From: Conor McCarthy cmccarthy@codeweavers.com
The descriptor component of struct d3d12_desc is replaced with a union containing a pointer which can be swapped out using InterlockedExchangePointer(). To make it safe to increment the refcount of such an object it is necessary to cache freed objects. Elimination of the descriptor mutexes on games which use multithreaded descriptor writes nearly doubles framerate on recent hardware. --- include/private/vkd3d_common.h | 17 ++ include/vkd3d_windows.h | 1 + libs/vkd3d/command.c | 77 +++++--- libs/vkd3d/device.c | 31 ++- libs/vkd3d/resource.c | 349 ++++++++++++++++----------------- libs/vkd3d/vkd3d_private.h | 133 +++++++------ 6 files changed, 332 insertions(+), 276 deletions(-)
diff --git a/include/private/vkd3d_common.h b/include/private/vkd3d_common.h index 017f3f41..5adc8795 100644 --- a/include/private/vkd3d_common.h +++ b/include/private/vkd3d_common.h @@ -255,6 +255,10 @@ static inline LONG InterlockedCompareExchange(LONG volatile *x, LONG xchg, LONG { return __sync_val_compare_and_swap(x, cmp, xchg); } +static inline PVOID InterlockedCompareExchangePointer(PVOID volatile *x, PVOID xchg, PVOID cmp) +{ + return __sync_val_compare_and_swap(x, cmp, xchg); +} # else # error "InterlockedCompareExchange() not implemented for this platform" # endif @@ -263,6 +267,10 @@ static inline LONG InterlockedExchange(LONG volatile *x, LONG val) { return __atomic_exchange_n(x, val, __ATOMIC_SEQ_CST); } +static inline PVOID InterlockedExchangePointer(PVOID volatile *x, PVOID val) +{ + return __atomic_exchange_n(x, val, __ATOMIC_SEQ_CST); +} # elif HAVE_SYNC_COMPARE_AND_SWAP static inline LONG InterlockedExchange(LONG volatile *x, LONG val) { @@ -273,6 +281,15 @@ static inline LONG InterlockedExchange(LONG volatile *x, LONG val) } while (__sync_val_compare_and_swap(x, i, val) != i); return i; } +static inline PVOID InterlockedExchangePointer(PVOID volatile *x, PVOID val) +{ + PVOID p; + do + { + p = *x; + } while (__sync_val_compare_and_swap(x, p, val) != p); + return p; +} # else # error "InterlockedExchange() not implemented for this platform" # endif diff --git a/include/vkd3d_windows.h b/include/vkd3d_windows.h index 002ff667..ba52231c 100644 --- a/include/vkd3d_windows.h +++ b/include/vkd3d_windows.h @@ -76,6 +76,7 @@ typedef int LONG; typedef unsigned int ULONG; typedef float FLOAT; typedef LONG BOOL; +typedef void *PVOID;
/* Assuming LP64 model */ typedef char INT8; diff --git a/libs/vkd3d/command.c b/libs/vkd3d/command.c index 3ba1f608..00a08253 100644 --- a/libs/vkd3d/command.c +++ b/libs/vkd3d/command.c @@ -2720,28 +2720,31 @@ static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_des unsigned int index, bool use_array) { uint32_t descriptor_range_magic = range->descriptor_magic; - const struct vkd3d_view *view = descriptor->s.u.view_info.view; + union d3d12_desc_object u = descriptor->s.u; uint32_t vk_binding = range->binding; + VkDescriptorType vk_descriptor_type; uint32_t set = range->set;
- if (descriptor->s.magic != descriptor_range_magic) + if (!u.header || u.header->magic != descriptor_range_magic) return false;
+ vk_descriptor_type = u.header->vk_descriptor_type; + vk_descriptor_write->sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; vk_descriptor_write->pNext = NULL; vk_descriptor_write->dstSet = vk_descriptor_sets[set]; vk_descriptor_write->dstBinding = use_array ? vk_binding : vk_binding + index; vk_descriptor_write->dstArrayElement = use_array ? index : 0; vk_descriptor_write->descriptorCount = 1; - vk_descriptor_write->descriptorType = descriptor->s.vk_descriptor_type; + vk_descriptor_write->descriptorType = vk_descriptor_type; vk_descriptor_write->pImageInfo = NULL; vk_descriptor_write->pBufferInfo = NULL; vk_descriptor_write->pTexelBufferView = NULL;
- switch (descriptor->s.magic) + switch (u.header->magic) { case VKD3D_DESCRIPTOR_MAGIC_CBV: - vk_descriptor_write->pBufferInfo = &descriptor->s.u.vk_cbv_info; + vk_descriptor_write->pBufferInfo = &u.cb_desc->vk_cbv_info; break;
case VKD3D_DESCRIPTOR_MAGIC_SRV: @@ -2752,8 +2755,8 @@ static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_des * in pairs in one set. */ if (range->descriptor_count == UINT_MAX) { - if (descriptor->s.vk_descriptor_type != VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER - && descriptor->s.vk_descriptor_type != VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER) + if (vk_descriptor_type != VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER + && vk_descriptor_type != VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER) { vk_descriptor_write->dstSet = vk_descriptor_sets[set + 1]; vk_descriptor_write->dstBinding = 0; @@ -2763,21 +2766,21 @@ static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_des { if (!use_array) vk_descriptor_write->dstBinding = vk_binding + 2 * index; - if (descriptor->s.vk_descriptor_type != VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER - && descriptor->s.vk_descriptor_type != VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER) + if (vk_descriptor_type != VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER + && vk_descriptor_type != VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER) ++vk_descriptor_write->dstBinding; }
- if (descriptor->s.vk_descriptor_type == VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER - || descriptor->s.vk_descriptor_type == VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER) + if (vk_descriptor_type == VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER + || vk_descriptor_type == VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER) { - vk_descriptor_write->pTexelBufferView = &view->u.vk_buffer_view; + vk_descriptor_write->pTexelBufferView = &u.view->v.u.vk_buffer_view; } else { vk_image_info->sampler = VK_NULL_HANDLE; - vk_image_info->imageView = view->u.vk_image_view; - vk_image_info->imageLayout = descriptor->s.magic == VKD3D_DESCRIPTOR_MAGIC_SRV + vk_image_info->imageView = u.view->v.u.vk_image_view; + vk_image_info->imageLayout = u.header->magic == VKD3D_DESCRIPTOR_MAGIC_SRV ? VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL : VK_IMAGE_LAYOUT_GENERAL;
vk_descriptor_write->pImageInfo = vk_image_info; @@ -2785,7 +2788,7 @@ static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_des break;
case VKD3D_DESCRIPTOR_MAGIC_SAMPLER: - vk_image_info->sampler = view->u.vk_sampler; + vk_image_info->sampler = u.view->v.u.vk_sampler; vk_image_info->imageView = VK_NULL_HANDLE; vk_image_info->imageLayout = VK_IMAGE_LAYOUT_UNDEFINED;
@@ -2793,7 +2796,7 @@ static bool vk_write_descriptor_set_from_d3d12_desc(VkWriteDescriptorSet *vk_des break;
default: - ERR("Invalid descriptor %#x.\n", descriptor->s.magic); + ERR("Invalid descriptor %#x.\n", u.header->magic); return false; }
@@ -2847,6 +2850,11 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list for (j = 0; j < descriptor_count; ++j, ++descriptor) { unsigned int register_idx = range->base_register_idx + j; + union d3d12_desc_object u = descriptor->s.u; + VkBufferView vk_counter_view; + + vk_counter_view = (u.header && u.header->magic == VKD3D_DESCRIPTOR_MAGIC_UAV) + ? u.view->v.vk_counter_view : VK_NULL_HANDLE;
/* Track UAV counters. */ if (range->descriptor_magic == VKD3D_DESCRIPTOR_MAGIC_UAV) @@ -2856,8 +2864,6 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list if (state->uav_counters.bindings[k].register_space == range->register_space && state->uav_counters.bindings[k].register_index == register_idx) { - VkBufferView vk_counter_view = descriptor->s.magic == VKD3D_DESCRIPTOR_MAGIC_UAV - ? descriptor->s.u.view_info.view->vk_counter_view : VK_NULL_HANDLE; if (bindings->vk_uav_counter_views[k] != vk_counter_view) bindings->uav_counters_dirty = true; bindings->vk_uav_counter_views[k] = vk_counter_view; @@ -2867,7 +2873,7 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list }
/* Not all descriptors are necessarily populated if the range is unbounded. */ - if (descriptor->s.magic == VKD3D_DESCRIPTOR_MAGIC_FREE) + if (!u.header) continue;
if (!vk_write_descriptor_set_from_d3d12_desc(current_descriptor_write, current_image_info, @@ -4836,7 +4842,7 @@ static void STDMETHODCALLTYPE d3d12_command_list_OMSetRenderTargets(ID3D12Graphi WARN("Failed to add view.\n"); }
- list->rtvs[i] = view->u.vk_image_view; + list->rtvs[i] = view->v.u.vk_image_view; list->fb_width = max(list->fb_width, rtv_desc->width); list->fb_height = max(list->fb_height, rtv_desc->height); list->fb_layer_count = max(list->fb_layer_count, rtv_desc->layer_count); @@ -4860,7 +4866,7 @@ static void STDMETHODCALLTYPE d3d12_command_list_OMSetRenderTargets(ID3D12Graphi list->dsv = VK_NULL_HANDLE; }
- list->dsv = view->u.vk_image_view; + list->dsv = view->v.u.vk_image_view; list->fb_width = max(list->fb_width, dsv_desc->width); list->fb_height = max(list->fb_height, dsv_desc->height); list->fb_layer_count = max(list->fb_layer_count, dsv_desc->layer_count); @@ -4952,7 +4958,7 @@ static void d3d12_command_list_clear(struct d3d12_command_list *list, fb_desc.flags = 0; fb_desc.renderPass = vk_render_pass; fb_desc.attachmentCount = 1; - fb_desc.pAttachments = &view->u.vk_image_view; + fb_desc.pAttachments = &view->v.u.vk_image_view; fb_desc.width = width; fb_desc.height = height; fb_desc.layers = layer_count; @@ -5155,13 +5161,14 @@ static void vkd3d_uav_clear_state_get_image_pipeline(const struct vkd3d_uav_clea }
static void d3d12_command_list_clear_uav(struct d3d12_command_list *list, - struct d3d12_resource *resource, struct vkd3d_view *view, const VkClearColorValue *clear_colour, + struct d3d12_resource *resource, struct vkd3d_view *descriptor, const VkClearColorValue *clear_colour, unsigned int rect_count, const D3D12_RECT *rects) { const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs; unsigned int i, miplevel_idx, layer_count; struct vkd3d_uav_clear_pipeline pipeline; struct vkd3d_uav_clear_args clear_args; + const struct vkd3d_resource_view *view; VkDescriptorImageInfo image_info; D3D12_RECT full_rect, curr_rect; VkWriteDescriptorSet write_set; @@ -5173,8 +5180,9 @@ static void d3d12_command_list_clear_uav(struct d3d12_command_list *list, d3d12_command_list_invalidate_bindings(list, list->state); d3d12_command_list_invalidate_root_parameters(list, VKD3D_PIPELINE_BIND_POINT_COMPUTE);
- if (!d3d12_command_allocator_add_view(list->allocator, view)) + if (!d3d12_command_allocator_add_view(list->allocator, descriptor)) WARN("Failed to add view.\n"); + view = &descriptor->v;
clear_args.colour = *clear_colour;
@@ -5287,10 +5295,11 @@ static void STDMETHODCALLTYPE d3d12_command_list_ClearUnorderedAccessViewUint(ID const UINT values[4], UINT rect_count, const D3D12_RECT *rects) { struct d3d12_command_list *list = impl_from_ID3D12GraphicsCommandList2(iface); + struct vkd3d_view *descriptor, *uint_view = NULL; struct d3d12_device *device = list->device; - struct vkd3d_view *view, *uint_view = NULL; struct vkd3d_texture_view_desc view_desc; const struct vkd3d_format *uint_format; + const struct vkd3d_resource_view *view; struct d3d12_resource *resource_impl; VkClearColorValue colour;
@@ -5298,7 +5307,9 @@ static void STDMETHODCALLTYPE d3d12_command_list_ClearUnorderedAccessViewUint(ID iface, gpu_handle.ptr, cpu_handle.ptr, resource, values, rect_count, rects);
resource_impl = unsafe_impl_from_ID3D12Resource(resource); - view = d3d12_desc_from_cpu_handle(cpu_handle)->s.u.view_info.view; + if (!(descriptor = d3d12_desc_from_cpu_handle(cpu_handle)->s.u.view)) + return; + view = &descriptor->v; memcpy(colour.uint32, values, sizeof(colour.uint32));
if (view->format->type != VKD3D_FORMAT_TYPE_UINT) @@ -5312,8 +5323,8 @@ static void STDMETHODCALLTYPE d3d12_command_list_ClearUnorderedAccessViewUint(ID
if (d3d12_resource_is_buffer(resource_impl)) { - if (!vkd3d_create_buffer_view(device, resource_impl->u.vk_buffer, uint_format, - view->info.buffer.offset, view->info.buffer.size, &uint_view)) + if (!vkd3d_create_buffer_view(device, VKD3D_DESCRIPTOR_MAGIC_UAV, resource_impl->u.vk_buffer, + uint_format, view->info.buffer.offset, view->info.buffer.size, &uint_view)) { ERR("Failed to create buffer view.\n"); return; @@ -5329,16 +5340,17 @@ static void STDMETHODCALLTYPE d3d12_command_list_ClearUnorderedAccessViewUint(ID view_desc.layer_idx = view->info.texture.layer_idx; view_desc.layer_count = view->info.texture.layer_count;
- if (!vkd3d_create_texture_view(device, resource_impl->u.vk_image, &view_desc, &uint_view)) + if (!vkd3d_create_texture_view(device, VKD3D_DESCRIPTOR_MAGIC_UAV, resource_impl->u.vk_image, &view_desc, + &uint_view)) { ERR("Failed to create image view.\n"); return; } } - view = uint_view; + descriptor = uint_view; }
- d3d12_command_list_clear_uav(list, resource_impl, view, &colour, rect_count, rects); + d3d12_command_list_clear_uav(list, resource_impl, descriptor, &colour, rect_count, rects);
if (uint_view) vkd3d_view_decref(uint_view, device); @@ -5357,7 +5369,8 @@ static void STDMETHODCALLTYPE d3d12_command_list_ClearUnorderedAccessViewFloat(I iface, gpu_handle.ptr, cpu_handle.ptr, resource, values, rect_count, rects);
resource_impl = unsafe_impl_from_ID3D12Resource(resource); - view = d3d12_desc_from_cpu_handle(cpu_handle)->s.u.view_info.view; + if (!(view = d3d12_desc_from_cpu_handle(cpu_handle)->s.u.view)) + return; memcpy(colour.float32, values, sizeof(colour.float32));
d3d12_command_list_clear_uav(list, resource_impl, view, &colour, rect_count, rects); diff --git a/libs/vkd3d/device.c b/libs/vkd3d/device.c index d9555d1d..93274bc9 100644 --- a/libs/vkd3d/device.c +++ b/libs/vkd3d/device.c @@ -2414,6 +2414,24 @@ static void vkd3d_init_descriptor_pool_sizes(VkDescriptorPoolSize *pool_sizes, VKD3D_MAX_VIRTUAL_HEAP_DESCRIPTORS_PER_TYPE); };
+static void vkd3d_desc_object_cache_init(struct vkd3d_desc_object_cache *cache, size_t size) +{ + cache->head = NULL; + cache->size = size; +} + +static void vkd3d_desc_object_cache_cleanup(struct vkd3d_desc_object_cache *cache) +{ + union d3d12_desc_object u; + void *next; + + for (u.object = cache->head; u.object; u.object = next) + { + next = u.header->next; + vkd3d_free(u.object); + } +} + /* ID3D12Device */ static inline struct d3d12_device *impl_from_ID3D12Device(ID3D12Device *iface) { @@ -2454,7 +2472,6 @@ static ULONG STDMETHODCALLTYPE d3d12_device_Release(ID3D12Device *iface) { struct d3d12_device *device = impl_from_ID3D12Device(iface); ULONG refcount = InterlockedDecrement(&device->refcount); - size_t i;
TRACE("%p decreasing refcount to %u.\n", device, refcount);
@@ -2474,8 +2491,8 @@ static ULONG STDMETHODCALLTYPE d3d12_device_Release(ID3D12Device *iface) vkd3d_render_pass_cache_cleanup(&device->render_pass_cache, device); d3d12_device_destroy_pipeline_cache(device); d3d12_device_destroy_vkd3d_queues(device); - for (i = 0; i < ARRAY_SIZE(device->desc_mutex); ++i) - vkd3d_mutex_destroy(&device->desc_mutex[i]); + vkd3d_desc_object_cache_cleanup(&device->view_desc_cache); + vkd3d_desc_object_cache_cleanup(&device->cbuffer_desc_cache); VK_CALL(vkDestroyDevice(device->vk_device, NULL)); if (device->parent) IUnknown_Release(device->parent); @@ -3411,8 +3428,7 @@ static void STDMETHODCALLTYPE d3d12_device_CopyDescriptors(ID3D12Device *iface,
for (; dst_idx < dst_range_size && src_idx < src_range_size; ++dst_idx, ++src_idx) { - if (dst[dst_idx].s.magic == src[src_idx].s.magic && (dst[dst_idx].s.magic & VKD3D_DESCRIPTOR_MAGIC_HAS_VIEW) - && dst[dst_idx].s.u.view_info.written_serial_id == src[src_idx].s.u.view_info.view->serial_id) + if (dst[dst_idx].s.u.object == src[src_idx].s.u.object) continue; d3d12_desc_copy(&dst[dst_idx], &src[src_idx], device); } @@ -3939,7 +3955,6 @@ static HRESULT d3d12_device_init(struct d3d12_device *device, { const struct vkd3d_vk_device_procs *vk_procs; HRESULT hr; - size_t i;
device->ID3D12Device_iface.lpVtbl = &d3d12_device_vtbl; device->refcount = 1; @@ -3982,8 +3997,8 @@ static HRESULT d3d12_device_init(struct d3d12_device *device, device->blocked_queue_count = 0; vkd3d_mutex_init(&device->blocked_queues_mutex);
- for (i = 0; i < ARRAY_SIZE(device->desc_mutex); ++i) - vkd3d_mutex_init(&device->desc_mutex[i]); + vkd3d_desc_object_cache_init(&device->view_desc_cache, sizeof(struct vkd3d_view)); + vkd3d_desc_object_cache_init(&device->cbuffer_desc_cache, sizeof(struct vkd3d_cbuffer_desc));
vkd3d_init_descriptor_pool_sizes(device->vk_pool_sizes, &device->vk_info.descriptor_limits);
diff --git a/libs/vkd3d/resource.c b/libs/vkd3d/resource.c index 260832e6..3e285900 100644 --- a/libs/vkd3d/resource.c +++ b/libs/vkd3d/resource.c @@ -2044,24 +2044,72 @@ ULONG vkd3d_resource_decref(ID3D12Resource *resource) return d3d12_resource_decref(impl_from_ID3D12Resource(resource)); }
-/* CBVs, SRVs, UAVs */ -static struct vkd3d_view *vkd3d_view_create(enum vkd3d_view_type type) +/* Objects are cached so that vkd3d_view_incref() can safely check the refcount + * of an object freed by another thread. */ +static void *vkd3d_desc_object_cache_get(struct vkd3d_desc_object_cache *cache) { - struct vkd3d_view *view; + union d3d12_desc_object u; + void *next;
- if ((view = vkd3d_malloc(sizeof(*view)))) + do { - view->refcount = 1; - view->type = type; - view->serial_id = InterlockedIncrement64(&object_global_serial_id); - view->vk_counter_view = VK_NULL_HANDLE; + u.object = cache->head; + if (!u.object) + return vkd3d_malloc(cache->size); + next = u.header->next; } - return view; + while (InterlockedCompareExchangePointer(&cache->head, next, u.object) != u.object); + + return u.object; +} + +static void vkd3d_desc_object_cache_push(struct vkd3d_desc_object_cache *cache, void *object) +{ + union d3d12_desc_object u = {object}; + void *head; + + do + { + head = cache->head; + u.header->next = head; + } + while (InterlockedCompareExchangePointer(&cache->head, u.object, head) != head); }
-void vkd3d_view_incref(struct vkd3d_view *view) +static struct vkd3d_cbuffer_desc *vkd3d_cbuffer_desc_create(struct d3d12_device *device) { - InterlockedIncrement(&view->refcount); + struct vkd3d_cbuffer_desc *desc; + + if (!(desc = vkd3d_desc_object_cache_get(&device->cbuffer_desc_cache))) + return NULL; + + desc->h.magic = VKD3D_DESCRIPTOR_MAGIC_CBV; + desc->h.vk_descriptor_type = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER; + desc->h.refcount = 1; + + return desc; +} + +static struct vkd3d_view *vkd3d_view_create(uint32_t magic, VkDescriptorType vk_descriptor_type, + enum vkd3d_view_type type, struct d3d12_device *device) +{ + struct vkd3d_view *view; + + assert(magic); + + if (!(view = vkd3d_desc_object_cache_get(&device->view_desc_cache))) + { + ERR("Failed to allocate descriptor object.\n"); + return NULL; + } + + view->h.magic = magic; + view->h.vk_descriptor_type = vk_descriptor_type; + view->h.refcount = 1; + view->v.type = type; + view->v.vk_counter_view = VK_NULL_HANDLE; + + return view; }
static void vkd3d_view_destroy(struct vkd3d_view *view, struct d3d12_device *device) @@ -2070,31 +2118,44 @@ static void vkd3d_view_destroy(struct vkd3d_view *view, struct d3d12_device *dev
TRACE("Destroying view %p.\n", view);
- switch (view->type) + switch (view->v.type) { case VKD3D_VIEW_TYPE_BUFFER: - VK_CALL(vkDestroyBufferView(device->vk_device, view->u.vk_buffer_view, NULL)); + VK_CALL(vkDestroyBufferView(device->vk_device, view->v.u.vk_buffer_view, NULL)); break; case VKD3D_VIEW_TYPE_IMAGE: - VK_CALL(vkDestroyImageView(device->vk_device, view->u.vk_image_view, NULL)); + VK_CALL(vkDestroyImageView(device->vk_device, view->v.u.vk_image_view, NULL)); break; case VKD3D_VIEW_TYPE_SAMPLER: - VK_CALL(vkDestroySampler(device->vk_device, view->u.vk_sampler, NULL)); + VK_CALL(vkDestroySampler(device->vk_device, view->v.u.vk_sampler, NULL)); break; default: - WARN("Unhandled view type %d.\n", view->type); + WARN("Unhandled view type %d.\n", view->v.type); }
- if (view->vk_counter_view) - VK_CALL(vkDestroyBufferView(device->vk_device, view->vk_counter_view, NULL)); + if (view->v.vk_counter_view) + VK_CALL(vkDestroyBufferView(device->vk_device, view->v.vk_counter_view, NULL));
- vkd3d_free(view); + vkd3d_desc_object_cache_push(&device->view_desc_cache, view); }
-void vkd3d_view_decref(struct vkd3d_view *view, struct d3d12_device *device) +void vkd3d_view_decref(void *view, struct d3d12_device *device) { - if (!InterlockedDecrement(&view->refcount)) - vkd3d_view_destroy(view, device); + union d3d12_desc_object u = {view}; + + if (InterlockedDecrement(&u.header->refcount)) + return; + + if (u.header->magic != VKD3D_DESCRIPTOR_MAGIC_CBV) + vkd3d_view_destroy(u.view, device); + else + vkd3d_desc_object_cache_push(&device->cbuffer_desc_cache, u.object); +} + +static inline void d3d12_desc_replace(struct d3d12_desc *dst, void *view, struct d3d12_device *device) +{ + if ((view = InterlockedExchangePointer(&dst->s.u.object, view))) + vkd3d_view_decref(view, device); }
static void d3d12_desc_write_vk_heap_null_descriptor(struct d3d12_descriptor_heap *descriptor_heap, @@ -2156,17 +2217,18 @@ static void d3d12_desc_write_vk_heap_null_descriptor(struct d3d12_descriptor_hea }
static void d3d12_desc_write_vk_heap(struct d3d12_descriptor_heap *descriptor_heap, unsigned int dst_array_element, - const struct d3d12_desc *src, struct d3d12_device *device) + void *object, struct d3d12_device *device) { struct d3d12_descriptor_heap_vk_set *descriptor_set; const struct vkd3d_vk_device_procs *vk_procs; VkWriteDescriptorSet vk_descriptor_writes[2]; + union d3d12_desc_object u = {object}; VkDescriptorImageInfo vk_image_info; unsigned int count = 1; VkDescriptorType type; bool is_null = false;
- type = src->s.vk_descriptor_type; + type = u.header->vk_descriptor_type; descriptor_set = &descriptor_heap->vk_descriptor_sets[vkd3d_vk_descriptor_set_index_from_vk_descriptor_type(type)]; vk_procs = &device->vk_procs;
@@ -2181,9 +2243,9 @@ static void d3d12_desc_write_vk_heap(struct d3d12_descriptor_heap *descriptor_he { case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER: vk_descriptor_writes[0].pImageInfo = NULL; - vk_descriptor_writes[0].pBufferInfo = &src->s.u.vk_cbv_info; + vk_descriptor_writes[0].pBufferInfo = &u.cb_desc->vk_cbv_info; vk_descriptor_writes[0].pTexelBufferView = NULL; - is_null = !src->s.u.vk_cbv_info.buffer; + is_null = !u.cb_desc->vk_cbv_info.buffer; break; case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE: case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: @@ -2191,7 +2253,7 @@ static void d3d12_desc_write_vk_heap(struct d3d12_descriptor_heap *descriptor_he vk_descriptor_writes[0].pBufferInfo = NULL; vk_descriptor_writes[0].pTexelBufferView = NULL; vk_image_info.sampler = VK_NULL_HANDLE; - is_null = !(vk_image_info.imageView = src->s.u.view_info.view->u.vk_image_view); + is_null = !(vk_image_info.imageView = u.view->v.u.vk_image_view); vk_image_info.imageLayout = (type == VK_DESCRIPTOR_TYPE_STORAGE_IMAGE) ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; break; @@ -2199,14 +2261,14 @@ static void d3d12_desc_write_vk_heap(struct d3d12_descriptor_heap *descriptor_he case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER: vk_descriptor_writes[0].pImageInfo = NULL; vk_descriptor_writes[0].pBufferInfo = NULL; - vk_descriptor_writes[0].pTexelBufferView = &src->s.u.view_info.view->u.vk_buffer_view; - is_null = !src->s.u.view_info.view->u.vk_buffer_view; + vk_descriptor_writes[0].pTexelBufferView = &u.view->v.u.vk_buffer_view; + is_null = !u.view->v.u.vk_buffer_view; break; case VK_DESCRIPTOR_TYPE_SAMPLER: vk_descriptor_writes[0].pImageInfo = &vk_image_info; vk_descriptor_writes[0].pBufferInfo = NULL; vk_descriptor_writes[0].pTexelBufferView = NULL; - vk_image_info.sampler = src->s.u.view_info.view->u.vk_sampler; + vk_image_info.sampler = u.view->v.u.vk_sampler; vk_image_info.imageView = VK_NULL_HANDLE; vk_image_info.imageLayout = VK_IMAGE_LAYOUT_UNDEFINED; break; @@ -2217,7 +2279,7 @@ static void d3d12_desc_write_vk_heap(struct d3d12_descriptor_heap *descriptor_he if (is_null && device->vk_info.EXT_robustness2) return d3d12_desc_write_vk_heap_null_descriptor(descriptor_heap, dst_array_element, device);
- if (src->s.magic == VKD3D_DESCRIPTOR_MAGIC_UAV && src->s.u.view_info.view->vk_counter_view) + if (u.header->magic == VKD3D_DESCRIPTOR_MAGIC_UAV && u.view->v.vk_counter_view) { descriptor_set = &descriptor_heap->vk_descriptor_sets[VKD3D_SET_INDEX_UAV_COUNTER]; vk_descriptor_writes[1].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET; @@ -2229,7 +2291,7 @@ static void d3d12_desc_write_vk_heap(struct d3d12_descriptor_heap *descriptor_he vk_descriptor_writes[1].descriptorCount = 1; vk_descriptor_writes[1].pImageInfo = NULL; vk_descriptor_writes[1].pBufferInfo = NULL; - vk_descriptor_writes[1].pTexelBufferView = &src->s.u.view_info.view->vk_counter_view; + vk_descriptor_writes[1].pTexelBufferView = &u.view->v.vk_counter_view; ++count; }
@@ -2239,8 +2301,7 @@ static void d3d12_desc_write_vk_heap(struct d3d12_descriptor_heap *descriptor_he void d3d12_desc_flush_vk_heap_updates(struct d3d12_descriptor_heap *descriptor_heap, struct d3d12_device *device) { struct d3d12_desc *descriptors, *src; - struct vkd3d_mutex *mutex; - struct d3d12_desc tmp; + union d3d12_desc_object u; unsigned int i, next;
i = InterlockedExchange(&descriptor_heap->dirty_list_head, UINT_MAX); @@ -2252,23 +2313,18 @@ void d3d12_desc_flush_vk_heap_updates(struct d3d12_descriptor_heap *descriptor_h src = &descriptors[i]; next = src->next >> 1;
- mutex = d3d12_device_get_descriptor_mutex(device, src); - vkd3d_mutex_lock(mutex); + u.object = d3d12_desc_get_object_ref(src, device);
- d3d12_desc_copy_raw(&tmp, src); - - if (tmp.s.magic & VKD3D_DESCRIPTOR_MAGIC_HAS_VIEW) - vkd3d_view_incref(tmp.s.u.view_info.view); - - vkd3d_mutex_unlock(mutex); - - if (tmp.s.magic) + if (!u.object) { - d3d12_desc_write_vk_heap(descriptor_heap, i, &tmp, device); - if (tmp.s.magic & VKD3D_DESCRIPTOR_MAGIC_HAS_VIEW) - vkd3d_view_decref(tmp.s.u.view_info.view, device); + InterlockedExchange(&src->next, 0); + continue; }
+ d3d12_desc_write_vk_heap(descriptor_heap, i, u.object, device); + + vkd3d_view_decref(u.object, device); + InterlockedExchange(&src->next, 0); } } @@ -2297,56 +2353,26 @@ static void d3d12_desc_mark_as_modified(struct d3d12_desc *dst) void d3d12_desc_write_atomic(struct d3d12_desc *dst, const struct d3d12_desc *src, struct d3d12_device *device) { - struct vkd3d_view *defunct_view = NULL; - struct vkd3d_mutex *mutex; - - mutex = d3d12_device_get_descriptor_mutex(device, dst); - vkd3d_mutex_lock(mutex); - - /* Nothing to do for VKD3D_DESCRIPTOR_MAGIC_CBV. */ - if ((dst->s.magic & VKD3D_DESCRIPTOR_MAGIC_HAS_VIEW) - && !InterlockedDecrement(&dst->s.u.view_info.view->refcount)) - defunct_view = dst->s.u.view_info.view; - - d3d12_desc_copy_raw(dst, src); - - vkd3d_mutex_unlock(mutex); + void *object = src->s.u.object;
- /* Destroy the view after unlocking to reduce wait time. */ - if (defunct_view) - vkd3d_view_destroy(defunct_view, device); - - if (device->use_vk_heaps && dst->s.magic && !dst->next) + d3d12_desc_replace(dst, object, device); + if (device->use_vk_heaps && object && !dst->next) d3d12_desc_mark_as_modified(dst); }
static void d3d12_desc_destroy(struct d3d12_desc *descriptor, struct d3d12_device *device) { - static const struct d3d12_desc null_desc = {0}; - - d3d12_desc_write_atomic(descriptor, &null_desc, device); + d3d12_desc_replace(descriptor, NULL, device); }
void d3d12_desc_copy(struct d3d12_desc *dst, const struct d3d12_desc *src, struct d3d12_device *device) { struct d3d12_desc tmp; - struct vkd3d_mutex *mutex;
assert(dst != src);
- /* Shadow of the Tomb Raider and possibly other titles sometimes destroy - * and rewrite a descriptor in another thread while it is being copied. */ - mutex = d3d12_device_get_descriptor_mutex(device, src); - vkd3d_mutex_lock(mutex); - - if (src->s.magic & VKD3D_DESCRIPTOR_MAGIC_HAS_VIEW) - vkd3d_view_incref(src->s.u.view_info.view); - - d3d12_desc_copy_raw(&tmp, src); - - vkd3d_mutex_unlock(mutex); - + tmp.s.u.object = d3d12_desc_get_object_ref(src, device); d3d12_desc_write_atomic(dst, &tmp, device); }
@@ -2408,8 +2434,9 @@ static bool vkd3d_create_vk_buffer_view(struct d3d12_device *device, return vr == VK_SUCCESS; }
-bool vkd3d_create_buffer_view(struct d3d12_device *device, VkBuffer vk_buffer, const struct vkd3d_format *format, - VkDeviceSize offset, VkDeviceSize size, struct vkd3d_view **view) +bool vkd3d_create_buffer_view(struct d3d12_device *device, uint32_t magic, VkBuffer vk_buffer, + const struct vkd3d_format *format, VkDeviceSize offset, VkDeviceSize size, + struct vkd3d_view **view) { const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs; VkBufferView vk_view = VK_NULL_HANDLE; @@ -2418,16 +2445,18 @@ bool vkd3d_create_buffer_view(struct d3d12_device *device, VkBuffer vk_buffer, c if (vk_buffer && !vkd3d_create_vk_buffer_view(device, vk_buffer, format, offset, size, &vk_view)) return false;
- if (!(object = vkd3d_view_create(VKD3D_VIEW_TYPE_BUFFER))) + if (!(object = vkd3d_view_create(magic, magic == VKD3D_DESCRIPTOR_MAGIC_UAV + ? VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER : VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER, + VKD3D_VIEW_TYPE_BUFFER, device))) { VK_CALL(vkDestroyBufferView(device->vk_device, vk_view, NULL)); return false; }
- object->u.vk_buffer_view = vk_view; - object->format = format; - object->info.buffer.offset = offset; - object->info.buffer.size = size; + object->v.u.vk_buffer_view = vk_view; + object->v.format = format; + object->v.info.buffer.offset = offset; + object->v.info.buffer.size = size; *view = object; return true; } @@ -2435,7 +2464,7 @@ bool vkd3d_create_buffer_view(struct d3d12_device *device, VkBuffer vk_buffer, c #define VKD3D_VIEW_RAW_BUFFER 0x1
static bool vkd3d_create_buffer_view_for_resource(struct d3d12_device *device, - struct d3d12_resource *resource, DXGI_FORMAT view_format, + uint32_t magic, struct d3d12_resource *resource, DXGI_FORMAT view_format, unsigned int offset, unsigned int size, unsigned int structure_stride, unsigned int flags, struct vkd3d_view **view) { @@ -2466,7 +2495,7 @@ 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, + return vkd3d_create_buffer_view(device, magic, resource->u.vk_buffer, format, offset * element_size, size * element_size, view); }
@@ -2694,7 +2723,7 @@ static void vkd3d_texture_view_desc_normalise(struct vkd3d_texture_view_desc *de desc->layer_count = max_layer_count; }
-bool vkd3d_create_texture_view(struct d3d12_device *device, VkImage vk_image, +bool vkd3d_create_texture_view(struct d3d12_device *device, uint32_t magic, VkImage vk_image, const struct vkd3d_texture_view_desc *desc, struct vkd3d_view **view) { const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs; @@ -2727,18 +2756,19 @@ bool vkd3d_create_texture_view(struct d3d12_device *device, VkImage vk_image, } }
- if (!(object = vkd3d_view_create(VKD3D_VIEW_TYPE_IMAGE))) + if (!(object = vkd3d_view_create(magic, magic == VKD3D_DESCRIPTOR_MAGIC_UAV ? VK_DESCRIPTOR_TYPE_STORAGE_IMAGE + : VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE, VKD3D_VIEW_TYPE_IMAGE, device))) { VK_CALL(vkDestroyImageView(device->vk_device, vk_view, NULL)); return false; }
- object->u.vk_image_view = vk_view; - object->format = format; - object->info.texture.vk_view_type = desc->view_type; - object->info.texture.miplevel_idx = desc->miplevel_idx; - object->info.texture.layer_idx = desc->layer_idx; - object->info.texture.layer_count = desc->layer_count; + object->v.u.vk_image_view = vk_view; + object->v.format = format; + object->v.info.texture.vk_view_type = desc->view_type; + object->v.info.texture.miplevel_idx = desc->miplevel_idx; + object->v.info.texture.layer_idx = desc->layer_idx; + object->v.info.texture.layer_count = desc->layer_count; *view = object; return true; } @@ -2747,6 +2777,7 @@ void d3d12_desc_create_cbv(struct d3d12_desc *descriptor, struct d3d12_device *device, const D3D12_CONSTANT_BUFFER_VIEW_DESC *desc) { struct VkDescriptorBufferInfo *buffer_info; + struct vkd3d_cbuffer_desc *cb_desc; struct d3d12_resource *resource;
if (!desc) @@ -2755,13 +2786,19 @@ void d3d12_desc_create_cbv(struct d3d12_desc *descriptor, return; }
+ if (!(cb_desc = vkd3d_cbuffer_desc_create(device))) + { + ERR("Failed to allocate descriptor object.\n"); + return; + } + if (desc->SizeInBytes & (D3D12_CONSTANT_BUFFER_DATA_PLACEMENT_ALIGNMENT - 1)) { WARN("Size is not %u bytes aligned.\n", D3D12_CONSTANT_BUFFER_DATA_PLACEMENT_ALIGNMENT); return; }
- buffer_info = &descriptor->s.u.vk_cbv_info; + buffer_info = &cb_desc->vk_cbv_info; if (desc->BufferLocation) { resource = vkd3d_gpu_va_allocator_dereference(&device->gpu_va_allocator, desc->BufferLocation); @@ -2777,8 +2814,7 @@ void d3d12_desc_create_cbv(struct d3d12_desc *descriptor, buffer_info->range = VK_WHOLE_SIZE; }
- descriptor->s.magic = VKD3D_DESCRIPTOR_MAGIC_CBV; - descriptor->s.vk_descriptor_type = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER; + descriptor->s.u.cb_desc = cb_desc; }
static unsigned int vkd3d_view_flags_from_d3d12_buffer_srv_flags(D3D12_BUFFER_SRV_FLAGS flags) @@ -2795,7 +2831,6 @@ static void vkd3d_create_null_srv(struct d3d12_desc *descriptor, { struct vkd3d_null_resources *null_resources = &device->null_resources; struct vkd3d_texture_view_desc vkd3d_desc; - struct vkd3d_view *view; VkImage vk_image;
if (!desc) @@ -2810,15 +2845,9 @@ static void vkd3d_create_null_srv(struct d3d12_desc *descriptor, if (!device->vk_info.EXT_robustness2) WARN("Creating NULL buffer SRV %#x.\n", desc->Format);
- if (vkd3d_create_buffer_view(device, null_resources->vk_buffer, + vkd3d_create_buffer_view(device, VKD3D_DESCRIPTOR_MAGIC_SRV, null_resources->vk_buffer, vkd3d_get_format(device, DXGI_FORMAT_R32_UINT, false), - 0, VKD3D_NULL_BUFFER_SIZE, &view)) - { - descriptor->s.magic = VKD3D_DESCRIPTOR_MAGIC_SRV; - descriptor->s.vk_descriptor_type = VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER; - descriptor->s.u.view_info.view = view; - descriptor->s.u.view_info.written_serial_id = view->serial_id; - } + 0, VKD3D_NULL_BUFFER_SIZE, &descriptor->s.u.view); return;
case D3D12_SRV_DIMENSION_TEXTURE2D: @@ -2857,20 +2886,13 @@ static void vkd3d_create_null_srv(struct d3d12_desc *descriptor, vkd3d_desc.components.a = VK_COMPONENT_SWIZZLE_ZERO; vkd3d_desc.allowed_swizzle = true;
- if (!vkd3d_create_texture_view(device, vk_image, &vkd3d_desc, &view)) - return; - - descriptor->s.magic = VKD3D_DESCRIPTOR_MAGIC_SRV; - descriptor->s.vk_descriptor_type = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE; - descriptor->s.u.view_info.view = view; - descriptor->s.u.view_info.written_serial_id = view->serial_id; + vkd3d_create_texture_view(device, VKD3D_DESCRIPTOR_MAGIC_SRV, vk_image, &vkd3d_desc, &descriptor->s.u.view); }
static void vkd3d_create_buffer_srv(struct d3d12_desc *descriptor, struct d3d12_device *device, struct d3d12_resource *resource, const D3D12_SHADER_RESOURCE_VIEW_DESC *desc) { - struct vkd3d_view *view; unsigned int flags;
if (!desc) @@ -2886,15 +2908,9 @@ static void vkd3d_create_buffer_srv(struct d3d12_desc *descriptor, }
flags = vkd3d_view_flags_from_d3d12_buffer_srv_flags(desc->u.Buffer.Flags); - if (!vkd3d_create_buffer_view_for_resource(device, resource, desc->Format, + vkd3d_create_buffer_view_for_resource(device, VKD3D_DESCRIPTOR_MAGIC_SRV, resource, desc->Format, desc->u.Buffer.FirstElement, desc->u.Buffer.NumElements, - desc->u.Buffer.StructureByteStride, flags, &view)) - return; - - descriptor->s.magic = VKD3D_DESCRIPTOR_MAGIC_SRV; - descriptor->s.vk_descriptor_type = VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER; - descriptor->s.u.view_info.view = view; - descriptor->s.u.view_info.written_serial_id = view->serial_id; + desc->u.Buffer.StructureByteStride, flags, &descriptor->s.u.view); }
static VkImageAspectFlags vk_image_aspect_flags_from_d3d12_plane_slice(const struct vkd3d_format *format, @@ -2923,7 +2939,6 @@ void d3d12_desc_create_srv(struct d3d12_desc *descriptor, const D3D12_SHADER_RESOURCE_VIEW_DESC *desc) { struct vkd3d_texture_view_desc vkd3d_desc; - struct vkd3d_view *view;
if (!resource) { @@ -3019,13 +3034,8 @@ void d3d12_desc_create_srv(struct d3d12_desc *descriptor, } }
- if (!vkd3d_create_texture_view(device, resource->u.vk_image, &vkd3d_desc, &view)) - return; - - descriptor->s.magic = VKD3D_DESCRIPTOR_MAGIC_SRV; - descriptor->s.vk_descriptor_type = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE; - descriptor->s.u.view_info.view = view; - descriptor->s.u.view_info.written_serial_id = view->serial_id; + vkd3d_create_texture_view(device, VKD3D_DESCRIPTOR_MAGIC_SRV, resource->u.vk_image, &vkd3d_desc, + &descriptor->s.u.view); }
static unsigned int vkd3d_view_flags_from_d3d12_buffer_uav_flags(D3D12_BUFFER_UAV_FLAGS flags) @@ -3042,7 +3052,6 @@ static void vkd3d_create_null_uav(struct d3d12_desc *descriptor, { struct vkd3d_null_resources *null_resources = &device->null_resources; struct vkd3d_texture_view_desc vkd3d_desc; - struct vkd3d_view *view; VkImage vk_image;
if (!desc) @@ -3057,15 +3066,9 @@ static void vkd3d_create_null_uav(struct d3d12_desc *descriptor, if (!device->vk_info.EXT_robustness2) WARN("Creating NULL buffer UAV %#x.\n", desc->Format);
- if (vkd3d_create_buffer_view(device, null_resources->vk_storage_buffer, + vkd3d_create_buffer_view(device, VKD3D_DESCRIPTOR_MAGIC_UAV, null_resources->vk_storage_buffer, vkd3d_get_format(device, DXGI_FORMAT_R32_UINT, false), - 0, VKD3D_NULL_BUFFER_SIZE, &view)) - { - descriptor->s.magic = VKD3D_DESCRIPTOR_MAGIC_UAV; - descriptor->s.vk_descriptor_type = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER; - descriptor->s.u.view_info.view = view; - descriptor->s.u.view_info.written_serial_id = view->serial_id; - } + 0, VKD3D_NULL_BUFFER_SIZE, &descriptor->s.u.view); return;
case D3D12_UAV_DIMENSION_TEXTURE2D: @@ -3103,13 +3106,7 @@ static void vkd3d_create_null_uav(struct d3d12_desc *descriptor, vkd3d_desc.components.a = VK_COMPONENT_SWIZZLE_A; vkd3d_desc.allowed_swizzle = false;
- if (!vkd3d_create_texture_view(device, vk_image, &vkd3d_desc, &view)) - return; - - descriptor->s.magic = VKD3D_DESCRIPTOR_MAGIC_UAV; - descriptor->s.vk_descriptor_type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE; - descriptor->s.u.view_info.view = view; - descriptor->s.u.view_info.written_serial_id = view->serial_id; + vkd3d_create_texture_view(device, VKD3D_DESCRIPTOR_MAGIC_UAV, vk_image, &vkd3d_desc, &descriptor->s.u.view); }
static void vkd3d_create_buffer_uav(struct d3d12_desc *descriptor, struct d3d12_device *device, @@ -3132,16 +3129,11 @@ static void vkd3d_create_buffer_uav(struct d3d12_desc *descriptor, struct d3d12_ }
flags = vkd3d_view_flags_from_d3d12_buffer_uav_flags(desc->u.Buffer.Flags); - if (!vkd3d_create_buffer_view_for_resource(device, resource, desc->Format, + if (!vkd3d_create_buffer_view_for_resource(device, VKD3D_DESCRIPTOR_MAGIC_UAV, resource, desc->Format, desc->u.Buffer.FirstElement, desc->u.Buffer.NumElements, desc->u.Buffer.StructureByteStride, flags, &view)) return;
- descriptor->s.magic = VKD3D_DESCRIPTOR_MAGIC_UAV; - descriptor->s.vk_descriptor_type = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER; - descriptor->s.u.view_info.view = view; - descriptor->s.u.view_info.written_serial_id = view->serial_id; - if (counter_resource) { const struct vkd3d_format *format; @@ -3151,13 +3143,16 @@ 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, sizeof(uint32_t), &view->v.vk_counter_view)) { WARN("Failed to create counter buffer view.\n"); - view->vk_counter_view = VK_NULL_HANDLE; - d3d12_desc_destroy(descriptor, device); + view->v.vk_counter_view = VK_NULL_HANDLE; + vkd3d_view_decref(view, device); + return; } } + + descriptor->s.u.view = view; }
static void vkd3d_create_texture_uav(struct d3d12_desc *descriptor, @@ -3165,7 +3160,6 @@ static void vkd3d_create_texture_uav(struct d3d12_desc *descriptor, const D3D12_UNORDERED_ACCESS_VIEW_DESC *desc) { struct vkd3d_texture_view_desc vkd3d_desc; - struct vkd3d_view *view;
if (!init_default_texture_view_desc(&vkd3d_desc, resource, desc ? desc->Format : 0)) return; @@ -3210,13 +3204,8 @@ static void vkd3d_create_texture_uav(struct d3d12_desc *descriptor, } }
- if (!vkd3d_create_texture_view(device, resource->u.vk_image, &vkd3d_desc, &view)) - return; - - descriptor->s.magic = VKD3D_DESCRIPTOR_MAGIC_UAV; - descriptor->s.vk_descriptor_type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE; - descriptor->s.u.view_info.view = view; - descriptor->s.u.view_info.written_serial_id = view->serial_id; + vkd3d_create_texture_view(device, VKD3D_DESCRIPTOR_MAGIC_UAV, resource->u.vk_image, &vkd3d_desc, + &descriptor->s.u.view); }
void d3d12_desc_create_uav(struct d3d12_desc *descriptor, struct d3d12_device *device, @@ -3365,21 +3354,21 @@ void d3d12_desc_create_sampler(struct d3d12_desc *sampler, FIXME("Ignoring border color {%.8e, %.8e, %.8e, %.8e}.\n", desc->BorderColor[0], desc->BorderColor[1], desc->BorderColor[2], desc->BorderColor[3]);
- if (!(view = vkd3d_view_create(VKD3D_VIEW_TYPE_SAMPLER))) + if (!(view = vkd3d_view_create(VKD3D_DESCRIPTOR_MAGIC_SAMPLER, VK_DESCRIPTOR_TYPE_SAMPLER, + VKD3D_VIEW_TYPE_SAMPLER, device))) return; + view->v.u.vk_sampler = VK_NULL_HANDLE; + view->v.format = NULL;
if (d3d12_create_sampler(device, desc->Filter, desc->AddressU, desc->AddressV, desc->AddressW, desc->MipLODBias, desc->MaxAnisotropy, - desc->ComparisonFunc, desc->MinLOD, desc->MaxLOD, &view->u.vk_sampler) < 0) + desc->ComparisonFunc, desc->MinLOD, desc->MaxLOD, &view->v.u.vk_sampler) < 0) { - vkd3d_free(view); + vkd3d_view_decref(view, device); return; }
- sampler->s.magic = VKD3D_DESCRIPTOR_MAGIC_SAMPLER; - sampler->s.vk_descriptor_type = VK_DESCRIPTOR_TYPE_SAMPLER; - sampler->s.u.view_info.view = view; - sampler->s.u.view_info.written_serial_id = view->serial_id; + sampler->s.u.view = view; }
HRESULT vkd3d_create_static_sampler(struct d3d12_device *device, @@ -3401,7 +3390,7 @@ HRESULT vkd3d_create_static_sampler(struct d3d12_device *device, /* RTVs */ static void d3d12_rtv_desc_destroy(struct d3d12_rtv_desc *rtv, struct d3d12_device *device) { - if (rtv->magic != VKD3D_DESCRIPTOR_MAGIC_RTV) + if (!rtv->view) return;
vkd3d_view_decref(rtv->view, device); @@ -3480,10 +3469,9 @@ void d3d12_rtv_desc_create_rtv(struct d3d12_rtv_desc *rtv_desc, struct d3d12_dev
assert(d3d12_resource_is_texture(resource));
- if (!vkd3d_create_texture_view(device, resource->u.vk_image, &vkd3d_desc, &view)) + if (!vkd3d_create_texture_view(device, VKD3D_DESCRIPTOR_MAGIC_RTV, resource->u.vk_image, &vkd3d_desc, &view)) return;
- rtv_desc->magic = VKD3D_DESCRIPTOR_MAGIC_RTV; rtv_desc->sample_count = vk_samples_from_dxgi_sample_desc(&resource->desc.SampleDesc); rtv_desc->format = vkd3d_desc.format; rtv_desc->width = d3d12_resource_desc_get_width(&resource->desc, vkd3d_desc.miplevel_idx); @@ -3496,7 +3484,7 @@ void d3d12_rtv_desc_create_rtv(struct d3d12_rtv_desc *rtv_desc, struct d3d12_dev /* DSVs */ static void d3d12_dsv_desc_destroy(struct d3d12_dsv_desc *dsv, struct d3d12_device *device) { - if (dsv->magic != VKD3D_DESCRIPTOR_MAGIC_DSV) + if (!dsv->view) return;
vkd3d_view_decref(dsv->view, device); @@ -3565,10 +3553,9 @@ void d3d12_dsv_desc_create_dsv(struct d3d12_dsv_desc *dsv_desc, struct d3d12_dev
assert(d3d12_resource_is_texture(resource));
- if (!vkd3d_create_texture_view(device, resource->u.vk_image, &vkd3d_desc, &view)) + if (!vkd3d_create_texture_view(device, VKD3D_DESCRIPTOR_MAGIC_DSV, resource->u.vk_image, &vkd3d_desc, &view)) return;
- dsv_desc->magic = VKD3D_DESCRIPTOR_MAGIC_DSV; dsv_desc->sample_count = vk_samples_from_dxgi_sample_desc(&resource->desc.SampleDesc); dsv_desc->format = vkd3d_desc.format; dsv_desc->width = d3d12_resource_desc_get_width(&resource->desc, vkd3d_desc.miplevel_idx); diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h index 2f3632bf..2dc86caf 100644 --- a/libs/vkd3d/vkd3d_private.h +++ b/libs/vkd3d/vkd3d_private.h @@ -44,13 +44,11 @@
#define VK_CALL(f) (vk_procs->f)
-#define VKD3D_DESCRIPTOR_MAGIC_HAS_VIEW 0x01000000u - #define VKD3D_DESCRIPTOR_MAGIC_FREE 0x00000000u #define VKD3D_DESCRIPTOR_MAGIC_CBV VKD3D_MAKE_TAG('C', 'B', 'V', 0) -#define VKD3D_DESCRIPTOR_MAGIC_SRV VKD3D_MAKE_TAG('S', 'R', 'V', 1) -#define VKD3D_DESCRIPTOR_MAGIC_UAV VKD3D_MAKE_TAG('U', 'A', 'V', 1) -#define VKD3D_DESCRIPTOR_MAGIC_SAMPLER VKD3D_MAKE_TAG('S', 'M', 'P', 1) +#define VKD3D_DESCRIPTOR_MAGIC_SRV VKD3D_MAKE_TAG('S', 'R', 'V', 0) +#define VKD3D_DESCRIPTOR_MAGIC_UAV VKD3D_MAKE_TAG('U', 'A', 'V', 0) +#define VKD3D_DESCRIPTOR_MAGIC_SAMPLER VKD3D_MAKE_TAG('S', 'M', 'P', 0) #define VKD3D_DESCRIPTOR_MAGIC_DSV VKD3D_MAKE_TAG('D', 'S', 'V', 0) #define VKD3D_DESCRIPTOR_MAGIC_RTV VKD3D_MAKE_TAG('R', 'T', 'V', 0)
@@ -662,11 +660,9 @@ enum vkd3d_view_type VKD3D_VIEW_TYPE_SAMPLER, };
-struct vkd3d_view +struct vkd3d_resource_view { - LONG refcount; enum vkd3d_view_type type; - uint64_t serial_id; union { VkBufferView vk_buffer_view; @@ -692,9 +688,6 @@ struct vkd3d_view } info; };
-void vkd3d_view_decref(struct vkd3d_view *view, struct d3d12_device *device); -void vkd3d_view_incref(struct vkd3d_view *view); - struct vkd3d_texture_view_desc { VkImageViewType view_type; @@ -708,33 +701,88 @@ struct vkd3d_texture_view_desc bool allowed_swizzle; };
-bool vkd3d_create_buffer_view(struct d3d12_device *device, VkBuffer vk_buffer, const struct vkd3d_format *format, - VkDeviceSize offset, VkDeviceSize size, struct vkd3d_view **view); -bool vkd3d_create_texture_view(struct d3d12_device *device, VkImage vk_image, +struct vkd3d_desc_header +{ + uint32_t magic; + LONG volatile refcount; + void *next; + VkDescriptorType vk_descriptor_type; +}; + +struct vkd3d_view +{ + struct vkd3d_desc_header h; + struct vkd3d_resource_view v; +}; + +bool vkd3d_create_buffer_view(struct d3d12_device *device, uint32_t magic, VkBuffer vk_buffer, + const struct vkd3d_format *format, VkDeviceSize offset, VkDeviceSize size, struct vkd3d_view **view); +bool vkd3d_create_texture_view(struct d3d12_device *device, uint32_t magic, VkImage vk_image, const struct vkd3d_texture_view_desc *desc, struct vkd3d_view **view);
-struct vkd3d_view_info +struct vkd3d_cbuffer_desc { - uint64_t written_serial_id; - struct vkd3d_view *view; + struct vkd3d_desc_header h; + VkDescriptorBufferInfo vk_cbv_info; };
struct d3d12_desc { struct { - uint32_t magic; - VkDescriptorType vk_descriptor_type; - union + union d3d12_desc_object { - VkDescriptorBufferInfo vk_cbv_info; - struct vkd3d_view_info view_info; + struct vkd3d_desc_header *header; + struct vkd3d_view *view; + struct vkd3d_cbuffer_desc *cb_desc; + void *object; } u; } s; unsigned int index; LONG next; };
+void vkd3d_view_decref(void *view, struct d3d12_device *device); + +static inline bool vkd3d_view_incref(void *desc) +{ + struct vkd3d_desc_header *h = desc; + LONG refcount; + + do + { + refcount = h->refcount; + /* Avoid incrementing a freed object. Reading the value is safe because objects are recycled. */ + if (refcount <= 0) + return false; + } + while (InterlockedCompareExchange(&h->refcount, refcount + 1, refcount) != refcount); + + return true; +} + +static inline void *d3d12_desc_get_object_ref(const volatile struct d3d12_desc *src, struct d3d12_device *device) +{ + void *view; + + /* Some games, e.g. Shadow of the Tomb Raider, GRID 2019, and Horizon Zero Dawn, write descriptors + * from multiple threads without syncronisation. This is apparently valid in Windows. */ + for (;;) + { + do + { + view = src->s.u.object; + } while (view && !vkd3d_view_incref(view)); + + /* Check if the object is still in src to handle the case where it was + * already freed and reused elsewhere when the refcount was incremented. */ + if (view == src->s.u.object) + return view; + + vkd3d_view_decref(view, device); + } +} + static inline struct d3d12_desc *d3d12_desc_from_cpu_handle(D3D12_CPU_DESCRIPTOR_HANDLE cpu_handle) { return (struct d3d12_desc *)cpu_handle.ptr; @@ -769,7 +817,6 @@ HRESULT vkd3d_create_static_sampler(struct d3d12_device *device,
struct d3d12_rtv_desc { - uint32_t magic; VkSampleCountFlagBits sample_count; const struct vkd3d_format *format; uint64_t width; @@ -789,7 +836,6 @@ void d3d12_rtv_desc_create_rtv(struct d3d12_rtv_desc *rtv_desc, struct d3d12_dev
struct d3d12_dsv_desc { - uint32_t magic; VkSampleCountFlagBits sample_count; const struct vkd3d_format *format; uint64_t width; @@ -862,7 +908,6 @@ struct d3d12_descriptor_heap struct d3d12_descriptor_heap_vk_set vk_descriptor_sets[VKD3D_SET_INDEX_COUNT];
LONG volatile dirty_list_head; - unsigned int padding;
BYTE DECLSPEC_ALIGN(8) descriptors[]; }; @@ -883,22 +928,6 @@ static inline unsigned int d3d12_desc_heap_range_size(const struct d3d12_desc *d HRESULT d3d12_descriptor_heap_create(struct d3d12_device *device, const D3D12_DESCRIPTOR_HEAP_DESC *desc, struct d3d12_descriptor_heap **descriptor_heap);
-struct d3d12_desc_copy_location -{ - struct d3d12_desc src; - struct d3d12_desc *dst; -}; - -struct d3d12_desc_copy_info -{ - unsigned int count; - bool uav_counter; -}; - -void d3d12_desc_copy_vk_heap_range(struct d3d12_desc_copy_location *locations, const struct d3d12_desc_copy_info *info, - struct d3d12_descriptor_heap *descriptor_heap, enum vkd3d_vk_descriptor_set_index set, - struct d3d12_device *device); - /* ID3D12QueryHeap */ struct d3d12_query_heap { @@ -1486,6 +1515,12 @@ struct vkd3d_uav_clear_state HRESULT vkd3d_uav_clear_state_init(struct vkd3d_uav_clear_state *state, struct d3d12_device *device); void vkd3d_uav_clear_state_cleanup(struct vkd3d_uav_clear_state *state, struct d3d12_device *device);
+struct vkd3d_desc_object_cache +{ + void * volatile head; + size_t size; +}; + #define VKD3D_DESCRIPTOR_POOL_COUNT 6
/* ID3D12Device */ @@ -1503,7 +1538,8 @@ struct d3d12_device struct vkd3d_gpu_va_allocator gpu_va_allocator;
struct vkd3d_mutex mutex; - struct vkd3d_mutex desc_mutex[8]; + struct vkd3d_desc_object_cache view_desc_cache; + struct vkd3d_desc_object_cache cbuffer_desc_cache; struct vkd3d_render_pass_cache render_pass_cache; VkPipelineCache vk_pipeline_cache;
@@ -1578,19 +1614,6 @@ static inline unsigned int d3d12_device_get_descriptor_handle_increment_size(str return ID3D12Device_GetDescriptorHandleIncrementSize(&device->ID3D12Device_iface, descriptor_type); }
-static inline struct vkd3d_mutex *d3d12_device_get_descriptor_mutex(struct d3d12_device *device, - const struct d3d12_desc *descriptor) -{ - STATIC_ASSERT(!(ARRAY_SIZE(device->desc_mutex) & (ARRAY_SIZE(device->desc_mutex) - 1))); - uintptr_t idx = (uintptr_t)descriptor; - - idx ^= idx >> 12; - idx ^= idx >> 6; - idx ^= idx >> 3; - - return &device->desc_mutex[idx & (ARRAY_SIZE(device->desc_mutex) - 1)]; -} - /* utils */ enum vkd3d_format_type {