Merge "Refactor denoiser frame buffer update"
diff --git a/test/datarate_test.cc b/test/datarate_test.cc
index 3d40148..573870e 100644
--- a/test/datarate_test.cc
+++ b/test/datarate_test.cc
@@ -246,6 +246,8 @@
for (int i = 0; i < 3; ++i) {
bits_total_[i] = 0;
}
+ denoiser_offon_test_ = 0;
+ denoiser_offon_period_ = -1;
}
//
@@ -316,6 +318,15 @@
if (video->frame() == 1)
encoder->Control(VP8E_SET_CPUUSED, set_cpu_used_);
+ if (denoiser_offon_test_) {
+ ASSERT_GT(denoiser_offon_period_, 0)
+ << "denoiser_offon_period_ is not positive.";
+ if ((video->frame() + 1) % denoiser_offon_period_ == 0) {
+ // Flip denoiser_on_ periodically
+ denoiser_on_ ^= 1;
+ }
+ }
+
encoder->Control(VP9E_SET_NOISE_SENSITIVITY, denoiser_on_);
if (cfg_.ts_number_layers > 1) {
@@ -399,6 +410,8 @@
vpx_codec_pts_t first_drop_;
int num_drops_;
int denoiser_on_;
+ int denoiser_offon_test_;
+ int denoiser_offon_period_;
};
// Check basic rate targeting,
@@ -653,6 +666,38 @@
ASSERT_LE(effective_datarate_[0], cfg_.rc_target_bitrate * 1.15)
<< " The datarate for the file is greater than target by too much!";
}
+
+// Check basic datarate targeting, for a single bitrate, when denoiser is off
+// and on.
+TEST_P(DatarateTestVP9Large, DenoiserOffon) {
+ cfg_.rc_buf_initial_sz = 500;
+ cfg_.rc_buf_optimal_sz = 500;
+ cfg_.rc_buf_sz = 1000;
+ cfg_.rc_dropframe_thresh = 1;
+ cfg_.rc_min_quantizer = 2;
+ cfg_.rc_max_quantizer = 56;
+ cfg_.rc_end_usage = VPX_CBR;
+ cfg_.g_lag_in_frames = 0;
+
+ ::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+ 30, 1, 0, 299);
+
+ // For the temporal denoiser (#if CONFIG_VP9_TEMPORAL_DENOISING),
+ // there is only one denoiser mode: denoiserYonly(which is 1),
+ // but may add more modes in the future.
+ cfg_.rc_target_bitrate = 300;
+ ResetModel();
+ // The denoiser is off by default.
+ denoiser_on_ = 0;
+ // Set the offon test flag.
+ denoiser_offon_test_ = 1;
+ denoiser_offon_period_ = 100;
+ ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+ ASSERT_GE(effective_datarate_[0], cfg_.rc_target_bitrate * 0.85)
+ << " The datarate for the file is lower than target by too much!";
+ ASSERT_LE(effective_datarate_[0], cfg_.rc_target_bitrate * 1.15)
+ << " The datarate for the file is greater than target by too much!";
+}
#endif // CONFIG_VP9_TEMPORAL_DENOISING
VP8_INSTANTIATE_TEST_CASE(DatarateTestLarge, ALL_TEST_MODES);
diff --git a/vp9/common/vp9_mfqe.c b/vp9/common/vp9_mfqe.c
index 110404f..e7ee903 100644
--- a/vp9/common/vp9_mfqe.c
+++ b/vp9/common/vp9_mfqe.c
@@ -136,13 +136,27 @@
}
}
+static void get_thr(BLOCK_SIZE bs, int qdiff, int *sad_thr, int *vdiff_thr) {
+ const int adj = qdiff >> MFQE_PRECISION;
+ if (bs == BLOCK_16X16) {
+ *sad_thr = 7 + adj;
+ } else if (bs == BLOCK_32X32) {
+ *sad_thr = 6 + adj;
+ } else { // BLOCK_64X64
+ *sad_thr = 5 + adj;
+ }
+ *vdiff_thr = 125 + qdiff;
+}
+
static void mfqe_block(BLOCK_SIZE bs, const uint8_t *y, const uint8_t *u,
const uint8_t *v, int y_stride, int uv_stride,
- uint8_t *yd, uint8_t *ud, uint8_t *vd,
- int yd_stride, int uvd_stride) {
- int sad, sad_thr, vdiff;
+ uint8_t *yd, uint8_t *ud, uint8_t *vd, int yd_stride,
+ int uvd_stride, int qdiff) {
+ int sad, sad_thr, vdiff, vdiff_thr;
uint32_t sse;
+ get_thr(bs, qdiff, &sad_thr, &vdiff_thr);
+
if (bs == BLOCK_16X16) {
vdiff = (vp9_variance16x16(y, y_stride, yd, yd_stride, &sse) + 128) >> 8;
sad = (vp9_sad16x16(y, y_stride, yd, yd_stride) + 128) >> 8;
@@ -154,23 +168,18 @@
sad = (vp9_sad64x64(y, y_stride, yd, yd_stride) + 2048) >> 12;
}
- if (bs == BLOCK_16X16) {
- sad_thr = 8;
- } else if (bs == BLOCK_32X32) {
- sad_thr = 7;
- } else { // BLOCK_64X64
- sad_thr = 6;
- }
-
- // TODO(jackychen): More experiments and remove magic numbers.
// vdiff > sad * 3 means vdiff should not be too small, otherwise,
// it might be a lighting change in smooth area. When there is a
// lighting change in smooth area, it is dangerous to do MFQE.
- if (sad > 1 && sad < sad_thr && vdiff > sad * 3 && vdiff < 150) {
- int weight = ((float)sad / (sad_thr - 1)) * ((float)vdiff / (150 - 1)) *
- (1 << MFQE_PRECISION);
- apply_ifactor(y, y_stride, yd, yd_stride, u, v, uv_stride,
- ud, vd, uvd_stride, bs, weight);
+ if (sad > 1 && vdiff > sad * 3) {
+ const int weight = 1 << MFQE_PRECISION;
+ int ifactor = weight * sad * vdiff / (sad_thr * vdiff_thr);
+ // When ifactor equals weight, no MFQE is done.
+ if (ifactor > weight) {
+ ifactor = weight;
+ }
+ apply_ifactor(y, y_stride, yd, yd_stride, u, v, uv_stride, ud, vd,
+ uvd_stride, bs, ifactor);
} else {
// Copy the block from current frame (i.e., no mfqe is done).
copy_block(y, u, v, y_stride, uv_stride, yd, ud, vd,
@@ -199,8 +208,7 @@
int yd_stride, int uvd_stride) {
int mi_offset, y_offset, uv_offset;
const BLOCK_SIZE cur_bs = mi->mbmi.sb_type;
- // TODO(jackychen): Consider how and whether to use qdiff in MFQE.
- // int qdiff = cm->base_qindex - cm->postproc_state.last_base_qindex;
+ const int qdiff = cm->base_qindex - cm->postproc_state.last_base_qindex;
const int bsl = b_width_log2_lookup[bs];
PARTITION_TYPE partition = partition_lookup[bsl][cur_bs];
const BLOCK_SIZE subsize = get_subsize(bs, partition);
@@ -235,18 +243,18 @@
if (mfqe_decision(mi, mfqe_bs)) {
// Do mfqe on the first square partition.
mfqe_block(bs_tmp, y, u, v, y_stride, uv_stride,
- yd, ud, vd, yd_stride, uvd_stride);
+ yd, ud, vd, yd_stride, uvd_stride, qdiff);
// Do mfqe on the second square partition.
mfqe_block(bs_tmp, y + y_offset, u + uv_offset, v + uv_offset,
y_stride, uv_stride, yd + y_offset, ud + uv_offset,
- vd + uv_offset, yd_stride, uvd_stride);
+ vd + uv_offset, yd_stride, uvd_stride, qdiff);
}
if (mfqe_decision(mi + mi_offset * cm->mi_stride, mfqe_bs)) {
// Do mfqe on the first square partition.
mfqe_block(bs_tmp, y + y_offset * y_stride, u + uv_offset * uv_stride,
v + uv_offset * uv_stride, y_stride, uv_stride,
yd + y_offset * yd_stride, ud + uv_offset * uvd_stride,
- vd + uv_offset * uvd_stride, yd_stride, uvd_stride);
+ vd + uv_offset * uvd_stride, yd_stride, uvd_stride, qdiff);
// Do mfqe on the second square partition.
mfqe_block(bs_tmp, y + y_offset * y_stride + y_offset,
u + uv_offset * uv_stride + uv_offset,
@@ -254,7 +262,7 @@
uv_stride, yd + y_offset * yd_stride + y_offset,
ud + uv_offset * uvd_stride + uv_offset,
vd + uv_offset * uvd_stride + uv_offset,
- yd_stride, uvd_stride);
+ yd_stride, uvd_stride, qdiff);
}
break;
case PARTITION_VERT:
@@ -268,18 +276,18 @@
if (mfqe_decision(mi, mfqe_bs)) {
// Do mfqe on the first square partition.
mfqe_block(bs_tmp, y, u, v, y_stride, uv_stride,
- yd, ud, vd, yd_stride, uvd_stride);
+ yd, ud, vd, yd_stride, uvd_stride, qdiff);
// Do mfqe on the second square partition.
mfqe_block(bs_tmp, y + y_offset * y_stride, u + uv_offset * uv_stride,
v + uv_offset * uv_stride, y_stride, uv_stride,
yd + y_offset * yd_stride, ud + uv_offset * uvd_stride,
- vd + uv_offset * uvd_stride, yd_stride, uvd_stride);
+ vd + uv_offset * uvd_stride, yd_stride, uvd_stride, qdiff);
}
if (mfqe_decision(mi + mi_offset, mfqe_bs)) {
// Do mfqe on the first square partition.
mfqe_block(bs_tmp, y + y_offset, u + uv_offset, v + uv_offset,
y_stride, uv_stride, yd + y_offset, ud + uv_offset,
- vd + uv_offset, yd_stride, uvd_stride);
+ vd + uv_offset, yd_stride, uvd_stride, qdiff);
// Do mfqe on the second square partition.
mfqe_block(bs_tmp, y + y_offset * y_stride + y_offset,
u + uv_offset * uv_stride + uv_offset,
@@ -287,14 +295,14 @@
uv_stride, yd + y_offset * yd_stride + y_offset,
ud + uv_offset * uvd_stride + uv_offset,
vd + uv_offset * uvd_stride + uv_offset,
- yd_stride, uvd_stride);
+ yd_stride, uvd_stride, qdiff);
}
break;
case PARTITION_NONE:
if (mfqe_decision(mi, cur_bs)) {
// Do mfqe on this partition.
mfqe_block(cur_bs, y, u, v, y_stride, uv_stride,
- yd, ud, vd, yd_stride, uvd_stride);
+ yd, ud, vd, yd_stride, uvd_stride, qdiff);
} else {
// Copy the block from current frame(i.e., no mfqe is done).
copy_block(y, u, v, y_stride, uv_stride, yd, ud, vd,
diff --git a/vp9/common/vp9_tile_common.c b/vp9/common/vp9_tile_common.c
index 8c4a303..7a20e0a 100644
--- a/vp9/common/vp9_tile_common.c
+++ b/vp9/common/vp9_tile_common.c
@@ -36,24 +36,24 @@
vp9_tile_set_col(tile, cm, col);
}
+static int get_min_log2_tile_cols(const int sb64_cols) {
+ int min_log2 = 0;
+ while ((MAX_TILE_WIDTH_B64 << min_log2) < sb64_cols)
+ ++min_log2;
+ return min_log2;
+}
+
+static int get_max_log2_tile_cols(const int sb64_cols) {
+ int max_log2 = 1;
+ while ((sb64_cols >> max_log2) >= MIN_TILE_WIDTH_B64)
+ ++max_log2;
+ return max_log2 - 1;
+}
+
void vp9_get_tile_n_bits(int mi_cols,
int *min_log2_tile_cols, int *max_log2_tile_cols) {
- const int sb_cols = mi_cols_aligned_to_sb(mi_cols) >> MI_BLOCK_SIZE_LOG2;
- int min_log2 = 0, max_log2 = 0;
-
- // max
- while ((sb_cols >> max_log2) >= MIN_TILE_WIDTH_B64)
- ++max_log2;
- --max_log2;
- if (max_log2 < 0)
- max_log2 = 0;
-
- // min
- while ((MAX_TILE_WIDTH_B64 << min_log2) < sb_cols)
- ++min_log2;
-
- assert(min_log2 <= max_log2);
-
- *min_log2_tile_cols = min_log2;
- *max_log2_tile_cols = max_log2;
+ const int sb64_cols = mi_cols_aligned_to_sb(mi_cols) >> MI_BLOCK_SIZE_LOG2;
+ *min_log2_tile_cols = get_min_log2_tile_cols(sb64_cols);
+ *max_log2_tile_cols = get_max_log2_tile_cols(sb64_cols);
+ assert(*min_log2_tile_cols <= *max_log2_tile_cols);
}
diff --git a/vp9/decoder/vp9_decodeframe.c b/vp9/decoder/vp9_decodeframe.c
index 9677173..3b8af84 100644
--- a/vp9/decoder/vp9_decodeframe.c
+++ b/vp9/decoder/vp9_decodeframe.c
@@ -1546,8 +1546,6 @@
vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
"Truncated packet or corrupt header length");
- init_macroblockd(cm, &pbi->mb);
-
cm->use_prev_frame_mvs = !cm->error_resilient_mode &&
cm->width == cm->last_width &&
cm->height == cm->last_height &&
diff --git a/vp9/encoder/vp9_dct.c b/vp9/encoder/vp9_dct.c
index 506f6de..41f72f8 100644
--- a/vp9/encoder/vp9_dct.c
+++ b/vp9/encoder/vp9_dct.c
@@ -170,7 +170,6 @@
vp9_fdct4x4_c(input, output, stride);
} else {
tran_low_t out[4 * 4];
- tran_low_t *outptr = &out[0];
int i, j;
tran_low_t temp_in[4], temp_out[4];
const transform_2d ht = FHT_4[tx_type];
@@ -183,7 +182,7 @@
temp_in[0] += 1;
ht.cols(temp_in, temp_out);
for (j = 0; j < 4; ++j)
- outptr[j * 4 + i] = temp_out[j];
+ out[j * 4 + i] = temp_out[j];
}
// Rows
@@ -711,7 +710,6 @@
vp9_fdct8x8_c(input, output, stride);
} else {
tran_low_t out[64];
- tran_low_t *outptr = &out[0];
int i, j;
tran_low_t temp_in[8], temp_out[8];
const transform_2d ht = FHT_8[tx_type];
@@ -722,7 +720,7 @@
temp_in[j] = input[j * stride + i] * 4;
ht.cols(temp_in, temp_out);
for (j = 0; j < 8; ++j)
- outptr[j * 8 + i] = temp_out[j];
+ out[j * 8 + i] = temp_out[j];
}
// Rows
@@ -1103,7 +1101,6 @@
vp9_fdct16x16_c(input, output, stride);
} else {
tran_low_t out[256];
- tran_low_t *outptr = &out[0];
int i, j;
tran_low_t temp_in[16], temp_out[16];
const transform_2d ht = FHT_16[tx_type];
@@ -1114,7 +1111,7 @@
temp_in[j] = input[j * stride + i] * 4;
ht.cols(temp_in, temp_out);
for (j = 0; j < 16; ++j)
- outptr[j * 16 + i] = (temp_out[j] + 1 + (temp_out[j] < 0)) >> 2;
+ out[j * 16 + i] = (temp_out[j] + 1 + (temp_out[j] < 0)) >> 2;
}
// Rows
diff --git a/vp9/encoder/vp9_denoiser.c b/vp9/encoder/vp9_denoiser.c
index dec340d..febad49 100644
--- a/vp9/encoder/vp9_denoiser.c
+++ b/vp9/encoder/vp9_denoiser.c
@@ -405,7 +405,10 @@
void vp9_denoiser_reset_frame_stats(PICK_MODE_CONTEXT *ctx) {
ctx->zeromv_sse = UINT_MAX;
- ctx->newmv_sse = UINT_MAX;
+ // This should be initialized as zero since mode search stage might skip
+ // NEWMV mode if inferred motion vector modes provide sufficiently good
+ // prediction quality.
+ ctx->newmv_sse = 0;
}
void vp9_denoiser_update_frame_stats(MB_MODE_INFO *mbmi, unsigned int sse,
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 628660b..230df1f 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -1780,9 +1780,7 @@
}
#if CONFIG_VP9_TEMPORAL_DENOISING
- if (cpi->oxcf.noise_sensitivity > 0) {
- vp9_denoiser_free(&(cpi->denoiser));
- }
+ vp9_denoiser_free(&(cpi->denoiser));
#endif
for (t = 0; t < cpi->num_workers; ++t) {
diff --git a/vp9/encoder/vp9_pickmode.c b/vp9/encoder/vp9_pickmode.c
index 60c7160..5acfcc5 100644
--- a/vp9/encoder/vp9_pickmode.c
+++ b/vp9/encoder/vp9_pickmode.c
@@ -605,6 +605,10 @@
mbmi->interp_filter = cm->interp_filter == SWITCHABLE ?
EIGHTTAP : cm->interp_filter;
+#if CONFIG_VP9_TEMPORAL_DENOISING
+ vp9_denoiser_reset_frame_stats(ctx);
+#endif
+
for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
x->pred_mv_sad[ref_frame] = INT_MAX;
frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
