turnip: Implement stream-out emit and vkApis for transform feedback
[mesa.git] / src / freedreno / vulkan / tu_private.h
index 4169d015ce1471e82b07f1b925c79bde5aeba30c..661529fa537be7fb2889e1115db5a743f41f4a55 100644 (file)
@@ -313,9 +313,17 @@ struct tu_physical_device
 
    unsigned gpu_id;
    uint32_t gmem_size;
+   uint64_t gmem_base;
    uint32_t tile_align_w;
    uint32_t tile_align_h;
 
+   struct {
+      uint32_t RB_UNKNOWN_8E04_blit;    /* for CP_BLIT's */
+      uint32_t RB_CCU_CNTL_gmem;        /* for GMEM */
+      uint32_t PC_UNKNOWN_9805;
+      uint32_t SP_UNKNOWN_A0F8;
+   } magic;
+
    /* This is the drivers on-disk cache used as a fallback as opposed to
     * the pipeline cache defined by apps.
     */
@@ -330,6 +338,8 @@ enum tu_debug_flags
    TU_DEBUG_NIR = 1 << 1,
    TU_DEBUG_IR3 = 1 << 2,
    TU_DEBUG_NOBIN = 1 << 3,
+   TU_DEBUG_SYSMEM = 1 << 4,
+   TU_DEBUG_FORCEBIN = 1 << 5,
 };
 
 struct tu_instance
@@ -453,6 +463,14 @@ struct tu_queue
    struct tu_fence submit_fence;
 };
 
+struct tu_bo
+{
+   uint32_t gem_handle;
+   uint64_t size;
+   uint64_t iova;
+   void *map;
+};
+
 struct tu_device
 {
    VK_LOADER_DATA _loader_data;
@@ -473,20 +491,17 @@ struct tu_device
    /* Backup in-memory cache to be used if the app doesn't provide one */
    struct tu_pipeline_cache *mem_cache;
 
+   struct tu_bo vsc_data;
+   struct tu_bo vsc_data2;
+   uint32_t vsc_data_pitch;
+   uint32_t vsc_data2_pitch;
+
    struct list_head shader_slabs;
    mtx_t shader_slab_mutex;
 
    struct tu_device_extension_table enabled_extensions;
 };
 
-struct tu_bo
-{
-   uint32_t gem_handle;
-   uint64_t size;
-   uint64_t iova;
-   void *map;
-};
-
 VkResult
 tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size);
 VkResult
@@ -555,6 +570,7 @@ struct tu_cs
    uint32_t *reserved_end;
    uint32_t *end;
 
+   struct tu_device *device;
    enum tu_cs_mode mode;
    uint32_t next_bo_size;
 
@@ -565,6 +581,10 @@ struct tu_cs
    struct tu_bo **bos;
    uint32_t bo_count;
    uint32_t bo_capacity;
+
+   /* state for cond_exec_start/cond_exec_end */
+   uint32_t cond_flags;
+   uint32_t *cond_dwords;
 };
 
 struct tu_device_memory
@@ -778,7 +798,6 @@ tu_get_perftest_option_name(int id);
 struct tu_descriptor_state
 {
    struct tu_descriptor_set *sets[MAX_SETS];
-   uint32_t dirty;
    uint32_t valid;
    struct tu_push_descriptor_set push_set;
    bool push_dirty;
@@ -810,6 +829,9 @@ struct tu_tiling_config
    /* pipe register values */
    uint32_t pipe_config[MAX_VSC_PIPES];
    uint32_t pipe_sizes[MAX_VSC_PIPES];
+
+   /* Whether sysmem rendering must be used */
+   bool force_sysmem;
 };
 
 enum tu_cmd_dirty_bits
@@ -819,6 +841,7 @@ enum tu_cmd_dirty_bits
    TU_CMD_DIRTY_VERTEX_BUFFERS = 1 << 2,
    TU_CMD_DIRTY_DESCRIPTOR_SETS = 1 << 3,
    TU_CMD_DIRTY_PUSH_CONSTANTS = 1 << 4,
+   TU_CMD_DIRTY_STREAMOUT_BUFFERS = 1 << 5,
 
    TU_CMD_DIRTY_DYNAMIC_LINE_WIDTH = 1 << 16,
    TU_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 17,
@@ -826,6 +849,14 @@ enum tu_cmd_dirty_bits
    TU_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE = 1 << 19,
 };
 
+struct tu_streamout_state {
+   uint16_t stride[IR3_MAX_SO_BUFFERS];
+   uint32_t ncomp[IR3_MAX_SO_BUFFERS];
+   uint32_t prog[IR3_MAX_SO_OUTPUTS * 2];
+   uint32_t prog_count;
+   uint32_t vpc_so_buf_cntl;
+};
+
 struct tu_cmd_state
 {
    uint32_t dirty;
@@ -842,6 +873,17 @@ struct tu_cmd_state
 
    struct tu_dynamic_state dynamic;
 
+   /* Stream output buffers */
+   struct
+   {
+      struct tu_buffer *buffers[IR3_MAX_SO_BUFFERS];
+      VkDeviceSize offsets[IR3_MAX_SO_BUFFERS];
+      VkDeviceSize sizes[IR3_MAX_SO_BUFFERS];
+   } streamout_buf;
+
+   uint8_t streamout_reset;
+   uint8_t streamout_enabled;
+
    /* Index buffer */
    struct tu_buffer *index_buffer;
    uint64_t index_offset;
@@ -855,7 +897,6 @@ struct tu_cmd_state
 
    struct tu_tiling_config tiling_config;
 
-   struct tu_cs_entry tile_load_ib;
    struct tu_cs_entry tile_store_ib;
 };
 
@@ -936,15 +977,14 @@ struct tu_cmd_buffer
    struct tu_bo_list bo_list;
    struct tu_cs cs;
    struct tu_cs draw_cs;
+   struct tu_cs draw_epilogue_cs;
    struct tu_cs sub_cs;
 
-   uint16_t marker_reg;
-   uint32_t marker_seqno;
-
    struct tu_bo scratch_bo;
    uint32_t scratch_seqno;
 #define VSC_OVERFLOW 0x8
 #define VSC_SCRATCH 0x10
+#define VSC_FLUSH   0x20
 
    struct tu_bo vsc_data;
    struct tu_bo vsc_data2;
@@ -955,6 +995,19 @@ struct tu_cmd_buffer
    bool wait_for_idle;
 };
 
+/* Temporary struct for tracking a register state to be written, used by
+ * a6xx-pack.h and tu_cs_emit_regs()
+ */
+struct tu_reg_value {
+   uint32_t reg;
+   uint64_t value;
+   bool is_address;
+   struct tu_bo *bo;
+   bool bo_write;
+   uint32_t bo_offset;
+   uint32_t bo_shift;
+};
+
 unsigned
 tu6_emit_event_write(struct tu_cmd_buffer *cmd,
                      struct tu_cs *cs,
@@ -1044,10 +1097,10 @@ struct tu_descriptor_map
    /* TODO: avoid fixed size array/justify the size */
    unsigned num; /* number of array entries */
    unsigned num_desc; /* Number of descriptors (sum of array_size[]) */
-   int set[64];
-   int binding[64];
-   int value[64];
-   int array_size[64];
+   int set[128];
+   int binding[128];
+   int value[128];
+   int array_size[128];
 };
 
 struct tu_shader
@@ -1120,6 +1173,8 @@ struct tu_pipeline
    bool need_indirect_descriptor_sets;
    VkShaderStageFlags active_stages;
 
+   struct tu_streamout_state streamout;
+
    struct
    {
       struct tu_bo binary_bo;
@@ -1230,17 +1285,24 @@ struct tu_graphics_pipeline_create_info
    uint32_t custom_blend_mode;
 };
 
+enum tu_supported_formats {
+   FMT_VERTEX = 1,
+   FMT_TEXTURE = 2,
+   FMT_COLOR = 4,
+};
+
 struct tu_native_format
 {
-   int vtx;      /* VFMTn_xxx or -1 */
-   int tex;      /* TFMTn_xxx or -1 */
-   int rb;       /* RBn_xxx or -1 */
-   int swap;     /* enum a3xx_color_swap */
-   bool present; /* internal only; always true to external users */
+   enum a6xx_format fmt : 8;
+   enum a3xx_color_swap swap : 8;
+   enum tu_supported_formats supported : 8;
 };
 
-const struct tu_native_format *
-tu6_get_native_format(VkFormat format);
+struct tu_native_format tu6_get_native_format(VkFormat format);
+struct tu_native_format tu6_format_vtx(VkFormat format);
+enum a6xx_format tu6_format_gmem(VkFormat format);
+struct tu_native_format tu6_format_color(VkFormat format, bool tiled);
+struct tu_native_format tu6_format_texture(VkFormat format, bool tiled);
 
 void
 tu_pack_clear_value(const VkClearValue *val,
@@ -1253,7 +1315,7 @@ tu_2d_clear_color(const VkClearColorValue *val, VkFormat format, uint32_t buf[4]
 void
 tu_2d_clear_zs(const VkClearDepthStencilValue *val, VkFormat format, uint32_t buf[4]);
 
-enum a6xx_2d_ifmt tu6_rb_fmt_to_ifmt(enum a6xx_color_fmt fmt);
+enum a6xx_2d_ifmt tu6_fmt_to_ifmt(enum a6xx_format fmt);
 enum a6xx_depth_format tu6_pipe2depth(VkFormat format);
 
 struct tu_image_level
@@ -1338,10 +1400,18 @@ tu_image_base(struct tu_image *image, int level, int layer)
       fdl_surface_offset(&image->layout, level, layer);
 }
 
+#define tu_image_base_ref(image, level, layer)                          \
+   .bo = image->bo,                                                     \
+   .bo_offset = (image->bo_offset + fdl_surface_offset(&image->layout,  \
+                                                       level, layer))
+
+#define tu_image_view_base_ref(iview)                                   \
+   tu_image_base_ref(iview->image, iview->base_mip, iview->base_layer)
+
 static inline VkDeviceSize
 tu_image_ubwc_size(struct tu_image *image, int level)
 {
-   return image->layout.ubwc_size;
+   return image->layout.ubwc_layer_size;
 }
 
 static inline uint32_t
@@ -1350,14 +1420,28 @@ tu_image_ubwc_pitch(struct tu_image *image, int level)
    return image->layout.ubwc_slices[level].pitch;
 }
 
+static inline uint64_t
+tu_image_ubwc_surface_offset(struct tu_image *image, int level, int layer)
+{
+   return image->layout.ubwc_slices[level].offset +
+      layer * tu_image_ubwc_size(image, level);
+}
+
 static inline uint64_t
 tu_image_ubwc_base(struct tu_image *image, int level, int layer)
 {
    return image->bo->iova + image->bo_offset +
-          image->layout.ubwc_slices[level].offset +
-          layer * tu_image_ubwc_size(image, level);
+      tu_image_ubwc_surface_offset(image, level, layer);
 }
 
+#define tu_image_ubwc_base_ref(image, level, layer)                     \
+   .bo = image->bo,                                                     \
+   .bo_offset = (image->bo_offset + tu_image_ubwc_surface_offset(image, \
+                                                                 level, layer))
+
+#define tu_image_view_ubwc_base_ref(iview) \
+   tu_image_ubwc_base_ref(iview->image, iview->base_mip, iview->base_layer)
+
 enum a6xx_tile_mode
 tu6_get_image_tile_mode(struct tu_image *image, int level);
 enum a3xx_msaa_samples
@@ -1413,9 +1497,9 @@ tu_image_view_init(struct tu_image_view *view,
 
 struct tu_buffer_view
 {
-   VkFormat vk_format;
-   uint64_t range; /**< VkBufferViewCreateInfo::range */
-   uint32_t state[4];
+   uint32_t descriptor[A6XX_TEX_CONST_DWORDS];
+
+   struct tu_buffer *buffer;
 };
 void
 tu_buffer_view_init(struct tu_buffer_view *view,
@@ -1514,12 +1598,11 @@ tu_device_finish_meta(struct tu_device *device);
 
 struct tu_query_pool
 {
+   VkQueryType type;
    uint32_t stride;
-   uint32_t availability_offset;
    uint64_t size;
-   char *ptr;
-   VkQueryType type;
-   uint32_t pipeline_stats_mask;
+   uint32_t pipeline_statistics;
+   struct tu_bo bo;
 };
 
 struct tu_semaphore
@@ -1565,6 +1648,9 @@ tu_drm_get_gpu_id(const struct tu_physical_device *dev, uint32_t *id);
 int
 tu_drm_get_gmem_size(const struct tu_physical_device *dev, uint32_t *size);
 
+int
+tu_drm_get_gmem_base(const struct tu_physical_device *dev, uint64_t *base);
+
 int
 tu_drm_submitqueue_new(const struct tu_device *dev,
                        int priority,
@@ -1588,6 +1674,21 @@ tu_gem_info_offset(const struct tu_device *dev, uint32_t gem_handle);
 uint64_t
 tu_gem_info_iova(const struct tu_device *dev, uint32_t gem_handle);
 
+
+void
+tu_clear_sysmem_attachment(struct tu_cmd_buffer *cmd,
+                           struct tu_cs *cs,
+                           uint32_t attachment,
+                           const VkClearValue *value,
+                           const VkClearRect *rect);
+
+void
+tu_clear_gmem_attachment(struct tu_cmd_buffer *cmd,
+                         struct tu_cs *cs,
+                         uint32_t attachment,
+                         uint8_t component_mask,
+                         const VkClearValue *value);
+
 #define TU_DEFINE_HANDLE_CASTS(__tu_type, __VkType)                          \
                                                                              \
    static inline struct __tu_type *__tu_type##_from_handle(__VkType _handle) \