contrib/benchmarks/Tensor/bench_reverse.cpp

// Benchmarks for Eigen TensorReverse.
// SPDX-FileCopyrightText: The Eigen Authors
// SPDX-License-Identifier: MPL-2.0

#include <benchmark/benchmark.h>
#include <contrib/Eigen/Tensor>

using namespace Eigen;

typedef float Scalar;

// --- Reverse only the inner-most (contiguous) dimension. SIMD preverse case. ---
static void BM_Reverse_Inner(benchmark::State& state) {
  const int M = state.range(0);
  const int N = state.range(1);

  Tensor<Scalar, 2> A(M, N);
  A.setRandom();

  array<bool, 2> dim_rev = {true, false};

  for (auto _ : state) {
    Tensor<Scalar, 2> B = A.reverse(dim_rev);
    benchmark::DoNotOptimize(B.data());
    benchmark::ClobberMemory();
  }
  state.SetBytesProcessed(state.iterations() * static_cast<int64_t>(M) * N * sizeof(Scalar));
}

// --- Reverse only an outer dimension. Inner dim stays contiguous. ---
static void BM_Reverse_Outer(benchmark::State& state) {
  const int M = state.range(0);
  const int N = state.range(1);

  Tensor<Scalar, 2> A(M, N);
  A.setRandom();

  array<bool, 2> dim_rev = {false, true};

  for (auto _ : state) {
    Tensor<Scalar, 2> B = A.reverse(dim_rev);
    benchmark::DoNotOptimize(B.data());
    benchmark::ClobberMemory();
  }
  state.SetBytesProcessed(state.iterations() * static_cast<int64_t>(M) * N * sizeof(Scalar));
}

// --- Reverse every dimension. ---
static void BM_Reverse_All(benchmark::State& state) {
  const int M = state.range(0);
  const int N = state.range(1);

  Tensor<Scalar, 2> A(M, N);
  A.setRandom();

  array<bool, 2> dim_rev = {true, true};

  for (auto _ : state) {
    Tensor<Scalar, 2> B = A.reverse(dim_rev);
    benchmark::DoNotOptimize(B.data());
    benchmark::ClobberMemory();
  }
  state.SetBytesProcessed(state.iterations() * static_cast<int64_t>(M) * N * sizeof(Scalar));
}

// --- 3D reverse with the inner dim reversed (typical CNN-style layout). ---
static void BM_Reverse_3D_Inner(benchmark::State& state) {
  const int D0 = state.range(0);
  const int D1 = state.range(1);
  const int D2 = state.range(2);

  Tensor<Scalar, 3> A(D0, D1, D2);
  A.setRandom();

  array<bool, 3> dim_rev = {true, false, false};

  for (auto _ : state) {
    Tensor<Scalar, 3> B = A.reverse(dim_rev);
    benchmark::DoNotOptimize(B.data());
    benchmark::ClobberMemory();
  }
  state.SetBytesProcessed(state.iterations() * static_cast<int64_t>(D0) * D1 * D2 * sizeof(Scalar));
}

// Sweep sizes that span L1 (~32 KB), L2 (~256 KB), and LLC (~MBs) for float
// tensors. Bytes per element = 4, so per-side sizes:
//   64x64    = 16 KB (L1)
//   256x256  = 256 KB (L2)
//   1024x1024 = 4 MB (LLC / DRAM)
// --- Writes through a reversed destination (lvalue writePacket path) ---
static void BM_ReverseWrite_Inner(benchmark::State& state) {
  const int M = state.range(0);
  const int N = state.range(1);

  Tensor<Scalar, 2> src(M, N);
  src.setRandom();
  Tensor<Scalar, 2> dst(M, N);
  dst.setZero();

  array<bool, 2> dim_rev = {true, false};

  for (auto _ : state) {
    dst.reverse(dim_rev) = src;
    benchmark::DoNotOptimize(dst.data());
    benchmark::ClobberMemory();
  }

  // Validate outside the timed loop.
  for (int i = 0; i < M; ++i) {
    for (int j = 0; j < N; ++j) {
      if (dst(M - 1 - i, j) != src(i, j)) {
        state.SkipWithError("validation failed");
        return;
      }
    }
  }
  state.SetBytesProcessed(state.iterations() * static_cast<int64_t>(M) * N * sizeof(Scalar));
}

static void BM_ReverseWrite_Outer(benchmark::State& state) {
  const int M = state.range(0);
  const int N = state.range(1);

  Tensor<Scalar, 2> src(M, N);
  src.setRandom();
  Tensor<Scalar, 2> dst(M, N);
  dst.setZero();

  array<bool, 2> dim_rev = {false, true};

  for (auto _ : state) {
    dst.reverse(dim_rev) = src;
    benchmark::DoNotOptimize(dst.data());
    benchmark::ClobberMemory();
  }

  for (int i = 0; i < M; ++i) {
    for (int j = 0; j < N; ++j) {
      if (dst(i, N - 1 - j) != src(i, j)) {
        state.SkipWithError("validation failed");
        return;
      }
    }
  }
  state.SetBytesProcessed(state.iterations() * static_cast<int64_t>(M) * N * sizeof(Scalar));
}

// --- Writes into a reversed destination from a block-preferring source. The
// shuffle routes the assignment through the tiled executor, so the reverse
// destination serves it via writeBlock instead of per-packet index math. ---
static void BM_ReverseWrite_FromShuffle_Inner(benchmark::State& state) {
  const int M = state.range(0);
  const int N = state.range(1);

  Tensor<Scalar, 2> src(N, M);
  src.setRandom();
  Tensor<Scalar, 2> dst(M, N);
  dst.setZero();

  Eigen::array<Index, 2> transpose = {1, 0};
  array<bool, 2> dim_rev = {true, false};

  for (auto _ : state) {
    dst.reverse(dim_rev) = src.shuffle(transpose);
    benchmark::DoNotOptimize(dst.data());
    benchmark::ClobberMemory();
  }

  for (int i = 0; i < M; ++i) {
    for (int j = 0; j < N; ++j) {
      if (dst(M - 1 - i, j) != src(j, i)) {
        state.SkipWithError("validation failed");
        return;
      }
    }
  }
  state.SetBytesProcessed(state.iterations() * static_cast<int64_t>(M) * N * sizeof(Scalar));
}

static void BM_ReverseWrite_FromShuffle_Outer(benchmark::State& state) {
  const int M = state.range(0);
  const int N = state.range(1);

  Tensor<Scalar, 2> src(N, M);
  src.setRandom();
  Tensor<Scalar, 2> dst(M, N);
  dst.setZero();

  Eigen::array<Index, 2> transpose = {1, 0};
  array<bool, 2> dim_rev = {false, true};

  for (auto _ : state) {
    dst.reverse(dim_rev) = src.shuffle(transpose);
    benchmark::DoNotOptimize(dst.data());
    benchmark::ClobberMemory();
  }

  for (int i = 0; i < M; ++i) {
    for (int j = 0; j < N; ++j) {
      if (dst(i, N - 1 - j) != src(j, i)) {
        state.SkipWithError("validation failed");
        return;
      }
    }
  }
  state.SetBytesProcessed(state.iterations() * static_cast<int64_t>(M) * N * sizeof(Scalar));
}

// A broadcast right-hand side is the motivating case for the reverse
// destination's writeBlock: broadcast's packet path pays heavy per-packet
// index math, while its block path is a bulk copy.
static void BM_ReverseWrite_FromBroadcast(benchmark::State& state) {
  const int M = state.range(0);
  const int N = state.range(1);

  Tensor<Scalar, 2> src(M / 4, N);
  src.setRandom();
  Tensor<Scalar, 2> dst(M, N);
  dst.setZero();

  Eigen::array<Index, 2> bcast = {4, 1};
  array<bool, 2> dim_rev = {true, false};

  for (auto _ : state) {
    dst.reverse(dim_rev) = src.broadcast(bcast);
    benchmark::DoNotOptimize(dst.data());
    benchmark::ClobberMemory();
  }

  for (int i = 0; i < M; ++i) {
    for (int j = 0; j < N; ++j) {
      if (dst(M - 1 - i, j) != src(i % (M / 4), j)) {
        state.SkipWithError("validation failed");
        return;
      }
    }
  }
  state.SetBytesProcessed(state.iterations() * static_cast<int64_t>(M) * N * sizeof(Scalar));
}

// Lazy right-hand-side blocks: writeBlock evaluates the re-reversed cwise
// expression straight into the destination box.
static void BM_ReverseWrite_FromShuffleExpr(benchmark::State& state) {
  const int M = state.range(0);
  const int N = state.range(1);

  Tensor<Scalar, 2> src(N, M);
  src.setRandom();
  Tensor<Scalar, 2> dst(M, N);
  dst.setZero();

  Eigen::array<Index, 2> transpose = {1, 0};
  array<bool, 2> dim_rev = {true, false};

  for (auto _ : state) {
    dst.reverse(dim_rev) = src.shuffle(transpose) + src.shuffle(transpose).constant(1.0f);
    benchmark::DoNotOptimize(dst.data());
    benchmark::ClobberMemory();
  }

  for (int i = 0; i < M; ++i) {
    for (int j = 0; j < N; ++j) {
      if (dst(M - 1 - i, j) != src(j, i) + 1.0f) {
        state.SkipWithError("validation failed");
        return;
      }
    }
  }
  state.SetBytesProcessed(state.iterations() * static_cast<int64_t>(M) * N * sizeof(Scalar));
}

// clang-format off
#define REVERSE_SIZES \
  ->Args({64, 64})->Args({256, 256})->Args({1024, 1024})

// 128 KB / 1 MB / 8 MB
#define REVERSE_3D_SIZES \
  ->Args({32, 32, 32})->Args({64, 64, 64})->Args({128, 128, 128})
// clang-format on

BENCHMARK(BM_Reverse_Inner) REVERSE_SIZES;
BENCHMARK(BM_Reverse_Outer) REVERSE_SIZES;
BENCHMARK(BM_Reverse_All) REVERSE_SIZES;
BENCHMARK(BM_Reverse_3D_Inner) REVERSE_3D_SIZES;
BENCHMARK(BM_ReverseWrite_Inner) REVERSE_SIZES;
BENCHMARK(BM_ReverseWrite_Outer) REVERSE_SIZES;
BENCHMARK(BM_ReverseWrite_FromShuffle_Inner) REVERSE_SIZES;
BENCHMARK(BM_ReverseWrite_FromShuffle_Outer) REVERSE_SIZES;
BENCHMARK(BM_ReverseWrite_FromBroadcast) REVERSE_SIZES;
BENCHMARK(BM_ReverseWrite_FromShuffleExpr) REVERSE_SIZES;