[PATCH v9 13/14] drm/amdgpu: enable compute/gfx usermode queue

[Alex Deucher]

On Fri, Apr 26, 2024 at 10:27 AM Shashank Sharma
<shashank.sharma at amd.com> wrote:
>
> From: Arvind Yadav <arvind.yadav at amd.com>
>
> This patch does the necessary changes required to
> enable compute workload support using the existing
> usermode queues infrastructure.
>
> Cc: Alex Deucher <alexander.deucher at amd.com>
> Cc: Christian Koenig <christian.koenig at amd.com>
> Signed-off-by: Arvind Yadav <arvind.yadav at amd.com>
> Signed-off-by: Shashank Sharma <shashank.sharma at amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c    |  3 ++-
>  drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c           |  2 ++
>  drivers/gpu/drm/amd/amdgpu/mes_v11_0_userqueue.c | 10 +++++++++-
>  include/uapi/drm/amdgpu_drm.h                    |  1 +
>  4 files changed, 14 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
> index e516487e8db9..78d34fa7a0b9 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
> @@ -189,7 +189,8 @@ amdgpu_userqueue_create(struct drm_file *filp, union drm_amdgpu_userq *args)
>         int qid, r = 0;
>
>         /* Usermode queues are only supported for GFX/SDMA engines as of now */
> -       if (args->in.ip_type != AMDGPU_HW_IP_GFX && args->in.ip_type != AMDGPU_HW_IP_DMA) {
> +       if (args->in.ip_type != AMDGPU_HW_IP_GFX && args->in.ip_type != AMDGPU_HW_IP_DMA
> +                       && args->in.ip_type != AMDGPU_HW_IP_COMPUTE) {
>                 DRM_ERROR("Usermode queue doesn't support IP type %u\n", args->in.ip_type);
>                 return -EINVAL;
>         }
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
> index 525bd0f4d3f7..27b86f7fe949 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
> @@ -1350,6 +1350,7 @@ static int gfx_v11_0_sw_init(void *handle)
>                 adev->gfx.mec.num_pipe_per_mec = 4;
>                 adev->gfx.mec.num_queue_per_pipe = 4;
>                 adev->userq_funcs[AMDGPU_HW_IP_GFX] = &userq_mes_v11_0_funcs;
> +               adev->userq_funcs[AMDGPU_HW_IP_COMPUTE] = &userq_mes_v11_0_funcs;
>                 break;
>         case IP_VERSION(11, 0, 1):
>         case IP_VERSION(11, 0, 4):
> @@ -1362,6 +1363,7 @@ static int gfx_v11_0_sw_init(void *handle)
>                 adev->gfx.mec.num_pipe_per_mec = 4;
>                 adev->gfx.mec.num_queue_per_pipe = 4;
>                 adev->userq_funcs[AMDGPU_HW_IP_GFX] = &userq_mes_v11_0_funcs;
> +               adev->userq_funcs[AMDGPU_HW_IP_COMPUTE] = &userq_mes_v11_0_funcs;
>                 break;
>         default:
>                 adev->gfx.me.num_me = 1;
> diff --git a/drivers/gpu/drm/amd/amdgpu/mes_v11_0_userqueue.c b/drivers/gpu/drm/amd/amdgpu/mes_v11_0_userqueue.c
> index a5e270eda37b..d61d80f86003 100644
> --- a/drivers/gpu/drm/amd/amdgpu/mes_v11_0_userqueue.c
> +++ b/drivers/gpu/drm/amd/amdgpu/mes_v11_0_userqueue.c
> @@ -183,7 +183,8 @@ static int mes_v11_0_userq_create_ctx_space(struct amdgpu_userq_mgr *uq_mgr,
>         }
>
>         /* We don't need to set other FW objects for SDMA queues */
> -       if (queue->queue_type == AMDGPU_HW_IP_DMA)
> +       if ((queue->queue_type == AMDGPU_HW_IP_DMA) ||
> +           (queue->queue_type == AMDGPU_HW_IP_COMPUTE))
>                 return 0;
>
>         /* Shadow and GDS objects come directly from userspace */
> @@ -246,6 +247,13 @@ static int mes_v11_0_userq_mqd_create(struct amdgpu_userq_mgr *uq_mgr,
>         userq_props->use_doorbell = true;
>         userq_props->doorbell_index = queue->doorbell_index;
>
> +       if (queue->queue_type == AMDGPU_HW_IP_COMPUTE) {
> +               userq_props->eop_gpu_addr = mqd_user->eop_va;
> +               userq_props->hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_NORMAL;
> +               userq_props->hqd_queue_priority = AMDGPU_GFX_QUEUE_PRIORITY_MINIMUM;
> +               userq_props->hqd_active = false;
> +       }
> +
>         queue->userq_prop = userq_props;
>
>         r = mqd_hw_default->init_mqd(adev, (void *)queue->mqd.cpu_ptr, userq_props);
> diff --git a/include/uapi/drm/amdgpu_drm.h b/include/uapi/drm/amdgpu_drm.h
> index 22f56a30f7cb..676792ad3618 100644
> --- a/include/uapi/drm/amdgpu_drm.h
> +++ b/include/uapi/drm/amdgpu_drm.h
> @@ -375,6 +375,7 @@ struct drm_amdgpu_userq_mqd {
>          * sized.
>          */
>         __u64   csa_va;
> +       __u64   eop_va;
>  };

Let's add a new mqd descriptor for compute since it's different from
gfx and sdma.  Also, can we handle the eop buffer as part of the
kernel metadata for compute user queues rather than having the user
specify it?

Alex

>
>  struct drm_amdgpu_userq_in {
> --
> 2.43.2
>