vkd3d: Optimise the GPU VA allocator.

The GPU VA allocator was allocating memory in a way where dereferencing
GPU VA required a lock + bsearch() to find the right VA range.

Rather than going this route, we turn the common case into O(1) and
lock-free by creating a slab allocator which allows us to lookup a
pointer directly from a GPU VA with (VA - Base) / PageSize.

The number of allocations in the fast path must be limited since we
cannot trivially grow the allocator while remaining lock-free for
dereferences.

Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Signed-off-by: Henri Verbeet <hverbeet@codeweavers.com>
Signed-off-by: Alexandre Julliard <julliard@winehq.org>

libs/vkd3d/device.c

@@ -1822,11 +1822,70 @@ static void d3d12_device_destroy_pipeline_cache(struct d3d12_device *device)
     pthread_mutex_destroy(&device->mutex);
 }
 
-D3D12_GPU_VIRTUAL_ADDRESS vkd3d_gpu_va_allocator_allocate(struct vkd3d_gpu_va_allocator *allocator,
+#define VKD3D_VA_FALLBACK_BASE      0x8000000000000000ull
-        size_t alignment, size_t size, void *ptr)
+#define VKD3D_VA_SLAB_BASE          0x0000001000000000ull
+#define VKD3D_VA_SLAB_SIZE_SHIFT    32
+#define VKD3D_VA_SLAB_SIZE          (1ull << VKD3D_VA_SLAB_SIZE_SHIFT)
+#define VKD3D_VA_SLAB_COUNT         (64 * 1024)
+
+static D3D12_GPU_VIRTUAL_ADDRESS vkd3d_gpu_va_allocator_allocate_slab(struct vkd3d_gpu_va_allocator *allocator,
+        size_t aligned_size, void *ptr)
+{
+    struct vkd3d_gpu_va_slab *slab;
+    D3D12_GPU_VIRTUAL_ADDRESS address;
+    unsigned slab_idx;
+
+    slab = allocator->free_slab;
+    allocator->free_slab = slab->ptr;
+    slab->size = aligned_size;
+    slab->ptr = ptr;
+
+    /* It is critical that the multiplication happens in 64-bit to not
+     * overflow. */
+    slab_idx = slab - allocator->slabs;
+    address = VKD3D_VA_SLAB_BASE + slab_idx * VKD3D_VA_SLAB_SIZE;
+
+    TRACE("Allocated address %#"PRIx64", slab %u, size %zu.\n", address, slab_idx, aligned_size);
+
+    return address;
+}
+
+static D3D12_GPU_VIRTUAL_ADDRESS vkd3d_gpu_va_allocator_allocate_fallback(struct vkd3d_gpu_va_allocator *allocator,
+        size_t alignment, size_t aligned_size, void *ptr)
 {
     struct vkd3d_gpu_va_allocation *allocation;
     D3D12_GPU_VIRTUAL_ADDRESS base, ceiling;
+
+    base = allocator->fallback_floor;
+    ceiling = ~(D3D12_GPU_VIRTUAL_ADDRESS)0;
+    ceiling -= alignment - 1;
+    if (aligned_size > ceiling || ceiling - aligned_size < base)
+        return 0;
+
+    base = (base + (alignment - 1)) & ~((D3D12_GPU_VIRTUAL_ADDRESS)alignment - 1);
+
+    if (!vkd3d_array_reserve((void **)&allocator->fallback_allocations, &allocator->fallback_allocations_size,
+            allocator->fallback_allocation_count + 1, sizeof(*allocator->fallback_allocations)))
+        return 0;
+
+    allocation = &allocator->fallback_allocations[allocator->fallback_allocation_count++];
+    allocation->base = base;
+    allocation->size = aligned_size;
+    allocation->ptr = ptr;
+
+    /* This pointer is bumped and never lowered on a free. However, this will
+     * only fail once we have exhausted 63 bits of address space. */
+    allocator->fallback_floor = base + aligned_size;
+
+    TRACE("Allocated address %#"PRIx64", size %zu.\n", base, aligned_size);
+
+    return base;
+}
+
+D3D12_GPU_VIRTUAL_ADDRESS vkd3d_gpu_va_allocator_allocate(struct vkd3d_gpu_va_allocator *allocator,
+        size_t alignment, size_t size, void *ptr)
+{
+    D3D12_GPU_VIRTUAL_ADDRESS address;
     int rc;
 
     if (size > ~(size_t)0 - (alignment - 1))
@@ -1839,34 +1898,41 @@ D3D12_GPU_VIRTUAL_ADDRESS vkd3d_gpu_va_allocator_allocate(struct vkd3d_gpu_va_al
         return 0;
     }
 
-    base = allocator->floor;
+    if (size <= VKD3D_VA_SLAB_SIZE && allocator->free_slab)
-    ceiling = ~(D3D12_GPU_VIRTUAL_ADDRESS)0;
+        address = vkd3d_gpu_va_allocator_allocate_slab(allocator, size, ptr);
-    ceiling -= alignment - 1;
+    else
-    if (size > ceiling || ceiling - size < base)
+        address = vkd3d_gpu_va_allocator_allocate_fallback(allocator, alignment, size, ptr);
-    {
-        pthread_mutex_unlock(&allocator->mutex);
-        return 0;
-    }
-
-    base = (base + (alignment - 1)) & ~((D3D12_GPU_VIRTUAL_ADDRESS)alignment - 1);
-
-    if (!vkd3d_array_reserve((void **)&allocator->allocations, &allocator->allocations_size,
-            allocator->allocation_count + 1, sizeof(*allocator->allocations)))
-    {
-        pthread_mutex_unlock(&allocator->mutex);
-        return 0;
-    }
-
-    allocation = &allocator->allocations[allocator->allocation_count++];
-    allocation->base = base;
-    allocation->size = size;
-    allocation->ptr = ptr;
-
-    allocator->floor = base + size;
 
     pthread_mutex_unlock(&allocator->mutex);
 
-    return allocation->base;
+    return address;
+}
+
+static void *vkd3d_gpu_va_allocator_dereference_slab(struct vkd3d_gpu_va_allocator *allocator,
+        D3D12_GPU_VIRTUAL_ADDRESS address)
+{
+    const struct vkd3d_gpu_va_slab *slab;
+    D3D12_GPU_VIRTUAL_ADDRESS base_offset;
+    unsigned int slab_idx;
+
+    base_offset = address - VKD3D_VA_SLAB_BASE;
+    slab_idx = base_offset >> VKD3D_VA_SLAB_SIZE_SHIFT;
+
+    if (slab_idx >= VKD3D_VA_SLAB_COUNT)
+    {
+        ERR("Invalid slab index %u for address %#"PRIx64".\n", slab_idx, address);
+        return NULL;
+    }
+
+    slab = &allocator->slabs[slab_idx];
+    base_offset -= slab_idx * VKD3D_VA_SLAB_SIZE;
+    if (base_offset >= slab->size)
+    {
+        ERR("Address %#"PRIx64" is %#"PRIx64" bytes into slab %u of size %zu.\n",
+                address, base_offset, slab_idx, slab->size);
+        return NULL;
+    }
+    return slab->ptr;
 }
 
 static int vkd3d_gpu_va_allocation_compare(const void *k, const void *e)
@@ -1881,30 +1947,94 @@ static int vkd3d_gpu_va_allocation_compare(const void *k, const void *e)
     return 0;
 }
 
-void *vkd3d_gpu_va_allocator_dereference(struct vkd3d_gpu_va_allocator *allocator,
+static void *vkd3d_gpu_va_allocator_dereference_fallback(struct vkd3d_gpu_va_allocator *allocator,
         D3D12_GPU_VIRTUAL_ADDRESS address)
 {
     struct vkd3d_gpu_va_allocation *allocation;
+
+    allocation = bsearch(&address, allocator->fallback_allocations, allocator->fallback_allocation_count,
+            sizeof(*allocation), vkd3d_gpu_va_allocation_compare);
+
+    return allocation ? allocation->ptr : NULL;
+}
+
+void *vkd3d_gpu_va_allocator_dereference(struct vkd3d_gpu_va_allocator *allocator,
+        D3D12_GPU_VIRTUAL_ADDRESS address)
+{
+    void *ret;
     int rc;
 
+    /* If we land in the non-fallback region, dereferencing VA is lock-less.
+     * The base pointer is immutable, and the only way we can have a data race
+     * is if some other thread is poking into the
+     * slab_mem_allocation[base_index] block. This can only happen if someone
+     * is trying to free the entry while we're dereferencing it, which would
+     * be a serious application bug. */
+    if (address < VKD3D_VA_FALLBACK_BASE)
+        return vkd3d_gpu_va_allocator_dereference_slab(allocator, address);
+
+    /* Slow fallback. */
     if ((rc = pthread_mutex_lock(&allocator->mutex)))
     {
         ERR("Failed to lock mutex, error %d.\n", rc);
         return NULL;
     }
 
-    allocation = bsearch(&address, allocator->allocations, allocator->allocation_count,
+    ret = vkd3d_gpu_va_allocator_dereference_fallback(allocator, address);
-            sizeof(*allocation), vkd3d_gpu_va_allocation_compare);
 
     pthread_mutex_unlock(&allocator->mutex);
 
-    return allocation ? allocation->ptr : NULL;
+    return ret;
+}
+
+static void vkd3d_gpu_va_allocator_free_slab(struct vkd3d_gpu_va_allocator *allocator,
+        D3D12_GPU_VIRTUAL_ADDRESS address)
+{
+    D3D12_GPU_VIRTUAL_ADDRESS base_offset;
+    struct vkd3d_gpu_va_slab *slab;
+    unsigned int slab_idx;
+
+    base_offset = address - VKD3D_VA_SLAB_BASE;
+    slab_idx = base_offset >> VKD3D_VA_SLAB_SIZE_SHIFT;
+
+    if (slab_idx >= VKD3D_VA_SLAB_COUNT)
+    {
+        ERR("Invalid slab index %u for address %#"PRIx64".\n", slab_idx, address);
+        return;
+    }
+
+    TRACE("Freeing address %#"PRIx64", slab %u.\n", address, slab_idx);
+
+    slab = &allocator->slabs[slab_idx];
+    slab->size = 0;
+    slab->ptr = allocator->free_slab;
+    allocator->free_slab = slab;
+}
+
+static void vkd3d_gpu_va_allocator_free_fallback(struct vkd3d_gpu_va_allocator *allocator,
+        D3D12_GPU_VIRTUAL_ADDRESS address)
+{
+    struct vkd3d_gpu_va_allocation *allocation;
+    unsigned int index;
+
+    allocation = bsearch(&address, allocator->fallback_allocations, allocator->fallback_allocation_count,
+            sizeof(*allocation), vkd3d_gpu_va_allocation_compare);
+
+    if (!allocation || allocation->base != address)
+    {
+        ERR("Address %#"PRIx64" does not match any allocation.\n", address);
+        return;
+    }
+
+    index = allocation - allocator->fallback_allocations;
+    --allocator->fallback_allocation_count;
+    if (index != allocator->fallback_allocation_count)
+        memmove(&allocator->fallback_allocations[index], &allocator->fallback_allocations[index + 1],
+                (allocator->fallback_allocation_count - index) * sizeof(*allocation));
 }
 
 void vkd3d_gpu_va_allocator_free(struct vkd3d_gpu_va_allocator *allocator, D3D12_GPU_VIRTUAL_ADDRESS address)
 {
-    struct vkd3d_gpu_va_allocation *allocation;
-    unsigned int index;
     int rc;
 
     if ((rc = pthread_mutex_lock(&allocator->mutex)))
@@ -1913,32 +2043,45 @@ void vkd3d_gpu_va_allocator_free(struct vkd3d_gpu_va_allocator *allocator, D3D12
         return;
     }
 
-    allocation = bsearch(&address, allocator->allocations, allocator->allocation_count,
+    if (address < VKD3D_VA_FALLBACK_BASE)
-            sizeof(*allocation), vkd3d_gpu_va_allocation_compare);
-    if (allocation && allocation->base == address)
     {
-        index = allocation - allocator->allocations;
+        vkd3d_gpu_va_allocator_free_slab(allocator, address);
-        --allocator->allocation_count;
+        pthread_mutex_unlock(&allocator->mutex);
-        if (index != allocator->allocation_count)
+        return;
-        {
-            memmove(&allocator->allocations[index], &allocator->allocations[index + 1],
-                    (allocator->allocation_count - index) * sizeof(*allocation));
-        }
     }
 
+    vkd3d_gpu_va_allocator_free_fallback(allocator, address);
+
     pthread_mutex_unlock(&allocator->mutex);
 }
 
 static bool vkd3d_gpu_va_allocator_init(struct vkd3d_gpu_va_allocator *allocator)
 {
+    unsigned int i;
     int rc;
 
     memset(allocator, 0, sizeof(*allocator));
-    allocator->floor = 0x1000;
+    allocator->fallback_floor = VKD3D_VA_FALLBACK_BASE;
+
+    /* To remain lock-less, we cannot grow the slabs array after the fact. If
+     * we commit to a maximum number of allocations here, we can dereference
+     * without taking a lock as the base pointer never changes. We would be
+     * able to grow more seamlessly using an array of pointers, but that would
+     * make dereferencing slightly less efficient. */
+    if (!(allocator->slabs = vkd3d_calloc(VKD3D_VA_SLAB_COUNT, sizeof(*allocator->slabs))))
+        return false;
+
+    /* Mark all slabs as free. */
+    allocator->free_slab = &allocator->slabs[0];
+    for (i = 0; i < VKD3D_VA_SLAB_COUNT - 1; ++i)
+    {
+        allocator->slabs[i].ptr = &allocator->slabs[i + 1];
+    }
 
     if ((rc = pthread_mutex_init(&allocator->mutex, NULL)))
     {
         ERR("Failed to initialize mutex, error %d.\n", rc);
+        vkd3d_free(allocator->slabs);
         return false;
     }
 
@@ -1954,7 +2097,8 @@ static void vkd3d_gpu_va_allocator_cleanup(struct vkd3d_gpu_va_allocator *alloca
         ERR("Failed to lock mutex, error %d.\n", rc);
         return;
     }
-    vkd3d_free(allocator->allocations);
+    vkd3d_free(allocator->slabs);
+    vkd3d_free(allocator->fallback_allocations);
     pthread_mutex_unlock(&allocator->mutex);
     pthread_mutex_destroy(&allocator->mutex);
 }

libs/vkd3d/vkd3d_private.h

@@ -202,20 +202,30 @@ HRESULT vkd3d_fence_worker_start(struct vkd3d_fence_worker *worker,
 HRESULT vkd3d_fence_worker_stop(struct vkd3d_fence_worker *worker,
         struct d3d12_device *device) DECLSPEC_HIDDEN;
 
+struct vkd3d_gpu_va_allocation
+{
+    D3D12_GPU_VIRTUAL_ADDRESS base;
+    SIZE_T size;
+    void *ptr;
+};
+
+struct vkd3d_gpu_va_slab
+{
+    SIZE_T size;
+    void *ptr;
+};
+
 struct vkd3d_gpu_va_allocator
 {
     pthread_mutex_t mutex;
 
-    D3D12_GPU_VIRTUAL_ADDRESS floor;
+    D3D12_GPU_VIRTUAL_ADDRESS fallback_floor;
+    struct vkd3d_gpu_va_allocation *fallback_allocations;
+    size_t fallback_allocations_size;
+    size_t fallback_allocation_count;
 
-    struct vkd3d_gpu_va_allocation
+    struct vkd3d_gpu_va_slab *slabs;
-    {
+    struct vkd3d_gpu_va_slab *free_slab;
-        D3D12_GPU_VIRTUAL_ADDRESS base;
-        SIZE_T size;
-        void *ptr;
-    } *allocations;
-    size_t allocations_size;
-    size_t allocation_count;
 };
 
 D3D12_GPU_VIRTUAL_ADDRESS vkd3d_gpu_va_allocator_allocate(struct vkd3d_gpu_va_allocator *allocator,