vo_gpu: vulkan: allow disabling async tf/comp

Async compute in particular seems to cause problems on some drivers, and
even when supprted the benefits are not that massive from the tests I
have seen, so it's probably safe to keep off by default.

Async transfer on the other hand seems to work better and offers a more
substantial improvement, so it's kept on.

4 files changed, 38 insertions, 4 deletions

diff --git a/DOCS/man/options.rst b/DOCS/man/options.rst
index f50f67a203..ed27598b8a 100644
--- a/DOCS/man/options.rst
+++ b/DOCS/man/options.rst
@@ -4298,6 +4298,23 @@ The following video options are currently all specific to ``--vo=gpu`` and
     1), but it can also slow things down on hardware where there's no true
     parallelism between queues. (Default: 1)
 
+``--vulkan-async-transfer``
+    Enables the use of async transfer queues on supported vulkan devices. Using
+    them allows transfer operations like texture uploads and blits to happen
+    concurrently with the actual rendering, thus improving overall throughput
+    and power consumption. Enabled by default, and should be relatively safe.
+
+``--vulkan-async-compute``
+    Enables the use of async compute queues on supported vulkan devices. Using
+    this, in theory, allows out-of-order scheduling of compute shaders with
+    graphics shaders, thus enabling the hardware to do more effective work while
+    waiting for pipeline bubbles and memory operations. Not beneficial on all
+    GPUs. It's worth noting that if async compute is enabled, and the device
+    supports more compute queues than graphics queues (bound by the restrictions
+    set by ``--vulkan-queue-count``), mpv will internally try and prefer the
+    use of compute shaders over fragment shaders wherever possible. Not enabled
+    by default, since it seems to cause issues with some drivers.
+
 ``--d3d11-warp=<yes|no|auto>``
     Use WARP (Windows Advanced Rasterization Platform) with the D3D11 GPU
     backend (default: auto). This is a high performance software renderer. By
diff --git a/video/out/vulkan/context.c b/video/out/vulkan/context.c
index 56bf496bb2..21ad5c30c7 100644
--- a/video/out/vulkan/context.c
+++ b/video/out/vulkan/context.c
@@ -104,9 +104,16 @@ const struct m_sub_options vulkan_conf = {
                    {"immediate",    SWAP_IMMEDIATE})),
         OPT_INTRANGE("vulkan-queue-count", dev_opts.queue_count, 0, 1, 8,
                      OPTDEF_INT(1)),
+        OPT_FLAG("vulkan-async-transfer", dev_opts.async_transfer, 0),
+        OPT_FLAG("vulkan-async-compute", dev_opts.async_compute, 0),
         {0}
     },
-    .size = sizeof(struct vulkan_opts)
+    .size = sizeof(struct vulkan_opts),
+    .defaults = &(struct vulkan_opts) {
+        .dev_opts = {
+            .async_transfer = 1,
+        },
+    },
 };
 
 struct priv {
diff --git a/video/out/vulkan/utils.c b/video/out/vulkan/utils.c
index 3dfb825032..5b9be3216f 100644
--- a/video/out/vulkan/utils.c
+++ b/video/out/vulkan/utils.c
@@ -455,9 +455,12 @@ bool mpvk_device_init(struct mpvk_ctx *vk, struct mpvk_device_opts opts)
                    (unsigned)qfs[i].queueFlags, (int)qfs[i].queueCount);
     }
 
-    int idx_gfx  = find_qf(qfs, qfnum, VK_QUEUE_GRAPHICS_BIT),
-        idx_comp = find_qf(qfs, qfnum, VK_QUEUE_COMPUTE_BIT),
-        idx_tf   = find_qf(qfs, qfnum, VK_QUEUE_TRANSFER_BIT);
+    int idx_gfx = -1, idx_comp = -1, idx_tf = -1;
+    idx_gfx = find_qf(qfs, qfnum, VK_QUEUE_GRAPHICS_BIT);
+    if (opts.async_compute)
+        idx_comp = find_qf(qfs, qfnum, VK_QUEUE_COMPUTE_BIT);
+    if (opts.async_transfer)
+        idx_tf = find_qf(qfs, qfnum, VK_QUEUE_TRANSFER_BIT);
 
     // Vulkan requires at least one GRAPHICS queue, so if this fails something
     // is horribly wrong.
@@ -477,6 +480,11 @@ bool mpvk_device_init(struct mpvk_ctx *vk, struct mpvk_device_opts opts)
     if (idx_comp >= 0 && idx_comp != idx_gfx)
         MP_VERBOSE(vk, "Using async compute (QF %d)\n", idx_comp);
 
+    // Fall back to supporting compute shaders via the graphics pool for
+    // devices which support compute shaders but not async compute.
+    if (idx_comp < 0 && qfs[idx_gfx].queueFlags & VK_QUEUE_COMPUTE_BIT)
+        idx_comp = idx_gfx;
+
     // Now that we know which QFs we want, we can create the logical device
     VkDeviceQueueCreateInfo *qinfos = NULL;
     int num_qinfos = 0;
diff --git a/video/out/vulkan/utils.h b/video/out/vulkan/utils.h
index de3a757be3..2962313257 100644
--- a/video/out/vulkan/utils.h
+++ b/video/out/vulkan/utils.h
@@ -55,6 +55,8 @@ bool mpvk_pick_surface_format(struct mpvk_ctx *vk);
 
 struct mpvk_device_opts {
     int queue_count;    // number of queues to use
+    int async_transfer; // enable async transfer
+    int async_compute;  // enable async compute
 };
 
 // Create a logical device and initialize the vk_cmdpools