parallel/queue: comments and type checks

parallel/spmc.hpp

@@ -17,12 +17,35 @@
 #ifndef CRUFT_UTIL_PARALLEL_SPMC_HPP
 #define CRUFT_UTIL_PARALLEL_SPMC_HPP
 
+#include "../maths.hpp"
+
 #include <atomic>
+#include <type_traits>
+
 
 namespace util::parallel {
+    /// A statically sized thread-safe, lock-free, multi-producer
+    /// multi-consumer queue for trivial types.
+    ///
+    /// \tparam ValueT the type of values to store
+    /// \tparam SizeV the number of elements in the underlying store
     template <typename ValueT, uint32_t SizeV = 32>
     class spmc {
     public:
+        // We optimise specifically for memcpyable types only
+        static_assert (
+            std::is_trivially_destructible_v<ValueT> && std::is_trivially_copyable_v<ValueT>
+        );
+
+        // We need to wrap indices to the bounds of the data array quite often.
+        // If it's a power of two then we can use bitmasks to avoid the
+        // modulus.
+        static_assert (util::is_pow2 (SizeV));
+
+        // We need at least 2 data slots; one for read, and one for write
+        // pointers. 4 gives a little safety margin for future optimisations.
+        static_assert (SizeV >= 4);
+
         spmc ():
             m_write (0),
             m_read { {0}, {0} }
@@ -59,8 +82,10 @@ namespace util::parallel {
                     return false;
 
                 // try to bump the write cursor to claim that slot
-                if (uint32_t orig = curr; !m_write.compare_exchange_weak (orig, next))
+                if (uint32_t orig = curr; !m_write.compare_exchange_weak (orig, next)) {
+                    asm volatile ("pause;");
                     continue;
+                }
 
                 // write the data to the buffer
                 m_data[wrap (curr)] = val;

test/parallel/spmc.cpp

@@ -1,6 +1,8 @@
 #include "parallel/spmc.hpp"
 #include "job/flag.hpp"
+#include "job/semaphore.hpp"
 #include "debug.hpp"
+#include "tap.hpp"
 
 #include <algorithm>
 #include <thread>
@@ -8,8 +10,8 @@
 
 
 static constexpr uint32_t slots = 4;
-static constexpr int parallelism = 15;
-static constexpr int chunk_size = 1<<16;
+static constexpr int parallelism = 8;
+static constexpr int chunk_size = 1<<12;
 
 util::job::flag start;
 
@@ -17,7 +19,22 @@ using queue_t = util::parallel::spmc<int,slots>;
 
 
 ///////////////////////////////////////////////////////////////////////////////
-void transfer (queue_t &src, std::vector<int> &dst)
+// push numbers in the range [first,first+chunk_size) to the queue
+void
+produce (queue_t &dst, const int first)
+{
+    start.wait ();
+
+    for (int i = first; i < first + chunk_size; ++i)
+        while (!dst.push (i))
+            ;
+}
+
+
+//-----------------------------------------------------------------------------
+// pop chunk_size elements from the queue to a vector
+void
+consume (queue_t &src, util::view<int*> dst)
 {
     (void)dst;
     start.wait ();
@@ -33,40 +50,54 @@ void transfer (queue_t &src, std::vector<int> &dst)
 
 
 ///////////////////////////////////////////////////////////////////////////////
+// * create n producers which push the c integers over [n*c,n*(c+1)) into a queue
+// * create n consumers which pop c integers from a queue into a store
+// * ensure all integers we expected are in the store
+//
+// we use an artificially small queue size to force contention.
+//
+// to stress test we increase the threads, increase the data length, and
+// decrease the queue size.
 int
 main ()
 {
+    util::TAP::logger tap;
     queue_t src;
 
-    std::vector<std::vector<int>> dst (parallelism + 1);
-    std::vector<std::thread> workers;
+    // start n consumers, and n producers that fill an array with integers
+    std::vector<int> dst (parallelism * chunk_size);
+    std::vector<std::thread> consumers;
+    std::vector<std::thread> producers;
 
     for (int i = 0; i < parallelism; ++i) {
-        dst[i].resize (chunk_size);
-        workers.emplace_back (transfer, std::ref (src), std::ref (dst[i]));
+        consumers.emplace_back (
+            consume,
+            std::ref (src),
+            util::view{
+                dst.data() + i * chunk_size,
+                chunk_size
+            }
+        );
+
+        producers.emplace_back (produce, std::ref (src), i * chunk_size);
     }
 
     start.notify ();
 
-    for (int i = 0; i < parallelism * chunk_size; ++i) {
-        while (!src.push (i))
-            ;
-    }
-
-    for (auto &t: workers)
+    // wait for everyone to complete
+    for (auto &t: producers)
+        t.join ();
+    for (auto &t: consumers)
         t.join ();
 
-    std::vector<int> tally;
-    for (auto &d: dst)
-        tally.insert (tally.end (), d.begin (), d.end ());
-
-    std::sort (tally.begin (), tally.end ());
+    // sort the result and ensure all values are present
+    std::sort (dst.begin (), dst.end ());
 
     int missing = 0;
     for (int i = 0; i < parallelism * chunk_size; ++i)
-        if (tally[i] != i)
+        if (dst[i] != i)
             ++missing;
 
-    std::clog << "missing: " << missing << '\n';
-    return 0;
+    tap.expect_eq (missing, 0, "no lost values in parallel produce/consume queue");
+    return tap.status ();
 }