llvmpipe: add grid launch

[mesa.git] / src / gallium / drivers / panfrost / pan_screen.h

diff --git a/src/gallium/drivers/panfrost/pan_screen.h b/src/gallium/drivers/panfrost/pan_screen.h
index 3a0e544c1a639dc0cb4325cbdc6074be586a7680..d8ea55912b4e39421268999fd2a867a1767d8a46 100644 (file)
--- a/src/gallium/drivers/panfrost/pan_screen.h
+++ b/src/gallium/drivers/panfrost/pan_screen.h
@@ -29,6 +29,7 @@
 #ifndef PAN_SCREEN_H
 #define PAN_SCREEN_H
 
+#include <xf86drm.h>
 #include "pipe/p_screen.h"
 #include "pipe/p_defines.h"
 #include "renderonly/renderonly.h"
@@ -42,6 +43,9 @@ struct panfrost_context;
 struct panfrost_resource;
 struct panfrost_screen;
 
+/* Driver limits */
+#define PAN_MAX_CONST_BUFFERS 16
+
 /* Flags for allocated memory */
 
 /* This memory region is executable */
@@ -78,13 +82,30 @@ struct panfrost_screen;
 
 #define MAX_TRANSIENT_SLABS (1024*1024 / TRANSIENT_SLAB_PAGES)
 
+/* How many power-of-two levels in the BO cache do we want? 2^12
+ * minimum chosen as it is the page size that all allocations are
+ * rounded to */
+
+#define MIN_BO_CACHE_BUCKET (12) /* 2^12 = 4KB */
+#define MAX_BO_CACHE_BUCKET (22) /* 2^22 = 4MB */
+
+/* Fencepost problem, hence the off-by-one */
+#define NR_BO_CACHE_BUCKETS (MAX_BO_CACHE_BUCKET - MIN_BO_CACHE_BUCKET + 1)
+
 struct panfrost_screen {
         struct pipe_screen base;
         int fd;
+
+        /* Properties of the GPU in use */
         unsigned gpu_id;
+        bool require_sfbd;
+
+        drmVersionPtr kernel_version;
 
         struct renderonly *ro;
 
+        pthread_mutex_t transient_lock;
+
         /* Transient memory management is based on borrowing fixed-size slabs
          * off the screen (loaning them out to the batch). Dynamic array
          * container of panfrost_bo */
@@ -94,11 +115,13 @@ struct panfrost_screen {
         /* Set of free transient BOs */
         BITSET_DECLARE(free_transient, MAX_TRANSIENT_SLABS);
 
-        /* While we're busy building up the job for frame N, the GPU is
-         * still busy executing frame N-1. So hold a reference to
-         * yesterjob */
-        int last_fragment_flushed;
-        struct panfrost_job *last_job;
+        pthread_mutex_t bo_cache_lock;
+
+        /* The BO cache is a set of buckets with power-of-two sizes ranging
+         * from 2^12 (4096, the page size) to 2^(12 + MAX_BO_CACHE_BUCKETS).
+         * Each bucket is a linked list of free panfrost_bo objects. */
+
+        struct list_head bo_cache[NR_BO_CACHE_BUCKETS];
 };
 
 static inline struct panfrost_screen *
@@ -134,14 +157,13 @@ panfrost_drm_create_bo(struct panfrost_screen *screen, size_t size,
 void
 panfrost_drm_mmap_bo(struct panfrost_screen *screen, struct panfrost_bo *bo);
 void
-panfrost_drm_release_bo(struct panfrost_screen *screen, struct panfrost_bo *bo);
+panfrost_drm_release_bo(struct panfrost_screen *screen, struct panfrost_bo *bo, bool cacheable);
 struct panfrost_bo *
 panfrost_drm_import_bo(struct panfrost_screen *screen, int fd);
 int
 panfrost_drm_export_bo(struct panfrost_screen *screen, const struct panfrost_bo *bo);
 int
-panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws,
-                              bool is_scanout);
+panfrost_drm_submit_vs_fs_job(struct panfrost_context *ctx, bool has_draws);
 void
 panfrost_drm_force_flush_fragment(struct panfrost_context *ctx,
                                   struct pipe_fence_handle **fence);
@@ -158,5 +180,18 @@ panfrost_drm_fence_finish(struct pipe_screen *pscreen,
                           struct pipe_context *ctx,
                           struct pipe_fence_handle *fence,
                           uint64_t timeout);
+struct panfrost_bo *
+panfrost_bo_cache_fetch(
+                struct panfrost_screen *screen,
+                size_t size, uint32_t flags);
+
+bool
+panfrost_bo_cache_put(
+                struct panfrost_screen *screen,
+                struct panfrost_bo *bo);
+
+void
+panfrost_bo_cache_evict_all(
+                struct panfrost_screen *screen);
 
 #endif /* PAN_SCREEN_H */