def write_divisor_initializer(myfile, dims, level):
level_dim = exp_dim(dims, level)
write_line(
myfile, 1, 'const std::array<unsigned,2> A_divisor={%d,%d};' %
(level_dim[0], level_dim[1]))
write_line(
myfile, 1, 'const std::array<unsigned,2> B_divisor={%d,%d};' %
(level_dim[1], level_dim[2]))
write_line(
myfile, 1, 'const std::array<unsigned,2> C_divisor={%d,%d};' %
(level_dim[1], level_dim[2]))
write_break(myfile)
def gen_abc_fmm(coeff_filename, dims, level, outfilename,
micro_kernel_filename, kernel_header_filename):
coeffs = read_coeffs(coeff_filename)
#print coeffs
#print coeffs[0][0]
#coeffs2= [ transpose( U2 ), transpose( V2 ), transpose( W2 ) ]
with open(outfilename, 'w') as myfile:
write_line(myfile, 0, '#include "%s"' % kernel_header_filename[10:])
write_line(myfile, 0, '#include "bl_dgemm.h"')
write_break(myfile)
cur_coeffs = generateCoeffs(coeffs, level)
#writeCoeffs( cur_coeffs )
#writeEquation( cur_coeffs, dims, level )
num_multiplies = len(cur_coeffs[0][0])
create_packm_functions(myfile, cur_coeffs)
my_micro_file = open(micro_kernel_filename, 'w')
create_micro_functions(my_micro_file, cur_coeffs,
kernel_header_filename[10:])
my_kernel_header = open(kernel_header_filename, 'w')
create_kernel_header(my_kernel_header, cur_coeffs)
create_macro_functions(myfile, cur_coeffs)
create_straprim_abc_functions(myfile, cur_coeffs, dims, level)
write_line(
myfile, 0,
'void bl_dgemm_strassen_abc( int m, int n, int k, double *XA, int lda, double *XB, int ldb, double *XC, int ldc )'
)
write_line(myfile, 0, '{')
write_abc_strassen_header(myfile)
writePartition(myfile, dims, level)
write_break(myfile)
write_line(myfile, 0, '#ifdef _PARALLEL_')
write_line(myfile, 1, '#pragma omp parallel num_threads( bl_ic_nt )')
write_line(myfile, 0, '#endif')
write_line(myfile, 1, '{')
create_straprim_caller(myfile, cur_coeffs, dims, num_multiplies, level)
write_line(myfile, 1, '}')
write_break(myfile)
level_dim = exp_dim(dims, level)
write_line(
myfile, 1,
'bl_dynamic_peeling( m, n, k, XA, lda, XB, ldb, XC, ldc, %d * DGEMM_MR, %d, %d * DGEMM_NR );'
% (level_dim[0], level_dim[1], level_dim[2]))
write_break(myfile)
write_line(myfile, 1, '//free( packA );')
write_line(myfile, 1, '//free( packB );')
write_line(myfile, 0, '}')