Redefined RadixThreshold to bound the array size instead of the number

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

diff --git a/lib/sorter/array.c b/lib/sorter/array.c
index 2eae93edbbe09352ffdadafd8ad68efd13cd3c59..5bc60276d5046fc8fca71dd29dbd6f40f4e21258 100644 (file)
--- a/lib/sorter/array.c
+++ b/lib/sorter/array.c
@@ -15,12 +15,12 @@
 #include <string.h>
 #include <alloca.h>
 
 #include <string.h>
 #include <alloca.h>
 

-#define ASORT_MIN_RADIX 5000           // FIXME: var?

 #define ASORT_MIN_SHIFT 2
 
 static void
 asort_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns hash_bits, uns swapped_output)
 {
 #define ASORT_MIN_SHIFT 2
 
 static void
 asort_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns hash_bits, uns swapped_output)
 {

+  // swap_output == 0 if result should be returned in `array', otherwise in `buffer'

   uns buckets = (1 << ctx->radix_bits);
   uns shift = (hash_bits > ctx->radix_bits) ? (hash_bits - ctx->radix_bits) : 0;
   uns cnt[buckets];
   uns buckets = (1 << ctx->radix_bits);
   uns shift = (hash_bits > ctx->radix_bits) ? (hash_bits - ctx->radix_bits) : 0;
   uns cnt[buckets];


@@ -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;
   for (uns i=0; i<buckets; i++)
     {
       uns n = cnt[i] - pos;

-      if (n < ASORT_MIN_RADIX || shift < ASORT_MIN_SHIFT)
+      if (n * cts->elt_size < sorter_radix_threshold || shift < ASORT_MIN_SHIFT)

        {
          ctx->quicksort(buffer, n);
          if (!swapped_output)
        {
          ctx->quicksort(buffer, n);
          if (!swapped_output)


@@ -214,16 +214,30 @@ rs_finish(struct worker_thread *thr UNUSED, struct work *ww)
 {
   struct rs_work *w = (struct rs_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 < ASORT_MIN_RADIX)
+  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 * ctx->elt_size < sorter_radix_threshold)

     {
     {

-      w->ctx->quicksort(w->out, w->num_elts);
-      if (!w->swap_output)
-       memcpy(w->in, w->out, w->num_elts * w->ctx->elt_size);
+      w->ctx->quicksort(w->in, w->num_elts);
+      if (w->swap_output)
+       memcpy(w->out, w->in, w->num_elts * w->ctx->elt_size);

     }
   else
     }
   else

-    asort_radix(w->ctx, w->out, w->in, w->num_elts, w->shift, !w->swap_output);
-  DBG("Thread %d: Finishing done", thr->id);
+    asort_radix(w->ctx, w->in, w->out, 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
 }
 
 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);
 
   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++)
   // Start parallel counting
   void *iptr = array;
   for (uns i=0; i<sorter_threads; i++)


@@ -244,12 +261,13 @@ rs_radix(struct asort_context *ctx, void *array, void *buffer, uns num_elts, uns
       w->w.go = rs_count;
       w->ctx = ctx;
       w->in = iptr;
       w->w.go = rs_count;
       w->ctx = ctx;
       w->in = iptr;

-      w->out = ctx->buffer;
+      w->out = buffer;

       w->num_elts = blksize;
       if (i == sorter_threads-1)
        w->num_elts += num_elts % sorter_threads;
       w->shift = shift;
       iptr += w->num_elts * ctx->elt_size;
       w->num_elts = blksize;
       if (i == sorter_threads-1)
        w->num_elts += num_elts % sorter_threads;
       w->shift = shift;
       iptr += w->num_elts * ctx->elt_size;

+      bzero(w->cnt, sizeof(uns) * buckets);

       work_submit(ctx->rs_work_queue, &w->w);
     }
 
       work_submit(ctx->rs_work_queue, &w->w);
     }
 


@@ -297,19 +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;
   for (uns i=0; i<buckets; i++)
     {
       uns n = cnt[i] - pos;

-      if (n < sorter_thread_threshold)
+      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;
        {
          struct rs_work *w = ep_alloc(ctx->eltpool);
          w->w.priority = 0;
          w->w.go = rs_finish;
          w->ctx = ctx;

-         w->in = array;
-         w->out = buffer;
+         w->in = buffer;
+         w->out = array;

          w->num_elts = n;
          w->shift = shift;
          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);
+         w->swap_output = !swapped_output;
+         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);
        }
       else
        rs_radix(ctx, buffer, array, n, shift, !swapped_output);


@@ -334,8 +363,7 @@ 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));
 
   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
   ctx->eltpool = ep_new(sizeof(struct rs_work), 1000);
 
   // Do the big splitting


@@ -345,12 +373,9 @@ threaded_radixsort(struct asort_context *ctx)
     big_free(ctx->rs_works[i], sizeof(struct rs_work) + sizeof(uns) * (1 << ctx->radix_bits));
 
   // Finish the small blocks
     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();
   ep_delete(ctx->eltpool);
   work_queue_cleanup(&q);
   asort_stop_threads();


@@ -371,7 +396,7 @@ asort_run(struct asort_context *ctx)
                              ctx->num_elts * ctx->elt_size >= sorter_thread_threshold &&
                              !(sorter_debug & SORT_DEBUG_ASORT_NO_THREADS));
 
                              ctx->num_elts * ctx->elt_size >= sorter_thread_threshold &&
                              !(sorter_debug & SORT_DEBUG_ASORT_NO_THREADS));
 

-  if (ctx->num_elts < ASORT_MIN_RADIX ||
+  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))
       ctx->hash_bits <= ASORT_MIN_SHIFT ||
       !ctx->radix_split ||
       (sorter_debug & SORT_DEBUG_ASORT_NO_RADIX))