use vulkano::buffer::{BufferUsage, CpuAccessibleBuffer, DeviceLocalBuffer, ImmutableBuffer, BufferAccess};
use vulkano::command_buffer::AutoCommandBufferBuilder;
use vulkano::descriptor::descriptor_set::{PersistentDescriptorSet, StdDescriptorPoolAlloc};
use vulkano::device::{Device, DeviceExtensions, QueuesIter, Queue};
use vulkano::instance::{Instance, InstanceExtensions, PhysicalDevice, QueueFamily};
use vulkano::pipeline::ComputePipeline;
use vulkano::sync::GpuFuture;
use vulkano::sync;
use std::time::SystemTime;
use std::sync::Arc;
use std::ffi::CStr;
use std::path::PathBuf;
use shade_runner as sr;
use image::{DynamicImage, ImageBuffer};
use image::GenericImageView;
use vulkano::descriptor::pipeline_layout::PipelineLayout;
use image::GenericImage;
use shade_runner::{ComputeLayout, CompileError};
use vulkano::descriptor::descriptor_set::PersistentDescriptorSetBuf;
use shaderc::CompileOptions;

pub struct VkProcessor<'a> {
    pub instance: Arc<Instance>,
    pub physical: PhysicalDevice<'a>,
    pub pipeline: Option<Arc<ComputePipeline<PipelineLayout<shade_runner::layouts::ComputeLayout>>>>,
    pub device: Arc<Device>,
    pub queues: QueuesIter,
    pub queue: Arc<Queue>,
    pub set: Option<Arc<PersistentDescriptorSet<std::sync::Arc<ComputePipeline<PipelineLayout<shade_runner::layouts::ComputeLayout>>>, ((((), PersistentDescriptorSetBuf<std::sync::Arc<vulkano::buffer::cpu_access::CpuAccessibleBuffer<[u8]>>>), PersistentDescriptorSetBuf<std::sync::Arc<vulkano::buffer::cpu_access::CpuAccessibleBuffer<[u8]>>>), PersistentDescriptorSetBuf<std::sync::Arc<vulkano::buffer::cpu_access::CpuAccessibleBuffer<[u32]>>>)>>>,
    pub image_buffer: Vec<u8>,
    pub img_buffers: Vec<Arc<CpuAccessibleBuffer<[u8]>>>,
    pub settings_buffer: Option<Arc<CpuAccessibleBuffer<[u32]>>>,
    pub xy: (u32, u32),
}

impl<'a> VkProcessor<'a> {
    pub fn new(instance : &'a Arc<Instance>) -> VkProcessor<'a> {

        let physical = PhysicalDevice::enumerate(instance).next().unwrap();
        let queue_family = physical.queue_families().find(|&q| q.supports_compute()).unwrap();
        let (device, mut queues) = Device::new(physical,
                                               physical.supported_features(),
                                               &DeviceExtensions::none(),
                                               [(queue_family, 0.5)].iter().cloned()).unwrap();

        VkProcessor {
            instance: instance.clone(),
            physical: physical.clone(),
            pipeline: Option::None,
            device: device,
            queue: queues.next().unwrap(),
            queues: queues,
            set: Option::None,
            image_buffer: Vec::new(),
            img_buffers: Vec::new(),
            settings_buffer: Option::None,
            xy: (0,0),
        }

    }

    pub fn compile_kernel(&mut self, filename: String) {

        let project_root =
            std::env::current_dir()
                .expect("failed to get root directory");

        let mut compute_path = project_root.clone();
        compute_path.push(PathBuf::from("resources/shaders/"));
        compute_path.push(PathBuf::from(filename));


        let mut options = CompileOptions::new().ok_or(CompileError::CreateCompiler).unwrap();
        options.add_macro_definition("SETTING_POS_X", Some("0"));
        options.add_macro_definition("SETTING_POS_Y", Some("1"));
        options.add_macro_definition("SETTING_BUCKETS_START", Some("2"));
        options.add_macro_definition("SETTING_BUCKETS_LEN", Some("2"));

        let shader =
            sr::load_compute_with_options(compute_path, options)
                .expect("Failed to compile");

        let vulkano_entry =
            sr::parse_compute(&shader)
                .expect("failed to parse");

        let x = unsafe {
            vulkano::pipeline::shader::ShaderModule::from_words(self.device.clone(), &shader.compute)
        }.unwrap();

        let pipeline = Arc::new({
            unsafe {
                ComputePipeline::new(self.device.clone(), &x.compute_entry_point(
                    CStr::from_bytes_with_nul_unchecked(b"main\0"),
                    vulkano_entry.compute_layout), &(),
                ).unwrap()
            }
        });

        self.pipeline = Some(pipeline);
    }

    pub fn load_buffers(&mut self, image_filename: String)
    {
        let project_root =
            std::env::current_dir()
                .expect("failed to get root directory");

        let mut compute_path = project_root.clone();
        compute_path.push(PathBuf::from("resources/images/"));
        compute_path.push(PathBuf::from(image_filename));

        let img = image::open(compute_path).expect("Couldn't find image");

        self.xy = img.dimensions();

        let data_length = self.xy.0 * self.xy.1 * 4;
        let pixel_count = img.raw_pixels().len();
        println!("Pixel count {}", pixel_count);

        if pixel_count != data_length as usize {
            println!("Creating apha channel...");
            for i in img.raw_pixels().iter() {
                if (self.image_buffer.len() + 1) % 4 == 0 {
                    self.image_buffer.push(255);
                }
                self.image_buffer.push(*i);
            }
            self.image_buffer.push(255);
        } else {
            self.image_buffer = img.raw_pixels();
        }

        println!("Buffer length {}", self.image_buffer.len());
        println!("Size {:?}", self.xy);

        println!("Allocating Buffers...");

        // Pull out the image data and place it in a buffer for the kernel to write to and for us to read from
        let write_buffer = {
            let mut buff = self.image_buffer.iter();
            let data_iter = (0..data_length).map(|n| *(buff.next().unwrap()));
            CpuAccessibleBuffer::from_iter(self.device.clone(), BufferUsage::all(), data_iter).unwrap()
        };


        // Pull out the image data and place it in a buffer for the kernel to read from
        let read_buffer = {
            let mut buff = self.image_buffer.iter();
            let data_iter = (0..data_length).map(|n| *(buff.next().unwrap()));
            CpuAccessibleBuffer::from_iter(self.device.clone(), BufferUsage::all(), data_iter).unwrap()
        };


        // A buffer to hold many i32 values to use as settings
        let settings_buffer = {
            let vec = vec![self.xy.0, self.xy.1];
            let mut buff = vec.iter();
            let data_iter =
                (0..2).map(|n| *(buff.next().unwrap()));
            CpuAccessibleBuffer::from_iter(self.device.clone(),
                                           BufferUsage::all(),
                                           data_iter).unwrap()
        };


        println!("Done");

        // Create the data descriptor set for our previously created shader pipeline
        let mut set =
            PersistentDescriptorSet::start(self.pipeline.clone().unwrap().clone(), 0)
            .add_buffer(write_buffer.clone()).unwrap()
            .add_buffer(read_buffer).unwrap()
            .add_buffer(settings_buffer).unwrap();


        self.set = Some(Arc::new(set.build().unwrap()));

        self.img_buffers.push(write_buffer);
    }

    pub fn run_kernel(&mut self) {

        println!("Running Kernel...");

        // The command buffer I think pretty much serves to define what runs where for how many times
        let command_buffer =
            AutoCommandBufferBuilder::primary_one_time_submit(self.device.clone(),self.queue.family()).unwrap()
            .dispatch([self.xy.0, self.xy.1, 1], self.pipeline.clone().unwrap().clone(), self.set.clone().unwrap().clone(), ()).unwrap()
            .build().unwrap();

        // Create a future for running the command buffer and then just fence it
        let future = sync::now(self.device.clone())
            .then_execute(self.queue.clone(), command_buffer).unwrap()
            .then_signal_fence_and_flush().unwrap();

        // I think this is redundant and returns immediately
        future.wait(None).unwrap();
        println!("Done running kernel");
    }

    pub fn read_image(&self) -> Vec<u8> {

        // The buffer is sync'd so we can just read straight from the handle
        let mut data_buffer_content = self.img_buffers.get(0).unwrap().read().unwrap();

        println!("Reading output");

        let mut image_buffer = Vec::new();

        for y in 0..self.xy.1 {
            for x in 0..self.xy.0 {

                let r = data_buffer_content[((self.xy.0 * y + x) * 4 + 0) as usize] as u8;
                let g = data_buffer_content[((self.xy.0 * y + x) * 4 + 1) as usize] as u8;
                let b = data_buffer_content[((self.xy.0 * y + x) * 4 + 2) as usize] as u8;
                let a = data_buffer_content[((self.xy.0 * y + x) * 4 + 3) as usize] as u8;

                image_buffer.push(r);
                image_buffer.push(g);
                image_buffer.push(b);
                image_buffer.push(a);

                //self.img.unwrap().put_pixel(x, y, image::Rgba([r, g, b, a]))
            }
        }

        image_buffer
    }

    pub fn save_image(&self) {
        println!("Saving output");

        let img_data = self.read_image();

        let img = ImageBuffer::from_fn(self.xy.0, self.xy.1, |x, y| {

            let r = img_data[((self.xy.0 * y + x) * 4 + 0) as usize] as u8;
            let g = img_data[((self.xy.0 * y + x) * 4 + 1) as usize] as u8;
            let b = img_data[((self.xy.0 * y + x) * 4 + 2) as usize] as u8;
            let a = img_data[((self.xy.0 * y + x) * 4 + 3) as usize] as u8;

            image::Rgba([r, g, b, a])
        });

        img.save(format!("output/{}.png", SystemTime::now().duration_since(SystemTime::UNIX_EPOCH).unwrap().as_secs()));
    }
}