[FFmpeg-devel] [PATCH 1/2] vulkan: synchronize access to execution pool fences

[Philip Langdale]

On Wed, 7 Jun 2023 01:22:25 +0200 (CEST)
Lynne <dev at lynne.ee> wrote:

> From c79aa3ed01033f515cbb21251e83cb5bafdf83d7 Mon Sep 17 00:00:00 2001
> From: Lynne <dev at lynne.ee>
> Date: Wed, 7 Jun 2023 00:24:43 +0200
> Subject: [PATCH 1/2] vulkan: synchronize access to execution pool
> fences
> 
> vkResetFences is specified as being user-synchronized
> (yet vkWaitFences, is not).
> ---
>  libavcodec/vulkan_decode.c |  2 +-
>  libavutil/vulkan.c         | 12 ++++++++++--
>  libavutil/vulkan.h         |  2 ++
>  3 files changed, 13 insertions(+), 3 deletions(-)
> 
> diff --git a/libavcodec/vulkan_decode.c b/libavcodec/vulkan_decode.c
> index 889c67a15f..9e3ebf6770 100644
> --- a/libavcodec/vulkan_decode.c
> +++ b/libavcodec/vulkan_decode.c
> @@ -1106,7 +1106,7 @@ int ff_vk_decode_init(AVCodecContext *avctx)
>  
>      /* Create decode exec context.
>       * 4 async contexts per thread seems like a good number. */
> -    err = ff_vk_exec_pool_init(s, &qf_dec, &ctx->exec_pool,
> 4*avctx->thread_count,
> +    err = ff_vk_exec_pool_init(s, &qf_dec, &ctx->exec_pool, 1,
>                                 nb_q,

Comment is now out of date?

> VK_QUERY_TYPE_RESULT_STATUS_ONLY_KHR, 0,
> session_create.pVideoProfile); if (err < 0)
> diff --git a/libavutil/vulkan.c b/libavutil/vulkan.c
> index bc4466e6c9..4b96c0c200 100644
> --- a/libavutil/vulkan.c
> +++ b/libavutil/vulkan.c
> @@ -241,6 +241,7 @@ void ff_vk_exec_pool_free(FFVulkanContext *s,
> FFVkExecPool *pool) vk->WaitForFences(s->hwctx->act_dev, 1,
> &e->fence, VK_TRUE, UINT64_MAX); vk->DestroyFence(s->hwctx->act_dev,
> e->fence, s->hwctx->alloc); }
> +        pthread_mutex_destroy(&e->lock);
>  
>          ff_vk_exec_discard_deps(s, e);
>  
> @@ -379,12 +380,17 @@ int ff_vk_exec_pool_init(FFVulkanContext *s,
> FFVkQueueFamilyCtx *qf, /* Init contexts */
>      for (int i = 0; i < pool->pool_size; i++) {
>          FFVkExecContext *e = &pool->contexts[i];
> -
> -        /* Fence */
>          VkFenceCreateInfo fence_create = {
>              .sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO,
>              .flags = VK_FENCE_CREATE_SIGNALED_BIT,
>          };
> +
> +        /* Mutex */
> +        err = pthread_mutex_init(&e->lock, NULL);
> +        if (err != 0)
> +            return AVERROR(err);
> +
> +        /* Fence */
>          ret = vk->CreateFence(s->hwctx->act_dev, &fence_create,
> s->hwctx->alloc, &e->fence);
>          if (ret != VK_SUCCESS) {
> @@ -489,8 +495,10 @@ int ff_vk_exec_start(FFVulkanContext *s,
> FFVkExecContext *e) };
>  
>      /* Create the fence and don't wait for it initially */
> +    pthread_mutex_lock(&e->lock);
>      vk->WaitForFences(s->hwctx->act_dev, 1, &e->fence, VK_TRUE,
> UINT64_MAX); vk->ResetFences(s->hwctx->act_dev, 1, &e->fence);
> +    pthread_mutex_unlock(&e->lock);

If WaitForFences doesn't require synchronisation, would it be desirable
to call it outside the mutex locking? 

Otherwise, LGTM.

--phil