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 is a modification of omxSP_FFTFwd_RToCCS_S32_Sfs_s.s |
| 11 @// to support float instead of SC32. |
| 12 @// |
| 13 |
| 14 @// |
| 15 @// Description: |
| 16 @// Compute FFT for a real signal |
| 17 @// |
| 18 @// |
| 19 |
| 20 |
| 21 @// Include standard headers |
| 22 |
| 23 #include "dl/api/armCOMM_s.h" |
| 24 #include "dl/api/omxtypes_s.h" |
| 25 |
| 26 |
| 27 @// Import symbols required from other files |
| 28 @// (For example tables) |
| 29 |
| 30 .extern armSP_FFTFwd_CToC_FC32_Radix2_fs_OutOfPlace_unsafe |
| 31 .extern armSP_FFTFwd_CToC_FC32_Radix4_fs_OutOfPlace_unsafe |
| 32 .extern armSP_FFTFwd_CToC_FC32_Radix8_fs_OutOfPlace_unsafe |
| 33 .extern armSP_FFTFwd_CToC_FC32_Radix4_OutOfPlace_unsafe |
| 34 .extern armSP_FFTFwd_CToC_FC32_Radix4_fs_OutOfPlace_unsafe |
| 35 .extern armSP_FFTFwd_CToC_FC32_Radix8_fs_OutOfPlace_unsafe |
| 36 .extern armSP_FFTFwd_CToC_FC32_Radix2_OutOfPlace_unsafe |
| 37 |
| 38 @// Set debugging level |
| 39 @//DEBUG_ON SETL {TRUE} |
| 40 |
| 41 |
| 42 |
| 43 @// Guarding implementation by the processor name |
| 44 |
| 45 |
| 46 |
| 47 @// Guarding implementation by the processor name |
| 48 |
| 49 @// Import symbols required from other files |
| 50 @// (For example tables) |
| 51 .extern armSP_FFTFwd_CToC_FC32_Radix4_ls_OutOfPlace_unsafe |
| 52 .extern armSP_FFTFwd_CToC_FC32_Radix2_ls_OutOfPlace_unsafe |
| 53 |
| 54 |
| 55 @//Input Registers |
| 56 |
| 57 #define pSrc r0 |
| 58 #define pDst r1 |
| 59 #define pFFTSpec r2 |
| 60 #define scale r3 |
| 61 |
| 62 |
| 63 @// Output registers |
| 64 #define result r0 |
| 65 |
| 66 @//Local Scratch Registers |
| 67 |
| 68 #define argTwiddle r1 |
| 69 #define argDst r2 |
| 70 #define argScale r4 |
| 71 #define tmpOrder r4 |
| 72 #define pTwiddle r4 |
| 73 #define pOut r5 |
| 74 #define subFFTSize r7 |
| 75 #define subFFTNum r6 |
| 76 #define N r6 |
| 77 #define order r14 |
| 78 #define diff r9 |
| 79 @// Total num of radix stages required to comple the FFT |
| 80 #define count r8 |
| 81 #define x0r r4 |
| 82 #define x0i r5 |
| 83 #define diffMinusOne r2 |
| 84 #define subFFTSizeTmp r6 |
| 85 #define step r3 |
| 86 #define step1 r4 |
| 87 #define twStep r8 |
| 88 #define zero r9 |
| 89 #define pTwiddleTmp r5 |
| 90 #define t0 r10 |
| 91 |
| 92 @// Neon registers |
| 93 |
| 94 #define dX0 d0.f32 |
| 95 #define dzero d1.f32 |
| 96 #define dZero d2.f32 |
| 97 #define dShift d3.f32 |
| 98 #define dX0r d2.f32 |
| 99 #define dX0i d3.f32 |
| 100 #define dX1r d4.f32 |
| 101 #define dX1i d5.f32 |
| 102 #define dT0 d6.f32 |
| 103 #define dT1 d7.f32 |
| 104 #define dT2 d8.f32 |
| 105 #define dT3 d9.f32 |
| 106 #define qT0 d10.f32 |
| 107 #define qT1 d12.f32 |
| 108 #define dW0r d14.f32 |
| 109 #define dW0i d15.f32 |
| 110 #define dW1r d16.f32 |
| 111 #define dW1i d17.f32 |
| 112 #define dY0r d14.f32 |
| 113 #define dY0i d15.f32 |
| 114 #define dY1r d16.f32 |
| 115 #define dY1i d17.f32 |
| 116 #define dY0rS64 d14.s64 |
| 117 #define dY0iS64 d15.s64 |
| 118 #define qT2 d18.f32 |
| 119 #define qT3 d20.f32 |
| 120 @// lastThreeelements |
| 121 #define dX1 d3.f32 |
| 122 #define dW0 d4.f32 |
| 123 #define dW1 d5.f32 |
| 124 #define dY0 d10.f32 |
| 125 #define dY1 d11.f32 |
| 126 #define dY2 d12.f32 |
| 127 #define dY3 d13.f32 |
| 128 |
| 129 #define half d0.f32 |
| 130 |
| 131 HALF: .float 0.5 |
| 132 |
| 133 @// Allocate stack memory required by the function |
| 134 |
| 135 @// Write function header |
| 136 M_START omxSP_FFTFwd_RToCCS_F32_Sfs,r11,d15 |
| 137 |
| 138 @ Structure offsets for the FFTSpec |
| 139 .set ARMsFFTSpec_N, 0 |
| 140 .set ARMsFFTSpec_pBitRev, 4 |
| 141 .set ARMsFFTSpec_pTwiddle, 8 |
| 142 .set ARMsFFTSpec_pBuf, 12 |
| 143 |
| 144 @// Define stack arguments |
| 145 |
| 146 @// Read the size from structure and take log |
| 147 LDR N, [pFFTSpec, #ARMsFFTSpec_N] |
| 148 |
| 149 @// Read other structure parameters |
| 150 LDR pTwiddle, [pFFTSpec, #ARMsFFTSpec_pTwiddle] |
| 151 LDR pOut, [pFFTSpec, #ARMsFFTSpec_pBuf] |
| 152 |
| 153 @// N=1 Treat seperately |
| 154 CMP N,#1 |
| 155 BGT sizeGreaterThanOne |
| 156 VLD1 dX0[0],[pSrc] |
| 157 MOV zero,#0 |
| 158 VMOV dzero[0],zero |
| 159 VMOV dZero[0],zero |
| 160 VST3 {dX0[0],dzero[0],dZero[0]},[pDst] |
| 161 |
| 162 B End |
| 163 |
| 164 |
| 165 |
| 166 sizeGreaterThanOne: |
| 167 @// Do a N/2 point complex FFT including the scaling |
| 168 |
| 169 MOV N,N,ASR #1 @// N/2 point complex FFT |
| 170 |
| 171 CLZ order,N @// N = 2^order |
| 172 RSB order,order,#31 |
| 173 MOV subFFTSize,#1 |
| 174 @//MOV subFFTNum,N |
| 175 |
| 176 CMP order,#3 |
| 177 BGT orderGreaterthan3 @// order > 3 |
| 178 |
| 179 CMP order,#1 |
| 180 BGE orderGreaterthan0 @// order > 0 |
| 181 VLD1 dX0,[pSrc] |
| 182 VST1 dX0,[pOut] |
| 183 MOV pSrc,pOut |
| 184 MOV argDst,pDst |
| 185 BLT FFTEnd |
| 186 |
| 187 orderGreaterthan0: |
| 188 @// set the buffers appropriately for various orders |
| 189 CMP order,#2 |
| 190 MOVEQ argDst,pDst |
| 191 MOVNE argDst,pOut |
| 192 @// Pass the first stage destination in RN5 |
| 193 MOVNE pOut,pDst |
| 194 MOV argTwiddle,pTwiddle |
| 195 |
| 196 CMP order,#1 |
| 197 BGT orderGreaterthan1 |
| 198 @// order = 1 |
| 199 BL armSP_FFTFwd_CToC_FC32_Radix2_fs_OutOfPlace_unsafe |
| 200 B FFTEnd |
| 201 |
| 202 orderGreaterthan1: |
| 203 CMP order,#2 |
| 204 BGT orderGreaterthan2 |
| 205 @// order =2 |
| 206 BL armSP_FFTFwd_CToC_FC32_Radix2_fs_OutOfPlace_unsafe |
| 207 BL armSP_FFTFwd_CToC_FC32_Radix2_ls_OutOfPlace_unsafe |
| 208 B FFTEnd |
| 209 |
| 210 orderGreaterthan2:@// order =3 |
| 211 BL armSP_FFTFwd_CToC_FC32_Radix2_fs_OutOfPlace_unsafe |
| 212 BL armSP_FFTFwd_CToC_FC32_Radix2_OutOfPlace_unsafe |
| 213 BL armSP_FFTFwd_CToC_FC32_Radix2_ls_OutOfPlace_unsafe |
| 214 |
| 215 B FFTEnd |
| 216 |
| 217 |
| 218 |
| 219 orderGreaterthan3: |
| 220 specialScaleCase: |
| 221 |
| 222 @// Set input args to fft stages |
| 223 TST order, #2 |
| 224 MOVEQ argDst,pDst |
| 225 MOVNE argDst,pOut |
| 226 @// Pass the first stage destination in RN5 |
| 227 MOVNE pOut,pDst |
| 228 MOV argTwiddle,pTwiddle |
| 229 |
| 230 @//check for even or odd order |
| 231 @// NOTE: The following combination of BL's would work fine even though |
| 232 @// the first BL would corrupt the flags. This is because the end of |
| 233 @// the "grpZeroSetLoop" loop inside |
| 234 @// armSP_FFTFwd_CToC_FC32_Radix4_fs_OutOfPlace_unsafe sets the Z flag |
| 235 @// to EQ |
| 236 |
| 237 TST order,#0x00000001 |
| 238 BLEQ armSP_FFTFwd_CToC_FC32_Radix4_fs_OutOfPlace_unsafe |
| 239 BLNE armSP_FFTFwd_CToC_FC32_Radix8_fs_OutOfPlace_unsafe |
| 240 |
| 241 CMP subFFTNum,#4 |
| 242 BLT FFTEnd |
| 243 |
| 244 |
| 245 unscaledRadix4Loop: |
| 246 BEQ lastStageUnscaledRadix4 |
| 247 BL armSP_FFTFwd_CToC_FC32_Radix4_OutOfPlace_unsafe |
| 248 CMP subFFTNum,#4 |
| 249 B unscaledRadix4Loop |
| 250 |
| 251 lastStageUnscaledRadix4: |
| 252 BL armSP_FFTFwd_CToC_FC32_Radix4_ls_OutOfPlace_unsafe |
| 253 B FFTEnd |
| 254 |
| 255 |
| 256 FFTEnd: |
| 257 finalComplexToRealFixup: |
| 258 |
| 259 |
| 260 @// F(0) = 1/2[Z(0) + Z'(0)] - j [Z(0) - Z'(0)] |
| 261 @// 1/2[(a+jb) + (a-jb)] - j [(a+jb) - (a-jb)] |
| 262 @// 1/2[2a+j0] - j [0+j2b] |
| 263 @// (a+b, 0) |
| 264 |
| 265 @// F(N/2) = 1/2[Z(0) + Z'(0)] + j [Z(0) - Z'(0)] |
| 266 @// 1/2[(a+jb) + (a-jb)] + j [(a+jb) - (a-jb)] |
| 267 @// 1/2[2a+j0] + j [0+j2b] |
| 268 @// (a-b, 0) |
| 269 |
| 270 @// F(0) and F(N/2) |
| 271 VLD2 {dX0r[0],dX0i[0]},[pSrc]! |
| 272 MOV zero,#0 |
| 273 VMOV dX0r[1],zero |
| 274 MOV step,subFFTSize,LSL #3 @// step = N/2 * 8 bytes |
| 275 VMOV dX0i[1],zero |
| 276 @// twStep = 3N/8 * 8 bytes pointing to W^1 |
| 277 SUB twStep,step,subFFTSize,LSL #1 |
| 278 |
| 279 VADD dY0r,dX0r,dX0i @// F(0) = ((Z0.r+Z0.i) , 0) |
| 280 MOV step1,subFFTSize,LSL #2 @// step1 = N/2 * 4 bytes |
| 281 VSUB dY0i,dX0r,dX0i @// F(N/2) = ((Z0.r-Z0.i) , 0) |
| 282 SUBS subFFTSize,subFFTSize,#2 |
| 283 |
| 284 VST1 dY0r,[argDst],step |
| 285 ADD pTwiddleTmp,argTwiddle,#8 @// W^2 |
| 286 VST1 dY0i,[argDst]! |
| 287 ADD argTwiddle,argTwiddle,twStep @// W^1 |
| 288 |
| 289 VDUP dzero,zero |
| 290 SUB argDst,argDst,step |
| 291 |
| 292 BLT End |
| 293 BEQ lastElement |
| 294 SUB step,step,#24 |
| 295 SUB step1,step1,#8 @// (N/4-1)*8 bytes |
| 296 |
| 297 @// F(k) = 1/2[Z(k) + Z'(N/2-k)] -j*W^(k) [Z(k) - Z'(N/2-k)] |
| 298 @// Note: W^k is stored as negative values in the table |
| 299 @// Process 4 elements at a time. E.g: F(1),F(2) and F(N/2-2),F(N/2-1) |
| 300 @// since both of them require Z(1),Z(2) and Z(N/2-2),Z(N/2-1) |
| 301 |
| 302 |
| 303 LDR t0, =HALF |
| 304 VLD1 half[0], [t0] |
| 305 |
| 306 evenOddButterflyLoop: |
| 307 |
| 308 |
| 309 VLD1 dW0r,[argTwiddle],step1 |
| 310 VLD1 dW1r,[argTwiddle]! |
| 311 |
| 312 VLD2 {dX0r,dX0i},[pSrc],step |
| 313 SUB argTwiddle,argTwiddle,step1 |
| 314 VLD2 {dX1r,dX1i},[pSrc]! |
| 315 |
| 316 |
| 317 |
| 318 SUB step1,step1,#8 @// (N/4-2)*8 bytes |
| 319 VLD1 dW0i,[pTwiddleTmp],step1 |
| 320 VLD1 dW1i,[pTwiddleTmp]! |
| 321 SUB pSrc,pSrc,step |
| 322 |
| 323 SUB pTwiddleTmp,pTwiddleTmp,step1 |
| 324 VREV64 dX1r,dX1r |
| 325 VREV64 dX1i,dX1i |
| 326 SUBS subFFTSize,subFFTSize,#4 |
| 327 |
| 328 |
| 329 |
| 330 VSUB dT2,dX0r,dX1r @// a-c |
| 331 SUB step1,step1,#8 |
| 332 VADD dT0,dX0r,dX1r @// a+c |
| 333 VSUB dT1,dX0i,dX1i @// b-d |
| 334 VADD dT3,dX0i,dX1i @// b+d |
| 335 VMUL dT0,dT0,half[0] |
| 336 VMUL dT1,dT1,half[0] |
| 337 VZIP dW1r,dW1i |
| 338 VZIP dW0r,dW0i |
| 339 |
| 340 |
| 341 VMUL qT0,dW1r,dT2 |
| 342 VMUL qT1,dW1r,dT3 |
| 343 VMUL qT2,dW0r,dT2 |
| 344 VMUL qT3,dW0r,dT3 |
| 345 |
| 346 VMLA qT0,dW1i,dT3 |
| 347 VMLS qT1,dW1i,dT2 |
| 348 |
| 349 VMLS qT2,dW0i,dT3 |
| 350 VMLA qT3,dW0i,dT2 |
| 351 |
| 352 |
| 353 VMUL dX1r,qT0,half[0] |
| 354 VMUL dX1i,qT1,half[0] |
| 355 |
| 356 VSUB dY1r,dT0,dX1i @// F(N/2 -1) |
| 357 VADD dY1i,dT1,dX1r |
| 358 VNEG dY1i,dY1i |
| 359 |
| 360 VREV64 dY1r,dY1r |
| 361 VREV64 dY1i,dY1i |
| 362 |
| 363 |
| 364 VMUL dX0r,qT2,half[0] |
| 365 VMUL dX0i,qT3,half[0] |
| 366 |
| 367 VSUB dY0r,dT0,dX0i @// F(1) |
| 368 VADD dY0i,dT1,dX0r |
| 369 |
| 370 |
| 371 VST2 {dY0r,dY0i},[argDst],step |
| 372 VST2 {dY1r,dY1i},[argDst]! |
| 373 SUB argDst,argDst,step |
| 374 SUB step,step,#32 @// (N/2-4)*8 bytes |
| 375 |
| 376 |
| 377 BGT evenOddButterflyLoop |
| 378 |
| 379 @// set both the ptrs to the last element |
| 380 SUB pSrc,pSrc,#8 |
| 381 SUB argDst,argDst,#8 |
| 382 |
| 383 |
| 384 |
| 385 @// Last element can be expanded as follows |
| 386 @// 1/2[Z(k) + Z'(k)] + j w^k [Z(k) - Z'(k)] |
| 387 @// 1/2[(a+jb) + (a-jb)] + j w^k [(a+jb) - (a-jb)] |
| 388 @// 1/2[2a+j0] + j (c+jd) [0+j2b] |
| 389 @// (a-bc, -bd) |
| 390 @// Since (c,d) = (0,1) for the last element, result is just (a,-b) |
| 391 |
| 392 lastElement: |
| 393 VLD1 dX0r,[pSrc] |
| 394 |
| 395 VST1 dX0r[0],[argDst]! |
| 396 VNEG dX0r,dX0r |
| 397 VST1 dX0r[1],[argDst]! |
| 398 |
| 399 End: |
| 400 @// Set return value |
| 401 MOV result, #OMX_Sts_NoErr |
| 402 |
| 403 @// Write function tail |
| 404 M_END |
| 405 |
| 406 .end |
OLD | NEW |