[PATCH 1011/1017] add pocl_get_distro_cpu_name()

in distro builds, map each kernellib_name to a specific llvm_cpu
which is then used as compilation target instead of the native cpu

this limits the possible code generation targets in distro builds
in order to be able to run tests against all of them with a minimal
number of different (physical) CPU models

Gbp-Pq: Name 1011-add-pocl_get_distro_cpu_name.patch

diff --git a/lib/CL/pocl_llvm.h b/lib/CL/pocl_llvm.h
index 2f5f625f990aa966d2693d4cdd4238130abb4362..be5a1a0804f98643fcd227f955f8b4e67abb5201 100644 (file)
--- a/lib/CL/pocl_llvm.h
+++ b/lib/CL/pocl_llvm.h
@@ -42,6 +42,10 @@ extern "C" {
   POCL_EXPORT
   const char *pocl_get_distro_kernellib_name ();
 
+  /* For distro builds, return the target cpu name for a kernellib name */
+  POCL_EXPORT
+  const char *pocl_get_distro_cpu_name (const char *kernellib_name);
+
   /* Returns if the cpu supports FMA instruction (uses LLVM). */
   int cpu_has_fma ();
 

diff --git a/lib/CL/pocl_llvm_utils.cc b/lib/CL/pocl_llvm_utils.cc
index 857773f311e0a332026ed5083428e3f5edcd097e..7bdfb387804b8a06a9d78da018b1125d71cd5ebb 100644 (file)
--- a/lib/CL/pocl_llvm_utils.cc
+++ b/lib/CL/pocl_llvm_utils.cc
@@ -146,29 +146,39 @@ char *pocl_get_llvm_cpu_name() {
 #ifdef KERNELLIB_HOST_DISTRO_VARIANTS
 const struct kernellib_features {
   const char *kernellib_name;
+  const char *cpu_name;
   const char *features[12];
 } kernellib_feature_map[] = {
 // order the entries s.t. if a cpu matches multiple entries, the "best" match
 // comes last
 #if defined(__x86_64__)
     "sse2",
+    "athlon64",
     {"sse2", NULL},
     "ssse3",
+    "core2",
     {"sse2", "ssse3", "cx16", NULL},
     "sse41",
+    "penryn",
     {"sse2", "sse4.1", "cx16", NULL},
     "avx",
+    "sandybridge",
     {"sse2", "avx", "cx16", "popcnt", NULL},
     "avx_f16c",
+    "ivybridge",
     {"sse2", "avx", "cx16", "popcnt", "f16c", NULL},
     "avx_fma4",
+    "bdver1",
     {"sse2", "avx", "cx16", "popcnt", "xop", "fma4", NULL},
     "avx2",
+    "haswell",
     {"sse2", "avx", "avx2", "cx16", "popcnt", "lzcnt", "f16c", "fma", "bmi",
      "bmi2", NULL},
     "avx512",
+    "skylake-avx512",
     {"sse2", "avx512f", NULL},
 #endif
+    NULL,
     NULL,
     {NULL}};
 
@@ -196,6 +206,25 @@ const char *pocl_get_distro_kernellib_name() {
 
   return best_match->kernellib_name;
 }
+
+/* for "distro" style kernel libs, return which target cpu to use for a given
+ * kernellib */
+const char *pocl_get_distro_cpu_name(const char *kernellib_name) {
+  StringMap<bool> Features;
+
+  if (!llvm::sys::getHostCPUFeatures(Features))
+    POCL_ABORT("LLVM can't get host CPU flags!\n");
+
+  const kernellib_features *best_match = NULL;
+  for (const kernellib_features *kf = kernellib_feature_map; kf->kernellib_name;
+       ++kf) {
+    if (!strcmp(kernellib_name, kf->kernellib_name))
+      return kf->cpu_name;
+  }
+
+  POCL_ABORT("Can't find a cpu name matching the kernellib (%s)\n",
+             kernellib_name);
+}
 #endif
 
 int bitcode_is_triple(const char *bitcode, size_t size, const char *triple) {