C was packed).
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ),
gemm_thread_sub_opackm( thread ) );
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
if( thread_am_ochief( thread ) )
bli_packm_release( c_pack, cntl_sub_packm_c( cntl ) );
if( thread_am_ichief( thread ) ){
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
bli_packm_release( b1_pack, cntl_sub_packm_b( cntl ) );
}
}
void bli_gemm_blk_var4f( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl,
gemm_thrinfo_t* thread )
{
extern packm_t* gemm3mh_packa_cntl_ro;
extern packm_t* gemm3mh_packa_cntl_io;
extern packm_t* gemm3mh_packa_cntl_rpi;
packm_t* packa_cntl_ro = gemm3mh_packa_cntl_ro;
packm_t* packa_cntl_io = gemm3mh_packa_cntl_io;
packm_t* packa_cntl_rpi = gemm3mh_packa_cntl_rpi;
//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;
dim_t m_trans;
if( thread_am_ochief( thread ) ) {
// Initialize object for packing B.
bli_obj_init_pack( &b_pack_s );
bli_packm_init( b, &b_pack_s,
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,
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,
cntl_sub_packm_b( cntl ),
gemm_thread_sub_opackm( thread ) );
// Query dimension in partitioning direction.
m_trans = bli_obj_length_after_trans( *a );
dim_t start, end;
bli_get_range_t2b( thread, 0, m_trans,
bli_blksz_get_mult_for_obj( a, cntl_blocksize( cntl ) ),
&start, &end );
// Partition along the m dimension.
for ( i = start; i < 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, end, a,
cntl_blocksize( cntl ) );
// 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,
packa_cntl_ro );
bli_packm_init( &c1, c1_pack,
cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
packa_cntl_ro,
gemm_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
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,
cntl_sub_gemm( cntl ),
gemm_thread_sub_gemm( thread ) );
thread_ibarrier( thread );
// Only apply beta within the first of three subproblems.
if ( thread_am_ichief( thread ) ) bli_obj_scalar_reset( c1_pack );
// Initialize objects for packing A1 and C1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
packa_cntl_io );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
packa_cntl_io,
gemm_thread_sub_ipackm( thread ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
c1_pack,
cntl_sub_gemm( cntl ),
gemm_thread_sub_gemm( thread ) );
thread_ibarrier( thread );
// Initialize objects for packing A1 and C1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
packa_cntl_rpi );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
packa_cntl_rpi,
gemm_thread_sub_ipackm( thread ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
c1_pack,
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,
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 ) ){
// It doesn't matter which packm cntl node we pass in, as long
// as it is valid, packm_release() will release the mem_t entry.
bli_packm_release( a1_pack, packa_cntl_ro );
bli_packm_release( c1_pack, cntl_sub_packm_c( cntl ) );
}
}
void bli_gemm_blk_var2( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
gemm_t* cntl )
{
obj_t a_pack_s;
obj_t b1_pack_s;
obj_t 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;
dim_t n_trans;
dim_t num_groups = bli_gemm_num_thread_groups( cntl->thread_info );
dim_t group_id = bli_gemm_group_id( cntl->thread_info );
if( bli_gemm_am_a_master( cntl->thread_info ) ) {
// Initialize object for packing A.
bli_obj_init_pack( &a_pack_s );
bli_packm_init( a, &a_pack_s,
cntl_sub_packm_a( cntl ) );
}
a_pack = bli_gemm_broadcast_a( cntl->thread_info, &a_pack_s );
// Pack A and scale by alpha (if instructed).
bli_packm_int( alpha,
a, a_pack,
cntl_sub_packm_a( cntl ) );
bli_gemm_a_barrier( cntl->thread_info );
if( bli_gemm_am_b_master( cntl->thread_info )) {
bli_obj_init_pack( &b1_pack_s );
}
b1_pack = bli_gemm_broadcast_b( cntl->thread_info, &b1_pack_s );
if( bli_gemm_am_c_master( cntl->thread_info )) {
bli_obj_init_pack( &c1_pack_s );
// Scale C by beta (if instructed).
bli_scalm_int( beta,
c,
cntl_sub_scalm( cntl ) );
}
c1_pack = bli_gemm_broadcast_c( cntl->thread_info, &c1_pack_s );
// Query dimension in partitioning direction.
n_trans = bli_obj_width_after_trans( *b );
dim_t n_pt = n_trans / num_groups;
n_pt = (n_pt * num_groups < n_trans) ? n_pt + 1 : n_pt;
n_pt = (n_pt % 8 == 0) ? n_pt : n_pt + 8 - (n_pt % 8);
dim_t start = group_id * n_pt;
dim_t end = bli_min( start + n_pt, n_trans );
// Partition along the n dimension.
for ( i = start; i < end; i += b_alg )
{
// Determine the current algorithmic blocksize.
// NOTE: Use of b (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, end, b,
cntl_blocksize( cntl ) );
// Acquire partitions for C1
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Acquire partitions for B1
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, b, &b1 );
if( bli_gemm_am_b_master( cntl->thread_info )) {
// Initialize objects for packing B1
bli_packm_init( &b1, &b1_pack_s,
cntl_sub_packm_b( cntl ) );
}
if( bli_gemm_am_c_master( cntl->thread_info )) {
// Initialize objects for packing C1
bli_packm_init( &c1, &c1_pack_s,
cntl_sub_packm_c( cntl ) );
}
bli_gemm_b_barrier( cntl->thread_info );
bli_gemm_c_barrier( cntl->thread_info );
// Pack B1 and scale by alpha (if instructed).
bli_packm_int( alpha,
&b1, b1_pack,
cntl_sub_packm_b( cntl ) );
// Pack C1 and scale by beta (if instructed).
bli_packm_int( beta,
&c1, c1_pack,
cntl_sub_packm_c( cntl ) );
// Packing must be done before computation
bli_gemm_b_barrier( cntl->thread_info );
bli_gemm_c_barrier( cntl->thread_info );
// Perform gemm subproblem.
bli_gemm_int( alpha,
a_pack,
b1_pack,
beta,
c1_pack,
cntl_sub_gemm( cntl ) );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_gemm_a_barrier( cntl->thread_info );
if( bli_gemm_am_a_master( cntl->thread_info ))
bli_obj_release_pack( &a_pack_s );
bli_gemm_b_barrier( cntl->thread_info );
if( bli_gemm_am_b_master( cntl->thread_info )) {
bli_obj_release_pack( &b1_pack_s );
}
bli_gemm_c_barrier( cntl->thread_info );
if( bli_gemm_am_c_master( cntl->thread_info )) {
bli_obj_release_pack( &c1_pack_s );
}
}
void bli_trmm_blk_var3b( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl,
trmm_thrinfo_t* thread )
{
obj_t c_pack_s;
obj_t a1_pack_s, b1_pack_s;
obj_t a1, b1;
obj_t* a1_pack = NULL;
obj_t* b1_pack = NULL;
obj_t* c_pack = NULL;
dim_t i;
dim_t b_alg;
dim_t k_trans;
if( thread_am_ochief( thread ) ){
// Initialize object for packing C
bli_obj_init_pack( &c_pack_s );
bli_packm_init( c, &c_pack_s,
cntl_sub_packm_c( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
}
c_pack = thread_obroadcast( thread, &c_pack_s );
// Initialize pack objects for A and B that are passed into packm_init().
if( thread_am_ichief( thread ) ){
bli_obj_init_pack( &a1_pack_s );
bli_obj_init_pack( &b1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
b1_pack = thread_ibroadcast( thread, &b1_pack_s );
// Pack C (if instructed).
bli_packm_int( c, c_pack,
cntl_sub_packm_c( cntl ),
trmm_thread_sub_opackm( thread ) );
// Query dimension in partitioning direction.
k_trans = bli_obj_width_after_trans( *a );
// Partition along the k dimension.
for ( i = 0; i < k_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
// NOTE: We call a trmm-specific function to determine the kc
// blocksize so that we can implement the "nudging" of kc to be
// a multiple of mr or nr, as needed.
b_alg = bli_trmm_determine_kc_b( i, k_trans, a, b,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and B1.
bli_acquire_mpart_r2l( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_b2t( BLIS_SUBPART1,
i, b_alg, b, &b1 );
// Initialize objects for packing A1 and B1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &b1, b1_pack,
cntl_sub_packm_b( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Pack B1 (if instructed).
bli_packm_int( &b1, b1_pack,
cntl_sub_packm_b( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Perform trmm subproblem.
bli_trmm_int( &BLIS_ONE,
a1_pack,
b1_pack,
&BLIS_ONE,
c_pack,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
thread_ibarrier( thread );
}
thread_obarrier( thread );
// Unpack C (if C was packed).
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ),
trmm_thread_sub_opackm( thread ) );
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
if( thread_am_ochief( thread ) ){
bli_packm_release( c_pack, cntl_sub_packm_c( cntl ) );
}
if( thread_am_ichief( thread ) ){
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
bli_packm_release( b1_pack, cntl_sub_packm_b( cntl ) );
}
}
void bli_gemm_blk_var4( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl )
{
obj_t a1, a1_pack;
obj_t b_pack;
obj_t c1, c1_pack;
dim_t i;
dim_t bm_alg;
dim_t m_trans;
// Initialize all pack objects that are passed into packm_init().
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &b_pack );
bli_obj_init_pack( &c1_pack );
// Query dimension in partitioning direction.
m_trans = bli_obj_length_after_trans( *a );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
// Initialize object for packing B.
bli_packm_init( b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Fuse the first iteration with incremental packing and computation.
{
obj_t b_inc, b_pack_inc;
obj_t c1_pack_inc;
dim_t j;
dim_t bn_inc;
dim_t n_trans;
// Query dimension in partitioning direction.
n_trans = bli_obj_width( b_pack );
// Determine the current algorithmic blocksize.
bm_alg = bli_determine_blocksize_f( 0, m_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
0, bm_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
0, bm_alg, c, &c1 );
// Initialize objects for packing A1 and C1.
bli_packm_init( &a1, &a1_pack, cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, &c1_pack, cntl_sub_packm_c( cntl ) );
// Pack A1 (if instructed).
bli_packm_int( &a1, &a1_pack, cntl_sub_packm_a( cntl ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, &c1_pack, cntl_sub_packm_c( cntl ) );
// Partition along the n dimension.
for ( j = 0; j < n_trans; j += bn_inc )
{
// Determine the current incremental packing blocksize.
bn_inc = bli_determine_blocksize_f( j, n_trans, b,
cntl_blocksize_aux( cntl ) );
// Acquire partitions.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, b, &b_inc );
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, &b_pack, &b_pack_inc );
bli_acquire_mpart_l2r( BLIS_SUBPART1,
j, bn_inc, &c1_pack, &c1_pack_inc );
// Pack B1 (if instructed).
bli_packm_int( &b_inc, &b_pack_inc, cntl_sub_packm_b( cntl ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
&a1_pack,
&b_pack_inc,
&BLIS_ONE,
&c1_pack_inc,
cntl_sub_gemm( cntl ) );
}
// Unpack C1 (if C1 was packed).
bli_unpackm_int( &c1_pack, &c1, cntl_sub_unpackm_c( cntl ) );
}
// Partition along the remaining portion of the m dimension.
for ( i = bm_alg; i < m_trans; i += bm_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.
bm_alg = bli_determine_blocksize_f( i, m_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, bm_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, bm_alg, c, &c1 );
// Initialize objects for packing A1 and C1.
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, &c1_pack,
cntl_sub_packm_c( cntl ) );
// Pack A1 (if instructed).
bli_packm_int( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, &c1_pack,
cntl_sub_packm_c( cntl ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
&a1_pack,
&b_pack,
&BLIS_ONE,
&c1_pack,
cntl_sub_gemm( cntl ) );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( &c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_obj_release_pack( &a1_pack );
bli_obj_release_pack( &b_pack );
bli_obj_release_pack( &c1_pack );
}
void bli_herk_blk_var3f( obj_t* a,
obj_t* ah,
obj_t* c,
gemm_t* cntl,
herk_thrinfo_t* thread )
{
obj_t c_pack_s;
obj_t a1_pack_s, ah1_pack_s;
obj_t a1, ah1;
obj_t* a1_pack = NULL;
obj_t* ah1_pack = NULL;
obj_t* c_pack = NULL;
dim_t i;
dim_t b_alg;
dim_t k_trans;
// Prune any zero region that exists along the partitioning dimension.
bli_herk_prune_unref_mparts_k( a, ah, c );
if( thread_am_ochief( thread ) ) {
// Initialize object for packing C.
bli_obj_init_pack( &c_pack_s );
bli_packm_init( c, &c_pack_s,
cntl_sub_packm_c( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
}
c_pack = thread_obroadcast( thread, &c_pack_s );
// Initialize all pack objects that are passed into packm_init().
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
bli_obj_init_pack( &ah1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
ah1_pack = thread_ibroadcast( thread, &ah1_pack_s );
// Pack C (if instructed).
bli_packm_int( c, c_pack,
cntl_sub_packm_c( cntl ),
herk_thread_sub_opackm( thread ) );
// Query dimension in partitioning direction.
k_trans = bli_obj_width_after_trans( *a );
// Partition along the k dimension.
for ( i = 0; i < k_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, k_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and A1'.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, ah, &ah1 );
// Initialize objects for packing A1 and A1'.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &ah1, ah1_pack,
cntl_sub_packm_b( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
herk_thread_sub_ipackm( thread ) );
// Pack B1 (if instructed).
bli_packm_int( &ah1, ah1_pack,
cntl_sub_packm_b( cntl ),
herk_thread_sub_ipackm( thread ) );
// Perform herk subproblem.
bli_herk_int( &BLIS_ONE,
a1_pack,
ah1_pack,
&BLIS_ONE,
c_pack,
cntl_sub_gemm( cntl ),
herk_thread_sub_herk( thread ) );
// This variant executes multiple rank-k updates. Therefore, if the
// internal beta scalar on matrix C is non-zero, we must use it
// only for the first iteration (and then BLIS_ONE for all others).
// And since c_pack is a local obj_t, we can simply overwrite the
// internal beta scalar with BLIS_ONE once it has been used in the
// first iteration.
thread_ibarrier( thread );
if ( i == 0 && thread_am_ichief( thread ) ) bli_obj_scalar_reset( c_pack );
}
thread_obarrier( thread );
// Unpack C (if C was packed).
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ),
herk_thread_sub_opackm( thread ) );
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
if( thread_am_ochief( thread ) ) {
bli_packm_release( c_pack, cntl_sub_packm_c( cntl ) );
}
if( thread_am_ichief( thread ) ) {
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
bli_packm_release( ah1_pack, cntl_sub_packm_b( cntl ) );
}
}
void bli_trmm_blk_var2b( obj_t* a,
obj_t* b,
obj_t* c,
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;
dim_t n_trans;
if( thread_am_ochief( thread ) ) {
// Initialize object for packing A
bli_obj_init_pack( &a_pack_s );
bli_packm_init( a, &a_pack_s,
cntl_sub_packm_a( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
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,
cntl_sub_packm_a( cntl ),
trmm_thread_sub_opackm( thread ) );
// Query dimension in partitioning direction.
n_trans = bli_obj_width_after_trans( *b );
dim_t start, end;
bli_get_range_weighted( thread, 0, n_trans,
bli_determine_reg_blocksize( b, cntl_blocksize( cntl ) ),
bli_obj_is_upper( *c ), &start, &end );
// Partition along the n dimension.
for ( i = start; i < end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_b( i, end, b,
cntl_blocksize( cntl ) );
// 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,
cntl_sub_packm_b( cntl ) );
bli_packm_init( &c1, c1_pack,
cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack B1 (if instructed).
bli_packm_int( &b1, b1_pack,
cntl_sub_packm_b( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
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,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
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_obj_release_pack( a_pack );
if( thread_am_ichief( thread ) ) {
bli_obj_release_pack( b1_pack );
bli_obj_release_pack( c1_pack );
}
}
void bli_herk_blk_var2f( obj_t* a,
obj_t* ah,
obj_t* c,
gemm_t* cntl,
herk_thrinfo_t* thread )
{
obj_t a_pack_s;
obj_t ah1_pack_s, c1S_pack_s;
obj_t ah1, c1, c1S;
obj_t aS_pack;
obj_t* a_pack;
obj_t* ah1_pack;
obj_t* c1S_pack;
dim_t i;
dim_t b_alg;
dim_t n_trans;
subpart_t stored_part;
// The upper and lower variants are identical, except for which
// merged subpartition is acquired in the loop body.
if ( bli_obj_is_lower( *c ) ) stored_part = BLIS_SUBPART1B;
else stored_part = BLIS_SUBPART1T;
if( thread_am_ochief( thread ) ) {
// Initialize object for packing A
bli_obj_init_pack( &a_pack_s );
bli_packm_init( a, &a_pack_s,
cntl_sub_packm_a( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
}
a_pack = thread_obroadcast( thread, &a_pack_s );
// Initialize pack objects for C and A' that are passed into packm_init().
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &ah1_pack_s );
bli_obj_init_pack( &c1S_pack_s );
}
ah1_pack = thread_ibroadcast( thread, &ah1_pack_s );
c1S_pack = thread_ibroadcast( thread, &c1S_pack_s );
// Pack A (if instructed).
bli_packm_int( a, a_pack,
cntl_sub_packm_a( cntl ),
herk_thread_sub_opackm( thread ) );
// Query dimension in partitioning direction.
n_trans = bli_obj_width_after_trans( *c );
dim_t start, end;
// Needs to be replaced with a weighted range because triangle
bli_get_range_weighted( thread, 0, n_trans,
bli_blksz_get_mult_for_obj( a, cntl_blocksize( cntl ) ),
bli_obj_is_lower( *c ), &start, &end );
// Partition along the n dimension.
for ( i = start; i < end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, end, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1' and C1.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, ah, &ah1 );
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Partition off the stored region of C1 and the corresponding region
// of A_pack.
bli_acquire_mpart_t2b( stored_part,
i, b_alg, &c1, &c1S );
bli_acquire_mpart_t2b( stored_part,
i, b_alg, a_pack, &aS_pack );
// Initialize objects for packing A1' and C1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &ah1, ah1_pack,
cntl_sub_packm_b( cntl ) );
bli_packm_init( &c1S, c1S_pack,
cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread ) ;
// Pack A1' (if instructed).
bli_packm_int( &ah1, ah1_pack,
cntl_sub_packm_b( cntl ),
herk_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1S, c1S_pack,
cntl_sub_packm_c( cntl ),
herk_thread_sub_ipackm( thread ) ) ;
// Perform herk subproblem.
bli_herk_int( &BLIS_ONE,
&aS_pack,
ah1_pack,
&BLIS_ONE,
c1S_pack,
cntl_sub_gemm( cntl ),
herk_thread_sub_herk( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1S_pack, &c1S,
cntl_sub_unpackm_c( cntl ),
herk_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( ah1_pack, cntl_sub_packm_b( cntl ) );
bli_packm_release( c1S_pack, cntl_sub_packm_c( cntl ) );
}
}
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_gemm_blk_var3f( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl )
{
obj_t a1, a1_pack;
obj_t b1, b1_pack;
obj_t c_pack;
dim_t i;
dim_t b_alg;
dim_t k_trans;
// Initialize all pack objects that are passed into packm_init().
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &b1_pack );
bli_obj_init_pack( &c_pack );
// Query dimension in partitioning direction.
k_trans = bli_obj_width_after_trans( *a );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
// Initialize object for packing C.
bli_packm_init( c, &c_pack,
cntl_sub_packm_c( cntl ) );
// Pack C (if instructed).
bli_packm_int( c, &c_pack,
cntl_sub_packm_c( cntl ) );
// Partition along the k dimension.
for ( i = 0; i < k_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
// NOTE: Use of b (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, k_trans, b,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and B1.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, b, &b1 );
// Initialize objects for packing A1 and B1.
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &b1, &b1_pack,
cntl_sub_packm_b( cntl ) );
// Pack A1 (if instructed).
bli_packm_int( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Pack B1 (if instructed).
bli_packm_int( &b1, &b1_pack,
cntl_sub_packm_b( cntl ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
&a1_pack,
&b1_pack,
&BLIS_ONE,
&c_pack,
cntl_sub_gemm( cntl ) );
// This variant executes multiple rank-k updates. Therefore, if the
// internal beta scalar on matrix C is non-zero, we must use it
// only for the first iteration (and then BLIS_ONE for all others).
// And since c_pack is a local obj_t, we can simply overwrite the
// internal beta scalar with BLIS_ONE once it has been used in the
// first iteration.
if ( i == 0 ) bli_obj_scalar_reset( &c_pack );
}
// Unpack C (if C was packed).
bli_unpackm_int( &c_pack, c,
cntl_sub_unpackm_c( cntl ) );
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_obj_release_pack( &a1_pack );
bli_obj_release_pack( &b1_pack );
bli_obj_release_pack( &c_pack );
}
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 ) );
}
}