[PATCH i-g-t v2 1/4] lib/intel_compute: move user-fence to allow async exec call

[Francois Dugast]

On Thu, May 22, 2025 at 09:55:12AM +0200, Zbigniew Kempczyński wrote:
> User-fence in bo_execenv_exec() limits its usage to synchronous call.
> Remove this limitation by migration of user-fence from function level
> to execenv structure. This allows to synchronize later what is useful
> for parallel submission scenarios.
> 
> Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
> Cc: Francois Dugast <francois.dugast at intel.com>

Reviewed-by: Francois Dugast <francois.dugast at intel.com>

> ---
>  lib/intel_compute.c | 77 ++++++++++++++++++++++++++++++---------------
>  1 file changed, 52 insertions(+), 25 deletions(-)
> 
> diff --git a/lib/intel_compute.c b/lib/intel_compute.c
> index 070bc0d2d8..5579bec85b 100644
> --- a/lib/intel_compute.c
> +++ b/lib/intel_compute.c
> @@ -89,6 +89,12 @@ struct bo_execenv {
>  	uint32_t exec_queue;
>  	uint32_t array_size;
>  
> +	/* Xe user-fence */
> +	uint32_t bo;
> +	size_t bo_size;
> +	struct bo_sync *bo_sync;
> +	struct drm_xe_sync sync;
> +
>  	/* i915 part */
>  	struct drm_i915_gem_execbuffer2 execbuf;
>  	struct drm_i915_gem_exec_object2 *obj;
> @@ -266,48 +272,69 @@ static void bo_execenv_unbind(struct bo_execenv *execenv,
>  	}
>  }
>  
> -static void bo_execenv_exec(struct bo_execenv *execenv, uint64_t start_addr)
> +static void __bo_execenv_exec(struct bo_execenv *execenv, uint64_t start_addr)
>  {
>  	int fd = execenv->fd;
>  
>  	if (execenv->driver == INTEL_DRIVER_XE) {
>  		uint32_t exec_queue = execenv->exec_queue;
> -		struct bo_sync *bo_sync;
> -		size_t bo_size = sizeof(*bo_sync);
> -		uint32_t bo = 0;
> -		struct drm_xe_sync sync = {
> -			.type = DRM_XE_SYNC_TYPE_USER_FENCE,
> -			.flags = DRM_XE_SYNC_FLAG_SIGNAL,
> -			.timeline_value = USER_FENCE_VALUE,
> -		};
> +		size_t bo_size = ALIGN(sizeof(struct bo_sync),
> +				       xe_get_default_alignment(fd));
>  
> -		bo_size = xe_bb_size(fd, bo_size);
> -		bo = xe_bo_create(fd, execenv->vm, bo_size, vram_if_possible(fd, 0),
> -				  DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM);
> -		bo_sync = xe_bo_map(fd, bo, bo_size);
> -		sync.addr = to_user_pointer(&bo_sync->sync);
> -		xe_vm_bind_async(fd, execenv->vm, 0, bo, 0, ADDR_SYNC, bo_size, &sync, 1);
> -		xe_wait_ufence(fd, &bo_sync->sync, USER_FENCE_VALUE, exec_queue, INT64_MAX);
> +		execenv->bo_size = bo_size;
> +		execenv->bo = xe_bo_create(fd, execenv->vm, bo_size, vram_if_possible(fd, 0),
> +					   DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM);
> +		execenv->bo_sync = xe_bo_map(fd, execenv->bo, bo_size);
> +		execenv->sync.type = DRM_XE_SYNC_TYPE_USER_FENCE;
> +		execenv->sync.flags = DRM_XE_SYNC_FLAG_SIGNAL;
> +		execenv->sync.timeline_value = USER_FENCE_VALUE;
> +		execenv->sync.addr = to_user_pointer(&execenv->bo_sync->sync);
> +		xe_vm_bind_async(fd, execenv->vm, 0, execenv->bo, 0, ADDR_SYNC,
> +				 bo_size, &execenv->sync, 1);
> +		xe_wait_ufence(fd, &execenv->bo_sync->sync, USER_FENCE_VALUE,
> +			       exec_queue, INT64_MAX);
>  
> -		sync.addr = ADDR_SYNC;
> -		bo_sync->sync = 0;
> +		execenv->sync.addr = ADDR_SYNC;
> +		execenv->bo_sync->sync = 0;
>  
> -		xe_exec_sync(fd, exec_queue, start_addr, &sync, 1);
> -		xe_wait_ufence(fd, &bo_sync->sync, USER_FENCE_VALUE, exec_queue, INT64_MAX);
> -
> -		munmap(bo_sync, bo_size);
> -		gem_close(fd, bo);
> +		xe_exec_sync(fd, exec_queue, start_addr, &execenv->sync, 1);
>  	} else {
>  		struct drm_i915_gem_execbuffer2 *execbuf = &execenv->execbuf;
> -		struct drm_i915_gem_exec_object2 *obj = execenv->obj;
> -		int num_objects = execbuf->buffer_count;
>  
>  		execbuf->flags = I915_EXEC_RENDER;
>  		gem_execbuf(fd, execbuf);
> +	}
> +}
> +
> +static void bo_execenv_sync(struct bo_execenv *execenv)
> +{
> +	int fd = execenv->fd;
> +
> +	if (execenv->driver == INTEL_DRIVER_XE) {
> +		xe_wait_ufence(fd, &execenv->bo_sync->sync,
> +			       USER_FENCE_VALUE, execenv->exec_queue, INT64_MAX);
> +		munmap(execenv->bo_sync, execenv->bo_size);
> +		gem_close(fd, execenv->bo);
> +	} else {
> +		struct drm_i915_gem_execbuffer2 *execbuf = &execenv->execbuf;
> +		struct drm_i915_gem_exec_object2 *obj = execenv->obj;
> +		int num_objects = execbuf->buffer_count;
> +
>  		gem_sync(fd, obj[num_objects - 1].handle); /* batch handle */
>  	}
>  }
>  
> +static void bo_execenv_exec_async(struct bo_execenv *execenv, uint64_t start_addr)
> +{
> +	__bo_execenv_exec(execenv, start_addr);
> +}
> +
> +static void bo_execenv_exec(struct bo_execenv *execenv, uint64_t start_addr)
> +{
> +	bo_execenv_exec_async(execenv, start_addr);
> +	bo_execenv_sync(execenv);
> +}
> +
>  static uint32_t size_thread_group_x(uint32_t work_size)
>  {
>  	return MAX(1, work_size / (ENQUEUED_LOCAL_SIZE_X *
> -- 
> 2.43.0
>