msm: vidc: improve heif decoder performance

Add V4L2 interface to indicate heif decoder, and use
below configurations to improve performance.
1. Set turbo session
2. Disable DCVS and batch
3. Increase input and output buffer count to 12
4. Modify input buffer size calculation
5. Avoid framerate calc and ts list maintenance
6. Skip setting bse-vpp delay

Change-Id: I9a47a88154238db3ac4285462c6030114546d333
Signed-off-by: Qiwei Liu <qiweil@codeaurora.org>

include/uapi/vidc/media/msm_vidc_utils.h

@@ -252,6 +252,8 @@ enum v4l2_mpeg_vidc_video_bitrate_savings_type {
 		(V4L2_CID_MPEG_MSM_VIDC_BASE + 132)
 #define V4L2_CID_MPEG_VIDC_VIDEO_LOWLATENCY_HINT \
 	(V4L2_CID_MPEG_MSM_VIDC_BASE + 133)
+#define V4L2_CID_MPEG_VIDC_VDEC_HEIF_MODE \
+		(V4L2_CID_MPEG_MSM_VIDC_BASE + 134)
 
 #define V4L2_CID_MPEG_VIDC_VIDEO_UNKNOWN \
 		(V4L2_CID_MPEG_MSM_VIDC_BASE + 0xFFF)

msm/vidc/msm_vdec.c

@@ -434,6 +434,15 @@ static struct msm_vidc_ctrl msm_vdec_ctrls[] = {
 		.default_value = 0,
 		.step = 1,
 	},
+	{
+		.id = V4L2_CID_MPEG_VIDC_VDEC_HEIF_MODE,
+		.name = "HEIF Decoder",
+		.type = V4L2_CTRL_TYPE_BOOLEAN,
+		.minimum = V4L2_MPEG_MSM_VIDC_DISABLE,
+		.maximum = V4L2_MPEG_MSM_VIDC_ENABLE,
+		.default_value = V4L2_MPEG_MSM_VIDC_DISABLE,
+		.step = 1,
+	},
 };
 
 #define NUM_CTRLS ARRAY_SIZE(msm_vdec_ctrls)
@@ -948,6 +957,15 @@ int msm_vdec_s_ctrl(struct msm_vidc_inst *inst, struct v4l2_ctrl *ctrl)
 		break;
 	case V4L2_CID_MPEG_VIDC_VIDEO_LOWLATENCY_HINT:
 		break;
+	case V4L2_CID_MPEG_VIDC_VDEC_HEIF_MODE:
+		if(get_v4l2_codec(inst) != V4L2_PIX_FMT_HEVC)
+			break;
+		inst->flags &= ~VIDC_TURBO;
+		if (ctrl->val)
+			inst->flags |= VIDC_TURBO;
+		if (inst->state < MSM_VIDC_LOAD_RESOURCES)
+			msm_vidc_calculate_buffer_counts(inst);
+		break;
 	default:
 		s_vpr_e(inst->sid, "Unknown control %#x\n", ctrl->id);
 		break;

msm/vidc/msm_vidc.c

@@ -432,7 +432,8 @@ int msm_vidc_qbuf(void *instance, struct media_device *mdev,
 
 	timestamp_us = (s64)((b->timestamp.tv_sec * 1000000) +
 		b->timestamp.tv_usec);
-	if (is_decode_session(inst) && b->type == INPUT_MPLANE) {
+	if (is_decode_session(inst) && b->type == INPUT_MPLANE &&
+		!is_heif_decoder(inst)) {
 		if (inst->flush_timestamps)
 			msm_comm_release_timestamps(inst);
 		inst->flush_timestamps = false;
@@ -513,7 +514,8 @@ int msm_vidc_dqbuf(void *instance, struct v4l2_buffer *b)
 	}
 	if (is_decode_session(inst) &&
 		b->type == OUTPUT_MPLANE &&
-		!(b->flags & V4L2_BUF_FLAG_CODECCONFIG))
+		!(b->flags & V4L2_BUF_FLAG_CODECCONFIG) &&
+		!is_heif_decoder(inst))
 		msm_comm_fetch_ts_framerate(inst, b);
 
 	return rc;

msm/vidc/msm_vidc_buffer_calculations.c

@@ -27,6 +27,10 @@
 /* extra output buffers for encoder HEIF usecase */
 #define HEIF_ENC_TOTAL_OUTPUT_BUFFERS 12
 
+/* extra buffer count for heif decoder */
+#define HEIF_DEC_TOTAL_INPUT_BUFFERS 12
+#define HEIF_DEC_EXTRA_OUTPUT_BUFFERS 8
+
 #define HFI_COLOR_FORMAT_YUV420_NV12_UBWC_Y_TILE_WIDTH 32
 #define HFI_COLOR_FORMAT_YUV420_NV12_UBWC_Y_TILE_HEIGHT 8
 #define HFI_COLOR_FORMAT_YUV420_NV12_UBWC_UV_TILE_WIDTH 16
@@ -789,6 +793,9 @@ static int msm_vidc_get_extra_input_buff_count(struct msm_vidc_inst *inst)
 
 	core = inst->core;
 
+	if (is_heif_decoder(inst))
+		return (HEIF_DEC_TOTAL_INPUT_BUFFERS - MIN_INPUT_BUFFERS);
+
 	/*
 	 * For thumbnail session, extra buffers are not required as
 	 * neither dcvs nor batching will be enabled.
@@ -821,6 +828,9 @@ static int msm_vidc_get_extra_output_buff_count(struct msm_vidc_inst *inst)
 
 	core = inst->core;
 
+	if (is_heif_decoder(inst))
+		return HEIF_DEC_EXTRA_OUTPUT_BUFFERS;
+
 	/*
 	 * For a non-realtime session, extra buffers are not required.
 	 * For thumbnail session, extra buffers are not required as
@@ -829,7 +839,7 @@ static int msm_vidc_get_extra_output_buff_count(struct msm_vidc_inst *inst)
 	if (!is_realtime_session(inst) || is_thumbnail_session(inst))
 		return extra_output_count;
 
-	/* For HEIF, we are increasing buffer count */
+	/* For HEIF encoder, we are increasing buffer count */
 	if (is_image_session(inst)) {
 		extra_output_count = (HEIF_ENC_TOTAL_OUTPUT_BUFFERS -
 			MIN_ENC_OUTPUT_BUFFERS);
@@ -889,6 +899,11 @@ u32 msm_vidc_calculate_dec_input_frame_size(struct msm_vidc_inst *inst)
 		base_res_mbs = inst->capability.cap[CAP_MBS_PER_FRAME].max;
 		div_factor = 1;
 	}
+	/* For HEIF image, use the actual resolution to calc buffer size */
+	if (is_heif_decoder(inst)) {
+		base_res_mbs = num_mbs;
+		div_factor = 1;
+	}
 
 	frame_size = base_res_mbs * MB_SIZE_IN_PIXEL * 3 / 2 / div_factor;
 

msm/vidc/msm_vidc_clocks.c

@@ -1293,6 +1293,7 @@ int msm_vidc_set_bse_vpp_delay(struct msm_vidc_inst *inst)
 	if (!inst->core->resources.has_vpp_delay ||
 		!is_decode_session(inst) ||
 		is_thumbnail_session(inst) ||
+		is_heif_decoder(inst) ||
 		inst->clk_data.work_mode != HFI_WORKMODE_2) {
 		s_vpr_hp(inst->sid, "%s: Skip bse-vpp\n", __func__);
 		return 0;

msm/vidc/msm_vidc_common.c

@@ -2839,6 +2839,7 @@ bool is_batching_allowed(struct msm_vidc_inst *inst)
 		is_decode_session(inst) &&
 		!is_thumbnail_session(inst) &&
 		is_realtime_session(inst) &&
+		!is_heif_decoder(inst) &&
 		!inst->clk_data.low_latency_mode &&
 		(op_pixelformat == V4L2_PIX_FMT_NV12_UBWC ||
 		 op_pixelformat	== V4L2_PIX_FMT_NV12_TP10_UBWC) &&

msm/vidc/msm_vidc_common.h

@@ -121,6 +121,17 @@ static inline bool is_grid_session(struct msm_vidc_inst *inst)
 	return 0;
 }
 
+static inline bool is_heif_decoder(struct msm_vidc_inst *inst)
+{
+	struct v4l2_ctrl *ctrl = NULL;
+	if (inst->session_type == MSM_VIDC_DECODER &&
+		get_v4l2_codec(inst) == V4L2_PIX_FMT_HEVC) {
+		ctrl = get_ctrl(inst, V4L2_CID_MPEG_VIDC_VDEC_HEIF_MODE);
+		return (ctrl->val > 0);
+	}
+	return 0;
+}
+
 static inline bool is_video_session(struct msm_vidc_inst *inst)
 {
 	return !is_grid_session(inst);