| 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_IMPL |
| 23 |
#error "This header should not be included directly" |
| 24 |
#endif /* PRIVATE_DSP_ARCH_X86_AVX_IMPL */ |
| 25 |
|
| 26 |
namespace lsp |
| 27 |
{ |
| 28 |
namespace avx |
| 29 |
{ |
| 30 |
static inline void FFT_PSCRAMBLE_SELF_DIRECT_NAME(float *dst, size_t rank) |
| 31 |
{ |
| 32 |
// Calculate number of items |
| 33 |
size_t items = (1 << rank) - 1; |
| 34 |
|
| 35 |
for (size_t i = 1; i < items; ++i) |
| 36 |
{ |
| 37 |
size_t j = reverse_bits(FFT_TYPE(i), rank); /* Reverse the order of the bits */ |
| 38 |
if (i >= j) |
| 39 |
continue; |
| 40 |
|
| 41 |
/* Copy the values from the reversed position */ |
| 42 |
/* Copy the values from the reversed position */ |
| 43 |
ARCH_X86_ASM |
| 44 |
( |
| 45 |
__ASM_EMIT("vmovlps (%[dst], %[i], 8), %%xmm0, %%xmm0") |
| 46 |
__ASM_EMIT("vmovlps (%[dst], %[j], 8), %%xmm1, %%xmm1") |
| 47 |
__ASM_EMIT("vmovlps %%xmm0, (%[dst], %[j], 8)") |
| 48 |
__ASM_EMIT("vmovlps %%xmm1, (%[dst], %[i], 8)") |
| 49 |
: |
| 50 |
: [dst] "r"(dst), [i] "r"(i), [j] "r"(j) |
| 51 |
: "memory", |
| 52 |
"%xmm0", "%xmm1" |
| 53 |
); |
| 54 |
} |
| 55 |
|
| 56 |
// Perform butterfly 8x |
| 57 |
items = 1 << (rank - 3); |
| 58 |
|
| 59 |
// Perform 4-element butterflies |
| 60 |
ARCH_X86_ASM |
| 61 |
( |
| 62 |
/* Loop 2x 4-element butterflies */ |
| 63 |
__ASM_EMIT("sub $2, %[items]") |
| 64 |
__ASM_EMIT("jb 2f") |
| 65 |
__ASM_EMIT("1:") |
| 66 |
/* Load data to registers */ |
| 67 |
__ASM_EMIT("vmovups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 i0 r1 i1 */ |
| 68 |
__ASM_EMIT("vmovups 0x10(%[dst]), %%xmm1") /* xmm1 = r2 i2 r3 i3 */ |
| 69 |
__ASM_EMIT("vmovups 0x20(%[dst]), %%xmm4") /* xmm4 = r4 i4 r5 i5 */ |
| 70 |
__ASM_EMIT("vmovups 0x30(%[dst]), %%xmm5") /* xmm5 = r6 i6 r7 i7 */ |
| 71 |
__ASM_EMIT("vinsertf128 $1, 0x40(%[dst]), %%ymm0, %%ymm0") /* ymm0 = r0 i0 r1 i1 r8 i8 r9 i9 */ |
| 72 |
__ASM_EMIT("vinsertf128 $1, 0x50(%[dst]), %%ymm1, %%ymm1") /* ymm1 = r2 i2 r3 i3 r10 i10 r11 i11 */ |
| 73 |
__ASM_EMIT("vinsertf128 $1, 0x60(%[dst]), %%ymm4, %%ymm4") /* ymm4 = r4 i4 r5 i5 r12 i12 r13 i13 */ |
| 74 |
__ASM_EMIT("vinsertf128 $1, 0x70(%[dst]), %%ymm5, %%ymm5") /* ymm5 = r6 i6 r7 i7 r14 i14 r15 i15 */ |
| 75 |
__ASM_EMIT("vshufps $0xdd, %%ymm1, %%ymm0, %%ymm2") /* ymm2 = i0 i1 i2 i3 i8 i9 i10 i11 */ |
| 76 |
__ASM_EMIT("vshufps $0x88, %%ymm1, %%ymm0, %%ymm0") /* ymm0 = r0 r1 r2 r3 r8 r9 r10 r11 */ |
| 77 |
__ASM_EMIT("vshufps $0xdd, %%ymm5, %%ymm4, %%ymm6") /* ymm6 = i4 i5 i6 i7 i12 i13 i14 i15 */ |
| 78 |
__ASM_EMIT("vshufps $0x88, %%ymm5, %%ymm4, %%ymm4") /* ymm4 = r4 r5 r6 r7 r12 r13 r14 r15 */ |
| 79 |
/* 1st-order 4x butterfly */ |
| 80 |
__ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' */ |
| 81 |
__ASM_EMIT("vhsubps %%ymm6, %%ymm2, %%ymm3") /* ymm3 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' */ |
| 82 |
__ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' */ |
| 83 |
__ASM_EMIT("vhaddps %%ymm6, %%ymm2, %%ymm2") /* ymm2 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' */ |
| 84 |
/* 2nd-order 4x butterfly */ |
| 85 |
__ASM_EMIT("vblendps $0xaa, %%ymm3, %%ymm1, %%ymm4") /* ymm4 = r1' i3' r5' i7' */ |
| 86 |
__ASM_EMIT("vblendps $0xaa, %%ymm1, %%ymm3, %%ymm5") /* ymm5 = i1' r3' i5' r7' */ |
| 87 |
__ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r3" r7" */ |
| 88 |
__ASM_EMIT("vhsubps %%ymm5, %%ymm2, %%ymm3") /* ymm3 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i1" i5" */ |
| 89 |
__ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r1" r5" */ |
| 90 |
__ASM_EMIT("vhaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i3" i7" */ |
| 91 |
__ASM_EMIT("vblendps $0xcc, %%ymm3, %%ymm2, %%ymm4") /* ymm4 = i0" i4" i1" i5" */ |
| 92 |
__ASM_EMIT("vblendps $0xcc, %%ymm2, %%ymm3, %%ymm5") /* ymm5 = i2" i6" i3" i7" */ |
| 93 |
__ASM_EMIT("vshufps $0x88, %%ymm1, %%ymm0, %%ymm2") /* ymm2 = r0" r1" r2" r3" */ |
| 94 |
__ASM_EMIT("vshufps $0xdd, %%ymm1, %%ymm0, %%ymm3") /* ymm3 = r4" r5" r6" r7" */ |
| 95 |
__ASM_EMIT("vshufps $0x88, %%ymm5, %%ymm4, %%ymm6") /* ymm6 = i0" i1" i2" i3" */ |
| 96 |
__ASM_EMIT("vshufps $0xdd, %%ymm5, %%ymm4, %%ymm7") /* ymm7 = i4" i5" i6" i7" */ |
| 97 |
/* 3rd-order 8x butterfly */ |
| 98 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm3, %%ymm4") /* ymm4 = x_im * b_re */ \ |
| 99 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm7, %%ymm5") /* ymm5 = x_im * b_im */ \ |
| 100 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%ymm3, %%ymm3", "")) /* ymm3 = x_re * b_re */ \ |
| 101 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%ymm7, %%ymm7", "")) /* ymm7 = x_re * b_im */ \ |
| 102 |
__ASM_EMIT(FFT_FMA("vaddps %%ymm5, %%ymm3, %%ymm5", "vfmadd231ps 0x00 + %[FFT_A], %%ymm3, %%ymm5")) /* ymm5 = c_re = x_re * b_re + x_im * b_im */ \ |
| 103 |
__ASM_EMIT(FFT_FMA("vsubps %%ymm4, %%ymm7, %%ymm4", "vfmsub231ps 0x00 + %[FFT_A], %%ymm7, %%ymm4")) /* ymm4 = c_im = x_re * b_im - x_im * b_re */ \ |
| 104 |
__ASM_EMIT("vsubps %%ymm5, %%ymm2, %%ymm0") /* ymm0 = a_re - c_re */ \ |
| 105 |
__ASM_EMIT("vsubps %%ymm4, %%ymm6, %%ymm1") /* ymm1 = a_im - c_im */ \ |
| 106 |
__ASM_EMIT("vaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = a_re + c_re */ \ |
| 107 |
__ASM_EMIT("vaddps %%ymm4, %%ymm6, %%ymm3") /* ymm3 = a_im + c_im */ \ |
| 108 |
/* Store */ |
| 109 |
__ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") |
| 110 |
__ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") |
| 111 |
__ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") |
| 112 |
__ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") |
| 113 |
__ASM_EMIT("vextractf128 $1, %%ymm2, 0x40(%[dst])") |
| 114 |
__ASM_EMIT("vextractf128 $1, %%ymm0, 0x50(%[dst])") |
| 115 |
__ASM_EMIT("vextractf128 $1, %%ymm3, 0x60(%[dst])") |
| 116 |
__ASM_EMIT("vextractf128 $1, %%ymm1, 0x70(%[dst])") |
| 117 |
/* Move pointers and repeat*/ |
| 118 |
__ASM_EMIT("add $0x80, %[dst]") |
| 119 |
__ASM_EMIT("sub $2, %[items]") |
| 120 |
__ASM_EMIT("jae 1b") |
| 121 |
__ASM_EMIT("2:") |
| 122 |
/* x4 scramble block */ |
| 123 |
__ASM_EMIT("add $1, %[items]") |
| 124 |
__ASM_EMIT("jl 4f") |
| 125 |
__ASM_EMIT("vmovups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 i0 r1 i1 */ |
| 126 |
__ASM_EMIT("vmovups 0x10(%[dst]), %%xmm1") /* xmm1 = r2 i2 r3 i3 */ |
| 127 |
__ASM_EMIT("vmovups 0x20(%[dst]), %%xmm4") /* xmm4 = r4 i4 r5 i5 */ |
| 128 |
__ASM_EMIT("vmovups 0x30(%[dst]), %%xmm5") /* xmm5 = r6 i6 r7 i7 */ |
| 129 |
__ASM_EMIT("vshufps $0xdd, %%xmm1, %%xmm0, %%xmm2") /* xmm2 = i0 i1 i2 i3 */ |
| 130 |
__ASM_EMIT("vshufps $0x88, %%xmm1, %%xmm0, %%xmm0") /* xmm0 = r0 r1 r2 r3 */ |
| 131 |
__ASM_EMIT("vshufps $0xdd, %%xmm5, %%xmm4, %%xmm6") /* xmm4 = i4 i5 i6 i7 */ |
| 132 |
__ASM_EMIT("vshufps $0x88, %%xmm5, %%xmm4, %%xmm4") /* xmm6 = r4 r5 r6 r7 */ |
| 133 |
/* 1st-order 4x butterfly */ |
| 134 |
__ASM_EMIT("vhsubps %%xmm4, %%xmm0, %%xmm1") /* xmm1 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' */ |
| 135 |
__ASM_EMIT("vhsubps %%xmm6, %%xmm2, %%xmm3") /* xmm3 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' */ |
| 136 |
__ASM_EMIT("vhaddps %%xmm4, %%xmm0, %%xmm0") /* xmm0 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' */ |
| 137 |
__ASM_EMIT("vhaddps %%xmm6, %%xmm2, %%xmm2") /* xmm2 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' */ |
| 138 |
/* 2nd-order 4x butterfly */ |
| 139 |
__ASM_EMIT("vblendps $0xaa, %%xmm3, %%xmm1, %%xmm4") /* xmm4 = r1' i3' r5' i7' */ |
| 140 |
__ASM_EMIT("vblendps $0xaa, %%xmm1, %%xmm3, %%xmm5") /* xmm5 = i1' r3' i5' r7' */ |
| 141 |
__ASM_EMIT("vhsubps %%xmm4, %%xmm0, %%xmm1") /* xmm1 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r3" r7" */ |
| 142 |
__ASM_EMIT("vhsubps %%xmm5, %%xmm2, %%xmm3") /* xmm3 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i1" i5" */ |
| 143 |
__ASM_EMIT("vhaddps %%xmm4, %%xmm0, %%xmm0") /* xmm0 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r1" r5" */ |
| 144 |
__ASM_EMIT("vhaddps %%xmm5, %%xmm2, %%xmm2") /* xmm2 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i3" i7" */ |
| 145 |
__ASM_EMIT("vblendps $0xcc, %%xmm3, %%xmm2, %%xmm4") /* xmm4 = i0" i4" i1" i5" */ |
| 146 |
__ASM_EMIT("vblendps $0xcc, %%xmm2, %%xmm3, %%xmm5") /* xmm5 = i2" i6" i3" i7" */ |
| 147 |
__ASM_EMIT("vshufps $0x88, %%xmm1, %%xmm0, %%xmm2") /* xmm2 = r0" r1" r2" r3" */ |
| 148 |
__ASM_EMIT("vshufps $0xdd, %%xmm1, %%xmm0, %%xmm3") /* xmm3 = r4" r5" r6" r7" */ |
| 149 |
__ASM_EMIT("vshufps $0x88, %%xmm5, %%xmm4, %%xmm6") /* xmm6 = i0" i1" i2" i3" */ |
| 150 |
__ASM_EMIT("vshufps $0xdd, %%xmm5, %%xmm4, %%xmm7") /* xmm7 = i4" i5" i6" i7" */ |
| 151 |
/* 3rd-order 8x butterfly */ |
| 152 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%xmm3, %%xmm4") /* xmm4 = x_im * b_re */ \ |
| 153 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%xmm7, %%xmm5") /* xmm5 = x_im * b_im */ \ |
| 154 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%xmm3, %%xmm3", "")) /* xmm3 = x_re * b_re */ \ |
| 155 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%xmm7, %%xmm7", "")) /* xmm7 = x_re * b_im */ \ |
| 156 |
__ASM_EMIT(FFT_FMA("vaddps %%xmm5, %%xmm3, %%xmm5", "vfmadd231ps 0x00 + %[FFT_A], %%xmm3, %%xmm5")) /* xmm5 = c_re = x_re * b_re + x_im * b_im */ \ |
| 157 |
__ASM_EMIT(FFT_FMA("vsubps %%xmm4, %%xmm7, %%xmm4", "vfmsub231ps 0x00 + %[FFT_A], %%xmm7, %%xmm4")) /* xmm4 = c_im = x_re * b_im - x_im * b_re */ \ |
| 158 |
__ASM_EMIT("vsubps %%xmm5, %%xmm2, %%xmm0") /* xmm0 = a_re - c_re */ \ |
| 159 |
__ASM_EMIT("vsubps %%xmm4, %%xmm6, %%xmm1") /* xmm1 = a_im - c_im */ \ |
| 160 |
__ASM_EMIT("vaddps %%xmm5, %%xmm2, %%xmm2") /* xmm2 = a_re + c_re */ \ |
| 161 |
__ASM_EMIT("vaddps %%xmm4, %%xmm6, %%xmm3") /* xmm3 = a_im + c_im */ \ |
| 162 |
/* Store */ |
| 163 |
__ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") |
| 164 |
__ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") |
| 165 |
__ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") |
| 166 |
__ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") |
| 167 |
__ASM_EMIT("4:") |
| 168 |
|
| 169 |
: [dst] "+r" (dst), [items] "+r" (items) |
| 170 |
: [FFT_A] "o" (FFT_A) |
| 171 |
: "cc", "memory", |
| 172 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 173 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 174 |
); |
| 175 |
} |
| 176 |
|
| 177 |
static inline void FFT_PSCRAMBLE_SELF_REVERSE_NAME(float *dst, size_t rank) |
| 178 |
{ |
| 179 |
// Calculate number of items |
| 180 |
size_t items = (1 << rank) - 1; |
| 181 |
|
| 182 |
for (size_t i = 1; i < items; ++i) |
| 183 |
{ |
| 184 |
size_t j = reverse_bits(FFT_TYPE(i), rank); /* Reverse the order of the bits */ |
| 185 |
if (i >= j) |
| 186 |
continue; |
| 187 |
|
| 188 |
/* Copy the values from the reversed position */ |
| 189 |
/* Copy the values from the reversed position */ |
| 190 |
ARCH_X86_ASM |
| 191 |
( |
| 192 |
__ASM_EMIT("vmovlps (%[dst], %[i], 8), %%xmm0, %%xmm0") |
| 193 |
__ASM_EMIT("vmovlps (%[dst], %[j], 8), %%xmm1, %%xmm1") |
| 194 |
__ASM_EMIT("vmovlps %%xmm0, (%[dst], %[j], 8)") |
| 195 |
__ASM_EMIT("vmovlps %%xmm1, (%[dst], %[i], 8)") |
| 196 |
: |
| 197 |
: [dst] "r"(dst), [i] "r"(i), [j] "r"(j) |
| 198 |
: "memory", |
| 199 |
"%xmm0", "%xmm1" |
| 200 |
); |
| 201 |
} |
| 202 |
|
| 203 |
// Perform butterfly 8x |
| 204 |
items = 1 << (rank - 3); |
| 205 |
|
| 206 |
// Perform 4-element butterflies |
| 207 |
ARCH_X86_ASM |
| 208 |
( |
| 209 |
/* Loop 2x 4-element butterflies */ |
| 210 |
__ASM_EMIT("sub $2, %[items]") |
| 211 |
__ASM_EMIT("jb 2f") |
| 212 |
__ASM_EMIT("1:") |
| 213 |
/* Load data to registers */ |
| 214 |
__ASM_EMIT("vmovups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 i0 r1 i1 */ |
| 215 |
__ASM_EMIT("vmovups 0x10(%[dst]), %%xmm1") /* xmm1 = r2 i2 r3 i3 */ |
| 216 |
__ASM_EMIT("vmovups 0x20(%[dst]), %%xmm4") /* xmm4 = r4 i4 r5 i5 */ |
| 217 |
__ASM_EMIT("vmovups 0x30(%[dst]), %%xmm5") /* xmm5 = r6 i6 r7 i7 */ |
| 218 |
__ASM_EMIT("vinsertf128 $1, 0x40(%[dst]), %%ymm0, %%ymm0") /* ymm0 = r0 i0 r1 i1 r8 i8 r9 i9 */ |
| 219 |
__ASM_EMIT("vinsertf128 $1, 0x50(%[dst]), %%ymm1, %%ymm1") /* ymm1 = r2 i2 r3 i3 r10 i10 r11 i11 */ |
| 220 |
__ASM_EMIT("vinsertf128 $1, 0x60(%[dst]), %%ymm4, %%ymm4") /* ymm4 = r4 i4 r5 i5 r12 i12 r13 i13 */ |
| 221 |
__ASM_EMIT("vinsertf128 $1, 0x70(%[dst]), %%ymm5, %%ymm5") /* ymm5 = r6 i6 r7 i7 r14 i14 r15 i15 */ |
| 222 |
__ASM_EMIT("vshufps $0xdd, %%ymm1, %%ymm0, %%ymm2") /* ymm2 = i0 i1 i2 i3 i8 i9 i10 i11 */ |
| 223 |
__ASM_EMIT("vshufps $0x88, %%ymm1, %%ymm0, %%ymm0") /* ymm0 = r0 r1 r2 r3 r8 r9 r10 r11 */ |
| 224 |
__ASM_EMIT("vshufps $0xdd, %%ymm5, %%ymm4, %%ymm6") /* ymm6 = i4 i5 i6 i7 i12 i13 i14 i15 */ |
| 225 |
__ASM_EMIT("vshufps $0x88, %%ymm5, %%ymm4, %%ymm4") /* ymm4 = r4 r5 r6 r7 r12 r13 r14 r15 */ |
| 226 |
/* 1st-order 4x butterfly */ |
| 227 |
__ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' */ |
| 228 |
__ASM_EMIT("vhsubps %%ymm6, %%ymm2, %%ymm3") /* ymm3 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' */ |
| 229 |
__ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' */ |
| 230 |
__ASM_EMIT("vhaddps %%ymm6, %%ymm2, %%ymm2") /* ymm2 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' */ |
| 231 |
/* 2nd-order 4x butterfly */ |
| 232 |
__ASM_EMIT("vblendps $0xaa, %%ymm3, %%ymm1, %%ymm4") /* ymm4 = r1' i3' r5' i7' */ |
| 233 |
__ASM_EMIT("vblendps $0xaa, %%ymm1, %%ymm3, %%ymm5") /* ymm5 = i1' r3' i5' r7' */ |
| 234 |
__ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r1" r5" */ |
| 235 |
__ASM_EMIT("vhsubps %%ymm5, %%ymm2, %%ymm3") /* ymm3 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i3" i7" */ |
| 236 |
__ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r3" r7" */ |
| 237 |
__ASM_EMIT("vhaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i1" i5" */ |
| 238 |
__ASM_EMIT("vblendps $0xcc, %%ymm1, %%ymm0, %%ymm4") /* ymm4 = r0" i4" r1" r5" */ |
| 239 |
__ASM_EMIT("vblendps $0xcc, %%ymm0, %%ymm1, %%ymm5") /* ymm5 = r2" r6" r3" r7" */ |
| 240 |
__ASM_EMIT("vshufps $0x88, %%ymm3, %%ymm2, %%ymm6") /* ymm6 = i0" i1" i2" i3" */ |
| 241 |
__ASM_EMIT("vshufps $0xdd, %%ymm3, %%ymm2, %%ymm7") /* ymm7 = i4" i5" i6" i7" */ |
| 242 |
__ASM_EMIT("vshufps $0x88, %%ymm5, %%ymm4, %%ymm2") /* ymm2 = r0" r1" r2" r3" */ |
| 243 |
__ASM_EMIT("vshufps $0xdd, %%ymm5, %%ymm4, %%ymm3") /* ymm3 = r4" r5" r6" r7" */ |
| 244 |
/* 3rd-order 8x butterfly */ |
| 245 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm3, %%ymm4") /* ymm4 = x_im * b_re */ \ |
| 246 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm7, %%ymm5") /* ymm5 = x_im * b_im */ \ |
| 247 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%ymm3, %%ymm3", "")) /* ymm3 = x_re * b_re */ \ |
| 248 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%ymm7, %%ymm7", "")) /* ymm7 = x_re * b_im */ \ |
| 249 |
__ASM_EMIT(FFT_FMA("vsubps %%ymm5, %%ymm3, %%ymm5", "vfmsub231ps 0x00 + %[FFT_A], %%ymm3, %%ymm5")) /* ymm5 = c_re = x_re * b_re - x_im * b_im */ \ |
| 250 |
__ASM_EMIT(FFT_FMA("vaddps %%ymm4, %%ymm7, %%ymm4", "vfmadd231ps 0x00 + %[FFT_A], %%ymm7, %%ymm4")) /* ymm4 = c_im = x_re * b_im + x_im * b_re */ \ |
| 251 |
__ASM_EMIT("vsubps %%ymm5, %%ymm2, %%ymm0") /* ymm0 = a_re - c_re */ \ |
| 252 |
__ASM_EMIT("vsubps %%ymm4, %%ymm6, %%ymm1") /* ymm1 = a_im - c_im */ \ |
| 253 |
__ASM_EMIT("vaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = a_re + c_re */ \ |
| 254 |
__ASM_EMIT("vaddps %%ymm4, %%ymm6, %%ymm3") /* ymm3 = a_im + c_im */ \ |
| 255 |
/* Store */ |
| 256 |
__ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") |
| 257 |
__ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") |
| 258 |
__ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") |
| 259 |
__ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") |
| 260 |
__ASM_EMIT("vextractf128 $1, %%ymm2, 0x40(%[dst])") |
| 261 |
__ASM_EMIT("vextractf128 $1, %%ymm0, 0x50(%[dst])") |
| 262 |
__ASM_EMIT("vextractf128 $1, %%ymm3, 0x60(%[dst])") |
| 263 |
__ASM_EMIT("vextractf128 $1, %%ymm1, 0x70(%[dst])") |
| 264 |
/* Move pointers and repeat*/ |
| 265 |
__ASM_EMIT("add $0x80, %[dst]") |
| 266 |
__ASM_EMIT("sub $2, %[items]") |
| 267 |
__ASM_EMIT("jae 1b") |
| 268 |
__ASM_EMIT("2:") |
| 269 |
/* x4 scramble block */ |
| 270 |
__ASM_EMIT("add $1, %[items]") |
| 271 |
__ASM_EMIT("jl 4f") |
| 272 |
__ASM_EMIT("vmovups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 i0 r1 i1 */ |
| 273 |
__ASM_EMIT("vmovups 0x10(%[dst]), %%xmm1") /* xmm1 = r2 i2 r3 i3 */ |
| 274 |
__ASM_EMIT("vmovups 0x20(%[dst]), %%xmm4") /* xmm4 = r4 i4 r5 i5 */ |
| 275 |
__ASM_EMIT("vmovups 0x30(%[dst]), %%xmm5") /* xmm5 = r6 i6 r7 i7 */ |
| 276 |
__ASM_EMIT("vshufps $0xdd, %%xmm1, %%xmm0, %%xmm2") /* xmm2 = i0 i1 i2 i3 */ |
| 277 |
__ASM_EMIT("vshufps $0x88, %%xmm1, %%xmm0, %%xmm0") /* xmm0 = r0 r1 r2 r3 */ |
| 278 |
__ASM_EMIT("vshufps $0xdd, %%xmm5, %%xmm4, %%xmm6") /* xmm4 = i4 i5 i6 i7 */ |
| 279 |
__ASM_EMIT("vshufps $0x88, %%xmm5, %%xmm4, %%xmm4") /* xmm6 = r4 r5 r6 r7 */ |
| 280 |
/* 1st-order 4x butterfly */ |
| 281 |
__ASM_EMIT("vhsubps %%xmm4, %%xmm0, %%xmm1") /* xmm1 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' */ |
| 282 |
__ASM_EMIT("vhsubps %%xmm6, %%xmm2, %%xmm3") /* xmm3 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' */ |
| 283 |
__ASM_EMIT("vhaddps %%xmm4, %%xmm0, %%xmm0") /* xmm0 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' */ |
| 284 |
__ASM_EMIT("vhaddps %%xmm6, %%xmm2, %%xmm2") /* xmm2 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' */ |
| 285 |
/* 2nd-order 4x butterfly */ |
| 286 |
__ASM_EMIT("vblendps $0xaa, %%xmm3, %%xmm1, %%xmm4") /* xmm4 = r1' i3' r5' i7' */ |
| 287 |
__ASM_EMIT("vblendps $0xaa, %%xmm1, %%xmm3, %%xmm5") /* xmm5 = i1' r3' i5' r7' */ |
| 288 |
__ASM_EMIT("vhsubps %%xmm4, %%xmm0, %%xmm1") /* xmm1 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r1" r5" */ |
| 289 |
__ASM_EMIT("vhsubps %%xmm5, %%xmm2, %%xmm3") /* xmm3 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i3" i7" */ |
| 290 |
__ASM_EMIT("vhaddps %%xmm4, %%xmm0, %%xmm0") /* xmm0 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r3" r7" */ |
| 291 |
__ASM_EMIT("vhaddps %%xmm5, %%xmm2, %%xmm2") /* xmm2 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i1" i5" */ |
| 292 |
__ASM_EMIT("vblendps $0xcc, %%xmm1, %%xmm0, %%xmm4") /* xmm4 = r0" i4" r1" r5" */ |
| 293 |
__ASM_EMIT("vblendps $0xcc, %%xmm0, %%xmm1, %%xmm5") /* xmm5 = r2" r6" r3" r7" */ |
| 294 |
__ASM_EMIT("vshufps $0x88, %%xmm3, %%xmm2, %%xmm6") /* xmm6 = i0" i1" i2" i3" */ |
| 295 |
__ASM_EMIT("vshufps $0xdd, %%xmm3, %%xmm2, %%xmm7") /* xmm7 = i4" i5" i6" i7" */ |
| 296 |
__ASM_EMIT("vshufps $0x88, %%xmm5, %%xmm4, %%xmm2") /* xmm2 = r0" r1" r2" r3" */ |
| 297 |
__ASM_EMIT("vshufps $0xdd, %%xmm5, %%xmm4, %%xmm3") /* xmm3 = r4" r5" r6" r7" */ |
| 298 |
/* 3rd-order 8x butterfly */ |
| 299 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%xmm3, %%xmm4") /* xmm4 = x_im * b_re */ \ |
| 300 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%xmm7, %%xmm5") /* xmm5 = x_im * b_im */ \ |
| 301 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%xmm3, %%xmm3", "")) /* xmm3 = x_re * b_re */ \ |
| 302 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%xmm7, %%xmm7", "")) /* xmm7 = x_re * b_im */ \ |
| 303 |
__ASM_EMIT(FFT_FMA("vsubps %%xmm5, %%xmm3, %%xmm5", "vfmsub231ps 0x00 + %[FFT_A], %%xmm3, %%xmm5")) /* xmm5 = c_re = x_re * b_re - x_im * b_im */ \ |
| 304 |
__ASM_EMIT(FFT_FMA("vaddps %%xmm4, %%xmm7, %%xmm4", "vfmadd231ps 0x00 + %[FFT_A], %%xmm7, %%xmm4")) /* xmm4 = c_im = x_re * b_im + x_im * b_re */ \ |
| 305 |
__ASM_EMIT("vsubps %%xmm5, %%xmm2, %%xmm0") /* xmm0 = a_re - c_re */ \ |
| 306 |
__ASM_EMIT("vsubps %%xmm4, %%xmm6, %%xmm1") /* xmm1 = a_im - c_im */ \ |
| 307 |
__ASM_EMIT("vaddps %%xmm5, %%xmm2, %%xmm2") /* xmm2 = a_re + c_re */ \ |
| 308 |
__ASM_EMIT("vaddps %%xmm4, %%xmm6, %%xmm3") /* xmm3 = a_im + c_im */ \ |
| 309 |
/* Store */ |
| 310 |
__ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") |
| 311 |
__ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") |
| 312 |
__ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") |
| 313 |
__ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") |
| 314 |
__ASM_EMIT("4:") |
| 315 |
|
| 316 |
: [dst] "+r" (dst), [items] "+r" (items) |
| 317 |
: [FFT_A] "o" (FFT_A) |
| 318 |
: "cc", "memory", |
| 319 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 320 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 321 |
); |
| 322 |
} |
| 323 |
|
| 324 |
static inline void FFT_PSCRAMBLE_COPY_DIRECT_NAME(float *dst, const float *src, size_t rank) |
| 325 |
{ |
| 326 |
size_t regs = 1 << rank; |
| 327 |
|
| 328 |
for (size_t i=0; i<regs; ++i) |
| 329 |
{ |
| 330 |
size_t index = reverse_bits(FFT_TYPE(i), rank); |
| 331 |
|
| 332 |
ARCH_X86_ASM |
| 333 |
( |
| 334 |
/* Load scalar values */ |
| 335 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm0, %%xmm0") /* xmm0 = r0 i0 x x */ |
| 336 |
__ASM_EMIT("add %[regs], %[index]") |
| 337 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm2, %%xmm2") /* xmm2 = r8 i8 x x */ |
| 338 |
__ASM_EMIT("add %[regs], %[index]") |
| 339 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm4, %%xmm4") /* xmm4 = r4 i4 x x */ |
| 340 |
__ASM_EMIT("add %[regs], %[index]") |
| 341 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm6, %%xmm6") /* xmm6 = r12 i12 x x */ |
| 342 |
__ASM_EMIT("add %[regs], %[index]") |
| 343 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm1, %%xmm1") /* xmm1 = r2 i2 x x */ |
| 344 |
__ASM_EMIT("add %[regs], %[index]") |
| 345 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm3, %%xmm3") /* xmm3 = r10 i10 x x */ |
| 346 |
__ASM_EMIT("add %[regs], %[index]") |
| 347 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm5, %%xmm5") /* xmm5 = r6 i6 x x */ |
| 348 |
__ASM_EMIT("add %[regs], %[index]") |
| 349 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm7, %%xmm7") /* xmm7 = r14 i14 x x */ |
| 350 |
__ASM_EMIT("add %[regs], %[index]") |
| 351 |
|
| 352 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm0, %%xmm0") /* xmm0 = r0 i0 r1 i1 */ |
| 353 |
__ASM_EMIT("add %[regs], %[index]") |
| 354 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm2, %%xmm2") /* xmm2 = r8 i8 r9 i9 */ |
| 355 |
__ASM_EMIT("add %[regs], %[index]") |
| 356 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm4, %%xmm4") /* xmm4 = r4 i4 r5 i5 */ |
| 357 |
__ASM_EMIT("add %[regs], %[index]") |
| 358 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm6, %%xmm6") /* xmm6 = r12 i12 r13 r14 */ |
| 359 |
__ASM_EMIT("add %[regs], %[index]") |
| 360 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm1, %%xmm1") /* xmm1 = r2 i2 r3 i3 */ |
| 361 |
__ASM_EMIT("add %[regs], %[index]") |
| 362 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm3, %%xmm3") /* xmm3 = r10 i10 r11 i11 */ |
| 363 |
__ASM_EMIT("add %[regs], %[index]") |
| 364 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm5, %%xmm5") /* xmm5 = r6 i6 r7 i7 */ |
| 365 |
__ASM_EMIT("add %[regs], %[index]") |
| 366 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm7, %%xmm7") /* xmm7 = r14 i14 r15 i15 */ |
| 367 |
__ASM_EMIT("add %[regs], %[index]") |
| 368 |
|
| 369 |
__ASM_EMIT("vinsertf128 $1, %%xmm2, %%ymm0, %%ymm0") /* ymm0 = r0 i0 r1 i1 r8 i8 r9 i9 */ |
| 370 |
__ASM_EMIT("vinsertf128 $1, %%xmm3, %%ymm1, %%ymm1") /* ymm1 = r2 i2 r3 i3 r10 i10 r11 i11 */ |
| 371 |
__ASM_EMIT("vinsertf128 $1, %%xmm6, %%ymm4, %%ymm4") /* ymm4 = r4 i4 r5 i5 r12 i12 r13 r14 */ |
| 372 |
__ASM_EMIT("vinsertf128 $1, %%xmm7, %%ymm5, %%ymm5") /* ymm5 = r6 i6 r7 i7 r14 i14 r15 i15 */ |
| 373 |
__ASM_EMIT("vshufps $0xdd, %%ymm1, %%ymm0, %%ymm2") /* ymm2 = i0 i1 i2 i3 i8 i9 i10 i11 */ |
| 374 |
__ASM_EMIT("vshufps $0x88, %%ymm1, %%ymm0, %%ymm0") /* ymm0 = r0 r1 r2 r3 r8 r9 r10 r11 */ |
| 375 |
__ASM_EMIT("vshufps $0xdd, %%ymm5, %%ymm4, %%ymm6") /* ymm6 = i4 i5 i6 i7 i12 i13 i14 i15 */ |
| 376 |
__ASM_EMIT("vshufps $0x88, %%ymm5, %%ymm4, %%ymm4") /* ymm4 = r4 r5 r6 r7 r12 r13 r14 r15 */ |
| 377 |
/* 1st-order 4x butterfly */ |
| 378 |
__ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' */ |
| 379 |
__ASM_EMIT("vhsubps %%ymm6, %%ymm2, %%ymm3") /* ymm3 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' */ |
| 380 |
__ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' */ |
| 381 |
__ASM_EMIT("vhaddps %%ymm6, %%ymm2, %%ymm2") /* ymm2 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' */ |
| 382 |
/* 2nd-order 4x butterfly */ |
| 383 |
__ASM_EMIT("vblendps $0xaa, %%ymm3, %%ymm1, %%ymm4") /* ymm4 = r1' i3' r5' i7' */ |
| 384 |
__ASM_EMIT("vblendps $0xaa, %%ymm1, %%ymm3, %%ymm5") /* ymm5 = i1' r3' i5' r7' */ |
| 385 |
__ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r3" r7" */ |
| 386 |
__ASM_EMIT("vhsubps %%ymm5, %%ymm2, %%ymm3") /* ymm3 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i1" i5" */ |
| 387 |
__ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r1" r5" */ |
| 388 |
__ASM_EMIT("vhaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i3" i7" */ |
| 389 |
__ASM_EMIT("vblendps $0xcc, %%ymm3, %%ymm2, %%ymm4") /* ymm4 = i0" i4" i1" i5" */ |
| 390 |
__ASM_EMIT("vblendps $0xcc, %%ymm2, %%ymm3, %%ymm5") /* ymm5 = i2" i6" i3" i7" */ |
| 391 |
__ASM_EMIT("vshufps $0x88, %%ymm1, %%ymm0, %%ymm2") /* ymm2 = r0" r1" r2" r3" */ |
| 392 |
__ASM_EMIT("vshufps $0xdd, %%ymm1, %%ymm0, %%ymm3") /* ymm3 = r4" r5" r6" r7" */ |
| 393 |
__ASM_EMIT("vshufps $0x88, %%ymm5, %%ymm4, %%ymm6") /* ymm6 = i0" i1" i2" i3" */ |
| 394 |
__ASM_EMIT("vshufps $0xdd, %%ymm5, %%ymm4, %%ymm7") /* ymm7 = i4" i5" i6" i7" */ |
| 395 |
/* 3rd-order 8x butterfly */ |
| 396 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm3, %%ymm4") /* ymm4 = x_im * b_re */ \ |
| 397 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm7, %%ymm5") /* ymm5 = x_im * b_im */ \ |
| 398 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%ymm3, %%ymm3", "")) /* ymm3 = x_re * b_re */ \ |
| 399 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%ymm7, %%ymm7", "")) /* ymm7 = x_re * b_im */ \ |
| 400 |
__ASM_EMIT(FFT_FMA("vaddps %%ymm5, %%ymm3, %%ymm5", "vfmadd231ps 0x00 + %[FFT_A], %%ymm3, %%ymm5")) /* ymm5 = c_re = x_re * b_re + x_im * b_im */ \ |
| 401 |
__ASM_EMIT(FFT_FMA("vsubps %%ymm4, %%ymm7, %%ymm4", "vfmsub231ps 0x00 + %[FFT_A], %%ymm7, %%ymm4")) /* ymm4 = c_im = x_re * b_im - x_im * b_re */ \ |
| 402 |
__ASM_EMIT("vsubps %%ymm5, %%ymm2, %%ymm0") /* ymm0 = a_re - c_re */ \ |
| 403 |
__ASM_EMIT("vsubps %%ymm4, %%ymm6, %%ymm1") /* ymm1 = a_im - c_im */ \ |
| 404 |
__ASM_EMIT("vaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = a_re + c_re */ \ |
| 405 |
__ASM_EMIT("vaddps %%ymm4, %%ymm6, %%ymm3") /* ymm3 = a_im + c_im */ \ |
| 406 |
/* Store */ |
| 407 |
__ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") |
| 408 |
__ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") |
| 409 |
__ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") |
| 410 |
__ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") |
| 411 |
__ASM_EMIT("vextractf128 $1, %%ymm2, 0x40(%[dst])") |
| 412 |
__ASM_EMIT("vextractf128 $1, %%ymm0, 0x50(%[dst])") |
| 413 |
__ASM_EMIT("vextractf128 $1, %%ymm3, 0x60(%[dst])") |
| 414 |
__ASM_EMIT("vextractf128 $1, %%ymm1, 0x70(%[dst])") |
| 415 |
__ASM_EMIT("add $0x80, %[dst]") |
| 416 |
|
| 417 |
: [dst] "+r" (dst), [index] "+r"(index) |
| 418 |
: [src] "r" (src), [regs] "r" (regs), |
| 419 |
[FFT_A] "o" (FFT_A) |
| 420 |
: "cc", "memory", |
| 421 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 422 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 423 |
); |
| 424 |
} |
| 425 |
} |
| 426 |
|
| 427 |
static inline void FFT_PSCRAMBLE_COPY_REVERSE_NAME(float *dst, const float *src, size_t rank) |
| 428 |
{ |
| 429 |
size_t regs = 1 << rank; |
| 430 |
|
| 431 |
for (size_t i=0; i<regs; ++i) |
| 432 |
{ |
| 433 |
size_t index = reverse_bits(FFT_TYPE(i), rank); |
| 434 |
|
| 435 |
ARCH_X86_ASM |
| 436 |
( |
| 437 |
/* Load scalar values */ |
| 438 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm0, %%xmm0") /* xmm0 = r0 i0 x x */ |
| 439 |
__ASM_EMIT("add %[regs], %[index]") |
| 440 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm2, %%xmm2") /* xmm2 = r8 i8 x x */ |
| 441 |
__ASM_EMIT("add %[regs], %[index]") |
| 442 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm4, %%xmm4") /* xmm4 = r4 i4 x x */ |
| 443 |
__ASM_EMIT("add %[regs], %[index]") |
| 444 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm6, %%xmm6") /* xmm6 = r12 i12 x x */ |
| 445 |
__ASM_EMIT("add %[regs], %[index]") |
| 446 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm1, %%xmm1") /* xmm1 = r2 i2 x x */ |
| 447 |
__ASM_EMIT("add %[regs], %[index]") |
| 448 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm3, %%xmm3") /* xmm3 = r10 i10 x x */ |
| 449 |
__ASM_EMIT("add %[regs], %[index]") |
| 450 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm5, %%xmm5") /* xmm5 = r6 i6 x x */ |
| 451 |
__ASM_EMIT("add %[regs], %[index]") |
| 452 |
__ASM_EMIT("vmovlps (%[src], %[index], 8), %%xmm7, %%xmm7") /* xmm7 = r14 i14 x x */ |
| 453 |
__ASM_EMIT("add %[regs], %[index]") |
| 454 |
|
| 455 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm0, %%xmm0") /* xmm0 = r0 i0 r1 i1 */ |
| 456 |
__ASM_EMIT("add %[regs], %[index]") |
| 457 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm2, %%xmm2") /* xmm2 = r8 i8 r9 i9 */ |
| 458 |
__ASM_EMIT("add %[regs], %[index]") |
| 459 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm4, %%xmm4") /* xmm4 = r4 i4 r5 i5 */ |
| 460 |
__ASM_EMIT("add %[regs], %[index]") |
| 461 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm6, %%xmm6") /* xmm6 = r12 i12 r13 r14 */ |
| 462 |
__ASM_EMIT("add %[regs], %[index]") |
| 463 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm1, %%xmm1") /* xmm1 = r2 i2 r3 i3 */ |
| 464 |
__ASM_EMIT("add %[regs], %[index]") |
| 465 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm3, %%xmm3") /* xmm3 = r10 i10 r11 i11 */ |
| 466 |
__ASM_EMIT("add %[regs], %[index]") |
| 467 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm5, %%xmm5") /* xmm5 = r6 i6 r7 i7 */ |
| 468 |
__ASM_EMIT("add %[regs], %[index]") |
| 469 |
__ASM_EMIT("vmovhps (%[src], %[index], 8), %%xmm7, %%xmm7") /* xmm7 = r14 i14 r15 i15 */ |
| 470 |
__ASM_EMIT("add %[regs], %[index]") |
| 471 |
|
| 472 |
__ASM_EMIT("vinsertf128 $1, %%xmm2, %%ymm0, %%ymm0") /* ymm0 = r0 i0 r1 i1 r8 i8 r9 i9 */ |
| 473 |
__ASM_EMIT("vinsertf128 $1, %%xmm3, %%ymm1, %%ymm1") /* ymm1 = r2 i2 r3 i3 r10 i10 r11 i11 */ |
| 474 |
__ASM_EMIT("vinsertf128 $1, %%xmm6, %%ymm4, %%ymm4") /* ymm4 = r4 i4 r5 i5 r12 i12 r13 r14 */ |
| 475 |
__ASM_EMIT("vinsertf128 $1, %%xmm7, %%ymm5, %%ymm5") /* ymm5 = r6 i6 r7 i7 r14 i14 r15 i15 */ |
| 476 |
__ASM_EMIT("vshufps $0xdd, %%ymm1, %%ymm0, %%ymm2") /* ymm2 = i0 i1 i2 i3 i8 i9 i10 i11 */ |
| 477 |
__ASM_EMIT("vshufps $0x88, %%ymm1, %%ymm0, %%ymm0") /* ymm0 = r0 r1 r2 r3 r8 r9 r10 r11 */ |
| 478 |
__ASM_EMIT("vshufps $0xdd, %%ymm5, %%ymm4, %%ymm6") /* ymm6 = i4 i5 i6 i7 i12 i13 i14 i15 */ |
| 479 |
__ASM_EMIT("vshufps $0x88, %%ymm5, %%ymm4, %%ymm4") /* ymm4 = r4 r5 r6 r7 r12 r13 r14 r15 */ |
| 480 |
/* 1st-order 4x butterfly */ |
| 481 |
__ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' */ |
| 482 |
__ASM_EMIT("vhsubps %%ymm6, %%ymm2, %%ymm3") /* ymm3 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' */ |
| 483 |
__ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' */ |
| 484 |
__ASM_EMIT("vhaddps %%ymm6, %%ymm2, %%ymm2") /* ymm2 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' */ |
| 485 |
/* 2nd-order 4x butterfly */ |
| 486 |
__ASM_EMIT("vblendps $0xaa, %%ymm3, %%ymm1, %%ymm4") /* ymm4 = r1' i3' r5' i7' */ |
| 487 |
__ASM_EMIT("vblendps $0xaa, %%ymm1, %%ymm3, %%ymm5") /* ymm5 = i1' r3' i5' r7' */ |
| 488 |
__ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r1" r5" */ |
| 489 |
__ASM_EMIT("vhsubps %%ymm5, %%ymm2, %%ymm3") /* ymm3 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i3" i7" */ |
| 490 |
__ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r3" r7" */ |
| 491 |
__ASM_EMIT("vhaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i1" i5" */ |
| 492 |
__ASM_EMIT("vblendps $0xcc, %%ymm1, %%ymm0, %%ymm4") /* ymm4 = r0" i4" r1" r5" */ |
| 493 |
__ASM_EMIT("vblendps $0xcc, %%ymm0, %%ymm1, %%ymm5") /* ymm5 = r2" r6" r3" r7" */ |
| 494 |
__ASM_EMIT("vshufps $0x88, %%ymm3, %%ymm2, %%ymm6") /* ymm6 = i0" i1" i2" i3" */ |
| 495 |
__ASM_EMIT("vshufps $0xdd, %%ymm3, %%ymm2, %%ymm7") /* ymm7 = i4" i5" i6" i7" */ |
| 496 |
__ASM_EMIT("vshufps $0x88, %%ymm5, %%ymm4, %%ymm2") /* ymm2 = r0" r1" r2" r3" */ |
| 497 |
__ASM_EMIT("vshufps $0xdd, %%ymm5, %%ymm4, %%ymm3") /* ymm3 = r4" r5" r6" r7" */ |
| 498 |
/* 3rd-order 8x butterfly */ |
| 499 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm3, %%ymm4") /* ymm4 = x_im * b_re */ \ |
| 500 |
__ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm7, %%ymm5") /* ymm5 = x_im * b_im */ \ |
| 501 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%ymm3, %%ymm3", "")) /* ymm3 = x_re * b_re */ \ |
| 502 |
__ASM_EMIT(FFT_FMA("vmulps 0x00 + %[FFT_A], %%ymm7, %%ymm7", "")) /* ymm7 = x_re * b_im */ \ |
| 503 |
__ASM_EMIT(FFT_FMA("vsubps %%ymm5, %%ymm3, %%ymm5", "vfmsub231ps 0x00 + %[FFT_A], %%ymm3, %%ymm5")) /* ymm5 = c_re = x_re * b_re - x_im * b_im */ \ |
| 504 |
__ASM_EMIT(FFT_FMA("vaddps %%ymm4, %%ymm7, %%ymm4", "vfmadd231ps 0x00 + %[FFT_A], %%ymm7, %%ymm4")) /* ymm4 = c_im = x_re * b_im + x_im * b_re */ \ |
| 505 |
__ASM_EMIT("vsubps %%ymm5, %%ymm2, %%ymm0") /* ymm0 = a_re - c_re */ \ |
| 506 |
__ASM_EMIT("vsubps %%ymm4, %%ymm6, %%ymm1") /* ymm1 = a_im - c_im */ \ |
| 507 |
__ASM_EMIT("vaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = a_re + c_re */ \ |
| 508 |
__ASM_EMIT("vaddps %%ymm4, %%ymm6, %%ymm3") /* ymm3 = a_im + c_im */ \ |
| 509 |
/* Store */ |
| 510 |
__ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") |
| 511 |
__ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") |
| 512 |
__ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") |
| 513 |
__ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") |
| 514 |
__ASM_EMIT("vextractf128 $1, %%ymm2, 0x40(%[dst])") |
| 515 |
__ASM_EMIT("vextractf128 $1, %%ymm0, 0x50(%[dst])") |
| 516 |
__ASM_EMIT("vextractf128 $1, %%ymm3, 0x60(%[dst])") |
| 517 |
__ASM_EMIT("vextractf128 $1, %%ymm1, 0x70(%[dst])") |
| 518 |
__ASM_EMIT("add $0x80, %[dst]") |
| 519 |
|
| 520 |
: [dst] "+r" (dst), [index] "+r"(index) |
| 521 |
: [src] "r" (src), [regs] "r" (regs), |
| 522 |
[FFT_A] "o" (FFT_A) |
| 523 |
: "cc", "memory", |
| 524 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 525 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 526 |
); |
| 527 |
} |
| 528 |
} |
| 529 |
} |
| 530 |
} |
| 531 |
|
| 532 |
#undef FFT_PSCRAMBLE_SELF_DIRECT_NAME |
| 533 |
#undef FFT_PSCRAMBLE_SELF_REVERSE_NAME |
| 534 |
#undef FFT_PSCRAMBLE_COPY_DIRECT_NAME |
| 535 |
#undef FFT_PSCRAMBLE_COPY_REVERSE_NAME |
| 536 |
#undef FFT_TYPE |
| 537 |
#undef FFT_FMA |