av1_error_neon.c (3693B)
1 /* 2 * Copyright (c) 2019, Alliance for Open Media. All rights reserved. 3 * 4 * This source code is subject to the terms of the BSD 2 Clause License and 5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License 6 * was not distributed with this source code in the LICENSE file, you can 7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open 8 * Media Patent License 1.0 was not distributed with this source code in the 9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent. 10 */ 11 12 #include <arm_neon.h> 13 #include <assert.h> 14 15 #include "config/aom_config.h" 16 #include "config/av1_rtcd.h" 17 18 #include "aom_dsp/aom_dsp_common.h" 19 #include "aom_dsp/arm/mem_neon.h" 20 #include "aom_dsp/arm/sum_neon.h" 21 22 int64_t av1_block_error_neon(const tran_low_t *coeff, const tran_low_t *dqcoeff, 23 intptr_t block_size, int64_t *ssz) { 24 uint64x2_t err_u64 = vdupq_n_u64(0); 25 int64x2_t ssz_s64 = vdupq_n_s64(0); 26 27 assert(block_size >= 16); 28 assert((block_size % 16) == 0); 29 30 do { 31 const int16x8_t c0 = load_tran_low_to_s16q(coeff); 32 const int16x8_t c1 = load_tran_low_to_s16q(coeff + 8); 33 const int16x8_t d0 = load_tran_low_to_s16q(dqcoeff); 34 const int16x8_t d1 = load_tran_low_to_s16q(dqcoeff + 8); 35 36 const uint16x8_t diff0 = vreinterpretq_u16_s16(vabdq_s16(c0, d0)); 37 const uint16x8_t diff1 = vreinterpretq_u16_s16(vabdq_s16(c1, d1)); 38 39 // By operating on unsigned integers we can store up to 4 squared diff in a 40 // 32-bit element before having to widen to 64 bits. 41 uint32x4_t err = vmull_u16(vget_low_u16(diff0), vget_low_u16(diff0)); 42 err = vmlal_u16(err, vget_high_u16(diff0), vget_high_u16(diff0)); 43 err = vmlal_u16(err, vget_low_u16(diff1), vget_low_u16(diff1)); 44 err = vmlal_u16(err, vget_high_u16(diff1), vget_high_u16(diff1)); 45 err_u64 = vpadalq_u32(err_u64, err); 46 47 // We can't do the same here as we're operating on signed integers, so we 48 // can only accumulate 2 squares. 49 int32x4_t ssz0 = vmull_s16(vget_low_s16(c0), vget_low_s16(c0)); 50 ssz0 = vmlal_s16(ssz0, vget_high_s16(c0), vget_high_s16(c0)); 51 ssz_s64 = vpadalq_s32(ssz_s64, ssz0); 52 53 int32x4_t ssz1 = vmull_s16(vget_low_s16(c1), vget_low_s16(c1)); 54 ssz1 = vmlal_s16(ssz1, vget_high_s16(c1), vget_high_s16(c1)); 55 ssz_s64 = vpadalq_s32(ssz_s64, ssz1); 56 57 coeff += 16; 58 dqcoeff += 16; 59 block_size -= 16; 60 } while (block_size != 0); 61 62 *ssz = horizontal_add_s64x2(ssz_s64); 63 return (int64_t)horizontal_add_u64x2(err_u64); 64 } 65 66 int64_t av1_block_error_lp_neon(const int16_t *coeff, const int16_t *dqcoeff, 67 intptr_t block_size) { 68 uint64x2_t err_u64 = vdupq_n_u64(0); 69 70 assert(block_size >= 16); 71 assert((block_size % 16) == 0); 72 73 do { 74 const int16x8_t c0 = vld1q_s16(coeff); 75 const int16x8_t c1 = vld1q_s16(coeff + 8); 76 const int16x8_t d0 = vld1q_s16(dqcoeff); 77 const int16x8_t d1 = vld1q_s16(dqcoeff + 8); 78 79 const uint16x8_t diff0 = vreinterpretq_u16_s16(vabdq_s16(c0, d0)); 80 const uint16x8_t diff1 = vreinterpretq_u16_s16(vabdq_s16(c1, d1)); 81 82 // By operating on unsigned integers we can store up to 4 squared diff in a 83 // 32-bit element before having to widen to 64 bits. 84 uint32x4_t err = vmull_u16(vget_low_u16(diff0), vget_low_u16(diff0)); 85 err = vmlal_u16(err, vget_high_u16(diff0), vget_high_u16(diff0)); 86 err = vmlal_u16(err, vget_low_u16(diff1), vget_low_u16(diff1)); 87 err = vmlal_u16(err, vget_high_u16(diff1), vget_high_u16(diff1)); 88 err_u64 = vpadalq_u32(err_u64, err); 89 90 coeff += 16; 91 dqcoeff += 16; 92 block_size -= 16; 93 } while (block_size != 0); 94 95 return (int64_t)horizontal_add_u64x2(err_u64); 96 }