radv: Fix assert that is too strict.
[mesa.git] / src / amd / vulkan / radv_image.c
index fdb8177abbc2f4395af7e13b891a16eea6460f27..fc0fb81175e50a19a8faf78bb4de973ecde47375 100644 (file)
@@ -98,7 +98,7 @@ radv_use_tc_compat_htile_for_image(struct radv_device *device,
        if (pCreateInfo->samples >= 2 &&
            (format == VK_FORMAT_D32_SFLOAT_S8_UINT ||
             (format == VK_FORMAT_D32_SFLOAT &&
-             device->physical_device->rad_info.chip_class == GFX10)))
+             device->physical_device->rad_info.chip_class >= GFX10)))
                return false;
 
        /* GFX9 supports both 32-bit and 16-bit depth surfaces, while GFX8 only
@@ -257,6 +257,13 @@ radv_use_fmask_for_image(const struct radv_image *image)
               image->usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
 }
 
+static inline bool
+radv_use_htile_for_image(const struct radv_image *image)
+{
+       return image->info.levels == 1 &&
+              image->info.width * image->info.height >= 8 * 8;
+}
+
 static bool
 radv_use_tc_compat_cmask_for_image(struct radv_device *device,
                                   struct radv_image *image)
@@ -472,6 +479,9 @@ radv_init_surface(struct radv_device *device,
 
        if (is_depth) {
                surface->flags |= RADEON_SURF_ZBUFFER;
+               if (!radv_use_htile_for_image(image) ||
+                   (device->instance->debug_flags & RADV_DEBUG_NO_HIZ))
+                       surface->flags |= RADEON_SURF_NO_HTILE;
                if (radv_use_tc_compat_htile_for_image(device, pCreateInfo, image_format))
                        surface->flags |= RADEON_SURF_TC_COMPATIBLE_HTILE;
        }
@@ -1220,61 +1230,38 @@ radv_image_override_offset_stride(struct radv_device *device,
 }
 
 static void
-radv_image_alloc_fmask(struct radv_device *device,
-                      struct radv_image *image)
+radv_image_alloc_single_sample_cmask(const struct radv_image *image,
+                                     struct radeon_surf *surf)
 {
-       unsigned fmask_alignment = image->planes[0].surface.fmask_alignment;
-
-       image->planes[0].surface.fmask_offset = align64(image->size, fmask_alignment);
-       image->size = image->planes[0].surface.fmask_offset + image->planes[0].surface.fmask_size;
-       image->alignment = MAX2(image->alignment, fmask_alignment);
-}
-
-static void
-radv_image_alloc_cmask(struct radv_device *device,
-                      struct radv_image *image)
-{
-       unsigned cmask_alignment = image->planes[0].surface.cmask_alignment;
-       unsigned cmask_size = image->planes[0].surface.cmask_size;
-       uint32_t clear_value_size = 0;
-
-       if (!cmask_size)
+       if (!surf->cmask_size || surf->cmask_offset || surf->bpe > 8 ||
+           image->info.levels > 1 || image->info.depth > 1 ||
+           radv_image_has_dcc(image) || !radv_image_use_fast_clear_for_image(image))
                return;
 
-       assert(cmask_alignment);
+       assert(image->info.storage_samples == 1);
 
-       image->planes[0].surface.cmask_offset = align64(image->size, cmask_alignment);
-       /* + 8 for storing the clear values */
-       if (!image->clear_value_offset) {
-               image->clear_value_offset = image->planes[0].surface.cmask_offset + cmask_size;
-               clear_value_size = 8;
-       }
-       image->size = image->planes[0].surface.cmask_offset + cmask_size + clear_value_size;
-       image->alignment = MAX2(image->alignment, cmask_alignment);
+       surf->cmask_offset = align64(surf->total_size, surf->cmask_alignment);
+       surf->total_size = surf->cmask_offset + surf->cmask_size;
+       surf->alignment = MAX2(surf->alignment, surf->cmask_alignment);
 }
 
 static void
-radv_image_alloc_dcc(struct radv_image *image)
+radv_image_alloc_values(const struct radv_device *device, struct radv_image *image)
 {
-       assert(image->plane_count == 1);
+       if (radv_image_has_dcc(image)) {
+               image->fce_pred_offset = image->size;
+               image->size += 8 * image->info.levels;
 
-       image->planes[0].surface.dcc_offset = align64(image->size, image->planes[0].surface.dcc_alignment);
-       /* + 24 for storing the clear values + fce pred + dcc pred for each mip */
-       image->clear_value_offset = image->planes[0].surface.dcc_offset + image->planes[0].surface.dcc_size;
-       image->fce_pred_offset = image->clear_value_offset + 8 * image->info.levels;
-       image->dcc_pred_offset = image->clear_value_offset + 16 * image->info.levels;
-       image->size = image->planes[0].surface.dcc_offset + image->planes[0].surface.dcc_size + 24 * image->info.levels;
-       image->alignment = MAX2(image->alignment, image->planes[0].surface.dcc_alignment);
-}
+               image->dcc_pred_offset = image->size;
+               image->size += 8 * image->info.levels;
+       }
 
-static void
-radv_image_alloc_htile(struct radv_device *device, struct radv_image *image)
-{
-       image->planes[0].surface.htile_offset = align64(image->size, image->planes[0].surface.htile_alignment);
+       if (radv_image_has_dcc(image) || radv_image_has_cmask(image) ||
+           radv_image_has_htile(image)) {
+               image->clear_value_offset = image->size;
+               image->size += 8 * image->info.levels;
+       }
 
-       /* + 8 for storing the clear values */
-       image->clear_value_offset = image->planes[0].surface.htile_offset + image->planes[0].surface.htile_size;
-       image->size = image->clear_value_offset + image->info.levels * 8;
        if (radv_image_is_tc_compat_htile(image) &&
            device->physical_device->rad_info.has_tc_compat_zrange_bug) {
                /* Metadata for the TC-compatible HTILE hardware bug which
@@ -1282,45 +1269,8 @@ radv_image_alloc_htile(struct radv_device *device, struct radv_image *image)
                 * fast depth clears to 0.0f.
                 */
                image->tc_compat_zrange_offset = image->size;
-               image->size = image->tc_compat_zrange_offset + image->info.levels * 4;
-       }
-       image->alignment = align64(image->alignment, image->planes[0].surface.htile_alignment);
-}
-
-static inline bool
-radv_image_can_enable_cmask(struct radv_image *image)
-{
-       if (image->planes[0].surface.bpe > 8 && image->info.samples == 1) {
-               /* Do not enable CMASK for non-MSAA images (fast color clear)
-                * because 128 bit formats are not supported, but FMASK might
-                * still be used.
-                */
-               return false;
+               image->size += image->info.levels * 4;
        }
-
-       return radv_image_use_fast_clear_for_image(image) &&
-              image->info.levels == 1 &&
-              image->info.depth == 1;
-}
-
-static inline bool
-radv_image_can_enable_htile(struct radv_image *image)
-{
-       return radv_image_has_htile(image) &&
-              image->info.levels == 1 &&
-              image->info.width * image->info.height >= 8 * 8;
-}
-
-static void radv_image_disable_dcc(struct radv_image *image)
-{
-       for (unsigned i = 0; i < image->plane_count; ++i)
-               image->planes[i].surface.dcc_size = 0;
-}
-
-static void radv_image_disable_htile(struct radv_image *image)
-{
-       for (unsigned i = 0; i < image->plane_count; ++i)
-               image->planes[i].surface.htile_size = 0;
 }
 
 VkResult
@@ -1362,52 +1312,45 @@ radv_image_create_layout(struct radv_device *device,
 
                device->ws->surface_init(device->ws, &info, &image->planes[plane].surface);
 
-               image->planes[plane].offset = align(image->size, image->planes[plane].surface.surf_alignment);
-               image->size = image->planes[plane].offset + image->planes[plane].surface.surf_size;
-               image->alignment = image->planes[plane].surface.surf_alignment;
+               if (!create_info.no_metadata_planes && image->plane_count == 1)
+                       radv_image_alloc_single_sample_cmask(image, &image->planes[plane].surface);
+
+               image->planes[plane].offset = align(image->size, image->planes[plane].surface.alignment);
+               image->size = image->planes[plane].offset + image->planes[plane].surface.total_size;
+               image->alignment = MAX2(image->alignment, image->planes[plane].surface.alignment);
 
                image->planes[plane].format = vk_format_get_plane_format(image->vk_format, plane);
        }
 
-       /* Try to enable DCC first. */
-       if (radv_image_has_dcc(image)) {
-               radv_image_alloc_dcc(image);
-               if (image->info.samples > 1) {
-                       /* CMASK should be enabled because DCC fast
-                        * clear with MSAA needs it.
-                        */
-                       assert(radv_image_can_enable_cmask(image));
-                       radv_image_alloc_cmask(device, image);
-               }
-       } else {
-               /* When DCC cannot be enabled, try CMASK. */
-               radv_image_disable_dcc(image);
-               if (radv_image_can_enable_cmask(image)) {
-                       radv_image_alloc_cmask(device, image);
-               }
-       }
+       image->tc_compatible_cmask = radv_image_has_cmask(image) &&
+                                    radv_use_tc_compat_cmask_for_image(device, image);
 
-       /* Try to enable FMASK for multisampled images. */
-       if (image->planes[0].surface.fmask_size) {
-               radv_image_alloc_fmask(device, image);
+       image->tc_compatible_htile = radv_image_has_htile(image) &&
+                                    image->planes[0].surface.flags & RADEON_SURF_TC_COMPATIBLE_HTILE;
 
-               if (radv_use_tc_compat_cmask_for_image(device, image))
-                       image->tc_compatible_cmask = true;
-       } else {
-               /* Otherwise, try to enable HTILE for depth surfaces. */
-               if (radv_image_can_enable_htile(image) &&
-                   !(device->instance->debug_flags & RADV_DEBUG_NO_HIZ)) {
-                       image->tc_compatible_htile = image->planes[0].surface.flags & RADEON_SURF_TC_COMPATIBLE_HTILE;
-                       radv_image_alloc_htile(device, image);
-               } else {
-                       radv_image_disable_htile(image);
-               }
-       }
+       radv_image_alloc_values(device, image);
 
        assert(image->planes[0].surface.surf_size);
        return VK_SUCCESS;
 }
 
+static void
+radv_destroy_image(struct radv_device *device,
+                  const VkAllocationCallbacks *pAllocator,
+                  struct radv_image *image)
+{
+       if ((image->flags & VK_IMAGE_CREATE_SPARSE_BINDING_BIT) && image->bo)
+               device->ws->buffer_destroy(image->bo);
+
+       if (image->owned_memory != VK_NULL_HANDLE) {
+               RADV_FROM_HANDLE(radv_device_memory, mem, image->owned_memory);
+               radv_free_memory(device, pAllocator, mem);
+       }
+
+       vk_object_base_finish(&image->base);
+       vk_free2(&device->vk.alloc, pAllocator, image);
+}
+
 VkResult
 radv_image_create(VkDevice _device,
                  const struct radv_image_create_info *create_info,
@@ -1497,7 +1440,7 @@ radv_image_create(VkDevice _device,
                image->bo = device->ws->buffer_create(device->ws, image->size, image->alignment,
                                                      0, RADEON_FLAG_VIRTUAL, RADV_BO_PRIORITY_VIRTUAL);
                if (!image->bo) {
-                       vk_free2(&device->vk.alloc, alloc, image);
+                       radv_destroy_image(device, alloc, image);
                        return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
                }
        }
@@ -1827,14 +1770,7 @@ radv_DestroyImage(VkDevice _device, VkImage _image,
        if (!image)
                return;
 
-       if (image->flags & VK_IMAGE_CREATE_SPARSE_BINDING_BIT)
-               device->ws->buffer_destroy(image->bo);
-
-       if (image->owned_memory != VK_NULL_HANDLE)
-               radv_FreeMemory(_device, image->owned_memory, pAllocator);
-
-       vk_object_base_finish(&image->base);
-       vk_free2(&device->vk.alloc, pAllocator, image);
+       radv_destroy_image(device, pAllocator, image);
 }
 
 void radv_GetImageSubresourceLayout(