This patch does the necessary changes required to
enable compute workload support using the existing
usermode queues infrastructure.
V9: Patch introduced
V10: Add custom IP specific mqd strcuture for compute (Alex)
V11: Rename drm_amdgpu_userq_mqd_compute_gfx_v11 to
drm_amdgpu_userq_mqd_compute_gfx11 (Marek)
Cc: Alex Deucher <alexander.deuc...@amd.com>
Cc: Christian Koenig <christian.koe...@amd.com>
Acked-by: Christian König <christian.koe...@amd.com>
Signed-off-by: Arvind Yadav <arvind.ya...@amd.com>
Signed-off-by: Shashank Sharma <shashank.sha...@amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c | 4 +++-
drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 2 ++
.../gpu/drm/amd/amdgpu/mes_v11_0_userqueue.c | 23 +++++++++++++++++++
include/uapi/drm/amdgpu_drm.h | 10 ++++++++
4 files changed, 38 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
index 2c5747cc492e..5173718c3848 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
@@ -189,7 +189,9 @@ amdgpu_userqueue_create(struct drm_file *filp, union
drm_amdgpu_userq *args)
int qid, r = 0;
/* Usermode queues are only supported for GFX IP 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 e68874fd0ff9..82a8df56240e 100644
--- a/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
@@ -1554,6 +1554,7 @@ 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):
@@ -1567,6 +1568,7 @@ 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 dc5359742774..e70b8e429e9c 100644
--- a/drivers/gpu/drm/amd/amdgpu/mes_v11_0_userqueue.c
+++ b/drivers/gpu/drm/amd/amdgpu/mes_v11_0_userqueue.c
@@ -268,6 +268,29 @@ 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) {
+ struct drm_amdgpu_userq_mqd_compute_gfx11 *compute_mqd;
+
+ if (mqd_user->mqd_size != sizeof(*compute_mqd)) {
+ DRM_ERROR("Invalid compute IP MQD size\n");
+ r = -EINVAL;
+ goto free_mqd;
+ }
+
+ compute_mqd = memdup_user(u64_to_user_ptr(mqd_user->mqd),
mqd_user->mqd_size);
+ if (IS_ERR(compute_mqd)) {
+ DRM_ERROR("Failed to read user MQD\n");
+ r = -ENOMEM;
+ goto free_mqd;
+ }
+
+ userq_props->eop_gpu_addr = compute_mqd->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;
+ kfree(compute_mqd);
+ }
+
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 3ea067242b19..6eac46e0f3fd 100644
--- a/include/uapi/drm/amdgpu_drm.h
+++ b/include/uapi/drm/amdgpu_drm.h
@@ -438,6 +438,16 @@ struct drm_amdgpu_userq_mqd_sdma_gfx11 {
__u64 csa_va;
};
+/* GFX V11 Compute IP specific MQD parameters */
+struct drm_amdgpu_userq_mqd_compute_gfx11 {
+ /**
+ * @eop_va: Virtual address of the GPU memory to hold the EOP buffer.
+ * This must be a from a separate GPU object, and must be at least 1
page
+ * sized.
+ */
+ __u64 eop_va;
+};
+
/* vm ioctl */
#define AMDGPU_VM_OP_RESERVE_VMID 1
#define AMDGPU_VM_OP_UNRESERVE_VMID 2
--
2.45.1
