void bli_trsm_blk_var2b( obj_t* a, obj_t* b, obj_t* c, trsm_t* cntl, trsm_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; // Initialize pack objects for A that are passed into packm_init(). if( thread_am_ochief( thread ) ) { bli_obj_init_pack( &a_pack_s ); // Initialize object for packing A. 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; num_t dt = bli_obj_execution_datatype( *a ); bli_get_range_r2l( thread, 0, n_trans, //bli_lcm( bli_info_get_default_nr( BLIS_TRSM, dt ), // bli_info_get_default_mr( BLIS_TRSM, dt ) ), bli_lcm( bli_blksz_get_nr( dt, cntl_blocksize( cntl ) ), bli_blksz_get_mr( dt, cntl_blocksize( cntl ) ) ), &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 ), trsm_thread_sub_ipackm( thread ) ); // Pack C1 (if instructed). bli_packm_int( &c1, c1_pack, cntl_sub_packm_c( cntl ), trsm_thread_sub_ipackm( thread ) ); // Perform trsm subproblem. bli_trsm_int( &BLIS_ONE, a_pack, b1_pack, &BLIS_ONE, c1_pack, cntl_sub_trsm( cntl ), trsm_thread_sub_trsm( thread ) ); thread_ibarrier( thread ); // Unpack C1 (if C1 was packed). bli_unpackm_int( c1_pack, &c1, cntl_sub_unpackm_c( cntl ), trsm_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 ) ); } }
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 ) ); } }
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_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 ) ); } }