From 1a9d0d91e08d0dee9b48c05ef0ac72cc732a42ab Mon Sep 17 00:00:00 2001
From: Dennis Kobert <dennis@kobert.dev>
Date: Mon, 6 Jan 2020 08:33:57 +0100
Subject: Backport to opencl 1.2

---
 src/solvers/opencl.rs | 42 ++++++++++++++++++++++++++++++------------
 1 file changed, 30 insertions(+), 12 deletions(-)

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

diff --git a/src/solvers/opencl.rs b/src/solvers/opencl.rs
index 1293295..8421d53 100644
--- a/src/solvers/opencl.rs
+++ b/src/solvers/opencl.rs
@@ -7,8 +7,11 @@ pub struct Job {
 }
 
 pub struct GpuSolver {
+    #[allow(unused)]
     platform: Platform,
+    #[allow(unused)]
     device: Device,
+    #[allow(unused)]
     context: Context,
     program: Program,
     queue: Queue,
@@ -17,19 +20,19 @@ pub struct GpuSolver {
     w: u32,
     wg_size: usize,
     permutations: Buffer<u64>,
-    rec_queues: Vec<ReqestBuffer>,
+    rec_queues: Vec<RequestBuffer>,
 }
 
-struct ReqestBuffer {
+struct RequestBuffer {
     mask_buff: Vec<u64>,
     row_buff: Vec<Vec<u32>>,
     pointer: usize,
     receiver: Receiver<Job>,
 }
 
-impl ReqestBuffer {
-    pub fn new(size: usize, receiver: Receiver<Job>) -> ReqestBuffer {
-        Self {
+impl RequestBuffer {
+    pub fn new(size: usize, receiver: Receiver<Job>) -> Self {
+        RequestBuffer {
             mask_buff: Vec::with_capacity(size),
             row_buff: Vec::with_capacity(size),
             pointer: 0,
@@ -48,13 +51,14 @@ impl ReqestBuffer {
         }
         None
     }
+    #[allow(unused)]
     pub fn get_rows(&self) -> &[Vec<u32>] {
         self.row_buff.as_ref()
     }
 }
 
 impl GpuSolver {
-    pub fn new(
+    pub fn launch_sevice(
         permutation_masks: &[u64],
         n: u32,
         h: u32,
@@ -86,7 +90,7 @@ impl GpuSolver {
         for _ in 0..h {
             let (sx, rx) = std::sync::mpsc::channel();
             senders.push(sx);
-            receivers.push(ReqestBuffer::new(wg_size, rx));
+            receivers.push(RequestBuffer::new(wg_size, rx));
         }
 
         let solver = Self {
@@ -108,6 +112,22 @@ impl GpuSolver {
         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)
+        }
+    }
+    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
+    }
+
     fn run(mut self) -> ! {
         let queues = self.rec_queues.len();
         let mut instruction_buffer = Vec::with_capacity((self.n - self.h) as usize);
@@ -123,10 +143,9 @@ impl GpuSolver {
                 .unwrap();
 
             instruction_buffer.push(buffer);
-            let dim = (i + 1) * chunk;
             let results: Buffer<u64> = Buffer::builder()
                 .queue(self.queue.clone())
-                .len(self.wg_size * dim / 64)
+                .len(self.wg_size * self.get_dim(i) / 64)
                 .flags(flags::MEM_READ_WRITE)
                 .build()
                 .unwrap();
@@ -135,8 +154,7 @@ impl GpuSolver {
         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();
-                println!("hello world");
-                let dim = (i + 1) * chunk;
+                let dim = self.get_dim(i);
 
                 let kernel = Kernel::builder()
                     .program(&self.program)
@@ -149,7 +167,7 @@ impl GpuSolver {
                     .arg_local::<u64>(self.wg_size)
                     .arg(self.n)
                     .arg(self.w)
-                    .arg((self.n as u64 - i as u64 - 1) * chunk as u64)
+                    .arg(self.get_off(i))
                     .build()
                     .unwrap();
 
-- 
cgit v1.2.3-54-g00ecf