使用降低精度进行优化

struct ReducedPrecisionSupportInfo {
  // Arithmetic support
  bool has_8_bit_int_ = false;
  bool has_16_bit_int_ = false;
  bool has_16_bit_float_ = false;
  // Storage support
  bool has_16_bit_SSBO_ = false;
  bool has_16_bit_UBO_ = false;
  bool has_16_bit_push_ = false;
  bool has_16_bit_input_output_ = false;
  // Use 16-bit floats if we have arithmetic
  // support and at least SSBO storage support.
  bool use_16bit_floats_ = false;
};

void CheckFormatSupport(VkPhysicalDevice physical_device,
    ReducedPrecisionSupportInfo &info) {

  // Retrieve the device extension list so we
  // can check for our desired extensions.
  uint32_t device_extension_count;
  vkEnumerateDeviceExtensionProperties(physical_device, nullptr,
      &device_extension_count, nullptr);
  std::vector<VkExtensionProperties> device_extensions(device_extension_count);
  vkEnumerateDeviceExtensionProperties(physical_device, nullptr,
      &device_extension_count, device_extensions.data());

  bool has_16_8_extension = HasDeviceExtension("VK_KHR_shader_float16_int8",
      device_extensions);

  // Initialize the device features structure and
  // chain the storage features structure and 8/16-bit
  // support structure if applicable.
  VkPhysicalDeviceFeatures2 device_features;
  memset(&device_features, 0, sizeof(device_features));
  device_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;

  VkPhysicalDeviceShaderFloat16Int8Features f16_int8_features;
  memset(&f16_int8_features, 0, sizeof(f16_int8_features));
  f16_int8_features.sType =
      VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR;

  VkPhysicalDevice16BitStorageFeatures storage_features;
  memset(&storage_features, 0, sizeof(storage_features));
  storage_features.sType =
      VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES;
  device_features.pNext = &storage_features;

  if (has_16_8_extension) {
    storage_features.pNext = &f16_int8_features;
  }

  vkGetPhysicalDeviceFeatures2(physical_device, &device_features);

  // Parse the storage features and determine
  // what kinds of 16-bit storage access are available.
  if (storage_features.storageBuffer16BitAccess ||
      storage_features.uniformAndStorageBuffer16BitAccess) {
    info.has_16_bit_SSBO_ = true;
  }
  info.has_16_bit_UBO_ = storage_features.uniformAndStorageBuffer16BitAccess;
  info.has_16_bit_push_ = storage_features.storagePushConstant16;
  info.has_16_bit_input_output_ = storage_features.storageInputOutput16;

  info.has_16_bit_int_ = device_features.features.shaderInt16;
  if (has_16_8_extension) {
    info.has_16_bit_float_ = f16_int8_features.shaderFloat16;
    info.has_8_bit_int_ = f16_int8_features.shaderInt8;
  }

  // Get arithmetic and at least some form of storage
  // support before enabling 16-bit float usage.
  if (info.has_16_bit_float_ && info.has_16_bit_SSBO_) {
    info.use_16bit_floats_ = true;
  }
}

// Enable 16-bit formats in storage and uniform buffers.
#extension GL_EXT_shader_16bit_storage : require
// Enable 8-bit formats in storage and uniform buffers.
#extension GL_EXT_shader_8bit_storage : require

mediump vec4 my_vector; // Suggest 16-bit half precision
highp mat4 my_matrix;   // Suggest 32-bit single precision

#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require

float16_t   f16vec2     f16vec3    f16vec4
f16mat2     f16mat3     f16mat4
f16mat2x2   f16mat2x3   f16mat2x4
f16mat3x2   f16mat3x3   f16mat3x4
f16mat4x2   f16mat4x3   f16mat4x4

int16_t     i16vec2     i16vec3    i16vec4
uint16_t    u16vec2     u16vec3    u16vec4

// Four element vector and four-by-four matrix types
min16float4 my_vector4;
min16float4x4 my_matrix4x4;

// Four element vector and four-by-four matrix types
half4 my_vector4;
half4x4 my_matrix4x4;