199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// 299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// Copyright (c) 2013 The WebRTC project authors. All Rights Reserved. 399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// 499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// Use of this source code is governed by a BSD-style license 599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// that can be found in the LICENSE file in the root of the source 699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// tree. An additional intellectual property rights grant can be found 799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// in the file PATENTS. All contributing project authors may 899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// be found in the AUTHORS file in the root of the source tree. 999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// 1099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// This is a modification of armSP_FFT_CToC_SC32_Radix2_unsafe_s.s 1199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// to support float instead of SC32. 1299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// 1399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 1499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// Description: 1599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// Compute a Radix 2 DIT in-order out-of-place FFT stage for an N point 1699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// complex signal. This handles the general stage, not the first or last 1799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// stage. 1899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// 1999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// 2099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 2199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 2299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// Include standard headers 2399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 24bdf981cb383b7ec472ee86d2fedb53937285f894rtoy@google.com#include "dl/api/arm/armCOMM_s.h" 25bdf981cb383b7ec472ee86d2fedb53937285f894rtoy@google.com#include "dl/api/arm/omxtypes_s.h" 2699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 2799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 2899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// Import symbols required from other files 2999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// (For example tables) 3099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 3199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 3299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 3399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// Set debugging level 3499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@//DEBUG_ON SETL {TRUE} 3599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 3699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 3799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 3899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// Guarding implementation by the processor name 3999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 4099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 4199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 4299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 4399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// Guarding implementation by the processor name 4499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 4599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 4699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@//Input Registers 4799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 4899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define pSrc r0 4999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define pDst r2 5099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define pTwiddle r1 5199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define subFFTNum r6 5299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define subFFTSize r7 5399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 5499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 5599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@//Output Registers 5699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 5799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 5899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@//Local Scratch Registers 5999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 6099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define outPointStep r3 6199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define pointStep r4 6299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define grpCount r5 6399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define setCount r8 6499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@//const RN 9 6599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define step r10 6699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define dstStep r11 6799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define pTable r9 6899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define pTmp r9 6999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 7099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com@// Neon Registers 7199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 7299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define dW D0.F32 7399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define dX0 D2.F32 7499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define dX1 D3.F32 7599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define dX2 D4.F32 7699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define dX3 D5.F32 7799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define dY0 D6.F32 7899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define dY1 D7.F32 7999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define dY2 D8.F32 8099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define dY3 D9.F32 8199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define qT0 D10.F32 8299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com#define qT1 D11.F32 8399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 8499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 8599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com .MACRO FFTSTAGE scaled, inverse, name 8699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 8799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// Define stack arguments 8899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 8999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 9099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// Update grpCount and grpSize rightaway inorder to reuse pGrpCount 9199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// and pGrpSize regs 9299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 9399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com LSR subFFTNum,subFFTNum,#1 @//grpSize 9499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com LSL grpCount,subFFTSize,#1 9599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 9699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 9799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// pT0+1 increments pT0 by 8 bytes 9899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// pT0+pointStep = increment of 8*pointStep bytes = 4*grpSize bytes 9999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com MOV pointStep,subFFTNum,LSL #2 10099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 10199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// update subFFTSize for the next stage 10299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com MOV subFFTSize,grpCount 10399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 10499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// pOut0+1 increments pOut0 by 8 bytes 10599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// pOut0+outPointStep == increment of 8*outPointStep bytes = 10699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// 4*size bytes 10799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com SMULBB outPointStep,grpCount,pointStep 10899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com LSL pointStep,pointStep,#1 10999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 11099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 11199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com RSB step,pointStep,#16 11299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com RSB dstStep,outPointStep,#16 11399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 11499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// Loop on the groups 11599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 11699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.comradix2GrpLoop\name : 11799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com MOV setCount,pointStep,LSR #3 11899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VLD1 dW,[pTwiddle],pointStep @//[wi | wr] 11999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 12099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 12199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// Loop on the sets 12299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 12399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 12499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.comradix2SetLoop\name : 12599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 12699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 12799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// point0: dX0-real part dX1-img part 12899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VLD2 {dX0,dX1},[pSrc],pointStep 12999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// point1: dX2-real part dX3-img part 13099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VLD2 {dX2,dX3},[pSrc],step 13199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 13299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com SUBS setCount,setCount,#2 13399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 13499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com .ifeqs "\inverse", "TRUE" 13599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VMUL qT0,dX2,dW[0] 13699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VMLA qT0,dX3,dW[1] @// real part 13799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VMUL qT1,dX3,dW[0] 13899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VMLS qT1,dX2,dW[1] @// imag part 13999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 14099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com .else 14199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 14299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VMUL qT0,dX2,dW[0] 14399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VMLS qT0,dX3,dW[1] @// real part 14499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VMUL qT1,dX3,dW[0] 14599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VMLA qT1,dX2,dW[1] @// imag part 14699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 14799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com .endif 14899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 14999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VSUB dY0,dX0,qT0 15099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VSUB dY1,dX1,qT1 15199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VADD dY2,dX0,qT0 15299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VADD dY3,dX1,qT1 15399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 15499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VST2 {dY0,dY1},[pDst],outPointStep 15599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// dstStep = -outPointStep + 16 15699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com VST2 {dY2,dY3},[pDst],dstStep 15799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 15899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com BGT radix2SetLoop\name 15999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 16099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com SUBS grpCount,grpCount,#2 16199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com ADD pSrc,pSrc,pointStep 16299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com BGT radix2GrpLoop\name 16399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 16499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 16599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// Reset and Swap pSrc and pDst for the next stage 16699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com MOV pTmp,pDst 16799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// pDst -= 4*size; pSrc -= 8*size bytes 16899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com SUB pDst,pSrc,outPointStep,LSL #1 16999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com SUB pSrc,pTmp,outPointStep 17099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 17199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// Reset pTwiddle for the next stage 17299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com @// pTwiddle -= 4*size bytes 17399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com SUB pTwiddle,pTwiddle,outPointStep 17499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 17599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 17699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com .endm 17799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 17899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 17999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 18099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com M_START armSP_FFTFwd_CToC_FC32_Radix2_OutOfPlace_unsafe,r4 18199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com FFTSTAGE "FALSE","FALSE",FWD 18299b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com M_END 18399b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 18499b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 18599b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 18699b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com M_START armSP_FFTInv_CToC_FC32_Radix2_OutOfPlace_unsafe,r4 18799b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com FFTSTAGE "FALSE","TRUE",INV 18899b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com M_END 18999b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 19099b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com 19199b31ea5d0f4629ceddf4852f96757c9c73654artoy@google.com .end 192