Debian Codegraph @

modules/lsp-dsp-lib/include/private/dsp/arch/x86/avx512/fft/butterfly.h

raw
1 /*
2 * Copyright (C) 2025 Linux Studio Plugins Project <https://lsp-plug.in/>
3 * (C) 2025 Vladimir Sadovnikov <[email protected]>
4 *
5 * This file is part of lsp-dsp-lib
6 * Created on: 23 февр. 2025 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_BUTTERFLY_H_
23 #define PRIVATE_DSP_ARCH_X86_AVX512_FFT_BUTTERFLY_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 #define FFT_BUTTERFLY_BODY16(fmadd_b, fmadd_a) \
34 __IF_32(float *ptr1, *ptr2);\
35 \
36 ARCH_X86_ASM( \
37 /* Prepare angle */ \
38 __ASM_EMIT32("mov %[fft_a], %[ptr2]") \
39 __ASM_EMIT32("mov %[dst_re], %[ptr1]") \
40 __ASM_EMIT("vmovaps 0x00(%[" __IF_32_64("ptr2", "fft_a") "]), %%zmm6") /* zmm6 = x_re */ \
41 __ASM_EMIT("vmovaps 0x40(%[" __IF_32_64("ptr2", "fft_a") "]), %%zmm7") /* zmm7 = x_im */ \
42 __ASM_EMIT32("mov %[dst_im], %[ptr2]") \
43 /* Start loop */ \
44 __ASM_EMIT("1:") \
45 __ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off1]), %%zmm0") /* zmm0 = a_re */ \
46 __ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off2]), %%zmm2") /* zmm2 = b_re */ \
47 __ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off1]), %%zmm1") /* zmm1 = a_im */ \
48 __ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off2]), %%zmm3") /* zmm3 = b_im */ \
49 /* Calculate complex multiplication */ \
50 __ASM_EMIT("vmulps %%zmm7, %%zmm2, %%zmm4") /* zmm4 = x_im * b_re */ \
51 __ASM_EMIT("vmulps %%zmm7, %%zmm3, %%zmm5") /* zmm5 = x_im * b_im */ \
52 __ASM_EMIT(fmadd_b " %%zmm6, %%zmm2, %%zmm5") /* zmm5 = c_re = x_re * b_re +- x_im * b_im */ \
53 __ASM_EMIT(fmadd_a " %%zmm6, %%zmm3, %%zmm4") /* zmm4 = c_im = x_re * b_im -+ x_im * b_re */ \
54 /* Perform butterfly */ \
55 __ASM_EMIT("vsubps %%zmm5, %%zmm0, %%zmm2") /* zmm2 = a_re - c_re */ \
56 __ASM_EMIT("vsubps %%zmm4, %%zmm1, %%zmm3") /* zmm3 = a_im - c_im */ \
57 __ASM_EMIT("vaddps %%zmm5, %%zmm0, %%zmm0") /* zmm0 = a_re + c_re */ \
58 __ASM_EMIT("vaddps %%zmm4, %%zmm1, %%zmm1") /* zmm1 = a_im + c_im */ \
59 /* Store values */ \
60 __ASM_EMIT("vmovups %%zmm0, 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off1])") \
61 __ASM_EMIT("vmovups %%zmm2, 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off2])") \
62 __ASM_EMIT("vmovups %%zmm1, 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off1])") \
63 __ASM_EMIT("vmovups %%zmm3, 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off2])") \
64 __ASM_EMIT("add $0x40, %[off1]") \
65 __ASM_EMIT("add $0x40, %[off2]") \
66 __ASM_EMIT32("subl $16, %[np]") \
67 __ASM_EMIT64("subq $16, %[np]") \
68 __ASM_EMIT("jz 2f") \
69 /* Rotate angle */ \
70 __ASM_EMIT32("mov %[fft_w], %[ptr2]") \
71 __ASM_EMIT("vmovaps 0x00(%[" __IF_32_64("ptr2", "fft_w") "]), %%zmm4") /* zmm4 = w_re */ \
72 __ASM_EMIT("vmovaps 0x40(%[" __IF_32_64("ptr2", "fft_w") "]), %%zmm5") /* zmm5 = w_im */ \
73 __ASM_EMIT32("mov %[dst_im], %[ptr2]") \
74 __ASM_EMIT("vmulps %%zmm5, %%zmm6, %%zmm2") /* zmm2 = w_im * x_re */ \
75 __ASM_EMIT("vmulps %%zmm5, %%zmm7, %%zmm3") /* zmm3 = w_im * x_im */ \
76 __ASM_EMIT("vfmsub132ps %%zmm4, %%zmm3, %%zmm6") /* zmm6 = x_re' = w_re * x_re - w_im * x_im */ \
77 __ASM_EMIT("vfmadd132ps %%zmm4, %%zmm2, %%zmm7") /* zmm7 = x_im' = w_re * x_im + w_im * x_re */ \
78 /* Repeat loop */ \
79 __ASM_EMIT("jmp 1b") \
80 __ASM_EMIT("2:") \
81 \
82 : __IF_32([ptr1] "=&r" (ptr1), [ptr2] "=&r" (ptr2), ) \
83 [off1] "+r" (off1), [off2] "+r" (off2), \
84 [np] X86_PGREG (np) \
85 : [dst_re] X86_GREG (dst_re), [dst_im] X86_GREG (dst_im), \
86 [fft_a] X86_GREG (fft_a), \
87 [fft_w] X86_GREG (fft_w) \
88 : "cc", "memory", \
89 "%xmm0", "%xmm1", "%xmm2", "%xmm3", \
90 "%xmm4", "%xmm5", "%xmm6", "%xmm7" \
91 );
92
93 static inline void butterfly_direct16p(float *dst_re, float *dst_im, size_t rank, size_t blocks)
94 {
95 size_t pairs = 1 << rank;
96 size_t off1 = 0, shift = sizeof(float) << rank;
97 const float *fft_a = &FFT_A[(rank - 2) << 5];
98 const float *fft_w = &FFT_DW[(rank - 3) << 5];
99
100 for (size_t b=0; b<blocks; ++b)
101 {
102 size_t off2 = off1 + shift;
103 size_t np = pairs;
104
105 FFT_BUTTERFLY_BODY16("vfmadd231ps", "vfmsub231ps");
106
107 off1 = off2;
108 }
109 }
110
111 static inline void butterfly_reverse16p(float *dst_re, float *dst_im, size_t rank, size_t blocks)
112 {
113 size_t pairs = 1 << rank;
114 size_t off1 = 0, shift = sizeof(float) << rank;
115 const float *fft_a = &FFT_A[(rank - 2) << 5];
116 const float *fft_w = &FFT_DW[(rank - 3) << 5];
117
118 for (size_t b=0; b<blocks; ++b)
119 {
120 size_t off2 = off1 + shift;
121 size_t np = pairs;
122
123 FFT_BUTTERFLY_BODY16("vfmsub231ps", "vfmadd231ps");
124
125 off1 = off2;
126 }
127 }
128
129 #undef FFT_BUTTERFLY_BODY16
130 } /* namespace avx512 */
131 } /* namespace lsp */
132
133 #endif /* PRIVATE_DSP_ARCH_X86_AVX512_FFT_BUTTERFLY_H_ */