Revise precision clamping in GM param refinement
This ensures that the parameter refinement never
results in a motion parameter value that exceeds the number
of alloted bits in the bitstream. It accounts for all of
the necessary precision shifts required to make global motion compatible
with the warped motion library. It also accounts for the
zero-centering that is applied to global motion parameters that are
naturally centered around one.
Change-Id: If429c93bb90b66fdff0edc07ecd9fc078077d303
diff --git a/av1/encoder/encodeframe.c b/av1/encoder/encodeframe.c
index 3733efc..522d6bb 100644
--- a/av1/encoder/encodeframe.c
+++ b/av1/encoder/encodeframe.c
@@ -4484,10 +4484,35 @@
#define MIN_TRANS_THRESH 8
#define GLOBAL_MOTION_ADVANTAGE_THRESH 0.60
#define GLOBAL_MOTION_MODEL ROTZOOM
-// TODO(sarahparker) This function needs to be adjusted
-// to accomodate changes in the paraemter integerization.
-// Commenting it out until the fix is made.
-/*
+
+// Adds some offset to a global motion parameter and handles
+// all of the necessary precision shifts, clamping, and
+// zero-centering.
+static int16_t add_param_offset(int param_index, int16_t param_value,
+ int16_t offset) {
+ int scale_vals[2] = { GM_ALPHA_PREC_DIFF, GM_TRANS_PREC_DIFF };
+ int clamp_vals[2] = { GM_ALPHA_MAX, GM_TRANS_MAX };
+ int is_trans_param = param_index < 2;
+ int is_one_centered = (!is_trans_param) && (param_index & 1);
+
+ // Make parameter zero-centered and offset the shift that was done to make
+ // it compatible with the warped model
+ param_value = (param_value - (is_one_centered << WARPEDMODEL_PREC_BITS)) >>
+ scale_vals[is_trans_param];
+ // Add desired offset to the rescaled/zero-centered parameter
+ param_value += offset;
+ // Clamp the parameter so it does not overflow the number of bits allotted
+ // to it in the bitstream
+ param_value = (int16_t)clamp(param_value, -clamp_vals[is_trans_param],
+ clamp_vals[is_trans_param]);
+ // Rescale the parameter to WARPEDMODEL_PRECIION_BITS so it is compatible
+ // with the warped motion library
+ param_value *= (1 << scale_vals[is_trans_param]);
+
+ // Undo the zero-centring step if necessary
+ return param_value + (is_one_centered << WARPEDMODEL_PREC_BITS);
+}
+
static void refine_integerized_param(WarpedMotionParams *wm,
#if CONFIG_AOM_HIGHBITDEPTH
int use_hbd, int bd,
@@ -4500,7 +4525,7 @@
int n_params = n_trans_model_params[wm->wmtype];
int16_t *param_mat = (int16_t *)wm->wmmat;
double step_error;
- int step;
+ int16_t step;
int16_t *param;
int16_t curr_param;
int16_t best_param;
@@ -4519,9 +4544,7 @@
best_param = curr_param;
for (i = 0; i < n_refinements; i++) {
// look to the left
- *param =
- (int16_t)clamp(curr_param - step, p < 2 ? GM_TRANS_MIN : GM_ALPHA_MIN,
- p < 2 ? GM_TRANS_MAX : GM_ALPHA_MAX);
+ *param = add_param_offset(p, curr_param, -step);
step_error =
av1_warp_erroradv(wm,
#if CONFIG_AOM_HIGHBITDEPTH
@@ -4538,9 +4561,7 @@
}
// look to the right
- *param =
- (int16_t)clamp(curr_param + step, p < 2 ? GM_TRANS_MIN : GM_ALPHA_MIN,
- p < 2 ? GM_TRANS_MAX : GM_ALPHA_MAX);
+ *param = add_param_offset(p, curr_param, step);
step_error =
av1_warp_erroradv(wm,
#if CONFIG_AOM_HIGHBITDEPTH
@@ -4564,7 +4585,6 @@
*param = best_param;
}
}
-*/
static void convert_to_params(const double *params, TransformationType type,
int16_t *model) {
@@ -4643,6 +4663,14 @@
convert_model_to_params(params, GLOBAL_MOTION_MODEL,
&cm->global_motion[frame]);
if (get_gmtype(&cm->global_motion[frame]) > GLOBAL_ZERO) {
+ refine_integerized_param(
+ &cm->global_motion[frame].motion_params,
+#if CONFIG_AOM_HIGHBITDEPTH
+ xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH, xd->bd,
+#endif // CONFIG_AOM_HIGHBITDEPTH
+ ref_buf->y_buffer, ref_buf->y_width, ref_buf->y_height,
+ ref_buf->y_stride, cpi->Source->y_buffer, cpi->Source->y_width,
+ cpi->Source->y_height, cpi->Source->y_stride, 3);
// compute the advantage of using gm parameters over 0 motion
erroradvantage = av1_warp_erroradv(
&cm->global_motion[frame].motion_params,
