1 /****************************************************************************** 2 * 3 * Copyright (C) 2014 The Android Open Source Project 4 * Copyright 2003 - 2004 Open Interface North America, Inc. All rights reserved. 5 * 6 * Licensed under the Apache License, Version 2.0 (the "License"); 7 * you may not use this file except in compliance with the License. 8 * You may obtain a copy of the License at: 9 * 10 * http://www.apache.org/licenses/LICENSE-2.0 11 * 12 * Unless required by applicable law or agreed to in writing, software 13 * distributed under the License is distributed on an "AS IS" BASIS, 14 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 15 * See the License for the specific language governing permissions and 16 * limitations under the License. 17 * 18 ******************************************************************************/ 19 20 /********************************************************************************** 21 $Revision: #1 $ 22 ***********************************************************************************/ 23 24 /** @file 25 @ingroup codec_internal 26 */ 27 28 /**@addgroup codec_internal*/ 29 /**@{*/ 30 31 /* 32 * Performs an 8-point Type-II scaled DCT using the Arai-Agui-Nakajima 33 * factorization. The scaling factors are folded into the windowing 34 * constants. 29 adds and 5 16x32 multiplies per 8 samples. 35 */ 36 37 #include "oi_codec_sbc_private.h" 38 39 #define AAN_C4_FIX (759250125)/* S1.30 759250125 0.707107*/ 40 41 #define AAN_C6_FIX (410903207)/* S1.30 410903207 0.382683*/ 42 43 #define AAN_Q0_FIX (581104888)/* S1.30 581104888 0.541196*/ 44 45 #define AAN_Q1_FIX (1402911301)/* S1.30 1402911301 1.306563*/ 46 47 /** Scales x by y bits to the right, adding a rounding factor. 48 */ 49 #ifndef SCALE 50 #define SCALE(x, y) (((x) + (1 <<((y)-1))) >> (y)) 51 #endif 52 53 /** 54 * Default C language implementation of a 32x32->32 multiply. This function may 55 * be replaced by a platform-specific version for speed. 56 * 57 * @param u A signed 32-bit multiplicand 58 * @param v A signed 32-bit multiplier 59 60 * @return A signed 32-bit value corresponding to the 32 most significant bits 61 * of the 64-bit product of u and v. 62 */ 63 INLINE OI_INT32 default_mul_32s_32s_hi(OI_INT32 u, OI_INT32 v) 64 { 65 OI_UINT32 u0, v0; 66 OI_INT32 u1, v1, w1, w2, t; 67 68 u0 = u & 0xFFFF; u1 = u >> 16; 69 v0 = v & 0xFFFF; v1 = v >> 16; 70 t = u0*v0; 71 t = u1*v0 + ((OI_UINT32)t >> 16); 72 w1 = t & 0xFFFF; 73 w2 = t >> 16; 74 w1 = u0*v1 + w1; 75 return u1*v1 + w2 + (w1 >> 16); 76 } 77 78 #define MUL_32S_32S_HI(_x, _y) default_mul_32s_32s_hi(_x, _y) 79 80 81 #ifdef DEBUG_DCT 82 PRIVATE void float_dct2_8(float * RESTRICT out, OI_INT32 const *RESTRICT in) 83 { 84 #define FIX(x,bits) (((int)floor(0.5f+((x)*((float)(1<<bits)))))/((float)(1<<bits))) 85 #define FLOAT_BUTTERFLY(x,y) x += y; y = x - (y*2); OI_ASSERT(VALID_INT32(x)); OI_ASSERT(VALID_INT32(y)); 86 #define FLOAT_MULT_DCT(K, sample) (FIX(K,20) * sample) 87 #define FLOAT_SCALE(x, y) (((x) / (double)(1 << (y)))) 88 89 double L00,L01,L02,L03,L04,L05,L06,L07; 90 double L25; 91 92 double in0,in1,in2,in3; 93 double in4,in5,in6,in7; 94 95 in0 = FLOAT_SCALE(in[0], DCTII_8_SHIFT_IN); OI_ASSERT(VALID_INT32(in0)); 96 in1 = FLOAT_SCALE(in[1], DCTII_8_SHIFT_IN); OI_ASSERT(VALID_INT32(in1)); 97 in2 = FLOAT_SCALE(in[2], DCTII_8_SHIFT_IN); OI_ASSERT(VALID_INT32(in2)); 98 in3 = FLOAT_SCALE(in[3], DCTII_8_SHIFT_IN); OI_ASSERT(VALID_INT32(in3)); 99 in4 = FLOAT_SCALE(in[4], DCTII_8_SHIFT_IN); OI_ASSERT(VALID_INT32(in4)); 100 in5 = FLOAT_SCALE(in[5], DCTII_8_SHIFT_IN); OI_ASSERT(VALID_INT32(in5)); 101 in6 = FLOAT_SCALE(in[6], DCTII_8_SHIFT_IN); OI_ASSERT(VALID_INT32(in6)); 102 in7 = FLOAT_SCALE(in[7], DCTII_8_SHIFT_IN); OI_ASSERT(VALID_INT32(in7)); 103 104 L00 = (in0 + in7); OI_ASSERT(VALID_INT32(L00)); 105 L01 = (in1 + in6); OI_ASSERT(VALID_INT32(L01)); 106 L02 = (in2 + in5); OI_ASSERT(VALID_INT32(L02)); 107 L03 = (in3 + in4); OI_ASSERT(VALID_INT32(L03)); 108 109 L04 = (in3 - in4); OI_ASSERT(VALID_INT32(L04)); 110 L05 = (in2 - in5); OI_ASSERT(VALID_INT32(L05)); 111 L06 = (in1 - in6); OI_ASSERT(VALID_INT32(L06)); 112 L07 = (in0 - in7); OI_ASSERT(VALID_INT32(L07)); 113 114 FLOAT_BUTTERFLY(L00, L03); 115 FLOAT_BUTTERFLY(L01, L02); 116 117 L02 += L03; OI_ASSERT(VALID_INT32(L02)); 118 119 L02 = FLOAT_MULT_DCT(AAN_C4_FLOAT, L02); OI_ASSERT(VALID_INT32(L02)); 120 121 FLOAT_BUTTERFLY(L00, L01); 122 123 out[0] = (float)FLOAT_SCALE(L00, DCTII_8_SHIFT_0); OI_ASSERT(VALID_INT16(out[0])); 124 out[4] = (float)FLOAT_SCALE(L01, DCTII_8_SHIFT_4); OI_ASSERT(VALID_INT16(out[4])); 125 126 FLOAT_BUTTERFLY(L03, L02); 127 out[6] = (float)FLOAT_SCALE(L02, DCTII_8_SHIFT_6); OI_ASSERT(VALID_INT16(out[6])); 128 out[2] = (float)FLOAT_SCALE(L03, DCTII_8_SHIFT_2); OI_ASSERT(VALID_INT16(out[2])); 129 130 L04 += L05; OI_ASSERT(VALID_INT32(L04)); 131 L05 += L06; OI_ASSERT(VALID_INT32(L05)); 132 L06 += L07; OI_ASSERT(VALID_INT32(L06)); 133 134 L04/=2; 135 L05/=2; 136 L06/=2; 137 L07/=2; 138 139 L05 = FLOAT_MULT_DCT(AAN_C4_FLOAT, L05); OI_ASSERT(VALID_INT32(L05)); 140 141 L25 = L06 - L04; OI_ASSERT(VALID_INT32(L25)); 142 L25 = FLOAT_MULT_DCT(AAN_C6_FLOAT, L25); OI_ASSERT(VALID_INT32(L25)); 143 144 L04 = FLOAT_MULT_DCT(AAN_Q0_FLOAT, L04); OI_ASSERT(VALID_INT32(L04)); 145 L04 -= L25; OI_ASSERT(VALID_INT32(L04)); 146 147 L06 = FLOAT_MULT_DCT(AAN_Q1_FLOAT, L06); OI_ASSERT(VALID_INT32(L06)); 148 L06 -= L25; OI_ASSERT(VALID_INT32(L25)); 149 150 FLOAT_BUTTERFLY(L07, L05); 151 152 FLOAT_BUTTERFLY(L05, L04); 153 out[3] = (float)(FLOAT_SCALE(L04, DCTII_8_SHIFT_3-1)); OI_ASSERT(VALID_INT16(out[3])); 154 out[5] = (float)(FLOAT_SCALE(L05, DCTII_8_SHIFT_5-1)); OI_ASSERT(VALID_INT16(out[5])); 155 156 FLOAT_BUTTERFLY(L07, L06); 157 out[7] = (float)(FLOAT_SCALE(L06, DCTII_8_SHIFT_7-1)); OI_ASSERT(VALID_INT16(out[7])); 158 out[1] = (float)(FLOAT_SCALE(L07, DCTII_8_SHIFT_1-1)); OI_ASSERT(VALID_INT16(out[1])); 159 } 160 #undef BUTTERFLY 161 #endif 162 163 164 /* 165 * This function calculates the AAN DCT. Its inputs are in S16.15 format, as 166 * returned by OI_SBC_Dequant. In practice, abs(in[x]) < 52429.0 / 1.38 167 * (1244918057 integer). The function it computes is an approximation to the array defined 168 * by: 169 * 170 * diag(aan_s) * AAN= C2 171 * 172 * or 173 * 174 * AAN = diag(1/aan_s) * C2 175 * 176 * where C2 is as it is defined in the comment at the head of this file, and 177 * 178 * aan_s[i] = aan_s = 1/(2*cos(i*pi/16)) with i = 1..7, aan_s[0] = 1; 179 * 180 * aan_s[i] = [ 1.000 0.510 0.541 0.601 0.707 0.900 1.307 2.563 ] 181 * 182 * The output ranges are shown as follows: 183 * 184 * Let Y[0..7] = AAN * X[0..7] 185 * 186 * Without loss of generality, assume the input vector X consists of elements 187 * between -1 and 1. The maximum possible value of a given output element occurs 188 * with some particular combination of input vector elements each of which is -1 189 * or 1. Consider the computation of Y[i]. Y[i] = sum t=0..7 of AAN[t,i]*X[i]. Y is 190 * maximized if the sign of X[i] matches the sign of AAN[t,i], ensuring a 191 * positive contribution to the sum. Equivalently, one may simply sum 192 * abs(AAN)[t,i] over t to get the maximum possible value of Y[i]. 193 * 194 * This yields approximately [8.00 10.05 9.66 8.52 8.00 5.70 4.00 2.00] 195 * 196 * Given the maximum magnitude sensible input value of +/-37992, this yields the 197 * following vector of maximum output magnitudes: 198 * 199 * [ 303936 381820 367003 323692 303936 216555 151968 75984 ] 200 * 201 * Ultimately, these values must fit into 16 bit signed integers, so they must 202 * be scaled. A non-uniform scaling helps maximize the kept precision. The 203 * relative number of extra bits of precision maintainable with respect to the 204 * largest value is given here: 205 * 206 * [ 0 0 0 0 0 0 1 2 ] 207 * 208 */ 209 PRIVATE void dct2_8(SBC_BUFFER_T * RESTRICT out, OI_INT32 const *RESTRICT in) 210 { 211 #define BUTTERFLY(x,y) x += y; y = x - (y<<1); 212 #define FIX_MULT_DCT(K, x) (MUL_32S_32S_HI(K,x)<<2) 213 214 OI_INT32 L00,L01,L02,L03,L04,L05,L06,L07; 215 OI_INT32 L25; 216 217 OI_INT32 in0,in1,in2,in3; 218 OI_INT32 in4,in5,in6,in7; 219 220 #if DCTII_8_SHIFT_IN != 0 221 in0 = SCALE(in[0], DCTII_8_SHIFT_IN); 222 in1 = SCALE(in[1], DCTII_8_SHIFT_IN); 223 in2 = SCALE(in[2], DCTII_8_SHIFT_IN); 224 in3 = SCALE(in[3], DCTII_8_SHIFT_IN); 225 in4 = SCALE(in[4], DCTII_8_SHIFT_IN); 226 in5 = SCALE(in[5], DCTII_8_SHIFT_IN); 227 in6 = SCALE(in[6], DCTII_8_SHIFT_IN); 228 in7 = SCALE(in[7], DCTII_8_SHIFT_IN); 229 #else 230 in0 = in[0]; 231 in1 = in[1]; 232 in2 = in[2]; 233 in3 = in[3]; 234 in4 = in[4]; 235 in5 = in[5]; 236 in6 = in[6]; 237 in7 = in[7]; 238 #endif 239 240 L00 = in0 + in7; 241 L01 = in1 + in6; 242 L02 = in2 + in5; 243 L03 = in3 + in4; 244 245 L04 = in3 - in4; 246 L05 = in2 - in5; 247 L06 = in1 - in6; 248 L07 = in0 - in7; 249 250 BUTTERFLY(L00, L03); 251 BUTTERFLY(L01, L02); 252 253 L02 += L03; 254 255 L02 = FIX_MULT_DCT(AAN_C4_FIX, L02); 256 257 BUTTERFLY(L00, L01); 258 259 out[0] = (OI_INT16)SCALE(L00, DCTII_8_SHIFT_0); 260 out[4] = (OI_INT16)SCALE(L01, DCTII_8_SHIFT_4); 261 262 BUTTERFLY(L03, L02); 263 out[6] = (OI_INT16)SCALE(L02, DCTII_8_SHIFT_6); 264 out[2] = (OI_INT16)SCALE(L03, DCTII_8_SHIFT_2); 265 266 L04 += L05; 267 L05 += L06; 268 L06 += L07; 269 270 L04/=2; 271 L05/=2; 272 L06/=2; 273 L07/=2; 274 275 L05 = FIX_MULT_DCT(AAN_C4_FIX, L05); 276 277 L25 = L06 - L04; 278 L25 = FIX_MULT_DCT(AAN_C6_FIX, L25); 279 280 L04 = FIX_MULT_DCT(AAN_Q0_FIX, L04); 281 L04 -= L25; 282 283 L06 = FIX_MULT_DCT(AAN_Q1_FIX, L06); 284 L06 -= L25; 285 286 BUTTERFLY(L07, L05); 287 288 BUTTERFLY(L05, L04); 289 out[3] = (OI_INT16)SCALE(L04, DCTII_8_SHIFT_3-1); 290 out[5] = (OI_INT16)SCALE(L05, DCTII_8_SHIFT_5-1); 291 292 BUTTERFLY(L07, L06); 293 out[7] = (OI_INT16)SCALE(L06, DCTII_8_SHIFT_7-1); 294 out[1] = (OI_INT16)SCALE(L07, DCTII_8_SHIFT_1-1); 295 #undef BUTTERFLY 296 297 #ifdef DEBUG_DCT 298 { 299 float float_out[8]; 300 float_dct2_8(float_out, in); 301 } 302 #endif 303 } 304 305 /**@}*/ 306
