/******************************************************************************
*
* Copyright (C) 2015 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at:
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*
*****************************************************************************
* Originally developed and contributed by Ittiam Systems Pvt. Ltd, Bangalore
*/
/****************************************************************************/
/* File Name : irc_rd_model.c */
/* */
/* Description : Implall the Functions to Model the */
/* Rate Distortion Behaviour of the Codec over the Last */
/* Few Frames. */
/* */
/* List of Functions : irc_update_frame_rd_model */
/* estimate_mpeg2_qp_for_resbits */
/* */
/* Issues / Problems : None */
/* */
/* Revision History : */
/* DD MM YYYY Author(s) Changes (Describe the changes made) */
/* 21 06 2006 Sarat Initial Version */
/****************************************************************************/
/* System include files */
#include <stdarg.h>
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include "math.h"
/* User include files */
#include "irc_datatypes.h"
#include "irc_common.h"
#include "irc_mem_req_and_acq.h"
#include "irc_rd_model.h"
#include "irc_rd_model_struct.h"
WORD32 irc_rd_model_num_fill_use_free_memtab(rc_rd_model_t **pps_rc_rd_model,
itt_memtab_t *ps_memtab,
ITT_FUNC_TYPE_E e_func_type)
{
WORD32 i4_mem_tab_idx = 0;
rc_rd_model_t s_rc_rd_model_temp;
/*
* Hack for al alloc, during which we don't have any state memory.
* Dereferencing can cause issues
*/
if(e_func_type == GET_NUM_MEMTAB || e_func_type == FILL_MEMTAB)
(*pps_rc_rd_model) = &s_rc_rd_model_temp;
/*for src rate control state structure*/
if(e_func_type != GET_NUM_MEMTAB)
{
fill_memtab(&ps_memtab[i4_mem_tab_idx], sizeof(rc_rd_model_t),
ALIGN_128_BYTE, PERSISTENT, DDR);
use_or_fill_base(&ps_memtab[0], (void**)pps_rc_rd_model, e_func_type);
}
i4_mem_tab_idx++;
return (i4_mem_tab_idx);
}
void irc_init_frm_rc_rd_model(rc_rd_model_t *ps_rd_model,
UWORD8 u1_max_frames_modelled)
{
ps_rd_model->u1_num_frms_in_model = 0;
ps_rd_model->u1_curr_frm_counter = 0;
ps_rd_model->u1_max_frms_to_model = u1_max_frames_modelled;
ps_rd_model->model_coeff_a_lin_wo_int = 0;
ps_rd_model->model_coeff_b_lin_wo_int = 0;
ps_rd_model->model_coeff_c_lin_wo_int = 0;
}
void irc_reset_frm_rc_rd_model(rc_rd_model_t *ps_rd_model)
{
ps_rd_model->u1_num_frms_in_model = 0;
ps_rd_model->u1_curr_frm_counter = 0;
ps_rd_model->model_coeff_a_lin_wo_int = 0;
ps_rd_model->model_coeff_b_lin_wo_int = 0;
ps_rd_model->model_coeff_c_lin_wo_int = 0;
}
static UWORD8 find_model_coeffs(UWORD32 *pi4_res_bits,
UWORD32 *pi4_sad_h264,
UWORD8 *pu1_num_skips,
UWORD8 *pui_avg_mpeg2_qp,
UWORD8 u1_num_frms,
UWORD8 u1_model_used,
WORD8 *pi1_frame_index,
model_coeff *pmc_model_coeff,
model_coeff *pmc_model_coeff_lin,
model_coeff *pmc_model_coeff_lin_wo_int,
rc_rd_model_t *ps_rd_model)
{
UWORD32 i;
UWORD8 u1_num_frms_used = 0;
UWORD8 u1_frm_indx;
float sum_y = 0;
float sum_x_y = 0;
float sum_x2_y = 0;
float sum_x = 0;
float sum_x2 = 0;
float sum_x3 = 0;
float sum_x4 = 0;
float x0, y0;
float model_coeff_a = 0.0, model_coeff_b = 0.0, model_coeff_c = 0.0;
#if !(ENABLE_QUAD_RC_MODEL||ENABLE_LIN_MODEL_WITH_INTERCEPT)
UNUSED(pu1_num_skips);
UNUSED(pmc_model_coeff);
UNUSED(pmc_model_coeff_lin);
#endif
for(i = 0; i < u1_num_frms; i++)
{
if(-1 == pi1_frame_index[i])
continue;
u1_frm_indx = (UWORD8)pi1_frame_index[i];
y0 = (float)(pi4_res_bits[u1_frm_indx]);
x0 = (float)(pi4_sad_h264[u1_frm_indx]
/ (float)pui_avg_mpeg2_qp[u1_frm_indx]);
sum_y += y0;
sum_x_y += x0 * y0;
sum_x2_y += x0 * x0 * y0;
sum_x += x0;
sum_x2 += x0 * x0;
sum_x3 += x0 * x0 * x0;
sum_x4 += x0 * x0 * x0 * x0;
u1_num_frms_used++;
}
sum_y /= u1_num_frms_used;
sum_x_y /= u1_num_frms_used;
sum_x2_y /= u1_num_frms_used;
sum_x /= u1_num_frms_used;
sum_x2 /= u1_num_frms_used;
sum_x3 /= u1_num_frms_used;
sum_x4 /= u1_num_frms_used;
{
UWORD8 u1_curr_frame_index;
UWORD8 u1_avgqp_prvfrm;
UWORD32 u4_prevfrm_bits, u4_prevfrm_sad;
u1_curr_frame_index = ps_rd_model->u1_curr_frm_counter;
if(0 == u1_curr_frame_index)
u1_curr_frame_index = (MAX_FRAMES_MODELLED - 1);
else
u1_curr_frame_index--;
u1_avgqp_prvfrm = ps_rd_model->pu1_avg_qp[u1_curr_frame_index];
u4_prevfrm_bits = ps_rd_model->pi4_res_bits[u1_curr_frame_index];
u4_prevfrm_sad = ps_rd_model->pi4_sad[u1_curr_frame_index];
if(0 != u4_prevfrm_sad)
model_coeff_a = (float)(u4_prevfrm_bits * u1_avgqp_prvfrm)
/ u4_prevfrm_sad;
else
model_coeff_a = 0;
model_coeff_b = 0;
model_coeff_c = 0;
pmc_model_coeff_lin_wo_int[0] = model_coeff_b;
pmc_model_coeff_lin_wo_int[1] = model_coeff_a;
pmc_model_coeff_lin_wo_int[2] = model_coeff_c;
}
return u1_model_used;
}
static void irc_update_frame_rd_model(rc_rd_model_t *ps_rd_model)
{
WORD8 pi1_frame_index[MAX_FRAMES_MODELLED],
pi1_frame_index_initial[MAX_FRAMES_MODELLED];
UWORD8 u1_num_skips_temp;
UWORD8 u1_avg_mpeg2_qp_temp, u1_min_mpeg2_qp, u1_max_mpeg2_qp;
UWORD8 u1_num_frms_input, u1_num_active_frames, u1_reject_frame;
UWORD32 u4_num_skips;
UWORD8 u1_min2_mpeg2_qp, u1_max2_mpeg2_qp;
UWORD8 u1_min_qp_frame_indx, u1_max_qp_frame_indx;
UWORD8 pu1_num_frames[MPEG2_QP_ELEM];
model_coeff model_coeff_array[3], model_coeff_array_lin[3],
model_coeff_array_lin_wo_int[3];
UWORD32 i;
UWORD8 u1_curr_frame_index;
u1_curr_frame_index = ps_rd_model->u1_curr_frm_counter;
ps_rd_model->u1_model_used = PREV_FRAME_MODEL;
if(0 == u1_curr_frame_index)
u1_curr_frame_index = (MAX_FRAMES_MODELLED - 1);
else
u1_curr_frame_index--;
/************************************************************************/
/* Rearrange data to be fed into a Linear Regression Module */
/* Module finds a,b,c such that */
/* y = ax + bx^2 + c */
/************************************************************************/
u4_num_skips = 0;
u1_num_frms_input = 0;
memset(pu1_num_frames, 0, MPEG2_QP_ELEM);
memset(pi1_frame_index, -1, MAX_FRAMES_MODELLED);
u1_min_mpeg2_qp = MAX_MPEG2_QP;
u1_max_mpeg2_qp = 0;
u1_num_active_frames = ps_rd_model->u1_num_frms_in_model;
if(u1_num_active_frames > MAX_ACTIVE_FRAMES)
{
u1_num_active_frames = MAX_ACTIVE_FRAMES;
}
/************************************************************************/
/* Choose the set of Points to be used for MSE fit of Quadratic model */
/* Points chosen are spread across the Qp range. Max of 2 points are */
/* chosen for a Qp. */
/************************************************************************/
for(i = 0; i < u1_num_active_frames; i++)
{
u1_reject_frame = 0;
u1_num_skips_temp = ps_rd_model->pu1_num_skips[u1_curr_frame_index];
u1_avg_mpeg2_qp_temp = ps_rd_model->pu1_avg_qp[u1_curr_frame_index];
if((0 == u4_num_skips) && (0 != u1_num_skips_temp))
u1_reject_frame = 1;
if((1 == u4_num_skips) && (u1_num_skips_temp > 1))
u1_reject_frame = 1;
if(pu1_num_frames[u1_avg_mpeg2_qp_temp] >= 2)
u1_reject_frame = 1;
if(0 == i)
u1_reject_frame = 0;
if(0 == u1_reject_frame)
{
pi1_frame_index[u1_num_frms_input] = (WORD8)u1_curr_frame_index;
pu1_num_frames[u1_avg_mpeg2_qp_temp] += 1;
if(u1_min_mpeg2_qp > u1_avg_mpeg2_qp_temp)
u1_min_mpeg2_qp = u1_avg_mpeg2_qp_temp;
if(u1_max_mpeg2_qp < u1_avg_mpeg2_qp_temp)
u1_max_mpeg2_qp = u1_avg_mpeg2_qp_temp;
u1_num_frms_input++;
}
if(0 == u1_curr_frame_index)
u1_curr_frame_index = (MAX_FRAMES_MODELLED - 1);
else
u1_curr_frame_index--;
}
/************************************************************************/
/* Add Pivot Points to the Data set to be used for finding Quadratic */
/* Model Coeffs. These will help in constraining the shape of Quadratic*/
/* to adapt too much to the Local deviations. */
/************************************************************************/
u1_min2_mpeg2_qp = u1_min_mpeg2_qp;
u1_max2_mpeg2_qp = u1_max_mpeg2_qp;
u1_min_qp_frame_indx = INVALID_FRAME_INDEX;
u1_max_qp_frame_indx = INVALID_FRAME_INDEX;
/* Loop runnning over the Stored Frame Level Data
to find frames of MinQp and MaxQp */
for(; i < ps_rd_model->u1_num_frms_in_model; i++)
{
u1_num_skips_temp = ps_rd_model->pu1_num_skips[u1_curr_frame_index];
u1_avg_mpeg2_qp_temp = ps_rd_model->pu1_avg_qp[u1_curr_frame_index];
if(((0 == u4_num_skips) && (0 != u1_num_skips_temp))
|| ((1 == u4_num_skips) && (u1_num_skips_temp > 1)))
continue;
if(u1_min2_mpeg2_qp > u1_avg_mpeg2_qp_temp)
{
u1_min2_mpeg2_qp = u1_avg_mpeg2_qp_temp;
u1_min_qp_frame_indx = u1_curr_frame_index;
}
if(u1_max2_mpeg2_qp < u1_avg_mpeg2_qp_temp)
{
u1_max2_mpeg2_qp = u1_avg_mpeg2_qp_temp;
u1_max_qp_frame_indx = u1_curr_frame_index;
}
if(0 == u1_curr_frame_index)
u1_curr_frame_index = (MAX_FRAMES_MODELLED - 1);
else
u1_curr_frame_index--;
}
/* Add the Chosen Points to the regression data set */
if(INVALID_FRAME_INDEX != u1_min_qp_frame_indx)
{
pi1_frame_index[u1_num_frms_input] = (WORD8)u1_min_qp_frame_indx;
u1_num_frms_input++;
}
if(INVALID_FRAME_INDEX != u1_max_qp_frame_indx)
{
pi1_frame_index[u1_num_frms_input] = (WORD8)u1_max_qp_frame_indx;
u1_num_frms_input++;
}
memcpy(pi1_frame_index_initial, pi1_frame_index, MAX_FRAMES_MODELLED);
/***** Call the Module to Return the Coeffs for the Fed Data *****/
ps_rd_model->u1_model_used = find_model_coeffs(ps_rd_model->pi4_res_bits,
ps_rd_model->pi4_sad,
ps_rd_model->pu1_num_skips,
ps_rd_model->pu1_avg_qp,
u1_num_frms_input,
ps_rd_model->u1_model_used,
pi1_frame_index,
model_coeff_array,
model_coeff_array_lin,
model_coeff_array_lin_wo_int,
ps_rd_model);
ps_rd_model->model_coeff_b_lin_wo_int = model_coeff_array_lin_wo_int[0];
ps_rd_model->model_coeff_a_lin_wo_int = model_coeff_array_lin_wo_int[1];
ps_rd_model->model_coeff_c_lin_wo_int = model_coeff_array_lin_wo_int[2];
}
UWORD32 irc_estimate_bits_for_qp(rc_rd_model_t *ps_rd_model,
UWORD32 u4_estimated_sad,
UWORD8 u1_avg_qp)
{
float fl_num_bits = 0;
fl_num_bits = ps_rd_model->model_coeff_a_lin_wo_int
* ((float)(u4_estimated_sad / u1_avg_qp));
return ((UWORD32)fl_num_bits);
}
UWORD8 irc_find_qp_for_target_bits(rc_rd_model_t *ps_rd_model,
UWORD32 u4_target_res_bits,
UWORD32 u4_estimated_sad,
UWORD8 u1_min_qp,
UWORD8 u1_max_qp)
{
UWORD8 u1_qp;
float x_value = 1.0, f_qp;
ps_rd_model->u1_model_used = PREV_FRAME_MODEL;
{
x_value = (float)u4_target_res_bits
/ ps_rd_model->model_coeff_a_lin_wo_int;
}
if(0 != x_value)
f_qp = u4_estimated_sad / x_value;
else
f_qp = 255;
if(f_qp > 255)
f_qp = 255;
/* Truncating the QP to the Max and Min Qp values possible */
if(f_qp < u1_min_qp)
f_qp = u1_min_qp;
if(f_qp > u1_max_qp)
f_qp = u1_max_qp;
u1_qp = (UWORD8)(f_qp + 0.5);
return u1_qp;
}
void irc_add_frame_to_rd_model(rc_rd_model_t *ps_rd_model,
UWORD32 i4_res_bits,
UWORD8 u1_avg_mp2qp,
UWORD32 i4_sad_h264,
UWORD8 u1_num_skips)
{
UWORD8 u1_curr_frame_index;
u1_curr_frame_index = ps_rd_model->u1_curr_frm_counter;
/*Insert the Present Frame Data into the RD Model State Memory*/
ps_rd_model->pi4_res_bits[u1_curr_frame_index] = i4_res_bits;
ps_rd_model->pi4_sad[u1_curr_frame_index] = i4_sad_h264;
ps_rd_model->pu1_num_skips[u1_curr_frame_index] = u1_num_skips;
ps_rd_model->pu1_avg_qp[u1_curr_frame_index] = u1_avg_mp2qp;
ps_rd_model->u1_curr_frm_counter++;
if(MAX_FRAMES_MODELLED == ps_rd_model->u1_curr_frm_counter)
ps_rd_model->u1_curr_frm_counter = 0;
if(ps_rd_model->u1_num_frms_in_model < ps_rd_model->u1_max_frms_to_model)
{
ps_rd_model->u1_num_frms_in_model++;
}
irc_update_frame_rd_model(ps_rd_model);
}
/*****************************************************************************
*Function Name : irc_calc_per_frm_bits
*Description :
*Inputs : pu2_num_pics_of_a_pic_type
* - pointer to RC api pointer
* pu2_num_pics_of_a_pic_type
* - N1, N2,...Nk
* pu1_update_pic_type_model
* - flag which tells whether or not to update model
* coefficients of a particular pic-type
* u1_num_pic_types
* - value of k
* pu4_num_skip_of_a_pic_type
* - the number of skips of that pic-type. It "may" be used to
* update the model coefficients at a later point. Right now
* it is not being used at all.
* u1_base_pic_type
* - base pic type index wrt which alpha & beta are calculated
* pfl_gamma
* - gamma_i = beta_i / alpha_i
* pfl_eta
* -
* u1_curr_pic_type
* - the current pic-type for which the targetted bits need to
* be computed
* u4_bits_for_sub_gop
* - the number of bits to be consumed for the remaining part of
* sub-gop
* u4_curr_estimated_sad
* -
* pu1_curr_pic_type_qp
* - output of this function
*****************************************************************************/
WORD32 irc_calc_per_frm_bits(rc_rd_model_t *ps_rd_model,
UWORD16 *pu2_num_pics_of_a_pic_type,
UWORD8 *pu1_update_pic_type_model,
UWORD8 u1_num_pic_types,
UWORD32 *pu4_num_skip_of_a_pic_type,
UWORD8 u1_base_pic_type,
float *pfl_gamma,
float *pfl_eta,
UWORD8 u1_curr_pic_type,
UWORD32 u4_bits_for_sub_gop,
UWORD32 u4_curr_estimated_sad,
UWORD8 *pu1_curr_pic_type_qp)
{
WORD32 i4_per_frm_bits_Ti;
UWORD8 u1_i;
rc_rd_model_t *ps_rd_model_of_pic_type;
UNUSED(pu4_num_skip_of_a_pic_type);
UNUSED(u1_base_pic_type);
/* First part of this function updates all the model coefficients */
/*for all the pic-types */
{
for(u1_i = 0; u1_i < u1_num_pic_types; u1_i++)
{
if((0 != pu2_num_pics_of_a_pic_type[u1_i])
&& (1 == pu1_update_pic_type_model[u1_i]))
{
irc_update_frame_rd_model(&ps_rd_model[u1_i]);
}
}
}
/*
* The second part of this function deals with solving the
* equation using all the pic-types models
*/
{
UWORD8 u1_combined_model_used;
/* solve the equation */
{
model_coeff eff_A;
float fl_sad_by_qp_base;
float fl_sad_by_qp_curr_frm = 1.0;
float fl_qp_curr_frm;
float fl_bits_for_curr_frm = 0;
/* If the combined chosen model is linear model without an intercept */
u1_combined_model_used = PREV_FRAME_MODEL;
{
eff_A = 0.0;
for(u1_i = 0; u1_i < u1_num_pic_types; u1_i++)
{
ps_rd_model_of_pic_type = ps_rd_model + u1_i;
eff_A += ((pfl_eta[u1_i]
+ pu2_num_pics_of_a_pic_type[u1_i]- 1)
* ps_rd_model_of_pic_type->model_coeff_a_lin_wo_int
* pfl_gamma[u1_i]);
}
fl_sad_by_qp_base = u4_bits_for_sub_gop / eff_A;
fl_sad_by_qp_curr_frm = fl_sad_by_qp_base
* pfl_gamma[u1_curr_pic_type]
* pfl_eta[u1_curr_pic_type];
ps_rd_model_of_pic_type = ps_rd_model + u1_curr_pic_type;
fl_bits_for_curr_frm =
ps_rd_model_of_pic_type->model_coeff_a_lin_wo_int
* fl_sad_by_qp_curr_frm;
}
/*
* Store the model that was finally used to calculate Qp.
* This is so that the same model is used in further calculations
* for this picture.
*/
ps_rd_model_of_pic_type = ps_rd_model + u1_curr_pic_type;
ps_rd_model_of_pic_type->u1_model_used = u1_combined_model_used;
i4_per_frm_bits_Ti = (WORD32)(fl_bits_for_curr_frm + 0.5);
if(fl_sad_by_qp_curr_frm > 0)
fl_qp_curr_frm = (float)u4_curr_estimated_sad
/ fl_sad_by_qp_curr_frm;
else
fl_qp_curr_frm = 255;
if(fl_qp_curr_frm > 255)
fl_qp_curr_frm = 255;
*pu1_curr_pic_type_qp = (fl_qp_curr_frm + 0.5);
}
}
return (i4_per_frm_bits_Ti);
}
model_coeff irc_get_linear_coefficient(rc_rd_model_t *ps_rd_model)
{
return (ps_rd_model->model_coeff_a_lin_wo_int);
}