Volk: added alignment prop to each machine. call volk_get_alignment() to get your machine's reqd alignment for malloc.

7 files changed, 41 insertions, 11 deletions

diff --git a/volk/gen/archs.xml b/volk/gen/archs.xml
index 977cc7924..0e8c22eea 100644
--- a/volk/gen/archs.xml
+++ b/volk/gen/archs.xml
@@ -7,6 +7,7 @@
 
 <arch name="altivec" type="powerpc">
   <flag>maltivec</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="32" type="x86" no_test="true" >
@@ -31,6 +32,7 @@
   <shift>31</shift>
   <flag>m3dnow</flag>
   <val>1</val>
+  <alignment>8</alignment>
 </arch>
 
 <arch name="abm" type="x86">
@@ -39,6 +41,7 @@
   <reg>d</reg>
   <shift>5</shift>
   <flag>sse4.2</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="popcount" type="x86">
@@ -55,6 +58,7 @@
   <reg>d</reg>
   <shift>23</shift>
   <flag>mmmx</flag>
+  <alignment>8</alignment>
 </arch>
 
 
@@ -66,6 +70,7 @@
   <flag>msse</flag>
   <environment>_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);</environment>
   <include>xmmintrin.h</include>
+  <alignment>16</alignment>
 </arch>
 
 
@@ -75,6 +80,7 @@
   <reg>d</reg>
   <shift>26</shift>
   <flag>msse2</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="orc" type="all">
@@ -91,6 +97,7 @@
   <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 +106,7 @@
   <reg>c</reg>
   <shift>9</shift>
   <flag>mssse3</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="sse4_a" type="x86">
@@ -107,6 +115,7 @@
   <reg>c</reg>
   <shift>6</shift>
   <flag>msse4a</flag>
+  <alignment>16</alignment>
 </arch>
 
 
@@ -116,6 +125,7 @@
   <reg>c</reg>
   <shift>19</shift>
   <flag>msse4.1</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="sse4_2" type="x86">
@@ -124,6 +134,7 @@
   <reg>c</reg>
   <shift>20</shift>
   <flag>msse4.2</flag>
+  <alignment>16</alignment>
 </arch>
 
 <arch name="avx" type="x86">
@@ -132,6 +143,7 @@
   <reg>c</reg>
   <shift>28</shift>
   <flag>mavx</flag>
+  <alignment>32</alignment>
 </arch>
 
 </grammar>
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_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..76f5f8c69 100644
--- a/volk/gen/make_h.py
+++ b/volk/gen/make_h.py
@@ -21,6 +21,9 @@ 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"
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 e87156fc8..76f41205b 100644
--- a/volk/gen/volk_register.py
+++ b/volk/gen/volk_register.py
@@ -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()
diff --git a/volk/lib/qa_utils.cc b/volk/lib/qa_utils.cc
index db606a472..7f86dd78b 100644
--- a/volk/lib/qa_utils.cc
+++ b/volk/lib/qa_utils.cc
@@ -233,7 +233,8 @@ bool icompare(t *in1, t *in2, unsigned int vlen, unsigned int tol) {
 
 class volk_qa_aligned_mem_pool{
 public:
-    void *get_new(size_t size, size_t alignment = 32){
+    void *get_new(size_t size){
+        size_t alignment = volk_get_alignment();
         _mems.push_back(std::vector<char>(size + alignment-1, 0));
         size_t ptr = size_t(&_mems.back().front());
         return (void *)((ptr + alignment-1) & ~(alignment-1));