#pragma once

#ifndef TEST_GENERIC_IN_LOCAL
#define TEST_GENERIC_IN_LOCAL 0
#endif

#include <algorithm>
#include <cassert>
#include <iostream>
#include <numeric>
#include <vector>

#include <sycl/detail/core.hpp>

#include <sycl/atomic_ref.hpp>
#include <sycl/usm.hpp>

using namespace sycl;

template <template <typename, memory_order, memory_scope, access::address_space>
          class AtomicRef,
          access::address_space space, typename T,
          memory_order order = memory_order::relaxed,
          memory_scope scope = memory_scope::device>
void min_local_test(queue q, size_t N) {
  T initial = std::numeric_limits<T>::max();
  T cum = initial;
  std::vector<T> output(N);
  std::fill(output.begin(), output.end(), T(123456));
  {
    buffer<T> cum_buf(&cum, 1);
    buffer<T> output_buf(output.data(), output.size());
    q.submit([&](handler &cgh) {
       auto cum = cum_buf.template get_access<access::mode::read_write>(cgh);
       auto out =
           output_buf.template get_access<access::mode::discard_write>(cgh);
       local_accessor<T, 1> loc(1, cgh);

       cgh.parallel_for(nd_range<1>(N, N), [=](nd_item<1> it) {
         int gid = it.get_global_id(0);
         if (gid == 0)
           loc[0] = initial;
         it.barrier(access::fence_space::local_space);
         auto atm = AtomicRef < T,
              (order == memory_order::acquire || order == memory_order::release)
                  ? memory_order::relaxed
                  : order,
              scope, space > (loc[0]);
         out[gid] = atm.fetch_min(T(gid), order);
         it.barrier(access::fence_space::local_space);
         if (gid == 0)
           cum[0] = loc[0];
       });
     }).wait_and_throw();
  }

  // Final value should be equal to 0
  assert(cum == 0);

  // Only one work-item should have received the initial value
  assert(std::count(output.begin(), output.end(), initial) == 1);

  // fetch_min returns original value
  // Intermediate values should all be <= initial value
  for (int i = 0; i < N; ++i) {
    assert(output[i] <= initial);
  }
}

template <template <typename, memory_order, memory_scope, access::address_space>
          class AtomicRef,
          access::address_space space, typename T,
          memory_order order = memory_order::relaxed,
          memory_scope scope = memory_scope::device>
void min_global_test(queue q, size_t N) {
  T initial = std::numeric_limits<T>::max();
  T val = initial;
  std::vector<T> output(N);
  std::fill(output.begin(), output.end(), 0);
  {
    buffer<T> val_buf(&val, 1);
    buffer<T> output_buf(output.data(), output.size());

    q.submit([&](handler &cgh) {
      auto val = val_buf.template get_access<access::mode::read_write>(cgh);
      auto out =
          output_buf.template get_access<access::mode::discard_write>(cgh);
      cgh.parallel_for(range<1>(N), [=](item<1> it) {
        int gid = it.get_id(0);
        auto atm = AtomicRef < T,
             (order == memory_order::acquire || order == memory_order::release)
                 ? memory_order::relaxed
                 : order,
             scope, space > (val[0]);
        out[gid] = atm.fetch_min(T(gid), order);
      });
    });
  }

  // Final value should be equal to 0
  assert(val == 0);

  // Only one work-item should have received the initial value
  assert(std::count(output.begin(), output.end(), initial) == 1);

  // fetch_min returns original value
  // Intermediate values should all be <= initial value
  for (int i = 0; i < N; ++i) {
    assert(output[i] <= initial);
  }
}

template <template <typename, memory_order, memory_scope, access::address_space>
          class AtomicRef,
          access::address_space space, typename T,
          memory_order order = memory_order::relaxed,
          memory_scope scope = memory_scope::device>
void min_global_test_usm_shared(queue q, size_t N) {
  T initial = std::numeric_limits<T>::max();
  T *val = malloc_shared<T>(1, q);
  val[0] = initial;
  T *output = malloc_shared<T>(N, q);
  T *output_begin = &output[0], *output_end = &output[N];
  std::fill(output_begin, output_end, 0);
  {
    q.submit([&](handler &cgh) {
       cgh.parallel_for(range<1>(N), [=](item<1> it) {
         int gid = it.get_id(0);
         auto atm = AtomicRef < T,
              (order == memory_order::acquire || order == memory_order::release)
                  ? memory_order::relaxed
                  : order,
              scope, space > (val[0]);
         output[gid] = atm.fetch_min(T(gid), order);
       });
     }).wait_and_throw();
  }

  // Final value should be equal to 0.
  assert(val[0] == 0);

  // Only one work-item should have received the initial value.
  assert(std::count(output_begin, output_end, initial) == 1);

  // fetch_min returns original value.
  // Intermediate values should all be <= initial value.
  for (int i = 0; i < N; ++i) {
    assert(output[i] <= initial && output[i] >= val[0]);
  }

  free(val, q);
  free(output, q);
}

template <access::address_space space, typename T,
          memory_order order = memory_order::relaxed,
          memory_scope scope = memory_scope::device>
void min_test(queue q, size_t N) {
  constexpr bool do_local_tests =
      space == access::address_space::local_space ||
      (space == access::address_space::generic_space && TEST_GENERIC_IN_LOCAL);
  constexpr bool do_global_tests =
      space == access::address_space::global_space ||
      (space == access::address_space::generic_space && !TEST_GENERIC_IN_LOCAL);
  constexpr bool do_ext_tests = space != access::address_space::generic_space;
  bool do_usm_tests = q.get_device().has(aspect::usm_shared_allocations);
  if constexpr (do_local_tests) {
#ifdef RUN_DEPRECATED
    if constexpr (do_ext_tests) {
      min_local_test<::sycl::ext::oneapi::atomic_ref, space, T, order, scope>(
          q, N);
    }
#else
    min_local_test<::sycl::atomic_ref, space, T, order, scope>(q, N);
#endif
  }
  if constexpr (do_global_tests) {
#ifdef RUN_DEPRECATED
    if constexpr (do_ext_tests) {
      min_global_test<::sycl::ext::oneapi::atomic_ref, space, T, order, scope>(
          q, N);
      if (do_usm_tests) {
        min_global_test_usm_shared<::sycl::ext::oneapi::atomic_ref, space, T,
                                   order, scope>(q, N);
      }
    }
#else
    min_global_test<::sycl::atomic_ref, space, T, order, scope>(q, N);
    if (do_usm_tests) {
      min_global_test_usm_shared<::sycl::atomic_ref, space, T, order, scope>(q,
                                                                             N);
    }
#endif
  }
}

template <access::address_space space, typename T,
          memory_order order = memory_order::relaxed>
void min_test_scopes(queue q, size_t N) {
  std::vector<memory_scope> scopes =
      q.get_device().get_info<info::device::atomic_memory_scope_capabilities>();
  if (std::find(scopes.begin(), scopes.end(), memory_scope::work_group) !=
      scopes.end()) {
    min_test<space, T, order, memory_scope::work_group>(q, N);
  }
  if (std::find(scopes.begin(), scopes.end(), memory_scope::sub_group) !=
      scopes.end()) {
    min_test<space, T, order, memory_scope::sub_group>(q, N);
  }
  min_test<space, T, order, memory_scope::device>(q, N);
}

template <access::address_space space, typename T>
void min_test_orders_scopes(queue q, size_t N) {
  std::vector<memory_order> orders =
      q.get_device().get_info<info::device::atomic_memory_order_capabilities>();
  if (std::find(orders.begin(), orders.end(), memory_order::acq_rel) !=
      orders.end()) {
    min_test_scopes<space, T, memory_order::acq_rel>(q, N);
  }
  if (std::find(orders.begin(), orders.end(), memory_order::acquire) !=
      orders.end()) {
    min_test_scopes<space, T, memory_order::acquire>(q, N);
  }
  if (std::find(orders.begin(), orders.end(), memory_order::release) !=
      orders.end()) {
    min_test_scopes<space, T, memory_order::release>(q, N);
  }
  min_test_scopes<space, T, memory_order::relaxed>(q, N);
}

template <access::address_space space> void min_test_all() {
  queue q;

  constexpr int N = 32;
#ifdef FULL_ATOMIC64_COVERAGE
  min_test_orders_scopes<space, double>(q, N);
  if constexpr (sizeof(long) == 8) {
    min_test_orders_scopes<space, long>(q, N);
    min_test_orders_scopes<space, unsigned long>(q, N);
  }
  if constexpr (sizeof(long long) == 8) {
    min_test_orders_scopes<space, long long>(q, N);
    min_test_orders_scopes<space, unsigned long long>(q, N);
  }
#endif
  min_test_orders_scopes<space, float>(q, N);
#ifdef FULL_ATOMIC32_COVERAGE
  min_test_orders_scopes<space, int>(q, N);
  min_test_orders_scopes<space, unsigned int>(q, N);
  if constexpr (sizeof(long) == 4) {
    min_test_orders_scopes<space, long>(q, N);
    min_test_orders_scopes<space, unsigned long>(q, N);
  }
#endif

#ifdef FULL_ATOMIC16_COVERAGE
  if (!q.get_device().has(aspect::ext_oneapi_atomic16)) {
    std::cout << "Skipping ext_oneapi_atomic16 tests\n";
    return;
  }

  min_test_orders_scopes<space, sycl::half>(q, N);
#endif

  std::cout << "Test passed." << std::endl;
}