X-Git-Url: https://git.libre-soc.org/?p=mesa.git;a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Ffreedreno%2Ffreedreno_batch.c;h=86e8b5b7834e038a6ceead8193a670fd8a03e051;hp=6d17a422fc44ef91ae547f6c2085455b1f167083;hb=28079970f2ce151673701c84ad765a235ab5ba50;hpb=eeafaf2d37cdc7f83f997e8babd8f770243ecf25

diff --git a/src/gallium/drivers/freedreno/freedreno_batch.c b/src/gallium/drivers/freedreno/freedreno_batch.c
index 6d17a422fc4..86e8b5b7834 100644
--- a/src/gallium/drivers/freedreno/freedreno_batch.c
+++ b/src/gallium/drivers/freedreno/freedreno_batch.c
@@ -25,109 +25,496 @@
  */
 
 #include "util/list.h"
+#include "util/set.h"
+#include "util/hash_table.h"
 #include "util/u_string.h"
 
 #include "freedreno_batch.h"
 #include "freedreno_context.h"
+#include "freedreno_fence.h"
 #include "freedreno_resource.h"
+#include "freedreno_query_hw.h"
+
+static struct fd_ringbuffer *
+alloc_ring(struct fd_batch *batch, unsigned sz, enum fd_ringbuffer_flags flags)
+{
+	struct fd_context *ctx = batch->ctx;
+
+	/* 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.
+	 *
+	 * Otherwise if supported, allocate a growable ring with initial
+	 * size of zero.
+	 */
+	if ((fd_device_version(ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS) &&
+			!(fd_mesa_debug & FD_DBG_NOGROW)){
+		flags |= FD_RINGBUFFER_GROWABLE;
+		sz = 0;
+	}
+
+	return fd_submit_new_ringbuffer(batch->submit, sz, flags);
+}
+
+static void
+batch_init(struct fd_batch *batch)
+{
+	struct fd_context *ctx = batch->ctx;
+
+	batch->submit = fd_submit_new(ctx->pipe);
+	if (batch->nondraw) {
+		batch->draw = alloc_ring(batch, 0x100000, FD_RINGBUFFER_PRIMARY);
+	} else {
+		batch->gmem = alloc_ring(batch, 0x100000, FD_RINGBUFFER_PRIMARY);
+		batch->draw = alloc_ring(batch, 0x100000, 0);
+
+		/* a6xx+ re-uses draw rb for both draw and binning pass: */
+		if (ctx->screen->gpu_id < 600) {
+			batch->binning = alloc_ring(batch, 0x100000, 0);
+		}
+	}
+
+	batch->in_fence_fd = -1;
+	batch->fence = fd_fence_create(batch);
+
+	batch->cleared = 0;
+	batch->fast_cleared = 0;
+	batch->invalidated = 0;
+	batch->restore = batch->resolve = 0;
+	batch->needs_flush = false;
+	batch->flushed = false;
+	batch->gmem_reason = 0;
+	batch->num_draws = 0;
+	batch->num_vertices = 0;
+	batch->num_bins_per_pipe = 0;
+	batch->prim_strm_bits = 0;
+	batch->draw_strm_bits = 0;
+	batch->stage = FD_STAGE_NULL;
+
+	fd_reset_wfi(batch);
+
+	util_dynarray_init(&batch->draw_patches, NULL);
+	util_dynarray_init(&batch->fb_read_patches, NULL);
+
+	if (is_a2xx(ctx->screen)) {
+		util_dynarray_init(&batch->shader_patches, NULL);
+		util_dynarray_init(&batch->gmem_patches, NULL);
+	}
+
+	if (is_a3xx(ctx->screen))
+		util_dynarray_init(&batch->rbrc_patches, NULL);
+
+	assert(batch->resources->entries == 0);
+
+	util_dynarray_init(&batch->samples, NULL);
+
+	list_inithead(&batch->log_chunks);
+}
 
 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);
-	static unsigned seqno = 0;
-	unsigned size = 0;
 
 	if (!batch)
 		return NULL;
 
+	DBG("%p", batch);
+
 	pipe_reference_init(&batch->reference, 1);
-	batch->seqno = ++seqno;
 	batch->ctx = ctx;
+	batch->nondraw = nondraw;
 
-	/* 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) {
-		size = 0x100000;
+	batch->resources = _mesa_set_create(NULL, _mesa_hash_pointer,
+			_mesa_key_pointer_equal);
+
+	batch_init(batch);
+
+	fd_screen_assert_locked(ctx->screen);
+	if (BATCH_DEBUG) {
+		_mesa_set_add(ctx->screen->live_batches, batch);
 	}
 
-	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);
+	return batch;
+}
+
+static void
+batch_fini(struct fd_batch *batch)
+{
+	DBG("%p", batch);
 
-	fd_ringbuffer_set_parent(batch->gmem, NULL);
-	fd_ringbuffer_set_parent(batch->draw, batch->gmem);
-	fd_ringbuffer_set_parent(batch->binning, batch->gmem);
+	pipe_resource_reference(&batch->query_buf, NULL);
 
-	list_inithead(&batch->used_resources);
+	if (batch->in_fence_fd != -1)
+		close(batch->in_fence_fd);
 
-	return batch;
+	/* in case batch wasn't flushed but fence was created: */
+	fd_fence_populate(batch->fence, 0, -1);
+
+	fd_fence_ref(&batch->fence, NULL);
+
+	fd_ringbuffer_del(batch->draw);
+	if (!batch->nondraw) {
+		if (batch->binning)
+			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;
+	}
+
+	if (batch->epilogue) {
+		fd_ringbuffer_del(batch->epilogue);
+		batch->epilogue = NULL;
+	}
+
+	if (batch->tile_setup) {
+		fd_ringbuffer_del(batch->tile_setup);
+		batch->tile_setup = NULL;
+	}
+
+	if (batch->tile_fini) {
+		fd_ringbuffer_del(batch->tile_fini);
+		batch->tile_fini = NULL;
+	}
+
+	if (batch->tessellation) {
+		fd_bo_del(batch->tessfactor_bo);
+		fd_bo_del(batch->tessparam_bo);
+		fd_ringbuffer_del(batch->tess_addrs_constobj);
+	}
+
+	fd_submit_del(batch->submit);
+
+	util_dynarray_fini(&batch->draw_patches);
+	util_dynarray_fini(&batch->fb_read_patches);
+
+	if (is_a2xx(batch->ctx->screen)) {
+		util_dynarray_fini(&batch->shader_patches);
+		util_dynarray_fini(&batch->gmem_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);
+
+	assert(list_is_empty(&batch->log_chunks));
+}
+
+static void
+batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
+{
+	struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
+	struct fd_batch *dep;
+
+	foreach_batch(dep, cache, batch->dependents_mask) {
+		if (flush)
+			fd_batch_flush(dep);
+		fd_batch_reference(&dep, NULL);
+	}
+
+	batch->dependents_mask = 0;
+}
+
+static void
+batch_reset_resources_locked(struct fd_batch *batch)
+{
+	fd_screen_assert_locked(batch->ctx->screen);
+
+	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_locked(&rsc->write_batch, NULL);
+	}
+}
+
+static void
+batch_reset_resources(struct fd_batch *batch)
+{
+	fd_screen_lock(batch->ctx->screen);
+	batch_reset_resources_locked(batch);
+	fd_screen_unlock(batch->ctx->screen);
+}
+
+static void
+batch_reset(struct fd_batch *batch)
+{
+	DBG("%p", batch);
+
+	batch_flush_reset_dependencies(batch, false);
+	batch_reset_resources(batch);
+
+	batch_fini(batch);
+	batch_init(batch);
+}
+
+void
+fd_batch_reset(struct fd_batch *batch)
+{
+	if (batch->needs_flush)
+		batch_reset(batch);
 }
 
 void
 __fd_batch_destroy(struct fd_batch *batch)
 {
-	fd_ringbuffer_del(batch->draw);
-	fd_ringbuffer_del(batch->binning);
-	fd_ringbuffer_del(batch->gmem);
+	struct fd_context *ctx = batch->ctx;
+
+	DBG("%p", batch);
+
+	fd_context_assert_locked(batch->ctx);
+
+	if (BATCH_DEBUG) {
+		_mesa_set_remove_key(ctx->screen->live_batches, batch);
+	}
+
+	fd_bc_invalidate_batch(batch, true);
+
+	batch_reset_resources_locked(batch);
+	debug_assert(batch->resources->entries == 0);
+	_mesa_set_destroy(batch->resources, NULL);
+
+	fd_context_unlock(ctx);
+	batch_flush_reset_dependencies(batch, false);
+	debug_assert(batch->dependents_mask == 0);
 
+	util_copy_framebuffer_state(&batch->framebuffer, NULL);
+	batch_fini(batch);
 	free(batch);
+	fd_context_lock(ctx);
 }
 
 void
 __fd_batch_describe(char* buf, const struct fd_batch *batch)
 {
-	util_sprintf(buf, "fd_batch<%u>", batch->seqno);
+	sprintf(buf, "fd_batch<%u>", batch->seqno);
 }
 
+static void
+batch_flush(struct fd_batch *batch)
+{
+	DBG("%p: needs_flush=%d", batch, batch->needs_flush);
+
+	if (batch->flushed)
+		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);
+
+	batch_flush_reset_dependencies(batch, true);
+
+	batch->flushed = true;
+
+	fd_fence_ref(&batch->ctx->last_fence, batch->fence);
+
+	fd_gmem_render_tiles(batch);
+	batch_reset_resources(batch);
+
+	debug_assert(batch->reference.count > 0);
+
+	fd_screen_lock(batch->ctx->screen);
+	fd_bc_invalidate_batch(batch, false);
+	fd_screen_unlock(batch->ctx->screen);
+}
+
+/* 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)
 {
-	struct fd_resource *rsc, *rsc_tmp;
+	struct fd_batch *tmp = NULL;
+
+	/* 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
+	 */
+	fd_batch_reference(&tmp, batch);
 
-	fd_gmem_render_tiles(batch->ctx);
+	batch_flush(tmp);
 
-	/* go through all the used resources and clear their reading flag */
-	LIST_FOR_EACH_ENTRY_SAFE(rsc, rsc_tmp, &batch->used_resources, list) {
-		debug_assert(rsc->pending_batch == batch);
-		debug_assert(rsc->status != 0);
-		rsc->status = 0;
-		fd_batch_reference(&rsc->pending_batch, NULL);
-		list_delinit(&rsc->list);
+	if (batch == batch->ctx->batch) {
+		fd_batch_reference(&batch->ctx->batch, NULL);
 	}
 
-	assert(LIST_IS_EMPTY(&batch->used_resources));
+	fd_batch_reference(&tmp, NULL);
+}
+
+/* find a batches dependents mask, including recursive dependencies: */
+static uint32_t
+recursive_dependents_mask(struct fd_batch *batch)
+{
+	struct fd_batch_cache *cache = &batch->ctx->screen->batch_cache;
+	struct fd_batch *dep;
+	uint32_t dependents_mask = batch->dependents_mask;
+
+	foreach_batch(dep, cache, batch->dependents_mask)
+		dependents_mask |= recursive_dependents_mask(dep);
+
+	return dependents_mask;
 }
 
 void
-fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc,
-		enum fd_resource_status status)
+fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep)
+{
+	fd_screen_assert_locked(batch->ctx->screen);
+
+	if (batch->dependents_mask & (1 << dep->idx))
+		return;
+
+	/* a loop should not be possible */
+	debug_assert(!((1 << batch->idx) & recursive_dependents_mask(dep)));
+
+	struct fd_batch *other = NULL;
+	fd_batch_reference_locked(&other, dep);
+	batch->dependents_mask |= (1 << dep->idx);
+	DBG("%p: added dependency on %p", batch, dep);
+}
+
+static void
+flush_write_batch(struct fd_resource *rsc)
 {
-	rsc->status |= status;
+	struct fd_batch *b = NULL;
+	fd_batch_reference_locked(&b, rsc->write_batch);
+
+	fd_screen_unlock(b->ctx->screen);
+	fd_batch_flush(b);
+	fd_screen_lock(b->ctx->screen);
+
+	fd_bc_invalidate_batch(b, false);
+	fd_batch_reference_locked(&b, NULL);
+}
+
+static void
+fd_batch_add_resource(struct fd_batch *batch, struct fd_resource *rsc)
+{
+
+	if (likely(fd_batch_references_resource(batch, rsc))) {
+		debug_assert(_mesa_set_search(batch->resources, rsc));
+		return;
+	}
+
+	debug_assert(!_mesa_set_search(batch->resources, rsc));
+
+	_mesa_set_add(batch->resources, rsc);
+	rsc->batch_mask |= (1 << batch->idx);
+}
+
+void
+fd_batch_resource_write(struct fd_batch *batch, struct fd_resource *rsc)
+{
+	fd_screen_assert_locked(batch->ctx->screen);
+
+	if (rsc->stencil)
+		fd_batch_resource_write(batch, rsc->stencil);
+
+	DBG("%p: write %p", batch, rsc);
+
+	rsc->valid = true;
+
+	/* note, invalidate write batch, to avoid further writes to rsc
+	 * resulting in a write-after-read hazard.
+	 */
+	/* if we are pending read or write by any other batch: */
+	if (unlikely(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)
+			flush_write_batch(rsc);
+
+		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);
+			fd_batch_add_dep(batch, b);
+			fd_bc_invalidate_batch(b, false);
+			fd_batch_reference_locked(&b, NULL);
+		}
+	}
+	fd_batch_reference_locked(&rsc->write_batch, batch);
+
+	fd_batch_add_resource(batch, rsc);
+}
+
+void
+fd_batch_resource_read_slowpath(struct fd_batch *batch, struct fd_resource *rsc)
+{
+	fd_screen_assert_locked(batch->ctx->screen);
 
 	if (rsc->stencil)
-		rsc->stencil->status |= status;
+		fd_batch_resource_read(batch, rsc->stencil);
 
-	/* TODO resources can actually be shared across contexts,
-	 * so I'm not sure a single list-head will do the trick?
+	DBG("%p: read %p", batch, rsc);
+
+	/* If reading a resource pending a write, go ahead and flush the
+	 * writer.  This avoids situations where we end up having to
+	 * flush the current batch in _resource_used()
 	 */
-	debug_assert((rsc->pending_batch == batch) || !rsc->pending_batch);
-	list_delinit(&rsc->list);
-	list_addtail(&rsc->list, &batch->used_resources);
-	fd_batch_reference(&rsc->pending_batch, batch);
+	if (unlikely(rsc->write_batch && rsc->write_batch != batch))
+		flush_write_batch(rsc);
+
+	fd_batch_add_resource(batch, rsc);
 }
 
 void
 fd_batch_check_size(struct fd_batch *batch)
 {
+	debug_assert(!batch->flushed);
+
+	if (unlikely(fd_mesa_debug & FD_DBG_FLUSH)) {
+		fd_batch_flush(batch);
+		return;
+	}
+
 	if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
 		return;
 
 	struct fd_ringbuffer *ring = batch->draw;
-	if (((ring->cur - ring->start) > (ring->size/4 - 0x1000)) ||
-			(fd_mesa_debug & FD_DBG_FLUSH))
-		fd_context_render(&batch->ctx->base);
+	if ((ring->cur - ring->start) > (ring->size/4 - 0x1000))
+		fd_batch_flush(batch);
+}
+
+/* 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;
+	}
 }