X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Ffreedreno%2Ffreedreno_batch.c;h=9918c259e48e9dac69ae8b36d012d1f9be679241;hb=f07154421ac5bb546dbd75fb74e8f2f8883c59c0;hp=2dd7eda72adfd1f9a507e9a9bc35a5e665249372;hpb=9f219c7047b51561f6f69274d445e6a6ec41c5f8;p=mesa.git

diff --git a/src/gallium/drivers/freedreno/freedreno_batch.c b/src/gallium/drivers/freedreno/freedreno_batch.c
index 2dd7eda72ad..9918c259e48 100644
--- a/src/gallium/drivers/freedreno/freedreno_batch.c
+++ b/src/gallium/drivers/freedreno/freedreno_batch.c
@@ -31,7 +31,9 @@
 
 #include "freedreno_batch.h"
 #include "freedreno_context.h"
+#include "freedreno_fence.h"
 #include "freedreno_resource.h"
+#include "freedreno_query_hw.h"
 
 static void
 batch_init(struct fd_batch *batch)
@@ -39,43 +41,59 @@ 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
 	 * suffer, but there is no good alternative.
 	 */
-	if (fd_device_version(ctx->screen->dev) < FD_VERSION_UNLIMITED_CMDS) {
+	if ((fd_device_version(ctx->screen->dev) < FD_VERSION_UNLIMITED_CMDS) ||
+			(fd_mesa_debug & FD_DBG_NOGROW)){
 		size = 0x100000;
 	}
 
-	batch->draw    = fd_ringbuffer_new(ctx->screen->pipe, size);
-	batch->binning = fd_ringbuffer_new(ctx->screen->pipe, size);
-	batch->gmem    = fd_ringbuffer_new(ctx->screen->pipe, size);
+	batch->draw    = fd_ringbuffer_new(ctx->pipe, size);
+	if (!batch->nondraw) {
+		batch->binning = fd_ringbuffer_new(ctx->pipe, size);
+		batch->gmem    = fd_ringbuffer_new(ctx->pipe, size);
 
-	fd_ringbuffer_set_parent(batch->gmem, NULL);
-	fd_ringbuffer_set_parent(batch->draw, batch->gmem);
-	fd_ringbuffer_set_parent(batch->binning, batch->gmem);
+		fd_ringbuffer_set_parent(batch->gmem, NULL);
+		fd_ringbuffer_set_parent(batch->draw, batch->gmem);
+		fd_ringbuffer_set_parent(batch->binning, batch->gmem);
+	} else {
+		fd_ringbuffer_set_parent(batch->draw, NULL);
+	}
+
+	batch->in_fence_fd = -1;
+	batch->fence = fd_fence_create(batch);
 
 	batch->cleared = batch->partial_cleared = 0;
 	batch->restore = batch->resolve = 0;
 	batch->needs_flush = false;
 	batch->gmem_reason = 0;
 	batch->num_draws = 0;
+	batch->stage = FD_STAGE_NULL;
+
+	fd_reset_wfi(batch);
 
 	/* reset maximal bounds: */
 	batch->max_scissor.minx = batch->max_scissor.miny = ~0;
 	batch->max_scissor.maxx = batch->max_scissor.maxy = 0;
 
-	util_dynarray_init(&batch->draw_patches);
+	util_dynarray_init(&batch->draw_patches, NULL);
 
 	if (is_a3xx(ctx->screen))
-		util_dynarray_init(&batch->rbrc_patches);
+		util_dynarray_init(&batch->rbrc_patches, NULL);
 
 	assert(batch->resources->entries == 0);
+
+	util_dynarray_init(&batch->samples, NULL);
 }
 
 struct fd_batch *
-fd_batch_create(struct fd_context *ctx)
+fd_batch_create(struct fd_context *ctx, bool nondraw)
 {
 	struct fd_batch *batch = CALLOC_STRUCT(fd_batch);
 
@@ -86,6 +104,7 @@ fd_batch_create(struct fd_context *ctx)
 
 	pipe_reference_init(&batch->reference, 1);
 	batch->ctx = ctx;
+	batch->nondraw = nondraw;
 
 	batch->resources = _mesa_set_create(NULL, _mesa_hash_pointer,
 			_mesa_key_pointer_equal);
@@ -98,14 +117,43 @@ fd_batch_create(struct fd_context *ctx)
 static void
 batch_fini(struct fd_batch *batch)
 {
+	pipe_resource_reference(&batch->query_buf, NULL);
+
+	if (batch->in_fence_fd != -1)
+		close(batch->in_fence_fd);
+
+	/* in case batch wasn't flushed but fence was created: */
+	fd_fence_populate(batch->fence, 0, -1);
+
+	fd_fence_ref(NULL, &batch->fence, NULL);
+
 	fd_ringbuffer_del(batch->draw);
-	fd_ringbuffer_del(batch->binning);
-	fd_ringbuffer_del(batch->gmem);
+	if (!batch->nondraw) {
+		fd_ringbuffer_del(batch->binning);
+		fd_ringbuffer_del(batch->gmem);
+	} else {
+		debug_assert(!batch->binning);
+		debug_assert(!batch->gmem);
+	}
+	if (batch->lrz_clear) {
+		fd_ringbuffer_del(batch->lrz_clear);
+		batch->lrz_clear = NULL;
+	}
 
 	util_dynarray_fini(&batch->draw_patches);
 
 	if (is_a3xx(batch->ctx->screen))
 		util_dynarray_fini(&batch->rbrc_patches);
+
+	while (batch->samples.size > 0) {
+		struct fd_hw_sample *samp =
+			util_dynarray_pop(&batch->samples, struct fd_hw_sample *);
+		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
@@ -116,7 +164,7 @@ 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, false);
 		fd_batch_reference(&dep, NULL);
 	}
 
@@ -124,25 +172,37 @@ batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 }
 
 static void
-batch_reset_resources(struct fd_batch *batch)
+batch_reset_resources_locked(struct fd_batch *batch)
 {
 	struct set_entry *entry;
 
+	pipe_mutex_assert_locked(batch->ctx->screen->lock);
+
 	set_foreach(batch->resources, entry) {
 		struct fd_resource *rsc = (struct fd_resource *)entry->key;
 		_mesa_set_remove(batch->resources, entry);
 		debug_assert(rsc->batch_mask & (1 << batch->idx));
 		rsc->batch_mask &= ~(1 << batch->idx);
 		if (rsc->write_batch == batch)
-			fd_batch_reference(&rsc->write_batch, NULL);
+			fd_batch_reference_locked(&rsc->write_batch, NULL);
 	}
 }
 
+static void
+batch_reset_resources(struct fd_batch *batch)
+{
+	mtx_lock(&batch->ctx->screen->lock);
+	batch_reset_resources_locked(batch);
+	mtx_unlock(&batch->ctx->screen->lock);
+}
+
 static void
 batch_reset(struct fd_batch *batch)
 {
 	DBG("%p", batch);
 
+	fd_batch_sync(batch);
+
 	batch_flush_reset_dependencies(batch, false);
 	batch_reset_resources(batch);
 
@@ -160,12 +220,14 @@ fd_batch_reset(struct fd_batch *batch)
 void
 __fd_batch_destroy(struct fd_batch *batch)
 {
-	fd_bc_invalidate_batch(batch, true);
-
 	DBG("%p", batch);
 
 	util_copy_framebuffer_state(&batch->framebuffer, NULL);
 
+	mtx_lock(&batch->ctx->screen->lock);
+	fd_bc_invalidate_batch(batch, true);
+	mtx_unlock(&batch->ctx->screen->lock);
+
 	batch_fini(batch);
 
 	batch_reset_resources(batch);
@@ -184,41 +246,101 @@ __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_fence_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);
+}
+
+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)
+batch_flush(struct fd_batch *batch, bool force)
 {
 	DBG("%p: needs_flush=%d", batch, batch->needs_flush);
 
-	if (!batch->needs_flush)
+	if (!batch->needs_flush) {
+		if (force) {
+			fd_gmem_render_noop(batch);
+			goto out;
+		}
 		return;
+	}
 
 	batch->needs_flush = false;
 
+	/* close out the draw cmds by making sure any active queries are
+	 * paused:
+	 */
+	fd_batch_set_stage(batch, FD_STAGE_NULL);
+
+	fd_context_all_dirty(batch->ctx);
 	batch_flush_reset_dependencies(batch, true);
 
-	fd_gmem_render_tiles(batch);
+	if (batch->ctx->screen->reorder) {
+		struct fd_batch *tmp = NULL;
+		fd_batch_reference(&tmp, batch);
 
-	batch_reset_resources(batch);
+		if (!util_queue_is_initialized(&batch->ctx->flush_queue))
+			util_queue_init(&batch->ctx->flush_queue, "flush_queue", 16, 1, 0);
+
+		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);
+	}
 
 	debug_assert(batch->reference.count > 0);
 
+out:
 	if (batch == batch->ctx->batch) {
 		batch_reset(batch);
 	} else {
+		mtx_lock(&batch->ctx->screen->lock);
 		fd_bc_invalidate_batch(batch, false);
+		mtx_unlock(&batch->ctx->screen->lock);
 	}
 }
 
+/* NOTE: could drop the last ref to batch
+ *
+ * @sync: synchronize with flush_queue, ensures batch is *actually* flushed
+ *   to kernel before this returns, as opposed to just being queued to be
+ *   flushed
+ * @force: force a flush even if no rendering, mostly useful if you need
+ *   a fence to sync on
+ */
 void
-fd_batch_flush(struct fd_batch *batch)
+fd_batch_flush(struct fd_batch *batch, bool sync, bool force)
 {
 	/* 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
 	 * up used_resources
 	 */
 	struct fd_batch *tmp = NULL;
+
 	fd_batch_reference(&tmp, batch);
-	batch_flush(tmp);
+	batch_flush(tmp, force);
+	if (sync)
+		fd_batch_sync(tmp);
 	fd_batch_reference(&tmp, NULL);
 }
 
@@ -239,8 +361,8 @@ batch_depends_on(struct fd_batch *batch, struct fd_batch *other)
 	return false;
 }
 
-static void
-batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
+void
+fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 {
 	if (batch->dependents_mask & (1 << dep->idx))
 		return;
@@ -250,10 +372,12 @@ 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);
+		mtx_unlock(&batch->ctx->screen->lock);
+		fd_batch_flush(dep, false, false);
+		mtx_lock(&batch->ctx->screen->lock);
 	} else {
 		struct fd_batch *other = NULL;
-		fd_batch_reference(&other, dep);
+		fd_batch_reference_locked(&other, dep);
 		batch->dependents_mask |= (1 << dep->idx);
 		DBG("%p: added dependency on %p", batch, dep);
 	}
@@ -262,11 +386,16 @@ batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
 void
 fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write)
 {
+	pipe_mutex_assert_locked(batch->ctx->screen->lock);
+
 	if (rsc->stencil)
 		fd_batch_resource_used(batch, rsc->stencil, write);
 
 	DBG("%p: %s %p", batch, write ? "write" : "read", rsc);
 
+	if (write)
+		rsc->valid = true;
+
 	/* note, invalidate write batch, to avoid further writes to rsc
 	 * resulting in a write-after-read hazard.
 	 */
@@ -276,22 +405,37 @@ fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool wri
 		if (rsc->batch_mask != (1 << batch->idx)) {
 			struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
 			struct fd_batch *dep;
+
+			if (rsc->write_batch && rsc->write_batch != batch) {
+				struct fd_batch *b = NULL;
+				fd_batch_reference(&b, rsc->write_batch);
+
+				mtx_unlock(&batch->ctx->screen->lock);
+				fd_batch_flush(b, true, false);
+				mtx_lock(&batch->ctx->screen->lock);
+
+				fd_bc_invalidate_batch(b, false);
+				fd_batch_reference_locked(&b, NULL);
+			}
+
 			foreach_batch(dep, cache, rsc->batch_mask) {
 				struct fd_batch *b = NULL;
+				if (dep == batch)
+					continue;
 				/* note that batch_add_dep could flush and unref dep, so
 				 * we need to hold a reference to keep it live for the
 				 * fd_bc_invalidate_batch()
 				 */
 				fd_batch_reference(&b, dep);
-				batch_add_dep(batch, b);
+				fd_batch_add_dep(batch, b);
 				fd_bc_invalidate_batch(b, false);
 				fd_batch_reference_locked(&b, NULL);
 			}
 		}
-		fd_batch_reference(&rsc->write_batch, batch);
+		fd_batch_reference_locked(&rsc->write_batch, batch);
 	} else {
 		if (rsc->write_batch) {
-			batch_add_dep(batch, rsc->write_batch);
+			fd_batch_add_dep(batch, rsc->write_batch);
 			fd_bc_invalidate_batch(rsc->write_batch, false);
 		}
 	}
@@ -314,5 +458,20 @@ 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, false);
+}
+
+/* emit a WAIT_FOR_IDLE only if needed, ie. if there has not already
+ * been one since last draw:
+ */
+void
+fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring)
+{
+	if (batch->needs_wfi) {
+		if (batch->ctx->screen->gpu_id >= 500)
+			OUT_WFI5(ring);
+		else
+			OUT_WFI(ring);
+		batch->needs_wfi = false;
+	}
 }