From 3a0d646ade02a6ca006a0d8cf6c0f60a1ece8272 Mon Sep 17 00:00:00 2001
From: Dennis Kobert <dennis@kobert.dev>
Date: Sun, 12 Jan 2020 07:08:30 +0100
Subject: Fix wrong workgroup size

---
 src/main.rs                |  2 +-
 src/solvers/gpu/host.rs    |  8 +-------
 src/solvers/gpu/manager.rs | 18 ++++++++++++++----
 src/solvers/gpu/output.rs  |  1 +
 4 files changed, 17 insertions(+), 12 deletions(-)

diff --git a/src/main.rs b/src/main.rs
index f18f2f8..c2643c3 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -3,7 +3,7 @@ mod solvers;
 mod structs;
 use crate::solvers::{IteratorSolver, Solver};
 
-pub static N: u32 = 4;
+pub static N: u32 = 8;
 fn main() {
     let mut solver = solvers::single::NormalSolver::new(N);
     solver.solve();
diff --git a/src/solvers/gpu/host.rs b/src/solvers/gpu/host.rs
index 39843c4..e354263 100644
--- a/src/solvers/gpu/host.rs
+++ b/src/solvers/gpu/host.rs
@@ -28,7 +28,7 @@ impl Host {
         n: u32,
         h: u32,
         w: u32,
-        mut wg_size: usize,
+        wg_size: usize,
         src: &str,
         output_sender: Sender<Message>,
     ) -> ocl::Result<(Sender<Message>, std::thread::JoinHandle<()>)> {
@@ -52,12 +52,6 @@ impl Host {
             .build()?;
 
         let (sender, receiver) = channel();
-        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"));
-        }
 
         let solver = Self {
             platform,
diff --git a/src/solvers/gpu/manager.rs b/src/solvers/gpu/manager.rs
index e210af2..a2253aa 100644
--- a/src/solvers/gpu/manager.rs
+++ b/src/solvers/gpu/manager.rs
@@ -45,10 +45,20 @@ impl OclManager {
         permutations_mask: &[u64],
         n: u32,
         // Workgroup size, set to 0 for max
-        wg_size: u32,
+        mut wg_size: usize,
     ) -> (Sender<Message>, JoinHandle<()>) {
         let (h, w) = crate::solvers::wall_stats(n);
         let src = include_str!("check.cl");
+
+        let platform = ocl::Platform::default();
+        let device = ocl::Device::first(platform).expect("failed to create opencl device");
+        let max_wg_size = device.max_wg_size().expect("failed to query max_wg_size");
+        if wg_size == 0 {
+            wg_size = max_wg_size;
+        } else if wg_size > max_wg_size {
+            println!("invalid workgroup size");
+        }
+
         let (output_sender, output_handle) =
             super::output::Output::launch_sevice(permutations, permutations_mask);
         let (host_sender, host_handle) = super::host::Host::launch_sevice(
@@ -56,7 +66,7 @@ impl OclManager {
             n,
             h,
             w,
-            wg_size as usize,
+            wg_size,
             src,
             output_sender.clone(),
         )
@@ -64,6 +74,7 @@ impl OclManager {
 
         let (sender, receiver) = channel();
 
+        println!("wg {}", wg_size);
         let mut buffers = Vec::with_capacity((n - h + 1) as usize);
         for _ in 0..=(n - h) {
             buffers.push(RequestBuffer::new(wg_size as usize));
@@ -94,7 +105,6 @@ impl OclManager {
             match self.receiver.recv().expect("Channel to GPU Manager broke") {
                 Message::CheckRequest(request) => {
                     let queue = request.queue;
-                    println!("{}", queue);
                     if let Some(buffer) = self.buffers[queue as usize].read(request) {
                         self.host_sender
                             .send(Message::HostMessage((self.job_id, queue, buffer.0.into())))
@@ -106,7 +116,7 @@ impl OclManager {
                     }
                 }
                 Message::Terminate => {
-                    panic!("flush buffers");
+                    //TODO panic!("flush buffers");
                     self.host_sender.send(Message::Terminate);
                     self.host_handle.join();
                     self.output_sender.send(Message::Terminate);
diff --git a/src/solvers/gpu/output.rs b/src/solvers/gpu/output.rs
index b35d072..43e0d98 100644
--- a/src/solvers/gpu/output.rs
+++ b/src/solvers/gpu/output.rs
@@ -74,6 +74,7 @@ pub struct RowResult {
 impl RowResult {
     fn new(mut rows: Vec<u32>) -> Self {
         rows.push(0);
+        rows.sort();
         Self { rows }
     }
     fn output(&self) {
-- 
cgit v1.2.3-54-g00ecf