OLD | NEW |
(Empty) | |
| 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_FFTInv_CCSToR_S32_Sfs_s.s |
| 16 @// OpenMAX DL: v1.0.2 |
| 17 @// Last Modified Revision: 7469 |
| 18 @// Last Modified Date: Thu, 20 Sep 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/armCOMM_s.h" |
| 33 #include "dl/api/omxtypes_s.h" |
| 34 |
| 35 |
| 36 @// Import symbols required from other files |
| 37 @// (For example tables) |
| 38 |
| 39 .extern armSP_FFTInv_CToC_SC32_Sfs_Radix2_fs_OutOfPlace_unsafe |
| 40 .extern armSP_FFTInv_CToC_SC32_Radix2_fs_OutOfPlace_unsafe |
| 41 .extern armSP_FFTInv_CToC_SC32_Radix4_fs_OutOfPlace_unsafe |
| 42 .extern armSP_FFTInv_CToC_SC32_Radix8_fs_OutOfPlace_unsafe |
| 43 .extern armSP_FFTInv_CToC_SC32_Radix4_OutOfPlace_unsafe |
| 44 .extern armSP_FFTInv_CToC_SC32_Sfs_Radix4_fs_OutOfPlace_unsafe |
| 45 .extern armSP_FFTInv_CToC_SC32_Sfs_Radix8_fs_OutOfPlace_unsafe |
| 46 .extern armSP_FFTInv_CToC_SC32_Sfs_Radix4_OutOfPlace_unsafe |
| 47 .extern armSP_FFTInv_CToC_SC32_Sfs_Radix2_OutOfPlace_unsafe |
| 48 .extern armSP_FFTInv_CToC_SC32_Radix2_OutOfPlace_unsafe |
| 49 .extern armSP_FFTInv_CCSToR_S32_Sfs_preTwiddleRadix2_unsafe |
| 50 .extern armSP_FFTInv_CCSToR_S32_preTwiddleRadix2_unsafe |
| 51 |
| 52 |
| 53 @// Set debugging level |
| 54 @//DEBUG_ON SETL {TRUE} |
| 55 |
| 56 |
| 57 |
| 58 @// Guarding implementation by the processor name |
| 59 |
| 60 |
| 61 |
| 62 @// Guarding implementation by the processor name |
| 63 |
| 64 @// Import symbols required from other files |
| 65 @// (For example tables) |
| 66 .extern armSP_FFTInv_CToC_SC32_Radix4_ls_OutOfPlace_unsafe |
| 67 .extern armSP_FFTInv_CToC_SC32_Radix2_ls_OutOfPlace_unsafe |
| 68 .extern armSP_FFTInv_CToC_SC32_Sfs_Radix2_ls_OutOfPlace_unsafe |
| 69 .extern armSP_FFTInv_CToC_SC32_Sfs_Radix4_ls_OutOfPlace_unsafe |
| 70 |
| 71 |
| 72 @//Input Registers |
| 73 |
| 74 #define pSrc r0 |
| 75 #define pDst r1 |
| 76 #define pFFTSpec r2 |
| 77 #define scale r3 |
| 78 |
| 79 |
| 80 @// Output registers |
| 81 #define result r0 |
| 82 |
| 83 @//Local Scratch Registers |
| 84 |
| 85 #define argTwiddle r1 |
| 86 #define argDst r2 |
| 87 #define argScale r4 |
| 88 #define tmpOrder r4 |
| 89 #define pTwiddle r4 |
| 90 #define pOut r5 |
| 91 #define subFFTSize r7 |
| 92 #define subFFTNum r6 |
| 93 #define N r6 |
| 94 #define order r14 |
| 95 #define diff r9 |
| 96 @// Total num of radix stages required to comple the FFT |
| 97 #define count r8 |
| 98 #define x0r r4 |
| 99 #define x0i r5 |
| 100 #define diffMinusOne r2 |
| 101 #define round r3 |
| 102 |
| 103 #define pOut1 r2 |
| 104 #define size r7 |
| 105 #define step r8 |
| 106 #define step1 r9 |
| 107 #define twStep r10 |
| 108 #define pTwiddleTmp r11 |
| 109 #define argTwiddle1 r12 |
| 110 #define zero r14 |
| 111 |
| 112 @// Neon registers |
| 113 |
| 114 #define dX0 D0.S32 |
| 115 #define dShift D1.S32 |
| 116 #define dX1 D1.S32 |
| 117 #define dY0 D2.S32 |
| 118 #define dY1 D3.S32 |
| 119 #define dX0r D0.S32 |
| 120 #define dX0i D1.S32 |
| 121 #define dX1r D2.S32 |
| 122 #define dX1i D3.S32 |
| 123 #define dW0r D4.S32 |
| 124 #define dW0i D5.S32 |
| 125 #define dW1r D6.S32 |
| 126 #define dW1i D7.S32 |
| 127 #define dT0 D8.S32 |
| 128 #define dT1 D9.S32 |
| 129 #define dT2 D10.S32 |
| 130 #define dT3 D11.S32 |
| 131 #define qT0 Q6.S64 |
| 132 #define qT1 Q7.S64 |
| 133 #define qT2 Q8.S64 |
| 134 #define qT3 Q9.S64 |
| 135 #define dY0r D4.S32 |
| 136 #define dY0i D5.S32 |
| 137 #define dY1r D6.S32 |
| 138 #define dY1i D7.S32 |
| 139 #define dzero D20.S32 |
| 140 |
| 141 #define dY2 D4.S32 |
| 142 #define dY3 D5.S32 |
| 143 #define dW0 D6.S32 |
| 144 #define dW1 D7.S32 |
| 145 #define dW0Tmp D10.S32 |
| 146 #define dW1Neg D11.S32 |
| 147 |
| 148 |
| 149 |
| 150 @// Allocate stack memory required by the function |
| 151 M_ALLOC4 diffOnStack, 4 |
| 152 |
| 153 @// Write function header |
| 154 M_START omxSP_FFTInv_CCSToR_S32_Sfs,r11,d15 |
| 155 |
| 156 @ Structure offsets for the FFTSpec |
| 157 .set ARMsFFTSpec_N, 0 |
| 158 .set ARMsFFTSpec_pBitRev, 4 |
| 159 .set ARMsFFTSpec_pTwiddle, 8 |
| 160 .set ARMsFFTSpec_pBuf, 12 |
| 161 |
| 162 @// Define stack arguments |
| 163 |
| 164 @// Read the size from structure and take log |
| 165 LDR N, [pFFTSpec, #ARMsFFTSpec_N] |
| 166 |
| 167 @// Read other structure parameters |
| 168 LDR pTwiddle, [pFFTSpec, #ARMsFFTSpec_pTwiddle] |
| 169 LDR pOut, [pFFTSpec, #ARMsFFTSpec_pBuf] |
| 170 |
| 171 @// N=1 Treat seperately |
| 172 CMP N,#1 |
| 173 BGT sizeGreaterThanOne |
| 174 VLD1 dX0[0],[pSrc] |
| 175 RSB scale,scale,#0 @// to use VRSHL for right
shift by a variable |
| 176 VMOV dShift[0],scale |
| 177 VRSHL dX0,dShift |
| 178 VST1 dX0[0],[pDst] |
| 179 |
| 180 B End |
| 181 |
| 182 sizeGreaterThanOne: |
| 183 |
| 184 @// Call the preTwiddle Radix2 stage before doing the compledIFFT |
| 185 |
| 186 @// The following conditional BL combination would work since |
| 187 @// evenOddButterflyLoop in the first call would set Z flag to zero |
| 188 |
| 189 CMP scale,#0 |
| 190 BLEQ armSP_FFTInv_CCSToR_S32_preTwiddleRadix2_unsafe |
| 191 BLGT armSP_FFTInv_CCSToR_S32_Sfs_preTwiddleRadix2_unsafe |
| 192 |
| 193 |
| 194 |
| 195 complexIFFT: |
| 196 |
| 197 ASR N,N,#1 @// N/2 point complex IFFT |
| 198 ADD pSrc,pOut,N,LSL #3 @// set pSrc as pOut1 |
| 199 |
| 200 CLZ order,N @// N = 2^order |
| 201 RSB order,order,#31 |
| 202 MOV subFFTSize,#1 |
| 203 @//MOV subFFTNum,N |
| 204 |
| 205 ADD scale,scale,order @// FFTInverse has a final s
caling factor by N |
| 206 |
| 207 CMP order,#3 |
| 208 BGT orderGreaterthan3 @// order > 3 |
| 209 |
| 210 CMP order,#1 |
| 211 BGE orderGreaterthan0 @// order > 0 |
| 212 M_STR scale, diffOnStack,LT @// order = 0 |
| 213 VLD1 dX0,[pSrc] |
| 214 VST1 dX0,[pDst] |
| 215 MOV pSrc,pDst |
| 216 BLT FFTEnd |
| 217 |
| 218 orderGreaterthan0: |
| 219 @// set the buffers appropriately for various orders |
| 220 CMP order,#2 |
| 221 MOVNE argDst,pDst |
| 222 MOVEQ argDst,pOut |
| 223 MOVEQ pOut,pDst @// Pass the first stage des
tination in RN5 |
| 224 MOV argTwiddle,pTwiddle |
| 225 @// Store the scale factor and scale at the end |
| 226 SUB diff,scale,order |
| 227 M_STR diff, diffOnStack |
| 228 BGE orderGreaterthan1 |
| 229 BLLT armSP_FFTInv_CToC_SC32_Sfs_Radix2_fs_OutOfPlace_unsafe @// orde
r = 1 |
| 230 B FFTEnd |
| 231 |
| 232 orderGreaterthan1: |
| 233 MOV tmpOrder,order @// tmpOrder = RN 4 |
| 234 BL armSP_FFTInv_CToC_SC32_Sfs_Radix2_fs_OutOfPlace_unsafe |
| 235 CMP tmpOrder,#2 |
| 236 BLGT armSP_FFTInv_CToC_SC32_Sfs_Radix2_OutOfPlace_unsafe |
| 237 BL armSP_FFTInv_CToC_SC32_Sfs_Radix2_ls_OutOfPlace_unsafe |
| 238 B FFTEnd |
| 239 |
| 240 |
| 241 orderGreaterthan3: |
| 242 @// check scale = 0 or scale = order |
| 243 SUBS diff, scale, order @// scale > order |
| 244 MOVGT scale,order |
| 245 BGE specialScaleCase @// scale = 0 or scale = orde
r |
| 246 CMP scale,#0 |
| 247 BEQ specialScaleCase |
| 248 B generalScaleCase |
| 249 |
| 250 specialScaleCase: @// scale = 0 or sc
ale = order and order >= 2 |
| 251 |
| 252 TST order, #2 @// Set input args to fft st
ages |
| 253 MOVNE argDst,pDst |
| 254 MOVEQ argDst,pOut |
| 255 MOVEQ pOut,pDst @// Pass the first stage des
tination in RN5 |
| 256 MOV argTwiddle,pTwiddle |
| 257 |
| 258 CMP diff,#0 |
| 259 M_STR diff, diffOnStack |
| 260 BGE scaleEqualsOrder |
| 261 |
| 262 @//check for even or odd order |
| 263 @// NOTE: The following combination of BL's would work fine eventhough t
he first |
| 264 @// BL would corrupt the flags. This is because the end of the "grpZeroS
etLoop" loop inside |
| 265 @// armSP_FFTInv_CToC_SC32_Radix4_fs_OutOfPlace_unsafe sets the Z flag t
o EQ |
| 266 |
| 267 TST order,#0x00000001 |
| 268 BLEQ armSP_FFTInv_CToC_SC32_Radix4_fs_OutOfPlace_unsafe |
| 269 BLNE armSP_FFTInv_CToC_SC32_Radix8_fs_OutOfPlace_unsafe |
| 270 |
| 271 CMP subFFTNum,#4 |
| 272 BLT FFTEnd |
| 273 |
| 274 |
| 275 unscaledRadix4Loop: |
| 276 BEQ lastStageUnscaledRadix4 |
| 277 BL armSP_FFTInv_CToC_SC32_Radix4_OutOfPlace_unsafe |
| 278 CMP subFFTNum,#4 |
| 279 B unscaledRadix4Loop |
| 280 |
| 281 lastStageUnscaledRadix4: |
| 282 BL armSP_FFTInv_CToC_SC32_Radix4_ls_OutOfPlace_unsafe |
| 283 B FFTEnd |
| 284 |
| 285 |
| 286 scaleEqualsOrder: |
| 287 @//check for even or odd order |
| 288 @// NOTE: The following combination of BL's would work fine eventhough t
he first |
| 289 @// BL would corrupt the flags. This is because the end of the "grpZeroS
etLoop" loop inside |
| 290 @// armSP_FFTInv_CToC_SC32_Radix4_fs_OutOfPlace_unsafe sets the Z flag t
o EQ |
| 291 |
| 292 TST order,#0x00000001 |
| 293 BLEQ armSP_FFTInv_CToC_SC32_Sfs_Radix4_fs_OutOfPlace_unsafe |
| 294 BLNE armSP_FFTInv_CToC_SC32_Sfs_Radix8_fs_OutOfPlace_unsafe |
| 295 |
| 296 CMP subFFTNum,#4 |
| 297 BLT FFTEnd |
| 298 |
| 299 |
| 300 scaledRadix4Loop: |
| 301 BEQ lastStageScaledRadix4 |
| 302 BL armSP_FFTInv_CToC_SC32_Sfs_Radix4_OutOfPlace_unsafe |
| 303 CMP subFFTNum,#4 |
| 304 B scaledRadix4Loop |
| 305 |
| 306 lastStageScaledRadix4: |
| 307 BL armSP_FFTInv_CToC_SC32_Sfs_Radix4_ls_OutOfPlace_unsafe |
| 308 B FFTEnd |
| 309 |
| 310 generalScaleCase: @// 0 < scale <
order and order >= 2 |
| 311 @// Determine the correct destination buffer |
| 312 SUB diff,order,scale |
| 313 TST diff,#0x01 |
| 314 ADDEQ count,scale,diff,LSR #1 @// count = scale + (order - sca
le)/2 |
| 315 MOVNE count,order |
| 316 TST count,#0x01 @// Is count even or odd ? |
| 317 |
| 318 MOVNE argDst,pDst @// Set input args to fft stages |
| 319 MOVEQ argDst,pOut |
| 320 MOVEQ pOut,pDst @// Pass the first stage destina
tion in RN5 |
| 321 MOV argTwiddle,pTwiddle |
| 322 |
| 323 M_STR diff, diffOnStack |
| 324 |
| 325 MOV argScale,scale @// Put scale in RN4 so as to sa
ve and restore |
| 326 BL armSP_FFTInv_CToC_SC32_Sfs_Radix2_fs_OutOfPlace_unsafe @// s
caled first stage |
| 327 SUBS argScale,argScale,#1 |
| 328 |
| 329 scaledRadix2Loop: |
| 330 BLGT armSP_FFTInv_CToC_SC32_Sfs_Radix2_OutOfPlace_unsafe |
| 331 SUBS argScale,argScale,#1 @// save and restore scale (RN4)
in the scaled stages |
| 332 BGT scaledRadix2Loop |
| 333 |
| 334 |
| 335 M_LDR diff, diffOnStack |
| 336 @//check for even or odd order |
| 337 TST diff,#0x00000001 |
| 338 BEQ generalUnscaledRadix4Loop |
| 339 B unscaledRadix2Loop |
| 340 |
| 341 generalUnscaledRadix4Loop: |
| 342 CMP subFFTNum,#4 |
| 343 BEQ generalLastStageUnscaledRadix4 |
| 344 BL armSP_FFTInv_CToC_SC32_Radix4_OutOfPlace_unsafe |
| 345 B generalUnscaledRadix4Loop |
| 346 |
| 347 generalLastStageUnscaledRadix4: |
| 348 BL armSP_FFTInv_CToC_SC32_Radix4_ls_OutOfPlace_unsafe |
| 349 B End |
| 350 |
| 351 |
| 352 unscaledRadix2Loop: |
| 353 CMP subFFTNum,#2 |
| 354 BEQ generalLastStageUnscaledRadix2 |
| 355 BL armSP_FFTInv_CToC_SC32_Radix2_OutOfPlace_unsafe |
| 356 B unscaledRadix2Loop |
| 357 |
| 358 generalLastStageUnscaledRadix2: |
| 359 BL armSP_FFTInv_CToC_SC32_Radix2_ls_OutOfPlace_unsafe |
| 360 B End |
| 361 |
| 362 |
| 363 FFTEnd: @// Does only the scaling |
| 364 |
| 365 M_LDR diff, diffOnStack |
| 366 CMP diff,#0 |
| 367 BLE End |
| 368 |
| 369 RSB diff,diff,#0 @// to use VRSHL for right s
hift by a variable |
| 370 VDUP dShift,diff |
| 371 |
| 372 scaleFFTData: @// N = subFFTSize ; da
taptr = pDst ; scale = diff |
| 373 VLD1 {dX0},[pSrc] @// pSrc contains pDst pointer |
| 374 SUBS subFFTSize,subFFTSize,#1 |
| 375 VRSHL dX0,dShift |
| 376 VST1 {dX0},[pSrc]! |
| 377 |
| 378 BGT scaleFFTData |
| 379 |
| 380 |
| 381 End: |
| 382 @// Set return value |
| 383 MOV result, #OMX_Sts_NoErr |
| 384 |
| 385 @// Write function tail |
| 386 M_END |
| 387 |
| 388 |
| 389 |
| 390 .end |
OLD | NEW |