From 1b703a5eeafe755c59aaa8db2e76845571af1a56 Mon Sep 17 00:00:00 2001
From: Sushanth Patil <sushanth.patil@ittiam.com>
Date: Fri, 9 Aug 2019 13:38:16 +0530
Subject: [PATCH] Optimisation changes in mps apply function

Added changes to remove redundant computation
loops in ixheaacd_mps_synt_calc,mps hybrid
analysis & synthesis functions based on number
of non-zero qmf bands.
---
 decoder/ixheaacd_mps_dec.c          | 24 ++++++++++++++++--------
 decoder/ixheaacd_mps_dec.h          | 11 ++++-------
 decoder/ixheaacd_mps_hybrid_filt.c  |  2 +-
 decoder/ixheaacd_mps_parse.c        |  3 +--
 decoder/ixheaacd_mps_poly_filt.c    | 13 +++++++------
 decoder/ixheaacd_mps_pre_mix.c      |  6 +++---
 decoder/ixheaacd_mps_temp_process.c | 13 +++++++++++--
 decoder/ixheaacd_mps_temp_reshape.c |  8 ++++----
 decoder/ixheaacd_process.c          |  7 +++++++
 decoder/ixheaacd_sbr_dec.c          | 10 +++++++++-
 decoder/ixheaacd_sbr_dec.h          |  1 +
 decoder/ixheaacd_sbrdec_initfuncs.c |  1 +
 12 files changed, 65 insertions(+), 34 deletions(-)

diff --git a/decoder/ixheaacd_mps_dec.c b/decoder/ixheaacd_mps_dec.c
index 98f6646..4446b51 100644
--- a/decoder/ixheaacd_mps_dec.c
+++ b/decoder/ixheaacd_mps_dec.c
@@ -71,6 +71,8 @@
 #include "ixheaacd_mps_huff_tab.h"
 #include "ixheaacd_error_standards.h"
 
+#define max(a,b) (((a) > (b)) ? (a) : (b))
+
 extern const ia_huff_pt0_nodes_struct ixheaacd_huff_part0_nodes;
 extern const ia_huff_ipd_nodes_struct ixheaacd_huff_ipd_nodes;
 extern const ia_huff_lav_nodes_struct ixheaacd_huff_lav_idx_nodes;
@@ -126,7 +128,7 @@ WORD32 ixheaacd_mps_create(ia_mps_dec_state_struct* self, WORD32 bs_frame_len,
   if ((self->residual_coding) && (self->res_bands > 0))
     ixheaacd_mps_qmf_hybrid_analysis_init(&self->hyb_filt_state[1]);
 
-  err_code = ixheaacd_mps_decor_init(&(self->mps_decor), self->hyb_band_count,
+  err_code = ixheaacd_mps_decor_init(&(self->mps_decor), self->hyb_band_count_max,
                                      self->config->bs_decorr_config);
   if (err_code != IA_NO_ERROR) return err_code;
 
@@ -163,12 +165,12 @@ static const FLOAT32 ixheaacd_tsd_mul_im[] = {
 
 VOID ixheaacd_mps_qmf_hyb_analysis(ia_mps_dec_state_struct* self) {
   ixheaacd_mps_qmf_hybrid_analysis(&self->hyb_filt_state[0], self->qmf_in[0],
-                                   self->qmf_band_count, self->time_slots,
+                                   self->band_count[0], self->time_slots,
                                    self->hyb_in[0]);
 
   if ((self->residual_coding) && (self->res_bands > 0)) {
     ixheaacd_mps_qmf_hybrid_analysis(&self->hyb_filt_state[self->in_ch_count],
-                                     self->qmf_in[1], self->qmf_band_count,
+                                     self->qmf_in[1], self->band_count[1],
                                      self->time_slots, self->hyb_res);
   }
 }
@@ -178,7 +180,7 @@ VOID ixheaacd_mps_qmf_hyb_synthesis(ia_mps_dec_state_struct* self) {
 
   for (ch = 0; ch < self->out_ch_count; ch++) {
     ixheaacd_mps_qmf_hybrid_synthesis(self->hyb_dir_out[ch],
-                                      self->qmf_band_count, self->time_slots,
+                                      self->band_count[0], self->time_slots,
                                       self->qmf_out_dir[ch]);
   }
 }
@@ -235,7 +237,7 @@ VOID ixheaacd_mps_mix_res_decor(ia_mps_dec_state_struct* self) {
   WORD32 ts, qs, row, indx;
 
   for (ts = 0; ts < self->time_slots; ts++) {
-    for (qs = 0; qs < self->hyb_band_count; qs++) {
+    for (qs = 0; qs < self->hyb_band_count_max; qs++) {
       indx = self->hyb_band_to_processing_band_table[qs];
 
       for (row = 0; row < self->dir_sig_count; row++) {
@@ -281,14 +283,19 @@ WORD32 ixheaacd_mps_apply(ia_mps_dec_state_struct* self,
   WORD32 time_slots = self->time_slots;
   WORD32 in_ch_count = self->in_ch_count + self->res_ch_count;
   WORD32 err = 0;
-
+  self->hyb_band_count[0] = self->band_count[0] - QMF_BANDS_TO_HYBRID + 10;
+  self->hyb_band_count[1] = self->band_count[1] - QMF_BANDS_TO_HYBRID + 10;
+  self->hyb_band_count_max =
+      max(self->hyb_band_count[0], self->hyb_band_count[1]);
+  self->mps_decor.decor_nrg_smooth.num_bins = self->hyb_band_count_max;
+  self->mps_decor.num_bins = self->hyb_band_count_max;
   self->output_buffer = output_buffer;
 
   assert(self->present_time_slot + time_slots <= self->time_slots);
 
   for (ts = 0; ts < time_slots; ts++) {
     for (ch = 0; ch < in_ch_count; ch++) {
-      for (qs = 0; qs < self->qmf_band_count; qs++) {
+      for (qs = 0; qs < self->band_count[ch]; qs++) {
         self->qmf_in[ch][self->present_time_slot + ts][qs].re =
             self->input_gain * input_buffer[2 * ch][ts][qs];
         self->qmf_in[ch][self->present_time_slot + ts][qs].im =
@@ -306,12 +313,13 @@ WORD32 ixheaacd_mps_apply(ia_mps_dec_state_struct* self,
   err = ixheaacd_mps_frame_decode(self);
 
   if (err != 0) return err;
-  ixheaacd_mps_qmf_hyb_analysis(self);
 
   ixheaacd_pre_and_mix_matrix_calculation(self);
 
   ixheaacd_mps_pre_matrix_mix_matrix_smoothing(self);
 
+  ixheaacd_mps_qmf_hyb_analysis(self);
+
   err = ixheaacd_mps_apply_pre_matrix(self);
   if (err < 0) return err;
 
diff --git a/decoder/ixheaacd_mps_dec.h b/decoder/ixheaacd_mps_dec.h
index 66d34f4..77b140e 100644
--- a/decoder/ixheaacd_mps_dec.h
+++ b/decoder/ixheaacd_mps_dec.h
@@ -189,7 +189,6 @@ typedef struct ia_mps_dec_state_struct {
 
   WORD32 time_slots;
   WORD32 present_time_slot;
-  WORD32 frame_len;
 
   WORD32 temp_shape_enable_ch_stp[2];
   WORD32 temp_shape_enable_ch_ges[2];
@@ -197,9 +196,11 @@ typedef struct ia_mps_dec_state_struct {
   FLOAT32 env_shape_data[2][MAX_TIME_SLOTS];
 
   WORD8 parse_nxt_frame;
-
+  WORD32 band_count[MAX_M1_INPUT];
+  WORD32 synth_count;
   WORD32 qmf_band_count;
-  WORD32 hyb_band_count;
+  WORD32 hyb_band_count[MAX_M1_INPUT];
+  WORD32 hyb_band_count_max;
   const WORD32 *hyb_band_to_processing_band_table;
 
   WORD32 res_ch_count;
@@ -342,10 +343,6 @@ VOID ixheaacd_pre_and_mix_matrix_calculation(ia_mps_dec_state_struct *self);
 WORD32 ixheaacd_mps_apply_pre_matrix(ia_mps_dec_state_struct *self);
 WORD32 ixheaacd_mps_apply_mix_matrix(ia_mps_dec_state_struct *self);
 
-VOID ixheaacd_mps_config(ia_mps_dec_state_struct *self, WORD32 frame_len,
-                         WORD32 residual_coding,
-                         ia_usac_dec_mps_config_struct *mps212_config);
-
 WORD32 ixheaacd_mps_frame_decode(ia_mps_dec_state_struct *self);
 
 WORD32 ixheaacd_mps_header_decode(ia_mps_dec_state_struct *self);
diff --git a/decoder/ixheaacd_mps_hybrid_filt.c b/decoder/ixheaacd_mps_hybrid_filt.c
index f76df28..190ddb2 100644
--- a/decoder/ixheaacd_mps_hybrid_filt.c
+++ b/decoder/ixheaacd_mps_hybrid_filt.c
@@ -243,7 +243,7 @@ VOID ixheaacd_mps_qmf_hybrid_analysis(
     }
   }
 
-  for (k = 0; k < num_bands - lf_qmf_bands; k++) {
+  for (k = 0; k < MAX_NUM_QMF_BANDS_SAC / 2 - lf_qmf_bands; k++) {
     for (n = 0; n < hf_samples_shift; n++) {
       handle->hf_buffer[k][n].re = handle->hf_buffer[k][n + num_samples].re;
       handle->hf_buffer[k][n].im = handle->hf_buffer[k][n + num_samples].im;
diff --git a/decoder/ixheaacd_mps_parse.c b/decoder/ixheaacd_mps_parse.c
index 81a8592..46baada 100644
--- a/decoder/ixheaacd_mps_parse.c
+++ b/decoder/ixheaacd_mps_parse.c
@@ -994,10 +994,9 @@ WORD32 ixheaacd_mps_frame_decode(ia_mps_dec_state_struct *self) {
 
 WORD32 ixheaacd_mps_header_decode(ia_mps_dec_state_struct *self) {
   self->time_slots = self->frame_length;
-  self->frame_len = self->time_slots * self->qmf_band_count;
   self->bs_param_bands = ixheaacd_freq_res_table[self->config->bs_freq_res];
 
-  self->hyb_band_count = self->qmf_band_count - QMF_BANDS_TO_HYBRID + 10;
+  self->hyb_band_count_max = self->qmf_band_count - QMF_BANDS_TO_HYBRID + 10;
 
   switch (self->bs_param_bands) {
     case 4:
diff --git a/decoder/ixheaacd_mps_poly_filt.c b/decoder/ixheaacd_mps_poly_filt.c
index 315f0c7..2f3c325 100644
--- a/decoder/ixheaacd_mps_poly_filt.c
+++ b/decoder/ixheaacd_mps_poly_filt.c
@@ -134,9 +134,9 @@ VOID ixheaacd_mps_synt_out_calc_dec(WORD32 resolution, WORD32 *out,
 VOID ixheaacd_mps_synt_calc(ia_mps_dec_state_struct *self) {
   WORD32 k, l, ts, ch;
   WORD64 acc;
-  WORD32 ptr_in[128];
-  WORD32 fin_re[128];
-  WORD32 fin_im[128];
+  WORD32 ptr_in[128] = {0};
+  WORD32 fin_re[128] = {0};
+  WORD32 fin_im[128] = {0};
   FLOAT32 temp;
   WORD32 *state, *tmp_state, *out;
   const WORD32 *filt_coeff;
@@ -152,18 +152,19 @@ VOID ixheaacd_mps_synt_calc(ia_mps_dec_state_struct *self) {
 
     for (ts = 0; ts < self->time_slots; ts++) {
       ixheaacd_float_to_int32(&self->qmf_out_dir[ch][ts][0].re, ptr_in, 10,
-                              resolution * 2);
+                              self->synth_count * 2);
 
       filt_coeff = ixheaacd_mps_polyphase_filter_coeff_fix;
 
       state -= (2 * resolution);
       (*ixheaacd_mps_synt_pre_twiddle)(ptr_in, ixheaacd_mps_pre_re,
-                                       ixheaacd_mps_pre_im, resolution);
+                                       ixheaacd_mps_pre_im, self->synth_count);
 
       (*ixheaacd_mps_complex_fft_64)(ptr_in, fin_re, fin_im, resolution);
 
       (*ixheaacd_mps_synt_post_twiddle)(ptr_in, ixheaacd_mps_post_re,
-                                        ixheaacd_mps_post_im, resolution);
+                                        ixheaacd_mps_post_im,
+                                        self->synth_count);
 
       (*ixheaacd_mps_complex_fft_64)(ptr_in, &fin_re[1], &fin_im[1],
                                      resolution);
diff --git a/decoder/ixheaacd_mps_pre_mix.c b/decoder/ixheaacd_mps_pre_mix.c
index 0e5fa92..9bb08e1 100644
--- a/decoder/ixheaacd_mps_pre_mix.c
+++ b/decoder/ixheaacd_mps_pre_mix.c
@@ -381,7 +381,7 @@ WORD32 ixheaacd_mps_apply_pre_matrix(ia_mps_dec_state_struct *self) {
         self->v[row][ts][qs].im = sum_imag;
       }
     }
-    for (qs = 2; qs < self->hyb_band_count; qs++) {
+    for (qs = 2; qs < self->hyb_band_count[0]; qs++) {
       WORD32 sign = 1;
       WORD32 indx = self->hyb_band_to_processing_band_table[qs];
       for (row = 0; row < (self->dir_sig_count + self->decor_sig_count);
@@ -469,7 +469,7 @@ WORD32 ixheaacd_mps_apply_mix_matrix(ia_mps_dec_state_struct *self) {
   }
 
   for (ts = 0; ts < self->time_slots; ts++) {
-    for (qs = 0; qs < self->hyb_band_count; qs++) {
+    for (qs = 0; qs < self->hyb_band_count_max; qs++) {
       WORD32 indx = self->hyb_band_to_processing_band_table[qs];
       for (row = 0; row < self->out_ch_count; row++) {
         FLOAT32 sum_re_dir = 0;
@@ -521,7 +521,7 @@ WORD32 ixheaacd_mps_apply_mix_matrix(ia_mps_dec_state_struct *self) {
           self->hyb_diff_out[row][ts][qs].im = sum_im_diff;
         }
       }
-      for (qs = 2; qs < self->hyb_band_count; qs++) {
+      for (qs = 2; qs < self->hyb_band_count_max; qs++) {
         WORD32 indx = self->hyb_band_to_processing_band_table[qs];
         for (row = 0; row < self->out_ch_count; row++) {
           FLOAT32 sum_re_dir = self->hyb_dir_out[row][ts][qs].re;
diff --git a/decoder/ixheaacd_mps_temp_process.c b/decoder/ixheaacd_mps_temp_process.c
index 736996c..964735a 100644
--- a/decoder/ixheaacd_mps_temp_process.c
+++ b/decoder/ixheaacd_mps_temp_process.c
@@ -182,7 +182,7 @@ static VOID ixheaacd_mps_subbandtp(ia_mps_dec_state_struct* self, WORD32 ts) {
       no_scaling = !self->temp_shape_enable_ch_stp[ch];
 
     if (no_scaling == 1) {
-      for (n = 0; n < self->hyb_band_count; n++) {
+      for (n = 0; n < self->hyb_band_count_max; n++) {
         self->hyb_dir_out[ch][ts][n].re += self->hyb_diff_out[ch][ts][n].re;
         self->hyb_dir_out[ch][ts][n].im += self->hyb_diff_out[ch][ts][n].im;
       }
@@ -202,7 +202,7 @@ static VOID ixheaacd_mps_subbandtp(ia_mps_dec_state_struct* self, WORD32 ts) {
         self->hyb_dir_out[ch][ts][n].im +=
             (self->hyb_diff_out[ch][ts][n].im * temp);
       }
-      for (; n < self->hyb_band_count; n++) {
+      for (; n < self->hyb_band_count_max; n++) {
         temp = (FLOAT32)(scale[ch]);
         self->hyb_dir_out[ch][ts][n].re +=
             (self->hyb_diff_out[ch][ts][n].re * temp);
@@ -216,6 +216,8 @@ static VOID ixheaacd_mps_subbandtp(ia_mps_dec_state_struct* self, WORD32 ts) {
 WORD32 ixheaacd_mps_temp_process(ia_mps_dec_state_struct* self) {
   WORD32 ch, ts, hyb;
   WORD32 err = 0;
+  ia_sbr_frame_info_data_struct* ptr_frame_data =
+      (ia_sbr_frame_info_data_struct*)self->p_sbr_frame[0];
 
   for (ch = 0; ch < self->out_ch_count; ch++) {
     for (ts = 0; ts < self->time_slots; ts++) {
@@ -239,6 +241,13 @@ WORD32 ixheaacd_mps_temp_process(ia_mps_dec_state_struct* self) {
     if (err) return err;
   }
 
+  if (ptr_frame_data->mps_sbr_flag) {
+    self->synth_count =
+        ptr_frame_data->pstr_sbr_header->pstr_freq_band_data->sub_band_end;
+  } else {
+    self->synth_count = self->band_count[0];
+  }
+
   ixheaacd_mps_synt_calc(self);
   return err;
 }
\ No newline at end of file
diff --git a/decoder/ixheaacd_mps_temp_reshape.c b/decoder/ixheaacd_mps_temp_reshape.c
index fd8e62c..1797539 100644
--- a/decoder/ixheaacd_mps_temp_reshape.c
+++ b/decoder/ixheaacd_mps_temp_reshape.c
@@ -67,7 +67,7 @@ static VOID ixheaacd_mps_est_normalized_envelope(ia_mps_dec_state_struct *self,
     case DIR_DIFF_IN:
       ch_offset = 0;
       for (ii = 0; ii < self->time_slots; ii++) {
-        for (jj = 0; jj < self->hyb_band_count; jj++) {
+        for (jj = 0; jj < self->hyb_band_count_max; jj++) {
           slot_energy[ii]
                      [ixheaacd_hybrid_band_71_to_processing_band_20_map[jj]] +=
               ((self->hyb_dir_out[ch][ii][jj].re +
@@ -84,7 +84,7 @@ static VOID ixheaacd_mps_est_normalized_envelope(ia_mps_dec_state_struct *self,
     case DOWNMIX_IN:
       ch_offset = self->out_ch_count;
       for (ii = 0; ii < self->time_slots; ii++) {
-        for (jj = 0; jj < self->hyb_band_count; jj++) {
+        for (jj = 0; jj < self->hyb_band_count_max; jj++) {
           slot_energy[ii]
                      [ixheaacd_hybrid_band_71_to_processing_band_20_map[jj]] +=
               self->hyb_in[ch][ii][jj].re * self->hyb_in[ch][ii][jj].re +
@@ -168,7 +168,7 @@ VOID ixheaacd_mps_time_env_shaping(ia_mps_dec_state_struct *self) {
         amp_direct = 0;
         amp_diff = 0;
 
-        for (jj = band_start; jj < self->hyb_band_count; jj++) {
+        for (jj = band_start; jj < self->hyb_band_count_max; jj++) {
           amp_direct += self->hyb_dir_out[ch][time_slot][jj].re *
                             self->hyb_dir_out[ch][time_slot][jj].re +
                         self->hyb_dir_out[ch][time_slot][jj].im *
@@ -184,7 +184,7 @@ VOID ixheaacd_mps_time_env_shaping(ia_mps_dec_state_struct *self) {
 
         ratio = min(max((gain + amp_ratio * (gain - 1)), 1 / LAMDA), LAMDA);
 
-        for (jj = band_start; jj < self->hyb_band_count; jj++) {
+        for (jj = band_start; jj < self->hyb_band_count_max; jj++) {
           self->hyb_dir_out[ch][time_slot][jj].re *= ratio;
           self->hyb_dir_out[ch][time_slot][jj].im *= ratio;
         }
diff --git a/decoder/ixheaacd_process.c b/decoder/ixheaacd_process.c
index 7ace0ca..672f4c6 100644
--- a/decoder/ixheaacd_process.c
+++ b/decoder/ixheaacd_process.c
@@ -383,6 +383,13 @@ WORD32 ixheaacd_usac_process(ia_dec_data_struct *pstr_dec_data,
               pstr_usac_data->usac_independency_flg, it_bit_buff);
           if (err) return err;
 
+          p_state_aac_dec->mps_dec_handle.band_count[0] =
+              pstr_usac_data->pstr_esbr_dec->pstr_sbr_channel[0]
+                  ->str_sbr_dec.band_count;
+          p_state_aac_dec->mps_dec_handle.band_count[1] =
+              pstr_usac_data->pstr_esbr_dec->pstr_sbr_channel[1]
+                  ->str_sbr_dec.band_count;
+
           for (ch = 0; ch < nr_core_coder_channels; ch++) {
             ptr_inp[2 * ch] =
                 pstr_usac_data->pstr_esbr_dec->pstr_sbr_channel[ch]
diff --git a/decoder/ixheaacd_sbr_dec.c b/decoder/ixheaacd_sbr_dec.c
index 671f4d6..d0bac14 100644
--- a/decoder/ixheaacd_sbr_dec.c
+++ b/decoder/ixheaacd_sbr_dec.c
@@ -735,7 +735,6 @@ WORD32 ixheaacd_sbr_dec(ia_sbr_dec_struct *ptr_sbr_dec, WORD16 *ptr_time_data,
             ptr_sbr_dec->p_hbe_txposer->max_stretch);
       }
     }
-    ixheaacd_qmf_enrg_calc(ptr_sbr_dec, upsample_ratio_idx, low_pow_flag);
 
     if (!mps_sbr_flag && apply_processing) {
       WORD32 err_code = 0;
@@ -752,6 +751,7 @@ WORD32 ixheaacd_sbr_dec(ia_sbr_dec_struct *ptr_sbr_dec, WORD16 *ptr_time_data,
       ptr_pvc_data->pvc_rate = ptr_header_data->upsamp_fac;
 
       if (sbr_mode == PVC_SBR) {
+        ixheaacd_qmf_enrg_calc(ptr_sbr_dec, upsample_ratio_idx, low_pow_flag);
         err_code = ixheaacd_pvc_process(
             ptr_pvc_data, ptr_header_data->pstr_freq_band_data->sub_band_start,
             ptr_frame_data->str_pvc_frame_info.border_vec[0],
@@ -788,6 +788,12 @@ WORD32 ixheaacd_sbr_dec(ia_sbr_dec_struct *ptr_sbr_dec, WORD16 *ptr_time_data,
       }
     }
 
+    if (!mps_sbr_flag) {
+      ptr_sbr_dec->band_count =
+          ptr_header_data->pstr_freq_band_data->sub_band_end;
+    } else
+      ptr_sbr_dec->band_count = ptr_sbr_dec->str_codec_qmf_bank.no_channels;
+
     ixheaacd_esbr_synthesis_filt_block(
         ptr_sbr_dec, ptr_header_data, ptr_frame_data, apply_processing,
         pp_qmf_buf_real, pp_qmf_buf_imag, stereo_config_idx, sbr_tables_ptr,
@@ -1135,6 +1141,8 @@ WORD32 ixheaacd_esbr_dec(ia_sbr_dec_struct *ptr_sbr_dec,
     memset(ptr_sbr_dec->sbr_qmf_out_imag[i], 0, 64 * sizeof(FLOAT32));
   }
 
+  ptr_sbr_dec->band_count = ptr_sbr_dec->str_codec_qmf_bank.no_channels;
+
   ixheaacd_esbr_synthesis_filt_block(
       ptr_sbr_dec, ptr_header_data, ptr_frame_data, apply_processing,
       pp_qmf_buf_real, pp_qmf_buf_imag, stereo_config_idx, ptr_sbr_tables,
diff --git a/decoder/ixheaacd_sbr_dec.h b/decoder/ixheaacd_sbr_dec.h
index 0461d39..8bafa92 100644
--- a/decoder/ixheaacd_sbr_dec.h
+++ b/decoder/ixheaacd_sbr_dec.h
@@ -67,6 +67,7 @@ typedef struct {
   ia_sbr_scale_fact_struct str_sbr_scale_fact;
 
   WORD32 max_samp_val;
+  WORD32 band_count;
   ia_esbr_hbe_txposer_struct *p_hbe_txposer;
 
   FLOAT32 core_sample_buf[2624];
diff --git a/decoder/ixheaacd_sbrdec_initfuncs.c b/decoder/ixheaacd_sbrdec_initfuncs.c
index f779c84..2793fe5 100644
--- a/decoder/ixheaacd_sbrdec_initfuncs.c
+++ b/decoder/ixheaacd_sbrdec_initfuncs.c
@@ -543,6 +543,7 @@ ia_handle_sbr_dec_inst_struct ixheaacd_init_sbr(
         audio_object_type);
 
     ptr_header_data[i]->status = 1;
+    ptr_sbr_dec[i]->band_count = 64;
 
     if (err) {
       return NULL;