ref: 1d13a6c321b3af595adb59f175894ebb9c36c1d6
parent: 13dc6d671ccd88a8af021b1b682400bf16b6c7c6
parent: b1072a793d4957e5a504d125ae5ded808b90060d
author: Sai Deng <[email protected]>
date: Tue Jun 4 12:10:51 EDT 2019
Merge "Hierarchical rdmult scaling when tune=ssim"
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -268,21 +268,38 @@
}
static void set_ssim_rdmult(VP9_COMP *const cpi, MACROBLOCK *const x,
- int mi_row, int mi_col, int *rdmult) {
+ const BLOCK_SIZE bsize, const int mi_row,
+ const int mi_col, int *const rdmult) {
const VP9_COMMON *const cm = &cpi->common;
- // SSIM rdmult scaling factors are currently 64x64 based.
- const int num_8x8_w = 8;
- const int num_8x8_h = 8;
+ const int bsize_base = BLOCK_16X16;
+ const int num_8x8_w = num_8x8_blocks_wide_lookup[bsize_base];
+ const int num_8x8_h = num_8x8_blocks_high_lookup[bsize_base];
const int num_cols = (cm->mi_cols + num_8x8_w - 1) / num_8x8_w;
- const int row = mi_row / num_8x8_h;
- const int col = mi_col / num_8x8_w;
- const int index = row * num_cols + col;
+ const int num_rows = (cm->mi_rows + num_8x8_h - 1) / num_8x8_h;
+ const int num_bcols =
+ (num_8x8_blocks_wide_lookup[bsize] + num_8x8_w - 1) / num_8x8_w;
+ const int num_brows =
+ (num_8x8_blocks_high_lookup[bsize] + num_8x8_h - 1) / num_8x8_h;
+ int row, col;
+ double num_of_mi = 0.0;
+ double geom_mean_of_scale = 0.0;
assert(cpi->oxcf.tuning == VP8_TUNE_SSIM);
- *rdmult =
- (int)((double)(*rdmult) * cpi->mi_ssim_rdmult_scaling_factors[index]);
- *rdmult = VPXMAX(*rdmult, 1);
+
+ for (row = mi_row / num_8x8_w;
+ row < num_rows && row < mi_row / num_8x8_w + num_brows; ++row) {
+ for (col = mi_col / num_8x8_h;
+ col < num_cols && col < mi_col / num_8x8_h + num_bcols; ++col) {
+ const int index = row * num_cols + col;
+ geom_mean_of_scale += log(cpi->mi_ssim_rdmult_scaling_factors[index]);
+ num_of_mi += 1.0;
+ }
+ }
+ geom_mean_of_scale = exp(geom_mean_of_scale / num_of_mi);
+
+ *rdmult = (int)((double)(*rdmult) * geom_mean_of_scale);
+ *rdmult = VPXMAX(*rdmult, 0);
set_error_per_bit(x, *rdmult);
vpx_clear_system_state();
}
@@ -323,7 +340,7 @@
x->rddiv = cpi->rd.RDDIV;
x->rdmult = cpi->rd.RDMULT;
if (oxcf->tuning == VP8_TUNE_SSIM) {
- set_ssim_rdmult(cpi, x, mi_row, mi_col, &x->rdmult);
+ set_ssim_rdmult(cpi, x, bsize, mi_row, mi_col, &x->rdmult);
}
// required by vp9_append_sub8x8_mvs_for_idx() and vp9_find_best_ref_mvs()
@@ -1970,7 +1987,7 @@
}
if (oxcf->tuning == VP8_TUNE_SSIM) {
- set_ssim_rdmult(cpi, x, mi_row, mi_col, &x->rdmult);
+ set_ssim_rdmult(cpi, x, bsize, mi_row, mi_col, &x->rdmult);
}
}
@@ -2221,7 +2238,7 @@
const VP9EncoderConfig *const oxcf = &cpi->oxcf;
x->rdmult = x->cb_rdmult;
if (oxcf->tuning == VP8_TUNE_SSIM) {
- set_ssim_rdmult(cpi, x, mi_row, mi_col, &x->rdmult);
+ set_ssim_rdmult(cpi, x, bsize, mi_row, mi_col, &x->rdmult);
}
}
@@ -3845,7 +3862,7 @@
const int rd_div = x->rddiv;
int partition_mul = x->cb_rdmult;
if (oxcf->tuning == VP8_TUNE_SSIM) {
- set_ssim_rdmult(cpi, x, mi_row, mi_col, &partition_mul);
+ set_ssim_rdmult(cpi, x, bsize, mi_row, mi_col, &partition_mul);
}
vp9_rd_cost_update(partition_mul, rd_div, &best_rdc);
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -2397,7 +2397,7 @@
vp9_set_speed_features_framesize_dependent(cpi, oxcf->speed);
{
- const int bsize = BLOCK_64X64;
+ const int bsize = BLOCK_16X16;
const int w = num_8x8_blocks_wide_lookup[bsize];
const int h = num_8x8_blocks_high_lookup[bsize];
const int num_cols = (cm->mi_cols + w - 1) / w;
@@ -4759,15 +4759,9 @@
// Some sample values are:
// (for midres test set)
// SSIM_VAR_SCALE avg_psnr ssim ms_ssim
-// 16.0 2.312 -3.062 -3.882
-// 32.0 0.852 -2.260 -2.821
-// 64.0 0.294 -1.606 -1.925
-// (for midres_10bd test set)
-// SSIM_VAR_SCALE avg_psnr ssim ms_ssim
-// 8.0 6.782 -3.872 -5.464
-// 16.0 3.189 -4.083 -5.258
-// 32.0 1.113 -3.423 -4.309
-// 64.0 0.241 -2.515 -3.074
+// 8.0 8.980 -5.767 -7.069
+// 16.0 4.315 -5.610 -6.411
+// 32.0 1.559 -4.541 -4.980
#define SSIM_VAR_SCALE 16.0
static void set_mb_ssim_rdmult_scaling(VP9_COMP *cpi) {
VP9_COMMON *cm = &cpi->common;
@@ -4776,7 +4770,7 @@
MACROBLOCKD *xd = &x->e_mbd;
uint8_t *y_buffer = cpi->Source->y_buffer;
const int y_stride = cpi->Source->y_stride;
- const int block_size = BLOCK_64X64;
+ const int block_size = BLOCK_16X16;
const int num_8x8_w = num_8x8_blocks_wide_lookup[block_size];
const int num_8x8_h = num_8x8_blocks_high_lookup[block_size];