radv: implement wsi's private transfer queue using SDMA

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13959>

src/amd/vulkan/radv_device.c

@@ -2635,9 +2635,9 @@ radv_get_queue_global_priority(const VkDeviceQueueGlobalPriorityCreateInfoEXT *p
    }
 }
 
-static int
-radv_queue_init(struct radv_device *device, struct radv_queue *queue,
-                int idx, const VkDeviceQueueCreateInfo *create_info,
+int
+radv_queue_init(struct radv_device *device, struct radv_queue *queue, int idx,
+                const VkDeviceQueueCreateInfo *create_info,
                 const VkDeviceQueueGlobalPriorityCreateInfoEXT *global_priority)
 {
    queue->device = device;
@@ -3105,6 +3105,7 @@ radv_CreateDevice(VkPhysicalDevice physicalDevice, const VkDeviceCreateInfo *pCr
             goto fail;
       }
    }
+   device->private_sdma_queue = VK_NULL_HANDLE;
 
    device->pbb_allowed = device->physical_device->rad_info.chip_class >= GFX9 &&
                          !(device->instance->debug_flags & RADV_DEBUG_NOBINNING);
@@ -3331,6 +3332,10 @@ radv_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
       if (device->queue_count[i])
          vk_free(&device->vk.alloc, device->queues[i]);
    }
+   if (device->private_sdma_queue != VK_NULL_HANDLE) {
+      radv_queue_finish(device->private_sdma_queue);
+      vk_free(&device->vk.alloc, device->private_sdma_queue);
+   }
 
    for (unsigned i = 0; i < RADV_NUM_HW_CTX; i++) {
       if (device->hw_ctx[i])

src/amd/vulkan/radv_image.c

@@ -531,6 +531,11 @@ radv_patch_image_from_extra_info(struct radv_device *device, struct radv_image *
 
          image->info.surf_index = NULL;
       }
+
+      if (create_info->prime_blit_src && device->physical_device->rad_info.chip_class == GFX9) {
+         /* Older SDMA hw can't handle DCC */
+         image->planes[plane].surface.flags |= RADEON_SURF_DISABLE_DCC;
+      }
    }
    return VK_SUCCESS;
 }
@@ -2310,11 +2315,13 @@ radv_CreateImage(VkDevice device, const VkImageCreateInfo *pCreateInfo,
    const struct wsi_image_create_info *wsi_info =
       vk_find_struct_const(pCreateInfo->pNext, WSI_IMAGE_CREATE_INFO_MESA);
    bool scanout = wsi_info && wsi_info->scanout;
+   bool prime_blit_src = wsi_info && wsi_info->prime_blit_src;
 
    return radv_image_create(device,
                             &(struct radv_image_create_info){
                                .vk_info = pCreateInfo,
                                .scanout = scanout,
+                               .prime_blit_src = prime_blit_src,
                             },
                             pAllocator, pImage);
 }

src/amd/vulkan/radv_private.h

@@ -858,6 +858,9 @@ struct radv_device {
    struct u_rwlock vs_prologs_lock;
    struct hash_table *vs_prologs;
 
+   /* Prime blit sdma queue */
+   struct radv_queue *private_sdma_queue;
+
    struct radv_shader_prolog *simple_vs_prologs[MAX_VERTEX_ATTRIBS];
    struct radv_shader_prolog *instance_rate_vs_prologs[816];
 };
@@ -2329,6 +2332,7 @@ struct radv_image_create_info {
    const VkImageCreateInfo *vk_info;
    bool scanout;
    bool no_metadata_planes;
+   bool prime_blit_src;
    const struct radeon_bo_metadata *bo_metadata;
 };
 
@@ -2541,6 +2545,10 @@ struct radv_query_pool {
 
 bool radv_queue_internal_submit(struct radv_queue *queue, struct radeon_cmdbuf *cs);
 
+int radv_queue_init(struct radv_device *device, struct radv_queue *queue, int idx,
+                    const VkDeviceQueueCreateInfo *create_info,
+                    const VkDeviceQueueGlobalPriorityCreateInfoEXT *global_priority);
+
 void radv_set_descriptor_set(struct radv_cmd_buffer *cmd_buffer, VkPipelineBindPoint bind_point,
                              struct radv_descriptor_set *set, unsigned idx);
 

src/amd/vulkan/radv_wsi.c

@@ -24,6 +24,7 @@
  */
 
 #include "util/macros.h"
+#include "radv_debug.h"
 #include "radv_meta.h"
 #include "radv_private.h"
 #include "vk_fence.h"
@@ -49,6 +50,39 @@ radv_wsi_set_memory_ownership(VkDevice _device, VkDeviceMemory _mem, VkBool32 ow
    }
 }
 
+static VkQueue
+radv_wsi_get_prime_blit_queue(VkDevice _device)
+{
+   RADV_FROM_HANDLE(radv_device, device, _device);
+
+   if (device->private_sdma_queue != VK_NULL_HANDLE)
+      return vk_queue_to_handle(&device->private_sdma_queue->vk);
+
+   if (device->physical_device->rad_info.chip_class >= GFX9 &&
+       !(device->physical_device->instance->debug_flags & RADV_DEBUG_NO_DMA_BLIT)) {
+      const VkDeviceQueueCreateInfo queue_create = {
+         .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
+         .queueFamilyIndex = RADV_QUEUE_TRANSFER,
+         .queueCount = 1,
+      };
+      device->private_sdma_queue = vk_zalloc(&device->vk.alloc, sizeof(struct radv_queue), 8,
+                                             VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
+
+      VkResult result = radv_queue_init(device, device->private_sdma_queue, 0, &queue_create, NULL);
+      if (result == VK_SUCCESS) {
+         /* Remove the queue from our queue list because it'll be cleared manually
+          * in radv_DestroyDevice.
+          */
+         list_delinit(&device->private_sdma_queue->vk.link);
+         return vk_queue_to_handle(&device->private_sdma_queue->vk);
+      } else {
+         vk_free(&device->vk.alloc, device->private_sdma_queue);
+         device->private_sdma_queue = VK_NULL_HANDLE;
+      }
+   }
+   return VK_NULL_HANDLE;
+}
+
 VkResult
 radv_init_wsi(struct radv_physical_device *physical_device)
 {
@@ -61,6 +95,7 @@ radv_init_wsi(struct radv_physical_device *physical_device)
 
    physical_device->wsi_device.supports_modifiers = physical_device->rad_info.chip_class >= GFX9;
    physical_device->wsi_device.set_memory_ownership = radv_wsi_set_memory_ownership;
+   physical_device->wsi_device.get_prime_blit_queue = radv_wsi_get_prime_blit_queue;
    physical_device->wsi_device.signal_semaphore_with_memory = true;
    physical_device->wsi_device.signal_fence_with_memory = true;