panfrost: Fix attribute buffer underallocation
[mesa.git] / src / gallium / drivers / panfrost / pan_job.h
index c6ae2a4eb9f4039ed149d69fb6b64f2baadf4dff..eb3cc58d573d8a257cde7050f54ae4dac8611d7a 100644 (file)
 
 #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"
 
-/* Used as a hash table key */
-
-struct panfrost_job_key {
-        struct pipe_surface *cbufs[4];
-        struct pipe_surface *zsbuf;
+/* 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
+ * one as want will certainly want to wait upon the fence after the batch has
+ * been submitted (which is when panfrost_batch objects are freed).
+ */
+struct panfrost_batch_fence {
+        /* Refcounting object for the fence. */
+        struct pipe_reference reference;
+
+        /* Batch that created this fence object. Will become NULL at batch
+         * submission time. This field is mainly here to know whether the
+         * batch has been flushed or not.
+         */
+        struct panfrost_batch *batch;
 };
 
 #define PAN_REQ_MSAA            (1 << 0)
 #define PAN_REQ_DEPTH_WRITE     (1 << 1)
 
-/* A panfrost_job corresponds to a bound FBO we're rendering to,
+/* A panfrost_batch corresponds to a bound FBO we're rendering to,
  * collecting over multiple draws. */
 
-struct panfrost_job {
+struct panfrost_batch {
         struct panfrost_context *ctx;
-        struct panfrost_job_key key;
+        struct pipe_framebuffer_state key;
 
         /* 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
@@ -60,6 +73,12 @@ struct panfrost_job {
         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;
@@ -70,128 +89,125 @@ struct panfrost_job {
         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
-         * SET_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;
+        /* BOs referenced not in the pool */
+        struct hash_table *bos;
 
-        /* (And the GPU versions; TODO maybe combine) */
-        struct util_dynarray gpu_headers;
+        /* Pool owned by this batch (released when the batch is released) used for temporary descriptors */
+        struct pan_pool pool;
 
-        /* The last job in the primary batch */
-        struct panfrost_transfer last_job;
+        /* 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;
 
-        /* The first/last tiler job */
-        struct panfrost_transfer first_tiler;
-        struct panfrost_transfer last_tiler;
+        /* Job scoreboarding state */
+        struct pan_scoreboard scoreboard;
 
-        /* The first vertex job used as the input to a tiler job */
-        struct panfrost_transfer first_vertex_for_tiler;
+        /* Polygon list bound to the batch, or NULL if none bound yet */
+        struct panfrost_bo *polygon_list;
 
-        /* The first job. Notice we've created a linked list */
-        struct panfrost_transfer first_job;
+        /* Scratchpad BO bound to the batch, or NULL if none bound yet */
+        struct panfrost_bo *scratchpad;
 
-        /* The number of jobs in the primary batch, essentially */
-        unsigned job_index;
+        /* Shared memory BO bound to the batch, or NULL if none bound yet */
+        struct panfrost_bo *shared_memory;
 
-        /* BOs referenced -- will be used for flushing logic */
-        struct set *bos;
+        /* Tiler heap BO bound to the batch, or NULL if none bound yet */
+        struct panfrost_bo *tiler_heap;
 
-        /* Indices of transient BOs referenced */
-        struct util_dynarray transient_indices;
+        /* Dummy tiler BO bound to the batch, or NULL if none bound yet */
+        struct panfrost_bo *tiler_dummy;
 
-        /* Within the topmost transient BO, how much has been used? */
-        unsigned transient_offset;
+        /* 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;
+
+        /* Batch dependencies */
+        struct util_dynarray dependencies;
 };
 
 /* Functions for managing the above */
 
-struct panfrost_job *
-panfrost_create_job(struct panfrost_context *ctx);
+void
+panfrost_batch_fence_unreference(struct panfrost_batch_fence *fence);
 
 void
-panfrost_free_job(struct panfrost_context *ctx, struct panfrost_job *job);
+panfrost_batch_fence_reference(struct panfrost_batch_fence *batch);
 
-struct panfrost_job *
-panfrost_get_job(struct panfrost_context *ctx,
-                 struct pipe_surface **cbufs, struct pipe_surface *zsbuf);
+struct panfrost_batch *
+panfrost_get_batch_for_fbo(struct panfrost_context *ctx);
 
-struct panfrost_job *
-panfrost_get_job_for_fbo(struct panfrost_context *ctx);
+struct panfrost_batch *
+panfrost_get_fresh_batch_for_fbo(struct panfrost_context *ctx);
 
 void
-panfrost_job_init(struct panfrost_context *ctx);
+panfrost_batch_init(struct panfrost_context *ctx);
 
 void
-panfrost_job_add_bo(struct panfrost_job *job, struct panfrost_bo *bo);
+panfrost_batch_add_bo(struct panfrost_batch *batch, struct panfrost_bo *bo,
+                      uint32_t flags);
 
-void
-panfrost_flush_jobs_writing_resource(struct panfrost_context *panfrost,
-                                     struct pipe_resource *prsc);
+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_jobs_reading_resource(struct panfrost_context *panfrost,
-                                     struct pipe_resource *prsc);
+panfrost_flush_all_batches(struct panfrost_context *ctx, uint32_t out_sync);
 
-void
-panfrost_job_submit(struct panfrost_context *ctx, struct panfrost_job *job);
+bool
+panfrost_pending_batches_access_bo(struct panfrost_context *ctx,
+                                   const struct panfrost_bo *bo);
 
 void
-panfrost_job_set_requirements(struct panfrost_context *ctx,
-                              struct panfrost_job *job);
+panfrost_flush_batches_accessing_bo(struct panfrost_context *ctx,
+                                    struct panfrost_bo *bo, bool flush_readers);
 
 void
-panfrost_job_clear(struct panfrost_context *ctx,
-                   struct panfrost_job *job,
-                   unsigned buffers,
-                   const union pipe_color_union *color,
-                   double depth, unsigned stencil);
+panfrost_batch_set_requirements(struct panfrost_batch *batch);
 
 void
-panfrost_job_union_scissor(struct panfrost_job *job,
-                           unsigned minx, unsigned miny,
-                           unsigned maxx, unsigned maxy);
+panfrost_batch_adjust_stack_size(struct panfrost_batch *batch);
 
-/* Scoreboarding */
+struct panfrost_bo *
+panfrost_batch_get_scratchpad(struct panfrost_batch *batch, unsigned size, unsigned thread_tls_alloc, unsigned core_count);
 
-void
-panfrost_scoreboard_queue_compute_job(
-        struct panfrost_job *batch,
-        struct panfrost_transfer job);
+struct panfrost_bo *
+panfrost_batch_get_shared_memory(struct panfrost_batch *batch, unsigned size, unsigned workgroup_count);
 
-void
-panfrost_scoreboard_queue_vertex_job(
-        struct panfrost_job *batch,
-        struct panfrost_transfer vertex,
-        bool requires_tiling);
+mali_ptr
+panfrost_batch_get_polygon_list(struct panfrost_batch *batch, unsigned size);
 
-void
-panfrost_scoreboard_queue_tiler_job(
-        struct panfrost_job *batch,
-        struct panfrost_transfer tiler);
+struct panfrost_bo *
+panfrost_batch_get_tiler_dummy(struct panfrost_batch *batch);
 
 void
-panfrost_scoreboard_queue_fused_job(
-        struct panfrost_job *batch,
-        struct panfrost_transfer vertex,
-        struct panfrost_transfer tiler);
+panfrost_batch_clear(struct panfrost_batch *batch,
+                     unsigned buffers,
+                     const union pipe_color_union *color,
+                     double depth, unsigned stencil);
+
 void
-panfrost_scoreboard_queue_fused_job_prepend(
-        struct panfrost_job *batch,
-        struct panfrost_transfer vertex,
-        struct panfrost_transfer tiler);
+panfrost_batch_union_scissor(struct panfrost_batch *batch,
+                             unsigned minx, unsigned miny,
+                             unsigned maxx, unsigned maxy);
 
 void
-panfrost_scoreboard_link_batch(struct panfrost_job *batch);
+panfrost_batch_intersection_scissor(struct panfrost_batch *batch,
+                                    unsigned minx, unsigned miny,
+                                    unsigned maxx, unsigned maxy);
+
+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