radv: add initial DCC support on GFX12

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33255>

src/amd/vulkan/radv_device_memory.c

@@ -11,6 +11,7 @@
 #include "radv_device_memory.h"
 #include "radv_android.h"
 #include "radv_buffer.h"
+#include "radv_debug.h"
 #include "radv_entrypoints.h"
 #include "radv_image.h"
 #include "radv_rmv.h"
@@ -201,6 +202,17 @@ radv_alloc_memory(struct radv_device *device, const VkMemoryAllocateInfo *pAlloc
       if (instance->drirc.zero_vram)
          flags |= RADEON_FLAG_ZERO_VRAM;
 
+      /* On GFX12, DCC is transparent to the userspace driver and PTE.DCC is
+       * set per buffer allocation. Only VRAM can have DCC. When the kernel
+       * moves a buffer from VRAM->GTT it decompresses. When the kernel moves
+       * it from GTT->VRAM it recompresses but only if WRITE_COMPRESS_DISABLE=0
+       * (see DCC tiling flags).
+       */
+      if (pdev->info.gfx_level >= GFX12 && pdev->info.gfx12_supports_dcc_write_compress_disable &&
+          domain == RADEON_DOMAIN_VRAM && !(instance->debug_flags & RADV_DEBUG_NO_DCC)) {
+         flags |= RADEON_FLAG_GFX12_ALLOW_DCC;
+      }
+
       if (device->overallocation_disallowed) {
          uint64_t total_size = pdev->memory_properties.memoryHeaps[heap_index].size;
 
@@ -226,6 +238,28 @@ radv_alloc_memory(struct radv_device *device, const VkMemoryAllocateInfo *pAlloc
          goto fail;
       }
 
+      if (flags & RADEON_FLAG_GFX12_ALLOW_DCC) {
+         if (mem->image) {
+            /* Set BO metadata (including DCC tiling flags) for dedicated
+             * allocations because compressed writes are enabled and the kernel
+             * requires a DCC view for recompression.
+             */
+            radv_image_bo_set_metadata(device, mem->image, mem->bo);
+         } else {
+            /* Otherwise, disable compressed writes to prevent recompression
+             * when the BO is moved back to VRAM because it's not yet possible
+             * to set DCC tiling flags per range for suballocations. The only
+             * problem is that we will loose DCC after migration but that
+             * should happen rarely.
+             */
+            struct radeon_bo_metadata md = {0};
+
+            md.u.gfx12.dcc_write_compress_disable = true;
+
+            device->ws->buffer_set_metadata(device->ws, mem->bo, &md);
+         }
+      }
+
       mem->heap_index = heap_index;
       mem->alloc_size = alloc_size;
    }

src/amd/vulkan/radv_image.c

@@ -1206,6 +1206,16 @@ radv_image_create_layout(struct radv_device *device, struct radv_image_create_in
             ac_surface_zero_dcc_fields(&image->planes[0].surface);
       }
 
+      if (pdev->info.gfx_level >= GFX12 &&
+          (!radv_surface_has_scanout(device, &create_info) || pdev->info.gfx12_supports_display_dcc)) {
+         const enum pipe_format format = vk_format_to_pipe_format(image->vk.format);
+
+         /* Set DCC tilings for both color and depth/stencil. */
+         image->planes[plane].surface.u.gfx9.color.dcc_number_type = ac_get_cb_number_type(format);
+         image->planes[plane].surface.u.gfx9.color.dcc_data_format = ac_get_cb_format(pdev->info.gfx_level, format);
+         image->planes[plane].surface.u.gfx9.color.dcc_write_compress_disable = false;
+      }
+
       if (create_info.bo_metadata && !mod_info &&
           !ac_surface_apply_umd_metadata(&pdev->info, &image->planes[plane].surface, image->vk.samples,
                                          image->vk.mip_levels, create_info.bo_metadata->size_metadata,

src/amd/vulkan/radv_image_view.c

@@ -55,19 +55,19 @@ radv_set_mutable_tex_desc_fields(struct radv_device *device, struct radv_image *
                                  bool is_storage_image, bool disable_compression, bool enable_write_compression,
                                  uint32_t *state, const struct ac_surf_nbc_view *nbc_view, uint64_t offset)
 {
+   const struct radv_physical_device *pdev = radv_device_physical(device);
    struct radv_image_plane *plane = &image->planes[plane_id];
    const uint32_t bind_idx = image->disjoint ? plane_id : 0;
    struct radv_image_binding *binding = &image->bindings[bind_idx];
    uint64_t gpu_address = binding->bo ? radv_image_get_va(image, bind_idx) + offset : 0;
-   const struct radv_physical_device *pdev = radv_device_physical(device);
+   const bool dcc_enabled = pdev->info.gfx_level >= GFX12 || radv_dcc_enabled(image, first_level);
 
    const struct ac_mutable_tex_state ac_state = {
       .surf = &plane->surface,
       .va = gpu_address,
       .gfx10 =
          {
-            .write_compress_enable =
+            .write_compress_enable = dcc_enabled && is_storage_image && enable_write_compression,
-               radv_dcc_enabled(image, first_level) && is_storage_image && enable_write_compression,
             .iterate_256 = radv_image_get_iterate256(device, image),
          },
       .gfx9 =
@@ -81,7 +81,7 @@ radv_set_mutable_tex_desc_fields(struct radv_device *device, struct radv_image *
             .block_width = block_width,
          },
       .is_stencil = is_stencil,
-      .dcc_enabled = !disable_compression && radv_dcc_enabled(image, first_level),
+      .dcc_enabled = !disable_compression && dcc_enabled,
       .tc_compat_htile_enabled = !disable_compression && radv_image_is_tc_compat_htile(image),
    };