X-Git-Url: http://mj.ucw.cz/gitweb/?a=blobdiff_plain;ds=sidebyside;f=lib%2Fsorter%2Farray.c;h=dd7659e70f04a2ccc18da447f6a9bdb269070eab;hb=a5ff98a53789157a6c96e58b2385bb898d688a22;hp=26b4a1b53048e1f11de89a653c36bcbdb513b292;hpb=7cf300f543b5023ad46a909ed6fde0feba5d1acd;p=libucw.git

diff --git a/lib/sorter/array.c b/lib/sorter/array.c
index 26b4a1b5..dd7659e7 100644
--- a/lib/sorter/array.c
+++ b/lib/sorter/array.c
@@ -48,7 +48,7 @@ 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->elt_size < sorter_radix_threshold || shift < ASORT_MIN_SHIFT)
 	{
 	  ctx->quicksort(buffer, n);
 	  if (!swapped_output)
@@ -182,7 +182,7 @@ 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 +195,7 @@ 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 +205,7 @@ 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);
 }
 
@@ -214,16 +214,30 @@ rs_finish(struct worker_thread *thr UNUSED, struct work *ww)
 {
   struct rs_work *w = (struct rs_work *) ww;
 
-  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 (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 * w->ctx->elt_size < sorter_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);
-  DBG("Thread %d: Finishing done", thr->id);
+    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);
+}
+
+static void
+rs_wait_small(struct asort_context *ctx)
+{
+  struct rs_work *w;
+
+  while (w = (struct rs_work *) work_wait(ctx->rs_work_queue))
+    {
+      DBG("Reaping small chunk of %d items", w->num_elts);
+      ep_free(ctx->eltpool, w);
+    }
 }
 
 static void
@@ -235,6 +249,9 @@ rs_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns
   uns blksize = num_elts / sorter_threads;
   DBG(">>> n=%d h=%d s=%d blk=%d sw=%d", num_elts, hash_bits, shift, blksize, swapped_output);
 
+  // If there are any small chunks in progress, wait for them to finish
+  rs_wait_small(ctx);
+
   // Start parallel counting
   void *iptr = array;
   for (uns i=0; i<sorter_threads; i++)
@@ -243,8 +260,8 @@ 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;
@@ -298,19 +315,30 @@ rs_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns
   for (uns i=0; i<buckets; i++)
     {
       uns n = cnt[i] - pos;
+      if (!n)
+	continue;
       if (n * ctx->elt_size < sorter_thread_threshold)
 	{
 	  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;
-	  clist_add_tail(&ctx->rs_bits, &w->w.n);
-	  DBG("Scheduling block %d+%d", pos, n);
+	  if (n * ctx->elt_size < sorter_thread_chunk)
+	    {
+	      DBG("Sorting block %d+%d inline", pos, n);
+	      rs_finish(NULL, &w->w);
+	      ep_free(ctx->eltpool, w);
+	    }
+	  else
+	    {
+	      DBG("Scheduling block %d+%d", pos, n);
+	      work_submit(ctx->rs_work_queue, &w->w);
+	    }
 	}
       else
 	rs_radix(ctx, buffer, array, n, shift, !swapped_output);
@@ -321,7 +349,7 @@ 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;
 
@@ -335,23 +363,18 @@ threaded_radixsort(struct asort_context *ctx)
   for (uns i=0; i<sorter_threads; i++)
     ctx->rs_works[i] = big_alloc(sizeof(struct rs_work) + sizeof(uns) * (1 << ctx->radix_bits));
 
-  // Prepare work structures for all remaining small bits which will be sorted later.
-  clist_init(&ctx->rs_bits);
+  // Prepare a pool for all remaining small bits which will be sorted on background.
   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));
 
   // Finish the small blocks
-  struct rs_work *w, *tmp;
-  CLIST_WALK_DELSAFE(w, ctx->rs_bits, tmp)
-    work_submit(&q, &w->w);
-  while (work_wait(&q))
-    ;
+  rs_wait_small(ctx);
 
+  ASSERT(!ctx->eltpool->num_allocated);
   ep_delete(ctx->eltpool);
   work_queue_cleanup(&q);
   asort_stop_threads();
@@ -364,6 +387,23 @@ 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->elt_size >= sorter_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)
 {
@@ -372,7 +412,7 @@ asort_run(struct asort_context *ctx)
 			      ctx->num_elts * ctx->elt_size >= sorter_thread_threshold &&
 			      !(sorter_debug & SORT_DEBUG_ASORT_NO_THREADS));
 
-  if (ctx->num_elts < sorter_radix_threshold ||
+  if (ctx->num_elts * ctx->elt_size < sorter_radix_threshold ||
       ctx->hash_bits <= ASORT_MIN_SHIFT ||
       !ctx->radix_split ||
       (sorter_debug & SORT_DEBUG_ASORT_NO_RADIX))
@@ -390,17 +430,19 @@ asort_run(struct asort_context *ctx)
     }
   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);
+	  SORT_XTRACE(12, "Decided to use parallel radix-sort (swap=%d)", swap);
+	  threaded_radixsort(ctx, swap);
 	  return;
 	}
 #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);
+      SORT_XTRACE(12, "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");