| 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_FASTCONV_PARSE_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_X86_SSE_FASTCONV_PARSE_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 |
namespace lsp |
| 30 |
{ |
| 31 |
namespace sse |
| 32 |
{ |
| 33 |
void fastconv_parse(float *dst, const float *src, size_t rank) |
| 34 |
{ |
| 35 |
// Prepare for butterflies |
| 36 |
const float *wk = &XFFT_W[(rank-3) << 3]; |
| 37 |
const float *ak = &XFFT_A[(rank-3) << 3]; |
| 38 |
size_t items = size_t(1) << (rank + 1); |
| 39 |
size_t bs = items; |
| 40 |
size_t n = bs >> 1; |
| 41 |
|
| 42 |
// Iterate first cycle |
| 43 |
if (n > 4) |
| 44 |
{ |
| 45 |
// ONE LARGE CYCLE |
| 46 |
// Set initial values of pointers |
| 47 |
float *a = dst; |
| 48 |
float *b = &a[n]; |
| 49 |
size_t k = n; |
| 50 |
|
| 51 |
ARCH_X86_ASM |
| 52 |
( |
| 53 |
__ASM_EMIT("movaps 0x00(%[ak]), %%xmm6") /* xmm6 = rA[i] */ |
| 54 |
__ASM_EMIT("movaps 0x10(%[ak]), %%xmm7") /* xmm7 = iA[i] */ |
| 55 |
__ASM_EMIT("xorps %%xmm5, %%xmm5") /* xmm5 = 0 */ |
| 56 |
: |
| 57 |
: [ak] "r"(ak) |
| 58 |
: |
| 59 |
"%xmm5", "%xmm6", "%xmm7" |
| 60 |
); |
| 61 |
|
| 62 |
ARCH_X86_ASM |
| 63 |
( |
| 64 |
// __ASM_EMIT(".align 16") |
| 65 |
__ASM_EMIT("1:") |
| 66 |
|
| 67 |
__ASM_EMIT("movups 0x00(%[src]), %%xmm0") /* xmm0 = s[i] */ |
| 68 |
__ASM_EMIT("xorps %%xmm3, %%xmm3") /* xmm3 = 0 */ |
| 69 |
__ASM_EMIT("movaps %%xmm0, %%xmm4") /* xmm4 = s[i] */ |
| 70 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = s[i] */ |
| 71 |
__ASM_EMIT("mulps %%xmm7, %%xmm4") /* xmm4 = s[i]*iA[i] */ |
| 72 |
__ASM_EMIT("mulps %%xmm6, %%xmm2") /* xmm2 = s[i]*rA[i] */ |
| 73 |
__ASM_EMIT("subps %%xmm4, %%xmm3") /* xmm3 = -s[i]*iA[i] */ |
| 74 |
|
| 75 |
__ASM_EMIT("movups %%xmm0, 0x00(%[a])") |
| 76 |
__ASM_EMIT("movups %%xmm5, 0x10(%[a])") |
| 77 |
__ASM_EMIT("movups %%xmm2, 0x00(%[b])") |
| 78 |
__ASM_EMIT("movups %%xmm3, 0x10(%[b])") |
| 79 |
|
| 80 |
__ASM_EMIT("add $0x10, %[src]") |
| 81 |
__ASM_EMIT("add $0x20, %[a]") |
| 82 |
__ASM_EMIT("add $0x20, %[b]") |
| 83 |
|
| 84 |
/* Repeat loop */ |
| 85 |
__ASM_EMIT("sub $8, %[k]") |
| 86 |
__ASM_EMIT("jz 2f") |
| 87 |
|
| 88 |
/* Rotate angle */ |
| 89 |
__ASM_EMIT("movaps 0x00(%[wk]), %%xmm0") /* xmm0 = rw[i] */ |
| 90 |
__ASM_EMIT("movaps 0x10(%[wk]), %%xmm1") /* xmm1 = iw[i] */ |
| 91 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = rw[i] */ |
| 92 |
__ASM_EMIT("movaps %%xmm1, %%xmm3") /* xmm3 = iw[i] */ |
| 93 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = rA[i] * iw[i] */ |
| 94 |
__ASM_EMIT("mulps %%xmm7, %%xmm1") /* xmm1 = iA[i] * iw[i] */ |
| 95 |
__ASM_EMIT("mulps %%xmm0, %%xmm6") /* xmm6 = rA[i] * rw[i] */ |
| 96 |
__ASM_EMIT("mulps %%xmm2, %%xmm7") /* xmm7 = iA[i] * rw[i] */ |
| 97 |
__ASM_EMIT("subps %%xmm1, %%xmm6") /* xmm6 = rA[i] * rw[i] - iA[i] * iw[i] */ |
| 98 |
__ASM_EMIT("addps %%xmm3, %%xmm7") /* xmm7 = rA[i] * iw[i] + iA[i] * rw[i] */ |
| 99 |
__ASM_EMIT("jmp 1b") |
| 100 |
|
| 101 |
__ASM_EMIT("2:") |
| 102 |
|
| 103 |
: [a] "+r" (a), [b] "+r" (b), [src] "+r" (src), [k] "+r" (k) |
| 104 |
: [wk] "r"(wk) |
| 105 |
: "cc", "memory", |
| 106 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 107 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 108 |
); |
| 109 |
|
| 110 |
wk -= 8; |
| 111 |
ak -= 8; |
| 112 |
n >>= 1; |
| 113 |
bs >>= 1; |
| 114 |
} |
| 115 |
else |
| 116 |
{ |
| 117 |
// Unpack 4x real to 4x split complex |
| 118 |
ARCH_X86_ASM |
| 119 |
( |
| 120 |
__ASM_EMIT("movups 0x00(%[src]), %%xmm0") /* xmm0 = s[i] */ |
| 121 |
__ASM_EMIT("xorps %%xmm1, %%xmm1") /* xmm1 = 0 */ |
| 122 |
__ASM_EMIT("movups %%xmm0, 0x00(%[dst])") |
| 123 |
__ASM_EMIT("movups %%xmm1, 0x10(%[dst])") |
| 124 |
: |
| 125 |
: [items] "r" (items), [src] "r" (src), [dst] "r" (dst) |
| 126 |
: "cc", "memory", |
| 127 |
"%xmm0", "%xmm1" |
| 128 |
); |
| 129 |
} |
| 130 |
|
| 131 |
// Iterate butterflies |
| 132 |
for (; n > 4; n >>= 1, bs >>= 1) |
| 133 |
{ |
| 134 |
for (size_t p=0; p<items; p += bs) |
| 135 |
{ |
| 136 |
// Set initial values of pointers |
| 137 |
float *a = &dst[p]; |
| 138 |
float *b = &a[n]; |
| 139 |
size_t k = n; |
| 140 |
|
| 141 |
ARCH_X86_ASM |
| 142 |
( |
| 143 |
__ASM_EMIT("movaps 0x00(%[ak]), %%xmm6") /* xmm6 = rA[i] */ |
| 144 |
__ASM_EMIT("movaps 0x10(%[ak]), %%xmm7") /* xmm7 = iA[i] */ |
| 145 |
: |
| 146 |
: [ak] "r"(ak) |
| 147 |
: "%xmm6", "%xmm7" |
| 148 |
); |
| 149 |
|
| 150 |
ARCH_X86_ASM |
| 151 |
( |
| 152 |
// __ASM_EMIT(".align 16") |
| 153 |
__ASM_EMIT("1:") |
| 154 |
|
| 155 |
__ASM_EMIT("movups 0x00(%[a]), %%xmm0") /* xmm0 = ra[i] */ |
| 156 |
__ASM_EMIT("movups 0x10(%[a]), %%xmm1") /* xmm1 = ia[i] */ |
| 157 |
__ASM_EMIT("movups 0x00(%[b]), %%xmm2") /* xmm2 = rb[i] */ |
| 158 |
__ASM_EMIT("movups 0x10(%[b]), %%xmm3") /* xmm3 = ib[i] */ |
| 159 |
|
| 160 |
__ASM_EMIT("movaps %%xmm0, %%xmm4") /* xmm4 = ra[i] */ |
| 161 |
__ASM_EMIT("movaps %%xmm1, %%xmm5") /* xmm5 = ia[i] */ |
| 162 |
__ASM_EMIT("addps %%xmm2, %%xmm0") /* xmm0 = ra[i]+rb[i] */ |
| 163 |
__ASM_EMIT("addps %%xmm3, %%xmm1") /* xmm1 = ia[i]+ib[i] */ |
| 164 |
__ASM_EMIT("subps %%xmm2, %%xmm4") /* xmm4 = ra[i]-rb[i] = rc[i] */ |
| 165 |
__ASM_EMIT("subps %%xmm3, %%xmm5") /* xmm5 = ia[i]-ib[i] = ic[i] */ |
| 166 |
__ASM_EMIT("movaps %%xmm4, %%xmm2") /* xmm2 = rc[i] */ |
| 167 |
__ASM_EMIT("movaps %%xmm5, %%xmm3") /* xmm3 = ic[i] */ |
| 168 |
__ASM_EMIT("mulps %%xmm6, %%xmm4") /* xmm4 = rA[i]*rc[i] */ |
| 169 |
__ASM_EMIT("mulps %%xmm7, %%xmm5") /* xmm5 = iA[i]*ic[i] */ |
| 170 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = rA[i]*ic[i] */ |
| 171 |
__ASM_EMIT("mulps %%xmm7, %%xmm2") /* xmm2 = iA[i]*rc[i] */ |
| 172 |
__ASM_EMIT("addps %%xmm5, %%xmm4") /* xmm4 = rA[i]*rc[i] + iA[i]*ic[i] */ |
| 173 |
__ASM_EMIT("subps %%xmm2, %%xmm3") /* xmm3 = rA[i]*ic[i] - iA[i]*rc[i] */ |
| 174 |
|
| 175 |
__ASM_EMIT("movups %%xmm0, 0x00(%[a])") |
| 176 |
__ASM_EMIT("movups %%xmm1, 0x10(%[a])") |
| 177 |
__ASM_EMIT("movups %%xmm4, 0x00(%[b])") |
| 178 |
__ASM_EMIT("movups %%xmm3, 0x10(%[b])") |
| 179 |
|
| 180 |
/* Repeat loop */ |
| 181 |
__ASM_EMIT("add $0x20, %[a]") |
| 182 |
__ASM_EMIT("add $0x20, %[b]") |
| 183 |
__ASM_EMIT("sub $8, %[k]") |
| 184 |
__ASM_EMIT("jz 2f") |
| 185 |
|
| 186 |
/* Rotate angle */ |
| 187 |
__ASM_EMIT("movaps 0x00(%[wk]), %%xmm0") /* xmm0 = rw */ |
| 188 |
__ASM_EMIT("movaps 0x10(%[wk]), %%xmm1") /* xmm1 = iw */ |
| 189 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = rw */ |
| 190 |
__ASM_EMIT("movaps %%xmm1, %%xmm3") /* xmm3 = iw */ |
| 191 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = ra * iw */ |
| 192 |
__ASM_EMIT("mulps %%xmm7, %%xmm1") /* xmm1 = ia * iw */ |
| 193 |
__ASM_EMIT("mulps %%xmm0, %%xmm6") /* xmm6 = ra * rw */ |
| 194 |
__ASM_EMIT("mulps %%xmm2, %%xmm7") /* xmm7 = ia * rw */ |
| 195 |
__ASM_EMIT("subps %%xmm1, %%xmm6") /* xmm6 = ra * rw - ia * iw */ |
| 196 |
__ASM_EMIT("addps %%xmm3, %%xmm7") /* xmm7 = ra * iw + ia * rw */ |
| 197 |
__ASM_EMIT("jmp 1b") |
| 198 |
|
| 199 |
__ASM_EMIT("2:") |
| 200 |
|
| 201 |
: [a] "+r" (a), [b] "+r" (b), [k] "+r" (k) |
| 202 |
: [wk] "r"(wk) |
| 203 |
: "cc", "memory", |
| 204 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 205 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 206 |
); |
| 207 |
} |
| 208 |
|
| 209 |
ak -= 8; |
| 210 |
wk -= 8; |
| 211 |
} |
| 212 |
|
| 213 |
ARCH_X86_ASM |
| 214 |
( |
| 215 |
// __ASM_EMIT(".align 16") |
| 216 |
__ASM_EMIT("1:") |
| 217 |
|
| 218 |
__ASM_EMIT("movups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 r1 r2 r3 */ |
| 219 |
__ASM_EMIT("movups 0x10(%[dst]), %%xmm1") /* xmm1 = i0 i1 i2 i3 */ |
| 220 |
__ASM_EMIT("movups 0x20(%[dst]), %%xmm4") |
| 221 |
__ASM_EMIT("movups 0x30(%[dst]), %%xmm5") |
| 222 |
|
| 223 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0 r1 r2 r3 */ |
| 224 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") |
| 225 |
__ASM_EMIT("unpcklps %%xmm1, %%xmm0") /* xmm0 = r0 i0 r1 i1 */ |
| 226 |
__ASM_EMIT("unpcklps %%xmm5, %%xmm4") |
| 227 |
__ASM_EMIT("unpckhps %%xmm1, %%xmm2") /* xmm2 = r2 i2 r3 i3 */ |
| 228 |
__ASM_EMIT("unpckhps %%xmm5, %%xmm6") |
| 229 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0 i0 r1 i1 */ |
| 230 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") |
| 231 |
|
| 232 |
__ASM_EMIT("addps %%xmm2, %%xmm0") /* xmm0 = r0+r2 i0+i2 r1+r3 i1+i3 = r0k i0k r2k i2k */ |
| 233 |
__ASM_EMIT("addps %%xmm6, %%xmm4") |
| 234 |
__ASM_EMIT("subps %%xmm2, %%xmm1") /* xmm1 = r0-r2 i0-i2 r1-r3 i1-i3 = r1k i1k r3k i3k */ |
| 235 |
__ASM_EMIT("subps %%xmm6, %%xmm5") |
| 236 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0k i0k r2k i2k */ |
| 237 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") |
| 238 |
__ASM_EMIT("unpcklps %%xmm1, %%xmm0") /* xmm0 = r0k r1k i0k i1k */ |
| 239 |
__ASM_EMIT("unpcklps %%xmm5, %%xmm4") |
| 240 |
__ASM_EMIT("unpckhps %%xmm1, %%xmm2") /* xmm2 = r2k r3k i2k i3k */ |
| 241 |
__ASM_EMIT("unpckhps %%xmm5, %%xmm6") |
| 242 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0k r1k i0k i1k */ |
| 243 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") |
| 244 |
__ASM_EMIT("shufps $0x6c, %%xmm2, %%xmm2") /* xmm2 = r2k i3k i2k r3k */ |
| 245 |
__ASM_EMIT("shufps $0x6c, %%xmm6, %%xmm6") |
| 246 |
|
| 247 |
__ASM_EMIT("addps %%xmm2, %%xmm0") /* xmm0 = r0k+r2k r1k+i3k i0k+i2k i1k+r3k */ |
| 248 |
__ASM_EMIT("addps %%xmm6, %%xmm4") |
| 249 |
__ASM_EMIT("subps %%xmm2, %%xmm1") /* xmm1 = r0k-r2k r1k-i3k i0k-i2k i1k-r3k */ |
| 250 |
__ASM_EMIT("subps %%xmm6, %%xmm5") |
| 251 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0k+r2k r1k+i3k i0k+i2k i1k+r3k */ |
| 252 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") |
| 253 |
__ASM_EMIT("unpcklps %%xmm1, %%xmm0") /* xmm0 = r0k+r2k r0k-r2k r1k+i3k r1k-i3k */ |
| 254 |
__ASM_EMIT("unpcklps %%xmm5, %%xmm4") |
| 255 |
__ASM_EMIT("unpckhps %%xmm1, %%xmm2") /* xmm2 = i0k+i2k i0k-i2k i1k+r3k i1k-r3k */ |
| 256 |
__ASM_EMIT("unpckhps %%xmm5, %%xmm6") |
| 257 |
__ASM_EMIT("shufps $0xb4, %%xmm2, %%xmm2") /* xmm2 = i0k+i2k i0k-i2k i1k-r3k i1k+r3k */ |
| 258 |
__ASM_EMIT("shufps $0xb4, %%xmm6, %%xmm6") |
| 259 |
|
| 260 |
__ASM_EMIT("movups %%xmm0, 0x00(%[dst])") |
| 261 |
__ASM_EMIT("movups %%xmm2, 0x10(%[dst])") |
| 262 |
__ASM_EMIT("movups %%xmm4, 0x20(%[dst])") |
| 263 |
__ASM_EMIT("movups %%xmm6, 0x30(%[dst])") |
| 264 |
|
| 265 |
__ASM_EMIT("add $0x40, %[dst]") |
| 266 |
__ASM_EMIT("sub $16, %[k]") |
| 267 |
__ASM_EMIT("jnz 1b") |
| 268 |
|
| 269 |
: [dst] "+r" (dst), [k] "+r" (items) |
| 270 |
: |
| 271 |
: "cc", "memory", |
| 272 |
"%xmm0", "%xmm1", "%xmm2", |
| 273 |
"%xmm4", "%xmm5", "%xmm6" |
| 274 |
); |
| 275 |
} |
| 276 |
|
| 277 |
static inline void fastconv_parse_internal(float *dst, const float *src, size_t rank) |
| 278 |
{ |
| 279 |
// Prepare for butterflies |
| 280 |
const float *wk = &XFFT_W[(rank-3) << 3]; |
| 281 |
const float *ak = &XFFT_A[(rank-3) << 3]; |
| 282 |
size_t items = size_t(1) << (rank + 1); |
| 283 |
size_t bs = items; |
| 284 |
size_t n = bs >> 1; |
| 285 |
|
| 286 |
// Iterate first cycle |
| 287 |
if (n > 4) |
| 288 |
{ |
| 289 |
// ONE LARGE CYCLE |
| 290 |
// Set initial values of pointers |
| 291 |
float *a = dst; |
| 292 |
float *b = &a[n]; |
| 293 |
size_t k = n; |
| 294 |
|
| 295 |
ARCH_X86_ASM |
| 296 |
( |
| 297 |
__ASM_EMIT("movaps 0x00(%[ak]), %%xmm6") /* xmm6 = rA[i] */ |
| 298 |
__ASM_EMIT("movaps 0x10(%[ak]), %%xmm7") /* xmm7 = iA[i] */ |
| 299 |
__ASM_EMIT("xorps %%xmm5, %%xmm5") /* xmm5 = 0 */ |
| 300 |
: |
| 301 |
: [ak] "r"(ak) |
| 302 |
: |
| 303 |
"%xmm5", "%xmm6", "%xmm7" |
| 304 |
); |
| 305 |
|
| 306 |
ARCH_X86_ASM |
| 307 |
( |
| 308 |
// __ASM_EMIT(".align 16") |
| 309 |
__ASM_EMIT("1:") |
| 310 |
|
| 311 |
__ASM_EMIT("movups 0x00(%[src]), %%xmm0") /* xmm0 = s[i] */ |
| 312 |
|
| 313 |
__ASM_EMIT("xorps %%xmm3, %%xmm3") /* xmm3 = 0 */ |
| 314 |
__ASM_EMIT("movaps %%xmm0, %%xmm4") /* xmm4 = s[i] */ |
| 315 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = s[i] */ |
| 316 |
__ASM_EMIT("mulps %%xmm7, %%xmm4") /* xmm4 = s[i]*iA[i] */ |
| 317 |
__ASM_EMIT("mulps %%xmm6, %%xmm2") /* xmm2 = s[i]*rA[i] */ |
| 318 |
__ASM_EMIT("subps %%xmm4, %%xmm3") /* xmm3 = -s[i]*iA[i] */ |
| 319 |
|
| 320 |
__ASM_EMIT("movups %%xmm0, 0x00(%[a])") |
| 321 |
__ASM_EMIT("movups %%xmm5, 0x10(%[a])") |
| 322 |
__ASM_EMIT("movups %%xmm2, 0x00(%[b])") |
| 323 |
__ASM_EMIT("movups %%xmm3, 0x10(%[b])") |
| 324 |
|
| 325 |
__ASM_EMIT("add $0x10, %[src]") |
| 326 |
__ASM_EMIT("add $0x20, %[a]") |
| 327 |
__ASM_EMIT("add $0x20, %[b]") |
| 328 |
|
| 329 |
/* Repeat loop */ |
| 330 |
__ASM_EMIT32("subl $8, %[k]") |
| 331 |
__ASM_EMIT64("sub $8, %[k]") |
| 332 |
__ASM_EMIT("jz 2f") |
| 333 |
|
| 334 |
/* Rotate angle */ |
| 335 |
__ASM_EMIT("movaps 0x00(%[wk]), %%xmm0") /* xmm0 = rw[i] */ |
| 336 |
__ASM_EMIT("movaps 0x10(%[wk]), %%xmm1") /* xmm1 = iw[i] */ |
| 337 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = rw[i] */ |
| 338 |
__ASM_EMIT("movaps %%xmm1, %%xmm3") /* xmm3 = iw[i] */ |
| 339 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = rA[i] * iw[i] */ |
| 340 |
__ASM_EMIT("mulps %%xmm7, %%xmm1") /* xmm1 = iA[i] * iw[i] */ |
| 341 |
__ASM_EMIT("mulps %%xmm0, %%xmm6") /* xmm6 = rA[i] * rw[i] */ |
| 342 |
__ASM_EMIT("mulps %%xmm2, %%xmm7") /* xmm7 = iA[i] * rw[i] */ |
| 343 |
__ASM_EMIT("subps %%xmm1, %%xmm6") /* xmm6 = rA[i] * rw[i] - iA[i] * iw[i] */ |
| 344 |
__ASM_EMIT("addps %%xmm3, %%xmm7") /* xmm7 = rA[i] * iw[i] + iA[i] * rw[i] */ |
| 345 |
__ASM_EMIT("jmp 1b") |
| 346 |
|
| 347 |
__ASM_EMIT("2:") |
| 348 |
|
| 349 |
: [a] "+r" (a), [b] "+r" (b), [src] "+r" (src), [k] "+r" (k) |
| 350 |
: [wk] "r" (wk) |
| 351 |
: "cc", "memory", |
| 352 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 353 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 354 |
); |
| 355 |
|
| 356 |
wk -= 8; |
| 357 |
ak -= 8; |
| 358 |
n >>= 1; |
| 359 |
bs >>= 1; |
| 360 |
} |
| 361 |
else |
| 362 |
{ |
| 363 |
// Unpack 4x real to 4x split complex |
| 364 |
ARCH_X86_ASM |
| 365 |
( |
| 366 |
__ASM_EMIT("movups 0x00(%[src]), %%xmm0") /* xmm0 = s[i] */ |
| 367 |
__ASM_EMIT("xorps %%xmm1, %%xmm1") /* xmm1 = 0 */ |
| 368 |
__ASM_EMIT("movups %%xmm0, 0x00(%[dst])") |
| 369 |
__ASM_EMIT("movups %%xmm1, 0x10(%[dst])") |
| 370 |
: |
| 371 |
: [src] "r" (src), [dst] "r" (dst) |
| 372 |
: "cc", "memory", |
| 373 |
"%xmm0", "%xmm1" |
| 374 |
); |
| 375 |
} |
| 376 |
|
| 377 |
// Iterate butterflies |
| 378 |
for (; n > 4; n >>= 1, bs >>= 1) |
| 379 |
{ |
| 380 |
for (size_t p=0; p<items; p += bs) |
| 381 |
{ |
| 382 |
// Set initial values of pointers |
| 383 |
float *a = &dst[p]; |
| 384 |
float *b = &a[n]; |
| 385 |
size_t k = n; |
| 386 |
|
| 387 |
ARCH_X86_ASM |
| 388 |
( |
| 389 |
__ASM_EMIT("movaps 0x00(%[ak]), %%xmm6") /* xmm6 = rA[i] */ |
| 390 |
__ASM_EMIT("movaps 0x10(%[ak]), %%xmm7") /* xmm7 = iA[i] */ |
| 391 |
: |
| 392 |
: [ak] "r"(ak) |
| 393 |
: "%xmm6", "%xmm7" |
| 394 |
); |
| 395 |
|
| 396 |
ARCH_X86_ASM |
| 397 |
( |
| 398 |
__ASM_EMIT("1:") |
| 399 |
|
| 400 |
__ASM_EMIT("movups 0x00(%[a]), %%xmm0") /* xmm0 = ra[i] */ |
| 401 |
__ASM_EMIT("movups 0x10(%[a]), %%xmm1") /* xmm1 = ia[i] */ |
| 402 |
__ASM_EMIT("movups 0x00(%[b]), %%xmm2") /* xmm2 = rb[i] */ |
| 403 |
__ASM_EMIT("movups 0x10(%[b]), %%xmm3") /* xmm3 = ib[i] */ |
| 404 |
|
| 405 |
__ASM_EMIT("movaps %%xmm0, %%xmm4") /* xmm4 = ra[i] */ |
| 406 |
__ASM_EMIT("movaps %%xmm1, %%xmm5") /* xmm5 = ia[i] */ |
| 407 |
__ASM_EMIT("addps %%xmm2, %%xmm0") /* xmm0 = ra[i]+rb[i] */ |
| 408 |
__ASM_EMIT("addps %%xmm3, %%xmm1") /* xmm1 = ia[i]+ib[i] */ |
| 409 |
__ASM_EMIT("subps %%xmm2, %%xmm4") /* xmm4 = ra[i]-rb[i] = rc[i] */ |
| 410 |
__ASM_EMIT("subps %%xmm3, %%xmm5") /* xmm5 = ia[i]-ib[i] = ic[i] */ |
| 411 |
__ASM_EMIT("movaps %%xmm4, %%xmm2") /* xmm2 = rc[i] */ |
| 412 |
__ASM_EMIT("movaps %%xmm5, %%xmm3") /* xmm3 = ic[i] */ |
| 413 |
__ASM_EMIT("mulps %%xmm6, %%xmm4") /* xmm4 = rA[i]*rc[i] */ |
| 414 |
__ASM_EMIT("mulps %%xmm7, %%xmm5") /* xmm5 = iA[i]*ic[i] */ |
| 415 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = rA[i]*ic[i] */ |
| 416 |
__ASM_EMIT("mulps %%xmm7, %%xmm2") /* xmm2 = iA[i]*rc[i] */ |
| 417 |
__ASM_EMIT("addps %%xmm5, %%xmm4") /* xmm4 = rA[i]*rc[i] + iA[i]*ic[i] */ |
| 418 |
__ASM_EMIT("subps %%xmm2, %%xmm3") /* xmm3 = rA[i]*ic[i] - iA[i]*rc[i] */ |
| 419 |
|
| 420 |
__ASM_EMIT("movups %%xmm0, 0x00(%[a])") |
| 421 |
__ASM_EMIT("movups %%xmm1, 0x10(%[a])") |
| 422 |
__ASM_EMIT("movups %%xmm4, 0x00(%[b])") |
| 423 |
__ASM_EMIT("movups %%xmm3, 0x10(%[b])") |
| 424 |
|
| 425 |
/* Repeat loop */ |
| 426 |
__ASM_EMIT("add $0x20, %[a]") |
| 427 |
__ASM_EMIT("add $0x20, %[b]") |
| 428 |
__ASM_EMIT("sub $8, %[k]") |
| 429 |
__ASM_EMIT("jz 2f") |
| 430 |
|
| 431 |
/* Rotate angle */ |
| 432 |
__ASM_EMIT("movaps 0x00(%[wk]), %%xmm0") /* xmm0 = rw */ |
| 433 |
__ASM_EMIT("movaps 0x10(%[wk]), %%xmm1") /* xmm1 = iw */ |
| 434 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = rw */ |
| 435 |
__ASM_EMIT("movaps %%xmm1, %%xmm3") /* xmm3 = iw */ |
| 436 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = ra * iw */ |
| 437 |
__ASM_EMIT("mulps %%xmm7, %%xmm1") /* xmm1 = ia * iw */ |
| 438 |
__ASM_EMIT("mulps %%xmm0, %%xmm6") /* xmm6 = ra * rw */ |
| 439 |
__ASM_EMIT("mulps %%xmm2, %%xmm7") /* xmm7 = ia * rw */ |
| 440 |
__ASM_EMIT("subps %%xmm1, %%xmm6") /* xmm6 = ra * rw - ia * iw */ |
| 441 |
__ASM_EMIT("addps %%xmm3, %%xmm7") /* xmm7 = ra * iw + ia * rw */ |
| 442 |
__ASM_EMIT("jmp 1b") |
| 443 |
|
| 444 |
__ASM_EMIT("2:") |
| 445 |
|
| 446 |
: [a] "+r" (a), [b] "+r" (b), [k] "+r" (k) |
| 447 |
: [wk] "r"(wk) |
| 448 |
: "cc", "memory", |
| 449 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 450 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 451 |
); |
| 452 |
} |
| 453 |
|
| 454 |
ak -= 8; |
| 455 |
wk -= 8; |
| 456 |
} |
| 457 |
} |
| 458 |
} |
| 459 |
} |
| 460 |
|
| 461 |
#endif /* PRIVATE_DSP_ARCH_X86_SSE_FASTCONV_PARSE_H_ */ |