[PATCH v6 1/2] media: amphion: Reduce decoding latency for HEVC decoder

[PATCH v6 1/2] media: amphion: Reduce decoding latency for HEVC decoder

[ming.qian]

From: Ming Qian <ming.qian@oss.nxp.com>

The amphion decoder firmware supports a low latency flush mode for the
HEVC format since v1.9.0. This feature, which is enabled when the
display delay is set to 0, can help to reduce the decoding latency by
appending some padding data to every frame.

Signed-off-by: Ming Qian <ming.qian@oss.nxp.com>
Reviewed-by: Nicolas Dufresne <nicolas.dufresne@collabora.com>
---
v6
- include bitfield.h
v5
- Apply FIELD_PREP() and FIELD_GET() in CHECK_VERSION
v4
- Add CHECK_VERSION macro
v3
- Improve commit message as recommended
v2
- Improve commit message
- Add firmware version check

 drivers/media/platform/amphion/vpu_malone.c | 25 ++++++++++++++++++---
 1 file changed, 22 insertions(+), 3 deletions(-)

diff --git a/drivers/media/platform/amphion/vpu_malone.c b/drivers/media/platform/amphion/vpu_malone.c
index 5c6b2a841b6f..112645b30e15 100644
--- a/drivers/media/platform/amphion/vpu_malone.c
+++ b/drivers/media/platform/amphion/vpu_malone.c
@@ -3,6 +3,7 @@
  * Copyright 2020-2021 NXP
  */
 
+#include <linux/bitfield.h>
 #include <linux/init.h>
 #include <linux/interconnect.h>
 #include <linux/ioctl.h>
@@ -68,6 +69,12 @@
 
 #define MALONE_DEC_FMT_RV_MASK			BIT(21)
 
+#define MALONE_VERSION_MASK			0xFFFFF
+#define MALONE_VERSION(maj, min, inc)		\
+		(FIELD_PREP(0xF0000, maj) | FIELD_PREP(0xFF00, min) | FIELD_PREP(0xFF, inc))
+#define CHECK_VERSION(iface, maj, min)		\
+		(FIELD_GET(MALONE_VERSION_MASK, (iface)->fw_version) >= MALONE_VERSION(maj, min, 0))
+
 enum vpu_malone_stream_input_mode {
 	INVALID_MODE = 0,
 	FRAME_LVL,
@@ -332,6 +339,8 @@ struct vpu_dec_ctrl {
 	u32 buf_addr[VID_API_NUM_STREAMS];
 };
 
+static const struct malone_padding_scode *get_padding_scode(u32 type, u32 fmt);
+
 u32 vpu_malone_get_data_size(void)
 {
 	return sizeof(struct vpu_dec_ctrl);
@@ -654,9 +663,15 @@ static int vpu_malone_set_params(struct vpu_shared_addr *shared,
 		hc->jpg[instance].jpg_mjpeg_interlaced = 0;
 	}
 
-	hc->codec_param[instance].disp_imm = params->display_delay_enable ? 1 : 0;
-	if (malone_format != MALONE_FMT_AVC)
+	if (params->display_delay_enable &&
+	    get_padding_scode(SCODE_PADDING_BUFFLUSH, params->codec_format))
+		hc->codec_param[instance].disp_imm = 1;
+	else
 		hc->codec_param[instance].disp_imm = 0;
+
+	if (params->codec_format == V4L2_PIX_FMT_HEVC && !CHECK_VERSION(iface, 1, 9))
+		hc->codec_param[instance].disp_imm = 0;
+
 	hc->codec_param[instance].dbglog_enable = 0;
 	iface->dbglog_desc.level = 0;
 
@@ -1024,6 +1039,7 @@ static const struct malone_padding_scode padding_scodes[] = {
 	{SCODE_PADDING_EOS,      V4L2_PIX_FMT_JPEG,        {0x0, 0x0}},
 	{SCODE_PADDING_BUFFLUSH, V4L2_PIX_FMT_H264,        {0x15010000, 0x0}},
 	{SCODE_PADDING_BUFFLUSH, V4L2_PIX_FMT_H264_MVC,    {0x15010000, 0x0}},
+	{SCODE_PADDING_BUFFLUSH, V4L2_PIX_FMT_HEVC,        {0x3e010000, 0x20}},
 };
 
 static const struct malone_padding_scode padding_scode_dft = {0x0, 0x0};
@@ -1058,8 +1074,11 @@ static int vpu_malone_add_padding_scode(struct vpu_buffer *stream_buffer,
 	int ret;
 
 	ps = get_padding_scode(scode_type, pixelformat);
-	if (!ps)
+	if (!ps) {
+		if (scode_type == SCODE_PADDING_BUFFLUSH)
+			return 0;
 		return -EINVAL;
+	}
 
 	wptr = readl(&str_buf->wptr);
 	if (wptr < stream_buffer->phys || wptr > stream_buffer->phys + stream_buffer->length)
-- 
2.43.0-rc1

[PATCH v6 2/2] media: amphion: Add a frame flush mode for decoder

[ming.qian]

From: Ming Qian <ming.qian@oss.nxp.com>

By default the amphion decoder will pre-parse 3 frames before starting
to decode the first frame. Alternatively, a block of flush padding data
can be appended to the frame, which will ensure that the decoder can
start decoding immediately after parsing the flush padding data, thus
potentially reducing decoding latency.

This mode was previously only enabled, when the display delay was set to
0. Allow the user to manually toggle the use of that mode via a module
parameter called low_latency, which enables the mode without
changing the display order.

Signed-off-by: Ming Qian <ming.qian@oss.nxp.com>
Reviewed-by: Nicolas Dufresne <nicolas.dufresne@collabora.com>
---
v4
- Improve the comment expressing
v3
- Improve commit message as recommended
- Add some comments to avoid code looks cryptic

 drivers/media/platform/amphion/vpu_malone.c | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)

diff --git a/drivers/media/platform/amphion/vpu_malone.c b/drivers/media/platform/amphion/vpu_malone.c
index 112645b30e15..1b59c77a84db 100644
--- a/drivers/media/platform/amphion/vpu_malone.c
+++ b/drivers/media/platform/amphion/vpu_malone.c
@@ -26,6 +26,10 @@
 #include "vpu_imx8q.h"
 #include "vpu_malone.h"
 
+static bool low_latency;
+module_param(low_latency, bool, 0644);
+MODULE_PARM_DESC(low_latency, "Set low latency frame flush mode: 0 (disable) or 1 (enable)");
+
 #define CMD_SIZE			25600
 #define MSG_SIZE			25600
 #define CODEC_SIZE			0x1000
@@ -1582,7 +1586,15 @@ static int vpu_malone_input_frame_data(struct vpu_malone_str_buffer __iomem *str
 
 	vpu_malone_update_wptr(str_buf, wptr);
 
-	if (disp_imm && !vpu_vb_is_codecconfig(vbuf)) {
+	/*
+	 * Enable the low latency flush mode if display delay is set to 0
+	 * or the low latency frame flush mode if it is set to 1.
+	 * The low latency flush mode requires some padding data to be appended to each frame,
+	 * but there must not be any padding data between the sequence header and the frame.
+	 * This module is currently only supported for the H264 and HEVC formats,
+	 * for other formats, vpu_malone_add_scode() will return 0.
+	 */
+	if ((disp_imm || low_latency) && !vpu_vb_is_codecconfig(vbuf)) {
 		ret = vpu_malone_add_scode(inst->core->iface,
 					   inst->id,
 					   &inst->stream_buffer,
-- 
2.43.0-rc1