ref: 86bb6df00515ab2263b41e506cdf55d90d76598f
parent: 4852a8023d0e8562fba338f20a064b40304f7858
author: Tero Rintaluoma <[email protected]>
date: Mon Jun 10 06:23:04 EDT 2013
Fixed point reference picture scaling Fixed point scaling factors are calculated once for each reference frame by using integer division. Otherwise fixed point scaling routines are used in all scaling calculations. This makes it possible to calculate fixed point scaling factors on device driver software and pass them to hardware and thus avoid division on hardware. TODO: - Missing check for maximum frame dimensions (currently scaling uses 14 bits) - Missing check for maximum scaling ratio (upscaling 16:1, downscaling 2:1) Problems: - Straightforward fixed point implementation can cause error +-1 compared to integer division (i.e. in x_step_q4). Should only be an issue for frames larger than 16k. Change-Id: I3cf4dabd610a4dc18da3bdb31ae244ebaf5d579c
--- a/vp9/common/vp9_blockd.h
+++ b/vp9/common/vp9_blockd.h
@@ -239,13 +239,12 @@
union b_mode_info bmi[4];
} MODE_INFO;
+#define VP9_REF_SCALE_SHIFT 14
struct scale_factors {- int x_num;
- int x_den;
+ int x_scale_fp; // horizontal fixed point scale factor
+ int y_scale_fp; // vertical fixed point scale factor
int x_offset_q4;
int x_step_q4;
- int y_num;
- int y_den;
int y_offset_q4;
int y_step_q4;
@@ -254,7 +253,7 @@
void (*set_scaled_offsets)(struct scale_factors *scale, int row, int col);
int_mv32 (*scale_mv_q3_to_q4)(const int_mv *src_mv,
const struct scale_factors *scale);
- int32_t (*scale_mv_component_q4)(int mv_q4, int num, int den, int offset_q4);
+ int32_t (*scale_mv_component_q4)(int mv_q4, int scale_fp, int offset_q4);
convolve_fn_t predict[2][2][2]; // horiz, vert, avg
};
--- a/vp9/common/vp9_reconinter.c
+++ b/vp9/common/vp9_reconinter.c
@@ -19,12 +19,12 @@
static int scale_value_x_with_scaling(int val,
const struct scale_factors *scale) {- return val * scale->x_num / scale->x_den;
+ return (val * scale->x_scale_fp >> VP9_REF_SCALE_SHIFT);
}
static int scale_value_y_with_scaling(int val,
const struct scale_factors *scale) {- return val * scale->y_num / scale->y_den;
+ return (val * scale->y_scale_fp >> VP9_REF_SCALE_SHIFT);
}
static int unscaled_value(int val, const struct scale_factors *scale) {@@ -39,10 +39,9 @@
const int32_t mv_row_q4 = src_mv->as_mv.row << 1;
const int32_t mv_col_q4 = src_mv->as_mv.col << 1;
- /* TODO(jkoleszar): make fixed point, or as a second multiply? */
- result.as_mv.row = mv_row_q4 * scale->y_num / scale->y_den
+ result.as_mv.row = (mv_row_q4 * scale->y_scale_fp >> VP9_REF_SCALE_SHIFT)
+ scale->y_offset_q4;
- result.as_mv.col = mv_col_q4 * scale->x_num / scale->x_den
+ result.as_mv.col = (mv_col_q4 * scale->x_scale_fp >> VP9_REF_SCALE_SHIFT)
+ scale->x_offset_q4;
return result;
}
@@ -57,19 +56,19 @@
return result;
}
-static int32_t mv_component_q4_with_scaling(int mv_q4, int num, int den,
+static int32_t mv_component_q4_with_scaling(int mv_q4, int scale_fp,
int offset_q4) {+ int32_t scaled_mv;
// returns the scaled and offset value of the mv component.
+ scaled_mv = (mv_q4 * scale_fp >> VP9_REF_SCALE_SHIFT) + offset_q4;
- /* TODO(jkoleszar): make fixed point, or as a second multiply? */
- return mv_q4 * num / den + offset_q4;
+ return scaled_mv;
}
-static int32_t mv_component_q4_without_scaling(int mv_q4, int num, int den,
+static int32_t mv_component_q4_without_scaling(int mv_q4, int scale_fp,
int offset_q4) {// returns the scaled and offset value of the mv component.
- (void)num;
- (void)den;
+ (void)scale_fp;
(void)offset_q4;
return mv_q4;
}
@@ -79,8 +78,8 @@
const int x_q4 = 16 * col;
const int y_q4 = 16 * row;
- scale->x_offset_q4 = (x_q4 * scale->x_num / scale->x_den) & 0xf;
- scale->y_offset_q4 = (y_q4 * scale->y_num / scale->y_den) & 0xf;
+ scale->x_offset_q4 = (x_q4 * scale->x_scale_fp >> VP9_REF_SCALE_SHIFT) & 0xf;
+ scale->y_offset_q4 = (y_q4 * scale->y_scale_fp >> VP9_REF_SCALE_SHIFT) & 0xf;
}
static void set_offsets_without_scaling(struct scale_factors *scale,
@@ -89,20 +88,26 @@
scale->y_offset_q4 = 0;
}
+static int get_fixed_point_scale_factor(int other_size, int this_size) {+ // Calculate scaling factor once for each reference frame
+ // and use fixed point scaling factors in decoding and encoding routines.
+ // Hardware implementations can calculate scale factor in device driver
+ // and use multiplication and shifting on hardware instead of division.
+ return (other_size << VP9_REF_SCALE_SHIFT) / this_size;
+}
+
void vp9_setup_scale_factors_for_frame(struct scale_factors *scale,
int other_w, int other_h,
int this_w, int this_h) {- scale->x_num = other_w;
- scale->x_den = this_w;
+ scale->x_scale_fp = get_fixed_point_scale_factor(other_w, this_w);
scale->x_offset_q4 = 0; // calculated per-mb
- scale->x_step_q4 = 16 * other_w / this_w;
+ scale->x_step_q4 = (16 * scale->x_scale_fp >> VP9_REF_SCALE_SHIFT);
- scale->y_num = other_h;
- scale->y_den = this_h;
+ scale->y_scale_fp = get_fixed_point_scale_factor(other_h, this_h);
scale->y_offset_q4 = 0; // calculated per-mb
- scale->y_step_q4 = 16 * other_h / this_h;
+ scale->y_step_q4 = (16 * scale->y_scale_fp >> VP9_REF_SCALE_SHIFT);
- if (scale->x_num == scale->x_den && scale->y_num == scale->y_den) {+ if ((other_w == this_w) && (other_h == this_h)) {scale->scale_value_x = unscaled_value;
scale->scale_value_y = unscaled_value;
scale->set_scaled_offsets = set_offsets_without_scaling;
@@ -303,12 +308,10 @@
int w, int h, int weight,
const struct subpix_fn_table *subpix) {const int scaled_mv_row_q4 = scale->scale_mv_component_q4(mv_q4->as_mv.row,
- scale->y_num,
- scale->y_den,
+ scale->y_scale_fp,
scale->y_offset_q4);
const int scaled_mv_col_q4 = scale->scale_mv_component_q4(mv_q4->as_mv.col,
- scale->x_num,
- scale->x_den,
+ scale->x_scale_fp,
scale->x_offset_q4);
const int subpel_x = scaled_mv_col_q4 & 15;
const int subpel_y = scaled_mv_row_q4 & 15;
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -1755,12 +1755,13 @@
// set up scaling factors
scale[frame_type] = cpi->common.active_ref_scale[frame_type - 1];
+
scale[frame_type].x_offset_q4 =
- (mi_col * MI_SIZE * scale[frame_type].x_num /
- scale[frame_type].x_den) & 0xf;
+ ROUND_POWER_OF_TWO(mi_col * MI_SIZE * scale[frame_type].x_scale_fp,
+ VP9_REF_SCALE_SHIFT) & 0xf;
scale[frame_type].y_offset_q4 =
- (mi_row * MI_SIZE * scale[frame_type].y_num /
- scale[frame_type].y_den) & 0xf;
+ ROUND_POWER_OF_TWO(mi_row * MI_SIZE * scale[frame_type].y_scale_fp,
+ VP9_REF_SCALE_SHIFT) & 0xf;
// TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
// use the UV scaling factors.
@@ -1783,8 +1784,8 @@
// Further refinement that is encode side only to test the top few candidates
// in full and choose the best as the centre point for subsequent searches.
// The current implementation doesn't support scaling.
- if (scale[frame_type].x_num == scale[frame_type].x_den &&
- scale[frame_type].y_num == scale[frame_type].y_den)
+ if (scale[frame_type].x_scale_fp == (1 << VP9_REF_SCALE_SHIFT) &&
+ scale[frame_type].y_scale_fp == (1 << VP9_REF_SCALE_SHIFT))
mv_pred(cpi, x, yv12_mb[frame_type][0].buf, yv12->y_stride,
frame_type, block_size);
}
@@ -2612,18 +2613,18 @@
// TODO(jingning, jkoleszar): scaling reference frame not supported for
// SPLITMV.
if (mbmi->ref_frame[0] > 0 &&
- (scale_factor[mbmi->ref_frame[0]].x_num !=
- scale_factor[mbmi->ref_frame[0]].x_den ||
- scale_factor[mbmi->ref_frame[0]].y_num !=
- scale_factor[mbmi->ref_frame[0]].y_den) &&
+ (scale_factor[mbmi->ref_frame[0]].x_scale_fp !=
+ (1 << VP9_REF_SCALE_SHIFT) ||
+ scale_factor[mbmi->ref_frame[0]].y_scale_fp !=
+ (1 << VP9_REF_SCALE_SHIFT)) &&
this_mode == SPLITMV)
continue;
if (mbmi->ref_frame[1] > 0 &&
- (scale_factor[mbmi->ref_frame[1]].x_num !=
- scale_factor[mbmi->ref_frame[1]].x_den ||
- scale_factor[mbmi->ref_frame[1]].y_num !=
- scale_factor[mbmi->ref_frame[1]].y_den) &&
+ (scale_factor[mbmi->ref_frame[1]].x_scale_fp !=
+ (1 << VP9_REF_SCALE_SHIFT) ||
+ scale_factor[mbmi->ref_frame[1]].y_scale_fp !=
+ (1 << VP9_REF_SCALE_SHIFT)) &&
this_mode == SPLITMV)
continue;