| 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_IMPL |
| 23 |
#error "This header should not be included directly" |
| 24 |
#endif /* PRIVATE_DSP_ARCH_X86_SSE_IMPL */ |
| 25 |
|
| 26 |
namespace lsp |
| 27 |
{ |
| 28 |
namespace sse |
| 29 |
{ |
| 30 |
// This is the SSE implementation of the scrambling functions for self data |
| 31 |
static inline void FFT_SCRAMBLE_SELF_DIRECT_NAME(float *dst, const float *src, size_t rank) |
| 32 |
{ |
| 33 |
// Calculate number of items |
| 34 |
size_t items = (1 << rank) - 1; |
| 35 |
|
| 36 |
for (size_t i = 1; i < items; ++i) |
| 37 |
{ |
| 38 |
size_t j = reverse_bits(FFT_TYPE(i), rank); /* Reverse the order of the bits */ |
| 39 |
if (i >= j) |
| 40 |
continue; |
| 41 |
|
| 42 |
/* Copy the values from the reversed position */ |
| 43 |
ARCH_X86_ASM |
| 44 |
( |
| 45 |
__ASM_EMIT("movlps (%[dst], %[i], 8), %%xmm0") |
| 46 |
__ASM_EMIT("movlps (%[dst], %[j], 8), %%xmm1") |
| 47 |
__ASM_EMIT("movlps %%xmm0, (%[dst], %[j], 8)") |
| 48 |
__ASM_EMIT("movlps %%xmm1, (%[dst], %[i], 8)") |
| 49 |
: |
| 50 |
: [dst] "r"(dst), [i] "r"(i), [j] "r"(j) |
| 51 |
: "memory", |
| 52 |
"%xmm0", "%xmm1" |
| 53 |
); |
| 54 |
} |
| 55 |
|
| 56 |
// Perform butterfly 8x |
| 57 |
items = 1 << (rank - 3); |
| 58 |
|
| 59 |
// Perform 4-element butterflies |
| 60 |
ARCH_X86_ASM |
| 61 |
( |
| 62 |
/* Loop */ |
| 63 |
__ASM_EMIT(".align 16") |
| 64 |
__ASM_EMIT("1:") |
| 65 |
|
| 66 |
/* Load data to registers */ |
| 67 |
__ASM_EMIT("movups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 i0 r1 i1 */ |
| 68 |
__ASM_EMIT("movups 0x10(%[dst]), %%xmm2") /* xmm2 = r2 i2 r3 i3 */ |
| 69 |
__ASM_EMIT("movups 0x20(%[dst]), %%xmm4") /* xmm4 = r4 i4 r5 i5 */ |
| 70 |
__ASM_EMIT("movups 0x30(%[dst]), %%xmm6") /* xmm6 = r6 i6 r7 i7 */ |
| 71 |
|
| 72 |
/* 1st-order 4x butterfly */ |
| 73 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0 i0 r1 i1 */ |
| 74 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") /* xmm5 = r4 i4 r5 i5 */ |
| 75 |
__ASM_EMIT("shufps $0x44, %%xmm2, %%xmm0") /* xmm0 = r0 i0 r2 i2 */ |
| 76 |
__ASM_EMIT("shufps $0x44, %%xmm6, %%xmm4") /* xmm4 = r4 i4 r6 i6 */ |
| 77 |
__ASM_EMIT("shufps $0xee, %%xmm2, %%xmm1") /* xmm1 = r1 i1 r3 i3 */ |
| 78 |
__ASM_EMIT("shufps $0xee, %%xmm6, %%xmm5") /* xmm5 = r5 i5 r7 i7 */ |
| 79 |
|
| 80 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0 i0 r2 i2 */ |
| 81 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") /* xmm6 = r4 i4 r6 i6 */ |
| 82 |
__ASM_EMIT("addps %%xmm1, %%xmm0") /* xmm0 = r0+r1 i0+i1 r2+r3 i2+i3 = r0' i0' r2' i2' */ |
| 83 |
__ASM_EMIT("addps %%xmm5, %%xmm4") /* xmm4 = r4+r5 i4+i5 r6+r7 i6+i7 = r4' i4' r6' i6' */ |
| 84 |
__ASM_EMIT("subps %%xmm1, %%xmm2") /* xmm2 = r0-r1 i0-i1 r2-r3 i2-i3 = r1' i1' r3' i3' */ |
| 85 |
__ASM_EMIT("subps %%xmm5, %%xmm6") /* xmm6 = r4-r5 i4-i5 r6-r7 i6-i7 = r5' i5' r7' i7' */ |
| 86 |
|
| 87 |
/* 2nd-order 4x butterfly */ |
| 88 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0' i0' r2' i2' */ |
| 89 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") /* xmm5 = r4' i4' r6' i6' */ |
| 90 |
__ASM_EMIT("shufps $0x44, %%xmm2, %%xmm0") /* xmm0 = r0' i0' r1' i1' */ |
| 91 |
__ASM_EMIT("shufps $0x44, %%xmm6, %%xmm4") /* xmm4 = r4' i4' r5' i5' */ |
| 92 |
__ASM_EMIT("shufps $0xbe, %%xmm2, %%xmm1") /* xmm1 = r2' i2' i3' r3' */ |
| 93 |
__ASM_EMIT("shufps $0xbe, %%xmm6, %%xmm5") /* xmm5 = r6' i6' i7' r7' */ |
| 94 |
|
| 95 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0' i0' r1' i1' */ |
| 96 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") /* xmm6 = r4' i4' r5' i5' */ |
| 97 |
__ASM_EMIT("addps %%xmm1, %%xmm0") /* xmm0 = r0'+r2' i0'+i2' r1'+i3' i1'+r3' = r0" i0" r1" i3" */ |
| 98 |
__ASM_EMIT("addps %%xmm5, %%xmm4") /* xmm4 = r4'+r6' i4'+i6' r5'+i7' i5'+r7' = r4" i4" r5" i7" */ |
| 99 |
__ASM_EMIT("subps %%xmm1, %%xmm2") /* xmm2 = r0'-r2' i0'-i2' r1'-i3' i1'-r3' = r2" i2" r3" i1" */ |
| 100 |
__ASM_EMIT("subps %%xmm5, %%xmm6") /* xmm6 = r4'-r6' i4'-i6' r5'-i7' i5'-r7' = r6" i6" r7" i5" */ |
| 101 |
|
| 102 |
/* Reorder and store */ |
| 103 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0" i0" r1" i3" */ |
| 104 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") /* xmm5 = r4" i4" r5" i7" */ |
| 105 |
__ASM_EMIT("shufps $0x88, %%xmm2, %%xmm0") /* xmm0 = r0" r1" r2" r3" */ |
| 106 |
__ASM_EMIT("shufps $0x88, %%xmm6, %%xmm4") /* xmm4 = r4" r5" r6" r7" */ |
| 107 |
__ASM_EMIT("shufps $0xdd, %%xmm2, %%xmm1") /* xmm1 = i0" i3" i2" i1" */ |
| 108 |
__ASM_EMIT("shufps $0xdd, %%xmm6, %%xmm5") /* xmm5 = i4" i7" i6" i5" */ |
| 109 |
__ASM_EMIT("shufps $0x6c, %%xmm1, %%xmm1") /* xmm1 = i0" i1" i2" i3" */ |
| 110 |
__ASM_EMIT("shufps $0x6c, %%xmm5, %%xmm5") /* xmm5 = i4" i5" i6" i7" */ |
| 111 |
|
| 112 |
__ASM_EMIT("movups %%xmm0, 0x00(%[dst])") |
| 113 |
__ASM_EMIT("movups %%xmm1, 0x10(%[dst])") |
| 114 |
__ASM_EMIT("movups %%xmm4, 0x20(%[dst])") |
| 115 |
__ASM_EMIT("movups %%xmm5, 0x30(%[dst])") |
| 116 |
|
| 117 |
/* Move pointers and repeat cycle */ |
| 118 |
__ASM_EMIT("add $0x40, %[dst]") |
| 119 |
__ASM_EMIT("dec %[items]") |
| 120 |
__ASM_EMIT("jnz 1b") |
| 121 |
|
| 122 |
: [dst] "+r"(dst), [items] "+r"(items) |
| 123 |
: |
| 124 |
: "cc", "memory", |
| 125 |
"%xmm0", "%xmm1", "%xmm2", "%xmm4", "%xmm5", "%xmm6" |
| 126 |
); |
| 127 |
} |
| 128 |
|
| 129 |
static inline void FFT_SCRAMBLE_SELF_REVERSE_NAME(float *dst, const float *src, size_t rank) |
| 130 |
{ |
| 131 |
// Calculate number of items |
| 132 |
size_t items = (1 << rank) - 1; |
| 133 |
|
| 134 |
for (size_t i = 1; i < items; ++i) |
| 135 |
{ |
| 136 |
size_t j = reverse_bits(FFT_TYPE(i), rank); /* Reverse the order of the bits */ |
| 137 |
if (i >= j) |
| 138 |
continue; |
| 139 |
|
| 140 |
/* Copy the values from the reversed position */ |
| 141 |
ARCH_X86_ASM |
| 142 |
( |
| 143 |
__ASM_EMIT("movlps (%[dst], %[i], 8), %%xmm0") |
| 144 |
__ASM_EMIT("movlps (%[dst], %[j], 8), %%xmm1") |
| 145 |
__ASM_EMIT("movlps %%xmm0, (%[dst], %[j], 8)") |
| 146 |
__ASM_EMIT("movlps %%xmm1, (%[dst], %[i], 8)") |
| 147 |
: |
| 148 |
: [dst] "r"(dst), [i] "r"(i), [j] "r"(j) |
| 149 |
: "memory", |
| 150 |
"%xmm0", "%xmm1" |
| 151 |
); |
| 152 |
} |
| 153 |
|
| 154 |
// Perform butterfly 8x |
| 155 |
items = 1 << (rank - 3); |
| 156 |
|
| 157 |
// Perform 4-element butterflies |
| 158 |
ARCH_X86_ASM |
| 159 |
( |
| 160 |
/* Loop */ |
| 161 |
__ASM_EMIT(".align 16") |
| 162 |
__ASM_EMIT("1:") |
| 163 |
|
| 164 |
/* Load data to registers */ |
| 165 |
__ASM_EMIT("movups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 i0 r1 i1 */ |
| 166 |
__ASM_EMIT("movups 0x10(%[dst]), %%xmm2") /* xmm2 = r2 i2 r3 i3 */ |
| 167 |
__ASM_EMIT("movups 0x20(%[dst]), %%xmm4") /* xmm4 = r4 i4 r5 i5 */ |
| 168 |
__ASM_EMIT("movups 0x30(%[dst]), %%xmm6") /* xmm6 = r6 i6 r7 i7 */ |
| 169 |
|
| 170 |
/* 1st-order 4x butterfly */ |
| 171 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0 i0 r1 i1 */ |
| 172 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") /* xmm5 = r4 i4 r5 i5 */ |
| 173 |
__ASM_EMIT("shufps $0x44, %%xmm2, %%xmm0") /* xmm0 = r0 i0 r2 i2 */ |
| 174 |
__ASM_EMIT("shufps $0x44, %%xmm6, %%xmm4") /* xmm4 = r4 i4 r6 i6 */ |
| 175 |
__ASM_EMIT("shufps $0xee, %%xmm2, %%xmm1") /* xmm1 = r1 i1 r3 i3 */ |
| 176 |
__ASM_EMIT("shufps $0xee, %%xmm6, %%xmm5") /* xmm5 = r5 i5 r7 i7 */ |
| 177 |
|
| 178 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0 i0 r2 i2 */ |
| 179 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") /* xmm6 = r4 i4 r6 i6 */ |
| 180 |
__ASM_EMIT("addps %%xmm1, %%xmm0") /* xmm0 = r0+r1 i0+i1 r2+r3 i2+i3 = r0' i0' r2' i2' */ |
| 181 |
__ASM_EMIT("addps %%xmm5, %%xmm4") /* xmm4 = r4+r5 i4+i5 r6+r7 i6+i7 = r4' i4' r6' i6' */ |
| 182 |
__ASM_EMIT("subps %%xmm1, %%xmm2") /* xmm2 = r0-r1 i0-i1 r2-r3 i2-i3 = r1' i1' r3' i3' */ |
| 183 |
__ASM_EMIT("subps %%xmm5, %%xmm6") /* xmm6 = r4-r5 i4-i5 r6-r7 i6-i7 = r5' i5' r7' i7' */ |
| 184 |
|
| 185 |
/* 2nd-order 4x butterfly */ |
| 186 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0' i0' r2' i2' */ |
| 187 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") /* xmm5 = r4' i4' r6' i6' */ |
| 188 |
__ASM_EMIT("shufps $0x44, %%xmm2, %%xmm0") /* xmm0 = r0' i0' r1' i1' */ |
| 189 |
__ASM_EMIT("shufps $0xbe, %%xmm2, %%xmm1") /* xmm1 = r2' i2' i3' r3' */ |
| 190 |
__ASM_EMIT("shufps $0x44, %%xmm6, %%xmm4") /* xmm4 = r4' i4' r5' i5' */ |
| 191 |
__ASM_EMIT("shufps $0xbe, %%xmm6, %%xmm5") /* xmm5 = r6' i6' i7' r7' */ |
| 192 |
|
| 193 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0' i0' r1' i1' */ |
| 194 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") /* xmm6 = r4' i4' r5' i5' */ |
| 195 |
__ASM_EMIT("addps %%xmm1, %%xmm0") /* xmm0 = r0'+r2' i0'+i2' r1'+i3' i1'+r3' = r0" i0" r3" i1" */ |
| 196 |
__ASM_EMIT("addps %%xmm5, %%xmm4") /* xmm4 = r4'+r6' i4'+i6' r5'+i7' i5'+r7' = r4" i4" r7" i5" */ |
| 197 |
__ASM_EMIT("subps %%xmm1, %%xmm2") /* xmm2 = r0'-r2' i0'-i2' r1'-i3' i1'-r3' = r2" i2" r1" i3" */ |
| 198 |
__ASM_EMIT("subps %%xmm5, %%xmm6") /* xmm6 = r4'-r6' i4'-i6' r5'-i7' i5'-r7' = r6" i6" r5" i7" */ |
| 199 |
|
| 200 |
/* Reorder and store */ |
| 201 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0" i0" r3" i1" */ |
| 202 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") /* xmm5 = r4" i4" r7" i5" */ |
| 203 |
__ASM_EMIT("shufps $0x88, %%xmm2, %%xmm0") /* xmm0 = r0" r3" r2" r1" */ |
| 204 |
__ASM_EMIT("shufps $0x88, %%xmm6, %%xmm4") /* xmm4 = r4" r7" r6" r5" */ |
| 205 |
__ASM_EMIT("shufps $0xdd, %%xmm2, %%xmm1") /* xmm1 = i0" i1" i2" i3" */ |
| 206 |
__ASM_EMIT("shufps $0xdd, %%xmm6, %%xmm5") /* xmm5 = i4" i5" i6" i7" */ |
| 207 |
__ASM_EMIT("shufps $0x6c, %%xmm0, %%xmm0") /* xmm0 = r0" r1" r2" r3" */ |
| 208 |
__ASM_EMIT("shufps $0x6c, %%xmm4, %%xmm4") /* xmm4 = r4" r5" r6" r7" */ |
| 209 |
|
| 210 |
__ASM_EMIT("movups %%xmm0, 0x00(%[dst])") |
| 211 |
__ASM_EMIT("movups %%xmm1, 0x10(%[dst])") |
| 212 |
__ASM_EMIT("movups %%xmm4, 0x20(%[dst])") |
| 213 |
__ASM_EMIT("movups %%xmm5, 0x30(%[dst])") |
| 214 |
|
| 215 |
/* Move pointers and repeat cycle */ |
| 216 |
__ASM_EMIT("add $0x40, %[dst]") |
| 217 |
__ASM_EMIT("dec %[items]") |
| 218 |
__ASM_EMIT("jnz 1b") |
| 219 |
|
| 220 |
: [dst] "+r"(dst), [items] "+r"(items) |
| 221 |
: |
| 222 |
: "cc", "memory", |
| 223 |
"%xmm0", "%xmm1", "%xmm2", "%xmm4", "%xmm5", "%xmm6" |
| 224 |
); |
| 225 |
} |
| 226 |
|
| 227 |
static inline void FFT_SCRAMBLE_COPY_DIRECT_NAME(float *dst, const float *src, size_t rank) |
| 228 |
{ |
| 229 |
size_t regs = 1 << rank; |
| 230 |
|
| 231 |
for (size_t i=0; i<regs; ++i) |
| 232 |
{ |
| 233 |
size_t index = reverse_bits(FFT_TYPE(i), rank); |
| 234 |
|
| 235 |
// Perform 4-element butterflies |
| 236 |
ARCH_X86_ASM |
| 237 |
( |
| 238 |
/* Load data to registers */ |
| 239 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm0") /* xmm0 = r0 i0 x x */ |
| 240 |
__ASM_EMIT("add %[regs], %[index]") |
| 241 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm4") /* xmm4 = r4 i4 x x */ |
| 242 |
__ASM_EMIT("add %[regs], %[index]") |
| 243 |
|
| 244 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm2") /* xmm2 = r2 i2 x x */ |
| 245 |
__ASM_EMIT("add %[regs], %[index]") |
| 246 |
__ASM_EMIT("movlhps %%xmm2, %%xmm0") /* xmm0 = r0 i0 r2 i2 */ |
| 247 |
|
| 248 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm6") /* xmm6 = r6 i6 x x */ |
| 249 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0 i0 r2 i2 */ |
| 250 |
__ASM_EMIT("add %[regs], %[index]") |
| 251 |
__ASM_EMIT("movlhps %%xmm6, %%xmm4") /* xmm4 = r4 i4 r6 i6 */ |
| 252 |
|
| 253 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm1") /* xmm0 = r1 i1 x x */ |
| 254 |
__ASM_EMIT("add %[regs], %[index]") |
| 255 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") /* xmm6 = r4 i4 r6 i6 */ |
| 256 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm5") /* xmm4 = r5 i5 x x */ |
| 257 |
__ASM_EMIT("add %[regs], %[index]") |
| 258 |
|
| 259 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm3") /* xmm1 = r3 i3 x x */ |
| 260 |
__ASM_EMIT("add %[regs], %[index]") |
| 261 |
__ASM_EMIT("movlhps %%xmm3, %%xmm1") /* xmm1 = r1 i1 r3 i3 */ |
| 262 |
|
| 263 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm7") /* xmm5 = r7 i7 x x */ |
| 264 |
__ASM_EMIT("addps %%xmm1, %%xmm0") /* xmm0 = r0+r1 i0+i1 r2+r3 i2+i3 = r0' i0' r2' i2' */ |
| 265 |
__ASM_EMIT("add %[regs], %[index]") |
| 266 |
__ASM_EMIT("movlhps %%xmm7, %%xmm5") /* xmm5 = r5 i5 r7 i7 */ |
| 267 |
|
| 268 |
/* 1st-order 4x butterfly */ |
| 269 |
/* xmm0 = r0 i0 r2 i2 */ |
| 270 |
/* xmm4 = r4 i4 r6 i6 */ |
| 271 |
/* xmm1 = r1 i1 r3 i3 */ |
| 272 |
/* xmm5 = r5 i5 r7 i7 */ |
| 273 |
__ASM_EMIT("subps %%xmm1, %%xmm2") /* xmm2 = r0-r1 i0-i1 r2-r3 i2-i3 = r1' i1' r3' i3' */ |
| 274 |
__ASM_EMIT("addps %%xmm5, %%xmm4") /* xmm4 = r4+r5 i4+i5 r6+r7 i6+i7 = r4' i4' r6' i6' */ |
| 275 |
__ASM_EMIT("subps %%xmm5, %%xmm6") /* xmm6 = r4-r5 i4-i5 r6-r7 i6-i7 = r5' i5' r7' i7' */ |
| 276 |
|
| 277 |
/* 2nd-order 4x butterfly */ |
| 278 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0' i0' r2' i2' */ |
| 279 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") /* xmm5 = r4' i4' r6' i6' */ |
| 280 |
__ASM_EMIT("shufps $0x44, %%xmm2, %%xmm0") /* xmm0 = r0' i0' r1' i1' */ |
| 281 |
__ASM_EMIT("shufps $0xbe, %%xmm2, %%xmm1") /* xmm1 = r2' i2' i3' r3' */ |
| 282 |
__ASM_EMIT("shufps $0x44, %%xmm6, %%xmm4") /* xmm4 = r4' i4' r5' i5' */ |
| 283 |
__ASM_EMIT("shufps $0xbe, %%xmm6, %%xmm5") /* xmm5 = r6' i6' i7' r7' */ |
| 284 |
|
| 285 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0' i0' r1' i1' */ |
| 286 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") /* xmm6 = r4' i4' r5' i5' */ |
| 287 |
__ASM_EMIT("addps %%xmm1, %%xmm0") /* xmm0 = r0'+r2' i0'+i2' r1'+i3' i1'+r3' = r0" i0" r1" i3" */ |
| 288 |
__ASM_EMIT("addps %%xmm5, %%xmm4") /* xmm4 = r4'+r6' i4'+i6' r5'+i7' i5'+r7' = r4" i4" r5" i7" */ |
| 289 |
__ASM_EMIT("subps %%xmm1, %%xmm2") /* xmm2 = r0'-r2' i0'-i2' r1'-i3' i1'-r3' = r2" i2" r3" i1" */ |
| 290 |
__ASM_EMIT("subps %%xmm5, %%xmm6") /* xmm6 = r4'-r6' i4'-i6' r5'-i7' i5'-r7' = r6" i6" r7" i5" */ |
| 291 |
|
| 292 |
/* Reorder and store */ |
| 293 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0" i0" r1" i3" */ |
| 294 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") /* xmm5 = r4" i4" r5" i7" */ |
| 295 |
__ASM_EMIT("shufps $0x88, %%xmm2, %%xmm0") /* xmm0 = r0" r1" r2" r3" */ |
| 296 |
__ASM_EMIT("shufps $0x88, %%xmm6, %%xmm4") /* xmm4 = r4" r5" r6" r7" */ |
| 297 |
__ASM_EMIT("shufps $0xdd, %%xmm2, %%xmm1") /* xmm1 = i0" i3" i2" i1" */ |
| 298 |
__ASM_EMIT("shufps $0xdd, %%xmm6, %%xmm5") /* xmm5 = i4" i7" i6" i5" */ |
| 299 |
__ASM_EMIT("shufps $0x6c, %%xmm1, %%xmm1") /* xmm1 = i0" i1" i2" i3" */ |
| 300 |
__ASM_EMIT("shufps $0x6c, %%xmm5, %%xmm5") /* xmm5 = i4" i5" i6" i7" */ |
| 301 |
|
| 302 |
__ASM_EMIT("movups %%xmm0, 0x00(%[dst])") |
| 303 |
__ASM_EMIT("movups %%xmm1, 0x10(%[dst])") |
| 304 |
__ASM_EMIT("movups %%xmm4, 0x20(%[dst])") |
| 305 |
__ASM_EMIT("movups %%xmm5, 0x30(%[dst])") |
| 306 |
|
| 307 |
/* Move pointers and repeat cycle */ |
| 308 |
__ASM_EMIT("add $0x40, %[dst]") |
| 309 |
|
| 310 |
: [dst] "+r"(dst), [src] "+r"(src), [index] "+r"(index) |
| 311 |
: [regs] "r" (regs) |
| 312 |
: "cc", "memory", |
| 313 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 314 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 315 |
); |
| 316 |
} |
| 317 |
} |
| 318 |
|
| 319 |
static inline void FFT_SCRAMBLE_COPY_REVERSE_NAME(float *dst, const float *src, size_t rank) |
| 320 |
{ |
| 321 |
size_t regs = 1 << rank; |
| 322 |
|
| 323 |
for (size_t i=0; i<regs; ++i) |
| 324 |
{ |
| 325 |
size_t index = reverse_bits(FFT_TYPE(i), rank); |
| 326 |
|
| 327 |
// Perform 4-element butterflies |
| 328 |
ARCH_X86_ASM |
| 329 |
( |
| 330 |
/* Load data to registers */ |
| 331 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm0") /* xmm0 = r0 i0 x x */ |
| 332 |
__ASM_EMIT("add %[regs], %[index]") |
| 333 |
|
| 334 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm4") /* xmm4 = r4 i4 x x */ |
| 335 |
__ASM_EMIT("add %[regs], %[index]") |
| 336 |
|
| 337 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm2") /* xmm2 = r2 i2 x x */ |
| 338 |
__ASM_EMIT("add %[regs], %[index]") |
| 339 |
__ASM_EMIT("movlhps %%xmm2, %%xmm0") /* xmm0 = r0 i0 r2 i2 */ |
| 340 |
|
| 341 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm6") /* xmm6 = r6 i6 x x */ |
| 342 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0 i0 r2 i2 */ |
| 343 |
__ASM_EMIT("add %[regs], %[index]") |
| 344 |
__ASM_EMIT("movlhps %%xmm6, %%xmm4") /* xmm4 = r4 i4 r6 i6 */ |
| 345 |
|
| 346 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm1") /* xmm1 = r1 i1 x x */ |
| 347 |
__ASM_EMIT("add %[regs], %[index]") |
| 348 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") /* xmm6 = r4 i4 r6 i6 */ |
| 349 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm5") /* xmm5 = r5 i5 x x */ |
| 350 |
__ASM_EMIT("add %[regs], %[index]") |
| 351 |
|
| 352 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm3") /* xmm3 = r3 i3 x x */ |
| 353 |
__ASM_EMIT("add %[regs], %[index]") |
| 354 |
__ASM_EMIT("movlhps %%xmm3, %%xmm1") /* xmm1 = r1 i1 r3 i3 */ |
| 355 |
|
| 356 |
__ASM_EMIT("movlps (%[src], %[index], 8), %%xmm7") /* xmm7 = r7 i7 x x */ |
| 357 |
__ASM_EMIT("addps %%xmm1, %%xmm0") /* xmm0 = r0+r1 i0+i1 r2+r3 i2+i3 = r0' i0' r2' i2' */ |
| 358 |
__ASM_EMIT("add %[regs], %[index]") |
| 359 |
__ASM_EMIT("movlhps %%xmm7, %%xmm5") /* xmm5 = r5 i5 r7 i7 */ |
| 360 |
|
| 361 |
/* 1st-order 4x butterfly */ |
| 362 |
/* xmm0 = r0 i0 r2 i2 */ |
| 363 |
/* xmm4 = r4 i4 r6 i6 */ |
| 364 |
/* xmm1 = r1 i1 r3 i3 */ |
| 365 |
/* xmm5 = r5 i5 r7 i7 */ |
| 366 |
__ASM_EMIT("subps %%xmm1, %%xmm2") /* xmm2 = r0-r1 i0-i1 r2-r3 i2-i3 = r1' i1' r3' i3' */ |
| 367 |
__ASM_EMIT("addps %%xmm5, %%xmm4") /* xmm4 = r4+r5 i4+i5 r6+r7 i6+i7 = r4' i4' r6' i6' */ |
| 368 |
__ASM_EMIT("subps %%xmm5, %%xmm6") /* xmm6 = r4-r5 i4-i5 r6-r7 i6-i7 = r5' i5' r7' i7' */ |
| 369 |
|
| 370 |
/* 2nd-order 4x butterfly */ |
| 371 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0' i0' r2' i2' */ |
| 372 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") /* xmm5 = r4' i4' r6' i6' */ |
| 373 |
__ASM_EMIT("shufps $0x44, %%xmm2, %%xmm0") /* xmm0 = r0' i0' r1' i1' */ |
| 374 |
__ASM_EMIT("shufps $0xbe, %%xmm2, %%xmm1") /* xmm1 = r2' i2' i3' r3' */ |
| 375 |
__ASM_EMIT("shufps $0x44, %%xmm6, %%xmm4") /* xmm4 = r4' i4' r5' i5' */ |
| 376 |
__ASM_EMIT("shufps $0xbe, %%xmm6, %%xmm5") /* xmm5 = r6' i6' i7' r7' */ |
| 377 |
|
| 378 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0' i0' r1' i1' */ |
| 379 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") /* xmm6 = r4' i4' r5' i5' */ |
| 380 |
__ASM_EMIT("addps %%xmm1, %%xmm0") /* xmm0 = r0'+r2' i0'+i2' r1'+i3' i1'+r3' = r0" i0" r3" i1" */ |
| 381 |
__ASM_EMIT("addps %%xmm5, %%xmm4") /* xmm4 = r4'+r6' i4'+i6' r5'+i7' i5'+r7' = r4" i4" r7" i5" */ |
| 382 |
__ASM_EMIT("subps %%xmm1, %%xmm2") /* xmm2 = r0'-r2' i0'-i2' r1'-i3' i1'-r3' = r2" i2" r1" i3" */ |
| 383 |
__ASM_EMIT("subps %%xmm5, %%xmm6") /* xmm6 = r4'-r6' i4'-i6' r5'-i7' i5'-r7' = r6" i6" r5" i7" */ |
| 384 |
|
| 385 |
/* Reorder and store */ |
| 386 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = r0" i0" r3" i1" */ |
| 387 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") /* xmm5 = r4" i4" r7" i5" */ |
| 388 |
__ASM_EMIT("shufps $0x88, %%xmm2, %%xmm0") /* xmm0 = r0" r3" r2" r1" */ |
| 389 |
__ASM_EMIT("shufps $0x88, %%xmm6, %%xmm4") /* xmm4 = r4" r7" r6" r5" */ |
| 390 |
__ASM_EMIT("shufps $0xdd, %%xmm2, %%xmm1") /* xmm1 = i0" i1" i2" i3" */ |
| 391 |
__ASM_EMIT("shufps $0xdd, %%xmm6, %%xmm5") /* xmm5 = i4" i5" i6" i7" */ |
| 392 |
__ASM_EMIT("shufps $0x6c, %%xmm0, %%xmm0") /* xmm0 = r0" r1" r2" r3" */ |
| 393 |
__ASM_EMIT("shufps $0x6c, %%xmm4, %%xmm4") /* xmm4 = r4" r5" r6" r7" */ |
| 394 |
|
| 395 |
__ASM_EMIT("movups %%xmm0, 0x00(%[dst])") |
| 396 |
__ASM_EMIT("movups %%xmm1, 0x10(%[dst])") |
| 397 |
__ASM_EMIT("movups %%xmm4, 0x20(%[dst])") |
| 398 |
__ASM_EMIT("movups %%xmm5, 0x30(%[dst])") |
| 399 |
|
| 400 |
/* Move pointers and repeat cycle */ |
| 401 |
__ASM_EMIT("add $0x40, %[dst]") |
| 402 |
|
| 403 |
: [dst] "+r"(dst), [src] "+r"(src), [index] "+r"(index) |
| 404 |
: [regs] "r" (regs) |
| 405 |
: "cc", "memory", |
| 406 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 407 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 408 |
); |
| 409 |
} |
| 410 |
} |
| 411 |
} |
| 412 |
} |
| 413 |
|
| 414 |
#undef FFT_SCRAMBLE_SELF_DIRECT_NAME |
| 415 |
#undef FFT_SCRAMBLE_COPY_DIRECT_NAME |
| 416 |
#undef FFT_SCRAMBLE_SELF_REVERSE_NAME |
| 417 |
#undef FFT_SCRAMBLE_COPY_REVERSE_NAME |
| 418 |
#undef FFT_REPACK |
| 419 |
#undef FFT_TYPE |
| 420 |
|