err_t bli_check_matrix_strides( dim_t m, dim_t n, inc_t rs, inc_t cs, inc_t is )
{
err_t e_val = BLIS_SUCCESS;
// Note: A lot of thought went into designing these checks. Do NOT change
// them unless you absolutely know what you are doing! Particularly, do
// not try to merge the general and row-/column-major sections. It might
// be possible, but it would be a lot less readable.
// Prohibit negative dimensions.
if ( m < 0 || n < 0 )
return BLIS_NEGATIVE_DIMENSION;
// Overwrite rs and cs with the absolute value of each. We can do this
// since the checks below are not dependent on the sign of the strides.
rs = bli_abs( rs );
cs = bli_abs( cs );
is = bli_abs( is );
// The default case (whereby we interpret rs == cs == 0 as a request for
// column-major order) is handled prior to calling this function, so the
// only time we should see zero strides here is if the matrix is empty.
if ( m == 0 || n == 0 ) return e_val;
// Disallow row, column, or imaginary strides of zero.
if ( ( rs == 0 || cs == 0 || is == 0 ) )
return BLIS_INVALID_DIM_STRIDE_COMBINATION;
// Check stride consistency in cases of general stride.
if ( rs != 1 && cs != 1 )
{
// We apply different tests depending on which way the strides
// "tilt".
if ( rs == cs )
{
// If rs == cs, then we must be dealing with an m-by-1 or a
// 1-by-n matrix and thus at least one of the dimensions, m
// or n, must be unit (even if the other is zero).
if ( m != 1 && n != 1 )
return BLIS_INVALID_DIM_STRIDE_COMBINATION;
}
else if ( rs < cs )
{
// For column-major tilt, cs must be equal or larger than m * rs.
if ( m * rs > cs )
return BLIS_INVALID_DIM_STRIDE_COMBINATION;
}
else if ( cs < rs )
{
// For row-major tilt, rs must be equal or larger than n * cs.
if ( n * cs > rs )
return BLIS_INVALID_DIM_STRIDE_COMBINATION;
}
}
else // check stride consistency of row-/column-storage cases.
{
if ( rs == 1 && cs == 1 )
{
// If rs == cs == 1, then we must be dealing with an m-by-1, a
// 1-by-n, or a 1-by-1 matrix and thus at least one of the
// dimensions, m or n, must be unit (even if the other is zero).
if ( m != 1 && n != 1 )
return BLIS_INVALID_DIM_STRIDE_COMBINATION;
}
else if ( rs == 1 )
{
// For column-major storage, don't allow the column stride to be
// less than the m dimension.
if ( cs < m )
return BLIS_INVALID_COL_STRIDE;
}
else if ( cs == 1 )
{
// For row-major storage, don't allow the row stride to be less
// than the n dimension.
if ( rs < n )
return BLIS_INVALID_ROW_STRIDE;
}
}
return e_val;
}
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_trsm_blk_var1f( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl,
trsm_thrinfo_t* thread )
{
obj_t b_pack_s;
obj_t a1_pack_s;
obj_t a1, c1;
obj_t* b_pack = NULL;
obj_t* a1_pack = NULL;
dim_t i;
dim_t b_alg;
dim_t m_trans;
dim_t offA;
// Initialize object for packing B.
if( thread_am_ochief( thread ) ) {
bli_obj_init_pack( &b_pack_s );
bli_packm_init( b, &b_pack_s,
cntl_sub_packm_b( cntl ) );
}
b_pack = thread_obroadcast( thread, &b_pack_s );
// Initialize object for packing B.
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
// Pack B1 (if instructed).
bli_packm_int( b, b_pack,
cntl_sub_packm_b( cntl ),
trsm_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 ( bli_obj_is_lower( *a ) )
offA = bli_abs( bli_obj_diag_offset_after_trans( *a ) );
dim_t start, end;
num_t dt = bli_obj_execution_datatype( *a );
bli_get_range_t2b( thread, offA, m_trans,
//bli_lcm( bli_info_get_default_nr( BLIS_TRSM, dt ), bli_info_get_default_mr( BLIS_TRSM, dt ) ),
bli_info_get_default_mc( BLIS_TRSM, dt ),
&start, &end );
// Partition along the remaining portion of the m 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_t2b( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize object for packing A1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
trsm_thread_sub_ipackm( thread ) );
// Perform trsm subproblem.
bli_trsm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
&c1,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
thread_ibarrier( 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 ) );
}
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_trsm_blk_var1f( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl )
{
obj_t a1, a1_pack;
obj_t b_pack;
obj_t c1;
dim_t i;
dim_t b_alg;
dim_t m_trans;
dim_t offA;
// Initialize all pack objects that are passed into packm_init().
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &b_pack );
// 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 ( bli_obj_is_lower( *a ) )
offA = 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 ) );
// Pack B1 (if instructed).
bli_packm_int( b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Partition along the remaining portion of the m dimension.
for ( i = offA; i < m_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_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, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize object for packing A1.
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Pack A1 (if instructed).
bli_packm_int( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Perform trsm subproblem.
bli_trsm_int( &BLIS_ONE,
&a1_pack,
&b_pack,
&BLIS_ONE,
&c1,
cntl_sub_trsm( 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_var1( 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 b_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 );
// If A is [upper] triangular, use the diagonal offset of A to determine
// the length of the non-zero region.
if ( bli_obj_is_triangular( *a ) )
mT_trans = bli_abs( bli_obj_diag_offset_after_trans( *a ) ) +
bli_obj_width_after_trans( *a );
else // if ( bli_obj_is_general( *a )
mT_trans = bli_obj_length_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 ) );
// Pack B and scale by alpha (if instructed).
bli_packm_int( alpha,
b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Partition along the m dimension.
for ( i = 0; i < mT_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_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, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_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.
bli_trmm_int( BLIS_LEFT,
alpha,
&a1_pack,
&b_pack,
beta,
&c1_pack,
cntl_sub_trmm( 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_obj_alloc_buffer( inc_t rs,
inc_t cs,
inc_t is,
obj_t* obj )
{
dim_t n_elem = 0;
dim_t m, n;
siz_t elem_size;
siz_t buffer_size;
void* p;
// Query the dimensions of the object we are allocating.
m = bli_obj_length( *obj );
n = bli_obj_width( *obj );
// Query the size of one element.
elem_size = bli_obj_elem_size( *obj );
// Adjust the strides, if needed, before doing anything else
// (particularly, before doing any error checking).
bli_adjust_strides( m, n, elem_size, &rs, &cs, &is );
if ( bli_error_checking_is_enabled() )
bli_obj_alloc_buffer_check( rs, cs, is, obj );
// Determine how much object to allocate.
if ( m == 0 || n == 0 )
{
// For empty objects, set n_elem to zero. Row and column strides
// should remain unchanged (because alignment is not needed).
n_elem = 0;
}
else
{
// The number of elements to allocate is given by the distance from
// the element with the lowest address (usually {0, 0}) to the element
// with the highest address (usually {m-1, n-1}), plus one for the
// highest element itself.
n_elem = (m-1) * bli_abs( rs ) + (n-1) * bli_abs( cs ) + 1;
}
// Handle the special case where imaginary stride is larger than
// normal.
if ( bli_obj_is_complex( *obj ) )
{
// Notice that adding is/2 works regardless of whether the
// imaginary stride is unit, something between unit and
// 2*n_elem, or something bigger than 2*n_elem.
n_elem = bli_abs( is ) / 2 + n_elem;
}
// Compute the size of the total buffer to be allocated, which includes
// padding if the leading dimension was increased for alignment purposes.
buffer_size = ( siz_t )n_elem * elem_size;
// Allocate the buffer.
p = bli_malloc_user( buffer_size );
// Set individual fields.
bli_obj_set_buffer( p, *obj );
bli_obj_set_strides( rs, cs, *obj );
bli_obj_set_imag_stride( is, *obj );
}
void bli_trmm_blk_var1( 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 b_alg;
dim_t m_trans;
dim_t offA;
// 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 );
// 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 );
// 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 ) );
// Pack B and scale by alpha (if instructed).
bli_packm_int( alpha,
b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Partition along the m dimension.
for ( i = offA; i < m_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_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, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_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.
bli_trmm_int( alpha,
&a1_pack,
&b_pack,
beta,
&c1_pack,
cntl_sub_trmm( 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 );
}
