vo_gpu: aggressively prefer async compute

On AMD devices, we only get one graphics pipe but several compute pipes
which can (in theory) run independently. As such, we should prefer
compute shaders over fragment shaders in scenarios where we expect them
to be better for parallelism.

This is amusingly trivial to do, and actually improves performance even
in a single-queue scenario.

3 files changed, 12 insertions, 1 deletions

diff --git a/video/out/gpu/ra.h b/video/out/gpu/ra.h
index ffb010960a..08ccdaee70 100644
--- a/video/out/gpu/ra.h
+++ b/video/out/gpu/ra.h
@@ -53,6 +53,7 @@ enum {
     RA_CAP_GLOBAL_UNIFORM = 1 << 8, // supports using "naked" uniforms (not UBO)
     RA_CAP_GATHER         = 1 << 9, // supports textureGather in GLSL
     RA_CAP_FRAGCOORD      = 1 << 10, // supports reading from gl_FragCoord
+    RA_CAP_PARALLEL_COMPUTE  = 1 << 11, // supports parallel compute shaders
 };
 
 enum ra_ctype {
diff --git a/video/out/gpu/video.c b/video/out/gpu/video.c
index 3f0959931d..1b50166dc4 100644
--- a/video/out/gpu/video.c
+++ b/video/out/gpu/video.c
@@ -1237,6 +1237,11 @@ static void finish_pass_tex(struct gl_video *p, struct ra_tex **dst_tex,
         return;
     }
 
+    // If RA_CAP_PARALLEL_COMPUTE is set, try to prefer compute shaders
+    // over fragment shaders wherever possible.
+    if (!p->pass_compute.active && (p->ra->caps & RA_CAP_PARALLEL_COMPUTE))
+        pass_is_compute(p, 16, 16);
+
     if (p->pass_compute.active) {
         gl_sc_uniform_image2D_wo(p->sc, "out_image", *dst_tex);
         if (!p->pass_compute.directly_writes)
diff --git a/video/out/vulkan/ra_vk.c b/video/out/vulkan/ra_vk.c
index 905fc89596..f0353629e6 100644
--- a/video/out/vulkan/ra_vk.c
+++ b/video/out/vulkan/ra_vk.c
@@ -208,8 +208,13 @@ struct ra *ra_create_vk(struct mpvk_ctx *vk, struct mp_log *log)
     ra->max_shmem = vk->limits.maxComputeSharedMemorySize;
     ra->max_pushc_size = vk->limits.maxPushConstantsSize;
 
-    if (vk->pool_compute)
+    if (vk->pool_compute) {
         ra->caps |= RA_CAP_COMPUTE;
+        // If we have more compute queues than graphics queues, we probably
+        // want to be using them. (This seems mostly relevant for AMD)
+        if (vk->pool_compute->num_queues > vk->pool_graphics->num_queues)
+            ra->caps |= RA_CAP_PARALLEL_COMPUTE;
+    }
 
     if (!vk_setup_formats(ra))
         goto error;