[CUDA][LIBCLC] Fix fence impl to satisfy SYCL 2020. (#12713)

- make relaxed fence a no op to satisfy the SYCL spec.
- make acquire/release/acq_rel use the lighter acq_rel fence for sm_70
instead of the seq_cst fence.

---------

Signed-off-by: JackAKirk <jack.kirk@codeplay.com>

libclc/ptx-nvidiacl/libspirv/synchronization/barrier.cl

@@ -9,14 +9,36 @@
 #include <spirv/spirv.h>
 #include <spirv/spirv_types.h>
 
+int __clc_nvvm_reflect_arch();
+
 _CLC_OVERLOAD _CLC_DEF void __spirv_MemoryBarrier(unsigned int memory,
                                                   unsigned int semantics) {
-  if (memory == CrossDevice) {
-    __nvvm_membar_sys();
-  } else if (memory == Device) {
-    __nvvm_membar_gl();
-  } else {
-    __nvvm_membar_cta();
+
+  // for sm_70 and above membar becomes semantically identical to fence.sc.
+  // However sm_70 and above also introduces a lightweight fence.acq_rel that
+  // can be used to form either acquire or release strong operations.
+  // Consult
+  // https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-membar-fence
+  // for details.
+
+  unsigned int order = semantics & 0x1F;
+  if (__clc_nvvm_reflect_arch() < 700 ||
+             order == SequentiallyConsistent) {
+    if (memory == CrossDevice) {
+      __nvvm_membar_sys();
+    } else if (memory == Device) {
+      __nvvm_membar_gl();
+    } else {
+      __nvvm_membar_cta();
+    }
+  } else if (order != None) {
+    if (memory == CrossDevice) {
+      __asm__ __volatile__("fence.acq_rel.sys;");
+    } else if (memory == Device) {
+      __asm__ __volatile__("fence.acq_rel.gpu;");
+    } else {
+      __asm__ __volatile__("fence.acq_rel.cta;");
+    }
   }
 }