#endif // !CONFIG_REALTIME_ONLY
// Test for whether to calculate metrics for the frame.
-static int is_psnr_calc_enabled(VP9_COMP *cpi) {
- VP9_COMMON *const cm = &cpi->common;
+static int is_psnr_calc_enabled(const VP9_COMP *cpi) {
+ const VP9_COMMON *const cm = &cpi->common;
const VP9EncoderConfig *const oxcf = &cpi->oxcf;
return cpi->b_calculate_psnr && (oxcf->pass != 1) && cm->show_frame;
}
CHECK_MEM_ERROR(cm, roi->roi_map, vpx_malloc(rows * cols));
- // Copy to ROI sturcture in the compressor.
+ // Copy to ROI structure in the compressor.
memcpy(roi->roi_map, map, rows * cols);
memcpy(&roi->delta_q, delta_q, MAX_SEGMENTS * sizeof(delta_q[0]));
memcpy(&roi->delta_lf, delta_lf, MAX_SEGMENTS * sizeof(delta_lf[0]));
}
}
-static void init_config(struct VP9_COMP *cpi, VP9EncoderConfig *oxcf) {
+static void init_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) {
VP9_COMMON *const cm = &cpi->common;
cpi->oxcf = *oxcf;
}
}
-VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
+static void update_initial_width(VP9_COMP *cpi, int use_highbitdepth,
+ int subsampling_x, int subsampling_y) {
+ VP9_COMMON *const cm = &cpi->common;
+#if !CONFIG_VP9_HIGHBITDEPTH
+ (void)use_highbitdepth;
+ assert(use_highbitdepth == 0);
+#endif
+
+ if (!cpi->initial_width ||
+#if CONFIG_VP9_HIGHBITDEPTH
+ cm->use_highbitdepth != use_highbitdepth ||
+#endif
+ cm->subsampling_x != subsampling_x ||
+ cm->subsampling_y != subsampling_y) {
+ cm->subsampling_x = subsampling_x;
+ cm->subsampling_y = subsampling_y;
+#if CONFIG_VP9_HIGHBITDEPTH
+ cm->use_highbitdepth = use_highbitdepth;
+#endif
+ alloc_util_frame_buffers(cpi);
+ cpi->initial_width = cm->width;
+ cpi->initial_height = cm->height;
+ cpi->initial_mbs = cm->MBs;
+ }
+}
+
+// TODO(angiebird): Check whether we can move this function to vpx_image.c
+static INLINE void vpx_img_chroma_subsampling(vpx_img_fmt_t fmt,
+ unsigned int *subsampling_x,
+ unsigned int *subsampling_y) {
+ switch (fmt) {
+ case VPX_IMG_FMT_I420:
+ case VPX_IMG_FMT_YV12:
+ case VPX_IMG_FMT_I422:
+ case VPX_IMG_FMT_I42016:
+ case VPX_IMG_FMT_I42216: *subsampling_x = 1; break;
+ default: *subsampling_x = 0; break;
+ }
+
+ switch (fmt) {
+ case VPX_IMG_FMT_I420:
+ case VPX_IMG_FMT_I440:
+ case VPX_IMG_FMT_YV12:
+ case VPX_IMG_FMT_I42016:
+ case VPX_IMG_FMT_I44016: *subsampling_y = 1; break;
+ default: *subsampling_y = 0; break;
+ }
+}
+
+// TODO(angiebird): Check whether we can move this function to vpx_image.c
+static INLINE int vpx_img_use_highbitdepth(vpx_img_fmt_t fmt) {
+ return fmt & VPX_IMG_FMT_HIGHBITDEPTH;
+}
+
+#if CONFIG_VP9_TEMPORAL_DENOISING
+static void setup_denoiser_buffer(VP9_COMP *cpi) {
+ VP9_COMMON *const cm = &cpi->common;
+ if (cpi->oxcf.noise_sensitivity > 0 &&
+ !cpi->denoiser.frame_buffer_initialized) {
+ if (vp9_denoiser_alloc(cm, &cpi->svc, &cpi->denoiser, cpi->use_svc,
+ cpi->oxcf.noise_sensitivity, cm->width, cm->height,
+ cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+ cm->use_highbitdepth,
+#endif
+ VP9_ENC_BORDER_IN_PIXELS))
+ vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+ "Failed to allocate denoiser");
+ }
+}
+#endif
+
+void vp9_update_compressor_with_img_fmt(VP9_COMP *cpi, vpx_img_fmt_t img_fmt) {
+ const VP9EncoderConfig *oxcf = &cpi->oxcf;
+ unsigned int subsampling_x, subsampling_y;
+ const int use_highbitdepth = vpx_img_use_highbitdepth(img_fmt);
+ vpx_img_chroma_subsampling(img_fmt, &subsampling_x, &subsampling_y);
+
+ update_initial_width(cpi, use_highbitdepth, subsampling_x, subsampling_y);
+#if CONFIG_VP9_TEMPORAL_DENOISING
+ setup_denoiser_buffer(cpi);
+#endif
+
+ assert(cpi->lookahead == NULL);
+ cpi->lookahead = vp9_lookahead_init(oxcf->width, oxcf->height, subsampling_x,
+ subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+ use_highbitdepth,
+#endif
+ oxcf->lag_in_frames);
+ alloc_raw_frame_buffers(cpi);
+}
+
+VP9_COMP *vp9_create_compressor(const VP9EncoderConfig *oxcf,
BufferPool *const pool) {
unsigned int i;
VP9_COMP *volatile const cpi = vpx_memalign(32, sizeof(VP9_COMP));
cm->error.setjmp = 0;
+#if CONFIG_RATE_CTRL
+ encode_command_init(&cpi->encode_command);
+#endif
+
return cpi;
}
#endif
}
-static void generate_psnr_packet(VP9_COMP *cpi) {
- struct vpx_codec_cx_pkt pkt;
- int i;
- PSNR_STATS psnr;
+int vp9_get_psnr(const VP9_COMP *cpi, PSNR_STATS *psnr) {
+ if (is_psnr_calc_enabled(cpi)) {
#if CONFIG_VP9_HIGHBITDEPTH
- vpx_calc_highbd_psnr(cpi->raw_source_frame, cpi->common.frame_to_show, &psnr,
- cpi->td.mb.e_mbd.bd, cpi->oxcf.input_bit_depth);
+ vpx_calc_highbd_psnr(cpi->raw_source_frame, cpi->common.frame_to_show, psnr,
+ cpi->td.mb.e_mbd.bd, cpi->oxcf.input_bit_depth);
#else
- vpx_calc_psnr(cpi->raw_source_frame, cpi->common.frame_to_show, &psnr);
+ vpx_calc_psnr(cpi->raw_source_frame, cpi->common.frame_to_show, psnr);
#endif
-
- for (i = 0; i < 4; ++i) {
- pkt.data.psnr.samples[i] = psnr.samples[i];
- pkt.data.psnr.sse[i] = psnr.sse[i];
- pkt.data.psnr.psnr[i] = psnr.psnr[i];
- }
- pkt.kind = VPX_CODEC_PSNR_PKT;
- if (!cpi->use_svc) {
- vpx_codec_pkt_list_add(cpi->output_pkt_list, &pkt);
+ return 1;
+ } else {
+ vp9_zero(*psnr);
+ return 0;
}
}
#endif // CONFIG_VP9_POSTPROC
}
-#if CONFIG_VP9_TEMPORAL_DENOISING
-static void setup_denoiser_buffer(VP9_COMP *cpi) {
- VP9_COMMON *const cm = &cpi->common;
- if (cpi->oxcf.noise_sensitivity > 0 &&
- !cpi->denoiser.frame_buffer_initialized) {
- if (vp9_denoiser_alloc(cm, &cpi->svc, &cpi->denoiser, cpi->use_svc,
- cpi->oxcf.noise_sensitivity, cm->width, cm->height,
- cm->subsampling_x, cm->subsampling_y,
-#if CONFIG_VP9_HIGHBITDEPTH
- cm->use_highbitdepth,
-#endif
- VP9_ENC_BORDER_IN_PIXELS))
- vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
- "Failed to allocate denoiser");
- }
-}
-#endif
-
static void init_motion_estimation(VP9_COMP *cpi) {
int y_stride = cpi->scaled_source.y_stride;
vp9_scale_references(cpi);
}
+#if CONFIG_RATE_CTRL
+ // TODO(angiebird): This is a hack for making sure the encoder use the
+ // external_quantize_index exactly. Avoid this kind of hack later.
+ if (cpi->encode_command.use_external_quantize_index) {
+ q = cpi->encode_command.external_quantize_index;
+ }
+#endif
+
vp9_set_quantizer(cm, q);
if (loop_count == 0) setup_frame(cpi);
if (frame_over_shoot_limit == 0) frame_over_shoot_limit = 1;
}
+#if CONFIG_RATE_CTRL
+ // This part needs to be after save_coding_context() because
+ // restore_coding_context will be called in the end of this function.
+ // TODO(angiebird): This is a hack for making sure the encoder use the
+ // external_quantize_index exactly. Avoid this kind of hack later.
+ if (cpi->encode_command.use_external_quantize_index) {
+ break;
+ }
+#endif
+
if (oxcf->rc_mode == VPX_Q) {
loop = 0;
} else {
// Special case reset for qlow for constrained quality.
// This should only trigger where there is very substantial
// undershoot on a frame and the auto cq level is above
- // the user passsed in value.
+ // the user passed in value.
if (oxcf->rc_mode == VPX_CQ && q < q_low) {
q_low = q;
}
}
if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF)
- if (loop || !enable_acl) restore_coding_context(cpi);
+ if (loop) restore_coding_context(cpi);
} while (loop);
#ifdef AGGRESSIVE_VBR
// Skip recoding, if model diff is below threshold
const int thresh = compute_context_model_thresh(cpi);
const int diff = compute_context_model_diff(cm);
- if (diff < thresh) {
- vpx_clear_system_state();
- restore_coding_context(cpi);
- return;
+ if (diff >= thresh) {
+ vp9_encode_frame(cpi);
}
-
- vp9_encode_frame(cpi);
+ }
+ if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF) {
vpx_clear_system_state();
restore_coding_context(cpi);
}
}
#ifdef ENABLE_KF_DENOISE
-// Baseline Kernal weights for denoise
+// Baseline kernel weights for denoise
static uint8_t dn_kernal_3[9] = { 1, 2, 1, 2, 4, 2, 1, 2, 1 };
static uint8_t dn_kernal_5[25] = { 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 2, 4,
2, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1 };
tmp_ptr += stride;
}
- // Select the kernal size.
+ // Select the kernel size.
if (max_diff > (strength + (strength >> 1))) {
kernal_size = 3;
half_k_size = 1;
}
kernal_ptr = (kernal_size == 3) ? dn_kernal_3 : dn_kernal_5;
- // Apply the kernal
+ // Apply the kernel
tmp_ptr = src_ptr - (stride * half_k_size) - half_k_size;
for (i = 0; i < kernal_size; ++i) {
for (j = 0; j < kernal_size; ++j) {
tmp_ptr += stride;
}
- // Select the kernal size.
+ // Select the kernel size.
if (max_diff > (strength + (strength >> 1))) {
kernal_size = 3;
half_k_size = 1;
}
kernal_ptr = (kernal_size == 3) ? dn_kernal_3 : dn_kernal_5;
- // Apply the kernal
+ // Apply the kernel
tmp_ptr = src_ptr - (stride * half_k_size) - half_k_size;
for (i = 0; i < kernal_size; ++i) {
for (j = 0; j < kernal_size; ++j) {
}
#endif // CONFIG_VP9_HIGHBITDEPTH
-// Apply thresholded spatial noise supression to a given buffer.
+// Apply thresholded spatial noise suppression to a given buffer.
static void spatial_denoise_buffer(VP9_COMP *cpi, uint8_t *buffer,
const int stride, const int width,
const int height, const int strength) {
}
}
-// Apply thresholded spatial noise supression to source.
+// Apply thresholded spatial noise suppression to source.
static void spatial_denoise_frame(VP9_COMP *cpi) {
YV12_BUFFER_CONFIG *src = cpi->Source;
const VP9EncoderConfig *const oxcf = &cpi->oxcf;
}
}
-// Implementation and modifications of C. Yeo, H. L. Tan, and Y. H. Tan, "On
-// rate distortion optimization using SSIM," Circuits and Systems for Video
-// Technology, IEEE Transactions on, vol. 23, no. 7, pp. 1170-1181, 2013.
-// SSIM_VAR_SCALE defines the strength of the bias towards SSIM in RDO.
-// Some sample values are:
-// (for midres test set)
-// SSIM_VAR_SCALE avg_psnr ssim ms_ssim
-// 8.0 9.421 -5.537 -6.898
-// 16.0 4.703 -5.378 -6.238
-// 32.0 1.929 -4.308 -4.807
-#define SSIM_VAR_SCALE 16.0
static void set_mb_ssim_rdmult_scaling(VP9_COMP *cpi) {
VP9_COMMON *cm = &cpi->common;
ThreadData *td = &cpi->td;
double log_sum = 0.0;
int row, col;
- const double c2 = 58.5225 * SSIM_VAR_SCALE; // 58.5225 = (.03*255)^2
-
// Loop through each 64x64 block.
for (row = 0; row < num_rows; ++row) {
for (col = 0; col < num_cols; ++col) {
}
}
var = var / num_of_var / 64.0;
- var = 2.0 * var + c2;
+
+ // Curve fitting with an exponential model on all 16x16 blocks from the
+ // Midres dataset.
+ var = 67.035434 * (1 - exp(-0.0021489 * var)) + 17.492222;
cpi->mi_ssim_rdmult_scaling_factors[index] = var;
log_sum += log(var);
}
mismatch_move_frame_idx_w();
#endif
encode_frame_to_data_rate(cpi, size, dest, frame_flags);
-
- vp9_twopass_postencode_update(cpi);
}
#endif // !CONFIG_REALTIME_ONLY
-static void update_initial_width(VP9_COMP *cpi, int use_highbitdepth,
- int subsampling_x, int subsampling_y) {
- VP9_COMMON *const cm = &cpi->common;
-#if !CONFIG_VP9_HIGHBITDEPTH
- (void)use_highbitdepth;
- assert(use_highbitdepth == 0);
-#endif
-
- if (!cpi->initial_width ||
-#if CONFIG_VP9_HIGHBITDEPTH
- cm->use_highbitdepth != use_highbitdepth ||
-#endif
- cm->subsampling_x != subsampling_x ||
- cm->subsampling_y != subsampling_y) {
- cm->subsampling_x = subsampling_x;
- cm->subsampling_y = subsampling_y;
-#if CONFIG_VP9_HIGHBITDEPTH
- cm->use_highbitdepth = use_highbitdepth;
-#endif
-
- cpi->initial_width = cm->width;
- cpi->initial_height = cm->height;
- cpi->initial_mbs = cm->MBs;
- }
-}
-
int vp9_receive_raw_frame(VP9_COMP *cpi, vpx_enc_frame_flags_t frame_flags,
YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
int64_t end_time) {
int step_param;
uint32_t bestsme = UINT_MAX;
const MvLimits tmp_mv_limits = x->mv_limits;
- // lambda is used to adjust the importance of motion vector consitency.
+ // lambda is used to adjust the importance of motion vector consistency.
// TODO(angiebird): Figure out lambda's proper value.
const int lambda = cpi->tpl_stats[frame_idx].lambda;
int_mv nb_full_mvs[NB_MVS_NUM];
assert(kMvPreCheckSize == (kMvPreCheckLines * (kMvPreCheckLines + 1)) >> 1);
// no new mv
- // diagnal scan order
+ // diagonal scan order
tmp_idx = 0;
for (idx = 0; idx < kMvPreCheckLines; ++idx) {
int r;
#endif // CONFIG_NON_GREEDY_MV
}
+static void init_encode_frame_result(ENCODE_FRAME_RESULT *encode_frame_result) {
+ encode_frame_result->show_idx = -1; // Actual encoding doesn't happen.
+}
+
+#if !CONFIG_REALTIME_ONLY
+static void update_encode_frame_result(
+ int show_idx, FRAME_UPDATE_TYPE update_type,
+ const YV12_BUFFER_CONFIG *source_frame,
+ const YV12_BUFFER_CONFIG *coded_frame, int quantize_index,
+ uint32_t bit_depth, uint32_t input_bit_depth,
+ ENCODE_FRAME_RESULT *encode_frame_result) {
+#if CONFIG_RATE_CTRL
+ PSNR_STATS psnr;
+#if CONFIG_VP9_HIGHBITDEPTH
+ vpx_calc_highbd_psnr(source_frame, coded_frame, &psnr, bit_depth,
+ input_bit_depth);
+#else // CONFIG_VP9_HIGHBITDEPTH
+ (void)bit_depth;
+ (void)input_bit_depth;
+ vpx_calc_psnr(source_frame, coded_frame, &psnr);
+#endif // CONFIG_VP9_HIGHBITDEPTH
+ encode_frame_result->psnr = psnr.psnr[0];
+ encode_frame_result->sse = psnr.sse[0];
+#else // CONFIG_RATE_CTRL
+ (void)bit_depth;
+ (void)input_bit_depth;
+ (void)source_frame;
+ (void)coded_frame;
+#endif // CONFIG_RATE_CTRL
+ encode_frame_result->show_idx = show_idx;
+ encode_frame_result->update_type = update_type;
+ encode_frame_result->quantize_index = quantize_index;
+}
+#endif // !CONFIG_REALTIME_ONLY
+
int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
size_t *size, uint8_t *dest, int64_t *time_stamp,
- int64_t *time_end, int flush) {
+ int64_t *time_end, int flush,
+ ENCODE_FRAME_RESULT *encode_frame_result) {
const VP9EncoderConfig *const oxcf = &cpi->oxcf;
VP9_COMMON *const cm = &cpi->common;
BufferPool *const pool = cm->buffer_pool;
int arf_src_index;
const int gf_group_index = cpi->twopass.gf_group.index;
int i;
+ init_encode_frame_result(encode_frame_result);
if (is_one_pass_cbr_svc(cpi)) {
vp9_one_pass_cbr_svc_start_layer(cpi);
*frame_flags = (source->flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
} else {
*size = 0;
-#if !CONFIG_REALTIME_ONLY
- if (flush && oxcf->pass == 1 && !cpi->twopass.first_pass_done) {
- vp9_end_first_pass(cpi); /* get last stats packet */
- cpi->twopass.first_pass_done = 1;
- }
-#endif // !CONFIG_REALTIME_ONLY
return -1;
}
vp9_first_pass(cpi, source);
} else if (oxcf->pass == 2 && !cpi->use_svc) {
Pass2Encode(cpi, size, dest, frame_flags);
+ // update_encode_frame_result() depends on twopass.gf_group.index and
+ // cm->new_fb_idx and cpi->Source are updated for current properly and have
+ // not been updated for the next frame yet.
+ // The update locations are as follows.
+ // 1) twopass.gf_group.index is initialized at define_gf_group by vp9_zero()
+ // for the first frame in the gf_group and is updated for the next frame at
+ // vp9_twopass_postencode_update().
+ // 2) cpi->Source is updated at the beginning of this function, i.e.
+ // vp9_get_compressed_data()
+ // 3) cm->new_fb_idx is updated at the beginning of this function by
+ // get_free_fb(cm)
+ // TODO(angiebird): Improve the codebase to make the update of frame
+ // dependent variables more robust.
+ update_encode_frame_result(
+ source->show_idx,
+ cpi->twopass.gf_group.update_type[cpi->twopass.gf_group.index],
+ cpi->Source, get_frame_new_buffer(cm), vp9_get_quantizer(cpi),
+ cpi->oxcf.input_bit_depth, cm->bit_depth, encode_frame_result);
+ vp9_twopass_postencode_update(cpi);
} else if (cpi->use_svc) {
SvcEncode(cpi, size, dest, frame_flags);
} else {
vpx_usec_timer_mark(&cmptimer);
cpi->time_compress_data += vpx_usec_timer_elapsed(&cmptimer);
- // Should we calculate metrics for the frame.
- if (is_psnr_calc_enabled(cpi)) generate_psnr_packet(cpi);
-
if (cpi->keep_level_stats && oxcf->pass != 1)
update_level_info(cpi, size, arf_src_index);
return;
}
-int vp9_get_quantizer(VP9_COMP *cpi) { return cpi->common.base_qindex; }
+int vp9_get_quantizer(const VP9_COMP *cpi) { return cpi->common.base_qindex; }
void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags) {
if (flags &