| 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_REPACK_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_X86_AVX512_FFT_P_REPACK_H_ |
| 24 |
|
| 25 |
#ifndef PRIVATE_DSP_ARCH_X86_AVX512_IMPL |
| 26 |
#error "This header should not be included directly" |
| 27 |
#endif /* PRIVATE_DSP_ARCH_X86_AVX512_IMPL */ |
| 28 |
|
| 29 |
namespace lsp |
| 30 |
{ |
| 31 |
namespace avx512 |
| 32 |
{ |
| 33 |
static inline void packed_fft_repack(float *dst, size_t rank) |
| 34 |
{ |
| 35 |
size_t blocks = 1 << rank; |
| 36 |
|
| 37 |
ARCH_X86_ASM |
| 38 |
( |
| 39 |
__ASM_EMIT("kmovw 0x00 + %[MASK], %%k4") |
| 40 |
__ASM_EMIT("kmovw 0x02 + %[MASK], %%k5") |
| 41 |
// 32x blocks |
| 42 |
__ASM_EMIT("sub $32, %[blocks]") |
| 43 |
__ASM_EMIT("jb 2f") |
| 44 |
__ASM_EMIT("1:") |
| 45 |
__ASM_EMIT("vexpandps 0x000(%[dst]), %%zmm0 %{%%k4%}%{z%}") /* zmm0 = r0 0 r1 0 r2 0 r3 0 r4 0 r5 0 r6 0 r7 0 */ |
| 46 |
__ASM_EMIT("vexpandps 0x020(%[dst]), %%zmm1 %{%%k4%}%{z%}") /* zmm1 = r8 0 r9 0 r10 0 r11 0 r12 0 r13 0 r14 0 r15 0 */ |
| 47 |
__ASM_EMIT("vexpandps 0x040(%[dst]), %%zmm0 %{%%k5%}") /* zmm0 = r0 i0 r1 i1 r2 i2 r3 i3 r4 i4 r5 i5 r6 i6 r7 i7 */ |
| 48 |
__ASM_EMIT("vexpandps 0x060(%[dst]), %%zmm1 %{%%k5%}") /* zmm1 = r8 i8 r9 i9 r10 i10 r11 i11 r12 i12 r13 i13 r14 i14 r15 i15 */ |
| 49 |
__ASM_EMIT("vexpandps 0x080(%[dst]), %%zmm2 %{%%k4%}%{z%}") |
| 50 |
__ASM_EMIT("vexpandps 0x0a0(%[dst]), %%zmm3 %{%%k4%}%{z%}") |
| 51 |
__ASM_EMIT("vexpandps 0x0c0(%[dst]), %%zmm2 %{%%k5%}") |
| 52 |
__ASM_EMIT("vmovups 0x0e0(%[dst]), %%ymm4") |
| 53 |
__ASM_EMIT("vexpandps %%zmm4, %%zmm3 %{%%k5%}") |
| 54 |
__ASM_EMIT("vmovups %%zmm0, 0x000(%[dst])") |
| 55 |
__ASM_EMIT("vmovups %%zmm1, 0x040(%[dst])") |
| 56 |
__ASM_EMIT("vmovups %%zmm2, 0x080(%[dst])") |
| 57 |
__ASM_EMIT("vmovups %%zmm3, 0x0c0(%[dst])") |
| 58 |
__ASM_EMIT("add $0x100, %[dst]") |
| 59 |
__ASM_EMIT("sub $32, %[blocks]") |
| 60 |
__ASM_EMIT("jae 1b") |
| 61 |
// 16x block |
| 62 |
__ASM_EMIT("2:") |
| 63 |
__ASM_EMIT("add $16, %[blocks]") |
| 64 |
__ASM_EMIT("jl 4f") |
| 65 |
__ASM_EMIT("vexpandps 0x000(%[dst]), %%zmm0 %{%%k4%}%{z%}") /* zmm0 = r0 0 r1 0 r2 0 r3 0 r4 0 r5 0 r6 0 r7 0 */ |
| 66 |
__ASM_EMIT("vexpandps 0x020(%[dst]), %%zmm1 %{%%k4%}%{z%}") /* zmm1 = r8 0 r9 0 r10 0 r11 0 r12 0 r13 0 r14 0 r15 0 */ |
| 67 |
__ASM_EMIT("vexpandps 0x040(%[dst]), %%zmm0 %{%%k5%}") /* zmm0 = r0 i0 r1 i1 r2 i2 r3 i3 r4 i4 r5 i5 r6 i6 r7 i7 */ |
| 68 |
__ASM_EMIT("vmovups 0x060(%[dst]), %%ymm4") |
| 69 |
__ASM_EMIT("vexpandps %%zmm4, %%zmm1 %{%%k5%}") /* zmm1 = r8 i8 r9 i9 r10 i10 r11 i11 r12 i12 r13 i13 r14 i14 r15 i15 */ |
| 70 |
__ASM_EMIT("vmovups %%zmm0, 0x000(%[dst])") |
| 71 |
__ASM_EMIT("vmovups %%zmm1, 0x040(%[dst])") |
| 72 |
__ASM_EMIT("sub $16, %[blocks]") |
| 73 |
__ASM_EMIT("add $0x100, %[dst]") |
| 74 |
// 8x block |
| 75 |
__ASM_EMIT("4:") |
| 76 |
__ASM_EMIT("add $8, %[blocks]") |
| 77 |
__ASM_EMIT("jl 6f") |
| 78 |
__ASM_EMIT("vmovups 0x00(%[dst]), %%ymm0") /* ymm0 = r0 r1 r2 r3 r4 r5 r6 r7 */ |
| 79 |
__ASM_EMIT("vmovups 0x20(%[dst]), %%ymm1") /* ymm1 = i0 i1 i2 i3 i4 i5 i6 i7 */ |
| 80 |
__ASM_EMIT("vunpcklps %%ymm1, %%ymm0, %%ymm4") /* ymm4 = r0 i0 r1 i1 r4 i4 r5 i5 */ |
| 81 |
__ASM_EMIT("vunpckhps %%ymm1, %%ymm0, %%ymm5") /* ymm5 = r2 i2 r3 i3 r6 i6 r7 i7 */ |
| 82 |
__ASM_EMIT("vmovups %%xmm4, 0x00(%[dst])") |
| 83 |
__ASM_EMIT("vmovups %%xmm5, 0x10(%[dst])") |
| 84 |
__ASM_EMIT("vextractf128 $1, %%ymm4, 0x20(%[dst])") |
| 85 |
__ASM_EMIT("vextractf128 $1, %%ymm5, 0x30(%[dst])") |
| 86 |
__ASM_EMIT("6:") |
| 87 |
|
| 88 |
: [dst] "+r"(dst), [blocks] "+r" (blocks) |
| 89 |
: [MASK] "o" (FFT_REPACK_MASKS) |
| 90 |
: "cc", "memory", |
| 91 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 92 |
"%xmm4", |
| 93 |
"%k4", "%k5" |
| 94 |
); |
| 95 |
} |
| 96 |
|
| 97 |
static inline void packed_fft_repack_normalize(float *dst, size_t rank) |
| 98 |
{ |
| 99 |
size_t blocks = 1 << rank; |
| 100 |
const float norm = 1.0f / float(blocks); |
| 101 |
|
| 102 |
// Perform 4-element butterflies |
| 103 |
ARCH_X86_ASM |
| 104 |
( |
| 105 |
__ASM_EMIT("vbroadcastss %[norm], %%zmm6") |
| 106 |
__ASM_EMIT("kmovw 0x00 + %[MASK], %%k4") |
| 107 |
__ASM_EMIT("kmovw 0x02 + %[MASK], %%k5") |
| 108 |
__ASM_EMIT("vmovaps %%zmm6, %%zmm7") |
| 109 |
// 32x blocks |
| 110 |
__ASM_EMIT("sub $32, %[blocks]") |
| 111 |
__ASM_EMIT("jb 2f") |
| 112 |
__ASM_EMIT("1:") |
| 113 |
__ASM_EMIT("vexpandps 0x000(%[dst]), %%zmm0 %{%%k4%}%{z%}") /* zmm0 = r0 0 r1 0 r2 0 r3 0 r4 0 r5 0 r6 0 r7 0 */ |
| 114 |
__ASM_EMIT("vexpandps 0x020(%[dst]), %%zmm1 %{%%k4%}%{z%}") /* zmm1 = r8 0 r9 0 r10 0 r11 0 r12 0 r13 0 r14 0 r15 0 */ |
| 115 |
__ASM_EMIT("vexpandps 0x040(%[dst]), %%zmm0 %{%%k5%}") /* zmm0 = r0 i0 r1 i1 r2 i2 r3 i3 r4 i4 r5 i5 r6 i6 r7 i7 */ |
| 116 |
__ASM_EMIT("vexpandps 0x060(%[dst]), %%zmm1 %{%%k5%}") /* zmm1 = r8 i8 r9 i9 r10 i10 r11 i11 r12 i12 r13 i13 r14 i14 r15 i15 */ |
| 117 |
__ASM_EMIT("vexpandps 0x080(%[dst]), %%zmm2 %{%%k4%}%{z%}") |
| 118 |
__ASM_EMIT("vexpandps 0x0a0(%[dst]), %%zmm3 %{%%k4%}%{z%}") |
| 119 |
__ASM_EMIT("vexpandps 0x0c0(%[dst]), %%zmm2 %{%%k5%}") |
| 120 |
__ASM_EMIT("vmovups 0x0e0(%[dst]), %%ymm4") |
| 121 |
__ASM_EMIT("vexpandps %%zmm4, %%zmm3 %{%%k5%}") |
| 122 |
__ASM_EMIT("vmulps %%zmm6, %%zmm0, %%zmm0") |
| 123 |
__ASM_EMIT("vmulps %%zmm7, %%zmm1, %%zmm1") |
| 124 |
__ASM_EMIT("vmulps %%zmm6, %%zmm2, %%zmm2") |
| 125 |
__ASM_EMIT("vmulps %%zmm7, %%zmm3, %%zmm3") |
| 126 |
__ASM_EMIT("vmovups %%zmm0, 0x000(%[dst])") |
| 127 |
__ASM_EMIT("vmovups %%zmm1, 0x040(%[dst])") |
| 128 |
__ASM_EMIT("vmovups %%zmm2, 0x080(%[dst])") |
| 129 |
__ASM_EMIT("vmovups %%zmm3, 0x0c0(%[dst])") |
| 130 |
__ASM_EMIT("add $0x100, %[dst]") |
| 131 |
__ASM_EMIT("sub $32, %[blocks]") |
| 132 |
__ASM_EMIT("jae 1b") |
| 133 |
// 16x block |
| 134 |
__ASM_EMIT("2:") |
| 135 |
__ASM_EMIT("add $16, %[blocks]") |
| 136 |
__ASM_EMIT("jl 4f") |
| 137 |
__ASM_EMIT("vexpandps 0x000(%[dst]), %%zmm0 %{%%k4%}%{z%}") /* zmm0 = r0 0 r1 0 r2 0 r3 0 r4 0 r5 0 r6 0 r7 0 */ |
| 138 |
__ASM_EMIT("vexpandps 0x020(%[dst]), %%zmm1 %{%%k4%}%{z%}") /* zmm1 = r8 0 r9 0 r10 0 r11 0 r12 0 r13 0 r14 0 r15 0 */ |
| 139 |
__ASM_EMIT("vexpandps 0x040(%[dst]), %%zmm0 %{%%k5%}") /* zmm0 = r0 i0 r1 i1 r2 i2 r3 i3 r4 i4 r5 i5 r6 i6 r7 i7 */ |
| 140 |
__ASM_EMIT("vmovups 0x060(%[dst]), %%ymm4") |
| 141 |
__ASM_EMIT("vexpandps %%zmm4, %%zmm1 %{%%k5%}") /* zmm1 = r8 i8 r9 i9 r10 i10 r11 i11 r12 i12 r13 i13 r14 i14 r15 i15 */ |
| 142 |
__ASM_EMIT("vmulps %%zmm6, %%zmm0, %%zmm0") |
| 143 |
__ASM_EMIT("vmulps %%zmm7, %%zmm1, %%zmm1") |
| 144 |
__ASM_EMIT("vmovups %%zmm0, 0x000(%[dst])") |
| 145 |
__ASM_EMIT("vmovups %%zmm1, 0x040(%[dst])") |
| 146 |
__ASM_EMIT("sub $16, %[blocks]") |
| 147 |
__ASM_EMIT("add $0x100, %[dst]") |
| 148 |
// 8x block |
| 149 |
__ASM_EMIT("4:") |
| 150 |
__ASM_EMIT("add $8, %[blocks]") |
| 151 |
__ASM_EMIT("jl 6f") |
| 152 |
__ASM_EMIT("vmulps 0x00(%[dst]), %%ymm6, %%ymm0") /* ymm0 = r0 r1 r2 r3 r4 r5 r6 r7 */ |
| 153 |
__ASM_EMIT("vmulps 0x20(%[dst]), %%ymm7, %%ymm1") /* ymm1 = i0 i1 i2 i3 i4 i5 i6 i7 */ |
| 154 |
__ASM_EMIT("vunpcklps %%ymm1, %%ymm0, %%ymm4") /* ymm4 = r0 i0 r1 i1 r4 i4 r5 i5 */ |
| 155 |
__ASM_EMIT("vunpckhps %%ymm1, %%ymm0, %%ymm5") /* ymm5 = r2 i2 r3 i3 r6 i6 r7 i7 */ |
| 156 |
__ASM_EMIT("vmovups %%xmm4, 0x00(%[dst])") |
| 157 |
__ASM_EMIT("vmovups %%xmm5, 0x10(%[dst])") |
| 158 |
__ASM_EMIT("vextractf128 $1, %%ymm4, 0x20(%[dst])") |
| 159 |
__ASM_EMIT("vextractf128 $1, %%ymm5, 0x30(%[dst])") |
| 160 |
__ASM_EMIT("6:") |
| 161 |
|
| 162 |
: [dst] "+r"(dst), [blocks] "+r" (blocks) |
| 163 |
: [norm] "m" (norm), |
| 164 |
[MASK] "o" (FFT_REPACK_MASKS) |
| 165 |
: "cc", "memory", |
| 166 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 167 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 168 |
); |
| 169 |
} |
| 170 |
} /* namespace avx512 */ |
| 171 |
} /* namespace lsp */ |
| 172 |
|
| 173 |
|
| 174 |
#endif /* PRIVATE_DSP_ARCH_X86_AVX512_FFT_P_REPACK_H_ */ |