comparison for load/store vs loadu/storeu

CMakeLists.txt

@@ -59,6 +59,7 @@ target_link_libraries(benchmarks
     benchmark
     benchmark_main
     pasta_bit_vector)
+
 add_test(
     NAME Benchmarks
     COMMAND bt_benchmarks --benchmark_out=bm_report.csv --benchmark_out_format=csv
@@ -104,6 +105,17 @@ target_link_libraries(test_rmm
     gtest_main)
 
 
+add_executable(alignment_comparison
+    src/alignment_comparison.cpp)
+
+target_include_directories(alignment_comparison
+    PUBLIC include
+)
+
+target_link_libraries(alignment_comparison
+    benchmark
+    benchmark_main
+    )
 FetchContent_Declare(
     doxygen-awesome-css
     URL https://github.com/jothepro/doxygen-awesome-css/archive/refs/heads/main.zip

src/alignment_comparison.cpp

@@ -0,0 +1,245 @@
+#include <benchmark/benchmark.h>
+#include <immintrin.h>
+
+#include <chrono>
+#include <random>
+#include <vector>
+
+#include "bits.h"
+
+alignas(64) uint8_t data[(1 << 29) + 1];
+
+#ifdef PIXIE_AVX512_SUPPORT
+
+static void BM_Loadu512_aligned(benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = (rng() & ((1 << k) - 1)) << 6;
+    const __m512i* ptr = reinterpret_cast<const __m512i*>(data + idx);
+
+    benchmark::DoNotOptimize(_mm512_loadu_si512(ptr));
+  }
+}
+
+static void BM_Loadu512_unaligned_crossing_64byte_border(
+    benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = ((rng() & ((1 << k) - 1)) << 6) + 48;
+    const __m512i* ptr = reinterpret_cast<const __m512i*>(data + idx);
+
+    benchmark::DoNotOptimize(_mm512_loadu_si512(ptr));
+  }
+}
+
+static void BM_Load512_aligned(benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = (rng() & ((1 << k) - 1)) << 6;
+    const __m512i* ptr = reinterpret_cast<const __m512i*>(data + idx);
+
+    benchmark::DoNotOptimize(_mm512_load_si512(ptr));
+  }
+}
+
+static void BM_Storeu512_aligned(benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = (rng() & ((1 << k) - 1)) << 6;
+    __m512i* ptr = reinterpret_cast<__m512i*>(data + idx);
+    __m512i value = _mm512_setzero_si512();
+
+    _mm512_storeu_si512(ptr, value);
+
+    benchmark::DoNotOptimize(ptr);
+  }
+}
+
+static void BM_Storeu512_unaligned_crossing_64byte_border(
+    benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = ((rng() & ((1 << k) - 1)) << 6) + 48;
+    __m512i* ptr = reinterpret_cast<__m512i*>(data + idx);
+    __m512i value = _mm512_setzero_si512();
+
+    _mm512_storeu_si512(ptr, value);
+
+    benchmark::DoNotOptimize(ptr);
+  }
+}
+
+static void BM_Store512_aligned(benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = (rng() & ((1 << k) - 1)) << 6;
+    __m512i* ptr = reinterpret_cast<__m512i*>(data + idx);
+    __m512i value = _mm512_setzero_si512();
+
+    _mm512_store_si512(ptr, value);
+
+    benchmark::DoNotOptimize(ptr);
+  }
+}
+
+BENCHMARK(BM_Loadu512_aligned)->ArgNames({"k"})->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Loadu512_unaligned_crossing_64byte_border)
+    ->ArgNames({"k"})
+    ->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Load512_aligned)->ArgNames({"k"})->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Storeu512_aligned)->ArgNames({"k"})->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Storeu512_unaligned_crossing_64byte_border)
+    ->ArgNames({"k"})
+    ->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Store512_aligned)->ArgNames({"k"})->DenseRange(1, 23, 2);
+
+#else
+
+static void BM_Loadu256_aligned(benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = (rng() & ((1 << k) - 1)) << 6;
+    const __m256i* ptr = reinterpret_cast<const __m256i*>(data + idx);
+
+    benchmark::DoNotOptimize(_mm256_loadu_si256(ptr));
+  }
+}
+
+static void BM_Loadu256_unaligned(benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = ((rng() & ((1 << k) - 1)) << 6) + 16;
+    const __m256i* ptr = reinterpret_cast<const __m256i*>(data + idx);
+
+    benchmark::DoNotOptimize(_mm256_loadu_si256(ptr));
+  }
+}
+
+static void BM_Loadu256_unaligned_crossing_64byte_border(
+    benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = ((rng() & ((1 << k) - 1)) << 6) + 48;
+    const __m256i* ptr = reinterpret_cast<const __m256i*>(data + idx);
+
+    benchmark::DoNotOptimize(_mm256_loadu_si256(ptr));
+  }
+}
+
+static void BM_Load256_aligned(benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = (rng() & ((1 << k) - 1)) << 6;
+    const __m256i* ptr = reinterpret_cast<const __m256i*>(data + idx);
+
+    benchmark::DoNotOptimize(_mm256_load_si256(ptr));
+  }
+}
+
+static void BM_Storeu256_aligned(benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = (rng() & ((1 << k) - 1)) << 6;
+    __m256i* ptr = reinterpret_cast<__m256i*>(data + idx);
+    __m256i value = _mm256_setzero_si256();
+
+    _mm256_storeu_si256(ptr, value);
+
+    benchmark::DoNotOptimize(ptr);
+  }
+}
+
+static void BM_Storeu256_unaligned(benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = ((rng() & ((1 << k) - 1)) << 6) + 16;
+    __m256i* ptr = reinterpret_cast<__m256i*>(data + idx);
+    __m256i value = _mm256_setzero_si256();
+
+    _mm256_storeu_si256(ptr, value);
+
+    benchmark::DoNotOptimize(ptr);
+  }
+}
+
+static void BM_Storeu256_unaligned_crossing_64byte_border(
+    benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = ((rng() & ((1 << k) - 1)) << 6) + 48;
+    __m256i* ptr = reinterpret_cast<__m256i*>(data + idx);
+    __m256i value = _mm256_setzero_si256();
+
+    _mm256_storeu_si256(ptr, value);
+
+    benchmark::DoNotOptimize(ptr);
+  }
+}
+
+static void BM_Store256_aligned(benchmark::State& state) {
+  size_t k = state.range(0);
+  std::mt19937_64 rng(42);
+
+  for (auto _ : state) {
+    size_t idx = (rng() & ((1 << k) - 1)) << 6;
+    __m256i* ptr = reinterpret_cast<__m256i*>(data + idx);
+    __m256i value = _mm256_setzero_si256();
+
+    _mm256_store_si256(ptr, value);
+
+    benchmark::DoNotOptimize(ptr);
+  }
+}
+
+BENCHMARK(BM_Loadu256_aligned)->ArgNames({"k"})->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Loadu256_unaligned)->ArgNames({"k"})->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Loadu256_unaligned_crossing_64byte_border)
+    ->ArgNames({"k"})
+    ->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Load256_aligned)->ArgNames({"k"})->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Storeu256_aligned)->ArgNames({"k"})->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Storeu256_unaligned)->ArgNames({"k"})->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Storeu256_unaligned_crossing_64byte_border)
+    ->ArgNames({"k"})
+    ->DenseRange(1, 23, 2);
+
+BENCHMARK(BM_Store256_aligned)->ArgNames({"k"})->DenseRange(1, 23, 2);
+
+#endif