Add remaining NDS-H queries to libcudf with CI validation

ci/run_cudf_benchmark_smoketests.sh

@@ -1,9 +1,11 @@
 #!/bin/bash
-# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 set -euo pipefail
 
+repo_root="$(dirname "$(realpath "${BASH_SOURCE[0]}")")/.."
+
 # Support customizing the benchmarks' install location
 # First, try the installed location (CI/conda environments)
 installed_benchmark_location="${INSTALL_PREFIX:-${CONDA_PREFIX:-/usr}}/bin/benchmarks/libcudf/"
@@ -22,12 +24,23 @@ else
 fi
 
 EXITCODE=0
+validation_dir="$(mktemp -d)"
+ndsh_scale_factor=1
+trap 'rm -rf "${validation_dir}"' EXIT
 # Run all nvbench benchmarks with --profile and rmm_mode=cuda
 for bench in *_NVBENCH; do
   if [[ -x "$bench" && -f "$bench" ]]; then
     start_time=$(date +%s)
     echo "Running $bench with --profile..."
-    "./$bench" --profile --devices 0 -q --rmm_mode cuda
+    args=(--profile --devices 0 -q --rmm_mode cuda)
+    if [[ "$bench" == NDSH_* ]]; then
+      args+=(--axis "scale_factor=${ndsh_scale_factor}")
+      if [[ "$bench" =~ ^NDSH_Q([0-9]{2})_NVBENCH$ ]]; then
+        # Validate small-scale NDS-H benchmark outputs against DuckDB
+        args+=(--output_directory "${validation_dir}/q${BASH_REMATCH[1]}")
+      fi
+    fi
+    "./$bench" "${args[@]}"
     SUITEERROR=$?
     end_time=$(date +%s)
     duration=$((end_time - start_time))
@@ -40,5 +53,10 @@ for bench in *_NVBENCH; do
   fi
 done
 
+python "${repo_root}/ci/validate_ndsh_benchmarks.py" \
+  --output-dir "${validation_dir}" \
+  --sql-dir "${repo_root}/cpp/libcudf_streaming/benchmarks/streaming/ndsh/sql" \
+  --scale-factor "${ndsh_scale_factor}"
+
 echo "Test script exiting with value: $EXITCODE"
 exit ${EXITCODE}

ci/validate_ndsh_benchmarks.py

@@ -0,0 +1,130 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+import argparse
+import math
+import numbers
+from pathlib import Path
+
+import duckdb
+
+QUERIES = (
+    "q01",
+    "q02",
+    "q03",
+    "q04",
+    "q05",
+    "q06",
+    "q07",
+    "q08",
+    "q09",
+    "q10",
+    "q11",
+    "q12",
+    "q13",
+    "q14",
+    "q15",
+    "q16",
+    "q17",
+    "q18",
+    "q19",
+    "q20",
+    "q21",
+    "q22",
+)
+EXPECTED_NAMES = {
+    "q18": [
+        "c_name",
+        "c_custkey",
+        "o_orderkey",
+        "o_orderdate",
+        "o_totalprice",
+        "sum(l_quantity)",
+    ]
+}
+
+
+def values_equal(actual, expected):
+    if actual is None or expected is None:
+        return actual is expected
+    if isinstance(actual, numbers.Number) and isinstance(
+        expected, numbers.Number
+    ):
+        return math.isclose(
+            float(actual), float(expected), rel_tol=0.0, abs_tol=0.01
+        )
+    return actual == expected
+
+
+def validate_query(query_name, sql_dir, output_dir, scale_factor=0.01):
+    connection = duckdb.connect()
+    for path in (output_dir / query_name / "input").glob("*.parquet"):
+        table_name = path.stem.replace('"', '""')
+        parquet_path = str(path).replace("'", "''")
+        connection.execute(
+            f'CREATE VIEW "{table_name}" AS '
+            f"SELECT * FROM read_parquet('{parquet_path}')"
+        )
+
+    parameters = (
+        {"scale_factor": scale_factor} if query_name == "q11" else None
+    )
+    expected = connection.execute(
+        (sql_dir / f"{query_name}.sql").read_text(), parameters
+    )
+    expected_names = [column[0] for column in expected.description]
+    expected_rows = expected.fetchall()
+
+    result_path = output_dir / query_name / "results" / f"{query_name}.parquet"
+    actual = connection.execute(
+        "SELECT * FROM read_parquet(?)", [str(result_path)]
+    )
+    actual_names = [column[0] for column in actual.description]
+    actual_rows = actual.fetchall()
+
+    expected_names = EXPECTED_NAMES.get(query_name, expected_names)
+    if actual_names != expected_names:
+        return f"column names differ: {actual_names} != {expected_names}"
+    if len(actual_rows) != len(expected_rows):
+        return f"row counts differ: {len(actual_rows)} != {len(expected_rows)}"
+
+    for row_index, (actual_row, expected_row) in enumerate(
+        zip(actual_rows, expected_rows, strict=True)
+    ):
+        for column_name, actual_value, expected_value in zip(
+            actual_names, actual_row, expected_row, strict=True
+        ):
+            if not values_equal(actual_value, expected_value):
+                return (
+                    f"row {row_index}, column {column_name} differs: "
+                    f"{actual_value!r} != {expected_value!r}"
+                )
+    return None
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Validate NDS-H benchmark Parquet results against DuckDB"
+    )
+    parser.add_argument("--output-dir", type=Path, required=True)
+    parser.add_argument("--sql-dir", type=Path, required=True)
+    parser.add_argument("--scale-factor", type=float, default=0.01)
+    args = parser.parse_args()
+
+    failed = False
+    for query_name in QUERIES:
+        error = validate_query(
+            query_name, args.sql_dir, args.output_dir, args.scale_factor
+        )
+        if error is None:
+            print(f"{query_name}: PASSED")
+        else:
+            failed = True
+            print(f"{query_name}: FAILED: {error}")
+
+    raise SystemExit(failed)
+
+
+if __name__ == "__main__":
+    main()

conda/environments/all_cuda-129_arch-aarch64.yaml

@@ -32,6 +32,7 @@ dependencies:
 - dask-cuda==26.10.*,>=0.0.0a0
 - dlpack>=0.8,<1.0
 - doxygen=1.9.1
+- duckdb
 - fastavro>=0.22.9
 - flatbuffers==24.3.25
 - fsspec>=0.6.0

conda/environments/all_cuda-129_arch-x86_64.yaml

@@ -32,6 +32,7 @@ dependencies:
 - dask-cuda==26.10.*,>=0.0.0a0
 - dlpack>=0.8,<1.0
 - doxygen=1.9.1
+- duckdb
 - fastavro>=0.22.9
 - flatbuffers==24.3.25
 - fsspec>=0.6.0

conda/environments/all_cuda-133_arch-aarch64.yaml

@@ -32,6 +32,7 @@ dependencies:
 - dask-cuda==26.10.*,>=0.0.0a0
 - dlpack>=0.8,<1.0
 - doxygen=1.9.1
+- duckdb
 - fastavro>=0.22.9
 - flatbuffers==24.3.25
 - fsspec>=0.6.0

conda/environments/all_cuda-133_arch-x86_64.yaml

@@ -32,6 +32,7 @@ dependencies:
 - dask-cuda==26.10.*,>=0.0.0a0
 - dlpack>=0.8,<1.0
 - doxygen=1.9.1
+- duckdb
 - fastavro>=0.22.9
 - flatbuffers==24.3.25
 - fsspec>=0.6.0

cpp/CMakeLists.txt

@@ -1600,6 +1600,10 @@ if(CUDF_BUILD_BENCHMARKS)
   add_subdirectory(benchmarks)
 endif()
 
+if(CUDF_BUILD_TESTS)
+  rapids_test_install_relocatable(INSTALL_COMPONENT_SET testing DESTINATION bin/gtests/libcudf)
+endif()
+
 # ##################################################################################################
 # * install targets -------------------------------------------------------------------------------
 rapids_cmake_install_lib_dir(lib_dir)

cpp/benchmarks/CMakeLists.txt

@@ -55,6 +55,29 @@ target_include_directories(
          "$<BUILD_INTERFACE:${CUDF_SOURCE_DIR}/src>"
 )
 
+if(CUDF_BUILD_TESTS)
+  add_executable(NDSH_DATA_GENERATOR_TEST common/ndsh_data_generator/ndsh_data_generator_test.cpp)
+  set_target_properties(
+    NDSH_DATA_GENERATOR_TEST
+    PROPERTIES RUNTIME_OUTPUT_DIRECTORY "$<BUILD_INTERFACE:${CUDF_BINARY_DIR}/gtests>"
+               INSTALL_RPATH "\$ORIGIN/../../../lib"
+               CXX_STANDARD 20
+               CXX_STANDARD_REQUIRED ON
+  )
+  target_link_libraries(
+    NDSH_DATA_GENERATOR_TEST PRIVATE ndsh_data_generator cudf::cudftestutil_objects
+                                     $<TARGET_NAME_IF_EXISTS:conda_env>
+  )
+  rapids_cuda_set_runtime(NDSH_DATA_GENERATOR_TEST USE_STATIC ON)
+  rapids_test_add(
+    NAME NDSH_DATA_GENERATOR_TEST
+    COMMAND NDSH_DATA_GENERATOR_TEST
+    GPUS 1
+    PERCENT 15
+    INSTALL_COMPONENT_SET testing
+  )
+endif()
+
 # ##################################################################################################
 # * compiler function -----------------------------------------------------------------------------
 
@@ -125,10 +148,27 @@ ConfigureNVBench(TRANSPOSE_NVBENCH transpose/transpose.cpp)
 # ##################################################################################################
 # * nds-h benchmark --------------------------------------------------------------------------------
 ConfigureNVBench(NDSH_Q01_NVBENCH ndsh/q01.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q02_NVBENCH ndsh/q02.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q03_NVBENCH ndsh/q03.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q04_NVBENCH ndsh/q04.cpp ndsh/utilities.cpp)
 ConfigureNVBench(NDSH_Q05_NVBENCH ndsh/q05.cpp ndsh/utilities.cpp)
 ConfigureNVBench(NDSH_Q06_NVBENCH ndsh/q06.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q07_NVBENCH ndsh/q07.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q08_NVBENCH ndsh/q08.cpp ndsh/utilities.cpp)
 ConfigureNVBench(NDSH_Q09_NVBENCH ndsh/q09.cpp ndsh/utilities.cpp)
 ConfigureNVBench(NDSH_Q10_NVBENCH ndsh/q10.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q11_NVBENCH ndsh/q11.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q12_NVBENCH ndsh/q12.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q13_NVBENCH ndsh/q13.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q14_NVBENCH ndsh/q14.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q15_NVBENCH ndsh/q15.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q16_NVBENCH ndsh/q16.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q17_NVBENCH ndsh/q17.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q18_NVBENCH ndsh/q18.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q19_NVBENCH ndsh/q19.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q20_NVBENCH ndsh/q20.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q21_NVBENCH ndsh/q21.cpp ndsh/utilities.cpp)
+ConfigureNVBench(NDSH_Q22_NVBENCH ndsh/q22.cpp ndsh/utilities.cpp)
 
 # ##################################################################################################
 # * filter benchmark -------------------------------------------------------------------

cpp/benchmarks/common/ndsh_data_generator/ndsh_data_generator_test.cpp

@@ -0,0 +1,57 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "ndsh_data_generator.hpp"
+
+#include <cudf_test/base_fixture.hpp>
+
+#include <cudf/reduction.hpp>
+#include <cudf/scalar/scalar.hpp>
+
+#include <gtest/gtest.h>
+
+struct NDSHDataGeneratorTest : public cudf::test::BaseFixture {};
+
+TEST_F(NDSHDataGeneratorTest, ScaleFactorPointZeroOne)
+{
+  constexpr double scale_factor = 0.01;
+
+  auto [orders, lineitem, part] = cudf::datagen::generate_orders_lineitem_part(scale_factor);
+  auto partsupp                 = cudf::datagen::generate_partsupp(scale_factor);
+  auto supplier                 = cudf::datagen::generate_supplier(scale_factor);
+  auto customer                 = cudf::datagen::generate_customer(scale_factor);
+  auto nation                   = cudf::datagen::generate_nation();
+  auto region                   = cudf::datagen::generate_region();
+
+  auto const expect_cardinality =
+    [](cudf::table const& table, cudf::size_type rows, cudf::size_type columns) {
+      EXPECT_EQ(table.num_rows(), rows);
+      EXPECT_EQ(table.num_columns(), columns);
+    };
+
+  expect_cardinality(*orders, 15'000, 9);
+  EXPECT_GE(lineitem->num_rows(), 15'000);
+  EXPECT_LE(lineitem->num_rows(), 105'000);
+  EXPECT_EQ(lineitem->num_columns(), 16);
+  expect_cardinality(*part, 2'000, 9);
+  expect_cardinality(*partsupp, 8'000, 5);
+  expect_cardinality(*supplier, 100, 7);
+  expect_cardinality(*customer, 1'500, 8);
+  expect_cardinality(*nation, 25, 4);
+  expect_cardinality(*region, 5, 3);
+
+  auto const expect_supplier_key_range = [](cudf::column_view const& keys,
+                                            cudf::size_type supplier_rows) {
+    EXPECT_EQ(keys.null_count(), 0);
+    auto const [minimum, maximum] = cudf::minmax(keys);
+    auto const min_key = static_cast<cudf::numeric_scalar<cudf::size_type> const*>(minimum.get());
+    auto const max_key = static_cast<cudf::numeric_scalar<cudf::size_type> const*>(maximum.get());
+    EXPECT_GE(min_key->value(), 1);
+    EXPECT_LE(max_key->value(), supplier_rows);
+  };
+
+  expect_supplier_key_range(lineitem->view().column(2), supplier->num_rows());
+  expect_supplier_key_range(partsupp->view().column(1), supplier->num_rows());
+}

cpp/benchmarks/common/ndsh_data_generator/table_helpers.cpp

@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION.
+ * SPDX-FileCopyrightText: Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * SPDX-License-Identifier: Apache-2.0
  */
 
@@ -145,13 +145,14 @@ std::unique_ptr<cudf::table> perform_left_join(cudf::table_view const& left_inpu
  * @param mr Device memory resource used to allocate the returned column's device memory
  */
 [[nodiscard]] std::unique_ptr<cudf::column> calculate_l_suppkey(cudf::column_view const& l_partkey,
-                                                                cudf::size_type scale_factor,
+                                                                double scale_factor,
                                                                 cudf::size_type num_rows,
                                                                 rmm::cuda_stream_view stream,
                                                                 rmm::device_async_resource_ref mr)
 {
   CUDF_BENCHMARK_RANGE();
   // Expression: (l_partkey + (i * (s/4 + (int)(l_partkey - 1)/s))) % s + 1
+  auto const supplier_count = static_cast<cudf::size_type>(scale_factor * 10'000);
 
   // Generate the `s` col
   auto s_empty = cudf::make_numeric_column(
@@ -160,7 +161,7 @@ std::unique_ptr<cudf::table> perform_left_join(cudf::table_view const& left_inpu
   auto s = cudf::fill(s_empty->view(),
                       0,
                       num_rows,
-                      cudf::numeric_scalar<cudf::size_type>(scale_factor * 10'000),
+                      cudf::numeric_scalar<cudf::size_type>(supplier_count),
                       stream,
                       mr);
 
@@ -217,13 +218,14 @@ std::unique_ptr<cudf::table> perform_left_join(cudf::table_view const& left_inpu
  */
 [[nodiscard]] std::unique_ptr<cudf::column> calculate_ps_suppkey(
   cudf::column_view const& ps_partkey,
-  cudf::size_type scale_factor,
+  double scale_factor,
   cudf::size_type num_rows,
   rmm::cuda_stream_view stream,
   rmm::device_async_resource_ref mr)
 {
   CUDF_BENCHMARK_RANGE();
   // Expression: ps_suppkey = (ps_partkey + (i * (s/4 + (int)(ps_partkey - 1)/s))) % s + 1
+  auto const supplier_count = static_cast<cudf::size_type>(scale_factor * 10'000);
 
   // Generate the `s` col
   auto s_empty = cudf::make_numeric_column(
@@ -232,7 +234,7 @@ std::unique_ptr<cudf::table> perform_left_join(cudf::table_view const& left_inpu
   auto s = cudf::fill(s_empty->view(),
                       0,
                       num_rows,
-                      cudf::numeric_scalar<cudf::size_type>(scale_factor * 10'000),
+                      cudf::numeric_scalar<cudf::size_type>(supplier_count),
                       stream,
                       mr);