[libc-commits] [libc] b42c332 - [libc] Use Atomics in GPU Benchmarks (#98842)
via libc-commits
libc-commits at lists.llvm.org
Mon Jul 15 05:08:27 PDT 2024
Author: jameshu15869
Date: 2024-07-15T07:08:23-05:00
New Revision: b42c332d734319c8a522fa3a24642550bac5d653
URL: https://github.com/llvm/llvm-project/commit/b42c332d734319c8a522fa3a24642550bac5d653
DIFF: https://github.com/llvm/llvm-project/commit/b42c332d734319c8a522fa3a24642550bac5d653.diff
LOG: [libc] Use Atomics in GPU Benchmarks (#98842)
This PR replaces our old method of reducing the benchmark results by
using an array to using atomics instead. This should help us implement
single threaded benchmarks.
Added:
Modified:
libc/benchmarks/gpu/CMakeLists.txt
libc/benchmarks/gpu/LibcGpuBenchmark.cpp
libc/benchmarks/gpu/LibcGpuBenchmark.h
Removed:
################################################################################
diff --git a/libc/benchmarks/gpu/CMakeLists.txt b/libc/benchmarks/gpu/CMakeLists.txt
index d167abcaf2db1..eaeecbdacd23e 100644
--- a/libc/benchmarks/gpu/CMakeLists.txt
+++ b/libc/benchmarks/gpu/CMakeLists.txt
@@ -43,6 +43,7 @@ add_unittest_framework_library(
libc.src.__support.CPP.functional
libc.src.__support.CPP.limits
libc.src.__support.CPP.algorithm
+ libc.src.__support.CPP.atomic
libc.src.__support.fixed_point.fx_rep
libc.src.__support.macros.properties.types
libc.src.__support.OSUtil.osutil
diff --git a/libc/benchmarks/gpu/LibcGpuBenchmark.cpp b/libc/benchmarks/gpu/LibcGpuBenchmark.cpp
index 1c1ba7639d0b1..23fff3e8180f7 100644
--- a/libc/benchmarks/gpu/LibcGpuBenchmark.cpp
+++ b/libc/benchmarks/gpu/LibcGpuBenchmark.cpp
@@ -1,6 +1,7 @@
#include "LibcGpuBenchmark.h"
#include "src/__support/CPP/algorithm.h"
#include "src/__support/CPP/array.h"
+#include "src/__support/CPP/atomic.h"
#include "src/__support/CPP/string.h"
#include "src/__support/FPUtil/sqrt.h"
#include "src/__support/GPU/utils.h"
@@ -12,41 +13,96 @@ namespace LIBC_NAMESPACE_DECL {
namespace benchmarks {
FixedVector<Benchmark *, 64> benchmarks;
-cpp::array<BenchmarkResult, 1024> results;
void Benchmark::add_benchmark(Benchmark *benchmark) {
benchmarks.push_back(benchmark);
}
-BenchmarkResult
-reduce_results(const cpp::array<BenchmarkResult, 1024> &results) {
- BenchmarkResult result;
- uint64_t cycles_sum = 0;
- double standard_deviation_sum = 0;
- uint64_t min = UINT64_MAX;
- uint64_t max = 0;
- uint32_t samples_sum = 0;
- uint32_t iterations_sum = 0;
- clock_t time_sum = 0;
- uint64_t num_threads = gpu::get_num_threads();
- for (uint64_t i = 0; i < num_threads; i++) {
- BenchmarkResult current_result = results[i];
- cycles_sum += current_result.cycles;
- standard_deviation_sum += current_result.standard_deviation;
- min = cpp::min(min, current_result.min);
- max = cpp::max(max, current_result.max);
- samples_sum += current_result.samples;
- iterations_sum += current_result.total_iterations;
- time_sum += current_result.total_time;
+struct AtomicBenchmarkSums {
+ cpp::Atomic<uint64_t> cycles_sum = 0;
+ cpp::Atomic<uint64_t> standard_deviation_sum = 0;
+ cpp::Atomic<uint64_t> min = UINT64_MAX;
+ cpp::Atomic<uint64_t> max = 0;
+ cpp::Atomic<uint32_t> samples_sum = 0;
+ cpp::Atomic<uint32_t> iterations_sum = 0;
+ cpp::Atomic<clock_t> time_sum = 0;
+ cpp::Atomic<uint64_t> active_threads = 0;
+
+ void reset() {
+ cpp::atomic_thread_fence(cpp::MemoryOrder::RELEASE);
+ active_threads.store(0, cpp::MemoryOrder::RELAXED);
+ cycles_sum.store(0, cpp::MemoryOrder::RELAXED);
+ standard_deviation_sum.store(0, cpp::MemoryOrder::RELAXED);
+ min.store(UINT64_MAX, cpp::MemoryOrder::RELAXED);
+ max.store(0, cpp::MemoryOrder::RELAXED);
+ samples_sum.store(0, cpp::MemoryOrder::RELAXED);
+ iterations_sum.store(0, cpp::MemoryOrder::RELAXED);
+ time_sum.store(0, cpp::MemoryOrder::RELAXED);
+ cpp::atomic_thread_fence(cpp::MemoryOrder::RELEASE);
}
- result.cycles = cycles_sum / num_threads;
- result.standard_deviation = standard_deviation_sum / num_threads;
- result.min = min;
- result.max = max;
- result.samples = samples_sum / num_threads;
- result.total_iterations = iterations_sum / num_threads;
- result.total_time = time_sum / num_threads;
- return result;
+
+ void update(const BenchmarkResult &result) {
+ cpp::atomic_thread_fence(cpp::MemoryOrder::RELEASE);
+ active_threads.fetch_add(1, cpp::MemoryOrder::RELAXED);
+
+ cycles_sum.fetch_add(result.cycles, cpp::MemoryOrder::RELAXED);
+ standard_deviation_sum.fetch_add(
+ static_cast<uint64_t>(result.standard_deviation),
+ cpp::MemoryOrder::RELAXED);
+
+ // Perform a CAS loop to atomically update the min
+ uint64_t orig_min = min.load(cpp::MemoryOrder::RELAXED);
+ while (!min.compare_exchange_strong(
+ orig_min, cpp::min(orig_min, result.min), cpp::MemoryOrder::ACQUIRE,
+ cpp::MemoryOrder::RELAXED))
+ ;
+
+ // Perform a CAS loop to atomically update the max
+ uint64_t orig_max = max.load(cpp::MemoryOrder::RELAXED);
+ while (!max.compare_exchange_strong(
+ orig_max, cpp::max(orig_max, result.max), cpp::MemoryOrder::ACQUIRE,
+ cpp::MemoryOrder::RELAXED))
+ ;
+
+ samples_sum.fetch_add(result.samples, cpp::MemoryOrder::RELAXED);
+ iterations_sum.fetch_add(result.total_iterations,
+ cpp::MemoryOrder::RELAXED);
+ time_sum.fetch_add(result.total_time, cpp::MemoryOrder::RELAXED);
+ cpp::atomic_thread_fence(cpp::MemoryOrder::RELEASE);
+ }
+};
+
+AtomicBenchmarkSums all_results;
+
+void print_results(Benchmark *b) {
+ constexpr auto GREEN = "\033[32m";
+ constexpr auto RESET = "\033[0m";
+
+ BenchmarkResult result;
+ cpp::atomic_thread_fence(cpp::MemoryOrder::RELEASE);
+ int num_threads = all_results.active_threads.load(cpp::MemoryOrder::RELAXED);
+ result.cycles =
+ all_results.cycles_sum.load(cpp::MemoryOrder::RELAXED) / num_threads;
+ result.standard_deviation =
+ all_results.standard_deviation_sum.load(cpp::MemoryOrder::RELAXED) /
+ num_threads;
+ result.min = all_results.min.load(cpp::MemoryOrder::RELAXED);
+ result.max = all_results.max.load(cpp::MemoryOrder::RELAXED);
+ result.samples =
+ all_results.samples_sum.load(cpp::MemoryOrder::RELAXED) / num_threads;
+ result.total_iterations =
+ all_results.iterations_sum.load(cpp::MemoryOrder::RELAXED) / num_threads;
+ result.total_time =
+ all_results.time_sum.load(cpp::MemoryOrder::RELAXED) / num_threads;
+ cpp::atomic_thread_fence(cpp::MemoryOrder::RELEASE);
+
+ log << GREEN << "[ RUN ] " << RESET << b->get_name() << '\n';
+ log << GREEN << "[ OK ] " << RESET << b->get_name() << ": "
+ << result.cycles << " cycles, " << result.min << " min, " << result.max
+ << " max, " << result.total_iterations << " iterations, "
+ << result.total_time << " ns, "
+ << static_cast<uint64_t>(result.standard_deviation)
+ << " stddev (num threads: " << num_threads << ")\n";
}
void Benchmark::run_benchmarks() {
@@ -54,19 +110,16 @@ void Benchmark::run_benchmarks() {
gpu::sync_threads();
for (Benchmark *b : benchmarks) {
- results[id] = b->run();
+ if (id == 0)
+ all_results.reset();
+
gpu::sync_threads();
- if (id == 0) {
- BenchmarkResult all_results = reduce_results(results);
- constexpr auto GREEN = "\033[32m";
- constexpr auto RESET = "\033[0m";
- log << GREEN << "[ RUN ] " << RESET << b->get_name() << '\n';
- log << GREEN << "[ OK ] " << RESET << b->get_name() << ": "
- << all_results.cycles << " cycles, " << all_results.min << " min, "
- << all_results.max << " max, " << all_results.total_iterations
- << " iterations, " << all_results.total_time << " ns, "
- << static_cast<long>(all_results.standard_deviation) << " stddev\n";
- }
+ auto current_result = b->run();
+ all_results.update(current_result);
+ gpu::sync_threads();
+
+ if (id == 0)
+ print_results(b);
}
gpu::sync_threads();
}
diff --git a/libc/benchmarks/gpu/LibcGpuBenchmark.h b/libc/benchmarks/gpu/LibcGpuBenchmark.h
index 26cb0fd30bc1c..1f813f8655de6 100644
--- a/libc/benchmarks/gpu/LibcGpuBenchmark.h
+++ b/libc/benchmarks/gpu/LibcGpuBenchmark.h
@@ -88,6 +88,7 @@ class Benchmark {
}
static void run_benchmarks();
+ const cpp::string_view get_name() const { return name; }
protected:
static void add_benchmark(Benchmark *benchmark);
@@ -97,13 +98,12 @@ class Benchmark {
BenchmarkOptions options;
return benchmark(options, func);
}
- const cpp::string_view get_name() const { return name; }
};
} // namespace benchmarks
} // namespace LIBC_NAMESPACE_DECL
#define BENCHMARK(SuiteName, TestName, Func) \
LIBC_NAMESPACE::benchmarks::Benchmark SuiteName##_##TestName##_Instance( \
- Func, #SuiteName "." #TestName);
+ Func, #SuiteName "." #TestName)
#endif
More information about the libc-commits
mailing list