freedreno: make fd_batch a one-shot thing
[mesa.git] / src / gallium / drivers / freedreno / freedreno_batch.c
index 76aaac8f0d967ac88ebea4782ba46a76b916fea7..18e4ca5fb03bff747ed7770a981247239a26d269 100644 (file)
@@ -31,6 +31,7 @@
 
 #include "freedreno_batch.h"
 #include "freedreno_context.h"
+#include "freedreno_fence.h"
 #include "freedreno_resource.h"
 #include "freedreno_query_hw.h"
 
@@ -48,21 +49,30 @@ batch_init(struct fd_batch *batch)
         * 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->flushed = false;
        batch->gmem_reason = 0;
        batch->num_draws = 0;
        batch->stage = FD_STAGE_NULL;
@@ -73,18 +83,18 @@ batch_init(struct fd_batch *batch)
        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);
+       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);
 
@@ -95,6 +105,7 @@ 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);
@@ -107,11 +118,30 @@ fd_batch_create(struct fd_context *ctx)
 static void
 batch_fini(struct fd_batch *batch)
 {
+       DBG("%p", 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);
 
@@ -137,7 +167,7 @@ batch_flush_reset_dependencies(struct fd_batch *batch, bool flush)
 
        foreach_batch(dep, cache, batch->dependents_mask) {
                if (flush)
-                       fd_batch_flush(dep, false);
+                       fd_batch_flush(dep, false, false);
                fd_batch_reference(&dep, NULL);
        }
 
@@ -164,9 +194,9 @@ batch_reset_resources_locked(struct fd_batch *batch)
 static void
 batch_reset_resources(struct fd_batch *batch)
 {
-       pipe_mutex_lock(batch->ctx->screen->lock);
+       mtx_lock(&batch->ctx->screen->lock);
        batch_reset_resources_locked(batch);
-       pipe_mutex_unlock(batch->ctx->screen->lock);
+       mtx_unlock(&batch->ctx->screen->lock);
 }
 
 static void
@@ -197,9 +227,9 @@ __fd_batch_destroy(struct fd_batch *batch)
 
        util_copy_framebuffer_state(&batch->framebuffer, NULL);
 
-       pipe_mutex_lock(batch->ctx->screen->lock);
+       mtx_lock(&batch->ctx->screen->lock);
        fd_bc_invalidate_batch(batch, true);
-       pipe_mutex_unlock(batch->ctx->screen->lock);
+       mtx_unlock(&batch->ctx->screen->lock);
 
        batch_fini(batch);
 
@@ -224,7 +254,7 @@ fd_batch_sync(struct fd_batch *batch)
 {
        if (!batch->ctx->screen->reorder)
                return;
-       util_queue_job_wait(&batch->flush_fence);
+       util_queue_fence_wait(&batch->flush_fence);
 }
 
 static void
@@ -232,9 +262,10 @@ batch_flush_func(void *job, int id)
 {
        struct fd_batch *batch = job;
 
+       DBG("%p", batch);
+
        fd_gmem_render_tiles(batch);
        batch_reset_resources(batch);
-       batch->ctx->last_fence = fd_ringbuffer_timestamp(batch->gmem);
 }
 
 static void
@@ -245,11 +276,11 @@ batch_cleanup_func(void *job, int id)
 }
 
 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->flushed)
                return;
 
        batch->needs_flush = false;
@@ -257,47 +288,76 @@ batch_flush(struct fd_batch *batch)
        /* close out the draw cmds by making sure any active queries are
         * paused:
         */
-       fd_hw_query_set_stage(batch, batch->draw, FD_STAGE_NULL);
+       fd_batch_set_stage(batch, FD_STAGE_NULL);
 
-       batch->ctx->dirty = ~0;
+       fd_context_all_dirty(batch->ctx);
        batch_flush_reset_dependencies(batch, true);
 
+       batch->flushed = true;
+
        if (batch->ctx->screen->reorder) {
                struct fd_batch *tmp = NULL;
                fd_batch_reference(&tmp, 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);
-               batch->ctx->last_fence = fd_ringbuffer_timestamp(batch->gmem);
        }
 
        debug_assert(batch->reference.count > 0);
 
-       if (batch == batch->ctx->batch) {
-               batch_reset(batch);
-       } else {
-               pipe_mutex_lock(batch->ctx->screen->lock);
-               fd_bc_invalidate_batch(batch, false);
-               pipe_mutex_unlock(batch->ctx->screen->lock);
-       }
+       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 */
+/* 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, bool sync)
+fd_batch_flush(struct fd_batch *batch, bool sync, bool force)
 {
+       struct fd_batch *tmp = NULL;
+       bool newbatch = false;
+
        /* 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);
+
+       if (batch == batch->ctx->batch) {
+               batch->ctx->batch = NULL;
+               newbatch = true;
+       }
+
+       batch_flush(tmp, force);
+
+       if (newbatch) {
+               struct fd_context *ctx = batch->ctx;
+               struct fd_batch *new_batch =
+                       fd_batch_from_fb(&ctx->screen->batch_cache, ctx, &batch->framebuffer);
+
+               util_copy_framebuffer_state(&new_batch->framebuffer, &batch->framebuffer);
+
+               fd_batch_reference(&batch, NULL);
+               ctx->batch = new_batch;
+       }
+
        if (sync)
                fd_batch_sync(tmp);
+
        fd_batch_reference(&tmp, NULL);
 }
 
@@ -318,26 +378,33 @@ 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;
 
-       /* if the new depedency already depends on us, we need to flush
-        * to avoid a loop in the dependency graph.
-        */
-       if (batch_depends_on(dep, batch)) {
-               DBG("%p: flush forced on %p!", batch, dep);
-               pipe_mutex_unlock(batch->ctx->screen->lock);
-               fd_batch_flush(dep, false);
-               pipe_mutex_lock(batch->ctx->screen->lock);
-       } else {
-               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);
-       }
+       /* a loop should not be possible */
+       debug_assert(!batch_depends_on(dep, batch));
+
+       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)
+{
+       struct fd_batch *b = NULL;
+       fd_batch_reference(&b, rsc->write_batch);
+
+       mtx_unlock(&b->ctx->screen->lock);
+       fd_batch_flush(b, true, false);
+       mtx_lock(&b->ctx->screen->lock);
+
+       fd_bc_invalidate_batch(b, false);
+       fd_batch_reference_locked(&b, NULL);
 }
 
 void
@@ -350,33 +417,44 @@ fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool wri
 
        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.
         */
 
        if (write) {
                /* if we are pending read or write by any other batch: */
-               if (rsc->batch_mask != (1 << batch->idx)) {
+               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)
+                               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);
-                               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_locked(&rsc->write_batch, batch);
        } else {
-               if (rsc->write_batch) {
-                       batch_add_dep(batch, rsc->write_batch);
-                       fd_bc_invalidate_batch(rsc->write_batch, false);
-               }
+               /* 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()
+                */
+               if (rsc->write_batch && rsc->write_batch != batch)
+                       flush_write_batch(rsc);
        }
 
        if (rsc->batch_mask & (1 << batch->idx))
@@ -391,11 +469,28 @@ fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool wri
 void
 fd_batch_check_size(struct fd_batch *batch)
 {
+       debug_assert(!batch->flushed);
+
        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_batch_flush(batch, true);
+               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;
+       }
 }