| 1 |
/* |
| 2 |
* Copyright (C) 2020 Linux Studio Plugins Project <https://lsp-plug.in/> |
| 3 |
* (C) 2020 Vladimir Sadovnikov <[email protected]> |
| 4 |
* |
| 5 |
* This file is part of lsp-dsp-lib |
| 6 |
* Created on: 31 мар. 2020 г. |
| 7 |
* |
| 8 |
* lsp-dsp-lib is free software: you can redistribute it and/or modify |
| 9 |
* it under the terms of the GNU Lesser General Public License as published by |
| 10 |
* the Free Software Foundation, either version 3 of the License, or |
| 11 |
* any later version. |
| 12 |
* |
| 13 |
* lsp-dsp-lib is distributed in the hope that it will be useful, |
| 14 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 16 |
* GNU Lesser General Public License for more details. |
| 17 |
* |
| 18 |
* You should have received a copy of the GNU Lesser General Public License |
| 19 |
* along with lsp-dsp-lib. If not, see <https://www.gnu.org/licenses/>. |
| 20 |
*/ |
| 21 |
|
| 22 |
#ifndef PRIVATE_DSP_ARCH_X86_SSE_FFT_BUTTERFLY_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_X86_SSE_FFT_BUTTERFLY_H_ |
| 24 |
|
| 25 |
#ifndef PRIVATE_DSP_ARCH_X86_SSE_IMPL |
| 26 |
#error "This header should not be included directly" |
| 27 |
#endif /* PRIVATE_DSP_ARCH_X86_SSE_IMPL */ |
| 28 |
|
| 29 |
/* This is the SSE implementation of the function: |
| 30 |
static void butterfly_direct_fft(float *dst_re, float *dst_im, size_t pairs, size_t blocks) |
| 31 |
{ |
| 32 |
float w[4], w_re[4], w_im[4], c_re[4], c_im[4]; |
| 33 |
size_t stride = pairs << 1; |
| 34 |
|
| 35 |
for (size_t p=0; p < pairs; p += 4) |
| 36 |
{ |
| 37 |
// Calculate sines and cosines |
| 38 |
float l = M_PI / pairs; |
| 39 |
w[0] = p * l; |
| 40 |
w[1] = (p + 1) * l; |
| 41 |
w[2] = (p + 2) * l; |
| 42 |
w[3] = (p + 3) * l; |
| 43 |
|
| 44 |
w_re[0] = cosf(w[0]); |
| 45 |
w_re[1] = cosf(w[1]); |
| 46 |
w_re[2] = cosf(w[2]); |
| 47 |
w_re[3] = cosf(w[3]); |
| 48 |
|
| 49 |
w_im[0] = sinf(w[0]); |
| 50 |
w_im[1] = sinf(w[1]); |
| 51 |
w_im[2] = sinf(w[2]); |
| 52 |
w_im[3] = sinf(w[3]); |
| 53 |
|
| 54 |
// Init pointers |
| 55 |
float *a_re = &dst_re[p]; |
| 56 |
float *a_im = &dst_im[p]; |
| 57 |
float *b_re = &a_re[pairs]; |
| 58 |
float *b_im = &a_im[pairs]; |
| 59 |
|
| 60 |
for (size_t b=0; b < blocks; ++b) |
| 61 |
{ |
| 62 |
// Perform 4x butterflies |
| 63 |
// Calculate complex c[i] = w[i] * b[i] |
| 64 |
c_re[0] = w_re[0] * b_re[0] + w_im[0] * b_im[0]; |
| 65 |
c_re[1] = w_re[1] * b_re[1] + w_im[1] * b_im[1]; |
| 66 |
c_re[2] = w_re[2] * b_re[2] + w_im[2] * b_im[2]; |
| 67 |
c_re[3] = w_re[3] * b_re[3] + w_im[3] * b_im[3]; |
| 68 |
|
| 69 |
c_im[0] = w_re[0] * b_im[0] - w_im[0] * b_re[0]; |
| 70 |
c_im[1] = w_re[1] * b_im[1] - w_im[1] * b_re[1]; |
| 71 |
c_im[2] = w_re[2] * b_im[2] - w_im[2] * b_re[2]; |
| 72 |
c_im[3] = w_re[3] * b_im[3] - w_im[3] * b_re[3]; |
| 73 |
|
| 74 |
// Calculate the output values: |
| 75 |
// a[i]' = a[i] + c[i] |
| 76 |
// b[i]' = a[i] - c[i] |
| 77 |
b_re[0] = a_re[0] - c_re[0]; |
| 78 |
b_re[1] = a_re[1] - c_re[1]; |
| 79 |
b_re[2] = a_re[2] - c_re[2]; |
| 80 |
b_re[3] = a_re[3] - c_re[3]; |
| 81 |
|
| 82 |
b_im[0] = a_im[0] - c_im[0]; |
| 83 |
b_im[1] = a_im[1] - c_im[1]; |
| 84 |
b_im[2] = a_im[2] - c_im[2]; |
| 85 |
b_im[3] = a_im[3] - c_im[3]; |
| 86 |
|
| 87 |
a_re[0] = a_re[0] + c_re[0]; |
| 88 |
a_re[1] = a_re[1] + c_re[1]; |
| 89 |
a_re[2] = a_re[2] + c_re[2]; |
| 90 |
a_re[3] = a_re[3] + c_re[3]; |
| 91 |
|
| 92 |
a_im[0] = a_im[0] + c_im[0]; |
| 93 |
a_im[1] = a_im[1] + c_im[1]; |
| 94 |
a_im[2] = a_im[2] + c_im[2]; |
| 95 |
a_im[3] = a_im[3] + c_im[3]; |
| 96 |
|
| 97 |
// Move pointers |
| 98 |
a_re += stride; |
| 99 |
a_im += stride; |
| 100 |
b_re += stride; |
| 101 |
b_im += stride; |
| 102 |
} |
| 103 |
} |
| 104 |
} |
| 105 |
*/ |
| 106 |
|
| 107 |
namespace lsp |
| 108 |
{ |
| 109 |
namespace sse |
| 110 |
{ |
| 111 |
#ifdef ARCH_I386 |
| 112 |
#define FFT_BUTTERFLY_BODY(add_b, add_a) \ |
| 113 |
/* Init pointers */ \ |
| 114 |
float *a_re = &dst_re[b*pairs*2]; \ |
| 115 |
float *a_im = &dst_im[b*pairs*2]; \ |
| 116 |
float *b_re = &a_re[pairs]; \ |
| 117 |
float *b_im = &a_im[pairs]; \ |
| 118 |
size_t p = pairs; \ |
| 119 |
size_t off = 0; \ |
| 120 |
float *tmp_re, *tmp_im; \ |
| 121 |
\ |
| 122 |
ARCH_X86_ASM \ |
| 123 |
( \ |
| 124 |
/* Prepare angle */ \ |
| 125 |
__ASM_EMIT("mov %[XFFT_A_RE], %[tmp_re]") \ |
| 126 |
__ASM_EMIT("mov %[XFFT_A_IM], %[tmp_im]") \ |
| 127 |
__ASM_EMIT("movaps (%[tmp_re], %[rank]), %%xmm6") /* xmm6 = angle_re[0..3] */ \ |
| 128 |
__ASM_EMIT("movaps (%[tmp_im], %[rank]), %%xmm7") /* xmm7 = angle_im[0..3] */ \ |
| 129 |
/* Start loop */ \ |
| 130 |
__ASM_EMIT(".align 16") \ |
| 131 |
__ASM_EMIT("1:") \ |
| 132 |
\ |
| 133 |
/* Load complex values */ \ |
| 134 |
/* predicate: xmm6 = w_re[0..3] */ \ |
| 135 |
/* predicate: xmm7 = w_im[0..3] */ \ |
| 136 |
__ASM_EMIT("mov %[a_re], %[tmp_re]") \ |
| 137 |
__ASM_EMIT("mov %[a_im], %[tmp_im]") \ |
| 138 |
__ASM_EMIT("movups (%[tmp_re], %[off]), %%xmm0") /* xmm0 = a_re[0..3] */ \ |
| 139 |
__ASM_EMIT("movups (%[tmp_im], %[off]), %%xmm1") /* xmm1 = a_im[0..3] */ \ |
| 140 |
__ASM_EMIT("mov %[b_re], %[tmp_re]") \ |
| 141 |
__ASM_EMIT("mov %[b_im], %[tmp_im]") \ |
| 142 |
__ASM_EMIT("movups (%[tmp_re], %[off]), %%xmm2") /* xmm2 = b_re[0..3] */ \ |
| 143 |
__ASM_EMIT("movups (%[tmp_im], %[off]), %%xmm3") /* xmm3 = b_im[0..3] */ \ |
| 144 |
\ |
| 145 |
/* Calculate complex multiplication */ \ |
| 146 |
__ASM_EMIT("movaps %%xmm2, %%xmm4") /* xmm4 = b_re[0..3] */ \ |
| 147 |
__ASM_EMIT("movaps %%xmm3, %%xmm5") /* xmm5 = b_im[0..3] */ \ |
| 148 |
__ASM_EMIT("mulps %%xmm6, %%xmm2") /* xmm2 = w_re[0..3] * b_re[0..3] */ \ |
| 149 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = w_re[0..3] * b_im[0..3] */ \ |
| 150 |
__ASM_EMIT("mulps %%xmm7, %%xmm4") /* xmm4 = w_im[0..3] * b_re[0..3] */ \ |
| 151 |
__ASM_EMIT("mulps %%xmm7, %%xmm5") /* xmm5 = w_im[0..3] * b_im[0..3] */ \ |
| 152 |
__ASM_EMIT(add_b " %%xmm5, %%xmm2") /* xmm2 = c_re[0..3] = w_re[0..3] * b_re[0..3] +- w_im[0..3] * b_im[0..3] */ \ |
| 153 |
__ASM_EMIT(add_a " %%xmm4, %%xmm3") /* xmm3 = c_im[0..3] = w_re[0..3] * b_im[0..3] -+ w_im[0..3] * b_re[0..3] */ \ |
| 154 |
\ |
| 155 |
/* Perform butterfly */ \ |
| 156 |
__ASM_EMIT("movaps %%xmm0, %%xmm4") /* xmm4 = a_re[0..3] */ \ |
| 157 |
__ASM_EMIT("movaps %%xmm1, %%xmm5") /* xmm5 = a_im[0..3] */ \ |
| 158 |
__ASM_EMIT("subps %%xmm2, %%xmm0") /* xmm0 = a_re[0..3] - c_re[0..3] */ \ |
| 159 |
__ASM_EMIT("subps %%xmm3, %%xmm1") /* xmm1 = a_im[0..3] - c_im[0..3] */ \ |
| 160 |
__ASM_EMIT("addps %%xmm4, %%xmm2") /* xmm2 = a_re[0..3] + c_re[0..3] */ \ |
| 161 |
__ASM_EMIT("addps %%xmm5, %%xmm3") /* xmm3 = a_im[0..3] + c_im[0..3] */ \ |
| 162 |
\ |
| 163 |
/* Store values */ \ |
| 164 |
__ASM_EMIT("movups %%xmm0, (%[tmp_re], %[off])") \ |
| 165 |
__ASM_EMIT("movups %%xmm1, (%[tmp_im], %[off])") \ |
| 166 |
__ASM_EMIT("mov %[a_re], %[tmp_re]") \ |
| 167 |
__ASM_EMIT("mov %[a_im], %[tmp_im]") \ |
| 168 |
__ASM_EMIT("movups %%xmm2, (%[tmp_re], %[off])") \ |
| 169 |
__ASM_EMIT("movups %%xmm3, (%[tmp_im], %[off])") \ |
| 170 |
__ASM_EMIT("add $0x10, %[off]") \ |
| 171 |
\ |
| 172 |
/* Repeat loop */ \ |
| 173 |
__ASM_EMIT("subl $4, %[p]") \ |
| 174 |
__ASM_EMIT("jz 2f") \ |
| 175 |
\ |
| 176 |
/* Rotate angle */ \ |
| 177 |
__ASM_EMIT("mov %[XFFT_W_RE], %[tmp_re]") \ |
| 178 |
__ASM_EMIT("mov %[XFFT_W_IM], %[tmp_im]") \ |
| 179 |
__ASM_EMIT("movaps (%[tmp_re], %[rank]), %%xmm0") /* xmm0 = w_re[0..3] */ \ |
| 180 |
__ASM_EMIT("movaps (%[tmp_im], %[rank]), %%xmm1") /* xmm1 = w_im[0..3] */ \ |
| 181 |
__ASM_EMIT("movaps %%xmm1, %%xmm3") /* xmm3 = w_im[0..3] */ \ |
| 182 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = w_re[0..3] */ \ |
| 183 |
\ |
| 184 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = a_re[0..3] * w_im[0..3] */ \ |
| 185 |
__ASM_EMIT("mulps %%xmm7, %%xmm1") /* xmm1 = a_im[0..3] * w_im[0..3] */ \ |
| 186 |
__ASM_EMIT("mulps %%xmm0, %%xmm6") /* xmm6 = a_re[0..3] * w_re[0..3] */ \ |
| 187 |
__ASM_EMIT("mulps %%xmm2, %%xmm7") /* xmm7 = a_im[0..3] * w_re[0..3] */ \ |
| 188 |
__ASM_EMIT("subps %%xmm1, %%xmm6") /* xmm6 = a_re[0..3] * w_re[0..3] + a_im[0..3] * w_im[0..3] */ \ |
| 189 |
__ASM_EMIT("addps %%xmm3, %%xmm7") /* xmm7 = a_im[0..3] * w_re[0..3] - a_re[0..3] * w_im[0..3] */ \ |
| 190 |
\ |
| 191 |
/* Repeat loop */ \ |
| 192 |
__ASM_EMIT("jmp 1b") \ |
| 193 |
__ASM_EMIT("2:") \ |
| 194 |
\ |
| 195 |
: [tmp_re] "=&r" (tmp_re), [tmp_im] "=&r" (tmp_im), \ |
| 196 |
[off] "+r" (off), [p] X86_PGREG (p) \ |
| 197 |
: [rank] "r" (rank), \ |
| 198 |
[a_re] "m" (a_re), [a_im] "m" (a_im), \ |
| 199 |
[b_re] "m" (b_re), [b_im] "m" (b_im), \ |
| 200 |
[XFFT_A_RE] "g" (&XFFT_A_RE[0]), [XFFT_A_IM] "g" (&XFFT_A_IM[0]), \ |
| 201 |
[XFFT_W_RE] "g" (&XFFT_W_RE[0]), [XFFT_W_IM] "g" (&XFFT_W_IM[0]) \ |
| 202 |
: "cc", "memory", \ |
| 203 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", \ |
| 204 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" \ |
| 205 |
); |
| 206 |
#else /* ARCH_X86_64 */ |
| 207 |
|
| 208 |
#define FFT_BUTTERFLY_BODY(add_b, add_a) \ |
| 209 |
/* Init pointers */ \ |
| 210 |
float *a_re = &dst_re[b*pairs*2]; \ |
| 211 |
float *a_im = &dst_im[b*pairs*2]; \ |
| 212 |
float *b_re = &a_re[pairs]; \ |
| 213 |
float *b_im = &a_im[pairs]; \ |
| 214 |
size_t p = pairs; \ |
| 215 |
size_t off = 0; \ |
| 216 |
\ |
| 217 |
ARCH_X86_ASM \ |
| 218 |
( \ |
| 219 |
/* Prepare angle */ \ |
| 220 |
__ASM_EMIT("movaps (%[XFFT_A_RE], %[rank]), %%xmm6") /* xmm6 = angle_re[0..3] */ \ |
| 221 |
__ASM_EMIT("movaps (%[XFFT_A_IM], %[rank]), %%xmm7") /* xmm7 = angle_im[0..3] */ \ |
| 222 |
/* Start loop */ \ |
| 223 |
__ASM_EMIT(".align 16") \ |
| 224 |
__ASM_EMIT("1:") \ |
| 225 |
\ |
| 226 |
/* Load complex values */ \ |
| 227 |
/* predicate: xmm6 = w_re[0..3] */ \ |
| 228 |
/* predicate: xmm7 = w_im[0..3] */ \ |
| 229 |
__ASM_EMIT("movups (%[a_re], %[off]), %%xmm0") /* xmm0 = a_re[0..3] */ \ |
| 230 |
__ASM_EMIT("movups (%[a_im], %[off]), %%xmm1") /* xmm1 = a_im[0..3] */ \ |
| 231 |
__ASM_EMIT("movups (%[b_re], %[off]), %%xmm2") /* xmm2 = b_re[0..3] */ \ |
| 232 |
__ASM_EMIT("movups (%[b_im], %[off]), %%xmm3") /* xmm3 = b_im[0..3] */ \ |
| 233 |
\ |
| 234 |
/* Calculate complex multiplication */ \ |
| 235 |
__ASM_EMIT("movaps %%xmm2, %%xmm4") /* xmm4 = b_re[0..3] */ \ |
| 236 |
__ASM_EMIT("movaps %%xmm3, %%xmm5") /* xmm5 = b_im[0..3] */ \ |
| 237 |
__ASM_EMIT("mulps %%xmm6, %%xmm2") /* xmm2 = w_re[0..3] * b_re[0..3] */ \ |
| 238 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = w_re[0..3] * b_im[0..3] */ \ |
| 239 |
__ASM_EMIT("mulps %%xmm7, %%xmm4") /* xmm4 = w_im[0..3] * b_re[0..3] */ \ |
| 240 |
__ASM_EMIT("mulps %%xmm7, %%xmm5") /* xmm5 = w_im[0..3] * b_im[0..3] */ \ |
| 241 |
__ASM_EMIT(add_b " %%xmm5, %%xmm2") /* xmm2 = c_re[0..3] = w_re[0..3] * b_re[0..3] +- w_im[0..3] * b_im[0..3] */ \ |
| 242 |
__ASM_EMIT(add_a " %%xmm4, %%xmm3") /* xmm3 = c_im[0..3] = w_re[0..3] * b_im[0..3] -+ w_im[0..3] * b_re[0..3] */ \ |
| 243 |
\ |
| 244 |
/* Perform butterfly */ \ |
| 245 |
__ASM_EMIT("movaps %%xmm0, %%xmm4") /* xmm4 = a_re[0..3] */ \ |
| 246 |
__ASM_EMIT("movaps %%xmm1, %%xmm5") /* xmm5 = a_im[0..3] */ \ |
| 247 |
__ASM_EMIT("subps %%xmm2, %%xmm0") /* xmm0 = a_re[0..3] - c_re[0..3] */ \ |
| 248 |
__ASM_EMIT("subps %%xmm3, %%xmm1") /* xmm1 = a_im[0..3] - c_im[0..3] */ \ |
| 249 |
__ASM_EMIT("addps %%xmm4, %%xmm2") /* xmm2 = a_re[0..3] + c_re[0..3] */ \ |
| 250 |
__ASM_EMIT("addps %%xmm5, %%xmm3") /* xmm3 = a_im[0..3] + c_im[0..3] */ \ |
| 251 |
\ |
| 252 |
/* Store values */ \ |
| 253 |
__ASM_EMIT("movups %%xmm2, (%[a_re], %[off])") \ |
| 254 |
__ASM_EMIT("movups %%xmm3, (%[a_im], %[off])") \ |
| 255 |
__ASM_EMIT("movups %%xmm0, (%[b_re], %[off])") \ |
| 256 |
__ASM_EMIT("movups %%xmm1, (%[b_im], %[off])") \ |
| 257 |
/* Update pointers */ \ |
| 258 |
__ASM_EMIT("add $0x10, %[off]") \ |
| 259 |
/* Repeat loop */ \ |
| 260 |
__ASM_EMIT("sub $4, %[p]") \ |
| 261 |
__ASM_EMIT("jz 2f") \ |
| 262 |
\ |
| 263 |
/* Rotate angle */ \ |
| 264 |
__ASM_EMIT("movaps (%[XFFT_W_IM], %[rank]), %%xmm1") /* xmm1 = w_im[0..3] */ \ |
| 265 |
__ASM_EMIT("movaps (%[XFFT_W_RE], %[rank]), %%xmm0") /* xmm0 = w_re[0..3] */ \ |
| 266 |
__ASM_EMIT("movaps %%xmm1, %%xmm3") /* xmm3 = w_im[0..3] */ \ |
| 267 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = w_re[0..3] */ \ |
| 268 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = a_re[0..3] * w_im[0..3] */ \ |
| 269 |
__ASM_EMIT("mulps %%xmm7, %%xmm1") /* xmm1 = a_im[0..3] * w_im[0..3] */ \ |
| 270 |
__ASM_EMIT("mulps %%xmm0, %%xmm6") /* xmm6 = a_re[0..3] * w_re[0..3] */ \ |
| 271 |
__ASM_EMIT("mulps %%xmm2, %%xmm7") /* xmm7 = a_im[0..3] * w_re[0..3] */ \ |
| 272 |
__ASM_EMIT("subps %%xmm1, %%xmm6") /* xmm6 = a_re[0..3] * w_re[0..3] + a_im[0..3] * w_im[0..3] */ \ |
| 273 |
__ASM_EMIT("addps %%xmm3, %%xmm7") /* xmm7 = a_im[0..3] * w_re[0..3] - a_re[0..3] * w_im[0..3] */ \ |
| 274 |
\ |
| 275 |
/* Repeat loop */ \ |
| 276 |
__ASM_EMIT("jmp 1b") \ |
| 277 |
__ASM_EMIT("2:") \ |
| 278 |
\ |
| 279 |
: [off] "+r" (off), [p] "+r" (p) \ |
| 280 |
: [a_re] "r"(a_re), [a_im] "r"(a_im), \ |
| 281 |
[b_re] "r"(b_re), [b_im] "r"(b_im), \ |
| 282 |
[rank] "r" (rank), \ |
| 283 |
[XFFT_A_RE] "r" (XFFT_A_RE), [XFFT_A_IM] "r" (XFFT_A_IM), \ |
| 284 |
[XFFT_W_RE] "r" (XFFT_W_RE), [XFFT_W_IM] "r"(XFFT_W_IM) \ |
| 285 |
: "cc", "memory", \ |
| 286 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", \ |
| 287 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" \ |
| 288 |
); |
| 289 |
#endif /* ARCH_X86_64 */ |
| 290 |
|
| 291 |
#define FFT_ANGLE_INIT \ |
| 292 |
ARCH_X86_ASM \ |
| 293 |
( \ |
| 294 |
__ASM_EMIT("sub $2, %[rank]") /* rank -= 2 */ \ |
| 295 |
__ASM_EMIT("shl $4, %[rank]") /* rank *= 16 */ \ |
| 296 |
__ASM_EMIT("movaps (%[XFFT_A_RE], %[rank]), %%xmm6") /* xmm6 = angle_re[0..3] */ \ |
| 297 |
__ASM_EMIT("movaps (%[XFFT_A_IM], %[rank]), %%xmm7") /* xmm7 = angle_im[0..3] */ \ |
| 298 |
: [rank] "+r" (rank) \ |
| 299 |
: \ |
| 300 |
[XFFT_A_RE] "r"(XFFT_A_RE), \ |
| 301 |
[XFFT_A_IM] "r"(XFFT_A_IM) \ |
| 302 |
: "cc", \ |
| 303 |
"%xmm6", "%xmm7" \ |
| 304 |
); |
| 305 |
|
| 306 |
#define FFT_ANGLE_ROTATE \ |
| 307 |
ARCH_X86_ASM \ |
| 308 |
( \ |
| 309 |
/* xmm6 = a_re[0..3] */ \ |
| 310 |
/* xmm7 = a_im[0..3] */ \ |
| 311 |
__ASM_EMIT("movaps (%[XFFT_W_IM], %[rank]), %%xmm1") /* xmm1 = w_im[0..3] */ \ |
| 312 |
__ASM_EMIT("movaps (%[XFFT_W_RE], %[rank]), %%xmm0") /* xmm0 = w_re[0..3] */ \ |
| 313 |
__ASM_EMIT("movaps %%xmm1, %%xmm3") /* xmm3 = w_im[0..3] */ \ |
| 314 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = w_re[0..3] */ \ |
| 315 |
\ |
| 316 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = a_re[0..3] * w_im[0..3] */ \ |
| 317 |
__ASM_EMIT("mulps %%xmm7, %%xmm1") /* xmm1 = a_im[0..3] * w_im[0..3] */ \ |
| 318 |
__ASM_EMIT("mulps %%xmm0, %%xmm6") /* xmm6 = a_re[0..3] * w_re[0..3] */ \ |
| 319 |
__ASM_EMIT("mulps %%xmm2, %%xmm7") /* xmm7 = a_im[0..3] * w_re[0..3] */ \ |
| 320 |
__ASM_EMIT("subps %%xmm1, %%xmm6") /* xmm6 = a_re[0..3] * w_re[0..3] + a_im[0..3] * w_im[0..3] */ \ |
| 321 |
__ASM_EMIT("addps %%xmm3, %%xmm7") /* xmm7 = a_im[0..3] * w_re[0..3] - a_re[0..3] * w_im[0..3] */ \ |
| 322 |
: \ |
| 323 |
: [rank] "r" (rank), \ |
| 324 |
[XFFT_W_RE] "r"(XFFT_W_RE), \ |
| 325 |
[XFFT_W_IM] "r"(XFFT_W_IM) \ |
| 326 |
: "cc", \ |
| 327 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", \ |
| 328 |
"%xmm6", "%xmm7" \ |
| 329 |
); |
| 330 |
|
| 331 |
static inline void butterfly_direct(float *dst_re, float *dst_im, size_t rank, size_t blocks) |
| 332 |
{ |
| 333 |
size_t pairs = 1 << rank; |
| 334 |
rank = (rank - 2) << 4; |
| 335 |
|
| 336 |
for (size_t b=0; b<blocks; ++b) |
| 337 |
{ |
| 338 |
FFT_BUTTERFLY_BODY("addps", "subps"); |
| 339 |
} |
| 340 |
} |
| 341 |
|
| 342 |
static inline void butterfly_reverse(float *dst_re, float *dst_im, size_t rank, size_t blocks) |
| 343 |
{ |
| 344 |
size_t pairs = 1 << rank; |
| 345 |
rank = (rank - 2) << 4; |
| 346 |
|
| 347 |
for (size_t b=0; b<blocks; ++b) |
| 348 |
{ |
| 349 |
FFT_BUTTERFLY_BODY("subps", "addps"); |
| 350 |
} |
| 351 |
} |
| 352 |
|
| 353 |
|
| 354 |
#undef FFT_ANGLE_INIT |
| 355 |
#undef FFT_ANGLE_ROTATE |
| 356 |
#undef FFT_BUTTERFLY_BODY |
| 357 |
|
| 358 |
} /* namespace sse */ |
| 359 |
} /* namespace lsp */ |
| 360 |
|
| 361 |
#endif /* PRIVATE_DSP_ARCH_X86_SSE_FFT_BUTTERFLY_H_ */ |