Greatly reduce VA allocations we have to make and makes returned VA more sensible, and better matches returned VAs we see on native drivers.
D3D12 usage flags for buffers seem generic enough that there is no obvious benefit to place smaller VkBuffers on top of VkDeviceMemory.
Ideally, physical_buffer_address is used here, but this works as a good fallback if that path is added later.
With this patch and previous VA optimization, I'm observing a 2.0-2.5% FPS uplift on SOTTR when CPU bound.
Signed-off-by: Hans-Kristian Arntzen post@arntzen-software.no --- libs/vkd3d/command.c | 14 ++-- libs/vkd3d/resource.c | 133 ++++++++++++++++++++++++++++++++----- libs/vkd3d/vkd3d_private.h | 2 + 3 files changed, 128 insertions(+), 21 deletions(-)
diff --git a/libs/vkd3d/command.c b/libs/vkd3d/command.c index 8a7ff66..8bff7ea 100644 --- a/libs/vkd3d/command.c +++ b/libs/vkd3d/command.c @@ -3081,8 +3081,8 @@ static void STDMETHODCALLTYPE d3d12_command_list_CopyBufferRegion(ID3D12Graphics
d3d12_command_list_end_current_render_pass(list);
- buffer_copy.srcOffset = src_offset; - buffer_copy.dstOffset = dst_offset; + buffer_copy.srcOffset = src_offset + src_resource->heap_offset; + buffer_copy.dstOffset = dst_offset + dst_resource->heap_offset; buffer_copy.size = byte_count;
VK_CALL(vkCmdCopyBuffer(list->vk_command_buffer, @@ -3395,6 +3395,7 @@ static void STDMETHODCALLTYPE d3d12_command_list_CopyTextureRegion(ID3D12Graphic
vk_image_buffer_copy_from_d3d12(&buffer_image_copy, &dst->u.PlacedFootprint, src->u.SubresourceIndex, &src_resource->desc, dst_format, src_box, dst_x, dst_y, dst_z); + buffer_image_copy.bufferOffset += dst_resource->heap_offset; VK_CALL(vkCmdCopyImageToBuffer(list->vk_command_buffer, src_resource->u.vk_image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, dst_resource->u.vk_buffer, 1, &buffer_image_copy)); @@ -3424,6 +3425,7 @@ static void STDMETHODCALLTYPE d3d12_command_list_CopyTextureRegion(ID3D12Graphic
vk_buffer_image_copy_from_d3d12(&buffer_image_copy, &src->u.PlacedFootprint, dst->u.SubresourceIndex, &dst_resource->desc, src_format, src_box, dst_x, dst_y, dst_z); + buffer_image_copy.bufferOffset += src_resource->heap_offset; VK_CALL(vkCmdCopyBufferToImage(list->vk_command_buffer, src_resource->u.vk_buffer, dst_resource->u.vk_image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &buffer_image_copy)); @@ -3504,8 +3506,8 @@ static void STDMETHODCALLTYPE d3d12_command_list_CopyResource(ID3D12GraphicsComm assert(d3d12_resource_is_buffer(src_resource)); assert(src_resource->desc.Width == dst_resource->desc.Width);
- vk_buffer_copy.srcOffset = 0; - vk_buffer_copy.dstOffset = 0; + vk_buffer_copy.srcOffset = src_resource->heap_offset; + vk_buffer_copy.dstOffset = dst_resource->heap_offset; vk_buffer_copy.size = dst_resource->desc.Width; VK_CALL(vkCmdCopyBuffer(list->vk_command_buffer, src_resource->u.vk_buffer, dst_resource->u.vk_buffer, 1, &vk_buffer_copy)); @@ -3962,8 +3964,8 @@ static void STDMETHODCALLTYPE d3d12_command_list_ResourceBarrier(ID3D12GraphicsC vk_barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED; vk_barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED; vk_barrier.buffer = resource->u.vk_buffer; - vk_barrier.offset = 0; - vk_barrier.size = VK_WHOLE_SIZE; + vk_barrier.offset = resource->heap_offset; + vk_barrier.size = resource->desc.Width;
VK_CALL(vkCmdPipelineBarrier(list->vk_command_buffer, src_stage_mask, dst_stage_mask, 0, 0, NULL, 1, &vk_barrier, 0, NULL)); diff --git a/libs/vkd3d/resource.c b/libs/vkd3d/resource.c index f40d986..ed2b18f 100644 --- a/libs/vkd3d/resource.c +++ b/libs/vkd3d/resource.c @@ -315,6 +315,8 @@ static ULONG STDMETHODCALLTYPE d3d12_heap_AddRef(ID3D12Heap *iface) return refcount; }
+static ULONG d3d12_resource_decref(struct d3d12_resource *resource); + static void d3d12_heap_destroy(struct d3d12_heap *heap) { struct d3d12_device *device = heap->device; @@ -322,6 +324,9 @@ static void d3d12_heap_destroy(struct d3d12_heap *heap)
TRACE("Destroying heap %p.\n", heap);
+ if (heap->buffer_resource) + d3d12_resource_decref(heap->buffer_resource); + vkd3d_private_store_destroy(&heap->private_store);
VK_CALL(vkFreeMemory(device->vk_device, heap->vk_memory, NULL)); @@ -562,6 +567,12 @@ static HRESULT validate_heap_desc(const D3D12_HEAP_DESC *desc, const struct d3d1 return S_OK; }
+static HRESULT d3d12_resource_create(struct d3d12_device *device, + const D3D12_HEAP_PROPERTIES *heap_properties, D3D12_HEAP_FLAGS heap_flags, + const D3D12_RESOURCE_DESC *desc, D3D12_RESOURCE_STATES initial_state, + const D3D12_CLEAR_VALUE *optimized_clear_value, bool placed, + struct d3d12_resource **resource); + static HRESULT d3d12_heap_init(struct d3d12_heap *heap, struct d3d12_device *device, const D3D12_HEAP_DESC *desc, const struct d3d12_resource *resource) { @@ -569,6 +580,9 @@ static HRESULT d3d12_heap_init(struct d3d12_heap *heap, VkDeviceSize vk_memory_size; HRESULT hr; int rc; + bool buffers_allowed; + D3D12_RESOURCE_DESC resource_desc; + D3D12_RESOURCE_STATES initial_resource_state;
heap->ID3D12Heap_iface.lpVtbl = &d3d12_heap_vtbl; heap->refcount = 1; @@ -579,6 +593,7 @@ static HRESULT d3d12_heap_init(struct d3d12_heap *heap,
heap->map_ptr = NULL; heap->map_count = 0; + heap->buffer_resource = NULL;
if (!heap->desc.Properties.CreationNodeMask) heap->desc.Properties.CreationNodeMask = 1; @@ -606,6 +621,53 @@ static HRESULT d3d12_heap_init(struct d3d12_heap *heap, return hr; }
+ buffers_allowed = !(heap->desc.Flags & D3D12_HEAP_FLAG_DENY_BUFFERS); + if (buffers_allowed && !resource) + { + /* Create a single omnipotent buffer which fills the entire heap. + * Whenever we place buffer resources on this heap, we'll just offset this VkBuffer. + * This allows us to keep VA space somewhat sane, and keeps number of (limited) VA allocations down. + * One possible downside is that the buffer might be slightly slower to access, + * but D3D12 has very lenient usage flags for buffers. */ + + memset(&resource_desc, 0, sizeof(resource_desc)); + resource_desc.Dimension = D3D12_RESOURCE_DIMENSION_BUFFER; + resource_desc.Width = desc->SizeInBytes; + resource_desc.Height = 1; + resource_desc.DepthOrArraySize = 1; + resource_desc.MipLevels = 1; + resource_desc.SampleDesc.Count = 1; + resource_desc.Layout = D3D12_TEXTURE_LAYOUT_ROW_MAJOR; + + switch (desc->Properties.Type) + { + case D3D12_HEAP_TYPE_UPLOAD: + initial_resource_state = D3D12_RESOURCE_STATE_GENERIC_READ; + break; + + case D3D12_HEAP_TYPE_READBACK: + initial_resource_state = D3D12_RESOURCE_STATE_COPY_DEST; + break; + + default: + /* Upload and readback heaps do not allow UAV access, only enable this flag for other heaps. */ + resource_desc.Flags |= D3D12_RESOURCE_FLAG_ALLOW_UNORDERED_ACCESS; + initial_resource_state = D3D12_RESOURCE_STATE_COMMON; + break; + } + + if (FAILED(hr = d3d12_resource_create(device, &desc->Properties, desc->Flags, + &resource_desc, initial_resource_state, + NULL, false, &heap->buffer_resource))) + { + heap->buffer_resource = NULL; + return hr; + } + /* This internal resource should not own a reference on the device. + * d3d12_resource_create takes a reference on the device. */ + d3d12_device_release(device); + } + if (resource) { if (d3d12_resource_is_buffer(resource)) @@ -623,8 +685,16 @@ static HRESULT d3d12_heap_init(struct d3d12_heap *heap,
heap->desc.SizeInBytes = vk_memory_size; } + else if (heap->buffer_resource) + { + hr = vkd3d_allocate_buffer_memory(device, heap->buffer_resource->u.vk_buffer, + &heap->desc.Properties, heap->desc.Flags, + &heap->vk_memory, &heap->vk_memory_type, &vk_memory_size); + } else { + /* Allocate generic memory which should hopefully match up with whatever resources + * we want to place here. */ memory_requirements.size = heap->desc.SizeInBytes; memory_requirements.alignment = heap->desc.Alignment; memory_requirements.memoryTypeBits = ~(uint32_t)0; @@ -637,6 +707,11 @@ static HRESULT d3d12_heap_init(struct d3d12_heap *heap, { vkd3d_private_store_destroy(&heap->private_store); pthread_mutex_destroy(&heap->mutex); + if (heap->buffer_resource) + { + d3d12_resource_decref(heap->buffer_resource); + heap->buffer_resource = NULL; + } return hr; }
@@ -1030,13 +1105,16 @@ static void d3d12_resource_destroy(struct d3d12_resource *resource, struct d3d12 if (resource->flags & VKD3D_RESOURCE_EXTERNAL) return;
- if (resource->gpu_address) - vkd3d_gpu_va_allocator_free(&device->gpu_va_allocator, resource->gpu_address); + if (!(resource->flags & VKD3D_RESOURCE_PLACED_BUFFER)) + { + if (resource->gpu_address) + vkd3d_gpu_va_allocator_free(&device->gpu_va_allocator, resource->gpu_address);
- if (d3d12_resource_is_buffer(resource)) - VK_CALL(vkDestroyBuffer(device->vk_device, resource->u.vk_buffer, NULL)); - else - VK_CALL(vkDestroyImage(device->vk_device, resource->u.vk_image, NULL)); + if (d3d12_resource_is_buffer(resource)) + VK_CALL(vkDestroyBuffer(device->vk_device, resource->u.vk_buffer, NULL)); + else + VK_CALL(vkDestroyImage(device->vk_device, resource->u.vk_image, NULL)); + }
if (resource->flags & VKD3D_RESOURCE_DEDICATED_HEAP) d3d12_heap_destroy(resource->heap); @@ -1738,7 +1816,7 @@ static bool d3d12_resource_validate_heap_properties(const struct d3d12_resource static HRESULT d3d12_resource_init(struct d3d12_resource *resource, struct d3d12_device *device, const D3D12_HEAP_PROPERTIES *heap_properties, D3D12_HEAP_FLAGS heap_flags, const D3D12_RESOURCE_DESC *desc, D3D12_RESOURCE_STATES initial_state, - const D3D12_CLEAR_VALUE *optimized_clear_value) + const D3D12_CLEAR_VALUE *optimized_clear_value, bool placed) { HRESULT hr;
@@ -1768,6 +1846,8 @@ static HRESULT d3d12_resource_init(struct d3d12_resource *resource, struct d3d12
resource->gpu_address = 0; resource->flags = 0; + if (placed && d3d12_resource_is_buffer(resource)) + resource->flags |= VKD3D_RESOURCE_PLACED_BUFFER;
if (FAILED(hr = d3d12_resource_validate_desc(&resource->desc, device))) return hr; @@ -1775,6 +1855,13 @@ static HRESULT d3d12_resource_init(struct d3d12_resource *resource, struct d3d12 switch (desc->Dimension) { case D3D12_RESOURCE_DIMENSION_BUFFER: + /* We'll inherit a VkBuffer reference from the heap with an implied offset. */ + if (placed) + { + resource->u.vk_buffer = VK_NULL_HANDLE; + break; + } + if (FAILED(hr = vkd3d_create_buffer(device, heap_properties, heap_flags, &resource->desc, &resource->u.vk_buffer))) return hr; @@ -1825,7 +1912,7 @@ static HRESULT d3d12_resource_init(struct d3d12_resource *resource, struct d3d12 static HRESULT d3d12_resource_create(struct d3d12_device *device, const D3D12_HEAP_PROPERTIES *heap_properties, D3D12_HEAP_FLAGS heap_flags, const D3D12_RESOURCE_DESC *desc, D3D12_RESOURCE_STATES initial_state, - const D3D12_CLEAR_VALUE *optimized_clear_value, struct d3d12_resource **resource) + const D3D12_CLEAR_VALUE *optimized_clear_value, bool placed, struct d3d12_resource **resource) { struct d3d12_resource *object; HRESULT hr; @@ -1834,7 +1921,7 @@ static HRESULT d3d12_resource_create(struct d3d12_device *device, return E_OUTOFMEMORY;
if (FAILED(hr = d3d12_resource_init(object, device, heap_properties, heap_flags, - desc, initial_state, optimized_clear_value))) + desc, initial_state, optimized_clear_value, placed))) { vkd3d_free(object); return hr; @@ -1876,7 +1963,7 @@ HRESULT d3d12_committed_resource_create(struct d3d12_device *device, }
if (FAILED(hr = d3d12_resource_create(device, heap_properties, heap_flags, - desc, initial_state, optimized_clear_value, &object))) + desc, initial_state, optimized_clear_value, false, &object))) return hr;
if (FAILED(hr = vkd3d_allocate_resource_memory(device, object, heap_properties, heap_flags))) @@ -1900,6 +1987,16 @@ static HRESULT vkd3d_bind_heap_memory(struct d3d12_device *device, VkMemoryRequirements requirements; VkResult vr;
+ if (resource->flags & VKD3D_RESOURCE_PLACED_BUFFER) + { + /* Just inherit the buffer from the heap. */ + resource->u.vk_buffer = heap->buffer_resource->u.vk_buffer; + resource->heap = heap; + resource->heap_offset = heap_offset; + resource->gpu_address = heap->buffer_resource->gpu_address + heap_offset; + return S_OK; + } + if (d3d12_resource_is_buffer(resource)) VK_CALL(vkGetBufferMemoryRequirements(vk_device, resource->u.vk_buffer, &requirements)); else @@ -1949,7 +2046,7 @@ HRESULT d3d12_placed_resource_create(struct d3d12_device *device, struct d3d12_h HRESULT hr;
if (FAILED(hr = d3d12_resource_create(device, &heap->desc.Properties, heap->desc.Flags, - desc, initial_state, optimized_clear_value, &object))) + desc, initial_state, optimized_clear_value, true, &object))) return hr;
if (FAILED(hr = vkd3d_bind_heap_memory(device, object, heap, heap_offset))) @@ -1973,7 +2070,7 @@ HRESULT d3d12_reserved_resource_create(struct d3d12_device *device, HRESULT hr;
if (FAILED(hr = d3d12_resource_create(device, NULL, 0, - desc, initial_state, optimized_clear_value, &object))) + desc, initial_state, optimized_clear_value, false, &object))) return hr;
TRACE("Created reserved resource %p.\n", object); @@ -2275,7 +2372,7 @@ static bool vkd3d_create_buffer_view_for_resource(struct d3d12_device *device, assert(d3d12_resource_is_buffer(resource));
return vkd3d_create_buffer_view(device, resource->u.vk_buffer, - format, offset * element_size, size * element_size, view); + format, resource->heap_offset + offset * element_size, size * element_size, view); }
static void vkd3d_set_view_swizzle_for_format(VkComponentMapping *components, @@ -2869,7 +2966,7 @@ static void vkd3d_create_buffer_uav(struct d3d12_desc *descriptor, struct d3d12_
format = vkd3d_get_format(device, DXGI_FORMAT_R32_UINT, false); if (!vkd3d_create_vk_buffer_view(device, counter_resource->u.vk_buffer, format, - desc->u.Buffer.CounterOffsetInBytes, sizeof(uint32_t), &view->vk_counter_view)) + desc->u.Buffer.CounterOffsetInBytes + resource->heap_offset, sizeof(uint32_t), &view->vk_counter_view)) { WARN("Failed to create counter buffer view.\n"); view->vk_counter_view = VK_NULL_HANDLE; @@ -2960,12 +3057,18 @@ bool vkd3d_create_raw_buffer_view(struct d3d12_device *device, { const struct vkd3d_format *format; struct d3d12_resource *resource; + uint64_t range; + uint64_t offset;
format = vkd3d_get_format(device, DXGI_FORMAT_R32_UINT, false); resource = vkd3d_gpu_va_allocator_dereference(&device->gpu_va_allocator, gpu_address); assert(d3d12_resource_is_buffer(resource)); + + offset = gpu_address - resource->gpu_address; + range = min(resource->desc.Width - offset, device->vk_info.device_limits.maxStorageBufferRange); + return vkd3d_create_vk_buffer_view(device, resource->u.vk_buffer, format, - gpu_address - resource->gpu_address, VK_WHOLE_SIZE, vk_buffer_view); + offset, range, vk_buffer_view); }
/* samplers */ diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h index 0c031d2..206e550 100644 --- a/libs/vkd3d/vkd3d_private.h +++ b/libs/vkd3d/vkd3d_private.h @@ -379,6 +379,7 @@ struct d3d12_heap unsigned int map_count; uint32_t vk_memory_type;
+ struct d3d12_resource *buffer_resource; struct d3d12_device *device;
struct vkd3d_private_store private_store; @@ -393,6 +394,7 @@ struct d3d12_heap *unsafe_impl_from_ID3D12Heap(ID3D12Heap *iface) DECLSPEC_HIDDE #define VKD3D_RESOURCE_EXTERNAL 0x00000004 #define VKD3D_RESOURCE_DEDICATED_HEAP 0x00000008 #define VKD3D_RESOURCE_LINEAR_TILING 0x00000010 +#define VKD3D_RESOURCE_PLACED_BUFFER 0x00000020
/* ID3D12Resource */ struct d3d12_resource