[NEON] highbd implementation of av1_highbd_convolve_y_sr_neon 2/3
Approximately 6.1x faster. This is the 2nd part of the 12-tap filter
implementation. Helper functions have been added.
Change-Id: Ibc7ae134a0dea7a231b87d47dd88eaf0d3e4b096
diff --git a/aom_dsp/arm/mem_neon.h b/aom_dsp/arm/mem_neon.h
index d27cff6..e07ea80 100644
--- a/aom_dsp/arm/mem_neon.h
+++ b/aom_dsp/arm/mem_neon.h
@@ -237,6 +237,36 @@
*s10 = vld1_s16(s);
}
+static INLINE void load_u16_4x11(const uint16_t *s, ptrdiff_t p,
+ uint16x4_t *const s0, uint16x4_t *const s1,
+ uint16x4_t *const s2, uint16x4_t *const s3,
+ uint16x4_t *const s4, uint16x4_t *const s5,
+ uint16x4_t *const s6, uint16x4_t *const s7,
+ uint16x4_t *const s8, uint16x4_t *const s9,
+ uint16x4_t *const s10) {
+ *s0 = vld1_u16(s);
+ s += p;
+ *s1 = vld1_u16(s);
+ s += p;
+ *s2 = vld1_u16(s);
+ s += p;
+ *s3 = vld1_u16(s);
+ s += p;
+ *s4 = vld1_u16(s);
+ s += p;
+ *s5 = vld1_u16(s);
+ s += p;
+ *s6 = vld1_u16(s);
+ s += p;
+ *s7 = vld1_u16(s);
+ s += p;
+ *s8 = vld1_u16(s);
+ s += p;
+ *s9 = vld1_u16(s);
+ s += p;
+ *s10 = vld1_u16(s);
+}
+
static INLINE void load_s16_4x8(const int16_t *s, ptrdiff_t p,
int16x4_t *const s0, int16x4_t *const s1,
int16x4_t *const s2, int16x4_t *const s3,
@@ -553,6 +583,36 @@
*s10 = vld1q_s16(s);
}
+static INLINE void load_u16_8x11(const uint16_t *s, ptrdiff_t p,
+ uint16x8_t *const s0, uint16x8_t *const s1,
+ uint16x8_t *const s2, uint16x8_t *const s3,
+ uint16x8_t *const s4, uint16x8_t *const s5,
+ uint16x8_t *const s6, uint16x8_t *const s7,
+ uint16x8_t *const s8, uint16x8_t *const s9,
+ uint16x8_t *const s10) {
+ *s0 = vld1q_u16(s);
+ s += p;
+ *s1 = vld1q_u16(s);
+ s += p;
+ *s2 = vld1q_u16(s);
+ s += p;
+ *s3 = vld1q_u16(s);
+ s += p;
+ *s4 = vld1q_u16(s);
+ s += p;
+ *s5 = vld1q_u16(s);
+ s += p;
+ *s6 = vld1q_u16(s);
+ s += p;
+ *s7 = vld1q_u16(s);
+ s += p;
+ *s8 = vld1q_u16(s);
+ s += p;
+ *s9 = vld1q_u16(s);
+ s += p;
+ *s10 = vld1q_u16(s);
+}
+
static INLINE void load_s16_8x8(const int16_t *s, ptrdiff_t p,
int16x8_t *const s0, int16x8_t *const s1,
int16x8_t *const s2, int16x8_t *const s3,
diff --git a/av1/common/arm/highbd_convolve_neon.c b/av1/common/arm/highbd_convolve_neon.c
index 3af834f..12bd22f 100644
--- a/av1/common/arm/highbd_convolve_neon.c
+++ b/av1/common/arm/highbd_convolve_neon.c
@@ -23,27 +23,197 @@
#include "av1/common/filter.h"
#include "av1/common/arm/highbd_convolve_neon.h"
+static INLINE void highbd_convolve_y_sr_12tap_neon(
+ const uint16_t *src_ptr, int src_stride, uint16_t *dst_ptr, int dst_stride,
+ int w, int h, const int16_t *y_filter_ptr, int bd) {
+ const uint16x8_t max = vdupq_n_u16((1 << bd) - 1);
+ const int16x8_t y_filter_0_7 = vld1q_s16(y_filter_ptr);
+ const int16x4_t y_filter_8_11 = vld1_s16(y_filter_ptr + 8);
+
+ if (w <= 4) {
+ uint16x4_t t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14;
+ int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14;
+ uint16x4_t d0, d1, d2, d3;
+ uint16x8_t d01, d23;
+
+ const uint16_t *s = src_ptr;
+ uint16_t *d = dst_ptr;
+
+ load_u16_4x11(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, &t8,
+ &t9, &t10);
+ s0 = vreinterpret_s16_u16(t0);
+ s1 = vreinterpret_s16_u16(t1);
+ s2 = vreinterpret_s16_u16(t2);
+ s3 = vreinterpret_s16_u16(t3);
+ s4 = vreinterpret_s16_u16(t4);
+ s5 = vreinterpret_s16_u16(t5);
+ s6 = vreinterpret_s16_u16(t6);
+ s7 = vreinterpret_s16_u16(t7);
+ s8 = vreinterpret_s16_u16(t8);
+ s9 = vreinterpret_s16_u16(t9);
+ s10 = vreinterpret_s16_u16(t10);
+
+ s += 11 * src_stride;
+
+ do {
+ load_u16_4x4(s, src_stride, &t11, &t12, &t13, &t14);
+ s11 = vreinterpret_s16_u16(t11);
+ s12 = vreinterpret_s16_u16(t12);
+ s13 = vreinterpret_s16_u16(t13);
+ s14 = vreinterpret_s16_u16(t14);
+
+ d0 = highbd_convolve12_y_4x4_s32_s16(s0, s1, s2, s3, s4, s5, s6, s7, s8,
+ s9, s10, s11, y_filter_0_7,
+ y_filter_8_11);
+ d1 = highbd_convolve12_y_4x4_s32_s16(s1, s2, s3, s4, s5, s6, s7, s8, s9,
+ s10, s11, s12, y_filter_0_7,
+ y_filter_8_11);
+ d2 = highbd_convolve12_y_4x4_s32_s16(s2, s3, s4, s5, s6, s7, s8, s9, s10,
+ s11, s12, s13, y_filter_0_7,
+ y_filter_8_11);
+ d3 = highbd_convolve12_y_4x4_s32_s16(s3, s4, s5, s6, s7, s8, s9, s10, s11,
+ s12, s13, s14, y_filter_0_7,
+ y_filter_8_11);
+
+ d01 = vcombine_u16(d0, d1);
+ d23 = vcombine_u16(d2, d3);
+
+ d01 = vminq_u16(d01, max);
+ d23 = vminq_u16(d23, max);
+
+ if (w == 2) {
+ store_u16q_2x1(d + 0 * dst_stride, d01, 0);
+ store_u16q_2x1(d + 1 * dst_stride, d01, 2);
+ if (h != 2) {
+ store_u16q_2x1(d + 2 * dst_stride, d23, 0);
+ store_u16q_2x1(d + 3 * dst_stride, d23, 2);
+ }
+ } else {
+ vst1_u16(d + 0 * dst_stride, vget_low_u16(d01));
+ vst1_u16(d + 1 * dst_stride, vget_high_u16(d01));
+ if (h != 2) {
+ vst1_u16(d + 2 * dst_stride, vget_low_u16(d23));
+ vst1_u16(d + 3 * dst_stride, vget_high_u16(d23));
+ }
+ }
+
+ s0 = s4;
+ s1 = s5;
+ s2 = s6;
+ s3 = s7;
+ s4 = s8;
+ s5 = s9;
+ s6 = s10;
+ s7 = s11;
+ s8 = s12;
+ s9 = s13;
+ s10 = s14;
+ s += 4 * src_stride;
+ d += 4 * dst_stride;
+ h -= 4;
+ } while (h > 0);
+ } else {
+ uint16x8_t t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11, t12, t13, t14;
+ uint16x8_t d0, d1, d2, d3;
+ int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14;
+
+ do {
+ const uint16_t *s = src_ptr;
+ uint16_t *d = dst_ptr;
+ int height = h;
+
+ load_u16_8x11(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, &t8,
+ &t9, &t10);
+ s0 = vreinterpretq_s16_u16(t0);
+ s1 = vreinterpretq_s16_u16(t1);
+ s2 = vreinterpretq_s16_u16(t2);
+ s3 = vreinterpretq_s16_u16(t3);
+ s4 = vreinterpretq_s16_u16(t4);
+ s5 = vreinterpretq_s16_u16(t5);
+ s6 = vreinterpretq_s16_u16(t6);
+ s7 = vreinterpretq_s16_u16(t7);
+ s8 = vreinterpretq_s16_u16(t8);
+ s9 = vreinterpretq_s16_u16(t9);
+ s10 = vreinterpretq_s16_u16(t10);
+
+ s += 11 * src_stride;
+
+ do {
+ load_u16_8x4(s, src_stride, &t11, &t12, &t13, &t14);
+ s11 = vreinterpretq_s16_u16(t11);
+ s12 = vreinterpretq_s16_u16(t12);
+ s13 = vreinterpretq_s16_u16(t13);
+ s14 = vreinterpretq_s16_u16(t14);
+
+ d0 = highbd_convolve12_y_8x4_s32_s16(s0, s1, s2, s3, s4, s5, s6, s7, s8,
+ s9, s10, s11, y_filter_0_7,
+ y_filter_8_11);
+ d1 = highbd_convolve12_y_8x4_s32_s16(s1, s2, s3, s4, s5, s6, s7, s8, s9,
+ s10, s11, s12, y_filter_0_7,
+ y_filter_8_11);
+ d2 = highbd_convolve12_y_8x4_s32_s16(s2, s3, s4, s5, s6, s7, s8, s9,
+ s10, s11, s12, s13, y_filter_0_7,
+ y_filter_8_11);
+ d3 = highbd_convolve12_y_8x4_s32_s16(s3, s4, s5, s6, s7, s8, s9, s10,
+ s11, s12, s13, s14, y_filter_0_7,
+ y_filter_8_11);
+
+ d0 = vminq_u16(d0, max);
+ d1 = vminq_u16(d1, max);
+ d2 = vminq_u16(d2, max);
+ d3 = vminq_u16(d3, max);
+
+ if (h == 2) {
+ store_u16_8x2(d, dst_stride, d0, d1);
+ } else {
+ store_u16_8x4(d, dst_stride, d0, d1, d2, d3);
+ }
+
+ s0 = s4;
+ s1 = s5;
+ s2 = s6;
+ s3 = s7;
+ s4 = s8;
+ s5 = s9;
+ s6 = s10;
+ s7 = s11;
+ s8 = s12;
+ s9 = s13;
+ s10 = s14;
+ s += 4 * src_stride;
+ d += 4 * dst_stride;
+ height -= 4;
+ } while (height > 0);
+
+ src_ptr += 8;
+ dst_ptr += 8;
+ w -= 8;
+ } while (w > 0);
+ }
+}
+
void av1_highbd_convolve_y_sr_neon(const uint16_t *src, int src_stride,
uint16_t *dst, int dst_stride, int w, int h,
const InterpFilterParams *filter_params_y,
const int subpel_y_qn, int bd) {
const int y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn);
- if (y_filter_taps > 8) {
- av1_highbd_convolve_y_sr_c(src, src_stride, dst, dst_stride, w, h,
- filter_params_y, subpel_y_qn, bd);
- return;
- }
const int vert_offset = filter_params_y->taps / 2 - 1;
- const uint16x8_t max = vdupq_n_u16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+ const uint16x8_t max = vdupq_n_u16((1 << bd) - 1);
const int16_t *y_filter_ptr = av1_get_interp_filter_subpel_kernel(
filter_params_y, subpel_y_qn & SUBPEL_MASK);
- const int16x8_t y_filter = vld1q_s16(y_filter_ptr);
-
src -= vert_offset * src_stride;
+ if (y_filter_taps > 8) {
+ highbd_convolve_y_sr_12tap_neon(src, src_stride, dst, dst_stride, w, h,
+ y_filter_ptr, bd);
+ return;
+ }
+
+ const int16x8_t y_filter = vld1q_s16(y_filter_ptr);
+
if (w <= 4) {
uint16x4_t t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
diff --git a/av1/common/arm/highbd_convolve_neon.h b/av1/common/arm/highbd_convolve_neon.h
index c683cf0..680c46e 100644
--- a/av1/common/arm/highbd_convolve_neon.h
+++ b/av1/common/arm/highbd_convolve_neon.h
@@ -82,4 +82,94 @@
vqrshrun_n_s32(sum1, COMPOUND_ROUND1_BITS));
}
+static INLINE int32x4_t highbd_convolve12_y_4x4_s32(
+ const int16x4_t s0, const int16x4_t s1, const int16x4_t s2,
+ const int16x4_t s3, const int16x4_t s4, const int16x4_t s5,
+ const int16x4_t s6, const int16x4_t s7, const int16x4_t s8,
+ const int16x4_t s9, const int16x4_t s10, const int16x4_t s11,
+ const int16x8_t y_filter_0_7, const int16x4_t y_filter_8_11) {
+ const int16x4_t y_filter_0_3 = vget_low_s16(y_filter_0_7);
+ const int16x4_t y_filter_4_7 = vget_high_s16(y_filter_0_7);
+
+ int32x4_t sum = vmull_lane_s16(s0, y_filter_0_3, 0);
+ sum = vmlal_lane_s16(sum, s1, y_filter_0_3, 1);
+ sum = vmlal_lane_s16(sum, s2, y_filter_0_3, 2);
+ sum = vmlal_lane_s16(sum, s3, y_filter_0_3, 3);
+ sum = vmlal_lane_s16(sum, s4, y_filter_4_7, 0);
+ sum = vmlal_lane_s16(sum, s5, y_filter_4_7, 1);
+ sum = vmlal_lane_s16(sum, s6, y_filter_4_7, 2);
+ sum = vmlal_lane_s16(sum, s7, y_filter_4_7, 3);
+ sum = vmlal_lane_s16(sum, s8, y_filter_8_11, 0);
+ sum = vmlal_lane_s16(sum, s9, y_filter_8_11, 1);
+ sum = vmlal_lane_s16(sum, s10, y_filter_8_11, 2);
+ sum = vmlal_lane_s16(sum, s11, y_filter_8_11, 3);
+
+ return sum;
+}
+
+static INLINE uint16x4_t highbd_convolve12_y_4x4_s32_s16(
+ const int16x4_t s0, const int16x4_t s1, const int16x4_t s2,
+ const int16x4_t s3, const int16x4_t s4, const int16x4_t s5,
+ const int16x4_t s6, const int16x4_t s7, const int16x4_t s8,
+ const int16x4_t s9, const int16x4_t s10, const int16x4_t s11,
+ const int16x8_t y_filter_0_7, const int16x4_t y_filter_8_11) {
+ int32x4_t sum =
+ highbd_convolve12_y_4x4_s32(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10,
+ s11, y_filter_0_7, y_filter_8_11);
+
+ return vqrshrun_n_s32(sum, FILTER_BITS);
+}
+
+static INLINE void highbd_convolve12_y_8x4_s32(
+ const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
+ const int16x8_t s3, const int16x8_t s4, const int16x8_t s5,
+ const int16x8_t s6, const int16x8_t s7, const int16x8_t s8,
+ const int16x8_t s9, const int16x8_t s10, const int16x8_t s11,
+ const int16x8_t y_filter_0_7, const int16x4_t y_filter_8_11,
+ int32x4_t *sum0, int32x4_t *sum1) {
+ const int16x4_t y_filter_0_3 = vget_low_s16(y_filter_0_7);
+ const int16x4_t y_filter_4_7 = vget_high_s16(y_filter_0_7);
+
+ *sum0 = vmull_lane_s16(vget_low_s16(s0), y_filter_0_3, 0);
+ *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s1), y_filter_0_3, 1);
+ *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s2), y_filter_0_3, 2);
+ *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s3), y_filter_0_3, 3);
+ *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s4), y_filter_4_7, 0);
+ *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s5), y_filter_4_7, 1);
+ *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s6), y_filter_4_7, 2);
+ *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s7), y_filter_4_7, 3);
+ *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s8), y_filter_8_11, 0);
+ *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s9), y_filter_8_11, 1);
+ *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s10), y_filter_8_11, 2);
+ *sum0 = vmlal_lane_s16(*sum0, vget_low_s16(s11), y_filter_8_11, 3);
+
+ *sum1 = vmull_lane_s16(vget_high_s16(s0), y_filter_0_3, 0);
+ *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s1), y_filter_0_3, 1);
+ *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s2), y_filter_0_3, 2);
+ *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s3), y_filter_0_3, 3);
+ *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s4), y_filter_4_7, 0);
+ *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s5), y_filter_4_7, 1);
+ *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s6), y_filter_4_7, 2);
+ *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s7), y_filter_4_7, 3);
+ *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s8), y_filter_8_11, 0);
+ *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s9), y_filter_8_11, 1);
+ *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s10), y_filter_8_11, 2);
+ *sum1 = vmlal_lane_s16(*sum1, vget_high_s16(s11), y_filter_8_11, 3);
+}
+
+static INLINE uint16x8_t highbd_convolve12_y_8x4_s32_s16(
+ const int16x8_t s0, const int16x8_t s1, const int16x8_t s2,
+ const int16x8_t s3, const int16x8_t s4, const int16x8_t s5,
+ const int16x8_t s6, const int16x8_t s7, const int16x8_t s8,
+ const int16x8_t s9, const int16x8_t s10, const int16x8_t s11,
+ const int16x8_t y_filter_0_7, const int16x4_t y_filter_8_11) {
+ int32x4_t sum0;
+ int32x4_t sum1;
+ highbd_convolve12_y_8x4_s32(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11,
+ y_filter_0_7, y_filter_8_11, &sum0, &sum1);
+
+ return vcombine_u16(vqrshrun_n_s32(sum0, FILTER_BITS),
+ vqrshrun_n_s32(sum1, FILTER_BITS));
+}
+
#endif // AOM_AV1_COMMON_ARM_HIGHBD_CONVOLVE_NEON_H_
