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_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_var2f( 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; // 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, 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 ) ); dim_t my_start, my_end; bli_get_range_weighted_l2r( thread, b, bli_blksz_get_mult_for_obj( b, cntl_blocksize( cntl ) ), &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_f( i, my_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 ), 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 ) ); 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( 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_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_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 ) ); } }