def write_stra_matC(myfile, coeff_idx, coeffs, idx, dim_name, dims, level): mat_name = (getName(coeff_idx))[0] nonzero_coeffs = [coeff for coeff in coeffs if is_nonzero(coeff)] #add = 'std::array<unsigned, ' + str(len(nonzero_coeffs)) + '> ' + mat_name + str(idx) + '_subid = {' #add += ', '.join(['%s' % getActualBlockIndex( coeff_idx, i, dims, level ) \ # for i, c in enumerate(coeffs) if is_nonzero(c)]) #add += '};' #write_line(myfile, 1, add) add = 'std::array<T*,' + str( len(nonzero_coeffs)) + '> ' + mat_name + str(idx) + '_list = {' #add += ', '.join( [ str(c) for i, c in enumerate(coeffs) if is_nonzero(c) ] ) add += ', '.join([ 'const_cast<T*>(%s)' % (getBlockName(coeff_idx, i, dims, level)) for i, c in enumerate(coeffs) if is_nonzero(c) ]) add += '};' write_line(myfile, 1, add) add = 'std::array<T,' + str( len(nonzero_coeffs)) + '> ' + mat_name + str(idx) + '_coeff_list = {' add += ', '.join([str(c) for i, c in enumerate(coeffs) if is_nonzero(c)]) add += '};' write_line(myfile, 1, add)
def write_stra_matAB(myfile, coeff_idx, coeffs, idx, dim_name, dims, level): mat_name = (getName(coeff_idx))[0] nonzero_coeffs = [coeff for coeff in coeffs if is_nonzero(coeff)] add = 'std::array<unsigned, ' + str( len(nonzero_coeffs)) + '> ' + mat_name + str(idx) + '_subid = {' add += ', '.join(['%s' % getActualBlockIndex( coeff_idx, i, dims, level ) \ for i, c in enumerate(coeffs) if is_nonzero(c)]) add += '};' write_line(myfile, 1, add) add = 'std::array<T,' + str( len(nonzero_coeffs)) + '> ' + mat_name + str(idx) + '_coeff_list = {' add += ', '.join([str(c) for i, c in enumerate(coeffs) if is_nonzero(c)]) add += '};' write_line(myfile, 1, add) add = 'stra_tensor_view<T,' + str(len(nonzero_coeffs)) + '> ' add += mat_name + 'v' + str(idx) add += '(my_len_' + dim_name[0] + ', ' add += 'my_len_' + dim_name[1] + ', ' add += mat_name + '_divisor, const_cast<T*>(' + mat_name + '), ' add += mat_name + str(idx) + '_subid, ' + mat_name + str( idx) + '_coeff_list, ' add += 'my_stride_' + mat_name + '_' + dim_name[0] + ',' add += ' my_stride_' + mat_name + '_' + dim_name[1] add += ');' write_line(myfile, 1, add)
def write_straprim_caller(myfile, index, a_coeffs, b_coeffs, c_coeffs, dims, num_multiplies, level=1): comment = '// M%d = (' % (index) comment += ' + '.join([str(c) + ' * %s' % getBlockName( 0, i, dims, level ) \ for i, c in enumerate(a_coeffs) if is_nonzero(c)]) comment += ') * (' comment += ' + '.join([str(c) + ' * %s' % getBlockName( 1, i, dims, level ) \ for i, c in enumerate(b_coeffs) if is_nonzero(c)]) comment += '); ' comment += '; '.join([ ' %s += %s * M%d' % (getBlockName(2, i, dims, level), c, index) for i, c in enumerate(c_coeffs) if is_nonzero(c) ]) comment += ';' write_line(myfile, 1, comment) write_stra_mat(myfile, 0, a_coeffs, index, 'ms, ks', dims, level) write_stra_mat(myfile, 1, b_coeffs, index, 'ks, ns', dims, level) write_stra_mat(myfile, 2, c_coeffs, index, 'ms, ns', dims, level) #add = 'stra_gemm(comm, cfg, alpha, Av{0}, Bv{0}, beta, Cv{0});'.format( index ) #add = 'straprim_naive(comm, cfg, alpha, Av{0}, Bv{0}, beta, Cv{0});'.format( index ) add = 'straprim_ab(comm, cfg, alpha, Av{0}, Bv{0}, beta, Cv{0});'.format( index) write_line(myfile, 1, add) write_line(myfile, 1, 'comm.barrier();') write_break(myfile)
def write_straprim_caller(myfile, index, a_coeffs, b_coeffs, c_coeffs, dims, num_multiplies, level=1): comment = '// M%d = (' % (index) comment += ' + '.join([str(c) + ' * %s' % getBlockName( 0, i, dims, level ) \ for i, c in enumerate(a_coeffs) if is_nonzero(c)]) comment += ') * (' comment += ' + '.join([str(c) + ' * %s' % getBlockName( 1, i, dims, level ) \ for i, c in enumerate(b_coeffs) if is_nonzero(c)]) comment += '); ' comment += '; '.join([ ' %s += %s * M%d' % (getBlockName(2, i, dims, level), c, index) for i, c in enumerate(c_coeffs) if is_nonzero(c) ]) comment += ';' write_line(myfile, 1, comment) write_stra_mat(myfile, 0, a_coeffs, index, ['AC', 'AB'], dims, level) write_stra_mat(myfile, 1, b_coeffs, index, ['AB', 'BC'], dims, level) write_stra_mat(myfile, 2, c_coeffs, index, ['AC', 'BC'], dims, level) myfile.write( \ '''\ if (Cv{0}.stride(!row_major) == 1) {{ Av{0}.transpose(); Bv{0}.transpose(); Cv{0}.transpose(); stra_gemm(comm, cfg, alpha, Bv{0}, Av{0}, beta, Cv{0}); }} else {{ stra_gemm(comm, cfg, alpha, Av{0}, Bv{0}, beta, Cv{0}); }} '''.format( index ) ) #Av{0}.swap(Bv{0}); #add = 'stra_gemm(comm, cfg, alpha, Av{0}, Bv{0}, beta, Cv{0});'.format( index ) #write_line( myfile, 1, add ) write_line(myfile, 1, 'comm.barrier();') write_line( myfile, 1, '//std::cout << "stra_internal/stra_mult_M{0}:" << std::endl;'.format( index)) write_line(myfile, 1, '//print_tensor_matrix( ct );') write_break(myfile)
def write_straprim_caller(myfile, index, a_coeffs, b_coeffs, c_coeffs, dims, num_multiplies, level=1): comment = '// M%d = (' % (index) comment += ' + '.join([str(c) + ' * %s' % getBlockName( 0, i, dims, level ) \ for i, c in enumerate(a_coeffs) if is_nonzero(c)]) comment += ') * (' comment += ' + '.join([str(c) + ' * %s' % getBlockName( 1, i, dims, level ) \ for i, c in enumerate(b_coeffs) if is_nonzero(c)]) comment += '); ' comment += '; '.join([' %s += %s * M%d' % ( getBlockName( 2, i, dims, level ), c, index ) for i, c in enumerate(c_coeffs) if is_nonzero(c)]) comment += ';' write_line(myfile, 1, comment) write_stra_mat( myfile, 0, a_coeffs, index, ['AC', 'AB'], dims, level ) write_stra_mat( myfile, 1, b_coeffs, index, ['AB', 'BC'], dims, level ) write_stra_mat( myfile, 2, c_coeffs, index, ['AC', 'BC'], dims, level ) myfile.write( \ '''\ //if (ct.stride(!row_major) == 1) //{{ // Av{0}.transpose(); // Bv{0}.transpose(); // Cv{0}.transpose(); // straprim_naive<T,{1},{2},{3}>(comm, cfg, my_sub_len_AB, my_sub_len_AC, my_sub_len_BC, // alpha, // B{0}_list, B{0}_coeff_list, my_stride_B_AB, my_stride_B_BC, // A{0}_list, A{0}_coeff_list, my_stride_A_AB, my_stride_A_AC, // beta, // C{0}_list, C{0}_coeff_list, my_stride_C_AC, my_stride_C_BC); //}} else {{ straprim_naive<T,{1},{2},{3}>(comm, cfg, my_sub_len_AB, my_sub_len_AC, my_sub_len_BC, alpha, A{0}_list, A{0}_coeff_list, my_stride_A_AB, my_stride_A_AC, B{0}_list, B{0}_coeff_list, my_stride_B_AB, my_stride_B_BC, beta, C{0}_list, C{0}_coeff_list, my_stride_C_AC, my_stride_C_BC); //}} '''.format( index, getNNZ(a_coeffs), getNNZ(b_coeffs), getNNZ(c_coeffs) ) ) #Av{0}.swap(Bv{0}); #add = 'stra_gemm(comm, cfg, alpha, Av{0}, Bv{0}, beta, Cv{0});'.format( index ) #write_line( myfile, 1, add ) write_line( myfile, 1, 'comm.barrier();' ) write_line( myfile, 1, '//std::cout << "stra_internal/stra_mult_M{0}:" << std::endl;'.format( index ) ) write_line( myfile, 1, '//print_tensor_matrix( ct );' ) write_break( myfile )
def write_stra_mat(myfile, coeff_idx, coeffs, idx, dim_name, dims, level): mat_name = (getName(coeff_idx))[0] nonzero_coeffs = [coeff for coeff in coeffs if is_nonzero(coeff)] add = 'stra_matrix_view<T,' + str(len(nonzero_coeffs)) + '> ' add += mat_name + 'v' + str(idx) + '({' + dim_name + '}, {' #add += ', '.join( ['const_cast<T*>(%s)' % (getSubMatName(coeff_idx, i, dims, level) ) for i, c in enumerate(coeffs) if is_nonzero(c)] ) add += ', '.join([ 'const_cast<T*>(%s)' % (getBlockName(coeff_idx, i, dims, level)) for i, c in enumerate(coeffs) if is_nonzero(c) ]) add += '}, {' add += ', '.join([str(c) for i, c in enumerate(coeffs) if is_nonzero(c)]) add += '}, {rs_' + mat_name + ', cs_' + mat_name + '});' write_line(myfile, 1, add)
def create_micro_functions(myfile, coeffs, kernel_header_filename): write_line(myfile, 0, '#include "%s"' % kernel_header_filename) write_break(myfile) abc_micro_kernel_gen.write_common_rankk_macro_assembly(myfile) write_break(myfile) abc_micro_kernel_gen.macro_initialize_assembly(myfile) #write_break( myfile ) #abc_micro_kernel_gen.macro_rankk_xor0_assembly( myfile ) #write_break( myfile ) #abc_micro_kernel_gen.macro_rankk_loopkiter_assembly( myfile ) #write_break( myfile ) #abc_micro_kernel_gen.macro_rankk_loopkleft_assembly( myfile ) #write_break( myfile ) #abc_micro_kernel_gen.macro_rankk_postaccum_assembly( myfile ) write_break(myfile) for i, coeff_set in enumerate(transpose(coeffs[2])): if len(coeff_set) > 0: nonzero_coeffs = [ coeff for coeff in coeff_set if is_nonzero(coeff) ] nnz = len(nonzero_coeffs) if nnz <= 23: abc_micro_kernel_gen.generate_micro_kernel( myfile, nonzero_coeffs, i) write_break(myfile)
def write_straprim_caller(myfile, index, a_coeffs, b_coeffs, c_coeffs, dims, num_multiplies, level=1): comment = '// M%d = (' % (index) comment += ' + '.join([str(c) + ' * %s' % getBlockName( 0, i, dims, level ) \ for i, c in enumerate(a_coeffs) if is_nonzero(c)]) comment += ') * (' comment += ' + '.join([str(c) + ' * %s' % getBlockName( 1, i, dims, level ) \ for i, c in enumerate(b_coeffs) if is_nonzero(c)]) comment += '); ' comment += '; '.join([ ' %s += %s * M%d' % (getBlockName(2, i, dims, level), c, index) for i, c in enumerate(c_coeffs) if is_nonzero(c) ]) comment += ';' write_line(myfile, 1, comment) add = 'bl_dgemm_straprim_abc%d( ms, ns, ks, ' % index add += ', '.join(['%s' % getBlockName( 0, i, dims, level ) \ for i, c in enumerate(a_coeffs) if is_nonzero(c)]) add += ', lda, ' add += ', '.join(['%s' % getBlockName( 1, i, dims, level ) \ for i, c in enumerate(b_coeffs) if is_nonzero(c)]) add += ', ldb, ' add += ', '.join(['%s' % getBlockName( 2, i, dims, level ) \ for i, c in enumerate(c_coeffs) if is_nonzero(c)]) add += ', ldc, packA, packB, bl_ic_nt );' write_line(myfile, 1, add)
def create_kernel_header(myfile, coeffs): #write_line( myfile, 0, '#include "bl_dgemm_kernel.h"' ) write_break(myfile) abc_micro_kernel_gen.write_header_start(myfile) for i, coeff_set in enumerate(transpose(coeffs[2])): if len(coeff_set) > 0: nonzero_coeffs = [ coeff for coeff in coeff_set if is_nonzero(coeff) ] nnz = len(nonzero_coeffs) abc_micro_kernel_gen.generate_kernel_header( myfile, nonzero_coeffs, i) write_break(myfile) abc_micro_kernel_gen.write_header_end(myfile)
def getNNZ ( coeffs ): nonzero_coeffs = [coeff for coeff in coeffs if is_nonzero(coeff)] nnz = len( nonzero_coeffs ) return nnz
def write_macro_func(myfile, coeffs, index, mat_name): ''' Write the add function for a set of coefficients. This is a custom add function used for a single multiply in a single fast algorithm. coeffs is the set of coefficients used for the add ''' nonzero_coeffs = [coeff for coeff in coeffs if is_nonzero(coeff)] nnz = len(nonzero_coeffs) # TODO(arbenson): put in a code-generated comment here add = 'inline void bl_macro_kernel_stra_abc%d( int m, int n, int k, double *packA, double *packB, ' % ( index) add += ', '.join(['double *%s%d' % (mat_name, i) for i in range(nnz)]) add += ', int ld%s ) {' % (mat_name) write_line(myfile, 0, add) write_line(myfile, 1, 'int i, j;') write_line(myfile, 1, 'aux_t aux;') write_line(myfile, 1, 'aux.b_next = packB;') write_line(myfile, 1, 'for ( j = 0; j < n; j += DGEMM_NR ) {') write_line(myfile, 1, ' aux.n = min( n - j, DGEMM_NR );') write_line(myfile, 1, ' for ( i = 0; i < m; i += DGEMM_MR ) {') write_line(myfile, 1, ' aux.m = min( m - i, DGEMM_MR );') write_line(myfile, 1, ' if ( i + DGEMM_MR >= m ) {') write_line(myfile, 1, ' aux.b_next += DGEMM_NR * k;') write_line(myfile, 1, ' }') #NEED to do: c_coeff -> pass in the parameters! #Generate the micro-kernel outside #abc_micro_kernel_gen.generate_kernel_header( my_kernel_header_file, nonzero_coeffs, index ) #abc_micro_kernel_gen.generate_micro_kernel( my_micro_kernel_file, nonzero_coeffs, index ) #generate the function caller #if nnz <= 23 and not contain_nontrivial( nonzero_coeffs ): # add = '( bl_dgemm_micro_kernel_stra_abc%d ) ( k, &packA[ i * k ], &packB[ j * k ], ' % index # add += '(unsigned long long) ld%s, ' % mat_name # add += ', '.join( ['&%s%d[ j * ld%s + i ]' % ( mat_name, i, mat_name ) for i in range( nnz )] ) # add += ', &aux );' # write_line(myfile, 3, add) #else: # write_mulstrassen_kernel_caller( myfile, nonzero_coeffs ) if nnz <= 23: if not contain_nontrivial(nonzero_coeffs): add = '( bl_dgemm_micro_kernel_stra_abc%d ) ( k, &packA[ i * k ], &packB[ j * k ], ' % index add += '(unsigned long long) ld%s, ' % mat_name add += ', '.join([ '&%s%d[ j * ld%s + i ]' % (mat_name, i, mat_name) for i in range(nnz) ]) add += ', &aux );' write_line(myfile, 3, add) else: write_line(myfile, 3, 'double alpha_list[%d];' % nnz) add = '; '.join([ 'alpha_list[%d]= (double)(%s)' % (j, coeff) for j, coeff in enumerate(nonzero_coeffs) ]) add += ';' write_line(myfile, 3, add) add = '( bl_dgemm_micro_kernel_stra_abc%d ) ( k, &packA[ i * k ], &packB[ j * k ], ' % index add += '(unsigned long long) ld%s, ' % mat_name add += ', '.join([ '&%s%d[ j * ld%s + i ]' % (mat_name, i, mat_name) for i in range(nnz) ]) add += ', alpha_list , &aux );' write_line(myfile, 3, add) else: write_mulstrassen_kernel_caller(myfile, nonzero_coeffs) #write_mulstrassen_kernel_caller( myfile, nonzero_coeffs ) write_line(myfile, 2, '}') write_line(myfile, 1, '}') write_line(myfile, 0, '}') # end of function
def write_straprim_abc_function(myfile, index, a_coeffs, b_coeffs, c_coeffs, dims, level): comment = '// M%d = (' % (index) comment += ' + '.join([str(c) + ' * %s' % getBlockName( 0, i, dims, level ) \ for i, c in enumerate(a_coeffs) if is_nonzero(c)]) comment += ') * (' comment += ' + '.join([str(c) + ' * %s' % getBlockName( 1, i, dims, level ) \ for i, c in enumerate(b_coeffs) if is_nonzero(c)]) comment += '); ' comment += '; '.join([ ' %s += %s * M%d' % (getBlockName(2, i, dims, level), c, index) for i, c in enumerate(c_coeffs) if is_nonzero(c) ]) comment += ';' write_line(myfile, 0, comment) add = 'void bl_dgemm_straprim_abc%d( int m, int n, int k, ' % index add += ', '.join( ['double* %s%d' % ('a', i) for i in range(getNNZ(a_coeffs))]) add += ', int lda, ' add += ', '.join( ['double* %s%d' % ('b', i) for i in range(getNNZ(b_coeffs))]) add += ', int ldb, ' add += ', '.join( ['double* %s%d' % ('c', i) for i in range(getNNZ(c_coeffs))]) add += ', int ldc, double *packA, double *packB, int bl_ic_nt ) {' write_line(myfile, 0, add) write_line(myfile, 1, 'int i, j, p, ic, ib, jc, jb, pc, pb;') write_line(myfile, 1, 'for ( jc = 0; jc < n; jc += DGEMM_NC ) {') write_line(myfile, 2, 'jb = min( n - jc, DGEMM_NC );') write_line(myfile, 2, 'for ( pc = 0; pc < k; pc += DGEMM_KC ) {') write_line(myfile, 3, 'pb = min( k - pc, DGEMM_KC );') #write_line( myfile, 0, '#ifdef _PARALLEL_') #write_line( myfile, 3, '#pragma omp parallel for num_threads( bl_ic_nt ) private( j )' ) #write_line( myfile, 0, '#endif') write_line(myfile, 3, '{') write_line(myfile, 4, 'int tid = omp_get_thread_num();') write_line(myfile, 4, 'int my_start;') write_line(myfile, 4, 'int my_end;') write_line(myfile, 4, 'bl_get_range( jb, DGEMM_NR, &my_start, &my_end );') write_line(myfile, 4, 'for ( j = my_start; j < my_end; j += DGEMM_NR ) {') add = 'packB_add_stra_abc%d( min( jb - j, DGEMM_NR ), pb, ' % index add += ', '.join([ '&%s%d[ pc + (jc+j)*ldb ]' % ('b', i) for i in range(getNNZ(b_coeffs)) ]) add += ', ldb, &packB[ j * pb ] );' write_line(myfile, 5, add) write_line(myfile, 4, '}') write_line(myfile, 3, '}') write_line(myfile, 0, '#ifdef _PARALLEL_') write_line(myfile, 0, '#pragma omp barrier') write_line(myfile, 0, '#endif') #write_line( myfile, 0, '#ifdef _PARALLEL_') #write_line( myfile, 3, '#pragma omp parallel num_threads( bl_ic_nt ) private( ic, ib, i )' ) #write_line( myfile, 0, '#endif') write_line(myfile, 3, '{') #write_line( myfile, 0, '#ifdef _PARALLEL_') write_line(myfile, 4, 'int tid = omp_get_thread_num();') write_line(myfile, 4, 'int my_start;') write_line(myfile, 4, 'int my_end;') write_line(myfile, 4, 'bl_get_range( m, DGEMM_MR, &my_start, &my_end );') #write_line( myfile, 0, '#else') #write_line( myfile, 4, 'int tid = 0;' ) #write_line( myfile, 4, 'int my_start = 0;' ) #write_line( myfile, 4, 'int my_end = m;' ) #write_line( myfile, 0, '#endif') write_line(myfile, 4, 'for ( ic = my_start; ic < my_end; ic += DGEMM_MC ) {') write_line(myfile, 5, 'ib = min( my_end - ic, DGEMM_MC );') write_line(myfile, 5, 'for ( i = 0; i < ib; i += DGEMM_MR ) {') add = 'packA_add_stra_abc%d( min( ib - i, DGEMM_MR ), pb, ' % index add += ', '.join([ '&%s%d[ pc*lda + (ic+i) ]' % ('a', i) for i in range(getNNZ(a_coeffs)) ]) add += ', lda, &packA[ tid * DGEMM_MC * pb + i * pb ] );' write_line(myfile, 6, add) write_line(myfile, 5, '}') add = 'bl_macro_kernel_stra_abc%d( ib, jb, pb, packA + tid * DGEMM_MC * pb, packB, ' % index add += ', '.join( ['&%s%d[ jc * ldc + ic ]' % ('c', i) for i in range(getNNZ(c_coeffs))]) add += ', ldc );' write_line(myfile, 5, add) write_line(myfile, 4, '}') write_line(myfile, 3, '}') write_line(myfile, 0, '#ifdef _PARALLEL_') write_line(myfile, 0, '#pragma omp barrier') write_line(myfile, 0, '#endif') write_line(myfile, 2, '}') write_line(myfile, 1, '}') write_line(myfile, 0, '#ifdef _PARALLEL_') write_line(myfile, 0, '#pragma omp barrier') write_line(myfile, 0, '#endif') write_line(myfile, 0, '}') write_break(myfile)
def write_packm_func(myfile, coeffs, index, mat_name): ''' Write the add function for a set of coefficients. This is a custom add function used for a single multiply in a single fast algorithm. coeffs is the set of coefficients used for the add ''' nonzero_coeffs = [coeff for coeff in coeffs if is_nonzero(coeff)] nnz = len(nonzero_coeffs) # TODO(arbenson): put in a code-generated comment here add = 'inline void pack%s_add_stra_abc%d( int m, int n, ' % (mat_name, index) add += ', '.join(['double *%s%d' % (mat_name, i) for i in range(nnz)]) add += ', int ld%s, double *pack%s ' % (mat_name, mat_name) add += ') {' write_line(myfile, 0, add) write_line(myfile, 1, 'int i, j;') add = 'double ' add += ', '.join(['*%s%d_pntr' % (mat_name, i) for i in range(nnz)]) add += ', *pack%s_pntr;' % mat_name write_line(myfile, 1, add) if (mat_name == 'A'): ldp = 'DGEMM_MR' incp = '1' ldm = 'ld%s' % mat_name incm = '1' elif (mat_name == 'B'): ldp = 'DGEMM_NR' incp = '1' ldm = '1' incm = 'ld%s' % mat_name else: print "Wrong mat_name!" #ldp = 'DGEMM_MR' if (mat_name == 'A') else 'DGEMM_NR' write_line(myfile, 1, 'for ( j = 0; j < n; ++j ) {') write_line(myfile, 2, 'pack%s_pntr = &pack%s[ %s * j ];' % (mat_name, mat_name, ldp)) if ldm == '1': add = ''.join([ '%s%d_pntr = &%s%d[ j ]; ' % (mat_name, i, mat_name, i) for i in range(nnz) ]) else: add = ''.join([ '%s%d_pntr = &%s%d[ %s * j ]; ' % (mat_name, i, mat_name, i, ldm) for i in range(nnz) ]) write_line(myfile, 2, add) write_line(myfile, 2, 'for ( i = 0; i < %s; ++i ) {' % ldp) add = 'pack%s_pntr[ i ]' % mat_name + ' =' for j, coeff in enumerate(nonzero_coeffs): ind = j add += arith_expression_pntr(coeff, mat_name, ind, incm) add += ';' write_line(myfile, 3, add) write_line(myfile, 2, '}') write_line(myfile, 1, '}') write_line(myfile, 0, '}') # end of function