Optimize HBD 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: Ic081be65d7a9e6b8c8711cf9778dd6ce4654336d
diff --git a/av1/common/arm/highbd_wiener_convolve_neon.c b/av1/common/arm/highbd_wiener_convolve_neon.c
index 588b4f8..a6bd6d3 100644
--- a/av1/common/arm/highbd_wiener_convolve_neon.c
+++ b/av1/common/arm/highbd_wiener_convolve_neon.c
@@ -136,6 +136,90 @@
#undef HBD_WIENER_7TAP_HORIZ
+#define HBD_WIENER_5TAP_VERT(name, shift) \
+ static INLINE uint16x8_t name##_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, const uint16x8_t res_max_val) { \
+ const int32x2_t y_filter_lo = vget_low_s32(vmovl_s16(y_filter)); \
+ const int32x2_t y_filter_hi = vget_high_s32(vmovl_s16(y_filter)); \
+ /* Wiener filter is symmetric so add mirrored source elements. */ \
+ int32x4_t s04_lo = vaddl_s16(vget_low_s16(s0), vget_low_s16(s4)); \
+ int32x4_t s13_lo = vaddl_s16(vget_low_s16(s1), vget_low_s16(s3)); \
+ \
+ /* y_filter[0] = 0. (5-tap filters are 0-padded to 7 taps.) */ \
+ int32x4_t sum_lo = vmlaq_lane_s32(round_vec, s04_lo, y_filter_lo, 1); \
+ sum_lo = vmlaq_lane_s32(sum_lo, s13_lo, y_filter_hi, 0); \
+ sum_lo = \
+ vmlaq_lane_s32(sum_lo, vmovl_s16(vget_low_s16(s2)), y_filter_hi, 1); \
+ \
+ int32x4_t s04_hi = vaddl_s16(vget_high_s16(s0), vget_high_s16(s4)); \
+ int32x4_t s13_hi = vaddl_s16(vget_high_s16(s1), vget_high_s16(s3)); \
+ \
+ int32x4_t sum_hi = vmlaq_lane_s32(round_vec, s04_hi, y_filter_lo, 1); \
+ sum_hi = vmlaq_lane_s32(sum_hi, s13_hi, y_filter_hi, 0); \
+ sum_hi = \
+ vmlaq_lane_s32(sum_hi, vmovl_s16(vget_high_s16(s2)), y_filter_hi, 1); \
+ \
+ uint16x4_t res_lo = vqrshrun_n_s32(sum_lo, shift); \
+ uint16x4_t res_hi = vqrshrun_n_s32(sum_hi, shift); \
+ \
+ return vminq_u16(vcombine_u16(res_lo, res_hi), res_max_val); \
+ } \
+ \
+ static INLINE void name##_convolve_add_src_5tap_vert( \
+ const uint16_t *src_ptr, ptrdiff_t src_stride, uint16_t *dst_ptr, \
+ ptrdiff_t dst_stride, int w, int h, const int16x4_t y_filter, \
+ const int32x4_t round_vec, const uint16x8_t res_max_val) { \
+ do { \
+ const int16_t *s = (int16_t *)src_ptr; \
+ uint16_t *d = dst_ptr; \
+ 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); \
+ \
+ uint16x8_t d0 = name##_wiener_convolve5_8_2d_v( \
+ s0, s1, s2, s3, s4, y_filter, round_vec, res_max_val); \
+ uint16x8_t d1 = name##_wiener_convolve5_8_2d_v( \
+ s1, s2, s3, s4, s5, y_filter, round_vec, res_max_val); \
+ uint16x8_t d2 = name##_wiener_convolve5_8_2d_v( \
+ s2, s3, s4, s5, s6, y_filter, round_vec, res_max_val); \
+ uint16x8_t d3 = name##_wiener_convolve5_8_2d_v( \
+ s3, s4, s5, s6, s7, y_filter, round_vec, res_max_val); \
+ \
+ store_u16_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); \
+ \
+ uint16x8_t d0 = name##_wiener_convolve5_8_2d_v( \
+ s0, s1, s2, s3, s4, y_filter, round_vec, res_max_val); \
+ \
+ vst1q_u16(d, d0); \
+ \
+ s += src_stride; \
+ d += dst_stride; \
+ } \
+ \
+ src_ptr += 8; \
+ dst_ptr += 8; \
+ w -= 8; \
+ } while (w != 0); \
+ }
+
+HBD_WIENER_5TAP_VERT(highbd, 2 * FILTER_BITS - WIENER_ROUND0_BITS)
+HBD_WIENER_5TAP_VERT(highbd_12, 2 * FILTER_BITS - WIENER_ROUND0_BITS - 2)
+
+#undef HBD_WIENER_5TAP_VERT
+
#define HBD_WIENER_7TAP_VERT(name, shift) \
static INLINE uint16x8_t name##_wiener_convolve7_8_2d_v( \
const int16x8_t s0, const int16x8_t s1, const int16x8_t s2, \
@@ -171,7 +255,7 @@
return vminq_u16(vcombine_u16(res_lo, res_hi), res_max_val); \
} \
\
- static INLINE void name##_convolve_add_src_vert_hip( \
+ static INLINE void name##_convolve_add_src_7tap_vert( \
const uint16_t *src_ptr, ptrdiff_t src_stride, uint16_t *dst_ptr, \
ptrdiff_t dst_stride, int w, int h, const int16x4_t y_filter, \
const int32x4_t round_vec, const uint16x8_t res_max_val) { \
@@ -250,6 +334,7 @@
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.)
@@ -257,9 +342,9 @@
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 im_h = h + y_filter_taps - 1;
const int horiz_offset = x_filter_taps / 2;
- const int vert_offset = WIENER_HALFWIN * (int)src_stride;
+ const int vert_offset = (y_filter_taps / 2) * (int)src_stride;
const int extraprec_clamp_limit =
WIENER_CLAMP_LIMIT(conv_params->round_0, bd);
@@ -284,9 +369,16 @@
im_h, x_filter_s16, horiz_round_vec, im_max_val);
}
- highbd_12_convolve_add_src_vert_hip(im_block, im_stride, dst, dst_stride, w,
- h, y_filter_s16, vert_round_vec,
- res_max_val);
+ if (y_filter_taps == WIENER_WIN_REDUCED) {
+ highbd_12_convolve_add_src_5tap_vert(im_block, im_stride, dst, dst_stride,
+ w, h, y_filter_s16, vert_round_vec,
+ res_max_val);
+ } else {
+ highbd_12_convolve_add_src_7tap_vert(im_block, im_stride, dst, dst_stride,
+ w, h, y_filter_s16, vert_round_vec,
+ res_max_val);
+ }
+
} else {
if (x_filter_taps == WIENER_WIN_REDUCED) {
highbd_convolve_add_src_5tap_horiz(
@@ -298,7 +390,14 @@
im_h, x_filter_s16, horiz_round_vec, im_max_val);
}
- highbd_convolve_add_src_vert_hip(im_block, im_stride, dst, dst_stride, w, h,
- y_filter_s16, vert_round_vec, res_max_val);
+ if (y_filter_taps == WIENER_WIN_REDUCED) {
+ highbd_convolve_add_src_5tap_vert(im_block, im_stride, dst, dst_stride, w,
+ h, y_filter_s16, vert_round_vec,
+ res_max_val);
+ } else {
+ highbd_convolve_add_src_7tap_vert(im_block, im_stride, dst, dst_stride, w,
+ h, y_filter_s16, vert_round_vec,
+ res_max_val);
+ }
}
}
