Merge "ppc: Add vpx_sadnxmx4d_vsx for n,m = {8, 16, 32 ,64}"

[libvpx] / vp8 / encoder / onyx_if.c

diff --git a/vp8/encoder/onyx_if.c b/vp8/encoder/onyx_if.c
index 87560f28b1e2d5ed22433158468a7f040f8c5151..9227f750d55979b572d6d6be054069c9e9ae972f 100644 (file)
--- a/vp8/encoder/onyx_if.c
+++ b/vp8/encoder/onyx_if.c
@@ -16,6 +16,7 @@
 #include "vp8/common/blockd.h"
 #include "onyx_int.h"
 #include "vp8/common/systemdependent.h"
+#include "vp8/common/skin_detection.h"
 #include "vp8/encoder/quantize.h"
 #include "vp8/common/alloccommon.h"
 #include "mcomp.h"
@@ -43,6 +44,7 @@
 #endif
 #include "encodeframe.h"
 
+#include <assert.h>
 #include <math.h>
 #include <stdio.h>
 #include <limits.h>
@@ -86,6 +88,9 @@ FILE *yuv_file;
 #ifdef OUTPUT_YUV_DENOISED
 FILE *yuv_denoised_file;
 #endif
+#ifdef OUTPUT_YUV_SKINMAP
+FILE *yuv_skinmap_file = NULL;
+#endif
 
 #if 0
 FILE *framepsnr;
@@ -446,6 +451,18 @@ static void dealloc_compressor_data(VP8_COMP *cpi) {
   cpi->mb.pip = 0;
 
 #if CONFIG_MULTITHREAD
+  /* De-allocate mutex */
+  if (cpi->pmutex != NULL) {
+    VP8_COMMON *const pc = &cpi->common;
+    int i;
+
+    for (i = 0; i < pc->mb_rows; ++i) {
+      pthread_mutex_destroy(&cpi->pmutex[i]);
+    }
+    vpx_free(cpi->pmutex);
+    cpi->pmutex = NULL;
+  }
+
   vpx_free(cpi->mt_current_mb_col);
   cpi->mt_current_mb_col = NULL;
 #endif
@@ -715,6 +732,7 @@ void vp8_set_speed_features(VP8_COMP *cpi) {
   SPEED_FEATURES *sf = &cpi->sf;
   int Mode = cpi->compressor_speed;
   int Speed = cpi->Speed;
+  int Speed2;
   int i;
   VP8_COMMON *cm = &cpi->common;
   int last_improved_quant = sf->improved_quant;
@@ -816,9 +834,16 @@ void vp8_set_speed_features(VP8_COMP *cpi) {
   cpi->mode_check_freq[THR_V_PRED] = cpi->mode_check_freq[THR_H_PRED] =
       cpi->mode_check_freq[THR_B_PRED] =
           speed_map(Speed, mode_check_freq_map_vhbpred);
-  cpi->mode_check_freq[THR_NEW1] = speed_map(Speed, mode_check_freq_map_new1);
+
+  // For real-time mode at speed 10 keep the mode_check_freq threshold
+  // for NEW1 similar to that of speed 9.
+  Speed2 = Speed;
+  if (cpi->Speed == 10 && Mode == 2) Speed2 = RT(9);
+  cpi->mode_check_freq[THR_NEW1] = speed_map(Speed2, mode_check_freq_map_new1);
+
   cpi->mode_check_freq[THR_NEW2] = cpi->mode_check_freq[THR_NEW3] =
       speed_map(Speed, mode_check_freq_map_new2);
+
   cpi->mode_check_freq[THR_SPLIT1] =
       speed_map(Speed, mode_check_freq_map_split1);
   cpi->mode_check_freq[THR_SPLIT2] = cpi->mode_check_freq[THR_SPLIT3] =
@@ -1075,6 +1100,9 @@ void vp8_alloc_compressor_data(VP8_COMP *cpi) {
 
   int width = cm->Width;
   int height = cm->Height;
+#if CONFIG_MULTITHREAD
+  int prev_mb_rows = cm->mb_rows;
+#endif
 
   if (vp8_alloc_frame_buffers(cm, width, height)) {
     vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
@@ -1164,6 +1192,25 @@ void vp8_alloc_compressor_data(VP8_COMP *cpi) {
   }
 
   if (cpi->oxcf.multi_threaded > 1) {
+    int i;
+
+    /* De-allocate and re-allocate mutex */
+    if (cpi->pmutex != NULL) {
+      for (i = 0; i < prev_mb_rows; ++i) {
+        pthread_mutex_destroy(&cpi->pmutex[i]);
+      }
+      vpx_free(cpi->pmutex);
+      cpi->pmutex = NULL;
+    }
+
+    CHECK_MEM_ERROR(cpi->pmutex,
+                    vpx_malloc(sizeof(*cpi->pmutex) * cm->mb_rows));
+    if (cpi->pmutex) {
+      for (i = 0; i < cm->mb_rows; ++i) {
+        pthread_mutex_init(&cpi->pmutex[i], NULL);
+      }
+    }
+
     vpx_free(cpi->mt_current_mb_col);
     CHECK_MEM_ERROR(cpi->mt_current_mb_col,
                     vpx_malloc(sizeof(*cpi->mt_current_mb_col) * cm->mb_rows));
@@ -1467,6 +1514,12 @@ void vp8_change_config(VP8_COMP *cpi, VP8_CONFIG *oxcf) {
   cpi->baseline_gf_interval =
       cpi->oxcf.alt_freq ? cpi->oxcf.alt_freq : DEFAULT_GF_INTERVAL;
 
+  // GF behavior for 1 pass CBR, used when error_resilience is off.
+  if (!cpi->oxcf.error_resilient_mode &&
+      cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER &&
+      cpi->oxcf.Mode == MODE_REALTIME)
+    cpi->baseline_gf_interval = cpi->gf_interval_onepass_cbr;
+
 #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
   cpi->oxcf.token_partitions = 3;
 #endif
@@ -1593,8 +1646,7 @@ void vp8_change_config(VP8_COMP *cpi, VP8_CONFIG *oxcf) {
   cm->sharpness_level = cpi->oxcf.Sharpness;
 
   if (cm->horiz_scale != NORMAL || cm->vert_scale != NORMAL) {
-    int UNINITIALIZED_IS_SAFE(hr), UNINITIALIZED_IS_SAFE(hs);
-    int UNINITIALIZED_IS_SAFE(vr), UNINITIALIZED_IS_SAFE(vs);
+    int hr, hs, vr, vs;
 
     Scale2Ratio(cm->horiz_scale, &hr, &hs);
     Scale2Ratio(cm->vert_scale, &vr, &vs);
@@ -1766,9 +1818,13 @@ struct VP8_COMP *vp8_create_compressor(VP8_CONFIG *oxcf) {
   cpi->mse_source_denoised = 0;
 
   /* Should we use the cyclic refresh method.
-   * Currently this is tied to error resilliant mode
+   * Currently there is no external control for this.
+   * Enable it for error_resilient_mode, or for 1 pass CBR mode.
    */
-  cpi->cyclic_refresh_mode_enabled = cpi->oxcf.error_resilient_mode;
+  cpi->cyclic_refresh_mode_enabled =
+      (cpi->oxcf.error_resilient_mode ||
+       (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER &&
+        cpi->oxcf.Mode <= 2));
   cpi->cyclic_refresh_mode_max_mbs_perframe =
       (cpi->common.mb_rows * cpi->common.mb_cols) / 7;
   if (cpi->oxcf.number_of_layers == 1) {
@@ -1781,6 +1837,23 @@ struct VP8_COMP *vp8_create_compressor(VP8_CONFIG *oxcf) {
   cpi->cyclic_refresh_mode_index = 0;
   cpi->cyclic_refresh_q = 32;
 
+  // GF behavior for 1 pass CBR, used when error_resilience is off.
+  cpi->gf_update_onepass_cbr = 0;
+  cpi->gf_noboost_onepass_cbr = 0;
+  if (!cpi->oxcf.error_resilient_mode &&
+      cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER && cpi->oxcf.Mode <= 2) {
+    cpi->gf_update_onepass_cbr = 1;
+    cpi->gf_noboost_onepass_cbr = 1;
+    cpi->gf_interval_onepass_cbr =
+        cpi->cyclic_refresh_mode_max_mbs_perframe > 0
+            ? (2 * (cpi->common.mb_rows * cpi->common.mb_cols) /
+               cpi->cyclic_refresh_mode_max_mbs_perframe)
+            : 10;
+    cpi->gf_interval_onepass_cbr =
+        VPXMIN(40, VPXMAX(6, cpi->gf_interval_onepass_cbr));
+    cpi->baseline_gf_interval = cpi->gf_interval_onepass_cbr;
+  }
+
   if (cpi->cyclic_refresh_mode_enabled) {
     CHECK_MEM_ERROR(cpi->cyclic_refresh_map,
                     vpx_calloc((cpi->common.mb_rows * cpi->common.mb_cols), 1));
@@ -1864,6 +1937,9 @@ struct VP8_COMP *vp8_create_compressor(VP8_CONFIG *oxcf) {
 #ifdef OUTPUT_YUV_DENOISED
   yuv_denoised_file = fopen("denoised.yuv", "ab");
 #endif
+#ifdef OUTPUT_YUV_SKINMAP
+  yuv_skinmap_file = fopen("skinmap.yuv", "ab");
+#endif
 
 #if 0
     framepsnr = fopen("framepsnr.stt", "a");
@@ -2029,11 +2105,7 @@ void vp8_remove_compressor(VP8_COMP **ptr) {
       double time_encoded =
           (cpi->last_end_time_stamp_seen - cpi->first_time_stamp_ever) /
           10000000.000;
-      double total_encode_time =
-          (cpi->time_receive_data + cpi->time_compress_data) / 1000.000;
       double dr = (double)cpi->bytes * 8.0 / 1000.0 / time_encoded;
-      const double target_rate = (double)cpi->oxcf.target_bandwidth / 1000;
-      const double rate_err = ((100.0 * (dr - target_rate)) / target_rate);
 
       if (cpi->b_calculate_psnr) {
         if (cpi->oxcf.number_of_layers > 1) {
@@ -2233,6 +2305,9 @@ void vp8_remove_compressor(VP8_COMP **ptr) {
 #ifdef OUTPUT_YUV_DENOISED
   fclose(yuv_denoised_file);
 #endif
+#ifdef OUTPUT_YUV_SKINMAP
+  fclose(yuv_skinmap_file);
+#endif
 
 #if 0
 
@@ -2409,10 +2484,11 @@ int vp8_update_entropy(VP8_COMP *cpi, int update) {
   return 0;
 }
 
-#if defined(OUTPUT_YUV_SRC) || defined(OUTPUT_YUV_DENOISED)
+#if defined(OUTPUT_YUV_SRC) || defined(OUTPUT_YUV_DENOISED) || \
+    defined(OUTPUT_YUV_SKINMAP)
 void vp8_write_yuv_frame(FILE *yuv_file, YV12_BUFFER_CONFIG *s) {
   unsigned char *src = s->y_buffer;
-  int h = s->y_height;
+  int h = s->y_crop_height;
 
   do {
     fwrite(src, s->y_width, 1, yuv_file);
@@ -2420,7 +2496,7 @@ void vp8_write_yuv_frame(FILE *yuv_file, YV12_BUFFER_CONFIG *s) {
   } while (--h);
 
   src = s->u_buffer;
-  h = s->uv_height;
+  h = s->uv_crop_height;
 
   do {
     fwrite(src, s->uv_width, 1, yuv_file);
@@ -2428,7 +2504,7 @@ void vp8_write_yuv_frame(FILE *yuv_file, YV12_BUFFER_CONFIG *s) {
   } while (--h);
 
   src = s->v_buffer;
-  h = s->uv_height;
+  h = s->uv_crop_height;
 
   do {
     fwrite(src, s->uv_width, 1, yuv_file);
@@ -2443,8 +2519,7 @@ static void scale_and_extend_source(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi) {
   /* are we resizing the image */
   if (cm->horiz_scale != 0 || cm->vert_scale != 0) {
 #if CONFIG_SPATIAL_RESAMPLING
-    int UNINITIALIZED_IS_SAFE(hr), UNINITIALIZED_IS_SAFE(hs);
-    int UNINITIALIZED_IS_SAFE(vr), UNINITIALIZED_IS_SAFE(vs);
+    int hr, hs, vr, vs;
     int tmp_height;
 
     if (cm->vert_scale == 3) {
@@ -2477,8 +2552,7 @@ static int resize_key_frame(VP8_COMP *cpi) {
    */
   if (cpi->oxcf.allow_spatial_resampling &&
       (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)) {
-    int UNINITIALIZED_IS_SAFE(hr), UNINITIALIZED_IS_SAFE(hs);
-    int UNINITIALIZED_IS_SAFE(vr), UNINITIALIZED_IS_SAFE(vs);
+    int hr, hs, vr, vs;
     int new_width, new_height;
 
     /* If we are below the resample DOWN watermark then scale down a
@@ -2994,6 +3068,7 @@ static int measure_square_diff_partial(YV12_BUFFER_CONFIG *source,
   }
   // Only return non-zero if we have at least ~1/16 samples for estimate.
   if (num_blocks > (tot_num_blocks >> 4)) {
+    assert(num_blocks != 0);
     return (Total / num_blocks);
   } else {
     return 0;
@@ -3925,7 +4000,6 @@ static void encode_frame_to_data_rate(VP8_COMP *cpi, size_t *size,
 #else
     /* transform / motion compensation build reconstruction frame */
     vp8_encode_frame(cpi);
-
     if (cpi->oxcf.screen_content_mode == 2) {
       if (vp8_drop_encodedframe_overshoot(cpi, Q)) return;
     }
@@ -4002,9 +4076,9 @@ static void encode_frame_to_data_rate(VP8_COMP *cpi, size_t *size,
 #if !CONFIG_REALTIME_ONLY
       top_index = cpi->active_worst_quality;
 #endif  // !CONFIG_REALTIME_ONLY
-        /* If we have updated the active max Q do not call
-         * vp8_update_rate_correction_factors() this loop.
-         */
+      /* If we have updated the active max Q do not call
+       * vp8_update_rate_correction_factors() this loop.
+       */
       active_worst_qchanged = 1;
     } else {
       active_worst_qchanged = 0;
@@ -4203,6 +4277,20 @@ static void encode_frame_to_data_rate(VP8_COMP *cpi, size_t *size,
     }
   } while (Loop == 1);
 
+#if defined(DROP_UNCODED_FRAMES)
+  /* if there are no coded macroblocks at all drop this frame */
+  if (cpi->common.MBs == cpi->mb.skip_true_count &&
+      (cpi->drop_frame_count & 7) != 7 && cm->frame_type != KEY_FRAME) {
+    cpi->common.current_video_frame++;
+    cpi->frames_since_key++;
+    cpi->drop_frame_count++;
+    // We advance the temporal pattern for dropped frames.
+    cpi->temporal_pattern_counter++;
+    return;
+  }
+  cpi->drop_frame_count = 0;
+#endif
+
 #if 0
     /* Experimental code for lagged and one pass
      * Update stats used for one pass GF selection
@@ -4344,6 +4432,12 @@ static void encode_frame_to_data_rate(VP8_COMP *cpi, size_t *size,
   }
 #endif
 
+#ifdef OUTPUT_YUV_SKINMAP
+  if (cpi->common.current_video_frame > 1) {
+    compute_skin_map(cpi, yuv_skinmap_file);
+  }
+#endif
+
 #if CONFIG_MULTITHREAD
   if (cpi->b_multi_threaded) {
     /* start loopfilter in separate thread */