venus: add some trace points

Add trace points for

 - vn_AcquireNextImage2KHR and vn_QueuePresentKHR
 - vn_AcquireImageANDROID and vn_QueueSignalReleaseImageANDROID
 - vn_BeginCommandBuffer and vn_EndCommandBuffer
 - vn_*Wait*
 - vn_Queue*
 - vn_instance_wait_roundtrip
 - shmem allocations and cache miss/skip

v2: fix cache miss/skip trace points (Ryan)

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Yiwei Zhang <zzyiwei@chromium.org> (v1)
Reviewed-by: Ryan Neph <ryanneph@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14215>

src/virtio/vulkan/vn_android.c

@@ -537,6 +537,7 @@ vn_AcquireImageANDROID(VkDevice device,
                        VkSemaphore semaphore,
                        VkFence fence)
 {
+   VN_TRACE_FUNC();
    struct vn_device *dev = vn_device_from_handle(device);
    VkResult result = VK_SUCCESS;
 
@@ -626,6 +627,7 @@ vn_QueueSignalReleaseImageANDROID(VkQueue queue,
                                   VkImage image,
                                   int *pNativeFenceFd)
 {
+   VN_TRACE_FUNC();
    struct vn_queue *que = vn_queue_from_handle(queue);
    struct vn_device *dev = que->device;
    const VkAllocationCallbacks *alloc = &dev->base.base.alloc;

src/virtio/vulkan/vn_command_buffer.c

@@ -623,6 +623,7 @@ VkResult
 vn_BeginCommandBuffer(VkCommandBuffer commandBuffer,
                       const VkCommandBufferBeginInfo *pBeginInfo)
 {
+   VN_TRACE_FUNC();
    struct vn_command_buffer *cmd =
       vn_command_buffer_from_handle(commandBuffer);
    struct vn_instance *instance = cmd->device->instance;
@@ -703,6 +704,7 @@ vn_cmd_submit(struct vn_command_buffer *cmd)
 VkResult
 vn_EndCommandBuffer(VkCommandBuffer commandBuffer)
 {
+   VN_TRACE_FUNC();
    struct vn_command_buffer *cmd =
       vn_command_buffer_from_handle(commandBuffer);
    struct vn_instance *instance = cmd->device->instance;

src/virtio/vulkan/vn_cs.c

@@ -184,6 +184,7 @@ vn_cs_encoder_grow_buffer_array(struct vn_cs_encoder *enc)
 bool
 vn_cs_encoder_reserve_internal(struct vn_cs_encoder *enc, size_t size)
 {
+   VN_TRACE_FUNC();
    if (unlikely(enc->storage_type == VN_CS_ENCODER_STORAGE_POINTER))
       return false;
 

src/virtio/vulkan/vn_device.c

@@ -420,6 +420,7 @@ vn_GetDeviceGroupPeerMemoryFeatures(
 VkResult
 vn_DeviceWaitIdle(VkDevice device)
 {
+   VN_TRACE_FUNC();
    struct vn_device *dev = vn_device_from_handle(device);
 
    for (uint32_t i = 0; i < dev->queue_count; i++) {

src/virtio/vulkan/vn_instance.c

@@ -332,6 +332,7 @@ void
 vn_instance_wait_roundtrip(struct vn_instance *instance,
                            uint32_t roundtrip_seqno)
 {
+   VN_TRACE_FUNC();
    const struct vn_ring *ring = &instance->ring.ring;
    const volatile atomic_uint *ptr = ring->shared.extra;
    uint32_t iter = 0;
@@ -478,6 +479,7 @@ static struct vn_cs_encoder *
 vn_instance_ring_cs_upload_locked(struct vn_instance *instance,
                                   const struct vn_cs_encoder *cs)
 {
+   VN_TRACE_FUNC();
    assert(cs->storage_type == VN_CS_ENCODER_STORAGE_POINTER &&
           cs->buffer_count == 1);
    const void *cs_data = cs->buffers[0].base;
@@ -554,6 +556,7 @@ vn_instance_get_reply_shmem_locked(struct vn_instance *instance,
                                    size_t size,
                                    void **out_ptr)
 {
+   VN_TRACE_FUNC();
    struct vn_renderer_shmem_pool *pool = &instance->reply_shmem_pool;
    const struct vn_renderer_shmem *saved_pool_shmem = pool->shmem;
 

src/virtio/vulkan/vn_queue.c

@@ -337,6 +337,7 @@ vn_QueueSubmit(VkQueue _queue,
                const VkSubmitInfo *pSubmits,
                VkFence fence)
 {
+   VN_TRACE_FUNC();
    struct vn_queue *queue = vn_queue_from_handle(_queue);
    struct vn_device *dev = queue->device;
 
@@ -396,6 +397,7 @@ vn_QueueBindSparse(VkQueue _queue,
                    const VkBindSparseInfo *pBindInfo,
                    VkFence fence)
 {
+   VN_TRACE_FUNC();
    struct vn_queue *queue = vn_queue_from_handle(_queue);
    struct vn_device *dev = queue->device;
 
@@ -421,6 +423,7 @@ vn_QueueBindSparse(VkQueue _queue,
 VkResult
 vn_QueueWaitIdle(VkQueue _queue)
 {
+   VN_TRACE_FUNC();
    struct vn_queue *queue = vn_queue_from_handle(_queue);
    VkDevice device = vn_device_to_handle(queue->device);
 
@@ -632,6 +635,7 @@ vn_WaitForFences(VkDevice device,
                  VkBool32 waitAll,
                  uint64_t timeout)
 {
+   VN_TRACE_FUNC();
    struct vn_device *dev = vn_device_from_handle(device);
    const VkAllocationCallbacks *alloc = &dev->base.base.alloc;
 
@@ -923,6 +927,7 @@ vn_WaitSemaphores(VkDevice device,
                   const VkSemaphoreWaitInfo *pWaitInfo,
                   uint64_t timeout)
 {
+   VN_TRACE_FUNC();
    struct vn_device *dev = vn_device_from_handle(device);
    const VkAllocationCallbacks *alloc = &dev->base.base.alloc;
 

src/virtio/vulkan/vn_renderer.h

@@ -277,6 +277,7 @@ vn_renderer_wait(struct vn_renderer *renderer,
 static inline struct vn_renderer_shmem *
 vn_renderer_shmem_create(struct vn_renderer *renderer, size_t size)
 {
+   VN_TRACE_FUNC();
    struct vn_renderer_shmem *shmem =
       renderer->shmem_ops.create(renderer, size);
    if (shmem) {

src/virtio/vulkan/vn_renderer_internal.c

@@ -124,6 +124,7 @@ vn_renderer_shmem_cache_get(struct vn_renderer_shmem_cache *cache,
    int idx;
    struct vn_renderer_shmem_bucket *bucket = choose_bucket(cache, size, &idx);
    if (!bucket) {
+      VN_TRACE_SCOPE("shmem cache skip");
       simple_mtx_lock(&cache->mutex);
       cache->debug.cache_skip_count++;
       simple_mtx_unlock(&cache->mutex);
@@ -144,6 +145,7 @@ vn_renderer_shmem_cache_get(struct vn_renderer_shmem_cache *cache,
 
       cache->debug.cache_hit_count++;
    } else {
+      VN_TRACE_SCOPE("shmem cache miss");
       cache->debug.cache_miss_count++;
    }
    simple_mtx_unlock(&cache->mutex);

src/virtio/vulkan/vn_renderer_util.c

@@ -68,6 +68,7 @@ vn_renderer_shmem_pool_grow(struct vn_renderer *renderer,
                             struct vn_renderer_shmem_pool *pool,
                             size_t size)
 {
+   VN_TRACE_FUNC();
    /* power-of-two to hit shmem cache */
    size_t alloc_size = pool->min_alloc_size;
    while (alloc_size < size) {

src/virtio/vulkan/vn_wsi.c

@@ -181,6 +181,7 @@ vn_DestroySwapchainKHR(VkDevice device,
 VkResult
 vn_QueuePresentKHR(VkQueue _queue, const VkPresentInfoKHR *pPresentInfo)
 {
+   VN_TRACE_FUNC();
    struct vn_queue *queue = vn_queue_from_handle(_queue);
 
    VkResult result =
@@ -206,6 +207,7 @@ vn_AcquireNextImage2KHR(VkDevice device,
                         const VkAcquireNextImageInfoKHR *pAcquireInfo,
                         uint32_t *pImageIndex)
 {
+   VN_TRACE_FUNC();
    struct vn_device *dev = vn_device_from_handle(device);
 
    VkResult result = wsi_common_acquire_next_image2(