Merge "Improve vp9_rb_bytes_read"
diff --git a/build/make/gen_msvs_proj.sh b/build/make/gen_msvs_proj.sh
index f1cc04e..3653309 100755
--- a/build/make/gen_msvs_proj.sh
+++ b/build/make/gen_msvs_proj.sh
@@ -137,7 +137,9 @@
;;
--lib) proj_kind="lib"
;;
- --src-path-bare=*) src_path_bare=$(fix_path "$optval")
+ --src-path-bare=*)
+ src_path_bare=$(fix_path "$optval")
+ src_path_bare=${src_path_bare%/}
;;
--static-crt) use_static_runtime=true
;;
@@ -151,9 +153,9 @@
esac
;;
-I*)
- opt="${opt%/}"
opt=${opt##-I}
opt=$(fix_path "$opt")
+ opt="${opt%/}"
incs="${incs}${incs:+;}"${opt}""
yasmincs="${yasmincs} -I"${opt}""
;;
@@ -414,7 +416,7 @@
vpx)
tag Tool \
Name="VCPreBuildEventTool" \
- CommandLine="call obj_int_extract.bat $src_path_bare $plat_no_ws\\\$(ConfigurationName)" \
+ CommandLine="call obj_int_extract.bat "$src_path_bare" $plat_no_ws\\\$(ConfigurationName)" \
tag Tool \
Name="VCCLCompilerTool" \
diff --git a/build/make/gen_msvs_vcxproj.sh b/build/make/gen_msvs_vcxproj.sh
index eee354d..23ef6a3 100755
--- a/build/make/gen_msvs_vcxproj.sh
+++ b/build/make/gen_msvs_vcxproj.sh
@@ -157,7 +157,9 @@
;;
--lib) proj_kind="lib"
;;
- --src-path-bare=*) src_path_bare=$(fix_path "$optval")
+ --src-path-bare=*)
+ src_path_bare=$(fix_path "$optval")
+ src_path_bare=${src_path_bare%/}
;;
--static-crt) use_static_runtime=true
;;
@@ -173,9 +175,9 @@
esac
;;
-I*)
- opt="${opt%/}"
opt=${opt##-I}
opt=$(fix_path "$opt")
+ opt="${opt%/}"
incs="${incs}${incs:+;}"${opt}""
yasmincs="${yasmincs} -I"${opt}""
;;
diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h
index a27868a..a80ab1b 100644
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -32,6 +32,7 @@
#include "vp9/encoder/vp9_mcomp.h"
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_ratectrl.h"
+#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_speed_features.h"
#include "vp9/encoder/vp9_svc_layercontext.h"
#include "vp9/encoder/vp9_tokenize.h"
@@ -46,9 +47,6 @@
#define DEFAULT_GF_INTERVAL 10
-#define MAX_MODES 30
-#define MAX_REFS 6
-
typedef struct {
int nmvjointcost[MV_JOINTS];
int nmvcosts[2][MV_VALS];
@@ -66,56 +64,6 @@
FRAME_CONTEXT fc;
} CODING_CONTEXT;
-// This enumerator type needs to be kept aligned with the mode order in
-// const MODE_DEFINITION vp9_mode_order[MAX_MODES] used in the rd code.
-typedef enum {
- THR_NEARESTMV,
- THR_NEARESTA,
- THR_NEARESTG,
-
- THR_DC,
-
- THR_NEWMV,
- THR_NEWA,
- THR_NEWG,
-
- THR_NEARMV,
- THR_NEARA,
- THR_COMP_NEARESTLA,
- THR_COMP_NEARESTGA,
-
- THR_TM,
-
- THR_COMP_NEARLA,
- THR_COMP_NEWLA,
- THR_NEARG,
- THR_COMP_NEARGA,
- THR_COMP_NEWGA,
-
- THR_ZEROMV,
- THR_ZEROG,
- THR_ZEROA,
- THR_COMP_ZEROLA,
- THR_COMP_ZEROGA,
-
- THR_H_PRED,
- THR_V_PRED,
- THR_D135_PRED,
- THR_D207_PRED,
- THR_D153_PRED,
- THR_D63_PRED,
- THR_D117_PRED,
- THR_D45_PRED,
-} THR_MODES;
-
-typedef enum {
- THR_LAST,
- THR_GOLD,
- THR_ALTR,
- THR_COMP_LA,
- THR_COMP_GA,
- THR_INTRA,
-} THR_MODES_SUB8X8;
typedef enum {
// encode_breakout is disabled.
@@ -293,32 +241,6 @@
return mode == ONE_PASS_BEST || mode == TWO_PASS_SECOND_BEST;
}
-typedef struct RD_OPT {
- // Thresh_mult is used to set a threshold for the rd score. A higher value
- // means that we will accept the best mode so far more often. This number
- // is used in combination with the current block size, and thresh_freq_fact
- // to pick a threshold.
- int thresh_mult[MAX_MODES];
- int thresh_mult_sub8x8[MAX_REFS];
-
- int threshes[MAX_SEGMENTS][BLOCK_SIZES][MAX_MODES];
- int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
-
- int64_t comp_pred_diff[REFERENCE_MODES];
- int64_t prediction_type_threshes[MAX_REF_FRAMES][REFERENCE_MODES];
- int64_t tx_select_diff[TX_MODES];
- // FIXME(rbultje) can this overflow?
- int tx_select_threshes[MAX_REF_FRAMES][TX_MODES];
-
- int64_t filter_diff[SWITCHABLE_FILTER_CONTEXTS];
- int64_t filter_threshes[MAX_REF_FRAMES][SWITCHABLE_FILTER_CONTEXTS];
- int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
- int64_t mask_filter;
-
- int RDMULT;
- int RDDIV;
-} RD_OPT;
-
typedef struct VP9_COMP {
QUANTS quants;
MACROBLOCK mb;
diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c
index 7cadb36..5e82bb3 100644
--- a/vp9/encoder/vp9_firstpass.c
+++ b/vp9/encoder/vp9_firstpass.c
@@ -594,8 +594,9 @@
// Other than for the first frame do a motion search.
if (cm->current_video_frame > 0) {
- int tmp_err, motion_error;
+ int tmp_err, motion_error, raw_motion_error;
int_mv mv, tmp_mv;
+ struct buf_2d unscaled_last_source_buf_2d;
xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
motion_error = get_prediction_error(bsize, &x->plane[0].src,
@@ -603,67 +604,82 @@
// Assume 0,0 motion with no mv overhead.
mv.as_int = tmp_mv.as_int = 0;
- // Test last reference frame using the previous best mv as the
- // starting point (best reference) for the search.
- first_pass_motion_search(cpi, x, &best_ref_mv.as_mv, &mv.as_mv,
- &motion_error);
- if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
- vp9_clear_system_state();
- motion_error = (int)(motion_error * error_weight);
- }
+ // Compute the motion error of the 0,0 motion using the last source
+ // frame as the reference. Skip the further motion search on
+ // reconstructed frame if this error is small.
+ unscaled_last_source_buf_2d.buf =
+ cpi->unscaled_last_source->y_buffer + recon_yoffset;
+ unscaled_last_source_buf_2d.stride =
+ cpi->unscaled_last_source->y_stride;
+ raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
+ &unscaled_last_source_buf_2d);
- // If the current best reference mv is not centered on 0,0 then do a 0,0
- // based search as well.
- if (best_ref_mv.as_int) {
- tmp_err = INT_MAX;
- first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
- &tmp_err);
+ // TODO(pengchong): Replace the hard-coded threshold
+ if (raw_motion_error > 25) {
+ // Test last reference frame using the previous best mv as the
+ // starting point (best reference) for the search.
+ first_pass_motion_search(cpi, x, &best_ref_mv.as_mv, &mv.as_mv,
+ &motion_error);
if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
vp9_clear_system_state();
- tmp_err = (int)(tmp_err * error_weight);
+ motion_error = (int)(motion_error * error_weight);
}
- if (tmp_err < motion_error) {
- motion_error = tmp_err;
- mv.as_int = tmp_mv.as_int;
- }
- }
+ // If the current best reference mv is not centered on 0,0 then do a
+ // 0,0 based search as well.
+ if (best_ref_mv.as_int) {
+ tmp_err = INT_MAX;
+ first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv, &tmp_err);
+ if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
+ vp9_clear_system_state();
+ tmp_err = (int)(tmp_err * error_weight);
+ }
- // Search in an older reference frame.
- if (cm->current_video_frame > 1 && gld_yv12 != NULL) {
- // Assume 0,0 motion with no mv overhead.
- int gf_motion_error;
-
- xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
- gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
- &xd->plane[0].pre[0]);
-
- first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
- &gf_motion_error);
- if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
- vp9_clear_system_state();
- gf_motion_error = (int)(gf_motion_error * error_weight);
+ if (tmp_err < motion_error) {
+ motion_error = tmp_err;
+ mv.as_int = tmp_mv.as_int;
+ }
}
- if (gf_motion_error < motion_error && gf_motion_error < this_error)
- ++second_ref_count;
+ // Search in an older reference frame.
+ if (cm->current_video_frame > 1 && gld_yv12 != NULL) {
+ // Assume 0,0 motion with no mv overhead.
+ int gf_motion_error;
- // Reset to last frame as reference buffer.
- xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
- xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
- xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
+ xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
+ gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
+ &xd->plane[0].pre[0]);
- // In accumulating a score for the older reference frame take the
- // best of the motion predicted score and the intra coded error
- // (just as will be done for) accumulation of "coded_error" for
- // the last frame.
- if (gf_motion_error < this_error)
- sr_coded_error += gf_motion_error;
- else
- sr_coded_error += this_error;
+ first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
+ &gf_motion_error);
+ if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
+ vp9_clear_system_state();
+ gf_motion_error = (int)(gf_motion_error * error_weight);
+ }
+
+ if (gf_motion_error < motion_error && gf_motion_error < this_error)
+ ++second_ref_count;
+
+ // Reset to last frame as reference buffer.
+ xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
+ xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
+ xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
+
+ // In accumulating a score for the older reference frame take the
+ // best of the motion predicted score and the intra coded error
+ // (just as will be done for) accumulation of "coded_error" for
+ // the last frame.
+ if (gf_motion_error < this_error)
+ sr_coded_error += gf_motion_error;
+ else
+ sr_coded_error += this_error;
+ } else {
+ sr_coded_error += motion_error;
+ }
} else {
sr_coded_error += motion_error;
}
+
// Start by assuming that intra mode is best.
best_ref_mv.as_int = 0;
diff --git a/vp9/encoder/vp9_rdopt.h b/vp9/encoder/vp9_rdopt.h
index 6e56317..fba54cc 100644
--- a/vp9/encoder/vp9_rdopt.h
+++ b/vp9/encoder/vp9_rdopt.h
@@ -13,7 +13,10 @@
#include <limits.h>
-#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/common/vp9_blockd.h"
+
+#include "vp9/encoder/vp9_block.h"
+#include "vp9/encoder/vp9_context_tree.h"
#ifdef __cplusplus
extern "C" {
@@ -30,21 +33,104 @@
#define INVALID_MV 0x80008000
+#define MAX_MODES 30
+#define MAX_REFS 6
+
+// This enumerator type needs to be kept aligned with the mode order in
+// const MODE_DEFINITION vp9_mode_order[MAX_MODES] used in the rd code.
+typedef enum {
+ THR_NEARESTMV,
+ THR_NEARESTA,
+ THR_NEARESTG,
+
+ THR_DC,
+
+ THR_NEWMV,
+ THR_NEWA,
+ THR_NEWG,
+
+ THR_NEARMV,
+ THR_NEARA,
+ THR_COMP_NEARESTLA,
+ THR_COMP_NEARESTGA,
+
+ THR_TM,
+
+ THR_COMP_NEARLA,
+ THR_COMP_NEWLA,
+ THR_NEARG,
+ THR_COMP_NEARGA,
+ THR_COMP_NEWGA,
+
+ THR_ZEROMV,
+ THR_ZEROG,
+ THR_ZEROA,
+ THR_COMP_ZEROLA,
+ THR_COMP_ZEROGA,
+
+ THR_H_PRED,
+ THR_V_PRED,
+ THR_D135_PRED,
+ THR_D207_PRED,
+ THR_D153_PRED,
+ THR_D63_PRED,
+ THR_D117_PRED,
+ THR_D45_PRED,
+} THR_MODES;
+
+typedef enum {
+ THR_LAST,
+ THR_GOLD,
+ THR_ALTR,
+ THR_COMP_LA,
+ THR_COMP_GA,
+ THR_INTRA,
+} THR_MODES_SUB8X8;
+
+typedef struct RD_OPT {
+ // Thresh_mult is used to set a threshold for the rd score. A higher value
+ // means that we will accept the best mode so far more often. This number
+ // is used in combination with the current block size, and thresh_freq_fact
+ // to pick a threshold.
+ int thresh_mult[MAX_MODES];
+ int thresh_mult_sub8x8[MAX_REFS];
+
+ int threshes[MAX_SEGMENTS][BLOCK_SIZES][MAX_MODES];
+ int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
+
+ int64_t comp_pred_diff[REFERENCE_MODES];
+ int64_t prediction_type_threshes[MAX_REF_FRAMES][REFERENCE_MODES];
+ int64_t tx_select_diff[TX_MODES];
+ // FIXME(rbultje) can this overflow?
+ int tx_select_threshes[MAX_REF_FRAMES][TX_MODES];
+
+ int64_t filter_diff[SWITCHABLE_FILTER_CONTEXTS];
+ int64_t filter_threshes[MAX_REF_FRAMES][SWITCHABLE_FILTER_CONTEXTS];
+ int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
+ int64_t mask_filter;
+
+ int RDMULT;
+ int RDDIV;
+} RD_OPT;
+
+
struct TileInfo;
+struct VP9_COMP;
+struct macroblock;
-int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex);
+int vp9_compute_rd_mult(const struct VP9_COMP *cpi, int qindex);
-void vp9_initialize_rd_consts(VP9_COMP *cpi);
+void vp9_initialize_rd_consts(struct VP9_COMP *cpi);
-void vp9_initialize_me_consts(VP9_COMP *cpi, int qindex);
+void vp9_initialize_me_consts(struct VP9_COMP *cpi, int qindex);
void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n,
unsigned int qstep, int *rate,
int64_t *dist);
-int vp9_get_switchable_rate(const VP9_COMP *cpi);
+int vp9_get_switchable_rate(const struct VP9_COMP *cpi);
-void vp9_setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
+void vp9_setup_buffer_inter(struct VP9_COMP *cpi, struct macroblock *x,
const TileInfo *const tile,
MV_REFERENCE_FRAME ref_frame,
BLOCK_SIZE block_size,
@@ -53,14 +139,14 @@
int_mv frame_near_mv[MAX_REF_FRAMES],
struct buf_2d yv12_mb[4][MAX_MB_PLANE]);
-const YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const VP9_COMP *cpi,
+const YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const struct VP9_COMP *cpi,
int ref_frame);
-void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
+void vp9_rd_pick_intra_mode_sb(struct VP9_COMP *cpi, struct macroblock *x,
int *r, int64_t *d, BLOCK_SIZE bsize,
PICK_MODE_CONTEXT *ctx, int64_t best_rd);
-int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
+int64_t vp9_rd_pick_inter_mode_sb(struct VP9_COMP *cpi, struct macroblock *x,
const struct TileInfo *const tile,
int mi_row, int mi_col,
int *returnrate,
@@ -69,7 +155,8 @@
PICK_MODE_CONTEXT *ctx,
int64_t best_rd_so_far);
-int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
+int64_t vp9_rd_pick_inter_mode_sub8x8(struct VP9_COMP *cpi,
+ struct macroblock *x,
const struct TileInfo *const tile,
int mi_row, int mi_col,
int *returnrate,
@@ -85,9 +172,9 @@
ENTROPY_CONTEXT t_above[16],
ENTROPY_CONTEXT t_left[16]);
-void vp9_set_rd_speed_thresholds(VP9_COMP *cpi);
+void vp9_set_rd_speed_thresholds(struct VP9_COMP *cpi);
-void vp9_set_rd_speed_thresholds_sub8x8(VP9_COMP *cpi);
+void vp9_set_rd_speed_thresholds_sub8x8(struct VP9_COMP *cpi);
static INLINE int rd_less_than_thresh(int64_t best_rd, int thresh,
int thresh_fact) {
