From 0f9b907deef5129141c0be52db12f6c0e8c054a4 Mon Sep 17 00:00:00 2001
From: Dennis Kobert <dennis@kobert.dev>
Date: Sun, 12 Jan 2020 12:40:13 +0100
Subject: ParallelIterator for results

---
 src/solvers/gpu/mod.rs    | 34 +++++++++++++++++++++++++++-------
 src/solvers/gpu/output.rs |  4 ++++
 src/solvers/gpusolver.rs  |  2 +-
 3 files changed, 32 insertions(+), 8 deletions(-)

diff --git a/src/solvers/gpu/mod.rs b/src/solvers/gpu/mod.rs
index 379416a..17bc964 100644
--- a/src/solvers/gpu/mod.rs
+++ b/src/solvers/gpu/mod.rs
@@ -3,6 +3,7 @@ pub mod manager;
 pub mod output;
 
 pub use manager::*;
+use rayon::prelude::*;
 
 type MaskMessage = (u64, u32, Vec<u64>);
 type RowMessage = (u64, Vec<Vec<u32>>);
@@ -40,14 +41,33 @@ impl ResultMessage {
     }
     fn valid_walls(&self) -> Vec<Vec<u32>> {
         let mut result = vec![Vec::new(); self.wg_size];
-        for (j, r) in self.data.iter().enumerate() {
-            for b in 0..64 {
-                if r & (1 << b) != 0 {
-                    let permutation = j / self.size + self.offset;
-                    let instruction = (j % self.size) * 64 + b;
-                    result[instruction].push(permutation as u32);
+        let tup: Vec<(usize, u32)> = self
+            .data
+            .par_iter()
+            .enumerate()
+            .filter_map(|(j, r)| {
+                if *r != 0 {
+                    Some(
+                        (0..64)
+                            .filter_map(|b| {
+                                if r & (1 << b) != 0 {
+                                    let permutation = j / self.size + self.offset;
+                                    let instruction = (j % self.size) * 64 + b;
+                                    Some((instruction, permutation as u32))
+                                } else {
+                                    None
+                                }
+                            })
+                            .collect(),
+                    )
+                } else {
+                    None
                 }
-            }
+            })
+            .flat_map(|a: Vec<_>| a.to_vec())
+            .collect();
+        for t in tup {
+            result[t.0 as usize].push(t.1);
         }
         result
     }
diff --git a/src/solvers/gpu/output.rs b/src/solvers/gpu/output.rs
index 6bf7bcb..9030586 100644
--- a/src/solvers/gpu/output.rs
+++ b/src/solvers/gpu/output.rs
@@ -19,6 +19,9 @@ impl InBuffer {
     }
     fn read(&mut self) -> Option<Vec<RowResult>> {
         loop {
+            //println!("{:?}", self.receiver.recv().unwrap());
+            //continue;
+
             match self
                 .receiver
                 .recv()
@@ -109,6 +112,7 @@ impl Output {
             if let Some(walls) = self.input.read() {
                 for wall in walls {
                     if !self.results.contains(&wall) {
+                        wall.output();
                         self.result_sender
                             .send(Message::RowResult(wall.clone()))
                             .or_else(|_| Err(println!("Failed to transmit result back")));
diff --git a/src/solvers/gpusolver.rs b/src/solvers/gpusolver.rs
index dc54bda..380eb6a 100644
--- a/src/solvers/gpusolver.rs
+++ b/src/solvers/gpusolver.rs
@@ -45,7 +45,7 @@ impl GpuSolver {
         let mut new_num = Vec::from(numbers);
         if index as usize == numbers.len() {
             let mut info = sys_info::mem_info().unwrap();
-            while info.avail < info.total / 8 {
+            while info.avail < info.total / 2 {
                 std::thread::sleep(std::time::Duration::from_millis(5));
                 info = sys_info::mem_info().unwrap();
                 println!("mem wait {:?}", info);
-- 
cgit v1.2.3-54-g00ecf