path: root/0001-wined3d-Initial-implementation-of-a-persistent-mappe.patch

From 1f69076549bf2351eb6d8d885b35a46b4dc69813 Mon Sep 17 00:00:00 2001
From: Andrew Comminos <andrew@comminos.com>
Date: Mon, 5 Mar 2018 15:38:35 -0800
Subject: [PATCH 1/9] wined3d: Initial implementation of a persistent mapped
 buffer allocator.

---
 dlls/wined3d-csmt/Makefile.in  |   1 +
 dlls/wined3d/Makefile.in       |   1 +
 dlls/wined3d/buffer_heap.c     | 508 +++++++++++++++++++++++++++++++++++++++++
 dlls/wined3d/cs.c              |   9 +
 dlls/wined3d/device.c          |  52 +++++
 dlls/wined3d/directx.c         |   3 +
 dlls/wined3d/query.c           |   2 +-
 dlls/wined3d/wined3d_gl.h      |   1 +
 dlls/wined3d/wined3d_private.h |  68 +++++-
 9 files changed, 641 insertions(+), 4 deletions(-)
 create mode 100644 dlls/wined3d/buffer_heap.c

diff --git a/dlls/wined3d-csmt/Makefile.in b/dlls/wined3d-csmt/Makefile.in
index 1d0458eb46..cb3a5484c6 100644
--- a/dlls/wined3d-csmt/Makefile.in
+++ b/dlls/wined3d-csmt/Makefile.in
@@ -8,6 +8,7 @@ C_SRCS = \
 	arb_program_shader.c \
 	ati_fragment_shader.c \
 	buffer.c \
+	buffer_heap.c \
 	context.c \
 	cs.c \
 	device.c \
diff --git a/dlls/wined3d/Makefile.in b/dlls/wined3d/Makefile.in
index b850ba6872..52ef8666fb 100644
--- a/dlls/wined3d/Makefile.in
+++ b/dlls/wined3d/Makefile.in
@@ -6,6 +6,7 @@ C_SRCS = \
 	arb_program_shader.c \
 	ati_fragment_shader.c \
 	buffer.c \
+	buffer_heap.c \
 	context.c \
 	cs.c \
 	device.c \
diff --git a/dlls/wined3d/buffer_heap.c b/dlls/wined3d/buffer_heap.c
new file mode 100644
index 0000000000..b133bd6893
--- /dev/null
+++ b/dlls/wined3d/buffer_heap.c
@@ -0,0 +1,508 @@
+/*
+ * Copyright 2018 Andrew Comminos
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
+ *
+ */
+
+#include "config.h"
+#include "wine/port.h"
+#include "wine/rbtree.h"
+#include "wined3d_private.h"
+
+WINE_DEFAULT_DEBUG_CHANNEL(d3d);
+WINE_DECLARE_DEBUG_CHANNEL(d3d_perf);
+
+struct wined3d_buffer_heap_element
+{
+    struct wined3d_map_range range;
+
+    // rbtree data
+    struct wine_rb_entry entry;
+
+    // Binned free list positions
+    struct wined3d_buffer_heap_element *next;
+    struct wined3d_buffer_heap_element *prev;
+};
+
+struct wined3d_buffer_heap_fenced_element
+{
+    struct wined3d_buffer_heap_bin_set free_list;
+    struct wined3d_fence *fence;
+
+    struct wined3d_buffer_heap_fenced_element *next;
+};
+
+static struct wined3d_buffer_heap_element* element_new(GLsizei offset, GLsizei size)
+{
+    struct wined3d_buffer_heap_element* elem;
+    elem = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct wined3d_buffer_heap_element));
+    if (!elem)
+        return NULL;
+    elem->range.offset = offset;
+    elem->range.size = size;
+    return elem;
+}
+
+static inline int bitwise_log2_floor(GLsizei size)
+{
+    // XXX(acomminos): I hope this gets unrolled.
+    for (int i = 8 * sizeof(GLsizei) - 1; i >= 0; i--)
+    {
+        if ((size >> i) & 1) {
+            return i;
+        }
+    }
+    return 0;
+}
+
+static inline int bitwise_log2_ceil(GLsizei size)
+{
+    // Add one to the floor of size if size isn't a power of two.
+    return bitwise_log2_floor(size) + !!(size & (size - 1));
+}
+
+static int element_bin(struct wined3d_buffer_heap_element *elem)
+{
+    return min(WINED3D_BUFFER_HEAP_BINS - 1, bitwise_log2_floor(elem->range.size));
+}
+
+// Inserts an element into the appropriate free list bin.
+static void element_insert_free_bin(struct wined3d_buffer_heap *heap, struct wined3d_buffer_heap_element *elem)
+{
+    int bin = element_bin(elem);
+
+    elem->prev = NULL;
+    elem->next = heap->free_list.bins[bin].head;
+    if (heap->free_list.bins[bin].head)
+        heap->free_list.bins[bin].head->prev = elem;
+    heap->free_list.bins[bin].head = elem;
+
+    if (!heap->free_list.bins[bin].tail)
+        heap->free_list.bins[bin].tail = elem;
+
+    TRACE("Inserted allocation at %p of size %lld into bin %d\n", elem->range.offset, elem->range.size, bin);
+}
+
+// Removes an element from the free tree, its bin, and the coalesce list.
+static void element_remove_free(struct wined3d_buffer_heap *heap, struct wined3d_buffer_heap_element *elem)
+{
+    int bin = element_bin(elem);
+
+    if (elem->prev)
+        elem->prev->next = elem->next;
+
+    if (elem->next)
+        elem->next->prev = elem->prev;
+
+    if (elem == heap->free_list.bins[bin].head)
+        heap->free_list.bins[bin].head = elem->next;
+
+    if (elem == heap->free_list.bins[bin].tail)
+        heap->free_list.bins[bin].tail = elem->prev;
+
+    elem->prev = NULL;
+    elem->next = NULL;
+
+    TRACE("Freed allocation at %p of size %lld from bin %d\n", elem->range.offset, elem->range.size, bin);
+}
+
+static struct wined3d_buffer_heap_fenced_element* fenced_element_new(struct wined3d_buffer_heap_bin_set bins, struct wined3d_fence* fence)
+{
+    struct wined3d_buffer_heap_fenced_element* elem;
+    elem = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct wined3d_buffer_heap_fenced_element));
+    if (!elem)
+        return NULL;
+    elem->free_list = bins;
+    elem->fence = fence;
+    elem->next = NULL;
+    return elem;
+}
+
+static int free_tree_compare(const void *key, const struct wine_rb_entry *entry)
+{
+    const GLsizei offset = *(const GLsizei*) key;
+    struct wined3d_buffer_heap_element *elem = WINE_RB_ENTRY_VALUE(entry, struct wined3d_buffer_heap_element, entry);
+
+    if (offset < elem->range.offset)
+        return -1;
+    if (offset > elem->range.offset)
+        return 1;
+    return 0;
+}
+
+/* Context activation is done by the caller. */
+HRESULT wined3d_buffer_heap_create(struct wined3d_context *context, GLsizeiptr size, GLsizeiptr alignment, BOOL write_only, struct wined3d_buffer_heap **buffer_heap)
+{
+    const struct wined3d_gl_info *gl_info = context->gl_info;
+    const GLenum buffer_target = GL_ARRAY_BUFFER;
+    GLbitfield access_flags;
+    GLbitfield storage_flags;
+    struct wined3d_buffer_heap_element *initial_elem;
+
+    struct wined3d_buffer_heap *object;
+
+    if ((alignment & (alignment - 1)) != 0)
+    {
+        return E_FAIL;
+    }
+
+    if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
+    {
+        return E_OUTOFMEMORY;
+    }
+
+    access_flags = GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT | GL_MAP_WRITE_BIT;
+    if (!write_only)
+    {
+        access_flags |= GL_MAP_READ_BIT;
+    }
+    storage_flags = access_flags;
+
+    // TODO(acomminos): where should we be checking for errors here?
+    GL_EXTCALL(glGenBuffers(1, &object->buffer_object));
+
+    context_bind_bo(context, buffer_target, object->buffer_object);
+
+    // TODO(acomminos): assert glBufferStorage supported?
+    GL_EXTCALL(glBufferStorage(buffer_target, size, NULL, storage_flags));
+
+    if (!(object->map_ptr = GL_EXTCALL(glMapBufferRange(buffer_target, 0, size, access_flags))))
+    {
+        ERR("Couldn't map persistent buffer.\n");
+        return -1; // FIXME(acomminos): proper error code, cleanup
+    }
+    context_bind_bo(context, buffer_target, 0);
+
+    object->fenced_head = object->fenced_tail = NULL;
+    object->alignment = alignment;
+    InitializeCriticalSection(&object->temp_lock);
+
+    initial_elem = element_new(0, size);
+    // Don't bother adding the initial allocation to the coalescing tree.
+    element_insert_free_bin(object, initial_elem);
+
+    *buffer_heap = object;
+
+    return WINED3D_OK;
+}
+
+/* Context activation is done by the caller. */
+HRESULT wined3d_buffer_heap_destroy(struct wined3d_buffer_heap *heap, struct wined3d_context *context)
+{
+    FIXME("Unimplemented, leaking buffer");
+    return WINED3D_OK;
+}
+
+HRESULT wined3d_buffer_heap_alloc(struct wined3d_buffer_heap *heap, GLsizeiptr size, struct wined3d_map_range *out_range)
+{
+    int initial_bin;
+    int initial_size = size;
+
+    EnterCriticalSection(&heap->temp_lock);
+
+    // After alignment, reduce fragmentation by rounding to next power of two.
+    // If the alignment is a power of two (which it should be), this should be
+    // no problem.
+    size = 1 << bitwise_log2_ceil(size);
+
+    // Align size values where possible.
+    if (heap->alignment && (size % heap->alignment != 0))
+        size += heap->alignment - (size % heap->alignment);
+
+    initial_bin = min(WINED3D_BUFFER_HEAP_BINS - 1, bitwise_log2_ceil(size));
+
+    for (int i = initial_bin; i < WINED3D_BUFFER_HEAP_BINS; i++)
+    {
+        struct wined3d_buffer_heap_element *elem = heap->free_list.bins[i].head;
+        if (elem)
+        {
+            struct wined3d_map_range remaining_range;
+            remaining_range.offset = elem->range.offset + size;
+            remaining_range.size = elem->range.size - size;
+
+            out_range->offset = elem->range.offset;
+            out_range->size = size;
+
+            TRACE_(d3d_perf)("Allocated %d (requested %d) at %p from bin %d (initial %d)\n", size, initial_size, elem->range.offset, i, initial_bin);
+
+            // Remove the element from its current free bin to move it to the correct list.
+            element_remove_free(heap, elem);
+
+            if (remaining_range.size > 0)
+            {
+                TRACE_(d3d_perf)("Imperfect fit allocated, fragmenting remainder of %lld at %p.\n", remaining_range.size, remaining_range.offset);
+
+                elem->range = remaining_range;
+                element_insert_free_bin(heap, elem);
+            }
+            else
+            {
+                HeapFree(GetProcessHeap(), 0, elem);
+            }
+
+            LeaveCriticalSection(&heap->temp_lock);
+            return WINED3D_OK;
+        }
+    }
+
+    LeaveCriticalSection(&heap->temp_lock);
+
+    FIXME_(d3d_perf)("Forcing coalesce, not enough free space in buffer heap.\n");
+    int num_coalesced;
+    if (SUCCEEDED(wined3d_buffer_heap_deferred_coalesce(heap, &num_coalesced)))
+    {
+        if (num_coalesced > 0)
+            return wined3d_buffer_heap_alloc(heap, size, out_range);
+    }
+
+    FIXME_(d3d_perf)("Coalescing did not create new blocks, failing.\n");
+
+    return WINED3DERR_OUTOFVIDEOMEMORY;
+}
+
+HRESULT wined3d_buffer_heap_free(struct wined3d_buffer_heap *heap, struct wined3d_map_range range)
+{
+    struct wined3d_buffer_heap_element *elem = element_new(range.offset, range.size);
+
+    if (!elem)
+        return E_OUTOFMEMORY;
+
+    EnterCriticalSection(&heap->temp_lock);
+
+    // Only insert the element into a free bin, coalescing will occur later.
+    element_insert_free_bin(heap, elem);
+
+    LeaveCriticalSection(&heap->temp_lock);
+
+    return WINED3D_OK;
+}
+
+HRESULT wined3d_buffer_heap_free_fenced(struct wined3d_buffer_heap *heap, struct wined3d_device *device, struct wined3d_map_range range)
+{
+    struct wined3d_buffer_heap_element *elem = element_new(range.offset, range.size);
+    int bin_index = element_bin(elem);
+    struct wined3d_buffer_heap_bin *bin = &heap->pending_fenced_bins.bins[bin_index];
+
+    if (bin->tail)
+    {
+        bin->tail->next = elem;
+        elem->prev = bin->tail;
+        bin->tail = elem;
+    }
+    else
+    {
+        bin->head = elem;
+        bin->tail = elem;
+    }
+
+    return WINED3D_OK;
+}
+
+HRESULT wined3d_buffer_heap_cs_fence_issue(struct wined3d_buffer_heap *heap, struct wined3d_device *device)
+{
+    struct wined3d_buffer_heap_fenced_element *fenced_elem;
+    struct wined3d_fence *fence;
+    HRESULT hr;
+
+    if (heap->fenced_head)
+    {
+        // XXX(acomminos): double or triple buffer this?
+        wined3d_buffer_heap_cs_fence_wait(heap, device);
+    }
+
+    if (FAILED(hr = wined3d_fence_create(device, &fence)))
+    {
+        ERR("Failed to create fence.\n");
+        return hr;
+    }
+
+    fenced_elem = fenced_element_new(heap->pending_fenced_bins, fence);
+    if (!fenced_elem)
+        return E_OUTOFMEMORY;
+
+    TRACE_(d3d_perf)("Dispatching fenced buffer set.\n");
+    memset(&heap->pending_fenced_bins, 0, sizeof(heap->pending_fenced_bins));
+
+    // Append to end of fenced list, which works well if you assume that buffers
+    // are freed in some ascending draw call ordering.
+    if (!heap->fenced_head)
+    {
+        heap->fenced_head = fenced_elem;
+        heap->fenced_tail = fenced_elem;
+    }
+    else
+    {
+        heap->fenced_tail->next = fenced_elem;
+        heap->fenced_tail = fenced_elem;
+    }
+
+    wined3d_fence_issue(fence, device);
+    return WINED3D_OK;
+}
+
+HRESULT wined3d_buffer_heap_cs_fence_wait(struct wined3d_buffer_heap *heap, struct wined3d_device *device)
+{
+    enum wined3d_fence_result res;
+    struct wined3d_buffer_heap_fenced_element *elem = heap->fenced_head;
+    if (!elem)
+        return WINED3D_OK;
+
+    res = wined3d_fence_wait(elem->fence, device);
+    switch (res)
+    {
+        case WINED3D_FENCE_OK:
+        case WINED3D_FENCE_NOT_STARTED:
+            {
+                TRACE_(d3d_perf)("Freed fence group.\n");
+
+                EnterCriticalSection(&heap->temp_lock);
+                for (int i = 0; i < WINED3D_BUFFER_HEAP_BINS; i++)
+                {
+                    struct wined3d_buffer_heap_bin *elem_bin = &elem->free_list.bins[i];
+                    if (!elem_bin->tail)
+                        continue;
+
+                    struct wined3d_buffer_heap_bin *heap_bin = &heap->free_list.bins[i];
+                    if (heap_bin->head)
+                    {
+                        // Insert to front.
+                        elem_bin->tail->next = heap_bin->head;
+                        heap_bin->head->prev = elem_bin->tail;
+
+                        elem_bin->head->prev = NULL;
+                        heap_bin->head = elem_bin->head;
+                    }
+                    else
+                    {
+                        elem_bin->head->prev = NULL;
+                        heap_bin->head = elem_bin->head;
+                        elem_bin->tail->next = NULL;
+                        heap_bin->tail = elem_bin->tail;
+                    }
+                }
+                LeaveCriticalSection(&heap->temp_lock);
+
+                wined3d_fence_destroy(elem->fence);
+
+                heap->fenced_head = elem->next;
+                HeapFree(GetProcessHeap(), 0, elem);
+                // TODO(acomminos): bother to null out fenced_tail?
+                break;
+            }
+        default:
+            return WINED3D_OK;
+    }
+
+    return WINED3D_OK;
+}
+
+HRESULT wined3d_buffer_heap_deferred_coalesce(struct wined3d_buffer_heap *heap, int *coalesced_count)
+{
+    struct wined3d_buffer_heap_element *elem = NULL;
+    struct wined3d_buffer_heap_element *next = NULL;
+    struct wine_rb_entry *entry;
+    struct wined3d_map_range coalesced_range;
+
+    struct wine_rb_tree free_tree;
+    int num_coalesced = 0;
+
+    wine_rb_init(&free_tree, free_tree_compare);
+
+    EnterCriticalSection(&heap->temp_lock);
+
+    // TODO(acomminos): on one hand, if there's a lot of elements in the list,
+    // it's highly fragmented. on the other, we can potentially waste a decent
+    // sum of time checking for uncoalesced bins.
+    for (int i = 0; i < WINED3D_BUFFER_HEAP_BINS; i++)
+    {
+        elem = heap->free_list.bins[i].head;
+        while (elem)
+        {
+            // Insert a sentry. FIXME(acomminos): can skip this with traversal.
+            if (wine_rb_put(&free_tree, &elem->range.offset, &elem->entry) == -1)
+            {
+                ERR("Failed to insert key %x in tree.\n", elem->range.offset);
+                elem = elem->next;
+                continue;
+            }
+
+            coalesced_range = elem->range;
+
+            // Coalesce right.
+            entry = wine_rb_next(&elem->entry);
+            if (entry)
+            {
+                TRACE("Coalesced right.\n");
+                struct wined3d_buffer_heap_element *right_elem = WINE_RB_ENTRY_VALUE(entry, struct wined3d_buffer_heap_element, entry);
+                if (elem->range.offset + elem->range.size == right_elem->range.offset)
+                {
+                    coalesced_range.size += right_elem->range.size;
+
+                    wine_rb_remove(&free_tree, entry);
+                    element_remove_free(heap, right_elem);
+                    HeapFree(GetProcessHeap(), 0, right_elem);
+
+                    num_coalesced++;
+                }
+            }
+
+            // Coalesce left.
+            entry = wine_rb_prev(&elem->entry);
+            if (entry)
+            {
+                TRACE("Coalesced left.\n");
+                struct wined3d_buffer_heap_element *left_elem = WINE_RB_ENTRY_VALUE(entry, struct wined3d_buffer_heap_element, entry);
+                if (left_elem->range.offset + left_elem->range.size == coalesced_range.offset)
+                {
+                    coalesced_range.offset = left_elem->range.offset;
+                    coalesced_range.size += left_elem->range.size;
+
+                    wine_rb_remove(&free_tree, entry);
+                    element_remove_free(heap, left_elem);
+                    HeapFree(GetProcessHeap(), 0, left_elem);
+
+                    num_coalesced++;
+                }
+            }
+
+            next = elem->next;
+
+            if (elem->range.size != coalesced_range.size)
+            {
+                FIXME_(d3d_perf)("Coalesced range from (%p, %ld) to (%p, %ld)\n", elem->range.offset, elem->range.size, coalesced_range.offset, coalesced_range.size);
+
+                wine_rb_remove(&free_tree, &elem->entry);
+
+                // Move to the correct free bin.
+                element_remove_free(heap, elem);
+                elem->range = coalesced_range;
+                element_insert_free_bin(heap, elem);
+
+                wine_rb_put(&free_tree, &elem->range.offset, &elem->entry);
+            }
+
+            elem = next;
+        }
+    }
+
+    LeaveCriticalSection(&heap->temp_lock);
+
+    FIXME_(d3d_perf)("Performed %d coalesces.\n", num_coalesced);
+    if (coalesced_count)
+        *coalesced_count = num_coalesced;
+
+    return WINED3D_OK;
+}
diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
index 3a7c95ddd8..50a4d041cd 100644
--- a/dlls/wined3d/cs.c
+++ b/dlls/wined3d/cs.c
@@ -472,6 +472,15 @@ static void wined3d_cs_exec_present(struct wined3d_cs *cs, const void *data)
     }
 
     InterlockedDecrement(&cs->pending_presents);
+
+    // FIXME(acomminos): is this the right place to put double-buffered frame
+    //                   timing based logic?
+    // FIXME(acomminos): this conditional sucks, replace with fancier feature check
+    if (cs->device->wo_buffer_heap && cs->device->cb_buffer_heap)
+    {
+        wined3d_buffer_heap_cs_fence_issue(cs->device->wo_buffer_heap, cs->device);
+        wined3d_buffer_heap_cs_fence_issue(cs->device->cb_buffer_heap, cs->device);
+    }
 }
 
 void wined3d_cs_emit_present(struct wined3d_cs *cs, struct wined3d_swapchain *swapchain,
diff --git a/dlls/wined3d/device.c b/dlls/wined3d/device.c
index e2b27e0cf4..785841a062 100644
--- a/dlls/wined3d/device.c
+++ b/dlls/wined3d/device.c
@@ -833,6 +833,53 @@ static void destroy_default_samplers(struct wined3d_device *device, struct wined
     device->null_sampler = NULL;
 }
 
+/* Context activation is done by the caller. */
+static void create_buffer_heap(struct wined3d_device *device, struct wined3d_context *context)
+{
+    const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
+    // TODO(acomminos): kill this magic number. perhaps base on vram.
+    GLsizeiptr geo_heap_size = 512 * 1024 * 1024;
+    // We choose a constant buffer size of 128MB, the same as NVIDIA claims to
+    // use in their Direct3D driver for discarded constant buffers.
+    GLsizeiptr cb_heap_size = 128 * 1024 * 1024;
+    GLint ub_alignment;
+    HRESULT hr;
+
+    if (gl_info->supported[ARB_BUFFER_STORAGE])
+    {
+        gl_info->gl_ops.gl.p_glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &ub_alignment);
+
+        // Align constant buffer heap size, in case GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT isn't a power of two (for some reason).
+        cb_heap_size -= cb_heap_size % ub_alignment;
+
+        if (FAILED(hr = wined3d_buffer_heap_create(context, geo_heap_size, 0, TRUE, &device->wo_buffer_heap)))
+        {
+            ERR("Failed to create write-only persistent buffer heap, hr %#x.\n", hr);
+        }
+
+        if (FAILED(hr = wined3d_buffer_heap_create(context, cb_heap_size, ub_alignment, TRUE, &device->cb_buffer_heap)))
+        {
+            ERR("Failed to create persistent buffer heap for constant buffers, hr %#x.\n", hr);
+        }
+
+        FIXME("Initialized PBA (geo_heap_size: %ld, cb_heap_size: %ld, ub_align: %d)\n", geo_heap_size, cb_heap_size, ub_alignment);
+    }
+    else
+    {
+        FIXME("Not using PBA, ARB_buffer_storage unsupported.\n");
+    }
+}
+
+/* Context activation is done by the caller. */
+static void destroy_buffer_heap(struct wined3d_device *device, struct wined3d_context *context)
+{
+    if (device->wo_buffer_heap)
+        wined3d_buffer_heap_destroy(device->wo_buffer_heap, context);
+
+    if (device->cb_buffer_heap)
+        wined3d_buffer_heap_destroy(device->cb_buffer_heap, context);
+}
+
 static LONG fullscreen_style(LONG style)
 {
     /* Make sure the window is managed, otherwise we won't get keyboard input. */
@@ -997,6 +1044,8 @@ static void wined3d_device_delete_opengl_contexts_cs(void *object)
     device->shader_backend->shader_free_private(device);
     destroy_dummy_textures(device, context);
     destroy_default_samplers(device, context);
+    destroy_buffer_heap(device, context);
+
     context_release(context);
 
     while (device->context_count)
@@ -1045,6 +1094,9 @@ static void wined3d_device_create_primary_opengl_context_cs(void *object)
     context = context_acquire(device, target, 0);
     create_dummy_textures(device, context);
     create_default_samplers(device, context);
+
+    create_buffer_heap(device, context);
+
     context_release(context);
 }
 
diff --git a/dlls/wined3d/directx.c b/dlls/wined3d/directx.c
index 8720fc7ad6..46c6a59536 100644
--- a/dlls/wined3d/directx.c
+++ b/dlls/wined3d/directx.c
@@ -111,6 +111,7 @@ static const struct wined3d_extension_map gl_extension_map[] =
     /* ARB */
     {"GL_ARB_base_instance",                ARB_BASE_INSTANCE             },
     {"GL_ARB_blend_func_extended",          ARB_BLEND_FUNC_EXTENDED       },
+    {"GL_ARB_buffer_storage",               ARB_BUFFER_STORAGE            },
     {"GL_ARB_clear_buffer_object",          ARB_CLEAR_BUFFER_OBJECT       },
     {"GL_ARB_clear_texture",                ARB_CLEAR_TEXTURE             },
     {"GL_ARB_clip_control",                 ARB_CLIP_CONTROL              },
@@ -2714,6 +2715,8 @@ static void load_gl_funcs(struct wined3d_gl_info *gl_info)
     /* GL_ARB_blend_func_extended */
     USE_GL_FUNC(glBindFragDataLocationIndexed)
     USE_GL_FUNC(glGetFragDataIndex)
+    /* GL_ARB_buffer_storage */
+    USE_GL_FUNC(glBufferStorage)
     /* GL_ARB_clear_buffer_object */
     USE_GL_FUNC(glClearBufferData)
     USE_GL_FUNC(glClearBufferSubData)
diff --git a/dlls/wined3d/query.c b/dlls/wined3d/query.c
index 5ea79b6e4a..f3ca1630e5 100644
--- a/dlls/wined3d/query.c
+++ b/dlls/wined3d/query.c
@@ -88,7 +88,7 @@ static BOOL wined3d_fence_supported(const struct wined3d_gl_info *gl_info)
     return gl_info->supported[ARB_SYNC] || gl_info->supported[NV_FENCE] || gl_info->supported[APPLE_FENCE];
 }
 
-static enum wined3d_fence_result wined3d_fence_test(const struct wined3d_fence *fence,
+enum wined3d_fence_result wined3d_fence_test(const struct wined3d_fence *fence,
         const struct wined3d_device *device, DWORD flags)
 {
     const struct wined3d_gl_info *gl_info;
diff --git a/dlls/wined3d/wined3d_gl.h b/dlls/wined3d/wined3d_gl.h
index 87283c850e..7626864ef2 100644
--- a/dlls/wined3d/wined3d_gl.h
+++ b/dlls/wined3d/wined3d_gl.h
@@ -44,6 +44,7 @@ enum wined3d_gl_extension
     /* ARB */
     ARB_BASE_INSTANCE,
     ARB_BLEND_FUNC_EXTENDED,
+    ARB_BUFFER_STORAGE,
     ARB_CLEAR_BUFFER_OBJECT,
     ARB_CLEAR_TEXTURE,
     ARB_CLIP_CONTROL,
diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
index 8aa61d811f..3d535f4e17 100644
--- a/dlls/wined3d/wined3d_private.h
+++ b/dlls/wined3d/wined3d_private.h
@@ -1712,6 +1712,9 @@ void wined3d_fence_destroy(struct wined3d_fence *fence) DECLSPEC_HIDDEN;
 void wined3d_fence_issue(struct wined3d_fence *fence, const struct wined3d_device *device) DECLSPEC_HIDDEN;
 enum wined3d_fence_result wined3d_fence_wait(const struct wined3d_fence *fence,
         const struct wined3d_device *device) DECLSPEC_HIDDEN;
+// XXX(acomminos): really expose this?
+enum wined3d_fence_result wined3d_fence_test(const struct wined3d_fence *fence,
+        const struct wined3d_device *device, DWORD flags) DECLSPEC_HIDDEN;
 
 /* Direct3D terminology with little modifications. We do not have an issued
  * state because only the driver knows about it, but we have a created state
@@ -2993,6 +2996,10 @@ struct wined3d_device
     /* Context management */
     struct wined3d_context **contexts;
     UINT context_count;
+
+    /* Dynamic buffer heap */
+    struct wined3d_buffer_heap *wo_buffer_heap;
+    struct wined3d_buffer_heap *cb_buffer_heap;
 };
 
 void device_clear_render_targets(struct wined3d_device *device, UINT rt_count, const struct wined3d_fb_state *fb,
@@ -3513,6 +3520,12 @@ void state_init(struct wined3d_state *state, struct wined3d_fb_state *fb,
         DWORD flags) DECLSPEC_HIDDEN;
 void state_unbind_resources(struct wined3d_state *state) DECLSPEC_HIDDEN;
 
+struct wined3d_map_range
+{
+    GLintptr offset;
+    GLsizeiptr size;
+};
+
 enum wined3d_cs_queue_id
 {
     WINED3D_CS_QUEUE_DEFAULT = 0,
@@ -3692,12 +3705,61 @@ enum wined3d_buffer_conversion_type
     CONV_POSITIONT,
 };
 
-struct wined3d_map_range
+struct wined3d_buffer_heap_element;
+struct wined3d_buffer_heap_fenced_element;
+
+// Number of power-of-two buckets to populate.
+#define WINED3D_BUFFER_HEAP_BINS 32
+
+struct wined3d_buffer_heap_bin
 {
-    UINT offset;
-    UINT size;
+    struct wined3d_buffer_heap_element *head;
+    struct wined3d_buffer_heap_element *tail;
+};
+
+struct wined3d_buffer_heap_bin_set
+{
+    struct wined3d_buffer_heap_bin bins[WINED3D_BUFFER_HEAP_BINS];
 };
 
+// A heap that manages allocations with a single GL buffer.
+struct wined3d_buffer_heap
+{
+    GLuint buffer_object;
+    void *map_ptr;
+    GLsizeiptr alignment;
+    CRITICAL_SECTION temp_lock; // Temporary lock while we implement the fenced free list.
+
+    struct wined3d_buffer_heap_bin_set free_list;
+
+    // Elements that need to be fenced, but haven't reached the required size.
+    struct wined3d_buffer_heap_bin_set pending_fenced_bins;
+
+    // List of sets of buffers behind a common fence, in FIFO order.
+    struct wined3d_buffer_heap_fenced_element *fenced_head;
+    struct wined3d_buffer_heap_fenced_element *fenced_tail;
+};
+
+HRESULT wined3d_buffer_heap_create(struct wined3d_context *context, GLsizeiptr size, GLsizeiptr alignment, BOOL write_only, struct wined3d_buffer_heap **heap) DECLSPEC_HIDDEN;
+HRESULT wined3d_buffer_heap_destroy(struct wined3d_buffer_heap *heap, struct wined3d_context *context) DECLSPEC_HIDDEN;
+// Fetches a buffer from the heap of at least the given size.
+// Attempts to coalesce blocks under memory pressure.
+HRESULT wined3d_buffer_heap_alloc(struct wined3d_buffer_heap *heap, GLsizeiptr size, struct wined3d_map_range* out_range) DECLSPEC_HIDDEN;
+// Immediately frees a heap-allocated buffer segment.
+HRESULT wined3d_buffer_heap_free(struct wined3d_buffer_heap *heap, struct wined3d_map_range range) DECLSPEC_HIDDEN;
+// Enqueues a buffer segment to return to the heap once its fence has been signaled.
+HRESULT wined3d_buffer_heap_free_fenced(struct wined3d_buffer_heap *heap, struct wined3d_device *device, struct wined3d_map_range range) DECLSPEC_HIDDEN;
+// Issues a fence for the current set of pending fenced buffers.
+// Double-buffered: if the last fence issued has not yet been triggered, waits
+// on it.
+HRESULT wined3d_buffer_heap_cs_fence_issue(struct wined3d_buffer_heap *heap, struct wined3d_device *device) DECLSPEC_HIDDEN;
+// Waits on the next issued fence in FIFO order. Frees the fenced buffers after
+// the fence has been triggered.
+HRESULT wined3d_buffer_heap_cs_fence_wait(struct wined3d_buffer_heap *heap, struct wined3d_device *device) DECLSPEC_HIDDEN;
+// Performs deferred coalescing of buffers. To be called under memory pressure.
+// Outputs the number of coalesced regions in `num_coalesced`.
+HRESULT wined3d_buffer_heap_deferred_coalesce(struct wined3d_buffer_heap *heap, int *num_coalesced) DECLSPEC_HIDDEN;
+
 struct wined3d_buffer
 {
     struct wined3d_resource resource;
-- 
2.16.2