made buffer loading much faster

src/main.rs

@@ -20,7 +20,7 @@ use sfml::system::Vector2 as sfVec2;
 use sfml::window::*;
 use sfml::window::{Event, Key, Style};
 
-use vulkano::buffer::{BufferUsage, CpuAccessibleBuffer};
+use vulkano::buffer::{BufferUsage, CpuAccessibleBuffer, DeviceLocalBuffer, ImmutableBuffer, BufferAccess};
 use vulkano::command_buffer::AutoCommandBufferBuilder;
 use vulkano::descriptor::descriptor_set::PersistentDescriptorSet;
 use vulkano::device::{Device, DeviceExtensions};
@@ -31,6 +31,7 @@ use vulkano::sync;
 use std::sync::Arc;
 use std::fs;
 use std::path::PathBuf;
+use std::result;
 
 use crate::input::Input;
 use crate::slider::Slider;
@@ -127,30 +128,6 @@ impl<'t> Effect for Edge<'t> {
 }
 // =================================================
 
-fn surrounding_pixels(x: u32, y: u32, img: &DynamicImage) -> Vec<image::Rgba<u8>> {
-
-    let mut pixels: Vec<image::Rgba<u8>> = Vec::new();
-
-    if img.in_bounds(x+1, y+1) {pixels.push(img.get_pixel(x+1, y+1))}
-    if img.in_bounds(x+1, y)   {pixels.push(img.get_pixel(x+1, y))}
-    if img.in_bounds(x,   y+1) {pixels.push(img.get_pixel(x, y+1))}
-
-    if x > 0 {
-        if img.in_bounds(x-1, y+1) {pixels.push(img.get_pixel(x-1, y+1))}
-        if img.in_bounds(x-1, y)   {pixels.push(img.get_pixel(x-1, y))}
-    }
-
-    if y > 0 {
-        if img.in_bounds(x+1, y-1) {pixels.push(img.get_pixel(x+1, y-1))}
-        if img.in_bounds(x,   y-1) {pixels.push(img.get_pixel(x, y-1))}
-        if x > 0 {
-            if img.in_bounds(x - 1, y - 1) { pixels.push(img.get_pixel(x - 1, y - 1)) }
-        }
-    }
-
-    pixels
-}
-
 fn main() {
 
     // Create the vulkan instance, device, and device queue
@@ -162,6 +139,7 @@ fn main() {
                                            &DeviceExtensions::none(),
                                            [(queue_family, 0.5)].iter().cloned()).unwrap();
     let queue = queues.next().unwrap();
+
     println!("Device initialized");
 
     let project_root = std::env::current_dir().expect("failed to get root directory");
@@ -187,39 +165,47 @@ fn main() {
     });
 
     // Load up the input image, determine some details
-    let mut img = image::open("resources/images/funky-bird.jpg").unwrap();
+    let mut img = image::open("resources/images/test.png").unwrap();
     let xy = img.dimensions();
     let data_length = xy.0*xy.1*4;
 
-    let mut image_buffer = Vec::new();
+    let pixel_count = img.raw_pixels().len();
+    println!("Pixel count {}", pixel_count);
 
-    for i in img.raw_pixels().iter() {
+    let mut image_buffer = Vec::new();
-        if (image_buffer.len() + 1) % 4 == 0 {
+    if pixel_count != data_length as usize {
-            image_buffer.push(255);
+        for i in img.raw_pixels().iter() {
+            if (image_buffer.len() + 1) % 4 == 0 {
+                image_buffer.push(255);
+            }
+            image_buffer.push(*i);
         }
-        image_buffer.push(*i);
+        image_buffer.push(255);
+    }
+    else {
+        image_buffer = img.raw_pixels();
     }
-    image_buffer.push(255);
 
-    println!("Buffer length {}", data_length);
     println!("Buffer length {}", image_buffer.len());
     println!("Size {:?}", xy);
 
+    println!("Allocating Buffers...");
     {
-        // Pull out the image data and place it in a sync'd CPU<->GPU buffer
+        // Pull out the image data and place it in a buffer for the kernel to read from
-        let data_buffer = {
+        let read_buffer = {
-
+            let q = ImmutableBuffer::from_data(image_buffer.clone(), BufferUsage::all(), queue.clone()).unwrap();
-            let mut buff = image_buffer.iter();
+            q.1.flush();
-            let data_iter = (0 .. data_length).map(|n| *(buff.next().unwrap()));
+            q.0
-            CpuAccessibleBuffer::from_iter(device.clone(), BufferUsage::all(), data_iter).unwrap()
         };
 
-        let out_buffer = {
+        // 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 = image_buffer.iter();
             let data_iter = (0 .. data_length).map(|n| *(buff.next().unwrap()));
-            CpuAccessibleBuffer::from_iter(device.clone(), BufferUsage::all(), data_iter).unwrap()
+            CpuAccessibleBuffer::from_data(device.clone(), BufferUsage::all(), image_buffer.clone()).unwrap()
         };
 
+        // A buffer to hold many i32 values to use as settings
         let settings_buffer = {
             let vec = vec![xy.0, xy.1];
             let mut buff = vec.iter();
@@ -227,14 +213,16 @@ fn main() {
             CpuAccessibleBuffer::from_iter(device.clone(), BufferUsage::all(), data_iter).unwrap()
         };
 
+        println!("Done");
         // Create the data descriptor set for our previously created shader pipeline
         let mut set = Arc::new(PersistentDescriptorSet::start(pipeline.clone(), 0)
-            .add_buffer(data_buffer.clone()).unwrap()
+            .add_buffer(read_buffer).unwrap()
-            .add_buffer(out_buffer.clone()).unwrap()
+            .add_buffer(write_buffer.clone()).unwrap()
             .add_buffer(settings_buffer.clone()).unwrap()
             .build().unwrap()
         );
 
+        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(device.clone(), queue.family()).unwrap()
             .dispatch([xy.0, xy.1, 1], pipeline.clone(), set.clone(), ()).unwrap()
@@ -248,9 +236,13 @@ fn main() {
         // I think this is redundant and returns immediately
         future.wait(None).unwrap();
 
+        println!("Done running kernel");
         // The buffer is sync'd so we can just read straight from the handle
-        let data_buffer_content = data_buffer.read().unwrap();
+        let data_buffer_content = write_buffer.read().unwrap();
+
+        let p = image::load_from_memory(data_buffer_content.as_slice()).unwrap();
 
+        println!("done2");
         for y in 0 .. xy.1 {
             for x in 0 .. xy.0 {
 
@@ -269,6 +261,7 @@ fn main() {
         }
     }
 
+    println!("done3");
     img.save(format!("output/{}.png", SystemTime::now().duration_since(SystemTime::UNIX_EPOCH).unwrap().as_secs()));
 
     println!("Stdddwafasddfqwefaarting");