highbd_avg_pred_neon.c (3869B)
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 #include <assert.h> 15 16 #include "config/aom_config.h" 17 #include "config/aom_dsp_rtcd.h" 18 19 #include "aom_dsp/arm/blend_neon.h" 20 #include "aom_dsp/arm/dist_wtd_avg_neon.h" 21 #include "aom_dsp/arm/mem_neon.h" 22 #include "aom_dsp/blend.h" 23 24 void aom_highbd_comp_avg_pred_neon(uint8_t *comp_pred8, const uint8_t *pred8, 25 int width, int height, const uint8_t *ref8, 26 int ref_stride) { 27 const uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); 28 const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); 29 uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); 30 31 int i = height; 32 if (width > 8) { 33 do { 34 int j = 0; 35 do { 36 const uint16x8_t p = vld1q_u16(pred + j); 37 const uint16x8_t r = vld1q_u16(ref + j); 38 39 uint16x8_t avg = vrhaddq_u16(p, r); 40 vst1q_u16(comp_pred + j, avg); 41 42 j += 8; 43 } while (j < width); 44 45 comp_pred += width; 46 pred += width; 47 ref += ref_stride; 48 } while (--i != 0); 49 } else if (width == 8) { 50 do { 51 const uint16x8_t p = vld1q_u16(pred); 52 const uint16x8_t r = vld1q_u16(ref); 53 54 uint16x8_t avg = vrhaddq_u16(p, r); 55 vst1q_u16(comp_pred, avg); 56 57 comp_pred += width; 58 pred += width; 59 ref += ref_stride; 60 } while (--i != 0); 61 } else { 62 assert(width == 4); 63 do { 64 const uint16x4_t p = vld1_u16(pred); 65 const uint16x4_t r = vld1_u16(ref); 66 67 uint16x4_t avg = vrhadd_u16(p, r); 68 vst1_u16(comp_pred, avg); 69 70 comp_pred += width; 71 pred += width; 72 ref += ref_stride; 73 } while (--i != 0); 74 } 75 } 76 77 void aom_highbd_comp_mask_pred_neon(uint8_t *comp_pred8, const uint8_t *pred8, 78 int width, int height, const uint8_t *ref8, 79 int ref_stride, const uint8_t *mask, 80 int mask_stride, int invert_mask) { 81 uint16_t *pred = CONVERT_TO_SHORTPTR(pred8); 82 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); 83 uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8); 84 85 const uint16_t *src0 = invert_mask ? pred : ref; 86 const uint16_t *src1 = invert_mask ? ref : pred; 87 const int src_stride0 = invert_mask ? width : ref_stride; 88 const int src_stride1 = invert_mask ? ref_stride : width; 89 90 if (width >= 8) { 91 do { 92 int j = 0; 93 94 do { 95 const uint16x8_t s0 = vld1q_u16(src0 + j); 96 const uint16x8_t s1 = vld1q_u16(src1 + j); 97 const uint16x8_t m0 = vmovl_u8(vld1_u8(mask + j)); 98 99 uint16x8_t blend_u16 = alpha_blend_a64_u16x8(m0, s0, s1); 100 101 vst1q_u16(comp_pred + j, blend_u16); 102 103 j += 8; 104 } while (j < width); 105 106 src0 += src_stride0; 107 src1 += src_stride1; 108 mask += mask_stride; 109 comp_pred += width; 110 } while (--height != 0); 111 } else { 112 assert(width == 4); 113 114 do { 115 const uint16x4_t s0 = vld1_u16(src0); 116 const uint16x4_t s1 = vld1_u16(src1); 117 const uint16x4_t m0 = vget_low_u16(vmovl_u8(load_unaligned_u8_4x1(mask))); 118 119 uint16x4_t blend_u16 = alpha_blend_a64_u16x4(m0, s0, s1); 120 121 vst1_u16(comp_pred, blend_u16); 122 123 src0 += src_stride0; 124 src1 += src_stride1; 125 mask += mask_stride; 126 comp_pred += 4; 127 } while (--height != 0); 128 } 129 }