video: rockchip: mpp: add calc hw real times per task

Signed-off-by: Yandong Lin <yandong.lin@rock-chips.com>
Change-Id: Ibf21463caae403f7c71884647c546fc46a614f06

drivers/video/rockchip/mpp/mpp_common.c

@@ -2351,6 +2351,28 @@ int mpp_time_diff(struct mpp_task *task)
 	return 0;
 }
 
+int mpp_time_diff_with_hw_time(struct mpp_task *task, u32 clk_hz)
+{
+	if (mpp_debug_unlikely(DEBUG_TIMING)) {
+		ktime_t end;
+		struct mpp_dev *mpp = mpp_get_task_used_device(task, task->session);
+
+		end = ktime_get();
+
+		if (clk_hz)
+			mpp_debug(DEBUG_TIMING, "%s:%d session %d:%d time: %lld us hw %d us\n",
+				dev_name(mpp->dev), task->core_id, task->session->pid,
+				task->session->index, ktime_us_delta(end, task->start),
+				task->hw_cycles / (clk_hz / 1000000));
+		else
+			mpp_debug(DEBUG_TIMING, "%s:%d session %d:%d time: %lld us\n",
+				dev_name(mpp->dev), task->core_id, task->session->pid,
+				task->session->index, ktime_us_delta(end, task->start));
+	}
+
+	return 0;
+}
+
 #define LOG_TIMING(state, id, stage, time, base) \
 	do { \
 		if (test_bit(id, &state)) \
@@ -2512,6 +2534,7 @@ int mpp_clk_set_rate(struct mpp_clk_info *clk_info,
 	if (clk_rate_hz) {
 		clk_info->used_rate_hz = clk_rate_hz;
 		clk_set_rate(clk_info->clk, clk_rate_hz);
+		clk_info->real_rate_hz = clk_get_rate(clk_info->clk);
 	}
 
 	return 0;

drivers/video/rockchip/mpp/mpp_common.h

@@ -285,6 +285,7 @@ struct mpp_clk_info {
 	u32 reduce_rate_hz;
 	/* record last used rate */
 	u32 used_rate_hz;
+	u32 real_rate_hz;
 };
 
 struct mpp_dev_var {
@@ -493,6 +494,8 @@ struct mpp_task {
 	/* for multi-core */
 	struct mpp_dev *mpp;
 	s32 core_id;
+	/* hw cycles */
+	u32 hw_cycles;
 };
 
 struct mpp_taskqueue {
@@ -707,6 +710,7 @@ int mpp_set_grf(struct mpp_grf_info *grf_info);
 
 int mpp_time_record(struct mpp_task *task);
 int mpp_time_diff(struct mpp_task *task);
+int mpp_time_diff_with_hw_time(struct mpp_task *task, u32 clk_hz);
 int mpp_time_part_diff(struct mpp_task *task);
 
 int mpp_write_req(struct mpp_dev *mpp, u32 *regs,

drivers/video/rockchip/mpp/mpp_rkvdec2.c

@@ -404,13 +404,15 @@ static int rkvdec2_isr(struct mpp_dev *mpp)
 	u32 err_mask;
 	struct rkvdec2_task *task = NULL;
 	struct mpp_task *mpp_task = mpp->cur_task;
+	struct rkvdec2_dev *dec = to_rkvdec2_dev(mpp);
 
 	/* FIXME use a spin lock here */
 	if (!mpp_task) {
 		dev_err(mpp->dev, "no current task\n");
 		return IRQ_HANDLED;
 	}
-	mpp_time_diff(mpp_task);
+	mpp_task->hw_cycles = mpp_read(mpp, RKVDEC_PERF_WORKING_CNT);
+	mpp_time_diff_with_hw_time(mpp_task, dec->core_clk_info.real_rate_hz);
 	mpp->cur_task = NULL;
 	task = to_rkvdec2_task(mpp_task);
 	task->irq_status = mpp->irq_status;

drivers/video/rockchip/mpp/mpp_rkvdec2.h

@@ -73,6 +73,7 @@
 					RKVDEC_BUF_EMPTY_STA |\
 					RKVDEC_TIMEOUT_STA |\
 					RKVDEC_ERROR_STA)
+#define RKVDEC_PERF_WORKING_CNT		0x41c
 
 /* perf sel reference register */
 #define RKVDEC_PERF_SEL_OFFSET		0x20000

drivers/video/rockchip/mpp/mpp_rkvdec2_link.c

@@ -71,6 +71,7 @@ struct rkvdec_link_info rkvdec_link_v2_hw_info = {
 		.reg_num = 28,
 	},
 	.tb_reg_int = 180,
+	.tb_reg_cycle = 195,
 	.hack_setup = 0,
 };
 
@@ -124,6 +125,7 @@ struct rkvdec_link_info rkvdec_link_rk356x_hw_info = {
 		.reg_num = 28,
 	},
 	.tb_reg_int = 164,
+	.tb_reg_cycle = 179,
 	.hack_setup = 1,
 };
 
@@ -538,6 +540,7 @@ static int rkvdec_link_isr_recv_task(struct mpp_dev *mpp,
 	struct rkvdec_link_info *info = link_dec->info;
 	u32 *table_base = (u32 *)link_dec->table->vaddr;
 	int i;
+	struct rkvdec2_dev *dec = to_rkvdec2_dev(mpp);
 
 	for (i = 0; i < count; i++) {
 		int idx = rkvdec_link_get_task_read(link_dec);
@@ -598,6 +601,8 @@ static int rkvdec_link_isr_recv_task(struct mpp_dev *mpp,
 		task = to_rkvdec2_task(mpp_task);
 		regs = table_base + idx * link_dec->link_reg_count;
 		irq_status = regs[info->tb_reg_int];
+		mpp_task->hw_cycles = regs[info->tb_reg_cycle];
+		mpp_time_diff_with_hw_time(mpp_task, dec->core_clk_info.real_rate_hz);
 		mpp_dbg_link_flow("slot %d rd task %d\n", idx,
 				  mpp_task->task_id);
 
@@ -1207,7 +1212,6 @@ static int mpp_task_queue(struct mpp_dev *mpp, struct mpp_task *task)
 	mpp_debug_enter();
 
 	rkvdec2_link_power_on(mpp);
-	mpp_time_record(task);
 	mpp_debug(DEBUG_TASK_INFO, "pid %d, start hw %s\n",
 		  task->session->pid, dev_name(mpp->dev));
 
@@ -1664,6 +1668,7 @@ static int rkvdec2_soft_ccu_dequeue(struct mpp_taskqueue *queue)
 				 &queue->running_list,
 				 queue_link) {
 		struct mpp_dev *mpp = mpp_get_task_used_device(mpp_task, mpp_task->session);
+		struct rkvdec2_dev *dec = to_rkvdec2_dev(mpp);
 		u32 irq_status = mpp->irq_status;
 		u32 timeout_flag = test_bit(TASK_STATE_TIMEOUT, &mpp_task->state);
 		u32 abort_flag = test_bit(TASK_STATE_ABORT, &mpp_task->state);
@@ -1685,7 +1690,8 @@ static int rkvdec2_soft_ccu_dequeue(struct mpp_taskqueue *queue)
 
 			set_bit(TASK_STATE_HANDLE, &mpp_task->state);
 			cancel_delayed_work(&mpp_task->timeout_work);
-			mpp_time_diff(mpp_task);
+			mpp_task->hw_cycles = mpp_read(mpp, RKVDEC_PERF_WORKING_CNT);
+			mpp_time_diff_with_hw_time(mpp_task, dec->core_clk_info.real_rate_hz);
 			task->irq_status = irq_status;
 			mpp_debug(DEBUG_IRQ_CHECK, "irq_status=%08x, timeout=%u, abort=%u\n",
 				  irq_status, timeout_flag, abort_flag);

drivers/video/rockchip/mpp/mpp_rkvdec2_link.h

@@ -112,6 +112,7 @@ struct rkvdec_link_info {
 
 	/* interrupt read back in table buffer */
 	u32 tb_reg_int;
+	u32 tb_reg_cycle;
 	bool hack_setup;
 };