| 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_P_BUTTERFLY_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_X86_SSE_FFT_P_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 |
namespace lsp |
| 30 |
{ |
| 31 |
namespace sse |
| 32 |
{ |
| 33 |
#define FFT_BUTTERFLY_BODY(add_b, add_a) \ |
| 34 |
/* Init pointers */ \ |
| 35 |
float *a = &dst[blk*pairs*2]; \ |
| 36 |
float *b = &a[pairs]; \ |
| 37 |
size_t p = pairs; \ |
| 38 |
__IF_32(size_t tmp1); \ |
| 39 |
\ |
| 40 |
ARCH_X86_ASM \ |
| 41 |
( \ |
| 42 |
/* Prepare angle */ \ |
| 43 |
__ASM_EMIT32("mov %[a], %[tmp1]") \ |
| 44 |
__ASM_EMIT32("mov %[XFFT_A], %[a]") \ |
| 45 |
__ASM_EMIT32("movaps 0x00(%[a], %[rank]), %%xmm6") /* xmm6 = angle_re[0..3] */ \ |
| 46 |
__ASM_EMIT32("movaps 0x10(%[a], %[rank]), %%xmm7") /* xmm7 = angle_im[0..3] */ \ |
| 47 |
__ASM_EMIT32("mov %[tmp1], %[a]") \ |
| 48 |
__ASM_EMIT64("movaps 0x00(%[XFFT_A], %[rank]), %%xmm6") /* xmm6 = angle_re[0..3] */ \ |
| 49 |
__ASM_EMIT64("movaps 0x10(%[XFFT_A], %[rank]), %%xmm7") /* xmm7 = angle_im[0..3] */ \ |
| 50 |
/* Start loop */ \ |
| 51 |
__ASM_EMIT(".align 16") \ |
| 52 |
__ASM_EMIT("1:") \ |
| 53 |
\ |
| 54 |
/* Load complex values */ \ |
| 55 |
/* predicate: xmm6 = w_re[0..3] */ \ |
| 56 |
/* predicate: xmm7 = w_im[0..3] */ \ |
| 57 |
__ASM_EMIT("movups 0x00(%[a]), %%xmm0") /* xmm0 = a_re[0..3] */ \ |
| 58 |
__ASM_EMIT("movups 0x10(%[a]), %%xmm1") /* xmm1 = a_im[0..3] */ \ |
| 59 |
__ASM_EMIT("movups 0x00(%[b]), %%xmm2") /* xmm2 = b_re[0..3] */ \ |
| 60 |
__ASM_EMIT("movups 0x10(%[b]), %%xmm3") /* xmm3 = b_im[0..3] */ \ |
| 61 |
\ |
| 62 |
/* Calculate complex multiplication */ \ |
| 63 |
__ASM_EMIT("movaps %%xmm2, %%xmm4") /* xmm4 = b_re[0..3] */ \ |
| 64 |
__ASM_EMIT("movaps %%xmm3, %%xmm5") /* xmm5 = b_im[0..3] */ \ |
| 65 |
__ASM_EMIT("mulps %%xmm6, %%xmm2") /* xmm2 = w_re[0..3] * b_re[0..3] */ \ |
| 66 |
__ASM_EMIT("mulps %%xmm7, %%xmm4") /* xmm4 = w_im[0..3] * b_re[0..3] */ \ |
| 67 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = w_re[0..3] * b_im[0..3] */ \ |
| 68 |
__ASM_EMIT("mulps %%xmm7, %%xmm5") /* xmm5 = w_im[0..3] * b_im[0..3] */ \ |
| 69 |
__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] */ \ |
| 70 |
__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] */ \ |
| 71 |
\ |
| 72 |
/* Perform butterfly */ \ |
| 73 |
__ASM_EMIT("movaps %%xmm0, %%xmm4") /* xmm4 = a_re[0..3] */ \ |
| 74 |
__ASM_EMIT("movaps %%xmm1, %%xmm5") /* xmm5 = a_im[0..3] */ \ |
| 75 |
__ASM_EMIT("subps %%xmm2, %%xmm0") /* xmm0 = a_re[0..3] - c_re[0..3] */ \ |
| 76 |
__ASM_EMIT("subps %%xmm3, %%xmm1") /* xmm1 = a_im[0..3] - c_im[0..3] */ \ |
| 77 |
__ASM_EMIT("addps %%xmm4, %%xmm2") /* xmm2 = a_re[0..3] + c_re[0..3] */ \ |
| 78 |
__ASM_EMIT("addps %%xmm5, %%xmm3") /* xmm3 = a_im[0..3] + c_im[0..3] */ \ |
| 79 |
\ |
| 80 |
/* Store values */ \ |
| 81 |
__ASM_EMIT("movups %%xmm2, 0x00(%[a])") \ |
| 82 |
__ASM_EMIT("movups %%xmm3, 0x10(%[a])") \ |
| 83 |
__ASM_EMIT("movups %%xmm0, 0x00(%[b])") \ |
| 84 |
__ASM_EMIT("movups %%xmm1, 0x10(%[b])") \ |
| 85 |
\ |
| 86 |
/* Update pointers */ \ |
| 87 |
__ASM_EMIT("add $0x20, %[a]") \ |
| 88 |
__ASM_EMIT("add $0x20, %[b]") \ |
| 89 |
\ |
| 90 |
/* Repeat loop */ \ |
| 91 |
__ASM_EMIT("sub $8, %[p]") \ |
| 92 |
__ASM_EMIT("jz 2f") \ |
| 93 |
\ |
| 94 |
/* Rotate angle */ \ |
| 95 |
__ASM_EMIT("movaps 0x00(%[XFFT_W], %[rank]), %%xmm0") /* xmm0 = w_re[0..3] */ \ |
| 96 |
__ASM_EMIT("movaps 0x10(%[XFFT_W], %[rank]), %%xmm1") /* xmm1 = w_im[0..3] */ \ |
| 97 |
\ |
| 98 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") /* xmm2 = w_re[0..3] */ \ |
| 99 |
__ASM_EMIT("movaps %%xmm1, %%xmm3") /* xmm3 = w_im[0..3] */ \ |
| 100 |
\ |
| 101 |
__ASM_EMIT("mulps %%xmm6, %%xmm3") /* xmm3 = a_re[0..3] * w_im[0..3] */ \ |
| 102 |
__ASM_EMIT("mulps %%xmm7, %%xmm1") /* xmm1 = a_im[0..3] * w_im[0..3] */ \ |
| 103 |
__ASM_EMIT("mulps %%xmm0, %%xmm6") /* xmm6 = a_re[0..3] * w_re[0..3] */ \ |
| 104 |
__ASM_EMIT("mulps %%xmm2, %%xmm7") /* xmm7 = a_im[0..3] * w_re[0..3] */ \ |
| 105 |
__ASM_EMIT("subps %%xmm1, %%xmm6") /* xmm6 = a_re[0..3] * w_re[0..3] + a_im[0..3] * w_im[0..3] */ \ |
| 106 |
__ASM_EMIT("addps %%xmm3, %%xmm7") /* xmm7 = a_im[0..3] * w_re[0..3] - a_re[0..3] * w_im[0..3] */ \ |
| 107 |
\ |
| 108 |
/* Repeat loop */ \ |
| 109 |
__ASM_EMIT("jmp 1b") \ |
| 110 |
__ASM_EMIT("2:") \ |
| 111 |
\ |
| 112 |
: [a] "+r" (a), [b] "+r"(b), [p] "+r" (p) \ |
| 113 |
: [rank] "r" (rank), [XFFT_W] "r"(XFFT_W), \ |
| 114 |
__IF_64([XFFT_A] "r" (XFFT_A)) \ |
| 115 |
__IF_32([XFFT_A] "g" (&XFFT_A[0]), [tmp1] "g" (&tmp1)) \ |
| 116 |
: "cc", "memory", \ |
| 117 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", \ |
| 118 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" \ |
| 119 |
); |
| 120 |
|
| 121 |
static inline void packed_butterfly_direct(float *dst, size_t rank, size_t blocks) |
| 122 |
{ |
| 123 |
size_t pairs = 1 << (rank + 1); |
| 124 |
rank = (rank - 2) << 5; |
| 125 |
|
| 126 |
for (size_t blk=0; blk<blocks; ++blk) |
| 127 |
{ |
| 128 |
FFT_BUTTERFLY_BODY("addps", "subps"); |
| 129 |
} |
| 130 |
} |
| 131 |
|
| 132 |
static inline void packed_butterfly_reverse(float *dst, size_t rank, size_t blocks) |
| 133 |
{ |
| 134 |
size_t pairs = 1 << (rank + 1); |
| 135 |
rank = (rank - 2) << 5; |
| 136 |
|
| 137 |
for (size_t blk=0; blk<blocks; ++blk) |
| 138 |
{ |
| 139 |
FFT_BUTTERFLY_BODY("subps", "addps"); |
| 140 |
} |
| 141 |
} |
| 142 |
|
| 143 |
#undef FFT_BUTTERFLY_BODY |
| 144 |
} /* namespace sse */ |
| 145 |
} /* namespace lsp */ |
| 146 |
|
| 147 |
#endif /* PRIVATE_DSP_ARCH_X86_SSE_FFT_P_BUTTERFLY_H_ */ |