From c36035c0667fdd224da914b50c30a9366e1a5c38 Mon Sep 17 00:00:00 2001
From: Dennis Kobert <dennis@kobert.dev>
Date: Tue, 7 Jan 2020 03:12:55 +0100
Subject: Fix global worgsize calculation

---
 src/solvers/opencl.rs | 75 +++++++++++++++++++++++++++++++++++++--------------
 1 file changed, 55 insertions(+), 20 deletions(-)

(limited to 'src/solvers/opencl.rs')

diff --git a/src/solvers/opencl.rs b/src/solvers/opencl.rs
index cdedd37..550ace9 100644
--- a/src/solvers/opencl.rs
+++ b/src/solvers/opencl.rs
@@ -6,6 +6,13 @@ pub struct Job {
     bitmask: u64,
 }
 
+impl Job {
+    pub fn new(rows: Vec<u32>, bitmask: u64) -> Self {
+        Self { rows, bitmask }
+    }
+}
+
+#[derive(Debug)]
 pub struct GpuSolver {
     #[allow(unused)]
     platform: Platform,
@@ -18,11 +25,14 @@ pub struct GpuSolver {
     n: u32,
     h: u32,
     w: u32,
+    /// Workgroup size, set to 0 for max
     wg_size: usize,
     permutations: Buffer<u64>,
     rec_queues: Vec<RequestBuffer>,
+    walls: Vec<Vec<u32>>,
 }
 
+#[derive(Debug)]
 struct RequestBuffer {
     mask_buff: Vec<u64>,
     row_buff: Vec<Vec<u32>>,
@@ -33,8 +43,8 @@ struct RequestBuffer {
 impl RequestBuffer {
     pub fn new(size: usize, receiver: Receiver<Job>) -> Self {
         RequestBuffer {
-            mask_buff: Vec::with_capacity(size),
-            row_buff: Vec::with_capacity(size),
+            mask_buff: vec![0; size],
+            row_buff: vec![Vec::new(); size],
             pointer: 0,
             receiver,
         }
@@ -85,15 +95,15 @@ impl GpuSolver {
             .len(permutation_masks.len())
             .build()?;
 
-        let mut senders = Vec::with_capacity(h as usize);
-        let mut receivers = Vec::with_capacity(h as usize);
+        let mut senders = Vec::with_capacity((n - h + 1) as usize);
+        let mut receivers = Vec::with_capacity((n - h + 1) as usize);
         let max_wg_size = device.max_wg_size()?;
         if wg_size == 0 {
             wg_size = max_wg_size;
         } else if wg_size > max_wg_size {
             return Err(ocl::Error::from("invalid workgroup size"));
         }
-        for _ in 0..h {
+        for _ in 0..=(n - h) {
             let (sx, rx) = std::sync::mpsc::channel();
             senders.push(sx);
             receivers.push(RequestBuffer::new(wg_size, rx));
@@ -111,34 +121,40 @@ impl GpuSolver {
             wg_size,
             permutations: buffer,
             rec_queues: receivers,
+            walls: Vec::new(),
         };
-        std::thread::spawn(move || {
-            solver.run();
-        });
+        std::thread::Builder::new()
+            .name("GPU Deamon".into())
+            .spawn(move || {
+                solver.run();
+            })
+            .unwrap();
+        println!("started gpu thread");
         Ok(senders)
     }
 
     fn get_dim(&self, queue: usize) -> usize {
         let chunk = self.permutations.len() / self.n as usize;
         let dim = (queue + 1) * chunk;
-        if dim % self.wg_size == 0 {
-            dim
-        } else {
-            let dim_small = dim / self.wg_size;
-            ((dim_small + 1) * chunk)
-        }
+        (dim + self.wg_size - 1) / self.wg_size * self.wg_size
     }
     fn get_off(&self, queue: usize) -> u64 {
         let chunk = self.permutations.len() / self.n as usize;
-        let dim = (queue + 1) * chunk;
-        ((self.n as usize - queue - 1) * chunk - self.wg_size + dim % self.wg_size) as u64
+        let off = self.permutations.len() - chunk - self.get_dim(queue);
+        off as u64
+    }
+    fn get_res(&self, queue: usize) -> usize {
+        let dim = self.get_dim(queue);
+        dim * self.get_res_save_dim()
+    }
+    fn get_res_save_dim(&self) -> usize {
+        (self.wg_size + 63) / 64
     }
 
     fn run(mut self) -> ! {
         let queues = self.rec_queues.len();
         let mut instruction_buffer = Vec::with_capacity((self.n - self.h) as usize);
         let mut result_buffer = Vec::with_capacity((self.n - self.h) as usize);
-        let chunk = self.permutations.len() / self.n as usize;
 
         for i in 0..queues {
             let buffer: Buffer<u64> = Buffer::builder()
@@ -151,17 +167,21 @@ impl GpuSolver {
             instruction_buffer.push(buffer);
             let results: Buffer<u64> = Buffer::builder()
                 .queue(self.queue.clone())
-                .len(self.wg_size * self.get_dim(i) / 64)
+                .len(self.get_res(i))
                 .flags(flags::MEM_READ_WRITE)
                 .build()
                 .unwrap();
             result_buffer.push(results);
         }
+        println!("finished gpu setup");
         for i in (0..self.rec_queues.len()).cycle() {
             if let Some(buffer) = self.rec_queues[i].read() {
                 instruction_buffer[i].write(buffer).enq().unwrap();
                 let dim = self.get_dim(i);
 
+                //println!("dim: {}", dim);
+                //println!("off: {}", self.get_off(i));
+                //println!("result size: {}", self.get_res_save_dim());
                 let kernel = Kernel::builder()
                     .program(&self.program)
                     .name("check")
@@ -189,7 +209,7 @@ impl GpuSolver {
                 }
 
                 // (5) Read results from the device into a vector (`::block` not shown):
-                let mut result = vec![0u64; dim * self.wg_size / 64];
+                let mut result = vec![0u64; self.get_res(i)];
                 result_buffer[i]
                     .cmd()
                     .queue(&self.queue)
@@ -197,7 +217,22 @@ impl GpuSolver {
                     .read(&mut result)
                     .enq()
                     .unwrap();
-                println!("{:?}", result);
+                for (j, r) in result.iter().enumerate() {
+                    if j == 0 {
+                        continue;
+                    }
+                    for b in 0..64 {
+                        if r & (1 << b) != 0 {
+                            let permutation =
+                                j / self.get_res_save_dim() + self.get_off(i) as usize;
+                            let instruction = (j % self.get_res_save_dim()) * 64 + b;
+                            let mut wall = self.rec_queues[i].get_rows()[instruction].clone();
+                            wall.push(permutation as u32);
+                            println!("{:?}", wall);
+                            self.walls.push(wall);
+                        }
+                    }
+                }
             }
         }
         panic!();
-- 
cgit v1.2.3-70-g09d2