Files
eigen/benchmarks/Core/bench_trmv.cpp
Rasmus Munk Larsen 662d5c21ff Optimize SYMV, SYR, SYR2, and TRMV product kernels
libeigen/eigen!2228

Co-authored-by: Rasmus Munk Larsen <rmlarsen@gmail.com>
2026-03-01 19:40:11 -08:00

47 lines
1.8 KiB
C++

// Benchmarks for triangular matrix-vector product (TRMV).
//
// Tests y += triangularView(A) * x for various modes and sizes.
// Exercises TriangularMatrixVector.h kernel.
#include <benchmark/benchmark.h>
#include <Eigen/Core>
using namespace Eigen;
template <typename Scalar>
double trmvFlops(Index n) {
// TRMV: ~n^2 multiply-adds
return (NumTraits<Scalar>::IsComplex ? 8.0 : 2.0) * n * n;
}
// y = triangularView<Mode>(A) * x
template <typename Scalar, unsigned int Mode>
static void BM_TRMV(benchmark::State& state) {
const Index n = state.range(0);
using Mat = Matrix<Scalar, Dynamic, Dynamic>;
using Vec = Matrix<Scalar, Dynamic, 1>;
Mat A = Mat::Random(n, n);
Vec x = Vec::Random(n);
Vec y(n);
for (auto _ : state) {
y.noalias() = A.template triangularView<Mode>() * x;
benchmark::DoNotOptimize(y.data());
benchmark::ClobberMemory();
}
state.counters["GFLOPS"] = benchmark::Counter(trmvFlops<Scalar>(n), benchmark::Counter::kIsIterationInvariantRate,
benchmark::Counter::kIs1000);
}
static void TrmvSizes(::benchmark::Benchmark* b) {
for (int n : {8, 16, 32, 64, 128, 256, 512, 1024, 2048}) b->Arg(n);
}
BENCHMARK(BM_TRMV<float, Lower>)->Apply(TrmvSizes)->Name("TRMV_float_Lower");
BENCHMARK(BM_TRMV<float, Upper>)->Apply(TrmvSizes)->Name("TRMV_float_Upper");
BENCHMARK(BM_TRMV<float, UnitLower>)->Apply(TrmvSizes)->Name("TRMV_float_UnitLower");
BENCHMARK(BM_TRMV<float, UnitUpper>)->Apply(TrmvSizes)->Name("TRMV_float_UnitUpper");
BENCHMARK(BM_TRMV<double, Lower>)->Apply(TrmvSizes)->Name("TRMV_double_Lower");
BENCHMARK(BM_TRMV<double, Upper>)->Apply(TrmvSizes)->Name("TRMV_double_Upper");
BENCHMARK(BM_TRMV<double, UnitLower>)->Apply(TrmvSizes)->Name("TRMV_double_UnitLower");
BENCHMARK(BM_TRMV<double, UnitUpper>)->Apply(TrmvSizes)->Name("TRMV_double_UnitUpper");