From 0a6d5547e28ec0e9f92dfdb231704f169b942b2f Mon Sep 17 00:00:00 2001 From: Dmitry Kovalev <dkovalev@google.com> Date: Fri, 28 Mar 2014 10:30:28 -0700 Subject: [PATCH] Adding vp9_speed_features.{h, c}. Change-Id: I7d9874da8ff78a2d7e0cf11073af9c30538bc9a6 --- vp9/encoder/vp9_onyx_if.c | 376 +----------------------------- vp9/encoder/vp9_onyx_int.h | 293 +----------------------- vp9/encoder/vp9_speed_features.c | 381 +++++++++++++++++++++++++++++++ vp9/encoder/vp9_speed_features.h | 319 ++++++++++++++++++++++++++ vp9/vp9cx.mk | 2 + 5 files changed, 708 insertions(+), 663 deletions(-) create mode 100644 vp9/encoder/vp9_speed_features.c create mode 100644 vp9/encoder/vp9_speed_features.h diff --git a/vp9/encoder/vp9_onyx_if.c b/vp9/encoder/vp9_onyx_if.c index 122477906a..0f1ca60222 100644 --- a/vp9/encoder/vp9_onyx_if.c +++ b/vp9/encoder/vp9_onyx_if.c @@ -44,12 +44,6 @@ #include "vp9/encoder/vp9_resize.h" #include "vp9/encoder/vp9_svc_layercontext.h" -#define ALL_INTRA_MODES 0x3FF -#define INTRA_DC_ONLY 0x01 -#define INTRA_DC_TM ((1 << TM_PRED) | (1 << DC_PRED)) -#define INTRA_DC_H_V ((1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED)) -#define INTRA_DC_TM_H_V (INTRA_DC_TM | (1 << V_PRED) | (1 << H_PRED)) - void vp9_coef_tree_initialize(); #define DEFAULT_INTERP_FILTER SWITCHABLE @@ -63,12 +57,6 @@ void vp9_coef_tree_initialize(); // now so that HIGH_PRECISION is always // chosen. -// Masks for partially or completely disabling split mode -#define DISABLE_ALL_SPLIT 0x3F -#define DISABLE_ALL_INTER_SPLIT 0x1F -#define DISABLE_COMPOUND_SPLIT 0x18 -#define LAST_AND_INTRA_SPLIT_ONLY 0x1E - // Max rate target for 1080P and below encodes under normal circumstances // (1920 * 1080 / (16 * 16)) * MAX_MB_RATE bits per MB #define MAX_MB_RATE 250 @@ -547,377 +535,23 @@ static void set_rd_speed_thresholds_sub8x8(VP9_COMP *cpi) { cpi->rd_thresh_mult_sub8x8[THR_COMP_GA] = INT_MAX; } -// Intra only frames, golden frames (except alt ref overlays) and -// alt ref frames tend to be coded at a higher than ambient quality -static INLINE int frame_is_boosted(const VP9_COMP *cpi) { - return frame_is_intra_only(&cpi->common) || cpi->refresh_alt_ref_frame || - (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref); -} - -static void set_good_speed_feature(VP9_COMP *cpi, - VP9_COMMON *cm, - SPEED_FEATURES *sf, - int speed) { - int i; - sf->adaptive_rd_thresh = 1; - sf->recode_loop = ((speed < 1) ? ALLOW_RECODE : ALLOW_RECODE_KFMAXBW); - sf->allow_skip_recode = 1; - - if (speed >= 1) { - sf->use_square_partition_only = !frame_is_intra_only(cm); - sf->less_rectangular_check = 1; - sf->tx_size_search_method = frame_is_boosted(cpi) - ? USE_FULL_RD : USE_LARGESTALL; - - if (MIN(cm->width, cm->height) >= 720) - sf->disable_split_mask = cm->show_frame ? - DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; - else - sf->disable_split_mask = DISABLE_COMPOUND_SPLIT; - sf->use_rd_breakout = 1; - sf->adaptive_motion_search = 1; - sf->auto_mv_step_size = 1; - sf->adaptive_rd_thresh = 2; - sf->subpel_iters_per_step = 1; - sf->mode_skip_start = 10; - sf->adaptive_pred_interp_filter = 1; - - sf->recode_loop = ALLOW_RECODE_KFARFGF; - sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V; - sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V; - sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V; - sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V; - } - // Additions or changes from speed 1 for speed >= 2. - if (speed >= 2) { - sf->tx_size_search_method = frame_is_boosted(cpi) - ? USE_FULL_RD : USE_LARGESTALL; - - if (MIN(cm->width, cm->height) >= 720) - sf->disable_split_mask = cm->show_frame ? - DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; - else - sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY; - - sf->adaptive_pred_interp_filter = 2; - - sf->reference_masking = 1; - sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH | - FLAG_SKIP_INTRA_BESTINTER | - FLAG_SKIP_COMP_BESTINTRA | - FLAG_SKIP_INTRA_LOWVAR; - sf->disable_filter_search_var_thresh = 100; - sf->comp_inter_joint_search_thresh = BLOCK_SIZES; - - sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX; - sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_LOW_MOTION; - sf->adjust_partitioning_from_last_frame = 1; - sf->last_partitioning_redo_frequency = 3; - } - // Additions or changes for speed 3 and above - if (speed >= 3) { - if (MIN(cm->width, cm->height) >= 720) - sf->disable_split_mask = DISABLE_ALL_SPLIT; - else - sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT; - - sf->recode_loop = ALLOW_RECODE_KFMAXBW; - - sf->adaptive_rd_thresh = 3; - sf->mode_skip_start = 6; - sf->use_fast_coef_updates = 2; - sf->use_fast_coef_costing = 1; - } - // Additions or changes for speed 3 and above - if (speed >= 4) { - sf->use_square_partition_only = 1; - sf->tx_size_search_method = USE_LARGESTALL; - sf->disable_split_mask = DISABLE_ALL_SPLIT; - - sf->adaptive_rd_thresh = 4; - - // Add a couple more skip flags - sf->mode_search_skip_flags |= FLAG_SKIP_COMP_REFMISMATCH | - FLAG_EARLY_TERMINATE; - - sf->disable_filter_search_var_thresh = 200; - - sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_ALL; - sf->use_lp32x32fdct = 1; - } - if (speed >= 5) { - sf->partition_search_type = FIXED_PARTITION; - sf->optimize_coefficients = 0; - sf->search_method = HEX; - sf->disable_filter_search_var_thresh = 500; - for (i = 0; i < TX_SIZES; i++) { - sf->intra_y_mode_mask[i] = INTRA_DC_ONLY; - sf->intra_uv_mode_mask[i] = INTRA_DC_ONLY; - } - cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED; - } -} - -static void set_rt_speed_feature(VP9_COMMON *cm, - SPEED_FEATURES *sf, - int speed) { - sf->static_segmentation = 0; - sf->adaptive_rd_thresh = 1; - sf->encode_breakout_thresh = 1; - sf->use_fast_coef_costing = 1; - - if (speed == 1) { - sf->use_square_partition_only = !frame_is_intra_only(cm); - sf->less_rectangular_check = 1; - sf->tx_size_search_method = - frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL; - - if (MIN(cm->width, cm->height) >= 720) - sf->disable_split_mask = cm->show_frame ? - DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; - else - sf->disable_split_mask = DISABLE_COMPOUND_SPLIT; - - sf->use_rd_breakout = 1; - sf->adaptive_motion_search = 1; - sf->adaptive_pred_interp_filter = 1; - sf->auto_mv_step_size = 1; - sf->adaptive_rd_thresh = 2; - sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V; - sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V; - sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V; - sf->encode_breakout_thresh = 8; - } - if (speed >= 2) { - sf->use_square_partition_only = !frame_is_intra_only(cm); - sf->less_rectangular_check = 1; - sf->tx_size_search_method = - frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL; - - if (MIN(cm->width, cm->height) >= 720) - sf->disable_split_mask = cm->show_frame ? - DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; - else - sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY; - - sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH - | FLAG_SKIP_INTRA_BESTINTER | FLAG_SKIP_COMP_BESTINTRA - | FLAG_SKIP_INTRA_LOWVAR; - - sf->use_rd_breakout = 1; - sf->adaptive_motion_search = 1; - sf->adaptive_pred_interp_filter = 2; - sf->auto_mv_step_size = 1; - sf->reference_masking = 1; - - sf->disable_filter_search_var_thresh = 50; - sf->comp_inter_joint_search_thresh = BLOCK_SIZES; - - sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX; - sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_LOW_MOTION; - sf->adjust_partitioning_from_last_frame = 1; - sf->last_partitioning_redo_frequency = 3; - - sf->adaptive_rd_thresh = 2; - sf->use_lp32x32fdct = 1; - sf->mode_skip_start = 11; - sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V; - sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V; - sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V; - sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V; - sf->encode_breakout_thresh = 200; - } - if (speed >= 3) { - sf->use_square_partition_only = 1; - if (MIN(cm->width, cm->height) >= 720) - sf->disable_split_mask = DISABLE_ALL_SPLIT; - else - sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT; - - sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH - | FLAG_SKIP_INTRA_BESTINTER | FLAG_SKIP_COMP_BESTINTRA - | FLAG_SKIP_INTRA_LOWVAR; - - sf->disable_filter_search_var_thresh = 100; - sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_ALL; - sf->use_uv_intra_rd_estimate = 1; - sf->skip_encode_sb = 1; - sf->subpel_iters_per_step = 1; - sf->use_fast_coef_updates = 2; - sf->adaptive_rd_thresh = 4; - sf->mode_skip_start = 6; - sf->encode_breakout_thresh = 400; - sf->allow_skip_recode = 0; - } - if (speed >= 4) { - sf->optimize_coefficients = 0; - sf->disable_split_mask = DISABLE_ALL_SPLIT; - sf->lpf_pick = LPF_PICK_FROM_Q; - sf->encode_breakout_thresh = 700; - } - if (speed >= 5) { - int i; - sf->last_partitioning_redo_frequency = 4; - sf->adaptive_rd_thresh = 5; - sf->use_fast_coef_costing = 0; - sf->auto_min_max_partition_size = STRICT_NEIGHBORING_MIN_MAX; - sf->adjust_partitioning_from_last_frame = - cm->last_frame_type != cm->frame_type || (0 == - (cm->current_video_frame + 1) % sf->last_partitioning_redo_frequency); - sf->subpel_force_stop = 1; - for (i = 0; i < TX_SIZES; i++) { - sf->intra_y_mode_mask[i] = INTRA_DC_H_V; - sf->intra_uv_mode_mask[i] = INTRA_DC_ONLY; - } - sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_ONLY; - sf->frame_parameter_update = 0; - sf->encode_breakout_thresh = 1000; - sf->search_method = FAST_HEX; - sf->disable_inter_mode_mask[BLOCK_32X32] = 1 << INTER_OFFSET(ZEROMV); - sf->disable_inter_mode_mask[BLOCK_32X64] = ~(1 << INTER_OFFSET(NEARESTMV)); - sf->disable_inter_mode_mask[BLOCK_64X32] = ~(1 << INTER_OFFSET(NEARESTMV)); - sf->disable_inter_mode_mask[BLOCK_64X64] = ~(1 << INTER_OFFSET(NEARESTMV)); - sf->max_intra_bsize = BLOCK_32X32; - sf->allow_skip_recode = 1; - } - if (speed >= 6) { - sf->max_partition_size = BLOCK_32X32; - sf->min_partition_size = BLOCK_8X8; - sf->partition_check = - (cm->current_video_frame % sf->last_partitioning_redo_frequency == 1); - sf->partition_search_type = REFERENCE_PARTITION; - sf->use_nonrd_pick_mode = 1; - sf->search_method = FAST_DIAMOND; - sf->allow_skip_recode = 0; - } - if (speed >= 7) { - sf->partition_search_type = VAR_BASED_FIXED_PARTITION; - sf->use_nonrd_pick_mode = 1; - sf->search_method = FAST_DIAMOND; - } - if (speed >= 8) { - int i; - for (i = 0; i < BLOCK_SIZES; ++i) - sf->disable_inter_mode_mask[i] = 14; // only search NEARESTMV (0) - } -} - -void vp9_set_speed_features(VP9_COMP *cpi) { - SPEED_FEATURES *sf = &cpi->sf; - VP9_COMMON *cm = &cpi->common; - int speed = cpi->speed; - int i; - - // Convert negative speed to positive - if (speed < 0) - speed = -speed; - +static void set_speed_features(VP9_COMP *cpi) { #if CONFIG_INTERNAL_STATS + int i; for (i = 0; i < MAX_MODES; ++i) cpi->mode_chosen_counts[i] = 0; #endif - // best quality defaults - sf->frame_parameter_update = 1; - sf->search_method = NSTEP; - sf->recode_loop = ALLOW_RECODE; - sf->subpel_search_method = SUBPEL_TREE; - sf->subpel_iters_per_step = 2; - sf->subpel_force_stop = 0; - sf->optimize_coefficients = !cpi->oxcf.lossless; - sf->reduce_first_step_size = 0; - sf->auto_mv_step_size = 0; - sf->max_step_search_steps = MAX_MVSEARCH_STEPS; - sf->comp_inter_joint_search_thresh = BLOCK_4X4; - sf->adaptive_rd_thresh = 0; - sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_OFF; - sf->tx_size_search_method = USE_FULL_RD; - sf->use_lp32x32fdct = 0; - sf->adaptive_motion_search = 0; - sf->adaptive_pred_interp_filter = 0; - sf->reference_masking = 0; - sf->partition_search_type = SEARCH_PARTITION; - sf->less_rectangular_check = 0; - sf->use_square_partition_only = 0; - sf->auto_min_max_partition_size = NOT_IN_USE; - sf->max_partition_size = BLOCK_64X64; - sf->min_partition_size = BLOCK_4X4; - sf->adjust_partitioning_from_last_frame = 0; - sf->last_partitioning_redo_frequency = 4; - sf->disable_split_mask = 0; - sf->mode_search_skip_flags = 0; - sf->disable_split_var_thresh = 0; - sf->disable_filter_search_var_thresh = 0; - for (i = 0; i < TX_SIZES; i++) { - sf->intra_y_mode_mask[i] = ALL_INTRA_MODES; - sf->intra_uv_mode_mask[i] = ALL_INTRA_MODES; - } - sf->use_rd_breakout = 0; - sf->skip_encode_sb = 0; - sf->use_uv_intra_rd_estimate = 0; - sf->allow_skip_recode = 0; - sf->lpf_pick = LPF_PICK_FROM_FULL_IMAGE; - sf->use_fast_coef_updates = 0; - sf->use_fast_coef_costing = 0; - sf->mode_skip_start = MAX_MODES; // Mode index at which mode skip mask set - sf->use_nonrd_pick_mode = 0; - sf->encode_breakout_thresh = 0; - for (i = 0; i < BLOCK_SIZES; ++i) - sf->disable_inter_mode_mask[i] = 0; - sf->max_intra_bsize = BLOCK_64X64; - // This setting only takes effect when partition_search_type is set - // to FIXED_PARTITION. - sf->always_this_block_size = BLOCK_16X16; - - switch (cpi->oxcf.mode) { - case MODE_BESTQUALITY: - case MODE_SECONDPASS_BEST: // This is the best quality mode. - cpi->diamond_search_sad = vp9_full_range_search; - break; - case MODE_FIRSTPASS: - case MODE_GOODQUALITY: - case MODE_SECONDPASS: - set_good_speed_feature(cpi, cm, sf, speed); - break; - case MODE_REALTIME: - set_rt_speed_feature(cm, sf, speed); - break; - }; /* switch */ + vp9_set_speed_features(cpi); // Set rd thresholds based on mode and speed setting set_rd_speed_thresholds(cpi); set_rd_speed_thresholds_sub8x8(cpi); - // Slow quant, dct and trellis not worthwhile for first pass - // so make sure they are always turned off. - if (cpi->pass == 1) { - sf->optimize_coefficients = 0; - } - - // No recode for 1 pass. - if (cpi->pass == 0) { - sf->recode_loop = DISALLOW_RECODE; - sf->optimize_coefficients = 0; - } - cpi->mb.fwd_txm4x4 = vp9_fdct4x4; if (cpi->oxcf.lossless || cpi->mb.e_mbd.lossless) { cpi->mb.fwd_txm4x4 = vp9_fwht4x4; } - - if (cpi->sf.subpel_search_method == SUBPEL_TREE) { - cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree; - cpi->find_fractional_mv_step_comp = vp9_find_best_sub_pixel_comp_tree; - } - - cpi->mb.optimize = cpi->sf.optimize_coefficients == 1 && cpi->pass != 1; - - if (cpi->encode_breakout && cpi->oxcf.mode == MODE_REALTIME && - sf->encode_breakout_thresh > cpi->encode_breakout) - cpi->encode_breakout = sf->encode_breakout_thresh; - - if (sf->disable_split_mask == DISABLE_ALL_SPLIT) - sf->adaptive_pred_interp_filter = 0; } static void alloc_raw_frame_buffers(VP9_COMP *cpi) { @@ -1736,7 +1370,7 @@ VP9_COMP *vp9_create_compressor(VP9_CONFIG *oxcf) { vp9_init_second_pass(cpi); } - vp9_set_speed_features(cpi); + set_speed_features(cpi); // Default rd threshold factors for mode selection for (i = 0; i < BLOCK_SIZES; ++i) { @@ -3042,7 +2676,7 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi, set_high_precision_mv(cpi, q < HIGH_PRECISION_MV_QTHRESH); } - vp9_set_speed_features(cpi); + set_speed_features(cpi); if (cpi->sf.recode_loop == DISALLOW_RECODE) { encode_without_recode_loop(cpi, size, dest, q); diff --git a/vp9/encoder/vp9_onyx_int.h b/vp9/encoder/vp9_onyx_int.h index 582fa0ea30..6dbe4d4742 100644 --- a/vp9/encoder/vp9_onyx_int.h +++ b/vp9/encoder/vp9_onyx_int.h @@ -31,6 +31,7 @@ #include "vp9/encoder/vp9_mcomp.h" #include "vp9/encoder/vp9_quantize.h" #include "vp9/encoder/vp9_ratectrl.h" +#include "vp9/encoder/vp9_speed_features.h" #include "vp9/encoder/vp9_svc_layercontext.h" #include "vp9/encoder/vp9_tokenize.h" #include "vp9/encoder/vp9_variance.h" @@ -114,75 +115,6 @@ typedef enum { THR_INTRA, } THR_MODES_SUB8X8; -typedef enum { - DIAMOND = 0, - NSTEP = 1, - HEX = 2, - BIGDIA = 3, - SQUARE = 4, - FAST_HEX = 5, - FAST_DIAMOND = 6 -} SEARCH_METHODS; - -typedef enum { - USE_FULL_RD = 0, - USE_LARGESTINTRA, - USE_LARGESTINTRA_MODELINTER, - USE_LARGESTALL -} TX_SIZE_SEARCH_METHOD; - -typedef enum { - NOT_IN_USE = 0, - RELAXED_NEIGHBORING_MIN_MAX = 1, - STRICT_NEIGHBORING_MIN_MAX = 2 -} AUTO_MIN_MAX_MODE; - -typedef enum { - // Terminate search early based on distortion so far compared to - // qp step, distortion in the neighborhood of the frame, etc. - FLAG_EARLY_TERMINATE = 1 << 0, - - // Skips comp inter modes if the best so far is an intra mode. - FLAG_SKIP_COMP_BESTINTRA = 1 << 1, - - // Skips comp inter modes if the best single intermode so far does - // not have the same reference as one of the two references being - // tested. - FLAG_SKIP_COMP_REFMISMATCH = 1 << 2, - - // Skips oblique intra modes if the best so far is an inter mode. - FLAG_SKIP_INTRA_BESTINTER = 1 << 3, - - // Skips oblique intra modes at angles 27, 63, 117, 153 if the best - // intra so far is not one of the neighboring directions. - FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4, - - // Skips intra modes other than DC_PRED if the source variance is small - FLAG_SKIP_INTRA_LOWVAR = 1 << 5, -} MODE_SEARCH_SKIP_LOGIC; - -typedef enum { - SUBPEL_TREE = 0, - // Other methods to come -} SUBPEL_SEARCH_METHODS; - -typedef enum { - LAST_FRAME_PARTITION_OFF = 0, - LAST_FRAME_PARTITION_LOW_MOTION = 1, - LAST_FRAME_PARTITION_ALL = 2 -} LAST_FRAME_PARTITION_METHOD; - -typedef enum { - // No recode. - DISALLOW_RECODE = 0, - // Allow recode for KF and exceeding maximum frame bandwidth. - ALLOW_RECODE_KFMAXBW = 1, - // Allow recode only for KF/ARF/GF frames. - ALLOW_RECODE_KFARFGF = 2, - // Allow recode for all frames based on bitrate constraints. - ALLOW_RECODE = 3, -} RECODE_LOOP_TYPE; - typedef enum { // encode_breakout is disabled. ENCODE_BREAKOUT_DISABLED = 0, @@ -192,227 +124,6 @@ typedef enum { ENCODE_BREAKOUT_LIMITED = 2 } ENCODE_BREAKOUT_TYPE; -typedef enum { - // Search partitions using RD/NONRD criterion - SEARCH_PARTITION = 0, - - // Always use a fixed size partition - FIXED_PARTITION = 1, - - // Use a fixed size partition in every 64X64 SB, where the size is - // determined based on source variance - VAR_BASED_FIXED_PARTITION = 2, - - REFERENCE_PARTITION = 3, - - // Use an arbitrary partitioning scheme based on source variance within - // a 64X64 SB - VAR_BASED_PARTITION -} PARTITION_SEARCH_TYPE; - -typedef enum { - // Try the full image with different values. - LPF_PICK_FROM_FULL_IMAGE, - // Try a small portion of the image with different values. - LPF_PICK_FROM_SUBIMAGE, - // Estimate the level based on quantizer and frame type - LPF_PICK_FROM_Q, -} LPF_PICK_METHOD; - -typedef struct { - // Frame level coding parameter update - int frame_parameter_update; - - // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc). - SEARCH_METHODS search_method; - - RECODE_LOOP_TYPE recode_loop; - - // Subpel_search_method can only be subpel_tree which does a subpixel - // logarithmic search that keeps stepping at 1/2 pixel units until - // you stop getting a gain, and then goes on to 1/4 and repeats - // the same process. Along the way it skips many diagonals. - SUBPEL_SEARCH_METHODS subpel_search_method; - - // Maximum number of steps in logarithmic subpel search before giving up. - int subpel_iters_per_step; - - // Control when to stop subpel search - int subpel_force_stop; - - // This parameter controls the number of steps we'll do in a diamond - // search. - int max_step_search_steps; - - // This parameter controls which step in the n-step process we start at. - // It's changed adaptively based on circumstances. - int reduce_first_step_size; - - // If this is set to 1, we limit the motion search range to 2 times the - // largest motion vector found in the last frame. - int auto_mv_step_size; - - // Trellis (dynamic programming) optimization of quantized values (+1, 0). - int optimize_coefficients; - - // Always set to 0. If on it enables 0 cost background transmission - // (except for the initial transmission of the segmentation). The feature is - // disabled because the addition of very large block sizes make the - // backgrounds very to cheap to encode, and the segmentation we have - // adds overhead. - int static_segmentation; - - // If 1 we iterate finding a best reference for 2 ref frames together - via - // a log search that iterates 4 times (check around mv for last for best - // error of combined predictor then check around mv for alt). If 0 we - // we just use the best motion vector found for each frame by itself. - int comp_inter_joint_search_thresh; - - // This variable is used to cap the maximum number of times we skip testing a - // mode to be evaluated. A high value means we will be faster. - int adaptive_rd_thresh; - - // Enables skipping the reconstruction step (idct, recon) in the - // intermediate steps assuming the last frame didn't have too many intra - // blocks and the q is less than a threshold. - int skip_encode_sb; - int skip_encode_frame; - // Speed feature to allow or disallow skipping of recode at block - // level within a frame. - int allow_skip_recode; - - // This variable allows us to reuse the last frames partition choices - // (64x64 v 32x32 etc) for this frame. It can be set to only use the last - // frame as a starting point in low motion scenes or always use it. If set - // we use last partitioning_redo frequency to determine how often to redo - // the partitioning from scratch. Adjust_partitioning_from_last_frame - // enables us to adjust up or down one partitioning from the last frames - // partitioning. - LAST_FRAME_PARTITION_METHOD use_lastframe_partitioning; - - // Determine which method we use to determine transform size. We can choose - // between options like full rd, largest for prediction size, largest - // for intra and model coefs for the rest. - TX_SIZE_SEARCH_METHOD tx_size_search_method; - - // Low precision 32x32 fdct keeps everything in 16 bits and thus is less - // precise but significantly faster than the non lp version. - int use_lp32x32fdct; - - // TODO(JBB): remove this as its no longer used. - - // After looking at the first set of modes (set by index here), skip - // checking modes for reference frames that don't match the reference frame - // of the best so far. - int mode_skip_start; - - // TODO(JBB): Remove this. - int reference_masking; - - PARTITION_SEARCH_TYPE partition_search_type; - - // Used if partition_search_type = FIXED_SIZE_PARTITION - BLOCK_SIZE always_this_block_size; - - // Skip rectangular partition test when partition type none gives better - // rd than partition type split. - int less_rectangular_check; - - // Disable testing non square partitions. (eg 16x32) - int use_square_partition_only; - - // Sets min and max partition sizes for this 64x64 region based on the - // same 64x64 in last encoded frame, and the left and above neighbor. - AUTO_MIN_MAX_MODE auto_min_max_partition_size; - - // Min and max partition size we enable (block_size) as per auto - // min max, but also used by adjust partitioning, and pick_partitioning. - BLOCK_SIZE min_partition_size; - BLOCK_SIZE max_partition_size; - - // Whether or not we allow partitions one smaller or one greater than the last - // frame's partitioning. Only used if use_lastframe_partitioning is set. - int adjust_partitioning_from_last_frame; - - // How frequently we re do the partitioning from scratch. Only used if - // use_lastframe_partitioning is set. - int last_partitioning_redo_frequency; - - // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable - // it always, to allow it for only Last frame and Intra, disable it for all - // inter modes or to enable it always. - int disable_split_mask; - - // TODO(jingning): combine the related motion search speed features - // This allows us to use motion search at other sizes as a starting - // point for this motion search and limits the search range around it. - int adaptive_motion_search; - - // Allows sub 8x8 modes to use the prediction filter that was determined - // best for 8x8 mode. If set to 0 we always re check all the filters for - // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter - // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected. - int adaptive_pred_interp_filter; - - // Search through variable block partition types in non-RD mode decision - // encoding process for RTC. - int partition_check; - - // Implements various heuristics to skip searching modes - // The heuristics selected are based on flags - // defined in the MODE_SEARCH_SKIP_HEURISTICS enum - unsigned int mode_search_skip_flags; - - // A source variance threshold below which the split mode is disabled - unsigned int disable_split_var_thresh; - - // A source variance threshold below which filter search is disabled - // Choose a very large value (UINT_MAX) to use 8-tap always - unsigned int disable_filter_search_var_thresh; - - // These bit masks allow you to enable or disable intra modes for each - // transform size separately. - int intra_y_mode_mask[TX_SIZES]; - int intra_uv_mode_mask[TX_SIZES]; - - // This variable enables an early break out of mode testing if the model for - // rd built from the prediction signal indicates a value that's much - // higher than the best rd we've seen so far. - int use_rd_breakout; - - // This enables us to use an estimate for intra rd based on dc mode rather - // than choosing an actual uv mode in the stage of encoding before the actual - // final encode. - int use_uv_intra_rd_estimate; - - // This feature controls how the loop filter level is determined. - LPF_PICK_METHOD lpf_pick; - - // This feature limits the number of coefficients updates we actually do - // by only looking at counts from 1/2 the bands. - int use_fast_coef_updates; // 0: 2-loop, 1: 1-loop, 2: 1-loop reduced - - // This flag controls the use of non-RD mode decision. - int use_nonrd_pick_mode; - - // This variable sets the encode_breakout threshold. Currently, it is only - // enabled in real time mode. - int encode_breakout_thresh; - - // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV - // modes are disabled in order from LSB to MSB for each BLOCK_SIZE. - int disable_inter_mode_mask[BLOCK_SIZES]; - - // This feature controls whether we do the expensive context update and - // calculation in the rd coefficient costing loop. - int use_fast_coef_costing; - - // This variable controls the maximum block size where intra blocks can be - // used in inter frames. - // TODO(aconverse): Fold this into one of the other many mode skips - BLOCK_SIZE max_intra_bsize; -} SPEED_FEATURES; - typedef enum { NORMAL = 0, FOURFIVE = 1, @@ -888,8 +599,6 @@ static INLINE YV12_BUFFER_CONFIG *get_ref_frame_buffer( .buf; } -void vp9_set_speed_features(VP9_COMP *cpi); - static INLINE int get_token_alloc(int mb_rows, int mb_cols) { // TODO(JBB): make this work for alpha channel and double check we can't // exceed this token count if we have a 32x32 transform crossing a boundary diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c new file mode 100644 index 0000000000..fd89d3304a --- /dev/null +++ b/vp9/encoder/vp9_speed_features.c @@ -0,0 +1,381 @@ +/* + * Copyright (c) 2010 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include <limits.h> + +#include "vp9/encoder/vp9_onyx_int.h" +#include "vp9/encoder/vp9_speed_features.h" + +#define ALL_INTRA_MODES 0x3FF +#define INTRA_DC_ONLY 0x01 +#define INTRA_DC_TM ((1 << TM_PRED) | (1 << DC_PRED)) +#define INTRA_DC_H_V ((1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED)) +#define INTRA_DC_TM_H_V (INTRA_DC_TM | (1 << V_PRED) | (1 << H_PRED)) + +// Masks for partially or completely disabling split mode +#define DISABLE_ALL_SPLIT 0x3F +#define DISABLE_ALL_INTER_SPLIT 0x1F +#define DISABLE_COMPOUND_SPLIT 0x18 +#define LAST_AND_INTRA_SPLIT_ONLY 0x1E + +// Intra only frames, golden frames (except alt ref overlays) and +// alt ref frames tend to be coded at a higher than ambient quality +static INLINE int frame_is_boosted(const VP9_COMP *cpi) { + return frame_is_intra_only(&cpi->common) || cpi->refresh_alt_ref_frame || + (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref); +} + +static void set_good_speed_feature(VP9_COMP *cpi, + VP9_COMMON *cm, + SPEED_FEATURES *sf, + int speed) { + int i; + sf->adaptive_rd_thresh = 1; + sf->recode_loop = ((speed < 1) ? ALLOW_RECODE : ALLOW_RECODE_KFMAXBW); + sf->allow_skip_recode = 1; + + if (speed >= 1) { + sf->use_square_partition_only = !frame_is_intra_only(cm); + sf->less_rectangular_check = 1; + sf->tx_size_search_method = frame_is_boosted(cpi) + ? USE_FULL_RD : USE_LARGESTALL; + + if (MIN(cm->width, cm->height) >= 720) + sf->disable_split_mask = cm->show_frame ? + DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; + else + sf->disable_split_mask = DISABLE_COMPOUND_SPLIT; + sf->use_rd_breakout = 1; + sf->adaptive_motion_search = 1; + sf->auto_mv_step_size = 1; + sf->adaptive_rd_thresh = 2; + sf->subpel_iters_per_step = 1; + sf->mode_skip_start = 10; + sf->adaptive_pred_interp_filter = 1; + + sf->recode_loop = ALLOW_RECODE_KFARFGF; + sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V; + sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V; + sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V; + sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V; + } + // Additions or changes from speed 1 for speed >= 2. + if (speed >= 2) { + sf->tx_size_search_method = frame_is_boosted(cpi) + ? USE_FULL_RD : USE_LARGESTALL; + + if (MIN(cm->width, cm->height) >= 720) + sf->disable_split_mask = cm->show_frame ? + DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; + else + sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY; + + sf->adaptive_pred_interp_filter = 2; + + sf->reference_masking = 1; + sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH | + FLAG_SKIP_INTRA_BESTINTER | + FLAG_SKIP_COMP_BESTINTRA | + FLAG_SKIP_INTRA_LOWVAR; + sf->disable_filter_search_var_thresh = 100; + sf->comp_inter_joint_search_thresh = BLOCK_SIZES; + + sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX; + sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_LOW_MOTION; + sf->adjust_partitioning_from_last_frame = 1; + sf->last_partitioning_redo_frequency = 3; + } + // Additions or changes for speed 3 and above + if (speed >= 3) { + if (MIN(cm->width, cm->height) >= 720) + sf->disable_split_mask = DISABLE_ALL_SPLIT; + else + sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT; + + sf->recode_loop = ALLOW_RECODE_KFMAXBW; + + sf->adaptive_rd_thresh = 3; + sf->mode_skip_start = 6; + sf->use_fast_coef_updates = 2; + sf->use_fast_coef_costing = 1; + } + // Additions or changes for speed 3 and above + if (speed >= 4) { + sf->use_square_partition_only = 1; + sf->tx_size_search_method = USE_LARGESTALL; + sf->disable_split_mask = DISABLE_ALL_SPLIT; + + sf->adaptive_rd_thresh = 4; + + // Add a couple more skip flags + sf->mode_search_skip_flags |= FLAG_SKIP_COMP_REFMISMATCH | + FLAG_EARLY_TERMINATE; + + sf->disable_filter_search_var_thresh = 200; + + sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_ALL; + sf->use_lp32x32fdct = 1; + } + if (speed >= 5) { + sf->partition_search_type = FIXED_PARTITION; + sf->optimize_coefficients = 0; + sf->search_method = HEX; + sf->disable_filter_search_var_thresh = 500; + for (i = 0; i < TX_SIZES; i++) { + sf->intra_y_mode_mask[i] = INTRA_DC_ONLY; + sf->intra_uv_mode_mask[i] = INTRA_DC_ONLY; + } + cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED; + } +} + +static void set_rt_speed_feature(VP9_COMMON *cm, + SPEED_FEATURES *sf, + int speed) { + sf->static_segmentation = 0; + sf->adaptive_rd_thresh = 1; + sf->encode_breakout_thresh = 1; + sf->use_fast_coef_costing = 1; + + if (speed == 1) { + sf->use_square_partition_only = !frame_is_intra_only(cm); + sf->less_rectangular_check = 1; + sf->tx_size_search_method = + frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL; + + if (MIN(cm->width, cm->height) >= 720) + sf->disable_split_mask = cm->show_frame ? + DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; + else + sf->disable_split_mask = DISABLE_COMPOUND_SPLIT; + + sf->use_rd_breakout = 1; + sf->adaptive_motion_search = 1; + sf->adaptive_pred_interp_filter = 1; + sf->auto_mv_step_size = 1; + sf->adaptive_rd_thresh = 2; + sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V; + sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V; + sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V; + sf->encode_breakout_thresh = 8; + } + if (speed >= 2) { + sf->use_square_partition_only = !frame_is_intra_only(cm); + sf->less_rectangular_check = 1; + sf->tx_size_search_method = + frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL; + + if (MIN(cm->width, cm->height) >= 720) + sf->disable_split_mask = cm->show_frame ? + DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; + else + sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY; + + sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH + | FLAG_SKIP_INTRA_BESTINTER | FLAG_SKIP_COMP_BESTINTRA + | FLAG_SKIP_INTRA_LOWVAR; + + sf->use_rd_breakout = 1; + sf->adaptive_motion_search = 1; + sf->adaptive_pred_interp_filter = 2; + sf->auto_mv_step_size = 1; + sf->reference_masking = 1; + + sf->disable_filter_search_var_thresh = 50; + sf->comp_inter_joint_search_thresh = BLOCK_SIZES; + + sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX; + sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_LOW_MOTION; + sf->adjust_partitioning_from_last_frame = 1; + sf->last_partitioning_redo_frequency = 3; + + sf->adaptive_rd_thresh = 2; + sf->use_lp32x32fdct = 1; + sf->mode_skip_start = 11; + sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V; + sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V; + sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V; + sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V; + sf->encode_breakout_thresh = 200; + } + if (speed >= 3) { + sf->use_square_partition_only = 1; + if (MIN(cm->width, cm->height) >= 720) + sf->disable_split_mask = DISABLE_ALL_SPLIT; + else + sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT; + + sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH + | FLAG_SKIP_INTRA_BESTINTER | FLAG_SKIP_COMP_BESTINTRA + | FLAG_SKIP_INTRA_LOWVAR; + + sf->disable_filter_search_var_thresh = 100; + sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_ALL; + sf->use_uv_intra_rd_estimate = 1; + sf->skip_encode_sb = 1; + sf->subpel_iters_per_step = 1; + sf->use_fast_coef_updates = 2; + sf->adaptive_rd_thresh = 4; + sf->mode_skip_start = 6; + sf->encode_breakout_thresh = 400; + sf->allow_skip_recode = 0; + } + if (speed >= 4) { + sf->optimize_coefficients = 0; + sf->disable_split_mask = DISABLE_ALL_SPLIT; + sf->lpf_pick = LPF_PICK_FROM_Q; + sf->encode_breakout_thresh = 700; + } + if (speed >= 5) { + int i; + sf->last_partitioning_redo_frequency = 4; + sf->adaptive_rd_thresh = 5; + sf->use_fast_coef_costing = 0; + sf->auto_min_max_partition_size = STRICT_NEIGHBORING_MIN_MAX; + sf->adjust_partitioning_from_last_frame = + cm->last_frame_type != cm->frame_type || (0 == + (cm->current_video_frame + 1) % sf->last_partitioning_redo_frequency); + sf->subpel_force_stop = 1; + for (i = 0; i < TX_SIZES; i++) { + sf->intra_y_mode_mask[i] = INTRA_DC_H_V; + sf->intra_uv_mode_mask[i] = INTRA_DC_ONLY; + } + sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_ONLY; + sf->frame_parameter_update = 0; + sf->encode_breakout_thresh = 1000; + sf->search_method = FAST_HEX; + sf->disable_inter_mode_mask[BLOCK_32X32] = 1 << INTER_OFFSET(ZEROMV); + sf->disable_inter_mode_mask[BLOCK_32X64] = ~(1 << INTER_OFFSET(NEARESTMV)); + sf->disable_inter_mode_mask[BLOCK_64X32] = ~(1 << INTER_OFFSET(NEARESTMV)); + sf->disable_inter_mode_mask[BLOCK_64X64] = ~(1 << INTER_OFFSET(NEARESTMV)); + sf->max_intra_bsize = BLOCK_32X32; + sf->allow_skip_recode = 1; + } + if (speed >= 6) { + sf->max_partition_size = BLOCK_32X32; + sf->min_partition_size = BLOCK_8X8; + sf->partition_check = + (cm->current_video_frame % sf->last_partitioning_redo_frequency == 1); + sf->partition_search_type = REFERENCE_PARTITION; + sf->use_nonrd_pick_mode = 1; + sf->search_method = FAST_DIAMOND; + sf->allow_skip_recode = 0; + } + if (speed >= 7) { + sf->partition_search_type = VAR_BASED_FIXED_PARTITION; + sf->use_nonrd_pick_mode = 1; + sf->search_method = FAST_DIAMOND; + } + if (speed >= 8) { + int i; + for (i = 0; i < BLOCK_SIZES; ++i) + sf->disable_inter_mode_mask[i] = 14; // only search NEARESTMV (0) + } +} + +void vp9_set_speed_features(VP9_COMP *cpi) { + SPEED_FEATURES *const sf = &cpi->sf; + VP9_COMMON *const cm = &cpi->common; + const int speed = cpi->speed < 0 ? -cpi->speed : cpi->speed; + int i; + + // best quality defaults + sf->frame_parameter_update = 1; + sf->search_method = NSTEP; + sf->recode_loop = ALLOW_RECODE; + sf->subpel_search_method = SUBPEL_TREE; + sf->subpel_iters_per_step = 2; + sf->subpel_force_stop = 0; + sf->optimize_coefficients = !cpi->oxcf.lossless; + sf->reduce_first_step_size = 0; + sf->auto_mv_step_size = 0; + sf->max_step_search_steps = MAX_MVSEARCH_STEPS; + sf->comp_inter_joint_search_thresh = BLOCK_4X4; + sf->adaptive_rd_thresh = 0; + sf->use_lastframe_partitioning = LAST_FRAME_PARTITION_OFF; + sf->tx_size_search_method = USE_FULL_RD; + sf->use_lp32x32fdct = 0; + sf->adaptive_motion_search = 0; + sf->adaptive_pred_interp_filter = 0; + sf->reference_masking = 0; + sf->partition_search_type = SEARCH_PARTITION; + sf->less_rectangular_check = 0; + sf->use_square_partition_only = 0; + sf->auto_min_max_partition_size = NOT_IN_USE; + sf->max_partition_size = BLOCK_64X64; + sf->min_partition_size = BLOCK_4X4; + sf->adjust_partitioning_from_last_frame = 0; + sf->last_partitioning_redo_frequency = 4; + sf->disable_split_mask = 0; + sf->mode_search_skip_flags = 0; + sf->disable_split_var_thresh = 0; + sf->disable_filter_search_var_thresh = 0; + for (i = 0; i < TX_SIZES; i++) { + sf->intra_y_mode_mask[i] = ALL_INTRA_MODES; + sf->intra_uv_mode_mask[i] = ALL_INTRA_MODES; + } + sf->use_rd_breakout = 0; + sf->skip_encode_sb = 0; + sf->use_uv_intra_rd_estimate = 0; + sf->allow_skip_recode = 0; + sf->lpf_pick = LPF_PICK_FROM_FULL_IMAGE; + sf->use_fast_coef_updates = 0; + sf->use_fast_coef_costing = 0; + sf->mode_skip_start = MAX_MODES; // Mode index at which mode skip mask set + sf->use_nonrd_pick_mode = 0; + sf->encode_breakout_thresh = 0; + for (i = 0; i < BLOCK_SIZES; ++i) + sf->disable_inter_mode_mask[i] = 0; + sf->max_intra_bsize = BLOCK_64X64; + // This setting only takes effect when partition_search_type is set + // to FIXED_PARTITION. + sf->always_this_block_size = BLOCK_16X16; + + switch (cpi->oxcf.mode) { + case MODE_BESTQUALITY: + case MODE_SECONDPASS_BEST: // This is the best quality mode. + cpi->diamond_search_sad = vp9_full_range_search; + break; + case MODE_FIRSTPASS: + case MODE_GOODQUALITY: + case MODE_SECONDPASS: + set_good_speed_feature(cpi, cm, sf, speed); + break; + case MODE_REALTIME: + set_rt_speed_feature(cm, sf, speed); + break; + } + + // Slow quant, dct and trellis not worthwhile for first pass + // so make sure they are always turned off. + if (cpi->pass == 1) { + sf->optimize_coefficients = 0; + } + + // No recode for 1 pass. + if (cpi->pass == 0) { + sf->recode_loop = DISALLOW_RECODE; + sf->optimize_coefficients = 0; + } + + if (cpi->sf.subpel_search_method == SUBPEL_TREE) { + cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree; + cpi->find_fractional_mv_step_comp = vp9_find_best_sub_pixel_comp_tree; + } + + cpi->mb.optimize = cpi->sf.optimize_coefficients == 1 && cpi->pass != 1; + + if (cpi->encode_breakout && cpi->oxcf.mode == MODE_REALTIME && + sf->encode_breakout_thresh > cpi->encode_breakout) + cpi->encode_breakout = sf->encode_breakout_thresh; + + if (sf->disable_split_mask == DISABLE_ALL_SPLIT) + sf->adaptive_pred_interp_filter = 0; +} diff --git a/vp9/encoder/vp9_speed_features.h b/vp9/encoder/vp9_speed_features.h new file mode 100644 index 0000000000..5091e2b06d --- /dev/null +++ b/vp9/encoder/vp9_speed_features.h @@ -0,0 +1,319 @@ +/* + * Copyright (c) 2010 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef VP9_ENCODER_VP9_SPEED_FEATURES_H_ +#define VP9_ENCODER_VP9_SPEED_FEATURES_H_ + +#include "vp9/common/vp9_enums.h" + +#ifdef __cplusplus +extern "C" { +#endif + +typedef enum { + DIAMOND = 0, + NSTEP = 1, + HEX = 2, + BIGDIA = 3, + SQUARE = 4, + FAST_HEX = 5, + FAST_DIAMOND = 6 +} SEARCH_METHODS; + +typedef enum { + // No recode. + DISALLOW_RECODE = 0, + // Allow recode for KF and exceeding maximum frame bandwidth. + ALLOW_RECODE_KFMAXBW = 1, + // Allow recode only for KF/ARF/GF frames. + ALLOW_RECODE_KFARFGF = 2, + // Allow recode for all frames based on bitrate constraints. + ALLOW_RECODE = 3, +} RECODE_LOOP_TYPE; + +typedef enum { + SUBPEL_TREE = 0, + // Other methods to come +} SUBPEL_SEARCH_METHODS; + +typedef enum { + LAST_FRAME_PARTITION_OFF = 0, + LAST_FRAME_PARTITION_LOW_MOTION = 1, + LAST_FRAME_PARTITION_ALL = 2 +} LAST_FRAME_PARTITION_METHOD; + +typedef enum { + USE_FULL_RD = 0, + USE_LARGESTINTRA, + USE_LARGESTINTRA_MODELINTER, + USE_LARGESTALL +} TX_SIZE_SEARCH_METHOD; + +typedef enum { + NOT_IN_USE = 0, + RELAXED_NEIGHBORING_MIN_MAX = 1, + STRICT_NEIGHBORING_MIN_MAX = 2 +} AUTO_MIN_MAX_MODE; + +typedef enum { + // Try the full image with different values. + LPF_PICK_FROM_FULL_IMAGE, + // Try a small portion of the image with different values. + LPF_PICK_FROM_SUBIMAGE, + // Estimate the level based on quantizer and frame type + LPF_PICK_FROM_Q, +} LPF_PICK_METHOD; + +typedef enum { + // Terminate search early based on distortion so far compared to + // qp step, distortion in the neighborhood of the frame, etc. + FLAG_EARLY_TERMINATE = 1 << 0, + + // Skips comp inter modes if the best so far is an intra mode. + FLAG_SKIP_COMP_BESTINTRA = 1 << 1, + + // Skips comp inter modes if the best single intermode so far does + // not have the same reference as one of the two references being + // tested. + FLAG_SKIP_COMP_REFMISMATCH = 1 << 2, + + // Skips oblique intra modes if the best so far is an inter mode. + FLAG_SKIP_INTRA_BESTINTER = 1 << 3, + + // Skips oblique intra modes at angles 27, 63, 117, 153 if the best + // intra so far is not one of the neighboring directions. + FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4, + + // Skips intra modes other than DC_PRED if the source variance is small + FLAG_SKIP_INTRA_LOWVAR = 1 << 5, +} MODE_SEARCH_SKIP_LOGIC; + +typedef enum { + // Search partitions using RD/NONRD criterion + SEARCH_PARTITION = 0, + + // Always use a fixed size partition + FIXED_PARTITION = 1, + + // Use a fixed size partition in every 64X64 SB, where the size is + // determined based on source variance + VAR_BASED_FIXED_PARTITION = 2, + + REFERENCE_PARTITION = 3, + + // Use an arbitrary partitioning scheme based on source variance within + // a 64X64 SB + VAR_BASED_PARTITION +} PARTITION_SEARCH_TYPE; + +typedef struct { + // Frame level coding parameter update + int frame_parameter_update; + + // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc). + SEARCH_METHODS search_method; + + RECODE_LOOP_TYPE recode_loop; + + // Subpel_search_method can only be subpel_tree which does a subpixel + // logarithmic search that keeps stepping at 1/2 pixel units until + // you stop getting a gain, and then goes on to 1/4 and repeats + // the same process. Along the way it skips many diagonals. + SUBPEL_SEARCH_METHODS subpel_search_method; + + // Maximum number of steps in logarithmic subpel search before giving up. + int subpel_iters_per_step; + + // Control when to stop subpel search + int subpel_force_stop; + + // This parameter controls the number of steps we'll do in a diamond + // search. + int max_step_search_steps; + + // This parameter controls which step in the n-step process we start at. + // It's changed adaptively based on circumstances. + int reduce_first_step_size; + + // If this is set to 1, we limit the motion search range to 2 times the + // largest motion vector found in the last frame. + int auto_mv_step_size; + + // Trellis (dynamic programming) optimization of quantized values (+1, 0). + int optimize_coefficients; + + // Always set to 0. If on it enables 0 cost background transmission + // (except for the initial transmission of the segmentation). The feature is + // disabled because the addition of very large block sizes make the + // backgrounds very to cheap to encode, and the segmentation we have + // adds overhead. + int static_segmentation; + + // If 1 we iterate finding a best reference for 2 ref frames together - via + // a log search that iterates 4 times (check around mv for last for best + // error of combined predictor then check around mv for alt). If 0 we + // we just use the best motion vector found for each frame by itself. + int comp_inter_joint_search_thresh; + + // This variable is used to cap the maximum number of times we skip testing a + // mode to be evaluated. A high value means we will be faster. + int adaptive_rd_thresh; + + // Enables skipping the reconstruction step (idct, recon) in the + // intermediate steps assuming the last frame didn't have too many intra + // blocks and the q is less than a threshold. + int skip_encode_sb; + int skip_encode_frame; + // Speed feature to allow or disallow skipping of recode at block + // level within a frame. + int allow_skip_recode; + + // This variable allows us to reuse the last frames partition choices + // (64x64 v 32x32 etc) for this frame. It can be set to only use the last + // frame as a starting point in low motion scenes or always use it. If set + // we use last partitioning_redo frequency to determine how often to redo + // the partitioning from scratch. Adjust_partitioning_from_last_frame + // enables us to adjust up or down one partitioning from the last frames + // partitioning. + LAST_FRAME_PARTITION_METHOD use_lastframe_partitioning; + + // Determine which method we use to determine transform size. We can choose + // between options like full rd, largest for prediction size, largest + // for intra and model coefs for the rest. + TX_SIZE_SEARCH_METHOD tx_size_search_method; + + // Low precision 32x32 fdct keeps everything in 16 bits and thus is less + // precise but significantly faster than the non lp version. + int use_lp32x32fdct; + + // TODO(JBB): remove this as its no longer used. + + // After looking at the first set of modes (set by index here), skip + // checking modes for reference frames that don't match the reference frame + // of the best so far. + int mode_skip_start; + + // TODO(JBB): Remove this. + int reference_masking; + + PARTITION_SEARCH_TYPE partition_search_type; + + // Used if partition_search_type = FIXED_SIZE_PARTITION + BLOCK_SIZE always_this_block_size; + + // Skip rectangular partition test when partition type none gives better + // rd than partition type split. + int less_rectangular_check; + + // Disable testing non square partitions. (eg 16x32) + int use_square_partition_only; + + // Sets min and max partition sizes for this 64x64 region based on the + // same 64x64 in last encoded frame, and the left and above neighbor. + AUTO_MIN_MAX_MODE auto_min_max_partition_size; + + // Min and max partition size we enable (block_size) as per auto + // min max, but also used by adjust partitioning, and pick_partitioning. + BLOCK_SIZE min_partition_size; + BLOCK_SIZE max_partition_size; + + // Whether or not we allow partitions one smaller or one greater than the last + // frame's partitioning. Only used if use_lastframe_partitioning is set. + int adjust_partitioning_from_last_frame; + + // How frequently we re do the partitioning from scratch. Only used if + // use_lastframe_partitioning is set. + int last_partitioning_redo_frequency; + + // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable + // it always, to allow it for only Last frame and Intra, disable it for all + // inter modes or to enable it always. + int disable_split_mask; + + // TODO(jingning): combine the related motion search speed features + // This allows us to use motion search at other sizes as a starting + // point for this motion search and limits the search range around it. + int adaptive_motion_search; + + // Allows sub 8x8 modes to use the prediction filter that was determined + // best for 8x8 mode. If set to 0 we always re check all the filters for + // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter + // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected. + int adaptive_pred_interp_filter; + + // Search through variable block partition types in non-RD mode decision + // encoding process for RTC. + int partition_check; + + // Implements various heuristics to skip searching modes + // The heuristics selected are based on flags + // defined in the MODE_SEARCH_SKIP_HEURISTICS enum + unsigned int mode_search_skip_flags; + + // A source variance threshold below which the split mode is disabled + unsigned int disable_split_var_thresh; + + // A source variance threshold below which filter search is disabled + // Choose a very large value (UINT_MAX) to use 8-tap always + unsigned int disable_filter_search_var_thresh; + + // These bit masks allow you to enable or disable intra modes for each + // transform size separately. + int intra_y_mode_mask[TX_SIZES]; + int intra_uv_mode_mask[TX_SIZES]; + + // This variable enables an early break out of mode testing if the model for + // rd built from the prediction signal indicates a value that's much + // higher than the best rd we've seen so far. + int use_rd_breakout; + + // This enables us to use an estimate for intra rd based on dc mode rather + // than choosing an actual uv mode in the stage of encoding before the actual + // final encode. + int use_uv_intra_rd_estimate; + + // This feature controls how the loop filter level is determined. + LPF_PICK_METHOD lpf_pick; + + // This feature limits the number of coefficients updates we actually do + // by only looking at counts from 1/2 the bands. + int use_fast_coef_updates; // 0: 2-loop, 1: 1-loop, 2: 1-loop reduced + + // This flag controls the use of non-RD mode decision. + int use_nonrd_pick_mode; + + // This variable sets the encode_breakout threshold. Currently, it is only + // enabled in real time mode. + int encode_breakout_thresh; + + // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV + // modes are disabled in order from LSB to MSB for each BLOCK_SIZE. + int disable_inter_mode_mask[BLOCK_SIZES]; + + // This feature controls whether we do the expensive context update and + // calculation in the rd coefficient costing loop. + int use_fast_coef_costing; + + // This variable controls the maximum block size where intra blocks can be + // used in inter frames. + // TODO(aconverse): Fold this into one of the other many mode skips + BLOCK_SIZE max_intra_bsize; +} SPEED_FEATURES; + +struct VP9_COMP; + +void vp9_set_speed_features(struct VP9_COMP *cpi); + +#ifdef __cplusplus +} // extern "C" +#endif + +#endif // VP9_ENCODER_VP9_SPEED_FEATURES_H_ + diff --git a/vp9/vp9cx.mk b/vp9/vp9cx.mk index e705991965..21a388e514 100644 --- a/vp9/vp9cx.mk +++ b/vp9/vp9cx.mk @@ -60,6 +60,8 @@ VP9_CX_SRCS-yes += encoder/vp9_pickmode.c VP9_CX_SRCS-yes += encoder/vp9_sad.c VP9_CX_SRCS-yes += encoder/vp9_segmentation.c VP9_CX_SRCS-yes += encoder/vp9_segmentation.h +VP9_CX_SRCS-yes += encoder/vp9_speed_features.c +VP9_CX_SRCS-yes += encoder/vp9_speed_features.h VP9_CX_SRCS-yes += encoder/vp9_subexp.c VP9_CX_SRCS-yes += encoder/vp9_subexp.h VP9_CX_SRCS-yes += encoder/vp9_svc_layercontext.c -- GitLab