Merge branch 'branch-24.10' into mixed_semi_join_refactor

.gitignore

@@ -80,7 +80,6 @@ build/
 cpp/build/
 cpp/examples/*/install/
 cpp/examples/*/build/
-cpp/examples/tpch/datagen/datafusion
 cpp/include/cudf/ipc_generated/*.h
 cpp/thirdparty/googletest/
 

README.md

@@ -79,7 +79,7 @@ pip install --extra-index-url=https://pypi.nvidia.com cudf-cu12
 
 ### Conda
 
-cuDF can be installed with conda (via [miniconda](https://docs.conda.io/projects/miniconda/en/latest/) or the full [Anaconda distribution](https://www.anaconda.com/download) from the `rapidsai` channel:
+cuDF can be installed with conda (via [miniforge](https://github.com/conda-forge/miniforge)) from the `rapidsai` channel:
 
 ```bash
 conda install -c rapidsai -c conda-forge -c nvidia \

cpp/CMakeLists.txt

@@ -384,6 +384,7 @@ add_library(
   src/io/json/nested_json_gpu.cu
   src/io/json/read_json.cu
   src/io/json/parser_features.cpp
+  src/io/json/process_tokens.cu
   src/io/json/write_json.cu
   src/io/orc/aggregate_orc_metadata.cpp
   src/io/orc/dict_enc.cu

cpp/benchmarks/CMakeLists.txt

@@ -36,25 +36,25 @@ target_include_directories(
 )
 
 add_library(
-  tpch_data_generator STATIC
-  common/tpch_data_generator/tpch_data_generator.cpp common/tpch_data_generator/table_helpers.cpp
-  common/tpch_data_generator/random_column_generator.cu
+  ndsh_data_generator STATIC
+  common/ndsh_data_generator/ndsh_data_generator.cpp common/ndsh_data_generator/table_helpers.cpp
+  common/ndsh_data_generator/random_column_generator.cu
 )
-target_compile_features(tpch_data_generator PUBLIC cxx_std_17 cuda_std_17)
+target_compile_features(ndsh_data_generator PUBLIC cxx_std_17 cuda_std_17)
 
 target_compile_options(
-  tpch_data_generator PUBLIC "$<$<COMPILE_LANGUAGE:CXX>:${CUDF_CXX_FLAGS}>"
+  ndsh_data_generator PUBLIC "$<$<COMPILE_LANGUAGE:CXX>:${CUDF_CXX_FLAGS}>"
                              "$<$<COMPILE_LANGUAGE:CUDA>:${CUDF_CUDA_FLAGS}>"
 )
 
 target_link_libraries(
-  tpch_data_generator
+  ndsh_data_generator
   PUBLIC cudf cudftestutil nvtx3::nvtx3-cpp
   PRIVATE $<TARGET_NAME_IF_EXISTS:conda_env>
 )
 
 target_include_directories(
-  tpch_data_generator
+  ndsh_data_generator
   PUBLIC "$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>" "$<BUILD_INTERFACE:${CUDF_SOURCE_DIR}>"
          "$<BUILD_INTERFACE:${CUDF_SOURCE_DIR}/src>"
 )
@@ -127,8 +127,8 @@ function(ConfigureNVBench CMAKE_BENCH_NAME)
                INSTALL_RPATH "\$ORIGIN/../../../lib"
   )
   target_link_libraries(
-    ${CMAKE_BENCH_NAME} PRIVATE cudf_benchmark_common cudf_datagen nvbench::nvbench
-                                $<TARGET_NAME_IF_EXISTS:conda_env>
+    ${CMAKE_BENCH_NAME} PRIVATE cudf_benchmark_common ndsh_data_generator cudf_datagen
+                                nvbench::nvbench $<TARGET_NAME_IF_EXISTS:conda_env>
   )
   install(
     TARGETS ${CMAKE_BENCH_NAME}
@@ -175,6 +175,14 @@ ConfigureBench(COPY_IF_ELSE_BENCH copying/copy_if_else.cpp)
 # * transpose benchmark ---------------------------------------------------------------------------
 ConfigureBench(TRANSPOSE_BENCH transpose/transpose.cpp)
 
+# ##################################################################################################
+# * nds-h benchmark --------------------------------------------------------------------------------
+ConfigureNVBench(NDSH_Q1 ndsh/q01.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q5 ndsh/q05.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q6 ndsh/q06.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q9 ndsh/q09.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q10 ndsh/q10.cpp ndsh/utilities.cpp)
+
 # ##################################################################################################
 # * stream_compaction benchmark -------------------------------------------------------------------
 ConfigureNVBench(

cpp/benchmarks/common/tpch_data_generator/tpch_data_generator.cpp → cpp/benchmarks/common/ndsh_data_generator/ndsh_data_generator.cpp

@@ -14,7 +14,7 @@
  * limitations under the License.
  */
 
-#include "tpch_data_generator.hpp"
+#include "ndsh_data_generator.hpp"
 
 #include "random_column_generator.hpp"
 #include "table_helpers.hpp"
@@ -435,46 +435,37 @@ std::unique_ptr<cudf::table> generate_lineitem_partial(cudf::table_view const& o
   columns.push_back(std::move(l_quantity));
   columns.push_back(std::move(l_discount));
   columns.push_back(std::move(l_tax));
+  columns.push_back(std::move(l_returnflag));
+  columns.push_back(std::move(l_linestatus));
   columns.push_back(std::move(l_shipdate_ts));
   columns.push_back(std::move(l_commitdate_ts));
   columns.push_back(std::move(l_receiptdate_ts));
-  columns.push_back(std::move(l_returnflag));
-  columns.push_back(std::move(l_linestatus));
   columns.push_back(std::move(l_shipinstruct));
   columns.push_back(std::move(l_shipmode));
   columns.push_back(std::move(l_comment));
   return std::make_unique<cudf::table>(std::move(columns));
 }
 
-std::unique_ptr<cudf::table> generate_orders_dependent(cudf::table_view const& lineitem,
+/**
+ * @brief Generate the part of the `orders` table dependent on the `lineitem` table
+ *
+ * @param lineitem_partial The partially generated `lineitem` table
+ * @param stream CUDA stream used for device memory operations and kernel launches
+ * @param mr Device memory resource used to allocate the returned column's device memory
+ */
+std::unique_ptr<cudf::table> generate_orders_dependent(cudf::table_view const& lineitem_partial,
                                                        rmm::cuda_stream_view stream,
                                                        rmm::device_async_resource_ref mr)
 {
   CUDF_FUNC_RANGE();
-  auto const l_linestatus_mask = lineitem.column(0);
-  auto const l_orderkey        = lineitem.column(1);
-  auto const l_discount        = lineitem.column(6);
-  auto const l_tax             = lineitem.column(7);
-  auto const l_extendedprice   = lineitem.column(16);
+  auto const l_linestatus_mask = lineitem_partial.column(0);
+  auto const l_orderkey        = lineitem_partial.column(1);
+  auto const l_extendedprice   = lineitem_partial.column(6);
+  auto const l_discount        = lineitem_partial.column(7);
+  auto const l_tax             = lineitem_partial.column(8);
 
   std::vector<std::unique_ptr<cudf::column>> orders_dependent_columns;
 
-  // Generate the `o_totalprice` column
-  // We calculate the `charge` column, which is a function of `l_extendedprice`,
-  // `l_tax`, and `l_discount` and then group by `l_orderkey` and sum the `charge`
-  auto const l_charge = calculate_charge(l_extendedprice, l_tax, l_discount, stream, mr);
-  auto o_totalprice   = [&]() {
-    auto const keys = cudf::table_view({l_orderkey});
-    cudf::groupby::groupby gb(keys);
-    std::vector<cudf::groupby::aggregation_request> requests;
-    requests.push_back(cudf::groupby::aggregation_request());
-    requests[0].aggregations.push_back(cudf::make_sum_aggregation<cudf::groupby_aggregation>());
-    requests[0].values = l_charge->view();
-    auto agg_result    = gb.aggregate(requests);
-    return cudf::round(agg_result.second[0].results[0]->view(), 2);
-  }();
-  orders_dependent_columns.push_back(std::move(o_totalprice));
-
   // Generate the `o_orderstatus` column
   auto o_orderstatus = [&]() {
     auto const keys = cudf::table_view({l_orderkey});
@@ -529,6 +520,22 @@ std::unique_ptr<cudf::table> generate_orders_dependent(cudf::table_view const& l
       cudf::string_scalar("P"), o_orderstatus_intermediate->view(), mask_b->view());
   }();
   orders_dependent_columns.push_back(std::move(o_orderstatus));
+
+  // Generate the `o_totalprice` column
+  // We calculate the `charge` column, which is a function of `l_extendedprice`,
+  // `l_tax`, and `l_discount` and then group by `l_orderkey` and sum the `charge`
+  auto const l_charge = calculate_charge(l_extendedprice, l_tax, l_discount, stream, mr);
+  auto o_totalprice   = [&]() {
+    auto const keys = cudf::table_view({l_orderkey});
+    cudf::groupby::groupby gb(keys);
+    std::vector<cudf::groupby::aggregation_request> requests;
+    requests.push_back(cudf::groupby::aggregation_request());
+    requests[0].aggregations.push_back(cudf::make_sum_aggregation<cudf::groupby_aggregation>());
+    requests[0].values = l_charge->view();
+    auto agg_result    = gb.aggregate(requests);
+    return cudf::round(agg_result.second[0].results[0]->view(), 2);
+  }();
+  orders_dependent_columns.push_back(std::move(o_totalprice));
   return std::make_unique<cudf::table>(std::move(orders_dependent_columns));
 }
 
@@ -730,9 +737,7 @@ generate_orders_lineitem_part(double scale_factor,
   // Generate the `part` table
   auto part = generate_part(scale_factor, stream, mr);
 
-  // Join the `part` and partial `lineitem` tables, then calculate the `l_extendedprice` column,
-  // add the column to the `lineitem` table, and write the `lineitem` table to a parquet file
-
+  // Join the `part` and partial `lineitem` tables, then calculate the `l_extendedprice` column
   auto l_extendedprice = [&]() {
     auto const left = cudf::table_view(
       {lineitem_partial->get_column(2).view(), lineitem_partial->get_column(5).view()});
@@ -752,8 +757,9 @@ generate_orders_lineitem_part(double scale_factor,
     return cudf::round(col->view(), 2);
   }();
 
+  // Insert the `l_extendedprice` column into the partial columns of the `lineitem` table
   auto lineitem_partial_columns = lineitem_partial->release();
-  lineitem_partial_columns.push_back(std::move(l_extendedprice));
+  lineitem_partial_columns.insert(lineitem_partial_columns.begin() + 6, std::move(l_extendedprice));
   auto lineitem_temp = std::make_unique<cudf::table>(std::move(lineitem_partial_columns));
 
   // Generate the dependent columns of the `orders` table
@@ -762,7 +768,7 @@ generate_orders_lineitem_part(double scale_factor,
 
   auto orders_independent_columns = orders_independent->release();
   auto orders_dependent_columns   = orders_dependent->release();
-  orders_independent_columns.insert(orders_independent_columns.end(),
+  orders_independent_columns.insert(orders_independent_columns.begin() + 2,
                                     std::make_move_iterator(orders_dependent_columns.begin()),
                                     std::make_move_iterator(orders_dependent_columns.end()));
 

cpp/benchmarks/hashing/hash.cpp

@@ -50,7 +50,7 @@ static void bench_hash(nvbench::state& state)
   state.add_global_memory_reads<nvbench::int64_t>(num_rows);
   // add memory read from bitmaks
   if (!no_nulls) {
-    state.add_global_memory_reads<nvbench::int8_t>(2 *
+    state.add_global_memory_reads<nvbench::int8_t>(2L *
                                                    cudf::bitmask_allocation_size_bytes(num_rows));
   }
   // memory written depends on used hash
@@ -63,37 +63,37 @@ static void bench_hash(nvbench::state& state)
     });
   } else if (hash_name == "md5") {
     // md5 creates a 32-byte string
-    state.add_global_memory_writes<nvbench::int8_t>(32 * num_rows);
+    state.add_global_memory_writes<nvbench::int8_t>(32L * num_rows);
 
     state.exec(nvbench::exec_tag::sync,
                [&](nvbench::launch& launch) { auto result = cudf::hashing::md5(data->view()); });
   } else if (hash_name == "sha1") {
     // sha1 creates a 40-byte string
-    state.add_global_memory_writes<nvbench::int8_t>(40 * num_rows);
+    state.add_global_memory_writes<nvbench::int8_t>(40L * num_rows);
 
     state.exec(nvbench::exec_tag::sync,
                [&](nvbench::launch& launch) { auto result = cudf::hashing::sha1(data->view()); });
   } else if (hash_name == "sha224") {
     // sha224 creates a 56-byte string
-    state.add_global_memory_writes<nvbench::int8_t>(56 * num_rows);
+    state.add_global_memory_writes<nvbench::int8_t>(56L * num_rows);
 
     state.exec(nvbench::exec_tag::sync,
                [&](nvbench::launch& launch) { auto result = cudf::hashing::sha224(data->view()); });
   } else if (hash_name == "sha256") {
     // sha256 creates a 64-byte string
-    state.add_global_memory_writes<nvbench::int8_t>(64 * num_rows);
+    state.add_global_memory_writes<nvbench::int8_t>(64L * num_rows);
 
     state.exec(nvbench::exec_tag::sync,
                [&](nvbench::launch& launch) { auto result = cudf::hashing::sha256(data->view()); });
   } else if (hash_name == "sha384") {
     // sha384 creates a 96-byte string
-    state.add_global_memory_writes<nvbench::int8_t>(96 * num_rows);
+    state.add_global_memory_writes<nvbench::int8_t>(96L * num_rows);
 
     state.exec(nvbench::exec_tag::sync,
                [&](nvbench::launch& launch) { auto result = cudf::hashing::sha384(data->view()); });
   } else if (hash_name == "sha512") {
     // sha512 creates a 128-byte string
-    state.add_global_memory_writes<nvbench::int8_t>(128 * num_rows);
+    state.add_global_memory_writes<nvbench::int8_t>(128L * num_rows);
 
     state.exec(nvbench::exec_tag::sync,
                [&](nvbench::launch& launch) { auto result = cudf::hashing::sha512(data->view()); });

cpp/benchmarks/ndsh/README.md

@@ -0,0 +1,11 @@
+# NDS-H Benchmarks for `libcudf`
+
+## Disclaimer
+
+NDS-H is derived from the TPC-H Benchmarks and as such any results obtained using NDS-H are not
+comparable to published TPC-H Benchmark results, as the results obtained from using NDS-H do not
+comply with the TPC-H Benchmarks.
+
+## Current Status
+
+For now, only Q1, Q5, Q6, Q9, and Q10 have been implemented

cpp/examples/tpch/q1.cpp → cpp/benchmarks/ndsh/q01.cpp

@@ -14,17 +14,19 @@
  * limitations under the License.
  */
 
-#include "../utilities/timer.hpp"
-#include "utils.hpp"
+#include "utilities.hpp"
 
 #include <cudf/ast/expressions.hpp>
+#include <cudf/binaryop.hpp>
 #include <cudf/column/column.hpp>
 #include <cudf/scalar/scalar.hpp>
 #include <cudf/utilities/memory_resource.hpp>
 
+#include <nvbench/nvbench.cuh>
+
 /**
- * @file q1.cpp
- * @brief Implement query 1 of the TPC-H benchmark.
+ * @file q01.cpp
+ * @brief Implement query 1 of the NDS-H benchmark.
  *
  * create view lineitem as select * from '/tables/scale-1/lineitem.parquet';
  *
@@ -59,7 +61,7 @@
  * @param stream The CUDA stream used for device memory operations and kernel launches.
  * @param mr Device memory resource used to allocate the returned column's device memory.
  */
-[[nodiscard]] std::unique_ptr<cudf::column> calc_disc_price(
+[[nodiscard]] std::unique_ptr<cudf::column> calculate_disc_price(
   cudf::column_view const& discount,
   cudf::column_view const& extendedprice,
   rmm::cuda_stream_view stream      = cudf::get_default_stream(),
@@ -86,7 +88,7 @@
  * @param stream The CUDA stream used for device memory operations and kernel launches.
  * @param mr Device memory resource used to allocate the returned column's device memory.
  */
-[[nodiscard]] std::unique_ptr<cudf::column> calc_charge(
+[[nodiscard]] std::unique_ptr<cudf::column> calculate_charge(
   cudf::column_view const& tax,
   cudf::column_view const& disc_price,
   rmm::cuda_stream_view stream      = cudf::get_default_stream(),
@@ -101,16 +103,9 @@
   return charge;
 }
 
-int main(int argc, char const** argv)
+void run_ndsh_q1(nvbench::state& state,
+                 std::unordered_map<std::string, parquet_device_buffer>& sources)
 {
-  auto const args = parse_args(argc, argv);
-
-  // Use a memory pool
-  auto resource = create_memory_resource(args.memory_resource_type);
-  cudf::set_current_device_resource(resource.get());
-
-  cudf::examples::timer timer;
-
   // Define the column projections and filter predicate for `lineitem` table
   std::vector<std::string> const lineitem_cols = {"l_returnflag",
                                                   "l_linestatus",
@@ -130,12 +125,12 @@ int main(int argc, char const** argv)
 
   // Read out the `lineitem` table from parquet file
   auto lineitem =
-    read_parquet(args.dataset_dir + "/lineitem.parquet", lineitem_cols, std::move(lineitem_pred));
+    read_parquet(sources["lineitem"].make_source_info(), lineitem_cols, std::move(lineitem_pred));
 
   // Calculate the discount price and charge columns and append to lineitem table
   auto disc_price =
-    calc_disc_price(lineitem->column("l_discount"), lineitem->column("l_extendedprice"));
-  auto charge = calc_charge(lineitem->column("l_tax"), disc_price->view());
+    calculate_disc_price(lineitem->column("l_discount"), lineitem->column("l_extendedprice"));
+  auto charge = calculate_charge(lineitem->column("l_tax"), disc_price->view());
   (*lineitem).append(disc_price, "disc_price").append(charge, "charge");
 
   // Perform the group by operation
@@ -167,9 +162,21 @@ int main(int argc, char const** argv)
                                              {"l_returnflag", "l_linestatus"},
                                              {cudf::order::ASCENDING, cudf::order::ASCENDING});
 
-  timer.print_elapsed_millis();
-
   // Write query result to a parquet file
   orderedby_table->to_parquet("q1.parquet");
-  return 0;
 }
+
+void ndsh_q1(nvbench::state& state)
+{
+  // Generate the required parquet files in device buffers
+  double const scale_factor = state.get_float64("scale_factor");
+  std::unordered_map<std::string, parquet_device_buffer> sources;
+  generate_parquet_data_sources(scale_factor, {"lineitem"}, sources);
+
+  auto stream = cudf::get_default_stream();
+  state.set_cuda_stream(nvbench::make_cuda_stream_view(stream.value()));
+  state.exec(nvbench::exec_tag::sync,
+             [&](nvbench::launch& launch) { run_ndsh_q1(state, sources); });
+}
+
+NVBENCH_BENCH(ndsh_q1).set_name("ndsh_q1").add_float64_axis("scale_factor", {0.01, 0.1, 1});