Media:wave5: wave5 v4l2 driver support gst/ffmpeg

Signed-off-by: Som Qin <som.qin@starfivetech.com>

[sw0312.kim: cherry-pick the commit 3f710b8fa137 from https://github.com/starfive-tech/linux/tree/JH7110_VisionFive2_6.1.y_devel
 - update sifive cache function name properly and add guard for its config option]
Signed-off-by: Seung-Woo Kim <sw0312.kim@samsung.com>

Change-Id: Ib264712e893c737a28c42a566607b7a55d7f7117

6 files changed, 180 insertions, 67 deletions

diff --git a/drivers/media/platform/chips-media/wave5/wave5-helper.c b/drivers/media/platform/chips-media/wave5/wave5-helper.c
index fb9c94d07853..09dfb20e9a20 100644
--- a/drivers/media/platform/chips-media/wave5/wave5-helper.c
+++ b/drivers/media/platform/chips-media/wave5/wave5-helper.c
@@ -47,7 +47,7 @@ int wave5_vpu_release_device(struct file *filp,
 				break;
 			if (fail_res != WAVE5_SYSERR_VPU_STILL_RUNNING)
 				break;
-			if (!wave5_vpu_wait_interrupt(inst, VPU_DEC_TIMEOUT))
+			if (!wave5_vpu_wait_interrupt(inst, VPU_DEC_TIMEOUT/10))
 				break;
 		} while (--retry_count);
 
@@ -106,7 +106,6 @@ int wave5_vpu_subscribe_event(struct v4l2_fh *fh, const struct v4l2_event_subscr
 {
 	struct vpu_instance *inst = wave5_to_vpu_inst(fh);
 	bool is_decoder = inst->type == VPU_INST_TYPE_DEC;
-	printk("wave5 subscribe event type: %d id: %d | flags: %d\n",sub->type, sub->id, sub->flags);
 
 	dev_dbg(inst->dev->dev, "%s: [%s] type: %u id: %u | flags: %u\n", __func__,
 		is_decoder ? "decoder" : "encoder", sub->type, sub->id, sub->flags);
@@ -130,8 +129,13 @@ int wave5_vpu_g_fmt_out(struct file *file, void *fh, struct v4l2_format *f)
 	struct vpu_instance *inst = wave5_to_vpu_inst(fh);
 	int i;
 
-	f->fmt.pix_mp.width = inst->src_fmt.width;
-	f->fmt.pix_mp.height = inst->src_fmt.height;
+	if (inst->state >= VPU_INST_STATE_INIT_SEQ){
+		f->fmt.pix_mp.width = inst->src_fmt.width - inst->crop_rect.right;
+		f->fmt.pix_mp.height = inst->src_fmt.height - inst->crop_rect.bottom;
+	} else {
+		f->fmt.pix_mp.width = inst->src_fmt.width;
+		f->fmt.pix_mp.height = inst->src_fmt.height;
+	}
 	f->fmt.pix_mp.pixelformat = inst->src_fmt.pixelformat;
 	f->fmt.pix_mp.field = inst->src_fmt.field;
 	f->fmt.pix_mp.flags = inst->src_fmt.flags;
diff --git a/drivers/media/platform/chips-media/wave5/wave5-helper.h b/drivers/media/platform/chips-media/wave5/wave5-helper.h
index d586d624275e..8a8dda8242f6 100644
--- a/drivers/media/platform/chips-media/wave5/wave5-helper.h
+++ b/drivers/media/platform/chips-media/wave5/wave5-helper.h
@@ -11,7 +11,7 @@
 #include "wave5-vpu.h"
 
 #define FMT_TYPES	2
-#define MAX_FMTS	6
+#define MAX_FMTS	3
 
 void wave5_cleanup_instance(struct vpu_instance *inst);
 int wave5_vpu_release_device(struct file *filp,
diff --git a/drivers/media/platform/chips-media/wave5/wave5-hw.c b/drivers/media/platform/chips-media/wave5/wave5-hw.c
index 3c8328699201..57bbc3862be2 100644
--- a/drivers/media/platform/chips-media/wave5/wave5-hw.c
+++ b/drivers/media/platform/chips-media/wave5/wave5-hw.c
@@ -584,6 +584,17 @@ static void wave5_get_dec_seq_result(struct vpu_instance *inst, struct dec_initi
 	info->f_rate_numerator = vpu_read_reg(inst->dev, W5_RET_DEC_FRAME_RATE_NR);
 	info->f_rate_denominator = vpu_read_reg(inst->dev, W5_RET_DEC_FRAME_RATE_DR);
 
+	if (info->f_rate_numerator > 0 && info->f_rate_denominator >0) {
+		if (inst->std == W_HEVC_DEC)
+			info->ns_per_frame = 1000000000 * (u64)info->f_rate_denominator / (u64)info->f_rate_numerator;
+		else if (inst->std == W_AVC_DEC)
+			info->ns_per_frame = 1000000000 * 2 * (u64)info->f_rate_denominator / (u64)info->f_rate_numerator;
+		else
+			info->ns_per_frame = 1000000000 / 30; //30fps
+	} else {
+		info->ns_per_frame = 1000000000 / 30; //30fps
+	}
+
 	reg_val = vpu_read_reg(inst->dev, W5_RET_DEC_COLOR_SAMPLE_INFO);
 	info->luma_bitdepth = reg_val & 0xf;
 	info->chroma_bitdepth = (reg_val >> 4) & 0xf;
@@ -820,10 +831,15 @@ int wave5_vpu_dec_register_framebuffer(struct vpu_instance *inst, struct frame_b
 				p_dec_info->vb_fbc_c_tbl[i] = vb_buf;
 			}
 		}
-		pic_size = (inst->display_fmt.width << 16) | (inst->display_fmt.height);
-		if (init_info->pic_width != inst->display_fmt.width ||
-				init_info->pic_height != inst->display_fmt.height)
+
+		if ((init_info->pic_width - init_info->pic_crop_rect.right != inst->display_fmt.width) ||
+			init_info->pic_height - init_info->pic_crop_rect.bottom != inst->display_fmt.height) {
+			pic_size = (inst->display_fmt.width << 16) | (inst->display_fmt.height);
 			scale_en = 1;
+		} else {
+			pic_size = (init_info->pic_width << 16) | (init_info->pic_height);
+			scale_en = 0;
+		}
 
 		// allocate task_buffer
 		vb_buf.size = (p_dec_info->vlc_buf_size * VLC_BUF_NUM) +
@@ -839,10 +855,14 @@ int wave5_vpu_dec_register_framebuffer(struct vpu_instance *inst, struct frame_b
 			      p_dec_info->vb_task.daddr);
 		vpu_write_reg(inst->dev, W5_CMD_SET_FB_TASK_BUF_SIZE, vb_buf.size);
 	} else {
-		pic_size = (inst->display_fmt.width << 16) | (inst->display_fmt.height);
-		if (init_info->pic_width != inst->display_fmt.width ||
-				init_info->pic_height != inst->display_fmt.height)
+		if ((init_info->pic_width - init_info->pic_crop_rect.right != inst->display_fmt.width) ||
+			init_info->pic_height - init_info->pic_crop_rect.bottom != inst->display_fmt.height) {
+			pic_size = (inst->display_fmt.width << 16) | (inst->display_fmt.height);
 			scale_en = 1;
+		} else {
+			pic_size = (init_info->pic_width << 16) | (init_info->pic_height);
+			scale_en = 0;
+		}
 	}
 	dev_dbg(inst->dev->dev, "set pic_size 0x%x\n", pic_size);
 	endian = wave5_vdi_convert_endian(inst->dev, fb_arr[0].endian);
@@ -859,7 +879,8 @@ int wave5_vpu_dec_register_framebuffer(struct vpu_instance *inst, struct frame_b
 		(color_format << 19) |
 		(nv21 << 17) |
 		(cbcr_interleave << 16) |
-		inst->display_fmt.width;
+		(scale_en ? inst->display_fmt.width : fb_arr[0].stride);
+		//inst->display_fmt.width;
 	dev_dbg(inst->dev->dev, "set W5_COMMON_PIC_INFO 0x%x\n",reg_val);
 	vpu_write_reg(inst->dev, W5_COMMON_PIC_INFO, reg_val);
 
diff --git a/drivers/media/platform/chips-media/wave5/wave5-vdi.c b/drivers/media/platform/chips-media/wave5/wave5-vdi.c
index ec8b6cb80361..e2b7dab60bed 100644
--- a/drivers/media/platform/chips-media/wave5/wave5-vdi.c
+++ b/drivers/media/platform/chips-media/wave5/wave5-vdi.c
@@ -10,7 +10,9 @@
 #include "wave5-vpu.h"
 #include "wave5-regdefine.h"
 #include <linux/delay.h>
+#ifdef CONFIG_SIFIVE_FLUSH
 #include <soc/sifive/sifive_l2_cache.h>
+#endif
 
 #define VDI_SRAM_BASE_ADDR		0x00
 
@@ -95,7 +97,9 @@ int wave5_vdi_clear_memory(struct vpu_device *vpu_dev, struct vpu_buf *vb)
 	}
 
 	memset(vb->vaddr, 0, vb->size);
-	sifive_l2_flush64_range(vb->daddr, vb->size);
+#ifdef CONFIG_SIFIVE_FLUSH
+	sifive_flush64_range(vb->daddr, vb->size);
+#endif
 	return vb->size;
 }
 
@@ -117,7 +121,9 @@ int wave5_vdi_write_memory(struct vpu_device *vpu_dev, struct vpu_buf *vb, size_
 
 	wave5_swap_endian(vpu_dev, data, len, endian);
 	memcpy(vb->vaddr + offset, data, len);
-	sifive_l2_flush64_range(vb->daddr + offset, len);
+#ifdef CONFIG_SIFIVE_FLUSH
+	sifive_flush64_range(vb->daddr + offset, len);
+#endif
 
 	return len;
 }
@@ -138,7 +144,9 @@ int wave5_vdi_allocate_dma_memory(struct vpu_device *vpu_dev, struct vpu_buf *vb
 	vb->vaddr = vaddr;
 	vb->daddr = daddr;
 
-	sifive_l2_flush64_range(daddr, vb->size);
+#ifdef CONFIG_SIFIVE_FLUSH
+	sifive_flush64_range(daddr, vb->size);
+#endif
 
 	return 0;
 }
diff --git a/drivers/media/platform/chips-media/wave5/wave5-vpu-dec.c b/drivers/media/platform/chips-media/wave5/wave5-vpu-dec.c
index bba48cbc8080..275db3e04c57 100644
--- a/drivers/media/platform/chips-media/wave5/wave5-vpu-dec.c
+++ b/drivers/media/platform/chips-media/wave5/wave5-vpu-dec.c
@@ -29,7 +29,7 @@ static const struct vpu_format dec_fmt_list[FMT_TYPES][MAX_FMTS] = {
 		},
 	},
 	[VPU_FMT_TYPE_RAW] = {
-		{
+		/*{
 			.v4l2_pix_fmt = V4L2_PIX_FMT_YUV420,
 			.max_width = 8192,
 			.min_width = 8,
@@ -49,7 +49,7 @@ static const struct vpu_format dec_fmt_list[FMT_TYPES][MAX_FMTS] = {
 			.min_width = 8,
 			.max_height = 4320,
 			.min_height = 8,
-		},
+		},*/
 		{
 			.v4l2_pix_fmt = V4L2_PIX_FMT_YUV420M,
 			.max_width = 8192,
@@ -202,6 +202,46 @@ static void wave5_update_pix_fmt(struct v4l2_pix_format_mplane *pix_mp, unsigned
 	case V4L2_PIX_FMT_NV12:
 	case V4L2_PIX_FMT_NV21:
 		pix_mp->width = round_up(width, 32);
+		pix_mp->height = round_up(height, 16);
+		pix_mp->plane_fmt[0].bytesperline = round_up(width, 32);
+		pix_mp->plane_fmt[0].sizeimage = width * height * 3 / 2;
+		break;
+	case V4L2_PIX_FMT_YUV420M:
+		pix_mp->width = round_up(width, 32);
+		pix_mp->height = round_up(height, 16);
+		pix_mp->plane_fmt[0].bytesperline = round_up(width, 32);
+		pix_mp->plane_fmt[0].sizeimage = width * height;
+		pix_mp->plane_fmt[1].bytesperline = round_up(width, 32) / 2;
+		pix_mp->plane_fmt[1].sizeimage = width * height / 4;
+		pix_mp->plane_fmt[2].bytesperline = round_up(width, 32) / 2;
+		pix_mp->plane_fmt[2].sizeimage = width * height / 4;
+		break;
+	case V4L2_PIX_FMT_NV12M:
+	case V4L2_PIX_FMT_NV21M:
+		pix_mp->width = round_up(width, 32);
+		pix_mp->height = round_up(height, 16);
+		pix_mp->plane_fmt[0].bytesperline = round_up(width, 32);
+		pix_mp->plane_fmt[0].sizeimage = width * height;
+		pix_mp->plane_fmt[1].bytesperline = round_up(width, 32);
+		pix_mp->plane_fmt[1].sizeimage = width * height / 2;
+		break;
+	default:
+		pix_mp->width = width;
+		pix_mp->height = height;
+		pix_mp->plane_fmt[0].bytesperline = 0;
+		pix_mp->plane_fmt[0].sizeimage = width * height;
+		break;
+	}
+}
+
+static void wave5_update_pix_fmt_r8(struct v4l2_pix_format_mplane *pix_mp, unsigned int width,
+				 unsigned int height)
+{
+	switch (pix_mp->pixelformat) {
+	case V4L2_PIX_FMT_YUV420:
+	case V4L2_PIX_FMT_NV12:
+	case V4L2_PIX_FMT_NV21:
+		pix_mp->width = round_up(width, 32);
 		pix_mp->height = round_up(height, 8);
 		pix_mp->plane_fmt[0].bytesperline = round_up(width, 32);
 		pix_mp->plane_fmt[0].sizeimage = width * height * 3 / 2;
@@ -261,6 +301,7 @@ static void wave5_vpu_dec_start_decode(struct vpu_instance *inst)
 
 		src_buf = v4l2_m2m_src_buf_remove(inst->v4l2_fh.m2m_ctx);
 		inst->state = VPU_INST_STATE_STOP;
+		//printk("%d wave5 state = %d\n",__LINE__, inst->state);
 		v4l2_m2m_buf_done(src_buf, VB2_BUF_STATE_ERROR);
 	}
 }
@@ -271,6 +312,7 @@ static void wave5_vpu_dec_stop_decode(struct vpu_instance *inst)
 	int ret;
 
 	inst->state = VPU_INST_STATE_STOP;
+	//printk("%d wave5 state = %d\n",__LINE__, inst->state);
 
 	ret = wave5_vpu_dec_update_bitstream_buffer(inst, 0);
 	if (ret) {
@@ -314,14 +356,12 @@ static void wave5_vpu_dec_finish_decode(struct vpu_instance *inst)
 			struct vb2_v4l2_buffer *dst_buf =
 				v4l2_m2m_dst_buf_remove_by_idx(inst->v4l2_fh.m2m_ctx,
 							       dec_output_info.index_frame_display);
-			int stride = dec_output_info.disp_frame.stride;
-			int height = dec_output_info.disp_pic_height -
-				dec_output_info.rc_display.bottom;
-			if (dec_output_info.disp_pic_height != inst->display_fmt.height)
-				height = inst->display_fmt.height;
+
+			int stride = inst->display_fmt.width;
+			int height  =inst->display_fmt.height;
 			dev_dbg(inst->dev->dev, "%s %d disp_pic_height %u rc_display.bottom %u\n",
 				__func__, __LINE__, dec_output_info.disp_pic_height, dec_output_info.rc_display.bottom);
-			dev_dbg(inst->dev->dev, "%s %d stride %u height %u\n", __func__, __LINE__, stride, height);
+			dev_dbg(inst->dev->dev, "%s %d stride %u height %u num %d\n", __func__, __LINE__, stride, height,inst->dst_fmt.num_planes);
 
 			if (inst->dst_fmt.num_planes == 1) {
 				vb2_set_plane_payload(&dst_buf->vb2_buf, 0,
@@ -340,7 +380,12 @@ static void wave5_vpu_dec_finish_decode(struct vpu_instance *inst)
 						      ((stride / 2) * (height / 2)));
 			}
 
-			dst_buf->vb2_buf.timestamp = inst->timestamp;
+			if (inst->timestamp) {
+				dst_buf->vb2_buf.timestamp = inst->timestamp;
+			} else {
+				dst_buf->vb2_buf.timestamp = inst->timestamp_cnt++ * inst->codec_info->dec_info.initial_info.ns_per_frame;
+			}
+
 			dst_buf->field = V4L2_FIELD_NONE;
 			v4l2_m2m_buf_done(dst_buf, VB2_BUF_STATE_DONE);
 
@@ -380,7 +425,7 @@ static void wave5_vpu_dec_finish_decode(struct vpu_instance *inst)
 			v4l2_m2m_buf_done(dst_buf, VB2_BUF_STATE_DONE);
 
 			inst->eos = TRUE;
-			pr_err("wave5 queue event type: %d id: %d\n",vpu_event_eos.type, vpu_event_eos.id);
+			//pr_err("wave5 queue event type: %d id: %d\n",vpu_event_eos.type, vpu_event_eos.id);
 			v4l2_event_queue_fh(&inst->v4l2_fh, &vpu_event_eos);
 
 			v4l2_m2m_job_finish(inst->v4l2_m2m_dev, inst->v4l2_fh.m2m_ctx);
@@ -453,7 +498,7 @@ static int wave5_vpu_dec_try_fmt_cap(struct file *file, void *fh, struct v4l2_fo
 	if (!vpu_fmt) {
 		f->fmt.pix_mp.pixelformat = inst->dst_fmt.pixelformat;
 		f->fmt.pix_mp.num_planes = inst->dst_fmt.num_planes;
-		wave5_update_pix_fmt(&f->fmt.pix_mp, inst->dst_fmt.width, inst->dst_fmt.height);
+		wave5_update_pix_fmt_r8(&f->fmt.pix_mp, inst->dst_fmt.width, inst->dst_fmt.height);
 	} else {
 		int width = clamp(f->fmt.pix_mp.width, vpu_fmt->min_width, vpu_fmt->max_width);
 		int height = clamp(f->fmt.pix_mp.height, vpu_fmt->min_height, vpu_fmt->max_height);
@@ -461,7 +506,7 @@ static int wave5_vpu_dec_try_fmt_cap(struct file *file, void *fh, struct v4l2_fo
 
 		f->fmt.pix_mp.pixelformat = vpu_fmt->v4l2_pix_fmt;
 		f->fmt.pix_mp.num_planes = info->mem_planes;
-		wave5_update_pix_fmt(&f->fmt.pix_mp, width, height);
+		wave5_update_pix_fmt_r8(&f->fmt.pix_mp, width, height);
 	}
 
 	f->fmt.pix_mp.flags = 0;
@@ -482,7 +527,7 @@ static int wave5_vpu_dec_s_fmt_cap(struct file *file, void *fh, struct v4l2_form
 	int i, ret;
 	unsigned int scalew, scaleh;
 
-	printk(
+	dev_dbg(inst->dev->dev,
 		"%s: fourcc: %u width: %u height: %u num_planes: %u colorspace: %u field: %u\n",
 		__func__, f->fmt.pix_mp.pixelformat, f->fmt.pix_mp.width, f->fmt.pix_mp.height,
 		f->fmt.pix_mp.num_planes, f->fmt.pix_mp.colorspace, f->fmt.pix_mp.field);
@@ -495,28 +540,30 @@ static int wave5_vpu_dec_s_fmt_cap(struct file *file, void *fh, struct v4l2_form
 	scalew = inst->src_fmt.width / f->fmt.pix_mp.width;
 	scaleh = inst->src_fmt.height / f->fmt.pix_mp.height;
 
-	if (scalew > 8 || scaleh > 8 || scalew < 1 || scaleh < 1) {
-		dev_err(inst->dev->dev,"Scaling should be 1 to 1/8 (down-scaling only)! Use input parameter. \n");
-		return -EINVAL;
-	}
-
-	inst->dst_fmt.width = f->fmt.pix_mp.width;
-	inst->dst_fmt.height = f->fmt.pix_mp.height;
+	//if (scalew > 8 || scaleh > 8 || scalew < 1 || scaleh < 1) {
+	//	dev_err(inst->dev->dev,"Scaling should be 1 to 1/8 (down-scaling only)! Use input parameter. \n");
+	//	return -EINVAL;
+	//}
+
+	inst->display_fmt.width = f->fmt.pix_mp.width;
+	inst->display_fmt.height = f->fmt.pix_mp.height;
+	inst->display_fmt.pixelformat = f->fmt.pix_mp.pixelformat;
+	inst->display_fmt.field = f->fmt.pix_mp.field;
+	inst->display_fmt.flags = f->fmt.pix_mp.flags;
+	inst->display_fmt.num_planes = f->fmt.pix_mp.num_planes;
 	inst->dst_fmt.pixelformat = f->fmt.pix_mp.pixelformat;
-	inst->dst_fmt.field = f->fmt.pix_mp.field;
-	inst->dst_fmt.flags = f->fmt.pix_mp.flags;
 	inst->dst_fmt.num_planes = f->fmt.pix_mp.num_planes;
-	for (i = 0; i < inst->dst_fmt.num_planes; i++) {
-		inst->dst_fmt.plane_fmt[i].bytesperline = f->fmt.pix_mp.plane_fmt[i].bytesperline;
-		inst->dst_fmt.plane_fmt[i].sizeimage = f->fmt.pix_mp.plane_fmt[i].sizeimage;
+	for (i = 0; i < inst->display_fmt.num_planes; i++) {
+		inst->display_fmt.plane_fmt[i].bytesperline = f->fmt.pix_mp.plane_fmt[i].bytesperline;
+		inst->display_fmt.plane_fmt[i].sizeimage = f->fmt.pix_mp.plane_fmt[i].sizeimage;
 	}
 
-	if (inst->dst_fmt.pixelformat == V4L2_PIX_FMT_NV12 ||
-	    inst->dst_fmt.pixelformat == V4L2_PIX_FMT_NV12M) {
+	if (inst->display_fmt.pixelformat == V4L2_PIX_FMT_NV12 ||
+	    inst->display_fmt.pixelformat == V4L2_PIX_FMT_NV12M) {
 		inst->cbcr_interleave = true;
 		inst->nv21 = false;
-	} else if (inst->dst_fmt.pixelformat == V4L2_PIX_FMT_NV21 ||
-		   inst->dst_fmt.pixelformat == V4L2_PIX_FMT_NV21M) {
+	} else if (inst->display_fmt.pixelformat == V4L2_PIX_FMT_NV21 ||
+		   inst->display_fmt.pixelformat == V4L2_PIX_FMT_NV21M) {
 		inst->cbcr_interleave = true;
 		inst->nv21 = true;
 	} else {
@@ -524,8 +571,6 @@ static int wave5_vpu_dec_s_fmt_cap(struct file *file, void *fh, struct v4l2_form
 		inst->nv21 = false;
 	}
 
-	memcpy((void *)&inst->display_fmt, (void *)&inst->dst_fmt, sizeof(struct v4l2_pix_format_mplane));
-
 	return 0;
 }
 
@@ -610,11 +655,12 @@ static int wave5_vpu_dec_s_fmt_out(struct file *file, void *fh, struct v4l2_form
 	struct vpu_instance *inst = wave5_to_vpu_inst(fh);
 	int i, ret;
 
-	printk(
+	dev_dbg(inst->dev->dev,
 		"%s: fourcc: %u width: %u height: %u num_planes: %u field: %u\n",
 		__func__, f->fmt.pix_mp.pixelformat, f->fmt.pix_mp.width, f->fmt.pix_mp.height,
 		f->fmt.pix_mp.num_planes, f->fmt.pix_mp.field);
 
+
 	ret = wave5_vpu_dec_try_fmt_out(file, fh, f);
 	if (ret)
 		return ret;
@@ -636,7 +682,9 @@ static int wave5_vpu_dec_s_fmt_out(struct file *file, void *fh, struct v4l2_form
 	inst->quantization = f->fmt.pix_mp.quantization;
 	inst->xfer_func = f->fmt.pix_mp.xfer_func;
 
-	wave5_update_pix_fmt(&inst->dst_fmt, f->fmt.pix_mp.width, f->fmt.pix_mp.height);
+
+	wave5_update_pix_fmt_r8(&inst->dst_fmt, f->fmt.pix_mp.width, f->fmt.pix_mp.height);
+	wave5_update_pix_fmt_r8(&inst->display_fmt, f->fmt.pix_mp.width, f->fmt.pix_mp.height);
 
 	return 0;
 }
@@ -713,7 +761,10 @@ static int wave5_vpu_dec_decoder_cmd(struct file *file, void *fh, struct v4l2_de
 
 	switch (dc->cmd) {
 	case V4L2_DEC_CMD_STOP:
+		wave5_handle_bitstream_buffer(inst);
+		inst->ops->start_process(inst);
 		inst->state = VPU_INST_STATE_STOP;
+		//printk("%d wave5 state = %d\n",__LINE__, inst->state);
 
 		ret = wave5_vpu_dec_update_bitstream_buffer(inst, 0);
 		if (ret) {
@@ -854,6 +905,8 @@ static int wave5_vpu_dec_queue_setup(struct vb2_queue *q, unsigned int *num_buff
 	if (inst->state == VPU_INST_STATE_NONE && q->type == V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE) {
 		struct dec_open_param open_param;
 
+		*num_buffers = 4;
+
 		memset(&open_param, 0, sizeof(struct dec_open_param));
 		wave5_set_default_dec_openparam(&open_param);
 
@@ -883,6 +936,7 @@ static int wave5_vpu_dec_queue_setup(struct vb2_queue *q, unsigned int *num_buff
 		}
 
 		inst->state = VPU_INST_STATE_OPEN;
+		//printk("wave5 state = %d\n",inst->state);
 
 		if (inst->thumbnail_mode)
 			wave5_vpu_dec_give_command(inst, ENABLE_DEC_THUMBNAIL_MODE, NULL);
@@ -893,9 +947,9 @@ static int wave5_vpu_dec_queue_setup(struct vb2_queue *q, unsigned int *num_buff
 		u32 fb_stride, fb_height;
 		u32 luma_size, chroma_size;
 
-		if (*num_buffers > inst->min_dst_buf_count &&
-		    *num_buffers < WAVE5_MAX_FBS)
-			inst->dst_buf_count = *num_buffers;
+		//if (*num_buffers > inst->min_dst_buf_count &&
+		//    *num_buffers < WAVE5_MAX_FBS)
+		//	inst->dst_buf_count = *num_buffers;
 
 		*num_buffers = inst->dst_buf_count;
 		non_linear_num = inst->dst_buf_count;
@@ -951,17 +1005,17 @@ static int wave5_vpu_dec_start_streaming_open(struct vpu_instance *inst)
 
 	ret = wave5_vpu_dec_issue_seq_init(inst);
 	if (ret) {
-		dev_dbg(inst->dev->dev, "%s: wave5_vpu_dec_issue_seq_init, fail: %d\n",
+		dev_err(inst->dev->dev, "%s: wave5_vpu_dec_issue_seq_init, fail: %d\n",
 			__func__, ret);
 		return ret;
 	}
 
 	if (wave5_vpu_wait_interrupt(inst, VPU_DEC_TIMEOUT) < 0)
-		dev_dbg(inst->dev->dev, "%s: failed to call vpu_wait_interrupt()\n", __func__);
+		dev_err(inst->dev->dev, "%s: failed to call vpu_wait_interrupt()\n", __func__);
 
 	ret = wave5_vpu_dec_complete_seq_init(inst, &initial_info);
 	if (ret) {
-		dev_dbg(inst->dev->dev, "%s: vpu_dec_complete_seq_init, fail: %d, reason: %u\n",
+		dev_err(inst->dev->dev, "%s: vpu_dec_complete_seq_init, fail: %d, reason: %u\n",
 			__func__, ret, initial_info.seq_init_err_reason);
 	} else {
 		static const struct v4l2_event vpu_event_src_ch = {
@@ -975,6 +1029,7 @@ static int wave5_vpu_dec_start_streaming_open(struct vpu_instance *inst)
 			initial_info.profile, initial_info.min_frame_buffer_count);
 
 		inst->state = VPU_INST_STATE_INIT_SEQ;
+		//printk("wave5 state = %d\n",inst->state);
 		inst->min_dst_buf_count = initial_info.min_frame_buffer_count + 1;
 		inst->dst_buf_count = inst->min_dst_buf_count;
 
@@ -988,11 +1043,20 @@ static int wave5_vpu_dec_start_streaming_open(struct vpu_instance *inst)
 
 		if (initial_info.pic_width != inst->src_fmt.width ||
 		    initial_info.pic_height != inst->src_fmt.height) {
-			wave5_update_pix_fmt(&inst->src_fmt, initial_info.pic_width,
-					     initial_info.pic_height);
-			wave5_update_pix_fmt(&inst->dst_fmt, initial_info.pic_width,
-					     initial_info.pic_height);
+			if (inst->std == W_AVC_DEC) {
+				wave5_update_pix_fmt(&inst->src_fmt, initial_info.pic_width,
+							initial_info.pic_height);
+				wave5_update_pix_fmt(&inst->dst_fmt, initial_info.pic_width,
+							initial_info.pic_height);
+			} else { //HEVC
+				wave5_update_pix_fmt_r8(&inst->src_fmt, initial_info.pic_width,
+							initial_info.pic_height);
+				wave5_update_pix_fmt_r8(&inst->dst_fmt, initial_info.pic_width,
+							initial_info.pic_height);
+			}
 		}
+		inst->crop_rect.right = initial_info.pic_crop_rect.right;
+		inst->crop_rect.bottom = initial_info.pic_crop_rect.bottom;
 
 		scalew = inst->dst_fmt.width / inst->display_fmt.width;
 		scaleh = inst->dst_fmt.height / inst->display_fmt.height;
@@ -1002,7 +1066,7 @@ static int wave5_vpu_dec_start_streaming_open(struct vpu_instance *inst)
 						inst->dst_fmt.height);
 		}
 
-		printk("wave5 queue event type: %d id: %d\n",vpu_event_src_ch.type, vpu_event_src_ch.id);
+		dev_dbg(inst->dev->dev, "wave5 queue event type: %d id: %d\n",vpu_event_src_ch.type, vpu_event_src_ch.id);
 		v4l2_event_queue_fh(&inst->v4l2_fh, &vpu_event_src_ch);
 
 		wave5_handle_src_buffer(inst);
@@ -1028,6 +1092,7 @@ static int wave5_vpu_dec_start_streaming_seek(struct vpu_instance *inst)
 
 		src_buf = v4l2_m2m_src_buf_remove(inst->v4l2_fh.m2m_ctx);
 		inst->state = VPU_INST_STATE_STOP;
+		//printk("%d wave5 state = %d\n",__LINE__, inst->state);
 		v4l2_m2m_buf_done(src_buf, VB2_BUF_STATE_ERROR);
 		dev_dbg(inst->dev->dev, "%s: wave5_vpu_dec_start_one_frame\n", __func__);
 		return ret;
@@ -1070,11 +1135,20 @@ static int wave5_vpu_dec_start_streaming_seek(struct vpu_instance *inst)
 
 		if (initial_info.pic_width != inst->src_fmt.width ||
 		    initial_info.pic_height != inst->src_fmt.height) {
-			wave5_update_pix_fmt(&inst->src_fmt, initial_info.pic_width,
-					     initial_info.pic_height);
-			wave5_update_pix_fmt(&inst->dst_fmt, initial_info.pic_width,
-					     initial_info.pic_height);
+			if (inst->std == W_AVC_DEC) {
+				wave5_update_pix_fmt(&inst->src_fmt, initial_info.pic_width,
+							initial_info.pic_height);
+				wave5_update_pix_fmt(&inst->dst_fmt, initial_info.pic_width,
+							initial_info.pic_height);
+			} else { //HEVC
+				wave5_update_pix_fmt_r8(&inst->src_fmt, initial_info.pic_width,
+							initial_info.pic_height);
+				wave5_update_pix_fmt_r8(&inst->dst_fmt, initial_info.pic_width,
+							initial_info.pic_height);
+			}
 		}
+		inst->crop_rect.right = initial_info.pic_crop_rect.right;
+		inst->crop_rect.bottom = initial_info.pic_crop_rect.bottom;
 
 		scalew = inst->dst_fmt.width / inst->display_fmt.width;
 		scaleh = inst->dst_fmt.height / inst->display_fmt.height;
@@ -1172,8 +1246,8 @@ static void wave5_vpu_dec_buf_queue(struct vb2_buffer *vb)
 	struct vpu_instance *inst = vb2_get_drv_priv(vb->vb2_queue);
 
 	dev_dbg(inst->dev->dev, "%s: type: %4u index: %4u size: ([0]=%4lu, [1]=%4lu, [2]=%4lu)\n",
-		__func__, vb->type, vb->index, vb2_plane_size(&vbuf->vb2_buf, 0),
-		vb2_plane_size(&vbuf->vb2_buf, 1), vb2_plane_size(&vbuf->vb2_buf, 2));
+		__func__, vb->type, vb->index, vb2_get_plane_payload(&vbuf->vb2_buf, 0),
+		vb2_get_plane_payload(&vbuf->vb2_buf, 1), vb2_get_plane_payload(&vbuf->vb2_buf, 2));
 
 	v4l2_m2m_buf_queue(inst->v4l2_fh.m2m_ctx, vbuf);
 
@@ -1215,6 +1289,7 @@ static void wave5_vpu_dec_stop_streaming(struct vb2_queue *q)
 {
 	struct vpu_instance *inst = vb2_get_drv_priv(q);
 	struct vb2_v4l2_buffer *buf;
+	int try_cnt = 0;
 	bool check_cmd = TRUE;
 
 	dev_dbg(inst->dev->dev, "%s: type: %u\n", __func__, q->type);
@@ -1222,7 +1297,6 @@ static void wave5_vpu_dec_stop_streaming(struct vb2_queue *q)
 	while (check_cmd) {
 		struct queue_status_info q_status;
 		struct dec_output_info dec_output_info;
-		int try_cnt = 0;
 
 		wave5_vpu_dec_give_command(inst, DEC_GET_QUEUE_STATUS, &q_status);
 
@@ -1231,7 +1305,7 @@ static void wave5_vpu_dec_stop_streaming(struct vb2_queue *q)
 
 		if (wave5_vpu_wait_interrupt(inst, 600) < 0){
 			try_cnt++;
-			if (try_cnt >= 100)
+			if (try_cnt >= 10)
 				break;
 			continue;
 		}
@@ -1288,6 +1362,7 @@ static void wave5_vpu_dec_stop_streaming(struct vb2_queue *q)
 		if (inst->eos) {
 			inst->eos = FALSE;
 			inst->state = VPU_INST_STATE_INIT_SEQ;
+			//printk("wave5 state = %d\n",inst->state);
 		}
 		inst->queued_dst_buf_num = 0;
 	}
@@ -1336,9 +1411,11 @@ static void wave5_vpu_dec_device_run(void *priv)
 {
 	struct vpu_instance *inst = priv;
 
+	wave5_handle_bitstream_buffer(inst);
 	inst->ops->start_process(inst);
 
 	inst->state = VPU_INST_STATE_PIC_RUN;
+	//printk("wave5 state = %d\n",inst->state);
 }
 
 static void wave5_vpu_dec_job_abort(void *priv)
diff --git a/drivers/media/platform/chips-media/wave5/wave5-vpuapi.h b/drivers/media/platform/chips-media/wave5/wave5-vpuapi.h
index 0a4dc20ddec3..dcd061b1aa9c 100644
--- a/drivers/media/platform/chips-media/wave5/wave5-vpuapi.h
+++ b/drivers/media/platform/chips-media/wave5/wave5-vpuapi.h
@@ -473,6 +473,7 @@ struct dec_initial_info {
 	u32 pic_height;
 	s32 f_rate_numerator; /* the numerator part of frame rate fraction */
 	s32 f_rate_denominator; /* the denominator part of frame rate fraction */
+	u64 ns_per_frame;
 	struct vpu_rect pic_crop_rect;
 	u32 min_frame_buffer_count; /* between 1 to 16 */
 	u32 frame_buf_delay;
@@ -1057,6 +1058,7 @@ struct vpu_instance {
 	enum vpu_instance_state state;
 	enum vpu_instance_type type;
 	const struct vpu_instance_ops *ops;
+	struct vpu_rect crop_rect;
 
 	enum wave_std		 std;
 	s32			 id;
@@ -1073,6 +1075,7 @@ struct vpu_instance {
 	u32 conf_win_width;
 	u32 conf_win_height;
 	u64 timestamp;
+	u64 timestamp_cnt;
 	bool cbcr_interleave;
 	bool nv21;
 	bool eos;