[ntuple] Check type in RValue::GetRef<T> and GetPtr<T> (version 2)

[hahnjo]

This is a bit tricky because RFieldBase.hxx cannot depend on RField specializations in RField.hxx. We considered factoring TypeName() into a separate RFieldTypeName trait, but that would have still required including a bunch of STL headers to specialize for all supported containers. In the end, we settled on taking the typeid and passing the std::type_info to Internal::GetRenormalizedTypeName.

However, this comes with a run time cost, especially for fundamental types. In a microbenchmark, I measure 24ns per (successful) call to REntry::GetPtr<std::int32_t> while a naive implementation of always calling GetRenormalizedTypeName takes ~190ns per RValue::GetPtr. For more complex types, for example ROOT::Math::PtEtaPhiMVector with a number of templates, the difference is less pronounced if at all measurable with around 2000ns both.

To avoid this cost on repeated calls, we cache the std::type_info of matching types and try if it's equal to the typeid, which is basically a pointer comparison. This brings the (amortized) cost down to 8ns for any type! In the future, we will want to reuse the caching approach of RValue::EnsureMatchingType also for REntry.

Closes #18316

[hahnjo]

For future reference, here are the microbenchmarks cited in the commit message:

#include <Math/Vector4D.h>

#include <ROOT/REntry.hxx>
#include <ROOT/RFieldBase.hxx>
#include <ROOT/RNTupleModel.hxx>

#include <benchmark/benchmark.h>

#include <cstdint>
#include <utility>

static void REntry_GetPtr_int32(benchmark::State& state) {
  auto model = ROOT::RNTupleModel::CreateBare();
  auto field = ROOT::RFieldBase::Create("i32", "std::int32_t").Unwrap();
  model->AddField(std::move(field));
  model->Freeze();
  auto token = model->GetToken("i32");
  auto entry = model->CreateEntry();
  for (auto _ : state) {
    benchmark::DoNotOptimize(entry->GetPtr<std::int32_t>(token));
  }
}
BENCHMARK(REntry_GetPtr_int32);

static void REntry_GetPtr_LorentzVector(benchmark::State& state) {
  auto model = ROOT::RNTupleModel::CreateBare();
  auto field = ROOT::RFieldBase::Create("v", "ROOT::Math::PtEtaPhiMVector").Unwrap();
  model->AddField(std::move(field));
  model->Freeze();
  auto token = model->GetToken("v");
  auto entry = model->CreateEntry();
  for (auto _ : state) {
    benchmark::DoNotOptimize(entry->GetPtr<ROOT::Math::PtEtaPhiMVector>(token));
  }
}
BENCHMARK(REntry_GetPtr_LorentzVector);

static void RValue_GetPtr_int32(benchmark::State& state) {
  auto field = ROOT::RFieldBase::Create("i32", "std::int32_t").Unwrap();
  auto value = field->CreateValue();
  for (auto _ : state) {
    benchmark::DoNotOptimize(value.GetPtr<std::int32_t>());
  }
}
BENCHMARK(RValue_GetPtr_int32);

static void RValue_GetPtr_LorentzVector(benchmark::State& state) {
  auto field = ROOT::RFieldBase::Create("v", "ROOT::Math::PtEtaPhiMVector").Unwrap();
  auto value = field->CreateValue();
  for (auto _ : state) {
    benchmark::DoNotOptimize(value.GetPtr<ROOT::Math::PtEtaPhiMVector>());
  }
}
BENCHMARK(RValue_GetPtr_LorentzVector);

BENCHMARK_MAIN();

[github-actions]

Test Results

    21 files      21 suites   3d 11h 23m 40s ⏱️
 3 250 tests  3 250 ✅ 0 💤 0 ❌
66 523 runs  66 523 ✅ 0 💤 0 ❌

Results for commit 5464821.

♻️ This comment has been updated with latest results.

[jblomer]

Nice!

[hahnjo]

The latest push also avoids extra type checks in RNTupleView<T>, which works around a renormalization bug that I'm fixing separately in #19550 (thanks @bellenot for confirming my suspicion).

[vepadulano]

Cool thanks!