venus: allow tls ring submission to utilize the entire ring shmem

Ring submissions on tls ring are synchronous and single threaded, thus
without perf degradation, a single cmd can use the entire ring shmem.

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27054>

src/virtio/vulkan/vn_common.c

@@ -276,10 +276,13 @@ vn_tls_get_ring(struct vn_instance *instance)
    /* only need a small ring for synchronous cmds on tls ring */
    static const size_t buf_size = 16 * 1024;
 
+   /* single cmd can use the entire ring shmem on tls ring */
+   static const uint8_t direct_order = 0;
+
    struct vn_ring_layout layout;
    vn_ring_get_layout(buf_size, extra_size, &layout);
 
-   tls_ring->ring = vn_ring_create(instance, &layout);
+   tls_ring->ring = vn_ring_create(instance, &layout, direct_order);
    if (!tls_ring->ring) {
       free(tls_ring);
       return NULL;

src/virtio/vulkan/vn_instance.c

@@ -20,8 +20,6 @@
 #include "vn_renderer.h"
 #include "vn_ring.h"
 
-#define VN_INSTANCE_RING_SIZE (128 * 1024)
-
 /*
  * Instance extensions add instance-level or physical-device-level
  * functionalities.  It seems renderer support is either unnecessary or
@@ -130,10 +128,17 @@ vn_instance_init_ring(struct vn_instance *instance)
 {
    /* 32-bit seqno for renderer roundtrips */
    static const size_t extra_size = sizeof(uint32_t);
-   struct vn_ring_layout layout;
-   vn_ring_get_layout(VN_INSTANCE_RING_SIZE, extra_size, &layout);
 
-   instance->ring.ring = vn_ring_create(instance, &layout);
+   /* default instance ring size */
+   static const size_t buf_size = 128 * 1024;
+
+   /* order of 4 for performant async cmd enqueue */
+   static const uint8_t direct_order = 4;
+
+   struct vn_ring_layout layout;
+   vn_ring_get_layout(buf_size, extra_size, &layout);
+
+   instance->ring.ring = vn_ring_create(instance, &layout, direct_order);
    if (!instance->ring.ring)
       return VK_ERROR_OUT_OF_HOST_MEMORY;
 

src/virtio/vulkan/vn_ring.c

@@ -40,6 +40,11 @@ struct vn_ring {
     */
    mtx_t mutex;
 
+   /* size limit for cmd submission via ring shmem, derived from
+    * (buffer_size >> direct_order) upon vn_ring_create
+    */
+   uint32_t direct_size;
+
    /* used for indirect submission of large command (non-VkCommandBuffer) */
    struct vn_cs_encoder upload;
 
@@ -251,7 +256,8 @@ vn_ring_get_layout(size_t buf_size,
 
 struct vn_ring *
 vn_ring_create(struct vn_instance *instance,
-               const struct vn_ring_layout *layout)
+               const struct vn_ring_layout *layout,
+               uint8_t direct_order)
 {
    VN_TRACE_FUNC();
 
@@ -289,6 +295,9 @@ vn_ring_create(struct vn_instance *instance,
 
    mtx_init(&ring->mutex, mtx_plain);
 
+   ring->direct_size = layout->buffer_size >> direct_order;
+   assert(ring->direct_size);
+
    vn_cs_encoder_init(&ring->upload, instance,
                       VN_CS_ENCODER_STORAGE_SHMEM_ARRAY, 1 * 1024 * 1024);
 
@@ -529,7 +538,7 @@ static inline bool
 vn_ring_submission_can_direct(const struct vn_ring *ring,
                               const struct vn_cs_encoder *cs)
 {
-   return vn_cs_encoder_get_len(cs) <= (ring->buffer_size >> 4);
+   return vn_cs_encoder_get_len(cs) <= ring->direct_size;
 }
 
 static struct vn_cs_encoder *

src/virtio/vulkan/vn_ring.h

@@ -46,7 +46,8 @@ vn_ring_get_layout(size_t buf_size,
 
 struct vn_ring *
 vn_ring_create(struct vn_instance *instance,
-               const struct vn_ring_layout *layout);
+               const struct vn_ring_layout *layout,
+               uint8_t direct_order);
 
 void
 vn_ring_destroy(struct vn_ring *ring);