devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Remove benchmark::internal::Benchmark* from all benchmarks

Browse Source 
libeigen/eigen!2332

Co-authored-by: Rasmus Munk Larsen <rlarsen@nvidia.com>
This commit is contained in:
Rasmus Munk Larsen
2026-03-20 17:42:07 -07:00
parent 8115b45e50
commit 9d1e5f3915
32 changed files with 477 additions and 576 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
benchmarks/Core/bench_gemv.cpp
View File

@@ -115,43 +115,28 @@ static void BM_GemvAdj(benchmark::State& state) {
// ---------- Size configurations ----------
// All sizes refer to the stored matrix A (m rows, n cols).
static void GemvSizes(::benchmark::Benchmark* b) {
// Square matrices: exercises balanced kernel behavior.
for (int size : {8, 32, 128, 512, 1024}) {
b->Args({size, size});
}
// Tall-thin (m >> n): in ColMajor kernel, the inner vectorized loop over rows
// is long while the outer column loop is short. In RowMajor kernel (transpose),
// there are many rows to process but short dot products.
for (int n : {1, 16}) {
for (int m : {256, 1024}) {
b->Args({m, n});
}
}
// Short-wide (m << n): in ColMajor kernel, the outer column loop is long but
// the inner vectorized loop over rows is short. In RowMajor kernel (transpose),
// there are few rows but long dot products.
for (int m : {1, 16}) {
for (int n : {256, 1024}) {
b->Args({m, n});
}
}
}
// ---------- Register benchmarks ----------
// clang-format off
// Square matrices; tall-thin (m >> n); short-wide (m << n).
#define GEMV_SIZES \
->Args({8, 8})->Args({32, 32})->Args({128, 128})->Args({512, 512})->Args({1024, 1024}) \
->Args({256, 1})->Args({1024, 1})->Args({256, 16})->Args({1024, 16}) \
->Args({1, 256})->Args({1, 1024})->Args({16, 256})->Args({16, 1024})
// Real types: Gemv and GemvTrans exercise the two kernel specializations.
// Conjugation is a no-op for real scalars.
BENCHMARK(BM_Gemv<float>)->Apply(GemvSizes)->Name("Gemv_float");
BENCHMARK(BM_Gemv<double>)->Apply(GemvSizes)->Name("Gemv_double");
BENCHMARK(BM_GemvTrans<float>)->Apply(GemvSizes)->Name("GemvTrans_float");
BENCHMARK(BM_GemvTrans<double>)->Apply(GemvSizes)->Name("GemvTrans_double");
BENCHMARK(BM_Gemv<float>) GEMV_SIZES ->Name("Gemv_float");
BENCHMARK(BM_Gemv<double>) GEMV_SIZES ->Name("Gemv_double");
BENCHMARK(BM_GemvTrans<float>) GEMV_SIZES ->Name("GemvTrans_float");
BENCHMARK(BM_GemvTrans<double>) GEMV_SIZES ->Name("GemvTrans_double");
// Complex types: all four variants exercise distinct kernel code paths.
// Only cfloat is benchmarked since cdouble exercises the same paths but slower.
BENCHMARK(BM_Gemv<std::complex<float>>) GEMV_SIZES ->Name("Gemv_cfloat");
BENCHMARK(BM_GemvTrans<std::complex<float>>) GEMV_SIZES ->Name("GemvTrans_cfloat");
BENCHMARK(BM_GemvConj<std::complex<float>>) GEMV_SIZES ->Name("GemvConj_cfloat");
BENCHMARK(BM_GemvAdj<std::complex<float>>) GEMV_SIZES ->Name("GemvAdj_cfloat");
BENCHMARK(BM_Gemv<std::complex<float>>)->Apply(GemvSizes)->Name("Gemv_cfloat");
BENCHMARK(BM_GemvTrans<std::complex<float>>)->Apply(GemvSizes)->Name("GemvTrans_cfloat");
BENCHMARK(BM_GemvConj<std::complex<float>>)->Apply(GemvSizes)->Name("GemvConj_cfloat");
BENCHMARK(BM_GemvAdj<std::complex<float>>)->Apply(GemvSizes)->Name("GemvAdj_cfloat");
#undef GEMV_SIZES
// clang-format on