diff options
Diffstat (limited to 'drivers/accel/ivpu/vpu_jsm_api.h')
| -rw-r--r-- | drivers/accel/ivpu/vpu_jsm_api.h | 303 |
1 files changed, 261 insertions, 42 deletions
diff --git a/drivers/accel/ivpu/vpu_jsm_api.h b/drivers/accel/ivpu/vpu_jsm_api.h index 33f462b1a25d..7215c144158c 100644 --- a/drivers/accel/ivpu/vpu_jsm_api.h +++ b/drivers/accel/ivpu/vpu_jsm_api.h @@ -22,7 +22,7 @@ /* * Minor version changes when API backward compatibility is preserved. */ -#define VPU_JSM_API_VER_MINOR 16 +#define VPU_JSM_API_VER_MINOR 25 /* * API header changed (field names, documentation, formatting) but API itself has not been changed @@ -36,7 +36,7 @@ /* * Number of Priority Bands for Hardware Scheduling - * Bands: RealTime, Focus, Normal, Idle + * Bands: Idle(0), Normal(1), Focus(2), RealTime(3) */ #define VPU_HWS_NUM_PRIORITY_BANDS 4 @@ -74,6 +74,7 @@ #define VPU_JSM_STATUS_MVNCI_INTERNAL_ERROR 0xCU /* Job status returned when the job was preempted mid-inference */ #define VPU_JSM_STATUS_PREEMPTED_MID_INFERENCE 0xDU +#define VPU_JSM_STATUS_MVNCI_CONTEXT_VIOLATION_HW 0xEU /* * Host <-> VPU IPC channels. @@ -86,18 +87,58 @@ /* * Job flags bit masks. */ -#define VPU_JOB_FLAGS_NULL_SUBMISSION_MASK 0x00000001 -#define VPU_JOB_FLAGS_PRIVATE_DATA_MASK 0xFF000000 +enum { + /* + * Null submission mask. + * When set, batch buffer's commands are not processed but returned as + * successful immediately, except fences and timestamps. + * When cleared, batch buffer's commands are processed normally. + * Used for testing and profiling purposes. + */ + VPU_JOB_FLAGS_NULL_SUBMISSION_MASK = (1 << 0U), + /* + * Inline command mask. + * When set, the object in job queue is an inline command (see struct vpu_inline_cmd below). + * When cleared, the object in job queue is a job (see struct vpu_job_queue_entry below). + */ + VPU_JOB_FLAGS_INLINE_CMD_MASK = (1 << 1U), + /* + * VPU private data mask. + * Reserved for the VPU to store private data about the job (or inline command) + * while being processed. + */ + VPU_JOB_FLAGS_PRIVATE_DATA_MASK = 0xFFFF0000U +}; /* - * Sizes of the reserved areas in jobs, in bytes. + * Job queue flags bit masks. */ -#define VPU_JOB_RESERVED_BYTES 8 +enum { + /* + * No job done notification mask. + * When set, indicates that no job done notification should be sent for any + * job from this queue. When cleared, indicates that job done notification + * should be sent for every job completed from this queue. + */ + VPU_JOB_QUEUE_FLAGS_NO_JOB_DONE_MASK = (1 << 0U), + /* + * Native fence usage mask. + * When set, indicates that job queue uses native fences (as inline commands + * in job queue). Such queues may also use legacy fences (as commands in batch buffers). + * When cleared, indicates the job queue only uses legacy fences. + * NOTE: For queues using native fences, VPU expects that all jobs in the queue + * are immediately followed by an inline command object. This object is expected + * to be a fence signal command in most cases, but can also be a NOP in case the host + * does not need per-job fence signalling. Other inline commands objects can be + * inserted between "job and inline command" pairs. + */ + VPU_JOB_QUEUE_FLAGS_USE_NATIVE_FENCE_MASK = (1 << 1U), -/* - * Sizes of the reserved areas in job queues, in bytes. - */ -#define VPU_JOB_QUEUE_RESERVED_BYTES 52 + /* + * Enable turbo mode for testing NPU performance; not recommended for regular usage. + */ + VPU_JOB_QUEUE_FLAGS_TURBO_MODE = (1 << 2U) +}; /* * Max length (including trailing NULL char) of trace entity name (e.g., the @@ -141,23 +182,112 @@ #define VPU_HWS_INVALID_CMDQ_HANDLE 0ULL /* + * Inline commands types. + */ +/* + * NOP. + * VPU does nothing other than consuming the inline command object. + */ +#define VPU_INLINE_CMD_TYPE_NOP 0x0 +/* + * Fence wait. + * VPU waits for the fence current value to reach monitored value. + * Fence wait operations are executed upon job dispatching. While waiting for + * the fence to be satisfied, VPU blocks fetching of the next objects in the queue. + * Jobs present in the queue prior to the fence wait object may be processed + * concurrently. + */ +#define VPU_INLINE_CMD_TYPE_FENCE_WAIT 0x1 +/* + * Fence signal. + * VPU sets the fence current value to the provided value. If new current value + * is equal to or higher than monitored value, VPU sends fence signalled notification + * to the host. Fence signal operations are executed upon completion of all the jobs + * present in the queue prior to them, and in-order relative to each other in the queue. + * But jobs in-between them may be processed concurrently and may complete out-of-order. + */ +#define VPU_INLINE_CMD_TYPE_FENCE_SIGNAL 0x2 + +/* + * Job scheduling priority bands for both hardware scheduling and OS scheduling. + */ +enum vpu_job_scheduling_priority_band { + VPU_JOB_SCHEDULING_PRIORITY_BAND_IDLE = 0, + VPU_JOB_SCHEDULING_PRIORITY_BAND_NORMAL = 1, + VPU_JOB_SCHEDULING_PRIORITY_BAND_FOCUS = 2, + VPU_JOB_SCHEDULING_PRIORITY_BAND_REALTIME = 3, + VPU_JOB_SCHEDULING_PRIORITY_BAND_COUNT = 4, +}; + +/* * Job format. + * Jobs defines the actual workloads to be executed by a given engine. */ struct vpu_job_queue_entry { - u64 batch_buf_addr; /**< Address of VPU commands batch buffer */ - u32 job_id; /**< Job ID */ - u32 flags; /**< Flags bit field, see VPU_JOB_FLAGS_* above */ - u64 root_page_table_addr; /**< Address of root page table to use for this job */ - u64 root_page_table_update_counter; /**< Page tables update events counter */ - u64 primary_preempt_buf_addr; + /**< Address of VPU commands batch buffer */ + u64 batch_buf_addr; + /**< Job ID */ + u32 job_id; + /**< Flags bit field, see VPU_JOB_FLAGS_* above */ + u32 flags; + /** + * Doorbell ring timestamp taken by KMD from SoC's global system clock, in + * microseconds. NPU can convert this value to its own fixed clock's timebase, + * to match other profiling timestamps. + */ + u64 doorbell_timestamp; + /**< Extra id for job tracking, used only in the firmware perf traces */ + u64 host_tracking_id; /**< Address of the primary preemption buffer to use for this job */ - u32 primary_preempt_buf_size; + u64 primary_preempt_buf_addr; /**< Size of the primary preemption buffer to use for this job */ - u32 secondary_preempt_buf_size; + u32 primary_preempt_buf_size; /**< Size of secondary preemption buffer to use for this job */ - u64 secondary_preempt_buf_addr; + u32 secondary_preempt_buf_size; /**< Address of secondary preemption buffer to use for this job */ - u8 reserved_0[VPU_JOB_RESERVED_BYTES]; + u64 secondary_preempt_buf_addr; + u64 reserved_0; +}; + +/* + * Inline command format. + * Inline commands are the commands executed at scheduler level (typically, + * synchronization directives). Inline command and job objects must be of + * the same size and have flags field at same offset. + */ +struct vpu_inline_cmd { + u64 reserved_0; + /* Inline command type, see VPU_INLINE_CMD_TYPE_* defines. */ + u32 type; + /* Flags bit field, see VPU_JOB_FLAGS_* above. */ + u32 flags; + /* Inline command payload. Depends on inline command type. */ + union { + /* Fence (wait and signal) commands' payload. */ + struct { + /* Fence object handle. */ + u64 fence_handle; + /* User VA of the current fence value. */ + u64 current_value_va; + /* User VA of the monitored fence value (read-only). */ + u64 monitored_value_va; + /* Value to wait for or write in fence location. */ + u64 value; + /* User VA of the log buffer in which to add log entry on completion. */ + u64 log_buffer_va; + } fence; + /* Other commands do not have a payload. */ + /* Payload definition for future inline commands can be inserted here. */ + u64 reserved_1[6]; + } payload; +}; + +/* + * Job queue slots can be populated either with job objects or inline command objects. + */ +union vpu_jobq_slot { + struct vpu_job_queue_entry job; + struct vpu_inline_cmd inline_cmd; }; /* @@ -167,7 +297,21 @@ struct vpu_job_queue_header { u32 engine_idx; u32 head; u32 tail; - u8 reserved_0[VPU_JOB_QUEUE_RESERVED_BYTES]; + u32 flags; + /* Set to 1 to indicate priority_band field is valid */ + u32 priority_band_valid; + /* + * Priority for the work of this job queue, valid only if the HWS is NOT used + * and the `priority_band_valid` is set to 1. It is applied only during + * the VPU_JSM_MSG_REGISTER_DB message processing. + * The device firmware might use the `priority_band` to optimize the power + * management logic, but it will not affect the order of jobs. + * Available priority bands: @see enum vpu_job_scheduling_priority_band + */ + u32 priority_band; + /* Inside realtime band assigns a further priority, limited to 0..31 range */ + u32 realtime_priority_level; + u32 reserved_0[9]; }; /* @@ -175,7 +319,7 @@ struct vpu_job_queue_header { */ struct vpu_job_queue { struct vpu_job_queue_header header; - struct vpu_job_queue_entry job[]; + union vpu_jobq_slot slot[]; }; /** @@ -197,9 +341,7 @@ enum vpu_trace_entity_type { struct vpu_hws_log_buffer_header { /* Written by VPU after adding a log entry. Initialised by host to 0. */ u32 first_free_entry_index; - /* Incremented by VPU every time the VPU overwrites the 0th entry; - * initialised by host to 0. - */ + /* Incremented by VPU every time the VPU writes the 0th entry; initialised by host to 0. */ u32 wraparound_count; /* * This is the number of buffers that can be stored in the log buffer provided by the host. @@ -230,14 +372,80 @@ struct vpu_hws_log_buffer_entry { u64 operation_data[2]; }; +/* Native fence log buffer types. */ +enum vpu_hws_native_fence_log_type { + VPU_HWS_NATIVE_FENCE_LOG_TYPE_WAITS = 1, + VPU_HWS_NATIVE_FENCE_LOG_TYPE_SIGNALS = 2 +}; + +/* HWS native fence log buffer header. */ +struct vpu_hws_native_fence_log_header { + union { + struct { + /* Index of the first free entry in buffer. */ + u32 first_free_entry_idx; + /* Incremented each time NPU wraps around the buffer to write next entry. */ + u32 wraparound_count; + }; + /* Field allowing atomic update of both fields above. */ + u64 atomic_wraparound_and_entry_idx; + }; + /* Log buffer type, see enum vpu_hws_native_fence_log_type. */ + u64 type; + /* Allocated number of entries in the log buffer. */ + u64 entry_nb; + u64 reserved[2]; +}; + +/* Native fence log operation types. */ +enum vpu_hws_native_fence_log_op { + VPU_HWS_NATIVE_FENCE_LOG_OP_SIGNAL_EXECUTED = 0, + VPU_HWS_NATIVE_FENCE_LOG_OP_WAIT_UNBLOCKED = 1 +}; + +/* HWS native fence log entry. */ +struct vpu_hws_native_fence_log_entry { + /* Newly signaled/unblocked fence value. */ + u64 fence_value; + /* Native fence object handle to which this operation belongs. */ + u64 fence_handle; + /* Operation type, see enum vpu_hws_native_fence_log_op. */ + u64 op_type; + u64 reserved_0; + /* + * VPU_HWS_NATIVE_FENCE_LOG_OP_WAIT_UNBLOCKED only: Timestamp at which fence + * wait was started (in NPU SysTime). + */ + u64 fence_wait_start_ts; + u64 reserved_1; + /* Timestamp at which fence operation was completed (in NPU SysTime). */ + u64 fence_end_ts; +}; + +/* Native fence log buffer. */ +struct vpu_hws_native_fence_log_buffer { + struct vpu_hws_native_fence_log_header header; + struct vpu_hws_native_fence_log_entry entry[]; +}; + /* * Host <-> VPU IPC messages types. */ enum vpu_ipc_msg_type { VPU_JSM_MSG_UNKNOWN = 0xFFFFFFFF, + /* IPC Host -> Device, Async commands */ VPU_JSM_MSG_ASYNC_CMD = 0x1100, VPU_JSM_MSG_ENGINE_RESET = VPU_JSM_MSG_ASYNC_CMD, + /** + * Preempt engine. The NPU stops (preempts) all the jobs currently + * executing on the target engine making the engine become idle and ready to + * execute new jobs. + * NOTE: The NPU does not remove unstarted jobs (if any) from job queues of + * the target engine, but it stops processing them (until the queue doorbell + * is rung again); the host is responsible to reset the job queue, either + * after preemption or when resubmitting jobs to the queue. + */ VPU_JSM_MSG_ENGINE_PREEMPT = 0x1101, VPU_JSM_MSG_REGISTER_DB = 0x1102, VPU_JSM_MSG_UNREGISTER_DB = 0x1103, @@ -323,9 +531,10 @@ enum vpu_ipc_msg_type { * NOTE: Please introduce new ASYNC commands before this one. * */ VPU_JSM_MSG_STATE_DUMP = 0x11FF, + /* IPC Host -> Device, General commands */ VPU_JSM_MSG_GENERAL_CMD = 0x1200, - VPU_JSM_MSG_BLOB_DEINIT = VPU_JSM_MSG_GENERAL_CMD, + VPU_JSM_MSG_BLOB_DEINIT_DEPRECATED = VPU_JSM_MSG_GENERAL_CMD, /** * Control dyndbg behavior by executing a dyndbg command; equivalent to * Linux command: `echo '<dyndbg_cmd>' > <debugfs>/dynamic_debug/control`. @@ -335,8 +544,12 @@ enum vpu_ipc_msg_type { * Perform the save procedure for the D0i3 entry */ VPU_JSM_MSG_PWR_D0I3_ENTER = 0x1202, + /* IPC Device -> Host, Job completion */ VPU_JSM_MSG_JOB_DONE = 0x2100, + /* IPC Device -> Host, Fence signalled */ + VPU_JSM_MSG_NATIVE_FENCE_SIGNALLED = 0x2101, + /* IPC Device -> Host, Async command completion */ VPU_JSM_MSG_ASYNC_CMD_DONE = 0x2200, VPU_JSM_MSG_ENGINE_RESET_DONE = VPU_JSM_MSG_ASYNC_CMD_DONE, @@ -422,6 +635,7 @@ enum vpu_ipc_msg_type { * NOTE: Please introduce new ASYNC responses before this one. * */ VPU_JSM_MSG_STATE_DUMP_RSP = 0x22FF, + /* IPC Device -> Host, General command completion */ VPU_JSM_MSG_GENERAL_CMD_DONE = 0x2300, VPU_JSM_MSG_BLOB_DEINIT_DONE = VPU_JSM_MSG_GENERAL_CMD_DONE, @@ -600,11 +814,6 @@ struct vpu_jsm_metric_streamer_update { u64 next_buffer_size; }; -struct vpu_ipc_msg_payload_blob_deinit { - /* 64-bit unique ID for the blob to be de-initialized. */ - u64 blob_id; -}; - struct vpu_ipc_msg_payload_job_done { /* Engine to which the job was submitted. */ u32 engine_idx; @@ -622,6 +831,21 @@ struct vpu_ipc_msg_payload_job_done { u64 cmdq_id; }; +/* + * Notification message upon native fence signalling. + * @see VPU_JSM_MSG_NATIVE_FENCE_SIGNALLED + */ +struct vpu_ipc_msg_payload_native_fence_signalled { + /* Engine ID. */ + u32 engine_idx; + /* Host SSID. */ + u32 host_ssid; + /* CMDQ ID */ + u64 cmdq_id; + /* Fence object handle. */ + u64 fence_handle; +}; + struct vpu_jsm_engine_reset_context { /* Host SSID */ u32 host_ssid; @@ -700,11 +924,6 @@ struct vpu_ipc_msg_payload_get_power_level_count_done { u8 power_limit[16]; }; -struct vpu_ipc_msg_payload_blob_deinit_done { - /* 64-bit unique ID for the blob de-initialized. */ - u64 blob_id; -}; - /* HWS priority band setup request / response */ struct vpu_ipc_msg_payload_hws_priority_band_setup { /* @@ -794,7 +1013,10 @@ struct vpu_ipc_msg_payload_hws_set_context_sched_properties { u32 reserved_0; /* Command queue id */ u64 cmdq_id; - /* Priority band to assign to work of this context */ + /* + * Priority band to assign to work of this context. + * Available priority bands: @see enum vpu_job_scheduling_priority_band + */ u32 priority_band; /* Inside realtime band assigns a further priority */ u32 realtime_priority_level; @@ -869,9 +1091,7 @@ struct vpu_ipc_msg_payload_hws_set_scheduling_log { */ u64 notify_index; /* - * Enable extra events to be output to log for debug of scheduling algorithm. - * Interpreted by VPU as a boolean to enable or disable, expected values are - * 0 and 1. + * Field is now deprecated, will be removed when KMD is updated to support removal */ u32 enable_extra_events; /* Zero Padding */ @@ -1243,10 +1463,10 @@ union vpu_ipc_msg_payload { struct vpu_jsm_metric_streamer_start metric_streamer_start; struct vpu_jsm_metric_streamer_stop metric_streamer_stop; struct vpu_jsm_metric_streamer_update metric_streamer_update; - struct vpu_ipc_msg_payload_blob_deinit blob_deinit; struct vpu_ipc_msg_payload_ssid_release ssid_release; struct vpu_jsm_hws_register_db hws_register_db; struct vpu_ipc_msg_payload_job_done job_done; + struct vpu_ipc_msg_payload_native_fence_signalled native_fence_signalled; struct vpu_ipc_msg_payload_engine_reset_done engine_reset_done; struct vpu_ipc_msg_payload_engine_preempt_done engine_preempt_done; struct vpu_ipc_msg_payload_register_db_done register_db_done; @@ -1254,7 +1474,6 @@ union vpu_ipc_msg_payload { struct vpu_ipc_msg_payload_query_engine_hb_done query_engine_hb_done; struct vpu_ipc_msg_payload_get_power_level_count_done get_power_level_count_done; struct vpu_jsm_metric_streamer_done metric_streamer_done; - struct vpu_ipc_msg_payload_blob_deinit_done blob_deinit_done; struct vpu_ipc_msg_payload_trace_config trace_config; struct vpu_ipc_msg_payload_trace_capability_rsp trace_capability; struct vpu_ipc_msg_payload_trace_get_name trace_get_name; |