freedreno: threaded batch flush

With the state accessed from GMEM+submit factored out of fd_context and
into fd_batch, now it is possible to punt this off to a helper thread.
And more importantly, since there are cases where one context might
force the batch-cache to flush another context's batches (ie. when there
are too many in-flight batches), using a per-context helper thread keeps
various different flushes for a given context serialized.

TODO as with batch-cache, there are a few places where we'll need a
mutex to protect critical sections, which is completely missing at the
moment.

Signed-off-by: Rob Clark <robdclark@gmail.com>

src/gallium/drivers/freedreno/freedreno_batch.c

@@ -40,6 +40,9 @@ batch_init(struct fd_batch *batch)
 	struct fd_context *ctx = batch->ctx;
 	unsigned size = 0;
 
+	if (ctx->screen->reorder)
+		util_queue_fence_init(&batch->flush_fence);
+
 	/* if kernel is too old to support unlimited # of cmd buffers, we
 	 * have no option but to allocate large worst-case sizes so that
 	 * we don't need to grow the ringbuffer.  Performance is likely to
@@ -119,6 +122,9 @@ batch_fini(struct fd_batch *batch)
 		fd_hw_sample_reference(batch->ctx, &samp, NULL);
 	}
 	util_dynarray_fini(&batch->samples);
+
+	if (batch->ctx->screen->reorder)
+		util_queue_fence_destroy(&batch->flush_fence);
 }
 
 static void
@@ -129,7 +135,7 @@ batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 
 	foreach_batch(dep, cache, batch->dependents_mask) {
 		if (flush)
-			fd_batch_flush(dep);
+			fd_batch_flush(dep, false);
 		fd_batch_reference(&dep, NULL);
 	}
 
@@ -156,6 +162,8 @@ batch_reset(struct fd_batch *batch)
 {
 	DBG("%p", batch);
 
+	fd_batch_sync(batch);
+
 	batch_flush_reset_dependencies(batch, false);
 	batch_reset_resources(batch);
 
@@ -197,6 +205,31 @@ __fd_batch_describe(char* buf, const struct fd_batch *batch)
 	util_sprintf(buf, "fd_batch<%u>", batch->seqno);
 }
 
+void
+fd_batch_sync(struct fd_batch *batch)
+{
+	if (!batch->ctx->screen->reorder)
+		return;
+	util_queue_job_wait(&batch->flush_fence);
+}
+
+static void
+batch_flush_func(void *job, int id)
+{
+	struct fd_batch *batch = job;
+
+	fd_gmem_render_tiles(batch);
+	batch_reset_resources(batch);
+	batch->ctx->last_fence = fd_ringbuffer_timestamp(batch->gmem);
+}
+
+static void
+batch_cleanup_func(void *job, int id)
+{
+	struct fd_batch *batch = job;
+	fd_batch_reference(&batch, NULL);
+}
+
 static void
 batch_flush(struct fd_batch *batch)
 {
@@ -207,11 +240,25 @@ batch_flush(struct fd_batch *batch)
 
 	batch->needs_flush = false;
 
+	/* close out the draw cmds by making sure any active queries are
+	 * paused:
+	 */
+	fd_hw_query_set_stage(batch, batch->draw, FD_STAGE_NULL);
+
+	batch->ctx->dirty = ~0;
 	batch_flush_reset_dependencies(batch, true);
 
-	fd_gmem_render_tiles(batch);
-
-	batch_reset_resources(batch);
+	if (batch->ctx->screen->reorder) {
+		struct fd_batch *tmp = NULL;
+		fd_batch_reference(&tmp, batch);
+		util_queue_add_job(&batch->ctx->flush_queue,
+				batch, &batch->flush_fence,
+				batch_flush_func, batch_cleanup_func);
+	} else {
+		fd_gmem_render_tiles(batch);
+		batch_reset_resources(batch);
+		batch->ctx->last_fence = fd_ringbuffer_timestamp(batch->gmem);
+	}
 
 	debug_assert(batch->reference.count > 0);
 
@@ -222,8 +269,9 @@ batch_flush(struct fd_batch *batch)
 	}
 }
 
+/* NOTE: could drop the last ref to batch */
 void
-fd_batch_flush(struct fd_batch *batch)
+fd_batch_flush(struct fd_batch *batch, bool sync)
 {
 	/* NOTE: we need to hold an extra ref across the body of flush,
 	 * since the last ref to this batch could be dropped when cleaning
@@ -232,6 +280,8 @@ fd_batch_flush(struct fd_batch *batch)
 	struct fd_batch *tmp = NULL;
 	fd_batch_reference(&tmp, batch);
 	batch_flush(tmp);
+	if (sync)
+		fd_batch_sync(tmp);
 	fd_batch_reference(&tmp, NULL);
 }
 
@@ -263,7 +313,7 @@ batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 	 */
 	if (batch_depends_on(dep, batch)) {
 		DBG("%p: flush forced on %p!", batch, dep);
-		fd_batch_flush(dep);
+		fd_batch_flush(dep, false);
 	} else {
 		struct fd_batch *other = NULL;
 		fd_batch_reference(&other, dep);
@@ -327,5 +377,5 @@ fd_batch_check_size(struct fd_batch *batch)
 	struct fd_ringbuffer *ring = batch->draw;
 	if (((ring->cur - ring->start) > (ring->size/4 - 0x1000)) ||
 			(fd_mesa_debug & FD_DBG_FLUSH))
-		fd_batch_flush(batch);
+		fd_batch_flush(batch, true);
 }

src/gallium/drivers/freedreno/freedreno_batch.h

@@ -28,6 +28,7 @@
 #define FREEDRENO_BATCH_H_
 
 #include "util/u_inlines.h"
+#include "util/u_queue.h"
 #include "util/list.h"
 
 #include "freedreno_util.h"
@@ -76,6 +77,8 @@ struct fd_batch {
 
 	struct fd_context *ctx;
 
+	struct util_queue_fence flush_fence;
+
 	/* do we need to mem2gmem before rendering.  We don't, if for example,
 	 * there was a glClear() that invalidated the entire previous buffer
 	 * contents.  Keep track of which buffer(s) are cleared, or needs
@@ -197,7 +200,8 @@ struct fd_batch {
 struct fd_batch * fd_batch_create(struct fd_context *ctx);
 
 void fd_batch_reset(struct fd_batch *batch);
-void fd_batch_flush(struct fd_batch *batch);
+void fd_batch_sync(struct fd_batch *batch);
+void fd_batch_flush(struct fd_batch *batch, bool sync);
 void fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write);
 void fd_batch_check_size(struct fd_batch *batch);
 

src/gallium/drivers/freedreno/freedreno_batch_cache.c

@@ -128,19 +128,24 @@ uint32_t
 fd_bc_flush(struct fd_batch_cache *cache, struct fd_context *ctx)
 {
 	struct hash_entry *entry;
-	uint32_t timestamp = 0;
+	struct fd_batch *last_batch = NULL;
 
 	hash_table_foreach(cache->ht, entry) {
 		struct fd_batch *batch = NULL;
 		fd_batch_reference(&batch, (struct fd_batch *)entry->data);
 		if (batch->ctx == ctx) {
-			fd_batch_flush(batch);
-			timestamp = MAX2(timestamp, fd_ringbuffer_timestamp(batch->gmem));
+			fd_batch_reference(&last_batch, batch);
+			fd_batch_flush(batch, false);
 		}
 		fd_batch_reference(&batch, NULL);
 	}
 
-	return timestamp;
+	if (last_batch) {
+		fd_batch_sync(last_batch);
+		fd_batch_reference(&last_batch, NULL);
+	}
+
+	return ctx->last_fence;
 }
 
 void
@@ -238,7 +243,7 @@ fd_bc_alloc_batch(struct fd_batch_cache *cache, struct fd_context *ctx)
 				fd_batch_reference(&flush_batch, cache->batches[i]);
 		}
 		DBG("%p: too many batches!  flush forced!", flush_batch);
-		fd_batch_flush(flush_batch);
+		fd_batch_flush(flush_batch, true);
 
 		/* While the resources get cleaned up automatically, the flush_batch
 		 * doesn't get removed from the dependencies of other batches, so

src/gallium/drivers/freedreno/freedreno_context.c

@@ -48,7 +48,7 @@ fd_context_flush(struct pipe_context *pctx, struct pipe_fence_handle **fence,
 	if (!ctx->screen->reorder) {
 		struct fd_batch *batch = NULL;
 		fd_batch_reference(&batch, ctx->batch);
-		fd_batch_flush(batch);
+		fd_batch_flush(batch, true);
 		timestamp = fd_ringbuffer_timestamp(batch->gmem);
 		fd_batch_reference(&batch, NULL);
 	} else {
@@ -103,6 +103,9 @@ fd_context_destroy(struct pipe_context *pctx)
 
 	DBG("");
 
+	if (ctx->screen->reorder)
+		util_queue_destroy(&ctx->flush_queue);
+
 	fd_batch_reference(&ctx->batch, NULL);  /* unref current batch */
 	fd_bc_invalidate_context(ctx);
 
@@ -179,8 +182,11 @@ fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
 	 * batches per compute job (since it isn't using tiling, so no point
 	 * in getting involved with the re-ordering madness)..
 	 */
-	if (!screen->reorder)
+	if (!screen->reorder) {
 		ctx->batch = fd_bc_alloc_batch(&screen->batch_cache, ctx);
+	} else {
+		util_queue_init(&ctx->flush_queue, "flush_queue", 16, 1);
+	}
 
 	fd_reset_wfi(ctx);
 

src/gallium/drivers/freedreno/freedreno_context.h

@@ -114,6 +114,8 @@ struct fd_context {
 	struct fd_device *dev;
 	struct fd_screen *screen;
 
+	struct util_queue flush_queue;
+
 	struct blitter_context *blitter;
 	struct primconvert_context *primconvert;
 
@@ -161,6 +163,8 @@ struct fd_context {
 	 */
 	struct fd_batch *batch;
 
+	uint32_t last_fence;
+
 	/* Are we in process of shadowing a resource? Used to detect recursion
 	 * in transfer_map, and skip unneeded synchronization.
 	 */

src/gallium/drivers/freedreno/freedreno_gmem.c

@@ -405,8 +405,6 @@ fd_gmem_render_tiles(struct fd_batch *batch)
 	fd_ringbuffer_flush(batch->gmem);
 
 	fd_reset_wfi(ctx);
-
-	ctx->dirty = ~0;
 }
 
 /* tile needs restore if it isn't completely contained within the

src/gallium/drivers/freedreno/freedreno_query_hw.c

@@ -238,7 +238,7 @@ fd_hw_get_query_result(struct fd_context *ctx, struct fd_query *q,
 			 * spin forever:
 			 */
 			if (hq->no_wait_cnt++ > 5)
-				fd_batch_flush(rsc->write_batch);
+				fd_batch_flush(rsc->write_batch, false);
 			return false;
 		}
 
@@ -266,7 +266,7 @@ fd_hw_get_query_result(struct fd_context *ctx, struct fd_query *q,
 		struct fd_resource *rsc = fd_resource(start->prsc);
 
 		if (rsc->write_batch)
-			fd_batch_flush(rsc->write_batch);
+			fd_batch_flush(rsc->write_batch, true);
 
 		/* some piglit tests at least do query with no draws, I guess: */
 		if (!rsc->bo)

src/gallium/drivers/freedreno/freedreno_resource.c

@@ -516,12 +516,18 @@ fd_resource_transfer_map(struct pipe_context *pctx,
 
 		if (needs_flush) {
 			if (usage & PIPE_TRANSFER_WRITE) {
-				struct fd_batch *batch;
-				foreach_batch(batch, &ctx->screen->batch_cache, rsc->batch_mask)
-					fd_batch_flush(batch);
+				struct fd_batch *batch, *last_batch = NULL;
+				foreach_batch(batch, &ctx->screen->batch_cache, rsc->batch_mask) {
+					fd_batch_reference(&last_batch, batch);
+					fd_batch_flush(batch, false);
+				}
+				if (last_batch) {
+					fd_batch_sync(last_batch);
+					fd_batch_reference(&last_batch, NULL);
+				}
 				assert(rsc->batch_mask == 0);
 			} else {
-				fd_batch_flush(rsc->write_batch);
+				fd_batch_flush(rsc->write_batch, true);
 			}
 			assert(!rsc->write_batch);
 		}
@@ -1080,7 +1086,7 @@ fd_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
 	struct fd_resource *rsc = fd_resource(prsc);
 
 	if (rsc->write_batch)
-		fd_batch_flush(rsc->write_batch);
+		fd_batch_flush(rsc->write_batch, true);
 
 	assert(!rsc->write_batch);
 }

src/gallium/drivers/freedreno/freedreno_state.c

@@ -137,14 +137,14 @@ fd_set_framebuffer_state(struct pipe_context *pctx,
 			 * multiple times to the same surface), so we might as
 			 * well go ahead and flush this one:
 			 */
-			fd_batch_flush(old_batch);
+			fd_batch_flush(old_batch, false);
 		}
 
 		fd_batch_reference(&old_batch, NULL);
 	} else {
 		DBG("%d: cbufs[0]=%p, zsbuf=%p", ctx->batch->needs_flush,
 				framebuffer->cbufs[0], framebuffer->zsbuf);
-		fd_batch_flush(ctx->batch);
+		fd_batch_flush(ctx->batch, false);
 	}
 
 	cso = &ctx->batch->framebuffer;