[vepu580]: Support intra frame twopass deflicker

NOTE: Only vepu580 support this feature.

Change-Id: I05768a5ca5f6d32eab3d0d9eccab0c471882cd0e
Signed-off-by: sayon.chen <sayon.chen@rock-chips.com>
Signed-off-by: Herman Chen <herman.chen@rock-chips.com>

inc/mpp_rc_defs.h

@@ -43,7 +43,18 @@ typedef union EncFrmStatus_u {
          * 1 - do not write the reconstructed frame pixel to memory
          */
         RK_U32          non_recn        : 1;
-        RK_U32          reserved0       : 2;
+
+        /*
+         * 0 - normal frame and normal dpb management
+         * 1 - save recon frame as first pass extra frame. Used in two pass mode
+         */
+        RK_U32          save_pass1      : 1;
+
+        /*
+         * 0 - use normal input source frame as input
+         * 1 - use the previously stored first pass recon frame as input frame
+         */
+        RK_U32          use_pass1       : 1;
 
         /* reference status flag */
         /*

mpp/base/inc/mpp_enc_refs.h

@@ -75,6 +75,11 @@ MPP_RET mpp_enc_refs_dryrun(MppEncRefs refs);
 MPP_RET mpp_enc_refs_stash(MppEncRefs refs);
 MPP_RET mpp_enc_refs_rollback(MppEncRefs refs);
 
+/* two-pass encoding functions */
+/* check next frame is intra or not */
+RK_S32  mpp_enc_refs_next_frm_is_intra(MppEncRefs refs);
+MPP_RET mpp_enc_refs_get_cpb_pass1(MppEncRefs refs, EncCpbStatus *status);
+
 #define dump_frm(frm)   _dump_frm(frm, __FUNCTION__, __LINE__)
 
 void _dump_frm(EncFrmStatus *frm, const char *func, RK_S32 line);

mpp/base/mpp_enc_refs.cpp

@@ -757,6 +757,7 @@ MPP_RET mpp_enc_refs_get_cpb(MppEncRefs refs, EncCpbStatus *status)
     RefsCnt *lt_cfg = cpb->lt_cnter;
     RK_S32 set_to_lt = 0;
     RK_S32 cleanup_cpb = 0;
+    RK_S32 prev_frm_is_pass1 = frm->save_pass1;
     RK_S32 i;
 
     /* step 1. check igop from cfg_set and force idr for usr_cfg */
@@ -868,6 +869,10 @@ MPP_RET mpp_enc_refs_get_cpb(MppEncRefs refs, EncCpbStatus *status)
     } else
         ref->val = 0;
 
+    /* step 5. check use previous pass one frame as input */
+    if (prev_frm_is_pass1)
+        frm->use_pass1 = 1;
+
     if (enc_refs_debug & MPP_ENC_REFS_DBG_FRM) {
         mpp_log_f("frm status:\n");
         dump_frm(frm);
@@ -891,6 +896,93 @@ MPP_RET mpp_enc_refs_get_cpb(MppEncRefs refs, EncCpbStatus *status)
     return MPP_OK;
 }
 
+RK_S32 mpp_enc_refs_next_frm_is_intra(MppEncRefs refs)
+{
+    if (NULL == refs) {
+        mpp_err_f("invalid NULL input refs\n");
+        return MPP_ERR_VALUE;
+    }
+
+    enc_refs_dbg_func("enter %p\n", refs);
+
+    MppEncRefsImpl *p = (MppEncRefsImpl *)refs;
+    EncVirtualCpb *cpb = &p->cpb;
+    MppEncRefFrmUsrCfg *usr_cfg = &p->usr_cfg;
+    RK_S32 is_intra = 0;
+
+    if (p->changed & ENC_REFS_IGOP_CHANGED)
+        is_intra = 1;
+
+    if (p->igop && cpb->seq_idx >= p->igop)
+        is_intra = 1;
+
+    if (usr_cfg->force_flag & ENC_FORCE_IDR)
+        is_intra = 1;
+
+    if (!cpb->frm_idx)
+        is_intra = 0;
+
+    enc_refs_dbg_func("leave %p\n", refs);
+
+    return is_intra;
+}
+
+MPP_RET mpp_enc_refs_get_cpb_pass1(MppEncRefs refs, EncCpbStatus *status)
+{
+    if (NULL == refs) {
+        mpp_err_f("invalid NULL input refs\n");
+        return MPP_ERR_VALUE;
+    }
+
+    enc_refs_dbg_func("enter %p\n", refs);
+
+    MppEncRefsImpl *p = (MppEncRefsImpl *)refs;
+    EncVirtualCpb *cpb = &p->cpb;
+    EncFrmStatus *frm = &status->curr;
+    EncFrmStatus *ref = &status->refr;
+
+    frm->valid = 1;
+    frm->save_pass1 = 1;
+    frm->is_non_ref = 1;
+    frm->is_lt_ref = 0;
+    frm->temporal_id = 0;
+    frm->ref_mode = REF_TO_PREV_REF_FRM;
+    frm->ref_arg = 0;
+    frm->non_recn = 0;
+
+    /* step 4. try find ref by the ref_mode */
+    EncFrmStatus *ref_found = get_ref_from_cpb(cpb, frm);
+    if (ref_found) {
+        RK_S32 cpb_idx = check_ref_cpb_pos(cpb, ref_found);
+
+        mpp_assert(cpb_idx >= 0);
+        cpb->list0[0].val = ref->val;
+        ref->val = ref_found->val;
+    } else
+        ref->val = 0;
+
+    if (enc_refs_debug & MPP_ENC_REFS_DBG_FRM) {
+        mpp_log_f("frm status:\n");
+        dump_frm(frm);
+        mpp_log_f("ref status:\n");
+        dump_frm(ref);
+    }
+
+    /* step 5. generate cpb init */
+    memset(status->init, 0, sizeof(status->init));
+    save_cpb_status(cpb, status->init);
+    // TODO: cpb_init must be the same to cpb_final
+
+    /* step 6. store frame according to status */
+    store_ref_to_cpb(cpb, frm);
+
+    /* step 7. generate cpb final */
+    memset(status->final, 0, sizeof(status->final));
+    save_cpb_status(cpb, status->final);
+
+    enc_refs_dbg_func("leave %p\n", refs);
+    return MPP_OK;
+}
 MPP_RET mpp_enc_refs_stash(MppEncRefs refs)
 {
     if (NULL == refs) {

mpp/codec/enc/h265/h265e_api.c

@@ -54,7 +54,6 @@ static MPP_RET h265e_init(void *ctx, EncImplCfg *ctrlCfg)
     p->cfg = ctrlCfg->cfg;
 
     memset(&p->syntax, 0, sizeof(p->syntax));
-    ctrlCfg->task_count = 1;
 
     p->extra_info = mpp_calloc(H265eExtraInfo, 1);
 

mpp/codec/enc/jpeg/jpege_api_v2.c

@@ -75,7 +75,6 @@ static MPP_RET jpege_init_v2(void *ctx, EncImplCfg *cfg)
     p->cfg = cfg->cfg;
 
     mpp_assert(cfg->coding = MPP_VIDEO_CodingMJPEG);
-    cfg->task_count = 1;
 
     {
         /* init default rc config */

mpp/codec/inc/enc_impl_api.h

@@ -32,9 +32,6 @@ typedef struct EncImplCfg_t {
     MppClientType   type;
     MppEncCfgSet    *cfg;
     MppEncRefs      refs;
-
-    // output
-    RK_S32          task_count;
 } EncImplCfg;
 
 /*

mpp/codec/inc/mpp_enc_impl.h

@@ -125,6 +125,10 @@ typedef struct MppEncImpl_t {
     MppEncRefs          refs;
     MppEncRefFrmUsrCfg  frm_cfg;
 
+    /* two-pass deflicker parameters */
+    RK_U32              support_hw_deflicker;
+    EncRcTaskInfo       rc_info_prev;
+
     /* Encoder configure set */
     MppEncCfgSet        cfg;
 } MppEncImpl;

mpp/codec/mpp_enc_impl.cpp

@@ -1201,6 +1201,74 @@ static MPP_RET mpp_enc_check_pkt_buf(MppEncImpl *enc)
     return MPP_OK;
 }
 
+static MPP_RET mpp_enc_proc_two_pass(Mpp *mpp, EncTask *task)
+{
+    MppEncImpl *enc = (MppEncImpl *)mpp->mEnc;
+    MPP_RET ret = MPP_OK;
+
+    if (mpp_enc_refs_next_frm_is_intra(enc->refs)) {
+        EncRcTask *rc_task = &enc->rc_task;
+        EncFrmStatus frm_bak = rc_task->frm;
+        EncRcTaskInfo rc_info = rc_task->info;
+        EncCpbStatus *cpb = &rc_task->cpb;
+        EncFrmStatus *frm = &rc_task->frm;
+        HalEncTask *hal_task = &task->info;
+        EncImpl impl = enc->impl;
+        MppEncHal hal = enc->enc_hal;
+        MppPacket packet = hal_task->packet;
+        RK_S32 task_len = hal_task->length;
+        RK_S32 hw_len = hal_task->hw_length;
+        RK_S32 pkt_len = mpp_packet_get_length(packet);
+
+        enc_dbg_detail("task %d two pass mode enter\n", frm->seq_idx);
+        rc_task->info = enc->rc_info_prev;
+
+        enc_dbg_detail("task %d enc proc dpb\n", frm->seq_idx);
+        mpp_enc_refs_get_cpb_pass1(enc->refs, cpb);
+
+        enc_dbg_frm_status("frm %d start ***********************************\n", cpb->curr.seq_idx);
+        ENC_RUN_FUNC2(enc_impl_proc_dpb, impl, hal_task, mpp, ret);
+
+        enc_dbg_detail("task %d enc proc hal\n", frm->seq_idx);
+        ENC_RUN_FUNC2(enc_impl_proc_hal, impl, hal_task, mpp, ret);
+
+        enc_dbg_detail("task %d hal get task\n", frm->seq_idx);
+        ENC_RUN_FUNC2(mpp_enc_hal_get_task, hal, hal_task, mpp, ret);
+
+        enc_dbg_detail("task %d hal generate reg\n", frm->seq_idx);
+        ENC_RUN_FUNC2(mpp_enc_hal_gen_regs, hal, hal_task, mpp, ret);
+
+        enc_dbg_detail("task %d hal start\n", frm->seq_idx);
+        ENC_RUN_FUNC2(mpp_enc_hal_start, hal, hal_task, mpp, ret);
+
+        enc_dbg_detail("task %d hal wait\n", frm->seq_idx);
+        ENC_RUN_FUNC2(mpp_enc_hal_wait,  hal, hal_task, mpp, ret);
+
+        enc_dbg_detail("task %d hal ret task\n", frm->seq_idx);
+        ENC_RUN_FUNC2(mpp_enc_hal_ret_task, hal, hal_task, mpp, ret);
+
+        //recover status & packet
+        mpp_packet_set_length(packet, pkt_len);
+        hal_task->hw_length = hw_len;
+        hal_task->length = task_len;
+
+        *frm = frm_bak;
+        rc_task->info = rc_info;
+
+        enc_dbg_detail("task %d two pass mode leave\n", frm->seq_idx);
+    }
+TASK_DONE:
+    return ret;
+}
+
+static void mpp_enc_rc_info_backup(MppEncImpl *enc)
+{
+    if (!enc->support_hw_deflicker || !enc->cfg.rc.debreath_en)
+        return;
+
+    enc->rc_info_prev = enc->rc_task.info;
+}
+
 static MPP_RET mpp_enc_normal(Mpp *mpp, EncTask *task)
 {
     MppEncImpl *enc = (MppEncImpl *)mpp->mEnc;
@@ -1215,6 +1283,12 @@ static MPP_RET mpp_enc_normal(Mpp *mpp, EncTask *task)
     MppPacket packet = hal_task->packet;
     MPP_RET ret = MPP_OK;
 
+    if (enc->support_hw_deflicker && enc->cfg.rc.debreath_en) {
+        ret = mpp_enc_proc_two_pass(mpp, task);
+        if (ret)
+            return ret;
+    }
+
     enc_dbg_detail("task %d enc proc dpb\n", frm->seq_idx);
     mpp_enc_refs_get_cpb(enc->refs, cpb);
 
@@ -1923,6 +1997,7 @@ TASK_DONE:
     mpp_task_meta_set_frame(enc->task_in, KEY_INPUT_FRAME, enc->frame);
     mpp_port_enqueue(enc->input, enc->task_in);
 
+    mpp_enc_rc_info_backup(enc);
     reset_enc_task(enc);
     task->status.val = 0;
 

mpp/codec/mpp_enc_v2.cpp

@@ -68,12 +68,12 @@ MPP_RET mpp_enc_init_v2(MppEnc *enc, MppEncInitCfg *cfg)
     enc_hal_cfg.cfg = &p->cfg;
     enc_hal_cfg.type = VPU_CLIENT_BUTT;
     enc_hal_cfg.dev = NULL;
+    enc_hal_cfg.cap_recn_out = 0;
 
     ctrl_cfg.coding = coding;
     ctrl_cfg.type = VPU_CLIENT_BUTT;
     ctrl_cfg.cfg = &p->cfg;
     ctrl_cfg.refs = p->refs;
-    ctrl_cfg.task_count = 2;
 
     ret = mpp_enc_hal_init(&enc_hal, &enc_hal_cfg);
     if (ret) {
@@ -82,7 +82,6 @@ MPP_RET mpp_enc_init_v2(MppEnc *enc, MppEncInitCfg *cfg)
     }
 
     ctrl_cfg.type = enc_hal_cfg.type;
-    ctrl_cfg.task_count = -1;
 
     ret = enc_impl_init(&impl, &ctrl_cfg);
     if (ret) {
@@ -107,6 +106,9 @@ MPP_RET mpp_enc_init_v2(MppEnc *enc, MppEncInitCfg *cfg)
     p->rc_cfg_size = SZ_1K;
     p->rc_cfg_info = mpp_calloc_size(char, p->rc_cfg_size);
 
+    if (enc_hal_cfg.cap_recn_out)
+        p->support_hw_deflicker = 1;
+
     {
         // create header packet storage
         size_t size = SZ_1K;

mpp/codec/rc/rc_ctx.h

@@ -21,7 +21,6 @@
 
 typedef struct RcModelV2Ctx_t {
     RcCfg           usr_cfg;
-    EncRcTaskInfo   hal_cfg;
 
     RK_U32          frame_type;
     RK_U32          last_frame_type;

mpp/codec/rc/rc_model_v2.c

@@ -474,10 +474,10 @@ MPP_RET calc_next_i_ratio(RcModelV2Ctx *ctx)
 MPP_RET calc_debreath_qp(RcModelV2Ctx *ctx)
 {
     rc_dbg_func("enter %p\n", ctx);
-    RK_S32 qp_start_sum = 0;
-    RK_U8 idx2 = ctx->pre_iblk4_prop >> 5;
-    RK_S32 new_start_qp = 0;
     RcDebreathCfg *debreath_cfg = &ctx->usr_cfg.debreath_cfg;
+    RK_S32 qp_start_sum = 0;
+    RK_S32 new_start_qp = 0;
+    RK_U8 idx2 = MPP_MIN(ctx->pre_iblk4_prop >> 5, (RK_S32)sizeof(intra_qp_map) - 1);
 
     static RK_S8 strength_map[36] = {
         0, 1, 1, 2,  2,  2,  3,  3,  3,  4,  4,  4,
@@ -485,7 +485,7 @@ MPP_RET calc_debreath_qp(RcModelV2Ctx *ctx)
         9, 9, 9, 10, 10, 10, 11, 11, 11, 12, 12, 12
     };
 
-    qp_start_sum = ctx->gop_qp_sum / ctx->gop_frm_cnt;
+    qp_start_sum = MPP_MIN(ctx->gop_qp_sum / ctx->gop_frm_cnt, (RK_S32)sizeof(strength_map) - 1);
 
     rc_dbg_qp("i start_qp %d, qp_start_sum = %d, intra_lv4_prop %d",
               ctx->start_qp, qp_start_sum, ctx->pre_iblk4_prop);
@@ -496,8 +496,7 @@ MPP_RET calc_debreath_qp(RcModelV2Ctx *ctx)
     else
         new_start_qp = qp_start_sum;
 
-    ctx->start_qp =  new_start_qp;
+    ctx->start_qp = mpp_clip(new_start_qp, ctx->usr_cfg.min_i_quality, ctx->usr_cfg.max_i_quality);
-    ctx->start_qp = mpp_clip(ctx->start_qp, 20, 51);
 
     rc_dbg_func("leave %p\n", ctx);
     return MPP_OK;
@@ -1264,11 +1263,6 @@ MPP_RET rc_model_v2_start(void *ctx, EncRcTask *task)
         info->quality_min = usr_cfg->min_quality;
     }
 
-    if (frm->is_idr) {
-        p->gop_frm_cnt = 0;
-        p->gop_qp_sum = 0;
-    }
-
     rc_dbg_rc("seq_idx %d intra %d\n", frm->seq_idx, frm->is_intra);
     rc_dbg_rc("bitrate [%d : %d : %d]\n", info->bit_min, info->bit_target, info->bit_max);
     rc_dbg_rc("quality [%d : %d : %d]\n", info->quality_min, info->quality_target, info->quality_max);
@@ -1407,9 +1401,13 @@ MPP_RET rc_model_v2_hal_start(void *ctx, EncRcTask *task)
 
             if (!p->reenc_cnt) {
                 p->cur_scale_qp = qp_scale;
-                if (p->usr_cfg.debreath_cfg.enable)
+                if (p->usr_cfg.debreath_cfg.enable) {
                     calc_debreath_qp(ctx);
                 }
+            }
+
+            p->gop_frm_cnt = 0;
+            p->gop_qp_sum = 0;
         } else {
             qp_scale = mpp_clip(qp_scale, (info->quality_min << 6), (info->quality_max << 6));
             p->cur_scale_qp = qp_scale;

mpp/hal/inc/mpp_enc_hal.h

@@ -29,6 +29,7 @@ typedef struct MppEncHalCfg_t {
     // output from enc_impl
     MppClientType   type;
     MppDev          dev;
+    RK_S32          cap_recn_out;
 } MppEncHalCfg;
 
 typedef struct MppEncHalApi_t {

mpp/hal/rkenc/h264e/hal_h264e_vepu580.c

@@ -49,6 +49,7 @@ typedef struct HalH264eVepu580Ctx_t {
     RK_S32                  pixel_buf_size;
     RK_S32                  thumb_buf_size;
     RK_S32                  max_buf_cnt;
+    MppDevRegOffCfgs        *offsets;
 
     /* external line buffer over 4K */
     MppBufferGroup          ext_line_buf_grp;
@@ -77,6 +78,9 @@ typedef struct HalH264eVepu580Ctx_t {
     /* finetune */
     void                    *tune;
 
+    /* two-pass deflicker */
+    MppBuffer               buf_pass1;
+
     /* register */
     HalVepu580RegSet        regs_set;
 } HalH264eVepu580Ctx;
@@ -146,6 +150,16 @@ static MPP_RET hal_h264e_vepu580_deinit(void *hal)
         p->hw_recn = NULL;
     }
 
+    if (p->offsets) {
+        mpp_dev_multi_offset_deinit(p->offsets);
+        p->offsets = NULL;
+    }
+
+    if (p->buf_pass1) {
+        mpp_buffer_put(p->buf_pass1);
+        p->buf_pass1 = NULL;
+    }
+
     if (p->tune) {
         vepu580_h264e_tune_deinit(p->tune);
         p->tune = NULL;
@@ -198,9 +212,13 @@ static MPP_RET hal_h264e_vepu580_init(void *hal, MppEncHalCfg *cfg)
         memcpy(hw->aq_step_i, h264_I_aq_step_default, sizeof(hw->aq_step_i));
         memcpy(hw->aq_step_p, h264_P_aq_step_default, sizeof(hw->aq_step_p));
     }
+    mpp_dev_multi_offset_init(&p->offsets, 24);
+    p->osd_cfg.reg_cfg = p->offsets;
 
     p->tune = vepu580_h264e_tune_init(p);
 
+    cfg->cap_recn_out = 1;
+
 DONE:
     if (ret)
         hal_h264e_vepu580_deinit(hal);
@@ -558,6 +576,71 @@ static MPP_RET setup_vepu580_prep(HalVepu580RegSet *regs, MppEncPrepCfg *prep)
     return ret;
 }
 
+static MPP_RET vepu580_h264e_save_pass1_patch(HalVepu580RegSet *regs, HalH264eVepu580Ctx *ctx)
+{
+    RK_S32 width_align = MPP_ALIGN(ctx->cfg->prep.width, 16);
+    RK_S32 height_align = MPP_ALIGN(ctx->cfg->prep.height, 16);
+
+    if (NULL == ctx->buf_pass1) {
+        mpp_buffer_get(NULL, &ctx->buf_pass1, width_align * height_align * 3 / 2);
+        if (!ctx->buf_pass1) {
+            mpp_err("buf_pass1 malloc fail, debreath invaild");
+            return MPP_NOK;
+        }
+    }
+
+    regs->reg_base.enc_pic.cur_frm_ref = 1;
+    regs->reg_base.rfpw_h_addr = mpp_buffer_get_fd(ctx->buf_pass1);
+    regs->reg_base.rfpw_b_addr = regs->reg_base.rfpw_h_addr;
+    regs->reg_base.enc_pic.rec_fbc_dis = 1;
+    mpp_dev_multi_offset_update(ctx->offsets, 164, width_align * height_align);
+
+    return MPP_OK;
+}
+
+static MPP_RET vepu580_h264e_use_pass1_patch(HalVepu580RegSet *regs, HalH264eVepu580Ctx *ctx)
+{
+    MppEncPrepCfg *prep = &ctx->cfg->prep;
+    RK_S32 hor_stride = MPP_ALIGN(prep->width, 16);
+    RK_S32 ver_stride = MPP_ALIGN(prep->height, 16);
+    RK_S32 frame_size = hor_stride * ver_stride;
+    RK_S32 fd_in = mpp_buffer_get_fd(ctx->buf_pass1);
+    RK_S32 y_stride;
+    RK_S32 c_stride;
+
+    hal_h264e_dbg_func("enter\n");
+
+    regs->reg_base.src_fmt.src_cfmt   = VEPU541_FMT_YUV420SP;
+    regs->reg_base.src_fmt.alpha_swap = 0;
+    regs->reg_base.src_fmt.rbuv_swap  = 0;
+    regs->reg_base.src_fmt.out_fmt    = 1;
+
+    y_stride = MPP_ALIGN(prep->width, 16);
+    c_stride = y_stride;
+
+    regs->reg_base.src_proc.afbcd_en   =  0;
+    regs->reg_base.src_strd0.src_strd0  = y_stride;
+    regs->reg_base.src_strd1.src_strd1  = c_stride;
+
+    regs->reg_base.src_proc.src_mirr   = 0;
+    regs->reg_base.src_proc.src_rot    = 0;
+    regs->reg_base.src_proc.txa_en     = 0;
+
+    regs->reg_base.pic_ofst.pic_ofst_y = 0;
+    regs->reg_base.pic_ofst.pic_ofst_x = 0;
+
+
+    regs->reg_base.adr_src0   = fd_in;
+    regs->reg_base.adr_src1   = fd_in;
+    regs->reg_base.adr_src2   = fd_in;
+
+    mpp_dev_multi_offset_update(ctx->offsets, 161, frame_size);
+    mpp_dev_multi_offset_update(ctx->offsets, 162, frame_size);
+
+    hal_h264e_dbg_func("leave\n");
+    return MPP_OK;
+}
+
 static void setup_vepu580_codec(HalVepu580RegSet *regs, H264eSps *sps,
                                 H264ePps *pps, H264eSlice *slice)
 {
@@ -1056,7 +1139,7 @@ static void setup_vepu580_rc_base(HalVepu580RegSet *regs, H264eSps *sps,
     hal_h264e_dbg_func("leave\n");
 }
 
-static void setup_vepu580_io_buf(HalVepu580RegSet *regs, MppDev dev,
+static void setup_vepu580_io_buf(HalVepu580RegSet *regs, MppDevRegOffCfgs *offsets,
                                  HalEncTask *task)
 {
     MppFrame frm = task->frame;
@@ -1131,23 +1214,10 @@ static void setup_vepu580_io_buf(HalVepu580RegSet *regs, MppDev dev,
         }
     }
 
-    MppDevRegOffsetCfg trans_cfg;
+    mpp_dev_multi_offset_update(offsets, 161, off_in[0]);
-
+    mpp_dev_multi_offset_update(offsets, 162, off_in[1]);
-    trans_cfg.reg_idx = 161;
+    mpp_dev_multi_offset_update(offsets, 172, siz_out);
-    trans_cfg.offset = off_in[0];
+    mpp_dev_multi_offset_update(offsets, 175, off_out);
-    mpp_dev_ioctl(dev, MPP_DEV_REG_OFFSET, &trans_cfg);
-
-    trans_cfg.reg_idx = 162;
-    trans_cfg.offset = off_in[1];
-    mpp_dev_ioctl(dev, MPP_DEV_REG_OFFSET, &trans_cfg);
-
-    trans_cfg.reg_idx = 172;
-    trans_cfg.offset = siz_out;
-    mpp_dev_ioctl(dev, MPP_DEV_REG_OFFSET, &trans_cfg);
-
-    trans_cfg.reg_idx = 175;
-    trans_cfg.offset = off_out;
-    mpp_dev_ioctl(dev, MPP_DEV_REG_OFFSET, &trans_cfg);
 
     hal_h264e_dbg_func("leave\n");
 }
@@ -1183,13 +1253,14 @@ static void setup_vepu580_roi(HalVepu580RegSet *regs, HalH264eVepu580Ctx *ctx)
     hal_h264e_dbg_func("leave\n");
 }
 
-static void setup_vepu580_recn_refr(HalVepu580RegSet *regs, MppDev dev,
+static void setup_vepu580_recn_refr(HalH264eVepu580Ctx *ctx, HalVepu580RegSet *regs)
-                                    H264eFrmInfo *frms, HalBufs bufs,
-                                    RK_S32 fbc_hdr_size)
 {
+    H264eFrmInfo *frms = ctx->frms;
+    HalBufs bufs = ctx->hw_recn;
+    RK_S32 fbc_hdr_size = ctx->pixel_buf_fbc_hdr_size;
+
     HalBuf *curr = hal_bufs_get_buf(bufs, frms->curr_idx);
     HalBuf *refr = hal_bufs_get_buf(bufs, frms->refr_idx);
-    MppDevRegOffsetCfg trans_cfg;
 
     hal_h264e_dbg_func("enter\n");
 
@@ -1204,10 +1275,6 @@ static void setup_vepu580_recn_refr(HalVepu580RegSet *regs, MppDev dev,
         regs->reg_base.rfpw_h_addr = fd;
         regs->reg_base.rfpw_b_addr = fd;
         regs->reg_base.dspw_addr = mpp_buffer_get_fd(buf_thumb);
-
-        trans_cfg.reg_idx = 164;
-        trans_cfg.offset = fbc_hdr_size;
-        mpp_dev_ioctl(dev, MPP_DEV_REG_OFFSET, &trans_cfg);
     }
 
     if (refr && refr->cnt) {
@@ -1221,12 +1288,11 @@ static void setup_vepu580_recn_refr(HalVepu580RegSet *regs, MppDev dev,
         regs->reg_base.rfpr_h_addr = fd;
         regs->reg_base.rfpr_b_addr = fd;
         regs->reg_base.dspr_addr = mpp_buffer_get_fd(buf_thumb);
-
-        trans_cfg.reg_idx = 166;
-        trans_cfg.offset = fbc_hdr_size;
-        mpp_dev_ioctl(dev, MPP_DEV_REG_OFFSET, &trans_cfg);
     }
 
+    mpp_dev_multi_offset_update(ctx->offsets, 164, fbc_hdr_size);
+    mpp_dev_multi_offset_update(ctx->offsets, 166, fbc_hdr_size);
+
     hal_h264e_dbg_func("leave\n");
 }
 
@@ -1620,20 +1686,21 @@ static void setup_vepu580_l2(HalVepu580RegSet *regs, H264eSlice *slice, MppEncHw
 
 static void setup_vepu580_ext_line_buf(HalVepu580RegSet *regs, HalH264eVepu580Ctx *ctx)
 {
+    RK_S32 offset = 0;
+
     if (ctx->ext_line_buf) {
-        MppDevRegOffsetCfg trans_cfg;
         RK_S32 fd = mpp_buffer_get_fd(ctx->ext_line_buf);
 
         regs->reg_base.ebuft_addr = fd;
         regs->reg_base.ebufb_addr = fd;
 
-        trans_cfg.reg_idx = 182;
+        offset = ctx->ext_line_buf_size;
-        trans_cfg.offset = ctx->ext_line_buf_size;
-        mpp_dev_ioctl(ctx->dev, MPP_DEV_REG_OFFSET, &trans_cfg);
     } else {
         regs->reg_base.ebufb_addr = 0;
         regs->reg_base.ebufb_addr = 0;
     }
+
+    mpp_dev_multi_offset_update(ctx->offsets, 182, offset);
 }
 
 static MPP_RET hal_h264e_vepu580_gen_regs(void *hal, HalEncTask *task)
@@ -1644,6 +1711,8 @@ static MPP_RET hal_h264e_vepu580_gen_regs(void *hal, HalEncTask *task)
     H264eSps *sps = ctx->sps;
     H264ePps *pps = ctx->pps;
     H264eSlice *slice = ctx->slice;
+    EncRcTask *rc_task = task->rc_task;
+    EncFrmStatus *frm = &rc_task->frm;
     MPP_RET ret = MPP_OK;
 
     hal_h264e_dbg_func("enter %p\n", hal);
@@ -1661,11 +1730,10 @@ static MPP_RET hal_h264e_vepu580_gen_regs(void *hal, HalEncTask *task)
     setup_vepu580_rdo_pred(regs, sps, pps, slice);
     setup_vepu580_rdo_cfg(&regs->reg_rdo);
     setup_vepu580_scl_cfg(&regs->reg_scl);
-    setup_vepu580_rc_base(regs, sps, slice, &cfg->hw, task->rc_task);
+    setup_vepu580_rc_base(regs, sps, slice, &cfg->hw, rc_task);
-    setup_vepu580_io_buf(regs, ctx->dev, task);
+    setup_vepu580_io_buf(regs, ctx->offsets, task);
     setup_vepu580_roi(regs, ctx);
-    setup_vepu580_recn_refr(regs, ctx->dev, ctx->frms, ctx->hw_recn,
+    setup_vepu580_recn_refr(ctx, regs);
-                            ctx->pixel_buf_fbc_hdr_size);
 
     regs->reg_base.meiw_addr = task->mv_info ? mpp_buffer_get_fd(task->mv_info) : 0;
 
@@ -1680,6 +1748,13 @@ static MPP_RET hal_h264e_vepu580_gen_regs(void *hal, HalEncTask *task)
     setup_vepu580_ext_line_buf(regs, ctx);
     vepu580_h264e_tune_reg_patch(ctx->tune);
 
+    /* two pass register patch */
+    if (frm->save_pass1)
+        vepu580_h264e_save_pass1_patch(regs, ctx);
+
+    if (frm->use_pass1)
+        vepu580_h264e_use_pass1_patch(regs, ctx);
+
     mpp_env_get_u32("dump_l1_reg", &dump_l1_reg, 0);
 
     if (dump_l1_reg) {
@@ -1789,6 +1864,12 @@ static MPP_RET hal_h264e_vepu580_start(void *hal, HalEncTask *task)
             break;
         }
 
+        ret = mpp_dev_ioctl(ctx->dev, MPP_DEV_REG_OFFS, ctx->offsets);
+        if (ret) {
+            mpp_err_f("set register offsets failed %d\n", ret);
+            break;
+        }
+
         rd_cfg.reg = &ctx->regs_set.reg_st;
         rd_cfg.size = sizeof(ctx->regs_set.reg_st);
         rd_cfg.offset = VEPU580_STATUS_OFFSET;
@@ -1901,6 +1982,8 @@ static MPP_RET hal_h264e_vepu580_ret_task(void *hal, HalEncTask *task)
 
     vepu580_h264e_tune_stat_update(ctx->tune);
 
+    mpp_dev_multi_offset_reset(ctx->offsets);
+
     hal_h264e_dbg_func("leave %p\n", hal);
 
     return MPP_OK;

mpp/hal/rkenc/h265e/hal_h265e_vepu580.c

@@ -87,6 +87,7 @@ typedef struct H265eV580HalContext_t {
     MppBufferGroup      tile_grp;
     MppBuffer           hw_tile_buf[2];
     MppBuffer           hw_tile_stream;
+    MppBuffer           buf_pass1;
 
     RK_U32              enc_mode;
     RK_U32              frame_size;
@@ -1172,6 +1173,9 @@ MPP_RET hal_h265e_v580_init(void *hal, MppEncHalCfg *cfg)
 
         ctx->tile_parall_en = cap->send_cmd > MPP_CMD_SET_SESSION_FD;
     }
+
+    cfg->cap_recn_out = 1;
+
     hal_h265e_leave();
     return ret;
 }
@@ -1211,6 +1215,11 @@ MPP_RET hal_h265e_v580_deinit(void *hal)
         ctx->tile_grp = NULL;
     }
 
+    if (ctx->buf_pass1) {
+        mpp_buffer_put(ctx->buf_pass1);
+        ctx->buf_pass1 = NULL;
+    }
+
     if (ctx->dev) {
         mpp_dev_deinit(ctx->dev);
         ctx->dev = NULL;
@@ -1308,18 +1317,14 @@ vepu580_h265_set_patch_info(MppDevRegOffCfgs *cfgs, H265eSyntax_new *syn,
     }
 
     /* input cb addr */
-    if (u_offset) {
     ret = mpp_dev_multi_offset_update(cfgs, 161, u_offset);
     if (ret)
         mpp_err_f("set input cb addr offset failed %d\n", ret);
-    }
 
     /* input cr addr */
-    if (v_offset) {
     ret = mpp_dev_multi_offset_update(cfgs, 162, v_offset);
     if (ret)
         mpp_err_f("set input cr addr offset failed %d\n", ret);
-    }
 
     return ret;
 }
@@ -1850,6 +1855,7 @@ void vepu580_h265_set_hw_address(H265eV580HalContext *ctx, hevc_vepu580_base *re
     if (!syn->sp.non_reference_flag) {
         regs->reg0163_rfpw_h_addr  = mpp_buffer_get_fd(recon_buf->buf[0]);
         regs->reg0164_rfpw_b_addr  = regs->reg0163_rfpw_h_addr;
+
         mpp_dev_multi_offset_update(ctx->reg_cfg, 164, ctx->fbc_header_len);
     }
     regs->reg0165_rfpr_h_addr = mpp_buffer_get_fd(ref_buf->buf[0]);
@@ -1899,15 +1905,76 @@ void vepu580_h265_set_hw_address(H265eV580HalContext *ctx, hevc_vepu580_base *re
 
     mpp_dev_multi_offset_update(ctx->reg_cfg, 175, mpp_packet_get_length(task->packet));
     mpp_dev_multi_offset_update(ctx->reg_cfg, 172, mpp_buffer_get_size(enc_task->output));
+}
 
-    regs->reg0204_pic_ofst.pic_ofst_y = mpp_frame_get_offset_y(task->frame);
+static MPP_RET vepu580_h265e_save_pass1_patch(H265eV580RegSet *regs, H265eV580HalContext *ctx,
-    regs->reg0204_pic_ofst.pic_ofst_x = mpp_frame_get_offset_x(task->frame);
+                                              RK_S32 tiles_enabled_flag)
+{
+    hevc_vepu580_base *reg_base = &regs->reg_base;
+    RK_S32 width_align = MPP_ALIGN(ctx->cfg->prep.width, 64);
+    RK_S32 height_align = MPP_ALIGN(ctx->cfg->prep.height, 64);
+
+    if (NULL == ctx->buf_pass1) {
+        mpp_buffer_get(NULL, &ctx->buf_pass1, width_align * height_align * 3 / 2);
+        if (!ctx->buf_pass1) {
+            mpp_err("buf_pass1 malloc fail, debreath invaild");
+            return MPP_NOK;
+        }
+    }
+
+    reg_base->reg0192_enc_pic.cur_frm_ref = 1;
+    reg_base->reg0163_rfpw_h_addr = mpp_buffer_get_fd(ctx->buf_pass1);
+    reg_base->reg0164_rfpw_b_addr = reg_base->reg0163_rfpw_h_addr;
+    reg_base->reg0192_enc_pic.rec_fbc_dis = 1;
+
+    if (tiles_enabled_flag)
+        reg_base->reg0238_synt_pps.lpf_fltr_acrs_til = 0;
+
+    mpp_dev_multi_offset_update(ctx->reg_cfg, 164, width_align * height_align);
+
+    return MPP_OK;
+}
+
+static MPP_RET vepu580_h265e_use_pass1_patch(H265eV580RegSet *regs, H265eV580HalContext *ctx,
+                                             H265eSyntax_new *syn)
+{
+    hevc_vepu580_control_cfg *reg_ctl = &regs->reg_ctl;
+    hevc_vepu580_base *reg_base = &regs->reg_base;
+    RK_U32 hor_stride = MPP_ALIGN(syn->pp.pic_width, 64);
+    RK_U32 ver_stride = MPP_ALIGN(syn->pp.pic_height, 64);
+    RK_U32 frame_size = hor_stride * ver_stride;
+    RK_S32 stridey = MPP_ALIGN(syn->pp.pic_width, 64);
+    VepuFmtCfg *fmt = (VepuFmtCfg *)ctx->input_fmt;
+    MPP_RET ret = MPP_OK;
+
+    reg_ctl->reg0012_dtrns_map.src_bus_edin = fmt->src_endian;
+    reg_base->reg0198_src_fmt.src_cfmt = VEPU541_FMT_YUV420SP;
+    reg_base->reg0198_src_fmt.out_fmt = 1;
+    reg_base->reg0205_src_strd0.src_strd0 = stridey;
+    reg_base->reg0206_src_strd1.src_strd1 = stridey;
+    reg_base->reg0160_adr_src0 = mpp_buffer_get_fd(ctx->buf_pass1);
+    reg_base->reg0161_adr_src1 = reg_base->reg0160_adr_src0;
+    reg_base->reg0162_adr_src2 = reg_base->reg0160_adr_src0;
+
+    /* input cb addr */
+    ret = mpp_dev_multi_offset_update(ctx->reg_cfg, 161, frame_size);
+    if (ret)
+        mpp_err_f("set input cb addr offset failed %d\n", ret);
+
+    /* input cr addr */
+    ret = mpp_dev_multi_offset_update(ctx->reg_cfg, 162, frame_size);
+    if (ret)
+        mpp_err_f("set input cr addr offset failed %d\n", ret);
+
+    return MPP_OK;
 }
 
 MPP_RET hal_h265e_v580_gen_regs(void *hal, HalEncTask *task)
 {
     H265eV580HalContext *ctx = (H265eV580HalContext *)hal;
     HalEncTask *enc_task = task;
+    EncRcTask *rc_task = enc_task->rc_task;
+    EncFrmStatus *frm = &rc_task->frm;
     H265eSyntax_new *syn = (H265eSyntax_new *)enc_task->syntax.data;
     H265eV580RegSet *regs = (H265eV580RegSet *)ctx->regs[0];
     RK_U32 pic_width_align8, pic_height_align8;
@@ -2041,6 +2108,13 @@ MPP_RET hal_h265e_v580_gen_regs(void *hal, HalEncTask *task)
 
     vepu580_h265e_tune_reg_patch(ctx->tune);
 
+    /* two pass register patch */
+    if (frm->save_pass1)
+        vepu580_h265e_save_pass1_patch(regs, ctx, syn->pp.tiles_enabled_flag);
+
+    if (frm->use_pass1)
+        vepu580_h265e_use_pass1_patch(regs, ctx, syn);
+
     ctx->frame_num++;
 
     hal_h265e_leave();
@@ -2131,11 +2205,14 @@ MPP_RET hal_h265e_v580_start(void *hal, HalEncTask *enc_task)
             hal_h265e_v580_set_uniform_tile(&hw_regs->reg_base, syn, k);
 
         if (k) {
-            if (!ctx->tile_parall_en) {
+            RK_U32 offset = 0;
-                RK_U32 offset = mpp_packet_get_length(enc_task->packet);
 
+            if (!ctx->tile_parall_en) {
+                offset = mpp_packet_get_length(enc_task->packet);
                 offset += stream_len;
+
                 reg_base->reg0173_bsbb_addr = mpp_buffer_get_fd(enc_task->output);
+
                 mpp_dev_multi_offset_update(ctx->reg_cfg, 175, offset);
                 mpp_dev_multi_offset_update(ctx->reg_cfg, 172, mpp_buffer_get_size(enc_task->output));
             } else {
@@ -2147,8 +2224,17 @@ MPP_RET hal_h265e_v580_start(void *hal, HalEncTask *enc_task)
                 mpp_dev_multi_offset_update(ctx->reg_cfg, 175, 0);
                 mpp_dev_multi_offset_update(ctx->reg_cfg, 172, mpp_buffer_get_size(ctx->hw_tile_stream));
             }
-            mpp_dev_multi_offset_update(ctx->reg_cfg, 164, ctx->fbc_header_len);
+
-            mpp_dev_multi_offset_update(ctx->reg_cfg, 166, ctx->fbc_header_len);
+            offset = ctx->fbc_header_len;
+
+            mpp_dev_multi_offset_update(ctx->reg_cfg, 166, offset);
+            mpp_dev_multi_offset_update(ctx->reg_cfg, 164, offset);
+
+            if (enc_task->rc_task->frm.save_pass1)
+                vepu580_h265e_save_pass1_patch(hw_regs, ctx, syn->pp.tiles_enabled_flag);
+
+            if (enc_task->rc_task->frm.use_pass1)
+                vepu580_h265e_use_pass1_patch(hw_regs, ctx, syn);
         }
         hal_h265e_v580_send_regs(ctx->dev, hw_regs, reg_out);
 

test/mpi_rc2_test.c

@@ -961,11 +961,6 @@ RET:
                             } \
                         } while (0)
 
-static const char *name_of_gop_mode[] = {
-    "normal_p",
-    "smart_p",
-};
-
 static void *rc2_pre_dec_thread(void *param)
 {
     MpiRc2TestCtx *ctx = (MpiRc2TestCtx *)param;
@@ -1057,16 +1052,19 @@ static MPP_RET mpi_rc_codec(MpiRc2TestCtx *ctx)
             ctx->frm_idx, elapsed_time / 1000, frame_rate);
 
     if (ctx->frm_idx) {
-        mpp_log("%s: %s: average: bps %d | psnr %5.2f | ssim %5.5f", ctx->enc_cmd->file_input,
+        MpiEncTestArgs* enc_cmd = ctx->enc_cmd;
-                name_of_gop_mode[ctx->enc_cmd->gop_mode],
+
+        mpp_log("%s: %s: average: bps %d | psnr %5.2f | ssim %5.5f",
+                enc_cmd->file_input, enc_cmd->gop_mode ? "smart_p" : "normal_p",
                 30 * (RK_U32)(ctx->total_bits / ctx->frm_idx),
                 ctx->total_psnrs / ctx->frm_idx, ctx->total_ssims / ctx->frm_idx);
         if (ctx->file.fp_stat)
-            fprintf(ctx->file.fp_stat, "%s: %s: average: bps %dk | psnr %5.2f | ssim %5.5f \n", ctx->enc_cmd->file_input,
+            fprintf(ctx->file.fp_stat, "%s: %s: average: bps %dk | psnr %5.2f | ssim %5.5f \n",
-                    name_of_gop_mode[ctx->enc_cmd->gop_mode],
+                    enc_cmd->file_input, enc_cmd->gop_mode ? "smart_p" : "normal_p",
                     30 * (RK_U32)(ctx->total_bits / ctx->frm_idx) / 1000,
                     ctx->total_psnrs / ctx->frm_idx, ctx->total_ssims / ctx->frm_idx);
     }
+
     CHECK_RET(ctx->enc_mpi->reset(ctx->enc_ctx));
     CHECK_RET(ctx->dec_mpi_pre->reset(ctx->dec_ctx_pre));
     CHECK_RET(ctx->dec_mpi_post->reset(ctx->dec_ctx_post));