[NEON] highbd implementation of av1_highbd_convolve_x_sr_neon 1/2
Approximately 4x faster, total gain ~1%. First part is tackling 8-tap filter.
Change-Id: I0de480a848183c1d990af607112e132fdc578671
diff --git a/aom_dsp/arm/mem_neon.h b/aom_dsp/arm/mem_neon.h
index 8f76f02..59de63c 100644
--- a/aom_dsp/arm/mem_neon.h
+++ b/aom_dsp/arm/mem_neon.h
@@ -194,6 +194,13 @@
*s6 = vld1_u16(s);
}
+static INLINE void load_u16_8x2(const uint16_t *s, const ptrdiff_t p,
+ uint16x8_t *const s0, uint16x8_t *const s1) {
+ *s0 = vld1q_u16(s);
+ s += p;
+ *s1 = vld1q_u16(s);
+}
+
static INLINE void load_u16_8x4(const uint16_t *s, const ptrdiff_t p,
uint16x8_t *const s0, uint16x8_t *const s1,
uint16x8_t *const s2, uint16x8_t *const s3) {
@@ -871,6 +878,24 @@
*s3 = vld1q_u8(s);
}
+static INLINE void load_u16_16x4(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) {
+ *s0 = vld1q_u16(s);
+ *s1 = vld1q_u16(s + 8);
+ s += p;
+ *s2 = vld1q_u16(s);
+ *s3 = vld1q_u16(s + 8);
+ s += p;
+ *s4 = vld1q_u16(s);
+ *s5 = vld1q_u16(s + 8);
+ s += p;
+ *s6 = vld1q_u16(s);
+ *s7 = vld1q_u16(s + 8);
+}
+
static INLINE void load_unaligned_u16_4x4(const uint16_t *buf, uint32_t stride,
uint16x8_t *tu0, uint16x8_t *tu1) {
uint64_t a;
diff --git a/av1/common/arm/highbd_convolve_neon.c b/av1/common/arm/highbd_convolve_neon.c
index 5a325c8..68c1358 100644
--- a/av1/common/arm/highbd_convolve_neon.c
+++ b/av1/common/arm/highbd_convolve_neon.c
@@ -473,3 +473,123 @@
} while (w > 0);
}
}
+
+void av1_highbd_convolve_x_sr_neon(const uint16_t *src, int src_stride,
+ uint16_t *dst, int dst_stride, int w, int h,
+ const InterpFilterParams *filter_params_x,
+ const int subpel_x_qn,
+ ConvolveParams *conv_params, int bd) {
+ const int x_filter_taps = get_filter_tap(filter_params_x, subpel_x_qn);
+
+ if (x_filter_taps > 8) {
+ av1_highbd_convolve_x_sr_c(src, src_stride, dst, dst_stride, w, h,
+ filter_params_x, subpel_x_qn, conv_params, bd);
+ return;
+ }
+
+ const int horiz_offset = filter_params_x->taps / 2 - 1;
+ const uint16x8_t max = vdupq_n_u16((1 << bd) - 1);
+ const int32x4_t shift_s32 = vdupq_n_s32(-conv_params->round_0);
+ const int bits = FILTER_BITS - conv_params->round_0;
+ const int16x8_t bits_s16 = vdupq_n_s16(-bits);
+ const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel(
+ filter_params_x, subpel_x_qn & SUBPEL_MASK);
+
+ const int16x8_t x_filter = vld1q_s16(x_filter_ptr);
+
+ src -= horiz_offset;
+
+ if (w <= 4) {
+ uint16x8_t t0, t1, t2, t3;
+ int16x8_t s0, s1, s2, s3;
+ uint16x4_t d0, d1;
+ uint16x8_t d01;
+
+ const uint16_t *s = src;
+ uint16_t *d = dst;
+
+ do {
+ load_u16_8x2(s, src_stride, &t0, &t2);
+ load_u16_8x2(s + 8, src_stride, &t1, &t3);
+ s0 = vreinterpretq_s16_u16(t0);
+ s1 = vreinterpretq_s16_u16(t1);
+ s2 = vreinterpretq_s16_u16(t2);
+ s3 = vreinterpretq_s16_u16(t3);
+
+ d0 = highbd_convolve8_horiz4_s32_s16(s0, s1, x_filter, shift_s32);
+ d1 = highbd_convolve8_horiz4_s32_s16(s2, s3, x_filter, shift_s32);
+
+ d01 = vcombine_u16(d0, d1);
+ d01 = vqrshlq_u16(d01, bits_s16);
+ d01 = vminq_u16(d01, max);
+
+ if (w == 2) {
+ store_u16q_2x1(d + 0 * dst_stride, d01, 0);
+ store_u16q_2x1(d + 1 * dst_stride, d01, 2);
+ } else {
+ vst1_u16(d + 0 * dst_stride, vget_low_u16(d01));
+ vst1_u16(d + 1 * dst_stride, vget_high_u16(d01));
+ }
+
+ s += 2 * src_stride;
+ d += 2 * dst_stride;
+ h -= 2;
+ } while (h > 0);
+ } else {
+ int height = h;
+ uint16x8_t t0, t1, t2, t3, t4, t5, t6, t7;
+ int16x8_t s0, s1, s2, s3, s4, s5, s6, s7;
+ uint16x8_t d0, d1, d2, d3;
+ do {
+ int width = w;
+ const uint16_t *s = src;
+ uint16_t *d = dst;
+ load_u16_8x4(s, src_stride, &t0, &t2, &t4, &t6);
+ s0 = vreinterpretq_s16_u16(t0);
+ s2 = vreinterpretq_s16_u16(t2);
+ s4 = vreinterpretq_s16_u16(t4);
+ s6 = vreinterpretq_s16_u16(t6);
+
+ s += 8;
+ do {
+ load_u16_8x4(s, src_stride, &t1, &t3, &t5, &t7);
+ s1 = vreinterpretq_s16_u16(t1);
+ s3 = vreinterpretq_s16_u16(t3);
+ s5 = vreinterpretq_s16_u16(t5);
+ s7 = vreinterpretq_s16_u16(t7);
+
+ d0 = highbd_convolve8_horiz8_s32_s16(s0, s1, x_filter, shift_s32);
+ d1 = highbd_convolve8_horiz8_s32_s16(s2, s3, x_filter, shift_s32);
+ d2 = highbd_convolve8_horiz8_s32_s16(s4, s5, x_filter, shift_s32);
+ d3 = highbd_convolve8_horiz8_s32_s16(s6, s7, x_filter, shift_s32);
+
+ d0 = vqrshlq_u16(d0, bits_s16);
+ d1 = vqrshlq_u16(d1, bits_s16);
+ d2 = vqrshlq_u16(d2, bits_s16);
+ d3 = vqrshlq_u16(d3, bits_s16);
+
+ 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 = s1;
+ s2 = s3;
+ s4 = s5;
+ s6 = s7;
+ s += 8;
+ d += 8;
+ width -= 8;
+ } while (width > 0);
+ src += 4 * src_stride;
+ dst += 4 * dst_stride;
+ height -= 4;
+ } while (height > 0);
+ }
+}
diff --git a/av1/common/arm/highbd_convolve_neon.h b/av1/common/arm/highbd_convolve_neon.h
index 5ab3425..f5b4508 100644
--- a/av1/common/arm/highbd_convolve_neon.h
+++ b/av1/common/arm/highbd_convolve_neon.h
@@ -233,4 +233,60 @@
vqrshrun_n_s32(sum1, FILTER_BITS));
}
+static INLINE int32x4_t highbd_convolve8_horiz4_s32(
+ const int16x8_t s0, const int16x8_t s1, const int16x8_t x_filter_0_7) {
+ const int16x8_t s2 = vextq_s16(s0, s1, 1);
+ const int16x8_t s3 = vextq_s16(s0, s1, 2);
+ const int16x8_t s4 = vextq_s16(s0, s1, 3);
+ const int16x4_t s0_lo = vget_low_s16(s0);
+ const int16x4_t s1_lo = vget_low_s16(s2);
+ const int16x4_t s2_lo = vget_low_s16(s3);
+ const int16x4_t s3_lo = vget_low_s16(s4);
+ const int16x4_t s4_lo = vget_high_s16(s0);
+ const int16x4_t s5_lo = vget_high_s16(s2);
+ const int16x4_t s6_lo = vget_high_s16(s3);
+ const int16x4_t s7_lo = vget_high_s16(s4);
+
+ return highbd_convolve8_4_s32(s0_lo, s1_lo, s2_lo, s3_lo, s4_lo, s5_lo, s6_lo,
+ s7_lo, x_filter_0_7);
+}
+
+static INLINE uint16x4_t highbd_convolve8_horiz4_s32_s16(
+ const int16x8_t s0, const int16x8_t s1, const int16x8_t x_filter_0_7,
+ const int32x4_t shift_s32) {
+ int32x4_t sum = highbd_convolve8_horiz4_s32(s0, s1, x_filter_0_7);
+
+ sum = vqrshlq_s32(sum, shift_s32);
+ return vqmovun_s32(sum);
+}
+
+static INLINE void highbd_convolve8_horiz8_s32(const int16x8_t s0,
+ const int16x8_t s0_hi,
+ const int16x8_t x_filter_0_7,
+ int32x4_t *sum0,
+ int32x4_t *sum1) {
+ const int16x8_t s1 = vextq_s16(s0, s0_hi, 1);
+ const int16x8_t s2 = vextq_s16(s0, s0_hi, 2);
+ const int16x8_t s3 = vextq_s16(s0, s0_hi, 3);
+ const int16x8_t s4 = vextq_s16(s0, s0_hi, 4);
+ const int16x8_t s5 = vextq_s16(s0, s0_hi, 5);
+ const int16x8_t s6 = vextq_s16(s0, s0_hi, 6);
+ const int16x8_t s7 = vextq_s16(s0, s0_hi, 7);
+
+ highbd_convolve8_8_s32(s0, s1, s2, s3, s4, s5, s6, s7, x_filter_0_7, sum0,
+ sum1);
+}
+
+static INLINE uint16x8_t highbd_convolve8_horiz8_s32_s16(
+ const int16x8_t s0, const int16x8_t s1, const int16x8_t x_filter_0_7,
+ const int32x4_t shift_s32) {
+ int32x4_t sum0, sum1;
+ highbd_convolve8_horiz8_s32(s0, s1, x_filter_0_7, &sum0, &sum1);
+
+ sum0 = vqrshlq_s32(sum0, shift_s32);
+ sum1 = vqrshlq_s32(sum1, shift_s32);
+
+ return vcombine_u16(vqmovun_s32(sum0), vqmovun_s32(sum1));
+}
+
#endif // AOM_AV1_COMMON_ARM_HIGHBD_CONVOLVE_NEON_H_
diff --git a/av1/common/av1_rtcd_defs.pl b/av1/common/av1_rtcd_defs.pl
index 845fd15..73003a9 100644
--- a/av1/common/av1_rtcd_defs.pl
+++ b/av1/common/av1_rtcd_defs.pl
@@ -596,7 +596,7 @@
specialize qw/av1_highbd_dist_wtd_convolve_y sse4_1 avx2/;
specialize qw/av1_highbd_dist_wtd_convolve_2d_copy sse4_1 avx2/;
specialize qw/av1_highbd_convolve_2d_sr ssse3 avx2/;
- specialize qw/av1_highbd_convolve_x_sr ssse3 avx2/;
+ specialize qw/av1_highbd_convolve_x_sr ssse3 avx2 neon/;
specialize qw/av1_highbd_convolve_y_sr ssse3 avx2 neon/;
specialize qw/av1_highbd_convolve_2d_scale sse4_1/;
}
diff --git a/test/av1_convolve_test.cc b/test/av1_convolve_test.cc
index 16ac687..1453d0d 100644
--- a/test/av1_convolve_test.cc
+++ b/test/av1_convolve_test.cc
@@ -535,6 +535,11 @@
BuildHighbdParams(av1_highbd_convolve_x_sr_avx2));
#endif
+#if HAVE_NEON
+INSTANTIATE_TEST_SUITE_P(NEON, AV1ConvolveXHighbdTest,
+ BuildHighbdParams(av1_highbd_convolve_x_sr_neon));
+#endif
+
#endif // CONFIG_AV1_HIGHBITDEPTH
////////////////////////////////////////////////////////
