Enable prefetching in cudf.pandas.install() (rapidsai#16439)

This PR enables `cudf.pandas` managed memory prefetching in
`cudf.pandas.install()`, to ensure that prefetching is enabled for all
methods of enabling `cudf.pandas`.

I also fixed a bug in libcudf's prefetching logic, where it tried to
compute the number of characters in a strings column view even if the
string column view's data is `nullptr`. This errors, so we must avoid
the `chars_size()` call and stop the prefetch attempt early.

cpp/src/column/column_view.cpp

@@ -45,7 +45,10 @@ void prefetch_col_data(ColumnView& col, void const* data_ptr, std::string_view k
         key, data_ptr, col.size() * size_of(col.type()), cudf::get_default_stream());
     } else if (col.type().id() == type_id::STRING) {
       strings_column_view scv{col};
-
+      if (data_ptr == nullptr) {
+        // Do not call chars_size if the data_ptr is nullptr.
+        return;
+      }
       cudf::experimental::prefetch::detail::prefetch_noexcept(
         key,
         data_ptr,

cpp/src/utilities/prefetch.cpp

@@ -51,6 +51,20 @@ cudaError_t prefetch_noexcept(std::string_view key,
                               rmm::cuda_stream_view stream,
                               rmm::cuda_device_id device_id) noexcept
 {
+  // Don't try to prefetch nullptrs or empty data. Sometimes libcudf has column
+  // views that use nullptrs with a nonzero size as an optimization.
+  if (ptr == nullptr) {
+    if (prefetch_config::instance().debug) {
+      std::cerr << "Skipping prefetch of nullptr" << std::endl;
+    }
+    return cudaSuccess;
+  }
+  if (size == 0) {
+    if (prefetch_config::instance().debug) {
+      std::cerr << "Skipping prefetch of size 0" << std::endl;
+    }
+    return cudaSuccess;
+  }
   if (prefetch_config::instance().get(key)) {
     if (prefetch_config::instance().debug) {
       std::cerr << "Prefetching " << size << " bytes for key " << key << " at location " << ptr

python/cudf/cudf/pandas/__init__.py

@@ -7,6 +7,8 @@
 
 import rmm.mr
 
+from cudf._lib import pylibcudf
+
 from .fast_slow_proxy import is_proxy_object
 from .magics import load_ipython_extension
 from .profiler import Profiler
@@ -16,6 +18,19 @@
 
 LOADED = False
 
+_SUPPORTED_PREFETCHES = {
+    "column_view::get_data",
+    "mutable_column_view::get_data",
+    "gather",
+    "hash_join",
+}
+
+
+def _enable_managed_prefetching(rmm_mode):
+    if "managed" in rmm_mode:
+        for key in _SUPPORTED_PREFETCHES:
+            pylibcudf.experimental.enable_prefetching(key)
+
 
 def install():
     """Enable Pandas Accelerator Mode."""
@@ -33,7 +48,7 @@ def install():
             f"cudf.pandas detected an already configured memory resource, ignoring 'CUDF_PANDAS_RMM_MODE'={str(rmm_mode)}",
             UserWarning,
         )
-        return rmm_mode
+        return
 
     free_memory, _ = rmm.mr.available_device_memory()
     free_memory = int(round(float(free_memory) * 0.80 / 256) * 256)
@@ -57,7 +72,7 @@ def install():
     elif rmm_mode != "cuda":
         raise ValueError(f"Unsupported {rmm_mode=}")
     rmm.mr.set_current_device_resource(new_mr)
-    return rmm_mode
+    _enable_managed_prefetching(rmm_mode)
 
 
 def pytest_load_initial_conftests(early_config, parser, args):

python/cudf/cudf/pandas/__main__.py

@@ -72,17 +72,7 @@ def main():
 
     args = parser.parse_args()
 
-    rmm_mode = install()
-    if "managed" in rmm_mode:
-        for key in {
-            "column_view::get_data",
-            "mutable_column_view::get_data",
-            "gather",
-            "hash_join",
-        }:
-            from cudf._lib import pylibcudf
-
-            pylibcudf.experimental.enable_prefetching(key)
+    install()
     with profile(args.profile, args.line_profile, args.args[0]) as fn:
         args.args[0] = fn
         if args.module: