From a25557c868443f22bcb549651e2b0021e029cf86 Mon Sep 17 00:00:00 2001
From: Carles Fernandez <carles.fernandez@gmail.com>
Date: Sat, 21 Jul 2018 09:18:37 +0200
Subject: [PATCH] Add detection of AVX2551CD and AVX512F

---
 .../volk_gnsssdr/gen/archs.xml                | 44 +++++++++++++++++++
 .../volk_gnsssdr/gen/machines.xml             | 10 +++++
 .../gen/volk_gnsssdr_kernel_defs.py           |  6 +--
 .../volk_gnsssdr/tmpl/volk_gnsssdr_cpu.tmpl.c |  9 ++++
 4 files changed, 66 insertions(+), 3 deletions(-)

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 @@
     <alignment>32</alignment>
 </arch>
 
+<arch name="avx512f">
+    <!-- check for AVX512F -->
+    <check name="cpuid_count_x86_bit">
+        <param>7</param>
+        <param>0</param>
+        <param>1</param>
+        <param>16</param>
+    </check>
+    <!-- check to make sure that xgetbv is enabled in OS -->
+    <check name="cpuid_x86_bit">
+        <param>2</param>
+        <param>0x00000001</param>
+        <param>27</param>
+    </check>
+    <!-- check to see that the OS has enabled AVX512 -->
+    <check name="get_avx512_enabled"></check>
+    <flag compiler="gnu">-mavx512f</flag>
+    <flag compiler="clang">-mavx512f</flag>
+    <flag compiler="msvc">/arch:AVX512F</flag>
+    <alignment>64</alignment>
+</arch>
+
+<arch name="avx512cd">
+    <!-- check for AVX512CD -->
+    <check name="cpuid_count_x86_bit">
+        <param>7</param>
+        <param>0</param>
+        <param>1</param>
+        <param>28</param>
+    </check>
+    <!-- check to make sure that xgetbv is enabled in OS -->
+    <check name="cpuid_x86_bit">
+        <param>2</param>
+        <param>0x00000001</param>
+        <param>27</param>
+    </check>
+    <!-- check to see that the OS has enabled AVX512 -->
+    <check name="get_avx512_enabled"></check>
+    <flag compiler="gnu">-mavx512cd</flag>
+    <flag compiler="clang">-mavx512cd</flag>
+    <flag compiler="msvc">/arch:AVX512CD</flag>
+    <alignment>64</alignment>
+</arch>
+
 </grammar>
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 @@
 <archs>generic 32|64| mmx| sse sse2 sse3 ssse3 sse4_1 sse4_2 popcount avx fma avx2 orc|</archs>
 </machine>
 
+<!-- trailing | bar means generate without either for MSVC -->
+<machine name="avx512f">
+<archs>generic 32|64| mmx| sse sse2 sse3 ssse3 sse4_1 sse4_2 popcount avx fma avx2 avx512f orc|</archs>
+</machine>
+
+<!-- trailing | bar means generate without either for MSVC -->
+<machine name="avx512cd">
+<archs>generic 32|64| mmx| sse sse2 sse3 ssse3 sse4_1 sse4_2 popcount avx fma avx2 avx512f avx512cd orc|</archs>
+</machine>
+
 </grammar>
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 <asm/hwcap.h>