Apply kmeans on log of wiener_variance

Change-Id: Ifc0068902e036cd5d94fc5d39f0349a5875469b7

3 files changed, 45 insertions, 10 deletions

diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index 3a526fe0a..22b5f06ac 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -3583,6 +3583,8 @@ static int wiener_var_rdmult(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
   int row, col;
   int64_t rdmult;
   int64_t wiener_variance = 0;
+  KMEANS_DATA *kmeans_data;
+  vpx_clear_system_state();
 
   assert(cpi->norm_wiener_variance > 0);
 
@@ -3590,10 +3592,12 @@ static int wiener_var_rdmult(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
     for (col = mb_col_start; col < mb_col_end; ++col)
       wiener_variance += cpi->mb_wiener_variance[row * cm->mb_cols + col];
 
+  kmeans_data = &cpi->kmeans_data_arr[cpi->kmeans_data_size++];
+  kmeans_data->value = log(1 + wiener_variance);
+  kmeans_data->pos = mi_row * cpi->kmeans_data_stride + mi_col;
   if (wiener_variance)
     wiener_variance /=
         (mb_row_end - mb_row_start) * (mb_col_end - mb_col_start);
-
   rdmult = (orig_rdmult * wiener_variance) / cpi->norm_wiener_variance;
 
   rdmult = VPXMIN(rdmult, orig_rdmult * 3);
@@ -5673,14 +5677,6 @@ static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
 }
 #endif
 
-#define MAX_KMEANS_GROUPS 8
-
-typedef struct KMEANS_DATA {
-  int64_t value;
-  int pos;
-  int group_idx;
-} KMEANS_DATA;
-
 static int compare_kmeans_data(const void *a, const void *b) {
   if (((const KMEANS_DATA *)a)->value > ((const KMEANS_DATA *)b)->value) {
     return 1;
@@ -5693,7 +5689,7 @@ static int compare_kmeans_data(const void *a, const void *b) {
 }
 
 void vp9_kmeans(double *ctr_ls, int k, KMEANS_DATA *arr, int size) {
-  int64_t min, max;
+  double min, max;
   double step;
   int i, j;
   int itr;
@@ -5873,6 +5869,11 @@ static void encode_frame_internal(VP9_COMP *cpi) {
     }
 #endif
 
+    if (cpi->sf.enable_wiener_variance && cm->show_frame) {
+      cpi->kmeans_data_size = 0;
+      cpi->kmeans_ctr_num = 5;
+    }
+
     if (!cpi->row_mt) {
       cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read_dummy;
       cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write_dummy;
@@ -5888,6 +5889,11 @@ static void encode_frame_internal(VP9_COMP *cpi) {
       vp9_encode_tiles_row_mt(cpi);
     }
 
+    if (cpi->sf.enable_wiener_variance && cm->show_frame) {
+      vp9_kmeans(cpi->kmeans_ctr_ls, cpi->kmeans_ctr_num, cpi->kmeans_data_arr,
+                 cpi->kmeans_data_size);
+    }
+
     vpx_usec_timer_mark(&emr_timer);
     cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
   }
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index fb462c0cb..5bf0d9610 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -2388,6 +2388,7 @@ VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
                                sizeof(*cpi->mb_wiener_variance)));
   }
 
+  cpi->kmeans_data_arr_alloc = 0;
 #if CONFIG_NON_GREEDY_MV
   cpi->feature_score_loc_alloc = 0;
   cpi->tpl_ready = 0;
@@ -2591,6 +2592,10 @@ void vp9_remove_compressor(VP9_COMP *cpi) {
   vp9_denoiser_free(&(cpi->denoiser));
 #endif
 
+  if (cpi->kmeans_data_arr_alloc) {
+    vpx_free(cpi->kmeans_data_arr);
+  }
+
 #if CONFIG_NON_GREEDY_MV
   vpx_free(cpi->feature_score_loc_arr);
   vpx_free(cpi->feature_score_loc_sort);
@@ -7254,6 +7259,16 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
     for (i = 0; i < REFS_PER_FRAME; ++i) cpi->scaled_ref_idx[i] = INVALID_IDX;
   }
 
+  if (cpi->kmeans_data_arr_alloc == 0) {
+    const int mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
+    const int mi_rows = mi_cols_aligned_to_sb(cm->mi_rows);
+    CHECK_MEM_ERROR(
+        cm, cpi->kmeans_data_arr,
+        vpx_calloc(mi_rows * mi_cols, sizeof(*cpi->kmeans_data_arr)));
+    cpi->kmeans_data_stride = mi_cols;
+    cpi->kmeans_data_arr_alloc = 1;
+  }
+
   if (gf_group_index == 1 &&
       cpi->twopass.gf_group.update_type[gf_group_index] == ARF_UPDATE &&
       cpi->sf.enable_tpl_model) {
diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h
index e578646b9..278408fd5 100644
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -567,6 +567,14 @@ typedef struct FEATURE_SCORE_LOC {
 } FEATURE_SCORE_LOC;
 #endif
 
+#define MAX_KMEANS_GROUPS 8
+
+typedef struct KMEANS_DATA {
+  double value;
+  int pos;
+  int group_idx;
+} KMEANS_DATA;
+
 typedef struct VP9_COMP {
   QUANTS quants;
   ThreadData td;
@@ -594,6 +602,12 @@ typedef struct VP9_COMP {
   TplDepFrame tpl_stats[MAX_ARF_GOP_SIZE];
   YV12_BUFFER_CONFIG *tpl_recon_frames[REF_FRAMES];
   EncFrameBuf enc_frame_buf[REF_FRAMES];
+  int kmeans_data_arr_alloc;
+  KMEANS_DATA *kmeans_data_arr;
+  int kmeans_data_size;
+  int kmeans_data_stride;
+  double kmeans_ctr_ls[MAX_KMEANS_GROUPS];
+  int kmeans_ctr_num;
 #if CONFIG_NON_GREEDY_MV
   int tpl_ready;
   int feature_score_loc_alloc;