| 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_SSE_CONVOLUTION_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_X86_SSE_CONVOLUTION_H_ |
| 24 |
|
| 25 |
#ifndef PRIVATE_DSP_ARCH_X86_SSE_IMPL |
| 26 |
#error "This header should not be included directly" |
| 27 |
#endif /* PRIVATE_DSP_ARCH_X86_SSE_IMPL */ |
| 28 |
|
| 29 |
namespace lsp |
| 30 |
{ |
| 31 |
namespace sse |
| 32 |
{ |
| 33 |
void convolve(float *dst, const float *src, const float *conv, size_t length, size_t count) |
| 34 |
{ |
| 35 |
const float *c; |
| 36 |
float *d; |
| 37 |
size_t clen; |
| 38 |
|
| 39 |
ARCH_X86_ASM( |
| 40 |
__ASM_EMIT("sub $4, %[count]") |
| 41 |
__ASM_EMIT("jb 20f") |
| 42 |
|
| 43 |
__ASM_EMIT("10:") |
| 44 |
__ASM_EMIT("xorps %%xmm7, %%xmm7") // xmm7 = 0 |
| 45 |
__ASM_EMIT("mov %[length], %[clen]") |
| 46 |
__ASM_EMIT("mov %[dst], %[d]") |
| 47 |
__ASM_EMIT("mov %[conv], %[c]") |
| 48 |
|
| 49 |
__ASM_EMIT("sub $4, %[clen]") |
| 50 |
__ASM_EMIT("jb 12f") |
| 51 |
// Load convolution kernel |
| 52 |
__ASM_EMIT("movups (%[k]), %%xmm0") // xmm0 = k0 k1 k2 k3 |
| 53 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") // xmm1 = k0 k1 k2 k3 |
| 54 |
__ASM_EMIT("movaps %%xmm0, %%xmm2") // xmm2 = k0 k1 k2 k3 |
| 55 |
__ASM_EMIT("movaps %%xmm1, %%xmm3") // xmm3 = k0 k1 k2 k3 |
| 56 |
__ASM_EMIT("shufps $0x00, %%xmm0, %%xmm0") // xmm0 = k0 k0 k0 k0 |
| 57 |
__ASM_EMIT("shufps $0x55, %%xmm1, %%xmm1") // xmm1 = k1 k1 k1 k1 |
| 58 |
__ASM_EMIT("shufps $0xaa, %%xmm2, %%xmm2") // xmm2 = k2 k2 k2 k2 |
| 59 |
__ASM_EMIT("shufps $0xff, %%xmm3, %%xmm3") // xmm3 = k2 k2 k2 k2 |
| 60 |
|
| 61 |
__ASM_EMIT("11:") |
| 62 |
__ASM_EMIT("movaps %%xmm7, %%xmm4") // xmm4 = p0 p1 p2 p3 |
| 63 |
__ASM_EMIT("movups (%[c]), %%xmm7") // xmm7 = c0 c1 c2 c3 |
| 64 |
__ASM_EMIT("movaps %%xmm4, %%xmm5") // xmm5 = p0 p1 p2 p3 |
| 65 |
__ASM_EMIT("shufps $0x4e, %%xmm7, %%xmm5") // xmm5 = p2 p3 c0 c1 (+) |
| 66 |
__ASM_EMIT("movaps %%xmm5, %%xmm6") // xmm6 = p2 p3 c0 c1 |
| 67 |
__ASM_EMIT("shufps $0x99, %%xmm7, %%xmm6") // xmm6 = p3 c0 c1 c2 (+) |
| 68 |
__ASM_EMIT("shufps $0x99, %%xmm5, %%xmm4") // xmm4 = p1 p2 p3 c0 |
| 69 |
|
| 70 |
// Untouchable: xmm0, xmm1, xmm2, xmm3, xmm7 |
| 71 |
__ASM_EMIT("mulps %%xmm2, %%xmm5") // xmm5 = V2 = k2*p2 k2*p3 k2*c0 k2*c1 |
| 72 |
__ASM_EMIT("mulps %%xmm1, %%xmm6") // xmm6 = V3 = k1*p3 k1*c0 k1*c1 k1*c2 |
| 73 |
__ASM_EMIT("mulps %%xmm3, %%xmm4") // xmm4 = V1 = k3*p1 k3*p2 k3*p3 k3*c0 |
| 74 |
__ASM_EMIT("addps %%xmm6, %%xmm5") // xmm5 = V2 + V3 |
| 75 |
__ASM_EMIT("movaps %%xmm7, %%xmm6") // xmm6 = c0 c1 c2 c3 |
| 76 |
__ASM_EMIT("addps %%xmm5, %%xmm4") // xmm4 = V1 + V2 + V3 |
| 77 |
__ASM_EMIT("mulps %%xmm0, %%xmm6") // xmm6 = V0 = k0*c0 k0*c1 k0*c2 k0*c3 |
| 78 |
__ASM_EMIT("movups (%[d]), %%xmm5") // xmm5 = D + d0 d1 d2 d3 |
| 79 |
__ASM_EMIT("addps %%xmm6, %%xmm4") // xmm4 = V0 + V1 + V2 + V3 |
| 80 |
__ASM_EMIT("addps %%xmm5, %%xmm4") // xmm4 = D + V0 + V1 + V2 + V3 |
| 81 |
__ASM_EMIT("movups %%xmm4, (%[d])") |
| 82 |
__ASM_EMIT("add $0x10, %[c]") // c += 4 |
| 83 |
__ASM_EMIT("add $0x10, %[d]") // d += 4 |
| 84 |
__ASM_EMIT("sub $4, %[clen]") // clen -= 4 |
| 85 |
__ASM_EMIT("jae 11b") |
| 86 |
|
| 87 |
// Apply tail: xmm7 = p0 p1 p2 p3 |
| 88 |
__ASM_EMIT("movaps %%xmm7, %%xmm5") // xmm5 = p0 p1 p2 p3 |
| 89 |
__ASM_EMIT("movhlps %%xmm7, %%xmm6") // xmm6 = p2 |
| 90 |
__ASM_EMIT("shufps $0xff, %%xmm5, %%xmm5") // xmm5 = p3 |
| 91 |
__ASM_EMIT("shufps $0x55, %%xmm7, %%xmm7") // xmm7 = p1 |
| 92 |
__ASM_EMIT("movss 0x04(%[k]), %%xmm0") // xmm0 = k1 |
| 93 |
__ASM_EMIT("movss 0x08(%[k]), %%xmm1") // xmm1 = k2 |
| 94 |
__ASM_EMIT("movss 0x0c(%[k]), %%xmm2") // xmm2 = k3 |
| 95 |
|
| 96 |
__ASM_EMIT("movaps %%xmm0, %%xmm3") // xmm3 = k1 |
| 97 |
__ASM_EMIT("movaps %%xmm1, %%xmm4") // xmm4 = k2 |
| 98 |
__ASM_EMIT("mulss %%xmm2, %%xmm7") // xmm7 = k3*p1 |
| 99 |
__ASM_EMIT("mulss %%xmm5, %%xmm3") // xmm3 = k1*p3 |
| 100 |
__ASM_EMIT("mulss %%xmm6, %%xmm4") // xmm4 = k2*p2 |
| 101 |
__ASM_EMIT("addss %%xmm7, %%xmm3") // xmm3 = k1*p3 + k3*p1 |
| 102 |
__ASM_EMIT("addss %%xmm4, %%xmm3") // xmm3 = k1*p3 + k2*p2 + k3*p1 |
| 103 |
__ASM_EMIT("movss 0x00(%[d]), %%xmm7") // xmm7 = d0 |
| 104 |
__ASM_EMIT("addss %%xmm3, %%xmm7") // xmm7 = d0 + k1*p3 + k2*p2 + k3*p1 |
| 105 |
__ASM_EMIT("movss %%xmm7, 0x00(%[d])") // xmm7 = d0 |
| 106 |
|
| 107 |
__ASM_EMIT("movss 0x04(%[d]), %%xmm7") // xmm4 = d1 |
| 108 |
__ASM_EMIT("mulss %%xmm2, %%xmm6") // xmm6 = k3*p2 |
| 109 |
__ASM_EMIT("mulss %%xmm5, %%xmm1") // xmm1 = k2*p3 |
| 110 |
__ASM_EMIT("addss %%xmm6, %%xmm7") // xmm7 = d1 + k3*p2 |
| 111 |
__ASM_EMIT("addss %%xmm1, %%xmm7") // xmm7 = d1 + k3*p2 + k2*p3 |
| 112 |
__ASM_EMIT("movss %%xmm7, 0x04(%[d])") |
| 113 |
|
| 114 |
__ASM_EMIT("movss 0x08(%[d]), %%xmm7") // xmm7 = d2 |
| 115 |
__ASM_EMIT("mulss %%xmm5, %%xmm2") // xmm2 = k3*p3 |
| 116 |
__ASM_EMIT("addss %%xmm2, %%xmm7") // xmm7 = d2 + k3*p3 |
| 117 |
__ASM_EMIT("movss %%xmm7, 0x08(%[d])") |
| 118 |
|
| 119 |
// Apply tail |
| 120 |
__ASM_EMIT("12:") |
| 121 |
__ASM_EMIT("add $3, %[clen]") // while (clen >= 0) |
| 122 |
__ASM_EMIT("jl 14f") |
| 123 |
__ASM_EMIT("movups 0x00(%[k]), %%xmm1") // xmm1 = k0 k1 k2 k3 |
| 124 |
|
| 125 |
__ASM_EMIT("15:") |
| 126 |
__ASM_EMIT("movss 0x00(%[c]), %%xmm0") // xmm0 = c0 |
| 127 |
__ASM_EMIT("shufps $0x00, %%xmm0, %%xmm0") // xmm0 = c0 c0 c0 c0 |
| 128 |
__ASM_EMIT("movups 0x00(%[d]), %%xmm2") // xmm2 = d0 d1 d2 d3 |
| 129 |
__ASM_EMIT("mulps %%xmm1, %%xmm0") // xmm0 = k0*c0 k1*c0 k2*c0 k3*c0 |
| 130 |
__ASM_EMIT("addps %%xmm2, %%xmm0") // xmm0 = d0+k0*c0 d1+k1*c0 d2+k2*c0 d3+k3*c0 |
| 131 |
__ASM_EMIT("movups %%xmm0, 0x00(%[d])") |
| 132 |
__ASM_EMIT("add $0x04, %[c]") // c++ |
| 133 |
__ASM_EMIT("add $0x04, %[d]") // d++ |
| 134 |
__ASM_EMIT("dec %[clen]") |
| 135 |
__ASM_EMIT("jge 15b") |
| 136 |
|
| 137 |
__ASM_EMIT("14:") |
| 138 |
__ASM_EMIT64("add $0x10, %[dst]") // dst += 4 |
| 139 |
__ASM_EMIT32("addl $0x10, %[dst]") |
| 140 |
__ASM_EMIT("add $0x10, %[k]") // k += 4 |
| 141 |
__ASM_EMIT("sub $0x04, %[count]") // count -= 4 |
| 142 |
__ASM_EMIT("jge 10b") |
| 143 |
|
| 144 |
__ASM_EMIT("20:") |
| 145 |
__ASM_EMIT("add $3, %[count]") |
| 146 |
__ASM_EMIT("jl 40f") |
| 147 |
__ASM_EMIT("21:") |
| 148 |
__ASM_EMIT("mov %[length], %[clen]") |
| 149 |
__ASM_EMIT("mov %[dst], %[d]") |
| 150 |
__ASM_EMIT("mov %[conv], %[c]") |
| 151 |
__ASM_EMIT("movss 0x00(%[k]), %%xmm0") // xmm0 = k0 |
| 152 |
__ASM_EMIT("sub $8, %[clen]") |
| 153 |
__ASM_EMIT("shufps $0x00, %%xmm0, %%xmm0") // xmm0 = k0 k0 k0 k0 |
| 154 |
__ASM_EMIT("jb 22f") |
| 155 |
__ASM_EMIT("movaps %%xmm0, %%xmm1") // xmm1 = k0 k0 k0 k0 |
| 156 |
__ASM_EMIT("23:") |
| 157 |
__ASM_EMIT("movups 0x00(%[c]), %%xmm2") // xmm2 = c0 c1 c2 c3 |
| 158 |
__ASM_EMIT("movups 0x10(%[c]), %%xmm3") |
| 159 |
__ASM_EMIT("movups 0x00(%[d]), %%xmm4") // xmm4 = d0 d1 d2 d3 |
| 160 |
__ASM_EMIT("movups 0x10(%[d]), %%xmm5") |
| 161 |
__ASM_EMIT("mulps %%xmm0, %%xmm2") // xmm2 = k0*c0 k0*c1 k0*c2 k0*c3 |
| 162 |
__ASM_EMIT("mulps %%xmm1, %%xmm3") |
| 163 |
__ASM_EMIT("addps %%xmm4, %%xmm2") // xmm2 = d0+k0*c0 d1+k0*c1 d2+k0*c2 d3+k0*c3 |
| 164 |
__ASM_EMIT("addps %%xmm5, %%xmm3") |
| 165 |
__ASM_EMIT("movups %%xmm2, 0x00(%[d])") |
| 166 |
__ASM_EMIT("movups %%xmm3, 0x10(%[d])") |
| 167 |
__ASM_EMIT("add $0x20, %[c]") // c += 8 |
| 168 |
__ASM_EMIT("add $0x20, %[d]") // d += 8 |
| 169 |
__ASM_EMIT("sub $8, %[clen]") // clen -= 8 |
| 170 |
__ASM_EMIT("jae 23b") |
| 171 |
|
| 172 |
// Apply convolution |
| 173 |
__ASM_EMIT("22:") |
| 174 |
__ASM_EMIT("add $4, %[clen]") |
| 175 |
__ASM_EMIT("jl 24f") |
| 176 |
__ASM_EMIT("movups 0x00(%[c]), %%xmm1") // xmm1 = c0 c1 c2 c3 |
| 177 |
__ASM_EMIT("movups 0x00(%[d]), %%xmm2") // xmm2 = d0 d1 d2 d3 |
| 178 |
__ASM_EMIT("mulps %%xmm0, %%xmm1") // xmm1 = k0*c0 k0*c1 k0*c2 k0*c3 |
| 179 |
__ASM_EMIT("addps %%xmm1, %%xmm2") // xmm2 = d0+k0*c0 d1+k0*c1 d2+k0*c2 d3+k0*c3 |
| 180 |
__ASM_EMIT("movups %%xmm2, 0x00(%[d])") |
| 181 |
__ASM_EMIT("add $0x10, %[c]") // c += 4 |
| 182 |
__ASM_EMIT("add $0x10, %[d]") // d += 4 |
| 183 |
__ASM_EMIT("sub $4, %[clen]") // clen -= 4 |
| 184 |
|
| 185 |
// Apply tail |
| 186 |
__ASM_EMIT("24:") |
| 187 |
__ASM_EMIT("add $3, %[clen]") |
| 188 |
__ASM_EMIT("jl 26f") |
| 189 |
__ASM_EMIT("25:") |
| 190 |
__ASM_EMIT("movss 0x00(%[c]), %%xmm1") // xmm1 = c0 |
| 191 |
__ASM_EMIT("movss 0x00(%[d]), %%xmm2") // xmm2 = d0 |
| 192 |
__ASM_EMIT("mulss %%xmm0, %%xmm1") // xmm1 = k0*c0 |
| 193 |
__ASM_EMIT("addss %%xmm1, %%xmm2") // xmm2 = d0+k0*c0 |
| 194 |
__ASM_EMIT("movss %%xmm2, 0x00(%[d])") |
| 195 |
__ASM_EMIT("add $0x04, %[c]") // c ++ |
| 196 |
__ASM_EMIT("add $0x04, %[d]") // d ++ |
| 197 |
__ASM_EMIT("dec %[clen]") |
| 198 |
__ASM_EMIT("jge 25b") |
| 199 |
|
| 200 |
__ASM_EMIT("26:") |
| 201 |
__ASM_EMIT("add $0x04, %[k]") |
| 202 |
__ASM_EMIT64("add $0x04, %[dst]") // dst++ |
| 203 |
__ASM_EMIT32("addl $0x04, %[dst]") |
| 204 |
__ASM_EMIT("dec %[count]") |
| 205 |
__ASM_EMIT("jge 21b") |
| 206 |
|
| 207 |
__ASM_EMIT("40:") |
| 208 |
: __IF_32([dst] "+m" (dst)) __IF_64([dst] "+r" (dst)), |
| 209 |
[k] "+r" (src), [count] "+r" (count), |
| 210 |
[c] "=&r" (c), [d] "=&r" (d), [clen] "=&r" (clen) |
| 211 |
: [conv] X86_GREG (conv), [length] X86_GREG (length) |
| 212 |
: "cc", "memory", |
| 213 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 214 |
); |
| 215 |
} |
| 216 |
} /* namespace sse */ |
| 217 |
} /* namespace lsp */ |
| 218 |
|
| 219 |
#endif /* PRIVATE_DSP_ARCH_X86_SSE_CONVOLUTION_H_ */ |