| 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_AVX_FFT_P_BUTTERFLY_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_X86_AVX_FFT_P_BUTTERFLY_H_ |
| 24 |
|
| 25 |
#ifndef PRIVATE_DSP_ARCH_X86_AVX_IMPL |
| 26 |
#error "This header should not be included directly" |
| 27 |
#endif /* PRIVATE_DSP_ARCH_X86_AVX_IMPL */ |
| 28 |
|
| 29 |
namespace lsp |
| 30 |
{ |
| 31 |
namespace avx |
| 32 |
{ |
| 33 |
#define FFT_BUTTERFLY_BODY8(add_b, add_a, FMA_SEL) \ |
| 34 |
ARCH_X86_ASM \ |
| 35 |
( \ |
| 36 |
/* Prepare angle */ \ |
| 37 |
__ASM_EMIT("vmovaps 0x00(%[fft_a]), %%ymm6") /* ymm6 = x_re */ \ |
| 38 |
__ASM_EMIT("vmovaps 0x20(%[fft_a]), %%ymm7") /* ymm7 = x_im */ \ |
| 39 |
/* Start loop */ \ |
| 40 |
__ASM_EMIT("1:") \ |
| 41 |
__ASM_EMIT("vmovups 0x00(%[dst], %[off1]), %%ymm0") /* ymm0 = a_re */ \ |
| 42 |
__ASM_EMIT("vmovups 0x20(%[dst], %[off1]), %%ymm1") /* ymm1 = a_im */ \ |
| 43 |
__ASM_EMIT("vmovups 0x00(%[dst], %[off2]), %%ymm2") /* ymm2 = b_re */ \ |
| 44 |
__ASM_EMIT("vmovups 0x20(%[dst], %[off2]), %%ymm3") /* ymm3 = b_im */ \ |
| 45 |
/* Calculate complex multiplication */ \ |
| 46 |
__ASM_EMIT("vmulps %%ymm7, %%ymm2, %%ymm4") /* ymm4 = x_im * b_re */ \ |
| 47 |
__ASM_EMIT("vmulps %%ymm7, %%ymm3, %%ymm5") /* ymm5 = x_im * b_im */ \ |
| 48 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm2, %%ymm2", "")) /* ymm2 = x_re * b_re */ \ |
| 49 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm3, %%ymm3", "")) /* ymm3 = x_re * b_im */ \ |
| 50 |
__ASM_EMIT(FMA_SEL(add_b " %%ymm5, %%ymm2, %%ymm5", add_b " %%ymm6, %%ymm2, %%ymm5")) /* ymm5 = c_re = x_re * b_re +- x_im * b_im */ \ |
| 51 |
__ASM_EMIT(FMA_SEL(add_a " %%ymm4, %%ymm3, %%ymm4", add_a " %%ymm6, %%ymm3, %%ymm4")) /* ymm4 = c_im = x_re * b_im -+ x_im * b_re */ \ |
| 52 |
/* Perform butterfly */ \ |
| 53 |
__ASM_EMIT("vsubps %%ymm5, %%ymm0, %%ymm2") /* ymm2 = a_re - c_re */ \ |
| 54 |
__ASM_EMIT("vsubps %%ymm4, %%ymm1, %%ymm3") /* ymm3 = a_im - c_im */ \ |
| 55 |
__ASM_EMIT("vaddps %%ymm5, %%ymm0, %%ymm0") /* ymm0 = a_re + c_re */ \ |
| 56 |
__ASM_EMIT("vaddps %%ymm4, %%ymm1, %%ymm1") /* ymm1 = a_im + c_im */ \ |
| 57 |
/* Store values */ \ |
| 58 |
__ASM_EMIT("vmovups %%ymm0, 0x00(%[dst], %[off1])") \ |
| 59 |
__ASM_EMIT("vmovups %%ymm1, 0x20(%[dst], %[off1])") \ |
| 60 |
__ASM_EMIT("vmovups %%ymm2, 0x00(%[dst], %[off2])") \ |
| 61 |
__ASM_EMIT("vmovups %%ymm3, 0x20(%[dst], %[off2])") \ |
| 62 |
__ASM_EMIT("add $0x40, %[off1]") \ |
| 63 |
__ASM_EMIT("add $0x40, %[off2]") \ |
| 64 |
__ASM_EMIT32("subl $8, %[np]") \ |
| 65 |
__ASM_EMIT64("subq $8, %[np]") \ |
| 66 |
__ASM_EMIT("jz 2f") \ |
| 67 |
/* Rotate angle */ \ |
| 68 |
__ASM_EMIT("vmovaps 0x00(%[fft_w]), %%ymm4") /* xmm4 = w_re */ \ |
| 69 |
__ASM_EMIT("vmovaps 0x20(%[fft_w]), %%ymm5") /* xmm5 = w_im */ \ |
| 70 |
__ASM_EMIT("vmulps %%ymm5, %%ymm6, %%ymm2") /* ymm2 = w_im * x_re */ \ |
| 71 |
__ASM_EMIT("vmulps %%ymm5, %%ymm7, %%ymm3") /* ymm3 = w_im * x_im */ \ |
| 72 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm6, %%ymm6", "")) /* ymm6 = w_re * x_re */ \ |
| 73 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm7, %%ymm7", "")) /* ymm7 = w_re * x_im */ \ |
| 74 |
__ASM_EMIT(FMA_SEL("vsubps %%ymm3, %%ymm6, %%ymm6", "vfmsub132ps %%ymm4, %%ymm3, %%ymm6")) /* ymm6 = x_re' = w_re * x_re - w_im * x_im */ \ |
| 75 |
__ASM_EMIT(FMA_SEL("vaddps %%ymm2, %%ymm7, %%ymm7", "vfmadd132ps %%ymm4, %%ymm2, %%ymm7")) /* ymm7 = x_im' = w_re * x_im + w_im * x_re */ \ |
| 76 |
/* Repeat loop */ \ |
| 77 |
__ASM_EMIT("jmp 1b") \ |
| 78 |
__ASM_EMIT("2:") \ |
| 79 |
\ |
| 80 |
: [off1] "+r" (off1), [off2] "+r" (off2), [np] __ASM_ARG_RW(np) \ |
| 81 |
: [dst] "r" (dst), [fft_a] "r" (fft_a), [fft_w] "r" (fft_w) \ |
| 82 |
: "cc", "memory", \ |
| 83 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", \ |
| 84 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" \ |
| 85 |
); |
| 86 |
|
| 87 |
#define FMA_OFF(a, b) a |
| 88 |
#define FMA_ON(a, b) b |
| 89 |
|
| 90 |
static inline void packed_butterfly_direct8p(float *dst, size_t rank, size_t blocks) |
| 91 |
{ |
| 92 |
size_t pairs = 1 << rank; |
| 93 |
size_t off1 = 0, shift = 8 << rank; //1 << (rank + 3); |
| 94 |
const float *fft_a = &FFT_A[(rank - 2) << 4]; |
| 95 |
const float *fft_w = &FFT_DW[(rank - 2) << 4]; |
| 96 |
|
| 97 |
for (size_t b=0; b<blocks; ++b) |
| 98 |
{ |
| 99 |
size_t off2 = off1 + shift; |
| 100 |
size_t np = pairs; |
| 101 |
|
| 102 |
FFT_BUTTERFLY_BODY8("vaddps", "vsubps", FMA_OFF); |
| 103 |
|
| 104 |
off1 = off2; |
| 105 |
} |
| 106 |
} |
| 107 |
|
| 108 |
static inline void packed_butterfly_reverse8p(float *dst, size_t rank, size_t blocks) |
| 109 |
{ |
| 110 |
size_t pairs = 1 << rank; |
| 111 |
size_t off1 = 0, shift = 8 << rank; // 1 << (rank + 3); |
| 112 |
const float *fft_a = &FFT_A[(rank - 2) << 4]; |
| 113 |
const float *fft_w = &FFT_DW[(rank - 2) << 4]; |
| 114 |
|
| 115 |
for (size_t b=0; b<blocks; ++b) |
| 116 |
{ |
| 117 |
size_t off2 = off1 + shift; |
| 118 |
size_t np = pairs; |
| 119 |
|
| 120 |
FFT_BUTTERFLY_BODY8("vsubps", "vaddps", FMA_OFF); |
| 121 |
|
| 122 |
off1 = off2; |
| 123 |
} |
| 124 |
} |
| 125 |
|
| 126 |
static inline void packed_butterfly_direct8p_fma3(float *dst, size_t rank, size_t blocks) |
| 127 |
{ |
| 128 |
size_t pairs = 1 << rank; |
| 129 |
size_t off1 = 0, shift = 8 << rank; // 1 << (rank + 3); |
| 130 |
const float *fft_a = &FFT_A[(rank - 2) << 4]; |
| 131 |
const float *fft_w = &FFT_DW[(rank - 2) << 4]; |
| 132 |
|
| 133 |
for (size_t b=0; b<blocks; ++b) |
| 134 |
{ |
| 135 |
size_t off2 = off1 + shift; |
| 136 |
size_t np = pairs; |
| 137 |
|
| 138 |
FFT_BUTTERFLY_BODY8("vfmadd231ps", "vfmsub231ps", FMA_ON); |
| 139 |
|
| 140 |
off1 = off2; |
| 141 |
} |
| 142 |
} |
| 143 |
|
| 144 |
static inline void packed_butterfly_reverse8p_fma3(float *dst, size_t rank, size_t blocks) |
| 145 |
{ |
| 146 |
size_t pairs = 1 << rank; |
| 147 |
size_t off1 = 0, shift = 8 << rank; // 1 << (rank + 3); |
| 148 |
const float *fft_a = &FFT_A[(rank - 2) << 4]; |
| 149 |
const float *fft_w = &FFT_DW[(rank - 2) << 4]; |
| 150 |
|
| 151 |
for (size_t b=0; b<blocks; ++b) |
| 152 |
{ |
| 153 |
size_t off2 = off1 + shift; |
| 154 |
size_t np = pairs; |
| 155 |
|
| 156 |
FFT_BUTTERFLY_BODY8("vfmsub231ps", "vfmadd231ps", FMA_ON); |
| 157 |
|
| 158 |
off1 = off2; |
| 159 |
} |
| 160 |
} |
| 161 |
|
| 162 |
#undef FMA_OFF |
| 163 |
#undef FMA_ON |
| 164 |
#undef FFT_BUTTERFLY_BODY8 |
| 165 |
} |
| 166 |
} |
| 167 |
|
| 168 |
#endif /* PRIVATE_DSP_ARCH_X86_AVX_FFT_P_BUTTERFLY_H_ */ |