// CUTLASS v4.5.1 teaching path. Coverage: parser_only, not GPU-observed.
#include <cutlass/cutlass.h>
#include <cutlass/gemm/device/gemm.h>
#include <cutlass/layout/matrix.h>
#include <cutlass/util/device_memory.h>

#include <iostream>

int main() {
  int const m = 128, n = 128, k = 128;
  cutlass::device_memory::allocation<float> a(m * k);
  cutlass::device_memory::allocation<float> b(k * n);
  cutlass::device_memory::allocation<float> c(m * n);

  using Gemm = cutlass::gemm::device::Gemm<
      float, cutlass::layout::RowMajor,
      float, cutlass::layout::RowMajor,
      float, cutlass::layout::RowMajor>;

  Gemm operation;
  Gemm::Arguments arguments(
      {m, n, k},
      {a.get(), k},
      {b.get(), n},
      {c.get(), n},
      {c.get(), n},
      {1.0F, 0.0F});

  auto status = operation.can_implement(arguments);
  if (status != cutlass::Status::kSuccess) return 2;
  status = operation(arguments);
  if (status != cutlass::Status::kSuccess) return 3;
  std::cout << "CUTLASS GEMM launch accepted; numerical receipt still required\n";
  return 0;
}