Merge pull request #534 from jvesely/opencl_extensions

Check for available OpenCL extensions
Theano · Sep 19, 2017 · ace45c6 · ace45c6
2 parents 4461ad8 + 4285e20
commit ace45c6
Show file tree

Hide file tree

Showing 2 changed files with 460 additions and 433 deletions.
diff --git a/src/cluda_opencl.h b/src/cluda_opencl.h
@@ -33,7 +33,10 @@
 #define ga_long long
 #define ga_ulong ulong
 #define ga_float float
+#ifdef cl_khr_fp64
+#pragma OPENCL EXTENSION cl_khr_fp64: enable
 #define ga_double double
+#endif
 #define ga_size ulong
 #define ga_ssize long
 #define GA_DECL_SHARED_PARAM(type, name) , __local type *name
@@ -109,6 +112,14 @@ static inline ga_half ga_float2half(ga_float f) {
 #define atom_add_Il(a, b) atomic_add(a, b)
 #define atom_xchg_Ig(a, b) atomic_xchg(a, b)
 #define atom_xchg_Il(a, b) atomic_xchg(a, b)
+/* ga_float */
+gen_atom32_add(atom_add_fg, ga_float, global)
+gen_atom32_add(atom_add_fl, ga_float, local)
+#define atom_xchg_fg(a, b) atomic_xchg(a, b)
+#define atom_xchg_fl(a, b) atomic_xchg(a, b)
+
+#ifdef cl_khr_int64_base_atomics
+#pragma OPENCL EXTENSION cl_khr_int64_base_atomics: enable
 /* ga_long */
 #define atom_add_lg(a, b) atom_add(a, b)
 #define atom_add_ll(a, b) atom_add(a, b)
@@ -119,16 +130,14 @@ static inline ga_half ga_float2half(ga_float f) {
 #define atom_add_Ll(a, b) atom_add(a, b)
 #define atom_xchg_Lg(a, b) atom_xchg(a, b)
 #define atom_xchg_Ll(a, b) atom_xchg(a, b)
-/* ga_float */
-gen_atom32_add(atom_add_fg, ga_float, global)
-gen_atom32_add(atom_add_fl, ga_float, local)
-#define atom_xchg_fg(a, b) atomic_xchg(a, b)
-#define atom_xchg_fl(a, b) atomic_xchg(a, b)
 /* ga_double */
+#ifdef cl_khr_fp64
 gen_atom64_add(atom_add_dg, ga_double, global)
 gen_atom64_add(atom_add_dl, ga_double, local)
 gen_atom64_xchg(atom_xchg_dg, ga_double, global)
 gen_atom64_xchg(atom_xchg_dl, ga_double, local)
+#endif
+#endif
 /* ga_half */
 #define gen_atomh_add(name, aspace)                                     \
   ga_half name(volatile aspace ga_half *addr, ga_half val);             \