anv: pCreateInfo->pApplicationInfo parameter to vkCreateInstance may be NULL
[mesa.git] / src / vulkan / anv_private.h
index 3acf9796a5e0bb23c9887fab692aca7ed0dbbbc3..e2ae011920090d595bbd868924eb83bd627e7394 100644 (file)
@@ -57,6 +57,7 @@ typedef uint32_t xcb_window_t;
 #define VK_PROTOTYPES
 #include <vulkan/vulkan.h>
 #include <vulkan/vulkan_intel.h>
+#include <vulkan/vk_icd.h>
 
 #include "anv_entrypoints.h"
 #include "anv_gen_macros.h"
@@ -67,12 +68,15 @@ typedef uint32_t xcb_window_t;
 extern "C" {
 #endif
 
-#define ICD_LOADER_MAGIC   0x01CDC0DE
-
-typedef union _VK_LOADER_DATA {
-  uintptr_t loaderMagic;
-  void *loaderData;
-} VK_LOADER_DATA;
+#define MAX_VBS         32
+#define MAX_SETS         8
+#define MAX_RTS          8
+#define MAX_VIEWPORTS   16
+#define MAX_SCISSORS    16
+#define MAX_PUSH_CONSTANTS_SIZE 128
+#define MAX_DYNAMIC_BUFFERS 16
+#define MAX_IMAGES 8
+#define MAX_SAMPLES_LOG2 4 /* SKL supports 16 samples */
 
 #define anv_noreturn __attribute__((__noreturn__))
 #define anv_printflike(a, b) __attribute__((__format__(__printf__, a, b)))
@@ -152,6 +156,8 @@ anv_clear_mask(uint32_t *inout_mask, uint32_t clear_mask)
    memcpy((dest), (src), (count) * sizeof(*(src))); \
 })
 
+#define zero(x) (memset(&(x), 0, sizeof(x)))
+
 /* Define no kernel as 1, since that's an illegal offset for a kernel */
 #define NO_KERNEL 1
 
@@ -412,22 +418,25 @@ struct anv_state_stream {
 #define CACHELINE_SIZE 64
 #define CACHELINE_MASK 63
 
-static void inline
-anv_state_clflush(struct anv_state state)
+static inline void
+anv_clflush_range(void *start, size_t size)
 {
-   /* state.map may not be cacheline aligned, so round down the start pointer
-    * to a cacheline boundary so we flush all pages that contain the state.
-    */
-   void *end = state.map + state.alloc_size;
-   void *p = (void *) (((uintptr_t) state.map) & ~CACHELINE_MASK);
+   void *p = (void *) (((uintptr_t) start) & ~CACHELINE_MASK);
+   void *end = start + size;
 
-   __builtin_ia32_sfence();
+   __builtin_ia32_mfence();
    while (p < end) {
       __builtin_ia32_clflush(p);
       p += CACHELINE_SIZE;
    }
 }
 
+static void inline
+anv_state_clflush(struct anv_state state)
+{
+   anv_clflush_range(state.map, state.alloc_size);
+}
+
 void anv_block_pool_init(struct anv_block_pool *pool,
                          struct anv_device *device, uint32_t block_size);
 void anv_block_pool_finish(struct anv_block_pool *pool);
@@ -535,6 +544,10 @@ struct anv_physical_device {
     struct isl_device                           isl_dev;
 };
 
+struct anv_wsi_interaface;
+
+#define VK_ICD_WSI_PLATFORM_MAX 5
+
 struct anv_instance {
     VK_LOADER_DATA                              _loader_data;
 
@@ -544,19 +557,33 @@ struct anv_instance {
     int                                         physicalDeviceCount;
     struct anv_physical_device                  physicalDevice;
 
-    void *                                      wayland_wsi;
+    struct anv_wsi_interface *                  wsi[VK_ICD_WSI_PLATFORM_MAX];
 };
 
 VkResult anv_init_wsi(struct anv_instance *instance);
 void anv_finish_wsi(struct anv_instance *instance);
 
 struct anv_meta_state {
+   VkAllocationCallbacks alloc;
+
+   /**
+    * Use array element `i` for images with `2^i` samples.
+    */
    struct {
-      struct anv_pipeline *color_pipeline;
+      /**
+       * Pipeline N is used to clear color attachment N of the current
+       * subpass.
+       *
+       * HACK: We use one pipeline per color attachment to work around the
+       * compiler's inability to dynamically set the render target index of
+       * the render target write message.
+       */
+      struct anv_pipeline *color_pipelines[MAX_RTS];
+
       struct anv_pipeline *depth_only_pipeline;
       struct anv_pipeline *stencil_only_pipeline;
       struct anv_pipeline *depthstencil_pipeline;
-   } clear;
+   } clear[1 + MAX_SAMPLES_LOG2];
 
    struct {
       VkRenderPass render_pass;
@@ -573,6 +600,15 @@ struct anv_meta_state {
       VkPipelineLayout                          pipeline_layout;
       VkDescriptorSetLayout                     ds_layout;
    } blit;
+
+   struct {
+      /** Pipeline [i] resolves an image with 2^(i+1) samples.  */
+      VkPipeline                                pipelines[MAX_SAMPLES_LOG2];
+
+      VkRenderPass                              pass;
+      VkPipelineLayout                          pipeline_layout;
+      VkDescriptorSetLayout                     ds_layout;
+   } resolve;
 };
 
 struct anv_queue {
@@ -587,11 +623,24 @@ struct anv_pipeline_cache {
    struct anv_device *                          device;
    struct anv_state_stream                      program_stream;
    pthread_mutex_t                              mutex;
+
+   uint32_t                                     total_size;
+   uint32_t                                     table_size;
+   uint32_t                                     kernel_count;
+   uint32_t                                    *table;
 };
 
 void anv_pipeline_cache_init(struct anv_pipeline_cache *cache,
                              struct anv_device *device);
 void anv_pipeline_cache_finish(struct anv_pipeline_cache *cache);
+uint32_t anv_pipeline_cache_search(struct anv_pipeline_cache *cache,
+                                   const unsigned char *sha1, void *prog_data);
+uint32_t anv_pipeline_cache_upload_kernel(struct anv_pipeline_cache *cache,
+                                          const unsigned char *sha1,
+                                          const void *kernel,
+                                          size_t kernel_size,
+                                          const void *prog_data,
+                                          size_t prog_data_size);
 
 struct anv_device {
     VK_LOADER_DATA                              _loader_data;
@@ -629,6 +678,14 @@ struct anv_device {
     pthread_mutex_t                             mutex;
 };
 
+VkResult gen7_init_device_state(struct anv_device *device);
+VkResult gen75_init_device_state(struct anv_device *device);
+VkResult gen8_init_device_state(struct anv_device *device);
+VkResult gen9_init_device_state(struct anv_device *device);
+
+void anv_device_get_cache_uuid(void *uuid);
+
+
 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);
@@ -643,6 +700,7 @@ int anv_gem_set_tiling(struct anv_device *device, uint32_t gem_handle,
 int anv_gem_create_context(struct anv_device *device);
 int anv_gem_destroy_context(struct anv_device *device, int context);
 int anv_gem_get_param(int fd, uint32_t param);
+bool anv_gem_get_bit6_swizzle(int fd, uint32_t tiling);
 int anv_gem_get_aperture(int fd, uint64_t *size);
 int anv_gem_handle_to_fd(struct anv_device *device, uint32_t gem_handle);
 uint32_t anv_gem_fd_to_handle(struct anv_device *device, int fd);
@@ -704,6 +762,8 @@ 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);
 
 struct anv_address {
    struct anv_bo *bo;
@@ -751,7 +811,7 @@ __gen_combine_address(struct anv_batch *batch, void *location,
       void *__dst = anv_batch_emit_dwords(batch, n);    \
       struct cmd __template = {                         \
          __anv_cmd_header(cmd),                         \
-        .DwordLength = n - __anv_cmd_length_bias(cmd),  \
+        .DWordLength = n - __anv_cmd_length_bias(cmd),  \
          __VA_ARGS__                                    \
       };                                                \
       __anv_cmd_pack(cmd)(batch, __dst, &__template);   \
@@ -884,9 +944,7 @@ struct anv_descriptor {
 
    union {
       struct {
-         union {
-            struct anv_image_view *image_view;
-         };
+         struct anv_image_view *image_view;
          struct anv_sampler *sampler;
       };
 
@@ -896,6 +954,7 @@ struct anv_descriptor {
 
 struct anv_descriptor_set {
    const struct anv_descriptor_set_layout *layout;
+   uint32_t buffer_count;
    struct anv_buffer_view *buffer_views;
    struct anv_descriptor descriptors[0];
 };
@@ -909,15 +968,6 @@ void
 anv_descriptor_set_destroy(struct anv_device *device,
                            struct anv_descriptor_set *set);
 
-#define MAX_VBS         32
-#define MAX_SETS         8
-#define MAX_RTS          8
-#define MAX_VIEWPORTS   16
-#define MAX_SCISSORS    16
-#define MAX_PUSH_CONSTANTS_SIZE 128
-#define MAX_DYNAMIC_BUFFERS 16
-#define MAX_IMAGES 8
-
 struct anv_pipeline_binding {
    /* The descriptor set this surface corresponds to */
    uint16_t set;
@@ -1073,6 +1123,7 @@ struct anv_attachment_state {
 struct anv_cmd_state {
    /* PIPELINE_SELECT.PipelineSelection */
    uint32_t                                     current_pipeline;
+   uint32_t                                     current_l3_config;
    uint32_t                                     vb_dirty;
    anv_cmd_dirty_mask_t                         dirty;
    anv_cmd_dirty_mask_t                         compute_dirty;
@@ -1232,24 +1283,41 @@ void anv_cmd_buffer_emit_state_base_address(struct anv_cmd_buffer *cmd_buffer);
 
 void anv_cmd_state_setup_attachments(struct anv_cmd_buffer *cmd_buffer,
                                      const VkRenderPassBeginInfo *info);
-void gen7_cmd_buffer_begin_subpass(struct anv_cmd_buffer *cmd_buffer,
-                                   struct anv_subpass *subpass);
 
-void gen8_cmd_buffer_begin_subpass(struct anv_cmd_buffer *cmd_buffer,
+void gen7_cmd_buffer_set_subpass(struct anv_cmd_buffer *cmd_buffer,
+                                   struct anv_subpass *subpass);
+void gen75_cmd_buffer_set_subpass(struct anv_cmd_buffer *cmd_buffer,
+                                  struct anv_subpass *subpass);
+void gen8_cmd_buffer_set_subpass(struct anv_cmd_buffer *cmd_buffer,
                                    struct anv_subpass *subpass);
-void gen9_cmd_buffer_begin_subpass(struct anv_cmd_buffer *cmd_buffer,
+void gen9_cmd_buffer_set_subpass(struct anv_cmd_buffer *cmd_buffer,
                                    struct anv_subpass *subpass);
-
-void anv_cmd_buffer_begin_subpass(struct anv_cmd_buffer *cmd_buffer,
+void anv_cmd_buffer_set_subpass(struct anv_cmd_buffer *cmd_buffer,
                                   struct anv_subpass *subpass);
 
+void gen7_flush_pipeline_select_3d(struct anv_cmd_buffer *cmd_buffer);
+void gen75_flush_pipeline_select_3d(struct anv_cmd_buffer *cmd_buffer);
+void gen8_flush_pipeline_select_3d(struct anv_cmd_buffer *cmd_buffer);
+void gen9_flush_pipeline_select_3d(struct anv_cmd_buffer *cmd_buffer);
+
+void gen7_cmd_buffer_flush_state(struct anv_cmd_buffer *cmd_buffer);
+void gen75_cmd_buffer_flush_state(struct anv_cmd_buffer *cmd_buffer);
+void gen8_cmd_buffer_flush_state(struct anv_cmd_buffer *cmd_buffer);
+void gen9_cmd_buffer_flush_state(struct anv_cmd_buffer *cmd_buffer);
+
+void gen7_cmd_buffer_flush_compute_state(struct anv_cmd_buffer *cmd_buffer);
+void gen75_cmd_buffer_flush_compute_state(struct anv_cmd_buffer *cmd_buffer);
+void gen8_cmd_buffer_flush_compute_state(struct anv_cmd_buffer *cmd_buffer);
+void gen9_cmd_buffer_flush_compute_state(struct anv_cmd_buffer *cmd_buffer);
+
 struct anv_state
 anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer,
                               gl_shader_stage stage);
 struct anv_state
 anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer);
 
-void anv_cmd_buffer_clear_attachments(struct anv_cmd_buffer *cmd_buffer);
+void anv_cmd_buffer_clear_subpass(struct anv_cmd_buffer *cmd_buffer);
+void anv_cmd_buffer_resolve_subpass(struct anv_cmd_buffer *cmd_buffer);
 
 const struct anv_image_view *
 anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer *cmd_buffer);
@@ -1264,7 +1332,7 @@ struct anv_fence {
 };
 
 struct anv_event {
-   uint32_t                                     semaphore;
+   uint64_t                                     semaphore;
    struct anv_state                             state;
 };
 
@@ -1273,10 +1341,16 @@ struct nir_shader;
 struct anv_shader_module {
    struct nir_shader *                          nir;
 
+   unsigned char                                sha1[20];
    uint32_t                                     size;
    char                                         data[0];
 };
 
+void anv_hash_shader(unsigned char *hash, const void *key, size_t key_size,
+                     struct anv_shader_module *module,
+                     const char *entrypoint,
+                     const VkSpecializationInfo *spec_info);
+
 static inline gl_shader_stage
 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
 {
@@ -1337,7 +1411,6 @@ struct anv_pipeline {
    uint32_t                                     ps_grf_start0;
    uint32_t                                     ps_grf_start2;
    uint32_t                                     gs_kernel;
-   uint32_t                                     gs_vertex_count;
    uint32_t                                     cs_simd;
 
    uint32_t                                     vb_used;
@@ -1366,6 +1439,12 @@ struct anv_pipeline {
 };
 
 struct anv_graphics_pipeline_create_info {
+   /**
+    * If non-negative, overrides the color attachment count of the pipeline's
+    * subpass.
+    */
+   int8_t color_attachment_count;
+
    bool                                         use_repclear;
    bool                                         disable_viewport;
    bool                                         disable_scissor;
@@ -1452,12 +1531,20 @@ gen9_compute_pipeline_create(VkDevice _device,
                              const VkAllocationCallbacks *alloc,
                              VkPipeline *pPipeline);
 
+struct anv_format_swizzle {
+   unsigned r:2;
+   unsigned g:2;
+   unsigned b:2;
+   unsigned a:2;
+};
+
 struct anv_format {
    const VkFormat vk_format;
    const char *name;
-   enum isl_format surface_format; /**< RENDER_SURFACE_STATE.SurfaceFormat */
+   enum isl_format isl_format; /**< RENDER_SURFACE_STATE.SurfaceFormat */
    const struct isl_format_layout *isl_layout;
-   uint16_t depth_format; /**< 3DSTATE_DEPTH_BUFFER.SurfaceFormat */
+   struct anv_format_swizzle swizzle;
+   bool has_depth;
    bool has_stencil;
 };
 
@@ -1466,18 +1553,18 @@ anv_format_for_vk_format(VkFormat format);
 
 enum isl_format
 anv_get_isl_format(VkFormat format, VkImageAspectFlags aspect,
-                   VkImageTiling tiling);
+                   VkImageTiling tiling, struct anv_format_swizzle *swizzle);
 
 static inline bool
 anv_format_is_color(const struct anv_format *format)
 {
-   return !format->depth_format && !format->has_stencil;
+   return !format->has_depth && !format->has_stencil;
 }
 
 static inline bool
 anv_format_is_depth_or_stencil(const struct anv_format *format)
 {
-   return format->depth_format || format->has_stencil;
+   return format->has_depth || format->has_stencil;
 }
 
 /**
@@ -1502,6 +1589,7 @@ struct anv_image {
    VkExtent3D extent;
    uint32_t levels;
    uint32_t array_size;
+   uint32_t samples; /**< VkImageCreateInfo::samples */
    VkImageUsageFlags usage; /**< Superset of VkImageCreateInfo::usage. */
    VkImageTiling tiling; /** VkImageCreateInfo::tiling */
 
@@ -1512,10 +1600,6 @@ struct anv_image {
    struct anv_bo *bo;
    VkDeviceSize offset;
 
-   bool needs_nonrt_surface_state:1;
-   bool needs_color_rt_surface_state:1;
-   bool needs_storage_surface_state:1;
-
    /**
     * Image subsurfaces
     *
@@ -1545,14 +1629,18 @@ struct anv_image_view {
 
    VkImageAspectFlags aspect_mask;
    VkFormat vk_format;
+   VkComponentMapping swizzle;
    enum isl_format format;
+   uint32_t base_layer;
+   uint32_t base_mip;
+   VkExtent3D level_0_extent; /**< Extent of ::image's level 0 adjusted for ::vk_format. */
    VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
 
    /** RENDER_SURFACE_STATE when using image as a color render target. */
    struct anv_state color_rt_surface_state;
 
-   /** RENDER_SURFACE_STATE when using image as a non render target. */
-   struct anv_state nonrt_surface_state;
+   /** RENDER_SURFACE_STATE when using image as a sampler surface. */
+   struct anv_state sampler_surface_state;
 
    /** RENDER_SURFACE_STATE when using image as a storage image. */
    struct anv_state storage_surface_state;
@@ -1576,31 +1664,34 @@ anv_image_get_surface_for_aspect_mask(struct anv_image *image,
 void anv_image_view_init(struct anv_image_view *view,
                          struct anv_device *device,
                          const VkImageViewCreateInfo* pCreateInfo,
-                         struct anv_cmd_buffer *cmd_buffer);
+                         struct anv_cmd_buffer *cmd_buffer,
+                         uint32_t offset);
 
 void
-gen7_image_view_init(struct anv_image_view *iview,
-                     struct anv_device *device,
-                     const VkImageViewCreateInfo* pCreateInfo,
-                     struct anv_cmd_buffer *cmd_buffer);
-
+anv_fill_image_surface_state(struct anv_device *device, struct anv_state state,
+                             struct anv_image_view *iview,
+                             const VkImageViewCreateInfo *pCreateInfo,
+                             VkImageUsageFlagBits usage);
 void
-gen75_image_view_init(struct anv_image_view *iview,
-                      struct anv_device *device,
-                      const VkImageViewCreateInfo* pCreateInfo,
-                      struct anv_cmd_buffer *cmd_buffer);
-
+gen7_fill_image_surface_state(struct anv_device *device, void *state_map,
+                              struct anv_image_view *iview,
+                              const VkImageViewCreateInfo *pCreateInfo,
+                              VkImageUsageFlagBits usage);
 void
-gen8_image_view_init(struct anv_image_view *iview,
-                     struct anv_device *device,
-                     const VkImageViewCreateInfo* pCreateInfo,
-                     struct anv_cmd_buffer *cmd_buffer);
-
+gen75_fill_image_surface_state(struct anv_device *device, void *state_map,
+                               struct anv_image_view *iview,
+                               const VkImageViewCreateInfo *pCreateInfo,
+                               VkImageUsageFlagBits usage);
+void
+gen8_fill_image_surface_state(struct anv_device *device, void *state_map,
+                              struct anv_image_view *iview,
+                              const VkImageViewCreateInfo *pCreateInfo,
+                              VkImageUsageFlagBits usage);
 void
-gen9_image_view_init(struct anv_image_view *iview,
-                     struct anv_device *device,
-                     const VkImageViewCreateInfo* pCreateInfo,
-                     struct anv_cmd_buffer *cmd_buffer);
+gen9_fill_image_surface_state(struct anv_device *device, void *state_map,
+                              struct anv_image_view *iview,
+                              const VkImageViewCreateInfo *pCreateInfo,
+                              VkImageUsageFlagBits usage);
 
 struct anv_buffer_view {
    enum isl_format format; /**< VkBufferViewCreateInfo::format */
@@ -1615,7 +1706,8 @@ struct anv_buffer_view {
 const struct anv_format *
 anv_format_for_descriptor_type(VkDescriptorType type);
 
-void anv_fill_buffer_surface_state(struct anv_device *device, void *state,
+void anv_fill_buffer_surface_state(struct anv_device *device,
+                                   struct anv_state state,
                                    enum isl_format format,
                                    uint32_t offset, uint32_t range,
                                    uint32_t stride);
@@ -1650,7 +1742,7 @@ struct anv_framebuffer {
    uint32_t                                     layers;
 
    uint32_t                                     attachment_count;
-   const struct anv_image_view *           attachments[0];
+   struct anv_image_view *                      attachments[0];
 };
 
 struct anv_subpass {
@@ -1660,6 +1752,9 @@ struct anv_subpass {
    uint32_t *                                   color_attachments;
    uint32_t *                                   resolve_attachments;
    uint32_t                                     depth_stencil_attachment;
+
+   /** Subpass has at least one resolve attachment */
+   bool                                         has_resolve;
 };
 
 struct anv_render_pass_attachment {