pvr: Add support for generating render pass hw setup data.

Signed-off-by: Rajnesh Kanwal <rajnesh.kanwal@imgtec.com>
Reviewed-by: Karmjit Mahil <Karmjit.Mahil@imgtec.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18707>

src/imagination/common/pvr_device_info.c

@@ -137,6 +137,7 @@ const struct pvr_device_features pvr_device_features_33_V_11_3 = {
    .has_num_clusters = true,
    .has_num_raster_pipes = true,
    .has_num_user_clip_planes = true,
+   .has_pbe2_in_xe = true,
    .has_roguexe = true,
    .has_screen_size8K = true,
    .has_simple_internal_parameter_format = true,
@@ -216,6 +217,7 @@ const struct pvr_device_features pvr_device_features_36_V_104_796 = {
    .has_num_raster_pipes = true,
    .has_num_user_clip_planes = true,
    .has_paired_tiles = true,
+   .has_pbe2_in_xe = true,
    .has_pds_ddmadt = true,
    .has_roguexe = true,
    .has_screen_size8K = true,

src/imagination/common/pvr_device_info.h

@@ -267,6 +267,7 @@ struct pvr_device_features {
    bool has_num_raster_pipes : 1;
    bool has_num_user_clip_planes : 1;
    bool has_paired_tiles : 1;
+   bool has_pbe2_in_xe : 1;
    bool has_pds_ddmadt : 1;
    bool has_robust_buffer_access : 1;
    bool has_roguexe : 1;
@@ -277,6 +278,7 @@ struct pvr_device_features {
    bool has_slc_cache_line_size_bits : 1;
    bool has_slc_mcu_cache_controls : 1;
    bool has_tf_bicubic_filter : 1;
+   bool has_tile_per_usc : 1;
    bool has_tile_size_16x16 : 1;
    bool has_tile_size_x : 1;
    bool has_tile_size_y : 1;

src/imagination/include/hwdef/rogue_hw_defs.h

@@ -125,4 +125,6 @@
  */
 #define ROGUE_MAX_OVERLAPPED_PIXEL_TASK_INSTANCES 7U
 
+#define PVR_NUM_PBE_EMIT_REGS 8U
+
 #endif /* ROGUE_HW_DEFS_H */

src/imagination/include/hwdef/rogue_hw_utils.h

@@ -216,12 +216,22 @@ rogue_max_compute_shared_registers(const struct pvr_device_info *dev_info)
    return 0U;
 }
 
+static inline uint32_t
+rogue_get_max_num_cores(const struct pvr_device_info *dev_info)
+{
+   if (PVR_HAS_FEATURE(dev_info, gpu_multicore_support) &&
+       PVR_HAS_FEATURE(dev_info, xpu_max_slaves)) {
+      return PVR_GET_FEATURE_VALUE(dev_info, xpu_max_slaves, 0U) + 1U;
+   }
+
+   return 1U;
+}
+
 static inline uint32_t
 rogue_get_cdm_context_resume_buffer_size(const struct pvr_device_info *dev_info)
 {
    if (PVR_HAS_FEATURE(dev_info, gpu_multicore_support)) {
-      const uint32_t max_num_cores =
-         PVR_GET_FEATURE_VALUE(dev_info, xpu_max_slaves, 0U) + 1U;
+      const uint32_t max_num_cores = rogue_get_max_num_cores(dev_info);
       const uint32_t cache_line_size = rogue_get_slc_cache_line_size(dev_info);
       const uint32_t cdm_context_resume_buffer_stride =
          ALIGN_POT(ROGUE_LLS_CDM_CONTEXT_RESUME_BUFFER_SIZE, cache_line_size);

src/imagination/vulkan/pvr_cmd_buffer.c

@@ -575,7 +575,7 @@ pvr_load_op_constants_create_and_upload(struct pvr_cmd_buffer *cmd_buffer,
    const struct pvr_render_pass *pass = render_pass_info->pass;
    const struct pvr_renderpass_hwsetup_render *hw_render =
       &pass->hw_setup->renders[idx];
-   ASSERTED const struct pvr_load_op *load_op = hw_render->client_data;
+   ASSERTED const struct pvr_load_op *load_op = hw_render->load_op;
    const struct pvr_renderpass_colorinit *color_init =
       &hw_render->color_init[0];
    const struct pvr_render_pass_attachment *attachment =
@@ -618,7 +618,7 @@ static VkResult pvr_load_op_pds_data_create_and_upload(
    const struct pvr_render_pass_info *render_pass_info =
       &cmd_buffer->state.render_pass_info;
    const struct pvr_load_op *load_op =
-      render_pass_info->pass->hw_setup->renders[idx].client_data;
+      render_pass_info->pass->hw_setup->renders[idx].load_op;
    struct pvr_device *device = cmd_buffer->device;
    const struct pvr_device_info *dev_info = &device->pdevice->dev_info;
    struct pvr_pds_pixel_shader_sa_program program = { 0 };
@@ -979,7 +979,7 @@ static VkResult pvr_sub_cmd_gfx_job_init(const struct pvr_device_info *dev_info,
 
    /* FIXME: Don't do this if there is a barrier load. */
    if (render_pass_info->enable_bg_tag) {
-      const struct pvr_load_op *load_op = hw_render->client_data;
+      const struct pvr_load_op *load_op = hw_render->load_op;
       struct pvr_pds_upload load_op_program;
 
       /* FIXME: Should we free the PDS pixel event data or let it be freed

src/imagination/vulkan/pvr_hw_pass.h

@@ -31,55 +31,64 @@
 struct pvr_device;
 struct pvr_render_pass;
 
-struct pvr_renderpass_hwsetup_subpass {
-   /* If >=0 then copy the depth into this pixel output for all fragment
-    * programs in the subpass.
-    */
-   int32_t z_replicate;
-
-   /* The operation to perform on the depth at the start of the subpass. Loads
-    * are deferred to subpasses when depth has been replicated
-    */
-   VkAttachmentLoadOp depth_initop;
-
-   /* If true then clear the stencil at the start of the subpass. */
-   bool stencil_clear;
-
-   /* Driver Id from the input pvr_render_subpass structure. */
-   uint32_t index;
-
-   /* For each color attachment to the subpass: the operation to perform at
-    * the start of the subpass.
-    */
-   VkAttachmentLoadOp *color_initops;
-
-   struct pvr_load_op *load_op;
-};
-
-struct pvr_renderpass_colorinit {
-   /* Source surface for the operation. */
-   uint32_t index;
-
-   /* Type of operation: either clear or load. */
-   VkAttachmentLoadOp op;
-};
-
-/* FIXME: Adding these USC enums and structures here for now to avoid adding
- * usc.h header. Needs to be moved to compiler specific header.
- */
 /* Specifies the location of render target writes. */
 enum usc_mrt_resource_type {
-   USC_MRT_RESOURCE_TYPE_INVALID = 0, /* explicitly treat 0 as invalid */
+   USC_MRT_RESOURCE_TYPE_INVALID = 0, /* explicitly treat 0 as invalid. */
    USC_MRT_RESOURCE_TYPE_OUTPUT_REG,
    USC_MRT_RESOURCE_TYPE_MEMORY,
 };
 
+enum pvr_resolve_type {
+   PVR_RESOLVE_TYPE_INVALID = 0, /* explicitly treat 0 as invalid. */
+   PVR_RESOLVE_TYPE_PBE,
+   PVR_RESOLVE_TYPE_TRANSFER,
+};
+
+enum pvr_renderpass_hwsetup_input_access {
+   /* The attachment must be loaded using a texture sample. */
+   PVR_RENDERPASS_HWSETUP_INPUT_ACCESS_OFFCHIP,
+   /* The attachment can be loaded from an output register or tile buffer. */
+   PVR_RENDERPASS_HWSETUP_INPUT_ACCESS_ONCHIP,
+   /* As _ONCHIP but the attachment is the result of a Z replicate in the same
+    * subpass.
+    */
+   PVR_RENDERPASS_HWSETUP_INPUT_ACCESS_ONCHIP_ZREPLICATE,
+};
+
+#define PVR_USC_RENDER_TARGET_MAXIMUM_SIZE_IN_DWORDS (4)
+
+struct usc_mrt_desc {
+   /* Size (in bytes) of the intermediate storage required for each pixel in the
+    * render target.
+    */
+   uint32_t intermediate_size;
+
+   /* Number of bytes allocated for each component in the output registers (as
+    * opposed to the pixel format).
+    */
+   uint32_t component_alignment;
+
+   /* Mask of the bits from each dword which are read by the PBE. */
+   uint32_t valid_mask[PVR_USC_RENDER_TARGET_MAXIMUM_SIZE_IN_DWORDS];
+
+   /* Higher number = higher priority. Used to decide which render targets get
+    * allocated dedicated output registers.
+    */
+   uint32_t priority;
+};
+
 struct usc_mrt_resource {
+   /* Input description of render target. */
+   struct usc_mrt_desc mrt_desc;
+
    /* Resource type allocated for render target. */
    enum usc_mrt_resource_type type;
 
+   /* Intermediate pixel size (in bytes). */
+   uint32_t intermediate_size;
+
    union {
-      /* If type == USC_MRT_RESOURCE_TYPE_OUTPUT_REGISTER. */
+      /* If type == USC_MRT_RESOURCE_TYPE_OUTPUT_REG. */
       struct {
          /* The output register to use. */
          uint32_t output_reg;
@@ -90,7 +99,7 @@ struct usc_mrt_resource {
 
       /* If type == USC_MRT_RESOURCE_TYPE_MEMORY. */
       struct {
-         /* The number of the tile buffer to use. */
+         /* The index of the tile buffer to use. */
          uint32_t tile_buffer;
 
          /* The offset in dwords within the tile buffer. */
@@ -103,16 +112,25 @@ struct usc_mrt_setup {
    /* Number of render targets present. */
    uint32_t num_render_targets;
 
+   /* Number of output registers used per-pixel (1, 2 or 4). */
+   uint32_t num_output_regs;
+
+   /* Number of tile buffers used. */
+   uint32_t num_tile_buffers;
+
+   /* Size of a tile buffer in bytes. */
+   uint32_t tile_buffer_size;
+
    /* Array of MRT resources allocated for each render target. The number of
-    * elements is determined by usc_mrt_setup::render_targets_count.
+    * elements is determined by usc_mrt_setup::num_render_targets.
     */
    struct usc_mrt_resource *mrt_resources;
-};
 
-enum pvr_resolve_type {
-   PVR_RESOLVE_TYPE_INVALID = 0, /* explicitly treat 0 as invalid */
-   PVR_RESOLVE_TYPE_PBE,
-   PVR_RESOLVE_TYPE_TRANSFER,
+   /* Don't set up source pos in emit. */
+   bool disable_source_pos_override;
+
+   /* Hash unique to this particular setup. */
+   uint32_t hash;
 };
 
 struct pvr_renderpass_hwsetup_eot_surface {
@@ -138,6 +156,51 @@ struct pvr_renderpass_hwsetup_eot_surface {
    uint32_t src_attachment_idx;
 };
 
+struct pvr_renderpass_hwsetup_subpass {
+   /* Mapping from fragment stage pixel outputs to hardware storage for all
+    * fragment programs in the subpass.
+    */
+   struct usc_mrt_setup setup;
+
+   /* If >=0 then copy the depth into this pixel output for all fragment
+    * programs in the subpass.
+    */
+   int32_t z_replicate;
+
+   /* The operation to perform on the depth at the start of the subpass. Loads
+    * are deferred to subpasses when depth has been replicated.
+    */
+   VkAttachmentLoadOp depth_initop;
+
+   /* If true then clear the stencil at the start of the subpass. */
+   bool stencil_clear;
+
+   /* Subpass index from the input pvr_render_subpass structure. */
+   uint32_t index;
+
+   /* For each color attachment to the subpass the operation to perform at
+    * the start of the subpass.
+    */
+   VkAttachmentLoadOp *color_initops;
+
+   struct pvr_load_op *load_op;
+
+   struct {
+      enum pvr_renderpass_hwsetup_input_access type;
+      uint32_t on_chip_rt;
+   } * input_access;
+
+   uint8_t output_register_mask;
+};
+
+struct pvr_renderpass_colorinit {
+   /* Source attachment for the operation. */
+   uint32_t index;
+
+   /* Type of operation either clear or load. */
+   VkAttachmentLoadOp op;
+};
+
 struct pvr_renderpass_hwsetup_render {
    /* Number of pixel output registers to allocate for this render. */
    uint32_t output_regs_count;
@@ -152,17 +215,17 @@ struct pvr_renderpass_hwsetup_render {
    struct pvr_renderpass_hwsetup_subpass *subpasses;
 
    /* The sample count of every color attachment (or depth attachment if
-    * z-only) in this render
+    * z-only) in this render.
     */
    uint32_t sample_count;
 
-   /* Driver Id for the surface to use for depth/stencil load/store in this
+   /* Index of the attachment to use for depth/stencil load/store in this
     * render.
     */
    int32_t ds_attach_idx;
 
    /* Operation on the on-chip depth at the start of the render.
-    * Either load from 'ds_surface_id', clear using 'ds_surface_id' or leave
+    * Either load from 'ds_attach_idx', clear using 'ds_attach_idx' or leave
     * uninitialized.
     */
    VkAttachmentLoadOp depth_init;
@@ -170,23 +233,33 @@ struct pvr_renderpass_hwsetup_render {
    /* Operation on the on-chip stencil at the start of the render. */
    VkAttachmentLoadOp stencil_init;
 
-   /* For each operation: the destination in the on-chip color storage. */
-   struct usc_mrt_setup init_setup;
-
    /* Count of operations on on-chip color storage at the start of the render.
     */
    uint32_t color_init_count;
 
+   /* For each operation: the destination in the on-chip color storage. */
+   struct usc_mrt_setup init_setup;
+
    /* How to initialize render targets at the start of the render. */
    struct pvr_renderpass_colorinit *color_init;
 
+   /* true to store depth to 'ds_attach_idx' at the end of the render. */
+   bool depth_store;
+   /* true to store stencil to 'ds_attach_idx' at the end of the render. */
+   bool stencil_store;
+
    /* Describes the location of the source data for each stored surface. */
    struct usc_mrt_setup eot_setup;
 
    struct pvr_renderpass_hwsetup_eot_surface *eot_surfaces;
    uint32_t eot_surface_count;
 
-   void *client_data;
+   uint32_t pbe_emits;
+
+   /* true if this HW render has lasting effects on its attachments. */
+   bool has_side_effects;
+
+   struct pvr_load_op *load_op;
 };
 
 struct pvr_renderpass_hw_map {
@@ -206,13 +279,18 @@ struct pvr_renderpass_hwsetup {
     * that render where the subpass is scheduled.
     */
    struct pvr_renderpass_hw_map *subpass_map;
+
+   bool *surface_allocate;
 };
 
-struct pvr_renderpass_hwsetup *
-pvr_create_renderpass_hwsetup(struct pvr_device *device,
-                              struct pvr_render_pass *pass,
-                              bool disable_merge);
-void pvr_destroy_renderpass_hwsetup(struct pvr_device *device,
+VkResult pvr_create_renderpass_hwsetup(
+   struct pvr_device *device,
+   const VkAllocationCallbacks *alloc,
+   struct pvr_render_pass *pass,
+   bool disable_merge,
+   struct pvr_renderpass_hwsetup **const hw_setup_out);
+
+void pvr_destroy_renderpass_hwsetup(const VkAllocationCallbacks *alloc,
                                     struct pvr_renderpass_hwsetup *hw_setup);
 
 #endif /* PVR_HW_PASS_H */

src/imagination/vulkan/pvr_limits.h

@@ -32,7 +32,7 @@
 #include "pvr_device_info.h"
 #include "util/u_math.h"
 
-#define PVR_MAX_COLOR_ATTACHMENTS 8U /* Number of PBE emit registers. */
+#define PVR_MAX_COLOR_ATTACHMENTS PVR_NUM_PBE_EMIT_REGS
 #define PVR_MAX_QUEUES 2U
 #define PVR_MAX_VIEWPORTS 1U
 #define PVR_MAX_NEG_OFFSCREEN_OFFSET 4096U

src/imagination/vulkan/pvr_pass.c

@@ -499,11 +499,14 @@ VkResult pvr_CreateRenderPass2(VkDevice _device,
    pass->max_tilebuffer_count =
       PVR_SPM_LOAD_IN_BUFFERS_COUNT(&device->pdevice->dev_info);
 
-   pass->hw_setup = pvr_create_renderpass_hwsetup(device, pass, false);
-   if (!pass->hw_setup) {
-      result = vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+   result =
+      pvr_create_renderpass_hwsetup(device,
+                                    pAllocator ? pAllocator : &device->vk.alloc,
+                                    pass,
+                                    false,
+                                    &pass->hw_setup);
+   if (result != VK_SUCCESS)
       goto err_free_pass;
-   }
 
    pvr_init_subpass_userpass_spawn(pass->hw_setup, pass, pass->subpasses);
 
@@ -516,7 +519,7 @@ VkResult pvr_CreateRenderPass2(VkDevice _device,
          pvr_finishme("Set up tile buffer table");
 
       if (!hw_render->color_init_count) {
-         assert(!hw_render->client_data);
+         assert(!hw_render->load_op);
          continue;
       }
 
@@ -527,7 +530,7 @@ VkResult pvr_CreateRenderPass2(VkDevice _device,
       if (result != VK_SUCCESS)
          goto err_load_op_destroy;
 
-      hw_render->client_data = load_op;
+      hw_render->load_op = load_op;
    }
 
    *pRenderPass = pvr_render_pass_to_handle(pass);
@@ -539,11 +542,12 @@ err_load_op_destroy:
       struct pvr_renderpass_hwsetup_render *hw_render =
          &pass->hw_setup->renders[i];
 
-      if (hw_render->client_data)
-         pvr_load_op_destroy(device, pAllocator, hw_render->client_data);
+      if (hw_render->load_op)
+         pvr_load_op_destroy(device, pAllocator, hw_render->load_op);
    }
 
-   pvr_destroy_renderpass_hwsetup(device, pass->hw_setup);
+   pvr_destroy_renderpass_hwsetup(pAllocator ? pAllocator : &device->vk.alloc,
+                                  pass->hw_setup);
 
 err_free_pass:
    vk_object_base_finish(&pass->base);
@@ -566,10 +570,11 @@ void pvr_DestroyRenderPass(VkDevice _device,
       struct pvr_renderpass_hwsetup_render *hw_render =
          &pass->hw_setup->renders[i];
 
-      pvr_load_op_destroy(device, pAllocator, hw_render->client_data);
+      pvr_load_op_destroy(device, pAllocator, hw_render->load_op);
    }
 
-   pvr_destroy_renderpass_hwsetup(device, pass->hw_setup);
+   pvr_destroy_renderpass_hwsetup(pAllocator ? pAllocator : &device->vk.alloc,
+                                  pass->hw_setup);
    vk_object_base_finish(&pass->base);
    vk_free2(&device->vk.alloc, pAllocator, pass);
 }