| 1 | /******************************************************************************
|
|---|
| 2 | * @file arm_vec_fft.h
|
|---|
| 3 | * @brief Private header file for CMSIS DSP Library
|
|---|
| 4 | * @version V1.7.0
|
|---|
| 5 | * @date 07. January 2020
|
|---|
| 6 | ******************************************************************************/
|
|---|
| 7 | /*
|
|---|
| 8 | * Copyright (c) 2010-2020 Arm Limited or its affiliates. All rights reserved.
|
|---|
| 9 | *
|
|---|
| 10 | * SPDX-License-Identifier: Apache-2.0
|
|---|
| 11 | *
|
|---|
| 12 | * Licensed under the Apache License, Version 2.0 (the License); you may
|
|---|
| 13 | * not use this file except in compliance with the License.
|
|---|
| 14 | * You may obtain a copy of the License at
|
|---|
| 15 | *
|
|---|
| 16 | * www.apache.org/licenses/LICENSE-2.0
|
|---|
| 17 | *
|
|---|
| 18 | * Unless required by applicable law or agreed to in writing, software
|
|---|
| 19 | * distributed under the License is distributed on an AS IS BASIS, WITHOUT
|
|---|
| 20 | * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|---|
| 21 | * See the License for the specific language governing permissions and
|
|---|
| 22 | * limitations under the License.
|
|---|
| 23 | */
|
|---|
| 24 |
|
|---|
| 25 | #ifndef _ARM_VEC_FFT_H_
|
|---|
| 26 | #define _ARM_VEC_FFT_H_
|
|---|
| 27 |
|
|---|
| 28 | #include "arm_math.h"
|
|---|
| 29 | #include "arm_helium_utils.h"
|
|---|
| 30 |
|
|---|
| 31 | #ifdef __cplusplus
|
|---|
| 32 | extern "C"
|
|---|
| 33 | {
|
|---|
| 34 | #endif
|
|---|
| 35 |
|
|---|
| 36 | #if (defined(ARM_MATH_MVEF) || defined(ARM_MATH_MVEI) || defined(ARM_MATH_HELIUM)) && !defined(ARM_MATH_AUTOVECTORIZE)
|
|---|
| 37 |
|
|---|
| 38 | #define MVE_CMPLX_ADD_A_ixB(A, B) vcaddq_rot90(A,B)
|
|---|
| 39 | #define MVE_CMPLX_SUB_A_ixB(A,B) vcaddq_rot270(A,B)
|
|---|
| 40 | #define MVE_CMPLX_MULT_FLT_AxB(A,B) vcmlaq_rot90(vcmulq(A, B), A, B)
|
|---|
| 41 | #define MVE_CMPLX_MULT_FLT_Conj_AxB(A,B) vcmlaq_rot270(vcmulq(A, B), A, B)
|
|---|
| 42 |
|
|---|
| 43 | #define MVE_CMPLX_MULT_FX_AxB(A,B,TyA) vqdmladhxq(vqdmlsdhq((TyA)vuninitializedq_s32(), A, B), A, B)
|
|---|
| 44 | #define MVE_CMPLX_MULT_FX_AxConjB(A,B,TyA) vqdmladhq(vqdmlsdhxq((TyA)vuninitializedq_s32(), A, B), A, B)
|
|---|
| 45 |
|
|---|
| 46 | #define MVE_CMPLX_ADD_FX_A_ixB(A, B) vhcaddq_rot90(A,B)
|
|---|
| 47 | #define MVE_CMPLX_SUB_FX_A_ixB(A,B) vhcaddq_rot270(A,B)
|
|---|
| 48 |
|
|---|
| 49 |
|
|---|
| 50 | /**
|
|---|
| 51 | @brief In-place 32 bit reversal function for helium
|
|---|
| 52 | @param[in,out] pSrc points to in-place buffer of unknown 32-bit data type
|
|---|
| 53 | @param[in] bitRevLen bit reversal table length
|
|---|
| 54 | @param[in] pBitRevTab points to bit reversal table
|
|---|
| 55 | @return none
|
|---|
| 56 | */
|
|---|
| 57 |
|
|---|
| 58 | __STATIC_INLINE void arm_bitreversal_32_inpl_mve(
|
|---|
| 59 | uint32_t *pSrc,
|
|---|
| 60 | const uint16_t bitRevLen,
|
|---|
| 61 | const uint16_t *pBitRevTab)
|
|---|
| 62 |
|
|---|
| 63 | {
|
|---|
| 64 | uint64_t *src = (uint64_t *) pSrc;
|
|---|
| 65 | int32_t blkCnt; /* loop counters */
|
|---|
| 66 | uint32x4_t bitRevTabOff;
|
|---|
| 67 | uint32x4_t one = vdupq_n_u32(1);
|
|---|
| 68 | uint64x2_t inLow, inHigh;
|
|---|
| 69 | uint64x2_t bitRevOff1Low, bitRevOff0Low;
|
|---|
| 70 | uint64x2_t bitRevOff1High, bitRevOff0High;
|
|---|
| 71 |
|
|---|
| 72 | /* load scheduling to increase gather load idx update / gather load distance */
|
|---|
| 73 | bitRevTabOff = vldrhq_u32(pBitRevTab);
|
|---|
| 74 | pBitRevTab += 4;
|
|---|
| 75 |
|
|---|
| 76 | bitRevOff0Low = vmullbq_int_u32(bitRevTabOff, one);
|
|---|
| 77 | bitRevOff0High = vmulltq_int_u32(bitRevTabOff, one);
|
|---|
| 78 |
|
|---|
| 79 |
|
|---|
| 80 | blkCnt = bitRevLen / 8;
|
|---|
| 81 | while (blkCnt > 0) {
|
|---|
| 82 | bitRevTabOff = vldrhq_u32(pBitRevTab);
|
|---|
| 83 | pBitRevTab += 4;
|
|---|
| 84 |
|
|---|
| 85 | /* 64-bit index expansion */
|
|---|
| 86 | bitRevOff1Low = vmullbq_int_u32(bitRevTabOff, one);
|
|---|
| 87 | bitRevOff1High = vmulltq_int_u32(bitRevTabOff, one);
|
|---|
| 88 |
|
|---|
| 89 | inLow = vldrdq_gather_offset_u64(src, bitRevOff0Low);
|
|---|
| 90 | inHigh = vldrdq_gather_offset_u64(src, bitRevOff0High);
|
|---|
| 91 |
|
|---|
| 92 | vstrdq_scatter_offset_u64(src, bitRevOff0Low, inHigh);
|
|---|
| 93 | vstrdq_scatter_offset_u64(src, bitRevOff0High, inLow);
|
|---|
| 94 |
|
|---|
| 95 |
|
|---|
| 96 | /* unrolled */
|
|---|
| 97 | bitRevTabOff = vldrhq_u32(pBitRevTab);
|
|---|
| 98 | pBitRevTab += 4;
|
|---|
| 99 |
|
|---|
| 100 | bitRevOff0Low = vmullbq_int_u32(bitRevTabOff, one);
|
|---|
| 101 | bitRevOff0High = vmulltq_int_u32(bitRevTabOff, one);
|
|---|
| 102 |
|
|---|
| 103 | inLow = vldrdq_gather_offset_u64(src, bitRevOff1Low);
|
|---|
| 104 | inHigh = vldrdq_gather_offset_u64(src, bitRevOff1High);
|
|---|
| 105 |
|
|---|
| 106 | vstrdq_scatter_offset_u64(src, bitRevOff1Low, inHigh);
|
|---|
| 107 | vstrdq_scatter_offset_u64(src, bitRevOff1High, inLow);
|
|---|
| 108 |
|
|---|
| 109 | /*
|
|---|
| 110 | * Decrement the blockSize loop counter
|
|---|
| 111 | */
|
|---|
| 112 | blkCnt--;
|
|---|
| 113 | }
|
|---|
| 114 |
|
|---|
| 115 | if (bitRevLen & 7) {
|
|---|
| 116 | /* FFT size = 16 */
|
|---|
| 117 | inLow = vldrdq_gather_offset_u64(src, bitRevOff0Low);
|
|---|
| 118 | inHigh = vldrdq_gather_offset_u64(src, bitRevOff0High);
|
|---|
| 119 |
|
|---|
| 120 | vstrdq_scatter_offset_u64(src, bitRevOff0Low, inHigh);
|
|---|
| 121 | vstrdq_scatter_offset_u64(src, bitRevOff0High, inLow);
|
|---|
| 122 | }
|
|---|
| 123 | }
|
|---|
| 124 |
|
|---|
| 125 |
|
|---|
| 126 |
|
|---|
| 127 | /**
|
|---|
| 128 | @brief In-place 16 bit reversal function for helium
|
|---|
| 129 | @param[in,out] pSrc points to in-place buffer of unknown 16-bit data type
|
|---|
| 130 | @param[in] bitRevLen bit reversal table length
|
|---|
| 131 | @param[in] pBitRevTab points to bit reversal table
|
|---|
| 132 | @return none
|
|---|
| 133 | */
|
|---|
| 134 |
|
|---|
| 135 | __STATIC_INLINE void arm_bitreversal_16_inpl_mve(
|
|---|
| 136 | uint16_t *pSrc,
|
|---|
| 137 | const uint16_t bitRevLen,
|
|---|
| 138 | const uint16_t *pBitRevTab)
|
|---|
| 139 |
|
|---|
| 140 | {
|
|---|
| 141 | uint32_t *src = (uint32_t *) pSrc;
|
|---|
| 142 | int32_t blkCnt; /* loop counters */
|
|---|
| 143 | uint32x4_t bitRevTabOff;
|
|---|
| 144 | uint16x8_t one = vdupq_n_u16(1);
|
|---|
| 145 | uint32x4_t bitRevOff1Low, bitRevOff0Low;
|
|---|
| 146 | uint32x4_t bitRevOff1High, bitRevOff0High;
|
|---|
| 147 | uint32x4_t inLow, inHigh;
|
|---|
| 148 |
|
|---|
| 149 | /* load scheduling to increase gather load idx update / gather load distance */
|
|---|
| 150 | bitRevTabOff = vldrhq_u16(pBitRevTab);
|
|---|
| 151 | pBitRevTab += 8;
|
|---|
| 152 |
|
|---|
| 153 | bitRevOff0Low = vmullbq_int_u16((uint16x8_t)bitRevTabOff, one);
|
|---|
| 154 | bitRevOff0High = vmulltq_int_u16((uint16x8_t)bitRevTabOff, one);
|
|---|
| 155 | bitRevOff0Low = vshrq_n_u16((uint16x8_t)bitRevOff0Low, 3);
|
|---|
| 156 | bitRevOff0High = vshrq_n_u16((uint16x8_t)bitRevOff0High, 3);
|
|---|
| 157 |
|
|---|
| 158 | blkCnt = (bitRevLen / 16);
|
|---|
| 159 | while (blkCnt > 0) {
|
|---|
| 160 | bitRevTabOff = vldrhq_u16(pBitRevTab);
|
|---|
| 161 | pBitRevTab += 8;
|
|---|
| 162 |
|
|---|
| 163 | bitRevOff1Low = vmullbq_int_u16((uint16x8_t)bitRevTabOff, one);
|
|---|
| 164 | bitRevOff1High = vmulltq_int_u16((uint16x8_t)bitRevTabOff, one);
|
|---|
| 165 | bitRevOff1Low = vshrq_n_u16((uint16x8_t)bitRevOff1Low, 3);
|
|---|
| 166 | bitRevOff1High = vshrq_n_u16((uint16x8_t)bitRevOff1High, 3);
|
|---|
| 167 |
|
|---|
| 168 | inLow = vldrwq_gather_shifted_offset_u32(src, bitRevOff0Low);
|
|---|
| 169 | inHigh = vldrwq_gather_shifted_offset_u32(src, bitRevOff0High);
|
|---|
| 170 |
|
|---|
| 171 | vstrwq_scatter_shifted_offset_u32(src, bitRevOff0Low, inHigh);
|
|---|
| 172 | vstrwq_scatter_shifted_offset_u32(src, bitRevOff0High, inLow);
|
|---|
| 173 |
|
|---|
| 174 | /* loop unrolling */
|
|---|
| 175 | bitRevTabOff = vldrhq_u16(pBitRevTab);
|
|---|
| 176 | pBitRevTab += 8;
|
|---|
| 177 |
|
|---|
| 178 | bitRevOff0Low = vmullbq_int_u16((uint16x8_t)bitRevTabOff, one);
|
|---|
| 179 | bitRevOff0High = vmulltq_int_u16((uint16x8_t)bitRevTabOff, one);
|
|---|
| 180 | bitRevOff0Low = vshrq_n_u16((uint16x8_t)bitRevOff0Low, 3);
|
|---|
| 181 | bitRevOff0High = vshrq_n_u16((uint16x8_t)bitRevOff0High, 3);
|
|---|
| 182 |
|
|---|
| 183 | inLow = vldrwq_gather_shifted_offset_u32(src, bitRevOff1Low);
|
|---|
| 184 | inHigh = vldrwq_gather_shifted_offset_u32(src, bitRevOff1High);
|
|---|
| 185 |
|
|---|
| 186 | vstrwq_scatter_shifted_offset_u32(src, bitRevOff1Low, inHigh);
|
|---|
| 187 | vstrwq_scatter_shifted_offset_u32(src, bitRevOff1High, inLow);
|
|---|
| 188 |
|
|---|
| 189 | blkCnt--;
|
|---|
| 190 | }
|
|---|
| 191 |
|
|---|
| 192 | /* tail handling */
|
|---|
| 193 | blkCnt = bitRevLen & 0xf;
|
|---|
| 194 | if (blkCnt == 8) {
|
|---|
| 195 | inLow = vldrwq_gather_shifted_offset_u32(src, bitRevOff0Low);
|
|---|
| 196 | inHigh = vldrwq_gather_shifted_offset_u32(src, bitRevOff0High);
|
|---|
| 197 |
|
|---|
| 198 | vstrwq_scatter_shifted_offset_u32(src, bitRevOff0Low, inHigh);
|
|---|
| 199 | vstrwq_scatter_shifted_offset_u32(src, bitRevOff0High, inLow);
|
|---|
| 200 | } else if (blkCnt == 12) {
|
|---|
| 201 | /* FFT 16 special case */
|
|---|
| 202 | mve_pred16_t p = vctp16q(4);
|
|---|
| 203 |
|
|---|
| 204 | bitRevTabOff = vldrhq_z_u16(pBitRevTab, p);
|
|---|
| 205 |
|
|---|
| 206 | inLow = vldrwq_gather_shifted_offset_u32(src, bitRevOff0Low);
|
|---|
| 207 | inHigh = vldrwq_gather_shifted_offset_u32(src, bitRevOff0High);
|
|---|
| 208 |
|
|---|
| 209 | vstrwq_scatter_shifted_offset_u32(src, bitRevOff0Low, inHigh);
|
|---|
| 210 | vstrwq_scatter_shifted_offset_u32(src, bitRevOff0High, inLow);
|
|---|
| 211 |
|
|---|
| 212 | bitRevOff0Low = vmullbq_int_u16((uint16x8_t)bitRevTabOff, one);
|
|---|
| 213 | bitRevOff0High = vmulltq_int_u16((uint16x8_t)bitRevTabOff, one);
|
|---|
| 214 | bitRevOff0Low = vshrq_n_u16((uint16x8_t)bitRevOff0Low, 3);
|
|---|
| 215 | bitRevOff0High = vshrq_n_u16((uint16x8_t)bitRevOff0High, 3);
|
|---|
| 216 |
|
|---|
| 217 | inLow = vldrwq_gather_shifted_offset_z_u32(src, bitRevOff0Low, p);
|
|---|
| 218 | inHigh = vldrwq_gather_shifted_offset_z_u32(src, bitRevOff0High, p);
|
|---|
| 219 |
|
|---|
| 220 | vstrwq_scatter_shifted_offset_p_u32(src, bitRevOff0Low, inHigh, p);
|
|---|
| 221 | vstrwq_scatter_shifted_offset_p_u32(src, bitRevOff0High, inLow, p);
|
|---|
| 222 | }
|
|---|
| 223 | }
|
|---|
| 224 |
|
|---|
| 225 | /**
|
|---|
| 226 | @brief Out-of-place 32 bit reversal function for helium
|
|---|
| 227 | @param[out] pDst points to destination buffer of unknown 32-bit data type
|
|---|
| 228 | @param[in] pSrc points to input buffer of unknown 32-bit data type
|
|---|
| 229 | @param[in] fftLen FFT length
|
|---|
| 230 | @return none
|
|---|
| 231 | */
|
|---|
| 232 | __STATIC_INLINE void arm_bitreversal_32_outpl_mve(void *pDst, void *pSrc, uint32_t fftLen)
|
|---|
| 233 | {
|
|---|
| 234 | uint32x4_t idxOffs0, idxOffs1, bitRevOffs0, bitRevOffs1;
|
|---|
| 235 | uint32_t bitRevPos, blkCnt;
|
|---|
| 236 | uint32_t *pDst32 = (uint32_t *) pDst;
|
|---|
| 237 |
|
|---|
| 238 | /* fwd indexes */
|
|---|
| 239 | idxOffs0 = vdupq_n_u32(0);
|
|---|
| 240 | idxOffs1 = vdupq_n_u32(0);
|
|---|
| 241 | idxOffs0[0] = 0; idxOffs0[2] = 4;
|
|---|
| 242 | idxOffs1[0] = 8; idxOffs1[2] = 12;
|
|---|
| 243 |
|
|---|
| 244 | bitRevPos = (31 - __CLZ(fftLen)) + 5;
|
|---|
| 245 | blkCnt = fftLen >> 2;
|
|---|
| 246 |
|
|---|
| 247 | /* issued earlier to increase gather load idx update / gather load distance */
|
|---|
| 248 | /* bit-reverse fwd indexes */
|
|---|
| 249 | bitRevOffs0 = vbrsrq(idxOffs0, bitRevPos);
|
|---|
| 250 | bitRevOffs1 = vbrsrq(idxOffs1, bitRevPos);
|
|---|
| 251 | while (blkCnt > 0) {
|
|---|
| 252 | uint64x2_t vecIn;
|
|---|
| 253 |
|
|---|
| 254 | vecIn = vldrdq_gather_offset_u64(pSrc, (uint64x2_t) bitRevOffs0);
|
|---|
| 255 | idxOffs0 = idxOffs0 + 16;
|
|---|
| 256 | vst1q(pDst32, (uint32x4_t) vecIn);
|
|---|
| 257 | pDst32 += 4;
|
|---|
| 258 | bitRevOffs0 = vbrsrq(idxOffs0, bitRevPos);
|
|---|
| 259 |
|
|---|
| 260 | vecIn = vldrdq_gather_offset_u64(pSrc, (uint64x2_t) bitRevOffs1);
|
|---|
| 261 | idxOffs1 = idxOffs1 + 16;
|
|---|
| 262 | vst1q(pDst32, (uint32x4_t) vecIn);
|
|---|
| 263 | pDst32 += 4;
|
|---|
| 264 | bitRevOffs1 = vbrsrq(idxOffs1, bitRevPos);
|
|---|
| 265 |
|
|---|
| 266 | blkCnt--;
|
|---|
| 267 | }
|
|---|
| 268 | }
|
|---|
| 269 |
|
|---|
| 270 |
|
|---|
| 271 | /**
|
|---|
| 272 | @brief Out-of-place 16 bit reversal function for helium
|
|---|
| 273 | @param[out] pDst points to destination buffer of unknown 16-bit data type
|
|---|
| 274 | @param[in] pSrc points to input buffer of unknown 16-bit data type
|
|---|
| 275 | @param[in] fftLen FFT length
|
|---|
| 276 | @return none
|
|---|
| 277 | */
|
|---|
| 278 |
|
|---|
| 279 | __STATIC_INLINE void arm_bitreversal_16_outpl_mve(void *pDst, void *pSrc, uint32_t fftLen)
|
|---|
| 280 | {
|
|---|
| 281 | uint32x4_t idxOffs0, idxOffs1, bitRevOffs0, bitRevOffs1;
|
|---|
| 282 | uint32_t bitRevPos, blkCnt;
|
|---|
| 283 | uint16_t *pDst16 = (uint16_t *) pDst;
|
|---|
| 284 | uint32_t incrIdx = 0;
|
|---|
| 285 |
|
|---|
| 286 | /* fwd indexes */
|
|---|
| 287 | idxOffs0 = vidupq_wb_u32(&incrIdx, 4); // {0, 4, 8, 12}
|
|---|
| 288 | idxOffs1 = vidupq_wb_u32(&incrIdx, 4); // {16, 20, 24, 28}
|
|---|
| 289 |
|
|---|
| 290 | bitRevPos = (31 - __CLZ(fftLen)) + 4;
|
|---|
| 291 | blkCnt = fftLen >> 3;
|
|---|
| 292 |
|
|---|
| 293 | /* issued earlier to increase gather load idx update / gather load distance */
|
|---|
| 294 | /* bit-reverse fwd indexes */
|
|---|
| 295 | bitRevOffs0 = vbrsrq(idxOffs0, bitRevPos);
|
|---|
| 296 | bitRevOffs1 = vbrsrq(idxOffs1, bitRevPos);
|
|---|
| 297 | while (blkCnt > 0) {
|
|---|
| 298 | uint32x4_t vecIn;
|
|---|
| 299 |
|
|---|
| 300 | vecIn = vldrwq_gather_offset_s32(pSrc, bitRevOffs0);
|
|---|
| 301 | idxOffs0 = idxOffs0 + 32;
|
|---|
| 302 | vst1q(pDst16, (uint16x8_t) vecIn);
|
|---|
| 303 | pDst16 += 8;
|
|---|
| 304 | bitRevOffs0 = vbrsrq(idxOffs0, bitRevPos);
|
|---|
| 305 |
|
|---|
| 306 | vecIn = vldrwq_gather_offset_s32(pSrc, bitRevOffs1);
|
|---|
| 307 | idxOffs1 = idxOffs1 + 32;
|
|---|
| 308 | vst1q(pDst16, (uint16x8_t) vecIn);
|
|---|
| 309 | pDst16 += 8;
|
|---|
| 310 | bitRevOffs1 = vbrsrq(idxOffs1, bitRevPos);
|
|---|
| 311 |
|
|---|
| 312 | blkCnt--;
|
|---|
| 313 | }
|
|---|
| 314 | }
|
|---|
| 315 |
|
|---|
| 316 |
|
|---|
| 317 | #endif /* (defined(ARM_MATH_MVEF) || defined(ARM_MATH_HELIUM)) && !defined(ARM_MATH_AUTOVECTORIZE)*/
|
|---|
| 318 |
|
|---|
| 319 |
|
|---|
| 320 | #ifdef __cplusplus
|
|---|
| 321 | }
|
|---|
| 322 | #endif
|
|---|
| 323 |
|
|---|
| 324 |
|
|---|
| 325 | #endif /* _ARM_VEC_FFT_H_ */
|
|---|