#include "util/u_dynarray.h"
#include "pipe/p_state.h"
-#include "pan_allocate.h"
+#include "pan_pool.h"
#include "pan_resource.h"
+#include "pan_scoreboard.h"
/* panfrost_batch_fence is the out fence of a batch that users or other batches
* might want to wait on. The batch fence lifetime is different from the batch
* batch has been flushed or not.
*/
struct panfrost_batch *batch;
-
- /* Context this fence is attached to. We need both ctx and batch, as
- * the batch will go away after it's been submitted, but the fence
- * will stay a bit longer.
- */
- struct panfrost_context *ctx;
-
- /* Sync object backing this fence. */
- uint32_t syncobj;
-
- /* Cached value of the signaled state to avoid calling WAIT_SYNCOBJs
- * when we know the fence has already been signaled.
- */
- bool signaled;
};
#define PAN_REQ_MSAA (1 << 0)
/* Buffers cleared (PIPE_CLEAR_* bitmask) */
unsigned clear;
+ /* Buffers drawn */
+ unsigned draws;
+
/* Packed clear values, indexed by both render target as well as word.
* Essentially, a single pixel is packed, with some padding to bring it
* up to a 32-bit interval; that pixel is then duplicated over to fill
float clear_depth;
unsigned clear_stencil;
+ /* Amount of thread local storage required per thread */
+ unsigned stack_size;
+
+ /* Amount of shared memory needed per workgroup (for compute) */
+ unsigned shared_size;
+
/* Whether this job uses the corresponding requirement (PAN_REQ_*
* bitmask) */
unsigned requirements;
unsigned minx, miny;
unsigned maxx, maxy;
- /* CPU pointers to the job descriptor headers. next_job is only
- * set at submit time (since only then are all the dependencies
- * known). The upshot is that this is append-only.
- *
- * These arrays contain the headers for the "primary batch", our jargon
- * referring to the part of the panfrost_job that actually contains
- * meaningful work. In an OpenGL ES setting, that means the
- * WRITE_VALUE/VERTEX/TILER jobs. Excluded is specifically the FRAGMENT
- * job, which is sent on as a secondary batch containing only a single
- * hardware job. Since there's one and only one FRAGMENT job issued per
- * panfrost_job, there is no need to do any scoreboarding / management;
- * it's easy enough to open-code it and it's not like we can get any
- * better anyway. */
- struct util_dynarray headers;
-
- /* (And the GPU versions; TODO maybe combine) */
- struct util_dynarray gpu_headers;
-
- /* The last job in the primary batch */
- struct panfrost_transfer last_job;
-
- /* The first/last tiler job */
- struct panfrost_transfer first_tiler;
- struct panfrost_transfer last_tiler;
-
- /* The first vertex job used as the input to a tiler job */
- struct panfrost_transfer first_vertex_for_tiler;
-
- /* The first job. Notice we've created a linked list */
- struct panfrost_transfer first_job;
-
- /* The number of jobs in the primary batch, essentially */
- unsigned job_index;
-
- /* BOs referenced -- will be used for flushing logic */
+ /* BOs referenced not in the pool */
struct hash_table *bos;
- /* Current transient BO */
- struct panfrost_bo *transient_bo;
+ /* Pool owned by this batch (released when the batch is released) used for temporary descriptors */
+ struct pan_pool pool;
- /* Within the topmost transient BO, how much has been used? */
- unsigned transient_offset;
+ /* Pool also owned by this batch that is not CPU mapped (created as
+ * INVISIBLE) used for private GPU-internal structures, particularly
+ * varyings */
+ struct pan_pool invisible_pool;
+
+ /* Job scoreboarding state */
+ struct pan_scoreboard scoreboard;
/* Polygon list bound to the batch, or NULL if none bound yet */
struct panfrost_bo *polygon_list;
- /* Scratchpath BO bound to the batch, or NULL if none bound yet */
+ /* Scratchpad BO bound to the batch, or NULL if none bound yet */
struct panfrost_bo *scratchpad;
+ /* Shared memory BO bound to the batch, or NULL if none bound yet */
+ struct panfrost_bo *shared_memory;
+
/* Tiler heap BO bound to the batch, or NULL if none bound yet */
struct panfrost_bo *tiler_heap;
/* Framebuffer descriptor. */
struct panfrost_transfer framebuffer;
+ /* Bifrost tiler meta descriptor. */
+ mali_ptr tiler_meta;
+
/* Output sync object. Only valid when submitted is true. */
struct panfrost_batch_fence *out_sync;
panfrost_batch_add_bo(struct panfrost_batch *batch, struct panfrost_bo *bo,
uint32_t flags);
-void panfrost_batch_add_fbo_bos(struct panfrost_batch *batch);
-
struct panfrost_bo *
panfrost_batch_create_bo(struct panfrost_batch *batch, size_t size,
uint32_t create_flags, uint32_t access_flags);
void
-panfrost_flush_all_batches(struct panfrost_context *ctx, bool wait);
+panfrost_flush_all_batches(struct panfrost_context *ctx, uint32_t out_sync);
bool
panfrost_pending_batches_access_bo(struct panfrost_context *ctx,
void
panfrost_flush_batches_accessing_bo(struct panfrost_context *ctx,
- struct panfrost_bo *bo, uint32_t flags);
+ struct panfrost_bo *bo, bool flush_readers);
void
panfrost_batch_set_requirements(struct panfrost_batch *batch);
-mali_ptr
-panfrost_batch_get_polygon_list(struct panfrost_batch *batch, unsigned size);
+void
+panfrost_batch_adjust_stack_size(struct panfrost_batch *batch);
struct panfrost_bo *
-panfrost_batch_get_scratchpad(struct panfrost_batch *batch);
+panfrost_batch_get_scratchpad(struct panfrost_batch *batch, unsigned size, unsigned thread_tls_alloc, unsigned core_count);
struct panfrost_bo *
-panfrost_batch_get_tiler_heap(struct panfrost_batch *batch);
+panfrost_batch_get_shared_memory(struct panfrost_batch *batch, unsigned size, unsigned workgroup_count);
+
+mali_ptr
+panfrost_batch_get_polygon_list(struct panfrost_batch *batch, unsigned size);
struct panfrost_bo *
panfrost_batch_get_tiler_dummy(struct panfrost_batch *batch);
unsigned minx, unsigned miny,
unsigned maxx, unsigned maxy);
-/* Scoreboarding */
-
-void
-panfrost_scoreboard_queue_compute_job(
- struct panfrost_batch *batch,
- struct panfrost_transfer job);
-
-void
-panfrost_scoreboard_queue_vertex_job(
- struct panfrost_batch *batch,
- struct panfrost_transfer vertex,
- bool requires_tiling);
-
-void
-panfrost_scoreboard_queue_tiler_job(
- struct panfrost_batch *batch,
- struct panfrost_transfer tiler);
-
-void
-panfrost_scoreboard_queue_fused_job(
- struct panfrost_batch *batch,
- struct panfrost_transfer vertex,
- struct panfrost_transfer tiler);
-void
-panfrost_scoreboard_queue_fused_job_prepend(
- struct panfrost_batch *batch,
- struct panfrost_transfer vertex,
- struct panfrost_transfer tiler);
-
-void
-panfrost_scoreboard_link_batch(struct panfrost_batch *batch);
-
bool
panfrost_batch_is_scanout(struct panfrost_batch *batch);
+mali_ptr
+panfrost_batch_get_tiler_meta(struct panfrost_batch *batch, unsigned vertex_count);
+
+mali_ptr
+panfrost_batch_reserve_framebuffer(struct panfrost_batch *batch);
+
#endif