| 1 |
/* |
| 2 |
* MP3 quantization, intrinsics functions |
| 3 |
* |
| 4 |
* Copyright (c) 2005-2006 Gabriel Bouvigne |
| 5 |
* |
| 6 |
* This library is free software; you can redistribute it and/or |
| 7 |
* modify it under the terms of the GNU Library General Public |
| 8 |
* License as published by the Free Software Foundation; either |
| 9 |
* version 2 of the License, or (at your option) any later version. |
| 10 |
* |
| 11 |
* This library is distributed in the hope that it will be useful, |
| 12 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 14 |
* Library General Public License for more details. |
| 15 |
* |
| 16 |
* You should have received a copy of the GNU Library General Public |
| 17 |
* License along with this library; if not, write to the |
| 18 |
* Free Software Foundation, Inc., 59 Temple Place - Suite 330, |
| 19 |
* Boston, MA 02111-1307, USA. |
| 20 |
*/ |
| 21 |
|
| 22 |
|
| 23 |
#ifdef HAVE_CONFIG_H |
| 24 |
# include <config.h> |
| 25 |
#endif |
| 26 |
|
| 27 |
#include "lame.h" |
| 28 |
#include "machine.h" |
| 29 |
#include "encoder.h" |
| 30 |
#include "util.h" |
| 31 |
#include "lame_intrin.h" |
| 32 |
|
| 33 |
|
| 34 |
|
| 35 |
#ifdef HAVE_XMMINTRIN_H |
| 36 |
|
| 37 |
#include <xmmintrin.h> |
| 38 |
|
| 39 |
typedef union { |
| 40 |
int32_t _i_32[4]; /* unions are initialized by its first member */ |
| 41 |
float _float[4]; |
| 42 |
__m128 _m128; |
| 43 |
} vecfloat_union; |
| 44 |
|
| 45 |
#define TRI_SIZE (5-1) /* 1024 = 4**5 */ |
| 46 |
static const FLOAT costab[TRI_SIZE * 2] = { |
| 47 |
9.238795325112867e-01, 3.826834323650898e-01, |
| 48 |
9.951847266721969e-01, 9.801714032956060e-02, |
| 49 |
9.996988186962042e-01, 2.454122852291229e-02, |
| 50 |
9.999811752826011e-01, 6.135884649154475e-03 |
| 51 |
}; |
| 52 |
|
| 53 |
|
| 54 |
/* make sure functions with SSE instructions maintain their own properly aligned stack */ |
| 55 |
#if defined (__GNUC__) && ((__GNUC__ > 4) || ((__GNUC__ == 4) && (__GNUC_MINOR__ >= 2))) |
| 56 |
#define REALIGN __attribute__((force_align_arg_pointer)) |
| 57 |
#define TARGET(x) __attribute__((target(x))) |
| 58 |
#else |
| 59 |
#define REALIGN |
| 60 |
#define TARGET(x) |
| 61 |
#endif |
| 62 |
|
| 63 |
#define SSE_FUNCTION REALIGN TARGET("sse2") |
| 64 |
|
| 65 |
|
| 66 |
SSE_FUNCTION void |
| 67 |
init_xrpow_core_sse(gr_info * const cod_info, FLOAT xrpow[576], int max_nz, FLOAT * sum) |
| 68 |
{ |
| 69 |
int i; |
| 70 |
float tmp_max = 0; |
| 71 |
float tmp_sum = 0; |
| 72 |
int upper = max_nz + 1; |
| 73 |
int upper4 = (upper / 4) * 4; |
| 74 |
int rest = upper-upper4; |
| 75 |
|
| 76 |
const vecfloat_union fabs_mask = {{ 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF }}; |
| 77 |
const __m128 vec_fabs_mask = _mm_loadu_ps(&fabs_mask._float[0]); |
| 78 |
vecfloat_union vec_xrpow_max; |
| 79 |
vecfloat_union vec_sum; |
| 80 |
vecfloat_union vec_tmp; |
| 81 |
|
| 82 |
_mm_prefetch((char const *) cod_info->xr, _MM_HINT_T0); |
| 83 |
_mm_prefetch((char const *) xrpow, _MM_HINT_T0); |
| 84 |
|
| 85 |
vec_xrpow_max._m128 = _mm_set_ps1(0); |
| 86 |
vec_sum._m128 = _mm_set_ps1(0); |
| 87 |
|
| 88 |
for (i = 0; i < upper4; i += 4) { |
| 89 |
vec_tmp._m128 = _mm_loadu_ps(&(cod_info->xr[i])); /* load */ |
| 90 |
vec_tmp._m128 = _mm_and_ps(vec_tmp._m128, vec_fabs_mask); /* fabs */ |
| 91 |
vec_sum._m128 = _mm_add_ps(vec_sum._m128, vec_tmp._m128); |
| 92 |
vec_tmp._m128 = _mm_sqrt_ps(_mm_mul_ps(vec_tmp._m128, _mm_sqrt_ps(vec_tmp._m128))); |
| 93 |
vec_xrpow_max._m128 = _mm_max_ps(vec_xrpow_max._m128, vec_tmp._m128); /* retrieve max */ |
| 94 |
_mm_storeu_ps(&(xrpow[i]), vec_tmp._m128); /* store into xrpow[] */ |
| 95 |
} |
| 96 |
vec_tmp._m128 = _mm_set_ps1(0); |
| 97 |
switch (rest) { |
| 98 |
case 3: vec_tmp._float[2] = cod_info->xr[upper4+2]; |
| 99 |
case 2: vec_tmp._float[1] = cod_info->xr[upper4+1]; |
| 100 |
case 1: vec_tmp._float[0] = cod_info->xr[upper4+0]; |
| 101 |
vec_tmp._m128 = _mm_and_ps(vec_tmp._m128, vec_fabs_mask); /* fabs */ |
| 102 |
vec_sum._m128 = _mm_add_ps(vec_sum._m128, vec_tmp._m128); |
| 103 |
vec_tmp._m128 = _mm_sqrt_ps(_mm_mul_ps(vec_tmp._m128, _mm_sqrt_ps(vec_tmp._m128))); |
| 104 |
vec_xrpow_max._m128 = _mm_max_ps(vec_xrpow_max._m128, vec_tmp._m128); /* retrieve max */ |
| 105 |
switch (rest) { |
| 106 |
case 3: xrpow[upper4+2] = vec_tmp._float[2]; |
| 107 |
case 2: xrpow[upper4+1] = vec_tmp._float[1]; |
| 108 |
case 1: xrpow[upper4+0] = vec_tmp._float[0]; |
| 109 |
default: |
| 110 |
break; |
| 111 |
} |
| 112 |
default: |
| 113 |
break; |
| 114 |
} |
| 115 |
tmp_sum = vec_sum._float[0] + vec_sum._float[1] + vec_sum._float[2] + vec_sum._float[3]; |
| 116 |
{ |
| 117 |
float ma = vec_xrpow_max._float[0] > vec_xrpow_max._float[1] |
| 118 |
? vec_xrpow_max._float[0] : vec_xrpow_max._float[1]; |
| 119 |
float mb = vec_xrpow_max._float[2] > vec_xrpow_max._float[3] |
| 120 |
? vec_xrpow_max._float[2] : vec_xrpow_max._float[3]; |
| 121 |
tmp_max = ma > mb ? ma : mb; |
| 122 |
} |
| 123 |
cod_info->xrpow_max = tmp_max; |
| 124 |
*sum = tmp_sum; |
| 125 |
} |
| 126 |
|
| 127 |
|
| 128 |
SSE_FUNCTION static void |
| 129 |
store4(__m128 v, float* f0, float* f1, float* f2, float* f3) |
| 130 |
{ |
| 131 |
vecfloat_union r; |
| 132 |
r._m128 = v; |
| 133 |
*f0 = r._float[0]; |
| 134 |
*f1 = r._float[1]; |
| 135 |
*f2 = r._float[2]; |
| 136 |
*f3 = r._float[3]; |
| 137 |
} |
| 138 |
|
| 139 |
|
| 140 |
SSE_FUNCTION void |
| 141 |
fht_SSE2(FLOAT * fz, int n) |
| 142 |
{ |
| 143 |
const FLOAT *tri = costab; |
| 144 |
int k4; |
| 145 |
FLOAT *fi, *gi; |
| 146 |
FLOAT const *fn; |
| 147 |
|
| 148 |
n <<= 1; /* to get BLKSIZE, because of 3DNow! ASM routine */ |
| 149 |
fn = fz + n; |
| 150 |
k4 = 4; |
| 151 |
do { |
| 152 |
FLOAT s1, c1; |
| 153 |
int i, k1, k2, k3, kx; |
| 154 |
kx = k4 >> 1; |
| 155 |
k1 = k4; |
| 156 |
k2 = k4 << 1; |
| 157 |
k3 = k2 + k1; |
| 158 |
k4 = k2 << 1; |
| 159 |
fi = fz; |
| 160 |
gi = fi + kx; |
| 161 |
do { |
| 162 |
FLOAT f0, f1, f2, f3; |
| 163 |
f1 = fi[0] - fi[k1]; |
| 164 |
f0 = fi[0] + fi[k1]; |
| 165 |
f3 = fi[k2] - fi[k3]; |
| 166 |
f2 = fi[k2] + fi[k3]; |
| 167 |
fi[k2] = f0 - f2; |
| 168 |
fi[0] = f0 + f2; |
| 169 |
fi[k3] = f1 - f3; |
| 170 |
fi[k1] = f1 + f3; |
| 171 |
f1 = gi[0] - gi[k1]; |
| 172 |
f0 = gi[0] + gi[k1]; |
| 173 |
f3 = SQRT2 * gi[k3]; |
| 174 |
f2 = SQRT2 * gi[k2]; |
| 175 |
gi[k2] = f0 - f2; |
| 176 |
gi[0] = f0 + f2; |
| 177 |
gi[k3] = f1 - f3; |
| 178 |
gi[k1] = f1 + f3; |
| 179 |
gi += k4; |
| 180 |
fi += k4; |
| 181 |
} while (fi < fn); |
| 182 |
c1 = tri[0]; |
| 183 |
s1 = tri[1]; |
| 184 |
for (i = 1; i < kx; i++) { |
| 185 |
__m128 v_s2; |
| 186 |
__m128 v_c2; |
| 187 |
__m128 v_c1; |
| 188 |
__m128 v_s1; |
| 189 |
FLOAT c2, s2, s1_2 = s1+s1; |
| 190 |
c2 = 1 - s1_2 * s1; |
| 191 |
s2 = s1_2 * c1; |
| 192 |
fi = fz + i; |
| 193 |
gi = fz + k1 - i; |
| 194 |
v_c1 = _mm_set_ps1(c1); |
| 195 |
v_s1 = _mm_set_ps1(s1); |
| 196 |
v_c2 = _mm_set_ps1(c2); |
| 197 |
v_s2 = _mm_set_ps1(s2); |
| 198 |
{ |
| 199 |
static const vecfloat_union sign_mask = {{0x80000000,0,0,0}}; |
| 200 |
v_c1 = _mm_xor_ps(sign_mask._m128, v_c1); /* v_c1 := {-c1, +c1, +c1, +c1} */ |
| 201 |
} |
| 202 |
{ |
| 203 |
static const vecfloat_union sign_mask = {{0,0x80000000,0,0}}; |
| 204 |
v_s1 = _mm_xor_ps(sign_mask._m128, v_s1); /* v_s1 := {+s1, -s1, +s1, +s1} */ |
| 205 |
} |
| 206 |
{ |
| 207 |
static const vecfloat_union sign_mask = {{0,0,0x80000000,0x80000000}}; |
| 208 |
v_c2 = _mm_xor_ps(sign_mask._m128, v_c2); /* v_c2 := {+c2, +c2, -c2, -c2} */ |
| 209 |
} |
| 210 |
do { |
| 211 |
__m128 p, q, r; |
| 212 |
|
| 213 |
q = _mm_setr_ps(fi[k1], fi[k3], gi[k1], gi[k3]); /* Q := {fi_k1,fi_k3,gi_k1,gi_k3}*/ |
| 214 |
p = _mm_mul_ps(v_s2, q); /* P := s2 * Q */ |
| 215 |
q = _mm_mul_ps(v_c2, q); /* Q := c2 * Q */ |
| 216 |
q = _mm_shuffle_ps(q, q, _MM_SHUFFLE(1,0,3,2)); /* Q := {-c2*gi_k1,-c2*gi_k3,c2*fi_k1,c2*fi_k3} */ |
| 217 |
p = _mm_add_ps(p, q); |
| 218 |
|
| 219 |
r = _mm_setr_ps(gi[0], gi[k2], fi[0], fi[k2]); /* R := {gi_0,gi_k2,fi_0,fi_k2} */ |
| 220 |
q = _mm_sub_ps(r, p); /* Q := {gi_0-p0,gi_k2-p1,fi_0-p2,fi_k2-p3} */ |
| 221 |
r = _mm_add_ps(r, p); /* R := {gi_0+p0,gi_k2+p1,fi_0+p2,fi_k2+p3} */ |
| 222 |
p = _mm_shuffle_ps(q, r, _MM_SHUFFLE(2,0,2,0)); /* P := {q0,q2,r0,r2} */ |
| 223 |
p = _mm_shuffle_ps(p, p, _MM_SHUFFLE(3,1,2,0)); /* P := {q0,r0,q2,r2} */ |
| 224 |
q = _mm_shuffle_ps(q, r, _MM_SHUFFLE(3,1,3,1)); /* Q := {q1,q3,r1,r3} */ |
| 225 |
r = _mm_mul_ps(v_c1, q); |
| 226 |
q = _mm_mul_ps(v_s1, q); |
| 227 |
q = _mm_shuffle_ps(q, q, _MM_SHUFFLE(0,1,2,3)); /* Q := {q3,q2,q1,q0} */ |
| 228 |
q = _mm_add_ps(q, r); |
| 229 |
|
| 230 |
store4(_mm_sub_ps(p, q), &gi[k3], &gi[k2], &fi[k3], &fi[k2]); |
| 231 |
store4(_mm_add_ps(p, q), &gi[k1], &gi[ 0], &fi[k1], &fi[ 0]); |
| 232 |
|
| 233 |
gi += k4; |
| 234 |
fi += k4; |
| 235 |
} while (fi < fn); |
| 236 |
c2 = c1; |
| 237 |
c1 = c2 * tri[0] - s1 * tri[1]; |
| 238 |
s1 = c2 * tri[1] + s1 * tri[0]; |
| 239 |
} |
| 240 |
tri += 2; |
| 241 |
} while (k4 < n); |
| 242 |
} |
| 243 |
|
| 244 |
#endif /* HAVE_XMMINTRIN_H */ |
| 245 |
|