/**************************************************************************
*
* Copyright 2018-2019 Alyssa Rosenzweig
- * Copyright 2018-2019 Collabora
+ * Copyright 2018-2019 Collabora, Ltd.
* All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
#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"
+#include "util/u_dynarray.h"
+#include "util/bitset.h"
#include <panfrost-misc.h>
#include "pan_allocate.h"
-#include "pan_trace.h"
+struct panfrost_batch;
struct panfrost_context;
struct panfrost_resource;
struct panfrost_screen;
-//#define DUMP_PERFORMANCE_COUNTERS
+/* Driver limits */
+#define PAN_MAX_CONST_BUFFERS 16
/* Flags for allocated memory */
-#define PAN_ALLOCATE_EXECUTE (1 << 0)
-#define PAN_ALLOCATE_GROWABLE (1 << 1)
-#define PAN_ALLOCATE_INVISIBLE (1 << 2)
-#define PAN_ALLOCATE_COHERENT_LOCAL (1 << 3)
-
-struct panfrost_driver {
- struct panfrost_bo * (*import_bo) (struct panfrost_screen *screen, struct winsys_handle *whandle);
-
- int (*submit_vs_fs_job) (struct panfrost_context *ctx, bool has_draws, bool is_scanout);
- void (*force_flush_fragment) (struct panfrost_context *ctx);
- void (*allocate_slab) (struct panfrost_screen *screen,
- struct panfrost_memory *mem,
- size_t pages,
- bool same_va,
- int extra_flags,
- int commit_count,
- int extent);
- void (*free_slab) (struct panfrost_screen *screen,
- struct panfrost_memory *mem);
- void (*free_imported_bo) (struct panfrost_screen *screen,
- struct panfrost_bo *bo);
- void (*enable_counters) (struct panfrost_screen *screen);
-};
+
+/* This memory region is executable */
+#define PAN_BO_EXECUTE (1 << 0)
+
+/* This memory region should be lazily allocated and grow-on-page-fault. Must
+ * be used in conjunction with INVISIBLE */
+#define PAN_BO_GROWABLE (1 << 1)
+
+/* This memory region should not be mapped to the CPU */
+#define PAN_BO_INVISIBLE (1 << 2)
+
+/* This memory region will be used for varyings and needs to have the cache
+ * bits twiddled accordingly */
+#define PAN_BO_COHERENT_LOCAL (1 << 3)
+
+/* This region may not be used immediately and will not mmap on allocate
+ * (semantically distinct from INVISIBLE, which cannot never be mmaped) */
+#define PAN_BO_DELAY_MMAP (1 << 4)
+
+/* Transient slab size. This is a balance between fragmentation against cache
+ * locality and ease of bookkeeping */
+
+#define TRANSIENT_SLAB_PAGES (32) /* 128kb */
+#define TRANSIENT_SLAB_SIZE (4096 * TRANSIENT_SLAB_PAGES)
+
+/* Maximum number of transient slabs so we don't need dynamic arrays. Most
+ * interesting Mali boards are 4GB RAM max, so if the entire RAM was filled
+ * with transient slabs, you could never exceed (4GB / TRANSIENT_SLAB_SIZE)
+ * allocations anyway. By capping, we can use a fixed-size bitset for tracking
+ * free slabs, eliminating quite a bit of complexity. We can pack the free
+ * state of 8 slabs into a single byte, so for 128kb transient slabs the bitset
+ * occupies a cheap 4kb of memory */
+
+#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;
- struct panfrost_driver *driver;
- struct panfrost_memory perf_counters;
+ pthread_mutex_t bo_cache_lock;
- /* Memory management is based on subdividing slabs with AMD's allocator */
- struct pb_slabs slabs;
-
- /* TODO: Where? */
- struct panfrost_resource *display_target;
+ /* 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. */
- int last_fragment_id;
- int last_fragment_flushed;
+ struct list_head bo_cache[NR_BO_CACHE_BUCKETS];
};
static inline struct panfrost_screen *
-panfrost_screen( struct pipe_screen *pipe )
+pan_screen(struct pipe_screen *p)
{
- return (struct panfrost_screen *)pipe;
+ return (struct panfrost_screen *)p;
}
+struct panfrost_fence *
+panfrost_fence_create(struct panfrost_context *ctx);
+
+struct panfrost_bo *
+panfrost_bo_create(struct panfrost_screen *screen, size_t size,
+ uint32_t flags);
+void
+panfrost_bo_mmap(struct panfrost_screen *screen, struct panfrost_bo *bo);
+void
+panfrost_bo_release(struct panfrost_screen *screen, struct panfrost_bo *bo,
+ bool cacheable);
+struct panfrost_bo *
+panfrost_bo_import(struct panfrost_screen *screen, int fd);
+int
+panfrost_bo_export(struct panfrost_screen *screen, const struct panfrost_bo *bo);
+
+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 */
projects / mesa.git / blobdiff