1*5f32b710SXin Li#!/usr/bin/env python 2*5f32b710SXin Li 3*5f32b710SXin Li 4*5f32b710SXin Liimport confu 5*5f32b710SXin Liparser = confu.standard_parser("FP16 configuration script") 6*5f32b710SXin Liparser.add_argument("--compare", dest="compare", action="store_true", 7*5f32b710SXin Li help="Enable performance comparison with other half-precision implementations") 8*5f32b710SXin Li 9*5f32b710SXin Lidef main(args): 10*5f32b710SXin Li options = parser.parse_args(args) 11*5f32b710SXin Li build = confu.Build.from_options(options) 12*5f32b710SXin Li 13*5f32b710SXin Li build.export_cpath("include", ["fp16.h"]) 14*5f32b710SXin Li 15*5f32b710SXin Li with build.options(source_dir="test", extra_include_dirs="test", deps=[build.deps.googletest, build.deps.psimd]): 16*5f32b710SXin Li fp16_tables = build.cxx("tables.cc") 17*5f32b710SXin Li build.unittest("ieee-to-fp32-bits", 18*5f32b710SXin Li [build.cxx("ieee-to-fp32-bits.cc"), fp16_tables]) 19*5f32b710SXin Li build.unittest("ieee-to-fp32-value", 20*5f32b710SXin Li [build.cxx("ieee-to-fp32-value.cc"), fp16_tables]) 21*5f32b710SXin Li build.unittest("ieee-from-fp32-value", 22*5f32b710SXin Li [build.cxx("ieee-from-fp32-value.cc"), fp16_tables]) 23*5f32b710SXin Li 24*5f32b710SXin Li build.unittest("alt-to-fp32-bits", 25*5f32b710SXin Li [build.cxx("alt-to-fp32-bits.cc"), fp16_tables]) 26*5f32b710SXin Li build.unittest("alt-to-fp32-value", 27*5f32b710SXin Li [build.cxx("alt-to-fp32-value.cc"), fp16_tables]) 28*5f32b710SXin Li build.unittest("alt-from-fp32-value", 29*5f32b710SXin Li [build.cxx("alt-from-fp32-value.cc"), fp16_tables]) 30*5f32b710SXin Li 31*5f32b710SXin Li if build.target.is_x86_64: 32*5f32b710SXin Li stubs = build.peachpy("peachpy/stubs.py") 33*5f32b710SXin Li build.unittest("alt-xmm-to-fp32-ymm-avx", [build.cxx("peachpy/alt-xmm-to-fp32-xmm-avx.cc"), stubs]) 34*5f32b710SXin Li build.unittest("alt-xmm-to-fp32-ymm-avx2", [build.cxx("peachpy/alt-xmm-to-fp32-ymm-avx2.cc"), stubs]) 35*5f32b710SXin Li 36*5f32b710SXin Li if not build.target.is_emscripten: 37*5f32b710SXin Li build.unittest("ieee-to-fp32-psimd", build.cxx("ieee-to-fp32-psimd.cc")) 38*5f32b710SXin Li build.unittest("alt-to-fp32-psimd", build.cxx("alt-to-fp32-psimd.cc")) 39*5f32b710SXin Li 40*5f32b710SXin Li build.unittest("ieee-to-fp32x2-psimd", build.cxx("ieee-to-fp32x2-psimd.cc")) 41*5f32b710SXin Li build.unittest("alt-to-fp32x2-psimd", build.cxx("alt-to-fp32x2-psimd.cc")) 42*5f32b710SXin Li 43*5f32b710SXin Li build.unittest("bitcasts", build.cxx("bitcasts.cc")) 44*5f32b710SXin Li 45*5f32b710SXin Li macros = ["BENCHMARK_HAS_NO_INLINE_ASSEMBLY"] 46*5f32b710SXin Li if options.compare: 47*5f32b710SXin Li macros.append("FP16_COMPARATIVE_BENCHMARKS") 48*5f32b710SXin Li with build.options(source_dir="bench", extra_include_dirs=".", macros=macros, 49*5f32b710SXin Li deps=[build.deps.googlebenchmark, build.deps.psimd]): 50*5f32b710SXin Li 51*5f32b710SXin Li build.benchmark("ieee-element-bench", build.cxx("ieee-element.cc")) 52*5f32b710SXin Li build.benchmark("alt-element-bench", build.cxx("alt-element.cc")) 53*5f32b710SXin Li 54*5f32b710SXin Li build.benchmark("from-ieee-array-bench", build.cxx("from-ieee-array.cc")) 55*5f32b710SXin Li build.benchmark("from-alt-array-bench", build.cxx("from-alt-array.cc")) 56*5f32b710SXin Li 57*5f32b710SXin Li build.benchmark("to-ieee-array-bench", build.cxx("to-ieee-array.cc")) 58*5f32b710SXin Li build.benchmark("to-alt-array-bench", build.cxx("to-alt-array.cc")) 59*5f32b710SXin Li 60*5f32b710SXin Li return build 61*5f32b710SXin Li 62*5f32b710SXin Li 63*5f32b710SXin Liif __name__ == "__main__": 64*5f32b710SXin Li import sys 65*5f32b710SXin Li main(sys.argv[1:]).generate() 66