1#!/usr/bin/env python
2
3
4import confu
5parser = confu.standard_parser("FP16 configuration script")
6parser.add_argument("--compare", dest="compare", action="store_true",
7    help="Enable performance comparison with other half-precision implementations")
8
9def main(args):
10    options = parser.parse_args(args)
11    build = confu.Build.from_options(options)
12
13    build.export_cpath("include", ["fp16.h"])
14
15    with build.options(source_dir="test", extra_include_dirs="test", deps=[build.deps.googletest, build.deps.psimd]):
16        fp16_tables = build.cxx("tables.cc")
17        build.unittest("ieee-to-fp32-bits",
18            [build.cxx("ieee-to-fp32-bits.cc"), fp16_tables])
19        build.unittest("ieee-to-fp32-value",
20            [build.cxx("ieee-to-fp32-value.cc"), fp16_tables])
21        build.unittest("ieee-from-fp32-value",
22            [build.cxx("ieee-from-fp32-value.cc"), fp16_tables])
23
24        build.unittest("alt-to-fp32-bits",
25            [build.cxx("alt-to-fp32-bits.cc"), fp16_tables])
26        build.unittest("alt-to-fp32-value",
27            [build.cxx("alt-to-fp32-value.cc"), fp16_tables])
28        build.unittest("alt-from-fp32-value",
29            [build.cxx("alt-from-fp32-value.cc"), fp16_tables])
30
31        if build.target.is_x86_64:
32            stubs = build.peachpy("peachpy/stubs.py")
33            build.unittest("alt-xmm-to-fp32-ymm-avx", [build.cxx("peachpy/alt-xmm-to-fp32-xmm-avx.cc"), stubs])
34            build.unittest("alt-xmm-to-fp32-ymm-avx2", [build.cxx("peachpy/alt-xmm-to-fp32-ymm-avx2.cc"), stubs])
35
36        if not build.target.is_emscripten:
37            build.unittest("ieee-to-fp32-psimd", build.cxx("ieee-to-fp32-psimd.cc"))
38            build.unittest("alt-to-fp32-psimd", build.cxx("alt-to-fp32-psimd.cc"))
39
40            build.unittest("ieee-to-fp32x2-psimd", build.cxx("ieee-to-fp32x2-psimd.cc"))
41            build.unittest("alt-to-fp32x2-psimd", build.cxx("alt-to-fp32x2-psimd.cc"))
42
43        build.unittest("bitcasts", build.cxx("bitcasts.cc"))
44
45    macros = ["BENCHMARK_HAS_NO_INLINE_ASSEMBLY"]
46    if options.compare:
47        macros.append("FP16_COMPARATIVE_BENCHMARKS")
48    with build.options(source_dir="bench", extra_include_dirs=".", macros=macros,
49            deps=[build.deps.googlebenchmark, build.deps.psimd]):
50
51        build.benchmark("ieee-element-bench", build.cxx("ieee-element.cc"))
52        build.benchmark("alt-element-bench", build.cxx("alt-element.cc"))
53
54        build.benchmark("from-ieee-array-bench", build.cxx("from-ieee-array.cc"))
55        build.benchmark("from-alt-array-bench", build.cxx("from-alt-array.cc"))
56
57        build.benchmark("to-ieee-array-bench", build.cxx("to-ieee-array.cc"))
58        build.benchmark("to-alt-array-bench", build.cxx("to-alt-array.cc"))
59
60    return build
61
62
63if __name__ == "__main__":
64    import sys
65    main(sys.argv[1:]).generate()
66