Optimize Neon vertical 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 vertical
Wiener filters, avoiding the redundant work of using the 7-tap path
with a 5-tap filter.
Change-Id: If53c96828551789ac20751f6d73f3caaecd82971
diff --git a/av1/common/arm/wiener_convolve_neon.c b/av1/common/arm/wiener_convolve_neon.c
index e4a59a1..6440c16 100644
--- a/av1/common/arm/wiener_convolve_neon.c
+++ b/av1/common/arm/wiener_convolve_neon.c
@@ -129,6 +129,77 @@
} while (--h != 0);
}
+static INLINE uint8x8_t wiener_convolve5_8_2d_v(
+ const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
+ const int16x8_t s3, const int16x8_t s4, const int16x4_t y_filter,
+ const int32x4_t round_vec) {
+ // Since the Wiener filter is symmetric about the middle tap (tap 2) add
+ // mirrored source elements before multiplying by filter coefficients.
+ int16x8_t s04 = vaddq_s16(s0, s4);
+ int16x8_t s13 = vaddq_s16(s1, s3);
+
+ int32x4_t sum_lo = vmlal_lane_s16(round_vec, vget_low_s16(s04), y_filter, 1);
+ sum_lo = vmlal_lane_s16(sum_lo, vget_low_s16(s13), y_filter, 2);
+ sum_lo = vmlal_lane_s16(sum_lo, vget_low_s16(s2), y_filter, 3);
+
+ int32x4_t sum_hi = vmlal_lane_s16(round_vec, vget_high_s16(s04), y_filter, 1);
+ sum_hi = vmlal_lane_s16(sum_hi, vget_high_s16(s13), y_filter, 2);
+ sum_hi = vmlal_lane_s16(sum_hi, vget_high_s16(s2), y_filter, 3);
+
+ int16x4_t res_lo = vshrn_n_s32(sum_lo, 2 * FILTER_BITS - WIENER_ROUND0_BITS);
+ int16x4_t res_hi = vshrn_n_s32(sum_hi, 2 * FILTER_BITS - WIENER_ROUND0_BITS);
+
+ return vqmovun_s16(vcombine_s16(res_lo, res_hi));
+}
+
+static INLINE void convolve_add_src_vert_5tap_neon(
+ const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst,
+ ptrdiff_t dst_stride, int w, int h, const int16x4_t y_filter,
+ const int32x4_t round_vec) {
+ do {
+ const int16_t *s = (int16_t *)src;
+ uint8_t *d = dst;
+ int height = h;
+
+ while (height > 3) {
+ int16x8_t s0, s1, s2, s3, s4, s5, s6, s7;
+ load_s16_8x8(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
+
+ uint8x8_t d0 =
+ wiener_convolve5_8_2d_v(s0, s1, s2, s3, s4, y_filter, round_vec);
+ uint8x8_t d1 =
+ wiener_convolve5_8_2d_v(s1, s2, s3, s4, s5, y_filter, round_vec);
+ uint8x8_t d2 =
+ wiener_convolve5_8_2d_v(s2, s3, s4, s5, s6, y_filter, round_vec);
+ uint8x8_t d3 =
+ wiener_convolve5_8_2d_v(s3, s4, s5, s6, s7, y_filter, round_vec);
+
+ store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
+
+ s += 4 * src_stride;
+ d += 4 * dst_stride;
+ height -= 4;
+ }
+
+ while (height-- != 0) {
+ int16x8_t s0, s1, s2, s3, s4;
+ load_s16_8x5(s, src_stride, &s0, &s1, &s2, &s3, &s4);
+
+ uint8x8_t d0 =
+ wiener_convolve5_8_2d_v(s0, s1, s2, s3, s4, y_filter, round_vec);
+
+ vst1_u8(d, d0);
+
+ d += dst_stride;
+ s += src_stride;
+ }
+
+ src += 8;
+ dst += 8;
+ w -= 8;
+ } while (w != 0);
+}
+
static INLINE uint8x8_t wiener_convolve7_8_2d_v(
const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
const int16x8_t s3, const int16x8_t s4, const int16x8_t s5,
@@ -155,18 +226,10 @@
return vqmovun_s16(vcombine_s16(res_lo, res_hi));
}
-static INLINE void convolve_add_src_vert_hip_neon(
+static INLINE void convolve_add_src_vert_7tap_neon(
const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst,
- ptrdiff_t dst_stride, const int16_t *y_filter_ptr, int w, int h) {
- const int bd = 8;
- const int32x4_t round_vec =
- vdupq_n_s32((1 << (2 * FILTER_BITS - WIENER_ROUND0_BITS - 1)) -
- (1 << (bd + (2 * FILTER_BITS - WIENER_ROUND0_BITS) - 1)));
-
- int16x4_t y_filter = vld1_s16(y_filter_ptr);
- // Add 128 to tap 3. (Needed for rounding.)
- y_filter = vadd_s16(y_filter, vcreate_s16((1ULL << FILTER_BITS) << 48));
-
+ ptrdiff_t dst_stride, int w, int h, const int16x4_t y_filter,
+ const int32x4_t round_vec) {
do {
const int16_t *s = (int16_t *)src;
uint8_t *d = dst;
@@ -244,22 +307,29 @@
im_block[(MAX_SB_SIZE + WIENER_WIN - 1) * MAX_SB_SIZE]);
const int x_filter_taps = get_wiener_filter_taps(x_filter);
+ const int y_filter_taps = get_wiener_filter_taps(y_filter);
int16x4_t x_filter_s16 = vld1_s16(x_filter);
+ int16x4_t y_filter_s16 = vld1_s16(y_filter);
// Add 128 to tap 3. (Needed for rounding.)
x_filter_s16 = vadd_s16(x_filter_s16, vcreate_s16(128ULL << 48));
+ y_filter_s16 = vadd_s16(y_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;
+ const int im_h = h + y_filter_taps - 1;
+ const int horiz_offset = x_filter_taps / 2;
+ const int vert_offset = (y_filter_taps / 2) * (int)src_stride;
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);
+ const int32x4_t horiz_round_vec = vdupq_n_s32(1 << (bd + FILTER_BITS - 1));
+
+ const int32x4_t vert_round_vec =
+ vdupq_n_s32((1 << (2 * FILTER_BITS - WIENER_ROUND0_BITS - 1)) -
+ (1 << (bd + (2 * 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,
+ convolve_add_src_horiz_5tap_neon(src - horiz_offset - vert_offset,
src_stride, im_block, im_stride, w, im_h,
x_filter_s16, horiz_round_vec, im_max_val);
} else {
@@ -268,6 +338,11 @@
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);
+ if (y_filter_taps == WIENER_WIN_REDUCED) {
+ convolve_add_src_vert_5tap_neon(im_block, im_stride, dst, dst_stride, w, h,
+ y_filter_s16, vert_round_vec);
+ } else {
+ convolve_add_src_vert_7tap_neon(im_block, im_stride, dst, dst_stride, w, h,
+ y_filter_s16, vert_round_vec);
+ }
}
