media: platform: mtk-mdp3: add support for parallel pipe to improve FPS

In some chips, MDP3 has the ability to utilize two pipelines to
parallelly process a single frame.
To enable this feature, multiple CMDQ clients and packets need to
be configured at the same time.

Signed-off-by: Moudy Ho <moudy.ho@mediatek.com>
Reviewed-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
Signed-off-by: Sebastian Fricke <sebastian.fricke@collabora.com>
Signed-off-by: Mauro Carvalho Chehab <mchehab@kernel.org>
This commit is contained in:
Moudy Ho 2023-12-20 11:18:38 +01:00 committed by Mauro Carvalho Chehab
parent 0e9bd2fcda
commit 9288eae430
9 changed files with 208 additions and 56 deletions

View File

@ -1060,6 +1060,11 @@ static const struct mdp_pipe_info mt8195_pipe_info[] = {
[MDP_PIPE_VPP0_SOUT] = {MDP_PIPE_VPP0_SOUT, MDP_MM_SUBSYS_1, 5},
};
static const struct v4l2_rect mt8195_mdp_pp_criteria = {
.width = 1920,
.height = 1080,
};
const struct mtk_mdp_driver_data mt8183_mdp_driver_data = {
.mdp_plat_id = MT8183,
.mdp_con_res = 0x14001000,
@ -1074,6 +1079,7 @@ const struct mtk_mdp_driver_data mt8183_mdp_driver_data = {
.def_limit = &mt8183_mdp_def_limit,
.pipe_info = mt8183_pipe_info,
.pipe_info_len = ARRAY_SIZE(mt8183_pipe_info),
.pp_used = MDP_PP_USED_1,
};
const struct mtk_mdp_driver_data mt8195_mdp_driver_data = {
@ -1090,6 +1096,8 @@ const struct mtk_mdp_driver_data mt8195_mdp_driver_data = {
.def_limit = &mt8195_mdp_def_limit,
.pipe_info = mt8195_pipe_info,
.pipe_info_len = ARRAY_SIZE(mt8195_pipe_info),
.pp_criteria = &mt8195_mdp_pp_criteria,
.pp_used = MDP_PP_USED_2,
};
s32 mdp_cfg_get_id_inner(struct mdp_dev *mdp_dev, enum mtk_mdp_comp_id id)

View File

@ -55,6 +55,16 @@ static struct mtk_mutex *__get_mutex(const struct mdp_dev *mdp_dev,
return mdp_dev->mm_subsys[p->sub_id].mdp_mutex[p->mutex_id];
}
static u8 __get_pp_num(enum mdp_stream_type type)
{
switch (type) {
case MDP_STREAM_TYPE_DUAL_BITBLT:
return MDP_PP_USED_2;
default:
return MDP_PP_USED_1;
}
}
static enum mdp_pipe_id __get_pipe(const struct mdp_dev *mdp_dev,
enum mtk_mdp_comp_id id)
{
@ -92,6 +102,44 @@ static enum mdp_pipe_id __get_pipe(const struct mdp_dev *mdp_dev,
return pipe_id;
}
static struct img_config *__get_config_offset(struct mdp_dev *mdp,
struct mdp_cmdq_param *param,
u8 pp_idx)
{
const int p_id = mdp->mdp_data->mdp_plat_id;
struct device *dev = &mdp->pdev->dev;
void *cfg_c, *cfg_n;
long bound = mdp->vpu.config_size;
if (pp_idx >= mdp->mdp_data->pp_used)
goto err_param;
if (CFG_CHECK(MT8183, p_id))
cfg_c = CFG_OFST(MT8183, param->config, pp_idx);
else if (CFG_CHECK(MT8195, p_id))
cfg_c = CFG_OFST(MT8195, param->config, pp_idx);
else
goto err_param;
if (CFG_CHECK(MT8183, p_id))
cfg_n = CFG_OFST(MT8183, param->config, pp_idx + 1);
else if (CFG_CHECK(MT8195, p_id))
cfg_n = CFG_OFST(MT8195, param->config, pp_idx + 1);
else
goto err_param;
if ((long)cfg_n - (long)mdp->vpu.config > bound) {
dev_err(dev, "config offset %ld OOB %ld\n", (long)cfg_n, bound);
cfg_c = ERR_PTR(-EFAULT);
}
return (struct img_config *)cfg_c;
err_param:
cfg_c = ERR_PTR(-EINVAL);
return (struct img_config *)cfg_c;
}
static int mdp_path_subfrm_require(const struct mdp_path *path,
struct mdp_cmdq_cmd *cmd,
struct mdp_pipe_info *p, u32 count)
@ -476,8 +524,19 @@ static void mdp_auto_release_work(struct work_struct *work)
mdp_comp_clocks_off(&mdp->pdev->dev, cmd->comps,
cmd->num_comps);
atomic_dec(&mdp->job_count);
wake_up(&mdp->callback_wq);
if (atomic_dec_and_test(&mdp->job_count)) {
if (cmd->mdp_ctx)
mdp_m2m_job_finish(cmd->mdp_ctx);
if (cmd->user_cmdq_cb) {
struct cmdq_cb_data user_cb_data;
user_cb_data.sta = cmd->data->sta;
user_cb_data.pkt = cmd->data->pkt;
cmd->user_cmdq_cb(user_cb_data);
}
wake_up(&mdp->callback_wq);
}
mdp_cmdq_pkt_destroy(&cmd->pkt);
kfree(cmd->comps);
@ -501,20 +560,10 @@ static void mdp_handle_cmdq_callback(struct mbox_client *cl, void *mssg)
data = (struct cmdq_cb_data *)mssg;
cmd = container_of(data->pkt, struct mdp_cmdq_cmd, pkt);
cmd->data = data;
mdp = cmd->mdp;
dev = &mdp->pdev->dev;
if (cmd->mdp_ctx)
mdp_m2m_job_finish(cmd->mdp_ctx);
if (cmd->user_cmdq_cb) {
struct cmdq_cb_data user_cb_data;
user_cb_data.sta = data->sta;
user_cb_data.pkt = data->pkt;
cmd->user_cmdq_cb(user_cb_data);
}
INIT_WORK(&cmd->auto_release_work, mdp_auto_release_work);
if (!queue_work(mdp->clock_wq, &cmd->auto_release_work)) {
struct mtk_mutex *mutex;
@ -526,8 +575,8 @@ static void mdp_handle_cmdq_callback(struct mbox_client *cl, void *mssg)
mdp_comp_clocks_off(&mdp->pdev->dev, cmd->comps,
cmd->num_comps);
atomic_dec(&mdp->job_count);
wake_up(&mdp->callback_wq);
if (atomic_dec_and_test(&mdp->job_count))
wake_up(&mdp->callback_wq);
mdp_cmdq_pkt_destroy(&cmd->pkt);
kfree(cmd->comps);
@ -537,31 +586,41 @@ static void mdp_handle_cmdq_callback(struct mbox_client *cl, void *mssg)
}
}
int mdp_cmdq_send(struct mdp_dev *mdp, struct mdp_cmdq_param *param)
static struct mdp_cmdq_cmd *mdp_cmdq_prepare(struct mdp_dev *mdp,
struct mdp_cmdq_param *param,
u8 pp_idx)
{
struct mdp_path *path = NULL;
struct mdp_cmdq_cmd *cmd = NULL;
struct mdp_comp *comps = NULL;
struct device *dev = &mdp->pdev->dev;
const int p_id = mdp->mdp_data->mdp_plat_id;
struct img_config *config;
struct mtk_mutex *mutex = NULL;
enum mdp_pipe_id pipe_id;
int i, ret;
u32 num_comp = 0;
int i, ret = -ECANCELED;
u32 num_comp;
atomic_inc(&mdp->job_count);
if (atomic_read(&mdp->suspended)) {
atomic_dec(&mdp->job_count);
return -ECANCELED;
config = __get_config_offset(mdp, param, pp_idx);
if (IS_ERR(config)) {
ret = PTR_ERR(config);
goto err_uninit;
}
if (CFG_CHECK(MT8183, p_id))
num_comp = CFG_GET(MT8183, config, num_components);
else if (CFG_CHECK(MT8195, p_id))
num_comp = CFG_GET(MT8195, config, num_components);
else
goto err_uninit;
cmd = kzalloc(sizeof(*cmd), GFP_KERNEL);
if (!cmd) {
ret = -ENOMEM;
goto err_cancel_job;
goto err_uninit;
}
ret = mdp_cmdq_pkt_create(mdp->cmdq_clt, &cmd->pkt, SZ_16K);
ret = mdp_cmdq_pkt_create(mdp->cmdq_clt[pp_idx], &cmd->pkt, SZ_16K);
if (ret)
goto err_free_cmd;
@ -586,7 +645,7 @@ int mdp_cmdq_send(struct mdp_dev *mdp, struct mdp_cmdq_param *param)
}
path->mdp_dev = mdp;
path->config = param->config;
path->config = config;
path->param = param->param;
for (i = 0; i < param->param->num_outputs; i++) {
path->bounds[i].left = 0;
@ -600,7 +659,7 @@ int mdp_cmdq_send(struct mdp_dev *mdp, struct mdp_cmdq_param *param)
}
ret = mdp_path_ctx_init(mdp, path);
if (ret) {
dev_err(dev, "mdp_path_ctx_init error\n");
dev_err(dev, "mdp_path_ctx_init error %d\n", pp_idx);
goto err_free_path;
}
@ -608,13 +667,13 @@ int mdp_cmdq_send(struct mdp_dev *mdp, struct mdp_cmdq_param *param)
mutex = __get_mutex(mdp, &mdp->mdp_data->pipe_info[pipe_id]);
ret = mtk_mutex_prepare(mutex);
if (ret) {
dev_err(dev, "Fail to enable mutex clk\n");
dev_err(dev, "Fail to enable mutex %d clk\n", pp_idx);
goto err_free_path;
}
ret = mdp_path_config(mdp, cmd, path);
if (ret) {
dev_err(dev, "mdp_path_config error\n");
dev_err(dev, "mdp_path_config error %d\n", pp_idx);
goto err_free_path;
}
cmdq_pkt_finalize(&cmd->pkt);
@ -633,7 +692,7 @@ int mdp_cmdq_send(struct mdp_dev *mdp, struct mdp_cmdq_param *param)
sizeof(struct mdp_comp));
}
mdp->cmdq_clt->client.rx_callback = mdp_handle_cmdq_callback;
mdp->cmdq_clt[pp_idx]->client.rx_callback = mdp_handle_cmdq_callback;
cmd->mdp = mdp;
cmd->user_cmdq_cb = param->cmdq_cb;
cmd->user_cb_data = param->cb_data;
@ -641,26 +700,9 @@ int mdp_cmdq_send(struct mdp_dev *mdp, struct mdp_cmdq_param *param)
cmd->num_comps = num_comp;
cmd->mdp_ctx = param->mdp_ctx;
ret = mdp_comp_clocks_on(&mdp->pdev->dev, cmd->comps, cmd->num_comps);
if (ret)
goto err_free_path;
dma_sync_single_for_device(mdp->cmdq_clt->chan->mbox->dev,
cmd->pkt.pa_base, cmd->pkt.cmd_buf_size,
DMA_TO_DEVICE);
ret = mbox_send_message(mdp->cmdq_clt->chan, &cmd->pkt);
if (ret < 0) {
dev_err(dev, "mbox send message fail %d!\n", ret);
goto err_clock_off;
}
mbox_client_txdone(mdp->cmdq_clt->chan, 0);
kfree(path);
return 0;
return cmd;
err_clock_off:
mdp_comp_clocks_off(&mdp->pdev->dev, cmd->comps,
cmd->num_comps);
err_free_path:
if (mutex)
mtk_mutex_unprepare(mutex);
@ -671,8 +713,58 @@ err_destroy_pkt:
mdp_cmdq_pkt_destroy(&cmd->pkt);
err_free_cmd:
kfree(cmd);
err_uninit:
return ERR_PTR(ret);
}
int mdp_cmdq_send(struct mdp_dev *mdp, struct mdp_cmdq_param *param)
{
struct mdp_cmdq_cmd *cmd[MDP_PP_MAX] = {NULL};
struct device *dev = &mdp->pdev->dev;
int i, ret;
u8 pp_used = __get_pp_num(param->param->type);
atomic_set(&mdp->job_count, pp_used);
if (atomic_read(&mdp->suspended)) {
atomic_set(&mdp->job_count, 0);
return -ECANCELED;
}
for (i = 0; i < pp_used; i++) {
cmd[i] = mdp_cmdq_prepare(mdp, param, i);
if (IS_ERR_OR_NULL(cmd[i])) {
ret = PTR_ERR(cmd[i]);
goto err_cancel_job;
}
}
for (i = 0; i < pp_used; i++) {
ret = mdp_comp_clocks_on(&mdp->pdev->dev, cmd[i]->comps, cmd[i]->num_comps);
if (ret)
goto err_clock_off;
}
for (i = 0; i < pp_used; i++) {
dma_sync_single_for_device(mdp->cmdq_clt[i]->chan->mbox->dev,
cmd[i]->pkt.pa_base, cmd[i]->pkt.cmd_buf_size,
DMA_TO_DEVICE);
ret = mbox_send_message(mdp->cmdq_clt[i]->chan, &cmd[i]->pkt);
if (ret < 0) {
dev_err(dev, "mbox send message fail %d!\n", ret);
i = pp_used;
goto err_clock_off;
}
mbox_client_txdone(mdp->cmdq_clt[i]->chan, 0);
}
return 0;
err_clock_off:
while (--i >= 0)
mdp_comp_clocks_off(&mdp->pdev->dev, cmd[i]->comps,
cmd[i]->num_comps);
err_cancel_job:
atomic_dec(&mdp->job_count);
atomic_set(&mdp->job_count, 0);
return ret;
}

View File

@ -29,6 +29,7 @@ struct mdp_cmdq_cmd {
struct cmdq_pkt pkt;
s32 *event;
struct mdp_dev *mdp;
struct cmdq_cb_data *data;
void (*user_cmdq_cb)(struct cmdq_cb_data data);
void *user_cb_data;
struct mdp_comp *comps;

View File

@ -142,6 +142,10 @@ void mdp_video_device_release(struct video_device *vdev)
struct mdp_dev *mdp = (struct mdp_dev *)video_get_drvdata(vdev);
int i;
for (i = 0; i < mdp->mdp_data->pp_used; i++)
if (mdp->cmdq_clt[i])
cmdq_mbox_destroy(mdp->cmdq_clt[i]);
scp_put(mdp->scp);
destroy_workqueue(mdp->job_wq);
@ -299,10 +303,12 @@ static int mdp_probe(struct platform_device *pdev)
mutex_init(&mdp->vpu_lock);
mutex_init(&mdp->m2m_lock);
mdp->cmdq_clt = cmdq_mbox_create(dev, 0);
if (IS_ERR(mdp->cmdq_clt)) {
ret = PTR_ERR(mdp->cmdq_clt);
goto err_put_scp;
for (i = 0; i < mdp->mdp_data->pp_used; i++) {
mdp->cmdq_clt[i] = cmdq_mbox_create(dev, i);
if (IS_ERR(mdp->cmdq_clt[i])) {
ret = PTR_ERR(mdp->cmdq_clt[i]);
goto err_mbox_destroy;
}
}
init_waitqueue_head(&mdp->callback_wq);
@ -331,8 +337,8 @@ success_return:
err_unregister_device:
v4l2_device_unregister(&mdp->v4l2_dev);
err_mbox_destroy:
cmdq_mbox_destroy(mdp->cmdq_clt);
err_put_scp:
while (--i >= 0)
cmdq_mbox_destroy(mdp->cmdq_clt[i]);
scp_put(mdp->scp);
err_destroy_clock_wq:
destroy_workqueue(mdp->clock_wq);

View File

@ -77,6 +77,14 @@ enum mdp_pipe_id {
MDP_PIPE_MAX
};
/* MDP parallel pipe control */
enum {
MDP_PP_USED_1 = 1,
MDP_PP_USED_2 = 2,
};
#define MDP_PP_MAX MDP_PP_USED_2
struct mtk_mdp_driver_data {
const int mdp_plat_id;
const resource_size_t mdp_con_res;
@ -91,6 +99,8 @@ struct mtk_mdp_driver_data {
const struct mdp_limit *def_limit;
const struct mdp_pipe_info *pipe_info;
unsigned int pipe_info_len;
const struct v4l2_rect *pp_criteria;
const u8 pp_used;
};
struct mdp_mm_subsys {
@ -115,7 +125,7 @@ struct mdp_dev {
s32 vpu_count;
u32 id_count;
struct ida mdp_ida;
struct cmdq_client *cmdq_clt;
struct cmdq_client *cmdq_clt[MDP_PP_MAX];
wait_queue_head_t callback_wq;
struct v4l2_device v4l2_dev;

View File

@ -87,6 +87,9 @@ static void mdp_m2m_device_run(void *priv)
dst_vb = v4l2_m2m_next_dst_buf(ctx->m2m_ctx);
mdp_set_dst_config(&param.outputs[0], frame, &dst_vb->vb2_buf);
if (mdp_check_pp_enable(ctx->mdp_dev, frame))
param.type = MDP_STREAM_TYPE_DUAL_BITBLT;
ret = mdp_vpu_process(&ctx->mdp_dev->vpu, &param);
if (ret) {
dev_err(&ctx->mdp_dev->pdev->dev,
@ -101,6 +104,18 @@ static void mdp_m2m_device_run(void *priv)
task.cb_data = NULL;
task.mdp_ctx = ctx;
if (atomic_read(&ctx->mdp_dev->job_count)) {
ret = wait_event_timeout(ctx->mdp_dev->callback_wq,
!atomic_read(&ctx->mdp_dev->job_count),
2 * HZ);
if (ret == 0) {
dev_err(&ctx->mdp_dev->pdev->dev,
"%d jobs not yet done\n",
atomic_read(&ctx->mdp_dev->job_count));
goto worker_end;
}
}
ret = mdp_cmdq_send(ctx->mdp_dev, &task);
if (ret) {
dev_err(&ctx->mdp_dev->pdev->dev,

View File

@ -304,6 +304,24 @@ int mdp_check_scaling_ratio(const struct v4l2_rect *crop,
return 0;
}
bool mdp_check_pp_enable(struct mdp_dev *mdp, struct mdp_frame *frame)
{
u32 s, r1, r2;
if (!mdp || !frame)
return false;
if (!mdp->mdp_data->pp_criteria)
return false;
s = mdp->mdp_data->pp_criteria->width *
mdp->mdp_data->pp_criteria->height;
r1 = frame->crop.c.width * frame->crop.c.height;
r2 = frame->compose.width * frame->compose.height;
return (r1 >= s || r2 >= s);
}
/* Stride that is accepted by MDP HW */
static u32 mdp_fmt_get_stride(const struct mdp_format *fmt,
u32 bytesperline, unsigned int plane)

View File

@ -368,6 +368,7 @@ int mdp_try_crop(struct mdp_m2m_ctx *ctx, struct v4l2_rect *r,
int mdp_check_scaling_ratio(const struct v4l2_rect *crop,
const struct v4l2_rect *compose, s32 rotation,
const struct mdp_limit *limit);
bool mdp_check_pp_enable(struct mdp_dev *mdp, struct mdp_frame *frame);
void mdp_set_src_config(struct img_input *in,
struct mdp_frame *frame, struct vb2_buffer *vb);
void mdp_set_dst_config(struct img_output *out,

View File

@ -198,6 +198,7 @@ int mdp_vpu_dev_init(struct mdp_vpu_dev *vpu, struct mtk_scp *scp,
};
struct mdp_dev *mdp = vpu_to_mdp(vpu);
int err;
u8 pp_num = mdp->mdp_data->pp_used;
init_completion(&vpu->ipi_acked);
vpu->scp = scp;
@ -211,7 +212,7 @@ int mdp_vpu_dev_init(struct mdp_vpu_dev *vpu, struct mtk_scp *scp,
mutex_lock(vpu->lock);
vpu->work_size = ALIGN(vpu->work_size, 64);
vpu->param_size = ALIGN(sizeof(struct img_ipi_frameparam), 64);
vpu->config_size = ALIGN(sizeof(struct img_config), 64);
vpu->config_size = ALIGN(sizeof(struct img_config) * pp_num, 64);
err = mdp_vpu_shared_mem_alloc(vpu);
mutex_unlock(vpu->lock);
if (err) {