Merge remote branch 'nick/volk_cmake'

8 files changed, 91 insertions, 13 deletions

diff --git a/volk/gen/archs.xml b/volk/gen/archs.xml
index 977cc7924..f6822871f 100644
--- a/volk/gen/archs.xml
+++ b/volk/gen/archs.xml
@@ -7,6 +7,12 @@
 
 <arch name="altivec" type="powerpc">
   <flag>maltivec</flag>
+  <alignment>16</alignment>
+</arch>
+
+<arch name="neon" type="arm">
+  <flag>mfpu=neon -mfloat-abi=softfp -funsafe-math-optimizations</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="32" type="x86" no_test="true" >
@@ -31,6 +37,7 @@
   <shift>31</shift>
   <flag>m3dnow</flag>
   <val>1</val>
+  <alignment>8</alignment>
 </arch>
 
 <arch name="abm" type="x86">
@@ -39,6 +46,7 @@
   <reg>d</reg>
   <shift>5</shift>
   <flag>sse4.2</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="popcount" type="x86">
@@ -55,6 +63,7 @@
   <reg>d</reg>
   <shift>23</shift>
   <flag>mmmx</flag>
+  <alignment>8</alignment>
 </arch>
 
 
@@ -66,6 +75,7 @@
   <flag>msse</flag>
   <environment>_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);</environment>
   <include>xmmintrin.h</include>
+  <alignment>16</alignment>
 </arch>
 
 
@@ -75,6 +85,7 @@
   <reg>d</reg>
   <shift>26</shift>
   <flag>msse2</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="orc" type="all">
@@ -91,6 +102,7 @@
   <flag>msse3</flag>
   <environment>_MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);</environment>
   <include>pmmintrin.h</include>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="ssse3" type="x86">
@@ -99,6 +111,7 @@
   <reg>c</reg>
   <shift>9</shift>
   <flag>mssse3</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="sse4_a" type="x86">
@@ -107,6 +120,7 @@
   <reg>c</reg>
   <shift>6</shift>
   <flag>msse4a</flag>
+  <alignment>16</alignment>
 </arch>
 
 
@@ -116,6 +130,7 @@
   <reg>c</reg>
   <shift>19</shift>
   <flag>msse4.1</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="sse4_2" type="x86">
@@ -124,6 +139,7 @@
   <reg>c</reg>
   <shift>20</shift>
   <flag>msse4.2</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="avx" type="x86">
@@ -132,6 +148,7 @@
   <reg>c</reg>
   <shift>28</shift>
   <flag>mavx</flag>
+  <alignment>32</alignment>
 </arch>
 
 </grammar>
diff --git a/volk/gen/machines.xml b/volk/gen/machines.xml
index 1f8b201ae..b872b9fb1 100644
--- a/volk/gen/machines.xml
+++ b/volk/gen/machines.xml
@@ -20,6 +20,11 @@ This machine is intended to support the MSVC compiler on x86/amd64.
 The MSVC compiler has intrinsic support for SSE and SSE2,
 however it does not support the gcc style inline assembly.
 -->
+
+<machine name="neon">
+<archs>generic neon</archs>
+</machine>
+
 <machine name="sse2_only">
 <archs>generic mmx sse sse2</archs>
 </machine>
diff --git a/volk/gen/make_c.py b/volk/gen/make_c.py
index e946152d0..19d679e71 100644
--- a/volk/gen/make_c.py
+++ b/volk/gen/make_c.py
@@ -66,6 +66,10 @@ struct volk_machine *get_machine(void) {
     }
 }
 
+unsigned int volk_get_alignment(void) {
+    return get_machine()->alignment;
+}
+
 """
     
     for i in range(len(functions)):
diff --git a/volk/gen/make_cpuid_c.py b/volk/gen/make_cpuid_c.py
index 2fdbaf304..eb88dcd7f 100644
--- a/volk/gen/make_cpuid_c.py
+++ b/volk/gen/make_cpuid_c.py
@@ -153,6 +153,45 @@ int i_can_has_%s () {
 }
 
 """ % (arch)
+
+        elif str(domarch.attributes["type"].value) == "arm":
+            arch = str(domarch.attributes["name"].value);
+            tempstring = tempstring + """\
+#if defined(__arm__) && defined(__linux__)
+#include <asm/hwcap.h>
+#include <linux/auxvec.h>
+#include <stdio.h>
+#define LOOK_FOR_NEON
+#endif
+
+int i_can_has_%s () {
+//it's linux-specific, but if you're compiling libvolk for NEON
+//on Windows you have other problems
+
+#ifdef LOOK_FOR_NEON
+    FILE *auxvec_f;
+    unsigned long auxvec[2];
+    unsigned int found_neon = 0;
+    auxvec_f = fopen("/proc/self/auxv", "rb");
+    if(!auxvec_f) return 0;
+    
+    //so auxv is basically 32b of ID and 32b of value
+    //so it goes like this
+    while(!found_neon && auxvec_f) {
+      fread(auxvec, sizeof(unsigned long), 2, auxvec_f);
+      if((auxvec[0] == AT_HWCAP) && (auxvec[1] & HWCAP_NEON))
+        found_neon = 1;
+    }
+    
+    fclose(auxvec_f);
+    return found_neon;
+
+#else
+    return 0;
+#endif
+}
+
+""" % (arch)
         
         elif str(domarch.attributes["type"].value) == "all":
             arch = str(domarch.attributes["name"].value);
diff --git a/volk/gen/make_each_machine_c.py b/volk/gen/make_each_machine_c.py
index d4f5f01de..44e2ef3f2 100644
--- a/volk/gen/make_each_machine_c.py
+++ b/volk/gen/make_each_machine_c.py
@@ -18,7 +18,7 @@
 from volk_regexp import *
 import string
 
-def _make_each_machine_struct(machine_name, archs, functions, fcountlist, taglist):
+def _make_each_machine_struct(machine_name, archs, functions, fcountlist, taglist, alignment):
 
     #make the machine fcountlist and taglist a subset given the archs list
     machine_fcountlists = list()
@@ -38,6 +38,7 @@ def _make_each_machine_struct(machine_name, archs, functions, fcountlist, taglis
     tempstring += "struct volk_machine volk_machine_" + machine_name + " = {\n"
     tempstring += "    " + ' | '.join(["(1 << LV_" + arch.swapcase() + ")" for arch in archs]) + ",\n"
     tempstring += "    \"%s\",\n"%machine_name
+    tempstring += "    %s,\n"%alignment
 
     #fill in the description for each function
     for i in range(len(functions)):
@@ -51,8 +52,7 @@ def _make_each_machine_struct(machine_name, archs, functions, fcountlist, taglis
     tempstring += "};\n"
     return tempstring
 
-def make_each_machine_c(machine_name, archs, functions, fcountlist, taglist):
-
+def make_each_machine_c(machine_name, archs, functions, fcountlist, taglist, alignment):
     tempstring = r"""
 // This file is automatically generated by make_each_machine_c.py.
 // Do not edit this file.
@@ -77,8 +77,8 @@ def make_each_machine_c(machine_name, archs, functions, fcountlist, taglist):
 %s
 #endif
 """%(
-    _make_each_machine_struct(machine_name, archs+["orc"], functions, fcountlist, taglist),
-    _make_each_machine_struct(machine_name, archs, functions, fcountlist, taglist)
+    _make_each_machine_struct(machine_name, archs+["orc"], functions, fcountlist, taglist, alignment),
+    _make_each_machine_struct(machine_name, archs, functions, fcountlist, taglist, alignment)
 )
 
     return tempstring
diff --git a/volk/gen/make_h.py b/volk/gen/make_h.py
index 3d5790de4..354e57258 100644
--- a/volk/gen/make_h.py
+++ b/volk/gen/make_h.py
@@ -21,9 +21,12 @@ struct volk_func_desc {
      const int *arch_defs;
      const int n_archs;
 };
+
+VOLK_API unsigned int volk_get_alignment(void);
+
 """
     for i in range(len(funclist)):
-        tempstring += "extern " + replace_volk.sub("p", funclist[i]) + " " + funclist[i] + ";\n"
+        tempstring += "extern VOLK_API " + replace_volk.sub("p", funclist[i]) + " " + funclist[i] + ";\n"
         tempstring += "extern VOLK_API void %s_manual%s;\n" % (funclist[i], arched_arglist[i])
         tempstring = strip_trailing(tempstring, " {")
         tempstring += "extern VOLK_API struct volk_func_desc %s_get_func_desc(void);\n" % (funclist[i])
diff --git a/volk/gen/make_machines_h.py b/volk/gen/make_machines_h.py
index 563de18a6..a48caa89c 100644
--- a/volk/gen/make_machines_h.py
+++ b/volk/gen/make_machines_h.py
@@ -33,6 +33,7 @@ __VOLK_DECL_BEGIN
 struct volk_machine {
    const unsigned int caps; //capabilities (i.e., archs compiled into this machine, in the volk_get_lvarch format)
    const char *name;
+   const unsigned int alignment; //the maximum byte alignment required for functions in this library
 """
     for function in functions:
         tempstring += "    const char *%s_name;\n"%function
diff --git a/volk/gen/volk_register.py b/volk/gen/volk_register.py
index 62ebba3c2..75e5eeb87 100644
--- a/volk/gen/volk_register.py
+++ b/volk/gen/volk_register.py
@@ -59,7 +59,7 @@ functions = [];
 
 
 for line in mfile:
-    subline = re.search(".*_(a16|u)\.h.*", line);
+    subline = re.search(".*_(a|u)\.h.*", line);
     if subline:
         subsubline = re.search("(?<=volk_).*", subline.group(0));
         if subsubline:
@@ -74,7 +74,7 @@ datatypes = set(datatypes);
 for line in mfile:
     for dt in datatypes:
         if dt in line:
-            subline = re.search("(volk_" + dt +"_.*(a16|u).*\.h)", line);
+            subline = re.search("(volk_" + dt +"_.*(a|u).*\.h)", line);
             if subline:
                 
                 subsubline = re.search(".+(?=\.h)", subline.group(0));
@@ -97,6 +97,12 @@ archflags_dict = {}
 for filearch in filearchs:
     archflags_dict[str(filearch.attributes["name"].value)] = str(filearch.getElementsByTagName("flag")[0].firstChild.data)
 
+archalign_dict = {}
+for filearch in filearchs:
+    alignelem = filearch.getElementsByTagName("alignment")
+    if(alignelem):
+        archalign_dict[str(filearch.attributes["name"].value)] = int(alignelem[0].firstChild.data)
+
 archs_or = "("
 for arch in archs:
     archs_or = archs_or + string.upper(arch) + "|";
@@ -132,11 +138,14 @@ for machine_name in machine_str_dict:
 
     if not already_done:
         machines[machine_name] = marchlist
- 
-#for machine_name in machines:
-#    print machine_name + ": " + str(machines[machine_name])
 
-#ok, now we have all the machines we're going to build. next step is to generate a Makefile.am where they're all laid out and compiled
+#get the maximum alignment for all archs in a machine
+machine_alignment_dict = {}
+for machine in machines:
+    machine_alignment_dict[machine] = max((archalign_dict.get(k, 1)) for k in machines[machine])
+
+#for machine in machine_alignment_dict:
+#    print machine + ": %d" % machine_alignment_dict[machine]
 
 taglist = [];
 fcountlist = [];
@@ -295,5 +304,5 @@ outfile_h.close()
 for machine in machines:
     machine_c_filename = os.path.join(gendir, "lib/volk_machine_" + machine + ".c")
     outfile_machine_c = open(machine_c_filename, "w")
-    outfile_machine_c.write(make_each_machine_c(machine, machines[machine], functions, fcountlist, taglist))
+    outfile_machine_c.write(make_each_machine_c(machine, machines[machine], functions, fcountlist, taglist, machine_alignment_dict[machine]))
     outfile_machine_c.close()