#include <string.h>
#include <alloca.h>
-#define ASORT_MIN_RADIX 5000 // FIXME: var?
#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)
{
for (uns i=0; i<buckets; i++)
{
uns n = cnt[i] - pos;
- if (n < ASORT_MIN_RADIX || shift < ASORT_MIN_SHIFT)
+ if (n < ctx->radix_threshold || shift < ASORT_MIN_SHIFT)
{
ctx->quicksort(buffer, n);
if (!swapped_output)
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)
{
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;
}
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;
}
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);
}
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;
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);
}
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);
}
{
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 < 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);
- 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
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++)
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;
for (uns i=0; i<buckets; i++)
{
uns n = cnt[i] - pos;
- if (n * ctx->elt_size < sorter_thread_threshold)
+ if (!n)
+ continue;
+ 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;
- clist_add_tail(&ctx->rs_bits, &w->w.n);
- DBG("Scheduling block %d+%d", pos, n);
+ if (n < ctx->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);
}
static void
-threaded_radixsort(struct asort_context *ctx)
+threaded_radixsort(struct asort_context *ctx, uns swap)
{
struct work_queue q;
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();
#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 < ASORT_MIN_RADIX ||
+ if (ctx->num_elts < ctx->radix_threshold ||
ctx->hash_bits <= ASORT_MIN_SHIFT ||
!ctx->radix_split ||
(sorter_debug & SORT_DEBUG_ASORT_NO_RADIX))
#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");
}