#!/usr/bin/env python # Copyright 2019 Google LLC # # This source code is licensed under the BSD-style license found in the # LICENSE file in the root directory of this source tree. import argparse import codecs import collections import math import os import re import sys import yaml sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) from primes import next_prime import xngen import xnncommon parser = argparse.ArgumentParser(description='XNNPACK generator') parser.add_argument("-k", "--ukernel", required=True, help="microkernel") parser.add_argument("-o", "--output", metavar="FILE", required=True, help='Output (C++ source) file') parser.set_defaults(defines=list()) def split_ukernel_name(name): common_name, target_name = name.split("__", 1) common_parts = common_name.split("_") param_spec = common_parts[-1] m = re.search(r'(\d+)f(\d+)m(\d+)l(\d+)c(\d+)s(\d+)r', param_spec) assert m first_pass_tile = int(m[1]) middle_pass_tile = int(m[2]) last_pass_tile = int(m[3]) channel_tile = int(m[4]) channel_subtile = int(m[5]) channel_round = int(m[6]) arch, isa, assembly = xnncommon.parse_target_name(target_name) requantization = common_parts[-3] if requantization not in ["fp32", "rndnu"]: requantization = None return (first_pass_tile, middle_pass_tile, last_pass_tile, channel_tile, channel_subtile, channel_round, requantization, arch, isa, assembly) DWCONV_CREATE_TESTS_CODE = """\ std::vector CreateTests( size_t c_block, size_t cr, size_t kr, size_t first_pass_tile, size_t middle_pass_tile, size_t last_pass_tile, size_t channel_subtile, size_t channel_round, std::function test_func, std::function isa_check = nullptr) { const std::string cbs = std::to_string(c_block); std::vector tests; tests.reserve(17); tests.push_back(DWConvTestParams( "c_eq_" + cbs + "_first_pass_plus_one", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .kernel_size(first_pass_tile + 1) .channels(c_block) , test_func, isa_check)); tests.push_back(DWConvTestParams( "c_eq_" + cbs + "_first_pass_and_last_pass", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .kernel_size(first_pass_tile + last_pass_tile) .channels(c_block) , test_func, isa_check)); tests.push_back(DWConvTestParams( "c_eq_" + cbs + "_multipass", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .channels(c_block) , test_func, isa_check) .loop_kernel_size( first_pass_tile + middle_pass_tile + last_pass_tile, first_pass_tile + middle_pass_tile * 2 + last_pass_tile)); if (c_block > 1) { tests.push_back(DWConvTestParams( "c_div_" + cbs + "_first_pass_plus_one", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .kernel_size(first_pass_tile + 1) , test_func, isa_check) .loop_channels(c_block * 2, cr * 16, cr * 3)); tests.push_back(DWConvTestParams( "c_div_" + cbs + "_first_pass_and_last_pass", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .kernel_size(first_pass_tile + last_pass_tile) , test_func, isa_check) .loop_channels(c_block * 2, cr * 16, cr * 3)); tests.push_back(DWConvTestParams( "c_div_" + cbs + "_multipass", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) , test_func, isa_check) .loop_channels(c_block * 2, cr * 16, cr * 3) .loop_kernel_size( first_pass_tile + middle_pass_tile + last_pass_tile, first_pass_tile + middle_pass_tile * 2 + last_pass_tile)); $if ACTIVATION == "MINMAX": tests.push_back(DWConvTestParams( "c_div_" + cbs + "_with_qmin", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .kernel_size(first_pass_tile + last_pass_tile) .qmin(128) , test_func, isa_check) .loop_channels(c_block * 2, cr * 16, cr * 3)); tests.push_back(DWConvTestParams( "c_div_" + cbs + "_with_qmax", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .kernel_size(first_pass_tile + last_pass_tile) .qmax(128) , test_func, isa_check) .loop_channels(c_block * 2, cr * 16, cr * 3)); } tests.push_back(DWConvTestParams( "c_gt_" + cbs + "_first_pass_plus_one", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .kernel_size(first_pass_tile + 1) , test_func, isa_check) .loop_channels(c_block + 1, c_block == 1 ? 10 : c_block * 2)); tests.push_back(DWConvTestParams( "c_gt_" + cbs + "_first_pass_and_last_pass", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .kernel_size(first_pass_tile + last_pass_tile) , test_func, isa_check) .loop_channels(c_block + 1, c_block == 1 ? 10 : c_block * 2)); tests.push_back(DWConvTestParams( "c_gt_" + cbs + "_multipass", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) , test_func, isa_check) .loop_channels(c_block + 1, c_block == 1 ? 10 : c_block * 2) .loop_kernel_size( first_pass_tile + middle_pass_tile + last_pass_tile, first_pass_tile + middle_pass_tile * 2 + last_pass_tile)); tests.push_back(DWConvTestParams( "c_eq_" + cbs + "_first_pass_plus_one_multipixel", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .kernel_size(first_pass_tile + 1) .width(3) , test_func, isa_check) .loop_channels(1, c_block * 5 + 1, std::max(size_t(1), c_block - 1))); tests.push_back(DWConvTestParams( "c_eq_" + cbs + "_first_pass_and_last_pass_multipixel", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .kernel_size(first_pass_tile + last_pass_tile) .width(3) , test_func, isa_check) .loop_channels(1, c_block * 5 + 1, std::max(size_t(1), c_block - 1))); tests.push_back(DWConvTestParams( "c_eq_" + cbs + "_multipass_multipixel", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .width(3) , test_func, isa_check) .loop_channels(1, c_block * 5 + 1, std::max(size_t(1), c_block - 1)) .loop_kernel_size( first_pass_tile + middle_pass_tile + last_pass_tile, first_pass_tile + middle_pass_tile * 2 + last_pass_tile)); tests.push_back(DWConvTestParams( "multipixel_with_step", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .width(3) , test_func, isa_check) .loop_channels(1, c_block * 5 + 1, std::max(size_t(1), c_block - 1)) .loop_kernel_size( first_pass_tile + middle_pass_tile + last_pass_tile, first_pass_tile + middle_pass_tile * 2 + last_pass_tile) .loop_step(2, kr + 1)); tests.push_back(DWConvTestParams( "multipixel_with_output_stride", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .width(5) .output_stride(xnnpack::NextPrime(cr * 5 + 1)) , test_func, isa_check) .loop_channels(1, c_block * 5 + 1, std::max(size_t(1), c_block - 1)) .loop_kernel_size(first_pass_tile + middle_pass_tile + last_pass_tile, first_pass_tile + middle_pass_tile * 2 + last_pass_tile)); tests.push_back(DWConvTestParams( "input_offset", DWConvMicrokernelTester() .first_pass_tile(first_pass_tile) .middle_pass_tile(middle_pass_tile) .last_pass_tile(last_pass_tile) .channel_tile(cr) .channel_subtile(channel_subtile) .channel_round(channel_round) .input_offset(xnnpack::NextPrime(cr + 1) * 16) , test_func, isa_check) .loop_channels(c_block * 2, cr * 16, cr * 3) .loop_kernel_size(first_pass_tile + middle_pass_tile + last_pass_tile, first_pass_tile + middle_pass_tile * 2 + last_pass_tile)); return tests; } """ TEST_TEMPLATE = """\ #define XNN_DWCONV_MULTIPASS(arch_flags, ukernel, first_pass_tile, middle_pass_tile, last_pass_tile, channel_tile, channel_subtile, channel_round, datatype, weights_type, buffer_type, params_type, init_params) INSTANTIATE_TEST_SUITE_P( ukernel, DWConvTest, testing::ValuesIn(CreateTests( channel_tile, channel_tile, first_pass_tile, first_pass_tile, middle_pass_tile, last_pass_tile, channel_subtile, channel_round, [](DWConvMicrokernelTester& tester) { TEST_REQUIRES_ARCH_FLAGS(arch_flags); tester.Test(${", ".join(TEST_ARGS)}); })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); """ def main(args): options = parser.parse_args(args) ukernel = options.ukernel test_header = """\ // Copyright 2022 Google LLC // // This source code is licensed under the BSD-style license found in the // LICENSE file in the root directory of this source tree. // // Auto-generated file. Do not edit! // Microkernel: {ukernel} // Generator: {generator} #include #include #include #include #include #include #include "xnnpack/common.h" #include "xnnpack/dwconv.h" #include "xnnpack/isa-checks.h" #include "xnnpack/microparams-init.h" #include "xnnpack/requantization.h" #include "dwconv-microkernel-tester.h" #include "next_prime.h" """.format(ukernel=ukernel, generator=sys.argv[0]) test_cases = "" parts = ukernel.split("-") datatype = parts[0] folder = datatype + "-dwconv" if parts[1] == "qc8w": folder = datatype + "-qc8w-dwconv" parts.pop(1) activation = "minmax" if "minmax" in parts else "linear" ukernel_type = "unipass" if "unipass" in parts else "multipass" requantization = "fp32" if "fp32" in parts else "rndnu" if "rndnu" in parts else None create_tests_args = { "UKERNEL_TYPE": ukernel_type.upper(), "DATATYPE": datatype, "ACTIVATION": activation.upper(), } create_tests = xngen.preprocess(DWCONV_CREATE_TESTS_CODE, create_tests_args) create_tests = ( "namespace {\n\n" + "\n".join([create_tests]) + "\n} // namespace\n" ) tests = test_header + "\n" + create_tests + "\n" + test_cases test_args = ["ukernel", "init_params"] if requantization: requantization_datatype = {"qc8": "qs8"}.get(datatype, datatype) test_args.append( "xnn_%s_requantize_%s" % (requantization_datatype, requantization) ) tests += xnncommon.make_multiline_macro(xngen.preprocess( TEST_TEMPLATE, { "TEST_ARGS": test_args, }, )) tests += f'#include "{xnncommon.xnnpack_src()}{folder}/{options.ukernel}.h"\n' tests += "#undef XNN_UKERNEL_WITH_PARAMS\n" xnncommon.overwrite_if_changed(options.output, tests) if __name__ == "__main__": main(sys.argv[1:])