video: tegra: host: restores sof/eof, TS feature

This commit restores the following,
[1] 890be46 video: tegra: host: dla: refactor and cleanup code
[2] 109f8e8 video: tegra: host: add prefence signal support
[3] 6f683c8 video: tegra: host: add sof/eof taskstatus support
[4] a9cb0c9 video: tegra: host: add sof/eof timestamp support
[5] dcdb0d6 video: tegra: host: increase fw minor version

Jira DLA-1992
Jira DLA-1993

Change-Id: I9844b141b72fc553ba7d886e706e340160586129
Signed-off-by: Arvind M <am@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/2135092
Tested-by: Mitch Harwell <mharwell@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: Prashant Gaikwad <pgaikwad@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>

drivers/video/tegra/host/nvdla/dla_fw_version.h

@@ -20,7 +20,7 @@
 #define _DLA_FW_VERSION_H_
 
 #define FIRMWARE_VERSION_MAJOR		0x1
-#define FIRMWARE_VERSION_MINOR		0x0
+#define FIRMWARE_VERSION_MINOR		0x1
 #define FIRMWARE_VERSION_SUBMINOR	0x0
 
 static inline uint32_t dla_version(void)

drivers/video/tegra/host/nvdla/dla_os_interface.h

@@ -74,18 +74,22 @@
 #define DLA_INT_ON_COMPLETE_SHIFT	8
 #define DLA_INT_ON_ERROR_SHIFT		9
 
-#define PREACTION_TERMINATE	0x0
+/* control actions */
-#define PREACTION_SEM_EQ	0x90
+#define ACTION_TERMINATE	0x0
-#define PREACTION_SEM_GE	0x92
-#define PREACTION_GOS_EQ	0xB0
-#define PREACTION_GOS_GE	0xB2
-#define PREACTION_TASK_STATUS	0xC0
 
-#define POSTACTION_TERMINATE	0x0
+/* conditional actions */
-#define POSTACTION_SEM		0x80
+#define ACTION_SEM_EQ	0x90
-#define POSTACTION_TS_SEM	0x83
+#define ACTION_SEM_GE	0x92
-#define POSTACTION_GOS		0xA0
+#define ACTION_GOS_EQ	0xB0
-#define POSTACTION_TASK_STATUS	0xC1
+#define ACTION_GOS_GE	0xB2
+#define ACTION_TASK_STATUS_EQ	0xC0
+
+/* write actions */
+#define ACTION_WRITE_SEM	0x80
+#define ACTION_WRITE_TS_SEM	0x83
+#define ACTION_WRITE_TIMESTAMP  	0x87
+#define ACTION_WRITE_GOS	0xA0
+#define ACTION_WRITE_TASK_STATUS	0xC1
 
 #define PING_DATA_SIZE		4
 #define BUFFER_MULTIPLIER	4
@@ -228,6 +232,17 @@ struct dla_action_task_status {
 	uint16_t status;
 } __attribute__ ((packed));
 
+/**
+ * Timestamp update action structure
+ *
+ * OPCODE = 0x87
+ *
+ * @address: Address to write timestamp
+ */
+struct dla_action_timestamp {
+	uint64_t address;
+} __attribute__ ((packed));
+
 /**
  * Status notifier structure
  *

drivers/video/tegra/host/nvdla/nvdla.h

@@ -99,8 +99,11 @@
  */
 #define MAX_NUM_NVDLA_PREFENCES		32
 #define MAX_NUM_NVDLA_POSTFENCES	32
+#define MAX_NUM_NVDLA_EMU_PREFENCES	16
+#define MAX_NUM_NVDLA_EMU_POSTFENCES	16
 #define MAX_NUM_NVDLA_IN_TASK_STATUS	MAX_NUM_NVDLA_PREFENCES
 #define MAX_NUM_NVDLA_OUT_TASK_STATUS	MAX_NUM_NVDLA_POSTFENCES
+#define MAX_NUM_NVDLA_OUT_TIMESTAMP  	32
 #define NUM_PROFILING_POSTACTION	1
 
 #define MAX_COMMANDS_PER_DEVICE		1
@@ -219,7 +222,9 @@ struct nvdla_device {
  *
  * @queue		Queue in which task submitted
  * @sp			pointer to syncpt
+ * @prefences		pointer to pre fences
  * @postfences		pointer to post fences
+ * @num_prefences	Number of prefences in task
  * @num_postfences	Number of postfences in task
  * @fence		Fence tracking for current task
  * @fence_counter	Counter used to track fence value
@@ -228,7 +233,9 @@ struct nvdla_device {
 struct nvdla_emu_task {
 	struct nvdla_queue *queue;
 	struct nvhost_syncpt *sp;
-	struct nvdev_fence postfences[MAX_NUM_NVDLA_POSTFENCES];
+	struct nvdev_fence prefences[MAX_NUM_NVDLA_EMU_PREFENCES];
+	struct nvdev_fence postfences[MAX_NUM_NVDLA_EMU_POSTFENCES];
+	u32 num_prefences;
 	u32 num_postfences;
 	u32 fence;
 	u32 fence_counter;
@@ -259,12 +266,18 @@ struct nvdla_task {
 	struct nvdev_fence prefences[MAX_NUM_NVDLA_PREFENCES];
 	struct nvdev_fence postfences[MAX_NUM_NVDLA_POSTFENCES];
 	struct nvdla_status_notify in_task_status[MAX_NUM_NVDLA_IN_TASK_STATUS];
-	struct nvdla_status_notify out_task_status[MAX_NUM_NVDLA_OUT_TASK_STATUS];
+	struct nvdla_status_notify sof_task_status[MAX_NUM_NVDLA_OUT_TASK_STATUS];
+	struct nvdla_status_notify eof_task_status[MAX_NUM_NVDLA_OUT_TASK_STATUS];
+	struct nvdla_mem_handle sof_timestamps[MAX_NUM_NVDLA_OUT_TIMESTAMP];
+	struct nvdla_mem_handle eof_timestamps[MAX_NUM_NVDLA_OUT_TIMESTAMP];
 	struct nvdla_mem_handle memory_handles[NVDLA_MAX_BUFFERS_PER_TASK];
 	u8 num_prefences;
 	u8 num_postfences;
 	u8 num_in_task_status;
-	u8 num_out_task_status;
+	u8 num_sof_task_status;
+	u8 num_eof_task_status;
+	u8 num_sof_timestamps;
+	u8 num_eof_timestamps;
 	u32 num_addresses;
 	u32 fence;
 	u32 fence_counter;
@@ -280,7 +293,10 @@ struct nvdla_task {
 	struct dma_buf *prefences_sem_dmabuf[MAX_NUM_NVDLA_PREFENCES];
 	struct dma_buf *in_task_status_dmabuf[MAX_NUM_NVDLA_IN_TASK_STATUS];
 	struct dma_buf *postfences_sem_dmabuf[MAX_NUM_NVDLA_POSTFENCES];
-	struct dma_buf *out_task_status_dmabuf[MAX_NUM_NVDLA_OUT_TASK_STATUS];
+	struct dma_buf *sof_task_status_dmabuf[MAX_NUM_NVDLA_OUT_TASK_STATUS];
+	struct dma_buf *eof_task_status_dmabuf[MAX_NUM_NVDLA_OUT_TASK_STATUS];
+	struct dma_buf *sof_timestamps_dmabuf[MAX_NUM_NVDLA_OUT_TIMESTAMP];
+	struct dma_buf *eof_timestamps_dmabuf[MAX_NUM_NVDLA_OUT_TIMESTAMP];
 };
 
 struct dla_mem_addr {
@@ -402,7 +418,7 @@ int nvdla_free_gcov_region(struct platform_device *pdev, bool update_region);
 int nvdla_emulator_submit(struct nvdla_queue *queue,
 				struct nvdla_emu_task *task);
 void task_free(struct kref *ref);
-int nvdla_get_postfences(struct nvdla_queue *queue, void *in_task);
+int nvdla_get_signal_fences(struct nvdla_queue *queue, void *in_task);
 int nvdla_send_gos_region(struct platform_device *pdev);
 
 #endif /* End of __NVHOST_NVDLA_H__ */

drivers/video/tegra/host/nvdla/nvdla_ioctl.c

@@ -376,13 +376,41 @@ static int nvdla_get_actions(struct nvdla_ioctl_submit_task *user_task,
 		goto fail;
 	}
 
-	/* get output task status */
+	/* get sof task status */
-	if (copy_from_user(task->out_task_status,
+	if (copy_from_user(task->sof_task_status,
-		(void __user *)user_task->output_task_status,
+		(void __user *)user_task->sof_task_status,
-		(task->num_out_task_status *
+		(task->num_sof_task_status *
 			sizeof(struct nvdla_status_notify)))) {
 		err = -EFAULT;
-		nvdla_dbg_err(pdev, "failed to copy output task status");
+		nvdla_dbg_err(pdev, "failed to copy sof task status");
+		goto fail;
+	}
+
+	/* get eof task status */
+	if (copy_from_user(task->eof_task_status,
+		(void __user *)user_task->eof_task_status,
+		(task->num_eof_task_status *
+			sizeof(struct nvdla_status_notify)))) {
+		err = -EFAULT;
+		nvdla_dbg_err(pdev, "failed to copy eof task status");
+		goto fail;
+	}
+
+	/* get sof timestamps */
+	if (copy_from_user(task->sof_timestamps,
+		(void __user*)user_task->sof_timestamps,
+		(task->num_sof_timestamps * sizeof(struct nvdla_mem_handle)))) {
+		err = -EFAULT;
+		nvdla_dbg_err(pdev, "failed to copy sof timestamps");
+		goto fail;
+	}
+
+	/* get eof timestamps */
+	if (copy_from_user(task->eof_timestamps,
+		(void __user*)user_task->eof_timestamps,
+		(task->num_eof_timestamps * sizeof(struct nvdla_mem_handle)))) {
+		err = -EFAULT;
+		nvdla_dbg_err(pdev, "failed to copy eof timestamps");
 		goto fail;
 	}
 
@@ -392,20 +420,66 @@ fail:
 	return err;
 }
 
-static int nvdla_send_emu_postfences(struct nvdla_emu_task *task,
+static int nvdla_send_emu_signal_fences(struct nvdla_emu_task *task,
 			struct nvdla_ioctl_emu_submit_task *user_task)
 {
 	int err = 0, i;
 	struct platform_device *dla_pdev = task->queue->pool->pdev;
 	struct platform_device *host_pdev =
 				to_platform_device(dla_pdev->dev.parent);
+	struct nvdev_fence __user *prefences =
+		(struct nvdev_fence __user *)(uintptr_t)user_task->prefences;
 	struct nvdev_fence __user *postfences =
 		(struct nvdev_fence __user *)(uintptr_t)user_task->postfences;
 	char fence_name[32];
 
-	nvdla_dbg_fn(dla_pdev, "sending post fences");
+	nvdla_dbg_fn(dla_pdev, "sending signal fences");
+
+	for (i = 0; i < task->num_prefences; i++) {
+		if (task->prefences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
+		if (task->prefences[i].type == NVDEV_FENCE_TYPE_SYNC_FD) {
+			struct nvhost_ctrl_sync_fence_info info;
+
+			info.id = task->prefences[i].syncpoint_index;
+			info.thresh = task->prefences[i].syncpoint_value;
+
+			nvdla_dbg_info(dla_pdev,
+					"creating pre sync fd [%d]:[%d]\n",
+					info.id, info.thresh);
+
+			/* create fence name format example: nvdla0_1_fence */
+			snprintf(fence_name, sizeof(fence_name),
+				"%s_%d_%d_prefence", dev_name(&dla_pdev->dev),
+				task->prefences[i].syncpoint_index, i);
+
+			err = nvhost_sync_create_fence_fd(host_pdev,
+				&info, 1, fence_name,
+				&task->prefences[i].sync_fd);
+
+			if (err) {
+				nvdla_dbg_err(dla_pdev,
+					"fail to create prefence syncfd\n");
+				goto fail;
+			}
+		}
+	}
+
+	nvdla_dbg_fn(dla_pdev, "copy prefences to user");
+	/* send pre fences */
+	if (copy_to_user(prefences, task->prefences,
+		(task->num_prefences * sizeof(struct nvdev_fence)))) {
+		err = -EFAULT;
+		nvdla_dbg_err(dla_pdev, "failed to send prefences");
+		goto fail;
+	}
+	nvdla_dbg_info(dla_pdev, "prefences sent");
 
 	for (i = 0; i < task->num_postfences; i++) {
+		if (task->postfences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
 		if (task->postfences[i].type == NVDEV_FENCE_TYPE_SYNC_FD) {
 			struct nvhost_ctrl_sync_fence_info info;
 
@@ -418,8 +492,8 @@ static int nvdla_send_emu_postfences(struct nvdla_emu_task *task,
 
 			/* create fence name format example: nvdla0_1_fence */
 			snprintf(fence_name, sizeof(fence_name),
-				"%s_%d_fence", dev_name(&dla_pdev->dev),
+				"%s_%d_%d_postfence", dev_name(&dla_pdev->dev),
-				task->postfences[i].syncpoint_index);
+				task->postfences[i].syncpoint_index, i);
 
 			err = nvhost_sync_create_fence_fd(host_pdev,
 				&info, 1, fence_name,
@@ -447,20 +521,67 @@ fail:
 	return err;
 }
 
-static int nvdla_update_postfences(struct nvdla_task *task,
+static int nvdla_update_signal_fences(struct nvdla_task *task,
 			struct nvdla_ioctl_submit_task *user_task)
 {
 	int err = 0, i;
 	struct platform_device *dla_pdev = task->queue->pool->pdev;
 	struct platform_device *host_pdev =
 				to_platform_device(dla_pdev->dev.parent);
+	struct nvdev_fence __user *prefences =
+		(struct nvdev_fence __user *)(uintptr_t)user_task->prefences;
 	struct nvdev_fence __user *postfences =
 		(struct nvdev_fence __user *)(uintptr_t)user_task->postfences;
 	char fence_name[32];
 
-	nvdla_dbg_fn(dla_pdev, "copy post fences for user");
+	nvdla_dbg_fn(dla_pdev, "copy fences for user");
 
+	/* update pre fence signals to users */
+	for (i = 0; i < task->num_prefences; i++) {
+		if (task->prefences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
+		if (task->prefences[i].type == NVDEV_FENCE_TYPE_SYNC_FD) {
+			struct nvhost_ctrl_sync_fence_info info;
+
+			info.id = task->prefences[i].syncpoint_index;
+			info.thresh = task->prefences[i].syncpoint_value;
+
+			nvdla_dbg_info(dla_pdev,
+					"creating pre sync fd [%d]:[%d]\n",
+					info.id, info.thresh);
+
+			/* create fence name format example: nvdla0_1_fence */
+			snprintf(fence_name, sizeof(fence_name),
+				"%s_%d_%d_prefence", dev_name(&dla_pdev->dev),
+				task->prefences[i].syncpoint_index, i);
+
+			err = nvhost_sync_create_fence_fd(host_pdev,
+				&info, 1, fence_name,
+				&task->prefences[i].sync_fd);
+
+			if (err) {
+				nvdla_dbg_err(dla_pdev,
+					"fail to create prefence syncfd\n");
+				goto fail;
+			}
+		}
+	}
+
+	nvdla_dbg_fn(dla_pdev, "copy prefences to user");
+	/* copy pre fences */
+	if (copy_to_user(prefences, task->prefences,
+				(task->num_prefences * sizeof(struct nvdev_fence)))) {
+		err = -EFAULT;
+		nvdla_dbg_err(dla_pdev, "failed to copy prefences");
+		goto fail;
+	}
+
+	/* update post fence signals to user */
 	for (i = 0; i < task->num_postfences; i++) {
+		if (task->postfences[i].action != NVDEV_FENCE_SIGNAL)
+			continue;
+
 		if (task->postfences[i].type == NVDEV_FENCE_TYPE_SYNC_FD) {
 			struct nvhost_ctrl_sync_fence_info info;
 
@@ -473,8 +594,8 @@ static int nvdla_update_postfences(struct nvdla_task *task,
 
 			/* create fence name format example: nvdla0_1_fence */
 			snprintf(fence_name, sizeof(fence_name),
-				"%s_%d_fence", dev_name(&dla_pdev->dev),
+				"%s_%d_%d_postfence", dev_name(&dla_pdev->dev),
-				task->postfences[i].syncpoint_index);
+				task->postfences[i].syncpoint_index, i);
 
 			err = nvhost_sync_create_fence_fd(host_pdev,
 				&info, 1, fence_name,
@@ -525,12 +646,30 @@ static int nvdla_val_task_submit_input(struct nvdla_ioctl_submit_task *in_task)
 			MAX_NUM_NVDLA_IN_TASK_STATUS);
 		return -EINVAL;
 	}
-	if (in_task->num_output_task_status > MAX_NUM_NVDLA_OUT_TASK_STATUS) {
+	if (in_task->num_sof_task_status > MAX_NUM_NVDLA_OUT_TASK_STATUS) {
-		pr_err("out task status[%u] crossing expected[%d]\n",
+		pr_err("sof task status[%u] crossing expected[%d]\n",
-			in_task->num_output_task_status,
+			in_task->num_sof_task_status,
 			MAX_NUM_NVDLA_OUT_TASK_STATUS);
 		return -EINVAL;
 	}
+	if (in_task->num_eof_task_status > MAX_NUM_NVDLA_OUT_TASK_STATUS) {
+		pr_err("eof task status[%u] crossing expected[%d]\n",
+			in_task->num_eof_task_status,
+			MAX_NUM_NVDLA_OUT_TASK_STATUS);
+		return -EINVAL;
+	}
+	if (in_task->num_sof_timestamps > MAX_NUM_NVDLA_OUT_TIMESTAMP) {
+		pr_err("sof timestamps[%u] crossing expected[%d]\n",
+			in_task->num_sof_timestamps,
+			MAX_NUM_NVDLA_OUT_TIMESTAMP);
+		return -EINVAL;
+	}
+	if (in_task->num_eof_timestamps > MAX_NUM_NVDLA_OUT_TIMESTAMP) {
+		pr_err("eof timestamps[%u] crossing expected[%d]\n",
+			in_task->num_eof_timestamps,
+			MAX_NUM_NVDLA_OUT_TIMESTAMP);
+		return -EINVAL;
+	}
 	if (in_task->num_addresses < 1) {
 		pr_err("num addresses[%u] should be min one\n",
 				in_task->num_addresses);
@@ -572,7 +711,10 @@ static int nvdla_fill_task(struct nvdla_queue *queue,
 	task->num_prefences = local_task->num_prefences;
 	task->num_postfences = local_task->num_postfences;
 	task->num_in_task_status = local_task->num_input_task_status;
-	task->num_out_task_status = local_task->num_output_task_status;
+	task->num_sof_task_status = local_task->num_sof_task_status;
+	task->num_eof_task_status = local_task->num_eof_task_status;
+	task->num_sof_timestamps = local_task->num_sof_timestamps;
+	task->num_eof_timestamps = local_task->num_eof_timestamps;
 	task->num_addresses = local_task->num_addresses;
 	task->timeout = local_task->timeout;
 
@@ -616,8 +758,14 @@ static void nvdla_dump_task(struct nvdla_task *task)
 	nvdla_dbg_info(pdev, "dumping input task [%p] parameters:", task);
 	nvdla_dbg_info(pdev, "num_prefences[%u] num_postfences[%u]",
 			task->num_prefences, task->num_postfences);
-	nvdla_dbg_info(pdev, "num_in_status[%u] num_out_task_status[%u]",
+	nvdla_dbg_info(pdev, "num_in_status[%u] num_sof_task_status[%u] "
-			task->num_in_task_status, task->num_out_task_status);
+			"num_eof_task_status[%u]",
+			task->num_in_task_status,
+			task->num_sof_task_status,
+			task->num_eof_task_status);
+	nvdla_dbg_info(pdev, "num_sof_timestamps[%u] num_eof_timestamps[%u]",
+			task->num_sof_timestamps,
+			task->num_eof_timestamps);
 	nvdla_dbg_info(pdev, "num_addresses[%u]", task->num_addresses);
 
 	for (i = 0; i < task->num_prefences; i++) {
@@ -654,12 +802,34 @@ static void nvdla_dump_task(struct nvdla_task *task)
 				task->in_task_status[i].status);
 	}
 
-	for (i = 0; i < task->num_out_task_status; i++) {
+	for (i = 0; i < task->num_sof_task_status; i++) {
-		nvdla_dbg_info(pdev, "Output task status[%d]:"
+		nvdla_dbg_info(pdev, "SOF task status[%d]:"
 				"handle[%u] offset[%u] status[%u]",
-				i, task->out_task_status[i].handle,
+				i, task->sof_task_status[i].handle,
-				task->out_task_status[i].offset,
+				task->sof_task_status[i].offset,
-				task->out_task_status[i].status);
+				task->sof_task_status[i].status);
+	}
+
+	for (i = 0; i < task->num_eof_task_status; i++) {
+		nvdla_dbg_info(pdev, "EOF task status[%d]:"
+				"handle[%u] offset[%u] status[%u]",
+				i, task->eof_task_status[i].handle,
+				task->eof_task_status[i].offset,
+				task->eof_task_status[i].status);
+	}
+
+	for (i = 0; i < task->num_sof_timestamps; i++) {
+		nvdla_dbg_info(pdev, "SOF timestamp[%d]:"
+				"handle[%u] offset[%u]",
+				i, task->sof_timestamps[i].handle,
+				task->sof_timestamps[i].offset);
+	}
+
+	for (i = 0; i < task->num_eof_timestamps; i++) {
+		nvdla_dbg_info(pdev, "EOF timestamp[%d]:"
+				"handle[%u] offset[%u]",
+				i, task->eof_timestamps[i].handle,
+				task->eof_timestamps[i].offset);
 	}
 
 	for (i = 0; i < task->num_addresses; i++) {
@@ -718,8 +888,18 @@ static int nvdla_emu_task_submit(struct nvdla_private *priv, void *arg)
 
 		nvdla_dbg_info(pdev, "submit [%d]th task", i + 1);
 
+		task.num_prefences = local_tasks[i].num_prefences;
 		task.num_postfences = local_tasks[i].num_postfences;
 
+		/* get pre fences */
+		if (copy_from_user(task.prefences,
+			(void __user *)local_tasks[i].prefences,
+			(task.num_prefences * sizeof(struct nvdev_fence)))) {
+			err = -EFAULT;
+			nvdla_dbg_err(pdev, "failed to copy prefences");
+			goto exit;
+		}
+
 		/* get post fences */
 		if (copy_from_user(task.postfences,
 			(void __user *)local_tasks[i].postfences,
@@ -736,13 +916,13 @@ static int nvdla_emu_task_submit(struct nvdla_private *priv, void *arg)
 		}
 		nvdla_dbg_info(pdev, "task[%d] submitted", i + 1);
 
-		/* send fences to user */
+		/* send signal fences to user */
-		err = nvdla_send_emu_postfences(&task, local_tasks + i);
+		err = nvdla_send_emu_signal_fences(&task, local_tasks + i);
 		if (err) {
-			nvdla_dbg_err(pdev, "fail to send postfence%d", i + 1);
+			nvdla_dbg_err(pdev, "fail to send sig fence%d", i + 1);
 			goto exit;
 		}
-		nvdla_dbg_info(pdev, "postfences of task[%d] sent", i + 1);
+		nvdla_dbg_info(pdev, "signal fences of task[%d] sent", i + 1);
 	}
 	nvdla_dbg_fn(pdev, "Emulator task submitted, done!");
 
@@ -825,8 +1005,8 @@ static int nvdla_submit(struct nvdla_private *priv, void *arg)
 		}
 		nvdla_dbg_info(pdev, "task[%d] desc filled", i + 1);
 
-		/* get expected postfences prior to submit */
+		/* get expected signal fences prior to submit */
-		err = nvdla_get_postfences(queue, task);
+		err = nvdla_get_signal_fences(queue, task);
 		if (err) {
 			nvdla_dbg_err(pdev, "fail to get fences%d", i + 1);
 			goto fail_to_get_fences;
@@ -834,7 +1014,7 @@ static int nvdla_submit(struct nvdla_private *priv, void *arg)
 		nvdla_dbg_info(pdev, "task[%d] got fences", i + 1);
 
 		/* update fences to user */
-		err = nvdla_update_postfences(task, local_tasks + i);
+		err = nvdla_update_signal_fences(task, local_tasks + i);
 		if (err) {
 			nvdla_dbg_err(pdev, "fail update postfence%d", i + 1);
 			goto fail_to_update_postfences;

include/uapi/linux/nvhost_nvdla_ioctl.h

@@ -3,7 +3,7 @@
  *
  * Tegra NvDLA Driver
  *
- * Copyright (c) 2016-2018, NVIDIA CORPORATION.  All rights reserved.
+ * Copyright (c) 2016-2019, NVIDIA CORPORATION.  All rights reserved.
  *
  * This program is free software; you can redistribute it and/or modify it
  * under the terms and conditions of the GNU General Public License,
@@ -79,7 +79,7 @@ struct nvdla_pin_unpin_args {
 struct nvdla_submit_args {
 	__u64 tasks;
 	__u16 num_tasks;
-#define MAX_TASKS_PER_SUBMIT		24
+#define MAX_TASKS_PER_SUBMIT		16
 #define NVDLA_SUBMIT_FLAGS_ATOMIC	(1 << 0)
 	__u16 flags;
 	__u32 version;
@@ -125,13 +125,19 @@ struct nvdla_mem_handle {
  * @num_prefences		number of pre-fences in task
  * @num_postfences		number of post-fences in task
  * @num_input_task_status	number of input task status
- * @num_output_task_status	number of output task status
+ * @num_sof_task_status  	number of sof task status
+ * @num_eof_task_status  	number of eof task status
+ * @num_sof_timestamps   	number of sof timestamp
+ * @num_eof_timestamps   	number of eof timestamp
  * @flags			flags for bitwise task info embeddeing
  * @reserved			reserved for future use
  * @prefences			pointer to pre-fence struct table
  * @postfences			pointer to post-fence struct table
  * @input_task_status		pointer to input task status struct table
- * @output_task_status		pointer to output task status struct table
+ * @sof_task_status  		pointer to sof task status struct table
+ * @eof_task_status  		pointer to eof task status struct table
+ * @sof_timestamps   		pointer to sof timestamp handle list
+ * @eof_timestamps   		pointer to eof timestamp handle list
  * @num_addresses		total number of addressed passed in structure
  * @address_list		pointer to address list
  * @timeout			task timeout
@@ -141,17 +147,24 @@ struct nvdla_ioctl_submit_task {
 	__u8 num_prefences;
 	__u8 num_postfences;
 	__u8 num_input_task_status;
-	__u8 num_output_task_status;
+	__u8 num_sof_task_status;
+	__u8 num_eof_task_status;
+	__u8 num_sof_timestamps;
+	__u8 num_eof_timestamps;
+	__u8 reserved0[1];
 #define NVDLA_MAX_BUFFERS_PER_TASK (6144)
 	__u32 num_addresses;
 	__u16 flags;
-	__u16 reserved;
+	__u16 reserved1;
 
 	__u64 prefences;
 	__u64 postfences;
 
 	__u64 input_task_status;
-	__u64 output_task_status;
+	__u64 sof_task_status;
+	__u64 eof_task_status;
+	__u64 sof_timestamps;
+	__u64 eof_timestamps;
 	__u64 address_list;
 	__u64 timeout;
 };
@@ -160,15 +173,17 @@ struct nvdla_ioctl_submit_task {
  * struct nvdla_ioctl_emu_submit_task structure for single emulator task
  * information
  *
+ * @num_prefences 		number of pre-fences in task
  * @num_postfences		number of post-fences in task
- * @reserved			reserved for padding and future use
+ * @prefences     		pointer to pre-fence struct table
  * @postfences    		pointer to post-fence struct table
  *
  */
 struct nvdla_ioctl_emu_submit_task {
+	__u32 num_prefences;
 	__u32 num_postfences;
-	__u32 reserved;
 
+	__u64 prefences;
 	__u64 postfences;
 };