Mirrors/overte-JulianGro
3
0
Fork 0
mirror of https://github.com/JulianGro/overte.git synced 2025-04-14 00:46:57 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

HRTF optimized for AVX512

Browse source
This commit is contained in:
Ken Cooke 2017-06-20 17:09:38 -07:00
parent b27b09ebc6
commit 1688ad5f0a
1 changed files with 94 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

94
libraries/audio/src/avx512/AudioHRTF_avx512.cpp Normal file
View file

@ -0,0 +1,94 @@
//
// AudioHRTF_avx512.cpp
// libraries/audio/src
//
// Created by Ken Cooke on 6/20/17.
// Copyright 2017 High Fidelity, Inc.
//
// Distributed under the Apache License, Version 2.0.
// See the accompanying file LICENSE or http://www.apache.org/licenses/LICENSE-2.0.html
//
#if defined(_M_IX86) || defined(_M_X64) || defined(__i386__) || defined(__x86_64__)
#include <assert.h>
#include <immintrin.h>
#include "../AudioHRTF.h"
#ifndef __AVX512F__
#error Must be compiled with /arch:AVX512 or -mavx512f.
#endif
#if defined(__GNUC__) && !defined(__clang__)
// for some reason, GCC -O2 results in poorly optimized code
#pragma GCC optimize("Os")
#endif
// 1 channel input, 4 channel output
void FIR_1x4_AVX512(float* src, float* dst0, float* dst1, float* dst2, float* dst3, float coef[4][HRTF_TAPS], int numFrames) {
float* coef0 = coef[0] + HRTF_TAPS - 1; // process backwards
float* coef1 = coef[1] + HRTF_TAPS - 1;
float* coef2 = coef[2] + HRTF_TAPS - 1;
float* coef3 = coef[3] + HRTF_TAPS - 1;
assert(numFrames % 16 == 0);
for (int i = 0; i < numFrames; i += 16) {
__m512 acc0 = _mm512_setzero_ps();
__m512 acc1 = _mm512_setzero_ps();
__m512 acc2 = _mm512_setzero_ps();
__m512 acc3 = _mm512_setzero_ps();
__m512 acc4 = _mm512_setzero_ps();
__m512 acc5 = _mm512_setzero_ps();
__m512 acc6 = _mm512_setzero_ps();
__m512 acc7 = _mm512_setzero_ps();
float* ps = &src[i - HRTF_TAPS + 1]; // process forwards
assert(HRTF_TAPS % 4 == 0);
for (int k = 0; k < HRTF_TAPS; k += 4) {
__m512 x0 = _mm512_loadu_ps(&ps[k+0]);
acc0 = _mm512_fmadd_ps(_mm512_set1_ps(coef0[-k-0]), x0, acc0); // vfmadd231ps acc0, x0, dword ptr [coef]{1to16}
acc1 = _mm512_fmadd_ps(_mm512_set1_ps(coef1[-k-0]), x0, acc1);
acc2 = _mm512_fmadd_ps(_mm512_set1_ps(coef2[-k-0]), x0, acc2);
acc3 = _mm512_fmadd_ps(_mm512_set1_ps(coef3[-k-0]), x0, acc3);
__m512 x1 = _mm512_loadu_ps(&ps[k+1]);
acc4 = _mm512_fmadd_ps(_mm512_set1_ps(coef0[-k-1]), x1, acc4);
acc5 = _mm512_fmadd_ps(_mm512_set1_ps(coef1[-k-1]), x1, acc5);
acc6 = _mm512_fmadd_ps(_mm512_set1_ps(coef2[-k-1]), x1, acc6);
acc7 = _mm512_fmadd_ps(_mm512_set1_ps(coef3[-k-1]), x1, acc7);
__m512 x2 = _mm512_loadu_ps(&ps[k+2]);
acc0 = _mm512_fmadd_ps(_mm512_set1_ps(coef0[-k-2]), x2, acc0);
acc1 = _mm512_fmadd_ps(_mm512_set1_ps(coef1[-k-2]), x2, acc1);
acc2 = _mm512_fmadd_ps(_mm512_set1_ps(coef2[-k-2]), x2, acc2);
acc3 = _mm512_fmadd_ps(_mm512_set1_ps(coef3[-k-2]), x2, acc3);
__m512 x3 = _mm512_loadu_ps(&ps[k+3]);
acc4 = _mm512_fmadd_ps(_mm512_set1_ps(coef0[-k-3]), x3, acc4);
acc5 = _mm512_fmadd_ps(_mm512_set1_ps(coef1[-k-3]), x3, acc5);
acc6 = _mm512_fmadd_ps(_mm512_set1_ps(coef2[-k-3]), x3, acc6);
acc7 = _mm512_fmadd_ps(_mm512_set1_ps(coef3[-k-3]), x3, acc7);
}
acc0 = _mm512_add_ps(acc0, acc4);
acc1 = _mm512_add_ps(acc1, acc5);
acc2 = _mm512_add_ps(acc2, acc6);
acc3 = _mm512_add_ps(acc3, acc7);
_mm512_storeu_ps(&dst0[i], acc0);
_mm512_storeu_ps(&dst1[i], acc1);
_mm512_storeu_ps(&dst2[i], acc2);
_mm512_storeu_ps(&dst3[i], acc3);
}
_mm256_zeroupper();
}
#endif