vo_gpu: vulkan: initial implementation

This time based on ra/vo_gpu. 2017 is the year of the vulkan desktop!

Current problems / limitations / improvement opportunities:

1. The swapchain/flipping code violates the vulkan spec, by assuming
   that the presentation queue will be bounded (in cases where rendering
   is significantly faster than vsync). But apparently, there's simply
   no better way to do this right now, to the point where even the
   stupid cube.c examples from LunarG etc. do it wrong.
   (cf. https://github.com/KhronosGroup/Vulkan-Docs/issues/370)

2. The memory allocator could be improved. (This is a universal
   constant)

3. Could explore using push descriptors instead of descriptor sets,
   especially since we expect to switch descriptors semi-often for some
   passes (like interpolation). Probably won't make a difference, but
   the synchronization overhead might be a factor. Who knows.

4. Parallelism across frames / async transfer is not well-defined, we
   either need to use a better semaphore / command buffer strategy or a
   resource pooling layer to safely handle cross-frame parallelism.
   (That said, I gave resource pooling a try and was not happy with the
   result at all - so I'm still exploring the semaphore strategy)

5. We aggressively use pipeline barriers where events would offer a much
   more fine-grained synchronization mechanism. As a result of this, we
   might be suffering from GPU bubbles due to too-short dependencies on
   objects. (That said, I'm also exploring the use of semaphores as a an
   ordering tactic which would allow cross-frame time slicing in theory)

Some minor changes to the vo_gpu and infrastructure, but nothing
consequential.

NOTE: For safety, all use of asynchronous commands / multiple command
pools is currently disabled completely. There are some left-over relics
of this in the code (e.g. the distinction between dev_poll and
pool_poll), but that is kept in place mostly because this will be
re-extended in the future (vulkan rev 2).

The queue count is also currently capped to 1, because of the lack of
cross-frame semaphores means we need the implicit synchronization from
the same-queue semantics to guarantee a correct result.

20 files changed, 3773 insertions, 16 deletions

diff --git a/DOCS/man/options.rst b/DOCS/man/options.rst
index b08150c6bb..80e7350292 100644
--- a/DOCS/man/options.rst
+++ b/DOCS/man/options.rst
@@ -4103,10 +4103,6 @@ The following video options are currently all specific to ``--vo=gpu`` and
     the video along the temporal axis. The filter used can be controlled using
     the ``--tscale`` setting.
 
-    Note that this relies on vsync to work, see ``--opengl-swapinterval`` for
-    more information. It should also only be used with an ``--fbo-format``
-    that has at least 16 bit precision.
-
 ``--interpolation-threshold=<0..1,-1>``
     Threshold below which frame ratio interpolation gets disabled (default:
     ``0.0001``). This is calculated as ``abs(disphz/vfps - 1) < threshold``,
@@ -4184,6 +4180,31 @@ The following video options are currently all specific to ``--vo=gpu`` and
     results, as can missing or incorrect display FPS information (see
     ``--display-fps``).
 
+``--vulkan-swap-mode=<mode>``
+    Controls the presentation mode of the vulkan swapchain. This is similar
+    to the ``--opengl-swapinterval`` option.
+
+    auto
+        Use the preferred swapchain mode for the vulkan context. (Default)
+    fifo
+        Non-tearing, vsync blocked. Similar to "VSync on".
+    fifo-relaxed
+        Tearing, vsync blocked. Late frames will tear instead of stuttering.
+    mailbox
+        Non-tearing, not vsync blocked. Similar to "triple buffering".
+    immediate
+        Tearing, not vsync blocked. Similar to "VSync off".
+
+``--vulkan-queue-count=<1..8>``
+    Controls the number of VkQueues used for rendering (limited by how many
+    your device supports). In theory, using more queues could enable some
+    parallelism between frames (when using a ``--swapchain-depth`` higher than
+    1). (Default: 1)
+
+    NOTE: Setting this to a value higher than 1 may cause graphical corruption,
+    as mpv's vulkan implementation currently does not try and protect textures
+    against concurrent access.
+
 ``--glsl-shaders=<file-list>``
     Custom GLSL hooks. These are a flexible way to add custom fragment shaders,
     which can be injected at almost arbitrary points in the rendering pipeline,
@@ -4590,7 +4611,7 @@ The following video options are currently all specific to ``--vo=gpu`` and
         on Nvidia and AMD. Newer Intel chips with the latest drivers may also
         work.
     x11
-        X11/GLX
+        X11/GLX, VK_KHR_xlib_surface
     x11probe
         For internal autoprobing, equivalent to ``x11`` otherwise. Don't use
         directly, it could be removed without warning as autoprobing is changed.
@@ -5020,7 +5041,10 @@ Miscellaneous
     Media files must use constant framerate. Section-wise VFR might work as well
     with some container formats (but not e.g. mkv). If the sync code detects
     severe A/V desync, or the framerate cannot be detected, the player
-    automatically reverts to ``audio`` mode for some time or permanently.
+    automatically reverts to ``audio`` mode for some time or permanently. These
+    modes also require a vsync blocked presentation mode. For OpenGL, this
+    translates to ``--opengl-swapinterval=1``. For Vulkan, it translates to
+    ``--vulkan-swap-mode=fifo`` (or ``fifo-relaxed``).
 
     The modes with ``desync`` in their names do not attempt to keep audio/video
     in sync. They will slowly (or quickly) desync, until e.g. the next seek
diff --git a/options/options.c b/options/options.c
index 1168cc196b..6467468691 100644
--- a/options/options.c
+++ b/options/options.c
@@ -89,6 +89,7 @@ extern const struct m_obj_list vo_obj_list;
 extern const struct m_obj_list ao_obj_list;
 
 extern const struct m_sub_options opengl_conf;
+extern const struct m_sub_options vulkan_conf;
 extern const struct m_sub_options angle_conf;
 extern const struct m_sub_options cocoa_conf;
 
@@ -690,6 +691,10 @@ const m_option_t mp_opts[] = {
     OPT_SUBSTRUCT("", opengl_opts, opengl_conf, 0),
 #endif
 
+#if HAVE_VULKAN
+    OPT_SUBSTRUCT("", vulkan_opts, vulkan_conf, 0),
+#endif
+
 #if HAVE_EGL_ANGLE_WIN32
     OPT_SUBSTRUCT("", angle_opts, angle_conf, 0),
 #endif
diff --git a/options/options.h b/options/options.h
index c02b7a34ca..63dee03612 100644
--- a/options/options.h
+++ b/options/options.h
@@ -329,6 +329,7 @@ typedef struct MPOpts {
     struct gl_video_opts *gl_video_opts;
     struct angle_opts *angle_opts;
     struct opengl_opts *opengl_opts;
+    struct vulkan_opts *vulkan_opts;
     struct cocoa_opts *cocoa_opts;
     struct dvd_opts *dvd_opts;
 
diff --git a/video/out/gpu/context.c b/video/out/gpu/context.c
index c5721c73b4..25e2a754bf 100644
--- a/video/out/gpu/context.c
+++ b/video/out/gpu/context.c
@@ -44,6 +44,7 @@ extern const struct ra_ctx_fns ra_ctx_dxgl;
 extern const struct ra_ctx_fns ra_ctx_rpi;
 extern const struct ra_ctx_fns ra_ctx_mali;
 extern const struct ra_ctx_fns ra_ctx_vdpauglx;
+extern const struct ra_ctx_fns ra_ctx_vulkan_xlib;
 
 static const struct ra_ctx_fns *contexts[] = {
 // OpenGL contexts:
@@ -83,6 +84,13 @@ static const struct ra_ctx_fns *contexts[] = {
 #if HAVE_VDPAU_GL_X11
     &ra_ctx_vdpauglx,
 #endif
+
+// Vulkan contexts:
+#if HAVE_VULKAN
+#if HAVE_X11
+    &ra_ctx_vulkan_xlib,
+#endif
+#endif
 };
 
 static bool get_help(struct mp_log *log, struct bstr param)
diff --git a/video/out/gpu/ra.h b/video/out/gpu/ra.h
index 10245b250e..7a2fa0e11c 100644
--- a/video/out/gpu/ra.h
+++ b/video/out/gpu/ra.h
@@ -146,6 +146,7 @@ enum ra_buf_type {
     RA_BUF_TYPE_TEX_UPLOAD,     // texture upload buffer (pixel buffer object)
     RA_BUF_TYPE_SHADER_STORAGE, // shader buffer (SSBO), for RA_VARTYPE_BUF_RW
     RA_BUF_TYPE_UNIFORM,        // uniform buffer (UBO), for RA_VARTYPE_BUF_RO
+    RA_BUF_TYPE_VERTEX,         // not publicly usable (RA-internal usage)
 };
 
 struct ra_buf_params {
@@ -369,10 +370,10 @@ struct ra_fns {
 
     void (*buf_destroy)(struct ra *ra, struct ra_buf *buf);
 
-    // Update the contents of a buffer, starting at a given offset and up to a
-    // given size, with the contents of *data. This is an extremely common
-    // operation. Calling this while the buffer is considered "in use" is an
-    // error. (See: buf_poll)
+    // Update the contents of a buffer, starting at a given offset (*must* be a
+    // multiple of 4) and up to a given size, with the contents of *data. This
+    // is an extremely common operation. Calling this while the buffer is
+    // considered "in use" is an error. (See: buf_poll)
     void (*buf_update)(struct ra *ra, struct ra_buf *buf, ptrdiff_t offset,
                        const void *data, size_t size);
 
diff --git a/video/out/vo_gpu.c b/video/out/vo_gpu.c
index bd245de05b..a26912e0d8 100644
--- a/video/out/vo_gpu.c
+++ b/video/out/vo_gpu.c
@@ -60,6 +60,7 @@ struct gpu_priv {
 static void resize(struct gpu_priv *p)
 {
     struct vo *vo = p->vo;
+    struct ra_swapchain *sw = p->ctx->swapchain;
 
     MP_VERBOSE(vo, "Resize: %dx%d\n", vo->dwidth, vo->dheight);
 
@@ -69,6 +70,11 @@ static void resize(struct gpu_priv *p)
 
     gl_video_resize(p->renderer, &src, &dst, &osd);
 
+    int fb_depth = sw->fns->color_depth ? sw->fns->color_depth(sw) : 0;
+    if (fb_depth)
+        MP_VERBOSE(p, "Reported display depth: %d\n", fb_depth);
+    gl_video_set_fb_depth(p->renderer, fb_depth);
+
     vo->want_redraw = true;
 }
 
@@ -289,7 +295,6 @@ static int preinit(struct vo *vo)
         goto err_out;
     assert(p->ctx->ra);
     assert(p->ctx->swapchain);
-    struct ra_swapchain *sw = p->ctx->swapchain;
 
     p->renderer = gl_video_init(p->ctx->ra, vo->log, vo->global);
     gl_video_set_osd_source(p->renderer, vo->osd);
@@ -305,11 +310,6 @@ static int preinit(struct vo *vo)
                              vo->hwdec_devs, vo->opts->gl_hwdec_interop);
     gl_video_set_hwdec(p->renderer, p->hwdec);
 
-    int fb_depth = sw->fns->color_depth ? sw->fns->color_depth(sw) : 0;
-    if (fb_depth)
-        MP_VERBOSE(p, "Reported display depth: %d\n", fb_depth);
-    gl_video_set_fb_depth(p->renderer, fb_depth);
-
     return 0;
 
 err_out:
diff --git a/video/out/vulkan/common.h b/video/out/vulkan/common.h
new file mode 100644
index 0000000000..4c0e783f0e
--- /dev/null
+++ b/video/out/vulkan/common.h
@@ -0,0 +1,51 @@
+#pragma once
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <stdint.h>
+#include <stdbool.h>
+#include <assert.h>
+
+#include "config.h"
+
+#include "common/common.h"
+#include "common/msg.h"
+
+// We need to define all platforms we want to support. Since we have
+// our own mechanism for checking this, we re-define the right symbols
+#if HAVE_X11
+#define VK_USE_PLATFORM_XLIB_KHR
+#endif
+
+#include <vulkan/vulkan.h>
+
+// Vulkan allows the optional use of a custom allocator. We don't need one but
+// mark this parameter with a better name in case we ever decide to change this
+// in the future. (And to make the code more readable)
+#define MPVK_ALLOCATOR NULL
+
+// A lot of things depend on streaming resources across frames. Depending on
+// how many frames we render ahead of time, we need to pick enough to avoid
+// any conflicts, so make all of these tunable relative to this constant in
+// order to centralize them.
+#define MPVK_MAX_STREAMING_DEPTH 8
+
+// Shared struct used to hold vulkan context information
+struct mpvk_ctx {
+    struct mp_log *log;
+    VkInstance inst;
+    VkPhysicalDevice physd;
+    VkDebugReportCallbackEXT dbg;
+    VkDevice dev;
+
+    // Surface, must be initialized fter the context itself
+    VkSurfaceKHR surf;
+    VkSurfaceFormatKHR surf_format; // picked at surface initialization time
+
+    struct vk_malloc *alloc; // memory allocator for this device
+    struct vk_cmdpool *pool; // primary command pool for this device
+    struct vk_cmd *last_cmd; // most recently submitted command
+
+    // Cached capabilities
+    VkPhysicalDeviceLimits limits;
+};
diff --git a/video/out/vulkan/context.c b/video/out/vulkan/context.c
new file mode 100644
index 0000000000..bd456d214c
--- /dev/null
+++ b/video/out/vulkan/context.c
@@ -0,0 +1,501 @@
+/*
+ * This file is part of mpv.
+ *
+ * mpv 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.
+ *
+ * mpv 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 mpv.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "options/m_config.h"
+#include "context.h"
+#include "ra_vk.h"
+#include "utils.h"
+
+enum {
+    SWAP_AUTO = 0,
+    SWAP_FIFO,
+    SWAP_FIFO_RELAXED,
+    SWAP_MAILBOX,
+    SWAP_IMMEDIATE,
+    SWAP_COUNT,
+};
+
+struct vulkan_opts {
+    struct mpvk_device_opts dev_opts; // logical device options
+    char *device; // force a specific GPU
+    int swap_mode;
+};
+
+static int vk_validate_dev(struct mp_log *log, const struct m_option *opt,
+                           struct bstr name, struct bstr param)
+{
+    int ret = M_OPT_INVALID;
+    VkResult res;
+
+    // Create a dummy instance to validate/list the devices
+    VkInstanceCreateInfo info = {
+        .sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
+    };
+
+    VkInstance inst;
+    VkPhysicalDevice *devices = NULL;
+    uint32_t num = 0;
+
+    res = vkCreateInstance(&info, MPVK_ALLOCATOR, &inst);
+    if (res != VK_SUCCESS)
+        goto error;
+
+    res = vkEnumeratePhysicalDevices(inst, &num, NULL);
+    if (res != VK_SUCCESS)
+        goto error;
+
+    devices = talloc_array(NULL, VkPhysicalDevice, num);
+    vkEnumeratePhysicalDevices(inst, &num, devices);
+    if (res != VK_SUCCESS)
+        goto error;
+
+    bool help = bstr_equals0(param, "help");
+    if (help) {
+        mp_info(log, "Available vulkan devices:\n");
+        ret = M_OPT_EXIT;
+    }
+
+    for (int i = 0; i < num; i++) {
+        VkPhysicalDeviceProperties prop;
+        vkGetPhysicalDeviceProperties(devices[i], &prop);
+
+        if (help) {
+            mp_info(log, "  '%s' (GPU %d, ID %x:%x)\n", prop.deviceName, i,
+                    (unsigned)prop.vendorID, (unsigned)prop.deviceID);
+        } else if (bstr_equals0(param, prop.deviceName)) {
+            ret = 0;
+            break;
+        }
+    }
+
+    if (!help)
+        mp_err(log, "No device with name '%.*s'!\n", BSTR_P(param));
+
+error:
+    talloc_free(devices);
+    return ret;
+}
+
+#define OPT_BASE_STRUCT struct vulkan_opts
+const struct m_sub_options vulkan_conf = {
+    .opts = (const struct m_option[]) {
+        OPT_STRING_VALIDATE("vulkan-device", device, 0, vk_validate_dev),
+        OPT_CHOICE("vulkan-swap-mode", swap_mode, 0,
+                   ({"auto",        SWAP_AUTO},
+                   {"fifo",         SWAP_FIFO},
+                   {"fifo-relaxed", SWAP_FIFO_RELAXED},
+                   {"mailbox",      SWAP_MAILBOX},
+                   {"immediate",    SWAP_IMMEDIATE})),
+        OPT_INTRANGE("vulkan-queue-count", dev_opts.queue_count, 0, 1,
+                     MPVK_MAX_QUEUES, OPTDEF_INT(1)),
+        {0}
+    },
+    .size = sizeof(struct vulkan_opts)
+};
+
+struct priv {
+    struct mpvk_ctx *vk;
+    struct vulkan_opts *opts;
+    // Swapchain metadata:
+    int w, h;                 // current size
+    VkSwapchainCreateInfoKHR protoInfo; // partially filled-in prototype
+    VkSwapchainKHR swapchain;
+    VkSwapchainKHR old_swapchain;
+    int frames_in_flight;
+    // state of the images:
+    struct ra_tex **images;   // ra_tex wrappers for the vkimages
+    int num_images;           // size of images
+    VkSemaphore *acquired;    // pool of semaphores used to synchronize images
+    int num_acquired;         // size of this pool
+    int idx_acquired;         // index of next free semaphore within this pool
+    int last_imgidx;          // the image index last acquired (for submit)
+};
+
+static bool update_swapchain_info(struct priv *p,
+                                  VkSwapchainCreateInfoKHR *info)
+{
+    struct mpvk_ctx *vk = p->vk;
+
+    // Query the supported capabilities and update this struct as needed
+    VkSurfaceCapabilitiesKHR caps;
+    VK(vkGetPhysicalDeviceSurfaceCapabilitiesKHR(vk->physd, vk->surf, &caps));
+
+    // Sorted by preference
+    static const VkCompositeAlphaFlagBitsKHR alphaModes[] = {
+        VK_COMPOSITE_ALPHA_POST_MULTIPLIED_BIT_KHR,
+        VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR,
+    };
+
+    for (int i = 0; i < MP_ARRAY_SIZE(alphaModes); i++) {
+        if (caps.supportedCompositeAlpha & alphaModes[i]) {
+            info->compositeAlpha = alphaModes[i];
+            break;
+        }
+    }
+
+    if (!info->compositeAlpha) {
+        MP_ERR(vk, "Failed picking alpha compositing mode (caps: 0x%x)\n",
+               caps.supportedCompositeAlpha);
+        goto error;
+    }
+
+    static const VkSurfaceTransformFlagBitsKHR rotModes[] = {
+        VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR,
+        VK_SURFACE_TRANSFORM_INHERIT_BIT_KHR,
+    };
+
+    for (int i = 0; i < MP_ARRAY_SIZE(rotModes); i++) {
+        if (caps.supportedTransforms & rotModes[i]) {
+            info->preTransform = rotModes[i];
+            break;
+        }
+    }
+
+    if (!info->preTransform) {
+        MP_ERR(vk, "Failed picking surface transform mode (caps: 0x%x)\n",
+               caps.supportedTransforms);
+        goto error;
+    }
+
+    // Image count as required
+    MP_VERBOSE(vk, "Requested image count: %d (min %d max %d)\n",
+               (int)info->minImageCount, (int)caps.minImageCount,
+               (int)caps.maxImageCount);
+
+    info->minImageCount = MPMAX(info->minImageCount, caps.minImageCount);
+    if (caps.maxImageCount)
+        info->minImageCount = MPMIN(info->minImageCount, caps.maxImageCount);
+
+    // Check the extent against the allowed parameters
+    if (caps.currentExtent.width != info->imageExtent.width &&
+        caps.currentExtent.width != 0xFFFFFFFF)
+    {
+        MP_WARN(vk, "Requested width %d does not match current width %d\n",
+                (int)info->imageExtent.width, (int)caps.currentExtent.width);
+        info->imageExtent.width = caps.currentExtent.width;
+    }
+
+    if (caps.currentExtent.height != info->imageExtent.height &&
+        caps.currentExtent.height != 0xFFFFFFFF)
+    {
+        MP_WARN(vk, "Requested height %d does not match current height %d\n",
+                (int)info->imageExtent.height, (int)caps.currentExtent.height);
+        info->imageExtent.height = caps.currentExtent.height;
+    }
+
+    if (caps.minImageExtent.width  > info->imageExtent.width ||
+        caps.minImageExtent.height > info->imageExtent.height)
+    {
+        MP_ERR(vk, "Requested size %dx%d smaller than device minimum %d%d\n",
+               (int)info->imageExtent.width, (int)info->imageExtent.height,
+               (int)caps.minImageExtent.width, (int)caps.minImageExtent.height);
+        goto error;
+    }
+
+    if (caps.maxImageExtent.width  < info->imageExtent.width ||
+        caps.maxImageExtent.height < info->imageExtent.height)
+    {
+        MP_ERR(vk, "Requested size %dx%d larger than device maximum %d%d\n",
+               (int)info->imageExtent.width, (int)info->imageExtent.height,
+               (int)caps.maxImageExtent.width, (int)caps.maxImageExtent.height);
+        goto error;
+    }
+
+    // We just request whatever usage we can, and let the ra_vk decide what
+    // ra_tex_params that translates to. This makes the images as flexible
+    // as possible.
+    info->imageUsage = caps.supportedUsageFlags;
+    return true;
+
+error:
+    return false;
+}
+
+void ra_vk_ctx_uninit(struct ra_ctx *ctx)
+{
+    if (ctx->ra) {
+        struct priv *p = ctx->swapchain->priv;
+        struct mpvk_ctx *vk = p->vk;
+
+        mpvk_pool_wait_idle(vk, vk->pool);
+
+        for (int i = 0; i < p->num_images; i++)
+            ra_tex_free(ctx->ra, &p->images[i]);
+        for (int i = 0; i < p->num_acquired; i++)
+            vkDestroySemaphore(vk->dev, p->acquired[i], MPVK_ALLOCATOR);
+
+        vkDestroySwapchainKHR(vk->dev, p->swapchain, MPVK_ALLOCATOR);
+
+        talloc_free(p->images);
+        talloc_free(p->acquired);
+        ctx->ra->fns->destroy(ctx->ra);
+        ctx->ra = NULL;
+    }
+
+    talloc_free(ctx->swapchain);
+    ctx->swapchain = NULL;
+}
+
+static const struct ra_swapchain_fns vulkan_swapchain;
+
+bool ra_vk_ctx_init(struct ra_ctx *ctx, struct mpvk_ctx *vk,
+                    VkPresentModeKHR preferred_mode)
+{
+    struct ra_swapchain *sw = ctx->swapchain = talloc_zero(NULL, struct ra_swapchain);
+    sw->ctx = ctx;
+    sw->fns = &vulkan_swapchain;
+
+    struct priv *p = sw->priv = talloc_zero(sw, struct priv);
+    p->vk = vk;
+    p->opts = mp_get_config_group(p, ctx->global, &vulkan_conf);
+
+    if (!mpvk_find_phys_device(vk, p->opts->device, ctx->opts.allow_sw))
+        goto error;
+    if (!mpvk_pick_surface_format(vk))
+        goto error;
+    if (!mpvk_device_init(vk, p->opts->dev_opts))
+        goto error;
+
+    ctx->ra = ra_create_vk(vk, ctx->log);
+    if (!ctx->ra)
+        goto error;
+
+    static const VkPresentModeKHR present_modes[SWAP_COUNT] = {
+        [SWAP_FIFO]         = VK_PRESENT_MODE_FIFO_KHR,
+        [SWAP_FIFO_RELAXED] = VK_PRESENT_MODE_FIFO_RELAXED_KHR,
+        [SWAP_MAILBOX]      = VK_PRESENT_MODE_MAILBOX_KHR,
+        [SWAP_IMMEDIATE]    = VK_PRESENT_MODE_IMMEDIATE_KHR,
+    };
+
+    p->protoInfo = (VkSwapchainCreateInfoKHR) {
+        .sType = VK_STRUCTURE_TYPE_SWAPCHAIN_CREATE_INFO_KHR,
+        .surface = vk->surf,
+        .imageFormat = vk->surf_format.format,
+        .imageColorSpace = vk->surf_format.colorSpace,
+        .imageArrayLayers = 1, // non-stereoscopic
+        .imageSharingMode = VK_SHARING_MODE_EXCLUSIVE,
+        .minImageCount = ctx->opts.swapchain_depth + 1, // +1 for FB
+        .presentMode = p->opts->swap_mode ? present_modes[p->opts->swap_mode]
+                                          : preferred_mode,
+        .clipped = true,
+    };
+
+    // Make sure the swapchain present mode is supported
+    int num_modes;
+    VK(vkGetPhysicalDeviceSurfacePresentModesKHR(vk->physd, vk->surf,
+                                                 &num_modes, NULL));
+    VkPresentModeKHR *modes = talloc_array(NULL, VkPresentModeKHR, num_modes);
+    VK(vkGetPhysicalDeviceSurfacePresentModesKHR(vk->physd, vk->surf,
+                                                 &num_modes, modes));
+    bool supported = false;
+    for (int i = 0; i < num_modes; i++)
+        supported |= (modes[i] == p->protoInfo.presentMode);
+    talloc_free(modes);
+
+    if (!supported) {
+        MP_ERR(ctx, "Requested swap mode unsupported by this device!\n");
+        goto error;
+    }
+
+    return true;
+
+error:
+    ra_vk_ctx_uninit(ctx);
+    return false;
+}
+
+static void destroy_swapchain(struct mpvk_ctx *vk, struct priv *p)
+{
+    assert(p->old_swapchain);
+    vkDestroySwapchainKHR(vk->dev, p->old_swapchain, MPVK_ALLOCATOR);
+    p->old_swapchain = NULL;
+}
+
+bool ra_vk_ctx_resize(struct ra_swapchain *sw, int w, int h)
+{
+    struct priv *p = sw->priv;
+    if (w == p->w && h == p->h)
+        return true;
+
+    struct ra *ra = sw->ctx->ra;
+    struct mpvk_ctx *vk = p->vk;
+    VkImage *vkimages = NULL;
+
+    // It's invalid to trigger another swapchain recreation while there's
+    // more than one swapchain already active, so we need to flush any pending
+    // asynchronous swapchain release operations that may be ongoing.
+    while (p->old_swapchain)
+        mpvk_dev_poll_cmds(vk, 100000); // 100μs
+
+    VkSwapchainCreateInfoKHR sinfo = p->protoInfo;
+    sinfo.imageExtent  = (VkExtent2D){ w, h };
+    sinfo.oldSwapchain = p->swapchain;
+
+    if (!update_swapchain_info(p, &sinfo))
+        goto error;
+
+    VK(vkCreateSwapchainKHR(vk->dev, &sinfo, MPVK_ALLOCATOR, &p->swapchain));
+    p->w = w;
+    p->h = h;
+
+    // Freeing the old swapchain while it's still in use is an error, so do
+    // it asynchronously once the device is idle.
+    if (sinfo.oldSwapchain) {
+        p->old_swapchain = sinfo.oldSwapchain;
+        vk_dev_callback(vk, (vk_cb) destroy_swapchain, vk, p);
+    }
+
+    // Get the new swapchain images
+    int num;
+    VK(vkGetSwapchainImagesKHR(vk->dev, p->swapchain, &num, NULL));
+    vkimages = talloc_array(NULL, VkImage, num);
+    VK(vkGetSwapchainImagesKHR(vk->dev, p->swapchain, &num, vkimages));
+
+    // If needed, allocate some more semaphores
+    while (num > p->num_acquired) {
+        VkSemaphore sem;
+        static const VkSemaphoreCreateInfo seminfo = {
+            .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
+        };
+        VK(vkCreateSemaphore(vk->dev, &seminfo, MPVK_ALLOCATOR, &sem));
+        MP_TARRAY_APPEND(NULL, p->acquired, p->num_acquired, sem);
+    }
+
+    // Recreate the ra_tex wrappers
+    for (int i = 0; i < p->num_images; i++)
+        ra_tex_free(ra, &p->images[i]);
+
+    p->num_images = num;
+    MP_TARRAY_GROW(NULL, p->images, p->num_images);
+    for (int i = 0; i < num; i++) {
+        p->images[i] = ra_vk_wrap_swapchain_img(ra, vkimages[i], sinfo);
+        if (!p->images[i])
+            goto error;
+    }
+
+    talloc_free(vkimages);
+    return true;
+
+error:
+    talloc_free(vkimages);
+    vkDestroySwapchainKHR(vk->dev, p->swapchain, MPVK_ALLOCATOR);
+    p->swapchain = NULL;
+    return false;
+}
+
+static int color_depth(struct ra_swapchain *sw)
+{
+    struct priv *p = sw->priv;
+    int bits = 0;
+
+    if (!p->num_images)
+        return bits;
+
+    // The channel with the most bits is probably the most authoritative about
+    // the actual color information (consider e.g. a2bgr10). Slight downside
+    // in that it results in rounding r/b for e.g. rgb565, but we don't pick
+    // surfaces with fewer than 8 bits anyway.
+    const struct ra_format *fmt = p->images[0]->params.format;
+    for (int i = 0; i < fmt->num_components; i++) {
+        int depth = fmt->component_depth[i];
+        bits = MPMAX(bits, depth ? depth : fmt->component_size[i]);
+    }
+
+    return bits;
+}
+
+static bool start_frame(struct ra_swapchain *sw, struct ra_fbo *out_fbo)
+{
+    struct priv *p = sw->priv;
+    struct mpvk_ctx *vk = p->vk;
+    if (!p->swapchain)
+        goto error;
+
+    uint32_t imgidx = 0;
+    MP_TRACE(vk, "vkAcquireNextImageKHR\n");
+    VkResult res = vkAcquireNextImageKHR(vk->dev, p->swapchain, UINT64_MAX,
+                                         p->acquired[p->idx_acquired], NULL,
+                                         &imgidx);
+    if (res == VK_ERROR_OUT_OF_DATE_KHR)
+        goto error; // just return in this case
+    VK_ASSERT(res, "Failed acquiring swapchain image");
+
+    p->last_imgidx = imgidx;
+    *out_fbo = (struct ra_fbo) {
+        .tex = p->images[imgidx],
+        .flip = false,
+    };
+    return true;
+
+error:
+    return false;
+}
+
+static bool submit_frame(struct ra_swapchain *sw, const struct vo_frame *frame)
+{
+    struct priv *p = sw->priv;
+    struct ra *ra = sw->ctx->ra;
+    struct mpvk_ctx *vk = p->vk;
+    if (!p->swapchain)
+        goto error;
+
+    VkSemaphore acquired = p->acquired[p->idx_acquired++];
+    p->idx_acquired %= p->num_acquired;
+
+    VkSemaphore done;
+    if (!ra_vk_submit(ra, p->images[p->last_imgidx], acquired, &done,
+                      &p->frames_in_flight))
+        goto error;
+
+    // For some reason, nvidia absolutely shits itself when presenting from a
+    // full queue - so advance all of the cmdpool indices first and then do the
+    // present on an "empty" queue
+    vk_cmd_cycle_queues(vk);
+    struct vk_cmdpool *pool = vk->pool;
+    VkQueue queue = pool->queues[pool->qindex];
+
+    VkPresentInfoKHR pinfo = {
+        .sType = VK_STRUCTURE_TYPE_PRESENT_INFO_KHR,
+        .waitSemaphoreCount = 1,
+        .pWaitSemaphores = &done,
+        .swapchainCount = 1,
+        .pSwapchains = &p->swapchain,
+        .pImageIndices = &p->last_imgidx,
+    };
+
+    VK(vkQueuePresentKHR(queue, &pinfo));
+    return true;
+
+error:
+    return false;
+}
+
+static void swap_buffers(struct ra_swapchain *sw)
+{
+    struct priv *p = sw->priv;
+
+    while (p->frames_in_flight >= sw->ctx->opts.swapchain_depth)
+        mpvk_dev_poll_cmds(p->vk, 100000); // 100μs
+}
+
+static const struct ra_swapchain_fns vulkan_swapchain = {
+    // .screenshot is not currently supported
+    .color_depth   = color_depth,
+    .start_frame   = start_frame,
+    .submit_frame  = submit_frame,
+    .swap_buffers  = swap_buffers,
+};
diff --git a/video/out/vulkan/context.h b/video/out/vulkan/context.h
new file mode 100644
index 0000000000..3f630bc10e
--- /dev/null
+++ b/video/out/vulkan/context.h
@@ -0,0 +1,10 @@
+#pragma once
+
+#include "video/out/gpu/context.h"
+#include "common.h"
+
+// Helpers for ra_ctx based on ra_vk. These initialize ctx->ra and ctx->swchain.
+void ra_vk_ctx_uninit(struct ra_ctx *ctx);
+bool ra_vk_ctx_init(struct ra_ctx *ctx, struct mpvk_ctx *vk,
+                    VkPresentModeKHR preferred_mode);
+bool ra_vk_ctx_resize(struct ra_swapchain *sw, int w, int h);
diff --git a/video/out/vulkan/context_xlib.c b/video/out/vulkan/context_xlib.c
new file mode 100644
index 0000000000..2611fbb706
--- /dev/null
+++ b/video/out/vulkan/context_xlib.c
@@ -0,0 +1,116 @@
+/*
+ * This file is part of mpv.
+ *
+ * mpv 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.
+ *
+ * mpv 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 mpv.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "video/out/gpu/context.h"
+#include "video/out/x11_common.h"
+
+#include "common.h"
+#include "context.h"
+#include "utils.h"
+
+struct priv {
+    struct mpvk_ctx vk;
+};
+
+static void xlib_uninit(struct ra_ctx *ctx)
+{
+    struct priv *p = ctx->priv;
+
+    ra_vk_ctx_uninit(ctx);
+    mpvk_uninit(&p->vk);
+    vo_x11_uninit(ctx->vo);
+}
+
+static bool xlib_init(struct ra_ctx *ctx)
+{
+    struct priv *p = ctx->priv = talloc_zero(ctx, struct priv);
+    struct mpvk_ctx *vk = &p->vk;
+    int msgl = ctx->opts.probing ? MSGL_V : MSGL_ERR;
+
+    if (!vo_x11_init(ctx->vo))
+        goto error;
+
+    if (!vo_x11_create_vo_window(ctx->vo, NULL, "mpvk"))
+        goto error;
+
+    if (!mpvk_instance_init(vk, ctx->log, ctx->opts.debug))
+        goto error;
+
+    VkXlibSurfaceCreateInfoKHR xinfo = {
+         .sType = VK_STRUCTURE_TYPE_XLIB_SURFACE_CREATE_INFO_KHR,
+         .dpy = ctx->vo->x11->display,
+         .window = ctx->vo->x11->window,
+    };
+
+    VkResult res = vkCreateXlibSurfaceKHR(vk->inst, &xinfo, MPVK_ALLOCATOR,
+                                          &vk->surf);
+    if (res != VK_SUCCESS) {
+        MP_MSG(ctx, msgl, "Failed creating Xlib surface: %s\n", vk_err(res));
+        goto error;
+    }
+
+    if (!ra_vk_ctx_init(ctx, vk, VK_PRESENT_MODE_FIFO_KHR))
+        goto error;
+
+    return true;
+
+error:
+    xlib_uninit(ctx);
+    return false;
+}
+
+static bool resize(struct ra_ctx *ctx)
+{
+    return ra_vk_ctx_resize(ctx->swapchain, ctx->vo->dwidth, ctx->vo->dheight);
+}
+
+static bool xlib_reconfig(struct ra_ctx *ctx)
+{
+    vo_x11_config_vo_window(ctx->vo);
+    return resize(ctx);
+}
+
+static int xlib_control(struct ra_ctx *ctx, int *events, int request, void *arg)
+{
+    int ret = vo_x11_control(ctx->vo, events, request, arg);
+    if (*events & VO_EVENT_RESIZE) {
+        if (!resize(ctx))
+            return VO_ERROR;
+    }
+    return ret;
+}
+
+static void xlib_wakeup(struct ra_ctx *ctx)
+{
+    vo_x11_wakeup(ctx->vo);
+}
+
+static void xlib_wait_events(struct ra_ctx *ctx, int64_t until_time_us)
+{
+    vo_x11_wait_events(ctx->vo, until_time_us);
+}
+
+const struct ra_ctx_fns ra_ctx_vulkan_xlib = {
+    .type           = "vulkan",
+    .name           = "x11",
+    .reconfig       = xlib_reconfig,
+    .control        = xlib_control,
+    .wakeup         = xlib_wakeup,
+    .wait_events    = xlib_wait_events,
+    .init           = xlib_init,
+    .uninit         = xlib_uninit,
+};
diff --git a/video/out/vulkan/formats.c b/video/out/vulkan/formats.c
new file mode 100644
index 0000000000..b44bead99c
--- /dev/null
+++ b/video/out/vulkan/formats.c
@@ -0,0 +1,55 @@
+#include "formats.h"
+
+const struct vk_format vk_formats[] = {
+    // Regular, byte-aligned integer formats
+    {"r8",       VK_FORMAT_R8_UNORM,                  1,  1,   {8             }, RA_CTYPE_UNORM },
+    {"rg8",      VK_FORMAT_R8G8_UNORM,                2,  2,   {8,  8         }, RA_CTYPE_UNORM },
+    {"rgb8",     VK_FORMAT_R8G8B8_UNORM,              3,  3,   {8,  8,  8     }, RA_CTYPE_UNORM },
+    {"rgba8",    VK_FORMAT_R8G8B8A8_UNORM,            4,  4,   {8,  8,  8,  8 }, RA_CTYPE_UNORM },
+    {"r16",      VK_FORMAT_R16_UNORM,                 1,  2,   {16            }, RA_CTYPE_UNORM },
+    {"rg16",     VK_FORMAT_R16G16_UNORM,              2,  4,   {16, 16        }, RA_CTYPE_UNORM },
+    {"rgb16",    VK_FORMAT_R16G16B16_UNORM,           3,  6,   {16, 16, 16    }, RA_CTYPE_UNORM },
+    {"rgba16",   VK_FORMAT_R16G16B16A16_UNORM,        4,  8,   {16, 16, 16, 16}, RA_CTYPE_UNORM },
+
+    // Special, integer-only formats
+    {"r32ui",    VK_FORMAT_R32_UINT,                  1,  4,   {32            }, RA_CTYPE_UINT },
+    {"rg32ui",   VK_FORMAT_R32G32_UINT,               2,  8,   {32, 32        }, RA_CTYPE_UINT },
+    {"rgb32ui",  VK_FORMAT_R32G32B32_UINT,            3,  12,  {32, 32, 32    }, RA_CTYPE_UINT },
+    {"rgba32ui", VK_FORMAT_R32G32B32A32_UINT,         4,  16,  {32, 32, 32, 32}, RA_CTYPE_UINT },
+    {"r64ui",    VK_FORMAT_R64_UINT,                  1,  8,   {64            }, RA_CTYPE_UINT },
+    {"rg64ui",   VK_FORMAT_R64G64_UINT,               2,  16,  {64, 64        }, RA_CTYPE_UINT },
+    {"rgb64ui",  VK_FORMAT_R64G64B64_UINT,            3,  24,  {64, 64, 64    }, RA_CTYPE_UINT },
+    {"rgba64ui", VK_FORMAT_R64G64B64A64_UINT,         4,  32,  {64, 64, 64, 64}, RA_CTYPE_UINT },
+
+    // Packed integer formats
+    {"rg4",      VK_FORMAT_R4G4_UNORM_PACK8,          2,  1,   {4,  4         }, RA_CTYPE_UNORM },
+    {"rgba4",    VK_FORMAT_R4G4B4A4_UNORM_PACK16,     4,  2,   {4,  4,  4,  4 }, RA_CTYPE_UNORM },
+    {"rgb565",   VK_FORMAT_R5G6B5_UNORM_PACK16,       3,  2,   {5,  6,  5     }, RA_CTYPE_UNORM },
+    {"rgb565a1", VK_FORMAT_R5G5B5A1_UNORM_PACK16,     4,  2,   {5,  5,  5,  1 }, RA_CTYPE_UNORM },
+
+    // Float formats (native formats, hf = half float, df = double float)
+    {"r16hf",    VK_FORMAT_R16_SFLOAT,                1,  2,   {16            }, RA_CTYPE_FLOAT },
+    {"rg16hf",   VK_FORMAT_R16G16_SFLOAT,             2,  4,   {16, 16        }, RA_CTYPE_FLOAT },
+    {"rgb16hf",  VK_FORMAT_R16G16B16_SFLOAT,          3,  6,   {16, 16, 16    }, RA_CTYPE_FLOAT },
+    {"rgba16hf", VK_FORMAT_R16G16B16A16_SFLOAT,       4,  8,   {16, 16, 16, 16}, RA_CTYPE_FLOAT },
+    {"r32f",     VK_FORMAT_R32_SFLOAT,                1,  4,   {32            }, RA_CTYPE_FLOAT },
+    {"rg32f",    VK_FORMAT_R32G32_SFLOAT,             2,  8,   {32, 32        }, RA_CTYPE_FLOAT },
+    {"rgb32f",   VK_FORMAT_R32G32B32_SFLOAT,          3, 12,   {32, 32, 32    }, RA_CTYPE_FLOAT },
+    {"rgba32f",  VK_FORMAT_R32G32B32A32_SFLOAT,       4, 16,   {32, 32, 32, 32}, RA_CTYPE_FLOAT },
+    {"r64df",    VK_FORMAT_R64_SFLOAT,                1,  8,   {64            }, RA_CTYPE_FLOAT },
+    {"rg64df",   VK_FORMAT_R64G64_SFLOAT,             2, 16,   {64, 64        }, RA_CTYPE_FLOAT },
+    {"rgb64df",  VK_FORMAT_R64G64B64_SFLOAT,          3, 24,   {64, 64, 64    }, RA_CTYPE_FLOAT },
+    {"rgba64df", VK_FORMAT_R64G64B64A64_SFLOAT,       4, 32,   {64, 64, 64, 64}, RA_CTYPE_FLOAT },
+
+    // "Swapped" component order images
+    {"bgr8",     VK_FORMAT_B8G8R8_UNORM,              3,  3,   {8,  8,  8     }, RA_CTYPE_UNORM, true },
+    {"bgra8",    VK_FORMAT_B8G8R8A8_UNORM,            4,  4,   {8,  8,  8,  8 }, RA_CTYPE_UNORM, true },
+    {"bgra4",    VK_FORMAT_B4G4R4A4_UNORM_PACK16,     4,  2,   {4,  4,  4,  4 }, RA_CTYPE_UNORM, true },
+    {"bgr565",   VK_FORMAT_B5G6R5_UNORM_PACK16,       3,  2,   {5,  6,  5     }, RA_CTYPE_UNORM, true },
+    {"bgr565a1", VK_FORMAT_B5G5R5A1_UNORM_PACK16,     4,  2,   {5,  5,  5,  1 }, RA_CTYPE_UNORM, true },
+    {"a1rgb5",   VK_FORMAT_A1R5G5B5_UNORM_PACK16,     4,  2,   {1,  5,  5,  5 }, RA_CTYPE_UNORM, true },
+    {"a2rgb10",  VK_FORMAT_A2R10G10B10_UNORM_PACK32,  4,  4,   {2,  10, 10, 10}, RA_CTYPE_UNORM, true },
+    {"a2bgr10",  VK_FORMAT_A2B10G10R10_UNORM_PACK32,  4,  4,   {2,  10, 10, 10}, RA_CTYPE_UNORM, true },
+    {"abgr8",    VK_FORMAT_A8B8G8R8_UNORM_PACK32,     4,  4,   {8,  8,  8,  8 }, RA_CTYPE_UNORM, true },
+    {0}
+};
diff --git a/video/out/vulkan/formats.h b/video/out/vulkan/formats.h
new file mode 100644
index 0000000000..22782a6958
--- /dev/null
+++ b/video/out/vulkan/formats.h
@@ -0,0 +1,16 @@
+#pragma once
+
+#include "video/out/gpu/ra.h"
+#include "common.h"
+
+struct vk_format {
+    const char *name;
+    VkFormat iformat;    // vulkan format enum
+    int components;      // how many components are there
+    int bytes;           // how many bytes is a texel
+    int bits[4];         // how many bits per component
+    enum ra_ctype ctype; // format representation type
+    bool fucked_order;   // used for formats which are not simply rgba
+};
+
+extern const struct vk_format vk_formats[];
diff --git a/video/out/vulkan/malloc.c b/video/out/vulkan/malloc.c
new file mode 100644
index 0000000000..31fcd36ddb
--- /dev/null
+++ b/video/out/vulkan/malloc.c
@@ -0,0 +1,424 @@
+#include "malloc.h"
+#include "utils.h"
+#include "osdep/timer.h"
+
+// Controls the multiplication factor for new slab allocations. The new slab
+// will always be allocated such that the size of the slab is this factor times
+// the previous slab. Higher values make it grow faster.
+#define MPVK_HEAP_SLAB_GROWTH_RATE 4
+
+// Controls the minimum slab size, to reduce the frequency at which very small
+// slabs would need to get allocated when allocating the first few buffers.
+// (Default: 1 MB)
+#define MPVK_HEAP_MINIMUM_SLAB_SIZE (1 << 20)
+
+// Controls the maximum slab size, to reduce the effect of unbounded slab
+// growth exhausting memory. If the application needs a single allocation
+// that's bigger than this value, it will be allocated directly from the
+// device. (Default: 512 MB)
+#define MPVK_HEAP_MAXIMUM_SLAB_SIZE (1 << 29)
+
+// Controls the minimum free region size, to reduce thrashing the free space
+// map with lots of small buffers during uninit. (Default: 1 KB)
+#define MPVK_HEAP_MINIMUM_REGION_SIZE (1 << 10)
+
+// Represents a region of available memory
+struct vk_region {
+    size_t start; // first offset in region
+    size_t end;   // first offset *not* in region
+};
+
+static inline size_t region_len(struct vk_region r)
+{
+    return r.end - r.start;
+}
+
+// A single slab represents a contiguous region of allocated memory. Actual
+// allocations are served as slices of this. Slabs are organized into linked
+// lists, which represent individual heaps.
+struct vk_slab {
+    VkDeviceMemory mem;   // underlying device allocation
+    size_t size;          // total size of `slab`
+    size_t used;          // number of bytes actually in use (for GC accounting)
+    bool dedicated;       // slab is allocated specifically for one object
+    // free space map: a sorted list of memory regions that are available
+    struct vk_region *regions;
+    int num_regions;
+    // optional, depends on the memory type:
+    VkBuffer buffer;      // buffer spanning the entire slab
+    void *data;           // mapped memory corresponding to `mem`
+};
+
+// Represents a single memory heap. We keep track of a vk_heap for each
+// combination of buffer type and memory selection parameters. This shouldn't
+// actually be that many in practice, because some combinations simply never
+// occur, and others will generally be the same for the same objects.
+struct vk_heap {
+    VkBufferUsageFlagBits usage;    // the buffer usage type (or 0)
+    VkMemoryPropertyFlagBits flags; // the memory type flags (or 0)
+    uint32_t typeBits;              // the memory type index requirements (or 0)
+    struct vk_slab **slabs;         // array of slabs sorted by size
+    int num_slabs;
+};
+
+// The overall state of the allocator, which keeps track of a vk_heap for each
+// memory type.
+struct vk_malloc {
+    VkPhysicalDeviceMemoryProperties props;
+    struct vk_heap *heaps;
+    int num_heaps;
+};
+
+static void slab_free(struct mpvk_ctx *vk, struct vk_slab *slab)
+{
+    if (!slab)
+        return;
+
+    assert(slab->used == 0);
+
+    int64_t start = mp_time_us();
+    vkDestroyBuffer(vk->dev, slab->buffer, MPVK_ALLOCATOR);
+    // also implicitly unmaps the memory if needed
+    vkFreeMemory(vk->dev, slab->mem, MPVK_ALLOCATOR);
+    int64_t stop = mp_time_us();
+
+    MP_VERBOSE(vk, "Freeing slab of size %zu took %lld μs.\n",
+               slab->size, (long long)(stop - start));
+
+    talloc_free(slab);
+}
+
+static bool find_best_memtype(struct mpvk_ctx *vk, uint32_t typeBits,
+                              VkMemoryPropertyFlagBits flags,
+                              VkMemoryType *out_type, int *out_index)
+{
+    struct vk_malloc *ma = vk->alloc;
+
+    // The vulkan spec requires memory types to be sorted in the "optimal"
+    // order, so the first matching type we find will be the best/fastest one.
+    for (int i = 0; i < ma->props.memoryTypeCount; i++) {
+        // The memory type flags must include our properties
+        if ((ma->props.memoryTypes[i].propertyFlags & flags) != flags)
+            continue;
+        // The memory type must be supported by the requirements (bitfield)
+        if (typeBits && !(typeBits & (1 << i)))
+            continue;
+        *out_type = ma->props.memoryTypes[i];
+        *out_index = i;
+        return true;
+    }
+
+    MP_ERR(vk, "Found no memory type matching property flags 0x%x and type "
+               "bits 0x%x!\n", flags, (unsigned)typeBits);
+    return false;
+}
+
+static struct vk_slab *slab_alloc(struct mpvk_ctx *vk, struct vk_heap *heap,
+                                  size_t size)
+{
+    struct vk_slab *slab = talloc_ptrtype(NULL, slab);
+    *slab = (struct vk_slab) {
+        .size = size,
+    };
+
+    MP_TARRAY_APPEND(slab, slab->regions, slab->num_regions, (struct vk_region) {
+        .start = 0,
+        .end   = slab->size,
+    });
+
+    VkMemoryAllocateInfo minfo = {
+        .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
+        .allocationSize = slab->size,
+    };
+
+    uint32_t typeBits = heap->typeBits ? heap->typeBits : UINT32_MAX;
+    if (heap->usage) {
+        VkBufferCreateInfo binfo = {
+            .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+            .size  = slab->size,
+            .usage = heap->usage,
+            .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+        };
+
+        VK(vkCreateBuffer(vk->dev, &binfo, MPVK_ALLOCATOR, &slab->buffer));
+
+        VkMemoryRequirements reqs;
+        vkGetBufferMemoryRequirements(vk->dev, slab->buffer, &reqs);
+        minfo.allocationSize = reqs.size; // this can be larger than slab->size
+        typeBits &= reqs.memoryTypeBits;  // this can restrict the types
+    }
+
+    VkMemoryType type;
+    int index;
+    if (!find_best_memtype(vk, typeBits, heap->flags, &type, &index))
+        goto error;
+
+    MP_VERBOSE(vk, "Allocating %zu memory of type 0x%x (id %d) in heap %d.\n",
+               slab->size, type.propertyFlags, index, (int)type.heapIndex);
+
+    minfo.memoryTypeIndex = index;
+    VK(vkAllocateMemory(vk->dev, &minfo, MPVK_ALLOCATOR, &slab->mem));
+
+    if (heap->flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT)
+        VK(vkMapMemory(vk->dev, slab->mem, 0, VK_WHOLE_SIZE, 0, &slab->data));
+
+    if (slab->buffer)
+        VK(vkBindBufferMemory(vk->dev, slab->buffer, slab->mem, 0));
+
+    return slab;
+
+error:
+    slab_free(vk, slab);
+    return NULL;
+}
+
+static void insert_region(struct vk_slab *slab, struct vk_region region)
+{
+    if (region.start == region.end)
+        return;
+
+    bool big_enough = region_len(region) >= MPVK_HEAP_MINIMUM_REGION_SIZE;
+
+    // Find the index of the first region that comes after this
+    for (int i = 0; i < slab->num_regions; i++) {
+        struct vk_region *r = &slab->regions[i];
+
+        // Check for a few special cases which can be coalesced
+        if (r->end == region.start) {
+            // The new region is at the tail of this region. In addition to
+            // modifying this region, we also need to coalesce all the following
+            // regions for as long as possible
+            r->end = region.end;
+
+            struct vk_region *next = &slab->regions[i+1];
+            while (i+1 < slab->num_regions && r->end == next->start) {
+                r->end = next->end;
+                MP_TARRAY_REMOVE_AT(slab->regions, slab->num_regions, i+1);
+            }
+            return;
+        }
+
+        if (r->start == region.end) {
+            // The new region is at the head of this region. We don't need to
+            // do anything special here - because if this could be further
+            // coalesced backwards, the previous loop iteration would already
+            // have caught it.
+            r->start = region.start;
+            return;
+        }
+
+        if (r->start > region.start) {
+            // The new region comes somewhere before this region, so insert
+            // it into this index in the array.
+            if (big_enough) {
+                MP_TARRAY_INSERT_AT(slab, slab->regions, slab->num_regions,
+                                    i, region);
+            }
+            return;
+        }
+    }
+
+    // If we've reached the end of this loop, then all of the regions
+    // come before the new region, and are disconnected - so append it
+    if (big_enough)
+        MP_TARRAY_APPEND(slab, slab->regions, slab->num_regions, region);
+}
+
+static void heap_uninit(struct mpvk_ctx *vk, struct vk_heap *heap)
+{
+    for (int i = 0; i < heap->num_slabs; i++)
+        slab_free(vk, heap->slabs[i]);
+
+    talloc_free(heap->slabs);
+    *heap = (struct vk_heap){0};
+}
+
+void vk_malloc_init(struct mpvk_ctx *vk)
+{
+    assert(vk->physd);
+    vk->alloc = talloc_zero(NULL, struct vk_malloc);
+    vkGetPhysicalDeviceMemoryProperties(vk->physd, &vk->alloc->props);
+}
+
+void vk_malloc_uninit(struct mpvk_ctx *vk)
+{
+    struct vk_malloc *ma = vk->alloc;
+    if (!ma)
+        return;
+
+    for (int i = 0; i < ma->num_heaps; i++)
+        heap_uninit(vk, &ma->heaps[i]);
+
+    talloc_free(ma);
+    vk->alloc = NULL;
+}
+
+void vk_free_memslice(struct mpvk_ctx *vk, struct vk_memslice slice)
+{
+    struct vk_slab *slab = slice.priv;
+    if (!slab)
+        return;
+
+    assert(slab->used >= slice.size);
+    slab->used -= slice.size;
+
+    MP_DBG(vk, "Freeing slice %zu + %zu from slab with size %zu\n",
+           slice.offset, slice.size, slab->size);
+
+    if (slab->dedicated) {
+        // If the slab was purpose-allocated for this memslice, we can just
+        // free it here
+        slab_free(vk, slab);
+    } else {
+        // Return the allocation to the free space map
+        insert_region(slab, (struct vk_region) {
+            .start = slice.offset,
+            .end   = slice.offset + slice.size,
+        });
+    }
+}
+
+// reqs: can be NULL
+static struct vk_heap *find_heap(struct mpvk_ctx *vk,
+                                 VkBufferUsageFlagBits usage,
+                                 VkMemoryPropertyFlagBits flags,
+                                 VkMemoryRequirements *reqs)
+{
+    struct vk_malloc *ma = vk->alloc;
+    int typeBits = reqs ? reqs->memoryTypeBits : 0;
+
+    for (int i = 0; i < ma->num_heaps; i++) {
+        if (ma->heaps[i].usage != usage)
+            continue;
+        if (ma->heaps[i].flags != flags)
+            continue;
+        if (ma->heaps[i].typeBits != typeBits)
+            continue;
+        return &ma->heaps[i];
+    }
+
+    // Not found => add it
+    MP_TARRAY_GROW(ma, ma->heaps, ma->num_heaps + 1);
+    struct vk_heap *heap = &ma->heaps[ma->num_heaps++];
+    *heap = (struct vk_heap) {
+        .usage    = usage,
+        .flags    = flags,
+        .typeBits = typeBits,
+    };
+    return heap;
+}
+
+static inline bool region_fits(struct vk_region r, size_t size, size_t align)
+{
+    return MP_ALIGN_UP(r.start, align) + size <= r.end;
+}
+
+// Finds the best-fitting region in a heap. If the heap is too small or too
+// fragmented, a new slab will be allocated under the hood.
+static bool heap_get_region(struct mpvk_ctx *vk, struct vk_heap *heap,
+                            size_t size, size_t align,
+                            struct vk_slab **out_slab, int *out_index)
+{
+    struct vk_slab *slab = NULL;
+
+    // If the allocation is very big, serve it directly instead of bothering
+    // with the heap
+    if (size > MPVK_HEAP_MAXIMUM_SLAB_SIZE) {
+        slab = slab_alloc(vk, heap, size);
+        *out_slab = slab;
+        *out_index = 0;
+        return !!slab;
+    }
+
+    for (int i = 0; i < heap->num_slabs; i++) {
+        slab = heap->slabs[i];
+        if (slab->size < size)
+            continue;
+
+        // Attempt a best fit search
+        int best = -1;
+        for (int n = 0; n < slab->num_regions; n++) {
+            struct vk_region r = slab->regions[n];
+            if (!region_fits(r, size, align))
+                continue;
+            if (best >= 0 && region_len(r) > region_len(slab->regions[best]))
+                continue;
+            best = n;
+        }
+
+        if (best >= 0) {
+            *out_slab = slab;
+            *out_index = best;
+            return true;
+        }
+    }
+
+    // Otherwise, allocate a new vk_slab and append it to the list.
+    size_t cur_size = MPMAX(size, slab ? slab->size : 0);
+    size_t slab_size = MPVK_HEAP_SLAB_GROWTH_RATE * cur_size;
+    slab_size = MPMAX(MPVK_HEAP_MINIMUM_SLAB_SIZE, slab_size);
+    slab_size = MPMIN(MPVK_HEAP_MAXIMUM_SLAB_SIZE, slab_size);
+    assert(slab_size >= size);
+    slab = slab_alloc(vk, heap, slab_size);
+    if (!slab)
+        return false;
+    MP_TARRAY_APPEND(NULL, heap->slabs, heap->num_slabs, slab);
+
+    // Return the only region there is in a newly allocated slab
+    assert(slab->num_regions == 1);
+    *out_slab = slab;
+    *out_index = 0;
+    return true;
+}
+
+static bool slice_heap(struct mpvk_ctx *vk, struct vk_heap *heap, size_t size,
+                       size_t alignment, struct vk_memslice *out)
+{
+    struct vk_slab *slab;
+    int index;
+    alignment = MP_ALIGN_UP(alignment, vk->limits.bufferImageGranularity);
+    if (!heap_get_region(vk, heap, size, alignment, &slab, &index))
+        return false;
+
+    struct vk_region reg = slab->regions[index];
+    MP_TARRAY_REMOVE_AT(slab->regions, slab->num_regions, index);
+    *out = (struct vk_memslice) {
+        .vkmem = slab->mem,
+        .offset = MP_ALIGN_UP(reg.start, alignment),
+        .size = size,
+        .priv = slab,
+    };
+
+    MP_DBG(vk, "Sub-allocating slice %zu + %zu from slab with size %zu\n",
+           out->offset, out->size, slab->size);
+
+    size_t out_end = out->offset + out->size;
+    insert_region(slab, (struct vk_region) { reg.start, out->offset });
+    insert_region(slab, (struct vk_region) { out_end, reg.end });
+
+    slab->used += size;
+    return true;
+}
+
+bool vk_malloc_generic(struct mpvk_ctx *vk, VkMemoryRequirements reqs,
+                       VkMemoryPropertyFlagBits flags, struct vk_memslice *out)
+{
+    struct vk_heap *heap = find_heap(vk, 0, flags, &reqs);
+    return slice_heap(vk, heap, reqs.size, reqs.alignment, out);
+}
+
+bool vk_malloc_buffer(struct mpvk_ctx *vk, VkBufferUsageFlagBits bufFlags,
+                      VkMemoryPropertyFlagBits memFlags, VkDeviceSize size,
+                      VkDeviceSize alignment, struct vk_bufslice *out)
+{
+    struct vk_heap *heap = find_heap(vk, bufFlags, memFlags, NULL);
+    if (!slice_heap(vk, heap, size, alignment, &out->mem))
+        return false;
+
+    struct vk_slab *slab = out->mem.priv;
+    out->buf = slab->buffer;
+    if (slab->data)
+        out->data = (void *)((uintptr_t)slab->data + (ptrdiff_t)out->mem.offset);
+
+    return true;
+}
diff --git a/video/out/vulkan/malloc.h b/video/out/vulkan/malloc.h
new file mode 100644
index 0000000000..65c1036929
--- /dev/null
+++ b/video/out/vulkan/malloc.h
@@ -0,0 +1,35 @@
+#pragma once
+
+#include "common.h"
+
+void vk_malloc_init(struct mpvk_ctx *vk);
+void vk_malloc_uninit(struct mpvk_ctx *vk);
+
+// Represents a single "slice" of generic (non-buffer) memory, plus some
+// metadata for accounting. This struct is essentially read-only.
+struct vk_memslice {
+    VkDeviceMemory vkmem;
+    size_t offset;
+    size_t size;
+    void *priv;
+};
+
+void vk_free_memslice(struct mpvk_ctx *vk, struct vk_memslice slice);
+bool vk_malloc_generic(struct mpvk_ctx *vk, VkMemoryRequirements reqs,
+                       VkMemoryPropertyFlagBits flags, struct vk_memslice *out);
+
+// Represents a single "slice" of a larger buffer
+struct vk_bufslice {
+    struct vk_memslice mem; // must be freed by the user when done
+    VkBuffer buf;           // the buffer this memory was sliced from
+    // For persistently mapped buffers, this points to the first usable byte of
+    // this slice.
+    void *data;
+};
+
+// Allocate a buffer slice. This is more efficient than vk_malloc_generic for
+// when the user needs lots of buffers, since it doesn't require
+// creating/destroying lots of (little) VkBuffers.
+bool vk_malloc_buffer(struct mpvk_ctx *vk, VkBufferUsageFlagBits bufFlags,
+                      VkMemoryPropertyFlagBits memFlags, VkDeviceSize size,
+                      VkDeviceSize alignment, struct vk_bufslice *out);
diff --git a/video/out/vulkan/ra_vk.c b/video/out/vulkan/ra_vk.c
new file mode 100644
index 0000000000..ce0cbc66e9
--- /dev/null
+++ b/video/out/vulkan/ra_vk.c
@@ -0,0 +1,1590 @@
+#include "ra_vk.h"
+#include "malloc.h"
+#include "video/out/opengl/utils.h"
+
+static struct ra_fns ra_fns_vk;
+
+// For ra.priv
+struct ra_vk {
+    struct mpvk_ctx *vk;
+    struct ra_tex *clear_tex; // stupid hack for clear()
+    struct vk_cmd *cmd;       // currently recording cmd
+};
+
+struct mpvk_ctx *ra_vk_get(struct ra *ra)
+{
+    if (ra->fns != &ra_fns_vk)
+        return NULL;
+
+    struct ra_vk *p = ra->priv;
+    return p->vk;
+}
+
+// Returns a command buffer, or NULL on error
+static struct vk_cmd *vk_require_cmd(struct ra *ra)
+{
+    struct ra_vk *p = ra->priv;
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+
+    if (!p->cmd)
+        p->cmd = vk_cmd_begin(vk, vk->pool);
+
+    return p->cmd;
+}
+
+// Note: This technically follows the flush() API, but we don't need
+// to expose that (and in fact, it's a bad idea) since we control flushing
+// behavior with ra_vk_present_frame already.
+static bool vk_flush(struct ra *ra, VkSemaphore *done)
+{
+    struct ra_vk *p = ra->priv;
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+
+    if (p->cmd) {
+        if (!vk_cmd_submit(vk, p->cmd, done))
+            return false;
+        p->cmd = NULL;
+    }
+
+    return true;
+}
+
+// The callback's *priv will always be set to `ra`
+static void vk_callback(struct ra *ra, vk_cb callback, void *arg)
+{
+    struct ra_vk *p = ra->priv;
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+
+    if (p->cmd) {
+        vk_cmd_callback(p->cmd, callback, ra, arg);
+    } else {
+        vk_dev_callback(vk, callback, ra, arg);
+    }
+}
+
+#define MAKE_LAZY_DESTRUCTOR(fun, argtype)                  \
+    static void fun##_lazy(struct ra *ra, argtype *arg) {   \
+        vk_callback(ra, (vk_cb) fun, arg);                  \
+    }
+
+static void vk_destroy_ra(struct ra *ra)
+{
+    struct ra_vk *p = ra->priv;
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+
+    vk_flush(ra, NULL);
+    mpvk_dev_wait_idle(vk);
+    ra_tex_free(ra, &p->clear_tex);
+
+    talloc_free(ra);
+}
+
+static bool vk_setup_formats(struct ra *ra)
+{
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+
+    for (const struct vk_format *vk_fmt = vk_formats; vk_fmt->name; vk_fmt++) {
+        VkFormatProperties prop;
+        vkGetPhysicalDeviceFormatProperties(vk->physd, vk_fmt->iformat, &prop);
+
+        // As a bare minimum, we need to sample from an allocated image
+        VkFormatFeatureFlags flags = prop.optimalTilingFeatures;
+        if (!(flags & VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT))
+            continue;
+
+        VkFormatFeatureFlags linear_bits, render_bits;
+        linear_bits = VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT;
+        render_bits = VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT |
+                      VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BLEND_BIT;
+
+        struct ra_format *fmt = talloc_zero(ra, struct ra_format);
+        *fmt = (struct ra_format) {
+            .name            = vk_fmt->name,
+            .priv            = (void *)vk_fmt,
+            .ctype           = vk_fmt->ctype,
+            .ordered         = !vk_fmt->fucked_order,
+            .num_components  = vk_fmt->components,
+            .pixel_size      = vk_fmt->bytes,
+            .linear_filter   = !!(flags & linear_bits),
+            .renderable      = !!(flags & render_bits),
+        };
+
+        for (int i = 0; i < 4; i++)
+            fmt->component_size[i] = fmt->component_depth[i] = vk_fmt->bits[i];
+
+        MP_TARRAY_APPEND(ra, ra->formats, ra->num_formats, fmt);
+    }
+
+    // Populate some other capabilities related to formats while we're at it
+    VkImageType imgType[3] = {
+        VK_IMAGE_TYPE_1D,
+        VK_IMAGE_TYPE_2D,
+        VK_IMAGE_TYPE_3D
+    };
+
+    // R8_UNORM is supported on literally every single vulkan implementation
+    const VkFormat testfmt = VK_FORMAT_R8_UNORM;
+
+    for (int d = 0; d < 3; d++) {
+        VkImageFormatProperties iprop;
+        VkResult res = vkGetPhysicalDeviceImageFormatProperties(vk->physd,
+                testfmt, imgType[d], VK_IMAGE_TILING_OPTIMAL,
+                VK_IMAGE_USAGE_SAMPLED_BIT, 0, &iprop);
+
+        switch (imgType[d]) {
+        case VK_IMAGE_TYPE_1D:
+            if (res == VK_SUCCESS)
+                ra->caps |= RA_CAP_TEX_1D;
+            break;
+        case VK_IMAGE_TYPE_2D:
+            // 2D formats must be supported by RA, so ensure this is the case
+            VK_ASSERT(res, "Querying 2D format limits");
+            ra->max_texture_wh = MPMIN(iprop.maxExtent.width, iprop.maxExtent.height);
+            break;
+        case VK_IMAGE_TYPE_3D:
+            if (res == VK_SUCCESS)
+                ra->caps |= RA_CAP_TEX_3D;
+            break;
+        }
+    }
+
+    // RA_CAP_BLIT implies both blitting between images as well as blitting
+    // directly to the swapchain image, so check for all three operations
+    bool blittable = true;
+    VkFormatProperties prop;
+    vkGetPhysicalDeviceFormatProperties(vk->physd, testfmt, &prop);
+    if (!(prop.optimalTilingFeatures & VK_FORMAT_FEATURE_BLIT_SRC_BIT))
+        blittable = false;
+    if (!(prop.optimalTilingFeatures & VK_FORMAT_FEATURE_BLIT_DST_BIT))
+        blittable = false;
+
+    vkGetPhysicalDeviceFormatProperties(vk->physd, vk->surf_format.format, &prop);
+    if (!(prop.optimalTilingFeatures & VK_FORMAT_FEATURE_BLIT_DST_BIT))
+        blittable = false;
+
+    if (blittable)
+        ra->caps |= RA_CAP_BLIT;
+
+    return true;
+
+error:
+    return false;
+}
+
+static struct ra_fns ra_fns_vk;
+
+struct ra *ra_create_vk(struct mpvk_ctx *vk, struct mp_log *log)
+{
+    assert(vk->dev);
+    assert(vk->alloc);
+
+    struct ra *ra = talloc_zero(NULL, struct ra);
+    ra->log = log;
+    ra->fns = &ra_fns_vk;
+
+    struct ra_vk *p = ra->priv = talloc_zero(ra, struct ra_vk);
+    p->vk = vk;
+
+    // There's no way to query the supported GLSL version from VK_NV_glsl_shader
+    // (thanks nvidia), so just pick the GL version that modern nvidia devices
+    // support..
+    ra->glsl_version = 450;
+    ra->glsl_vulkan = true;
+    ra->max_shmem = vk->limits.maxComputeSharedMemorySize;
+    ra->caps = RA_CAP_NESTED_ARRAY;
+
+    if (vk->pool->props.queueFlags & VK_QUEUE_COMPUTE_BIT)
+        ra->caps |= RA_CAP_COMPUTE;
+
+    if (!vk_setup_formats(ra))
+        goto error;
+
+    // UBO support is required
+    ra->caps |= RA_CAP_BUF_RO;
+
+    // Try creating a shader storage buffer
+    struct ra_buf_params ssbo_params = {
+        .type = RA_BUF_TYPE_SHADER_STORAGE,
+        .size = 16,
+    };
+
+    struct ra_buf *ssbo = ra_buf_create(ra, &ssbo_params);
+    if (ssbo) {
+        ra->caps |= RA_CAP_BUF_RW;
+        ra_buf_free(ra, &ssbo);
+    }
+
+    // To support clear() by region, we need to allocate a dummy 1x1 image that
+    // will be used as the source of blit operations
+    struct ra_tex_params clear_params = {
+        .dimensions = 1, // no point in using a 2D image if height = 1
+        .w = 1,
+        .h = 1,
+        .d = 1,
+        .format = ra_find_float16_format(ra, 4),
+        .blit_src = 1,
+        .host_mutable = 1,
+    };
+
+    p->clear_tex = ra_tex_create(ra, &clear_params);
+    if (!p->clear_tex) {
+        MP_ERR(ra, "Failed creating 1x1 dummy texture for clear()!\n");
+        goto error;
+    }
+
+    return ra;
+
+error:
+    vk_destroy_ra(ra);
+    return NULL;
+}
+
+// Boilerplate wrapper around vkCreateRenderPass to ensure passes remain
+// compatible
+static VkResult vk_create_render_pass(VkDevice dev, const struct ra_format *fmt,
+                                      bool load_fbo, VkRenderPass *out)
+{
+    struct vk_format *vk_fmt = fmt->priv;
+    assert(fmt->renderable);
+
+    VkRenderPassCreateInfo rinfo = {
+        .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
+        .attachmentCount = 1,
+        .pAttachments = &(VkAttachmentDescription) {
+            .format = vk_fmt->iformat,
+            .samples = VK_SAMPLE_COUNT_1_BIT,
+            .loadOp = load_fbo ? VK_ATTACHMENT_LOAD_OP_LOAD
+                               : VK_ATTACHMENT_LOAD_OP_DONT_CARE,
+            .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
+            .initialLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
+            .finalLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
+        },
+        .subpassCount = 1,
+        .pSubpasses = &(VkSubpassDescription) {
+            .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
+            .colorAttachmentCount = 1,
+            .pColorAttachments = &(VkAttachmentReference) {
+                .attachment = 0,
+                .layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
+            },
+        },
+    };
+
+    return vkCreateRenderPass(dev, &rinfo, MPVK_ALLOCATOR, out);
+}
+
+// For ra_tex.priv
+struct ra_tex_vk {
+    bool external_img;
+    VkImageType type;
+    VkImage img;
+    struct vk_memslice mem;
+    // for sampling
+    VkImageView view;
+    VkSampler sampler;
+    // for rendering
+    VkFramebuffer framebuffer;
+    VkRenderPass dummyPass;
+    // for uploading
+    struct ra_buf_pool pbo;
+    // "current" metadata, can change during the course of execution
+    VkImageLayout current_layout;
+    VkPipelineStageFlagBits current_stage;
+    VkAccessFlagBits current_access;
+};
+
+// Small helper to ease image barrier creation. if `discard` is set, the contents
+// of the image will be undefined after the barrier
+static void tex_barrier(struct vk_cmd *cmd, struct ra_tex_vk *tex_vk,
+                        VkPipelineStageFlagBits newStage,
+                        VkAccessFlagBits newAccess, VkImageLayout newLayout,
+                        bool discard)
+{
+    VkImageMemoryBarrier imgBarrier = {
+        .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+        .oldLayout = tex_vk->current_layout,
+        .newLayout = newLayout,
+        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+        .srcAccessMask = tex_vk->current_access,
+        .dstAccessMask = newAccess,
+        .image = tex_vk->img,
+        .subresourceRange = vk_range,
+    };
+
+    if (discard) {
+        imgBarrier.oldLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+        imgBarrier.srcAccessMask = 0;
+    }
+
+    if (imgBarrier.oldLayout != imgBarrier.newLayout ||
+        imgBarrier.srcAccessMask != imgBarrier.dstAccessMask)
+    {
+        vkCmdPipelineBarrier(cmd->buf, tex_vk->current_stage, newStage, 0,
+                             0, NULL, 0, NULL, 1, &imgBarrier);
+    }
+
+    tex_vk->current_stage = newStage;
+    tex_vk->current_layout = newLayout;
+    tex_vk->current_access = newAccess;
+}
+
+static void vk_tex_destroy(struct ra *ra, struct ra_tex *tex)
+{
+    if (!tex)
+        return;
+
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+    struct ra_tex_vk *tex_vk = tex->priv;
+
+    ra_buf_pool_uninit(ra, &tex_vk->pbo);
+    vkDestroyFramebuffer(vk->dev, tex_vk->framebuffer, MPVK_ALLOCATOR);
+    vkDestroyRenderPass(vk->dev, tex_vk->dummyPass, MPVK_ALLOCATOR);
+    vkDestroySampler(vk->dev, tex_vk->sampler, MPVK_ALLOCATOR);
+    vkDestroyImageView(vk->dev, tex_vk->view, MPVK_ALLOCATOR);
+    if (!tex_vk->external_img) {
+        vkDestroyImage(vk->dev, tex_vk->img, MPVK_ALLOCATOR);
+        vk_free_memslice(vk, tex_vk->mem);
+    }
+
+    talloc_free(tex);
+}
+
+MAKE_LAZY_DESTRUCTOR(vk_tex_destroy, struct ra_tex);
+
+// Initializes non-VkImage values like the image view, samplers, etc.
+static bool vk_init_image(struct ra *ra, struct ra_tex *tex)
+{
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+
+    struct ra_tex_params *params = &tex->params;
+    struct ra_tex_vk *tex_vk = tex->priv;
+    assert(tex_vk->img);
+
+    tex_vk->current_layout = VK_IMAGE_LAYOUT_UNDEFINED;
+    tex_vk->current_stage = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+    tex_vk->current_access = 0;
+
+    if (params->render_src || params->render_dst) {
+        static const VkImageViewType viewType[] = {
+            [VK_IMAGE_TYPE_1D] = VK_IMAGE_VIEW_TYPE_1D,
+            [VK_IMAGE_TYPE_2D] = VK_IMAGE_VIEW_TYPE_2D,
+            [VK_IMAGE_TYPE_3D] = VK_IMAGE_VIEW_TYPE_3D,
+        };
+
+        const struct vk_format *fmt = params->format->priv;
+        VkImageViewCreateInfo vinfo = {
+            .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
+            .image = tex_vk->img,
+            .viewType = viewType[tex_vk->type],
+            .format = fmt->iformat,
+            .subresourceRange = vk_range,
+        };
+
+        VK(vkCreateImageView(vk->dev, &vinfo, MPVK_ALLOCATOR, &tex_vk->view));
+    }
+
+    if (params->render_src) {
+        assert(params->format->linear_filter || !params->src_linear);
+        VkFilter filter = params->src_linear
+            ? VK_FILTER_LINEAR
+            : VK_FILTER_NEAREST;
+        VkSamplerAddressMode wrap = params->src_repeat
+            ? VK_SAMPLER_ADDRESS_MODE_REPEAT
+            : VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
+        VkSamplerCreateInfo sinfo = {
+            .sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO,
+            .magFilter = filter,
+            .minFilter = filter,
+            .addressModeU = wrap,
+            .addressModeV = wrap,
+            .addressModeW = wrap,
+            .maxAnisotropy = 1.0,
+        };
+
+        VK(vkCreateSampler(vk->dev, &sinfo, MPVK_ALLOCATOR, &tex_vk->sampler));
+    }
+
+    if (params->render_dst) {
+        // Framebuffers need to be created against a specific render pass
+        // layout, so we need to temporarily create a skeleton/dummy render
+        // pass for vulkan to figure out the compatibility
+        VK(vk_create_render_pass(vk->dev, params->format, false, &tex_vk->dummyPass));
+
+        VkFramebufferCreateInfo finfo = {
+            .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
+            .renderPass = tex_vk->dummyPass,
+            .attachmentCount = 1,
+            .pAttachments = &tex_vk->view,
+            .width = tex->params.w,
+            .height = tex->params.h,
+            .layers = 1,
+        };
+
+        VK(vkCreateFramebuffer(vk->dev, &finfo, MPVK_ALLOCATOR,
+                               &tex_vk->framebuffer));
+
+        // NOTE: Normally we would free the dummyPass again here, but a bug
+        // in the nvidia vulkan driver causes a segfault if you do.
+    }
+
+    return true;
+
+error:
+    return false;
+}
+
+static struct ra_tex *vk_tex_create(struct ra *ra,
+                                    const struct ra_tex_params *params)
+{
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+
+    struct ra_tex *tex = talloc_zero(NULL, struct ra_tex);
+    tex->params = *params;
+    tex->params.initial_data = NULL;
+
+    struct ra_tex_vk *tex_vk = tex->priv = talloc_zero(tex, struct ra_tex_vk);
+
+    const struct vk_format *fmt = params->format->priv;
+    switch (params->dimensions) {
+    case 1: tex_vk->type = VK_IMAGE_TYPE_1D; break;
+    case 2: tex_vk->type = VK_IMAGE_TYPE_2D; break;
+    case 3: tex_vk->type = VK_IMAGE_TYPE_3D; break;
+    default: abort();
+    }
+
+    VkImageUsageFlags usage = 0;
+    if (params->render_src)
+        usage |= VK_IMAGE_USAGE_SAMPLED_BIT;
+    if (params->render_dst)
+        usage |= VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
+    if (params->storage_dst)
+        usage |= VK_IMAGE_USAGE_STORAGE_BIT;
+    if (params->blit_src)
+        usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
+    if (params->host_mutable || params->blit_dst || params->initial_data)
+        usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT;
+
+    // Double-check image usage support and fail immediately if invalid
+    VkImageFormatProperties iprop;
+    VkResult res = vkGetPhysicalDeviceImageFormatProperties(vk->physd,
+            fmt->iformat, tex_vk->type, VK_IMAGE_TILING_OPTIMAL, usage, 0,
+            &iprop);
+    if (res == VK_ERROR_FORMAT_NOT_SUPPORTED) {
+        return NULL;
+    } else {
+        VK_ASSERT(res, "Querying image format properties");
+    }
+
+    VkFormatProperties prop;
+    vkGetPhysicalDeviceFormatProperties(vk->physd, fmt->iformat, &prop);
+    VkFormatFeatureFlags flags = prop.optimalTilingFeatures;
+
+    bool has_blit_src   = flags & VK_FORMAT_FEATURE_BLIT_SRC_BIT,
+         has_src_linear = flags & VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT;
+
+    if (params->w > iprop.maxExtent.width ||
+        params->h > iprop.maxExtent.height ||
+        params->d > iprop.maxExtent.depth ||
+        (params->blit_src && !has_blit_src) ||
+        (params->src_linear && !has_src_linear))
+    {
+        return NULL;
+    }
+
+    VkImageCreateInfo iinfo = {
+        .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
+        .imageType = tex_vk->type,
+        .format = fmt->iformat,
+        .extent = (VkExtent3D) { params->w, params->h, params->d },
+        .mipLevels = 1,
+        .arrayLayers = 1,
+        .samples = VK_SAMPLE_COUNT_1_BIT,
+        .tiling = VK_IMAGE_TILING_OPTIMAL,
+        .usage = usage,
+        .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
+        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+        .queueFamilyIndexCount = 1,
+        .pQueueFamilyIndices = &vk->pool->qf,
+    };
+
+    VK(vkCreateImage(vk->dev, &iinfo, MPVK_ALLOCATOR, &tex_vk->img));
+
+    VkMemoryPropertyFlagBits memFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+    VkMemoryRequirements reqs;
+    vkGetImageMemoryRequirements(vk->dev, tex_vk->img, &reqs);
+
+    struct vk_memslice *mem = &tex_vk->mem;
+    if (!vk_malloc_generic(vk, reqs, memFlags, mem))
+        goto error;
+
+    VK(vkBindImageMemory(vk->dev, tex_vk->img, mem->vkmem, mem->offset));
+
+    if (!vk_init_image(ra, tex))
+        goto error;
+
+    if (params->initial_data) {
+        struct ra_tex_upload_params ul_params = {
+            .tex = tex,
+            .invalidate = true,
+            .src = params->initial_data,
+            .stride = params->w * fmt->bytes,
+        };
+        if (!ra->fns->tex_upload(ra, &ul_params))
+            goto error;
+    }
+
+    return tex;
+
+error:
+    vk_tex_destroy(ra, tex);
+    return NULL;
+}
+
+struct ra_tex *ra_vk_wrap_swapchain_img(struct ra *ra, VkImage vkimg,
+                                        VkSwapchainCreateInfoKHR info)
+{
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+    struct ra_tex *tex = NULL;
+
+    const struct ra_format *format = NULL;
+    for (int i = 0; i < ra->num_formats; i++) {
+        const struct vk_format *fmt = ra->formats[i]->priv;
+        if (fmt->iformat == vk->surf_format.format) {
+            format = ra->formats[i];
+            break;
+        }
+    }
+
+    if (!format) {
+        MP_ERR(ra, "Could not find ra_format suitable for wrapped swchain image "
+                   "with surface format 0x%x\n", vk->surf_format.format);
+        goto error;
+    }
+
+    tex = talloc_zero(NULL, struct ra_tex);
+    tex->params = (struct ra_tex_params) {
+        .format = format,
+        .dimensions = 2,
+        .w = info.imageExtent.width,
+        .h = info.imageExtent.height,
+        .d = 1,
+        .blit_src    = !!(info.imageUsage & VK_IMAGE_USAGE_TRANSFER_SRC_BIT),
+        .blit_dst    = !!(info.imageUsage & VK_IMAGE_USAGE_TRANSFER_DST_BIT),
+        .render_src  = !!(info.imageUsage & VK_IMAGE_USAGE_SAMPLED_BIT),
+        .render_dst  = !!(info.imageUsage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT),
+        .storage_dst = !!(info.imageUsage & VK_IMAGE_USAGE_STORAGE_BIT),
+    };
+
+    struct ra_tex_vk *tex_vk = tex->priv = talloc_zero(tex, struct ra_tex_vk);
+    tex_vk->type = VK_IMAGE_TYPE_2D;
+    tex_vk->external_img = true;
+    tex_vk->img = vkimg;
+
+    if (!vk_init_image(ra, tex))
+        goto error;
+
+    return tex;
+
+error:
+    vk_tex_destroy(ra, tex);
+    return NULL;
+}
+
+// For ra_buf.priv
+struct ra_buf_vk {
+    struct vk_bufslice slice;
+    int refcount; // 1 = object allocated but not in use, > 1 = in use
+    bool needsflush;
+    // "current" metadata, can change during course of execution
+    VkPipelineStageFlagBits current_stage;
+    VkAccessFlagBits current_access;
+};
+
+static void vk_buf_deref(struct ra *ra, struct ra_buf *buf)
+{
+    if (!buf)
+        return;
+
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+    struct ra_buf_vk *buf_vk = buf->priv;
+
+    if (--buf_vk->refcount == 0) {
+        vk_free_memslice(vk, buf_vk->slice.mem);
+        talloc_free(buf);
+    }
+}
+
+static void buf_barrier(struct ra *ra, struct vk_cmd *cmd, struct ra_buf *buf,
+                        VkPipelineStageFlagBits newStage,
+                        VkAccessFlagBits newAccess, int offset, size_t size)
+{
+    struct ra_buf_vk *buf_vk = buf->priv;
+
+    VkBufferMemoryBarrier buffBarrier = {
+        .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
+        .srcAccessMask = buf_vk->current_access,
+        .dstAccessMask = newAccess,
+        .buffer = buf_vk->slice.buf,
+        .offset = offset,
+        .size = size,
+    };
+
+    if (buf_vk->needsflush || buf->params.host_mapped) {
+        buffBarrier.srcAccessMask = VK_ACCESS_HOST_WRITE_BIT;
+        buf_vk->current_stage = VK_PIPELINE_STAGE_HOST_BIT;
+        buf_vk->needsflush = false;
+    }
+
+    if (buffBarrier.srcAccessMask != buffBarrier.dstAccessMask) {
+        vkCmdPipelineBarrier(cmd->buf, buf_vk->current_stage, newStage, 0,
+                             0, NULL, 1, &buffBarrier, 0, NULL);
+    }
+
+    buf_vk->current_stage = newStage;
+    buf_vk->current_access = newAccess;
+    buf_vk->refcount++;
+    vk_cmd_callback(cmd, (vk_cb) vk_buf_deref, ra, buf);
+}
+
+#define vk_buf_destroy vk_buf_deref
+MAKE_LAZY_DESTRUCTOR(vk_buf_destroy, struct ra_buf);
+
+static void vk_buf_update(struct ra *ra, struct ra_buf *buf, ptrdiff_t offset,
+                          const void *data, size_t size)
+{
+    assert(buf->params.host_mutable || buf->params.initial_data);
+    struct ra_buf_vk *buf_vk = buf->priv;
+
+    // For host-mapped buffers, we can just directly memcpy the buffer contents.
+    // Otherwise, we can update the buffer from the GPU using a command buffer.
+    if (buf_vk->slice.data) {
+        assert(offset + size <= buf->params.size);
+        uintptr_t addr = (uintptr_t)buf_vk->slice.data + offset;
+        memcpy((void *)addr, data, size);
+        buf_vk->needsflush = true;
+    } else {
+        struct vk_cmd *cmd = vk_require_cmd(ra);
+        if (!cmd) {
+            MP_ERR(ra, "Failed updating buffer!\n");
+            return;
+        }
+
+        buf_barrier(ra, cmd, buf, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                    VK_ACCESS_TRANSFER_WRITE_BIT, offset, size);
+
+        VkDeviceSize bufOffset = buf_vk->slice.mem.offset + offset;
+        assert(bufOffset == MP_ALIGN_UP(bufOffset, 4));
+        vkCmdUpdateBuffer(cmd->buf, buf_vk->slice.buf, bufOffset, size, data);
+    }
+}
+
+static struct ra_buf *vk_buf_create(struct ra *ra,
+                                    const struct ra_buf_params *params)
+{
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+
+    struct ra_buf *buf = talloc_zero(NULL, struct ra_buf);
+    buf->params = *params;
+
+    struct ra_buf_vk *buf_vk = buf->priv = talloc_zero(buf, struct ra_buf_vk);
+    buf_vk->current_stage = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+    buf_vk->current_access = 0;
+    buf_vk->refcount = 1;
+
+    VkBufferUsageFlagBits bufFlags = 0;
+    VkMemoryPropertyFlagBits memFlags = 0;
+    VkDeviceSize align = 4; // alignment 4 is needed for buf_update
+
+    switch (params->type) {
+    case RA_BUF_TYPE_TEX_UPLOAD:
+        bufFlags |= VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
+        memFlags |= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT;
+        break;
+    case RA_BUF_TYPE_UNIFORM:
+        bufFlags |= VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT;
+        memFlags |= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+        align = MP_ALIGN_UP(align, vk->limits.minUniformBufferOffsetAlignment);
+        break;
+    case RA_BUF_TYPE_SHADER_STORAGE:
+        bufFlags |= VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
+        memFlags |= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+        align = MP_ALIGN_UP(align, vk->limits.minStorageBufferOffsetAlignment);
+        break;
+    case RA_BUF_TYPE_VERTEX:
+        bufFlags |= VK_BUFFER_USAGE_VERTEX_BUFFER_BIT;
+        memFlags |= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+        break;
+    default: abort();
+    }
+
+    if (params->host_mutable || params->initial_data) {
+        bufFlags |= VK_BUFFER_USAGE_TRANSFER_DST_BIT;
+        align = MP_ALIGN_UP(align, vk->limits.optimalBufferCopyOffsetAlignment);
+    }
+
+    if (params->host_mapped) {
+        memFlags |= VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
+                    VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
+                    VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
+    }
+
+    if (!vk_malloc_buffer(vk, bufFlags, memFlags, params->size, align,
+                          &buf_vk->slice))
+    {
+        goto error;
+    }
+
+    if (params->host_mapped)
+        buf->data = buf_vk->slice.data;
+
+    if (params->initial_data)
+        vk_buf_update(ra, buf, 0, params->initial_data, params->size);
+
+    buf->params.initial_data = NULL; // do this after vk_buf_update
+    return buf;
+
+error:
+    vk_buf_destroy(ra, buf);
+    return NULL;
+}
+
+static bool vk_buf_poll(struct ra *ra, struct ra_buf *buf)
+{
+    struct ra_buf_vk *buf_vk = buf->priv;
+    return buf_vk->refcount == 1;
+}
+
+static bool vk_tex_upload(struct ra *ra,
+                          const struct ra_tex_upload_params *params)
+{
+    struct ra_tex *tex = params->tex;
+    struct ra_tex_vk *tex_vk = tex->priv;
+
+    if (!params->buf)
+        return ra_tex_upload_pbo(ra, &tex_vk->pbo, params);
+
+    assert(!params->src);
+    assert(params->buf);
+    struct ra_buf *buf = params->buf;
+    struct ra_buf_vk *buf_vk = buf->priv;
+
+    VkBufferImageCopy region = {
+        .bufferOffset = buf_vk->slice.mem.offset + params->buf_offset,
+        .bufferRowLength = tex->params.w,
+        .bufferImageHeight = tex->params.h,
+        .imageSubresource = vk_layers,
+        .imageExtent = (VkExtent3D){tex->params.w, tex->params.h, tex->params.d},
+    };
+
+    if (tex->params.dimensions == 2) {
+        int pix_size = tex->params.format->pixel_size;
+        region.bufferRowLength = params->stride / pix_size;
+        if (region.bufferRowLength * pix_size != params->stride) {
+            MP_ERR(ra, "Texture upload strides must be a multiple of the texel "
+                       "size!\n");
+            goto error;
+        }
+
+        if (params->rc) {
+            struct mp_rect *rc = params->rc;
+            region.imageOffset = (VkOffset3D){rc->x0, rc->y0, 0};
+            region.imageExtent = (VkExtent3D){mp_rect_w(*rc), mp_rect_h(*rc), 1};
+        }
+    }
+
+    uint64_t size = region.bufferRowLength * region.bufferImageHeight *
+                    region.imageExtent.depth;
+
+    struct vk_cmd *cmd = vk_require_cmd(ra);
+    if (!cmd)
+        goto error;
+
+    buf_barrier(ra, cmd, buf, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                VK_ACCESS_TRANSFER_READ_BIT, region.bufferOffset, size);
+
+    tex_barrier(cmd, tex_vk, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                VK_ACCESS_TRANSFER_WRITE_BIT,
+                VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                params->invalidate);
+
+    vkCmdCopyBufferToImage(cmd->buf, buf_vk->slice.buf, tex_vk->img,
+                           tex_vk->current_layout, 1, &region);
+
+    return true;
+
+error:
+    return false;
+}
+
+#define MPVK_NUM_DS MPVK_MAX_STREAMING_DEPTH
+
+// For ra_renderpass.priv
+struct ra_renderpass_vk {
+    // Compiled shaders
+    VkShaderModule vert;
+    VkShaderModule frag;
+    VkShaderModule comp;
+    // Pipeline / render pass
+    VkPipeline pipe;
+    VkPipelineLayout pipeLayout;
+    VkPipelineCache pipeCache;
+    VkRenderPass renderPass;
+    // Descriptor set (bindings)
+    VkDescriptorSetLayout dsLayout;
+    VkDescriptorPool dsPool;
+    VkDescriptorSet dss[MPVK_NUM_DS];
+    int dindex;
+    // Vertex buffers (vertices)
+    struct ra_buf_pool vbo;
+
+    // For updating
+    VkWriteDescriptorSet *dswrite;
+    VkDescriptorImageInfo *dsiinfo;
+    VkDescriptorBufferInfo *dsbinfo;
+};
+
+static void vk_renderpass_destroy(struct ra *ra, struct ra_renderpass *pass)
+{
+    if (!pass)
+        return;
+
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+    struct ra_renderpass_vk *pass_vk = pass->priv;
+
+    ra_buf_pool_uninit(ra, &pass_vk->vbo);
+    vkDestroyPipeline(vk->dev, pass_vk->pipe, MPVK_ALLOCATOR);
+    vkDestroyPipelineCache(vk->dev, pass_vk->pipeCache, MPVK_ALLOCATOR);
+    vkDestroyRenderPass(vk->dev, pass_vk->renderPass, MPVK_ALLOCATOR);
+    vkDestroyPipelineLayout(vk->dev, pass_vk->pipeLayout, MPVK_ALLOCATOR);
+    vkDestroyDescriptorPool(vk->dev, pass_vk->dsPool, MPVK_ALLOCATOR);
+    vkDestroyDescriptorSetLayout(vk->dev, pass_vk->dsLayout, MPVK_ALLOCATOR);
+    vkDestroyShaderModule(vk->dev, pass_vk->vert, MPVK_ALLOCATOR);
+    vkDestroyShaderModule(vk->dev, pass_vk->frag, MPVK_ALLOCATOR);
+    vkDestroyShaderModule(vk->dev, pass_vk->comp, MPVK_ALLOCATOR);
+
+    talloc_free(pass);
+}
+
+MAKE_LAZY_DESTRUCTOR(vk_renderpass_destroy, struct ra_renderpass);
+
+static const VkDescriptorType dsType[] = {
+    [RA_VARTYPE_TEX]    = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
+    [RA_VARTYPE_IMG_W]  = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+    [RA_VARTYPE_BUF_RO] = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
+    [RA_VARTYPE_BUF_RW] = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+};
+
+static bool vk_get_input_format(struct ra *ra, struct ra_renderpass_input *inp,
+                                VkFormat *out_fmt)
+{
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+
+    enum ra_ctype ctype;
+    switch (inp->type) {
+    case RA_VARTYPE_FLOAT:      ctype = RA_CTYPE_FLOAT; break;
+    case RA_VARTYPE_BYTE_UNORM: ctype = RA_CTYPE_UNORM; break;
+    default: abort();
+    }
+
+    assert(inp->dim_m == 1);
+    for (const struct vk_format *fmt = vk_formats; fmt->name; fmt++) {
+        if (fmt->ctype != ctype)
+            continue;
+        if (fmt->components != inp->dim_v)
+            continue;
+        if (fmt->bytes != ra_renderpass_input_layout(inp).size)
+            continue;
+
+        // Ensure this format is valid for vertex attributes
+        VkFormatProperties prop;
+        vkGetPhysicalDeviceFormatProperties(vk->physd, fmt->iformat, &prop);
+        if (!(prop.bufferFeatures & VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT))
+            continue;
+
+        *out_fmt = fmt->iformat;
+        return true;
+    }
+
+    return false;
+}
+
+static const VkPipelineStageFlagBits stageFlags[] = {
+    [RA_RENDERPASS_TYPE_RASTER]  = VK_SHADER_STAGE_FRAGMENT_BIT,
+    [RA_RENDERPASS_TYPE_COMPUTE] = VK_SHADER_STAGE_COMPUTE_BIT,
+};
+
+static struct ra_renderpass *vk_renderpass_create(struct ra *ra,
+                                    const struct ra_renderpass_params *params)
+{
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+
+    struct ra_renderpass *pass = talloc_zero(NULL, struct ra_renderpass);
+    pass->params = *ra_renderpass_params_copy(pass, params);
+    pass->params.cached_program = (bstr){0};
+    struct ra_renderpass_vk *pass_vk = pass->priv =
+        talloc_zero(pass, struct ra_renderpass_vk);
+
+    static int dsCount[RA_VARTYPE_COUNT] = {0};
+    VkDescriptorSetLayoutBinding *bindings = NULL;
+    int num_bindings = 0;
+
+    for (int i = 0; i < params->num_inputs; i++) {
+        struct ra_renderpass_input *inp = &params->inputs[i];
+        switch (inp->type) {
+        case RA_VARTYPE_TEX:
+        case RA_VARTYPE_IMG_W:
+        case RA_VARTYPE_BUF_RO:
+        case RA_VARTYPE_BUF_RW: {
+            VkDescriptorSetLayoutBinding desc = {
+                .binding = inp->binding,
+                .descriptorType = dsType[inp->type],
+                .descriptorCount = 1,
+                .stageFlags = stageFlags[params->type],
+            };
+
+            MP_TARRAY_APPEND(pass, bindings, num_bindings, desc);
+            dsCount[inp->type]++;
+            break;
+        }
+        default: abort();
+        }
+    }
+
+    VkDescriptorPoolSize *dsPoolSizes = NULL;
+    int poolSizeCount = 0;
+    for (enum ra_vartype t = 0; t < RA_VARTYPE_COUNT; t++) {
+        if (dsCount[t] > 0) {
+            VkDescriptorPoolSize dssize = {
+                .type = dsType[t],
+                .descriptorCount = dsCount[t] * MPVK_NUM_DS,
+            };
+
+            MP_TARRAY_APPEND(pass, dsPoolSizes, poolSizeCount, dssize);
+        }
+    }
+
+    VkDescriptorPoolCreateInfo pinfo = {
+        .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
+        .maxSets = MPVK_NUM_DS,
+        .pPoolSizes = dsPoolSizes,
+        .poolSizeCount = poolSizeCount,
+    };
+
+    VK(vkCreateDescriptorPool(vk->dev, &pinfo, MPVK_ALLOCATOR, &pass_vk->dsPool));
+    talloc_free(dsPoolSizes);
+
+    pass_vk->dswrite = talloc_array(pass, VkWriteDescriptorSet, num_bindings);
+    pass_vk->dsiinfo = talloc_array(pass, VkDescriptorImageInfo, num_bindings);
+    pass_vk->dsbinfo = talloc_array(pass, VkDescriptorBufferInfo, num_bindings);
+
+    VkDescriptorSetLayoutCreateInfo dinfo = {
+        .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
+        .pBindings = bindings,
+        .bindingCount = num_bindings,
+    };
+
+    VK(vkCreateDescriptorSetLayout(vk->dev, &dinfo, MPVK_ALLOCATOR,
+                                   &pass_vk->dsLayout));
+
+    VkDescriptorSetLayout layouts[MPVK_NUM_DS];
+    for (int i = 0; i < MPVK_NUM_DS; i++)
+        layouts[i] = pass_vk->dsLayout;
+
+    VkDescriptorSetAllocateInfo ainfo = {
+        .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO,
+        .descriptorPool = pass_vk->dsPool,
+        .descriptorSetCount = MPVK_NUM_DS,
+        .pSetLayouts = layouts,
+    };
+
+    VK(vkAllocateDescriptorSets(vk->dev, &ainfo, pass_vk->dss));
+
+    VkPipelineLayoutCreateInfo linfo = {
+        .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
+        .setLayoutCount = 1,
+        .pSetLayouts = &pass_vk->dsLayout,
+    };
+
+    VK(vkCreatePipelineLayout(vk->dev, &linfo, MPVK_ALLOCATOR,
+                              &pass_vk->pipeLayout));
+
+    VkPipelineCacheCreateInfo pcinfo = {
+        .sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO,
+        .pInitialData = params->cached_program.start,
+        .initialDataSize = params->cached_program.len,
+    };
+
+    VK(vkCreatePipelineCache(vk->dev, &pcinfo, MPVK_ALLOCATOR, &pass_vk->pipeCache));
+
+    VkShaderModuleCreateInfo sinfo = {
+        .sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO,
+    };
+
+    switch (params->type) {
+    case RA_RENDERPASS_TYPE_RASTER: {
+        sinfo.pCode = (uint32_t *)params->vertex_shader;
+        sinfo.codeSize = strlen(params->vertex_shader);
+        VK(vkCreateShaderModule(vk->dev, &sinfo, MPVK_ALLOCATOR, &pass_vk->vert));
+
+        sinfo.pCode = (uint32_t *)params->frag_shader;
+        sinfo.codeSize = strlen(params->frag_shader);
+        VK(vkCreateShaderModule(vk->dev, &sinfo, MPVK_ALLOCATOR, &pass_vk->frag));
+
+        VK(vk_create_render_pass(vk->dev, params->target_format,
+                                 params->enable_blend, &pass_vk->renderPass));
+
+        VkPipelineShaderStageCreateInfo stages[] = {
+            {
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+                .stage = VK_SHADER_STAGE_VERTEX_BIT,
+                .module = pass_vk->vert,
+                .pName = "main",
+            },
+            {
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+                .stage = VK_SHADER_STAGE_FRAGMENT_BIT,
+                .module = pass_vk->frag,
+                .pName = "main",
+            }
+        };
+
+        VkVertexInputAttributeDescription *attrs = talloc_array(pass,
+                VkVertexInputAttributeDescription, params->num_vertex_attribs);
+
+        for (int i = 0; i < params->num_vertex_attribs; i++) {
+            struct ra_renderpass_input *inp = &params->vertex_attribs[i];
+            attrs[i] = (VkVertexInputAttributeDescription) {
+                .location = i,
+                .binding = 0,
+                .offset = inp->offset,
+            };
+
+            if (!vk_get_input_format(ra, inp, &attrs[i].format)) {
+                MP_ERR(ra, "No suitable VkFormat for vertex attrib '%s'!\n",
+                       inp->name);
+                goto error;
+            }
+        }
+
+        static const VkBlendFactor blendFactors[] = {
+            [RA_BLEND_ZERO]                = VK_BLEND_FACTOR_ZERO,
+            [RA_BLEND_ONE]                 = VK_BLEND_FACTOR_ONE,
+            [RA_BLEND_SRC_ALPHA]           = VK_BLEND_FACTOR_SRC_ALPHA,
+            [RA_BLEND_ONE_MINUS_SRC_ALPHA] = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA,
+        };
+
+        VkPipelineColorBlendAttachmentState binfo = {
+            .blendEnable = params->enable_blend,
+            .colorBlendOp = VK_BLEND_OP_ADD,
+            .srcColorBlendFactor = blendFactors[params->blend_src_rgb],
+            .dstColorBlendFactor = blendFactors[params->blend_dst_rgb],
+            .alphaBlendOp = VK_BLEND_OP_ADD,
+            .srcAlphaBlendFactor = blendFactors[params->blend_src_alpha],
+            .dstAlphaBlendFactor = blendFactors[params->blend_dst_alpha],
+            .colorWriteMask = VK_COLOR_COMPONENT_R_BIT |
+                              VK_COLOR_COMPONENT_G_BIT |
+                              VK_COLOR_COMPONENT_B_BIT |
+                              VK_COLOR_COMPONENT_A_BIT,
+        };
+
+        VkGraphicsPipelineCreateInfo cinfo = {
+            .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
+            .stageCount = MP_ARRAY_SIZE(stages),
+            .pStages = &stages[0],
+            .pVertexInputState = &(VkPipelineVertexInputStateCreateInfo) {
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
+                .vertexBindingDescriptionCount = 1,
+                .pVertexBindingDescriptions = &(VkVertexInputBindingDescription) {
+                    .binding = 0,
+                    .stride = params->vertex_stride,
+                    .inputRate = VK_VERTEX_INPUT_RATE_VERTEX,
+                },
+                .vertexAttributeDescriptionCount = params->num_vertex_attribs,
+                .pVertexAttributeDescriptions = attrs,
+            },
+            .pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
+                .topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST,
+            },
+            .pViewportState = &(VkPipelineViewportStateCreateInfo) {
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
+                .viewportCount = 1,
+                .scissorCount = 1,
+            },
+            .pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
+                .polygonMode = VK_POLYGON_MODE_FILL,
+                .cullMode = VK_CULL_MODE_NONE,
+                .lineWidth = 1.0f,
+            },
+            .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
+                .rasterizationSamples = VK_SAMPLE_COUNT_1_BIT,
+            },
+            .pColorBlendState = &(VkPipelineColorBlendStateCreateInfo) {
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
+                .attachmentCount = 1,
+                .pAttachments = &binfo,
+            },
+            .pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
+                .dynamicStateCount = 2,
+                .pDynamicStates = (VkDynamicState[]){
+                    VK_DYNAMIC_STATE_VIEWPORT,
+                    VK_DYNAMIC_STATE_SCISSOR,
+                },
+            },
+            .layout = pass_vk->pipeLayout,
+            .renderPass = pass_vk->renderPass,
+        };
+
+        VK(vkCreateGraphicsPipelines(vk->dev, pass_vk->pipeCache, 1, &cinfo,
+                                     MPVK_ALLOCATOR, &pass_vk->pipe));
+        break;
+    }
+    case RA_RENDERPASS_TYPE_COMPUTE: {
+        sinfo.pCode = (uint32_t *)params->compute_shader;
+        sinfo.codeSize = strlen(params->compute_shader);
+        VK(vkCreateShaderModule(vk->dev, &sinfo, MPVK_ALLOCATOR, &pass_vk->comp));
+
+        VkComputePipelineCreateInfo cinfo = {
+            .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
+            .stage = {
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+                .stage = VK_SHADER_STAGE_COMPUTE_BIT,
+                .module = pass_vk->comp,
+                .pName = "main",
+            },
+            .layout = pass_vk->pipeLayout,
+        };
+
+        VK(vkCreateComputePipelines(vk->dev, pass_vk->pipeCache, 1, &cinfo,
+                                    MPVK_ALLOCATOR, &pass_vk->pipe));
+        break;
+    }
+    }
+
+    // Update cached program
+    bstr *prog = &pass->params.cached_program;
+    VK(vkGetPipelineCacheData(vk->dev, pass_vk->pipeCache, &prog->len, NULL));
+    prog->start = talloc_size(pass, prog->len);
+    VK(vkGetPipelineCacheData(vk->dev, pass_vk->pipeCache, &prog->len, prog->start));
+
+    return pass;
+
+error:
+    vk_renderpass_destroy(ra, pass);
+    return NULL;
+}
+
+static void vk_update_descriptor(struct ra *ra, struct vk_cmd *cmd,
+                                 struct ra_renderpass *pass,
+                                 struct ra_renderpass_input_val val,
+                                 VkDescriptorSet ds, int idx)
+{
+    struct ra_renderpass_vk *pass_vk = pass->priv;
+    struct ra_renderpass_input *inp = &pass->params.inputs[val.index];
+
+    VkWriteDescriptorSet *wds = &pass_vk->dswrite[idx];
+    *wds = (VkWriteDescriptorSet) {
+        .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+        .dstSet = ds,
+        .dstBinding = inp->binding,
+        .descriptorCount = 1,
+        .descriptorType = dsType[inp->type],
+    };
+
+    static const VkPipelineStageFlags passStages[] = {
+        [RA_RENDERPASS_TYPE_RASTER]  = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
+        [RA_RENDERPASS_TYPE_COMPUTE] = VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+    };
+
+    switch (inp->type) {
+    case RA_VARTYPE_TEX: {
+        struct ra_tex *tex = *(struct ra_tex **)val.data;
+        struct ra_tex_vk *tex_vk = tex->priv;
+
+        assert(tex->params.render_src);
+        tex_barrier(cmd, tex_vk, passStages[pass->params.type],
+                    VK_ACCESS_SHADER_READ_BIT,
+                    VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, false);
+
+        VkDescriptorImageInfo *iinfo = &pass_vk->dsiinfo[idx];
+        *iinfo = (VkDescriptorImageInfo) {
+            .sampler = tex_vk->sampler,
+            .imageView = tex_vk->view,
+            .imageLayout = tex_vk->current_layout,
+        };
+
+        wds->pImageInfo = iinfo;
+        break;
+    }
+    case RA_VARTYPE_IMG_W: {
+        struct ra_tex *tex = *(struct ra_tex **)val.data;
+        struct ra_tex_vk *tex_vk = tex->priv;
+
+        assert(tex->params.storage_dst);
+        tex_barrier(cmd, tex_vk, passStages[pass->params.type],
+                    VK_ACCESS_SHADER_WRITE_BIT,
+                    VK_IMAGE_LAYOUT_GENERAL, false);
+
+        VkDescriptorImageInfo *iinfo = &pass_vk->dsiinfo[idx];
+        *iinfo = (VkDescriptorImageInfo) {
+            .imageView = tex_vk->view,
+            .imageLayout = tex_vk->current_layout,
+        };
+
+        wds->pImageInfo = iinfo;
+        break;
+    }
+    case RA_VARTYPE_BUF_RO:
+    case RA_VARTYPE_BUF_RW: {
+        struct ra_buf *buf = *(struct ra_buf **)val.data;
+        struct ra_buf_vk *buf_vk = buf->priv;
+
+        VkBufferUsageFlags access = VK_ACCESS_SHADER_READ_BIT;
+        if (inp->type == RA_VARTYPE_BUF_RW)
+            access |= VK_ACCESS_SHADER_WRITE_BIT;
+
+        buf_barrier(ra, cmd, buf, passStages[pass->params.type],
+                    access, buf_vk->slice.mem.offset, buf->params.size);
+
+        VkDescriptorBufferInfo *binfo = &pass_vk->dsbinfo[idx];
+        *binfo = (VkDescriptorBufferInfo) {
+            .buffer = buf_vk->slice.buf,
+            .offset = buf_vk->slice.mem.offset,
+            .range = buf->params.size,
+        };
+
+        wds->pBufferInfo = binfo;
+        break;
+    }
+    }
+}
+
+static void vk_renderpass_run(struct ra *ra,
+                              const struct ra_renderpass_run_params *params)
+{
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+    struct ra_renderpass *pass = params->pass;
+    struct ra_renderpass_vk *pass_vk = pass->priv;
+
+    struct vk_cmd *cmd = vk_require_cmd(ra);
+    if (!cmd)
+        goto error;
+
+    static const VkPipelineBindPoint bindPoint[] = {
+        [RA_RENDERPASS_TYPE_RASTER]  = VK_PIPELINE_BIND_POINT_GRAPHICS,
+        [RA_RENDERPASS_TYPE_COMPUTE] = VK_PIPELINE_BIND_POINT_COMPUTE,
+    };
+
+    vkCmdBindPipeline(cmd->buf, bindPoint[pass->params.type], pass_vk->pipe);
+
+    VkDescriptorSet ds = pass_vk->dss[pass_vk->dindex++];
+    pass_vk->dindex %= MPVK_NUM_DS;
+
+    for (int i = 0; i < params->num_values; i++)
+        vk_update_descriptor(ra, cmd, pass, params->values[i], ds, i);
+
+    if (params->num_values > 0) {
+        vkUpdateDescriptorSets(vk->dev, params->num_values, pass_vk->dswrite,
+                               0, NULL);
+    }
+
+    vkCmdBindDescriptorSets(cmd->buf, bindPoint[pass->params.type],
+                            pass_vk->pipeLayout, 0, 1, &ds, 0, NULL);
+
+    switch (pass->params.type) {
+    case RA_RENDERPASS_TYPE_COMPUTE:
+        vkCmdDispatch(cmd->buf, params->compute_groups[0],
+                      params->compute_groups[1],
+                      params->compute_groups[2]);
+        break;
+    case RA_RENDERPASS_TYPE_RASTER: {
+        struct ra_tex *tex = params->target;
+        struct ra_tex_vk *tex_vk = tex->priv;
+        assert(tex->params.render_dst);
+
+        struct ra_buf_params buf_params = {
+            .type = RA_BUF_TYPE_VERTEX,
+            .size = params->vertex_count * pass->params.vertex_stride,
+            .host_mutable = true,
+        };
+
+        struct ra_buf *buf = ra_buf_pool_get(ra, &pass_vk->vbo, &buf_params);
+        if (!buf) {
+            MP_ERR(ra, "Failed allocating vertex buffer!\n");
+            goto error;
+        }
+        struct ra_buf_vk *buf_vk = buf->priv;
+
+        vk_buf_update(ra, buf, 0, params->vertex_data, buf_params.size);
+
+        buf_barrier(ra, cmd, buf, VK_PIPELINE_STAGE_VERTEX_INPUT_BIT,
+                    VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT,
+                    buf_vk->slice.mem.offset, buf_params.size);
+
+        vkCmdBindVertexBuffers(cmd->buf, 0, 1, &buf_vk->slice.buf,
+                               &buf_vk->slice.mem.offset);
+
+        tex_barrier(cmd, tex_vk, VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+                    VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+                    VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, false);
+
+        VkViewport viewport = {
+            .x = params->viewport.x0,
+            .y = params->viewport.y0,
+            .width  = mp_rect_w(params->viewport),
+            .height = mp_rect_h(params->viewport),
+        };
+
+        VkRect2D scissor = {
+            .offset = {params->scissors.x0, params->scissors.y0},
+            .extent = {mp_rect_w(params->scissors), mp_rect_h(params->scissors)},
+        };
+
+        vkCmdSetViewport(cmd->buf, 0, 1, &viewport);
+        vkCmdSetScissor(cmd->buf, 0, 1, &scissor);
+
+        VkRenderPassBeginInfo binfo = {
+            .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+            .renderPass = pass_vk->renderPass,
+            .framebuffer = tex_vk->framebuffer,
+            .renderArea = (VkRect2D){{0, 0}, {tex->params.w, tex->params.h}},
+        };
+
+        vkCmdBeginRenderPass(cmd->buf, &binfo, VK_SUBPASS_CONTENTS_INLINE);
+        vkCmdDraw(cmd->buf, params->vertex_count, 1, 0, 0);
+        vkCmdEndRenderPass(cmd->buf);
+        break;
+    }
+    default: abort();
+    };
+
+error:
+    return;
+}
+
+static void vk_blit(struct ra *ra, struct ra_tex *dst, struct ra_tex *src,
+                    struct mp_rect *dst_rc, struct mp_rect *src_rc)
+{
+    assert(src->params.blit_src);
+    assert(dst->params.blit_dst);
+
+    struct ra_tex_vk *src_vk = src->priv;
+    struct ra_tex_vk *dst_vk = dst->priv;
+
+    struct vk_cmd *cmd = vk_require_cmd(ra);
+    if (!cmd)
+        return;
+
+    tex_barrier(cmd, src_vk, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                VK_ACCESS_TRANSFER_READ_BIT,
+                VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+                false);
+
+    bool discard = dst_rc->x0 == 0 &&
+                   dst_rc->y0 == 0 &&
+                   dst_rc->x1 == dst->params.w &&
+                   dst_rc->y1 == dst->params.h;
+
+    tex_barrier(cmd, dst_vk, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                VK_ACCESS_TRANSFER_WRITE_BIT,
+                VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                discard);
+
+    VkImageBlit region = {
+        .srcSubresource = vk_layers,
+        .srcOffsets = {{src_rc->x0, src_rc->y0, 0}, {src_rc->x1, src_rc->y1, 1}},
+        .dstSubresource = vk_layers,
+        .dstOffsets = {{dst_rc->x0, dst_rc->y0, 0}, {dst_rc->x1, dst_rc->y1, 1}},
+    };
+
+    vkCmdBlitImage(cmd->buf, src_vk->img, src_vk->current_layout, dst_vk->img,
+                   dst_vk->current_layout, 1, &region, VK_FILTER_NEAREST);
+}
+
+static void vk_clear(struct ra *ra, struct ra_tex *tex, float color[4],
+                     struct mp_rect *rc)
+{
+    struct ra_vk *p = ra->priv;
+    struct ra_tex_vk *tex_vk = tex->priv;
+    assert(tex->params.blit_dst);
+
+    struct vk_cmd *cmd = vk_require_cmd(ra);
+    if (!cmd)
+        return;
+
+    struct mp_rect full = {0, 0, tex->params.w, tex->params.h};
+    if (!rc || mp_rect_equals(rc, &full)) {
+        // To clear the entire image, we can use the efficient clear command
+        tex_barrier(cmd, tex_vk, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                    VK_ACCESS_TRANSFER_WRITE_BIT,
+                    VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, true);
+
+        VkClearColorValue clearColor = {0};
+        for (int c = 0; c < 4; c++)
+            clearColor.float32[c] = color[c];
+
+        vkCmdClearColorImage(cmd->buf, tex_vk->img, tex_vk->current_layout,
+                             &clearColor, 1, &vk_range);
+    } else {
+        // To simulate per-region clearing, we blit from a 1x1 texture instead
+        struct ra_tex_upload_params ul_params = {
+            .tex = p->clear_tex,
+            .invalidate = true,
+            .src = &color[0],
+        };
+        vk_tex_upload(ra, &ul_params);
+        vk_blit(ra, tex, p->clear_tex, rc, &(struct mp_rect){0, 0, 1, 1});
+    }
+}
+
+#define VK_QUERY_POOL_SIZE (MPVK_MAX_STREAMING_DEPTH * 4)
+
+struct vk_timer {
+    VkQueryPool pool;
+    int index;
+    uint64_t result;
+};
+
+static void vk_timer_destroy(struct ra *ra, ra_timer *ratimer)
+{
+    if (!ratimer)
+        return;
+
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+    struct vk_timer *timer = ratimer;
+
+    vkDestroyQueryPool(vk->dev, timer->pool, MPVK_ALLOCATOR);
+
+    talloc_free(timer);
+}
+
+MAKE_LAZY_DESTRUCTOR(vk_timer_destroy, ra_timer);
+
+static ra_timer *vk_timer_create(struct ra *ra)
+{
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+
+    struct vk_timer *timer = talloc_zero(NULL, struct vk_timer);
+
+    struct VkQueryPoolCreateInfo qinfo = {
+        .sType = VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO,
+        .queryType = VK_QUERY_TYPE_TIMESTAMP,
+        .queryCount = VK_QUERY_POOL_SIZE,
+    };
+
+    VK(vkCreateQueryPool(vk->dev, &qinfo, MPVK_ALLOCATOR, &timer->pool));
+
+    return (ra_timer *)timer;
+
+error:
+    vk_timer_destroy(ra, timer);
+    return NULL;
+}
+
+static void vk_timer_record(struct ra *ra, VkQueryPool pool, int index,
+                            VkPipelineStageFlags stage)
+{
+    struct vk_cmd *cmd = vk_require_cmd(ra);
+    if (!cmd)
+        return;
+
+    vkCmdWriteTimestamp(cmd->buf, stage, pool, index);
+}
+
+static void vk_timer_start(struct ra *ra, ra_timer *ratimer)
+{
+    struct mpvk_ctx *vk = ra_vk_get(ra);
+    struct vk_timer *timer = ratimer;
+
+    timer->index = (timer->index + 2) % VK_QUERY_POOL_SIZE;
+
+    uint64_t out[2];
+    VkResult res = vkGetQueryPoolResults(vk->dev, timer->pool, timer->index, 2,
+                                         sizeof(out), &out[0], sizeof(uint64_t),
+                                         VK_QUERY_RESULT_64_BIT);
+    switch (res) {
+    case VK_SUCCESS:
+        timer->result = (out[1] - out[0]) * vk->limits.timestampPeriod;
+        break;
+    case VK_NOT_READY:
+        timer->result = 0;
+        break;
+    default:
+        MP_WARN(vk, "Failed reading timer query result: %s\n", vk_err(res));
+        return;
+    };
+
+    vk_timer_record(ra, timer->pool, timer->index,
+                    VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT);
+}
+
+static uint64_t vk_timer_stop(struct ra *ra, ra_timer *ratimer)
+{
+    struct vk_timer *timer = ratimer;
+    vk_timer_record(ra, timer->pool, timer->index + 1,
+                    VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT);
+
+    return timer->result;
+}
+
+static struct ra_fns ra_fns_vk = {
+    .destroy                = vk_destroy_ra,
+    .tex_create             = vk_tex_create,
+    .tex_destroy            = vk_tex_destroy_lazy,
+    .tex_upload             = vk_tex_upload,
+    .buf_create             = vk_buf_create,
+    .buf_destroy            = vk_buf_destroy_lazy,
+    .buf_update             = vk_buf_update,
+    .buf_poll               = vk_buf_poll,
+    .clear                  = vk_clear,
+    .blit                   = vk_blit,
+    .uniform_layout         = std140_layout,
+    .renderpass_create      = vk_renderpass_create,
+    .renderpass_destroy     = vk_renderpass_destroy_lazy,
+    .renderpass_run         = vk_renderpass_run,
+    .timer_create           = vk_timer_create,
+    .timer_destroy          = vk_timer_destroy_lazy,
+    .timer_start            = vk_timer_start,
+    .timer_stop             = vk_timer_stop,
+};
+
+static void present_cb(void *priv, int *inflight)
+{
+    *inflight -= 1;
+}
+
+bool ra_vk_submit(struct ra *ra, struct ra_tex *tex, VkSemaphore acquired,
+                  VkSemaphore *done, int *inflight)
+{
+    struct vk_cmd *cmd = vk_require_cmd(ra);
+    if (!cmd)
+        goto error;
+
+    if (inflight) {
+        *inflight += 1;
+        vk_cmd_callback(cmd, (vk_cb)present_cb, NULL, inflight);
+    }
+
+    struct ra_tex_vk *tex_vk = tex->priv;
+    assert(tex_vk->external_img);
+    tex_barrier(cmd, tex_vk, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, 0,
+                VK_IMAGE_LAYOUT_PRESENT_SRC_KHR, false);
+
+    // These are the only two stages that we use/support for actually
+    // outputting to swapchain imagechain images, so just add a dependency
+    // on both of them. In theory, we could maybe come up with some more
+    // advanced mechanism of tracking dynamic dependencies, but that seems
+    // like overkill.
+    vk_cmd_dep(cmd, acquired,
+               VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT |
+               VK_PIPELINE_STAGE_TRANSFER_BIT);
+
+    return vk_flush(ra, done);
+
+error:
+    return false;
+}
diff --git a/video/out/vulkan/ra_vk.h b/video/out/vulkan/ra_vk.h
new file mode 100644
index 0000000000..893421bc59
--- /dev/null
+++ b/video/out/vulkan/ra_vk.h
@@ -0,0 +1,31 @@
+#pragma once
+
+#include "video/out/gpu/ra.h"
+
+#include "common.h"
+#include "utils.h"
+
+struct ra *ra_create_vk(struct mpvk_ctx *vk, struct mp_log *log);
+
+// Access to the VkDevice is needed for swapchain creation
+VkDevice ra_vk_get_dev(struct ra *ra);
+
+// Allocates a ra_tex that wraps a swapchain image. The contents of the image
+// will be invalidated, and access to it will only be internally synchronized.
+// So the calling could should not do anything else with the VkImage.
+struct ra_tex *ra_vk_wrap_swapchain_img(struct ra *ra, VkImage vkimg,
+                                        VkSwapchainCreateInfoKHR info);
+
+// This function flushes the command buffers, transitions `tex` (which must be
+// a wrapped swapchain image) into a format suitable for presentation, and
+// submits the current rendering commands. The indicated semaphore must fire
+// before the submitted command can run. If `done` is non-NULL, it will be
+// set to a semaphore that fires once the command completes. If `inflight`
+// is non-NULL, it will be incremented when the command starts and decremented
+// when it completes.
+bool ra_vk_submit(struct ra *ra, struct ra_tex *tex, VkSemaphore acquired,
+                  VkSemaphore *done, int *inflight);
+
+// May be called on a struct ra of any type. Returns NULL if the ra is not
+// a vulkan ra.
+struct mpvk_ctx *ra_vk_get(struct ra *ra);
diff --git a/video/out/vulkan/utils.c b/video/out/vulkan/utils.c
new file mode 100644
index 0000000000..43e446bc36
--- /dev/null
+++ b/video/out/vulkan/utils.c
@@ -0,0 +1,726 @@
+#include <libavutil/macros.h>
+
+#include "utils.h"
+#include "malloc.h"
+
+const char* vk_err(VkResult res)
+{
+    switch (res) {
+    // These are technically success codes, but include them nonetheless
+    case VK_SUCCESS:     return "VK_SUCCESS";
+    case VK_NOT_READY:   return "VK_NOT_READY";
+    case VK_TIMEOUT:     return "VK_TIMEOUT";
+    case VK_EVENT_SET:   return "VK_EVENT_SET";
+    case VK_EVENT_RESET: return "VK_EVENT_RESET";
+    case VK_INCOMPLETE:  return "VK_INCOMPLETE";
+    case VK_SUBOPTIMAL_KHR: return "VK_SUBOPTIMAL_KHR";
+
+    // Actual error codes
+    case VK_ERROR_OUT_OF_HOST_MEMORY:    return "VK_ERROR_OUT_OF_HOST_MEMORY";
+    case VK_ERROR_OUT_OF_DEVICE_MEMORY:  return "VK_ERROR_OUT_OF_DEVICE_MEMORY";
+    case VK_ERROR_INITIALIZATION_FAILED: return "VK_ERROR_INITIALIZATION_FAILED";
+    case VK_ERROR_DEVICE_LOST:           return "VK_ERROR_DEVICE_LOST";
+    case VK_ERROR_MEMORY_MAP_FAILED:     return "VK_ERROR_MEMORY_MAP_FAILED";
+    case VK_ERROR_LAYER_NOT_PRESENT:     return "VK_ERROR_LAYER_NOT_PRESENT";
+    case VK_ERROR_EXTENSION_NOT_PRESENT: return "VK_ERROR_EXTENSION_NOT_PRESENT";
+    case VK_ERROR_FEATURE_NOT_PRESENT:   return "VK_ERROR_FEATURE_NOT_PRESENT";
+    case VK_ERROR_INCOMPATIBLE_DRIVER:   return "VK_ERROR_INCOMPATIBLE_DRIVER";
+    case VK_ERROR_TOO_MANY_OBJECTS:      return "VK_ERROR_TOO_MANY_OBJECTS";
+    case VK_ERROR_FORMAT_NOT_SUPPORTED:  return "VK_ERROR_FORMAT_NOT_SUPPORTED";
+    case VK_ERROR_FRAGMENTED_POOL:       return "VK_ERROR_FRAGMENTED_POOL";
+    case VK_ERROR_INVALID_SHADER_NV:     return "VK_ERROR_INVALID_SHADER_NV";
+    case VK_ERROR_OUT_OF_DATE_KHR:       return "VK_ERROR_OUT_OF_DATE_KHR";
+    case VK_ERROR_SURFACE_LOST_KHR:      return "VK_ERROR_SURFACE_LOST_KHR";
+    }
+
+    return "Unknown error!";
+}
+
+static const char* vk_dbg_type(VkDebugReportObjectTypeEXT type)
+{
+    switch (type) {
+    case VK_DEBUG_REPORT_OBJECT_TYPE_INSTANCE_EXT:
+        return "VkInstance";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_PHYSICAL_DEVICE_EXT:
+        return "VkPhysicalDevice";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT:
+        return "VkDevice";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_QUEUE_EXT:
+        return "VkQueue";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_SEMAPHORE_EXT:
+        return "VkSemaphore";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_COMMAND_BUFFER_EXT:
+        return "VkCommandBuffer";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_FENCE_EXT:
+        return "VkFence";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_MEMORY_EXT:
+        return "VkDeviceMemory";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_BUFFER_EXT:
+        return "VkBuffer";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_IMAGE_EXT:
+        return "VkImage";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_EVENT_EXT:
+        return "VkEvent";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_QUERY_POOL_EXT:
+        return "VkQueryPool";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_BUFFER_VIEW_EXT:
+        return "VkBufferView";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_IMAGE_VIEW_EXT:
+        return "VkImageView";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_SHADER_MODULE_EXT:
+        return "VkShaderModule";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_PIPELINE_CACHE_EXT:
+        return "VkPipelineCache";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_PIPELINE_LAYOUT_EXT:
+        return "VkPipelineLayout";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_RENDER_PASS_EXT:
+        return "VkRenderPass";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_PIPELINE_EXT:
+        return "VkPipeline";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_DESCRIPTOR_SET_LAYOUT_EXT:
+        return "VkDescriptorSetLayout";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_SAMPLER_EXT:
+        return "VkSampler";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_DESCRIPTOR_POOL_EXT:
+        return "VkDescriptorPool";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_DESCRIPTOR_SET_EXT:
+        return "VkDescriptorSet";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_FRAMEBUFFER_EXT:
+        return "VkFramebuffer";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_COMMAND_POOL_EXT:
+        return "VkCommandPool";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_SURFACE_KHR_EXT:
+        return "VkSurfaceKHR";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_SWAPCHAIN_KHR_EXT:
+        return "VkSwapchainKHR";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_DEBUG_REPORT_EXT:
+        return "VkDebugReportCallbackEXT";
+    case VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT:
+    default:
+        return "unknown object";
+    }
+}
+
+static VkBool32 vk_dbg_callback(VkDebugReportFlagsEXT flags,
+                                VkDebugReportObjectTypeEXT objType,
+                                uint64_t obj, size_t loc, int32_t msgCode,
+                                const char *layer, const char *msg, void *priv)
+{
+    struct mpvk_ctx *vk = priv;
+    int lev = MSGL_V;
+
+    switch (flags) {
+    case VK_DEBUG_REPORT_ERROR_BIT_EXT:               lev = MSGL_ERR;   break;
+    case VK_DEBUG_REPORT_WARNING_BIT_EXT:             lev = MSGL_WARN;  break;
+    case VK_DEBUG_REPORT_INFORMATION_BIT_EXT:         lev = MSGL_TRACE; break;
+    case VK_DEBUG_REPORT_PERFORMANCE_WARNING_BIT_EXT: lev = MSGL_WARN;  break;
+    case VK_DEBUG_REPORT_DEBUG_BIT_EXT:               lev = MSGL_DEBUG; break;
+    };
+
+    MP_MSG(vk, lev, "vk [%s] %d: %s (obj 0x%llx (%s), loc 0x%zx)\n",
+           layer, (int)msgCode, msg, (unsigned long long)obj,
+           vk_dbg_type(objType), loc);
+
+    // The return value of this function determines whether the call will
+    // be explicitly aborted (to prevent GPU errors) or not. In this case,
+    // we generally want this to be on for the errors.
+    return (flags & VK_DEBUG_REPORT_ERROR_BIT_EXT);
+}
+
+static void vk_cmdpool_uninit(struct mpvk_ctx *vk, struct vk_cmdpool *pool)
+{
+    if (!pool)
+        return;
+
+    // also frees associated command buffers
+    vkDestroyCommandPool(vk->dev, pool->pool, MPVK_ALLOCATOR);
+    for (int n = 0; n < MPVK_MAX_CMDS; n++) {
+        vkDestroyFence(vk->dev, pool->cmds[n].fence, MPVK_ALLOCATOR);
+        vkDestroySemaphore(vk->dev, pool->cmds[n].done, MPVK_ALLOCATOR);
+        talloc_free(pool->cmds[n].callbacks);
+    }
+    talloc_free(pool);
+}
+
+void mpvk_uninit(struct mpvk_ctx *vk)
+{
+    if (!vk->inst)
+        return;
+
+    if (vk->dev) {
+        vk_cmdpool_uninit(vk, vk->pool);
+        vk_malloc_uninit(vk);
+        vkDestroyDevice(vk->dev, MPVK_ALLOCATOR);
+    }
+
+    if (vk->dbg) {
+        // Same deal as creating the debug callback, we need to load this
+        // first.
+        VK_LOAD_PFN(vkDestroyDebugReportCallbackEXT)
+        pfn_vkDestroyDebugReportCallbackEXT(vk->inst, vk->dbg, MPVK_ALLOCATOR);
+    }
+
+    vkDestroySurfaceKHR(vk->inst, vk->surf, MPVK_ALLOCATOR);
+    vkDestroyInstance(vk->inst, MPVK_ALLOCATOR);
+
+    *vk = (struct mpvk_ctx){0};
+}
+
+bool mpvk_instance_init(struct mpvk_ctx *vk, struct mp_log *log, bool debug)
+{
+    *vk = (struct mpvk_ctx) {
+        .log = log,
+    };
+
+    VkInstanceCreateInfo info = {
+        .sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
+    };
+
+    if (debug) {
+        // Enables the LunarG standard validation layer, which
+        // is a meta-layer that loads lots of other validators
+        static const char* layers[] = {
+            "VK_LAYER_LUNARG_standard_validation",
+        };
+
+        info.ppEnabledLayerNames = layers;
+        info.enabledLayerCount = MP_ARRAY_SIZE(layers);
+    }
+
+    // Enable whatever extensions were compiled in.
+    static const char *extensions[] = {
+        VK_KHR_SURFACE_EXTENSION_NAME,
+#if HAVE_X11
+        VK_KHR_XLIB_SURFACE_EXTENSION_NAME,
+#endif
+
+        // Extra extensions only used for debugging. These are toggled by
+        // decreasing the enabledExtensionCount, so the number needs to be
+        // synchronized with the code below.
+        VK_EXT_DEBUG_REPORT_EXTENSION_NAME,
+    };
+
+    const int debugExtensionCount = 1;
+
+    info.ppEnabledExtensionNames = extensions;
+    info.enabledExtensionCount = MP_ARRAY_SIZE(extensions);
+
+    if (!debug)
+        info.enabledExtensionCount -= debugExtensionCount;
+
+    MP_VERBOSE(vk, "Creating instance with extensions:\n");
+    for (int i = 0; i < info.enabledExtensionCount; i++)
+        MP_VERBOSE(vk, "    %s\n", info.ppEnabledExtensionNames[i]);
+
+    VkResult res = vkCreateInstance(&info, MPVK_ALLOCATOR, &vk->inst);
+    if (res != VK_SUCCESS) {
+        MP_VERBOSE(vk, "Failed creating instance: %s\n", vk_err(res));
+        return false;
+    }
+
+    if (debug) {
+        // Set up a debug callback to catch validation messages
+        VkDebugReportCallbackCreateInfoEXT dinfo = {
+            .sType = VK_STRUCTURE_TYPE_DEBUG_REPORT_CALLBACK_CREATE_INFO_EXT,
+            .flags = VK_DEBUG_REPORT_INFORMATION_BIT_EXT |
+                     VK_DEBUG_REPORT_WARNING_BIT_EXT |
+                     VK_DEBUG_REPORT_PERFORMANCE_WARNING_BIT_EXT |
+                     VK_DEBUG_REPORT_ERROR_BIT_EXT |
+                     VK_DEBUG_REPORT_DEBUG_BIT_EXT,
+            .pfnCallback = vk_dbg_callback,
+            .pUserData = vk,
+        };
+
+        // Since this is not part of the core spec, we need to load it. This
+        // can't fail because we've already successfully created an instance
+        // with this extension enabled.
+        VK_LOAD_PFN(vkCreateDebugReportCallbackEXT)
+        pfn_vkCreateDebugReportCallbackEXT(vk->inst, &dinfo, MPVK_ALLOCATOR,
+                                           &vk->dbg);
+    }
+
+    return true;
+}
+
+#define MPVK_MAX_DEVICES 16
+
+static bool physd_supports_surface(struct mpvk_ctx *vk, VkPhysicalDevice physd)
+{
+    uint32_t qfnum;
+    vkGetPhysicalDeviceQueueFamilyProperties(physd, &qfnum, NULL);
+
+    for (int i = 0; i < qfnum; i++) {
+        VkBool32 sup;
+        VK(vkGetPhysicalDeviceSurfaceSupportKHR(physd, i, vk->surf, &sup));
+        if (sup)
+            return true;
+    }
+
+error:
+    return false;
+}
+
+bool mpvk_find_phys_device(struct mpvk_ctx *vk, const char *name, bool sw)
+{
+    assert(vk->surf);
+
+    MP_VERBOSE(vk, "Probing for vulkan devices:\n");
+
+    VkPhysicalDevice *devices = NULL;
+    uint32_t num = 0;
+    VK(vkEnumeratePhysicalDevices(vk->inst, &num, NULL));
+    devices = talloc_array(NULL, VkPhysicalDevice, num);
+    VK(vkEnumeratePhysicalDevices(vk->inst, &num, devices));
+
+    // Sorted by "priority". Reuses some m_opt code for convenience
+    static const struct m_opt_choice_alternatives types[] = {
+        {"discrete",   VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU},
+        {"integrated", VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU},
+        {"virtual",    VK_PHYSICAL_DEVICE_TYPE_VIRTUAL_GPU},
+        {"software",   VK_PHYSICAL_DEVICE_TYPE_CPU},
+        {"unknown",    VK_PHYSICAL_DEVICE_TYPE_OTHER},
+        {0}
+    };
+
+    VkPhysicalDeviceProperties props[MPVK_MAX_DEVICES];
+    for (int i = 0; i < num; i++) {
+        vkGetPhysicalDeviceProperties(devices[i], &props[i]);
+        MP_VERBOSE(vk, "    GPU %d: %s (%s)\n", i, props[i].deviceName,
+                   m_opt_choice_str(types, props[i].deviceType));
+    }
+
+    // Iterate through each type in order of decreasing preference
+    for (int t = 0; types[t].name; t++) {
+        // Disallow SW rendering unless explicitly enabled
+        if (types[t].value == VK_PHYSICAL_DEVICE_TYPE_CPU && !sw)
+            continue;
+
+        for (int i = 0; i < num; i++) {
+            VkPhysicalDeviceProperties prop = props[i];
+            if (prop.deviceType != types[t].value)
+                continue;
+            if (name && strcmp(name, prop.deviceName) != 0)
+                continue;
+            if (!physd_supports_surface(vk, devices[i]))
+                continue;
+
+            MP_VERBOSE(vk, "Chose device:\n");
+            MP_VERBOSE(vk, "    Device Name: %s\n", prop.deviceName);
+            MP_VERBOSE(vk, "    Device ID: %x:%x\n",
+                       (unsigned)prop.vendorID, (unsigned)prop.deviceID);
+            MP_VERBOSE(vk, "    Driver version: %d\n", (int)prop.driverVersion);
+            MP_VERBOSE(vk, "    API version: %d.%d.%d\n",
+                    (int)VK_VERSION_MAJOR(prop.apiVersion),
+                    (int)VK_VERSION_MINOR(prop.apiVersion),
+                    (int)VK_VERSION_PATCH(prop.apiVersion));
+            vk->physd = devices[i];
+            vk->limits = prop.limits;
+            talloc_free(devices);
+            return true;
+        }
+    }
+
+error:
+    MP_VERBOSE(vk, "Found no suitable device, giving up.\n");
+    talloc_free(devices);
+    return false;
+}
+
+bool mpvk_pick_surface_format(struct mpvk_ctx *vk)
+{
+    assert(vk->physd);
+
+    VkSurfaceFormatKHR *formats = NULL;
+    int num;
+
+    // Enumerate through the surface formats and find one that we can map to
+    // a ra_format
+    VK(vkGetPhysicalDeviceSurfaceFormatsKHR(vk->physd, vk->surf, &num, NULL));
+    formats = talloc_array(NULL, VkSurfaceFormatKHR, num);
+    VK(vkGetPhysicalDeviceSurfaceFormatsKHR(vk->physd, vk->surf, &num, formats));
+
+    for (int i = 0; i < num; i++) {
+        // A value of VK_FORMAT_UNDEFINED means we can pick anything we want
+        if (formats[i].format == VK_FORMAT_UNDEFINED) {
+            vk->surf_format = (VkSurfaceFormatKHR) {
+                .colorSpace = VK_COLOR_SPACE_SRGB_NONLINEAR_KHR,
+                .format = VK_FORMAT_R16G16B16A16_UNORM,
+            };
+            break;
+        }
+
+        if (formats[i].colorSpace != VK_COLOR_SPACE_SRGB_NONLINEAR_KHR)
+            continue;
+
+        // Format whitelist, since we want only >= 8 bit _UNORM formats
+        switch (formats[i].format) {
+        case VK_FORMAT_R8G8B8_UNORM:
+        case VK_FORMAT_B8G8R8_UNORM:
+        case VK_FORMAT_R8G8B8A8_UNORM:
+        case VK_FORMAT_B8G8R8A8_UNORM:
+        case VK_FORMAT_A8B8G8R8_UNORM_PACK32:
+        case VK_FORMAT_A2R10G10B10_UNORM_PACK32:
+        case VK_FORMAT_A2B10G10R10_UNORM_PACK32:
+        case VK_FORMAT_R16G16B16_UNORM:
+        case VK_FORMAT_R16G16B16A16_UNORM:
+             break; // accept
+        default: continue;
+        }
+
+        vk->surf_format = formats[i];
+        break;
+    }
+
+    talloc_free(formats);
+
+    if (!vk->surf_format.format)
+        goto error;
+
+    return true;
+
+error:
+    MP_ERR(vk, "Failed picking surface format!\n");
+    talloc_free(formats);
+    return false;
+}
+
+static bool vk_cmdpool_init(struct mpvk_ctx *vk, VkDeviceQueueCreateInfo qinfo,
+                            VkQueueFamilyProperties props,
+                            struct vk_cmdpool **out)
+{
+    struct vk_cmdpool *pool = *out = talloc_ptrtype(NULL, pool);
+    *pool = (struct vk_cmdpool) {
+        .qf = qinfo.queueFamilyIndex,
+        .props = props,
+        .qcount = qinfo.queueCount,
+    };
+
+    for (int n = 0; n < pool->qcount; n++)
+        vkGetDeviceQueue(vk->dev, pool->qf, n, &pool->queues[n]);
+
+    VkCommandPoolCreateInfo cinfo = {
+        .sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
+        .flags = VK_COMMAND_POOL_CREATE_TRANSIENT_BIT |
+                 VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT,
+        .queueFamilyIndex = pool->qf,
+    };
+
+    VK(vkCreateCommandPool(vk->dev, &cinfo, MPVK_ALLOCATOR, &pool->pool));
+
+    VkCommandBufferAllocateInfo ainfo = {
+        .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
+        .commandPool = pool->pool,
+        .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
+        .commandBufferCount = MPVK_MAX_CMDS,
+    };
+
+    VkCommandBuffer cmdbufs[MPVK_MAX_CMDS];
+    VK(vkAllocateCommandBuffers(vk->dev, &ainfo, cmdbufs));
+
+    for (int n = 0; n < MPVK_MAX_CMDS; n++) {
+        struct vk_cmd *cmd = &pool->cmds[n];
+        cmd->pool = pool;
+        cmd->buf = cmdbufs[n];
+
+        VkFenceCreateInfo finfo = {
+            .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
+            .flags = VK_FENCE_CREATE_SIGNALED_BIT,
+        };
+
+        VK(vkCreateFence(vk->dev, &finfo, MPVK_ALLOCATOR, &cmd->fence));
+
+        VkSemaphoreCreateInfo sinfo = {
+            .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
+        };
+
+        VK(vkCreateSemaphore(vk->dev, &sinfo, MPVK_ALLOCATOR, &cmd->done));
+    }
+
+    return true;
+
+error:
+    return false;
+}
+
+bool mpvk_device_init(struct mpvk_ctx *vk, struct mpvk_device_opts opts)
+{
+    assert(vk->physd);
+
+    VkQueueFamilyProperties *qfs = NULL;
+    int qfnum;
+
+    // Enumerate the queue families and find suitable families for each task
+    vkGetPhysicalDeviceQueueFamilyProperties(vk->physd, &qfnum, NULL);
+    qfs = talloc_array(NULL, VkQueueFamilyProperties, qfnum);
+    vkGetPhysicalDeviceQueueFamilyProperties(vk->physd, &qfnum, qfs);
+
+    MP_VERBOSE(vk, "Queue families supported by device:\n");
+
+    for (int i = 0; i < qfnum; i++) {
+        MP_VERBOSE(vk, "QF %d: flags 0x%x num %d\n", i,
+                   (unsigned)qfs[i].queueFlags, (int)qfs[i].queueCount);
+    }
+
+    // For most of our rendering operations, we want to use one "primary" pool,
+    // so just pick the queue family with the most features.
+    int idx = -1;
+    for (int i = 0; i < qfnum; i++) {
+        if (!(qfs[i].queueFlags & VK_QUEUE_GRAPHICS_BIT))
+            continue;
+
+        // QF supports more features
+        if (idx < 0 || qfs[i].queueFlags > qfs[idx].queueFlags)
+            idx = i;
+
+        // QF supports more queues (at the same specialization level)
+        if (qfs[i].queueFlags == qfs[idx].queueFlags &&
+            qfs[i].queueCount > qfs[idx].queueCount)
+        {
+            idx = i;
+        }
+    }
+
+    // Vulkan requires at least one GRAPHICS queue, so if this fails something
+    // is horribly wrong.
+    assert(idx >= 0);
+
+    // Ensure we can actually present to the surface using this queue
+    VkBool32 sup;
+    VK(vkGetPhysicalDeviceSurfaceSupportKHR(vk->physd, idx, vk->surf, &sup));
+    if (!sup) {
+        MP_ERR(vk, "Queue family does not support surface presentation!\n");
+        goto error;
+    }
+
+    // Now that we know which queue families we want, we can create the logical
+    // device
+    assert(opts.queue_count <= MPVK_MAX_QUEUES);
+    static const float priorities[MPVK_MAX_QUEUES] = {0};
+    VkDeviceQueueCreateInfo qinfo = {
+        .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
+        .queueFamilyIndex = idx,
+        .queueCount = MPMIN(qfs[idx].queueCount, opts.queue_count),
+        .pQueuePriorities = priorities,
+    };
+
+    static const char *exts[] = {
+        VK_KHR_SWAPCHAIN_EXTENSION_NAME,
+        VK_NV_GLSL_SHADER_EXTENSION_NAME,
+    };
+
+    VkDeviceCreateInfo dinfo = {
+        .sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
+        .queueCreateInfoCount = 1,
+        .pQueueCreateInfos = &qinfo,
+        .ppEnabledExtensionNames = exts,
+        .enabledExtensionCount = MP_ARRAY_SIZE(exts),
+    };
+
+    MP_VERBOSE(vk, "Creating vulkan device...\n");
+    VK(vkCreateDevice(vk->physd, &dinfo, MPVK_ALLOCATOR, &vk->dev));
+
+    vk_malloc_init(vk);
+
+    // Create the vk_cmdpools and all required queues / synchronization objects
+    if (!vk_cmdpool_init(vk, qinfo, qfs[idx], &vk->pool))
+        goto error;
+
+    talloc_free(qfs);
+    return true;
+
+error:
+    MP_ERR(vk, "Failed creating logical device!\n");
+    talloc_free(qfs);
+    return false;
+}
+
+static void run_callbacks(struct mpvk_ctx *vk, struct vk_cmd *cmd)
+{
+    for (int i = 0; i < cmd->num_callbacks; i++) {
+        struct vk_callback *cb = &cmd->callbacks[i];
+        cb->run(cb->priv, cb->arg);
+        *cb = (struct vk_callback){0};
+    }
+
+    cmd->num_callbacks = 0;
+
+    // Also reset vk->last_cmd in case this was the last command to run
+    if (vk->last_cmd == cmd)
+        vk->last_cmd = NULL;
+}
+
+static void wait_for_cmds(struct mpvk_ctx *vk, struct vk_cmd cmds[], int num)
+{
+    if (!num)
+        return;
+
+    VkFence fences[MPVK_MAX_CMDS];
+    for (int i = 0; i < num; i++)
+        fences[i] = cmds[i].fence;
+
+    vkWaitForFences(vk->dev, num, fences, true, UINT64_MAX);
+
+    for (int i = 0; i < num; i++)
+        run_callbacks(vk, &cmds[i]);
+}
+
+void mpvk_pool_wait_idle(struct mpvk_ctx *vk, struct vk_cmdpool *pool)
+{
+    if (!pool)
+        return;
+
+    int idx = pool->cindex, pidx = pool->cindex_pending;
+    if (pidx < idx) { // range doesn't wrap
+        wait_for_cmds(vk, &pool->cmds[pidx], idx - pidx);
+    } else if (pidx > idx) { // range wraps
+        wait_for_cmds(vk, &pool->cmds[pidx], MPVK_MAX_CMDS - pidx);
+        wait_for_cmds(vk, &pool->cmds[0], idx);
+    }
+    pool->cindex_pending = pool->cindex;
+}
+
+void mpvk_dev_wait_idle(struct mpvk_ctx *vk)
+{
+    mpvk_pool_wait_idle(vk, vk->pool);
+}
+
+void mpvk_pool_poll_cmds(struct mpvk_ctx *vk, struct vk_cmdpool *pool,
+                         uint64_t timeout)
+{
+    if (!pool)
+        return;
+
+    // If requested, hard block until at least one command completes
+    if (timeout > 0 && pool->cindex_pending != pool->cindex) {
+        vkWaitForFences(vk->dev, 1, &pool->cmds[pool->cindex_pending].fence,
+                        true, timeout);
+    }
+
+    // Lazily garbage collect the commands based on their status
+    while (pool->cindex_pending != pool->cindex) {
+        struct vk_cmd *cmd = &pool->cmds[pool->cindex_pending];
+        VkResult res = vkGetFenceStatus(vk->dev, cmd->fence);
+        if (res != VK_SUCCESS)
+            break;
+        run_callbacks(vk, cmd);
+        pool->cindex_pending++;
+        pool->cindex_pending %= MPVK_MAX_CMDS;
+    }
+}
+
+void mpvk_dev_poll_cmds(struct mpvk_ctx *vk, uint32_t timeout)
+{
+    mpvk_pool_poll_cmds(vk, vk->pool, timeout);
+}
+
+void vk_dev_callback(struct mpvk_ctx *vk, vk_cb callback, void *p, void *arg)
+{
+    if (vk->last_cmd) {
+        vk_cmd_callback(vk->last_cmd, callback, p, arg);
+    } else {
+        // The device was already idle, so we can just immediately call it
+        callback(p, arg);
+    }
+}
+
+void vk_cmd_callback(struct vk_cmd *cmd, vk_cb callback, void *p, void *arg)
+{
+    MP_TARRAY_GROW(NULL, cmd->callbacks, cmd->num_callbacks);
+    cmd->callbacks[cmd->num_callbacks++] = (struct vk_callback) {
+        .run  = callback,
+        .priv = p,
+        .arg  = arg,
+    };
+}
+
+void vk_cmd_dep(struct vk_cmd *cmd, VkSemaphore dep,
+                VkPipelineStageFlagBits depstage)
+{
+    assert(cmd->num_deps < MPVK_MAX_CMD_DEPS);
+    cmd->deps[cmd->num_deps] = dep;
+    cmd->depstages[cmd->num_deps++] = depstage;
+}
+
+struct vk_cmd *vk_cmd_begin(struct mpvk_ctx *vk, struct vk_cmdpool *pool)
+{
+    // Garbage collect the cmdpool first
+    mpvk_pool_poll_cmds(vk, pool, 0);
+
+    int next = (pool->cindex + 1) % MPVK_MAX_CMDS;
+    if (next == pool->cindex_pending) {
+        MP_ERR(vk, "No free command buffers!\n");
+        goto error;
+    }
+
+    struct vk_cmd *cmd = &pool->cmds[pool->cindex];
+    pool->cindex = next;
+
+    VK(vkResetCommandBuffer(cmd->buf, 0));
+
+    VkCommandBufferBeginInfo binfo = {
+        .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
+        .flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
+    };
+
+    VK(vkBeginCommandBuffer(cmd->buf, &binfo));
+
+    return cmd;
+
+error:
+    return NULL;
+}
+
+bool vk_cmd_submit(struct mpvk_ctx *vk, struct vk_cmd *cmd, VkSemaphore *done)
+{
+    VK(vkEndCommandBuffer(cmd->buf));
+
+    struct vk_cmdpool *pool = cmd->pool;
+    VkQueue queue = pool->queues[pool->qindex];
+
+    VkSubmitInfo sinfo = {
+        .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
+        .commandBufferCount = 1,
+        .pCommandBuffers = &cmd->buf,
+        .waitSemaphoreCount = cmd->num_deps,
+        .pWaitSemaphores = cmd->deps,
+        .pWaitDstStageMask = cmd->depstages,
+    };
+
+    if (done) {
+        sinfo.signalSemaphoreCount = 1;
+        sinfo.pSignalSemaphores = &cmd->done;
+        *done = cmd->done;
+    }
+
+    VK(vkResetFences(vk->dev, 1, &cmd->fence));
+    VK(vkQueueSubmit(queue, 1, &sinfo, cmd->fence));
+    MP_TRACE(vk, "Submitted command on queue %p (QF %d)\n", (void *)queue,
+             pool->qf);
+
+    for (int i = 0; i < cmd->num_deps; i++)
+        cmd->deps[i] = NULL;
+    cmd->num_deps = 0;
+
+    vk->last_cmd = cmd;
+    return true;
+
+error:
+    return false;
+}
+
+void vk_cmd_cycle_queues(struct mpvk_ctx *vk)
+{
+    struct vk_cmdpool *pool = vk->pool;
+    pool->qindex = (pool->qindex + 1) % pool->qcount;
+}
+
+const VkImageSubresourceRange vk_range = {
+    .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+    .levelCount = 1,
+    .layerCount = 1,
+};
+
+const VkImageSubresourceLayers vk_layers = {
+    .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+    .layerCount = 1,
+};
diff --git a/video/out/vulkan/utils.h b/video/out/vulkan/utils.h
new file mode 100644
index 0000000000..5bde48089d
--- /dev/null
+++ b/video/out/vulkan/utils.h
@@ -0,0 +1,153 @@
+#pragma once
+
+#include "video/out/vo.h"
+#include "video/out/gpu/context.h"
+#include "video/mp_image.h"
+
+#include "common.h"
+#include "formats.h"
+
+#define VK_LOAD_PFN(name) PFN_##name pfn_##name = (PFN_##name) \
+                            vkGetInstanceProcAddr(vk->inst, #name);
+
+// Return a human-readable name for various struct mpvk_ctx enums
+const char* vk_err(VkResult res);
+
+// Convenience macros to simplify a lot of common boilerplate
+#define VK_ASSERT(res, str)                               \
+    do {                                                  \
+        if (res != VK_SUCCESS) {                          \
+            MP_ERR(vk, str ": %s\n", vk_err(res));        \
+            goto error;                                   \
+        }                                                 \
+    } while (0)
+
+#define VK(cmd)                                           \
+    do {                                                  \
+        MP_TRACE(vk, #cmd "\n");                          \
+        VkResult res ## __LINE__ = (cmd);                 \
+        VK_ASSERT(res ## __LINE__, #cmd);                 \
+    } while (0)
+
+// Uninits everything in the correct order
+void mpvk_uninit(struct mpvk_ctx *vk);
+
+// Initialization functions: As a rule of thumb, these need to be called in
+// this order, followed by vk_malloc_init, followed by RA initialization, and
+// finally followed by vk_swchain initialization.
+
+// Create a vulkan instance. Returns VK_NULL_HANDLE on failure
+bool mpvk_instance_init(struct mpvk_ctx *vk, struct mp_log *log, bool debug);
+
+// Generate a VkSurfaceKHR usable for video output. Returns VK_NULL_HANDLE on
+// failure. Must be called after mpvk_instance_init.
+bool mpvk_surface_init(struct vo *vo, struct mpvk_ctx *vk);
+
+// Find a suitable physical device for use with rendering and which supports
+// the surface.
+// name: only match a device with this name
+// sw: also allow software/virtual devices
+bool mpvk_find_phys_device(struct mpvk_ctx *vk, const char *name, bool sw);
+
+// Pick a suitable surface format that's supported by this physical device.
+bool mpvk_pick_surface_format(struct mpvk_ctx *vk);
+
+struct mpvk_device_opts {
+    int queue_count;    // number of queues to use
+};
+
+// Create a logical device and initialize the vk_cmdpools
+bool mpvk_device_init(struct mpvk_ctx *vk, struct mpvk_device_opts opts);
+
+// Wait until all commands submitted to all queues have completed
+void mpvk_pool_wait_idle(struct mpvk_ctx *vk, struct vk_cmdpool *pool);
+void mpvk_dev_wait_idle(struct mpvk_ctx *vk);
+
+// Wait until at least one command submitted to any queue has completed, and
+// process the callbacks. Good for event loops that need to delay until a
+// command completes. Will block at most `timeout` nanoseconds. If used with
+// 0, it only garbage collects completed commands without blocking.
+void mpvk_pool_poll_cmds(struct mpvk_ctx *vk, struct vk_cmdpool *pool,
+                         uint64_t timeout);
+void mpvk_dev_poll_cmds(struct mpvk_ctx *vk, uint32_t timeout);
+
+// Since lots of vulkan operations need to be done lazily once the affected
+// resources are no longer in use, provide an abstraction for tracking these.
+// In practice, these are only checked and run when submitting new commands, so
+// the actual execution may be delayed by a frame.
+typedef void (*vk_cb)(void *priv, void *arg);
+
+struct vk_callback {
+    vk_cb run;
+    void *priv;
+    void *arg; // as a convenience, you also get to pass an arg for "free"
+};
+
+// Associate a callback with the completion of all currently pending commands.
+// This will essentially run once the device is completely idle.
+void vk_dev_callback(struct mpvk_ctx *vk, vk_cb callback, void *p, void *arg);
+
+#define MPVK_MAX_CMD_DEPS 8
+
+// Helper wrapper around command buffers that also track dependencies,
+// callbacks and synchronization primitives
+struct vk_cmd {
+    struct vk_cmdpool *pool; // pool it was allocated from
+    VkCommandBuffer buf;
+    VkFence fence; // the fence guards cmd buffer reuse
+    VkSemaphore done; // the semaphore signals when execution is done
+    // The semaphores represent dependencies that need to complete before
+    // this command can be executed. These are *not* owned by the vk_cmd
+    VkSemaphore deps[MPVK_MAX_CMD_DEPS];
+    VkPipelineStageFlags depstages[MPVK_MAX_CMD_DEPS];
+    int num_deps;
+    // Since VkFences are useless, we have to manually track "callbacks"
+    // to fire once the VkFence completes. These are used for multiple purposes,
+    // ranging from garbage collection (resource deallocation) to fencing.
+    struct vk_callback *callbacks;
+    int num_callbacks;
+};
+
+// Associate a callback with the completion of the current command. This
+// bool will be set to `true` once the command completes, or shortly thereafter.
+void vk_cmd_callback(struct vk_cmd *cmd, vk_cb callback, void *p, void *arg);
+
+// Associate a dependency for the current command. This semaphore must signal
+// by the corresponding stage before the command may execute.
+void vk_cmd_dep(struct vk_cmd *cmd, VkSemaphore dep,
+                VkPipelineStageFlagBits depstage);
+
+#define MPVK_MAX_QUEUES 8
+#define MPVK_MAX_CMDS 64
+
+// Command pool / queue family hybrid abstraction
+struct vk_cmdpool {
+    VkQueueFamilyProperties props;
+    uint32_t qf; // queue family index
+    VkCommandPool pool;
+    VkQueue queues[MPVK_MAX_QUEUES];
+    int qcount;
+    int qindex;
+    // Command buffers associated with this queue
+    struct vk_cmd cmds[MPVK_MAX_CMDS];
+    int cindex;
+    int cindex_pending;
+};
+
+// Fetch the next command buffer from a command pool and begin recording to it.
+// Returns NULL on failure.
+struct vk_cmd *vk_cmd_begin(struct mpvk_ctx *vk, struct vk_cmdpool *pool);
+
+// Finish the currently recording command buffer and submit it for execution.
+// If `done` is not NULL, it will be set to a semaphore that will signal once
+// the command completes. (And MUST have a corresponding semaphore wait)
+// Returns whether successful.
+bool vk_cmd_submit(struct mpvk_ctx *vk, struct vk_cmd *cmd, VkSemaphore *done);
+
+// Rotate the queues for each vk_cmdpool. Call this once per frame to ensure
+// good parallelism between frames when using multiple queues
+void vk_cmd_cycle_queues(struct mpvk_ctx *vk);
+
+// Predefined structs for a simple non-layered, non-mipped image
+extern const VkImageSubresourceRange vk_range;
+extern const VkImageSubresourceLayers vk_layers;
diff --git a/wscript b/wscript
index dd47956392..964b7878c7 100644
--- a/wscript
+++ b/wscript
@@ -803,6 +803,10 @@ video_output_features = [
                 "Aborting. If you really mean to compile without OpenGL " +
                 "video outputs use --disable-gl.",
     }, {
+        'name': '--vulkan',
+        'desc':  'Vulkan context support',
+        'func': check_cc(header_name='vulkan/vulkan.h', lib='vulkan'),
+    }, {
         'name': 'egl-helpers',
         'desc': 'EGL helper functions',
         'deps': 'egl-x11 || mali-fbdev || rpi || gl-wayland || egl-drm || ' +
diff --git a/wscript_build.py b/wscript_build.py
index 68cfafb94f..86b51daaa2 100644
--- a/wscript_build.py
+++ b/wscript_build.py
@@ -445,6 +445,12 @@ def build(ctx):
         ( "video/out/w32_common.c",              "win32-desktop" ),
         ( "video/out/win32/displayconfig.c",     "win32-desktop" ),
         ( "video/out/win32/droptarget.c",        "win32-desktop" ),
+        ( "video/out/vulkan/utils.c",            "vulkan" ),
+        ( "video/out/vulkan/malloc.c",           "vulkan" ),
+        ( "video/out/vulkan/formats.c",          "vulkan" ),
+        ( "video/out/vulkan/ra_vk.c",            "vulkan" ),
+        ( "video/out/vulkan/context.c",          "vulkan" ),
+        ( "video/out/vulkan/context_xlib.c",     "vulkan && x11" ),
         ( "video/out/win32/exclusive_hack.c",    "gl-win32" ),
         ( "video/out/wayland_common.c",          "wayland" ),
         ( "video/out/wayland/buffer.c",          "wayland" ),