switched ast benchmarks from googlebench to nvbench

rapidsai · Sep 27, 2024 · cdf4ede · cdf4ede
1 parent 6b3d57d
commit cdf4ede
Show file tree

Hide file tree

Showing 4 changed files with 43 additions and 46 deletions.
diff --git a/cpp/benchmarks/CMakeLists.txt b/cpp/benchmarks/CMakeLists.txt
@@ -330,7 +330,7 @@ ConfigureNVBench(CSV_WRITER_NVBENCH io/csv/csv_writer.cpp)
 
 # ##################################################################################################
 # * ast benchmark ---------------------------------------------------------------------------------
-ConfigureBench(AST_BENCH ast/transform.cpp)
+ConfigureNVBench(AST_BENCH ast/transform.cpp)
 
 # ##################################################################################################
 # * binaryop benchmark ----------------------------------------------------------------------------

diff --git a/cpp/benchmarks/ast/transform.cpp b/cpp/benchmarks/ast/transform.cpp
@@ -15,14 +15,18 @@
  */
 
 #include <benchmarks/common/generate_input.hpp>
-#include <benchmarks/fixture/benchmark_fixture.hpp>
 #include <benchmarks/synchronization/synchronization.hpp>
 
 #include <cudf/transform.hpp>
 #include <cudf/types.hpp>
 
+#include <rmm/cuda_stream_view.hpp>
+
 #include <thrust/iterator/counting_iterator.h>
 
+#include <nvbench/exec_tag.cuh>
+#include <nvbench/nvbench.cuh>
+
 #include <algorithm>
 #include <list>
 #include <memory>
@@ -35,13 +39,10 @@ enum class TreeType {
 };
 
 template <typename key_type, TreeType tree_type, bool reuse_columns, bool Nullable>
-class AST : public cudf::benchmark {};
-
-template <typename key_type, TreeType tree_type, bool reuse_columns, bool Nullable>
-static void BM_ast_transform(benchmark::State& state)
+static void BM_ast_transform(nvbench::state& state)
 {
-  auto const table_size{static_cast<cudf::size_type>(state.range(0))};
-  auto const tree_levels{static_cast<cudf::size_type>(state.range(1))};
+  auto const table_size  = static_cast<cudf::size_type>(state.get_int64("TableSize"));
+  auto const tree_levels = static_cast<cudf::size_type>(state.get_int64("TreeLevels"));
 
   // Create table data
   auto const n_cols = reuse_columns ? 1 : tree_levels + 1;
@@ -86,38 +87,25 @@ static void BM_ast_transform(benchmark::State& state)
 
   auto const& expression_tree_root = expressions.back();
 
-  // Execute benchmark
-  for (auto _ : state) {
-    cuda_event_timer raii(state, true);  // flush_l2_cache = true, stream = 0
+  state.exec(nvbench::exec_tag::sync, [&](nvbench::launch&) {
+    flush_device_L2_cache(rmm::cuda_stream_view{state.get_cuda_stream().get_stream()});
     cudf::compute_column(table, expression_tree_root);
-  }
+  });
 
   // Use the number of bytes read from global memory
-  state.SetBytesProcessed(static_cast<int64_t>(state.iterations()) * state.range(0) *
-                          (tree_levels + 1) * sizeof(key_type));
-}
-
-static void CustomRanges(benchmark::internal::Benchmark* b)
-{
-  auto row_counts       = std::vector<cudf::size_type>{100'000, 1'000'000, 10'000'000, 100'000'000};
-  auto operation_counts = std::vector<cudf::size_type>{1, 5, 10};
-  for (auto const& row_count : row_counts) {
-    for (auto const& operation_count : operation_counts) {
-      b->Args({row_count, operation_count});
-    }
-  }
+  state.add_global_memory_reads(static_cast<int64_t>(state.get_summaries().size()) * table_size *
+                                (tree_levels + 1) * sizeof(key_type));
 }
 
 #define AST_TRANSFORM_BENCHMARK_DEFINE(name, key_type, tree_type, reuse_columns, nullable) \
-  BENCHMARK_TEMPLATE_DEFINE_F(AST, name, key_type, tree_type, reuse_columns, nullable)     \
-  (::benchmark::State & st)                                                                \
+  static void name(::nvbench::state& st)                                                   \
   {                                                                                        \
-    BM_ast_transform<key_type, tree_type, reuse_columns, nullable>(st);                    \
+    ::BM_ast_transform<key_type, tree_type, reuse_columns, nullable>(st);                  \
   }                                                                                        \
-  BENCHMARK_REGISTER_F(AST, name)                                                          \
-    ->Apply(CustomRanges)                                                                  \
-    ->Unit(benchmark::kMillisecond)                                                        \
-    ->UseManualTime();
+  NVBENCH_BENCH(name)                                                                      \
+    .set_name(#name)                                                                       \
+    .add_int64_axis("TableSize", {100'000, 1'000'000, 10'000'000, 100'000'000})            \
+    .add_int64_axis("TreeLevels", {1, 5, 10})
 
 AST_TRANSFORM_BENCHMARK_DEFINE(
   ast_int32_imbalanced_unique, int32_t, TreeType::IMBALANCED_LEFT, false, false);

diff --git a/cpp/benchmarks/synchronization/synchronization.cpp b/cpp/benchmarks/synchronization/synchronization.cpp
@@ -21,26 +21,29 @@
 #include <rmm/cuda_stream_view.hpp>
 #include <rmm/device_buffer.hpp>
 
+void flush_device_L2_cache(rmm::cuda_stream_view stream)
+{
+  int current_device = 0;
+  CUDF_CUDA_TRY(cudaGetDevice(&current_device));
+
+  int l2_cache_bytes = 0;
+  CUDF_CUDA_TRY(cudaDeviceGetAttribute(&l2_cache_bytes, cudaDevAttrL2CacheSize, current_device));
+
+  if (l2_cache_bytes > 0) {
+    int const memset_value = 0;
+    rmm::device_buffer l2_cache_buffer(l2_cache_bytes, stream);
+    CUDF_CUDA_TRY(
+      cudaMemsetAsync(l2_cache_buffer.data(), memset_value, l2_cache_bytes, stream.value()));
+  }
+}
+
 cuda_event_timer::cuda_event_timer(benchmark::State& state,
                                    bool flush_l2_cache,
                                    rmm::cuda_stream_view stream)
   : stream(stream), p_state(&state)
 {
   // flush all of L2$
-  if (flush_l2_cache) {
-    int current_device = 0;
-    CUDF_CUDA_TRY(cudaGetDevice(&current_device));
-
-    int l2_cache_bytes = 0;
-    CUDF_CUDA_TRY(cudaDeviceGetAttribute(&l2_cache_bytes, cudaDevAttrL2CacheSize, current_device));
-
-    if (l2_cache_bytes > 0) {
-      int const memset_value = 0;
-      rmm::device_buffer l2_cache_buffer(l2_cache_bytes, stream);
-      CUDF_CUDA_TRY(
-        cudaMemsetAsync(l2_cache_buffer.data(), memset_value, l2_cache_bytes, stream.value()));
-    }
-  }
+  if (flush_l2_cache) { flush_device_L2_cache(stream); }
 
   CUDF_CUDA_TRY(cudaEventCreate(&start));
   CUDF_CUDA_TRY(cudaEventCreate(&stop));

diff --git a/cpp/benchmarks/synchronization/synchronization.hpp b/cpp/benchmarks/synchronization/synchronization.hpp
@@ -69,6 +69,12 @@
 #include <benchmark/benchmark.h>
 #include <driver_types.h>
 
+/**
+ * @brief clears the L2$ by cudaMemset'ing a buffer of L2$ size
+ * @param stream CUDA stream used for device memory operations and kernel launches
+ */
+void flush_device_L2_cache(rmm::cuda_stream_view stream = cudf::get_default_stream());
+
 class cuda_event_timer {
  public:
   /**