| 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_FASTCONV_BUTTERFLY_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_X86_AVX_FASTCONV_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 __SKIP(x) |
| 34 |
|
| 35 |
// c = a - b |
| 36 |
// a' = a + b |
| 37 |
// b' = c * w |
| 38 |
#define FASTCONV_DIRECT_BUTTERFLY_BODY8(add_re, add_im, FMA_SEL) \ |
| 39 |
ARCH_X86_ASM \ |
| 40 |
( \ |
| 41 |
/* Prepare angle */ \ |
| 42 |
__ASM_EMIT("vmovaps 0x00(%[ak]), %%ymm6") /* ymm6 = x_re */ \ |
| 43 |
__ASM_EMIT("vmovaps 0x20(%[ak]), %%ymm7") /* ymm7 = x_im */ \ |
| 44 |
/* Start loop */ \ |
| 45 |
__ASM_EMIT("1:") \ |
| 46 |
__ASM_EMIT("vmovups 0x00(%[dst], %[off1]), %%ymm0") /* ymm0 = a_re */ \ |
| 47 |
__ASM_EMIT("vmovups 0x20(%[dst], %[off1]), %%ymm1") /* ymm1 = a_im */ \ |
| 48 |
__ASM_EMIT("vmovups 0x00(%[dst], %[off2]), %%ymm2") /* ymm2 = b_re */ \ |
| 49 |
__ASM_EMIT("vmovups 0x20(%[dst], %[off2]), %%ymm3") /* ymm3 = b_im */ \ |
| 50 |
/* Perform butterfly */ \ |
| 51 |
__ASM_EMIT("vsubps %%ymm2, %%ymm0, %%ymm4") /* ymm4 = c_re = a_re - b_re */ \ |
| 52 |
__ASM_EMIT("vsubps %%ymm3, %%ymm1, %%ymm5") /* ymm5 = c_im = a_im - b_im */ \ |
| 53 |
__ASM_EMIT("vaddps %%ymm2, %%ymm0, %%ymm0") /* ymm0 = a_re' = a_re + b_re */ \ |
| 54 |
__ASM_EMIT("vaddps %%ymm3, %%ymm1, %%ymm1") /* ymm1 = a_im' = a_im + b_im */ \ |
| 55 |
__ASM_EMIT("vmulps %%ymm7, %%ymm4, %%ymm2") /* ymm2 = x_im * c_re */ \ |
| 56 |
__ASM_EMIT("vmulps %%ymm7, %%ymm5, %%ymm3") /* ymm3 = x_im * c_im */ \ |
| 57 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm4, %%ymm4", "")) /* ymm4 = x_re * c_re */ \ |
| 58 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm5, %%ymm5", "")) /* ymm5 = x_re * c_im */ \ |
| 59 |
__ASM_EMIT(FMA_SEL(add_re " %%ymm3, %%ymm4, %%ymm4", add_re " %%ymm6, %%ymm3, %%ymm4")) /* ymm4 = b_re = x_re * c_re +- x_im * c_im */ \ |
| 60 |
__ASM_EMIT(FMA_SEL(add_im " %%ymm2, %%ymm5, %%ymm5", add_im " %%ymm6, %%ymm2, %%ymm5")) /* ymm5 = b_im = x_re * c_im -+ x_im * c_re */ \ |
| 61 |
/* Store values */ \ |
| 62 |
__ASM_EMIT("vmovups %%ymm0, 0x00(%[dst], %[off1])") \ |
| 63 |
__ASM_EMIT("vmovups %%ymm1, 0x20(%[dst], %[off1])") \ |
| 64 |
__ASM_EMIT("vmovups %%ymm4, 0x00(%[dst], %[off2])") \ |
| 65 |
__ASM_EMIT("vmovups %%ymm5, 0x20(%[dst], %[off2])") \ |
| 66 |
__ASM_EMIT("add $0x40, %[off1]") \ |
| 67 |
__ASM_EMIT("add $0x40, %[off2]") \ |
| 68 |
__ASM_EMIT32("subl $8, %[np]") \ |
| 69 |
__ASM_EMIT64("subq $8, %[np]") \ |
| 70 |
__ASM_EMIT("jz 2f") \ |
| 71 |
/* Rotate angle */ \ |
| 72 |
__ASM_EMIT("vmovaps 0x00(%[wk]), %%ymm4") /* xmm4 = w_re */ \ |
| 73 |
__ASM_EMIT("vmovaps 0x20(%[wk]), %%ymm5") /* xmm5 = w_im */ \ |
| 74 |
__ASM_EMIT("vmulps %%ymm5, %%ymm6, %%ymm2") /* ymm2 = w_im * x_re */ \ |
| 75 |
__ASM_EMIT("vmulps %%ymm5, %%ymm7, %%ymm3") /* ymm3 = w_im * x_im */ \ |
| 76 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm6, %%ymm6", "")) /* ymm6 = w_re * x_re */ \ |
| 77 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm7, %%ymm7", "")) /* ymm7 = w_re * x_im */ \ |
| 78 |
__ASM_EMIT(FMA_SEL("vsubps %%ymm3, %%ymm6, %%ymm6", "vfmsub132ps %%ymm4, %%ymm3, %%ymm6")) /* ymm6 = x_re' = w_re * x_re - w_im * x_im */ \ |
| 79 |
__ASM_EMIT(FMA_SEL("vaddps %%ymm2, %%ymm7, %%ymm7", "vfmadd132ps %%ymm4, %%ymm2, %%ymm7")) /* ymm7 = x_im' = w_re * x_im + w_im * x_re */ \ |
| 80 |
/* Repeat loop */ \ |
| 81 |
__ASM_EMIT("jmp 1b") \ |
| 82 |
__ASM_EMIT("2:") \ |
| 83 |
\ |
| 84 |
: [off1] "+r" (off1), [off2] "+r" (off2), [np] __ASM_ARG_RW(np) \ |
| 85 |
: [dst] "r" (dst), [ak] "r" (ak), [wk] "r" (wk) \ |
| 86 |
: "cc", "memory", \ |
| 87 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", \ |
| 88 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" \ |
| 89 |
); |
| 90 |
|
| 91 |
|
| 92 |
#define FASTCONV_DIRECT_BUTTERFLY_LAST(add_re, add_im, FMA_SEL) \ |
| 93 |
ARCH_X86_ASM( \ |
| 94 |
/* Loop 2x 4-element butterflies */ \ |
| 95 |
__ASM_EMIT("vmovaps 0x00 + %[FFT_A], %%ymm6") /* ymm6 = x_re */ \ |
| 96 |
__ASM_EMIT("vmovaps 0x20 + %[FFT_A], %%ymm7") /* ymm7 = x_im */ \ |
| 97 |
__ASM_EMIT("sub $2, %[nb]") \ |
| 98 |
__ASM_EMIT("jb 2f") \ |
| 99 |
__ASM_EMIT("1:") \ |
| 100 |
/* Load data to registers */ \ |
| 101 |
__ASM_EMIT("vmovups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 r1 r2 r3 */ \ |
| 102 |
__ASM_EMIT("vmovups 0x10(%[dst]), %%xmm2") /* xmm2 = r4 r5 r6 r7 */ \ |
| 103 |
__ASM_EMIT("vmovups 0x20(%[dst]), %%xmm1") /* xmm1 = i0 i1 i2 i3 */ \ |
| 104 |
__ASM_EMIT("vmovups 0x30(%[dst]), %%xmm3") /* xmm3 = i4 i5 i6 i7 */ \ |
| 105 |
__ASM_EMIT("vinsertf128 $1, 0x40(%[dst]), %%ymm0, %%ymm0") /* ymm0 = a_re = r0 r1 r2 r3 r8 r9 r10 r11 */ \ |
| 106 |
__ASM_EMIT("vinsertf128 $1, 0x50(%[dst]), %%ymm2, %%ymm2") /* ymm2 = b_re = r4 r5 r6 r7 r12 r13 r14 r15 */ \ |
| 107 |
__ASM_EMIT("vinsertf128 $1, 0x60(%[dst]), %%ymm1, %%ymm1") /* ymm1 = a_im = i0 i1 i2 i3 i8 i9 i10 i11 */ \ |
| 108 |
__ASM_EMIT("vinsertf128 $1, 0x70(%[dst]), %%ymm3, %%ymm3") /* ymm3 = b_im = i4 i5 i6 i7 i12 i13 i14 i15 */ \ |
| 109 |
/* Perform 3rd-order butterflies */ \ |
| 110 |
__ASM_EMIT("vsubps %%ymm2, %%ymm0, %%ymm4") /* ymm4 = c_re = a_re - b_re */ \ |
| 111 |
__ASM_EMIT("vsubps %%ymm3, %%ymm1, %%ymm5") /* ymm5 = c_im = a_im - b_im */ \ |
| 112 |
__ASM_EMIT("vaddps %%ymm2, %%ymm0, %%ymm0") /* ymm0 = a_re' = a_re + b_re */ \ |
| 113 |
__ASM_EMIT("vaddps %%ymm3, %%ymm1, %%ymm1") /* ymm1 = a_im' = a_im + b_im */ \ |
| 114 |
__ASM_EMIT("vmulps %%ymm7, %%ymm4, %%ymm2") /* ymm2 = x_im * c_re */ \ |
| 115 |
__ASM_EMIT("vmulps %%ymm7, %%ymm5, %%ymm3") /* ymm3 = x_im * c_im */ \ |
| 116 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm4, %%ymm4", "")) /* ymm4 = x_re * c_re */ \ |
| 117 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm5, %%ymm5", "")) /* ymm5 = x_re * c_im */ \ |
| 118 |
__ASM_EMIT(FMA_SEL(add_re " %%ymm3, %%ymm4, %%ymm4", add_re " %%ymm6, %%ymm3, %%ymm4")) /* ymm4 = b_re = x_re * c_re +- x_im * c_im */ \ |
| 119 |
__ASM_EMIT(FMA_SEL(add_im " %%ymm2, %%ymm5, %%ymm5", add_im " %%ymm6, %%ymm2, %%ymm5")) /* ymm5 = b_im = x_re * c_im -+ x_im * c_re */ \ |
| 120 |
/* 2nd-order butterflies */ \ |
| 121 |
/* s0" = (r0 + r2) + j*(i0 + i2) + (r1 + r3) + j*(i1 + i3) */ \ |
| 122 |
/* s1" = (r0 + r2) + j*(i0 + i2) - (r1 + r3) - j*(i1 + i3) */ \ |
| 123 |
/* s2" = (r0 - r2) + j*(i0 - i2) + (i1 - i3) - j*(r1 - r3) */ \ |
| 124 |
/* s3" = (r0 - r2) + j*(i0 - i2) - (i1 - i3) + j*(r1 - r3) */ \ |
| 125 |
/* ymm0 = r0 r1 r2 r3 ... */ \ |
| 126 |
/* ymm1 = i0 i1 i2 i3 ... */ \ |
| 127 |
/* ymm4 = r4 r5 r6 r7 ... */ \ |
| 128 |
/* ymm5 = i4 i5 i6 i7 ... */ \ |
| 129 |
__ASM_EMIT("vshufps $0xd8, %%ymm0, %%ymm0, %%ymm0") /* ymm0 = r0 r2 r1 r3 */ \ |
| 130 |
__ASM_EMIT("vshufps $0xd8, %%ymm1, %%ymm1, %%ymm1") /* ymm1 = i0 i2 i1 i3 */ \ |
| 131 |
__ASM_EMIT("vshufps $0xd8, %%ymm4, %%ymm4, %%ymm4") \ |
| 132 |
__ASM_EMIT("vshufps $0xd8, %%ymm5, %%ymm5, %%ymm5") \ |
| 133 |
__ASM_EMIT("vhsubps %%ymm1, %%ymm0, %%ymm2") /* ymm2 = r0-r2 r1-r3 i0-i2 i1-i3 = r1' r3' i1' i3' */ \ |
| 134 |
__ASM_EMIT("vhsubps %%ymm5, %%ymm4, %%ymm3") \ |
| 135 |
__ASM_EMIT("vhaddps %%ymm1, %%ymm0, %%ymm0") /* ymm0 = r0+r2 r1+r3 i0+i2 i1+i3 = r0' r2' i0' i2' */ \ |
| 136 |
__ASM_EMIT("vhaddps %%ymm5, %%ymm4, %%ymm4") \ |
| 137 |
/* 1st-order 8x butterfly */ \ |
| 138 |
__ASM_EMIT("vshufps $0x6e, %%ymm2, %%ymm0, %%ymm1") /* ymm0 = i0' i2' i1' r3' */ \ |
| 139 |
__ASM_EMIT("vshufps $0x6e, %%ymm3, %%ymm4, %%ymm5") \ |
| 140 |
__ASM_EMIT("vshufps $0xc4, %%ymm2, %%ymm0, %%ymm0") /* ymm1 = r0' r2' r1' i3' */ \ |
| 141 |
__ASM_EMIT("vshufps $0xc4, %%ymm3, %%ymm4, %%ymm4") \ |
| 142 |
__ASM_EMIT("vhsubps %%ymm1, %%ymm0, %%ymm2") /* ymm2 = r0'-r2' r1'-i3' i0'-i2' i1'-r3' = r1" r3" i1" i2" */ \ |
| 143 |
__ASM_EMIT("vhsubps %%ymm5, %%ymm4, %%ymm3") \ |
| 144 |
__ASM_EMIT("vhaddps %%ymm1, %%ymm0, %%ymm0") /* ymm0 = r0'+r2' r1'+i3' i0'+i2' i1'+r3' = r0" r2" i0" i3" */ \ |
| 145 |
__ASM_EMIT("vhaddps %%ymm5, %%ymm4, %%ymm4") \ |
| 146 |
__ASM_EMIT("vblendps $0x88, %%ymm0, %%ymm2, %%ymm1") /* ymm1 = r1" r3" i1" i3" */ \ |
| 147 |
__ASM_EMIT("vblendps $0x88, %%ymm4, %%ymm3, %%ymm5") \ |
| 148 |
__ASM_EMIT("vblendps $0x88, %%ymm2, %%ymm0, %%ymm0") /* ymm0 = r0" r2" i0" i2" */ \ |
| 149 |
__ASM_EMIT("vblendps $0x88, %%ymm3, %%ymm4, %%ymm4") \ |
| 150 |
__ASM_EMIT("vunpckhps %%ymm1, %%ymm0, %%ymm3") /* ymm3 = r0" r1" r2" r3" */ \ |
| 151 |
__ASM_EMIT("vunpcklps %%ymm1, %%ymm0, %%ymm2") /* ymm2 = i0" i1" i2" i3" */ \ |
| 152 |
__ASM_EMIT("vunpckhps %%ymm5, %%ymm4, %%ymm1") \ |
| 153 |
__ASM_EMIT("vunpcklps %%ymm5, %%ymm4, %%ymm0") \ |
| 154 |
/* Store */ \ |
| 155 |
__ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") \ |
| 156 |
__ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") \ |
| 157 |
__ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") \ |
| 158 |
__ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") \ |
| 159 |
__ASM_EMIT("vextractf128 $1, %%ymm2, 0x40(%[dst])") \ |
| 160 |
__ASM_EMIT("vextractf128 $1, %%ymm0, 0x50(%[dst])") \ |
| 161 |
__ASM_EMIT("vextractf128 $1, %%ymm3, 0x60(%[dst])") \ |
| 162 |
__ASM_EMIT("vextractf128 $1, %%ymm1, 0x70(%[dst])") \ |
| 163 |
/* Move pointers and repeat*/ \ |
| 164 |
__ASM_EMIT("add $0x80, %[dst]") \ |
| 165 |
__ASM_EMIT("sub $2, %[nb]") \ |
| 166 |
__ASM_EMIT("jae 1b") \ |
| 167 |
__ASM_EMIT("2:") \ |
| 168 |
/* 1x 4-element butterflies */ \ |
| 169 |
__ASM_EMIT("add $1, %[nb]") \ |
| 170 |
__ASM_EMIT("jl 4f") \ |
| 171 |
__ASM_EMIT("vmovups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 r1 r2 r3 */ \ |
| 172 |
__ASM_EMIT("vmovups 0x10(%[dst]), %%xmm2") /* xmm2 = r4 r5 r6 r7 */ \ |
| 173 |
__ASM_EMIT("vmovups 0x20(%[dst]), %%xmm1") /* xmm1 = i0 i1 i2 i3 */ \ |
| 174 |
__ASM_EMIT("vmovups 0x30(%[dst]), %%xmm3") /* xmm3 = i4 i5 i6 i7 */ \ |
| 175 |
/* Perform 3rd-order 8x butterfly */ \ |
| 176 |
__ASM_EMIT("vsubps %%xmm2, %%xmm0, %%xmm4") /* xmm4 = c_re = a_re - b_re */ \ |
| 177 |
__ASM_EMIT("vsubps %%xmm3, %%xmm1, %%xmm5") /* xmm5 = c_im = a_im - b_im */ \ |
| 178 |
__ASM_EMIT("vaddps %%xmm2, %%xmm0, %%xmm0") /* xmm0 = a_re' = a_re + b_re */ \ |
| 179 |
__ASM_EMIT("vaddps %%xmm3, %%xmm1, %%xmm1") /* xmm1 = a_im' = a_im + b_im */ \ |
| 180 |
__ASM_EMIT("vmulps %%xmm7, %%xmm4, %%xmm2") /* xmm2 = x_im * c_re */ \ |
| 181 |
__ASM_EMIT("vmulps %%xmm7, %%xmm5, %%xmm3") /* xmm3 = x_im * c_im */ \ |
| 182 |
__ASM_EMIT(FMA_SEL("vmulps %%xmm6, %%xmm4, %%xmm4", "")) /* xmm4 = x_re * c_re */ \ |
| 183 |
__ASM_EMIT(FMA_SEL("vmulps %%xmm6, %%xmm5, %%xmm5", "")) /* xmm5 = x_re * c_im */ \ |
| 184 |
__ASM_EMIT(FMA_SEL(add_re " %%xmm3, %%xmm4, %%xmm4", add_re " %%xmm6, %%xmm3, %%xmm4")) /* xmm4 = b_re = x_re * c_re +- x_im * c_im */ \ |
| 185 |
__ASM_EMIT(FMA_SEL(add_im " %%xmm2, %%xmm5, %%xmm5", add_im " %%xmm6, %%xmm2, %%xmm5")) /* xmm5 = b_im = x_re * c_im -+ x_im * c_re */ \ |
| 186 |
/* 2nd-order butterflies */ \ |
| 187 |
/* s0" = (r0 + r2) + j*(i0 + i2) + (r1 + r3) + j*(i1 + i3) */ \ |
| 188 |
/* s1" = (r0 + r2) + j*(i0 + i2) - (r1 + r3) - j*(i1 + i3) */ \ |
| 189 |
/* s2" = (r0 - r2) + j*(i0 - i2) + (i1 - i3) - j*(r1 - r3) */ \ |
| 190 |
/* s3" = (r0 - r2) + j*(i0 - i2) - (i1 - i3) + j*(r1 - r3) */ \ |
| 191 |
/* xmm0 = r0 r1 r2 r3 ... */ \ |
| 192 |
/* xmm1 = i0 i1 i2 i3 ... */ \ |
| 193 |
/* xmm4 = r4 r5 r6 r7 ... */ \ |
| 194 |
/* xmm5 = i4 i5 i6 i7 ... */ \ |
| 195 |
__ASM_EMIT("vshufps $0xd8, %%xmm0, %%xmm0, %%xmm0") /* xmm0 = r0 r2 r1 r3 */ \ |
| 196 |
__ASM_EMIT("vshufps $0xd8, %%xmm1, %%xmm1, %%xmm1") /* xmm1 = i0 i2 i1 i3 */ \ |
| 197 |
__ASM_EMIT("vshufps $0xd8, %%xmm4, %%xmm4, %%xmm4") \ |
| 198 |
__ASM_EMIT("vshufps $0xd8, %%xmm5, %%xmm5, %%xmm5") \ |
| 199 |
__ASM_EMIT("vhsubps %%xmm1, %%xmm0, %%xmm2") /* xmm2 = r0-r2 r1-r3 i0-i2 i1-i3 = r1' r3' i1' i3' */ \ |
| 200 |
__ASM_EMIT("vhsubps %%xmm5, %%xmm4, %%xmm3") \ |
| 201 |
__ASM_EMIT("vhaddps %%xmm1, %%xmm0, %%xmm0") /* xmm0 = r0+r2 r1+r3 i0+i2 i1+i3 = r0' r2' i0' i2' */ \ |
| 202 |
__ASM_EMIT("vhaddps %%xmm5, %%xmm4, %%xmm4") \ |
| 203 |
/* 1st-order butterflies */ \ |
| 204 |
__ASM_EMIT("vshufps $0x6e, %%xmm2, %%xmm0, %%xmm1") /* xmm0 = i0' i2' i1' r3' */ \ |
| 205 |
__ASM_EMIT("vshufps $0x6e, %%xmm3, %%xmm4, %%xmm5") \ |
| 206 |
__ASM_EMIT("vshufps $0xc4, %%xmm2, %%xmm0, %%xmm0") /* xmm1 = r0' r2' r1' i3' */ \ |
| 207 |
__ASM_EMIT("vshufps $0xc4, %%xmm3, %%xmm4, %%xmm4") \ |
| 208 |
__ASM_EMIT("vmovups %%xmm0, 0x00(%[dst])") \ |
| 209 |
__ASM_EMIT("vmovups %%xmm1, 0x10(%[dst])") \ |
| 210 |
__ASM_EMIT("vmovups %%xmm4, 0x20(%[dst])") \ |
| 211 |
__ASM_EMIT("vmovups %%xmm5, 0x30(%[dst])") \ |
| 212 |
__ASM_EMIT("vhsubps %%xmm1, %%xmm0, %%xmm2") /* xmm2 = r0'-r2' r1'-i3' i0'-i2' i1'-r3' = r1" r3" i1" i2" */ \ |
| 213 |
__ASM_EMIT("vhsubps %%xmm5, %%xmm4, %%xmm3") \ |
| 214 |
__ASM_EMIT("vhaddps %%xmm1, %%xmm0, %%xmm0") /* xmm0 = r0'+r2' r1'+i3' i0'+i2' i1'+r3' = r0" r2" i0" i3" */ \ |
| 215 |
__ASM_EMIT("vhaddps %%xmm5, %%xmm4, %%xmm4") \ |
| 216 |
__ASM_EMIT("vblendps $0x88, %%xmm0, %%xmm2, %%xmm1") /* xmm1 = r1" r3" i1" i3" */ \ |
| 217 |
__ASM_EMIT("vblendps $0x88, %%xmm4, %%xmm3, %%xmm5") \ |
| 218 |
__ASM_EMIT("vblendps $0x88, %%xmm2, %%xmm0, %%xmm0") /* xmm0 = r0" r2" i0" i2" */ \ |
| 219 |
__ASM_EMIT("vblendps $0x88, %%xmm3, %%xmm4, %%xmm4") \ |
| 220 |
__ASM_EMIT("vunpckhps %%xmm1, %%xmm0, %%xmm3") /* xmm3 = i0" i1" i2" i3" */ \ |
| 221 |
__ASM_EMIT("vunpcklps %%xmm1, %%xmm0, %%xmm2") /* xmm2 = r0" r1" r2" r3" */ \ |
| 222 |
__ASM_EMIT("vunpckhps %%xmm5, %%xmm4, %%xmm1") \ |
| 223 |
__ASM_EMIT("vunpcklps %%xmm5, %%xmm4, %%xmm0") \ |
| 224 |
/* Store */ \ |
| 225 |
__ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") \ |
| 226 |
__ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") \ |
| 227 |
__ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") \ |
| 228 |
__ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") \ |
| 229 |
__ASM_EMIT("4:") \ |
| 230 |
\ |
| 231 |
: [dst] "+r" (dst), [nb] "+r" (nb) \ |
| 232 |
: [FFT_A] "o" (FFT_A) \ |
| 233 |
: "cc", "memory", \ |
| 234 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", \ |
| 235 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" \ |
| 236 |
); |
| 237 |
|
| 238 |
#define FASTCONV_REVERSE_BUTTERFLY_BODY8(add_re, add_im, FMA_SEL) \ |
| 239 |
ARCH_X86_ASM \ |
| 240 |
( \ |
| 241 |
/* Prepare angle */ \ |
| 242 |
__ASM_EMIT("vmovaps 0x00(%[ak]), %%ymm6") /* ymm6 = x_re */ \ |
| 243 |
__ASM_EMIT("vmovaps 0x20(%[ak]), %%ymm7") /* ymm7 = x_im */ \ |
| 244 |
/* Start loop */ \ |
| 245 |
__ASM_EMIT("1:") \ |
| 246 |
__ASM_EMIT("vmovups 0x00(%[dst], %[off1]), %%ymm0") /* ymm0 = a_re */ \ |
| 247 |
__ASM_EMIT("vmovups 0x20(%[dst], %[off1]), %%ymm1") /* ymm1 = a_im */ \ |
| 248 |
__ASM_EMIT("vmovups 0x00(%[dst], %[off2]), %%ymm2") /* ymm2 = b_re */ \ |
| 249 |
__ASM_EMIT("vmovups 0x20(%[dst], %[off2]), %%ymm3") /* ymm3 = b_im */ \ |
| 250 |
/* Calculate complex multiplication */ \ |
| 251 |
__ASM_EMIT("vmulps %%ymm7, %%ymm2, %%ymm4") /* ymm4 = x_im * b_re */ \ |
| 252 |
__ASM_EMIT("vmulps %%ymm7, %%ymm3, %%ymm5") /* ymm5 = x_im * b_im */ \ |
| 253 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm2, %%ymm2", "")) /* ymm2 = x_re * b_re */ \ |
| 254 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm3, %%ymm3", "")) /* ymm3 = x_re * b_im */ \ |
| 255 |
__ASM_EMIT(FMA_SEL(add_re " %%ymm5, %%ymm2, %%ymm5", add_re " %%ymm6, %%ymm2, %%ymm5")) /* ymm5 = c_re = x_re * b_re +- x_im * b_im */ \ |
| 256 |
__ASM_EMIT(FMA_SEL(add_im " %%ymm4, %%ymm3, %%ymm4", add_im " %%ymm6, %%ymm3, %%ymm4")) /* ymm4 = c_im = x_re * b_im -+ x_im * b_re */ \ |
| 257 |
/* Perform butterfly */ \ |
| 258 |
__ASM_EMIT("vsubps %%ymm5, %%ymm0, %%ymm2") /* ymm2 = a_re - c_re */ \ |
| 259 |
__ASM_EMIT("vsubps %%ymm4, %%ymm1, %%ymm3") /* ymm3 = a_im - c_im */ \ |
| 260 |
__ASM_EMIT("vaddps %%ymm5, %%ymm0, %%ymm0") /* ymm0 = a_re + c_re */ \ |
| 261 |
__ASM_EMIT("vaddps %%ymm4, %%ymm1, %%ymm1") /* ymm1 = a_im + c_im */ \ |
| 262 |
/* Store values */ \ |
| 263 |
__ASM_EMIT("vmovups %%ymm0, 0x00(%[dst], %[off1])") \ |
| 264 |
__ASM_EMIT("vmovups %%ymm1, 0x20(%[dst], %[off1])") \ |
| 265 |
__ASM_EMIT("vmovups %%ymm2, 0x00(%[dst], %[off2])") \ |
| 266 |
__ASM_EMIT("vmovups %%ymm3, 0x20(%[dst], %[off2])") \ |
| 267 |
__ASM_EMIT("add $0x40, %[off1]") \ |
| 268 |
__ASM_EMIT("add $0x40, %[off2]") \ |
| 269 |
__ASM_EMIT32("subl $8, %[np]") \ |
| 270 |
__ASM_EMIT64("subq $8, %[np]") \ |
| 271 |
__ASM_EMIT("jz 2f") \ |
| 272 |
/* Rotate angle */ \ |
| 273 |
__ASM_EMIT("vmovaps 0x00(%[wk]), %%ymm4") /* xmm4 = w_re */ \ |
| 274 |
__ASM_EMIT("vmovaps 0x20(%[wk]), %%ymm5") /* xmm5 = w_im */ \ |
| 275 |
__ASM_EMIT("vmulps %%ymm5, %%ymm6, %%ymm2") /* ymm2 = w_im * x_re */ \ |
| 276 |
__ASM_EMIT("vmulps %%ymm5, %%ymm7, %%ymm3") /* ymm3 = w_im * x_im */ \ |
| 277 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm6, %%ymm6", "")) /* ymm6 = w_re * x_re */ \ |
| 278 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm7, %%ymm7", "")) /* ymm7 = w_re * x_im */ \ |
| 279 |
__ASM_EMIT(FMA_SEL("vsubps %%ymm3, %%ymm6, %%ymm6", "vfmsub132ps %%ymm4, %%ymm3, %%ymm6")) /* ymm6 = x_re' = w_re * x_re - w_im * x_im */ \ |
| 280 |
__ASM_EMIT(FMA_SEL("vaddps %%ymm2, %%ymm7, %%ymm7", "vfmadd132ps %%ymm4, %%ymm2, %%ymm7")) /* ymm7 = x_im' = w_re * x_im + w_im * x_re */ \ |
| 281 |
/* Repeat loop */ \ |
| 282 |
__ASM_EMIT("jmp 1b") \ |
| 283 |
__ASM_EMIT("2:") \ |
| 284 |
\ |
| 285 |
: [off1] "+r" (off1), [off2] "+r" (off2), [np] __ASM_ARG_RW(np) \ |
| 286 |
: [dst] "r" (dst), [ak] "r" (ak), [wk] "r" (wk) \ |
| 287 |
: "cc", "memory", \ |
| 288 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", \ |
| 289 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" \ |
| 290 |
); |
| 291 |
|
| 292 |
#define FASTCONV_REVERSE_BUTTERFLY_BODY_LAST(add_re, add_im, FMA_SEL, IF_ADD) \ |
| 293 |
size_t off; \ |
| 294 |
float norm = 0.5f / np; \ |
| 295 |
ARCH_X86_ASM \ |
| 296 |
( \ |
| 297 |
/* Prepare angle */ \ |
| 298 |
__ASM_EMIT("vbroadcastss %[norm], %%ymm1") /* ymm1 = k */ \ |
| 299 |
__ASM_EMIT("lea (,%[np], 4), %[off]") /* off = np * 8 */ \ |
| 300 |
__ASM_EMIT("vmovaps 0x00(%[ak]), %%ymm6") /* ymm6 = x_re */ \ |
| 301 |
__ASM_EMIT("vmovaps 0x20(%[ak]), %%ymm7") /* ymm7 = x_im */ \ |
| 302 |
__ASM_EMIT("vmovaps 0x00(%[wk]), %%ymm4") /* xmm4 = w_re */ \ |
| 303 |
__ASM_EMIT("vmovaps 0x20(%[wk]), %%ymm5") /* xmm5 = w_im */ \ |
| 304 |
/* Start loop */ \ |
| 305 |
__ASM_EMIT("1:") \ |
| 306 |
__ASM_EMIT("vmovups 0x00(%[src]), %%ymm0") /* ymm0 = a_re */ \ |
| 307 |
__ASM_EMIT("vmovups 0x00(%[src], %[off], 2), %%ymm2") /* ymm2 = b_re */ \ |
| 308 |
__ASM_EMIT("vmovups 0x20(%[src], %[off], 2), %%ymm3") /* ymm3 = b_im */ \ |
| 309 |
/* Calculate complex multiplication */ \ |
| 310 |
__ASM_EMIT("vmulps %%ymm7, %%ymm3, %%ymm3") /* ymm3 = x_im * b_im */ \ |
| 311 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm2, %%ymm2", "")) /* ymm2 = x_re * b_re */ \ |
| 312 |
__ASM_EMIT(FMA_SEL(add_re " %%ymm3, %%ymm2, %%ymm3", add_re " %%ymm6, %%ymm2, %%ymm3")) /* ymm3 = c_re = x_re * b_re +- x_im * b_im */ \ |
| 313 |
/* Perform butterfly */ \ |
| 314 |
__ASM_EMIT("vsubps %%ymm3, %%ymm0, %%ymm2") /* ymm2 = a_re - c_re */ \ |
| 315 |
__ASM_EMIT("vaddps %%ymm3, %%ymm0, %%ymm0") /* ymm0 = a_re + c_re */ \ |
| 316 |
__ASM_EMIT("vmulps %%ymm1, %%ymm2, %%ymm2") \ |
| 317 |
__ASM_EMIT("vmulps %%ymm1, %%ymm0, %%ymm0") \ |
| 318 |
/* Store values */ \ |
| 319 |
__ASM_EMIT(IF_ADD("vaddps 0x00(%[dst]), %%ymm0, %%ymm0")) \ |
| 320 |
__ASM_EMIT(IF_ADD("vaddps 0x00(%[dst], %[off]), %%ymm2, %%ymm2")) \ |
| 321 |
__ASM_EMIT("vmovups %%ymm0, 0x00(%[dst])") \ |
| 322 |
__ASM_EMIT("vmovups %%ymm2, 0x00(%[dst], %[off])") \ |
| 323 |
__ASM_EMIT("add $0x40, %[src]") \ |
| 324 |
__ASM_EMIT("add $0x20, %[dst]") \ |
| 325 |
__ASM_EMIT32("subl $8, %[np]") \ |
| 326 |
__ASM_EMIT64("subq $8, %[np]") \ |
| 327 |
__ASM_EMIT("jbe 2f") \ |
| 328 |
/* Rotate angle */ \ |
| 329 |
__ASM_EMIT("vmulps %%ymm5, %%ymm6, %%ymm2") /* ymm2 = w_im * x_re */ \ |
| 330 |
__ASM_EMIT("vmulps %%ymm5, %%ymm7, %%ymm3") /* ymm3 = w_im * x_im */ \ |
| 331 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm6, %%ymm6", "")) /* ymm6 = w_re * x_re */ \ |
| 332 |
__ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm7, %%ymm7", "")) /* ymm7 = w_re * x_im */ \ |
| 333 |
__ASM_EMIT(FMA_SEL("vsubps %%ymm3, %%ymm6, %%ymm6", "vfmsub132ps %%ymm4, %%ymm3, %%ymm6")) /* ymm6 = x_re' = w_re * x_re - w_im * x_im */ \ |
| 334 |
__ASM_EMIT(FMA_SEL("vaddps %%ymm2, %%ymm7, %%ymm7", "vfmadd132ps %%ymm4, %%ymm2, %%ymm7")) /* ymm7 = x_im' = w_re * x_im + w_im * x_re */ \ |
| 335 |
/* Repeat loop */ \ |
| 336 |
__ASM_EMIT("jmp 1b") \ |
| 337 |
__ASM_EMIT("2:") \ |
| 338 |
\ |
| 339 |
: [off] "=&r" (off), [np] __ASM_ARG_RW(np) \ |
| 340 |
: [dst] "r" (dst), [src] "r" (src), [ak] "r" (ak), [wk] "r" (wk), \ |
| 341 |
[norm] "o" (norm) \ |
| 342 |
: "cc", "memory", \ |
| 343 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", \ |
| 344 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" \ |
| 345 |
); |
| 346 |
|
| 347 |
#define FMA_OFF(a, b) a |
| 348 |
#define FMA_ON(a, b) b |
| 349 |
#define FASTCONV_SET(x) |
| 350 |
#define FASTCONV_ADD(x) x |
| 351 |
|
| 352 |
static inline void fastconv_direct_butterfly(float *dst, const float *ak, const float *wk, size_t pairs, size_t nb) |
| 353 |
{ |
| 354 |
size_t off1, off2, np; |
| 355 |
off1 = 0; |
| 356 |
size_t step = pairs << 3; |
| 357 |
for (size_t i=0; i<nb; ++i) |
| 358 |
{ |
| 359 |
off2 = off1 + step; |
| 360 |
np = pairs; |
| 361 |
|
| 362 |
FASTCONV_DIRECT_BUTTERFLY_BODY8("vaddps", "vsubps", FMA_OFF); |
| 363 |
|
| 364 |
off1 = off2; |
| 365 |
} |
| 366 |
} |
| 367 |
|
| 368 |
static inline void fastconv_direct_butterfly_last(float *dst, size_t nb) |
| 369 |
{ |
| 370 |
FASTCONV_DIRECT_BUTTERFLY_LAST("vaddps", "vsubps", FMA_OFF); |
| 371 |
} |
| 372 |
|
| 373 |
static inline void fastconv_direct_butterfly_fma3(float *dst, const float *ak, const float *wk, size_t pairs, size_t nb) |
| 374 |
{ |
| 375 |
size_t off1, off2, np; |
| 376 |
off1 = 0; |
| 377 |
size_t step = pairs << 3; |
| 378 |
for (size_t i=0; i<nb; ++i) |
| 379 |
{ |
| 380 |
off2 = off1 + step; |
| 381 |
np = pairs; |
| 382 |
|
| 383 |
FASTCONV_DIRECT_BUTTERFLY_BODY8("vfmadd132ps", "vfmsub132ps", FMA_ON); |
| 384 |
|
| 385 |
off1 = off2; |
| 386 |
} |
| 387 |
} |
| 388 |
|
| 389 |
static inline void fastconv_direct_butterfly_last_fma3(float *dst, size_t nb) |
| 390 |
{ |
| 391 |
FASTCONV_DIRECT_BUTTERFLY_LAST("vfmadd132ps", "vfmsub132ps", FMA_ON); |
| 392 |
} |
| 393 |
|
| 394 |
static inline void fastconv_reverse_butterfly(float *dst, const float *ak, const float *wk, size_t pairs, size_t nb) |
| 395 |
{ |
| 396 |
size_t off1, off2, np; |
| 397 |
off1 = 0; |
| 398 |
size_t step = pairs << 3; |
| 399 |
for (size_t i=0; i<nb; ++i) |
| 400 |
{ |
| 401 |
off2 = off1 + step; |
| 402 |
np = pairs; |
| 403 |
|
| 404 |
FASTCONV_REVERSE_BUTTERFLY_BODY8("vsubps", "vaddps", FMA_OFF); |
| 405 |
|
| 406 |
off1 = off2; |
| 407 |
} |
| 408 |
} |
| 409 |
|
| 410 |
static inline void fastconv_reverse_butterfly_last(float *dst, const float *src, const float *ak, const float *wk, size_t np) |
| 411 |
{ |
| 412 |
FASTCONV_REVERSE_BUTTERFLY_BODY_LAST("vsubps", "vaddps", FMA_OFF, FASTCONV_SET); |
| 413 |
} |
| 414 |
|
| 415 |
static inline void fastconv_reverse_butterfly_last_adding(float *dst, const float *src, const float *ak, const float *wk, size_t np) |
| 416 |
{ |
| 417 |
FASTCONV_REVERSE_BUTTERFLY_BODY_LAST("vsubps", "vaddps", FMA_OFF, FASTCONV_ADD); |
| 418 |
} |
| 419 |
|
| 420 |
static inline void fastconv_reverse_butterfly_fma3(float *dst, const float *ak, const float *wk, size_t pairs, size_t nb) |
| 421 |
{ |
| 422 |
size_t off1, off2, np; |
| 423 |
off1 = 0; |
| 424 |
size_t step = pairs << 3; |
| 425 |
for (size_t i=0; i<nb; ++i) |
| 426 |
{ |
| 427 |
off2 = off1 + step; |
| 428 |
np = pairs; |
| 429 |
|
| 430 |
FASTCONV_REVERSE_BUTTERFLY_BODY8("vfmsub231ps", "vfmadd231ps", FMA_ON); |
| 431 |
|
| 432 |
off1 = off2; |
| 433 |
} |
| 434 |
} |
| 435 |
|
| 436 |
static inline void fastconv_reverse_butterfly_last_fma3(float *dst, const float *src, const float *ak, const float *wk, size_t np) |
| 437 |
{ |
| 438 |
FASTCONV_REVERSE_BUTTERFLY_BODY_LAST("vfmsub231ps", "vfmadd231ps", FMA_ON, FASTCONV_SET); |
| 439 |
} |
| 440 |
|
| 441 |
static inline void fastconv_reverse_butterfly_last_adding_fma3(float *dst, const float *src, const float *ak, const float *wk, size_t np) |
| 442 |
{ |
| 443 |
FASTCONV_REVERSE_BUTTERFLY_BODY_LAST("vfmsub231ps", "vfmadd231ps", FMA_ON, FASTCONV_ADD); |
| 444 |
} |
| 445 |
|
| 446 |
#undef FASTCONV_DIRECT_BUTTERFLY_BODY8 |
| 447 |
#undef FASTCONV_DIRECT_BUTTERFLY_LAST |
| 448 |
#undef FASTCONV_REVERSE_BUTTERFLY_BODY8 |
| 449 |
#undef FASTCONV_REVERSE_BUTTERFLY_LAST |
| 450 |
#undef FASTCONV_SET |
| 451 |
#undef FASTCONV_ADD |
| 452 |
#undef FMA_OFF |
| 453 |
#undef FMA_ON |
| 454 |
} |
| 455 |
} |
| 456 |
|
| 457 |
#endif /* PRIVATE_DSP_ARCH_X86_AVX_FASTCONV_BUTTERFLY_H_ */ |