void bli_gemm_blk_var1f( obj_t* a, obj_t* b, obj_t* c, 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; 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( thread, 0, m_trans, bli_determine_reg_blocksize( 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, 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 ), 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 ) ); // 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_obj_release_pack( b_pack ); if( thread_am_ichief( thread ) ){ bli_obj_release_pack( a1_pack ); bli_obj_release_pack( c1_pack ); } }
void bli_trsm_blk_var2f( 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 datatype = bli_obj_execution_datatype( *a ); bli_get_range( thread, 0, n_trans, bli_lcm( bli_info_get_default_nr( datatype ), bli_info_get_default_mr( datatype ) ), &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, b, cntl_blocksize( cntl ) ); // Acquire partitions for B1 and C1. bli_acquire_mpart_l2r( BLIS_SUBPART1, i, b_alg, b, &b1 ); bli_acquire_mpart_l2r( 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 ) ); // 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_obj_release_pack( a_pack ); if( thread_am_ichief( thread ) ) { bli_obj_release_pack( b1_pack ); bli_obj_release_pack( c1_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 datatype = bli_obj_execution_datatype( *a ); bli_get_range( thread, offA, m_trans, //bli_lcm( bli_info_get_default_nr( datatype ), bli_info_get_default_mr( datatype ) ), bli_info_get_default_mc( datatype ), &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 ) ); } // 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( b_pack ); if( thread_am_ichief( thread ) ) bli_obj_release_pack( a1_pack ); }