Merge branch 'fix_ldg_error' into '0.6.15'

add __ldg fun definition

See merge request aicomponent/torch-sparce!5

csrc/cuda/utils.cuh

@@ -16,3 +16,16 @@ __device__ __inline__ at::Half __shfl_down_sync(const unsigned mask,
                                                 const unsigned int delta) {
   return __shfl_down_sync(mask, var.operator __half(), delta);
 }
+
+#ifdef USE_ROCM
+__device__ __inline__ at::Half __ldg(const at::Half* ptr) {
+  return __ldg(reinterpret_cast<const __half*>(ptr));
+}
+#define SHFL_UP_SYNC(mask, var, delta) __shfl_up(var, delta)
+#define SHFL_DOWN_SYNC(mask, var, delta) __shfl_down(var, delta)
+#define SHFL_SYNC(mask, var, delta) __shfl(var, delta)
+#else
+#define SHFL_UP_SYNC __shfl_up_sync
+#define SHFL_DOWN_SYNC __shfl_down_sync
+#define SHFL_SYNC __shfl_sync
+#endif