radv: add support for NV_dedicated_allocation

This adds initial support for NV_dedicated_allocation, then
uses it for the wsi image/memory allocation paths internally
in the driver.

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Signed-off-by: Dave Airlie <airlied@redhat.com>

src/amd/vulkan/radv_device.c

@@ -33,7 +33,7 @@
 #include "radv_cs.h"
 #include "util/disk_cache.h"
 #include "util/strtod.h"
-
+#include "util/vk_util.h"
 #include <xf86drm.h>
 #include <amdgpu.h>
 #include <amdgpu_drm.h>
@@ -111,6 +111,10 @@ static const VkExtensionProperties common_device_extensions[] = {
 		.extensionName = VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
 		.specVersion = 1,
 	},
+	{
+		.extensionName = VK_NV_DEDICATED_ALLOCATION_EXTENSION_NAME,
+		.specVersion = 1,
+	},
 };
 
 static VkResult
@@ -1635,6 +1639,21 @@ PFN_vkVoidFunction radv_GetDeviceProcAddr(
 	return radv_lookup_entrypoint(pName);
 }
 
+bool radv_get_memory_fd(struct radv_device *device,
+			struct radv_device_memory *memory,
+			int *pFD)
+{
+	struct radeon_bo_metadata metadata;
+
+	if (memory->image) {
+		radv_init_metadata(device, memory->image, &metadata);
+		device->ws->buffer_set_metadata(memory->bo, &metadata);
+	}
+
+	return device->ws->buffer_get_fd(device->ws, memory->bo,
+					 pFD);
+}
+
 VkResult radv_AllocateMemory(
 	VkDevice                                    _device,
 	const VkMemoryAllocateInfo*                 pAllocateInfo,
@@ -1646,6 +1665,7 @@ VkResult radv_AllocateMemory(
 	VkResult result;
 	enum radeon_bo_domain domain;
 	uint32_t flags = 0;
+	const VkDedicatedAllocationMemoryAllocateInfoNV *dedicate_info = NULL;
 	assert(pAllocateInfo->sType == VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO);
 
 	if (pAllocateInfo->allocationSize == 0) {
@@ -1654,11 +1674,29 @@ VkResult radv_AllocateMemory(
 		return VK_SUCCESS;
 	}
 
+	vk_foreach_struct(ext, pAllocateInfo->pNext) {
+		switch (ext->sType) {
+		case VK_STRUCTURE_TYPE_DEDICATED_ALLOCATION_MEMORY_ALLOCATE_INFO_NV:
+			dedicate_info = (const VkDedicatedAllocationMemoryAllocateInfoNV *)ext;
+			break;
+		default:
+			break;
+		}
+	}
+
 	mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
 			  VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 	if (mem == NULL)
 		return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 
+	if (dedicate_info) {
+		mem->image = radv_image_from_handle(dedicate_info->image);
+		mem->buffer = radv_buffer_from_handle(dedicate_info->buffer);
+	} else {
+		mem->image = NULL;
+		mem->buffer = NULL;
+	}
+
 	uint64_t alloc_size = align_u64(pAllocateInfo->allocationSize, 4096);
 	if (pAllocateInfo->memoryTypeIndex == RADV_MEM_TYPE_GTT_WRITE_COMBINE ||
 	    pAllocateInfo->memoryTypeIndex == RADV_MEM_TYPE_GTT_CACHED)

src/amd/vulkan/radv_entrypoints_gen.py

@@ -29,6 +29,7 @@ max_api_version = 1.0
 
 supported_extensions = [
    'VK_AMD_draw_indirect_count',
+   'VK_NV_dedicated_allocation',
    'VK_KHR_get_physical_device_properties2',
    'VK_KHR_maintenance1',
    'VK_KHR_sampler_mirror_clamp_to_edge',

src/amd/vulkan/radv_private.h

@@ -523,6 +523,9 @@ struct radv_device {
 
 struct radv_device_memory {
 	struct radeon_winsys_bo                      *bo;
+	/* for dedicated allocations */
+	struct radv_image                            *image;
+	struct radv_buffer                           *buffer;
 	uint32_t                                     type_index;
 	VkDeviceSize                                 map_size;
 	void *                                       map;
@@ -814,7 +817,9 @@ void radv_fill_buffer(struct radv_cmd_buffer *cmd_buffer,
 		      struct radeon_winsys_bo *bo,
 		      uint64_t offset, uint64_t size, uint32_t value);
 void radv_cmd_buffer_trace_emit(struct radv_cmd_buffer *cmd_buffer);
-
+bool radv_get_memory_fd(struct radv_device *device,
+			struct radv_device_memory *memory,
+			int *pFD);
 /*
  * Takes x,y,z as exact numbers of invocations, instead of blocks.
  *

src/amd/vulkan/radv_wsi.c

@@ -148,12 +148,10 @@ radv_wsi_image_create(VkDevice device_h,
 		      uint32_t *offset,
 		      uint32_t *row_pitch, int *fd_p)
 {
-	struct radv_device *device = radv_device_from_handle(device_h);
 	VkResult result = VK_SUCCESS;
 	struct radeon_surf *surface;
 	VkImage image_h;
 	struct radv_image *image;
-	bool bret;
 	int fd;
 
 	result = radv_image_create(device_h,
@@ -183,42 +181,42 @@ radv_wsi_image_create(VkDevice device_h,
 		return result;
 
 	image = radv_image_from_handle(image_h);
+
 	VkDeviceMemory memory_h;
-	struct radv_device_memory *memory;
+
+	const VkDedicatedAllocationMemoryAllocateInfoNV ded_alloc = {
+		.sType = VK_STRUCTURE_TYPE_DEDICATED_ALLOCATION_MEMORY_ALLOCATE_INFO_NV,
+		.pNext = NULL,
+		.buffer = NULL,
+		.image = image_h
+	};
 
 	result = radv_AllocateMemory(device_h,
 				     &(VkMemoryAllocateInfo) {
 					     .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
-						     .allocationSize = image->size,
-						     .memoryTypeIndex = linear ? 1 : 0,
-						     },
+					     .pNext = &ded_alloc,
+					     .allocationSize = image->size,
+					     .memoryTypeIndex = linear ? 1 : 0,
+				     },
 				     NULL /* XXX: pAllocator */,
 				     &memory_h);
 	if (result != VK_SUCCESS)
 		goto fail_create_image;
 
-	memory = radv_device_memory_from_handle(memory_h);
-
-	radv_BindImageMemory(VK_NULL_HANDLE, image_h, memory_h, 0);
+	radv_BindImageMemory(device_h, image_h, memory_h, 0);
 
 	/*
 	 * return the fd for the image in the no copy mode,
 	 * or the fd for the linear image if a copy is required.
 	 */
 	if (!needs_linear_copy || (needs_linear_copy && linear)) {
-		bret = device->ws->buffer_get_fd(device->ws,
-						 memory->bo, &fd);
-		if (bret == false)
+		RADV_FROM_HANDLE(radv_device, device, device_h);
+		RADV_FROM_HANDLE(radv_device_memory, memory, memory_h);
+		if (!radv_get_memory_fd(device, memory, &fd))
 			goto fail_alloc_memory;
 		*fd_p = fd;
 	}
 
-	{
-		struct radeon_bo_metadata metadata;
-		radv_init_metadata(device, image, &metadata);
-		device->ws->buffer_set_metadata(memory->bo, &metadata);
-	}
-
 	surface = &image->surface;
 
 	*image_p = image_h;