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_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 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 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 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_common_start_assembly(myfile, nnz):
myfile.write( \
'''\
void* b_next = bli_auxinfo_next_b( data );
uint64_t k_iter = k / 4;
uint64_t k_left = k % 4;
''' )
add = 'double '
add += ', '.join(
['*coeff%d = &coeff_list[%d]' % (i, i) for i in range(nnz)])
add += ';'
write_line(myfile, 1, add)
add = 'double '
add += ', '.join(['*c%d = c_list[%d]' % (i, i) for i in range(nnz)])
add += ';'
write_line(myfile, 1, add)
write_break(myfile)
myfile.write( \
'''\
__asm__ volatile
(
" \\n\\t"
" \\n\\t"
"movq %[a], %%rax \\n\\t" // load address of a. ( v )
"movq %[b], %%rbx \\n\\t" // load address of b. ( v )
"movq %[b_next], %%r15 \\n\\t" // load address of b_next. ( v )
"addq $-4 * 64, %%r15 \\n\\t" // ( ? )
" \\n\\t"
"vmovapd 0 * 32(%%rax), %%ymm0 \\n\\t" // initialize loop by pre-loading
"vmovapd 0 * 32(%%rbx), %%ymm2 \\n\\t" // elements of a and b.
"vpermilpd $0x5, %%ymm2, %%ymm3 \\n\\t"
" \\n\\t"
" \\n\\t"
"movq %[cs_c], %%rdi \\n\\t" // load cs_c
"leaq (,%%rdi,8), %%rdi \\n\\t" // cs_c * sizeof(double)
''' )
def gen_abc_fmm(coeff_filename, dims, level, outfilename):
coeffs = read_coeffs(coeff_filename)
#print coeffs
#print coeffs[0][0]
#coeffs2= [ transpose( U2 ), transpose( V2 ), transpose( W2 ) ]
with open(outfilename, 'w') as myfile:
cur_coeffs = generateCoeffs(coeffs, level)
#writeCoeffs( cur_coeffs )
#writeEquation( cur_coeffs, dims, level )
num_multiplies = len(cur_coeffs[0][0])
writePartition(myfile, dims, level)
write_break(myfile)
create_straprim_caller(myfile, cur_coeffs, dims, num_multiplies, level)
def write_abc_strassen_header(myfile):
write_line(myfile, 1, 'double *packA, *packB;')
write_break(myfile)
write_line(myfile, 1,
'int bl_ic_nt = bl_read_nway_from_env( "BLISLAB_IC_NT" );')
write_break(myfile)
write_line(myfile, 1, '//// Allocate packing buffers')
write_line(
myfile, 1,
'//packA = bl_malloc_aligned( DGEMM_KC, ( DGEMM_MC + 1 ) * bl_ic_nt, sizeof(double) );'
)
write_line(
myfile, 1,
'//packB = bl_malloc_aligned( DGEMM_KC, ( DGEMM_NC + 1 ) , sizeof(double) );'
)
write_line(myfile, 1,
'bl_malloc_packing_pool( &packA, &packB, n, bl_ic_nt );')
write_break(myfile)
def create_macro_functions(myfile, coeffs):
for i, coeff_set in enumerate(transpose(coeffs[2])):
if len(coeff_set) > 0:
write_macro_func(myfile, coeff_set, i, 'C')
write_break(myfile)
def all_adds(coeffs, name):
for i, coeff_set in enumerate(coeffs):
if len(coeff_set) > 0:
write_packm_func(myfile, coeff_set, i, name)
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, '}')
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)
