mirror of
https://gitlab.winehq.org/wine/vkd3d.git
synced 2024-11-21 16:46:41 -08:00
vkd3d: Optimise descriptor copying for Vulkan-backed heaps.
Source descriptors are copied to separate arrays to facilitate use of pre-initialised Vulkan structures, and allow arrayed writes where possible. Signed-off-by: Conor McCarthy <cmccarthy@codeweavers.com> Signed-off-by: Henri Verbeet <hverbeet@codeweavers.com> Signed-off-by: Alexandre Julliard <julliard@winehq.org>
This commit is contained in:
parent
2b71ea406f
commit
5e4f1e1ead
@ -3556,6 +3556,127 @@ static void STDMETHODCALLTYPE d3d12_device_CreateSampler(ID3D12Device *iface,
|
|||||||
d3d12_desc_write_atomic(d3d12_desc_from_cpu_handle(descriptor), &tmp, device);
|
d3d12_desc_write_atomic(d3d12_desc_from_cpu_handle(descriptor), &tmp, device);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
static void flush_desc_writes(struct d3d12_desc_copy_location locations[][VKD3D_DESCRIPTOR_WRITE_BUFFER_SIZE],
|
||||||
|
struct d3d12_desc_copy_info *infos, struct d3d12_descriptor_heap *descriptor_heap, struct d3d12_device *device)
|
||||||
|
{
|
||||||
|
enum vkd3d_vk_descriptor_set_index set;
|
||||||
|
for (set = 0; set < VKD3D_SET_INDEX_COUNT; ++set)
|
||||||
|
{
|
||||||
|
if (!infos[set].count)
|
||||||
|
continue;
|
||||||
|
d3d12_desc_copy_vk_heap_range(locations[set], &infos[set], descriptor_heap, set, device);
|
||||||
|
infos[set].count = 0;
|
||||||
|
infos[set].uav_counter = false;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
static void d3d12_desc_buffered_copy_atomic(struct d3d12_desc *dst, const struct d3d12_desc *src,
|
||||||
|
struct d3d12_desc_copy_location locations[][VKD3D_DESCRIPTOR_WRITE_BUFFER_SIZE],
|
||||||
|
struct d3d12_desc_copy_info *infos, struct d3d12_descriptor_heap *descriptor_heap, struct d3d12_device *device)
|
||||||
|
{
|
||||||
|
struct d3d12_desc_copy_location *location;
|
||||||
|
enum vkd3d_vk_descriptor_set_index set;
|
||||||
|
struct vkd3d_mutex *mutex;
|
||||||
|
|
||||||
|
mutex = d3d12_device_get_descriptor_mutex(device, src);
|
||||||
|
vkd3d_mutex_lock(mutex);
|
||||||
|
|
||||||
|
if (src->magic == VKD3D_DESCRIPTOR_MAGIC_FREE)
|
||||||
|
{
|
||||||
|
/* Source must be unlocked first, and therefore can't be used as a null source. */
|
||||||
|
static const struct d3d12_desc null = {0};
|
||||||
|
vkd3d_mutex_unlock(mutex);
|
||||||
|
d3d12_desc_write_atomic(dst, &null, device);
|
||||||
|
return;
|
||||||
|
}
|
||||||
|
|
||||||
|
set = vkd3d_vk_descriptor_set_index_from_vk_descriptor_type(src->vk_descriptor_type);
|
||||||
|
location = &locations[set][infos[set].count++];
|
||||||
|
|
||||||
|
location->src = *src;
|
||||||
|
|
||||||
|
if (location->src.magic & VKD3D_DESCRIPTOR_MAGIC_HAS_VIEW)
|
||||||
|
vkd3d_view_incref(location->src.u.view);
|
||||||
|
|
||||||
|
vkd3d_mutex_unlock(mutex);
|
||||||
|
|
||||||
|
infos[set].uav_counter |= (location->src.magic == VKD3D_DESCRIPTOR_MAGIC_UAV)
|
||||||
|
& !!location->src.u.view->vk_counter_view;
|
||||||
|
location->dst = dst;
|
||||||
|
|
||||||
|
if (infos[set].count == ARRAY_SIZE(locations[0]))
|
||||||
|
{
|
||||||
|
d3d12_desc_copy_vk_heap_range(locations[set], &infos[set], descriptor_heap, set, device);
|
||||||
|
infos[set].count = 0;
|
||||||
|
infos[set].uav_counter = false;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
/* Some games, e.g. Control, copy a large number of descriptors per frame, so the
|
||||||
|
* speed of this function is critical. */
|
||||||
|
static void d3d12_device_vk_heaps_copy_descriptors(struct d3d12_device *device,
|
||||||
|
UINT dst_descriptor_range_count, const D3D12_CPU_DESCRIPTOR_HANDLE *dst_descriptor_range_offsets,
|
||||||
|
const UINT *dst_descriptor_range_sizes,
|
||||||
|
UINT src_descriptor_range_count, const D3D12_CPU_DESCRIPTOR_HANDLE *src_descriptor_range_offsets,
|
||||||
|
const UINT *src_descriptor_range_sizes)
|
||||||
|
{
|
||||||
|
struct d3d12_desc_copy_location locations[VKD3D_SET_INDEX_COUNT][VKD3D_DESCRIPTOR_WRITE_BUFFER_SIZE];
|
||||||
|
unsigned int dst_range_idx, dst_idx, src_range_idx, src_idx;
|
||||||
|
/* The locations array is relatively large, and often mostly empty. Keeping these
|
||||||
|
* values together in a separate array will likely result in fewer cache misses. */
|
||||||
|
struct d3d12_desc_copy_info infos[VKD3D_SET_INDEX_COUNT];
|
||||||
|
struct d3d12_descriptor_heap *descriptor_heap = NULL;
|
||||||
|
const struct d3d12_desc *src, *heap_base, *heap_end;
|
||||||
|
unsigned int dst_range_size, src_range_size;
|
||||||
|
struct d3d12_desc *dst;
|
||||||
|
|
||||||
|
descriptor_heap = vkd3d_gpu_descriptor_allocator_heap_from_descriptor(&device->gpu_descriptor_allocator,
|
||||||
|
d3d12_desc_from_cpu_handle(dst_descriptor_range_offsets[0]));
|
||||||
|
heap_base = (const struct d3d12_desc *)descriptor_heap->descriptors;
|
||||||
|
heap_end = heap_base + descriptor_heap->desc.NumDescriptors;
|
||||||
|
|
||||||
|
memset(infos, 0, sizeof(infos));
|
||||||
|
dst_range_idx = dst_idx = 0;
|
||||||
|
src_range_idx = src_idx = 0;
|
||||||
|
while (dst_range_idx < dst_descriptor_range_count && src_range_idx < src_descriptor_range_count)
|
||||||
|
{
|
||||||
|
dst_range_size = dst_descriptor_range_sizes ? dst_descriptor_range_sizes[dst_range_idx] : 1;
|
||||||
|
src_range_size = src_descriptor_range_sizes ? src_descriptor_range_sizes[src_range_idx] : 1;
|
||||||
|
|
||||||
|
dst = d3d12_desc_from_cpu_handle(dst_descriptor_range_offsets[dst_range_idx]);
|
||||||
|
src = d3d12_desc_from_cpu_handle(src_descriptor_range_offsets[src_range_idx]);
|
||||||
|
|
||||||
|
if (dst < heap_base || dst >= heap_end)
|
||||||
|
{
|
||||||
|
flush_desc_writes(locations, infos, descriptor_heap, device);
|
||||||
|
descriptor_heap = vkd3d_gpu_descriptor_allocator_heap_from_descriptor(&device->gpu_descriptor_allocator,
|
||||||
|
dst);
|
||||||
|
heap_base = (const struct d3d12_desc *)descriptor_heap->descriptors;
|
||||||
|
heap_end = heap_base + descriptor_heap->desc.NumDescriptors;
|
||||||
|
}
|
||||||
|
|
||||||
|
for (; dst_idx < dst_range_size && src_idx < src_range_size; src_idx++, dst_idx++)
|
||||||
|
{
|
||||||
|
d3d12_desc_buffered_copy_atomic(&dst[dst_idx], &src[src_idx], locations, infos, descriptor_heap, device);
|
||||||
|
}
|
||||||
|
|
||||||
|
if (dst_idx >= dst_range_size)
|
||||||
|
{
|
||||||
|
++dst_range_idx;
|
||||||
|
dst_idx = 0;
|
||||||
|
}
|
||||||
|
if (src_idx >= src_range_size)
|
||||||
|
{
|
||||||
|
++src_range_idx;
|
||||||
|
src_idx = 0;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
flush_desc_writes(locations, infos, descriptor_heap, device);
|
||||||
|
}
|
||||||
|
|
||||||
|
#define VKD3D_DESCRIPTOR_OPTIMISED_COPY_MIN_COUNT 8
|
||||||
|
|
||||||
static void STDMETHODCALLTYPE d3d12_device_CopyDescriptors(ID3D12Device *iface,
|
static void STDMETHODCALLTYPE d3d12_device_CopyDescriptors(ID3D12Device *iface,
|
||||||
UINT dst_descriptor_range_count, const D3D12_CPU_DESCRIPTOR_HANDLE *dst_descriptor_range_offsets,
|
UINT dst_descriptor_range_count, const D3D12_CPU_DESCRIPTOR_HANDLE *dst_descriptor_range_offsets,
|
||||||
const UINT *dst_descriptor_range_sizes,
|
const UINT *dst_descriptor_range_sizes,
|
||||||
@ -3584,6 +3705,18 @@ static void STDMETHODCALLTYPE d3d12_device_CopyDescriptors(ID3D12Device *iface,
|
|||||||
return;
|
return;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
if (!dst_descriptor_range_count)
|
||||||
|
return;
|
||||||
|
|
||||||
|
if (device->use_vk_heaps && (dst_descriptor_range_count > 1 || (dst_descriptor_range_sizes
|
||||||
|
&& dst_descriptor_range_sizes[0] >= VKD3D_DESCRIPTOR_OPTIMISED_COPY_MIN_COUNT)))
|
||||||
|
{
|
||||||
|
d3d12_device_vk_heaps_copy_descriptors(device, dst_descriptor_range_count, dst_descriptor_range_offsets,
|
||||||
|
dst_descriptor_range_sizes, src_descriptor_range_count, src_descriptor_range_offsets,
|
||||||
|
src_descriptor_range_sizes);
|
||||||
|
return;
|
||||||
|
}
|
||||||
|
|
||||||
dst_range_idx = dst_idx = 0;
|
dst_range_idx = dst_idx = 0;
|
||||||
src_range_idx = src_idx = 0;
|
src_range_idx = src_idx = 0;
|
||||||
while (dst_range_idx < dst_descriptor_range_count && src_range_idx < src_descriptor_range_count)
|
while (dst_range_idx < dst_descriptor_range_count && src_range_idx < src_descriptor_range_count)
|
||||||
@ -3620,6 +3753,17 @@ static void STDMETHODCALLTYPE d3d12_device_CopyDescriptorsSimple(ID3D12Device *i
|
|||||||
iface, descriptor_count, dst_descriptor_range_offset.ptr, src_descriptor_range_offset.ptr,
|
iface, descriptor_count, dst_descriptor_range_offset.ptr, src_descriptor_range_offset.ptr,
|
||||||
descriptor_heap_type);
|
descriptor_heap_type);
|
||||||
|
|
||||||
|
if (descriptor_count >= VKD3D_DESCRIPTOR_OPTIMISED_COPY_MIN_COUNT)
|
||||||
|
{
|
||||||
|
struct d3d12_device *device = impl_from_ID3D12Device(iface);
|
||||||
|
if (device->use_vk_heaps)
|
||||||
|
{
|
||||||
|
d3d12_device_vk_heaps_copy_descriptors(device, 1, &dst_descriptor_range_offset,
|
||||||
|
&descriptor_count, 1, &src_descriptor_range_offset, &descriptor_count);
|
||||||
|
return;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
d3d12_device_CopyDescriptors(iface, 1, &dst_descriptor_range_offset, &descriptor_count,
|
d3d12_device_CopyDescriptors(iface, 1, &dst_descriptor_range_offset, &descriptor_count,
|
||||||
1, &src_descriptor_range_offset, &descriptor_count, descriptor_heap_type);
|
1, &src_descriptor_range_offset, &descriptor_count, descriptor_heap_type);
|
||||||
}
|
}
|
||||||
|
@ -2128,6 +2128,53 @@ void vkd3d_view_decref(struct vkd3d_view *view, struct d3d12_device *device)
|
|||||||
vkd3d_view_destroy(view, device);
|
vkd3d_view_destroy(view, device);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
static void d3d12_descriptor_heap_write_vk_descriptor_range(struct d3d12_descriptor_heap_vk_set *descriptor_set,
|
||||||
|
struct d3d12_desc_copy_location *locations, unsigned int write_count)
|
||||||
|
{
|
||||||
|
unsigned int i, info_index = 0, write_index = 0;
|
||||||
|
|
||||||
|
switch (locations[0].src.vk_descriptor_type)
|
||||||
|
{
|
||||||
|
case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
|
||||||
|
for (; write_index < write_count; ++write_index)
|
||||||
|
{
|
||||||
|
descriptor_set->vk_descriptor_writes[write_index].pBufferInfo = &descriptor_set->vk_buffer_infos[info_index];
|
||||||
|
for (i = 0; i < descriptor_set->vk_descriptor_writes[write_index].descriptorCount; ++i, ++info_index)
|
||||||
|
descriptor_set->vk_buffer_infos[info_index] = locations[info_index].src.u.vk_cbv_info;
|
||||||
|
}
|
||||||
|
break;
|
||||||
|
case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
|
||||||
|
case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
|
||||||
|
for (; write_index < write_count; ++write_index)
|
||||||
|
{
|
||||||
|
descriptor_set->vk_descriptor_writes[write_index].pImageInfo = &descriptor_set->vk_image_infos[info_index];
|
||||||
|
for (i = 0; i < descriptor_set->vk_descriptor_writes[write_index].descriptorCount; ++i, ++info_index)
|
||||||
|
descriptor_set->vk_image_infos[info_index].imageView = locations[info_index].src.u.view->u.vk_image_view;
|
||||||
|
}
|
||||||
|
break;
|
||||||
|
case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
|
||||||
|
case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
|
||||||
|
for (; write_index < write_count; ++write_index)
|
||||||
|
{
|
||||||
|
descriptor_set->vk_descriptor_writes[write_index].pTexelBufferView = &descriptor_set->vk_buffer_views[info_index];
|
||||||
|
for (i = 0; i < descriptor_set->vk_descriptor_writes[write_index].descriptorCount; ++i, ++info_index)
|
||||||
|
descriptor_set->vk_buffer_views[info_index] = locations[info_index].src.u.view->u.vk_buffer_view;
|
||||||
|
}
|
||||||
|
break;
|
||||||
|
case VK_DESCRIPTOR_TYPE_SAMPLER:
|
||||||
|
for (; write_index < write_count; ++write_index)
|
||||||
|
{
|
||||||
|
descriptor_set->vk_descriptor_writes[write_index].pImageInfo = &descriptor_set->vk_image_infos[info_index];
|
||||||
|
for (i = 0; i < descriptor_set->vk_descriptor_writes[write_index].descriptorCount; ++i, ++info_index)
|
||||||
|
descriptor_set->vk_image_infos[info_index].sampler = locations[info_index].src.u.view->u.vk_sampler;
|
||||||
|
}
|
||||||
|
break;
|
||||||
|
default:
|
||||||
|
ERR("Unhandled descriptor type %#x.\n", locations[0].src.vk_descriptor_type);
|
||||||
|
break;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
/* dst and src contain the same data unless another thread overwrites dst. The array index is
|
/* dst and src contain the same data unless another thread overwrites dst. The array index is
|
||||||
* calculated from dst, and src is thread safe. */
|
* calculated from dst, and src is thread safe. */
|
||||||
static void d3d12_desc_write_vk_heap(const struct d3d12_desc *dst, const struct d3d12_desc *src,
|
static void d3d12_desc_write_vk_heap(const struct d3d12_desc *dst, const struct d3d12_desc *src,
|
||||||
@ -2144,42 +2191,67 @@ static void d3d12_desc_write_vk_heap(const struct d3d12_desc *dst, const struct
|
|||||||
|
|
||||||
vkd3d_mutex_lock(&descriptor_heap->vk_sets_mutex);
|
vkd3d_mutex_lock(&descriptor_heap->vk_sets_mutex);
|
||||||
|
|
||||||
descriptor_set->vk_descriptor_write.dstArrayElement = dst
|
descriptor_set->vk_descriptor_writes[0].dstArrayElement = dst
|
||||||
- (const struct d3d12_desc *)descriptor_heap->descriptors;
|
- (const struct d3d12_desc *)descriptor_heap->descriptors;
|
||||||
|
descriptor_set->vk_descriptor_writes[0].descriptorCount = 1;
|
||||||
switch (src->vk_descriptor_type)
|
switch (src->vk_descriptor_type)
|
||||||
{
|
{
|
||||||
case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
|
case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
|
||||||
descriptor_set->vk_descriptor_write.pBufferInfo = &src->u.vk_cbv_info;
|
descriptor_set->vk_descriptor_writes[0].pBufferInfo = &src->u.vk_cbv_info;
|
||||||
break;
|
break;
|
||||||
case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
|
case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
|
||||||
case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
|
case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
|
||||||
descriptor_set->vk_image_info.imageView = src->u.view->u.vk_image_view;
|
descriptor_set->vk_image_infos[0].imageView = src->u.view->u.vk_image_view;
|
||||||
break;
|
break;
|
||||||
case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
|
case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
|
||||||
case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
|
case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
|
||||||
descriptor_set->vk_descriptor_write.pTexelBufferView = &src->u.view->u.vk_buffer_view;
|
descriptor_set->vk_descriptor_writes[0].pTexelBufferView = &src->u.view->u.vk_buffer_view;
|
||||||
break;
|
break;
|
||||||
case VK_DESCRIPTOR_TYPE_SAMPLER:
|
case VK_DESCRIPTOR_TYPE_SAMPLER:
|
||||||
descriptor_set->vk_image_info.sampler = src->u.view->u.vk_sampler;
|
descriptor_set->vk_image_infos[0].sampler = src->u.view->u.vk_sampler;
|
||||||
break;
|
break;
|
||||||
default:
|
default:
|
||||||
ERR("Unhandled descriptor type %#x.\n", src->vk_descriptor_type);
|
ERR("Unhandled descriptor type %#x.\n", src->vk_descriptor_type);
|
||||||
break;
|
break;
|
||||||
}
|
}
|
||||||
VK_CALL(vkUpdateDescriptorSets(device->vk_device, 1, &descriptor_set->vk_descriptor_write, 0, NULL));
|
VK_CALL(vkUpdateDescriptorSets(device->vk_device, 1, descriptor_set->vk_descriptor_writes, 0, NULL));
|
||||||
|
|
||||||
if (src->magic == VKD3D_DESCRIPTOR_MAGIC_UAV && src->u.view->vk_counter_view)
|
if (src->magic == VKD3D_DESCRIPTOR_MAGIC_UAV && src->u.view->vk_counter_view)
|
||||||
{
|
{
|
||||||
descriptor_set = &descriptor_heap->vk_descriptor_sets[VKD3D_SET_INDEX_UAV_COUNTER];
|
descriptor_set = &descriptor_heap->vk_descriptor_sets[VKD3D_SET_INDEX_UAV_COUNTER];
|
||||||
descriptor_set->vk_descriptor_write.dstArrayElement = dst
|
descriptor_set->vk_descriptor_writes[0].dstArrayElement = dst
|
||||||
- (const struct d3d12_desc *)descriptor_heap->descriptors;
|
- (const struct d3d12_desc *)descriptor_heap->descriptors;
|
||||||
descriptor_set->vk_descriptor_write.pTexelBufferView = &src->u.view->vk_counter_view;
|
descriptor_set->vk_descriptor_writes[0].descriptorCount = 1;
|
||||||
VK_CALL(vkUpdateDescriptorSets(device->vk_device, 1, &descriptor_set->vk_descriptor_write, 0, NULL));
|
descriptor_set->vk_descriptor_writes[0].pTexelBufferView = &src->u.view->vk_counter_view;
|
||||||
|
VK_CALL(vkUpdateDescriptorSets(device->vk_device, 1, descriptor_set->vk_descriptor_writes, 0, NULL));
|
||||||
}
|
}
|
||||||
|
|
||||||
vkd3d_mutex_unlock(&descriptor_heap->vk_sets_mutex);
|
vkd3d_mutex_unlock(&descriptor_heap->vk_sets_mutex);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
static void d3d12_desc_write_atomic_d3d12_only(struct d3d12_desc *dst, const struct d3d12_desc *src, struct d3d12_device *device)
|
||||||
|
{
|
||||||
|
struct vkd3d_view *defunct_view;
|
||||||
|
struct vkd3d_mutex *mutex;
|
||||||
|
|
||||||
|
mutex = d3d12_device_get_descriptor_mutex(device, dst);
|
||||||
|
vkd3d_mutex_lock(mutex);
|
||||||
|
|
||||||
|
if (!(dst->magic & VKD3D_DESCRIPTOR_MAGIC_HAS_VIEW) || InterlockedDecrement(&dst->u.view->refcount))
|
||||||
|
{
|
||||||
|
*dst = *src;
|
||||||
|
vkd3d_mutex_unlock(mutex);
|
||||||
|
return;
|
||||||
|
}
|
||||||
|
|
||||||
|
defunct_view = dst->u.view;
|
||||||
|
*dst = *src;
|
||||||
|
vkd3d_mutex_unlock(mutex);
|
||||||
|
|
||||||
|
/* Destroy the view after unlocking to reduce wait time. */
|
||||||
|
vkd3d_view_destroy(defunct_view, device);
|
||||||
|
}
|
||||||
|
|
||||||
void d3d12_desc_write_atomic(struct d3d12_desc *dst, const struct d3d12_desc *src,
|
void d3d12_desc_write_atomic(struct d3d12_desc *dst, const struct d3d12_desc *src,
|
||||||
struct d3d12_device *device)
|
struct d3d12_device *device)
|
||||||
{
|
{
|
||||||
@ -2213,6 +2285,56 @@ static void d3d12_desc_destroy(struct d3d12_desc *descriptor, struct d3d12_devic
|
|||||||
d3d12_desc_write_atomic(descriptor, &null_desc, device);
|
d3d12_desc_write_atomic(descriptor, &null_desc, device);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
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)
|
||||||
|
{
|
||||||
|
struct d3d12_descriptor_heap_vk_set *descriptor_set = &descriptor_heap->vk_descriptor_sets[set];
|
||||||
|
const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs;
|
||||||
|
unsigned int i, write_count;
|
||||||
|
|
||||||
|
vkd3d_mutex_lock(&descriptor_heap->vk_sets_mutex);
|
||||||
|
|
||||||
|
for (i = 0, write_count = 0; i < info->count; ++i)
|
||||||
|
{
|
||||||
|
d3d12_desc_write_atomic_d3d12_only(locations[i].dst, &locations[i].src, device);
|
||||||
|
|
||||||
|
if (i && locations[i].dst == locations[i - 1].dst + 1)
|
||||||
|
{
|
||||||
|
++descriptor_set->vk_descriptor_writes[write_count - 1].descriptorCount;
|
||||||
|
continue;
|
||||||
|
}
|
||||||
|
descriptor_set->vk_descriptor_writes[write_count].dstArrayElement = locations[i].dst
|
||||||
|
- (const struct d3d12_desc *)descriptor_heap->descriptors;
|
||||||
|
descriptor_set->vk_descriptor_writes[write_count++].descriptorCount = 1;
|
||||||
|
}
|
||||||
|
d3d12_descriptor_heap_write_vk_descriptor_range(descriptor_set, locations, write_count);
|
||||||
|
/* We could pass a VkCopyDescriptorSet array instead, but that would require also storing a src array index
|
||||||
|
* for each location, which means querying the src descriptor heap. Contiguous copies require contiguous src
|
||||||
|
* descriptors as well as dst, which is less likely to occur. And client race conditions may break it. */
|
||||||
|
VK_CALL(vkUpdateDescriptorSets(device->vk_device, write_count, descriptor_set->vk_descriptor_writes, 0, NULL));
|
||||||
|
|
||||||
|
if (!info->uav_counter)
|
||||||
|
goto done;
|
||||||
|
|
||||||
|
descriptor_set = &descriptor_heap->vk_descriptor_sets[VKD3D_SET_INDEX_UAV_COUNTER];
|
||||||
|
|
||||||
|
for (i = 0, write_count = 0; i < info->count; ++i)
|
||||||
|
{
|
||||||
|
if (!locations[i].src.u.view->vk_counter_view)
|
||||||
|
continue;
|
||||||
|
descriptor_set->vk_buffer_views[write_count] = locations[i].src.u.view->vk_counter_view;
|
||||||
|
descriptor_set->vk_descriptor_writes[write_count].pTexelBufferView = &descriptor_set->vk_buffer_views[write_count];
|
||||||
|
descriptor_set->vk_descriptor_writes[write_count].dstArrayElement = locations[i].dst
|
||||||
|
- (const struct d3d12_desc *)descriptor_heap->descriptors;
|
||||||
|
descriptor_set->vk_descriptor_writes[write_count++].descriptorCount = 1;
|
||||||
|
}
|
||||||
|
VK_CALL(vkUpdateDescriptorSets(device->vk_device, write_count, descriptor_set->vk_descriptor_writes, 0, NULL));
|
||||||
|
|
||||||
|
done:
|
||||||
|
vkd3d_mutex_unlock(&descriptor_heap->vk_sets_mutex);
|
||||||
|
}
|
||||||
|
|
||||||
void d3d12_desc_copy(struct d3d12_desc *dst, const struct d3d12_desc *src,
|
void d3d12_desc_copy(struct d3d12_desc *dst, const struct d3d12_desc *src,
|
||||||
struct d3d12_device *device)
|
struct d3d12_device *device)
|
||||||
{
|
{
|
||||||
@ -3694,6 +3816,7 @@ static HRESULT d3d12_descriptor_heap_create_descriptor_set(struct d3d12_descript
|
|||||||
const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs;
|
const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs;
|
||||||
VkDescriptorSetVariableDescriptorCountAllocateInfoEXT set_size;
|
VkDescriptorSetVariableDescriptorCountAllocateInfoEXT set_size;
|
||||||
VkDescriptorSetAllocateInfo set_desc;
|
VkDescriptorSetAllocateInfo set_desc;
|
||||||
|
unsigned int i;
|
||||||
VkResult vr;
|
VkResult vr;
|
||||||
|
|
||||||
set_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO;
|
set_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO;
|
||||||
@ -3707,7 +3830,8 @@ static HRESULT d3d12_descriptor_heap_create_descriptor_set(struct d3d12_descript
|
|||||||
set_size.pDescriptorCounts = &variable_binding_size;
|
set_size.pDescriptorCounts = &variable_binding_size;
|
||||||
if ((vr = VK_CALL(vkAllocateDescriptorSets(device->vk_device, &set_desc, &descriptor_set->vk_set))) >= 0)
|
if ((vr = VK_CALL(vkAllocateDescriptorSets(device->vk_device, &set_desc, &descriptor_set->vk_set))) >= 0)
|
||||||
{
|
{
|
||||||
descriptor_set->vk_descriptor_write.dstSet = descriptor_set->vk_set;
|
for (i = 0; i < ARRAY_SIZE(descriptor_set->vk_descriptor_writes); ++i)
|
||||||
|
descriptor_set->vk_descriptor_writes[i].dstSet = descriptor_set->vk_set;
|
||||||
return S_OK;
|
return S_OK;
|
||||||
}
|
}
|
||||||
|
|
||||||
@ -3735,15 +3859,18 @@ static HRESULT d3d12_descriptor_heap_vk_descriptor_sets_init(struct d3d12_descri
|
|||||||
for (set = 0; set < ARRAY_SIZE(descriptor_heap->vk_descriptor_sets); ++set)
|
for (set = 0; set < ARRAY_SIZE(descriptor_heap->vk_descriptor_sets); ++set)
|
||||||
{
|
{
|
||||||
struct d3d12_descriptor_heap_vk_set *descriptor_set = &descriptor_heap->vk_descriptor_sets[set];
|
struct d3d12_descriptor_heap_vk_set *descriptor_set = &descriptor_heap->vk_descriptor_sets[set];
|
||||||
|
unsigned int i;
|
||||||
|
|
||||||
descriptor_set->vk_descriptor_write.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
|
for (i = 0; i < ARRAY_SIZE(descriptor_set->vk_descriptor_writes); ++i)
|
||||||
descriptor_set->vk_descriptor_write.pNext = NULL;
|
{
|
||||||
descriptor_set->vk_descriptor_write.dstBinding = 0;
|
descriptor_set->vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
|
||||||
descriptor_set->vk_descriptor_write.descriptorCount = 1;
|
descriptor_set->vk_descriptor_writes[i].pNext = NULL;
|
||||||
descriptor_set->vk_descriptor_write.descriptorType = device->vk_descriptor_heap_layouts[set].type;
|
descriptor_set->vk_descriptor_writes[i].dstBinding = 0;
|
||||||
descriptor_set->vk_descriptor_write.pImageInfo = NULL;
|
descriptor_set->vk_descriptor_writes[i].descriptorType = device->vk_descriptor_heap_layouts[set].type;
|
||||||
descriptor_set->vk_descriptor_write.pBufferInfo = NULL;
|
descriptor_set->vk_descriptor_writes[i].pImageInfo = NULL;
|
||||||
descriptor_set->vk_descriptor_write.pTexelBufferView = NULL;
|
descriptor_set->vk_descriptor_writes[i].pBufferInfo = NULL;
|
||||||
|
descriptor_set->vk_descriptor_writes[i].pTexelBufferView = NULL;
|
||||||
|
}
|
||||||
switch (device->vk_descriptor_heap_layouts[set].type)
|
switch (device->vk_descriptor_heap_layouts[set].type)
|
||||||
{
|
{
|
||||||
case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
|
case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
|
||||||
@ -3751,19 +3878,28 @@ static HRESULT d3d12_descriptor_heap_vk_descriptor_sets_init(struct d3d12_descri
|
|||||||
case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
|
case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
|
||||||
break;
|
break;
|
||||||
case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
|
case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
|
||||||
descriptor_set->vk_descriptor_write.pImageInfo = &descriptor_set->vk_image_info;
|
descriptor_set->vk_descriptor_writes[0].pImageInfo = &descriptor_set->vk_image_infos[0];
|
||||||
descriptor_set->vk_image_info.sampler = VK_NULL_HANDLE;
|
for (i = 0; i < ARRAY_SIZE(descriptor_set->vk_image_infos); ++i)
|
||||||
descriptor_set->vk_image_info.imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
|
{
|
||||||
|
descriptor_set->vk_image_infos[i].sampler = VK_NULL_HANDLE;
|
||||||
|
descriptor_set->vk_image_infos[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
|
||||||
|
}
|
||||||
break;
|
break;
|
||||||
case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
|
case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
|
||||||
descriptor_set->vk_descriptor_write.pImageInfo = &descriptor_set->vk_image_info;
|
descriptor_set->vk_descriptor_writes[0].pImageInfo = &descriptor_set->vk_image_infos[0];
|
||||||
descriptor_set->vk_image_info.sampler = VK_NULL_HANDLE;
|
for (i = 0; i < ARRAY_SIZE(descriptor_set->vk_image_infos); ++i)
|
||||||
descriptor_set->vk_image_info.imageLayout = VK_IMAGE_LAYOUT_GENERAL;
|
{
|
||||||
|
descriptor_set->vk_image_infos[i].sampler = VK_NULL_HANDLE;
|
||||||
|
descriptor_set->vk_image_infos[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
|
||||||
|
}
|
||||||
break;
|
break;
|
||||||
case VK_DESCRIPTOR_TYPE_SAMPLER:
|
case VK_DESCRIPTOR_TYPE_SAMPLER:
|
||||||
descriptor_set->vk_descriptor_write.pImageInfo = &descriptor_set->vk_image_info;
|
descriptor_set->vk_descriptor_writes[0].pImageInfo = &descriptor_set->vk_image_infos[0];
|
||||||
descriptor_set->vk_image_info.imageView = VK_NULL_HANDLE;
|
for (i = 0; i < ARRAY_SIZE(descriptor_set->vk_image_infos); ++i)
|
||||||
descriptor_set->vk_image_info.imageLayout = VK_IMAGE_LAYOUT_UNDEFINED;
|
{
|
||||||
|
descriptor_set->vk_image_infos[i].imageView = VK_NULL_HANDLE;
|
||||||
|
descriptor_set->vk_image_infos[i].imageLayout = VK_IMAGE_LAYOUT_UNDEFINED;
|
||||||
|
}
|
||||||
break;
|
break;
|
||||||
default:
|
default:
|
||||||
ERR("Unhandled descriptor type %#x.\n", device->vk_descriptor_heap_layouts[set].type);
|
ERR("Unhandled descriptor type %#x.\n", device->vk_descriptor_heap_layouts[set].type);
|
||||||
|
@ -805,11 +805,15 @@ struct vkd3d_vk_descriptor_heap_layout
|
|||||||
VkDescriptorSetLayout vk_set_layout;
|
VkDescriptorSetLayout vk_set_layout;
|
||||||
};
|
};
|
||||||
|
|
||||||
|
#define VKD3D_DESCRIPTOR_WRITE_BUFFER_SIZE 64
|
||||||
|
|
||||||
struct d3d12_descriptor_heap_vk_set
|
struct d3d12_descriptor_heap_vk_set
|
||||||
{
|
{
|
||||||
VkDescriptorSet vk_set;
|
VkDescriptorSet vk_set;
|
||||||
VkDescriptorImageInfo vk_image_info;
|
VkDescriptorBufferInfo vk_buffer_infos[VKD3D_DESCRIPTOR_WRITE_BUFFER_SIZE];
|
||||||
VkWriteDescriptorSet vk_descriptor_write;
|
VkBufferView vk_buffer_views[VKD3D_DESCRIPTOR_WRITE_BUFFER_SIZE];
|
||||||
|
VkDescriptorImageInfo vk_image_infos[VKD3D_DESCRIPTOR_WRITE_BUFFER_SIZE];
|
||||||
|
VkWriteDescriptorSet vk_descriptor_writes[VKD3D_DESCRIPTOR_WRITE_BUFFER_SIZE];
|
||||||
};
|
};
|
||||||
|
|
||||||
/* ID3D12DescriptorHeap */
|
/* ID3D12DescriptorHeap */
|
||||||
@ -835,6 +839,22 @@ struct d3d12_descriptor_heap
|
|||||||
HRESULT d3d12_descriptor_heap_create(struct d3d12_device *device,
|
HRESULT d3d12_descriptor_heap_create(struct d3d12_device *device,
|
||||||
const D3D12_DESCRIPTOR_HEAP_DESC *desc, struct d3d12_descriptor_heap **descriptor_heap);
|
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 */
|
/* ID3D12QueryHeap */
|
||||||
struct d3d12_query_heap
|
struct d3d12_query_heap
|
||||||
{
|
{
|
||||||
|
Loading…
Reference in New Issue
Block a user