Array sorter thresholds are now configured as 64-bit numbers of bytes,

[libucw.git] / lib / sorter / array.c

diff --git a/lib/sorter/array.c b/lib/sorter/array.c
index efa314a8702c1fd01339c9467cc473b5c70d96f3..a612e8f7841b66b9cd1c02fafc0089d72b15050d 100644 (file)
--- a/lib/sorter/array.c
+++ b/lib/sorter/array.c
@@ -17,6 +17,9 @@
 
 #define ASORT_MIN_SHIFT 2
 
+#define ASORT_TRACE(x...) ASORT_XTRACE(1, x)
+#define ASORT_XTRACE(level, x...) do { if (sorter_trace_array >= level) msg(L_DEBUG, x); } while(0)
+
 static void
 asort_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns hash_bits, uns swapped_output)
 {
@@ -48,7 +51,7 @@ asort_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts,
   for (uns i=0; i<buckets; i++)
     {
       uns n = cnt[i] - pos;
-      if (n < sorter_radix_threshold || shift < ASORT_MIN_SHIFT)
+      if (n < ctx->radix_threshold || shift < ASORT_MIN_SHIFT)
        {
          ctx->quicksort(buffer, n);
          if (!swapped_output)
@@ -72,6 +75,15 @@ static uns asort_threads_use_count;
 static uns asort_threads_ready;
 static struct worker_pool asort_thread_pool;
 
+static uns
+rs_estimate_stack(void)
+{
+  // Stack space needed by the recursive radix-sorter
+  uns ctrsize = sizeof(uns) * (1 << CONFIG_UCW_RADIX_SORTER_BITS);
+  uns maxdepth = (64 / CONFIG_UCW_RADIX_SORTER_BITS) + 1;
+  return ctrsize * maxdepth;
+}
+
 void
 asort_start_threads(uns run)
 {
@@ -79,8 +91,11 @@ asort_start_threads(uns run)
   asort_threads_use_count++;
   if (run && !asort_threads_ready)
     {
-      SORT_XTRACE(2, "Initializing thread pool (%d threads)", sorter_threads);
+      // XXX: If somebody overrides the radix-sorter parameters to insane values,
+      // he also should override the stack size to insane values.
+      asort_thread_pool.stack_size = default_thread_stack_size + rs_estimate_stack();
       asort_thread_pool.num_threads = sorter_threads;
+      ASORT_TRACE("Initializing thread pool (%d threads, %dK stack)", sorter_threads, asort_thread_pool.stack_size >> 10);
       worker_pool_init(&asort_thread_pool);
       asort_threads_ready = 1;
     }
@@ -93,7 +108,7 @@ asort_stop_threads(void)
   ucwlib_lock();
   if (!--asort_threads_use_count && asort_threads_ready)
     {
-      SORT_XTRACE(2, "Shutting down thread pool");
+      ASORT_TRACE("Shutting down thread pool");
       worker_pool_cleanup(&asort_thread_pool);
       asort_threads_ready = 0;
     }
@@ -113,15 +128,16 @@ static void
 qs_handle_work(struct worker_thread *thr UNUSED, struct work *ww)
 {
   struct qs_work *w = (struct qs_work *) ww;
+  struct asort_context *ctx = w->ctx;
 
   DBG("Thread %d: got %d elts", thr->id, w->num_elts);
-  if (w->num_elts * w->ctx->elt_size < sorter_thread_threshold)
+  if (w->num_elts < ctx->thread_threshold)
     {
-      w->ctx->quicksort(w->array, w->num_elts);
+      ctx->quicksort(w->array, w->num_elts);
       w->left = w->right = LR_UNDEF;
     }
   else
-    w->ctx->quicksplit(w->array, w->num_elts, &w->left, &w->right);
+    ctx->quicksplit(w->array, w->num_elts, &w->left, &w->right);
   DBG("Thread %d: returning l=%d r=%d", thr->id, w->left, w->right);
 }
 
@@ -182,7 +198,7 @@ threaded_quicksort(struct asort_context *ctx)
 struct rs_work {
   struct work w;
   struct asort_context *ctx;
-  void *in, *out;
+  void *array, *buffer;                // Like asort_radix().
   uns num_elts;
   uns shift;
   uns swap_output;
@@ -195,7 +211,7 @@ rs_count(struct worker_thread *thr UNUSED, struct work *ww)
   struct rs_work *w = (struct rs_work *) ww;
 
   DBG("Thread %d: Counting %d items, shift=%d", thr->id, w->num_elts, w->shift);
-  w->ctx->radix_count(w->in, w->num_elts, w->cnt, w->shift);
+  w->ctx->radix_count(w->array, w->num_elts, w->cnt, w->shift);
   DBG("Thread %d: Counting done", thr->id);
 }
 
@@ -205,7 +221,7 @@ rs_split(struct worker_thread *thr UNUSED, struct work *ww)
   struct rs_work *w = (struct rs_work *) ww;
 
   DBG("Thread %d: Splitting %d items, shift=%d", thr->id, w->num_elts, w->shift);
-  w->ctx->radix_split(w->in, w->out, w->num_elts, w->cnt, w->shift);
+  w->ctx->radix_split(w->array, w->buffer, w->num_elts, w->cnt, w->shift);
   DBG("Thread %d: Splitting done", thr->id);
 }
 
@@ -216,14 +232,14 @@ rs_finish(struct worker_thread *thr UNUSED, struct work *ww)
 
   if (thr)
     DBG("Thread %d: Finishing %d items, shift=%d", thr->id, w->num_elts, w->shift);
-  if (w->shift < ASORT_MIN_SHIFT || w->num_elts < sorter_radix_threshold)
+  if (w->shift < ASORT_MIN_SHIFT || w->num_elts < w->ctx->radix_threshold)
     {
-      w->ctx->quicksort(w->in, w->num_elts);
+      w->ctx->quicksort(w->array, w->num_elts);
       if (w->swap_output)
-       memcpy(w->out, w->in, w->num_elts * w->ctx->elt_size);
+       memcpy(w->buffer, w->array, w->num_elts * w->ctx->elt_size);
     }
   else
-    asort_radix(w->ctx, w->in, w->out, w->num_elts, w->shift, w->swap_output);
+    asort_radix(w->ctx, w->array, w->buffer, w->num_elts, w->shift, w->swap_output);
   if (thr)
     DBG("Thread %d: Finishing done", thr->id);
 }
@@ -260,8 +276,8 @@ rs_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns
       w->w.priority = 0;
       w->w.go = rs_count;
       w->ctx = ctx;
-      w->in = iptr;
-      w->out = buffer;
+      w->array = iptr;
+      w->buffer = buffer;
       w->num_elts = blksize;
       if (i == sorter_threads-1)
        w->num_elts += num_elts % sorter_threads;
@@ -317,18 +333,18 @@ rs_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns
       uns n = cnt[i] - pos;
       if (!n)
        continue;
-      if (n * ctx->elt_size < sorter_thread_threshold)
+      if (n < ctx->thread_threshold || shift < ASORT_MIN_SHIFT)
        {
          struct rs_work *w = ep_alloc(ctx->eltpool);
          w->w.priority = 0;
          w->w.go = rs_finish;
          w->ctx = ctx;
-         w->in = buffer;
-         w->out = array;
+         w->array = buffer;
+         w->buffer = array;
          w->num_elts = n;
          w->shift = shift;
          w->swap_output = !swapped_output;
-         if (n * ctx->elt_size < sorter_thread_chunk)
+         if (n < ctx->thread_chunk)
            {
              DBG("Sorting block %d+%d inline", pos, n);
              rs_finish(NULL, &w->w);
@@ -349,7 +365,7 @@ rs_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns
 }
 
 static void
-threaded_radixsort(struct asort_context *ctx)
+threaded_radixsort(struct asort_context *ctx, uns swap)
 {
   struct work_queue q;
 
@@ -367,8 +383,7 @@ threaded_radixsort(struct asort_context *ctx)
   ctx->eltpool = ep_new(sizeof(struct rs_work), 1000);
 
   // Do the big splitting
-  // FIXME: Set the swap bit carefully.
-  rs_radix(ctx, ctx->array, ctx->buffer, ctx->num_elts, ctx->hash_bits, 0);
+  rs_radix(ctx, ctx->array, ctx->buffer, ctx->num_elts, ctx->hash_bits, swap);
   for (uns i=0; i<sorter_threads; i++)
     big_free(ctx->rs_works[i], sizeof(struct rs_work) + sizeof(uns) * (1 << ctx->radix_bits));
 
@@ -388,15 +403,38 @@ void asort_stop_threads(void) { }
 
 #endif
 
+static uns
+predict_swap(struct asort_context *ctx)
+{
+  uns bits = ctx->radix_bits;
+  uns elts = ctx->num_elts;
+  uns swap = 0;
+
+  while (elts >= ctx->radix_threshold && bits >= ASORT_MIN_SHIFT)
+    {
+      DBG("Predicting pass: %d elts, %d bits", elts, bits);
+      swap = !swap;
+      elts >>= ctx->radix_bits;
+      bits = MAX(bits, ctx->radix_bits) - ctx->radix_bits;
+    }
+  return swap;
+}
+
 void
 asort_run(struct asort_context *ctx)
 {
-  SORT_XTRACE(10, "Array-sorting %d items per %d bytes, hash_bits=%d", ctx->num_elts, ctx->elt_size, ctx->hash_bits);
+  ctx->thread_threshold = MIN(sorter_thread_threshold / ctx->elt_size, ~0U);
+  ctx->thread_chunk = MIN(sorter_thread_chunk / ctx->elt_size, ~0U);
+  ctx->radix_threshold = MIN(sorter_radix_threshold / ctx->elt_size, ~0U);
+
+  ASORT_TRACE("Array-sorting %d items per %d bytes, hash_bits=%d", ctx->num_elts, ctx->elt_size, ctx->hash_bits);
+  ASORT_XTRACE(2, "Limits: thread_threshold=%u, thread_chunk=%u, radix_threshold=%u",
+       ctx->thread_threshold, ctx->thread_chunk, ctx->radix_threshold);
   uns allow_threads UNUSED = (sorter_threads > 1 &&
-                             ctx->num_elts * ctx->elt_size >= sorter_thread_threshold &&
+                             ctx->num_elts >= ctx->thread_threshold &&
                              !(sorter_debug & SORT_DEBUG_ASORT_NO_THREADS));
 
-  if (ctx->num_elts < sorter_radix_threshold ||
+  if (ctx->num_elts < ctx->radix_threshold ||
       ctx->hash_bits <= ASORT_MIN_SHIFT ||
       !ctx->radix_split ||
       (sorter_debug & SORT_DEBUG_ASORT_NO_RADIX))
@@ -404,28 +442,35 @@ asort_run(struct asort_context *ctx)
 #ifdef CONFIG_UCW_THREADS
       if (allow_threads)
        {
-         SORT_XTRACE(12, "Decided to use parallel quicksort");
+         ASORT_XTRACE(2, "Decided to use parallel quicksort");
          threaded_quicksort(ctx);
-         return;
        }
+      else
 #endif
-      SORT_XTRACE(12, "Decided to use sequential quicksort");
-      ctx->quicksort(ctx->array, ctx->num_elts);
+       {
+         ASORT_XTRACE(2, "Decided to use sequential quicksort");
+         ctx->quicksort(ctx->array, ctx->num_elts);
+       }
     }
   else
     {
+      uns swap = predict_swap(ctx);
 #ifdef CONFIG_UCW_THREADS
       if (allow_threads)
        {
-         SORT_XTRACE(12, "Decided to use parallel radix-sort");
-         threaded_radixsort(ctx);
+         ASORT_XTRACE(2, "Decided to use parallel radix-sort (swap=%d)", swap);
+         threaded_radixsort(ctx, swap);
          return;
        }
+      else
 #endif
-      SORT_XTRACE(12, "Decided to use sequential radix-sort");
-      // FIXME: select dest buffer
-      asort_radix(ctx, ctx->array, ctx->buffer, ctx->num_elts, ctx->hash_bits, 0);
+       {
+         ASORT_XTRACE(2, "Decided to use sequential radix-sort (swap=%d)", swap);
+         asort_radix(ctx, ctx->array, ctx->buffer, ctx->num_elts, ctx->hash_bits, swap);
+       }
+      if (swap)
+       ctx->array = ctx->buffer;
     }
 
-  SORT_XTRACE(11, "Array-sort finished");
+  ASORT_XTRACE(2, "Array-sort finished");
 }