[PATCH v2 2/4] vkd3d: Delay writing Vulkan descriptors until submitted to a queue.

[Conor McCarthy]

From: Conor McCarthy

Eliminates vk_sets_mutex. Performance on average may be lower until
the descriptor mutexes are replaced and Vulkan writes are buffered
to reduce thunk calls.
---
 configure.ac                   |   2 +
 include/private/vkd3d_common.h |  27 +++
 libs/vkd3d/command.c           |  50 ++++-
 libs/vkd3d/device.c            | 155 +---------------
 libs/vkd3d/resource.c          | 329 +++++++++++++--------------------
 libs/vkd3d/vkd3d_private.h     |  14 +-
 6 files changed, 210 insertions(+), 367 deletions(-)

diff --git a/configure.ac b/configure.ac
index 29225c8c..93c3d585 100644
--- a/configure.ac
+++ b/configure.ac
@@ -142,6 +142,8 @@ VKD3D_CHECK_FUNC([HAVE_BUILTIN_POPCOUNT], [__builtin_popcount], [__builtin_popco
 VKD3D_CHECK_FUNC([HAVE_BUILTIN_ADD_OVERFLOW], [__builtin_add_overflow], [__builtin_add_overflow(0, 0, (int *)0)])
 VKD3D_CHECK_FUNC([HAVE_SYNC_ADD_AND_FETCH], [__sync_add_and_fetch], [__sync_add_and_fetch((int *)0, 0)])
 VKD3D_CHECK_FUNC([HAVE_SYNC_SUB_AND_FETCH], [__sync_sub_and_fetch], [__sync_sub_and_fetch((int *)0, 0)])
+VKD3D_CHECK_FUNC([HAVE_SYNC_COMPARE_AND_SWAP], [__sync_val_compare_and_swap], [__sync_val_compare_and_swap((int *)0, 0, 0)])
+VKD3D_CHECK_FUNC([HAVE_ATOMIC_EXCHANGE_N], [__atomic_exchange_n], [__atomic_exchange_n((int *)0, 0)])

dnl Makefiles
 case $host_os in
diff --git a/include/private/vkd3d_common.h b/include/private/vkd3d_common.h
index 78a15c19..017f3f41 100644
--- a/include/private/vkd3d_common.h
+++ b/include/private/vkd3d_common.h
@@ -249,6 +249,33 @@ static inline LONG InterlockedDecrement(LONG volatile *x)
 # else
 #  error "InterlockedDecrement() not implemented for this platform"
 # endif
+
+# if HAVE_SYNC_COMPARE_AND_SWAP
+static inline LONG InterlockedCompareExchange(LONG volatile *x, LONG xchg, LONG cmp)
+{
+    return __sync_val_compare_and_swap(x, cmp, xchg);
+}
+# else
+#  error "InterlockedCompareExchange() not implemented for this platform"
+# endif
+# if HAVE_ATOMIC_EXCHANGE_N
+static inline LONG InterlockedExchange(LONG volatile *x, LONG 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)
+{
+    LONG i;
+    do
+    {
+        i = *x;
+    } while (__sync_val_compare_and_swap(x, i, val) != i);
+    return i;
+}
+# else
+#   error "InterlockedExchange() not implemented for this platform"
+# endif
 #endif  /* _WIN32 */

static inline void vkd3d_parse_version(const char *version, int *major, int *minor)
diff --git a/libs/vkd3d/command.c b/libs/vkd3d/command.c
index 6eddcfa2..3ba1f608 100644
--- a/libs/vkd3d/command.c
+++ b/libs/vkd3d/command.c
@@ -3177,12 +3177,6 @@ static void d3d12_command_list_bind_descriptor_heap(struct d3d12_command_list *l
         bindings->sampler_heap_id = heap->serial_id;
     }

-    /* These sets can be shared across multiple command lists, and therefore binding must
-     * be synchronised. On an experimental branch in which caching of Vk descriptor writes
-     * greatly increased the chance of multiple threads arriving here at the same time,
-     * GRID 2019 crashed without the mutex lock. */
-    vkd3d_mutex_lock(&heap->vk_sets_mutex);
-
     for (set = 0; set < ARRAY_SIZE(heap->vk_descriptor_sets); ++set)
     {
         VkDescriptorSet vk_descriptor_set = heap->vk_descriptor_sets[set].vk_set;
@@ -3193,8 +3187,6 @@ static void d3d12_command_list_bind_descriptor_heap(struct d3d12_command_list *l
         VK_CALL(vkCmdBindDescriptorSets(list->vk_command_buffer, bindings->vk_bind_point, rs->vk_pipeline_layout,
                 rs->vk_set_count + set, 1, &vk_descriptor_set, 0, NULL));
     }
-
-    vkd3d_mutex_unlock(&heap->vk_sets_mutex);
 }

static void d3d12_command_list_update_heap_descriptors(struct d3d12_command_list *list,
@@ -6166,6 +6158,44 @@ static void d3d12_command_queue_submit_locked(struct d3d12_command_queue *queue)
     }
 }

+static bool contains_heap(const struct d3d12_descriptor_heap **heap_array, unsigned int count,
+        const struct d3d12_descriptor_heap *query)
+{
+    unsigned int i;
+
+    for (i = 0; i < count; ++i)
+        if (heap_array[i] == query)
+            return true;
+    return false;
+}
+
+static void pipeline_bindings_flush_vk_heap_updates(struct vkd3d_pipeline_bindings *bindings,
+        struct d3d12_device *device)
+{
+    /* Only two heaps are strictly allowed, but more could be supported with a hack. */
+    const struct d3d12_descriptor_heap *heap_array[3];
+    struct d3d12_descriptor_heap *descriptor_heap;
+    unsigned int i, count;
+    uint64_t mask;
+
+    mask = bindings->descriptor_table_active_mask & bindings->root_signature->descriptor_table_mask;
+
+    for (i = 0, count = 0; i < ARRAY_SIZE(bindings->descriptor_tables); ++i)
+    {
+        if (!(mask & (1ull << i)) || !bindings->descriptor_tables[i])
+            continue;
+
+        descriptor_heap = d3d12_desc_get_descriptor_heap(bindings->descriptor_tables[i]);
+        /* Another thread could be writing unused descriptors, so try to check each heap only once. Flushing
+         * any updates added after the first flush will only delay execution of the command list. */
+        if (contains_heap(heap_array, count, descriptor_heap))
+            continue;
+        if (count < ARRAY_SIZE(heap_array))
+            heap_array[count++] = descriptor_heap;
+        d3d12_desc_flush_vk_heap_updates(descriptor_heap, device);
+    }
+}
+
 static void STDMETHODCALLTYPE d3d12_command_queue_ExecuteCommandLists(ID3D12CommandQueue *iface,
         UINT command_list_count, ID3D12CommandList * const *command_lists)
 {
@@ -6199,6 +6229,10 @@ static void STDMETHODCALLTYPE d3d12_command_queue_ExecuteCommandLists(ID3D12Comm
             return;
         }

+        if (cmd_list->state)
+            pipeline_bindings_flush_vk_heap_updates(&cmd_list->pipeline_bindings[cmd_list->state->vk_bind_point],
+                    cmd_list->device);
+
         buffers[i] = cmd_list->vk_command_buffer;
     }

diff --git a/libs/vkd3d/device.c b/libs/vkd3d/device.c
index 39a5ca01..d9555d1d 100644
--- a/libs/vkd3d/device.c
+++ b/libs/vkd3d/device.c
@@ -3368,132 +3368,6 @@ static void STDMETHODCALLTYPE d3d12_device_CreateSampler(ID3D12Device *iface,
     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->s.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->s.vk_descriptor_type);
-    location = &locations[set][infos[set].count++];
-
-    location->src.s = src->s;
-
-    if (location->src.s.magic & VKD3D_DESCRIPTOR_MAGIC_HAS_VIEW)
-        vkd3d_view_incref(location->src.s.u.view_info.view);
-
-    vkd3d_mutex_unlock(mutex);
-
-    infos[set].uav_counter |= (location->src.s.magic == VKD3D_DESCRIPTOR_MAGIC_UAV)
-            && !!location->src.s.u.view_info.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 = d3d12_desc_get_descriptor_heap(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 = d3d12_desc_get_descriptor_heap(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++)
-        {
-            /* We don't need to lock either descriptor for the identity check. The descriptor
-             * mutex is only intended to prevent use-after-free of the vkd3d_view caused by a
-             * race condition in the calling app. It is unnecessary to protect this test as it's
-             * the app's race condition, not ours. */
-            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)
-                continue;
-            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,
         UINT dst_descriptor_range_count, const D3D12_CPU_DESCRIPTOR_HANDLE *dst_descriptor_range_offsets,
         const UINT *dst_descriptor_range_sizes,
@@ -3525,15 +3399,6 @@ static void STDMETHODCALLTYPE d3d12_device_CopyDescriptors(ID3D12Device *iface,
     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;
     src_range_idx = src_idx = 0;
     while (dst_range_idx < dst_descriptor_range_count && src_range_idx < src_descriptor_range_count)
@@ -3544,8 +3409,13 @@ static void STDMETHODCALLTYPE d3d12_device_CopyDescriptors(ID3D12Device *iface,
         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]);

-        while (dst_idx < dst_range_size && src_idx < src_range_size)
-            d3d12_desc_copy(&dst[dst_idx++], &src[src_idx++], device);
+        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)
+                continue;
+            d3d12_desc_copy(&dst[dst_idx], &src[src_idx], device);
+        }

if (dst_idx >= dst_range_size)
         {
@@ -3570,17 +3440,6 @@ static void STDMETHODCALLTYPE d3d12_device_CopyDescriptorsSimple(ID3D12Device *i
             iface, descriptor_count, dst_descriptor_range_offset.ptr, src_descriptor_range_offset.ptr,
             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,
             1, &src_descriptor_range_offset, &descriptor_count, descriptor_heap_type);
 }
diff --git a/libs/vkd3d/resource.c b/libs/vkd3d/resource.c
index c45406c5..719da18c 100644
--- a/libs/vkd3d/resource.c
+++ b/libs/vkd3d/resource.c
@@ -2097,179 +2097,201 @@ void vkd3d_view_decref(struct vkd3d_view *view, struct d3d12_device *device)
         vkd3d_view_destroy(view, device);
 }

-/* TODO: write null descriptors to all applicable sets (invalid behaviour workaround). */
-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.s.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.s.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.s.u.view_info.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.s.u.view_info.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.s.u.view_info.view->u.vk_sampler;
-            }
-            break;
-        default:
-            ERR("Unhandled descriptor type %#x.\n", locations[0].src.s.vk_descriptor_type);
-            break;
-    }
-}
-
 static void d3d12_desc_write_vk_heap_null_descriptor(struct d3d12_descriptor_heap *descriptor_heap,
-        uint32_t dst_array_element, const struct d3d12_device *device)
+        uint32_t dst_array_element, struct d3d12_device *device)
 {
+    VkWriteDescriptorSet vk_descriptor_writes[VKD3D_SET_INDEX_STORAGE_IMAGE + 1];
     const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs;
     struct d3d12_descriptor_heap_vk_set *descriptor_set;
     VkBufferView vk_buffer_view = VK_NULL_HANDLE;
+    VkDescriptorImageInfo vk_image_infos[2];
     enum vkd3d_vk_descriptor_set_index i;
     VkDescriptorBufferInfo vk_cbv_info;

vk_cbv_info.buffer = VK_NULL_HANDLE;
     vk_cbv_info.offset = 0;
     vk_cbv_info.range = VK_WHOLE_SIZE;
+    memset(vk_image_infos, 0, sizeof(vk_image_infos));
+    vk_image_infos[0].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+    vk_image_infos[1].imageLayout = VK_IMAGE_LAYOUT_GENERAL;

/* Binding a shader with the wrong null descriptor type works in Windows.
      * To support that here we must write one to all applicable Vulkan sets. */
     for (i = VKD3D_SET_INDEX_UNIFORM_BUFFER; i <= VKD3D_SET_INDEX_STORAGE_IMAGE; ++i)
     {
         descriptor_set = &descriptor_heap->vk_descriptor_sets[i];
-        descriptor_set->vk_descriptor_writes[0].dstArrayElement = dst_array_element;
-        descriptor_set->vk_descriptor_writes[0].descriptorCount = 1;
+        vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+        vk_descriptor_writes[i].pNext = NULL;
+        vk_descriptor_writes[i].dstSet = descriptor_set->vk_set;
+        vk_descriptor_writes[i].dstBinding = 0;
+        vk_descriptor_writes[i].dstArrayElement = dst_array_element;
+        vk_descriptor_writes[i].descriptorCount = 1;
+        vk_descriptor_writes[i].descriptorType = descriptor_set->vk_type;
         switch (i)
         {
             case VKD3D_SET_INDEX_UNIFORM_BUFFER:
-                descriptor_set->vk_descriptor_writes[0].pBufferInfo = &vk_cbv_info;
+                vk_descriptor_writes[i].pImageInfo = NULL;
+                vk_descriptor_writes[i].pBufferInfo = &vk_cbv_info;
+                vk_descriptor_writes[i].pTexelBufferView = NULL;
                 break;
             case VKD3D_SET_INDEX_SAMPLED_IMAGE:
             case VKD3D_SET_INDEX_STORAGE_IMAGE:
-                descriptor_set->vk_image_infos[0].imageView = VK_NULL_HANDLE;
+                vk_descriptor_writes[i].pImageInfo = &vk_image_infos[i == VKD3D_SET_INDEX_STORAGE_IMAGE];
+                vk_descriptor_writes[i].pBufferInfo = NULL;
+                vk_descriptor_writes[i].pTexelBufferView = NULL;
                 break;
             case VKD3D_SET_INDEX_UNIFORM_TEXEL_BUFFER:
             case VKD3D_SET_INDEX_STORAGE_TEXEL_BUFFER:
-                descriptor_set->vk_descriptor_writes[0].pTexelBufferView = &vk_buffer_view;
+                vk_descriptor_writes[i].pImageInfo = NULL;
+                vk_descriptor_writes[i].pBufferInfo = NULL;
+                vk_descriptor_writes[i].pTexelBufferView = &vk_buffer_view;
                 break;
             default:
                 assert(false);
                 break;
         }
-        VK_CALL(vkUpdateDescriptorSets(device->vk_device, 1, descriptor_set->vk_descriptor_writes, 0, NULL));
     }
+
+    VK_CALL(vkUpdateDescriptorSets(device->vk_device, i, vk_descriptor_writes, 0, NULL));
 }

-/* dst and src contain the same data unless another thread overwrites dst. The array index is
- * 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,
-        struct d3d12_device *device)
+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)
 {
     struct d3d12_descriptor_heap_vk_set *descriptor_set;
-    struct d3d12_descriptor_heap *descriptor_heap;
     const struct vkd3d_vk_device_procs *vk_procs;
+    VkWriteDescriptorSet vk_descriptor_writes[2];
+    VkDescriptorImageInfo vk_image_info;
+    unsigned int count = 1;
+    VkDescriptorType type;
     bool is_null = false;

-    descriptor_heap = d3d12_desc_get_descriptor_heap(dst);
-    descriptor_set = &descriptor_heap->vk_descriptor_sets[vkd3d_vk_descriptor_set_index_from_vk_descriptor_type(
-            src->s.vk_descriptor_type)];
+    type = src->s.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;

-    vkd3d_mutex_lock(&descriptor_heap->vk_sets_mutex);
-
-    descriptor_set->vk_descriptor_writes[0].dstArrayElement = dst->index;
-    descriptor_set->vk_descriptor_writes[0].descriptorCount = 1;
-    switch (src->s.vk_descriptor_type)
+    vk_descriptor_writes[0].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+    vk_descriptor_writes[0].pNext = NULL;
+    vk_descriptor_writes[0].dstSet = descriptor_set->vk_set;
+    vk_descriptor_writes[0].dstBinding = 0;
+    vk_descriptor_writes[0].dstArrayElement = dst_array_element;
+    vk_descriptor_writes[0].descriptorCount = 1;
+    vk_descriptor_writes[0].descriptorType = type;
+    switch (type)
     {
         case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
-            descriptor_set->vk_descriptor_writes[0].pBufferInfo = &src->s.u.vk_cbv_info;
+            vk_descriptor_writes[0].pImageInfo = NULL;
+            vk_descriptor_writes[0].pBufferInfo = &src->s.u.vk_cbv_info;
+            vk_descriptor_writes[0].pTexelBufferView = NULL;
             is_null = !src->s.u.vk_cbv_info.buffer;
             break;
         case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
         case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
-            is_null = !(descriptor_set->vk_image_infos[0].imageView = src->s.u.view_info.view->u.vk_image_view);
+            vk_descriptor_writes[0].pImageInfo = &vk_image_info;
+            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);
+            vk_image_info.imageLayout = (type == VK_DESCRIPTOR_TYPE_STORAGE_IMAGE)
+                    ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
             break;
         case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
         case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
-            descriptor_set->vk_descriptor_writes[0].pTexelBufferView = &src->s.u.view_info.view->u.vk_buffer_view;
+            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;
             break;
         case VK_DESCRIPTOR_TYPE_SAMPLER:
-            descriptor_set->vk_image_infos[0].sampler = src->s.u.view_info.view->u.vk_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.imageView = VK_NULL_HANDLE;
+            vk_image_info.imageLayout = VK_IMAGE_LAYOUT_UNDEFINED;
             break;
         default:
-            ERR("Unhandled descriptor type %#x.\n", src->s.vk_descriptor_type);
+            ERR("Unhandled descriptor type %#x.\n", type);
             break;
     }
     if (is_null && device->vk_info.EXT_robustness2)
-    {
-        d3d12_desc_write_vk_heap_null_descriptor(descriptor_heap,
-                descriptor_set->vk_descriptor_writes[0].dstArrayElement, device);
-        vkd3d_mutex_unlock(&descriptor_heap->vk_sets_mutex);
-        return;
-    }
-
-    VK_CALL(vkUpdateDescriptorSets(device->vk_device, 1, descriptor_set->vk_descriptor_writes, 0, NULL));
+        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)
     {
         descriptor_set = &descriptor_heap->vk_descriptor_sets[VKD3D_SET_INDEX_UAV_COUNTER];
-        descriptor_set->vk_descriptor_writes[0].dstArrayElement = dst->index;
-        descriptor_set->vk_descriptor_writes[0].descriptorCount = 1;
-        descriptor_set->vk_descriptor_writes[0].pTexelBufferView = &src->s.u.view_info.view->vk_counter_view;
-        VK_CALL(vkUpdateDescriptorSets(device->vk_device, 1, descriptor_set->vk_descriptor_writes, 0, NULL));
+        vk_descriptor_writes[1].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+        vk_descriptor_writes[1].pNext = NULL;
+        vk_descriptor_writes[1].dstSet = descriptor_set->vk_set;
+        vk_descriptor_writes[1].dstBinding = 0;
+        vk_descriptor_writes[1].dstArrayElement = dst_array_element;
+        vk_descriptor_writes[1].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER;
+        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;
+        ++count;
     }

-    vkd3d_mutex_unlock(&descriptor_heap->vk_sets_mutex);
+    VK_CALL(vkUpdateDescriptorSets(device->vk_device, count, vk_descriptor_writes, 0, NULL));
 }

-static void d3d12_desc_write_atomic_d3d12_only(struct d3d12_desc *dst, const struct d3d12_desc *src, struct d3d12_device *device)
+void d3d12_desc_flush_vk_heap_updates(struct d3d12_descriptor_heap *descriptor_heap, struct d3d12_device *device)
 {
-    struct vkd3d_view *defunct_view;
+    struct d3d12_desc *descriptors, *src;
     struct vkd3d_mutex *mutex;
+    struct d3d12_desc tmp;
+    unsigned int i, next;

-    mutex = d3d12_device_get_descriptor_mutex(device, dst);
-    vkd3d_mutex_lock(mutex);
+    i = InterlockedExchange(&descriptor_heap->dirty_list_head, UINT_MAX);
+
+    descriptors = (struct d3d12_desc *)descriptor_heap->descriptors;

-    if (!(dst->s.magic & VKD3D_DESCRIPTOR_MAGIC_HAS_VIEW) || InterlockedDecrement(&dst->s.u.view_info.view->refcount))
+    for (; i != UINT_MAX; i = next)
     {
-        d3d12_desc_copy_raw(dst, src);
+        src = &descriptors[i];
+        next = src->next >> 1;
+
+        mutex = d3d12_device_get_descriptor_mutex(device, src);
+        vkd3d_mutex_lock(mutex);
+
+        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);
-        return;
+
+        if (tmp.s.magic)
+        {
+            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);
     }
+}

-    defunct_view = dst->s.u.view_info.view;
-    d3d12_desc_copy_raw(dst, src);
-    vkd3d_mutex_unlock(mutex);
+static void d3d12_desc_mark_as_modified(struct d3d12_desc *dst)
+{
+    struct d3d12_descriptor_heap *descriptor_heap;
+    unsigned int i;
+    LONG head;

-    /* Destroy the view after unlocking to reduce wait time. */
-    vkd3d_view_destroy(defunct_view, device);
+    i = dst->index;
+    descriptor_heap = d3d12_desc_get_descriptor_heap(dst);
+    head = descriptor_heap->dirty_list_head;
+
+    /* Only one thread can swap the value away from zero. */
+    if (InterlockedCompareExchange(&dst->next, (head << 1) | 1, 0))
+        return;
+    /* Now it is safe to modify 'next' to another nonzero value if necessary. */
+    while (InterlockedCompareExchange(&descriptor_heap->dirty_list_head, i, head) != head)
+    {
+        head = descriptor_heap->dirty_list_head;
+        InterlockedExchange(&dst->next, (head << 1) | 1);
+    }
 }

void d3d12_desc_write_atomic(struct d3d12_desc *dst, const struct d3d12_desc *src,
@@ -2294,8 +2316,8 @@ void d3d12_desc_write_atomic(struct d3d12_desc *dst, const struct d3d12_desc *sr
     if (defunct_view)
         vkd3d_view_destroy(defunct_view, device);

-    if (device->use_vk_heaps && dst->s.magic)
-        d3d12_desc_write_vk_heap(dst, src, device);
+    if (device->use_vk_heaps && dst->s.magic && !dst->next)
+        d3d12_desc_mark_as_modified(dst);
 }

static void d3d12_desc_destroy(struct d3d12_desc *descriptor, struct d3d12_device *device)
@@ -2305,58 +2327,6 @@ static void d3d12_desc_destroy(struct d3d12_desc *descriptor, struct d3d12_devic
     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;
-        }
-        /* Accessing dst->index will be slow if a cache miss occurs, so calculate instead. */
-        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.s.u.view_info.view->vk_counter_view)
-            continue;
-        descriptor_set->vk_buffer_views[write_count] = locations[i].src.s.u.view_info.view->vk_counter_view;
-        descriptor_set->vk_descriptor_writes[write_count].pTexelBufferView = &descriptor_set->vk_buffer_views[write_count];
-        /* Accessing dst->index will be slow if a cache miss occurs, so calculate instead. */
-        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,
         struct d3d12_device *device)
 {
@@ -3705,7 +3675,6 @@ static ULONG STDMETHODCALLTYPE d3d12_descriptor_heap_Release(ID3D12DescriptorHea
         }

VK_CALL(vkDestroyDescriptorPool(device->vk_device, heap->vk_descriptor_pool, NULL));
-        vkd3d_mutex_destroy(&heap->vk_sets_mutex);

vkd3d_free(heap);

@@ -3866,7 +3835,6 @@ static HRESULT d3d12_descriptor_heap_create_descriptor_set(struct d3d12_descript
     const struct vkd3d_vk_device_procs *vk_procs = &device->vk_procs;
     VkDescriptorSetVariableDescriptorCountAllocateInfoEXT set_size;
     VkDescriptorSetAllocateInfo set_desc;
-    unsigned int i;
     VkResult vr;

set_desc.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO;
@@ -3880,8 +3848,7 @@ static HRESULT d3d12_descriptor_heap_create_descriptor_set(struct d3d12_descript
     set_size.pDescriptorCounts = &variable_binding_size;
     if ((vr = VK_CALL(vkAllocateDescriptorSets(device->vk_device, &set_desc, &descriptor_set->vk_set))) >= 0)
     {
-        for (i = 0; i < ARRAY_SIZE(descriptor_set->vk_descriptor_writes); ++i)
-            descriptor_set->vk_descriptor_writes[i].dstSet = descriptor_set->vk_set;
+        descriptor_set->vk_type = device->vk_descriptor_heap_layouts[set].type;
         return S_OK;
     }

@@ -3897,7 +3864,6 @@ static HRESULT d3d12_descriptor_heap_vk_descriptor_sets_init(struct d3d12_descri

descriptor_heap->vk_descriptor_pool = VK_NULL_HANDLE;
     memset(descriptor_heap->vk_descriptor_sets, 0, sizeof(descriptor_heap->vk_descriptor_sets));
-    vkd3d_mutex_init(&descriptor_heap->vk_sets_mutex);

if (!device->use_vk_heaps || (desc->Type != D3D12_DESCRIPTOR_HEAP_TYPE_CBV_SRV_UAV
             && desc->Type != D3D12_DESCRIPTOR_HEAP_TYPE_SAMPLER))
@@ -3908,53 +3874,6 @@ static HRESULT d3d12_descriptor_heap_vk_descriptor_sets_init(struct d3d12_descri

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];
-        unsigned int i;
-
-        for (i = 0; i < ARRAY_SIZE(descriptor_set->vk_descriptor_writes); ++i)
-        {
-            descriptor_set->vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
-            descriptor_set->vk_descriptor_writes[i].pNext = NULL;
-            descriptor_set->vk_descriptor_writes[i].dstBinding = 0;
-            descriptor_set->vk_descriptor_writes[i].descriptorType = device->vk_descriptor_heap_layouts[set].type;
-            descriptor_set->vk_descriptor_writes[i].pImageInfo = 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)
-        {
-            case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
-            case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
-            case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
-                break;
-            case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
-                descriptor_set->vk_descriptor_writes[0].pImageInfo = &descriptor_set->vk_image_infos[0];
-                for (i = 0; i < ARRAY_SIZE(descriptor_set->vk_image_infos); ++i)
-                {
-                    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;
-            case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
-                descriptor_set->vk_descriptor_writes[0].pImageInfo = &descriptor_set->vk_image_infos[0];
-                for (i = 0; i < ARRAY_SIZE(descriptor_set->vk_image_infos); ++i)
-                {
-                    descriptor_set->vk_image_infos[i].sampler = VK_NULL_HANDLE;
-                    descriptor_set->vk_image_infos[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
-                }
-                break;
-            case VK_DESCRIPTOR_TYPE_SAMPLER:
-                descriptor_set->vk_descriptor_writes[0].pImageInfo = &descriptor_set->vk_image_infos[0];
-                for (i = 0; i < ARRAY_SIZE(descriptor_set->vk_image_infos); ++i)
-                {
-                    descriptor_set->vk_image_infos[i].imageView = VK_NULL_HANDLE;
-                    descriptor_set->vk_image_infos[i].imageLayout = VK_IMAGE_LAYOUT_UNDEFINED;
-                }
-                break;
-            default:
-                ERR("Unhandled descriptor type %#x.\n", device->vk_descriptor_heap_layouts[set].type);
-                return E_FAIL;
-        }
         if (device->vk_descriptor_heap_layouts[set].applicable_heap_type == desc->Type
                 && FAILED(hr = d3d12_descriptor_heap_create_descriptor_set(descriptor_heap, device, set)))
             return hr;
@@ -4032,7 +3951,9 @@ HRESULT d3d12_descriptor_heap_create(struct d3d12_device *device,
         {
             memset(&dst[i].s, 0, sizeof(dst[i].s));
             dst[i].index = i;
+            dst[i].next = 0;
         }
+        object->dirty_list_head = UINT_MAX;
     }
     else
     {
diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h
index 1a277a47..24eebf30 100644
--- a/libs/vkd3d/vkd3d_private.h
+++ b/libs/vkd3d/vkd3d_private.h
@@ -732,6 +732,7 @@ struct d3d12_desc
         } u;
     } s;
     unsigned int index;
+    LONG next;
 };

static inline struct d3d12_desc *d3d12_desc_from_cpu_handle(D3D12_CPU_DESCRIPTOR_HANDLE cpu_handle)
@@ -838,15 +839,10 @@ struct vkd3d_vk_descriptor_heap_layout
     VkDescriptorSetLayout vk_set_layout;
 };

-#define VKD3D_DESCRIPTOR_WRITE_BUFFER_SIZE 64
-
 struct d3d12_descriptor_heap_vk_set
 {
     VkDescriptorSet vk_set;
-    VkDescriptorBufferInfo vk_buffer_infos[VKD3D_DESCRIPTOR_WRITE_BUFFER_SIZE];
-    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];
+    VkDescriptorType vk_type;
 };

/* ID3D12DescriptorHeap */
@@ -864,11 +860,15 @@ struct d3d12_descriptor_heap

VkDescriptorPool vk_descriptor_pool;
     struct d3d12_descriptor_heap_vk_set vk_descriptor_sets[VKD3D_SET_INDEX_COUNT];
-    struct vkd3d_mutex vk_sets_mutex;
+
+    LONG volatile dirty_list_head;
+    unsigned int padding;

BYTE descriptors[];
 };

+void d3d12_desc_flush_vk_heap_updates(struct d3d12_descriptor_heap *descriptor_heap, struct d3d12_device *device);
+
 static inline struct d3d12_descriptor_heap *d3d12_desc_get_descriptor_heap(const struct d3d12_desc *descriptor)
 {
     return CONTAINING_RECORD(descriptor - descriptor->index, struct d3d12_descriptor_heap, descriptors);