diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/archs.xml b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/archs.xml index a8bee46a2..1f741ac02 100644 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/archs.xml +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/archs.xml @@ -238,4 +238,48 @@ 32 + + + + 7 + 0 + 1 + 16 + + + + 2 + 0x00000001 + 27 + + + + -mavx512f + -mavx512f + /arch:AVX512F + 64 + + + + + + 7 + 0 + 1 + 28 + + + + 2 + 0x00000001 + 27 + + + + -mavx512cd + -mavx512cd + /arch:AVX512CD + 64 + + diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/machines.xml b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/machines.xml index e7c6eaf20..d787a6b14 100644 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/machines.xml +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/machines.xml @@ -51,4 +51,14 @@ generic 32|64| mmx| sse sse2 sse3 ssse3 sse4_1 sse4_2 popcount avx fma avx2 orc| + + +generic 32|64| mmx| sse sse2 sse3 ssse3 sse4_1 sse4_2 popcount avx fma avx2 avx512f orc| + + + + +generic 32|64| mmx| sse sse2 sse3 ssse3 sse4_1 sse4_2 popcount avx fma avx2 avx512f avx512cd orc| + + diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/volk_gnsssdr_kernel_defs.py b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/volk_gnsssdr_kernel_defs.py index 9daddda73..fe03ae701 100644 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/volk_gnsssdr_kernel_defs.py +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/gen/volk_gnsssdr_kernel_defs.py @@ -121,7 +121,7 @@ def flatten_section_text(sections): class impl_class: def __init__(self, kern_name, header, body): #extract LV_HAVE_* - self.deps = set(map(str.lower, re.findall('LV_HAVE_(\w+)', header))) + self.deps = set(res.lower() for res in re.findall('LV_HAVE_(\w+)', header)) #extract function suffix and args body = flatten_section_text(body) try: @@ -153,7 +153,7 @@ def extract_lv_haves(code): haves = list() for line in code.splitlines(): if not line.strip().startswith('#'): continue - have_set = set(map(str.lower, re.findall('LV_HAVE_(\w+)', line))) + have_set = set(res.lower() for res in re.findall('LV_HAVE_(\w+)', line)) if have_set: haves.append(have_set) return haves @@ -164,7 +164,7 @@ class kernel_class: def __init__(self, kernel_file): self.name = os.path.splitext(os.path.basename(kernel_file))[0] self.pname = self.name.replace('volk_gnsssdr_', 'p_') - code = open(kernel_file, 'r').read() + code = open(kernel_file, 'rb').read().decode("utf-8") code = comment_remover(code) sections = split_into_nested_ifdef_sections(code) self._impls = list() diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/tmpl/volk_gnsssdr_cpu.tmpl.c b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/tmpl/volk_gnsssdr_cpu.tmpl.c index 4aba13014..420b5c91d 100644 --- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/tmpl/volk_gnsssdr_cpu.tmpl.c +++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/tmpl/volk_gnsssdr_cpu.tmpl.c @@ -123,6 +123,15 @@ static inline unsigned int get_avx2_enabled(void) #endif } +static inline unsigned int get_avx512_enabled(void) +{ +#if defined(VOLK_CPU_x86) + return __xgetbv() & 0xE6; //check for zmm, xmm and ymm regs +#else + return 0; +#endif +} + //neon detection is linux specific #if defined(__arm__) && defined(__linux__) #include