1cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org/* 2cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org * Copyright (c) 2014 The WebRTC project authors. All Rights Reserved. 3cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org * 4cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org * Use of this source code is governed by a BSD-style license 5cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org * that can be found in the LICENSE file in the root of the source 6cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org * tree. An additional intellectual property rights grant can be found 7cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org * in the file PATENTS. All contributing project authors may 8cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org * be found in the AUTHORS file in the root of the source tree. 9cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org */ 10cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org 11cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org/* 12cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org * The rdft AEC algorithm, neon version of speed-critical functions. 13cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org * 14cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org * Based on the sse2 version. 15cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org */ 16cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org 17cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org 18cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org#include "webrtc/modules/audio_processing/aec/aec_rdft.h" 19cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org 20cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org#include <arm_neon.h> 21cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org 22cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.orgstatic const ALIGN16_BEG float ALIGN16_END 23cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org k_swap_sign[4] = {-1.f, 1.f, -1.f, 1.f}; 24cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org 25cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.orgstatic void cft1st_128_neon(float* a) { 26cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org const float32x4_t vec_swap_sign = vld1q_f32((float32_t*)k_swap_sign); 27cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org int j, k2; 28cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org 29cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org for (k2 = 0, j = 0; j < 128; j += 16, k2 += 4) { 30cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org float32x4_t a00v = vld1q_f32(&a[j + 0]); 31cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org float32x4_t a04v = vld1q_f32(&a[j + 4]); 32cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org float32x4_t a08v = vld1q_f32(&a[j + 8]); 33cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org float32x4_t a12v = vld1q_f32(&a[j + 12]); 34cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org float32x4_t a01v = vcombine_f32(vget_low_f32(a00v), vget_low_f32(a08v)); 35cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org float32x4_t a23v = vcombine_f32(vget_high_f32(a00v), vget_high_f32(a08v)); 36cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org float32x4_t a45v = vcombine_f32(vget_low_f32(a04v), vget_low_f32(a12v)); 37cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org float32x4_t a67v = vcombine_f32(vget_high_f32(a04v), vget_high_f32(a12v)); 38cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org const float32x4_t wk1rv = vld1q_f32(&rdft_wk1r[k2]); 39cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org const float32x4_t wk1iv = vld1q_f32(&rdft_wk1i[k2]); 40cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org const float32x4_t wk2rv = vld1q_f32(&rdft_wk2r[k2]); 41cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org const float32x4_t wk2iv = vld1q_f32(&rdft_wk2i[k2]); 42cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org const float32x4_t wk3rv = vld1q_f32(&rdft_wk3r[k2]); 43cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org const float32x4_t wk3iv = vld1q_f32(&rdft_wk3i[k2]); 44cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org float32x4_t x0v = vaddq_f32(a01v, a23v); 45cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org const float32x4_t x1v = vsubq_f32(a01v, a23v); 46cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org const float32x4_t x2v = vaddq_f32(a45v, a67v); 47cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org const float32x4_t x3v = vsubq_f32(a45v, a67v); 48cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org const float32x4_t x3w = vrev64q_f32(x3v); 49cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org float32x4_t x0w; 50cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org a01v = vaddq_f32(x0v, x2v); 51cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org x0v = vsubq_f32(x0v, x2v); 52cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org x0w = vrev64q_f32(x0v); 53cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org a45v = vmulq_f32(wk2rv, x0v); 54cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org a45v = vmlaq_f32(a45v, wk2iv, x0w); 55cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org x0v = vmlaq_f32(x1v, x3w, vec_swap_sign); 56cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org x0w = vrev64q_f32(x0v); 57cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org a23v = vmulq_f32(wk1rv, x0v); 58cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org a23v = vmlaq_f32(a23v, wk1iv, x0w); 59cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org x0v = vmlsq_f32(x1v, x3w, vec_swap_sign); 60cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org x0w = vrev64q_f32(x0v); 61cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org a67v = vmulq_f32(wk3rv, x0v); 62cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org a67v = vmlaq_f32(a67v, wk3iv, x0w); 63cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org a00v = vcombine_f32(vget_low_f32(a01v), vget_low_f32(a23v)); 64cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org a04v = vcombine_f32(vget_low_f32(a45v), vget_low_f32(a67v)); 65cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org a08v = vcombine_f32(vget_high_f32(a01v), vget_high_f32(a23v)); 66cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org a12v = vcombine_f32(vget_high_f32(a45v), vget_high_f32(a67v)); 67cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org vst1q_f32(&a[j + 0], a00v); 68cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org vst1q_f32(&a[j + 4], a04v); 69cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org vst1q_f32(&a[j + 8], a08v); 70cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org vst1q_f32(&a[j + 12], a12v); 71cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org } 72cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org} 73cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org 7459adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.orgstatic void cftmdl_128_neon(float* a) { 7559adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org int j; 7659adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const int l = 8; 7759adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t vec_swap_sign = vld1q_f32((float32_t*)k_swap_sign); 7859adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org float32x4_t wk1rv = vld1q_f32(cftmdl_wk1r); 7959adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org 8059adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org for (j = 0; j < l; j += 2) { 8159adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_00 = vld1_f32(&a[j + 0]); 8259adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_08 = vld1_f32(&a[j + 8]); 8359adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_32 = vld1_f32(&a[j + 32]); 8459adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_40 = vld1_f32(&a[j + 40]); 8559adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t a_00_32 = vcombine_f32(a_00, a_32); 8659adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t a_08_40 = vcombine_f32(a_08, a_40); 8759adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x0r0_0i0_0r1_x0i1 = vaddq_f32(a_00_32, a_08_40); 8859adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x1r0_1i0_1r1_x1i1 = vsubq_f32(a_00_32, a_08_40); 8959adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_16 = vld1_f32(&a[j + 16]); 9059adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_24 = vld1_f32(&a[j + 24]); 9159adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_48 = vld1_f32(&a[j + 48]); 9259adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_56 = vld1_f32(&a[j + 56]); 9359adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t a_16_48 = vcombine_f32(a_16, a_48); 9459adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t a_24_56 = vcombine_f32(a_24, a_56); 9559adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x2r0_2i0_2r1_x2i1 = vaddq_f32(a_16_48, a_24_56); 9659adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x3r0_3i0_3r1_x3i1 = vsubq_f32(a_16_48, a_24_56); 9759adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t xx0 = vaddq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1); 9859adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t xx1 = vsubq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1); 9959adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x3i0_3r0_3i1_x3r1 = vrev64q_f32(x3r0_3i0_3r1_x3i1); 10059adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x1_x3_add = 10159adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vmlaq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1); 10259adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x1_x3_sub = 10359adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vmlsq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1); 10459adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t yy0_a = vdup_lane_f32(vget_high_f32(x1_x3_add), 0); 10559adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t yy0_s = vdup_lane_f32(vget_high_f32(x1_x3_sub), 0); 10659adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t yy0_as = vcombine_f32(yy0_a, yy0_s); 10759adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t yy1_a = vdup_lane_f32(vget_high_f32(x1_x3_add), 1); 10859adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t yy1_s = vdup_lane_f32(vget_high_f32(x1_x3_sub), 1); 10959adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t yy1_as = vcombine_f32(yy1_a, yy1_s); 11059adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t yy0 = vmlaq_f32(yy0_as, vec_swap_sign, yy1_as); 11159adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t yy4 = vmulq_f32(wk1rv, yy0); 11259adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t xx1_rev = vrev64q_f32(xx1); 11359adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t yy4_rev = vrev64q_f32(yy4); 11459adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org 11559adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 0], vget_low_f32(xx0)); 11659adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 32], vget_high_f32(xx0)); 11759adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 16], vget_low_f32(xx1)); 11859adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 48], vget_high_f32(xx1_rev)); 11959adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org 12059adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org a[j + 48] = -a[j + 48]; 12159adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org 12259adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 8], vget_low_f32(x1_x3_add)); 12359adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 24], vget_low_f32(x1_x3_sub)); 12459adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 40], vget_low_f32(yy4)); 12559adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 56], vget_high_f32(yy4_rev)); 12659adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org } 12759adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org 12859adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org { 12959adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const int k = 64; 13059adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const int k1 = 2; 13159adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const int k2 = 2 * k1; 13259adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t wk2rv = vld1q_f32(&rdft_wk2r[k2 + 0]); 13359adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t wk2iv = vld1q_f32(&rdft_wk2i[k2 + 0]); 13459adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t wk1iv = vld1q_f32(&rdft_wk1i[k2 + 0]); 13559adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t wk3rv = vld1q_f32(&rdft_wk3r[k2 + 0]); 13659adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t wk3iv = vld1q_f32(&rdft_wk3i[k2 + 0]); 13759adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org wk1rv = vld1q_f32(&rdft_wk1r[k2 + 0]); 13859adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org for (j = k; j < l + k; j += 2) { 13959adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_00 = vld1_f32(&a[j + 0]); 14059adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_08 = vld1_f32(&a[j + 8]); 14159adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_32 = vld1_f32(&a[j + 32]); 14259adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_40 = vld1_f32(&a[j + 40]); 14359adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t a_00_32 = vcombine_f32(a_00, a_32); 14459adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t a_08_40 = vcombine_f32(a_08, a_40); 14559adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x0r0_0i0_0r1_x0i1 = vaddq_f32(a_00_32, a_08_40); 14659adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x1r0_1i0_1r1_x1i1 = vsubq_f32(a_00_32, a_08_40); 14759adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_16 = vld1_f32(&a[j + 16]); 14859adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_24 = vld1_f32(&a[j + 24]); 14959adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_48 = vld1_f32(&a[j + 48]); 15059adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x2_t a_56 = vld1_f32(&a[j + 56]); 15159adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t a_16_48 = vcombine_f32(a_16, a_48); 15259adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t a_24_56 = vcombine_f32(a_24, a_56); 15359adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x2r0_2i0_2r1_x2i1 = vaddq_f32(a_16_48, a_24_56); 15459adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x3r0_3i0_3r1_x3i1 = vsubq_f32(a_16_48, a_24_56); 15559adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t xx = vaddq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1); 15659adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t xx1 = vsubq_f32(x0r0_0i0_0r1_x0i1, x2r0_2i0_2r1_x2i1); 15759adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x3i0_3r0_3i1_x3r1 = vrev64q_f32(x3r0_3i0_3r1_x3i1); 15859adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x1_x3_add = 15959adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vmlaq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1); 16059adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org const float32x4_t x1_x3_sub = 16159adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vmlsq_f32(x1r0_1i0_1r1_x1i1, vec_swap_sign, x3i0_3r0_3i1_x3r1); 16259adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org float32x4_t xx4 = vmulq_f32(wk2rv, xx1); 16359adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org float32x4_t xx12 = vmulq_f32(wk1rv, x1_x3_add); 16459adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org float32x4_t xx22 = vmulq_f32(wk3rv, x1_x3_sub); 16559adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org xx4 = vmlaq_f32(xx4, wk2iv, vrev64q_f32(xx1)); 16659adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org xx12 = vmlaq_f32(xx12, wk1iv, vrev64q_f32(x1_x3_add)); 16759adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org xx22 = vmlaq_f32(xx22, wk3iv, vrev64q_f32(x1_x3_sub)); 16859adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org 16959adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 0], vget_low_f32(xx)); 17059adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 32], vget_high_f32(xx)); 17159adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 16], vget_low_f32(xx4)); 17259adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 48], vget_high_f32(xx4)); 17359adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 8], vget_low_f32(xx12)); 17459adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 40], vget_high_f32(xx12)); 17559adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 24], vget_low_f32(xx22)); 17659adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org vst1_f32(&a[j + 56], vget_high_f32(xx22)); 17759adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org } 17859adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org } 17959adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org} 18059adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org 181e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org__inline static float32x4_t reverse_order_f32x4(float32x4_t in) { 182e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // A B C D -> C D A B 183e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t rev = vcombine_f32(vget_high_f32(in), vget_low_f32(in)); 184e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // C D A B -> D C B A 185e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org return vrev64q_f32(rev); 186e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org} 187e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org 188e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.orgstatic void rftfsub_128_neon(float* a) { 189e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float* c = rdft_w + 32; 190ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org int j1, j2; 191e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t mm_half = vdupq_n_f32(0.5f); 192e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org 193e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // Vectorized code (four at once). 194e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // Note: commented number are indexes for the first iteration of the loop. 195e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org for (j1 = 1, j2 = 2; j2 + 7 < 64; j1 += 4, j2 += 8) { 196e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // Load 'wk'. 197ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t c_j1 = vld1q_f32(&c[j1]); // 1, 2, 3, 4, 198ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t c_k1 = vld1q_f32(&c[29 - j1]); // 28, 29, 30, 31, 199ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t wkrt = vsubq_f32(mm_half, c_k1); // 28, 29, 30, 31, 200ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t wkr_ = reverse_order_f32x4(wkrt); // 31, 30, 29, 28, 201ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t wki_ = c_j1; // 1, 2, 3, 4, 202e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // Load and shuffle 'a'. 203e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 2, 4, 6, 8, 3, 5, 7, 9 204e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org float32x4x2_t a_j2_p = vld2q_f32(&a[0 + j2]); 205e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 120, 122, 124, 126, 121, 123, 125, 127, 206e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4x2_t k2_0_4 = vld2q_f32(&a[122 - j2]); 207e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 126, 124, 122, 120 208e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t a_k2_p0 = reverse_order_f32x4(k2_0_4.val[0]); 209e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 127, 125, 123, 121 210e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t a_k2_p1 = reverse_order_f32x4(k2_0_4.val[1]); 211e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // Calculate 'x'. 212e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t xr_ = vsubq_f32(a_j2_p.val[0], a_k2_p0); 213e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 2-126, 4-124, 6-122, 8-120, 214e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t xi_ = vaddq_f32(a_j2_p.val[1], a_k2_p1); 215e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 3-127, 5-125, 7-123, 9-121, 216e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // Calculate product into 'y'. 217e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // yr = wkr * xr - wki * xi; 218e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // yi = wkr * xi + wki * xr; 219e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t a_ = vmulq_f32(wkr_, xr_); 220e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t b_ = vmulq_f32(wki_, xi_); 221e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t c_ = vmulq_f32(wkr_, xi_); 222e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t d_ = vmulq_f32(wki_, xr_); 223e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t yr_ = vsubq_f32(a_, b_); // 2-126, 4-124, 6-122, 8-120, 224e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t yi_ = vaddq_f32(c_, d_); // 3-127, 5-125, 7-123, 9-121, 225e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // Update 'a'. 226e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // a[j2 + 0] -= yr; 227e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // a[j2 + 1] -= yi; 228e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // a[k2 + 0] += yr; 229e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // a[k2 + 1] -= yi; 230e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 126, 124, 122, 120, 231e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t a_k2_p0n = vaddq_f32(a_k2_p0, yr_); 232e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 127, 125, 123, 121, 233e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t a_k2_p1n = vsubq_f32(a_k2_p1, yi_); 234e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // Shuffle in right order and store. 235e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t a_k2_p0nr = vrev64q_f32(a_k2_p0n); 236e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4_t a_k2_p1nr = vrev64q_f32(a_k2_p1n); 237e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 124, 125, 126, 127, 120, 121, 122, 123 238e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org const float32x4x2_t a_k2_n = vzipq_f32(a_k2_p0nr, a_k2_p1nr); 239e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 2, 4, 6, 8, 240e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org a_j2_p.val[0] = vsubq_f32(a_j2_p.val[0], yr_); 241e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 3, 5, 7, 9, 242e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org a_j2_p.val[1] = vsubq_f32(a_j2_p.val[1], yi_); 243e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // 2, 3, 4, 5, 6, 7, 8, 9, 244e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org vst2q_f32(&a[0 + j2], a_j2_p); 245e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org 246e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org vst1q_f32(&a[122 - j2], a_k2_n.val[1]); 247e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org vst1q_f32(&a[126 - j2], a_k2_n.val[0]); 248e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org } 249e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org 250e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org // Scalar code for the remaining items. 251e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org for (; j2 < 64; j1 += 1, j2 += 2) { 252ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const int k2 = 128 - j2; 253ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const int k1 = 32 - j1; 254ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float wkr = 0.5f - c[k1]; 255ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float wki = c[j1]; 256ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float xr = a[j2 + 0] - a[k2 + 0]; 257ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float xi = a[j2 + 1] + a[k2 + 1]; 258ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float yr = wkr * xr - wki * xi; 259ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float yi = wkr * xi + wki * xr; 260e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org a[j2 + 0] -= yr; 261e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org a[j2 + 1] -= yi; 262e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org a[k2 + 0] += yr; 263e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org a[k2 + 1] -= yi; 264e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org } 265e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org} 266e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org 267ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.orgstatic void rftbsub_128_neon(float* a) { 268ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float* c = rdft_w + 32; 269ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org int j1, j2; 270ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t mm_half = vdupq_n_f32(0.5f); 271ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org 272ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org a[1] = -a[1]; 273ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // Vectorized code (four at once). 274ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // Note: commented number are indexes for the first iteration of the loop. 275ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org for (j1 = 1, j2 = 2; j2 + 7 < 64; j1 += 4, j2 += 8) { 276ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // Load 'wk'. 277ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t c_j1 = vld1q_f32(&c[j1]); // 1, 2, 3, 4, 278ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t c_k1 = vld1q_f32(&c[29 - j1]); // 28, 29, 30, 31, 279ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t wkrt = vsubq_f32(mm_half, c_k1); // 28, 29, 30, 31, 280ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t wkr_ = reverse_order_f32x4(wkrt); // 31, 30, 29, 28, 281ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t wki_ = c_j1; // 1, 2, 3, 4, 282ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // Load and shuffle 'a'. 283ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 2, 4, 6, 8, 3, 5, 7, 9 284ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org float32x4x2_t a_j2_p = vld2q_f32(&a[0 + j2]); 285ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 120, 122, 124, 126, 121, 123, 125, 127, 286ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4x2_t k2_0_4 = vld2q_f32(&a[122 - j2]); 287ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 126, 124, 122, 120 288ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t a_k2_p0 = reverse_order_f32x4(k2_0_4.val[0]); 289ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 127, 125, 123, 121 290ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t a_k2_p1 = reverse_order_f32x4(k2_0_4.val[1]); 291ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // Calculate 'x'. 292ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t xr_ = vsubq_f32(a_j2_p.val[0], a_k2_p0); 293ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 2-126, 4-124, 6-122, 8-120, 294ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t xi_ = vaddq_f32(a_j2_p.val[1], a_k2_p1); 295ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 3-127, 5-125, 7-123, 9-121, 296ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // Calculate product into 'y'. 297ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // yr = wkr * xr - wki * xi; 298ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // yi = wkr * xi + wki * xr; 299ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t a_ = vmulq_f32(wkr_, xr_); 300ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t b_ = vmulq_f32(wki_, xi_); 301ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t c_ = vmulq_f32(wkr_, xi_); 302ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t d_ = vmulq_f32(wki_, xr_); 303ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t yr_ = vaddq_f32(a_, b_); // 2-126, 4-124, 6-122, 8-120, 304ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t yi_ = vsubq_f32(c_, d_); // 3-127, 5-125, 7-123, 9-121, 305ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // Update 'a'. 306ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // a[j2 + 0] -= yr; 307ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // a[j2 + 1] -= yi; 308ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // a[k2 + 0] += yr; 309ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // a[k2 + 1] -= yi; 310ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 126, 124, 122, 120, 311ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t a_k2_p0n = vaddq_f32(a_k2_p0, yr_); 312ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 127, 125, 123, 121, 313ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t a_k2_p1n = vsubq_f32(yi_, a_k2_p1); 314ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // Shuffle in right order and store. 315ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 2, 3, 4, 5, 6, 7, 8, 9, 316ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t a_k2_p0nr = vrev64q_f32(a_k2_p0n); 317ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4_t a_k2_p1nr = vrev64q_f32(a_k2_p1n); 318ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 124, 125, 126, 127, 120, 121, 122, 123 319ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float32x4x2_t a_k2_n = vzipq_f32(a_k2_p0nr, a_k2_p1nr); 320ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 2, 4, 6, 8, 321ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org a_j2_p.val[0] = vsubq_f32(a_j2_p.val[0], yr_); 322ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 3, 5, 7, 9, 323ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org a_j2_p.val[1] = vsubq_f32(yi_, a_j2_p.val[1]); 324ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // 2, 3, 4, 5, 6, 7, 8, 9, 325ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org vst2q_f32(&a[0 + j2], a_j2_p); 326ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org 327ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org vst1q_f32(&a[122 - j2], a_k2_n.val[1]); 328ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org vst1q_f32(&a[126 - j2], a_k2_n.val[0]); 329ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org } 330ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org 331ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org // Scalar code for the remaining items. 332ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org for (; j2 < 64; j1 += 1, j2 += 2) { 333ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const int k2 = 128 - j2; 334ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const int k1 = 32 - j1; 335ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float wkr = 0.5f - c[k1]; 336ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float wki = c[j1]; 337ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float xr = a[j2 + 0] - a[k2 + 0]; 338ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float xi = a[j2 + 1] + a[k2 + 1]; 339ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float yr = wkr * xr + wki * xi; 340ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org const float yi = wkr * xi - wki * xr; 341ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org a[j2 + 0] = a[j2 + 0] - yr; 342ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org a[j2 + 1] = yi - a[j2 + 1]; 343ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org a[k2 + 0] = yr + a[k2 + 0]; 344ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org a[k2 + 1] = yi - a[k2 + 1]; 345ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org } 346ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org a[65] = -a[65]; 347ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org} 348ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org 349cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.orgvoid aec_rdft_init_neon(void) { 350cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org cft1st_128 = cft1st_128_neon; 35159adb1dcd70c81ba7133dcad15518098cb856edabjornv@webrtc.org cftmdl_128 = cftmdl_128_neon; 352e55641d4f70024118a0af989fbf46a77ccd1a241bjornv@webrtc.org rftfsub_128 = rftfsub_128_neon; 353ac800c8004c718a87aa46e0795f3f654e29245f9bjornv@webrtc.org rftbsub_128 = rftbsub_128_neon; 354cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org} 355cd9b90ab539859034f0ffcac63b65e5939fb883ebjornv@webrtc.org 356