Statistics cleanup

cpp/include/cudf/detail/utilities/device_operators.cuh

@@ -14,8 +14,7 @@
  * limitations under the License.
  */
 
-#ifndef DEVICE_OPERATORS_CUH
-#define DEVICE_OPERATORS_CUH
+#pragma once
 
 /**
  * @brief definition of the device operators
@@ -24,6 +23,7 @@
 
 #include <cudf/fixed_point/fixed_point.hpp>
 #include <cudf/scalar/scalar.hpp>
+#include <cudf/strings/string_view.cuh>
 #include <cudf/types.hpp>
 #include <cudf/utilities/error.hpp>
 #include <cudf/utilities/traits.hpp>
@@ -84,16 +84,6 @@ struct DeviceCount {
   }
 };
 
-/**
- * @brief string value for sentinel which is used in min, max reduction
- * operators
- * This sentinel string value is the highest possible valid UTF-8 encoded
- * character. This serves as identity value for maximum operator on string
- * values. Also, this char pointer serves as valid device pointer of identity
- * value for minimum operator on string values.
- */
-__constant__ char max_string_sentinel[5]{"\xF7\xBF\xBF\xBF"};
-
 /* @brief binary `min` operator */
 struct DeviceMin {
   template <typename T>
@@ -123,13 +113,7 @@ struct DeviceMin {
             typename std::enable_if_t<std::is_same<T, cudf::string_view>::value>* = nullptr>
   CUDA_HOST_DEVICE_CALLABLE static constexpr T identity()
   {
-    const char* psentinel{nullptr};
-#if defined(__CUDA_ARCH__)
-    psentinel = &max_string_sentinel[0];
-#else
-    CUDA_TRY(cudaGetSymbolAddress((void**)&psentinel, max_string_sentinel));
-#endif
-    return T(psentinel, 4);
+    return string_view::max();
   }
 
   template <typename T, typename std::enable_if_t<cudf::is_dictionary<T>()>* = nullptr>
@@ -167,13 +151,7 @@ struct DeviceMax {
             typename std::enable_if_t<std::is_same<T, cudf::string_view>::value>* = nullptr>
   CUDA_HOST_DEVICE_CALLABLE static constexpr T identity()
   {
-    const char* psentinel{nullptr};
-#if defined(__CUDA_ARCH__)
-    psentinel = &max_string_sentinel[0];
-#else
-    CUDA_TRY(cudaGetSymbolAddress((void**)&psentinel, max_string_sentinel));
-#endif
-    return T(psentinel, 0);
+    return string_view::min();
   }
 
   template <typename T, typename std::enable_if_t<cudf::is_dictionary<T>()>* = nullptr>
@@ -242,5 +220,3 @@ struct DeviceLeadLag {
 };
 
 }  // namespace cudf
-
-#endif

cpp/include/cudf/detail/utilities/int_fastdiv.h

@@ -14,8 +14,7 @@
  *  limitations under the License.
  */
 
-#ifndef _INT_FASTDIV_KJGIUHFG
-#define _INT_FASTDIV_KJGIUHFG
+#pragma once
 
 class int_fastdiv {
  public:
@@ -172,5 +171,3 @@ __host__ __device__ __forceinline__ int operator%(const unsigned char n, const i
 {
   return ((int)n) % divisor;
 }
-
-#endif

cpp/include/cudf/strings/string.cuh

@@ -105,6 +105,7 @@ __device__ bool is_float(string_view const& d_str)
   }
   return result;
 }
+
 /** @} */  // end of group
 }  // namespace string
 }  // namespace strings

cpp/include/cudf/strings/string_view.cuh

@@ -17,6 +17,7 @@
 #pragma once
 
 #include <cudf/strings/string_view.hpp>
+#include <cudf/utilities/error.hpp>
 
 #include <thrust/count.h>
 #include <thrust/find.h>
@@ -43,9 +44,50 @@ __device__ inline size_type characters_in_string(const char* str, size_type byte
   return thrust::count_if(
     thrust::seq, ptr, ptr + bytes, [](uint8_t chr) { return is_begin_utf8_char(chr); });
 }
+
+/**
+ * @brief string value for sentinel which is used in min, max reduction
+ * operators
+ *
+ * This sentinel string value is the highest possible valid UTF-8 encoded
+ * character. This serves as identity value for maximum operator on string
+ * values. Also, this char pointer serves as valid device pointer of identity
+ * value for minimum operator on string values.
+ */
+static __constant__ char max_string_sentinel[5]{"\xF7\xBF\xBF\xBF"};
 }  // namespace detail
 }  // namespace strings
 
+/**
+ * @brief Return minimum value associated with the string type
+ *
+ * This function is needed to be host callable because it is called by a host
+ * calleable function DeviceMax::identity<string_view>()
+ *
+ * @return An empty string
+ */
+CUDA_HOST_DEVICE_CALLABLE string_view string_view::min() { return string_view(); }
+
+/**
+ * @brief Return maximum value associated with the string type
+ *
+ * This function is needed to be host callable because it is called by a host
+ * calleable function DeviceMin::identity<string_view>()
+ *
+ * @return A string value which represents the highest possible valid UTF-8 encoded
+ * character.
+ */
+CUDA_HOST_DEVICE_CALLABLE string_view string_view::max()
+{
+  const char* psentinel{nullptr};
+#if defined(__CUDA_ARCH__)
+  psentinel = &cudf::strings::detail::max_string_sentinel[0];
+#else
+  CUDA_TRY(cudaGetSymbolAddress((void**)&psentinel, cudf::strings::detail::max_string_sentinel));
+#endif
+  return string_view(psentinel, 4);
+}
+
 __device__ inline size_type string_view::length() const
 {
   if (_length == UNKNOWN_STRING_LENGTH)

cpp/include/cudf/strings/string_view.hpp

@@ -290,6 +290,21 @@ class string_view {
    */
   CUDA_DEVICE_CALLABLE string_view substr(size_type start, size_type length) const;
 
+  /**
+   * @brief Return minimum value associated with the string type
+   *
+   * @return An empty string
+   */
+  CUDA_HOST_DEVICE_CALLABLE static string_view min();
+
+  /**
+   * @brief Return maximum value associated with the string type
+   *
+   * @return A string that represents the highest possible valid UTF-8 encoded
+   * character.
+   */
+  CUDA_HOST_DEVICE_CALLABLE static string_view max();
+
   /**
    * @brief Default constructor represents an empty string.
    */

cpp/src/io/orc/stats_enc.cu

@@ -51,7 +51,7 @@ __global__ void __launch_bounds__(init_threads_per_block)
   uint32_t t              = threadIdx.x;
   statistics_group *group = &group_g[threadIdx.y];
   if (chunk_id < num_rowgroups and t == 0) {
-    uint32_t num_rows = cols[col_id].num_rows;
+    uint32_t num_rows = cols[col_id].leaf_column->size();
     group->col        = &cols[col_id];
     group->start_row  = chunk_id * row_index_stride;
     group->num_rows = min(num_rows - min(chunk_id * row_index_stride, num_rows), row_index_stride);

cpp/src/io/orc/writer_impl.cu

@@ -21,6 +21,8 @@
 
 #include "writer_impl.hpp"
 
+#include <io/utilities/column_utils.cuh>
+
 #include <cudf/null_mask.hpp>
 #include <cudf/strings/strings_column_view.hpp>
 
@@ -38,6 +40,7 @@ namespace detail {
 namespace orc {
 using namespace cudf::io::orc;
 using namespace cudf::io;
+using cudf::io::orc::gpu::nvstrdesc_s;
 
 struct row_group_index_info {
   int32_t pos       = -1;  // Position
@@ -725,6 +728,7 @@ std::vector<StripeInformation> writer::impl::gather_stripes(
 }
 
 std::vector<std::vector<uint8_t>> writer::impl::gather_statistic_blobs(
+  const table_device_view &table,
   orc_column_view const *columns,
   size_t num_columns,
   size_t num_rows,
@@ -787,6 +791,10 @@ std::vector<std::vector<uint8_t>> writer::impl::gather_statistic_blobs(
   }
   stat_desc.host_to_device(stream);
   stat_merge.host_to_device(stream);
+
+  rmm::device_uvector<column_device_view> leaf_column_views =
+    cudf::io::create_leaf_column_device_views(stat_desc, table, stream);
+
   gpu::orc_init_statistics_groups(stat_groups.data().get(),
                                   stat_desc.device_ptr(),
                                   num_columns,
@@ -1019,6 +1027,8 @@ void writer::impl::write(table_view const &table)
       "be specified");
   }
 
+  auto device_columns = table_device_view::create(table);
+
   // Wrapper around cudf columns to attach ORC-specific type info
   std::vector<orc_column_view> orc_columns;
   orc_columns.reserve(num_columns);  // Avoids unnecessary re-allocation
@@ -1116,8 +1126,14 @@ void writer::impl::write(table_view const &table)
   // Gather column statistics
   std::vector<std::vector<uint8_t>> column_stats;
   if (enable_statistics_ && num_columns > 0 && num_rows > 0) {
-    column_stats = gather_statistic_blobs(
-      orc_columns.data(), num_columns, num_rows, num_rowgroups, stripe_list, stripes, chunks);
+    column_stats = gather_statistic_blobs(*device_columns,
+                                          orc_columns.data(),
+                                          num_columns,
+                                          num_rows,
+                                          num_rowgroups,
+                                          stripe_list,
+                                          stripes,
+                                          chunks);
   }
 
   // Allocate intermediate output stream buffer

cpp/src/io/orc/writer_impl.hpp

@@ -26,6 +26,7 @@
 #include <cudf/io/detail/orc.hpp>
 #include <cudf/io/orc.hpp>
 #include <cudf/table/table.hpp>
+#include <cudf/table/table_device_view.cuh>
 #include <cudf/utilities/error.hpp>
 
 #include <rmm/cuda_stream_view.hpp>
@@ -215,6 +216,7 @@ class writer::impl {
    * @brief Returns per-stripe and per-file column statistics encoded
    * in ORC protobuf format
    *
+   * @param table Table information to be written
    * @param columns List of columns
    * @param num_columns Total number of columns
    * @param num_rows Total number of rows
@@ -226,6 +228,7 @@ class writer::impl {
    * @return The statistic blobs
    */
   std::vector<std::vector<uint8_t>> gather_statistic_blobs(
+    const table_device_view& table,
     orc_column_view const* columns,
     size_t num_columns,
     size_t num_rows,

cpp/src/io/parquet/page_enc.cu

@@ -1961,41 +1961,6 @@ void InitPageFragments(PageFragment *frag,
     frag, col_desc, num_fragments, num_columns, fragment_size, num_rows);
 }
 
-/**
- * @copydoc void init_column_device_views(EncColumnDesc *col_desc,
- *            column_device_view *leaf_column_views,
- *            const table_device_view &parent_table_device_view,
- *            rmm::cuda_stream_view stream)
- */
-void init_column_device_views(EncColumnDesc *col_desc,
-                              column_device_view *leaf_column_views,
-                              const table_device_view &parent_column_table_device_view,
-                              rmm::cuda_stream_view stream)
-{
-  cudf::detail::device_single_thread(
-    [col_desc,
-     parent_col_view = parent_column_table_device_view,
-     leaf_column_views] __device__() mutable {
-      for (size_type i = 0; i < parent_col_view.num_columns(); ++i) {
-        column_device_view col = parent_col_view.column(i);
-        if (col.type().id() == type_id::LIST) {
-          col_desc[i].parent_column = parent_col_view.begin() + i;
-        } else {
-          col_desc[i].parent_column = nullptr;
-        }
-        // traverse till leaf column
-        while (col.type().id() == type_id::LIST) {
-          col = col.child(lists_column_view::child_column_index);
-        }
-        // Store leaf_column to device storage
-        column_device_view *leaf_col_ptr = leaf_column_views + i;
-        *leaf_col_ptr                    = col;
-        col_desc[i].leaf_column          = leaf_col_ptr;
-      }
-    },
-    stream);
-}
-
 /**
  * @brief Launches kernel for initializing fragment statistics groups
  *

cpp/src/io/parquet/parquet_gpu.hpp

@@ -229,9 +229,6 @@ struct EncColumnDesc : stats_column_desc {
   size_type const *level_offsets;  //!< Offset array for per-row pre-calculated rep/def level values
   uint8_t const *rep_values;       //!< Pre-calculated repetition level values
   uint8_t const *def_values;       //!< Pre-calculated definition level values
-
-  column_device_view *leaf_column;    //!< Pointer to leaf column
-  column_device_view *parent_column;  //!< Pointer to parent column. Is nullptr if not list type.
 };
 
 constexpr int max_page_fragment_size = 5000;  //!< Max number of rows in a page fragment
@@ -448,19 +445,6 @@ void InitPageFragments(PageFragment *frag,
                        uint32_t num_rows,
                        rmm::cuda_stream_view stream);
 
-/**
- * @brief Set column_device_view pointers in column description array
- *
- * @param[out] col_desc Column description array [column_id]
- * @param[out] leaf_column_views Device array to store leaf columns
- * @param[in] parent_table_device_view Table device view containing parent columns
- * @param[in] stream CUDA stream to use, default 0
- */
-void init_column_device_views(EncColumnDesc *col_desc,
-                              column_device_view *leaf_column_views,
-                              const table_device_view &parent_table_device_view,
-                              rmm::cuda_stream_view stream);
-
 /**
  * @brief Launches kernel for initializing fragment statistics groups
  *

cpp/src/io/parquet/writer_impl.cu

@@ -22,6 +22,7 @@
 #include "writer_impl.hpp"
 
 #include <io/parquet/compact_protocol_writer.hpp>
+#include <io/utilities/column_utils.cuh>
 
 #include <cudf/column/column_device_view.cuh>
 #include <cudf/lists/lists_column_view.hpp>
@@ -495,17 +496,6 @@ class parquet_column_view {
   uint8_t _decimal_precision = 0;
 };
 
-rmm::device_uvector<column_device_view> writer::impl::create_leaf_column_device_views(
-  hostdevice_vector<gpu::EncColumnDesc> &col_desc,
-  const table_device_view &parent_table_device_view)
-{
-  rmm::device_uvector<column_device_view> leaf_column_views(parent_table_device_view.num_columns(),
-                                                            stream);
-  gpu::init_column_device_views(
-    col_desc.device_ptr(), leaf_column_views.data(), parent_table_device_view, stream);
-  return leaf_column_views;
-}
-
 void writer::impl::init_page_fragments(hostdevice_vector<gpu::PageFragment> &frag,
                                        hostdevice_vector<gpu::EncColumnDesc> &col_desc,
                                        uint32_t num_columns,
@@ -919,7 +909,8 @@ void writer::impl::write(table_view const &table)
   if (fragments.size() != 0) {
     // Move column info to device
     col_desc.host_to_device(stream);
-    leaf_column_views = create_leaf_column_device_views(col_desc, *parent_column_table_device_view);
+    leaf_column_views =
+      cudf::io::create_leaf_column_device_views(col_desc, *parent_column_table_device_view, stream);
 
     init_page_fragments(fragments, col_desc, num_columns, num_fragments, num_rows, fragment_size);
   }

cpp/src/io/parquet/writer_impl.hpp

@@ -136,18 +136,6 @@ class writer::impl {
                            uint32_t num_rows,
                            uint32_t fragment_size);
 
-  /**
-   * @brief Create column_device_view pointers from leaf columns
-   *
-   * @param col_desc Column description array
-   * @param parent_table_device_view Table device view containing parent columns
-   *
-   * @return Device array containing leaf column device views
-   */
-  rmm::device_uvector<column_device_view> create_leaf_column_device_views(
-    hostdevice_vector<gpu::EncColumnDesc>& col_desc,
-    const table_device_view& parent_table_device_view);
-
   /**
    * @brief Gather per-fragment statistics
    *