X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fintel%2Fvulkan%2Fanv_private.h;h=5b8ac491ce570dfd84bd6021f1438110fbba54d9;hb=0709c0f6b40b1e365104b248464ffefa746b5052;hp=081606deda2278b573e151497bb870e974b7445e;hpb=5c664dff7593959d43f89d3fadcc148303134675;p=mesa.git diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h index 081606deda2..5b8ac491ce5 100644 --- a/src/intel/vulkan/anv_private.h +++ b/src/intel/vulkan/anv_private.h @@ -49,10 +49,10 @@ #include "dev/gen_device_info.h" #include "blorp/blorp.h" #include "compiler/brw_compiler.h" +#include "util/bitset.h" #include "util/macros.h" #include "util/hash_table.h" #include "util/list.h" -#include "util/set.h" #include "util/sparse_array.h" #include "util/u_atomic.h" #include "util/u_vector.h" @@ -69,6 +69,7 @@ typedef struct xcb_connection_t xcb_connection_t; typedef uint32_t xcb_visualid_t; typedef uint32_t xcb_window_t; +struct anv_batch; struct anv_buffer; struct anv_buffer_view; struct anv_image_view; @@ -91,6 +92,8 @@ struct gen_perf_config; #include "common/intel_log.h" #include "wsi_common.h" +#define NSEC_PER_SEC 1000000000ull + /* anv Virtual Memory Layout * ========================= * @@ -297,6 +300,20 @@ vk_to_isl_color(VkClearColorValue color) }; } +static inline void *anv_unpack_ptr(uintptr_t ptr, int bits, int *flags) +{ + uintptr_t mask = (1ull << bits) - 1; + *flags = ptr & mask; + return (void *) (ptr & ~mask); +} + +static inline uintptr_t anv_pack_ptr(void *ptr, int bits, int flags) +{ + uintptr_t value = (uintptr_t) ptr; + uintptr_t mask = (1ull << bits) - 1; + return value | (mask & flags); +} + #define for_each_bit(b, dword) \ for (uint32_t __dword = (dword); \ (b) = __builtin_ffs(__dword) - 1, __dword; \ @@ -608,6 +625,9 @@ struct anv_bo { */ uint32_t index; + /* Index for use with util_sparse_array_free_list */ + uint32_t free_index; + /* Last known offset. This value is provided by the kernel when we * execbuf and is used as the presumed offset for the next bunch of * relocations. @@ -644,22 +664,6 @@ struct anv_bo { bool from_host_ptr:1; }; -static inline void -anv_bo_init(struct anv_bo *bo, uint32_t gem_handle, uint64_t size) -{ - bo->gem_handle = gem_handle; - bo->refcount = 1; - bo->index = 0; - bo->offset = -1; - bo->size = size; - bo->map = NULL; - bo->flags = 0; - bo->is_external = false; - bo->is_wrapper = false; - bo->has_fixed_address = false; - bo->from_host_ptr = false; -} - static inline struct anv_bo * anv_bo_unwrap(struct anv_bo *bo) { @@ -701,7 +705,9 @@ struct anv_block_state { }; #define anv_block_pool_foreach_bo(bo, pool) \ - for (struct anv_bo *bo = (pool)->bos; bo != &(pool)->bos[(pool)->nbos]; bo++) + for (struct anv_bo **_pp_bo = (pool)->bos, *bo; \ + _pp_bo != &(pool)->bos[(pool)->nbos] && (bo = *_pp_bo, true); \ + _pp_bo++) #define ANV_MAX_BLOCK_POOL_BOS 20 @@ -716,7 +722,7 @@ struct anv_block_pool { */ struct anv_bo wrapper_bo; - struct anv_bo bos[ANV_MAX_BLOCK_POOL_BOS]; + struct anv_bo *bos[ANV_MAX_BLOCK_POOL_BOS]; struct anv_bo *bo; uint32_t nbos; @@ -896,24 +902,19 @@ struct anv_bo_pool { uint64_t bo_flags; - void *free_list[16]; + struct util_sparse_array_free_list free_list[16]; }; void anv_bo_pool_init(struct anv_bo_pool *pool, struct anv_device *device, uint64_t bo_flags); void anv_bo_pool_finish(struct anv_bo_pool *pool); -VkResult anv_bo_pool_alloc(struct anv_bo_pool *pool, struct anv_bo *bo, - uint32_t size); -void anv_bo_pool_free(struct anv_bo_pool *pool, const struct anv_bo *bo); - -struct anv_scratch_bo { - bool exists; - struct anv_bo bo; -}; +VkResult anv_bo_pool_alloc(struct anv_bo_pool *pool, uint32_t size, + struct anv_bo **bo_out); +void anv_bo_pool_free(struct anv_bo_pool *pool, struct anv_bo *bo); struct anv_scratch_pool { /* Indexed by Per-Thread Scratch Space number (the hardware value) and stage */ - struct anv_scratch_bo bos[16][MESA_SHADER_STAGES]; + struct anv_bo *bos[16][MESA_SHADER_STAGES]; }; void anv_scratch_pool_init(struct anv_device *device, @@ -1063,11 +1064,63 @@ uint32_t anv_physical_device_api_version(struct anv_physical_device *dev); bool anv_physical_device_extension_supported(struct anv_physical_device *dev, const char *name); +struct anv_queue_submit { + struct anv_cmd_buffer * cmd_buffer; + + uint32_t fence_count; + uint32_t fence_array_length; + struct drm_i915_gem_exec_fence * fences; + + uint32_t temporary_semaphore_count; + uint32_t temporary_semaphore_array_length; + struct anv_semaphore_impl * temporary_semaphores; + + /* Semaphores to be signaled with a SYNC_FD. */ + struct anv_semaphore ** sync_fd_semaphores; + uint32_t sync_fd_semaphore_count; + uint32_t sync_fd_semaphore_array_length; + + /* Allocated only with non shareable timelines. */ + struct anv_timeline ** wait_timelines; + uint32_t wait_timeline_count; + uint32_t wait_timeline_array_length; + uint64_t * wait_timeline_values; + + struct anv_timeline ** signal_timelines; + uint32_t signal_timeline_count; + uint32_t signal_timeline_array_length; + uint64_t * signal_timeline_values; + + int in_fence; + bool need_out_fence; + int out_fence; + + uint32_t fence_bo_count; + uint32_t fence_bo_array_length; + /* An array of struct anv_bo pointers with lower bit used as a flag to + * signal we will wait on that BO (see anv_(un)pack_ptr). + */ + uintptr_t * fence_bos; + + const VkAllocationCallbacks * alloc; + VkSystemAllocationScope alloc_scope; + + struct anv_bo * simple_bo; + uint32_t simple_bo_size; + + struct list_head link; +}; + struct anv_queue { VK_LOADER_DATA _loader_data; struct anv_device * device; + /* + * A list of struct anv_queue_submit to be submitted to i915. + */ + struct list_head queued_submits; + VkDeviceQueueCreateFlags flags; }; @@ -1175,9 +1228,9 @@ struct anv_device { struct anv_state_pool binding_table_pool; struct anv_state_pool surface_state_pool; - struct anv_bo workaround_bo; - struct anv_bo trivial_batch_bo; - struct anv_bo hiz_clear_bo; + struct anv_bo * workaround_bo; + struct anv_bo * trivial_batch_bo; + struct anv_bo * hiz_clear_bo; struct anv_pipeline_cache default_pipeline_cache; struct blorp_context blorp; @@ -1190,12 +1243,9 @@ struct anv_device { struct anv_scratch_pool scratch_pool; - uint32_t default_mocs; - uint32_t external_mocs; - pthread_mutex_t mutex; pthread_cond_t queue_submit; - bool _lost; + int _lost; struct gen_batch_decode_ctx decoder_ctx; /* @@ -1237,30 +1287,34 @@ static inline uint32_t anv_mocs_for_bo(const struct anv_device *device, const struct anv_bo *bo) { if (bo->is_external) - return device->external_mocs; + return device->isl_dev.mocs.external; else - return device->default_mocs; + return device->isl_dev.mocs.internal; } void anv_device_init_blorp(struct anv_device *device); void anv_device_finish_blorp(struct anv_device *device); +void _anv_device_set_all_queue_lost(struct anv_device *device); VkResult _anv_device_set_lost(struct anv_device *device, const char *file, int line, const char *msg, ...) anv_printflike(4, 5); +VkResult _anv_queue_set_lost(struct anv_queue *queue, + const char *file, int line, + const char *msg, ...) + anv_printflike(4, 5); #define anv_device_set_lost(dev, ...) \ _anv_device_set_lost(dev, __FILE__, __LINE__, __VA_ARGS__) +#define anv_queue_set_lost(queue, ...) \ + _anv_queue_set_lost(queue, __FILE__, __LINE__, __VA_ARGS__) static inline bool anv_device_is_lost(struct anv_device *device) { - return unlikely(device->_lost); + return unlikely(p_atomic_read(&device->_lost)); } -VkResult anv_device_execbuf(struct anv_device *device, - struct drm_i915_gem_execbuffer2 *execbuf, - struct anv_bo **execbuf_bos); VkResult anv_device_query_status(struct anv_device *device); @@ -1324,6 +1378,16 @@ VkResult anv_device_bo_busy(struct anv_device *device, struct anv_bo *bo); VkResult anv_device_wait(struct anv_device *device, struct anv_bo *bo, int64_t timeout); +VkResult anv_queue_init(struct anv_device *device, struct anv_queue *queue); +void anv_queue_finish(struct anv_queue *queue); + +VkResult anv_queue_execbuf_locked(struct anv_queue *queue, struct anv_queue_submit *submit); +VkResult anv_queue_submit_simple_batch(struct anv_queue *queue, + struct anv_batch *batch); + +uint64_t anv_gettime_ns(void); +uint64_t anv_get_absolute_timeout(uint64_t timeout); + void* anv_gem_mmap(struct anv_device *device, uint32_t gem_handle, uint64_t offset, uint64_t size, uint32_t flags); void anv_gem_munmap(void *p, uint64_t size); @@ -1374,14 +1438,13 @@ int anv_gem_syncobj_wait(struct anv_device *device, bool anv_vma_alloc(struct anv_device *device, struct anv_bo *bo); void anv_vma_free(struct anv_device *device, struct anv_bo *bo); -VkResult anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size); - struct anv_reloc_list { uint32_t num_relocs; uint32_t array_length; struct drm_i915_gem_relocation_entry * relocs; struct anv_bo ** reloc_bos; - struct set * deps; + uint32_t dep_words; + BITSET_WORD * deps; }; VkResult anv_reloc_list_init(struct anv_reloc_list *list, @@ -1398,7 +1461,7 @@ struct anv_batch_bo { /* Link in the anv_cmd_buffer.owned_batch_bos list */ struct list_head link; - struct anv_bo bo; + struct anv_bo * bo; /* Bytes actually consumed in this batch BO */ uint32_t length; @@ -1435,8 +1498,6 @@ void *anv_batch_emit_dwords(struct anv_batch *batch, int num_dwords); void anv_batch_emit_batch(struct anv_batch *batch, struct anv_batch *other); uint64_t anv_batch_emit_reloc(struct anv_batch *batch, void *location, struct anv_bo *bo, uint32_t offset); -VkResult anv_device_submit_simple_batch(struct anv_device *device, - struct anv_batch *batch); static inline VkResult anv_batch_set_error(struct anv_batch *batch, VkResult error) @@ -1570,56 +1631,6 @@ _anv_combine_address(struct anv_batch *batch, void *location, _dst = NULL; \ })) -/* MEMORY_OBJECT_CONTROL_STATE: - * .GraphicsDataTypeGFDT = 0, - * .LLCCacheabilityControlLLCCC = 0, - * .L3CacheabilityControlL3CC = 1, - */ -#define GEN7_MOCS 1 - -/* MEMORY_OBJECT_CONTROL_STATE: - * .LLCeLLCCacheabilityControlLLCCC = 0, - * .L3CacheabilityControlL3CC = 1, - */ -#define GEN75_MOCS 1 - -/* MEMORY_OBJECT_CONTROL_STATE: - * .MemoryTypeLLCeLLCCacheabilityControl = WB, - * .TargetCache = L3DefertoPATforLLCeLLCselection, - * .AgeforQUADLRU = 0 - */ -#define GEN8_MOCS 0x78 - -/* MEMORY_OBJECT_CONTROL_STATE: - * .MemoryTypeLLCeLLCCacheabilityControl = UCwithFenceifcoherentcycle, - * .TargetCache = L3DefertoPATforLLCeLLCselection, - * .AgeforQUADLRU = 0 - */ -#define GEN8_EXTERNAL_MOCS 0x18 - -/* Skylake: MOCS is now an index into an array of 62 different caching - * configurations programmed by the kernel. - */ - -/* TC=LLC/eLLC, LeCC=WB, LRUM=3, L3CC=WB */ -#define GEN9_MOCS (2 << 1) - -/* TC=LLC/eLLC, LeCC=WB, LRUM=3, L3CC=WB */ -#define GEN9_EXTERNAL_MOCS (1 << 1) - -/* Cannonlake MOCS defines are duplicates of Skylake MOCS defines. */ -#define GEN10_MOCS GEN9_MOCS -#define GEN10_EXTERNAL_MOCS GEN9_EXTERNAL_MOCS - -/* Ice Lake MOCS defines are duplicates of Skylake MOCS defines. */ -#define GEN11_MOCS GEN9_MOCS -#define GEN11_EXTERNAL_MOCS GEN9_EXTERNAL_MOCS - -/* TigerLake MOCS */ -#define GEN12_MOCS GEN9_MOCS -/* TC=1/LLC Only, LeCC=1/Uncacheable, LRUM=0, L3CC=1/Uncacheable */ -#define GEN12_EXTERNAL_MOCS (3 << 1) - struct anv_device_memory { struct list_head link; @@ -1887,7 +1898,7 @@ struct anv_descriptor_pool { uint32_t next; uint32_t free_list; - struct anv_bo bo; + struct anv_bo *bo; struct util_vma_heap bo_heap; struct anv_state_stream surface_state_stream; @@ -2001,25 +2012,32 @@ anv_descriptor_set_destroy(struct anv_device *device, #define ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS UINT8_MAX struct anv_pipeline_binding { - /* The descriptor set this surface corresponds to. The special value of - * ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS indicates that the offset refers - * to a color attachment and not a regular descriptor. + /** Index in the descriptor set + * + * This is a flattened index; the descriptor set layout is already taken + * into account. */ - uint8_t set; + uint32_t index; - /* Binding in the descriptor set */ - uint32_t binding; + /** The descriptor set this surface corresponds to. + * + * The special ANV_DESCRIPTOR_SET_* values above indicates that this + * binding is not a normal descriptor set but something else. + */ + uint8_t set; - /* Index in the binding */ - uint32_t index; + union { + /** Plane in the binding index for images */ + uint8_t plane; - /* Plane in the binding index */ - uint8_t plane; + /** Input attachment index (relative to the subpass) */ + uint8_t input_attachment_index; - /* Input attachment index (relative to the subpass) */ - uint8_t input_attachment_index; + /** Dynamic offset index (for dynamic UBOs and SSBOs) */ + uint8_t dynamic_offset_index; + }; - /* For a storage image, whether it is write-only */ + /** For a storage image, whether it is write-only */ bool write_only; }; @@ -2429,6 +2447,7 @@ struct anv_attachment_state { struct anv_surface_state input; VkImageLayout current_layout; + VkImageLayout current_stencil_layout; VkImageAspectFlags pending_clear_aspects; VkImageAspectFlags pending_load_aspects; bool fast_clear; @@ -2456,7 +2475,6 @@ struct anv_attachment_state { */ struct anv_cmd_pipeline_state { struct anv_pipeline *pipeline; - struct anv_pipeline_layout *layout; struct anv_descriptor_set *descriptors[MAX_SETS]; uint32_t dynamic_offsets[MAX_DYNAMIC_BUFFERS]; @@ -2643,11 +2661,13 @@ void anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer *cmd_buffer); void anv_cmd_buffer_add_secondary(struct anv_cmd_buffer *primary, struct anv_cmd_buffer *secondary); void anv_cmd_buffer_prepare_execbuf(struct anv_cmd_buffer *cmd_buffer); -VkResult anv_cmd_buffer_execbuf(struct anv_device *device, +VkResult anv_cmd_buffer_execbuf(struct anv_queue *queue, struct anv_cmd_buffer *cmd_buffer, const VkSemaphore *in_semaphores, + const uint64_t *in_wait_values, uint32_t num_in_semaphores, const VkSemaphore *out_semaphores, + const uint64_t *out_signal_values, uint32_t num_out_semaphores, VkFence fence); @@ -2737,7 +2757,7 @@ struct anv_fence_impl { * will say it's idle in this case. */ struct { - struct anv_bo bo; + struct anv_bo *bo; enum anv_bo_fence_state state; } bo; @@ -2777,6 +2797,32 @@ enum anv_semaphore_type { ANV_SEMAPHORE_TYPE_BO, ANV_SEMAPHORE_TYPE_SYNC_FILE, ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ, + ANV_SEMAPHORE_TYPE_TIMELINE, +}; + +struct anv_timeline_point { + struct list_head link; + + uint64_t serial; + + /* Number of waiter on this point, when > 0 the point should not be garbage + * collected. + */ + int waiting; + + /* BO used for synchronization. */ + struct anv_bo *bo; +}; + +struct anv_timeline { + pthread_mutex_t mutex; + pthread_cond_t cond; + + uint64_t highest_past; + uint64_t highest_pending; + + struct list_head points; + struct list_head free_points; }; struct anv_semaphore_impl { @@ -2801,10 +2847,18 @@ struct anv_semaphore_impl { * import so we don't need to bother with a userspace cache. */ uint32_t syncobj; + + /* Non shareable timeline semaphore + * + * Used when kernel don't have support for timeline semaphores. + */ + struct anv_timeline timeline; }; }; struct anv_semaphore { + uint32_t refcount; + /* Permanent semaphore state. Every semaphore has some form of permanent * state (type != ANV_SEMAPHORE_TYPE_NONE). This may be a BO to fence on * (for cross-process semaphores0 or it could just be a dummy for use @@ -3761,6 +3815,9 @@ struct anv_subpass_attachment { VkImageUsageFlagBits usage; uint32_t attachment; VkImageLayout layout; + + /* Used only with attachment containing stencil data. */ + VkImageLayout stencil_layout; }; struct anv_subpass { @@ -3811,6 +3868,9 @@ struct anv_render_pass_attachment { VkImageLayout final_layout; VkImageLayout first_subpass_layout; + VkImageLayout stencil_initial_layout; + VkImageLayout stencil_final_layout; + /* The subpass id in which the attachment will be used last. */ uint32_t last_subpass_idx; }; @@ -3833,7 +3893,7 @@ struct anv_query_pool { uint32_t stride; /** Number of slots in this query pool */ uint32_t slots; - struct anv_bo bo; + struct anv_bo * bo; }; int anv_get_instance_entrypoint_index(const char *name);