| 1 |
/* |
| 2 |
* Copyright (C) 2026 Linux Studio Plugins Project <https://lsp-plug.in/> |
| 3 |
* (C) 2026 Vladimir Sadovnikov <[email protected]> |
| 4 |
* |
| 5 |
* This file is part of lsp-dsp-lib |
| 6 |
* Created on: 18 мар. 2026 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_AVX512_FFT_P_BUTTERFLY_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_X86_AVX512_FFT_P_BUTTERFLY_H_ |
| 24 |
|
| 25 |
namespace lsp |
| 26 |
{ |
| 27 |
namespace avx512 |
| 28 |
{ |
| 29 |
#define FFT_BUTTERFLY_BODY16(fmadd_b, fmadd_a) \ |
| 30 |
ARCH_X86_ASM( \ |
| 31 |
/* Prepare angle */ \ |
| 32 |
__ASM_EMIT("vmovaps 0x00(%[fft_a]), %%zmm6") /* zmm6 = x_re */ \ |
| 33 |
__ASM_EMIT("vmovaps 0x40(%[fft_a]), %%zmm7") /* zmm7 = x_im */ \ |
| 34 |
/* Start loop */ \ |
| 35 |
__ASM_EMIT("1:") \ |
| 36 |
__ASM_EMIT("vmovups 0x00(%[dst], %[off1]), %%zmm0") /* zmm0 = a_re */ \ |
| 37 |
__ASM_EMIT("vmovups 0x40(%[dst], %[off1]), %%zmm1") /* zmm1 = a_im */ \ |
| 38 |
__ASM_EMIT("vmovups 0x00(%[dst], %[off2]), %%zmm2") /* zmm2 = b_re */ \ |
| 39 |
__ASM_EMIT("vmovups 0x40(%[dst], %[off2]), %%zmm3") /* zmm3 = b_im */ \ |
| 40 |
/* Calculate complex multiplication */ \ |
| 41 |
__ASM_EMIT("vmulps %%zmm7, %%zmm2, %%zmm4") /* zmm4 = x_im * b_re */ \ |
| 42 |
__ASM_EMIT("vmulps %%zmm7, %%zmm3, %%zmm5") /* zmm5 = x_im * b_im */ \ |
| 43 |
__ASM_EMIT(fmadd_b " %%zmm6, %%zmm2, %%zmm5") /* zmm5 = c_re = x_re * b_re +- x_im * b_im */ \ |
| 44 |
__ASM_EMIT(fmadd_a " %%zmm6, %%zmm3, %%zmm4") /* zmm4 = c_im = x_re * b_im -+ x_im * b_re */ \ |
| 45 |
/* Perform butterfly */ \ |
| 46 |
__ASM_EMIT("vsubps %%zmm5, %%zmm0, %%zmm2") /* zmm2 = a_re - c_re */ \ |
| 47 |
__ASM_EMIT("vsubps %%zmm4, %%zmm1, %%zmm3") /* zmm3 = a_im - c_im */ \ |
| 48 |
__ASM_EMIT("vaddps %%zmm5, %%zmm0, %%zmm0") /* zmm0 = a_re + c_re */ \ |
| 49 |
__ASM_EMIT("vaddps %%zmm4, %%zmm1, %%zmm1") /* zmm1 = a_im + c_im */ \ |
| 50 |
/* Store values */ \ |
| 51 |
__ASM_EMIT("vmovups %%zmm0, 0x00(%[dst], %[off1])") \ |
| 52 |
__ASM_EMIT("vmovups %%zmm1, 0x40(%[dst], %[off1])") \ |
| 53 |
__ASM_EMIT("vmovups %%zmm2, 0x00(%[dst], %[off2])") \ |
| 54 |
__ASM_EMIT("vmovups %%zmm3, 0x40(%[dst], %[off2])") \ |
| 55 |
__ASM_EMIT("add $0x80, %[off1]") \ |
| 56 |
__ASM_EMIT("add $0x80, %[off2]") \ |
| 57 |
__ASM_EMIT32("subl $16, %[np]") \ |
| 58 |
__ASM_EMIT64("subq $16, %[np]") \ |
| 59 |
__ASM_EMIT("jz 2f") \ |
| 60 |
/* Rotate angle */ \ |
| 61 |
__ASM_EMIT("vmovaps 0x00(%[fft_w]), %%zmm4") /* zmm4 = w_re */ \ |
| 62 |
__ASM_EMIT("vmovaps 0x40(%[fft_w]), %%zmm5") /* zmm5 = w_im */ \ |
| 63 |
__ASM_EMIT("vmulps %%zmm5, %%zmm6, %%zmm2") /* zmm2 = w_im * x_re */ \ |
| 64 |
__ASM_EMIT("vmulps %%zmm5, %%zmm7, %%zmm3") /* zmm3 = w_im * x_im */ \ |
| 65 |
__ASM_EMIT("vfmsub132ps %%zmm4, %%zmm3, %%zmm6") /* zmm6 = x_re' = w_re * x_re - w_im * x_im */ \ |
| 66 |
__ASM_EMIT("vfmadd132ps %%zmm4, %%zmm2, %%zmm7") /* zmm7 = x_im' = w_re * x_im + w_im * x_re */ \ |
| 67 |
/* Repeat loop */ \ |
| 68 |
__ASM_EMIT("jmp 1b") \ |
| 69 |
__ASM_EMIT("2:") \ |
| 70 |
\ |
| 71 |
: [off1] "+r" (off1), [off2] "+r" (off2), [np] __ASM_ARG_RW(np) \ |
| 72 |
: [dst] "r" (dst), [fft_a] "r" (fft_a), [fft_w] "r" (fft_w) \ |
| 73 |
: "cc", "memory", \ |
| 74 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", \ |
| 75 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" \ |
| 76 |
); |
| 77 |
|
| 78 |
static inline void packed_butterfly_direct16p(float *dst, size_t rank, size_t blocks) |
| 79 |
{ |
| 80 |
size_t pairs = 1 << rank; |
| 81 |
size_t off1 = 0, shift = 8 << rank; |
| 82 |
const float *fft_a = &FFT_A[(rank - 2) << 5]; |
| 83 |
const float *fft_w = &FFT_DW[(rank - 3) << 5]; |
| 84 |
|
| 85 |
for (size_t b=0; b<blocks; ++b) |
| 86 |
{ |
| 87 |
size_t off2 = off1 + shift; |
| 88 |
size_t np = pairs; |
| 89 |
|
| 90 |
FFT_BUTTERFLY_BODY16("vfmadd231ps", "vfmsub231ps"); |
| 91 |
|
| 92 |
off1 = off2; |
| 93 |
} |
| 94 |
} |
| 95 |
|
| 96 |
static inline void packed_butterfly_reverse16p(float *dst, size_t rank, size_t blocks) |
| 97 |
{ |
| 98 |
size_t pairs = 1 << rank; |
| 99 |
size_t off1 = 0, shift = 8 << rank; |
| 100 |
const float *fft_a = &FFT_A[(rank - 2) << 5]; |
| 101 |
const float *fft_w = &FFT_DW[(rank - 3) << 5]; |
| 102 |
|
| 103 |
for (size_t b=0; b<blocks; ++b) |
| 104 |
{ |
| 105 |
size_t off2 = off1 + shift; |
| 106 |
size_t np = pairs; |
| 107 |
|
| 108 |
FFT_BUTTERFLY_BODY16("vfmsub231ps", "vfmadd231ps"); |
| 109 |
|
| 110 |
off1 = off2; |
| 111 |
} |
| 112 |
} |
| 113 |
|
| 114 |
#undef FFT_BUTTERFLY_BODY16 |
| 115 |
} /* namespace avx512 */ |
| 116 |
} /* namespace lsp */ |
| 117 |
|
| 118 |
#endif /* PRIVATE_DSP_ARCH_X86_AVX512_FFT_P_BUTTERFLY_H_ */ |