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_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_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_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 ); }