mat_a = DenseMatrix(num_rows=56,
num_cols=56,
addressing=Addressing.NONE,
bbox=[0, 0, 20, 9])
mat_b = DenseMatrix(num_rows=56,
num_cols=56,
addressing=Addressing.STRIDED,
bbox=[0, 0, 9, 20])
tmp1 = generate_tmp_matrix(mat_a, mat_b)
gemm_list = [GemmDescr(trans_a=False,
trans_b=False,
a=mat_a, b=mat_b, c=tmp1),
GemmDescr(trans_a=False,
trans_b=True,
a=tmp1, b=mat_b, c=mat_q)]
vm = vm_factory(name='nvidia',
sub_name='sm_60',
fp_type=FloatingPointType.FLOAT)
generator = Generator(gemm_list, vm)
generator.generate()
print(generator.get_launcher())
print()
print(generator.get_header())
print()
print(generator.get_kernel())
def main():
cmd = argparse.ArgumentParser()
cmd.add_argument('-i', '--input', type=str, help="input file")
cmd.add_argument('-c', '--config', type=str, help="config file")
cmd.add_argument('-a', '--arch', type=str, help='gpu arch (nvidia, amd)')
cmd.add_argument('-s',
'--sub_arch',
type=str,
help='sub architecture (nvidia, amd)')
args = cmd.parse_args()
try:
does_file_exist(args.input, format='.cf')
does_file_exist(args.config, format='.yaml')
except ValueError as err:
print(f'{err}')
sys.exit(-1)
curr_dir = path.join(path.dirname(__file__))
with open(f'{curr_dir}/{args.input}', 'r') as file:
program = file.read()
# get AST and append symbol table with temporaries
parser = Parser()
ast, symbol_table = parser.parse(translation_unit=program)
# convert AST to lists of gemms
symbol_table.add_scope()
processor = PostProcessor(ast, symbol_table)
gemm_dicts = processor.process()
stream = open(args.config, 'r')
config = yaml.safe_load(stream)
vm = vm_factory(name=args.arch,
sub_name=args.sub_arch,
fp_type=FloatingPointType.str2enum(config['fp_type']))
kernels = []
launchers = []
headers = []
benchmarks_src = []
benchmarks_names = []
for bench_name, gemm_list in gemm_dicts.items():
gpu_generator = Generator(gemm_list, vm)
gpu_generator.set_kernel_name(bench_name)
gpu_generator.generate()
# write kernel, launcher and header to files
kernels.append(gpu_generator.get_kernel())
launchers.append(gpu_generator.get_launcher())
headers.append(gpu_generator.get_header())
call_site = Aux.get_call_site(gpu_generator, gemm_list)
bench_generator = BenchGenerator(bench_name, deepcopy(symbol_table),
gemm_list, config, call_site,
args.arch)
benchmarks_names.append(bench_name)
benchmarks_src.append(bench_generator.generate())
# generate main file
enty_point = EnryPointGenerator(benchmarks_names, benchmarks_src, config)
main_src = enty_point.generate()
tmp_dir = make_tmp_folder(file_name='tmp')
with open(path.join(tmp_dir, 'main.cu'), 'w') as file:
file.write(main_src)
# write kernel, launcher and header to files
with open(path.join(tmp_dir, 'kernel.cu'), 'w') as file:
file.write('#include \"chainforge_aux.h\"\n')
for kernel, launcher in zip(kernels, launchers):
file.write(kernel)
file.write(launcher)
with open(path.join(tmp_dir, 'kernel.h'), 'w') as file:
file.write('#ifndef KERNEL_H\n')
file.write('#define KERNEL_H\n')
for header in headers:
file.write(header)
file.write('#endif\n')
with open(path.join(tmp_dir, 'cmake_params.cmake'), 'w') as file:
real_size = 8 if config['fp_type'] == 'double' else 4
file.write(f'set(REAL_SIZE {real_size})\n')
file.write(f'set(ARCH {args.arch})\n')
file.write(f'set(SM_ARCH {args.sub_arch})\n')