| 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 |
#define _SSE_FFT_NAME(id) id |
| 27 |
#define _SSE_CALL_NAME(id, bits) id ## bits |
| 28 |
|
| 29 |
#define SSE_FFT_NAME(id) _SSE_FFT_NAME(id) |
| 30 |
#define SSE_CALL_NAME(id, bits) _SSE_CALL_NAME(id, bits)(dst, src, rank) |
| 31 |
|
| 32 |
namespace lsp |
| 33 |
{ |
| 34 |
namespace sse |
| 35 |
{ |
| 36 |
static inline void SSE_FFT_NAME(FFT_SCRAMBLE_DIRECT_NAME)(float *dst, const float *src, size_t rank) |
| 37 |
{ |
| 38 |
// Scramble the order of samples |
| 39 |
if (dst == src) |
| 40 |
{ |
| 41 |
if (rank <= 8) |
| 42 |
SSE_CALL_NAME(FFT_SCRAMBLE_SELF_DIRECT_NAME, 8); |
| 43 |
else //if (rank <= 16) |
| 44 |
SSE_CALL_NAME(FFT_SCRAMBLE_SELF_DIRECT_NAME, 16); |
| 45 |
} |
| 46 |
else |
| 47 |
{ |
| 48 |
rank -= 3; |
| 49 |
|
| 50 |
if (rank <= 8) |
| 51 |
SSE_CALL_NAME(FFT_SCRAMBLE_COPY_DIRECT_NAME, 8); |
| 52 |
else //if (rank <= 16) |
| 53 |
SSE_CALL_NAME(FFT_SCRAMBLE_COPY_DIRECT_NAME, 16); |
| 54 |
} |
| 55 |
} |
| 56 |
|
| 57 |
static inline void SSE_FFT_NAME(FFT_SCRAMBLE_REVERSE_NAME)(float *dst, const float *src, size_t rank) |
| 58 |
{ |
| 59 |
// Scramble the order of samples |
| 60 |
if (dst == src) |
| 61 |
{ |
| 62 |
if (rank <= 8) |
| 63 |
SSE_CALL_NAME(FFT_SCRAMBLE_SELF_REVERSE_NAME, 8); |
| 64 |
else //if (rank <= 16) |
| 65 |
SSE_CALL_NAME(FFT_SCRAMBLE_SELF_REVERSE_NAME, 16); |
| 66 |
} |
| 67 |
else |
| 68 |
{ |
| 69 |
rank -= 3; |
| 70 |
|
| 71 |
if (rank <= 8) |
| 72 |
SSE_CALL_NAME(FFT_SCRAMBLE_COPY_REVERSE_NAME, 8); |
| 73 |
else //if (rank <= 16) |
| 74 |
SSE_CALL_NAME(FFT_SCRAMBLE_COPY_REVERSE_NAME, 16); |
| 75 |
} |
| 76 |
} |
| 77 |
|
| 78 |
static inline void SSE_FFT_NAME(FFT_REPACK)(float *dst, size_t rank) |
| 79 |
{ |
| 80 |
size_t blocks = 1 << (rank-3); |
| 81 |
|
| 82 |
// Perform 4-element butterflies |
| 83 |
ARCH_X86_ASM |
| 84 |
( |
| 85 |
__ASM_EMIT(".align 16") |
| 86 |
__ASM_EMIT("1:") |
| 87 |
|
| 88 |
/* Load data */ |
| 89 |
__ASM_EMIT("movups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 r1 r2 r3 */ |
| 90 |
__ASM_EMIT("movups 0x10(%[dst]), %%xmm1") /* xmm1 = i0 i1 i2 i3 */ |
| 91 |
__ASM_EMIT("movups 0x20(%[dst]), %%xmm4") /* xmm4 = r4 r5 r6 r7 */ |
| 92 |
__ASM_EMIT("movups 0x30(%[dst]), %%xmm5") /* xmm5 = i4 i5 i6 i7 */ |
| 93 |
|
| 94 |
/* 1st step: repack pairs */ |
| 95 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = r0 r1 r2 r3 */ |
| 96 |
__ASM_EMIT("movaps %%xmm4, %%xmm6") /* xmm6 = r4 r5 r6 r7 */ |
| 97 |
__ASM_EMIT("unpcklps %%xmm1, %%xmm0") /* xmm0 = r0 i0 r1 i1 */ |
| 98 |
__ASM_EMIT("unpcklps %%xmm5, %%xmm4") /* xmm4 = r4 i4 r5 i5 */ |
| 99 |
__ASM_EMIT("unpckhps %%xmm1, %%xmm2") /* xmm2 = r2 i2 r3 i3 */ |
| 100 |
__ASM_EMIT("unpckhps %%xmm5, %%xmm6") /* xmm6 = r6 i6 r7 i7 */ |
| 101 |
|
| 102 |
/* Store data */ |
| 103 |
__ASM_EMIT("movups %%xmm0, 0x00(%[dst])") |
| 104 |
__ASM_EMIT("movups %%xmm2, 0x10(%[dst])") |
| 105 |
__ASM_EMIT("movups %%xmm4, 0x20(%[dst])") |
| 106 |
__ASM_EMIT("movups %%xmm6, 0x30(%[dst])") |
| 107 |
|
| 108 |
/* Move pointers and repeat cycle */ |
| 109 |
__ASM_EMIT("add $0x40, %[dst]") |
| 110 |
__ASM_EMIT("dec %[blocks]") |
| 111 |
__ASM_EMIT("jnz 1b") |
| 112 |
|
| 113 |
: [dst] "+r"(dst), [blocks] "+r"(blocks) |
| 114 |
: |
| 115 |
: "cc", "memory", |
| 116 |
"%xmm0", "%xmm1", "%xmm2", |
| 117 |
"%xmm4", "%xmm5", "%xmm6" |
| 118 |
); |
| 119 |
} |
| 120 |
|
| 121 |
static inline void SSE_FFT_NAME(FFT_REPACK_NORMALIZE)(float *dst, size_t rank) |
| 122 |
{ |
| 123 |
size_t blocks = 1 << (rank-3); |
| 124 |
float k = 0.125f/blocks; |
| 125 |
|
| 126 |
// Perform 4-element butterflies |
| 127 |
ARCH_X86_ASM |
| 128 |
( |
| 129 |
__ASM_EMIT("shufps $0x00, %%xmm0, %%xmm0") /* xmm0 = k k k k */ |
| 130 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") /* xmm1 = k k k k */ |
| 131 |
__ASM_EMIT(".align 16") |
| 132 |
__ASM_EMIT("1:") |
| 133 |
|
| 134 |
/* Load data */ |
| 135 |
__ASM_EMIT("movups 0x00(%[dst]), %%xmm2") /* xmm2 = r0 r1 r2 r3 */ |
| 136 |
__ASM_EMIT("movups 0x10(%[dst]), %%xmm3") /* xmm3 = i0 i1 i2 i3 */ |
| 137 |
__ASM_EMIT("movups 0x20(%[dst]), %%xmm5") /* xmm5 = r4 r5 r6 r7 */ |
| 138 |
__ASM_EMIT("movups 0x30(%[dst]), %%xmm6") /* xmm6 = i4 i5 i6 i7 */ |
| 139 |
|
| 140 |
/* 1st step: repack pairs */ |
| 141 |
__ASM_EMIT("movaps %%xmm2, %%xmm4") /* xmm4 = r0 r1 r2 r3 */ |
| 142 |
__ASM_EMIT("movaps %%xmm5, %%xmm7") /* xmm7 = r4 r5 r6 r7 */ |
| 143 |
__ASM_EMIT("unpcklps %%xmm3, %%xmm2") /* xmm2 = r0 i0 r1 i1 */ |
| 144 |
__ASM_EMIT("unpcklps %%xmm6, %%xmm5") /* xmm5 = r4 i4 r5 i5 */ |
| 145 |
__ASM_EMIT("unpckhps %%xmm3, %%xmm4") /* xmm4 = r2 i2 r3 i3 */ |
| 146 |
__ASM_EMIT("unpckhps %%xmm6, %%xmm7") /* xmm7 = r6 i6 r7 i7 */ |
| 147 |
|
| 148 |
/* 3rd step: apply normalization */ |
| 149 |
__ASM_EMIT("mulps %%xmm0, %%xmm2") |
| 150 |
__ASM_EMIT("mulps %%xmm1, %%xmm4") |
| 151 |
__ASM_EMIT("mulps %%xmm0, %%xmm5") |
| 152 |
__ASM_EMIT("mulps %%xmm1, %%xmm7") |
| 153 |
|
| 154 |
/* Store data */ |
| 155 |
__ASM_EMIT("movups %%xmm2, 0x00(%[dst])") |
| 156 |
__ASM_EMIT("movups %%xmm4, 0x10(%[dst])") |
| 157 |
__ASM_EMIT("movups %%xmm5, 0x20(%[dst])") |
| 158 |
__ASM_EMIT("movups %%xmm7, 0x30(%[dst])") |
| 159 |
|
| 160 |
/* Move pointers and repeat cycle */ |
| 161 |
__ASM_EMIT("add $0x40, %[dst]") |
| 162 |
__ASM_EMIT("dec %[blocks]") |
| 163 |
__ASM_EMIT("jnz 1b") |
| 164 |
|
| 165 |
: [dst] "+r"(dst), [blocks] "+r"(blocks), "+Yz"(k) |
| 166 |
: |
| 167 |
: "cc", "memory", |
| 168 |
"%xmm1", "%xmm2", "%xmm3", |
| 169 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 170 |
); |
| 171 |
} |
| 172 |
} |
| 173 |
} |
| 174 |
|
| 175 |
#undef SSE_FFT_NAME |
| 176 |
#undef SSE_CALL_NAME |
| 177 |
#undef _SSE_FFT_NAME |
| 178 |
#undef _SSE_CALL_NAME |
| 179 |
|
| 180 |
#undef FFT_SCRAMBLE_SELF_DIRECT_NAME |
| 181 |
#undef FFT_SCRAMBLE_COPY_DIRECT_NAME |
| 182 |
#undef FFT_SCRAMBLE_SELF_REVERSE_NAME |
| 183 |
#undef FFT_SCRAMBLE_COPY_REVERSE_NAME |
| 184 |
#undef LS_RE |
| 185 |
|
| 186 |
#undef FFT_SCRAMBLE_DIRECT_NAME |
| 187 |
#undef FFT_SCRAMBLE_REVERSE_NAME |
| 188 |
#undef FFT_REPACK |
| 189 |
#undef FFT_REPACK_NORMALIZED |
| 190 |
#undef FFT_MODE |
| 191 |
|