kernel/0163-fix-Update-the-NPU-DSP-driver-code.patch

From 45c9e8bf67e031e2d1e49b5b3c02f5da0dfa87db Mon Sep 17 00:00:00 2001
From: denglei <denglei@eswincomputing.com>
Date: Mon, 2 Sep 2024 10:17:35 +0800
Subject: [PATCH 163/219] fix:Update the NPU & DSP driver code.

Changelogs:
	Merge the driver code of npu sdk & dsp sdk to dev branch.

Signed-off-by: denglei <denglei@eswincomputing.com>
---
 drivers/soc/eswin/ai_driver/dsp/dsp_hw_if.h   |  32 +++++
 drivers/soc/eswin/ai_driver/dsp/dsp_main.c    |  47 ++++---
 drivers/soc/eswin/ai_driver/dsp/dsp_main.h    |   6 +-
 .../soc/eswin/ai_driver/dsp/dsp_platform.c    |  64 ++++++++-
 .../soc/eswin/ai_driver/dsp/dsp_platform.h    |   2 +-
 drivers/soc/eswin/ai_driver/dsp/dsp_proc.c    | 124 +++++++++++++++---
 .../eswin/ai_driver/include/es_dsp_internal.h |   9 ++
 drivers/soc/eswin/ai_driver/npu/npu_main.c    |  17 ++-
 8 files changed, 252 insertions(+), 49 deletions(-)

diff --git a/drivers/soc/eswin/ai_driver/dsp/dsp_hw_if.h b/drivers/soc/eswin/ai_driver/dsp/dsp_hw_if.h
index 27e3313bfb5f..00b8539b1d47 100644
--- a/drivers/soc/eswin/ai_driver/dsp/dsp_hw_if.h
+++ b/drivers/soc/eswin/ai_driver/dsp/dsp_hw_if.h
@@ -22,6 +22,38 @@
 #define DSP_OP_LIB_DIR "/lib/firmware/dsp_kernels"
 #endif
 
+/*
+ * bit7-0: dsp fw state.
+ * 0: dsp fw not inited.
+ * 1: dsp is alive
+ * 2: dsp is died.
+ * bit15-8: npu task state.
+ * 0: no npu task process.
+ * 1: npu task is processing.
+ * bit23-16: dsp task state.
+ * 0: no dsp task process.
+ * 1: dsp task is processing.
+ * bit31-24: task func state:
+ * 0: no func exec
+ * 1: prepare func is executing.
+ * 2: prepare func is done.
+ * 3: eval func is executing.
+ * 4: eval func is done
+ */
+struct dsp_fw_state_t {
+    ES_U32 exccause;
+    ES_U32 ps;
+    ES_U32 pc;
+    union {
+        struct {
+            ES_U32 fw_state : 8;
+            ES_U32 npu_task_state : 8;
+            ES_U32 dsp_task_state : 8;
+            ES_U32 func_state : 8;
+        };
+        ES_U32 val;
+    };
+};
 
 typedef struct es_dsp_buffer_group_t {
     es_dsp_buffer* buffers;
diff --git a/drivers/soc/eswin/ai_driver/dsp/dsp_main.c b/drivers/soc/eswin/ai_driver/dsp/dsp_main.c
index 1fa928df25c2..dcfd23a6f205 100644
--- a/drivers/soc/eswin/ai_driver/dsp/dsp_main.c
+++ b/drivers/soc/eswin/ai_driver/dsp/dsp_main.c
@@ -56,7 +56,7 @@
 #define DSP_SUBSYS_HILOAD_CLK 1040000000
 #define DSP_SUBSYS_LOWLOAD_CLK 5200000
 
-#define ES_DSP_DEFAULT_TIMEOUT (100 * 2)
+#define ES_DSP_DEFAULT_TIMEOUT (100* 6)
 
 #ifdef DEBUG
 #pragma GCC optimize("O0")
@@ -236,8 +236,15 @@ static void dsp_process_expire_work(struct work_struct *work)
 	dsp_request_t *req;
 	unsigned long flags;
 	int ret;
-
-	dsp_err("%s, %d, task timeout.\n", __func__, __LINE__);
+	struct dsp_fw_state_t *dsp_fw_state =
+		(struct dsp_fw_state_t *)dsp->dsp_fw_state_base;
+
+	dsp_err("%s, %d, task timeout, dsp fw state=0x%x, excause=0x%x, ps=0x%x, pc=0x%x, dsp_task=0x%x"
+		"npu_task=0x%x, func_state=0x%x\n",
+		__func__, __LINE__, dsp_fw_state->fw_state,
+		dsp_fw_state->exccause, dsp_fw_state->ps, dsp_fw_state->pc,
+		dsp_fw_state->dsp_task_state, dsp_fw_state->npu_task_state,
+		dsp_fw_state->func_state);
 	ret = es_dsp_reboot_core(dsp->hw_arg);
 	if (ret < 0) {
 		dsp_err("reboot dsp core failed.\n");
@@ -695,7 +702,7 @@ int dsp_boot_firmware(struct es_dsp *dsp)
 	return 0;
 }
 
-int dsp_suspend(struct device *dev)
+int __maybe_unused dsp_suspend(struct device *dev)
 {
 	struct es_dsp *dsp = dev_get_drvdata(dev);
 	int ret;
@@ -716,11 +723,12 @@ int dsp_suspend(struct device *dev)
 	win2030_tbu_power(dsp->dev, false);
 	es_dsp_core_clk_disable(dsp);
 	dsp_disable_mbox_clock(dsp);
-	dsp_debug("%s, %d, dsp core%d generic suspend done.\n", __func__, __LINE__, dsp->process_id);
+	dsp_debug("%s, %d, dsp core%d generic suspend done.\n", __func__,
+		  __LINE__, dsp->process_id);
 	return 0;
 }
 
-int dsp_resume(struct device *dev)
+int __maybe_unused dsp_resume(struct device *dev)
 {
 	struct es_dsp *dsp = dev_get_drvdata(dev);
 	int ret;
@@ -760,7 +768,8 @@ int dsp_resume(struct device *dev)
 
 	pm_runtime_mark_last_busy(dsp->dev);
 	pm_runtime_put_autosuspend(dsp->dev);
-	dsp_debug("dsp_core%d Generic resume ok, dsp->off=%d.\n", dsp->process_id, dsp->off);
+	dsp_debug("dsp_core%d Generic resume ok, dsp->off=%d.\n",
+		  dsp->process_id, dsp->off);
 	return 0;
 err_firm:
 	es_dsp_hw_uninit(dsp);
@@ -773,16 +782,18 @@ int dsp_resume(struct device *dev)
 	return ret;
 }
 
-int dsp_runtime_suspend(struct device *dev)
+int __maybe_unused dsp_runtime_suspend(struct device *dev)
 {
 	struct es_dsp *dsp = dev_get_drvdata(dev);
-	dsp_debug("%s, dsp core%d runtime suspend.\n", __func__, dsp->process_id);
+	dsp_debug("%s, dsp core%d runtime suspend.\n", __func__,
+		  dsp->process_id);
+	win2030_tbu_power(dev, false);
 	es_dsp_core_clk_disable(dsp);
 	return 0;
 }
 EXPORT_SYMBOL(dsp_runtime_suspend);
 
-int dsp_runtime_resume(struct device *dev)
+int __maybe_unused dsp_runtime_resume(struct device *dev)
 {
 	struct es_dsp *dsp = dev_get_drvdata(dev);
 	int ret = 0;
@@ -798,6 +809,7 @@ int dsp_runtime_resume(struct device *dev)
 		dev_err(dsp->dev, "couldn't enable DSP\n");
 		goto out;
 	}
+	win2030_tbu_power(dev, true);
 	dsp_debug("dsp core%d, runtime resume ok.\n", dsp->process_id);
 out:
 	return ret;
@@ -965,7 +977,8 @@ static int es_dsp_hw_probe(struct platform_device *pdev)
 
 	ret = es_dsp_map_resource(dsp);
 	if (ret < 0) {
-		dsp_err("%s, %d, dsp map resource err, ret=%d.\n", __func__, __LINE__, ret);
+		dsp_err("%s, %d, dsp map resource err, ret=%d.\n", __func__,
+			__LINE__, ret);
 		goto err_map_res;
 	}
 	init_waitqueue_head(&dsp->hd_ready_wait);
@@ -983,7 +996,8 @@ static int es_dsp_hw_probe(struct platform_device *pdev)
 
 	ret = dsp_enable_mbox_clock(dsp);
 	if (ret < 0) {
-		dsp_err("%s, %d, enable mbox clock err, ret = %d.\n", __func__, __LINE__, ret);
+		dsp_err("%s, %d, enable mbox clock err, ret = %d.\n", __func__,
+			__LINE__, ret);
 		goto err_mbox_clk;
 	}
 	ret = es_dsp_clk_enable(dsp);
@@ -1097,10 +1111,9 @@ static int es_dsp_hw_remove(struct platform_device *pdev)
 	return 0;
 }
 
-
-static const struct dev_pm_ops es_dsp_hw_pm_ops = {
-	SYSTEM_SLEEP_PM_OPS(dsp_suspend, dsp_resume)
-	SET_RUNTIME_PM_OPS(dsp_runtime_suspend, dsp_runtime_resume, NULL) };
+static const struct dev_pm_ops es_dsp_hw_pm_ops = { SYSTEM_SLEEP_PM_OPS(
+	dsp_suspend, dsp_resume) SET_RUNTIME_PM_OPS(dsp_runtime_suspend,
+						    dsp_runtime_resume, NULL) };
 
 static struct platform_driver es_dsp_hw_driver = {
 	.probe   = es_dsp_hw_probe,
@@ -1108,7 +1121,7 @@ static struct platform_driver es_dsp_hw_driver = {
 	.driver  = {
 		.name = DRIVER_NAME,
 		.of_match_table = of_match_ptr(es_dsp_hw_match),
-		.pm = &es_dsp_hw_pm_ops,
+		.pm = pm_ptr(&es_dsp_hw_pm_ops),
 	},
 };
 
diff --git a/drivers/soc/eswin/ai_driver/dsp/dsp_main.h b/drivers/soc/eswin/ai_driver/dsp/dsp_main.h
index 7d2b9e77ee57..308a5b1758fb 100644
--- a/drivers/soc/eswin/ai_driver/dsp/dsp_main.h
+++ b/drivers/soc/eswin/ai_driver/dsp/dsp_main.h
@@ -128,6 +128,9 @@ struct es_dsp {
 	struct work_struct task_work;
 
 	struct es_dsp_stats *stats;
+	void __iomem *perf_reg_base;
+	u32 __iomem *dsp_fw_state_base;
+	void __iomem *flat_base;
 
 	struct timer_list task_timer;
 	struct work_struct expire_work;
@@ -147,7 +150,6 @@ struct es_dsp {
 	struct resource *mbox_tx_res;
 	struct resource *mbox_rx_res;
 
-
 	u64 send_time;
 	u64 done_time;
 
@@ -179,7 +181,7 @@ struct es_dsp {
 
 /* 0xFF9B_0000 -- 0xFFFB_0000 for SPAD */
 #define DSP_IDDR_IOVA 0xff9b0000
-#define DSP_IDDR_IOVA_SIZE 0x400000
+#define DSP_IDDR_IOVA_SIZE 0x600000
 
 #define DSP_DEVICE_AUX_E31_IOVA 0xfffb0000  // unused iova
 #define DSP_DEVICE_AUX_E31_IOVA_SIZE 0x1000
diff --git a/drivers/soc/eswin/ai_driver/dsp/dsp_platform.c b/drivers/soc/eswin/ai_driver/dsp/dsp_platform.c
index 998538a44f46..14c8ed7c2350 100644
--- a/drivers/soc/eswin/ai_driver/dsp/dsp_platform.c
+++ b/drivers/soc/eswin/ai_driver/dsp/dsp_platform.c
@@ -79,6 +79,13 @@
 #define SCU_DSPT_DIV_SEL BIT_ULL(19)
 #define SCU_DSPT_DIV_EN BIT_ULL(23)
 
+// from eswin/dsp/framework/lsp/memmap.xmm .dram1.perfdata(0x2813ffc8)
+#define DSP_FLAT_ADDR 0x5b13fe70
+#define DSP_FW_STATE_ADDR 0x5b13ffb0
+#define DSP_PERF_START_ADDR 0x5b13ffc8
+#define DIE_BASE_INTERVAL 0x20000000
+#define DSP_CORE_INTERVAL 0x40000
+
 enum dsp_irq_mode {
 	DSP_IRQ_NONE,
 	DSP_IRQ_LEVEL,
@@ -339,6 +346,17 @@ void es_dsp_release(struct es_dsp_hw *hw)
 	return;
 }
 
+static int check_dsp_fw_state(struct es_dsp *dsp)
+{
+	struct dsp_fw_state_t *dsp_state = (struct dsp_fw_state_t *)dsp->dsp_fw_state_base;
+	if (dsp_state == NULL) {
+		return 0;
+	}
+	dsp_err("%s, %d, exccause=0x%x, ps=0x%x, pc=0x%x.\n", __func__, __LINE__, dsp_state->exccause, dsp_state->ps, dsp_state->pc);
+	dsp_err("%s, %d, fw_state=%d, npu_state=%d, dsp_state=%d, func_state=%d.\n", __func__, __LINE__, dsp_state->fw_state, dsp_state->npu_task_state, dsp_state->dsp_task_state, dsp_state->func_state);
+	return 0;
+}
+
 static int dsp_send_msg_by_mbx(struct es_dsp *dsp, void *data)
 {
 	unsigned long flags;
@@ -351,6 +369,9 @@ static int dsp_send_msg_by_mbx(struct es_dsp *dsp, void *data)
 	while (true) {
 		if (count > 3) {
 			spin_unlock_irqrestore(&dsp->mbox_lock, flags);
+			dsp_err("%s, %d, tx mbxlock = 0x%x, fifo status=0x%x.\n", __func__, __LINE__, readl(dsp->mbox_tx_base + ESWIN_MBOX_WR_LOCK),
+					readl(dsp->mbox_tx_base + ESWIN_MBOX_FIFO_STATUS));
+			check_dsp_fw_state(dsp);
 			return -EBUSY;
 		}
 		writel(dsp->mbox_lock_bit,
@@ -390,14 +411,15 @@ void dsp_send_invalid_code_seg(struct es_dsp_hw *hw, struct dsp_op_desc *op)
 	h2d_msg.command = DSP_CMD_INVALID_ICACHE;
 	h2d_msg.size = op->op_shared_seg_size >> DSP_2M_SHIFT;
 	h2d_msg.iova_ptr = op->iova_base;
-
+	dsp_debug("DSP, send inv iov=0x%x.\n", op->iova_base);
 	ret = dsp_send_msg_by_mbx(hw->es_dsp, (void *)&h2d_msg);
-	if (ret < 0)
+	if (ret < 0) {
 		dev_err(NULL, "Failed to send message via mailbox\n");
+	}
 	return;
 }
 
-void es_dsp_send_irq(struct es_dsp_hw *hw, dsp_request_t *req)
+int es_dsp_send_irq(struct es_dsp_hw *hw, dsp_request_t *req)
 {
 	es_dsp_h2d_msg h2d_msg;
 	int ret;
@@ -406,11 +428,12 @@ void es_dsp_send_irq(struct es_dsp_hw *hw, dsp_request_t *req)
 	h2d_msg.poll_mode = req->poll_mode;
 	h2d_msg.sync_cache = req->sync_cache;
 	h2d_msg.iova_ptr = req->dsp_flat1_iova;
-
+	dsp_debug("DSP, send irq iova=0x%x.\n", req->dsp_flat1_iova);
 	ret = dsp_send_msg_by_mbx(hw->es_dsp, (void *)&h2d_msg);
-	if (ret < 0)
-		dev_err(NULL, "Failed to send message via mailbox\n");
-	return;
+	if (ret < 0) {
+		dev_err(NULL, " dsp mailbox send message busy, try again.\n");
+	}
+	return ret;
 }
 
 /*
@@ -1047,6 +1070,7 @@ int es_dsp_map_resource(struct es_dsp *dsp)
 {
 	struct es_dsp_hw *hw = (struct es_dsp_hw *)dsp->hw_arg;
 	int ret;
+	unsigned long base;
 
 	hw->flat_dma_pool = dma_pool_create("dsp_flat_dma", dsp->dev,
 					    sizeof(struct es_dsp_flat1_desc) +
@@ -1079,6 +1103,32 @@ int es_dsp_map_resource(struct es_dsp *dsp)
 		ret = -ENOMEM;
 		return ret;
 	}
+
+	base = DSP_PERF_START_ADDR + dsp->numa_id * DIE_BASE_INTERVAL +
+	       DSP_CORE_INTERVAL * dsp->process_id;
+
+	dsp->perf_reg_base = devm_ioremap(dsp->dev, base, sizeof(es_dsp_perf_info));
+	if (!dsp->perf_reg_base) {
+		dev_err(dsp->dev, "ioremap for perf reg err.\n");
+		return -ENOMEM;
+	}
+
+	base = DSP_FW_STATE_ADDR + dsp->numa_id * DIE_BASE_INTERVAL +
+	       DSP_CORE_INTERVAL * dsp->process_id;
+	dsp->dsp_fw_state_base = devm_ioremap(dsp->dev, base, 0x18);
+	if (!dsp->dsp_fw_state_base) {
+		dev_err(dsp->dev, "ioremap for dsp fw state err.\n");
+		return -ENOMEM;
+	}
+
+	base = DSP_FLAT_ADDR + dsp->numa_id * DIE_BASE_INTERVAL +
+	       DSP_CORE_INTERVAL * dsp->process_id;
+	dsp->flat_base = devm_ioremap(dsp->dev, base, 0x140);
+	if (!dsp->flat_base) {
+		dev_err(dsp->dev, "ioremap for dsp flat base err.\n");
+		return -ENOMEM;
+	}
+
 	return 0;
 }
 
diff --git a/drivers/soc/eswin/ai_driver/dsp/dsp_platform.h b/drivers/soc/eswin/ai_driver/dsp/dsp_platform.h
index 82343e910bc2..b01f21bc7055 100644
--- a/drivers/soc/eswin/ai_driver/dsp/dsp_platform.h
+++ b/drivers/soc/eswin/ai_driver/dsp/dsp_platform.h
@@ -11,7 +11,7 @@
 #include <linux/platform_device.h>
 #include "dsp_main.h"
 struct es_dsp_hw;
-void es_dsp_send_irq(struct es_dsp_hw *, dsp_request_t *);
+int es_dsp_send_irq(struct es_dsp_hw *, dsp_request_t *);
 int es_dsp_reboot_core(struct es_dsp_hw *);
 int es_dsp_enable(struct es_dsp_hw *);
 void es_dsp_disable(struct es_dsp_hw *);
diff --git a/drivers/soc/eswin/ai_driver/dsp/dsp_proc.c b/drivers/soc/eswin/ai_driver/dsp/dsp_proc.c
index b12258767c43..27e39de0ccb0 100644
--- a/drivers/soc/eswin/ai_driver/dsp/dsp_proc.c
+++ b/drivers/soc/eswin/ai_driver/dsp/dsp_proc.c
@@ -9,51 +9,51 @@
 #include <linux/proc_fs.h>
 #include <linux/kernel.h>
 #include <linux/seq_file.h>
+#include <linux/iommu.h>
 #include "dsp_main.h"
 
 static struct proc_dir_entry *proc_es_dsp;
 extern int dsp_log_level;
 int dsp_perf_enable = 0;
 
-// from eswin/dsp/framework/lsp/memmap.xmm .dram1.perfdata(0x2813ffc0)
-#define DSP_PERF_START_ADDR 0x5b13ffc0
-#define DIE_BASE_INTERVAL 0x20000000
-#define DSP_CORE_INTERVAL 0x40000
-
 void get_dsp_perf_info(es_dsp_perf_info *perf_info, int die_num, int dsp_num)
 {
 	struct es_dsp *dsp = NULL;
-	void *iomem = NULL;
-	unsigned long phys;
 
 	memset((void *)perf_info, 0, sizeof(es_dsp_perf_info));
 	dsp = es_proc_get_dsp(die_num, dsp_num);
 	if (!dsp) {
 		return;
 	}
+	memcpy((void *)perf_info, dsp->perf_reg_base, sizeof(es_dsp_perf_info));
+}
 
-	phys = DSP_PERF_START_ADDR + die_num * DIE_BASE_INTERVAL +
-	       DSP_CORE_INTERVAL * dsp_num;
-	iomem = ioremap(phys, sizeof(es_dsp_perf_info));
-	if (!iomem) {
-		return;
-	}
-	memcpy((void *)perf_info, iomem, sizeof(es_dsp_perf_info));
+static char *fw_state[] = {
+	"Not_ready",
+	"Alive",
+	"Hang",
+};
 
-	iounmap(iomem);
-}
+static char *func_state[] = {
+	"No_task", "prep_run", "prep_done", "eval_run", "eval_done",
+};
 
 static int stats_show(struct seq_file *m, void *p)
 {
 	struct es_dsp *dsp;
 	int i, j;
 	dsp_request_t req;
+	dsp_request_t *myreq;
 	struct timespec64 ts;
 	const int die_cnt = 2;
 	const int dsp_cnt = 4;
 	es_dsp_perf_info perf_info;
 	int k;
-
+	u32 state, cause, ps, pc;
+	u32 fw_val, npu_task, dsp_task, func_val;
+	struct dsp_fw_state_t *dsp_fw_state;
+	struct iommu_domain *domain;
+	phys_addr_t phys;
 	seq_printf(
 		m,
 		"--------------------------------DSP PARAM INFO----------------------------------\n");
@@ -75,6 +75,96 @@ static int stats_show(struct seq_file *m, void *p)
 		}
 	}
 
+	seq_printf(m, "\n");
+	seq_printf(
+		m,
+		"--------------------------DSP FW STATE--------------------------------------\n");
+	seq_printf(m,
+		   "  %-5s %-6s\t %-8s\t %-8s   %-8s"
+		   "   %-8s\t    %-8s   %-8s   %-10s\n",
+		   "DieId", "CoreId", "fw_state", "cause", "ps", "pc",
+		   "npu_task", "dsp_task", "func_state");
+	for (j = 0; j < die_cnt; j++) {
+		for (i = 0; i < dsp_cnt; i++) {
+			dsp = es_proc_get_dsp(j, i);
+			if (dsp == NULL) {
+				continue;
+			}
+			if (!dsp->dsp_fw_state_base) {
+				continue;
+			}
+			dsp_fw_state =
+				(struct dsp_fw_state_t *)dsp->dsp_fw_state_base;
+			cause = dsp_fw_state->exccause;
+			ps = dsp_fw_state->ps;
+			pc = dsp_fw_state->pc;
+			fw_val = dsp_fw_state->fw_state;
+			dsp_task = dsp_fw_state->dsp_task_state;
+			npu_task = dsp_fw_state->npu_task_state;
+			func_val = dsp_fw_state->func_state;
+			seq_printf(
+				m,
+				"  %-5d %-6d\t %-8s\t 0x%-8x 0x%-8x 0x%-8x\t %-8s %-8s %-10s\n",
+				j, i, fw_state[fw_val], cause, ps, pc,
+				npu_task ? "run" : "no_task",
+				dsp_task ? "run" : "no_task",
+				func_state[func_val]);
+		}
+	}
+
+	seq_printf(
+		m,
+		"--------------------------dsp hw flat content--------------------------------------\n");
+	for (j = 0; j < die_cnt; j++) {
+		struct dsp_op_desc *opdesc;
+		struct dsp_hw_flat_test *hw_flat;
+		for (i = 0; i < dsp_cnt; i++) {
+			dsp = es_proc_get_dsp(j, i);
+			if (dsp == NULL) {
+				continue;
+			}
+			domain = iommu_get_domain_for_dev(dsp->dev);
+
+			hw_flat = dsp->flat_base;
+			if (hw_flat->flat_iova == 0) {
+				continue;
+			}
+			phys = iommu_iova_to_phys(domain, hw_flat->flat_iova);
+			seq_printf(m, "die=%d, core=%d, flat iova = 0x%x, phys=0x%lx.\n", j, i, hw_flat->flat_iova, phys);
+			seq_printf(m, "num_buf=%d, input_idx=%d, out_idx=%d.\n", hw_flat->num_buffer, hw_flat->input_index, hw_flat->output_index);
+		}
+	}
+	seq_printf(
+		m,
+		"--------------------------DSP Current Task--------------------------------------\n");
+
+	for (j = 0; j < die_cnt; j++) {
+		struct dsp_op_desc *opdesc;
+		for (i = 0; i < dsp_cnt; i++) {
+			dsp = es_proc_get_dsp(j, i);
+			if (dsp == NULL) {
+				continue;
+			}
+			if (!dsp->current_task) {
+				continue;
+			}
+			myreq = dsp->current_task;
+			opdesc = (struct dsp_op_desc *)myreq->handle;
+			seq_printf(m, "die=%d, dspcore=%d, opname=%s", j, i, opdesc->name);
+			seq_printf(
+				m,
+				"\tdie=%d core=%d flat_iova=%x num_buf=%x input_idx=%d output_idx=%d\n", j, i, myreq->dsp_flat1_iova, myreq->flat_virt->num_buffer, myreq->flat_virt->input_index,
+				myreq->flat_virt->output_index);
+			domain = iommu_get_domain_for_dev(dsp->dev);
+			seq_printf(m, "domain_type=0x%x.\n", domain->type);
+			for (k = 0; k < myreq->flat_virt->num_buffer; k++) {
+				phys = iommu_iova_to_phys(domain, myreq->flat_virt->buffers[k].addr);
+				seq_printf(m, "buffer[%d]=0x%x. size=0x%x, phys=0x%llx\n", k, myreq->flat_virt->buffers[k].addr, myreq->flat_virt->buffers[k].size, phys);
+			}
+		}
+	}
+
+
 	seq_printf(m, "\n");
 	seq_printf(
 		m,
diff --git a/drivers/soc/eswin/ai_driver/include/es_dsp_internal.h b/drivers/soc/eswin/ai_driver/include/es_dsp_internal.h
index ec2c202ed2d8..3fd15afae0c8 100644
--- a/drivers/soc/eswin/ai_driver/include/es_dsp_internal.h
+++ b/drivers/soc/eswin/ai_driver/include/es_dsp_internal.h
@@ -60,6 +60,7 @@ struct operator_funcs {
 };
 #endif
 
+
 typedef struct es_dsp_buffer_t {
     ES_U32 addr;
     ES_U32 size;
@@ -133,6 +134,14 @@ typedef struct {
      */
     ES_U32 reserved : 11;
 } es_dsp_h2d_msg;
+struct dsp_hw_flat_test {
+    struct operator_funcs funcs;
+    ES_U32 num_buffer;
+    ES_U32 input_index;
+    ES_U32 output_index;
+    ES_U32 flat_iova;
+    es_dsp_buffer buffers[0];
+};
 
 typedef struct {
     /**
diff --git a/drivers/soc/eswin/ai_driver/npu/npu_main.c b/drivers/soc/eswin/ai_driver/npu/npu_main.c
index e83f7c9c85f3..309e6b6bafef 100644
--- a/drivers/soc/eswin/ai_driver/npu/npu_main.c
+++ b/drivers/soc/eswin/ai_driver/npu/npu_main.c
@@ -558,7 +558,7 @@ static int32_t __exit edla_remove(struct platform_device *pdev)
 	return 0;
 }
 
-int npu_runtime_suspend(struct device *dev)
+int __maybe_unused npu_runtime_suspend(struct device *dev)
 {
 	struct nvdla_device *ndev = dev_get_drvdata(dev);
 	int ret;
@@ -567,12 +567,14 @@ int npu_runtime_suspend(struct device *dev)
 		dla_error("%s, %d, ndev is null.\n", __func__, __LINE__);
 		return -EIO;
 	}
+
+	npu_tbu_power(dev, false);
 	ret = npu_disable_clock(ndev);
 	dla_debug("%s, %d, ret=%d.\n", __func__, __LINE__, ret);
 	return ret;
 }
 
-int npu_runtime_resume(struct device *dev)
+int __maybe_unused npu_runtime_resume(struct device *dev)
 {
 	struct nvdla_device *ndev = dev_get_drvdata(dev);
 	int ret;
@@ -582,11 +584,16 @@ int npu_runtime_resume(struct device *dev)
 		return -EIO;
 	}
 	ret = npu_enable_clock(ndev);
+    if (ret) {
+        dla_error("%s, %d, enable clock err, ret = %d.\n", __func__, __LINE__, ret);
+        return ret;
+    }
+	npu_tbu_power(dev, true);
 	dla_debug("%s, %d, ret=%d.\n", __func__, __LINE__, ret);
 	return ret;
 }
 
-int npu_suspend(struct device *dev)
+int __maybe_unused npu_suspend(struct device *dev)
 {
 	int ret;
 	struct nvdla_device *nvdla_dev = dev_get_drvdata(dev);
@@ -616,7 +623,7 @@ int npu_suspend(struct device *dev)
 	return 0;
 }
 
-int npu_resume(struct device *dev)
+int __maybe_unused npu_resume(struct device *dev)
 {
 	int ret;
 	struct nvdla_device *ndev = dev_get_drvdata(dev);
@@ -683,7 +690,7 @@ static struct platform_driver edla_driver =
 	{
 		.name = DRIVER_NAME,
 		.of_match_table = of_match_ptr(edla_of_match),
-		.pm = &npu_hw_pm_ops,
+		.pm = pm_ptr(&npu_hw_pm_ops),
 	},
 };
 
-- 
2.47.0