| 1 |
/* |
| 2 |
* Copyright (C) 2026 Linux Studio Plugins Project <https://lsp-plug.in/> |
| 3 |
* (C) 2026 Vladimir Sadovnikov <[email protected]> |
| 4 |
* |
| 5 |
* This file is part of lsp-dsp-lib |
| 6 |
* Created on: 17 апр. 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_AVX2_PMATH_LANCZOS_H_ |
| 23 |
#define PRIVATE_DSP_ARCH_X86_AVX2_PMATH_LANCZOS_H_ |
| 24 |
|
| 25 |
#ifndef PRIVATE_DSP_ARCH_X86_AVX2_IMPL |
| 26 |
#error "This header should not be included directly" |
| 27 |
#endif /* PRIVATE_DSP_ARCH_X86_AVX2_IMPL */ |
| 28 |
|
| 29 |
#include <private/dsp/arch/x86/avx2/pmath/sin.h> |
| 30 |
|
| 31 |
namespace lsp |
| 32 |
{ |
| 33 |
namespace avx2 |
| 34 |
{ |
| 35 |
IF_ARCH_X86( |
| 36 |
static const uint32_t lanczos_const[] __lsp_aligned32 = |
| 37 |
{ |
| 38 |
LSP_DSP_VEC8(0x7fffffff), // +0x00: Mask for fabsf |
| 39 |
LSP_DSP_VEC8(0x38d1b717), // +0x20: Sinc threshold = 1e-4 |
| 40 |
LSP_DSP_VEC8(0x3f800000), // +0x40: 1.0 |
| 41 |
}; |
| 42 |
) |
| 43 |
|
| 44 |
typedef struct lanczos_gen_t |
| 45 |
{ |
| 46 |
float k[8]; // +0x00: Multiplier |
| 47 |
float p[8]; // +0x20: Initial phase |
| 48 |
float a[8]; // +0x40: Number of lobes |
| 49 |
float i[8]; // +0x60: Indices |
| 50 |
float x1[8]; // +0x80: Computed X1 |
| 51 |
float d[8]; // +0xa0: Denominator |
| 52 |
float t[8]; // +0xc0: Time range |
| 53 |
} lanczos_gen_t; |
| 54 |
|
| 55 |
#define LANCZOS_GEN_FUNC_X8 \ |
| 56 |
/* ymm0 = x1 */ \ |
| 57 |
__ASM_EMIT("vmulps 0x40(%[state]), %%ymm0, %%ymm4") /* ymm4 = x2 = x1*a */ \ |
| 58 |
__ASM_EMIT("vmovaps %%ymm0, 0x80(%[state])") /* *x1 = x1 */ \ |
| 59 |
__ASM_EMIT("vmulps %%ymm0, %%ymm4, %%ymm2") /* ymm2 = d = x1 * x2 */ \ |
| 60 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm0, %%ymm0") /* ymm0 = x1 + PI/2 */ \ |
| 61 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm4, %%ymm4") /* ymm4 = x2 + PI/2 */ \ |
| 62 |
__ASM_EMIT("vmovaps %%ymm2, 0xa0(%[state])") /* *d = d */ \ |
| 63 |
SINF_X_PLUS_PI_2_CORE_X16 /* ymm0 = sinf(x1), ymm4 = sinf(x2) */ \ |
| 64 |
__ASM_EMIT("vmovaps 0x80(%[state]), %%ymm1") /* ymm1 = x1 */ \ |
| 65 |
__ASM_EMIT("vmulps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = sinf(x1)*sinf(x2) */ \ |
| 66 |
__ASM_EMIT("vandps 0x00 + %[LC], %%ymm1, %%ymm1") /* ymm1 = fabsf(x) */ \ |
| 67 |
__ASM_EMIT("vdivps 0xa0(%[state]), %%ymm0, %%ymm0") /* ymm0 = F = (sinf(x1)*sinf(x2)) / (x1 * x2) */ \ |
| 68 |
__ASM_EMIT("vcmpps $1, 0x20 + %[LC], %%ymm1, %%ymm2") /* ymm2 = [ fabsf(x) < 1e-4 ] */ \ |
| 69 |
__ASM_EMIT("vcmpps $1, 0xc0(%[state]), %%ymm1, %%ymm1") /* ymm1 = [ fabsf(x) < t ] */ \ |
| 70 |
__ASM_EMIT("vblendvps %%ymm2, 0x40 + %[LC], %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x) >= 1e-4 ] ? f : 1.0 */ \ |
| 71 |
__ASM_EMIT("vandps %%ymm1, %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x) < t ] ? ([ fabsf(x) >= 1e-4 ] ? f : 1.0) : 0.0 */ |
| 72 |
|
| 73 |
#define LANCZOS_GEN_FUNC_X4 \ |
| 74 |
/* xmm0 = x1 */ \ |
| 75 |
__ASM_EMIT("vmulps 0x40(%[state]), %%xmm0, %%xmm4" ) /* xmm4 = x2 = x1*a */ \ |
| 76 |
__ASM_EMIT("vmovaps %%xmm0, 0x80(%[state])") /* *x1 = x1 */ \ |
| 77 |
__ASM_EMIT("vmulps %%xmm0, %%xmm4, %%xmm2") /* xmm2 = d = x1 * x2 */ \ |
| 78 |
__ASM_EMIT("vinsertf128 $1, %%xmm4, %%ymm0, %%ymm0") /* ymm0 = x1 x1 x1 x1 x2 x2 x2 x2 */ \ |
| 79 |
__ASM_EMIT("vmovaps %%xmm2, 0xa0(%[state])") /* *d = d */ \ |
| 80 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm0, %%ymm0") /* xmm0 = x1 + PI/2 ... x2 + PI/2 ... */ \ |
| 81 |
SINF_X_PLUS_PI_2_CORE_X8 /* ymm0 = sinf(x1) ... sinf(x2) ... */ \ |
| 82 |
__ASM_EMIT("vextractf128 $1, %%ymm0, %%xmm4") /* xmm4 = sif(x2) */ \ |
| 83 |
__ASM_EMIT("vmovaps 0x80(%[state]), %%xmm1") /* xmm1 = x1 */ \ |
| 84 |
__ASM_EMIT("vmulps %%xmm4, %%xmm0, %%xmm0") /* xmm0 = sinf(x1)*sinf(x2) */ \ |
| 85 |
__ASM_EMIT("vandps 0x00 + %[LC], %%xmm1, %%xmm1") /* xmm1 = fabsf(x) */ \ |
| 86 |
__ASM_EMIT("vdivps 0xa0(%[state]), %%xmm0, %%xmm0") /* xmm0 = F = (sinf(x1)*sinf(x2)) / (x1 * x2) */ \ |
| 87 |
__ASM_EMIT("vcmpps $1, 0x20 + %[LC], %%xmm1, %%xmm2") /* xmm2 = [ fabsf(x) < 1e-4 ] */ \ |
| 88 |
__ASM_EMIT("vcmpps $1, 0xc0(%[state]), %%xmm1, %%xmm1") /* xmm1 = [ fabsf(x) < t ] */ \ |
| 89 |
__ASM_EMIT("vblendvps %%xmm2, 0x40 + %[LC], %%xmm0, %%xmm0") /* xmm0 = [ fabsf(x) >= 1e-4 ] ? f : 1.0 */ \ |
| 90 |
__ASM_EMIT("vandps %%xmm1, %%xmm0, %%xmm0") /* xmm0 = [ fabsf(x) < t ] ? ([ fabsf(x) >= 1e-4 ] ? f : 1.0) : 0.0 */ |
| 91 |
|
| 92 |
|
| 93 |
void lanczos1(float *dst, float k, float p, float t, float a, size_t count) |
| 94 |
{ |
| 95 |
IF_ARCH_X86( |
| 96 |
lanczos_gen_t state __lsp_aligned32; |
| 97 |
float stub[8] __lsp_aligned16; |
| 98 |
); |
| 99 |
|
| 100 |
ARCH_X86_ASM( |
| 101 |
// Prepare |
| 102 |
__ASM_EMIT("vbroadcastss %[k], %%ymm4") // ymm4 = k |
| 103 |
__ASM_EMIT("vbroadcastss %[p], %%ymm5") // ymm5 = p |
| 104 |
__ASM_EMIT("vbroadcastss %[a], %%ymm6") // ymm6 = a |
| 105 |
__ASM_EMIT("vbroadcastss %[t], %%ymm7") // ymm7 = t |
| 106 |
__ASM_EMIT("vmovaps %%ymm4, 0x00(%[state])") |
| 107 |
__ASM_EMIT("vmovaps %%ymm5, 0x20(%[state])") |
| 108 |
__ASM_EMIT("vmovaps %%ymm6, 0x40(%[state])") |
| 109 |
__ASM_EMIT("vmovaps %%ymm7, 0xc0(%[state])") |
| 110 |
__ASM_EMIT("vmovaps 0x00 + %[LGEN], %%ymm1") // ymm1 = i = 0 1 2 3 4 5 6 7 |
| 111 |
// x8 blocks |
| 112 |
__ASM_EMIT("sub $8, %[count]") |
| 113 |
__ASM_EMIT("jb 2f") |
| 114 |
__ASM_EMIT("1:") |
| 115 |
__ASM_EMIT("vmulps 0x00(%[state]), %%ymm1, %%ymm0") // ymm0 = k*i |
| 116 |
__ASM_EMIT("vaddps 0x60 + %[LGEN], %%ymm1, %%ymm1") // ymm1 = i' = i + step |
| 117 |
__ASM_EMIT("vsubps 0x20(%[state]), %%ymm0, %%ymm0") // ymm0 = x1 = k*i - p |
| 118 |
__ASM_EMIT("vmovaps %%ymm1, 0x60(%[state])") // *i = i' |
| 119 |
LANCZOS_GEN_FUNC_X8 |
| 120 |
__ASM_EMIT("vmovaps 0x60(%[state]), %%ymm1") // ymm1 = i' |
| 121 |
__ASM_EMIT("vmovups %%ymm0, 0x00(%[dst])") |
| 122 |
__ASM_EMIT("add $0x20, %[dst]") |
| 123 |
__ASM_EMIT("sub $8, %[count]") |
| 124 |
__ASM_EMIT("jae 1b") |
| 125 |
__ASM_EMIT("2:") |
| 126 |
// x4 block |
| 127 |
__ASM_EMIT("add $4, %[count]") |
| 128 |
__ASM_EMIT("jl 4f") |
| 129 |
__ASM_EMIT("vmulps 0x00(%[state]), %%xmm1, %%xmm0") // xmm0 = k*i |
| 130 |
__ASM_EMIT("vaddps 0x80 + %[LGEN], %%xmm1, %%xmm1") // xmm1 = i' = i + step |
| 131 |
__ASM_EMIT("vsubps 0x20(%[state]), %%xmm0, %%xmm0") // xmm0 = x1 = k*i - p |
| 132 |
__ASM_EMIT("vmovaps %%xmm1, 0x60(%[state])") // *i = i' |
| 133 |
LANCZOS_GEN_FUNC_X4 |
| 134 |
__ASM_EMIT("vmovaps 0x60(%[state]), %%xmm1") // xmm1 = i' |
| 135 |
__ASM_EMIT("vmovups %%xmm0, 0x00(%[dst])") |
| 136 |
__ASM_EMIT("add $0x10, %[dst]") |
| 137 |
__ASM_EMIT("sub $4, %[count]") |
| 138 |
__ASM_EMIT("4:") |
| 139 |
// Tail: 1x-3x block |
| 140 |
__ASM_EMIT("add $4, %[count]") |
| 141 |
__ASM_EMIT("jle 8f") |
| 142 |
__ASM_EMIT("vmulps 0x00(%[state]), %%xmm1, %%xmm0") // xmm0 = k*i |
| 143 |
__ASM_EMIT("vsubps 0x20(%[state]), %%xmm0, %%xmm0") // xmm0 = x1 = k*i - p |
| 144 |
LANCZOS_GEN_FUNC_X4 |
| 145 |
__ASM_EMIT("test $1, %[count]") |
| 146 |
__ASM_EMIT("jz 6f") |
| 147 |
__ASM_EMIT("vmovss %%xmm0, 0x00(%[dst])") |
| 148 |
__ASM_EMIT("vshufps $0x39, %%xmm0, %%xmm0, %%xmm0") |
| 149 |
__ASM_EMIT("add $4, %[dst]") |
| 150 |
__ASM_EMIT("6:") |
| 151 |
__ASM_EMIT("test $2, %[count]") |
| 152 |
__ASM_EMIT("jz 8f") |
| 153 |
__ASM_EMIT("vmovlps %%xmm0, 0x00(%[dst])") |
| 154 |
// End |
| 155 |
__ASM_EMIT("8:") |
| 156 |
|
| 157 |
: [dst] "+r" (dst), [count] "+r" (count) |
| 158 |
: [state] "r" (&state), |
| 159 |
[S2C] "o" (sinf_const), |
| 160 |
[LGEN] "o" (kp_gen_const), |
| 161 |
[LC] "o" (lanczos_const), |
| 162 |
[stub] "o" (stub), |
| 163 |
[k] "m" (k), |
| 164 |
[p] "m" (p), |
| 165 |
[a] "m" (a), |
| 166 |
[t] "m" (t) |
| 167 |
: "cc", "memory", |
| 168 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 169 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 170 |
); |
| 171 |
} |
| 172 |
|
| 173 |
#define LANCZOS_GEN_X64_FUNC_X8 \ |
| 174 |
/* ymm8 = k */ \ |
| 175 |
/* ymm9 = p */ \ |
| 176 |
/* ymm10 = t */ \ |
| 177 |
/* ymm11 = a */ \ |
| 178 |
/* ymm12 = x1 */ \ |
| 179 |
/* ymm13 = d */ \ |
| 180 |
/* ymm14 = i */ \ |
| 181 |
/* ymm15 = step */ \ |
| 182 |
__ASM_EMIT("vmulps %%ymm11, %%ymm12, %%ymm4") /* ymm12 = x2 = x1*a */ \ |
| 183 |
__ASM_EMIT("vmulps %%ymm12, %%ymm4, %%ymm13") /* ymm13 = d = x1 * x2 */ \ |
| 184 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm12, %%ymm0") /* ymm0 = x1 + PI/2 */ \ |
| 185 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm4, %%ymm4") /* ymm4 = x2 + PI/2 */ \ |
| 186 |
SINF_X_PLUS_PI_2_CORE_X16 /* ymm0 = sinf(x1), ymm4 = sinf(x2) */ \ |
| 187 |
__ASM_EMIT("vmulps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = sinf(x1)*sinf(x2) */ \ |
| 188 |
__ASM_EMIT("vandps 0x00 + %[LC], %%ymm12, %%ymm12") /* ymm12 = fabsf(x1) */ \ |
| 189 |
__ASM_EMIT("vdivps %%ymm13, %%ymm0, %%ymm0") /* ymm0 = F = (sinf(x1)*sinf(x2)) / (x1 * x2) */ \ |
| 190 |
__ASM_EMIT("vcmpps $1, 0x20 + %[LC], %%ymm12, %%ymm2") /* ymm2 = [ fabsf(x1) < 1e-4 ] */ \ |
| 191 |
__ASM_EMIT("vcmpps $1, %%ymm10, %%ymm12, %%ymm1") /* ymm1 = [ fabsf(x1) < t ] */ \ |
| 192 |
__ASM_EMIT("vblendvps %%ymm2, 0x40 + %[LC], %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x1) >= 1e-4 ] ? f : 1.0 */ \ |
| 193 |
__ASM_EMIT("vandps %%ymm1, %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x1) < t ] ? ([ fabsf(x1) >= 1e-4 ] ? f : 1.0) : 0.0 */ |
| 194 |
|
| 195 |
#define LANCZOS_GEN_X64_FUNC_X4 \ |
| 196 |
/* xmm0 = x1 */ \ |
| 197 |
__ASM_EMIT("vmulps %%xmm11, %%xmm12, %%xmm4") /* xmm12 = x2 = x1*a */ \ |
| 198 |
__ASM_EMIT("vmulps %%xmm12, %%xmm4, %%xmm13") /* xmm13 = d = x1 * x2 */ \ |
| 199 |
__ASM_EMIT("vinsertf128 $1, %%xmm4, %%ymm12, %%ymm0") /* ymm0 = x1 x1 x1 x1 x2 x2 x2 x2 */ \ |
| 200 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm0, %%ymm0") /* ymm0 = x1 + PI/2 ... x2 + PI/2 */ \ |
| 201 |
SINF_X_PLUS_PI_2_CORE_X8 /* ymm0 = sinf(x1) ... sinf(x2) */ \ |
| 202 |
__ASM_EMIT("vextractf128 $1, %%ymm0, %%xmm4") /* xmm4 = sif(x2) */ \ |
| 203 |
__ASM_EMIT("vmulps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = sinf(x1)*sinf(x2) */ \ |
| 204 |
__ASM_EMIT("vandps 0x00 + %[LC], %%ymm12, %%ymm12") /* ymm12 = fabsf(x1) */ \ |
| 205 |
__ASM_EMIT("vdivps %%ymm13, %%ymm0, %%ymm0") /* ymm0 = F = (sinf(x1)*sinf(x2)) / (x1 * x2) */ \ |
| 206 |
__ASM_EMIT("vcmpps $1, 0x20 + %[LC], %%ymm12, %%ymm2") /* ymm2 = [ fabsf(x1) < 1e-4 ] */ \ |
| 207 |
__ASM_EMIT("vcmpps $1, %%ymm10, %%ymm12, %%ymm1") /* ymm1 = [ fabsf(x1) < t ] */ \ |
| 208 |
__ASM_EMIT("vblendvps %%ymm2, 0x40 + %[LC], %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x1) >= 1e-4 ] ? f : 1.0 */ \ |
| 209 |
__ASM_EMIT("vandps %%ymm1, %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x1) < t ] ? ([ fabsf(x1) >= 1e-4 ] ? f : 1.0) : 0.0 */ |
| 210 |
|
| 211 |
void x64_lanczos1(float *dst, float k, float p, float t, float a, size_t count) |
| 212 |
{ |
| 213 |
ARCH_X86_64_ASM( |
| 214 |
// Prepare |
| 215 |
__ASM_EMIT("vbroadcastss %[k], %%ymm8") // ymm8 = k |
| 216 |
__ASM_EMIT("vbroadcastss %[p], %%ymm9") // ymm9 = p |
| 217 |
__ASM_EMIT("vbroadcastss %[t], %%ymm10") // ymm10 = t |
| 218 |
__ASM_EMIT("vbroadcastss %[a], %%ymm11") // ymm11 = a |
| 219 |
__ASM_EMIT("vmovaps 0x00 + %[LGEN], %%ymm14") // ymm14 = i = 0 1 2 3 4 5 6 7 |
| 220 |
// x8 blocks |
| 221 |
__ASM_EMIT("sub $8, %[count]") |
| 222 |
__ASM_EMIT("jb 2f") |
| 223 |
__ASM_EMIT("vmovaps 0x60 + %[LGEN], %%ymm15") // ymm15 = step |
| 224 |
__ASM_EMIT("1:") |
| 225 |
__ASM_EMIT("vmulps %%ymm8, %%ymm14, %%ymm12") // ymm12 = k*i |
| 226 |
__ASM_EMIT("vaddps %%ymm15, %%ymm14, %%ymm14") // ymm14 = i' = i + step |
| 227 |
__ASM_EMIT("vsubps %%ymm9, %%ymm12, %%ymm12") // ymm12 = x1 = k*i - p |
| 228 |
LANCZOS_GEN_X64_FUNC_X8 |
| 229 |
__ASM_EMIT("vmovups %%ymm0, 0x00(%[dst])") |
| 230 |
__ASM_EMIT("add $0x20, %[dst]") |
| 231 |
__ASM_EMIT("sub $8, %[count]") |
| 232 |
__ASM_EMIT("jae 1b") |
| 233 |
__ASM_EMIT("2:") |
| 234 |
// x4 block |
| 235 |
__ASM_EMIT("add $4, %[count]") |
| 236 |
__ASM_EMIT("jl 4f") |
| 237 |
__ASM_EMIT("vmulps %%xmm8, %%xmm14, %%xmm12") // xmm12 = k*i |
| 238 |
__ASM_EMIT("vaddps 0x80 + %[LGEN], %%xmm14, %%xmm14") // xmm14 = i' = i + step |
| 239 |
__ASM_EMIT("vsubps %%xmm9, %%xmm12, %%xmm12") // xmm12 = x1 = k*i - p |
| 240 |
LANCZOS_GEN_X64_FUNC_X4 |
| 241 |
__ASM_EMIT("vmovups %%xmm0, 0x00(%[dst])") |
| 242 |
__ASM_EMIT("add $0x10, %[dst]") |
| 243 |
__ASM_EMIT("sub $4, %[count]") |
| 244 |
__ASM_EMIT("4:") |
| 245 |
// Tail: 1x-3x block |
| 246 |
__ASM_EMIT("add $4, %[count]") |
| 247 |
__ASM_EMIT("jle 8f") |
| 248 |
__ASM_EMIT("vmulps %%xmm8, %%xmm14, %%xmm12") // xmm12 = k*i |
| 249 |
__ASM_EMIT("vsubps %%xmm9, %%xmm12, %%xmm12") // xmm12 = x1 = k*i - p |
| 250 |
LANCZOS_GEN_X64_FUNC_X4 |
| 251 |
__ASM_EMIT("test $1, %[count]") |
| 252 |
__ASM_EMIT("jz 6f") |
| 253 |
__ASM_EMIT("vmovss %%xmm0, 0x00(%[dst])") |
| 254 |
__ASM_EMIT("vshufps $0x39, %%xmm0, %%xmm0, %%xmm0") |
| 255 |
__ASM_EMIT("add $4, %[dst]") |
| 256 |
__ASM_EMIT("6:") |
| 257 |
__ASM_EMIT("test $2, %[count]") |
| 258 |
__ASM_EMIT("jz 8f") |
| 259 |
__ASM_EMIT("vmovlps %%xmm0, 0x00(%[dst])") |
| 260 |
// End |
| 261 |
__ASM_EMIT("8:") |
| 262 |
|
| 263 |
: [dst] "+r" (dst), [count] "+r" (count) |
| 264 |
: [S2C] "o" (sinf_const), |
| 265 |
[LGEN] "o" (kp_gen_const), |
| 266 |
[LC] "o" (lanczos_const), |
| 267 |
[k] "m" (k), |
| 268 |
[p] "m" (p), |
| 269 |
[a] "m" (a), |
| 270 |
[t] "m" (t) |
| 271 |
: "cc", "memory", |
| 272 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 273 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7", |
| 274 |
"%xmm8", "%xmm9", "%xmm10", "%xmm11", |
| 275 |
"%xmm12", "%xmm13", "%xmm14", "%xmm15" |
| 276 |
); |
| 277 |
} |
| 278 |
|
| 279 |
|
| 280 |
#define LANCZOS_GEN_FUNC_X8_FMA3 \ |
| 281 |
/* ymm0 = x1 */ \ |
| 282 |
__ASM_EMIT("vmulps 0x40(%[state]), %%ymm0, %%ymm4") /* ymm4 = x2 = x1*a */ \ |
| 283 |
__ASM_EMIT("vmovaps %%ymm0, 0x80(%[state])") /* *x1 = x1 */ \ |
| 284 |
__ASM_EMIT("vmulps %%ymm0, %%ymm4, %%ymm2") /* ymm2 = d = x1 * x2 */ \ |
| 285 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm0, %%ymm0") /* ymm0 = x1 + PI/2 */ \ |
| 286 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm4, %%ymm4") /* ymm4 = x2 + PI/2 */ \ |
| 287 |
__ASM_EMIT("vmovaps %%ymm2, 0xa0(%[state])") /* *d = d */ \ |
| 288 |
SINF_X_PLUS_PI_2_CORE_X16_FMA3 /* ymm0 = sinf(x1), ymm4 = sinf(x2) */ \ |
| 289 |
__ASM_EMIT("vmovaps 0x80(%[state]), %%ymm1") /* ymm1 = x1 */ \ |
| 290 |
__ASM_EMIT("vmulps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = sinf(x1)*sinf(x2) */ \ |
| 291 |
__ASM_EMIT("vandps 0x00 + %[LC], %%ymm1, %%ymm1") /* ymm1 = fabsf(x) */ \ |
| 292 |
__ASM_EMIT("vdivps 0xa0(%[state]), %%ymm0, %%ymm0") /* ymm0 = F = (sinf(x1)*sinf(x2)) / (x1 * x2) */ \ |
| 293 |
__ASM_EMIT("vcmpps $1, 0x20 + %[LC], %%ymm1, %%ymm2") /* ymm2 = [ fabsf(x) < 1e-4 ] */ \ |
| 294 |
__ASM_EMIT("vcmpps $1, 0xc0(%[state]), %%ymm1, %%ymm1") /* ymm1 = [ fabsf(x) < t ] */ \ |
| 295 |
__ASM_EMIT("vblendvps %%ymm2, 0x40 + %[LC], %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x) >= 1e-4 ] ? f : 1.0 */ \ |
| 296 |
__ASM_EMIT("vandps %%ymm1, %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x) < t ] ? ([ fabsf(x) >= 1e-4 ] ? f : 1.0) : 0.0 */ |
| 297 |
|
| 298 |
#define LANCZOS_GEN_FUNC_X4_FMA3 \ |
| 299 |
/* xmm0 = x1 */ \ |
| 300 |
__ASM_EMIT("vmulps 0x40(%[state]), %%xmm0, %%xmm4") /* xmm4 = x2 = x1*a */ \ |
| 301 |
__ASM_EMIT("vmovaps %%xmm0, 0x80(%[state])") /* *x1 = x1 */ \ |
| 302 |
__ASM_EMIT("vmulps %%xmm0, %%xmm4, %%xmm2") /* xmm2 = d = x1 * x2 */ \ |
| 303 |
__ASM_EMIT("vinsertf128 $1, %%xmm4, %%ymm0, %%ymm0") /* ymm0 = x1 x1 x1 x1 x2 x2 x2 x2 */ \ |
| 304 |
__ASM_EMIT("vmovaps %%xmm2, 0xa0(%[state])") /* *d = d */ \ |
| 305 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm0, %%ymm0") /* xmm0 = x1 + PI/2 ... x2 + PI/2 ... */ \ |
| 306 |
SINF_X_PLUS_PI_2_CORE_X8_FMA3 /* ymm0 = sinf(x1) ... sinf(x2) */ \ |
| 307 |
__ASM_EMIT("vextractf128 $1, %%ymm0, %%xmm4") /* xmm4 = sif(x2) */ \ |
| 308 |
__ASM_EMIT("vmovaps 0x80(%[state]), %%xmm1") /* xmm1 = x1 */ \ |
| 309 |
__ASM_EMIT("vmulps %%xmm4, %%xmm0, %%xmm0") /* xmm0 = sinf(x1)*sinf(x2) */ \ |
| 310 |
__ASM_EMIT("vandps 0x00 + %[LC], %%xmm1, %%xmm1") /* xmm1 = fabsf(x) */ \ |
| 311 |
__ASM_EMIT("vdivps 0xa0(%[state]), %%xmm0, %%xmm0") /* xmm0 = F = (sinf(x1)*sinf(x2)) / (x1 * x2) */ \ |
| 312 |
__ASM_EMIT("vcmpps $1, 0x20 + %[LC], %%xmm1, %%xmm2") /* xmm2 = [ fabsf(x) < 1e-4 ] */ \ |
| 313 |
__ASM_EMIT("vcmpps $1,0xc0(%[state]), %%xmm1, %%xmm1") /* xmm1 = [ fabsf(x) < t ] */ \ |
| 314 |
__ASM_EMIT("vblendvps %%xmm2, 0x40 + %[LC], %%xmm0, %%xmm0") /* xmm0 = [ fabsf(x) >= 1e-4 ] ? f : 1.0 */ \ |
| 315 |
__ASM_EMIT("vandps %%xmm1, %%xmm0, %%xmm0") /* xmm0 = [ fabsf(x) < t ] ? ([ fabsf(x) >= 1e-4 ] ? f : 1.0) : 0.0 */ |
| 316 |
|
| 317 |
void lanczos1_fma3(float *dst, float k, float p, float t, float a, size_t count) |
| 318 |
{ |
| 319 |
IF_ARCH_X86( |
| 320 |
lanczos_gen_t state __lsp_aligned32; |
| 321 |
float stub[8] __lsp_aligned16; |
| 322 |
); |
| 323 |
|
| 324 |
ARCH_X86_ASM( |
| 325 |
// Prepare |
| 326 |
__ASM_EMIT("vbroadcastss %[k], %%ymm4") // ymm4 = k |
| 327 |
__ASM_EMIT("vbroadcastss %[p], %%ymm5") // ymm5 = p |
| 328 |
__ASM_EMIT("vbroadcastss %[a], %%ymm6") // ymm6 = a |
| 329 |
__ASM_EMIT("vbroadcastss %[t], %%ymm7") // ymm7 = t |
| 330 |
__ASM_EMIT("vmovaps %%ymm4, 0x00(%[state])") |
| 331 |
__ASM_EMIT("vmovaps %%ymm5, 0x20(%[state])") |
| 332 |
__ASM_EMIT("vmovaps %%ymm6, 0x40(%[state])") |
| 333 |
__ASM_EMIT("vmovaps %%ymm7, 0xc0(%[state])") |
| 334 |
__ASM_EMIT("vmovaps 0x00 + %[LGEN], %%ymm1") // ymm1 = i = 0 1 2 3 4 5 6 7 |
| 335 |
// x8 blocks |
| 336 |
__ASM_EMIT("sub $8, %[count]") |
| 337 |
__ASM_EMIT("jb 2f") |
| 338 |
__ASM_EMIT("1:") |
| 339 |
__ASM_EMIT("vmulps 0x00(%[state]), %%ymm1, %%ymm0") // ymm0 = k*i |
| 340 |
__ASM_EMIT("vaddps 0x60 + %[LGEN], %%ymm1, %%ymm1") // ymm1 = i' = i + step |
| 341 |
__ASM_EMIT("vsubps 0x20(%[state]), %%ymm0, %%ymm0") // ymm0 = x1 = k*i - p |
| 342 |
__ASM_EMIT("vmovaps %%ymm1, 0x60(%[state])") // *i = i' |
| 343 |
LANCZOS_GEN_FUNC_X8_FMA3 |
| 344 |
__ASM_EMIT("vmovaps 0x60(%[state]), %%ymm1") // ymm1 = i' |
| 345 |
__ASM_EMIT("vmovups %%ymm0, 0x00(%[dst])") |
| 346 |
__ASM_EMIT("add $0x20, %[dst]") |
| 347 |
__ASM_EMIT("sub $8, %[count]") |
| 348 |
__ASM_EMIT("jae 1b") |
| 349 |
__ASM_EMIT("2:") |
| 350 |
// x4 block |
| 351 |
__ASM_EMIT("add $4, %[count]") |
| 352 |
__ASM_EMIT("jl 4f") |
| 353 |
__ASM_EMIT("vmulps 0x00(%[state]), %%xmm1, %%xmm0") // xmm0 = k*i |
| 354 |
__ASM_EMIT("vaddps 0x80 + %[LGEN], %%xmm1, %%xmm1") // xmm1 = i' = i + step |
| 355 |
__ASM_EMIT("vsubps 0x20(%[state]), %%xmm0, %%xmm0") // xmm0 = x1 = k*i - p |
| 356 |
__ASM_EMIT("vmovaps %%xmm1, 0x60(%[state])") // *i = i' |
| 357 |
LANCZOS_GEN_FUNC_X4_FMA3 |
| 358 |
__ASM_EMIT("vmovaps 0x60(%[state]), %%xmm1") // xmm1 = i' |
| 359 |
__ASM_EMIT("vmovups %%xmm0, 0x00(%[dst])") |
| 360 |
__ASM_EMIT("add $0x10, %[dst]") |
| 361 |
__ASM_EMIT("sub $4, %[count]") |
| 362 |
__ASM_EMIT("4:") |
| 363 |
// Tail: 1x-3x block |
| 364 |
__ASM_EMIT("add $4, %[count]") |
| 365 |
__ASM_EMIT("jle 8f") |
| 366 |
__ASM_EMIT("vmulps 0x00(%[state]), %%xmm1, %%xmm0") // xmm0 = k*i |
| 367 |
__ASM_EMIT("vsubps 0x20(%[state]), %%xmm0, %%xmm0") // xmm0 = x1 = k*i - p |
| 368 |
LANCZOS_GEN_FUNC_X4_FMA3 |
| 369 |
__ASM_EMIT("test $1, %[count]") |
| 370 |
__ASM_EMIT("jz 6f") |
| 371 |
__ASM_EMIT("vmovss %%xmm0, 0x00(%[dst])") |
| 372 |
__ASM_EMIT("vshufps $0x39, %%xmm0, %%xmm0, %%xmm0") |
| 373 |
__ASM_EMIT("add $4, %[dst]") |
| 374 |
__ASM_EMIT("6:") |
| 375 |
__ASM_EMIT("test $2, %[count]") |
| 376 |
__ASM_EMIT("jz 8f") |
| 377 |
__ASM_EMIT("vmovlps %%xmm0, 0x00(%[dst])") |
| 378 |
// End |
| 379 |
__ASM_EMIT("8:") |
| 380 |
|
| 381 |
: [dst] "+r" (dst), [count] "+r" (count) |
| 382 |
: [state] "r" (&state), |
| 383 |
[S2C] "o" (sinf_const), |
| 384 |
[LGEN] "o" (kp_gen_const), |
| 385 |
[LC] "o" (lanczos_const), |
| 386 |
[stub] "o" (stub), |
| 387 |
[k] "m" (k), |
| 388 |
[p] "m" (p), |
| 389 |
[a] "m" (a), |
| 390 |
[t] "m" (t) |
| 391 |
: "cc", "memory", |
| 392 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 393 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7" |
| 394 |
); |
| 395 |
} |
| 396 |
|
| 397 |
#define LANCZOS_GEN_X64_FUNC_X8_FMA3 \ |
| 398 |
/* ymm8 = k */ \ |
| 399 |
/* ymm9 = p */ \ |
| 400 |
/* ymm10 = t */ \ |
| 401 |
/* ymm11 = a */ \ |
| 402 |
/* ymm12 = x1 */ \ |
| 403 |
/* ymm13 = d */ \ |
| 404 |
/* ymm14 = i */ \ |
| 405 |
/* ymm15 = step */ \ |
| 406 |
__ASM_EMIT("vmulps %%ymm11, %%ymm12, %%ymm4") /* ymm12 = x2 = x1*a */ \ |
| 407 |
__ASM_EMIT("vmulps %%ymm12, %%ymm4, %%ymm13") /* ymm13 = d = x1 * x2 */ \ |
| 408 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm12, %%ymm0") /* ymm0 = x1 + PI/2 */ \ |
| 409 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm4, %%ymm4") /* ymm4 = x2 + PI/2 */ \ |
| 410 |
SINF_X_PLUS_PI_2_CORE_X16_FMA3 /* ymm0 = sinf(x1), ymm4 = sinf(x2) */ \ |
| 411 |
__ASM_EMIT("vmulps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = sinf(x1)*sinf(x2) */ \ |
| 412 |
__ASM_EMIT("vandps 0x00 + %[LC], %%ymm12, %%ymm12") /* ymm12 = fabsf(x1) */ \ |
| 413 |
__ASM_EMIT("vdivps %%ymm13, %%ymm0, %%ymm0") /* ymm0 = F = (sinf(x1)*sinf(x2)) / (x1 * x2) */ \ |
| 414 |
__ASM_EMIT("vcmpps $1, 0x20 + %[LC], %%ymm12, %%ymm2") /* ymm2 = [ fabsf(x1) < 1e-4 ] */ \ |
| 415 |
__ASM_EMIT("vcmpps $1, %%ymm10, %%ymm12, %%ymm1") /* ymm1 = [ fabsf(x1) < t ] */ \ |
| 416 |
__ASM_EMIT("vblendvps %%ymm2, 0x40 + %[LC], %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x1) >= 1e-4 ] ? f : 1.0 */ \ |
| 417 |
__ASM_EMIT("vandps %%ymm1, %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x1) < t ] ? ([ fabsf(x1) >= 1e-4 ] ? f : 1.0) : 0.0 */ |
| 418 |
|
| 419 |
#define LANCZOS_GEN_X64_FUNC_X4_FMA3 \ |
| 420 |
/* xmm0 = x1 */ \ |
| 421 |
__ASM_EMIT("vmulps %%xmm11, %%xmm12, %%xmm4") /* xmm12 = x2 = x1*a */ \ |
| 422 |
__ASM_EMIT("vmulps %%xmm12, %%xmm4, %%xmm13") /* xmm13 = d = x1 * x2 */ \ |
| 423 |
__ASM_EMIT("vinsertf128 $1, %%xmm4, %%ymm12, %%ymm0") /* ymm0 = x1 x1 x1 x1 x2 x2 x2 x2 */ \ |
| 424 |
__ASM_EMIT("vaddps 0x00 + %[S2C], %%ymm0, %%ymm0") /* ymm0 = x1 + PI/2 ... x2 + PI/2 */ \ |
| 425 |
SINF_X_PLUS_PI_2_CORE_X8_FMA3 /* ymm0 = sinf(x1) ... sinf(x2) */ \ |
| 426 |
__ASM_EMIT("vextractf128 $1, %%ymm0, %%xmm4") /* xmm4 = sif(x2) */ \ |
| 427 |
__ASM_EMIT("vmulps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = sinf(x1)*sinf(x2) */ \ |
| 428 |
__ASM_EMIT("vandps 0x00 + %[LC], %%ymm12, %%ymm12") /* ymm12 = fabsf(x1) */ \ |
| 429 |
__ASM_EMIT("vdivps %%ymm13, %%ymm0, %%ymm0") /* ymm0 = F = (sinf(x1)*sinf(x2)) / (x1 * x2) */ \ |
| 430 |
__ASM_EMIT("vcmpps $1, 0x20 + %[LC], %%ymm12, %%ymm2") /* ymm2 = [ fabsf(x1) < 1e-4 ] */ \ |
| 431 |
__ASM_EMIT("vcmpps $1, %%ymm10, %%ymm12, %%ymm1") /* ymm1 = [ fabsf(x1) < t ] */ \ |
| 432 |
__ASM_EMIT("vblendvps %%ymm2, 0x40 + %[LC], %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x1) >= 1e-4 ] ? f : 1.0 */ \ |
| 433 |
__ASM_EMIT("vandps %%ymm1, %%ymm0, %%ymm0") /* ymm0 = [ fabsf(x1) < t ] ? ([ fabsf(x1) >= 1e-4 ] ? f : 1.0) : 0.0 */ |
| 434 |
|
| 435 |
void x64_lanczos1_fma3(float *dst, float k, float p, float t, float a, size_t count) |
| 436 |
{ |
| 437 |
ARCH_X86_64_ASM( |
| 438 |
// Prepare |
| 439 |
__ASM_EMIT("vbroadcastss %[k], %%ymm8") // ymm8 = k |
| 440 |
__ASM_EMIT("vbroadcastss %[p], %%ymm9") // ymm9 = p |
| 441 |
__ASM_EMIT("vbroadcastss %[t], %%ymm10") // ymm10 = t |
| 442 |
__ASM_EMIT("vbroadcastss %[a], %%ymm11") // ymm11 = a |
| 443 |
__ASM_EMIT("vmovaps 0x00 + %[LGEN], %%ymm14") // ymm14 = i = 0 1 2 3 4 5 6 7 |
| 444 |
// x8 blocks |
| 445 |
__ASM_EMIT("sub $8, %[count]") |
| 446 |
__ASM_EMIT("jb 2f") |
| 447 |
__ASM_EMIT("vmovaps 0x60 + %[LGEN], %%ymm15") // ymm15 = step |
| 448 |
__ASM_EMIT("1:") |
| 449 |
__ASM_EMIT("vmulps %%ymm8, %%ymm14, %%ymm12") // ymm12 = k*i |
| 450 |
__ASM_EMIT("vaddps %%ymm15, %%ymm14, %%ymm14") // ymm14 = i' = i + step |
| 451 |
__ASM_EMIT("vsubps %%ymm9, %%ymm12, %%ymm12") // ymm12 = x1 = k*i - p |
| 452 |
LANCZOS_GEN_X64_FUNC_X8_FMA3 |
| 453 |
__ASM_EMIT("vmovups %%ymm0, 0x00(%[dst])") |
| 454 |
__ASM_EMIT("add $0x20, %[dst]") |
| 455 |
__ASM_EMIT("sub $8, %[count]") |
| 456 |
__ASM_EMIT("jae 1b") |
| 457 |
__ASM_EMIT("2:") |
| 458 |
// x4 block |
| 459 |
__ASM_EMIT("add $4, %[count]") |
| 460 |
__ASM_EMIT("jl 4f") |
| 461 |
__ASM_EMIT("vmulps %%xmm8, %%xmm14, %%xmm12") // xmm12 = k*i |
| 462 |
__ASM_EMIT("vaddps 0x80 + %[LGEN], %%xmm14, %%xmm14") // xmm14 = i' = i + step |
| 463 |
__ASM_EMIT("vsubps %%xmm9, %%xmm12, %%xmm12") // xmm12 = x1 = k*i - p |
| 464 |
LANCZOS_GEN_X64_FUNC_X4_FMA3 |
| 465 |
__ASM_EMIT("vmovups %%xmm0, 0x00(%[dst])") |
| 466 |
__ASM_EMIT("add $0x10, %[dst]") |
| 467 |
__ASM_EMIT("sub $4, %[count]") |
| 468 |
__ASM_EMIT("4:") |
| 469 |
// Tail: 1x-3x block |
| 470 |
__ASM_EMIT("add $4, %[count]") |
| 471 |
__ASM_EMIT("jle 8f") |
| 472 |
__ASM_EMIT("vmulps %%xmm8, %%xmm14, %%xmm12") // xmm12 = k*i |
| 473 |
__ASM_EMIT("vsubps %%xmm9, %%xmm12, %%xmm12") // xmm12 = x1 = k*i - p |
| 474 |
LANCZOS_GEN_X64_FUNC_X4_FMA3 |
| 475 |
__ASM_EMIT("test $1, %[count]") |
| 476 |
__ASM_EMIT("jz 6f") |
| 477 |
__ASM_EMIT("vmovss %%xmm0, 0x00(%[dst])") |
| 478 |
__ASM_EMIT("vshufps $0x39, %%xmm0, %%xmm0, %%xmm0") |
| 479 |
__ASM_EMIT("add $4, %[dst]") |
| 480 |
__ASM_EMIT("6:") |
| 481 |
__ASM_EMIT("test $2, %[count]") |
| 482 |
__ASM_EMIT("jz 8f") |
| 483 |
__ASM_EMIT("vmovlps %%xmm0, 0x00(%[dst])") |
| 484 |
// End |
| 485 |
__ASM_EMIT("8:") |
| 486 |
|
| 487 |
: [dst] "+r" (dst), [count] "+r" (count) |
| 488 |
: [S2C] "o" (sinf_const), |
| 489 |
[LGEN] "o" (kp_gen_const), |
| 490 |
[LC] "o" (lanczos_const), |
| 491 |
[k] "m" (k), |
| 492 |
[p] "m" (p), |
| 493 |
[a] "m" (a), |
| 494 |
[t] "m" (t) |
| 495 |
: "cc", "memory", |
| 496 |
"%xmm0", "%xmm1", "%xmm2", "%xmm3", |
| 497 |
"%xmm4", "%xmm5", "%xmm6", "%xmm7", |
| 498 |
"%xmm8", "%xmm9", "%xmm10", "%xmm11", |
| 499 |
"%xmm12", "%xmm13", "%xmm14", "%xmm15" |
| 500 |
); |
| 501 |
} |
| 502 |
|
| 503 |
} /* namespace avx2 */ |
| 504 |
} /* namespace lsp */ |
| 505 |
|
| 506 |
|
| 507 |
#endif /* PRIVATE_DSP_ARCH_X86_AVX2_PMATH_LANCZOS_H_ */ |