Debian Codegraph @

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

raw
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_BUTTERFLY_H_
23 #define PRIVATE_DSP_ARCH_X86_AVX_FFT_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 __IF_32(float *ptr1, *ptr2);\
35 \
36 ARCH_X86_ASM \
37 ( \
38 /* Prepare angle */ \
39 __ASM_EMIT32("mov %[fft_a], %[ptr2]") \
40 __ASM_EMIT32("mov %[dst_re], %[ptr1]") \
41 __ASM_EMIT("vmovaps 0x00(%[" __IF_32_64("ptr2", "fft_a") "]), %%ymm6") /* ymm6 = x_re */ \
42 __ASM_EMIT("vmovaps 0x20(%[" __IF_32_64("ptr2", "fft_a") "]), %%ymm7") /* ymm7 = x_im */ \
43 __ASM_EMIT32("mov %[dst_im], %[ptr2]") \
44 /* Start loop */ \
45 __ASM_EMIT("1:") \
46 __ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off1]), %%ymm0") /* ymm0 = a_re */ \
47 __ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off2]), %%ymm2") /* ymm2 = b_re */ \
48 __ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off1]), %%ymm1") /* ymm1 = a_im */ \
49 __ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off2]), %%ymm3") /* ymm3 = b_im */ \
50 /* Calculate complex multiplication */ \
51 __ASM_EMIT("vmulps %%ymm7, %%ymm2, %%ymm4") /* ymm4 = x_im * b_re */ \
52 __ASM_EMIT("vmulps %%ymm7, %%ymm3, %%ymm5") /* ymm5 = x_im * b_im */ \
53 __ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm2, %%ymm2", "")) /* ymm2 = x_re * b_re */ \
54 __ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm3, %%ymm3", "")) /* ymm3 = x_re * b_im */ \
55 __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 */ \
56 __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 */ \
57 /* Perform butterfly */ \
58 __ASM_EMIT("vsubps %%ymm5, %%ymm0, %%ymm2") /* ymm2 = a_re - c_re */ \
59 __ASM_EMIT("vsubps %%ymm4, %%ymm1, %%ymm3") /* ymm3 = a_im - c_im */ \
60 __ASM_EMIT("vaddps %%ymm5, %%ymm0, %%ymm0") /* ymm0 = a_re + c_re */ \
61 __ASM_EMIT("vaddps %%ymm4, %%ymm1, %%ymm1") /* ymm1 = a_im + c_im */ \
62 /* Store values */ \
63 __ASM_EMIT("vmovups %%ymm0, 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off1])") \
64 __ASM_EMIT("vmovups %%ymm2, 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off2])") \
65 __ASM_EMIT("vmovups %%ymm1, 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off1])") \
66 __ASM_EMIT("vmovups %%ymm3, 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off2])") \
67 __ASM_EMIT("add $0x20, %[off1]") \
68 __ASM_EMIT("add $0x20, %[off2]") \
69 __ASM_EMIT32("subl $8, %[np]") \
70 __ASM_EMIT64("subq $8, %[np]") \
71 __ASM_EMIT("jz 2f") \
72 /* Rotate angle */ \
73 __ASM_EMIT32("mov %[fft_w], %[ptr2]") \
74 __ASM_EMIT("vmovaps 0x00(%[" __IF_32_64("ptr2", "fft_w") "]), %%ymm4") /* xmm4 = w_re */ \
75 __ASM_EMIT("vmovaps 0x20(%[" __IF_32_64("ptr2", "fft_w") "]), %%ymm5") /* xmm5 = w_im */ \
76 __ASM_EMIT32("mov %[dst_im], %[ptr2]") \
77 __ASM_EMIT("vmulps %%ymm5, %%ymm6, %%ymm2") /* ymm2 = w_im * x_re */ \
78 __ASM_EMIT("vmulps %%ymm5, %%ymm7, %%ymm3") /* ymm3 = w_im * x_im */ \
79 __ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm6, %%ymm6", "")) /* ymm6 = w_re * x_re */ \
80 __ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm7, %%ymm7", "")) /* ymm7 = w_re * x_im */ \
81 __ASM_EMIT(FMA_SEL("vsubps %%ymm3, %%ymm6, %%ymm6", "vfmsub132ps %%ymm4, %%ymm3, %%ymm6")) /* ymm6 = x_re' = w_re * x_re - w_im * x_im */ \
82 __ASM_EMIT(FMA_SEL("vaddps %%ymm2, %%ymm7, %%ymm7", "vfmadd132ps %%ymm4, %%ymm2, %%ymm7")) /* ymm7 = x_im' = w_re * x_im + w_im * x_re */ \
83 /* Repeat loop */ \
84 __ASM_EMIT("jmp 1b") \
85 __ASM_EMIT("2:") \
86 \
87 : __IF_32([ptr1] "=&r" (ptr1), [ptr2] "=&r" (ptr2), ) \
88 [off1] "+r" (off1), [off2] "+r" (off2), \
89 [np] X86_PGREG (np) \
90 : [dst_re] X86_GREG (dst_re), [dst_im] X86_GREG (dst_im), [fft_a] X86_GREG (fft_a), [fft_w] X86_GREG (fft_w) \
91 : "cc", "memory", \
92 "%xmm0", "%xmm1", "%xmm2", "%xmm3", \
93 "%xmm4", "%xmm5", "%xmm6", "%xmm7" \
94 );
95
96 #define FMA_OFF(a, b) a
97 #define FMA_ON(a, b) b
98
99 static inline void butterfly_direct8p(float *dst_re, float *dst_im, size_t rank, size_t blocks)
100 {
101 size_t pairs = 1 << rank;
102 size_t off1 = 0, shift = 4 << rank; //1 << (rank + 2);
103 const float *fft_a = &FFT_A[(rank - 2) << 4];
104 const float *fft_w = &FFT_DW[(rank - 2) << 4];
105
106 for (size_t b=0; b<blocks; ++b)
107 {
108 size_t off2 = off1 + shift;
109 size_t np = pairs;
110
111 FFT_BUTTERFLY_BODY8("vaddps", "vsubps", FMA_OFF);
112
113 off1 = off2;
114 }
115 }
116
117 static inline void butterfly_reverse8p(float *dst_re, float *dst_im, size_t rank, size_t blocks)
118 {
119 size_t pairs = 1 << rank;
120 size_t off1 = 0, shift = 4 << rank; // 1 << (rank + 2);
121 const float *fft_a = &FFT_A[(rank - 2) << 4];
122 const float *fft_w = &FFT_DW[(rank - 2) << 4];
123
124 for (size_t b=0; b<blocks; ++b)
125 {
126 size_t off2 = off1 + shift;
127 size_t np = pairs;
128
129 FFT_BUTTERFLY_BODY8("vsubps", "vaddps", FMA_OFF);
130
131 off1 = off2;
132 }
133 }
134
135 static inline void butterfly_direct8p_fma3(float *dst_re, float *dst_im, size_t rank, size_t blocks)
136 {
137 size_t pairs = 1 << rank;
138 size_t off1 = 0, shift = 4 << rank; // 1 << (rank + 2);
139 const float *fft_a = &FFT_A[(rank - 2) << 4];
140 const float *fft_w = &FFT_DW[(rank - 2) << 4];
141
142 for (size_t b=0; b<blocks; ++b)
143 {
144 size_t off2 = off1 + shift;
145 size_t np = pairs;
146
147 FFT_BUTTERFLY_BODY8("vfmadd231ps", "vfmsub231ps", FMA_ON);
148
149 off1 = off2;
150 }
151 }
152
153 static inline void butterfly_reverse8p_fma3(float *dst_re, float *dst_im, size_t rank, size_t blocks)
154 {
155 size_t pairs = 1 << rank;
156 size_t off1 = 0, shift = 4 << rank; // 1 << (rank + 2);
157 const float *fft_a = &FFT_A[(rank - 2) << 4];
158 const float *fft_w = &FFT_DW[(rank - 2) << 4];
159
160 for (size_t b=0; b<blocks; ++b)
161 {
162 size_t off2 = off1 + shift;
163 size_t np = pairs;
164
165 FFT_BUTTERFLY_BODY8("vfmsub231ps", "vfmadd231ps", FMA_ON);
166
167 off1 = off2;
168 }
169 }
170
171 #undef FMA_OFF
172 #undef FMA_ON
173 #undef FFT_BUTTERFLY_BODY8
174 }
175 }
176
177 #endif /* PRIVATE_DSP_ARCH_X86_AVX_FFT_BUTTERFLY_H_ */