aom_convolve.c: extract functions 'scalar_product'
And make variable consts when possible.
Change-Id: I823e56327e338ba669a0edd68c6c6d67077ebb7e
diff --git a/aom_dsp/aom_convolve.c b/aom_dsp/aom_convolve.c
index b6aac45..8621a93 100644
--- a/aom_dsp/aom_convolve.c
+++ b/aom_dsp/aom_convolve.c
@@ -20,6 +20,19 @@
#include "aom_dsp/aom_filter.h"
#include "aom_ports/mem.h"
+static INLINE int horz_scalar_product(const uint8_t *a, const int16_t *b) {
+ int sum = 0;
+ for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k] * b[k];
+ return sum;
+}
+
+static INLINE int vert_scalar_product(const uint8_t *a, ptrdiff_t a_stride,
+ const int16_t *b) {
+ int sum = 0;
+ for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k * a_stride] * b[k];
+ return sum;
+}
+
static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
uint8_t *dst, ptrdiff_t dst_stride,
const InterpKernel *x_filters, int x0_q4,
@@ -30,8 +43,7 @@
for (int x = 0; x < w; ++x) {
const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+ const int sum = horz_scalar_product(src_x, x_filter);
dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
x_q4 += x_step_q4;
}
@@ -52,8 +64,7 @@
const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
assert(x_filter_idx < SUBPEL_SHIFTS);
const int16_t *const x_filter = x_filters[x_filter_idx];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+ const int sum = horz_scalar_product(src_x, x_filter);
dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
x_qn += x_step_qn;
}
@@ -72,8 +83,7 @@
for (int x = 0; x < w; ++x) {
const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+ const int sum = horz_scalar_product(src_x, x_filter);
dst[x] = ROUND_POWER_OF_TWO(
dst[x] + clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
x_q4 += x_step_q4;
@@ -95,8 +105,7 @@
const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
assert(x_filter_idx < SUBPEL_SHIFTS);
const int16_t *const x_filter = x_filters[x_filter_idx];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+ const int sum = horz_scalar_product(src_x, x_filter);
dst[x] = ROUND_POWER_OF_TWO(
dst[x] + clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
x_qn += x_step_qn;
@@ -117,9 +126,7 @@
for (int y = 0; y < h; ++y) {
const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k)
- sum += src_y[k * src_stride] * y_filter[k];
+ const int sum = vert_scalar_product(src_y, src_stride, y_filter);
dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
y_q4 += y_step_q4;
}
@@ -141,9 +148,7 @@
&src[(y_qn >> SCALE_SUBPEL_BITS) * src_stride];
const int16_t *const y_filter =
y_filters[(y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k)
- sum += src_y[k * src_stride] * y_filter[k];
+ const int sum = vert_scalar_product(src_y, src_stride, y_filter);
dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
y_qn += y_step_qn;
}
@@ -163,9 +168,7 @@
for (int y = 0; y < h; ++y) {
const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k)
- sum += src_y[k * src_stride] * y_filter[k];
+ const int sum = vert_scalar_product(src_y, src_stride, y_filter);
dst[y * dst_stride] = ROUND_POWER_OF_TWO(
dst[y * dst_stride] +
clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)),
@@ -190,9 +193,7 @@
&src[(y_qn >> SCALE_SUBPEL_BITS) * src_stride];
const int16_t *const y_filter =
y_filters[(y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k)
- sum += src_y[k * src_stride] * y_filter[k];
+ const int sum = vert_scalar_product(src_y, src_stride, y_filter);
dst[y * dst_stride] = ROUND_POWER_OF_TWO(
dst[y * dst_stride] +
clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)),
@@ -547,6 +548,16 @@
filter_y, y_step_q4, w, h);
}
+#if CONFIG_HIGHBITDEPTH || CONFIG_LOOP_RESTORATION
+static INLINE int highbd_vert_scalar_product(const uint16_t *a,
+ ptrdiff_t a_stride,
+ const int16_t *b) {
+ int sum = 0;
+ for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k * a_stride] * b[k];
+ return sum;
+}
+#endif
+
// TODO(afergs): Make sure this works too
#if CONFIG_LOOP_RESTORATION
static void convolve_add_src_horiz(const uint8_t *src, ptrdiff_t src_stride,
@@ -559,8 +570,8 @@
for (int x = 0; x < w; ++x) {
const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+
+ const int sum = horz_scalar_product(src_x, x_filter);
dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS) +
src_x[SUBPEL_TAPS / 2 - 1]);
x_q4 += x_step_q4;
@@ -581,9 +592,7 @@
for (int y = 0; y < h; ++y) {
const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k)
- sum += src_y[k * src_stride] * y_filter[k];
+ const int sum = vert_scalar_product(src_y, src_stride, y_filter);
dst[y * dst_stride] =
clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS) +
src_y[(SUBPEL_TAPS / 2 - 1) * src_stride]);
@@ -672,9 +681,9 @@
for (int x = 0; x < w; ++x) {
const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
- int sum = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS) +
- (1 << (bd + FILTER_BITS - 1));
- for (int k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+ const int rounding = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS) +
+ (1 << (bd + FILTER_BITS - 1));
+ const int sum = horz_scalar_product(src_x, x_filter) + rounding;
dst[x] =
(uint16_t)clamp(ROUND_POWER_OF_TWO(sum, FILTER_BITS - EXTRAPREC_BITS),
0, EXTRAPREC_CLAMP_LIMIT(bd) - 1);
@@ -697,11 +706,11 @@
for (int y = 0; y < h; ++y) {
const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
- int sum =
+ const int rounding =
((int)src_y[(SUBPEL_TAPS / 2 - 1) * src_stride] << FILTER_BITS) -
(1 << (bd + FILTER_BITS + EXTRAPREC_BITS - 1));
- for (int k = 0; k < SUBPEL_TAPS; ++k)
- sum += src_y[k * src_stride] * y_filter[k];
+ const int sum =
+ highbd_vert_scalar_product(src_y, src_stride, y_filter) + rounding;
dst[y * dst_stride] =
clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS + EXTRAPREC_BITS));
y_q4 += y_step_q4;
@@ -783,6 +792,14 @@
// TODO(afergs): Make sure this works too
#if CONFIG_HIGHBITDEPTH
+
+static INLINE int highbd_horz_scalar_product(const uint16_t *a,
+ const int16_t *b) {
+ int sum = 0;
+ for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k] * b[k];
+ return sum;
+}
+
static void highbd_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride,
uint8_t *dst8, ptrdiff_t dst_stride,
const InterpKernel *x_filters, int x0_q4,
@@ -795,8 +812,7 @@
for (int x = 0; x < w; ++x) {
const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+ const int sum = highbd_horz_scalar_product(src_x, x_filter);
dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
x_q4 += x_step_q4;
}
@@ -817,8 +833,7 @@
for (int x = 0; x < w; ++x) {
const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+ const int sum = highbd_horz_scalar_product(src_x, x_filter);
dst[x] = ROUND_POWER_OF_TWO(
dst[x] + clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
1);
@@ -841,9 +856,7 @@
for (int y = 0; y < h; ++y) {
const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k)
- sum += src_y[k * src_stride] * y_filter[k];
+ const int sum = highbd_vert_scalar_product(src_y, src_stride, y_filter);
dst[y * dst_stride] =
clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
y_q4 += y_step_q4;
@@ -865,9 +878,7 @@
for (int y = 0; y < h; ++y) {
const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k)
- sum += src_y[k * src_stride] * y_filter[k];
+ const int sum = highbd_vert_scalar_product(src_y, src_stride, y_filter);
dst[y * dst_stride] = ROUND_POWER_OF_TWO(
dst[y * dst_stride] +
clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
@@ -1058,8 +1069,7 @@
for (int x = 0; x < w; ++x) {
const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+ const int sum = highbd_horz_scalar_product(src_x, x_filter);
dst[x] = clip_pixel_highbd(
ROUND_POWER_OF_TWO(sum, FILTER_BITS) + src_x[SUBPEL_TAPS / 2 - 1],
bd);
@@ -1084,9 +1094,7 @@
for (int y = 0; y < h; ++y) {
const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
- int sum = 0;
- for (int k = 0; k < SUBPEL_TAPS; ++k)
- sum += src_y[k * src_stride] * y_filter[k];
+ const int sum = highbd_vert_scalar_product(src_y, src_stride, y_filter);
dst[y * dst_stride] =
clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS) +
src_y[(SUBPEL_TAPS / 2 - 1) * src_stride],
@@ -1190,9 +1198,9 @@
for (int x = 0; x < w; ++x) {
const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
- int sum = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS) +
- (1 << (bd + FILTER_BITS - 1));
- for (int k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+ const int rounding = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS) +
+ (1 << (bd + FILTER_BITS - 1));
+ const int sum = highbd_horz_scalar_product(src_x, x_filter) + rounding;
dst[x] =
(uint16_t)clamp(ROUND_POWER_OF_TWO(sum, FILTER_BITS - EXTRAPREC_BITS),
0, extraprec_clamp_limit - 1);
@@ -1214,11 +1222,11 @@
for (int y = 0; y < h; ++y) {
const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
- int sum =
+ const int rounding =
((int)src_y[(SUBPEL_TAPS / 2 - 1) * src_stride] << FILTER_BITS) -
(1 << (bd + FILTER_BITS + EXTRAPREC_BITS - 1));
- for (int k = 0; k < SUBPEL_TAPS; ++k)
- sum += src_y[k * src_stride] * y_filter[k];
+ const int sum =
+ highbd_vert_scalar_product(src_y, src_stride, y_filter) + rounding;
dst[y * dst_stride] = clip_pixel_highbd(
ROUND_POWER_OF_TWO(sum, FILTER_BITS + EXTRAPREC_BITS), bd);
y_q4 += y_step_q4;
