| 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_RESTORE_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_X86_SSE_FASTCONV_RESTORE_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_restore(float *dst, float *tmp, size_t rank) |
| 34 |
{ |
| 35 |
// Prepare for butterflies |
| 36 |
size_t last = size_t(1) << rank; |
| 37 |
size_t items = last; |
| 38 |
float *ptr = tmp; |
| 39 |
|
| 40 |
ARCH_X86_ASM |
| 41 |
( |
| 42 |
__ASM_EMIT("1:") |
| 43 |
|
| 44 |
// Load data |
| 45 |
__ASM_EMIT("movups 0x00(%[ptr]), %%xmm1") |
| 46 |
__ASM_EMIT("movups 0x10(%[ptr]), %%xmm3") |
| 47 |
__ASM_EMIT("movups 0x20(%[ptr]), %%xmm5") |
| 48 |
__ASM_EMIT("movups 0x30(%[ptr]), %%xmm7") |
| 49 |
|
| 50 |
// Do x4 reverse butterflies |
| 51 |
// xmm1 = r0 r1 r2 r3 |
| 52 |
// xmm3 = i0 i1 i2 i3 |
| 53 |
// xmm5 = r4 r5 r6 r7 |
| 54 |
// xmm7 = i4 i5 i6 i7 |
| 55 |
__ASM_EMIT("movaps %%xmm1, %%xmm0") /* xmm0 = r0 r1 r2 r3 */ |
| 56 |
__ASM_EMIT("movaps %%xmm5, %%xmm4") |
| 57 |
__ASM_EMIT("shufps $0x88, %%xmm3, %%xmm0") /* xmm0 = r0 r2 i0 i2 */ |
| 58 |
__ASM_EMIT("shufps $0x88, %%xmm7, %%xmm4") |
| 59 |
__ASM_EMIT("shufps $0xdd, %%xmm3, %%xmm1") /* xmm1 = r1 r3 i1 i3 */ |
| 60 |
__ASM_EMIT("shufps $0xdd, %%xmm7, %%xmm5") |
| 61 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0 r2 i0 i2 */ |
| 62 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") |
| 63 |
__ASM_EMIT("addps %%xmm1, %%xmm0") /* xmm0 = r0+r1 r2+r3 i0+i1 i2+i3 = r0k r2k i0k i2k */ |
| 64 |
__ASM_EMIT("addps %%xmm5, %%xmm4") |
| 65 |
__ASM_EMIT("subps %%xmm1, %%xmm2") /* xmm2 = r0-r1 r2-r3 i0-i1 i2-i3 = r1k r3k i1k i3k */ |
| 66 |
__ASM_EMIT("subps %%xmm5, %%xmm6") |
| 67 |
|
| 68 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0k r2k i0k i2k */ |
| 69 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") |
| 70 |
__ASM_EMIT("shufps $0x88, %%xmm2, %%xmm0") /* xmm0 = r0k i0k r1k i1k */ |
| 71 |
__ASM_EMIT("shufps $0x88, %%xmm6, %%xmm4") |
| 72 |
__ASM_EMIT("shufps $0x7d, %%xmm2, %%xmm1") /* xmm1 = r2k i2k i3k r3k */ |
| 73 |
__ASM_EMIT("shufps $0x7d, %%xmm6, %%xmm5") |
| 74 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0k i0k r1k i1k */ |
| 75 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") |
| 76 |
__ASM_EMIT("addps %%xmm1, %%xmm0") /* xmm0 = r0k+r2k i0k+i2k r1k+i3k i1k+r3k = d0 d4 d3 d5 */ |
| 77 |
__ASM_EMIT("addps %%xmm5, %%xmm4") |
| 78 |
__ASM_EMIT("subps %%xmm1, %%xmm2") /* xmm2 = r0k-r2k i0k-i2k r1k-i3k i1k-r3k = d2 d6 d1 d7 */ |
| 79 |
__ASM_EMIT("subps %%xmm5, %%xmm6") |
| 80 |
|
| 81 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = d0 d4 d3 d5 */ |
| 82 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") |
| 83 |
__ASM_EMIT("shufps $0x88, %%xmm2, %%xmm0") /* xmm0 = d0 d3 d2 d1 */ |
| 84 |
__ASM_EMIT("shufps $0x88, %%xmm6, %%xmm4") |
| 85 |
__ASM_EMIT("shufps $0xdd, %%xmm2, %%xmm1") /* xmm1 = d4 d5 d6 d7 */ |
| 86 |
__ASM_EMIT("shufps $0xdd, %%xmm6, %%xmm5") |
| 87 |
__ASM_EMIT("shufps $0x6c, %%xmm0, %%xmm0") /* xmm0 = d0 d1 d2 d3 */ |
| 88 |
__ASM_EMIT("shufps $0x6c, %%xmm4, %%xmm4") |
| 89 |
|
| 90 |
// Finally store result |
| 91 |
__ASM_EMIT("movups %%xmm0, 0x00(%[ptr])") |
| 92 |
__ASM_EMIT("movups %%xmm1, 0x10(%[ptr])") |
| 93 |
__ASM_EMIT("movups %%xmm4, 0x20(%[ptr])") |
| 94 |
__ASM_EMIT("movups %%xmm5, 0x30(%[ptr])") |
| 95 |
|
| 96 |
__ASM_EMIT("add $0x40, %[ptr]") |
| 97 |
__ASM_EMIT("sub $8, %[k]") |
| 98 |
__ASM_EMIT("jnz 1b") |
| 99 |
|
| 100 |
: [ptr] "+r" (ptr), [k] "+r" (items) |
| 101 |
: |
| 102 |
: "cc", "memory", |
| 103 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 104 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 105 |
); |
| 106 |
|
| 107 |
items = last << 1; |
| 108 |
size_t n = 8; |
| 109 |
size_t bs = n << 1; |
| 110 |
|
| 111 |
const float *wk = XFFT_W; |
| 112 |
const float *ak = XFFT_A; |
| 113 |
|
| 114 |
// Iterate butterflies |
| 115 |
while (n < last) |
| 116 |
{ |
| 117 |
for (size_t p=0; p<items; p += bs) |
| 118 |
{ |
| 119 |
// Set initial values of pointers |
| 120 |
float *a = &tmp[p]; |
| 121 |
float *b = &a[n]; |
| 122 |
size_t k = n; |
| 123 |
|
| 124 |
ARCH_X86_ASM |
| 125 |
( |
| 126 |
__ASM_EMIT("movaps 0x00(%[ak]), %%xmm6") /* xmm6 = rak[i] */ |
| 127 |
__ASM_EMIT("movaps 0x10(%[ak]), %%xmm7") /* xmm7 = iak[i] */ |
| 128 |
: |
| 129 |
: [ak] "r"(ak) |
| 130 |
: "%xmm6", "%xmm7" |
| 131 |
); |
| 132 |
|
| 133 |
ARCH_X86_ASM |
| 134 |
( |
| 135 |
// __ASM_EMIT(".align 16") |
| 136 |
__ASM_EMIT("1:") |
| 137 |
|
| 138 |
__ASM_EMIT("movups 0x00(%[a]), %%xmm0") /* xmm0 = ra[i] */ |
| 139 |
__ASM_EMIT("movups 0x10(%[a]), %%xmm1") /* xmm1 = ia[i] */ |
| 140 |
__ASM_EMIT("movups 0x00(%[b]), %%xmm2") /* xmm2 = rb[i] */ |
| 141 |
__ASM_EMIT("movups 0x10(%[b]), %%xmm3") /* xmm3 = ib[i] */ |
| 142 |
|
| 143 |
__ASM_EMIT("movaps %%xmm2, %%xmm4") /* xmm4 = rb[i] */ |
| 144 |
__ASM_EMIT("movaps %%xmm3, %%xmm5") /* xmm5 = ib[i] */ |
| 145 |
__ASM_EMIT("mulps %%xmm6, %%xmm2") /* xmm2 = rak[i]*rb[i] */ |
| 146 |
__ASM_EMIT("mulps %%xmm7, %%xmm4") /* xmm4 = iak[i]*rb[i] */ |
| 147 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = rak[i]*ib[i] */ |
| 148 |
__ASM_EMIT("mulps %%xmm7, %%xmm5") /* xmm5 = iak[i]*ib[i] */ |
| 149 |
__ASM_EMIT("addps %%xmm4, %%xmm3") /* xmm3 = rak[i]*ib[i] + iak[i]*rb[i] = ic[i] */ |
| 150 |
__ASM_EMIT("subps %%xmm5, %%xmm2") /* xmm2 = rak[i]*rb[i] - iak[i]*ib[i] = rc[i] */ |
| 151 |
__ASM_EMIT("movaps %%xmm0, %%xmm4") /* xmm4 = ra[i] */ |
| 152 |
__ASM_EMIT("movaps %%xmm1, %%xmm5") /* xmm5 = ia[i] */ |
| 153 |
__ASM_EMIT("subps %%xmm2, %%xmm0") /* xmm0 = ra[i] - rc[i] */ |
| 154 |
__ASM_EMIT("subps %%xmm3, %%xmm1") /* xmm1 = ia[i] - ic[i] */ |
| 155 |
__ASM_EMIT("addps %%xmm4, %%xmm2") /* xmm2 = ra[i] + rc[i] */ |
| 156 |
__ASM_EMIT("addps %%xmm5, %%xmm3") /* xmm3 = ia[i] + ic[i] */ |
| 157 |
|
| 158 |
__ASM_EMIT("movups %%xmm2, 0x00(%[a])") |
| 159 |
__ASM_EMIT("movups %%xmm3, 0x10(%[a])") |
| 160 |
__ASM_EMIT("movups %%xmm0, 0x00(%[b])") |
| 161 |
__ASM_EMIT("movups %%xmm1, 0x10(%[b])") |
| 162 |
|
| 163 |
__ASM_EMIT("add $0x20, %[a]") |
| 164 |
__ASM_EMIT("add $0x20, %[b]") |
| 165 |
__ASM_EMIT("sub $8, %[k]") |
| 166 |
__ASM_EMIT("jz 2f") |
| 167 |
|
| 168 |
/* Rotate angle */ |
| 169 |
__ASM_EMIT("movaps 0x00(%[wk]), %%xmm0") /* xmm0 = rw */ |
| 170 |
__ASM_EMIT("movaps 0x10(%[wk]), %%xmm1") /* xmm1 = iw */ |
| 171 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = rw */ |
| 172 |
__ASM_EMIT("movaps %%xmm1, %%xmm3") /* xmm3 = iw */ |
| 173 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = ra * iw */ |
| 174 |
__ASM_EMIT("mulps %%xmm7, %%xmm1") /* xmm1 = ia * iw */ |
| 175 |
__ASM_EMIT("mulps %%xmm0, %%xmm6") /* xmm6 = ra * rw */ |
| 176 |
__ASM_EMIT("mulps %%xmm2, %%xmm7") /* xmm7 = ia * rw */ |
| 177 |
__ASM_EMIT("subps %%xmm1, %%xmm6") /* xmm6 = ra * rw - ia * iw */ |
| 178 |
__ASM_EMIT("addps %%xmm3, %%xmm7") /* xmm7 = ra * iw + ia * rw */ |
| 179 |
__ASM_EMIT("jmp 1b") |
| 180 |
|
| 181 |
__ASM_EMIT("2:") |
| 182 |
|
| 183 |
: [a] "+r"(a), [b] "+r"(b), [k] "+r" (k) |
| 184 |
: [wk] "r"(wk) |
| 185 |
: "cc", "memory", |
| 186 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 187 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 188 |
); |
| 189 |
} |
| 190 |
|
| 191 |
wk += 8; |
| 192 |
ak += 8; |
| 193 |
n <<= 1; |
| 194 |
bs <<= 1; |
| 195 |
} |
| 196 |
|
| 197 |
if (n < items) |
| 198 |
{ |
| 199 |
// ONE LARGE CYCLE |
| 200 |
// Set initial values of pointers |
| 201 |
float kn = 1.0f / last; |
| 202 |
size_t k = n; |
| 203 |
|
| 204 |
ARCH_X86_ASM |
| 205 |
( |
| 206 |
__ASM_EMIT("movaps 0x00(%[ak]), %%xmm6") /* xmm6 = rak[i] */ |
| 207 |
__ASM_EMIT("movaps 0x10(%[ak]), %%xmm7") /* xmm7 = iak[i] */ |
| 208 |
__ASM_EMIT("shufps $0x00, %%xmm0, %%xmm0") /* xmm0 = kn */ |
| 209 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = kn */ |
| 210 |
: "+Yz" (kn) |
| 211 |
: [ak] "r" (ak) |
| 212 |
: "%xmm1", "%xmm6", "%xmm7" |
| 213 |
); |
| 214 |
|
| 215 |
ARCH_X86_ASM |
| 216 |
( |
| 217 |
// __ASM_EMIT(".align 16") |
| 218 |
__ASM_EMIT("1:") |
| 219 |
|
| 220 |
__ASM_EMIT("movups 0x00(%[tmp]), %%xmm4") /* xmm4 = ra[i] */ |
| 221 |
__ASM_EMIT("movups 0x00(%[tmp],%[n],4), %%xmm2") /* xmm2 = rb[i] */ |
| 222 |
__ASM_EMIT("movups 0x10(%[tmp],%[n],4), %%xmm3") /* xmm3 = ib[i] */ |
| 223 |
|
| 224 |
__ASM_EMIT("mulps %%xmm6, %%xmm2") /* xmm2 = rak[i]*rb[i] */ |
| 225 |
__ASM_EMIT("mulps %%xmm7, %%xmm3") /* xmm3 = iak[i]*ib[i] */ |
| 226 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") /* xmm5 = ra[i] */ |
| 227 |
__ASM_EMIT("subps %%xmm3, %%xmm2") /* xmm2 = rak[i]*rb[i] - iak[i]*ib[i] = rc[i] */ |
| 228 |
__ASM_EMIT("addps %%xmm2, %%xmm4") /* xmm4 = ra[i] + rc[i] */ |
| 229 |
__ASM_EMIT("subps %%xmm2, %%xmm5") /* xmm5 = ra[i] - rc[i] */ |
| 230 |
__ASM_EMIT("mulps %%xmm0, %%xmm4") /* xmm4 = kn*(ra[i] + rc[i]) */ |
| 231 |
__ASM_EMIT("mulps %%xmm1, %%xmm5") /* xmm5 = kn*(ra[i] - rc[i]) */ |
| 232 |
|
| 233 |
__ASM_EMIT("movups %%xmm4, 0x00(%[dst])") |
| 234 |
__ASM_EMIT("movups %%xmm5, 0x00(%[dst],%[n],2)") |
| 235 |
|
| 236 |
__ASM_EMIT("add $0x20, %[tmp]") |
| 237 |
__ASM_EMIT("add $0x10, %[dst]") |
| 238 |
__ASM_EMIT("sub $8, %[k]") |
| 239 |
__ASM_EMIT("jz 2f") |
| 240 |
|
| 241 |
/* Rotate angle */ |
| 242 |
__ASM_EMIT("movaps 0x00(%[wk]), %%xmm4") /* xmm4 = rw */ |
| 243 |
__ASM_EMIT("movaps 0x10(%[wk]), %%xmm5") /* xmm5 = iw */ |
| 244 |
__ASM_EMIT("movaps %%xmm4, %%xmm2") /* xmm2 = rw */ |
| 245 |
__ASM_EMIT("movaps %%xmm5, %%xmm3") /* xmm3 = iw */ |
| 246 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = ra * iw */ |
| 247 |
__ASM_EMIT("mulps %%xmm7, %%xmm5") /* xmm5 = ia * iw */ |
| 248 |
__ASM_EMIT("mulps %%xmm4, %%xmm6") /* xmm6 = ra * rw */ |
| 249 |
__ASM_EMIT("mulps %%xmm2, %%xmm7") /* xmm7 = ia * rw */ |
| 250 |
__ASM_EMIT("subps %%xmm5, %%xmm6") /* xmm6 = ra * rw - ia * iw */ |
| 251 |
__ASM_EMIT("addps %%xmm3, %%xmm7") /* xmm7 = ra * iw + ia * rw */ |
| 252 |
__ASM_EMIT("jmp 1b") |
| 253 |
|
| 254 |
__ASM_EMIT("2:") |
| 255 |
|
| 256 |
: [tmp] "+r"(tmp), [dst] "+r"(dst), [n] "+r"(n), [k] "+r" (k) |
| 257 |
: [wk] "r"(wk) |
| 258 |
: "cc", "memory", |
| 259 |
"%xmm0", "%xmm5", "%xmm2", "%xmm3", |
| 260 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 261 |
); |
| 262 |
} |
| 263 |
else |
| 264 |
{ |
| 265 |
// Add real result to the target (ignore complex result) |
| 266 |
float kn = 1.0f / last; |
| 267 |
|
| 268 |
// Unpack 4x split complex to 4x real |
| 269 |
ARCH_X86_ASM |
| 270 |
( |
| 271 |
__ASM_EMIT("shufps $0x00, %%xmm0, %%xmm0") /* xmm0 = kn */ |
| 272 |
__ASM_EMIT("movups 0x00(%[tmp]), %%xmm1") /* xmm0 = s[i] */ |
| 273 |
__ASM_EMIT("movups 0x00(%[dst]), %%xmm2") |
| 274 |
__ASM_EMIT("mulps %%xmm0, %%xmm1") |
| 275 |
__ASM_EMIT("addps %%xmm1, %%xmm2") |
| 276 |
__ASM_EMIT("movups %%xmm2, 0x00(%[dst])") |
| 277 |
: "+Yz" (kn) |
| 278 |
: [tmp] "r" (tmp), [dst] "r" (dst) |
| 279 |
: "cc", "memory", |
| 280 |
"%xmm1", "%xmm2" |
| 281 |
); |
| 282 |
} |
| 283 |
|
| 284 |
} |
| 285 |
} |
| 286 |
} |
| 287 |
|
| 288 |
#endif /* PRIVATE_DSP_ARCH_X86_SSE_FASTCONV_RESTORE_H_ */ |