/*
 * Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 *
 * SPDX-License-Identifier: MIT
 */

#include <hip_test_common.hh>
#include <hip_array_common.hh>
#include <vector>
#include <iostream>

#pragma clang diagnostic ignored "-Wunused-variable"
#pragma clang diagnostic ignored "-Wunused-parameter"

template <typename T> __global__ void tex1dKernelFetch(T* val, hipTextureObject_t obj, int N) {
#if !__HIP_NO_IMAGE_SUPPORT
  int k = blockIdx.x * blockDim.x + threadIdx.x;
  if (k < N) {
    val[k] = tex1Dfetch<T>(obj, k);
  }
#endif
}

template <typename T, typename std::enable_if<rank<T>() == 1>::type* = nullptr>
static inline void printVector(T& val) {
  using B = decltype(T::x);
  constexpr bool isChar = std::is_same<B, char>::value || std::is_same<B, unsigned char>::value;
  std::cout << "(";
  std::cout << (isChar ? static_cast<int>(val.x) : val.x);
  std::cout << ")";
}

template <typename T, typename std::enable_if<rank<T>() == 2>::type* = nullptr>
static inline void printVector(T& val) {
  using B = decltype(T::x);
  constexpr bool isChar = std::is_same<B, char>::value || std::is_same<B, unsigned char>::value;
  std::cout << "(";
  std::cout << (isChar ? static_cast<int>(val.x) : val.x);
  std::cout << ", " << (isChar ? static_cast<int>(val.y) : val.y);
  std::cout << ")";
}

template <typename T, typename std::enable_if<rank<T>() == 4>::type* = nullptr>
static inline void printVector(T& val) {
  using B = decltype(T::x);
  constexpr bool isChar = std::is_same<B, char>::value || std::is_same<B, unsigned char>::value;
  std::cout << "(";
  std::cout << (isChar ? static_cast<int>(val.x) : val.x);
  std::cout << ", " << (isChar ? static_cast<int>(val.y) : val.y);
  std::cout << ", " << (isChar ? static_cast<int>(val.z) : val.z);
  std::cout << ", " << (isChar ? static_cast<int>(val.w) : val.w);
  std::cout << ")";
}

template <typename T> bool runTest() {
  const int N = 1024;
  bool testResult = true;
  // Allocating the required buffer on gpu device
  T *texBuf, *texBufOut;
  T val[N], output[N];
  auto err = hipGetLastError();  // Clear err due to negative tests
  memset(output, 0, sizeof(output));
  std::srand(std::time(nullptr));  // use current time as seed for random generator

  for (int i = 0; i < N; i++) {
    initVal<T>(val[i]);
  }

  HIP_CHECK(hipMalloc(&texBuf, N * sizeof(T)));
  HIP_CHECK(hipMalloc(&texBufOut, N * sizeof(T)));
  HIP_CHECK(hipMemcpy(texBuf, val, N * sizeof(T), hipMemcpyHostToDevice));
  HIP_CHECK(hipMemset(texBufOut, 0, N * sizeof(T)));
  hipResourceDesc resDescLinear;

  memset(&resDescLinear, 0, sizeof(resDescLinear));
  resDescLinear.resType = hipResourceTypeLinear;
  resDescLinear.res.linear.devPtr = texBuf;
  resDescLinear.res.linear.desc = hipCreateChannelDesc<T>();
  resDescLinear.res.linear.sizeInBytes = N * sizeof(T);

  hipTextureDesc texDesc;
  memset(&texDesc, 0, sizeof(texDesc));
  texDesc.readMode = hipReadModeElementType;
  texDesc.addressMode[0] = hipAddressModeClamp;

  // Creating texture object
  hipTextureObject_t texObj = 0;
  HIP_CHECK(hipCreateTextureObject(&texObj, &resDescLinear, &texDesc, NULL));

  dim3 dimBlock(64, 1, 1);
  dim3 dimGrid((N + dimBlock.x - 1) / dimBlock.x, 1, 1);

  hipLaunchKernelGGL(tex1dKernelFetch<T>, dimGrid, dimBlock, 0, 0, texBufOut, texObj, N);
  HIP_CHECK(hipGetLastError());
  HIP_CHECK(hipDeviceSynchronize());

  HIP_CHECK(hipMemcpy(output, texBufOut, N * sizeof(T), hipMemcpyDeviceToHost));

  for (int i = 0; i < N; i++) {
    if (!isEqual(output[i], val[i])) {
      std::cout << "output[" << i << "]= ";
      printVector<T>(output[i]);
      std::cout << ", expected val[" << i << "]= ";
      printVector<T>(val[i]);
      std::cout << "\n";
      testResult = false;
      break;
    }
  }
  HIP_CHECK(hipDestroyTextureObject(texObj));
  HIP_CHECK(hipFree(texBuf));
  HIP_CHECK(hipFree(texBufOut));

  REQUIRE(testResult == true);
  return testResult;
}

HIP_TEST_CASE(Unit_hipTextureFetch_vector) {
  CHECK_IMAGE_SUPPORT

  // test for char
  runTest<char1>();
  runTest<char2>();
  runTest<char4>();

  // test for uchar
  runTest<uchar1>();
  runTest<uchar2>();
  runTest<uchar4>();

  // test for short
  runTest<short1>();
  runTest<short2>();
  runTest<short4>();

  // test for ushort
  runTest<ushort1>();
  runTest<ushort2>();
  runTest<ushort4>();

  // test for int
  runTest<int1>();
  runTest<int2>();
  runTest<int4>();

  // test for unsigned int
  runTest<uint1>();
  runTest<uint2>();
  runTest<uint4>();

  // test for float
  runTest<float1>();
  runTest<float2>();
  runTest<float4>();
}