From bb2abb9d63c40d2db07be4c1e01d1ee53bf08249 Mon Sep 17 00:00:00 2001
From: Basit Ayantunde <rlamarrr@gmail.com>
Date: Mon, 13 Jan 2025 17:59:58 +0000
Subject: [PATCH] added random scalar generation & cleanup of polynomials
 benchmark

---
 cpp/benchmarks/CMakeLists.txt            |  5 +++--
 cpp/benchmarks/ast/polynomials.cpp       | 17 ++++++++++++-----
 cpp/benchmarks/binaryop/polynomials.cpp  | 19 ++++++++++++++-----
 cpp/benchmarks/transform/polynomials.cpp | 16 +++++++++++-----
 4 files changed, 40 insertions(+), 17 deletions(-)

diff --git a/cpp/benchmarks/CMakeLists.txt b/cpp/benchmarks/CMakeLists.txt
index 468115942fb..4d32a1a1f3e 100644
--- a/cpp/benchmarks/CMakeLists.txt
+++ b/cpp/benchmarks/CMakeLists.txt
@@ -344,7 +344,7 @@ ConfigureNVBench(CSV_WRITER_NVBENCH io/csv/csv_writer.cpp)
 
 # ##################################################################################################
 # * ast benchmark ---------------------------------------------------------------------------------
-ConfigureNVBench(AST_NVBENCH ast/transform.cpp ast/polynomials.cpp)
+ConfigureNVBench(AST_NVBENCH ast/polynomials.cpp ast/transform.cpp)
 
 # ##################################################################################################
 # * binaryop benchmark ----------------------------------------------------------------------------
@@ -353,7 +353,8 @@ ConfigureNVBench(
 )
 
 # ##################################################################################################
-# * ast benchmark ---------------------------------------------------------------------------------
+# * transform benchmark
+# ---------------------------------------------------------------------------------
 ConfigureNVBench(TRANSFORM_NVBENCH transform/polynomials.cpp)
 
 # ##################################################################################################
diff --git a/cpp/benchmarks/ast/polynomials.cpp b/cpp/benchmarks/ast/polynomials.cpp
index c71f3dfedc9..b8e4ca46b72 100644
--- a/cpp/benchmarks/ast/polynomials.cpp
+++ b/cpp/benchmarks/ast/polynomials.cpp
@@ -28,6 +28,8 @@
 #include <nvbench/nvbench.cuh>
 #include <nvbench/types.cuh>
 
+#include <random>
+
 template <typename key_type>
 static void BM_ast_polynomials(nvbench::state& state)
 {
@@ -45,11 +47,16 @@ static void BM_ast_polynomials(nvbench::state& state)
   auto column_view = table->get_column(0);
 
   std::vector<cudf::numeric_scalar<key_type>> constants;
-
-  std::transform(thrust::make_counting_iterator(0),
-                 thrust::make_counting_iterator(order + 1),
-                 std::back_inserter(constants),
-                 [](int) { return cudf::numeric_scalar<key_type>(1); });
+  {
+    std::random_device random_device;
+    std::mt19937 generator;
+    std::uniform_real_distribution<key_type> distribution{0, 1};
+
+    std::transform(thrust::make_counting_iterator(0),
+                   thrust::make_counting_iterator(order + 1),
+                   std::back_inserter(constants),
+                   [&](int) { return distribution(generator); });
+  }
 
   cudf::ast::tree tree{};
 
diff --git a/cpp/benchmarks/binaryop/polynomials.cpp b/cpp/benchmarks/binaryop/polynomials.cpp
index 7a6489eb2c9..782ae1db927 100644
--- a/cpp/benchmarks/binaryop/polynomials.cpp
+++ b/cpp/benchmarks/binaryop/polynomials.cpp
@@ -28,6 +28,7 @@
 #include <nvbench/nvbench.cuh>
 
 #include <algorithm>
+#include <random>
 
 template <typename key_type>
 static void BM_binaryop_polynomials(nvbench::state& state)
@@ -46,11 +47,16 @@ static void BM_binaryop_polynomials(nvbench::state& state)
   auto column_view = table->get_column(0);
 
   std::vector<cudf::numeric_scalar<key_type>> constants;
+  {
+    std::random_device random_device;
+    std::mt19937 generator;
+    std::uniform_real_distribution<key_type> distribution{0, 1};
 
-  std::transform(thrust::make_counting_iterator(0),
-                 thrust::make_counting_iterator(order + 1),
-                 std::back_inserter(constants),
-                 [](int) { return cudf::numeric_scalar<key_type>(0.8F); });
+    std::transform(thrust::make_counting_iterator(0),
+                   thrust::make_counting_iterator(order + 1),
+                   std::back_inserter(constants),
+                   [&](int) { return cudf::numeric_scalar<key_type>(distribution(generator)); });
+  }
 
   // Use the number of bytes read from global memory
   state.add_global_memory_reads<key_type>(num_rows);
@@ -60,6 +66,7 @@ static void BM_binaryop_polynomials(nvbench::state& state)
     // computes polynomials: (((ax + b)x + c)x + d)x + e... = ax**4 + bx**3 + cx**2 + dx + e....
     cudf::scoped_range range{"benchmark_iteration"};
     rmm::cuda_stream_view stream{launch.get_stream().get_stream()};
+    std::vector<std::unique_ptr<cudf::column>> intermediates;
 
     auto result = cudf::make_column_from_scalar(constants[0], num_rows, stream);
 
@@ -74,7 +81,9 @@ static void BM_binaryop_polynomials(nvbench::state& state)
                                         cudf::binary_operator::ADD,
                                         cudf::data_type{cudf::type_to_id<key_type>()},
                                         stream);
-      result       = std::move(sum);
+      intermediates.push_back(std::move(product));
+      intermediates.push_back(std::move(result));
+      result = std::move(sum);
     }
   });
 }
diff --git a/cpp/benchmarks/transform/polynomials.cpp b/cpp/benchmarks/transform/polynomials.cpp
index 32fbeb0af45..f23789f5460 100644
--- a/cpp/benchmarks/transform/polynomials.cpp
+++ b/cpp/benchmarks/transform/polynomials.cpp
@@ -47,10 +47,16 @@ static void BM_transform_polynomials(nvbench::state& state)
 
   std::vector<key_type> constants;
 
-  std::transform(thrust::make_counting_iterator(0),
-                 thrust::make_counting_iterator(order + 1),
-                 std::back_inserter(constants),
-                 [](int) { return 0.8F; });
+  {
+    std::random_device random_device;
+    std::mt19937 generator;
+    std::uniform_real_distribution<key_type> distribution{0, 1};
+
+    std::transform(thrust::make_counting_iterator(0),
+                   thrust::make_counting_iterator(order + 1),
+                   std::back_inserter(constants),
+                   [&](int) { return distribution(generator); });
+  }
 
   // Use the number of bytes read from global memory
   state.add_global_memory_reads<key_type>(num_rows);
@@ -71,7 +77,7 @@ static void BM_transform_polynomials(nvbench::state& state)
     std::string type = std::is_same_v<key_type, float> ? "float" : "double";
 
     std::string udf = R"***(
-__device__ inline void    fdsf   (
+__device__ inline void    compute_polynomial   (
        )***" + type + R"***(* out,
        )***" + type + R"***( x
 )