From 8dd16259287f58f9273002717ec4d27e97127719 Mon Sep 17 00:00:00 2001 From: Daniel Baumann Date: Wed, 12 Jun 2024 07:43:14 +0200 Subject: Merging upstream version 127.0. Signed-off-by: Daniel Baumann --- third_party/aom/aom_dsp/arm/aom_convolve8_neon.c | 401 ++++++++++++------- .../aom/aom_dsp/arm/aom_convolve8_neon_dotprod.c | 428 ++++++++++----------- .../aom/aom_dsp/arm/aom_convolve8_neon_i8mm.c | 334 ++++++++-------- 3 files changed, 619 insertions(+), 544 deletions(-) (limited to 'third_party/aom/aom_dsp/arm') diff --git a/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c b/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c index 7441108b01..6a177b2e6b 100644 --- a/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c +++ b/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c @@ -20,6 +20,7 @@ #include "aom/aom_integer.h" #include "aom_dsp/aom_dsp_common.h" #include "aom_dsp/aom_filter.h" +#include "aom_dsp/arm/aom_filter.h" #include "aom_dsp/arm/mem_neon.h" #include "aom_dsp/arm/transpose_neon.h" #include "aom_ports/mem.h" @@ -31,14 +32,14 @@ static INLINE int16x4_t convolve8_4(const int16x4_t s0, const int16x4_t s1, const int16x8_t filter) { const int16x4_t filter_lo = vget_low_s16(filter); const int16x4_t filter_hi = vget_high_s16(filter); - int16x4_t sum; - sum = vmul_lane_s16(s0, filter_lo, 0); + int16x4_t sum = vmul_lane_s16(s0, filter_lo, 0); sum = vmla_lane_s16(sum, s1, filter_lo, 1); sum = vmla_lane_s16(sum, s2, filter_lo, 2); sum = vmla_lane_s16(sum, s5, filter_hi, 1); sum = vmla_lane_s16(sum, s6, filter_hi, 2); sum = vmla_lane_s16(sum, s7, filter_hi, 3); + sum = vqadd_s16(sum, vmul_lane_s16(s3, filter_lo, 3)); sum = vqadd_s16(sum, vmul_lane_s16(s4, filter_hi, 0)); return sum; @@ -51,65 +52,56 @@ static INLINE uint8x8_t convolve8_8(const int16x8_t s0, const int16x8_t s1, const int16x8_t filter) { const int16x4_t filter_lo = vget_low_s16(filter); const int16x4_t filter_hi = vget_high_s16(filter); - int16x8_t sum; - sum = vmulq_lane_s16(s0, filter_lo, 0); + int16x8_t sum = vmulq_lane_s16(s0, filter_lo, 0); sum = vmlaq_lane_s16(sum, s1, filter_lo, 1); sum = vmlaq_lane_s16(sum, s2, filter_lo, 2); sum = vmlaq_lane_s16(sum, s5, filter_hi, 1); sum = vmlaq_lane_s16(sum, s6, filter_hi, 2); sum = vmlaq_lane_s16(sum, s7, filter_hi, 3); + sum = vqaddq_s16(sum, vmulq_lane_s16(s3, filter_lo, 3)); sum = vqaddq_s16(sum, vmulq_lane_s16(s4, filter_hi, 0)); return vqrshrun_n_s16(sum, FILTER_BITS); } -void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, - uint8_t *dst, ptrdiff_t dst_stride, - const int16_t *filter_x, int x_step_q4, - const int16_t *filter_y, int y_step_q4, int w, - int h) { +static INLINE void convolve8_horiz_8tap_neon(const uint8_t *src, + ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, + const int16_t *filter_x, int w, + int h) { const int16x8_t filter = vld1q_s16(filter_x); - assert((intptr_t)dst % 4 == 0); - assert(dst_stride % 4 == 0); - - (void)x_step_q4; - (void)filter_y; - (void)y_step_q4; - - src -= ((SUBPEL_TAPS / 2) - 1); - if (h == 4) { - uint8x8_t t0, t1, t2, t3, d01, d23; - int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3; - + uint8x8_t t0, t1, t2, t3; load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3); transpose_elems_inplace_u8_8x4(&t0, &t1, &t2, &t3); - s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0))); - s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1))); - s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2))); - s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3))); - s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0))); - s5 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t1))); - s6 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t2))); + + int16x4_t s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0))); + int16x4_t s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1))); + int16x4_t s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2))); + int16x4_t s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3))); + int16x4_t s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0))); + int16x4_t s5 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t1))); + int16x4_t s6 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t2))); src += 7; do { load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3); transpose_elems_inplace_u8_8x4(&t0, &t1, &t2, &t3); - s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0))); - s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1))); - s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2))); - s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3))); - - d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filter); - d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filter); - d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filter); - d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filter); - d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS); - d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS); + + int16x4_t s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0))); + int16x4_t s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1))); + int16x4_t s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2))); + int16x4_t s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3))); + + int16x4_t d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filter); + int16x4_t d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filter); + int16x4_t d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filter); + int16x4_t d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filter); + uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS); + uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS); transpose_elems_inplace_u8_4x4(&d01, &d23); @@ -123,39 +115,40 @@ void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, s4 = s8; s5 = s9; s6 = s10; + src += 4; dst += 4; w -= 4; } while (w != 0); } else { - uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7, d0, d1, d2, d3; - int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10; - if (w == 4) { do { + uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7; load_u8_8x8(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7); transpose_elems_inplace_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7); - s0 = vreinterpretq_s16_u16(vmovl_u8(t0)); - s1 = vreinterpretq_s16_u16(vmovl_u8(t1)); - s2 = vreinterpretq_s16_u16(vmovl_u8(t2)); - s3 = vreinterpretq_s16_u16(vmovl_u8(t3)); - s4 = vreinterpretq_s16_u16(vmovl_u8(t4)); - s5 = vreinterpretq_s16_u16(vmovl_u8(t5)); - s6 = vreinterpretq_s16_u16(vmovl_u8(t6)); + + int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0)); + int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t1)); + int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2)); + int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t3)); + int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t4)); + int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t5)); + int16x8_t s6 = vreinterpretq_s16_u16(vmovl_u8(t6)); load_u8_8x8(src + 7, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7); transpose_elems_u8_4x8(t0, t1, t2, t3, t4, t5, t6, t7, &t0, &t1, &t2, &t3); - s7 = vreinterpretq_s16_u16(vmovl_u8(t0)); - s8 = vreinterpretq_s16_u16(vmovl_u8(t1)); - s9 = vreinterpretq_s16_u16(vmovl_u8(t2)); - s10 = vreinterpretq_s16_u16(vmovl_u8(t3)); - d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter); - d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter); - d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter); - d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter); + int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(t0)); + int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t1)); + int16x8_t s9 = vreinterpretq_s16_u16(vmovl_u8(t2)); + int16x8_t s10 = vreinterpretq_s16_u16(vmovl_u8(t3)); + + uint8x8_t d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter); + uint8x8_t d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter); + uint8x8_t d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter); + uint8x8_t d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter); transpose_elems_inplace_u8_8x4(&d0, &d1, &d2, &d3); @@ -169,48 +162,49 @@ void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, h -= 8; } while (h > 0); } else { - uint8x8_t d4, d5, d6, d7; - int16x8_t s11, s12, s13, s14; - int width; - const uint8_t *s; - uint8_t *d; - do { - load_u8_8x8(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7); + int width = w; + const uint8_t *s = src; + uint8_t *d = dst; + + uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7; + load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7); transpose_elems_inplace_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7); - s0 = vreinterpretq_s16_u16(vmovl_u8(t0)); - s1 = vreinterpretq_s16_u16(vmovl_u8(t1)); - s2 = vreinterpretq_s16_u16(vmovl_u8(t2)); - s3 = vreinterpretq_s16_u16(vmovl_u8(t3)); - s4 = vreinterpretq_s16_u16(vmovl_u8(t4)); - s5 = vreinterpretq_s16_u16(vmovl_u8(t5)); - s6 = vreinterpretq_s16_u16(vmovl_u8(t6)); - - width = w; - s = src + 7; - d = dst; + + int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0)); + int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t1)); + int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2)); + int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t3)); + int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t4)); + int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t5)); + int16x8_t s6 = vreinterpretq_s16_u16(vmovl_u8(t6)); + + s += 7; do { load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7); transpose_elems_inplace_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7); - s7 = vreinterpretq_s16_u16(vmovl_u8(t0)); - s8 = vreinterpretq_s16_u16(vmovl_u8(t1)); - s9 = vreinterpretq_s16_u16(vmovl_u8(t2)); - s10 = vreinterpretq_s16_u16(vmovl_u8(t3)); - s11 = vreinterpretq_s16_u16(vmovl_u8(t4)); - s12 = vreinterpretq_s16_u16(vmovl_u8(t5)); - s13 = vreinterpretq_s16_u16(vmovl_u8(t6)); - s14 = vreinterpretq_s16_u16(vmovl_u8(t7)); - - d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter); - d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter); - d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter); - d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter); - d4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filter); - d5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filter); - d6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filter); - d7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filter); + + int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(t0)); + int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t1)); + int16x8_t s9 = vreinterpretq_s16_u16(vmovl_u8(t2)); + int16x8_t s10 = vreinterpretq_s16_u16(vmovl_u8(t3)); + int16x8_t s11 = vreinterpretq_s16_u16(vmovl_u8(t4)); + int16x8_t s12 = vreinterpretq_s16_u16(vmovl_u8(t5)); + int16x8_t s13 = vreinterpretq_s16_u16(vmovl_u8(t6)); + int16x8_t s14 = vreinterpretq_s16_u16(vmovl_u8(t7)); + + uint8x8_t d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter); + uint8x8_t d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter); + uint8x8_t d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter); + uint8x8_t d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter); + uint8x8_t d4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filter); + uint8x8_t d5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filter); + uint8x8_t d6 = + convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filter); + uint8x8_t d7 = + convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filter); transpose_elems_inplace_u8_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7); @@ -224,6 +218,7 @@ void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, s4 = s12; s5 = s13; s6 = s14; + s += 8; d += 8; width -= 8; @@ -236,6 +231,137 @@ void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, } } +static INLINE int16x4_t convolve4_4(const int16x4_t s0, const int16x4_t s1, + const int16x4_t s2, const int16x4_t s3, + const int16x4_t filter) { + int16x4_t sum = vmul_lane_s16(s0, filter, 0); + sum = vmla_lane_s16(sum, s1, filter, 1); + sum = vmla_lane_s16(sum, s2, filter, 2); + sum = vmla_lane_s16(sum, s3, filter, 3); + + return sum; +} + +static INLINE uint8x8_t convolve4_8(const int16x8_t s0, const int16x8_t s1, + const int16x8_t s2, const int16x8_t s3, + const int16x4_t filter) { + int16x8_t sum = vmulq_lane_s16(s0, filter, 0); + sum = vmlaq_lane_s16(sum, s1, filter, 1); + sum = vmlaq_lane_s16(sum, s2, filter, 2); + sum = vmlaq_lane_s16(sum, s3, filter, 3); + + // We halved the filter values so -1 from right shift. + return vqrshrun_n_s16(sum, FILTER_BITS - 1); +} + +static INLINE void convolve8_horiz_4tap_neon(const uint8_t *src, + ptrdiff_t src_stride, uint8_t *dst, + ptrdiff_t dst_stride, + const int16_t *filter_x, int w, + int h) { + // All filter values are even, halve to reduce intermediate precision + // requirements. + const int16x4_t filter = vshr_n_s16(vld1_s16(filter_x + 2), 1); + + if (w == 4) { + do { + int16x8_t t0 = + vreinterpretq_s16_u16(vmovl_u8(vld1_u8(src + 0 * src_stride))); + int16x8_t t1 = + vreinterpretq_s16_u16(vmovl_u8(vld1_u8(src + 1 * src_stride))); + + int16x4_t s0[4], s1[4]; + s0[0] = vget_low_s16(t0); + s0[1] = vget_low_s16(vextq_s16(t0, t0, 1)); + s0[2] = vget_low_s16(vextq_s16(t0, t0, 2)); + s0[3] = vget_low_s16(vextq_s16(t0, t0, 3)); + + s1[0] = vget_low_s16(t1); + s1[1] = vget_low_s16(vextq_s16(t1, t1, 1)); + s1[2] = vget_low_s16(vextq_s16(t1, t1, 2)); + s1[3] = vget_low_s16(vextq_s16(t1, t1, 3)); + + int16x4_t d0 = convolve4_4(s0[0], s0[1], s0[2], s0[3], filter); + int16x4_t d1 = convolve4_4(s1[0], s1[1], s1[2], s1[3], filter); + // We halved the filter values so -1 from right shift. + uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS - 1); + + store_u8x4_strided_x2(dst + 0 * dst_stride, dst_stride, d01); + + src += 2 * src_stride; + dst += 2 * dst_stride; + h -= 2; + } while (h > 0); + } else { + do { + int width = w; + const uint8_t *s = src; + uint8_t *d = dst; + + int16x8_t t0 = + vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s + 0 * src_stride))); + int16x8_t t1 = + vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s + 1 * src_stride))); + + s += 8; + do { + int16x8_t t2 = + vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s + 0 * src_stride))); + int16x8_t t3 = + vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s + 1 * src_stride))); + + int16x8_t s0[4], s1[4]; + s0[0] = t0; + s0[1] = vextq_s16(t0, t2, 1); + s0[2] = vextq_s16(t0, t2, 2); + s0[3] = vextq_s16(t0, t2, 3); + + s1[0] = t1; + s1[1] = vextq_s16(t1, t3, 1); + s1[2] = vextq_s16(t1, t3, 2); + s1[3] = vextq_s16(t1, t3, 3); + + uint8x8_t d0 = convolve4_8(s0[0], s0[1], s0[2], s0[3], filter); + uint8x8_t d1 = convolve4_8(s1[0], s1[1], s1[2], s1[3], filter); + + store_u8_8x2(d, dst_stride, d0, d1); + + t0 = t2; + t1 = t3; + + s += 8; + d += 8; + width -= 8; + } while (width != 0); + src += 2 * src_stride; + dst += 2 * dst_stride; + h -= 2; + } while (h > 0); + } +} + +void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, int w, + int h) { + assert((intptr_t)dst % 4 == 0); + assert(dst_stride % 4 == 0); + + (void)x_step_q4; + (void)filter_y; + (void)y_step_q4; + + src -= ((SUBPEL_TAPS / 2) - 1); + + if (get_filter_taps_convolve8(filter_x) <= 4) { + convolve8_horiz_4tap_neon(src + 2, src_stride, dst, dst_stride, filter_x, w, + h); + } else { + convolve8_horiz_8tap_neon(src, src_stride, dst, dst_stride, filter_x, w, h); + } +} + void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, @@ -253,33 +379,33 @@ void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride, src -= ((SUBPEL_TAPS / 2) - 1) * src_stride; if (w == 4) { - uint8x8_t t0, t1, t2, t3, t4, t5, t6, d01, d23; - int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3; - + uint8x8_t t0, t1, t2, t3, t4, t5, t6; load_u8_8x7(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6); - s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0))); - s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1))); - s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2))); - s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3))); - s4 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t4))); - s5 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t5))); - s6 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t6))); + + int16x4_t s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0))); + int16x4_t s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1))); + int16x4_t s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2))); + int16x4_t s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3))); + int16x4_t s4 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t4))); + int16x4_t s5 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t5))); + int16x4_t s6 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t6))); src += 7 * src_stride; do { load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3); - s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0))); - s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1))); - s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2))); - s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3))); - - d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filter); - d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filter); - d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filter); - d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filter); - d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS); - d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS); + + int16x4_t s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0))); + int16x4_t s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1))); + int16x4_t s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2))); + int16x4_t s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3))); + + int16x4_t d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filter); + int16x4_t d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filter); + int16x4_t d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filter); + int16x4_t d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filter); + uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS); + uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS); store_u8x4_strided_x2(dst + 0 * dst_stride, dst_stride, d01); store_u8x4_strided_x2(dst + 2 * dst_stride, dst_stride, d23); @@ -291,42 +417,40 @@ void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride, s4 = s8; s5 = s9; s6 = s10; + src += 4 * src_stride; dst += 4 * dst_stride; h -= 4; } while (h != 0); } else { - uint8x8_t t0, t1, t2, t3, t4, t5, t6, d0, d1, d2, d3; - int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10; - int height; - const uint8_t *s; - uint8_t *d; - do { + uint8x8_t t0, t1, t2, t3, t4, t5, t6; load_u8_8x7(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6); - s0 = vreinterpretq_s16_u16(vmovl_u8(t0)); - s1 = vreinterpretq_s16_u16(vmovl_u8(t1)); - s2 = vreinterpretq_s16_u16(vmovl_u8(t2)); - s3 = vreinterpretq_s16_u16(vmovl_u8(t3)); - s4 = vreinterpretq_s16_u16(vmovl_u8(t4)); - s5 = vreinterpretq_s16_u16(vmovl_u8(t5)); - s6 = vreinterpretq_s16_u16(vmovl_u8(t6)); - - height = h; - s = src + 7 * src_stride; - d = dst; + + int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0)); + int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t1)); + int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2)); + int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t3)); + int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t4)); + int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t5)); + int16x8_t s6 = vreinterpretq_s16_u16(vmovl_u8(t6)); + + int height = h; + const uint8_t *s = src + 7 * src_stride; + uint8_t *d = dst; do { load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3); - s7 = vreinterpretq_s16_u16(vmovl_u8(t0)); - s8 = vreinterpretq_s16_u16(vmovl_u8(t1)); - s9 = vreinterpretq_s16_u16(vmovl_u8(t2)); - s10 = vreinterpretq_s16_u16(vmovl_u8(t3)); - d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter); - d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter); - d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter); - d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter); + int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(t0)); + int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t1)); + int16x8_t s9 = vreinterpretq_s16_u16(vmovl_u8(t2)); + int16x8_t s10 = vreinterpretq_s16_u16(vmovl_u8(t3)); + + uint8x8_t d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter); + uint8x8_t d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter); + uint8x8_t d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter); + uint8x8_t d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter); store_u8_8x4(d, dst_stride, d0, d1, d2, d3); @@ -337,6 +461,7 @@ void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride, s4 = s8; s5 = s9; s6 = s10; + s += 4 * src_stride; d += 4 * dst_stride; height -= 4; diff --git a/third_party/aom/aom_dsp/arm/aom_convolve8_neon_dotprod.c b/third_party/aom/aom_dsp/arm/aom_convolve8_neon_dotprod.c index c82125ba17..120c479798 100644 --- a/third_party/aom/aom_dsp/arm/aom_convolve8_neon_dotprod.c +++ b/third_party/aom/aom_dsp/arm/aom_convolve8_neon_dotprod.c @@ -24,81 +24,72 @@ #include "aom_dsp/arm/transpose_neon.h" #include "aom_ports/mem.h" -DECLARE_ALIGNED(16, static const uint8_t, dot_prod_permute_tbl[48]) = { +// Filter values always sum to 128. +#define FILTER_WEIGHT 128 + +DECLARE_ALIGNED(16, static const uint8_t, kDotProdPermuteTbl[48]) = { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10, 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 }; -DECLARE_ALIGNED(16, static const uint8_t, dot_prod_tran_concat_tbl[32]) = { - 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27, - 4, 12, 20, 28, 5, 13, 21, 29, 6, 14, 22, 30, 7, 15, 23, 31 -}; - -DECLARE_ALIGNED(16, static const uint8_t, dot_prod_merge_block_tbl[48]) = { - /* Shift left and insert new last column in transposed 4x4 block. */ +DECLARE_ALIGNED(16, static const uint8_t, kDotProdMergeBlockTbl[48]) = { + // Shift left and insert new last column in transposed 4x4 block. 1, 2, 3, 16, 5, 6, 7, 20, 9, 10, 11, 24, 13, 14, 15, 28, - /* Shift left and insert two new columns in transposed 4x4 block. */ + // Shift left and insert two new columns in transposed 4x4 block. 2, 3, 16, 17, 6, 7, 20, 21, 10, 11, 24, 25, 14, 15, 28, 29, - /* Shift left and insert three new columns in transposed 4x4 block. */ + // Shift left and insert three new columns in transposed 4x4 block. 3, 16, 17, 18, 7, 20, 21, 22, 11, 24, 25, 26, 15, 28, 29, 30 }; -static INLINE int16x4_t convolve8_4_sdot(uint8x16_t samples, - const int8x8_t filter, - const int32x4_t correction, - const uint8x16_t range_limit, - const uint8x16x2_t permute_tbl) { - int8x16_t clamped_samples, permuted_samples[2]; - int32x4_t sum; - - /* Clamp sample range to [-128, 127] for 8-bit signed dot product. */ - clamped_samples = vreinterpretq_s8_u8(vsubq_u8(samples, range_limit)); - - /* Permute samples ready for dot product. */ - /* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */ - permuted_samples[0] = vqtbl1q_s8(clamped_samples, permute_tbl.val[0]); - /* { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } */ - permuted_samples[1] = vqtbl1q_s8(clamped_samples, permute_tbl.val[1]); - - /* Accumulate dot product into 'correction' to account for range clamp. */ - sum = vdotq_lane_s32(correction, permuted_samples[0], filter, 0); - sum = vdotq_lane_s32(sum, permuted_samples[1], filter, 1); - - /* Further narrowing and packing is performed by the caller. */ +static INLINE int16x4_t convolve8_4_h(const uint8x16_t samples, + const int8x8_t filters, + const uint8x16x2_t permute_tbl) { + // Transform sample range to [-128, 127] for 8-bit signed dot product. + int8x16_t samples_128 = + vreinterpretq_s8_u8(vsubq_u8(samples, vdupq_n_u8(128))); + + // Permute samples ready for dot product. + // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } + // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } + int8x16_t perm_samples[2] = { vqtbl1q_s8(samples_128, permute_tbl.val[0]), + vqtbl1q_s8(samples_128, permute_tbl.val[1]) }; + + // Accumulate into 128 * FILTER_WEIGHT to account for range transform. + int32x4_t acc = vdupq_n_s32(128 * FILTER_WEIGHT); + int32x4_t sum = vdotq_lane_s32(acc, perm_samples[0], filters, 0); + sum = vdotq_lane_s32(sum, perm_samples[1], filters, 1); + + // Further narrowing and packing is performed by the caller. return vqmovn_s32(sum); } -static INLINE uint8x8_t convolve8_8_sdot(uint8x16_t samples, - const int8x8_t filter, - const int32x4_t correction, - const uint8x16_t range_limit, - const uint8x16x3_t permute_tbl) { - int8x16_t clamped_samples, permuted_samples[3]; - int32x4_t sum0, sum1; - int16x8_t sum; - - /* Clamp sample range to [-128, 127] for 8-bit signed dot product. */ - clamped_samples = vreinterpretq_s8_u8(vsubq_u8(samples, range_limit)); - - /* Permute samples ready for dot product. */ - /* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */ - permuted_samples[0] = vqtbl1q_s8(clamped_samples, permute_tbl.val[0]); - /* { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } */ - permuted_samples[1] = vqtbl1q_s8(clamped_samples, permute_tbl.val[1]); - /* { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 } */ - permuted_samples[2] = vqtbl1q_s8(clamped_samples, permute_tbl.val[2]); - - /* Accumulate dot product into 'correction' to account for range clamp. */ - /* First 4 output values. */ - sum0 = vdotq_lane_s32(correction, permuted_samples[0], filter, 0); - sum0 = vdotq_lane_s32(sum0, permuted_samples[1], filter, 1); - /* Second 4 output values. */ - sum1 = vdotq_lane_s32(correction, permuted_samples[1], filter, 0); - sum1 = vdotq_lane_s32(sum1, permuted_samples[2], filter, 1); - - /* Narrow and re-pack. */ - sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1)); +static INLINE uint8x8_t convolve8_8_h(const uint8x16_t samples, + const int8x8_t filters, + const uint8x16x3_t permute_tbl) { + // Transform sample range to [-128, 127] for 8-bit signed dot product. + int8x16_t samples_128 = + vreinterpretq_s8_u8(vsubq_u8(samples, vdupq_n_u8(128))); + + // Permute samples ready for dot product. + // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } + // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } + // { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 } + int8x16_t perm_samples[3] = { vqtbl1q_s8(samples_128, permute_tbl.val[0]), + vqtbl1q_s8(samples_128, permute_tbl.val[1]), + vqtbl1q_s8(samples_128, permute_tbl.val[2]) }; + + // Accumulate into 128 * FILTER_WEIGHT to account for range transform. + int32x4_t acc = vdupq_n_s32(128 * FILTER_WEIGHT); + // First 4 output values. + int32x4_t sum0 = vdotq_lane_s32(acc, perm_samples[0], filters, 0); + sum0 = vdotq_lane_s32(sum0, perm_samples[1], filters, 1); + // Second 4 output values. + int32x4_t sum1 = vdotq_lane_s32(acc, perm_samples[1], filters, 0); + sum1 = vdotq_lane_s32(sum1, perm_samples[2], filters, 1); + + // Narrow and re-pack. + int16x8_t sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1)); return vqrshrun_n_s16(sum, FILTER_BITS); } @@ -108,10 +99,6 @@ void aom_convolve8_horiz_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, const int16_t *filter_y, int y_step_q4, int w, int h) { const int8x8_t filter = vmovn_s16(vld1q_s16(filter_x)); - const int16x8_t correct_tmp = vmulq_n_s16(vld1q_s16(filter_x), 128); - const int32x4_t correction = vdupq_n_s32((int32_t)vaddvq_s16(correct_tmp)); - const uint8x16_t range_limit = vdupq_n_u8(128); - uint8x16_t s0, s1, s2, s3; assert((intptr_t)dst % 4 == 0); assert(dst_stride % 4 == 0); @@ -123,19 +110,17 @@ void aom_convolve8_horiz_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, src -= ((SUBPEL_TAPS / 2) - 1); if (w == 4) { - const uint8x16x2_t perm_tbl = vld1q_u8_x2(dot_prod_permute_tbl); + const uint8x16x2_t perm_tbl = vld1q_u8_x2(kDotProdPermuteTbl); do { - int16x4_t t0, t1, t2, t3; - uint8x8_t d01, d23; - + uint8x16_t s0, s1, s2, s3; load_u8_16x4(src, src_stride, &s0, &s1, &s2, &s3); - t0 = convolve8_4_sdot(s0, filter, correction, range_limit, perm_tbl); - t1 = convolve8_4_sdot(s1, filter, correction, range_limit, perm_tbl); - t2 = convolve8_4_sdot(s2, filter, correction, range_limit, perm_tbl); - t3 = convolve8_4_sdot(s3, filter, correction, range_limit, perm_tbl); - d01 = vqrshrun_n_s16(vcombine_s16(t0, t1), FILTER_BITS); - d23 = vqrshrun_n_s16(vcombine_s16(t2, t3), FILTER_BITS); + int16x4_t d0 = convolve8_4_h(s0, filter, perm_tbl); + int16x4_t d1 = convolve8_4_h(s1, filter, perm_tbl); + int16x4_t d2 = convolve8_4_h(s2, filter, perm_tbl); + int16x4_t d3 = convolve8_4_h(s3, filter, perm_tbl); + uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS); + uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS); store_u8x4_strided_x2(dst + 0 * dst_stride, dst_stride, d01); store_u8x4_strided_x2(dst + 2 * dst_stride, dst_stride, d23); @@ -145,23 +130,20 @@ void aom_convolve8_horiz_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, h -= 4; } while (h > 0); } else { - const uint8x16x3_t perm_tbl = vld1q_u8_x3(dot_prod_permute_tbl); - const uint8_t *s; - uint8_t *d; - int width; - uint8x8_t d0, d1, d2, d3; + const uint8x16x3_t perm_tbl = vld1q_u8_x3(kDotProdPermuteTbl); do { - width = w; - s = src; - d = dst; + int width = w; + const uint8_t *s = src; + uint8_t *d = dst; do { + uint8x16_t s0, s1, s2, s3; load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3); - d0 = convolve8_8_sdot(s0, filter, correction, range_limit, perm_tbl); - d1 = convolve8_8_sdot(s1, filter, correction, range_limit, perm_tbl); - d2 = convolve8_8_sdot(s2, filter, correction, range_limit, perm_tbl); - d3 = convolve8_8_sdot(s3, filter, correction, range_limit, perm_tbl); + uint8x8_t d0 = convolve8_8_h(s0, filter, perm_tbl); + uint8x8_t d1 = convolve8_8_h(s1, filter, perm_tbl); + uint8x8_t d2 = convolve8_8_h(s2, filter, perm_tbl); + uint8x8_t d3 = convolve8_8_h(s3, filter, perm_tbl); store_u8_8x4(d, dst_stride, d0, d1, d2, d3); @@ -177,83 +159,88 @@ void aom_convolve8_horiz_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, } static INLINE void transpose_concat_4x4(int8x8_t a0, int8x8_t a1, int8x8_t a2, - int8x8_t a3, int8x16_t *b, - const uint8x16_t permute_tbl) { - /* Transpose 8-bit elements and concatenate result rows as follows: - * a0: 00, 01, 02, 03, XX, XX, XX, XX - * a1: 10, 11, 12, 13, XX, XX, XX, XX - * a2: 20, 21, 22, 23, XX, XX, XX, XX - * a3: 30, 31, 32, 33, XX, XX, XX, XX - * - * b: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33 - * - * The 'permute_tbl' is always 'dot_prod_tran_concat_tbl' above. Passing it - * as an argument is preferable to loading it directly from memory as this - * inline helper is called many times from the same parent function. - */ - - int8x16x2_t samples = { { vcombine_s8(a0, a1), vcombine_s8(a2, a3) } }; - *b = vqtbl2q_s8(samples, permute_tbl); + int8x8_t a3, int8x16_t *b) { + // Transpose 8-bit elements and concatenate result rows as follows: + // a0: 00, 01, 02, 03, XX, XX, XX, XX + // a1: 10, 11, 12, 13, XX, XX, XX, XX + // a2: 20, 21, 22, 23, XX, XX, XX, XX + // a3: 30, 31, 32, 33, XX, XX, XX, XX + // + // b: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33 + + int8x16_t a0q = vcombine_s8(a0, vdup_n_s8(0)); + int8x16_t a1q = vcombine_s8(a1, vdup_n_s8(0)); + int8x16_t a2q = vcombine_s8(a2, vdup_n_s8(0)); + int8x16_t a3q = vcombine_s8(a3, vdup_n_s8(0)); + + int8x16_t a01 = vzipq_s8(a0q, a1q).val[0]; + int8x16_t a23 = vzipq_s8(a2q, a3q).val[0]; + + int16x8_t a0123 = + vzipq_s16(vreinterpretq_s16_s8(a01), vreinterpretq_s16_s8(a23)).val[0]; + + *b = vreinterpretq_s8_s16(a0123); } static INLINE void transpose_concat_8x4(int8x8_t a0, int8x8_t a1, int8x8_t a2, int8x8_t a3, int8x16_t *b0, - int8x16_t *b1, - const uint8x16x2_t permute_tbl) { - /* Transpose 8-bit elements and concatenate result rows as follows: - * a0: 00, 01, 02, 03, 04, 05, 06, 07 - * a1: 10, 11, 12, 13, 14, 15, 16, 17 - * a2: 20, 21, 22, 23, 24, 25, 26, 27 - * a3: 30, 31, 32, 33, 34, 35, 36, 37 - * - * b0: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33 - * b1: 04, 14, 24, 34, 05, 15, 25, 35, 06, 16, 26, 36, 07, 17, 27, 37 - * - * The 'permute_tbl' is always 'dot_prod_tran_concat_tbl' above. Passing it - * as an argument is preferable to loading it directly from memory as this - * inline helper is called many times from the same parent function. - */ - - int8x16x2_t samples = { { vcombine_s8(a0, a1), vcombine_s8(a2, a3) } }; - *b0 = vqtbl2q_s8(samples, permute_tbl.val[0]); - *b1 = vqtbl2q_s8(samples, permute_tbl.val[1]); + int8x16_t *b1) { + // Transpose 8-bit elements and concatenate result rows as follows: + // a0: 00, 01, 02, 03, 04, 05, 06, 07 + // a1: 10, 11, 12, 13, 14, 15, 16, 17 + // a2: 20, 21, 22, 23, 24, 25, 26, 27 + // a3: 30, 31, 32, 33, 34, 35, 36, 37 + // + // b0: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33 + // b1: 04, 14, 24, 34, 05, 15, 25, 35, 06, 16, 26, 36, 07, 17, 27, 37 + + int8x16_t a0q = vcombine_s8(a0, vdup_n_s8(0)); + int8x16_t a1q = vcombine_s8(a1, vdup_n_s8(0)); + int8x16_t a2q = vcombine_s8(a2, vdup_n_s8(0)); + int8x16_t a3q = vcombine_s8(a3, vdup_n_s8(0)); + + int8x16_t a01 = vzipq_s8(a0q, a1q).val[0]; + int8x16_t a23 = vzipq_s8(a2q, a3q).val[0]; + + int16x8x2_t a0123 = + vzipq_s16(vreinterpretq_s16_s8(a01), vreinterpretq_s16_s8(a23)); + + *b0 = vreinterpretq_s8_s16(a0123.val[0]); + *b1 = vreinterpretq_s8_s16(a0123.val[1]); } -static INLINE int16x4_t convolve8_4_sdot_partial(const int8x16_t samples_lo, - const int8x16_t samples_hi, - const int32x4_t correction, - const int8x8_t filter) { - /* Sample range-clamping and permutation are performed by the caller. */ - int32x4_t sum; +static INLINE int16x4_t convolve8_4_v(const int8x16_t samples_lo, + const int8x16_t samples_hi, + const int8x8_t filters) { + // The sample range transform and permutation are performed by the caller. - /* Accumulate dot product into 'correction' to account for range clamp. */ - sum = vdotq_lane_s32(correction, samples_lo, filter, 0); - sum = vdotq_lane_s32(sum, samples_hi, filter, 1); + // Accumulate into 128 * FILTER_WEIGHT to account for range transform. + int32x4_t acc = vdupq_n_s32(128 * FILTER_WEIGHT); + int32x4_t sum = vdotq_lane_s32(acc, samples_lo, filters, 0); + sum = vdotq_lane_s32(sum, samples_hi, filters, 1); - /* Further narrowing and packing is performed by the caller. */ + // Further narrowing and packing is performed by the caller. return vqmovn_s32(sum); } -static INLINE uint8x8_t convolve8_8_sdot_partial(const int8x16_t samples0_lo, - const int8x16_t samples0_hi, - const int8x16_t samples1_lo, - const int8x16_t samples1_hi, - const int32x4_t correction, - const int8x8_t filter) { - /* Sample range-clamping and permutation are performed by the caller. */ - int32x4_t sum0, sum1; - int16x8_t sum; - - /* Accumulate dot product into 'correction' to account for range clamp. */ - /* First 4 output values. */ - sum0 = vdotq_lane_s32(correction, samples0_lo, filter, 0); - sum0 = vdotq_lane_s32(sum0, samples0_hi, filter, 1); - /* Second 4 output values. */ - sum1 = vdotq_lane_s32(correction, samples1_lo, filter, 0); - sum1 = vdotq_lane_s32(sum1, samples1_hi, filter, 1); - - /* Narrow and re-pack. */ - sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1)); +static INLINE uint8x8_t convolve8_8_v(const int8x16_t samples0_lo, + const int8x16_t samples0_hi, + const int8x16_t samples1_lo, + const int8x16_t samples1_hi, + const int8x8_t filters) { + // The sample range transform and permutation are performed by the caller. + + // Accumulate into 128 * FILTER_WEIGHT to account for range transform. + int32x4_t acc = vdupq_n_s32(128 * FILTER_WEIGHT); + // First 4 output values. + int32x4_t sum0 = vdotq_lane_s32(acc, samples0_lo, filters, 0); + sum0 = vdotq_lane_s32(sum0, samples0_hi, filters, 1); + // Second 4 output values. + int32x4_t sum1 = vdotq_lane_s32(acc, samples1_lo, filters, 0); + sum1 = vdotq_lane_s32(sum1, samples1_hi, filters, 1); + + // Narrow and re-pack. + int16x8_t sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1)); return vqrshrun_n_s16(sum, FILTER_BITS); } @@ -263,10 +250,7 @@ void aom_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, const int16_t *filter_y, int y_step_q4, int w, int h) { const int8x8_t filter = vmovn_s16(vld1q_s16(filter_y)); - const int16x8_t correct_tmp = vmulq_n_s16(vld1q_s16(filter_y), 128); - const int32x4_t correction = vdupq_n_s32((int32_t)vaddvq_s16(correct_tmp)); - const uint8x8_t range_limit = vdup_n_u8(128); - const uint8x16x3_t merge_block_tbl = vld1q_u8_x3(dot_prod_merge_block_tbl); + const uint8x16x3_t merge_block_tbl = vld1q_u8_x3(kDotProdMergeBlockTbl); int8x16x2_t samples_LUT; assert((intptr_t)dst % 4 == 0); @@ -279,62 +263,58 @@ void aom_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, src -= ((SUBPEL_TAPS / 2) - 1) * src_stride; if (w == 4) { - const uint8x16_t tran_concat_tbl = vld1q_u8(dot_prod_tran_concat_tbl); - uint8x8_t t0, t1, t2, t3, t4, t5, t6; load_u8_8x7(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6); src += 7 * src_stride; - /* Clamp sample range to [-128, 127] for 8-bit signed dot product. */ - int8x8_t s0 = vreinterpret_s8_u8(vsub_u8(t0, range_limit)); - int8x8_t s1 = vreinterpret_s8_u8(vsub_u8(t1, range_limit)); - int8x8_t s2 = vreinterpret_s8_u8(vsub_u8(t2, range_limit)); - int8x8_t s3 = vreinterpret_s8_u8(vsub_u8(t3, range_limit)); - int8x8_t s4 = vreinterpret_s8_u8(vsub_u8(t4, range_limit)); - int8x8_t s5 = vreinterpret_s8_u8(vsub_u8(t5, range_limit)); - int8x8_t s6 = vreinterpret_s8_u8(vsub_u8(t6, range_limit)); - - /* This operation combines a conventional transpose and the sample permute - * (see horizontal case) required before computing the dot product. - */ + // Clamp sample range to [-128, 127] for 8-bit signed dot product. + int8x8_t s0 = vreinterpret_s8_u8(vsub_u8(t0, vdup_n_u8(128))); + int8x8_t s1 = vreinterpret_s8_u8(vsub_u8(t1, vdup_n_u8(128))); + int8x8_t s2 = vreinterpret_s8_u8(vsub_u8(t2, vdup_n_u8(128))); + int8x8_t s3 = vreinterpret_s8_u8(vsub_u8(t3, vdup_n_u8(128))); + int8x8_t s4 = vreinterpret_s8_u8(vsub_u8(t4, vdup_n_u8(128))); + int8x8_t s5 = vreinterpret_s8_u8(vsub_u8(t5, vdup_n_u8(128))); + int8x8_t s6 = vreinterpret_s8_u8(vsub_u8(t6, vdup_n_u8(128))); + + // This operation combines a conventional transpose and the sample permute + // (see horizontal case) required before computing the dot product. int8x16_t s0123, s1234, s2345, s3456; - transpose_concat_4x4(s0, s1, s2, s3, &s0123, tran_concat_tbl); - transpose_concat_4x4(s1, s2, s3, s4, &s1234, tran_concat_tbl); - transpose_concat_4x4(s2, s3, s4, s5, &s2345, tran_concat_tbl); - transpose_concat_4x4(s3, s4, s5, s6, &s3456, tran_concat_tbl); + transpose_concat_4x4(s0, s1, s2, s3, &s0123); + transpose_concat_4x4(s1, s2, s3, s4, &s1234); + transpose_concat_4x4(s2, s3, s4, s5, &s2345); + transpose_concat_4x4(s3, s4, s5, s6, &s3456); do { uint8x8_t t7, t8, t9, t10; load_u8_8x4(src, src_stride, &t7, &t8, &t9, &t10); - int8x8_t s7 = vreinterpret_s8_u8(vsub_u8(t7, range_limit)); - int8x8_t s8 = vreinterpret_s8_u8(vsub_u8(t8, range_limit)); - int8x8_t s9 = vreinterpret_s8_u8(vsub_u8(t9, range_limit)); - int8x8_t s10 = vreinterpret_s8_u8(vsub_u8(t10, range_limit)); + int8x8_t s7 = vreinterpret_s8_u8(vsub_u8(t7, vdup_n_u8(128))); + int8x8_t s8 = vreinterpret_s8_u8(vsub_u8(t8, vdup_n_u8(128))); + int8x8_t s9 = vreinterpret_s8_u8(vsub_u8(t9, vdup_n_u8(128))); + int8x8_t s10 = vreinterpret_s8_u8(vsub_u8(t10, vdup_n_u8(128))); int8x16_t s4567, s5678, s6789, s78910; - transpose_concat_4x4(s7, s8, s9, s10, &s78910, tran_concat_tbl); + transpose_concat_4x4(s7, s8, s9, s10, &s78910); - /* Merge new data into block from previous iteration. */ + // Merge new data into block from previous iteration. samples_LUT.val[0] = s3456; samples_LUT.val[1] = s78910; s4567 = vqtbl2q_s8(samples_LUT, merge_block_tbl.val[0]); s5678 = vqtbl2q_s8(samples_LUT, merge_block_tbl.val[1]); s6789 = vqtbl2q_s8(samples_LUT, merge_block_tbl.val[2]); - int16x4_t d0 = convolve8_4_sdot_partial(s0123, s4567, correction, filter); - int16x4_t d1 = convolve8_4_sdot_partial(s1234, s5678, correction, filter); - int16x4_t d2 = convolve8_4_sdot_partial(s2345, s6789, correction, filter); - int16x4_t d3 = - convolve8_4_sdot_partial(s3456, s78910, correction, filter); + int16x4_t d0 = convolve8_4_v(s0123, s4567, filter); + int16x4_t d1 = convolve8_4_v(s1234, s5678, filter); + int16x4_t d2 = convolve8_4_v(s2345, s6789, filter); + int16x4_t d3 = convolve8_4_v(s3456, s78910, filter); uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS); uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS); store_u8x4_strided_x2(dst + 0 * dst_stride, dst_stride, d01); store_u8x4_strided_x2(dst + 2 * dst_stride, dst_stride, d23); - /* Prepare block for next iteration - re-using as much as possible. */ - /* Shuffle everything up four rows. */ + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. s0123 = s4567; s1234 = s5678; s2345 = s6789; @@ -345,8 +325,6 @@ void aom_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, h -= 4; } while (h != 0); } else { - const uint8x16x2_t tran_concat_tbl = vld1q_u8_x2(dot_prod_tran_concat_tbl); - do { int height = h; const uint8_t *s = src; @@ -356,44 +334,38 @@ void aom_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, load_u8_8x7(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6); s += 7 * src_stride; - /* Clamp sample range to [-128, 127] for 8-bit signed dot product. */ - int8x8_t s0 = vreinterpret_s8_u8(vsub_u8(t0, range_limit)); - int8x8_t s1 = vreinterpret_s8_u8(vsub_u8(t1, range_limit)); - int8x8_t s2 = vreinterpret_s8_u8(vsub_u8(t2, range_limit)); - int8x8_t s3 = vreinterpret_s8_u8(vsub_u8(t3, range_limit)); - int8x8_t s4 = vreinterpret_s8_u8(vsub_u8(t4, range_limit)); - int8x8_t s5 = vreinterpret_s8_u8(vsub_u8(t5, range_limit)); - int8x8_t s6 = vreinterpret_s8_u8(vsub_u8(t6, range_limit)); - - /* This operation combines a conventional transpose and the sample permute - * (see horizontal case) required before computing the dot product. - */ + // Clamp sample range to [-128, 127] for 8-bit signed dot product. + int8x8_t s0 = vreinterpret_s8_u8(vsub_u8(t0, vdup_n_u8(128))); + int8x8_t s1 = vreinterpret_s8_u8(vsub_u8(t1, vdup_n_u8(128))); + int8x8_t s2 = vreinterpret_s8_u8(vsub_u8(t2, vdup_n_u8(128))); + int8x8_t s3 = vreinterpret_s8_u8(vsub_u8(t3, vdup_n_u8(128))); + int8x8_t s4 = vreinterpret_s8_u8(vsub_u8(t4, vdup_n_u8(128))); + int8x8_t s5 = vreinterpret_s8_u8(vsub_u8(t5, vdup_n_u8(128))); + int8x8_t s6 = vreinterpret_s8_u8(vsub_u8(t6, vdup_n_u8(128))); + + // This operation combines a conventional transpose and the sample permute + // (see horizontal case) required before computing the dot product. int8x16_t s0123_lo, s0123_hi, s1234_lo, s1234_hi, s2345_lo, s2345_hi, s3456_lo, s3456_hi; - transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi, - tran_concat_tbl); - transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi, - tran_concat_tbl); - transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi, - tran_concat_tbl); - transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi, - tran_concat_tbl); + transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi); + transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi); + transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi); + transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi); do { uint8x8_t t7, t8, t9, t10; load_u8_8x4(s, src_stride, &t7, &t8, &t9, &t10); - int8x8_t s7 = vreinterpret_s8_u8(vsub_u8(t7, range_limit)); - int8x8_t s8 = vreinterpret_s8_u8(vsub_u8(t8, range_limit)); - int8x8_t s9 = vreinterpret_s8_u8(vsub_u8(t9, range_limit)); - int8x8_t s10 = vreinterpret_s8_u8(vsub_u8(t10, range_limit)); + int8x8_t s7 = vreinterpret_s8_u8(vsub_u8(t7, vdup_n_u8(128))); + int8x8_t s8 = vreinterpret_s8_u8(vsub_u8(t8, vdup_n_u8(128))); + int8x8_t s9 = vreinterpret_s8_u8(vsub_u8(t9, vdup_n_u8(128))); + int8x8_t s10 = vreinterpret_s8_u8(vsub_u8(t10, vdup_n_u8(128))); int8x16_t s4567_lo, s4567_hi, s5678_lo, s5678_hi, s6789_lo, s6789_hi, s78910_lo, s78910_hi; - transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi, - tran_concat_tbl); + transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi); - /* Merge new data into block from previous iteration. */ + // Merge new data into block from previous iteration. samples_LUT.val[0] = s3456_lo; samples_LUT.val[1] = s78910_lo; s4567_lo = vqtbl2q_s8(samples_LUT, merge_block_tbl.val[0]); @@ -406,19 +378,19 @@ void aom_convolve8_vert_neon_dotprod(const uint8_t *src, ptrdiff_t src_stride, s5678_hi = vqtbl2q_s8(samples_LUT, merge_block_tbl.val[1]); s6789_hi = vqtbl2q_s8(samples_LUT, merge_block_tbl.val[2]); - uint8x8_t d0 = convolve8_8_sdot_partial(s0123_lo, s4567_lo, s0123_hi, - s4567_hi, correction, filter); - uint8x8_t d1 = convolve8_8_sdot_partial(s1234_lo, s5678_lo, s1234_hi, - s5678_hi, correction, filter); - uint8x8_t d2 = convolve8_8_sdot_partial(s2345_lo, s6789_lo, s2345_hi, - s6789_hi, correction, filter); - uint8x8_t d3 = convolve8_8_sdot_partial(s3456_lo, s78910_lo, s3456_hi, - s78910_hi, correction, filter); + uint8x8_t d0 = + convolve8_8_v(s0123_lo, s4567_lo, s0123_hi, s4567_hi, filter); + uint8x8_t d1 = + convolve8_8_v(s1234_lo, s5678_lo, s1234_hi, s5678_hi, filter); + uint8x8_t d2 = + convolve8_8_v(s2345_lo, s6789_lo, s2345_hi, s6789_hi, filter); + uint8x8_t d3 = + convolve8_8_v(s3456_lo, s78910_lo, s3456_hi, s78910_hi, filter); store_u8_8x4(d, dst_stride, d0, d1, d2, d3); - /* Prepare block for next iteration - re-using as much as possible. */ - /* Shuffle everything up four rows. */ + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. s0123_lo = s4567_lo; s0123_hi = s4567_hi; s1234_lo = s5678_lo; diff --git a/third_party/aom/aom_dsp/arm/aom_convolve8_neon_i8mm.c b/third_party/aom/aom_dsp/arm/aom_convolve8_neon_i8mm.c index df6e4d2ab5..68e031461d 100644 --- a/third_party/aom/aom_dsp/arm/aom_convolve8_neon_i8mm.c +++ b/third_party/aom/aom_dsp/arm/aom_convolve8_neon_i8mm.c @@ -23,69 +23,60 @@ #include "aom_dsp/arm/transpose_neon.h" #include "aom_ports/mem.h" -DECLARE_ALIGNED(16, static const uint8_t, dot_prod_permute_tbl[48]) = { +DECLARE_ALIGNED(16, static const uint8_t, kDotProdPermuteTbl[48]) = { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10, 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 }; -DECLARE_ALIGNED(16, static const uint8_t, dot_prod_tran_concat_tbl[32]) = { - 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27, - 4, 12, 20, 28, 5, 13, 21, 29, 6, 14, 22, 30, 7, 15, 23, 31 -}; - -DECLARE_ALIGNED(16, static const uint8_t, dot_prod_merge_block_tbl[48]) = { - /* Shift left and insert new last column in transposed 4x4 block. */ +DECLARE_ALIGNED(16, static const uint8_t, kDotProdMergeBlockTbl[48]) = { + // Shift left and insert new last column in transposed 4x4 block. 1, 2, 3, 16, 5, 6, 7, 20, 9, 10, 11, 24, 13, 14, 15, 28, - /* Shift left and insert two new columns in transposed 4x4 block. */ + // Shift left and insert two new columns in transposed 4x4 block. 2, 3, 16, 17, 6, 7, 20, 21, 10, 11, 24, 25, 14, 15, 28, 29, - /* Shift left and insert three new columns in transposed 4x4 block. */ + // Shift left and insert three new columns in transposed 4x4 block. 3, 16, 17, 18, 7, 20, 21, 22, 11, 24, 25, 26, 15, 28, 29, 30 }; -static INLINE int16x4_t convolve8_4_usdot(const uint8x16_t samples, - const int8x8_t filter, - const uint8x16x2_t permute_tbl) { - uint8x16_t permuted_samples[2]; - int32x4_t sum; - - /* Permute samples ready for dot product. */ - /* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */ - permuted_samples[0] = vqtbl1q_u8(samples, permute_tbl.val[0]); - /* { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } */ - permuted_samples[1] = vqtbl1q_u8(samples, permute_tbl.val[1]); +static INLINE int16x4_t convolve8_4_h(const uint8x16_t samples, + const int8x8_t filters, + const uint8x16x2_t permute_tbl) { + // Permute samples ready for dot product. + // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } + // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } + uint8x16_t permuted_samples[2] = { vqtbl1q_u8(samples, permute_tbl.val[0]), + vqtbl1q_u8(samples, permute_tbl.val[1]) }; - sum = vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[0], filter, 0); - sum = vusdotq_lane_s32(sum, permuted_samples[1], filter, 1); + int32x4_t sum = + vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[0], filters, 0); + sum = vusdotq_lane_s32(sum, permuted_samples[1], filters, 1); - /* Further narrowing and packing is performed by the caller. */ + // Further narrowing and packing is performed by the caller. return vqmovn_s32(sum); } -static INLINE uint8x8_t convolve8_8_usdot(const uint8x16_t samples, - const int8x8_t filter, - const uint8x16x3_t permute_tbl) { - uint8x16_t permuted_samples[3]; - int32x4_t sum0, sum1; - int16x8_t sum; - - /* Permute samples ready for dot product. */ - /* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */ - permuted_samples[0] = vqtbl1q_u8(samples, permute_tbl.val[0]); - /* { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } */ - permuted_samples[1] = vqtbl1q_u8(samples, permute_tbl.val[1]); - /* { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 } */ - permuted_samples[2] = vqtbl1q_u8(samples, permute_tbl.val[2]); - - /* First 4 output values. */ - sum0 = vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[0], filter, 0); - sum0 = vusdotq_lane_s32(sum0, permuted_samples[1], filter, 1); - /* Second 4 output values. */ - sum1 = vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[1], filter, 0); - sum1 = vusdotq_lane_s32(sum1, permuted_samples[2], filter, 1); - - /* Narrow and re-pack. */ - sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1)); +static INLINE uint8x8_t convolve8_8_h(const uint8x16_t samples, + const int8x8_t filters, + const uint8x16x3_t permute_tbl) { + // Permute samples ready for dot product. + // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } + // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } + // { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 } + uint8x16_t permuted_samples[3] = { vqtbl1q_u8(samples, permute_tbl.val[0]), + vqtbl1q_u8(samples, permute_tbl.val[1]), + vqtbl1q_u8(samples, permute_tbl.val[2]) }; + + // First 4 output values. + int32x4_t sum0 = + vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[0], filters, 0); + sum0 = vusdotq_lane_s32(sum0, permuted_samples[1], filters, 1); + // Second 4 output values. + int32x4_t sum1 = + vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[1], filters, 0); + sum1 = vusdotq_lane_s32(sum1, permuted_samples[2], filters, 1); + + // Narrow and re-pack. + int16x8_t sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1)); return vqrshrun_n_s16(sum, FILTER_BITS); } @@ -95,7 +86,6 @@ void aom_convolve8_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, const int16_t *filter_y, int y_step_q4, int w, int h) { const int8x8_t filter = vmovn_s16(vld1q_s16(filter_x)); - uint8x16_t s0, s1, s2, s3; assert((intptr_t)dst % 4 == 0); assert(dst_stride % 4 == 0); @@ -107,19 +97,17 @@ void aom_convolve8_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, src -= ((SUBPEL_TAPS / 2) - 1); if (w == 4) { - const uint8x16x2_t perm_tbl = vld1q_u8_x2(dot_prod_permute_tbl); + const uint8x16x2_t perm_tbl = vld1q_u8_x2(kDotProdPermuteTbl); do { - int16x4_t t0, t1, t2, t3; - uint8x8_t d01, d23; - + uint8x16_t s0, s1, s2, s3; load_u8_16x4(src, src_stride, &s0, &s1, &s2, &s3); - t0 = convolve8_4_usdot(s0, filter, perm_tbl); - t1 = convolve8_4_usdot(s1, filter, perm_tbl); - t2 = convolve8_4_usdot(s2, filter, perm_tbl); - t3 = convolve8_4_usdot(s3, filter, perm_tbl); - d01 = vqrshrun_n_s16(vcombine_s16(t0, t1), FILTER_BITS); - d23 = vqrshrun_n_s16(vcombine_s16(t2, t3), FILTER_BITS); + int16x4_t d0 = convolve8_4_h(s0, filter, perm_tbl); + int16x4_t d1 = convolve8_4_h(s1, filter, perm_tbl); + int16x4_t d2 = convolve8_4_h(s2, filter, perm_tbl); + int16x4_t d3 = convolve8_4_h(s3, filter, perm_tbl); + uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS); + uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS); store_u8x4_strided_x2(dst + 0 * dst_stride, dst_stride, d01); store_u8x4_strided_x2(dst + 2 * dst_stride, dst_stride, d23); @@ -129,23 +117,20 @@ void aom_convolve8_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, h -= 4; } while (h > 0); } else { - const uint8x16x3_t perm_tbl = vld1q_u8_x3(dot_prod_permute_tbl); - const uint8_t *s; - uint8_t *d; - int width; - uint8x8_t d0, d1, d2, d3; + const uint8x16x3_t perm_tbl = vld1q_u8_x3(kDotProdPermuteTbl); do { - width = w; - s = src; - d = dst; + int width = w; + const uint8_t *s = src; + uint8_t *d = dst; do { + uint8x16_t s0, s1, s2, s3; load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3); - d0 = convolve8_8_usdot(s0, filter, perm_tbl); - d1 = convolve8_8_usdot(s1, filter, perm_tbl); - d2 = convolve8_8_usdot(s2, filter, perm_tbl); - d3 = convolve8_8_usdot(s3, filter, perm_tbl); + uint8x8_t d0 = convolve8_8_h(s0, filter, perm_tbl); + uint8x8_t d1 = convolve8_8_h(s1, filter, perm_tbl); + uint8x8_t d2 = convolve8_8_h(s2, filter, perm_tbl); + uint8x8_t d3 = convolve8_8_h(s3, filter, perm_tbl); store_u8_8x4(d, dst_stride, d0, d1, d2, d3); @@ -162,79 +147,83 @@ void aom_convolve8_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, static INLINE void transpose_concat_4x4(uint8x8_t a0, uint8x8_t a1, uint8x8_t a2, uint8x8_t a3, - uint8x16_t *b, - const uint8x16_t permute_tbl) { - /* Transpose 8-bit elements and concatenate result rows as follows: - * a0: 00, 01, 02, 03, XX, XX, XX, XX - * a1: 10, 11, 12, 13, XX, XX, XX, XX - * a2: 20, 21, 22, 23, XX, XX, XX, XX - * a3: 30, 31, 32, 33, XX, XX, XX, XX - * - * b: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33 - * - * The 'permute_tbl' is always 'dot_prod_tran_concat_tbl' above. Passing it - * as an argument is preferable to loading it directly from memory as this - * inline helper is called many times from the same parent function. - */ - - uint8x16x2_t samples = { { vcombine_u8(a0, a1), vcombine_u8(a2, a3) } }; - *b = vqtbl2q_u8(samples, permute_tbl); + uint8x16_t *b) { + // Transpose 8-bit elements and concatenate result rows as follows: + // a0: 00, 01, 02, 03, XX, XX, XX, XX + // a1: 10, 11, 12, 13, XX, XX, XX, XX + // a2: 20, 21, 22, 23, XX, XX, XX, XX + // a3: 30, 31, 32, 33, XX, XX, XX, XX + // + // b: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33 + + uint8x16_t a0q = vcombine_u8(a0, vdup_n_u8(0)); + uint8x16_t a1q = vcombine_u8(a1, vdup_n_u8(0)); + uint8x16_t a2q = vcombine_u8(a2, vdup_n_u8(0)); + uint8x16_t a3q = vcombine_u8(a3, vdup_n_u8(0)); + + uint8x16_t a01 = vzipq_u8(a0q, a1q).val[0]; + uint8x16_t a23 = vzipq_u8(a2q, a3q).val[0]; + + uint16x8_t a0123 = + vzipq_u16(vreinterpretq_u16_u8(a01), vreinterpretq_u16_u8(a23)).val[0]; + + *b = vreinterpretq_u8_u16(a0123); } static INLINE void transpose_concat_8x4(uint8x8_t a0, uint8x8_t a1, uint8x8_t a2, uint8x8_t a3, - uint8x16_t *b0, uint8x16_t *b1, - const uint8x16x2_t permute_tbl) { - /* Transpose 8-bit elements and concatenate result rows as follows: - * a0: 00, 01, 02, 03, 04, 05, 06, 07 - * a1: 10, 11, 12, 13, 14, 15, 16, 17 - * a2: 20, 21, 22, 23, 24, 25, 26, 27 - * a3: 30, 31, 32, 33, 34, 35, 36, 37 - * - * b0: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33 - * b1: 04, 14, 24, 34, 05, 15, 25, 35, 06, 16, 26, 36, 07, 17, 27, 37 - * - * The 'permute_tbl' is always 'dot_prod_tran_concat_tbl' above. Passing it - * as an argument is preferable to loading it directly from memory as this - * inline helper is called many times from the same parent function. - */ - - uint8x16x2_t samples = { { vcombine_u8(a0, a1), vcombine_u8(a2, a3) } }; - *b0 = vqtbl2q_u8(samples, permute_tbl.val[0]); - *b1 = vqtbl2q_u8(samples, permute_tbl.val[1]); + uint8x16_t *b0, uint8x16_t *b1) { + // Transpose 8-bit elements and concatenate result rows as follows: + // a0: 00, 01, 02, 03, 04, 05, 06, 07 + // a1: 10, 11, 12, 13, 14, 15, 16, 17 + // a2: 20, 21, 22, 23, 24, 25, 26, 27 + // a3: 30, 31, 32, 33, 34, 35, 36, 37 + // + // b0: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33 + // b1: 04, 14, 24, 34, 05, 15, 25, 35, 06, 16, 26, 36, 07, 17, 27, 37 + + uint8x16_t a0q = vcombine_u8(a0, vdup_n_u8(0)); + uint8x16_t a1q = vcombine_u8(a1, vdup_n_u8(0)); + uint8x16_t a2q = vcombine_u8(a2, vdup_n_u8(0)); + uint8x16_t a3q = vcombine_u8(a3, vdup_n_u8(0)); + + uint8x16_t a01 = vzipq_u8(a0q, a1q).val[0]; + uint8x16_t a23 = vzipq_u8(a2q, a3q).val[0]; + + uint16x8x2_t a0123 = + vzipq_u16(vreinterpretq_u16_u8(a01), vreinterpretq_u16_u8(a23)); + + *b0 = vreinterpretq_u8_u16(a0123.val[0]); + *b1 = vreinterpretq_u8_u16(a0123.val[1]); } -static INLINE int16x4_t convolve8_4_usdot_partial(const uint8x16_t samples_lo, - const uint8x16_t samples_hi, - const int8x8_t filter) { - /* Sample permutation is performed by the caller. */ - int32x4_t sum; - - sum = vusdotq_lane_s32(vdupq_n_s32(0), samples_lo, filter, 0); - sum = vusdotq_lane_s32(sum, samples_hi, filter, 1); +static INLINE int16x4_t convolve8_4_v(const uint8x16_t samples_lo, + const uint8x16_t samples_hi, + const int8x8_t filters) { + // Sample permutation is performed by the caller. + int32x4_t sum = vusdotq_lane_s32(vdupq_n_s32(0), samples_lo, filters, 0); + sum = vusdotq_lane_s32(sum, samples_hi, filters, 1); - /* Further narrowing and packing is performed by the caller. */ + // Further narrowing and packing is performed by the caller. return vqmovn_s32(sum); } -static INLINE uint8x8_t convolve8_8_usdot_partial(const uint8x16_t samples0_lo, - const uint8x16_t samples0_hi, - const uint8x16_t samples1_lo, - const uint8x16_t samples1_hi, - const int8x8_t filter) { - /* Sample permutation is performed by the caller. */ - int32x4_t sum0, sum1; - int16x8_t sum; - - /* First 4 output values. */ - sum0 = vusdotq_lane_s32(vdupq_n_s32(0), samples0_lo, filter, 0); - sum0 = vusdotq_lane_s32(sum0, samples0_hi, filter, 1); - /* Second 4 output values. */ - sum1 = vusdotq_lane_s32(vdupq_n_s32(0), samples1_lo, filter, 0); - sum1 = vusdotq_lane_s32(sum1, samples1_hi, filter, 1); - - /* Narrow and re-pack. */ - sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1)); +static INLINE uint8x8_t convolve8_8_v(const uint8x16_t samples0_lo, + const uint8x16_t samples0_hi, + const uint8x16_t samples1_lo, + const uint8x16_t samples1_hi, + const int8x8_t filters) { + // Sample permutation is performed by the caller. + + // First 4 output values. + int32x4_t sum0 = vusdotq_lane_s32(vdupq_n_s32(0), samples0_lo, filters, 0); + sum0 = vusdotq_lane_s32(sum0, samples0_hi, filters, 1); + // Second 4 output values. + int32x4_t sum1 = vusdotq_lane_s32(vdupq_n_s32(0), samples1_lo, filters, 0); + sum1 = vusdotq_lane_s32(sum1, samples1_hi, filters, 1); + + // Narrow and re-pack. + int16x8_t sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1)); return vqrshrun_n_s16(sum, FILTER_BITS); } @@ -244,7 +233,7 @@ void aom_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, const int16_t *filter_y, int y_step_q4, int w, int h) { const int8x8_t filter = vmovn_s16(vld1q_s16(filter_y)); - const uint8x16x3_t merge_block_tbl = vld1q_u8_x3(dot_prod_merge_block_tbl); + const uint8x16x3_t merge_block_tbl = vld1q_u8_x3(kDotProdMergeBlockTbl); uint8x16x2_t samples_LUT; assert((intptr_t)dst % 4 == 0); @@ -257,47 +246,44 @@ void aom_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, src -= ((SUBPEL_TAPS / 2) - 1) * src_stride; if (w == 4) { - const uint8x16_t tran_concat_tbl = vld1q_u8(dot_prod_tran_concat_tbl); - uint8x8_t s0, s1, s2, s3, s4, s5, s6; load_u8_8x7(src, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); src += 7 * src_stride; - /* This operation combines a conventional transpose and the sample permute - * (see horizontal case) required before computing the dot product. - */ + // This operation combines a conventional transpose and the sample permute + // (see horizontal case) required before computing the dot product. uint8x16_t s0123, s1234, s2345, s3456; - transpose_concat_4x4(s0, s1, s2, s3, &s0123, tran_concat_tbl); - transpose_concat_4x4(s1, s2, s3, s4, &s1234, tran_concat_tbl); - transpose_concat_4x4(s2, s3, s4, s5, &s2345, tran_concat_tbl); - transpose_concat_4x4(s3, s4, s5, s6, &s3456, tran_concat_tbl); + transpose_concat_4x4(s0, s1, s2, s3, &s0123); + transpose_concat_4x4(s1, s2, s3, s4, &s1234); + transpose_concat_4x4(s2, s3, s4, s5, &s2345); + transpose_concat_4x4(s3, s4, s5, s6, &s3456); do { uint8x8_t s7, s8, s9, s10; load_u8_8x4(src, src_stride, &s7, &s8, &s9, &s10); uint8x16_t s4567, s5678, s6789, s78910; - transpose_concat_4x4(s7, s8, s9, s10, &s78910, tran_concat_tbl); + transpose_concat_4x4(s7, s8, s9, s10, &s78910); - /* Merge new data into block from previous iteration. */ + // Merge new data into block from previous iteration. samples_LUT.val[0] = s3456; samples_LUT.val[1] = s78910; s4567 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]); s5678 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]); s6789 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]); - int16x4_t d0 = convolve8_4_usdot_partial(s0123, s4567, filter); - int16x4_t d1 = convolve8_4_usdot_partial(s1234, s5678, filter); - int16x4_t d2 = convolve8_4_usdot_partial(s2345, s6789, filter); - int16x4_t d3 = convolve8_4_usdot_partial(s3456, s78910, filter); + int16x4_t d0 = convolve8_4_v(s0123, s4567, filter); + int16x4_t d1 = convolve8_4_v(s1234, s5678, filter); + int16x4_t d2 = convolve8_4_v(s2345, s6789, filter); + int16x4_t d3 = convolve8_4_v(s3456, s78910, filter); uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS); uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS); store_u8x4_strided_x2(dst + 0 * dst_stride, dst_stride, d01); store_u8x4_strided_x2(dst + 2 * dst_stride, dst_stride, d23); - /* Prepare block for next iteration - re-using as much as possible. */ - /* Shuffle everything up four rows. */ + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. s0123 = s4567; s1234 = s5678; s2345 = s6789; @@ -308,8 +294,6 @@ void aom_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, h -= 4; } while (h != 0); } else { - const uint8x16x2_t tran_concat_tbl = vld1q_u8_x2(dot_prod_tran_concat_tbl); - do { int height = h; const uint8_t *s = src; @@ -319,19 +303,14 @@ void aom_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, load_u8_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6); s += 7 * src_stride; - /* This operation combines a conventional transpose and the sample permute - * (see horizontal case) required before computing the dot product. - */ + // This operation combines a conventional transpose and the sample permute + // (see horizontal case) required before computing the dot product. uint8x16_t s0123_lo, s0123_hi, s1234_lo, s1234_hi, s2345_lo, s2345_hi, s3456_lo, s3456_hi; - transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi, - tran_concat_tbl); - transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi, - tran_concat_tbl); - transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi, - tran_concat_tbl); - transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi, - tran_concat_tbl); + transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi); + transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi); + transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi); + transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi); do { uint8x8_t s7, s8, s9, s10; @@ -339,10 +318,9 @@ void aom_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, uint8x16_t s4567_lo, s4567_hi, s5678_lo, s5678_hi, s6789_lo, s6789_hi, s78910_lo, s78910_hi; - transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi, - tran_concat_tbl); + transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi); - /* Merge new data into block from previous iteration. */ + // Merge new data into block from previous iteration. samples_LUT.val[0] = s3456_lo; samples_LUT.val[1] = s78910_lo; s4567_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]); @@ -355,19 +333,19 @@ void aom_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride, s5678_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]); s6789_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]); - uint8x8_t d0 = convolve8_8_usdot_partial(s0123_lo, s4567_lo, s0123_hi, - s4567_hi, filter); - uint8x8_t d1 = convolve8_8_usdot_partial(s1234_lo, s5678_lo, s1234_hi, - s5678_hi, filter); - uint8x8_t d2 = convolve8_8_usdot_partial(s2345_lo, s6789_lo, s2345_hi, - s6789_hi, filter); - uint8x8_t d3 = convolve8_8_usdot_partial(s3456_lo, s78910_lo, s3456_hi, - s78910_hi, filter); + uint8x8_t d0 = + convolve8_8_v(s0123_lo, s4567_lo, s0123_hi, s4567_hi, filter); + uint8x8_t d1 = + convolve8_8_v(s1234_lo, s5678_lo, s1234_hi, s5678_hi, filter); + uint8x8_t d2 = + convolve8_8_v(s2345_lo, s6789_lo, s2345_hi, s6789_hi, filter); + uint8x8_t d3 = + convolve8_8_v(s3456_lo, s78910_lo, s3456_hi, s78910_hi, filter); store_u8_8x4(d, dst_stride, d0, d1, d2, d3); - /* Prepare block for next iteration - re-using as much as possible. */ - /* Shuffle everything up four rows. */ + // Prepare block for next iteration - re-using as much as possible. + // Shuffle everything up four rows. s0123_lo = s4567_lo; s0123_hi = s4567_hi; s1234_lo = s5678_lo; -- cgit v1.2.3