-
Notifications
You must be signed in to change notification settings - Fork 257
cuda.bindings latency benchmarks #1736
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
base: main
Are you sure you want to change the base?
Changes from all commits
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,10 @@ | ||
| # Build artifacts | ||
| .build/ | ||
| __pycache__/ | ||
|
|
||
| # Benchmark results | ||
| *.json | ||
| .benchmarks/ | ||
|
|
||
| # Pixi environments | ||
| .pixi/ |
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,42 @@ | ||
| # cuda.bindings Benchmarks | ||
|
|
||
| ## Usage | ||
|
|
||
| Requires pixi. | ||
|
|
||
| There are a couple of environments defined based on how `cuda.bindings` is installed: | ||
|
|
||
| - `wheel`: Installs from conda packages | ||
| - `source`: Installs from source | ||
|
|
||
| There are a couple of tasks defined: | ||
|
|
||
| - `bench`: Runs the Python benchmarks | ||
| - `bench-cpp`: Runs the C++ benchmarks | ||
|
|
||
| To run the benchmarks combine the environment and task: | ||
|
|
||
| ```bash | ||
|
|
||
| # Run the Python benchmarks in the wheel environment | ||
| pixi run -e wheel bench | ||
|
|
||
| # Run the Python benchmarks in the source environment | ||
| pixi run -e source bench | ||
|
|
||
| # Run the C++ benchmarks (environment is irrelavant here) | ||
| pixi run -e wheel bench-cpp | ||
| ``` | ||
|
|
||
| ## pyperf JSON | ||
|
|
||
| The benchmarks are run using [pyperf](https://pyperf.readthedocs.io/en/latest/). | ||
| The results are written to a JSON file in the format expected by pyperf. | ||
|
|
||
| The C++ benchmarks also generate a valid JSON file, in the same format. | ||
|
|
||
| ``` | ||
| pixi run -e wheel bench-cpp -0 cpp.json | ||
|
|
||
| pixi run -e wheel pyperf stats cpp.json | ||
| ``` | ||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,17 @@ | ||
| # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. | ||
|
Contributor
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Maybe move this file into a |
||
| # | ||
| # SPDX-License-Identifier: Apache-2.0 | ||
|
|
||
| from cuda.bindings import driver as cuda | ||
|
|
||
| from runner.runtime import alloc_persistent, assert_drv | ||
|
|
||
|
|
||
| # Allocate memory used by the tests | ||
| PTR = alloc_persistent(1 << 18) | ||
| ATTRIBUTE = cuda.CUpointer_attribute.CU_POINTER_ATTRIBUTE_MEMORY_TYPE | ||
|
|
||
|
|
||
| def bench_pointer_get_attribute() -> None: | ||
| err, _ = cuda.cuPointerGetAttribute(ATTRIBUTE, PTR) | ||
|
Contributor
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. When this is refactored to do its own timing measurement, the |
||
| assert_drv(err) | ||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,48 @@ | ||
| # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. | ||
| # | ||
| # SPDX-License-Identifier: Apache-2.0 | ||
|
|
||
| cmake_minimum_required(VERSION 3.24) | ||
| project(cuda_bindings_cpp_benchmarks LANGUAGES CXX) | ||
|
|
||
| set(CMAKE_CXX_STANDARD 17) | ||
| set(CMAKE_CXX_STANDARD_REQUIRED ON) | ||
| set(CMAKE_CXX_EXTENSIONS OFF) | ||
|
|
||
| set(CUDA_HOME_HINT "$ENV{CUDA_HOME}") | ||
| set(CONDA_PREFIX_HINT "$ENV{CONDA_PREFIX}") | ||
|
|
||
| # Find cuda.h (driver API header) | ||
| find_path( | ||
| CUDA_DRIVER_INCLUDE_DIR | ||
| cuda.h | ||
| HINTS | ||
| "${CUDA_HOME_HINT}/include" | ||
| "${CONDA_PREFIX_HINT}/targets/x86_64-linux/include" | ||
| "${CONDA_PREFIX_HINT}/include" | ||
| ) | ||
|
|
||
| # Find libcuda (driver API library) — lives on the system, not in toolkit | ||
| find_library( | ||
| CUDA_DRIVER_LIBRARY | ||
| NAMES cuda | ||
| HINTS | ||
| "/usr/lib/x86_64-linux-gnu" | ||
| "/usr/lib64" | ||
| "${CUDA_HOME_HINT}/lib64/stubs" | ||
| "${CUDA_HOME_HINT}/lib/stubs" | ||
| "${CONDA_PREFIX_HINT}/targets/x86_64-linux/lib/stubs" | ||
| "${CONDA_PREFIX_HINT}/lib/stubs" | ||
| ) | ||
|
|
||
| if(NOT CUDA_DRIVER_INCLUDE_DIR) | ||
| message(FATAL_ERROR "Could not find cuda.h. Ensure CUDA_HOME is set or install cuda-crt-dev.") | ||
| endif() | ||
|
|
||
| if(NOT CUDA_DRIVER_LIBRARY) | ||
| message(FATAL_ERROR "Could not find libcuda. Ensure the NVIDIA driver is installed.") | ||
| endif() | ||
|
|
||
| add_executable(bench_pointer_attributes_cpp bench_pointer_attributes.cpp) | ||
|
Contributor
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Did you forget to commit |
||
| target_include_directories(bench_pointer_attributes_cpp PRIVATE "${CUDA_DRIVER_INCLUDE_DIR}") | ||
| target_link_libraries(bench_pointer_attributes_cpp PRIVATE "${CUDA_DRIVER_LIBRARY}") | ||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,218 @@ | ||
| // SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. | ||
| // | ||
| // SPDX-License-Identifier: Apache-2.0 | ||
|
|
||
| #pragma once | ||
|
|
||
| #include <chrono> | ||
| #include <cmath> | ||
| #include <cstdint> | ||
| #include <cstdlib> | ||
| #include <ctime> | ||
| #include <fstream> | ||
| #include <iomanip> | ||
| #include <iostream> | ||
| #include <string> | ||
| #include <vector> | ||
|
|
||
| namespace bench { | ||
|
|
||
| struct Options { | ||
| std::uint64_t loops = 1000; | ||
| std::uint64_t warmups = 5; | ||
| std::uint64_t values = 20; | ||
| std::uint64_t runs = 20; | ||
| std::string output_path; | ||
| std::string benchmark_name; | ||
| }; | ||
|
|
||
| // A single run result: warmup values and timed values (seconds per loop) | ||
| struct RunResult { | ||
| std::string date; | ||
| double duration_sec; | ||
| std::vector<double> warmup_values; // seconds per loop | ||
| std::vector<double> values; // seconds per loop | ||
| }; | ||
|
|
||
| inline Options parse_args(int argc, char** argv) { | ||
| Options options; | ||
| for (int i = 1; i < argc; ++i) { | ||
| const std::string arg(argv[i]); | ||
| if (arg == "--loops" && i + 1 < argc) { | ||
| options.loops = std::strtoull(argv[++i], nullptr, 10); | ||
| continue; | ||
| } | ||
| if (arg == "--warmups" && i + 1 < argc) { | ||
| options.warmups = std::strtoull(argv[++i], nullptr, 10); | ||
| continue; | ||
| } | ||
| if (arg == "--values" && i + 1 < argc) { | ||
| options.values = std::strtoull(argv[++i], nullptr, 10); | ||
| continue; | ||
| } | ||
| if (arg == "--runs" && i + 1 < argc) { | ||
| options.runs = std::strtoull(argv[++i], nullptr, 10); | ||
| continue; | ||
| } | ||
| if ((arg == "-o" || arg == "--output") && i + 1 < argc) { | ||
| options.output_path = argv[++i]; | ||
| continue; | ||
| } | ||
| if (arg == "--name" && i + 1 < argc) { | ||
| options.benchmark_name = argv[++i]; | ||
| continue; | ||
| } | ||
| if (arg == "--help" || arg == "-h") { | ||
| std::cout << "Usage: benchmark [options]\n" | ||
| << " --loops N Loop iterations per value (default: 1000)\n" | ||
| << " --warmups N Warmup values per run (default: 5)\n" | ||
| << " --values N Timed values per run (default: 20)\n" | ||
| << " --runs N Number of runs (default: 20)\n" | ||
| << " -o, --output F Write pyperf-compatible JSON to file\n" | ||
| << " --name S Benchmark name (overrides default)\n"; | ||
| std::exit(0); | ||
| } | ||
|
|
||
| std::cerr << "Unknown argument: " << arg << '\n'; | ||
| std::exit(2); | ||
| } | ||
| return options; | ||
| } | ||
|
|
||
| inline std::string iso_now() { | ||
| const auto now = std::chrono::system_clock::now(); | ||
| const std::time_t t = std::chrono::system_clock::to_time_t(now); | ||
| std::tm tm{}; | ||
| gmtime_r(&t, &tm); | ||
| char buf[64]; | ||
| std::strftime(buf, sizeof(buf), "%Y-%m-%d %H:%M:%S", &tm); | ||
| return std::string(buf); | ||
| } | ||
|
|
||
| // Run a benchmark function. The function signature is: void fn() — one call = one operation. | ||
| // The harness calls fn() in a tight loop `loops` times per value. | ||
| template <typename Fn> | ||
| std::vector<RunResult> run_benchmark(const Options& options, Fn&& fn) { | ||
| std::vector<RunResult> results; | ||
| results.reserve(options.runs); | ||
|
|
||
| for (std::uint64_t r = 0; r < options.runs; ++r) { | ||
| RunResult run; | ||
| run.date = iso_now(); | ||
| const auto run_start = std::chrono::steady_clock::now(); | ||
|
|
||
| // Warmups | ||
| for (std::uint64_t w = 0; w < options.warmups; ++w) { | ||
| const auto t0 = std::chrono::steady_clock::now(); | ||
| for (std::uint64_t i = 0; i < options.loops; ++i) { | ||
| fn(); | ||
| } | ||
| const auto t1 = std::chrono::steady_clock::now(); | ||
| const double elapsed = std::chrono::duration<double>(t1 - t0).count(); | ||
| run.warmup_values.push_back(elapsed / static_cast<double>(options.loops)); | ||
| } | ||
|
|
||
| // Timed values | ||
| for (std::uint64_t v = 0; v < options.values; ++v) { | ||
| const auto t0 = std::chrono::steady_clock::now(); | ||
| for (std::uint64_t i = 0; i < options.loops; ++i) { | ||
| fn(); | ||
| } | ||
| const auto t1 = std::chrono::steady_clock::now(); | ||
| const double elapsed = std::chrono::duration<double>(t1 - t0).count(); | ||
| run.values.push_back(elapsed / static_cast<double>(options.loops)); | ||
| } | ||
|
|
||
| const auto run_end = std::chrono::steady_clock::now(); | ||
| run.duration_sec = std::chrono::duration<double>(run_end - run_start).count(); | ||
| results.push_back(std::move(run)); | ||
| } | ||
|
|
||
| return results; | ||
| } | ||
|
|
||
| inline void print_summary(const std::string& name, const std::vector<RunResult>& results) { | ||
| // Collect all timed values | ||
| std::vector<double> all_values; | ||
| for (const auto& run : results) { | ||
| for (double v : run.values) { | ||
| all_values.push_back(v); | ||
| } | ||
| } | ||
| if (all_values.empty()) return; | ||
|
|
||
| double sum = 0; | ||
| for (double v : all_values) sum += v; | ||
| double mean = sum / static_cast<double>(all_values.size()); | ||
|
|
||
| double sq_sum = 0; | ||
| for (double v : all_values) { | ||
| double diff = v - mean; | ||
| sq_sum += diff * diff; | ||
| } | ||
| double stdev = std::sqrt(sq_sum / static_cast<double>(all_values.size())); | ||
|
|
||
| std::cout << name << ": Mean +- std dev: " | ||
| << std::fixed << std::setprecision(0) | ||
| << (mean * 1e9) << " ns +- " | ||
| << (stdev * 1e9) << " ns\n"; | ||
| } | ||
|
|
||
| // Escape a JSON string (minimal — no control chars expected) | ||
| inline std::string json_str(const std::string& s) { | ||
| return "\"" + s + "\""; | ||
| } | ||
|
|
||
| inline void write_pyperf_json( | ||
| const std::string& output_path, | ||
| const std::string& name, | ||
| std::uint64_t loops, | ||
| const std::vector<RunResult>& results | ||
| ) { | ||
| std::ofstream out(output_path); | ||
| if (!out) { | ||
| std::cerr << "Failed to open output file: " << output_path << '\n'; | ||
| std::exit(3); | ||
| } | ||
|
|
||
| out << std::setprecision(17); | ||
|
|
||
| out << "{\"version\": \"1.0\", "; | ||
| out << "\"metadata\": {"; | ||
| out << "\"name\": " << json_str(name) << ", "; | ||
| out << "\"loops\": " << loops << ", "; | ||
| out << "\"unit\": \"second\""; | ||
| out << "}, "; | ||
|
|
||
| out << "\"benchmarks\": [{\"runs\": ["; | ||
|
|
||
| for (std::size_t r = 0; r < results.size(); ++r) { | ||
| const auto& run = results[r]; | ||
| if (r > 0) out << ", "; | ||
|
|
||
| out << "{\"metadata\": {"; | ||
| out << "\"date\": " << json_str(run.date) << ", "; | ||
| out << "\"duration\": " << run.duration_sec; | ||
| out << "}, "; | ||
|
|
||
| // Warmups: array of [loops, value] pairs | ||
| out << "\"warmups\": ["; | ||
| for (std::size_t w = 0; w < run.warmup_values.size(); ++w) { | ||
| if (w > 0) out << ", "; | ||
| out << "[" << loops << ", " << run.warmup_values[w] << "]"; | ||
| } | ||
| out << "], "; | ||
|
|
||
| // Values | ||
| out << "\"values\": ["; | ||
| for (std::size_t v = 0; v < run.values.size(); ++v) { | ||
| if (v > 0) out << ", "; | ||
| out << run.values[v]; | ||
| } | ||
| out << "]}"; | ||
| } | ||
|
|
||
| out << "]}]}\n"; | ||
| } | ||
|
|
||
| } // namespace bench |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Maybe mention pyperf system tune here?