amass/FaceAccess
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

add neon code.

Browse Source
This commit is contained in:
luocai
2024-09-06 11:01:01 +08:00
parent fe7ab957c8
commit c6debcc62a
11 changed files with 1312 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

88
VocieProcess/common_audio/signal_processing/cross_correlation_neon.c Normal file
View File

@ -0,0 +1,88 @@
/*
* Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "common_audio/signal_processing/include/signal_processing_library.h"
#include "rtc_base/system/arch.h"
#include <arm_neon.h>
static inline void DotProductWithScaleNeon(int32_t* cross_correlation,
const int16_t* vector1,
const int16_t* vector2,
size_t length,
int scaling) {
size_t i = 0;
size_t len1 = length >> 3;
size_t len2 = length & 7;
int64x2_t sum0 = vdupq_n_s64(0);
int64x2_t sum1 = vdupq_n_s64(0);
for (i = len1; i > 0; i -= 1) {
int16x8_t seq1_16x8 = vld1q_s16(vector1);
int16x8_t seq2_16x8 = vld1q_s16(vector2);
#if defined(WEBRTC_ARCH_ARM64)
int32x4_t tmp0 = vmull_s16(vget_low_s16(seq1_16x8),
vget_low_s16(seq2_16x8));
int32x4_t tmp1 = vmull_high_s16(seq1_16x8, seq2_16x8);
#else
int32x4_t tmp0 = vmull_s16(vget_low_s16(seq1_16x8),
vget_low_s16(seq2_16x8));
int32x4_t tmp1 = vmull_s16(vget_high_s16(seq1_16x8),
vget_high_s16(seq2_16x8));
#endif
sum0 = vpadalq_s32(sum0, tmp0);
sum1 = vpadalq_s32(sum1, tmp1);
vector1 += 8;
vector2 += 8;
}
// Calculate the rest of the samples.
int64_t sum_res = 0;
for (i = len2; i > 0; i -= 1) {
sum_res += WEBRTC_SPL_MUL_16_16(*vector1, *vector2);
vector1++;
vector2++;
}
sum0 = vaddq_s64(sum0, sum1);
#if defined(WEBRTC_ARCH_ARM64)
int64_t sum2 = vaddvq_s64(sum0);
*cross_correlation = (int32_t)((sum2 + sum_res) >> scaling);
#else
int64x1_t shift = vdup_n_s64(-scaling);
int64x1_t sum2 = vadd_s64(vget_low_s64(sum0), vget_high_s64(sum0));
sum2 = vadd_s64(sum2, vdup_n_s64(sum_res));
sum2 = vshl_s64(sum2, shift);
vst1_lane_s32(cross_correlation, vreinterpret_s32_s64(sum2), 0);
#endif
}
/* NEON version of WebRtcSpl_CrossCorrelation() for ARM32/64 platforms. */
void WebRtcSpl_CrossCorrelationNeon(int32_t* cross_correlation,
const int16_t* seq1,
const int16_t* seq2,
size_t dim_seq,
size_t dim_cross_correlation,
int right_shifts,
int step_seq2) {
int i = 0;
for (i = 0; i < (int)dim_cross_correlation; i++) {
const int16_t* seq1_ptr = seq1;
const int16_t* seq2_ptr = seq2 + (step_seq2 * i);
DotProductWithScaleNeon(cross_correlation,
seq1_ptr,
seq2_ptr,
dim_seq,
right_shifts);
cross_correlation++;
}
}

224
VocieProcess/common_audio/signal_processing/downsample_fast_neon.c Normal file
View File

@ -0,0 +1,224 @@
/*
* Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include <arm_neon.h>
#include "common_audio/signal_processing/include/signal_processing_library.h"
#include "rtc_base/checks.h"
// NEON intrinsics version of WebRtcSpl_DownsampleFast()
// for ARM 32-bit/64-bit platforms.
int WebRtcSpl_DownsampleFastNeon(const int16_t* data_in,
size_t data_in_length,
int16_t* data_out,
size_t data_out_length,
const int16_t* __restrict coefficients,
size_t coefficients_length,
int factor,
size_t delay) {
// Using signed indexes to be able to compute negative i-j that
// is used to index data_in.
int i = 0;
int j = 0;
int32_t out_s32 = 0;
int endpos = delay + factor * (data_out_length - 1) + 1;
size_t res = data_out_length & 0x7;
int endpos1 = endpos - factor * res;
// Return error if any of the running conditions doesn't meet.
if (data_out_length == 0 || coefficients_length == 0
|| (int)data_in_length < endpos) {
return -1;
}
RTC_DCHECK_GE(endpos, 0);
RTC_DCHECK_GE(endpos1, 0);
// First part, unroll the loop 8 times, with 3 subcases
// (factor == 2, 4, others).
switch (factor) {
case 2: {
for (i = delay; i < endpos1; i += 16) {
// Round value, 0.5 in Q12.
int32x4_t out32x4_0 = vdupq_n_s32(2048);
int32x4_t out32x4_1 = vdupq_n_s32(2048);
#if defined(WEBRTC_ARCH_ARM64)
// Unroll the loop 2 times.
for (j = 0; j < (int)coefficients_length - 1; j += 2) {
int32x2_t coeff32 = vld1_dup_s32((int32_t*)&coefficients[j]);
int16x4_t coeff16x4 = vreinterpret_s16_s32(coeff32);
int16x8x2_t in16x8x2 = vld2q_s16(&data_in[i - j - 1]);
// Mul and accumulate low 64-bit data.
int16x4_t in16x4_0 = vget_low_s16(in16x8x2.val[0]);
int16x4_t in16x4_1 = vget_low_s16(in16x8x2.val[1]);
out32x4_0 = vmlal_lane_s16(out32x4_0, in16x4_0, coeff16x4, 1);
out32x4_0 = vmlal_lane_s16(out32x4_0, in16x4_1, coeff16x4, 0);
// Mul and accumulate high 64-bit data.
// TODO: vget_high_s16 need extra cost on ARM64. This could be
// replaced by vmlal_high_lane_s16. But for the interface of
// vmlal_high_lane_s16, there is a bug in gcc 4.9.
// This issue need to be tracked in the future.
int16x4_t in16x4_2 = vget_high_s16(in16x8x2.val[0]);
int16x4_t in16x4_3 = vget_high_s16(in16x8x2.val[1]);
out32x4_1 = vmlal_lane_s16(out32x4_1, in16x4_2, coeff16x4, 1);
out32x4_1 = vmlal_lane_s16(out32x4_1, in16x4_3, coeff16x4, 0);
}
for (; j < (int)coefficients_length; j++) {
int16x4_t coeff16x4 = vld1_dup_s16(&coefficients[j]);
int16x8x2_t in16x8x2 = vld2q_s16(&data_in[i - j]);
// Mul and accumulate low 64-bit data.
int16x4_t in16x4_0 = vget_low_s16(in16x8x2.val[0]);
out32x4_0 = vmlal_lane_s16(out32x4_0, in16x4_0, coeff16x4, 0);
// Mul and accumulate high 64-bit data.
// TODO: vget_high_s16 need extra cost on ARM64. This could be
// replaced by vmlal_high_lane_s16. But for the interface of
// vmlal_high_lane_s16, there is a bug in gcc 4.9.
// This issue need to be tracked in the future.
int16x4_t in16x4_1 = vget_high_s16(in16x8x2.val[0]);
out32x4_1 = vmlal_lane_s16(out32x4_1, in16x4_1, coeff16x4, 0);
}
#else
// On ARMv7, the loop unrolling 2 times results in performance
// regression.
for (j = 0; j < (int)coefficients_length; j++) {
int16x4_t coeff16x4 = vld1_dup_s16(&coefficients[j]);
int16x8x2_t in16x8x2 = vld2q_s16(&data_in[i - j]);
// Mul and accumulate.
int16x4_t in16x4_0 = vget_low_s16(in16x8x2.val[0]);
int16x4_t in16x4_1 = vget_high_s16(in16x8x2.val[0]);
out32x4_0 = vmlal_lane_s16(out32x4_0, in16x4_0, coeff16x4, 0);
out32x4_1 = vmlal_lane_s16(out32x4_1, in16x4_1, coeff16x4, 0);
}
#endif
// Saturate and store the output.
int16x4_t out16x4_0 = vqshrn_n_s32(out32x4_0, 12);
int16x4_t out16x4_1 = vqshrn_n_s32(out32x4_1, 12);
vst1q_s16(data_out, vcombine_s16(out16x4_0, out16x4_1));
data_out += 8;
}
break;
}
case 4: {
for (i = delay; i < endpos1; i += 32) {
// Round value, 0.5 in Q12.
int32x4_t out32x4_0 = vdupq_n_s32(2048);
int32x4_t out32x4_1 = vdupq_n_s32(2048);
// Unroll the loop 4 times.
for (j = 0; j < (int)coefficients_length - 3; j += 4) {
int16x4_t coeff16x4 = vld1_s16(&coefficients[j]);
int16x8x4_t in16x8x4 = vld4q_s16(&data_in[i - j - 3]);
// Mul and accumulate low 64-bit data.
int16x4_t in16x4_0 = vget_low_s16(in16x8x4.val[0]);
int16x4_t in16x4_2 = vget_low_s16(in16x8x4.val[1]);
int16x4_t in16x4_4 = vget_low_s16(in16x8x4.val[2]);
int16x4_t in16x4_6 = vget_low_s16(in16x8x4.val[3]);
out32x4_0 = vmlal_lane_s16(out32x4_0, in16x4_0, coeff16x4, 3);
out32x4_0 = vmlal_lane_s16(out32x4_0, in16x4_2, coeff16x4, 2);
out32x4_0 = vmlal_lane_s16(out32x4_0, in16x4_4, coeff16x4, 1);
out32x4_0 = vmlal_lane_s16(out32x4_0, in16x4_6, coeff16x4, 0);
// Mul and accumulate high 64-bit data.
// TODO: vget_high_s16 need extra cost on ARM64. This could be
// replaced by vmlal_high_lane_s16. But for the interface of
// vmlal_high_lane_s16, there is a bug in gcc 4.9.
// This issue need to be tracked in the future.
int16x4_t in16x4_1 = vget_high_s16(in16x8x4.val[0]);
int16x4_t in16x4_3 = vget_high_s16(in16x8x4.val[1]);
int16x4_t in16x4_5 = vget_high_s16(in16x8x4.val[2]);
int16x4_t in16x4_7 = vget_high_s16(in16x8x4.val[3]);
out32x4_1 = vmlal_lane_s16(out32x4_1, in16x4_1, coeff16x4, 3);
out32x4_1 = vmlal_lane_s16(out32x4_1, in16x4_3, coeff16x4, 2);
out32x4_1 = vmlal_lane_s16(out32x4_1, in16x4_5, coeff16x4, 1);
out32x4_1 = vmlal_lane_s16(out32x4_1, in16x4_7, coeff16x4, 0);
}
for (; j < (int)coefficients_length; j++) {
int16x4_t coeff16x4 = vld1_dup_s16(&coefficients[j]);
int16x8x4_t in16x8x4 = vld4q_s16(&data_in[i - j]);
// Mul and accumulate low 64-bit data.
int16x4_t in16x4_0 = vget_low_s16(in16x8x4.val[0]);
out32x4_0 = vmlal_lane_s16(out32x4_0, in16x4_0, coeff16x4, 0);
// Mul and accumulate high 64-bit data.
// TODO: vget_high_s16 need extra cost on ARM64. This could be
// replaced by vmlal_high_lane_s16. But for the interface of
// vmlal_high_lane_s16, there is a bug in gcc 4.9.
// This issue need to be tracked in the future.
int16x4_t in16x4_1 = vget_high_s16(in16x8x4.val[0]);
out32x4_1 = vmlal_lane_s16(out32x4_1, in16x4_1, coeff16x4, 0);
}
// Saturate and store the output.
int16x4_t out16x4_0 = vqshrn_n_s32(out32x4_0, 12);
int16x4_t out16x4_1 = vqshrn_n_s32(out32x4_1, 12);
vst1q_s16(data_out, vcombine_s16(out16x4_0, out16x4_1));
data_out += 8;
}
break;
}
default: {
for (i = delay; i < endpos1; i += factor * 8) {
// Round value, 0.5 in Q12.
int32x4_t out32x4_0 = vdupq_n_s32(2048);
int32x4_t out32x4_1 = vdupq_n_s32(2048);
for (j = 0; j < (int)coefficients_length; j++) {
int16x4_t coeff16x4 = vld1_dup_s16(&coefficients[j]);
int16x4_t in16x4_0 = vld1_dup_s16(&data_in[i - j]);
in16x4_0 = vld1_lane_s16(&data_in[i + factor - j], in16x4_0, 1);
in16x4_0 = vld1_lane_s16(&data_in[i + factor * 2 - j], in16x4_0, 2);
in16x4_0 = vld1_lane_s16(&data_in[i + factor * 3 - j], in16x4_0, 3);
int16x4_t in16x4_1 = vld1_dup_s16(&data_in[i + factor * 4 - j]);
in16x4_1 = vld1_lane_s16(&data_in[i + factor * 5 - j], in16x4_1, 1);
in16x4_1 = vld1_lane_s16(&data_in[i + factor * 6 - j], in16x4_1, 2);
in16x4_1 = vld1_lane_s16(&data_in[i + factor * 7 - j], in16x4_1, 3);
// Mul and accumulate.
out32x4_0 = vmlal_lane_s16(out32x4_0, in16x4_0, coeff16x4, 0);
out32x4_1 = vmlal_lane_s16(out32x4_1, in16x4_1, coeff16x4, 0);
}
// Saturate and store the output.
int16x4_t out16x4_0 = vqshrn_n_s32(out32x4_0, 12);
int16x4_t out16x4_1 = vqshrn_n_s32(out32x4_1, 12);
vst1q_s16(data_out, vcombine_s16(out16x4_0, out16x4_1));
data_out += 8;
}
break;
}
}
// Second part, do the rest iterations (if any).
for (; i < endpos; i += factor) {
out_s32 = 2048; // Round value, 0.5 in Q12.
for (j = 0; j < (int)coefficients_length; j++) {
out_s32 = WebRtc_MulAccumW16(coefficients[j], data_in[i - j], out_s32);
}
// Saturate and store the output.
out_s32 >>= 12;
*data_out++ = WebRtcSpl_SatW32ToW16(out_s32);
}
return 0;
}

333
VocieProcess/common_audio/signal_processing/min_max_operations_neon.c Normal file
View File

@ -0,0 +1,333 @@
/*
* Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include <arm_neon.h>
#include <stdlib.h>
#include "rtc_base/checks.h"
#include "common_audio/signal_processing/include/signal_processing_library.h"
// Maximum absolute value of word16 vector. C version for generic platforms.
int16_t WebRtcSpl_MaxAbsValueW16Neon(const int16_t* vector, size_t length) {
int absolute = 0, maximum = 0;
RTC_DCHECK_GT(length, 0);
const int16_t* p_start = vector;
size_t rest = length & 7;
const int16_t* p_end = vector + length - rest;
int16x8_t v;
uint16x8_t max_qv;
max_qv = vdupq_n_u16(0);
while (p_start < p_end) {
v = vld1q_s16(p_start);
// Note vabs doesn't change the value of -32768.
v = vabsq_s16(v);
// Use u16 so we don't lose the value -32768.
max_qv = vmaxq_u16(max_qv, vreinterpretq_u16_s16(v));
p_start += 8;
}
#ifdef WEBRTC_ARCH_ARM64
maximum = (int)vmaxvq_u16(max_qv);
#else
uint16x4_t max_dv;
max_dv = vmax_u16(vget_low_u16(max_qv), vget_high_u16(max_qv));
max_dv = vpmax_u16(max_dv, max_dv);
max_dv = vpmax_u16(max_dv, max_dv);
maximum = (int)vget_lane_u16(max_dv, 0);
#endif
p_end = vector + length;
while (p_start < p_end) {
absolute = abs((int)(*p_start));
if (absolute > maximum) {
maximum = absolute;
}
p_start++;
}
// Guard the case for abs(-32768).
if (maximum > WEBRTC_SPL_WORD16_MAX) {
maximum = WEBRTC_SPL_WORD16_MAX;
}
return (int16_t)maximum;
}
// Maximum absolute value of word32 vector. NEON intrinsics version for
// ARM 32-bit/64-bit platforms.
int32_t WebRtcSpl_MaxAbsValueW32Neon(const int32_t* vector, size_t length) {
// Use uint32_t for the local variables, to accommodate the return value
// of abs(0x80000000), which is 0x80000000.
uint32_t absolute = 0, maximum = 0;
size_t i = 0;
size_t residual = length & 0x7;
RTC_DCHECK_GT(length, 0);
const int32_t* p_start = vector;
uint32x4_t max32x4_0 = vdupq_n_u32(0);
uint32x4_t max32x4_1 = vdupq_n_u32(0);
// First part, unroll the loop 8 times.
for (i = 0; i < length - residual; i += 8) {
int32x4_t in32x4_0 = vld1q_s32(p_start);
p_start += 4;
int32x4_t in32x4_1 = vld1q_s32(p_start);
p_start += 4;
in32x4_0 = vabsq_s32(in32x4_0);
in32x4_1 = vabsq_s32(in32x4_1);
// vabs doesn't change the value of 0x80000000.
// Use u32 so we don't lose the value 0x80000000.
max32x4_0 = vmaxq_u32(max32x4_0, vreinterpretq_u32_s32(in32x4_0));
max32x4_1 = vmaxq_u32(max32x4_1, vreinterpretq_u32_s32(in32x4_1));
}
uint32x4_t max32x4 = vmaxq_u32(max32x4_0, max32x4_1);
#if defined(WEBRTC_ARCH_ARM64)
maximum = vmaxvq_u32(max32x4);
#else
uint32x2_t max32x2 = vmax_u32(vget_low_u32(max32x4), vget_high_u32(max32x4));
max32x2 = vpmax_u32(max32x2, max32x2);
maximum = vget_lane_u32(max32x2, 0);
#endif
// Second part, do the remaining iterations (if any).
for (i = residual; i > 0; i--) {
absolute = abs((int)(*p_start));
if (absolute > maximum) {
maximum = absolute;
}
p_start++;
}
// Guard against the case for 0x80000000.
maximum = WEBRTC_SPL_MIN(maximum, WEBRTC_SPL_WORD32_MAX);
return (int32_t)maximum;
}
// Maximum value of word16 vector. NEON intrinsics version for
// ARM 32-bit/64-bit platforms.
int16_t WebRtcSpl_MaxValueW16Neon(const int16_t* vector, size_t length) {
int16_t maximum = WEBRTC_SPL_WORD16_MIN;
size_t i = 0;
size_t residual = length & 0x7;
RTC_DCHECK_GT(length, 0);
const int16_t* p_start = vector;
int16x8_t max16x8 = vdupq_n_s16(WEBRTC_SPL_WORD16_MIN);
// First part, unroll the loop 8 times.
for (i = 0; i < length - residual; i += 8) {
int16x8_t in16x8 = vld1q_s16(p_start);
max16x8 = vmaxq_s16(max16x8, in16x8);
p_start += 8;
}
#if defined(WEBRTC_ARCH_ARM64)
maximum = vmaxvq_s16(max16x8);
#else
int16x4_t max16x4 = vmax_s16(vget_low_s16(max16x8), vget_high_s16(max16x8));
max16x4 = vpmax_s16(max16x4, max16x4);
max16x4 = vpmax_s16(max16x4, max16x4);
maximum = vget_lane_s16(max16x4, 0);
#endif
// Second part, do the remaining iterations (if any).
for (i = residual; i > 0; i--) {
if (*p_start > maximum)
maximum = *p_start;
p_start++;
}
return maximum;
}
// Maximum value of word32 vector. NEON intrinsics version for
// ARM 32-bit/64-bit platforms.
int32_t WebRtcSpl_MaxValueW32Neon(const int32_t* vector, size_t length) {
int32_t maximum = WEBRTC_SPL_WORD32_MIN;
size_t i = 0;
size_t residual = length & 0x7;
RTC_DCHECK_GT(length, 0);
const int32_t* p_start = vector;
int32x4_t max32x4_0 = vdupq_n_s32(WEBRTC_SPL_WORD32_MIN);
int32x4_t max32x4_1 = vdupq_n_s32(WEBRTC_SPL_WORD32_MIN);
// First part, unroll the loop 8 times.
for (i = 0; i < length - residual; i += 8) {
int32x4_t in32x4_0 = vld1q_s32(p_start);
p_start += 4;
int32x4_t in32x4_1 = vld1q_s32(p_start);
p_start += 4;
max32x4_0 = vmaxq_s32(max32x4_0, in32x4_0);
max32x4_1 = vmaxq_s32(max32x4_1, in32x4_1);
}
int32x4_t max32x4 = vmaxq_s32(max32x4_0, max32x4_1);
#if defined(WEBRTC_ARCH_ARM64)
maximum = vmaxvq_s32(max32x4);
#else
int32x2_t max32x2 = vmax_s32(vget_low_s32(max32x4), vget_high_s32(max32x4));
max32x2 = vpmax_s32(max32x2, max32x2);
maximum = vget_lane_s32(max32x2, 0);
#endif
// Second part, do the remaining iterations (if any).
for (i = residual; i > 0; i--) {
if (*p_start > maximum)
maximum = *p_start;
p_start++;
}
return maximum;
}
// Minimum value of word16 vector. NEON intrinsics version for
// ARM 32-bit/64-bit platforms.
int16_t WebRtcSpl_MinValueW16Neon(const int16_t* vector, size_t length) {
int16_t minimum = WEBRTC_SPL_WORD16_MAX;
size_t i = 0;
size_t residual = length & 0x7;
RTC_DCHECK_GT(length, 0);
const int16_t* p_start = vector;
int16x8_t min16x8 = vdupq_n_s16(WEBRTC_SPL_WORD16_MAX);
// First part, unroll the loop 8 times.
for (i = 0; i < length - residual; i += 8) {
int16x8_t in16x8 = vld1q_s16(p_start);
min16x8 = vminq_s16(min16x8, in16x8);
p_start += 8;
}
#if defined(WEBRTC_ARCH_ARM64)
minimum = vminvq_s16(min16x8);
#else
int16x4_t min16x4 = vmin_s16(vget_low_s16(min16x8), vget_high_s16(min16x8));
min16x4 = vpmin_s16(min16x4, min16x4);
min16x4 = vpmin_s16(min16x4, min16x4);
minimum = vget_lane_s16(min16x4, 0);
#endif
// Second part, do the remaining iterations (if any).
for (i = residual; i > 0; i--) {
if (*p_start < minimum)
minimum = *p_start;
p_start++;
}
return minimum;
}
// Minimum value of word32 vector. NEON intrinsics version for
// ARM 32-bit/64-bit platforms.
int32_t WebRtcSpl_MinValueW32Neon(const int32_t* vector, size_t length) {
int32_t minimum = WEBRTC_SPL_WORD32_MAX;
size_t i = 0;
size_t residual = length & 0x7;
RTC_DCHECK_GT(length, 0);
const int32_t* p_start = vector;
int32x4_t min32x4_0 = vdupq_n_s32(WEBRTC_SPL_WORD32_MAX);
int32x4_t min32x4_1 = vdupq_n_s32(WEBRTC_SPL_WORD32_MAX);
// First part, unroll the loop 8 times.
for (i = 0; i < length - residual; i += 8) {
int32x4_t in32x4_0 = vld1q_s32(p_start);
p_start += 4;
int32x4_t in32x4_1 = vld1q_s32(p_start);
p_start += 4;
min32x4_0 = vminq_s32(min32x4_0, in32x4_0);
min32x4_1 = vminq_s32(min32x4_1, in32x4_1);
}
int32x4_t min32x4 = vminq_s32(min32x4_0, min32x4_1);
#if defined(WEBRTC_ARCH_ARM64)
minimum = vminvq_s32(min32x4);
#else
int32x2_t min32x2 = vmin_s32(vget_low_s32(min32x4), vget_high_s32(min32x4));
min32x2 = vpmin_s32(min32x2, min32x2);
minimum = vget_lane_s32(min32x2, 0);
#endif
// Second part, do the remaining iterations (if any).
for (i = residual; i > 0; i--) {
if (*p_start < minimum)
minimum = *p_start;
p_start++;
}
return minimum;
}
// Finds both the minimum and maximum elements in an array of 16-bit integers.
void WebRtcSpl_MinMaxW16Neon(const int16_t* vector, size_t length,
int16_t* min_val, int16_t* max_val) {
int16_t minimum = WEBRTC_SPL_WORD16_MAX;
int16_t maximum = WEBRTC_SPL_WORD16_MIN;
size_t i = 0;
size_t residual = length & 0x7;
RTC_DCHECK_GT(length, 0);
const int16_t* p_start = vector;
int16x8_t min16x8 = vdupq_n_s16(WEBRTC_SPL_WORD16_MAX);
int16x8_t max16x8 = vdupq_n_s16(WEBRTC_SPL_WORD16_MIN);
// First part, unroll the loop 8 times.
for (i = 0; i < length - residual; i += 8) {
int16x8_t in16x8 = vld1q_s16(p_start);
min16x8 = vminq_s16(min16x8, in16x8);
max16x8 = vmaxq_s16(max16x8, in16x8);
p_start += 8;
}
#if defined(WEBRTC_ARCH_ARM64)
minimum = vminvq_s16(min16x8);
maximum = vmaxvq_s16(max16x8);
#else
int16x4_t min16x4 = vmin_s16(vget_low_s16(min16x8), vget_high_s16(min16x8));
min16x4 = vpmin_s16(min16x4, min16x4);
min16x4 = vpmin_s16(min16x4, min16x4);
minimum = vget_lane_s16(min16x4, 0);
int16x4_t max16x4 = vmax_s16(vget_low_s16(max16x8), vget_high_s16(max16x8));
max16x4 = vpmax_s16(max16x4, max16x4);
max16x4 = vpmax_s16(max16x4, max16x4);
maximum = vget_lane_s16(max16x4, 0);
#endif
// Second part, do the remaining iterations (if any).
for (i = residual; i > 0; i--) {
if (*p_start < minimum)
minimum = *p_start;
if (*p_start > maximum)
maximum = *p_start;
p_start++;
}
*min_val = minimum;
*max_val = maximum;
}