highbd_avg_neon.c (4730B)
1 /* 2 * Copyright (c) 2023 The WebM project authors. All rights reserved. 3 * Copyright (c) 2023, Alliance for Open Media. All rights reserved. 4 * 5 * This source code is subject to the terms of the BSD 2 Clause License and 6 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License 7 * was not distributed with this source code in the LICENSE file, you can 8 * obtain it at www.aomedia.org/license/software. If the Alliance for Open 9 * Media Patent License 1.0 was not distributed with this source code in the 10 * PATENTS file, you can obtain it at www.aomedia.org/license/patent. 11 */ 12 13 #include <arm_neon.h> 14 15 #include "config/aom_config.h" 16 #include "config/aom_dsp_rtcd.h" 17 #include "aom/aom_integer.h" 18 #include "aom_dsp/arm/mem_neon.h" 19 #include "aom_dsp/arm/sum_neon.h" 20 #include "aom_ports/mem.h" 21 22 uint32_t aom_highbd_avg_4x4_neon(const uint8_t *a, int a_stride) { 23 const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a); 24 uint16x4_t sum, a0, a1, a2, a3; 25 26 load_u16_4x4(a_ptr, a_stride, &a0, &a1, &a2, &a3); 27 28 sum = vadd_u16(a0, a1); 29 sum = vadd_u16(sum, a2); 30 sum = vadd_u16(sum, a3); 31 32 return (horizontal_add_u16x4(sum) + (1 << 3)) >> 4; 33 } 34 35 uint32_t aom_highbd_avg_8x8_neon(const uint8_t *a, int a_stride) { 36 const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a); 37 uint16x8_t sum, a0, a1, a2, a3, a4, a5, a6, a7; 38 39 load_u16_8x8(a_ptr, a_stride, &a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7); 40 41 sum = vaddq_u16(a0, a1); 42 sum = vaddq_u16(sum, a2); 43 sum = vaddq_u16(sum, a3); 44 sum = vaddq_u16(sum, a4); 45 sum = vaddq_u16(sum, a5); 46 sum = vaddq_u16(sum, a6); 47 sum = vaddq_u16(sum, a7); 48 49 return (horizontal_add_u16x8(sum) + (1 << 5)) >> 6; 50 } 51 52 void aom_highbd_minmax_8x8_neon(const uint8_t *s8, int p, const uint8_t *d8, 53 int dp, int *min, int *max) { 54 const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(s8); 55 const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(d8); 56 57 const uint16x8_t a0 = vld1q_u16(a_ptr + 0 * p); 58 const uint16x8_t a1 = vld1q_u16(a_ptr + 1 * p); 59 const uint16x8_t a2 = vld1q_u16(a_ptr + 2 * p); 60 const uint16x8_t a3 = vld1q_u16(a_ptr + 3 * p); 61 const uint16x8_t a4 = vld1q_u16(a_ptr + 4 * p); 62 const uint16x8_t a5 = vld1q_u16(a_ptr + 5 * p); 63 const uint16x8_t a6 = vld1q_u16(a_ptr + 6 * p); 64 const uint16x8_t a7 = vld1q_u16(a_ptr + 7 * p); 65 66 const uint16x8_t b0 = vld1q_u16(b_ptr + 0 * dp); 67 const uint16x8_t b1 = vld1q_u16(b_ptr + 1 * dp); 68 const uint16x8_t b2 = vld1q_u16(b_ptr + 2 * dp); 69 const uint16x8_t b3 = vld1q_u16(b_ptr + 3 * dp); 70 const uint16x8_t b4 = vld1q_u16(b_ptr + 4 * dp); 71 const uint16x8_t b5 = vld1q_u16(b_ptr + 5 * dp); 72 const uint16x8_t b6 = vld1q_u16(b_ptr + 6 * dp); 73 const uint16x8_t b7 = vld1q_u16(b_ptr + 7 * dp); 74 75 const uint16x8_t abs_diff0 = vabdq_u16(a0, b0); 76 const uint16x8_t abs_diff1 = vabdq_u16(a1, b1); 77 const uint16x8_t abs_diff2 = vabdq_u16(a2, b2); 78 const uint16x8_t abs_diff3 = vabdq_u16(a3, b3); 79 const uint16x8_t abs_diff4 = vabdq_u16(a4, b4); 80 const uint16x8_t abs_diff5 = vabdq_u16(a5, b5); 81 const uint16x8_t abs_diff6 = vabdq_u16(a6, b6); 82 const uint16x8_t abs_diff7 = vabdq_u16(a7, b7); 83 84 const uint16x8_t max01 = vmaxq_u16(abs_diff0, abs_diff1); 85 const uint16x8_t max23 = vmaxq_u16(abs_diff2, abs_diff3); 86 const uint16x8_t max45 = vmaxq_u16(abs_diff4, abs_diff5); 87 const uint16x8_t max67 = vmaxq_u16(abs_diff6, abs_diff7); 88 89 const uint16x8_t max0123 = vmaxq_u16(max01, max23); 90 const uint16x8_t max4567 = vmaxq_u16(max45, max67); 91 const uint16x8_t max07 = vmaxq_u16(max0123, max4567); 92 93 const uint16x8_t min01 = vminq_u16(abs_diff0, abs_diff1); 94 const uint16x8_t min23 = vminq_u16(abs_diff2, abs_diff3); 95 const uint16x8_t min45 = vminq_u16(abs_diff4, abs_diff5); 96 const uint16x8_t min67 = vminq_u16(abs_diff6, abs_diff7); 97 98 const uint16x8_t min0123 = vminq_u16(min01, min23); 99 const uint16x8_t min4567 = vminq_u16(min45, min67); 100 const uint16x8_t min07 = vminq_u16(min0123, min4567); 101 102 #if AOM_ARCH_AARCH64 103 *max = (int)vmaxvq_u16(max07); 104 *min = (int)vminvq_u16(min07); 105 #else 106 // Split into 64-bit vectors and execute pairwise min/max. 107 uint16x4_t ab_max = vmax_u16(vget_high_u16(max07), vget_low_u16(max07)); 108 uint16x4_t ab_min = vmin_u16(vget_high_u16(min07), vget_low_u16(min07)); 109 110 // Enough runs of vpmax/min propagate the max/min values to every position. 111 ab_max = vpmax_u16(ab_max, ab_max); 112 ab_min = vpmin_u16(ab_min, ab_min); 113 114 ab_max = vpmax_u16(ab_max, ab_max); 115 ab_min = vpmin_u16(ab_min, ab_min); 116 117 ab_max = vpmax_u16(ab_max, ab_max); 118 ab_min = vpmin_u16(ab_min, ab_min); 119 120 *min = *max = 0; // Clear high bits 121 // Store directly to avoid costly neon->gpr transfer. 122 vst1_lane_u16((uint16_t *)max, ab_max, 0); 123 vst1_lane_u16((uint16_t *)min, ab_min, 0); 124 #endif 125 }