src/third_party/openmax_dl/dl/sp/src/arm/neon/omxSP_FFTFwd_CToC_SC16_Sfs_s.S

   1 @//
   2 @//  Copyright (c) 2013 The WebRTC project authors. All Rights Reserved.
   3 @//
   4 @//  Use of this source code is governed by a BSD-style license
   5 @//  that can be found in the LICENSE file in the root of the source
   6 @//  tree. An additional intellectual property rights grant can be found
   7 @//  in the file PATENTS.  All contributing project authors may
   8 @//  be found in the AUTHORS file in the root of the source tree.
   9 @//
  10 @//  This file was originally licensed as follows. It has been
  11 @//  relicensed with permission from the copyright holders.
  12
  13 @//
  14 @//
  15 @// File Name:  omxSP_FFTFwd_CToC_SC16_Sfs_s.s
  16 @// OpenMAX DL: v1.0.2
  17 @// Last Modified Revision:   6729
  18 @// Last Modified Date:       Tue, 17 Jul 2007
  19 @//
  20 @// (c) Copyright 2007-2008 ARM Limited. All Rights Reserved.
  21 @//
  22 @//
  23 @//
  24 @// Description:
  25 @// Compute an inverse FFT for a complex signal
  26 @//
  27 @//
  28
  29
  30 @// Include standard headers
  31
  32 #include "dl/api/arm/armCOMM_s.h"
  33 #include "dl/api/arm/omxtypes_s.h"
  34
  35
  36 @// Import symbols required from other files
  37 @// (For example tables)
  38
  39         .extern  armSP_FFTFwd_CToC_SC16_Sfs_Radix2_fs_OutOfPlace_unsafe
  40         .extern  armSP_FFTFwd_CToC_SC16_Radix2_fs_OutOfPlace_unsafe
  41         .extern  armSP_FFTFwd_CToC_SC16_Radix4_fs_OutOfPlace_unsafe
  42         .extern  armSP_FFTFwd_CToC_SC16_Radix4_ls_OutOfPlace_unsafe
  43         .extern  armSP_FFTFwd_CToC_SC16_Radix8_fs_OutOfPlace_unsafe
  44         .extern  armSP_FFTFwd_CToC_SC16_Radix4_OutOfPlace_unsafe
  45         .extern  armSP_FFTFwd_CToC_SC16_Sfs_Radix4_fs_OutOfPlace_unsafe
  46         .extern  armSP_FFTFwd_CToC_SC16_Sfs_Radix4_ls_OutOfPlace_unsafe
  47         .extern  armSP_FFTFwd_CToC_SC16_Sfs_Radix8_fs_OutOfPlace_unsafe
  48         .extern  armSP_FFTFwd_CToC_SC16_Sfs_Radix4_OutOfPlace_unsafe
  49         .extern  armSP_FFTFwd_CToC_SC16_Sfs_Radix2_OutOfPlace_unsafe
  50         .extern  armSP_FFTFwd_CToC_SC16_Radix2_OutOfPlace_unsafe
  51         .extern  armSP_FFTFwd_CToC_SC16_Sfs_Radix2_ls_OutOfPlace_unsafe
  52         .extern  armSP_FFTFwd_CToC_SC16_Radix2_ls_OutOfPlace_unsafe
  53
  54 @// Set debugging level
  55 @//DEBUG_ON    SETL {TRUE}
  56
  57
  58
  59 @// Guarding implementation by the processor name
  60
  61
  62
  63 @// Guarding implementation by the processor name
  64
  65
  66     .extern  armSP_FFTFwd_CToC_SC16_Radix2_ps_OutOfPlace_unsafe
  67     .extern  armSP_FFTFwd_CToC_SC16_Sfs_Radix2_ps_OutOfPlace_unsafe
  68
  69 @//Input Registers
  70
  71 #define pSrc            r0
  72 #define pDst            r1
  73 #define pFFTSpec                r2
  74 #define scale           r3
  75
  76
  77 @// Output registers
  78 #define result          r0
  79
  80 @//Local Scratch Registers
  81 #define argTwiddle              r1
  82 #define argDst          r2
  83 #define argScale                r4
  84 #define pTwiddle                r4
  85 #define tmpOrder                r4
  86 #define pOut            r5
  87 #define subFFTSize              r7
  88 #define subFFTNum               r6
  89 #define N               r6
  90 #define order           r14
  91 #define diff            r9
  92 @// Total num of radix stages required to comple the FFT
  93 #define count           r8
  94 #define x0r             r4
  95 #define x0i             r5
  96 #define diffMinusOne            r2
  97 #define round           r3
  98
  99 @// Neon registers
 100
 101 #define dX0     D0.S16
 102 #define dShift  D1.S16
 103 #define dX0S32  D0.S32
 104
 105
 106
 107     @// Allocate stack memory required by the function
 108         M_ALLOC4        diffOnStack, 4
 109
 110     @// Write function header
 111         M_START     omxSP_FFTFwd_CToC_SC16_Sfs,r11,d15
 112
 113 @ Structure offsets for the FFTSpec
 114         .set    ARMsFFTSpec_N, 0
 115         .set    ARMsFFTSpec_pBitRev, 4
 116         .set    ARMsFFTSpec_pTwiddle, 8
 117         .set    ARMsFFTSpec_pBuf, 12
 118
 119         @// Define stack arguments
 120
 121         @// Read the size from structure and take log
 122         LDR     N, [pFFTSpec, #ARMsFFTSpec_N]
 123
 124         @// Read other structure parameters
 125         LDR     pTwiddle, [pFFTSpec, #ARMsFFTSpec_pTwiddle]
 126         LDR     pOut, [pFFTSpec, #ARMsFFTSpec_pBuf]
 127
 128         CLZ     order,N                             @// N = 2^order
 129         RSB     order,order,#31
 130         MOV     subFFTSize,#1
 131         @//MOV     subFFTNum,N
 132
 133         CMP     order,#3
 134         BGT     orderGreaterthan3                   @// order > 3
 135
 136         CMP     order,#1
 137         BGE     orderGreaterthan0                   @// order > 0
 138         M_STR   scale, diffOnStack,LT               @// order = 0
 139         LDRLT   x0r,[pSrc]
 140         STRLT   x0r,[pDst]
 141         MOVLT   pSrc,pDst
 142         BLT     FFTEnd
 143
 144 orderGreaterthan0:
 145         @// set the buffers appropriately for various orders
 146         CMP     order,#2
 147         MOVNE   argDst,pDst
 148         MOVEQ   argDst,pOut
 149         MOVEQ   pOut,pDst                           @// Pass the first stage destination in RN5
 150         MOV     argTwiddle,pTwiddle
 151
 152         SUBS     diff,scale,order
 153         M_STR   diff,diffOnStack
 154         MOVGT   scale,order
 155         @// Now scale <= order
 156
 157         CMP     order,#1
 158         BGT     orderGreaterthan1
 159         SUBS    scale,scale,#1
 160         BLEQ    armSP_FFTFwd_CToC_SC16_Sfs_Radix2_fs_OutOfPlace_unsafe  @// order = 1
 161         BLLT    armSP_FFTFwd_CToC_SC16_Radix2_fs_OutOfPlace_unsafe      @// order = 1
 162         B       FFTEnd
 163
 164 orderGreaterthan1:
 165         CMP     order,#2
 166         MOV     argScale,scale
 167         BGT     orderGreaterthan2
 168         SUBS    argScale,argScale,#1
 169         BLGE    armSP_FFTFwd_CToC_SC16_Sfs_Radix2_fs_OutOfPlace_unsafe      @// order =2
 170         BLLT    armSP_FFTFwd_CToC_SC16_Radix2_fs_OutOfPlace_unsafe
 171         SUBS    argScale,argScale,#1
 172         BLEQ    armSP_FFTFwd_CToC_SC16_Sfs_Radix2_ls_OutOfPlace_unsafe
 173         BLLT    armSP_FFTFwd_CToC_SC16_Radix2_ls_OutOfPlace_unsafe
 174         B       FFTEnd
 175
 176 orderGreaterthan2:                                                                     @// order =3
 177         SUBS    argScale,argScale,#1
 178         BLGE    armSP_FFTFwd_CToC_SC16_Sfs_Radix2_fs_OutOfPlace_unsafe
 179         BLLT    armSP_FFTFwd_CToC_SC16_Radix2_fs_OutOfPlace_unsafe
 180         SUBS    argScale,argScale,#1
 181         BLGE    armSP_FFTFwd_CToC_SC16_Sfs_Radix2_ps_OutOfPlace_unsafe
 182         BLLT    armSP_FFTFwd_CToC_SC16_Radix2_ps_OutOfPlace_unsafe
 183         SUBS    argScale,argScale,#1
 184         BLEQ    armSP_FFTFwd_CToC_SC16_Sfs_Radix2_ls_OutOfPlace_unsafe
 185         BLLT    armSP_FFTFwd_CToC_SC16_Radix2_ls_OutOfPlace_unsafe
 186         B       FFTEnd
 187
 188
 189 orderGreaterthan3:
 190         @// check scale = 0 or scale = order
 191         SUBS    diff, scale, order                 @// scale > order
 192         MOVGT   scale,order
 193         BGE     specialScaleCase                   @// scale = 0 or scale = order
 194         CMP     scale,#0
 195         BEQ     specialScaleCase
 196         B       generalScaleCase
 197
 198 specialScaleCase:                                           @//  scale = 0 or scale = order  and order > 3
 199
 200         TST     order, #2                           @// Set input args to fft stages
 201         MOVNE   argDst,pDst
 202         MOVEQ   argDst,pOut
 203         MOVEQ   pOut,pDst                           @// Pass the first stage destination in RN5
 204         MOV     argTwiddle,pTwiddle
 205
 206         CMP      diff,#0
 207         M_STR    diff, diffOnStack
 208         BGE      scaleEqualsOrder
 209
 210         @//check for even or odd order
 211         @// NOTE: The following combination of BL's would work fine eventhough the first
 212         @// BL would corrupt the flags. This is because the end of the "grpZeroSetLoop" loop inside
 213         @// armSP_FFTFwd_CToC_SC16_Radix4_fs_OutOfPlace_unsafe sets the Z flag to EQ
 214
 215         TST     order,#0x00000001
 216         BLEQ    armSP_FFTFwd_CToC_SC16_Radix4_fs_OutOfPlace_unsafe
 217         BLNE    armSP_FFTFwd_CToC_SC16_Radix8_fs_OutOfPlace_unsafe
 218
 219         CMP        subFFTNum,#4
 220         BLT     FFTEnd
 221
 222 unscaledRadix4Loop:
 223         BEQ        lastStageUnscaledRadix4
 224         BL        armSP_FFTFwd_CToC_SC16_Radix4_OutOfPlace_unsafe
 225          CMP        subFFTNum,#4
 226          B        unscaledRadix4Loop
 227
 228 lastStageUnscaledRadix4:
 229         BL      armSP_FFTFwd_CToC_SC16_Radix4_ls_OutOfPlace_unsafe
 230         B        FFTEnd
 231
 232 scaleEqualsOrder:
 233         @//check for even or odd order
 234         @// NOTE: The following combination of BL's would work fine eventhough the first
 235         @// BL would corrupt the flags. This is because the end of the "grpZeroSetLoop" loop inside
 236         @// armSP_FFTFwd_CToC_SC32_Radix4_fs_OutOfPlace_unsafe sets the Z flag to EQ
 237
 238         TST     order,#0x00000001
 239         BLEQ    armSP_FFTFwd_CToC_SC16_Sfs_Radix4_fs_OutOfPlace_unsafe
 240         BLNE    armSP_FFTFwd_CToC_SC16_Sfs_Radix8_fs_OutOfPlace_unsafe
 241
 242         CMP        subFFTNum,#4
 243         BLT     FFTEnd
 244
 245 scaledRadix4Loop:
 246         BEQ        lastStageScaledRadix4
 247         BL        armSP_FFTFwd_CToC_SC16_Sfs_Radix4_OutOfPlace_unsafe
 248          CMP        subFFTNum,#4
 249          B        scaledRadix4Loop
 250
 251 lastStageScaledRadix4:
 252         BL      armSP_FFTFwd_CToC_SC16_Sfs_Radix4_ls_OutOfPlace_unsafe
 253         B        FFTEnd
 254
 255
 256
 257 generalScaleCase:                                               @// 0 < scale < order and order > 3
 258         @// Determine the correct destination buffer
 259         SUB     diff,order,scale
 260         TST     diff,#0x01
 261         ADDEQ   count,scale,diff,LSR #1         @// count = scale + (order - scale)/2
 262         MOVNE   count,order
 263         TST     count,#0x01                     @// Is count even or odd ?
 264
 265         MOVNE   argDst,pDst                     @// Set input args to fft stages
 266         MOVEQ   argDst,pOut
 267         MOVEQ   pOut,pDst                       @// Pass the first stage destination in RN5
 268         MOV     argTwiddle,pTwiddle
 269
 270         CMP     diff,#1
 271         M_STR   diff, diffOnStack
 272         BEQ     scaleps                         @// scaling including a radix2_ps stage
 273
 274         MOV     argScale,scale                  @// Put scale in RN4 so as to save and restore
 275         BL      armSP_FFTFwd_CToC_SC16_Sfs_Radix2_fs_OutOfPlace_unsafe     @// scaled first stage
 276         SUBS    argScale,argScale,#1
 277
 278 scaledRadix2Loop:
 279         BLGT    armSP_FFTFwd_CToC_SC16_Sfs_Radix2_OutOfPlace_unsafe
 280         SUBS    argScale,argScale,#1            @// save and restore scale (RN4) in the scaled stages
 281         BGT     scaledRadix2Loop
 282         B       outScale
 283
 284 scaleps:
 285         SUB     argScale,scale,#1                   @// order>3 and diff=1 => scale >= 3
 286         BL      armSP_FFTFwd_CToC_SC16_Sfs_Radix2_fs_OutOfPlace_unsafe     @// scaled first stage
 287         SUBS    argScale,argScale,#1
 288
 289 scaledRadix2psLoop:
 290         BEQ     scaledRadix2psStage
 291         BLGT    armSP_FFTFwd_CToC_SC16_Sfs_Radix2_OutOfPlace_unsafe
 292         SUBS    argScale,argScale,#1            @// save and restore scale (RN4) in the scaled stages
 293         BGE     scaledRadix2psLoop
 294
 295 scaledRadix2psStage:
 296         BL      armSP_FFTFwd_CToC_SC16_Sfs_Radix2_ps_OutOfPlace_unsafe
 297         B       generalLastStageUnscaledRadix2
 298
 299
 300 outScale:
 301         M_LDR   diff, diffOnStack
 302         @//check for even or odd order
 303         TST     diff,#0x00000001
 304         BEQ     generalUnscaledRadix4Loop
 305         B       unscaledRadix2Loop
 306
 307 generalUnscaledRadix4Loop:
 308         CMP        subFFTNum,#4
 309          BEQ        generalLastStageUnscaledRadix4
 310          BL        armSP_FFTFwd_CToC_SC16_Radix4_OutOfPlace_unsafe
 311          B        generalUnscaledRadix4Loop
 312
 313 generalLastStageUnscaledRadix4:
 314         BL      armSP_FFTFwd_CToC_SC16_Radix4_ls_OutOfPlace_unsafe
 315         B        End
 316
 317 unscaledRadix2Loop:
 318         CMP        subFFTNum,#4
 319          BEQ        generalLastTwoStagesUnscaledRadix2
 320          BL        armSP_FFTFwd_CToC_SC16_Radix2_OutOfPlace_unsafe
 321          B        unscaledRadix2Loop
 322
 323 generalLastTwoStagesUnscaledRadix2:
 324         BL      armSP_FFTFwd_CToC_SC16_Radix2_ps_OutOfPlace_unsafe
 325 generalLastStageUnscaledRadix2:
 326         BL      armSP_FFTFwd_CToC_SC16_Radix2_ls_OutOfPlace_unsafe
 327         B        End
 328
 329
 330 FFTEnd:                                               @// Does only the scaling
 331
 332         M_LDR   diff, diffOnStack
 333         CMP     diff,#0
 334         BLE     End
 335
 336         RSB     diff,diff,#0                        @// to use VRSHL for right shift by a variable
 337         VDUP    dShift,diff
 338
 339 scaleFFTData:                                           @// N = subFFTSize  ; dataptr = pDst  ; scale = diff
 340         VLD1    {dX0S32[0]},[pSrc]                        @// pSrc contains pDst pointer
 341         SUBS    subFFTSize,subFFTSize,#1
 342         VRSHL   dX0,dShift
 343         VST1    {dX0S32[0]},[pSrc]!
 344
 345         BGT     scaleFFTData
 346
 347
 348
 349 End:
 350         @// Set return value
 351         MOV     result, #OMX_Sts_NoErr
 352
 353         @// Write function tail
 354         M_END
 355
 356     .end