Optimize Neon horizontal Wiener convolve for 5-tap filters
Wiener convolution filters can be either 5 or 7 taps. The current Neon
implementation pads 5-tap filters with zeros to treat them as 7-tap
filters. This patch adds a Neon path specialised for 5-tap horizontal
Wiener filters, avoiding the redundant work of using the 7-tap path
with a 5-tap filter.
Change-Id: I87ec5e1a546e2a112447f0bca1ba523b738aef2f
diff --git a/av1/common/arm/wiener_convolve_neon.c b/av1/common/arm/wiener_convolve_neon.c
index c127dd1..e4a59a1 100644
--- a/av1/common/arm/wiener_convolve_neon.c
+++ b/av1/common/arm/wiener_convolve_neon.c
@@ -22,6 +22,58 @@
#include "av1/common/common.h"
#include "av1/common/restoration.h"
+static INLINE uint16x8_t wiener_convolve5_8_2d_h(
+ const uint8x8_t t0, const uint8x8_t t1, const uint8x8_t t2,
+ const uint8x8_t t3, const uint8x8_t t4, const int16x4_t x_filter,
+ const int32x4_t round_vec, const uint16x8_t im_max_val) {
+ // Since the Wiener filter is symmetric about the middle tap (tap 2) add
+ // mirrored source elements before multiplying filter coefficients.
+ int16x8_t s04 = vreinterpretq_s16_u16(vaddl_u8(t0, t4));
+ int16x8_t s13 = vreinterpretq_s16_u16(vaddl_u8(t1, t3));
+ int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+
+ // x_filter[0] = 0. (5-tap filters are 0-padded to 7 taps.)
+ int32x4_t sum_lo = vmlal_lane_s16(round_vec, vget_low_s16(s04), x_filter, 1);
+ sum_lo = vmlal_lane_s16(sum_lo, vget_low_s16(s13), x_filter, 2);
+ sum_lo = vmlal_lane_s16(sum_lo, vget_low_s16(s2), x_filter, 3);
+
+ int32x4_t sum_hi = vmlal_lane_s16(round_vec, vget_high_s16(s04), x_filter, 1);
+ sum_hi = vmlal_lane_s16(sum_hi, vget_high_s16(s13), x_filter, 2);
+ sum_hi = vmlal_lane_s16(sum_hi, vget_high_s16(s2), x_filter, 3);
+
+ uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum_lo, WIENER_ROUND0_BITS),
+ vqrshrun_n_s32(sum_hi, WIENER_ROUND0_BITS));
+
+ return vminq_u16(res, im_max_val);
+}
+
+static INLINE void convolve_add_src_horiz_5tap_neon(
+ const uint8_t *src_ptr, ptrdiff_t src_stride, uint16_t *dst_ptr,
+ ptrdiff_t dst_stride, int w, int h, const int16x4_t x_filter,
+ const int32x4_t round_vec, const uint16x8_t im_max_val) {
+ do {
+ const uint8_t *s = src_ptr;
+ uint16_t *d = dst_ptr;
+ int width = w;
+
+ do {
+ uint8x8_t s0, s1, s2, s3, s4;
+ load_u8_8x5(s, 1, &s0, &s1, &s2, &s3, &s4);
+
+ uint16x8_t d0 = wiener_convolve5_8_2d_h(s0, s1, s2, s3, s4, x_filter,
+ round_vec, im_max_val);
+
+ vst1q_u16(d, d0);
+
+ s += 8;
+ d += 8;
+ width -= 8;
+ } while (width != 0);
+ src_ptr += src_stride;
+ dst_ptr += dst_stride;
+ } while (--h != 0);
+}
+
static INLINE uint16x8_t wiener_convolve7_8_2d_h(
const uint8x8_t t0, const uint8x8_t t1, const uint8x8_t t2,
const uint8x8_t t3, const uint8x8_t t4, const uint8x8_t t5,
@@ -50,18 +102,10 @@
return vminq_u16(res, im_max_val);
}
-static INLINE void convolve_add_src_horiz_hip_neon(
+static INLINE void convolve_add_src_horiz_7tap_neon(
const uint8_t *src_ptr, ptrdiff_t src_stride, uint16_t *dst_ptr,
- ptrdiff_t dst_stride, const int16_t *x_filter_ptr, int w, int h) {
- const int bd = 8;
- const int32x4_t round_vec = vdupq_n_s32(1 << (bd + FILTER_BITS - 1));
- const uint16x8_t im_max_val =
- vdupq_n_u16((1 << (bd + 1 + FILTER_BITS - WIENER_ROUND0_BITS)) - 1);
-
- int16x4_t x_filter = vld1_s16(x_filter_ptr);
- // Add 128 to tap 3. (Needed for rounding.)
- x_filter = vadd_s16(x_filter, vcreate_s16((1ULL << FILTER_BITS) << 48));
-
+ ptrdiff_t dst_stride, int w, int h, const int16x4_t x_filter,
+ const int32x4_t round_vec, const uint16x8_t im_max_val) {
do {
const uint8_t *s = src_ptr;
uint16_t *d = dst_ptr;
@@ -168,6 +212,14 @@
} while (w != 0);
}
+static AOM_INLINE int get_wiener_filter_taps(const int16_t *filter) {
+ assert(filter[7] == 0);
+ if (filter[0] == 0 && filter[6] == 0) {
+ return WIENER_WIN_REDUCED;
+ }
+ return WIENER_WIN;
+}
+
// Wiener filter 2D
// Apply horizontal filter and store in a temporary buffer. When applying
// vertical filter, overwrite the original pixel values.
@@ -191,13 +243,31 @@
DECLARE_ALIGNED(16, uint16_t,
im_block[(MAX_SB_SIZE + WIENER_WIN - 1) * MAX_SB_SIZE]);
+ const int x_filter_taps = get_wiener_filter_taps(x_filter);
+ int16x4_t x_filter_s16 = vld1_s16(x_filter);
+ // Add 128 to tap 3. (Needed for rounding.)
+ x_filter_s16 = vadd_s16(x_filter_s16, vcreate_s16(128ULL << 48));
+
const int im_stride = MAX_SB_SIZE;
const int im_h = h + WIENER_WIN - 1;
const int horiz_offset = WIENER_HALFWIN;
const int vert_offset = WIENER_HALFWIN * (int)src_stride;
- convolve_add_src_horiz_hip_neon(src - horiz_offset - vert_offset, src_stride,
- im_block, im_stride, x_filter, w, im_h);
+ const int bd = 8;
+ const int32x4_t horiz_round_vec = vdupq_n_s32(1 << (bd + FILTER_BITS - 1));
+ const uint16x8_t im_max_val =
+ vdupq_n_u16((1 << (bd + 1 + FILTER_BITS - WIENER_ROUND0_BITS)) - 1);
+
+ if (x_filter_taps == WIENER_WIN_REDUCED) {
+ convolve_add_src_horiz_5tap_neon(src - horiz_offset - vert_offset + 1,
+ src_stride, im_block, im_stride, w, im_h,
+ x_filter_s16, horiz_round_vec, im_max_val);
+ } else {
+ convolve_add_src_horiz_7tap_neon(src - horiz_offset - vert_offset,
+ src_stride, im_block, im_stride, w, im_h,
+ x_filter_s16, horiz_round_vec, im_max_val);
+ }
+
convolve_add_src_vert_hip_neon(im_block, im_stride, dst, dst_stride, y_filter,
w, h);
}
