summaryrefslogtreecommitdiffstats
path: root/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c
diff options
context:
space:
mode:
Diffstat (limited to 'third_party/aom/aom_dsp/arm/aom_convolve8_neon.c')
-rw-r--r--third_party/aom/aom_dsp/arm/aom_convolve8_neon.c401
1 files changed, 263 insertions, 138 deletions
diff --git a/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c b/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c
index 7441108b01..6a177b2e6b 100644
--- a/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c
+++ b/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c
@@ -20,6 +20,7 @@
#include "aom/aom_integer.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/arm/aom_filter.h"
#include "aom_dsp/arm/mem_neon.h"
#include "aom_dsp/arm/transpose_neon.h"
#include "aom_ports/mem.h"
@@ -31,14 +32,14 @@ static INLINE int16x4_t convolve8_4(const int16x4_t s0, const int16x4_t s1,
const int16x8_t filter) {
const int16x4_t filter_lo = vget_low_s16(filter);
const int16x4_t filter_hi = vget_high_s16(filter);
- int16x4_t sum;
- sum = vmul_lane_s16(s0, filter_lo, 0);
+ int16x4_t sum = vmul_lane_s16(s0, filter_lo, 0);
sum = vmla_lane_s16(sum, s1, filter_lo, 1);
sum = vmla_lane_s16(sum, s2, filter_lo, 2);
sum = vmla_lane_s16(sum, s5, filter_hi, 1);
sum = vmla_lane_s16(sum, s6, filter_hi, 2);
sum = vmla_lane_s16(sum, s7, filter_hi, 3);
+
sum = vqadd_s16(sum, vmul_lane_s16(s3, filter_lo, 3));
sum = vqadd_s16(sum, vmul_lane_s16(s4, filter_hi, 0));
return sum;
@@ -51,65 +52,56 @@ static INLINE uint8x8_t convolve8_8(const int16x8_t s0, const int16x8_t s1,
const int16x8_t filter) {
const int16x4_t filter_lo = vget_low_s16(filter);
const int16x4_t filter_hi = vget_high_s16(filter);
- int16x8_t sum;
- sum = vmulq_lane_s16(s0, filter_lo, 0);
+ int16x8_t sum = vmulq_lane_s16(s0, filter_lo, 0);
sum = vmlaq_lane_s16(sum, s1, filter_lo, 1);
sum = vmlaq_lane_s16(sum, s2, filter_lo, 2);
sum = vmlaq_lane_s16(sum, s5, filter_hi, 1);
sum = vmlaq_lane_s16(sum, s6, filter_hi, 2);
sum = vmlaq_lane_s16(sum, s7, filter_hi, 3);
+
sum = vqaddq_s16(sum, vmulq_lane_s16(s3, filter_lo, 3));
sum = vqaddq_s16(sum, vmulq_lane_s16(s4, filter_hi, 0));
return vqrshrun_n_s16(sum, FILTER_BITS);
}
-void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
- uint8_t *dst, ptrdiff_t dst_stride,
- const int16_t *filter_x, int x_step_q4,
- const int16_t *filter_y, int y_step_q4, int w,
- int h) {
+static INLINE void convolve8_horiz_8tap_neon(const uint8_t *src,
+ ptrdiff_t src_stride, uint8_t *dst,
+ ptrdiff_t dst_stride,
+ const int16_t *filter_x, int w,
+ int h) {
const int16x8_t filter = vld1q_s16(filter_x);
- assert((intptr_t)dst % 4 == 0);
- assert(dst_stride % 4 == 0);
-
- (void)x_step_q4;
- (void)filter_y;
- (void)y_step_q4;
-
- src -= ((SUBPEL_TAPS / 2) - 1);
-
if (h == 4) {
- uint8x8_t t0, t1, t2, t3, d01, d23;
- int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3;
-
+ uint8x8_t t0, t1, t2, t3;
load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
transpose_elems_inplace_u8_8x4(&t0, &t1, &t2, &t3);
- s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
- s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
- s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
- s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
- s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
- s5 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
- s6 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+
+ int16x4_t s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ int16x4_t s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ int16x4_t s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+ int16x4_t s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
+ int16x4_t s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ int16x4_t s5 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ int16x4_t s6 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
src += 7;
do {
load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
transpose_elems_inplace_u8_8x4(&t0, &t1, &t2, &t3);
- s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
- s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
- s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
- s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
-
- d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filter);
- d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filter);
- d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filter);
- d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filter);
- d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
- d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
+
+ int16x4_t s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ int16x4_t s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ int16x4_t s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+ int16x4_t s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
+
+ int16x4_t d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filter);
+ int16x4_t d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filter);
+ int16x4_t d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filter);
+ int16x4_t d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filter);
+ uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
+ uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
transpose_elems_inplace_u8_4x4(&d01, &d23);
@@ -123,39 +115,40 @@ void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
s4 = s8;
s5 = s9;
s6 = s10;
+
src += 4;
dst += 4;
w -= 4;
} while (w != 0);
} else {
- uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7, d0, d1, d2, d3;
- int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
-
if (w == 4) {
do {
+ uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7;
load_u8_8x8(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
transpose_elems_inplace_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
- s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
- s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
- s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
- s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
- s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
- s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
- s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+ int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ int16x8_t s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
load_u8_8x8(src + 7, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6,
&t7);
transpose_elems_u8_4x8(t0, t1, t2, t3, t4, t5, t6, t7, &t0, &t1, &t2,
&t3);
- s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
- s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
- s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
- s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
- d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter);
- d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter);
- d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter);
- d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter);
+ int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ int16x8_t s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ int16x8_t s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
+
+ uint8x8_t d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter);
+ uint8x8_t d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter);
+ uint8x8_t d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter);
+ uint8x8_t d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter);
transpose_elems_inplace_u8_8x4(&d0, &d1, &d2, &d3);
@@ -169,48 +162,49 @@ void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
h -= 8;
} while (h > 0);
} else {
- uint8x8_t d4, d5, d6, d7;
- int16x8_t s11, s12, s13, s14;
- int width;
- const uint8_t *s;
- uint8_t *d;
-
do {
- load_u8_8x8(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ int width = w;
+ const uint8_t *s = src;
+ uint8_t *d = dst;
+
+ uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7;
+ load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
transpose_elems_inplace_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
- s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
- s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
- s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
- s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
- s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
- s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
- s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
-
- width = w;
- s = src + 7;
- d = dst;
+
+ int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ int16x8_t s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+ s += 7;
do {
load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
transpose_elems_inplace_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6,
&t7);
- s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
- s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
- s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
- s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
- s11 = vreinterpretq_s16_u16(vmovl_u8(t4));
- s12 = vreinterpretq_s16_u16(vmovl_u8(t5));
- s13 = vreinterpretq_s16_u16(vmovl_u8(t6));
- s14 = vreinterpretq_s16_u16(vmovl_u8(t7));
-
- d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter);
- d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter);
- d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter);
- d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter);
- d4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filter);
- d5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filter);
- d6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filter);
- d7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filter);
+
+ int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ int16x8_t s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ int16x8_t s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ int16x8_t s11 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ int16x8_t s12 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ int16x8_t s13 = vreinterpretq_s16_u16(vmovl_u8(t6));
+ int16x8_t s14 = vreinterpretq_s16_u16(vmovl_u8(t7));
+
+ uint8x8_t d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter);
+ uint8x8_t d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter);
+ uint8x8_t d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter);
+ uint8x8_t d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter);
+ uint8x8_t d4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filter);
+ uint8x8_t d5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filter);
+ uint8x8_t d6 =
+ convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filter);
+ uint8x8_t d7 =
+ convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filter);
transpose_elems_inplace_u8_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6,
&d7);
@@ -224,6 +218,7 @@ void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
s4 = s12;
s5 = s13;
s6 = s14;
+
s += 8;
d += 8;
width -= 8;
@@ -236,6 +231,137 @@ void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
}
}
+static INLINE int16x4_t convolve4_4(const int16x4_t s0, const int16x4_t s1,
+ const int16x4_t s2, const int16x4_t s3,
+ const int16x4_t filter) {
+ int16x4_t sum = vmul_lane_s16(s0, filter, 0);
+ sum = vmla_lane_s16(sum, s1, filter, 1);
+ sum = vmla_lane_s16(sum, s2, filter, 2);
+ sum = vmla_lane_s16(sum, s3, filter, 3);
+
+ return sum;
+}
+
+static INLINE uint8x8_t convolve4_8(const int16x8_t s0, const int16x8_t s1,
+ const int16x8_t s2, const int16x8_t s3,
+ const int16x4_t filter) {
+ int16x8_t sum = vmulq_lane_s16(s0, filter, 0);
+ sum = vmlaq_lane_s16(sum, s1, filter, 1);
+ sum = vmlaq_lane_s16(sum, s2, filter, 2);
+ sum = vmlaq_lane_s16(sum, s3, filter, 3);
+
+ // We halved the filter values so -1 from right shift.
+ return vqrshrun_n_s16(sum, FILTER_BITS - 1);
+}
+
+static INLINE void convolve8_horiz_4tap_neon(const uint8_t *src,
+ ptrdiff_t src_stride, uint8_t *dst,
+ ptrdiff_t dst_stride,
+ const int16_t *filter_x, int w,
+ int h) {
+ // All filter values are even, halve to reduce intermediate precision
+ // requirements.
+ const int16x4_t filter = vshr_n_s16(vld1_s16(filter_x + 2), 1);
+
+ if (w == 4) {
+ do {
+ int16x8_t t0 =
+ vreinterpretq_s16_u16(vmovl_u8(vld1_u8(src + 0 * src_stride)));
+ int16x8_t t1 =
+ vreinterpretq_s16_u16(vmovl_u8(vld1_u8(src + 1 * src_stride)));
+
+ int16x4_t s0[4], s1[4];
+ s0[0] = vget_low_s16(t0);
+ s0[1] = vget_low_s16(vextq_s16(t0, t0, 1));
+ s0[2] = vget_low_s16(vextq_s16(t0, t0, 2));
+ s0[3] = vget_low_s16(vextq_s16(t0, t0, 3));
+
+ s1[0] = vget_low_s16(t1);
+ s1[1] = vget_low_s16(vextq_s16(t1, t1, 1));
+ s1[2] = vget_low_s16(vextq_s16(t1, t1, 2));
+ s1[3] = vget_low_s16(vextq_s16(t1, t1, 3));
+
+ int16x4_t d0 = convolve4_4(s0[0], s0[1], s0[2], s0[3], filter);
+ int16x4_t d1 = convolve4_4(s1[0], s1[1], s1[2], s1[3], filter);
+ // We halved the filter values so -1 from right shift.
+ uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS - 1);
+
+ store_u8x4_strided_x2(dst + 0 * dst_stride, dst_stride, d01);
+
+ src += 2 * src_stride;
+ dst += 2 * dst_stride;
+ h -= 2;
+ } while (h > 0);
+ } else {
+ do {
+ int width = w;
+ const uint8_t *s = src;
+ uint8_t *d = dst;
+
+ int16x8_t t0 =
+ vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s + 0 * src_stride)));
+ int16x8_t t1 =
+ vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s + 1 * src_stride)));
+
+ s += 8;
+ do {
+ int16x8_t t2 =
+ vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s + 0 * src_stride)));
+ int16x8_t t3 =
+ vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s + 1 * src_stride)));
+
+ int16x8_t s0[4], s1[4];
+ s0[0] = t0;
+ s0[1] = vextq_s16(t0, t2, 1);
+ s0[2] = vextq_s16(t0, t2, 2);
+ s0[3] = vextq_s16(t0, t2, 3);
+
+ s1[0] = t1;
+ s1[1] = vextq_s16(t1, t3, 1);
+ s1[2] = vextq_s16(t1, t3, 2);
+ s1[3] = vextq_s16(t1, t3, 3);
+
+ uint8x8_t d0 = convolve4_8(s0[0], s0[1], s0[2], s0[3], filter);
+ uint8x8_t d1 = convolve4_8(s1[0], s1[1], s1[2], s1[3], filter);
+
+ store_u8_8x2(d, dst_stride, d0, d1);
+
+ t0 = t2;
+ t1 = t3;
+
+ s += 8;
+ d += 8;
+ width -= 8;
+ } while (width != 0);
+ src += 2 * src_stride;
+ dst += 2 * dst_stride;
+ h -= 2;
+ } while (h > 0);
+ }
+}
+
+void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int x_step_q4,
+ const int16_t *filter_y, int y_step_q4, int w,
+ int h) {
+ assert((intptr_t)dst % 4 == 0);
+ assert(dst_stride % 4 == 0);
+
+ (void)x_step_q4;
+ (void)filter_y;
+ (void)y_step_q4;
+
+ src -= ((SUBPEL_TAPS / 2) - 1);
+
+ if (get_filter_taps_convolve8(filter_x) <= 4) {
+ convolve8_horiz_4tap_neon(src + 2, src_stride, dst, dst_stride, filter_x, w,
+ h);
+ } else {
+ convolve8_horiz_8tap_neon(src, src_stride, dst, dst_stride, filter_x, w, h);
+ }
+}
+
void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
uint8_t *dst, ptrdiff_t dst_stride,
const int16_t *filter_x, int x_step_q4,
@@ -253,33 +379,33 @@ void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
src -= ((SUBPEL_TAPS / 2) - 1) * src_stride;
if (w == 4) {
- uint8x8_t t0, t1, t2, t3, t4, t5, t6, d01, d23;
- int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3;
-
+ uint8x8_t t0, t1, t2, t3, t4, t5, t6;
load_u8_8x7(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6);
- s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
- s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
- s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
- s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
- s4 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t4)));
- s5 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t5)));
- s6 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t6)));
+
+ int16x4_t s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ int16x4_t s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ int16x4_t s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+ int16x4_t s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
+ int16x4_t s4 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t4)));
+ int16x4_t s5 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t5)));
+ int16x4_t s6 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t6)));
src += 7 * src_stride;
do {
load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
- s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
- s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
- s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
- s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
-
- d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filter);
- d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filter);
- d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filter);
- d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filter);
- d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
- d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
+
+ int16x4_t s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ int16x4_t s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ int16x4_t s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+ int16x4_t s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
+
+ int16x4_t d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filter);
+ int16x4_t d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filter);
+ int16x4_t d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filter);
+ int16x4_t d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filter);
+ uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
+ uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
store_u8x4_strided_x2(dst + 0 * dst_stride, dst_stride, d01);
store_u8x4_strided_x2(dst + 2 * dst_stride, dst_stride, d23);
@@ -291,42 +417,40 @@ void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
s4 = s8;
s5 = s9;
s6 = s10;
+
src += 4 * src_stride;
dst += 4 * dst_stride;
h -= 4;
} while (h != 0);
} else {
- uint8x8_t t0, t1, t2, t3, t4, t5, t6, d0, d1, d2, d3;
- int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
- int height;
- const uint8_t *s;
- uint8_t *d;
-
do {
+ uint8x8_t t0, t1, t2, t3, t4, t5, t6;
load_u8_8x7(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6);
- s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
- s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
- s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
- s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
- s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
- s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
- s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
-
- height = h;
- s = src + 7 * src_stride;
- d = dst;
+
+ int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ int16x8_t s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+ int height = h;
+ const uint8_t *s = src + 7 * src_stride;
+ uint8_t *d = dst;
do {
load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3);
- s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
- s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
- s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
- s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
- d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter);
- d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter);
- d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter);
- d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter);
+ int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ int16x8_t s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ int16x8_t s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
+
+ uint8x8_t d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter);
+ uint8x8_t d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter);
+ uint8x8_t d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter);
+ uint8x8_t d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter);
store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
@@ -337,6 +461,7 @@ void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
s4 = s8;
s5 = s9;
s6 = s10;
+
s += 4 * src_stride;
d += 4 * dst_stride;
height -= 4;
generated by cgit v1.2.3 (git 2.39.1) at 2024-08-28 16:42:16 +0000