1 /******************************************************************************
2 *
3 * Copyright 1999-2012 Broadcom Corporation
4 *
5 * Licensed under the Apache License, Version 2.0 (the "License");
6 * you may not use this file except in compliance with the License.
7 * You may obtain a copy of the License at:
8 *
9 * http://www.apache.org/licenses/LICENSE-2.0
10 *
11 * Unless required by applicable law or agreed to in writing, software
12 * distributed under the License is distributed on an "AS IS" BASIS,
13 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 * See the License for the specific language governing permissions and
15 * limitations under the License.
16 *
17 ******************************************************************************/
18
19 /******************************************************************************
20 *
21 * This file contains the code that performs Analysis of the input audio
22 * stream.
23 *
24 ******************************************************************************/
25 #include <string.h>
26 #include "sbc_enc_func_declare.h"
27 #include "sbc_encoder.h"
28 /*#include <math.h>*/
29
30 #if (SBC_IS_64_MULT_IN_WINDOW_ACCU == TRUE)
31 #define WIND_4_SUBBANDS_0_1 \
32 (int32_t)0x01659F45 /* gas32CoeffFor4SBs[8] = -gas32CoeffFor4SBs[32] = \
33 0x01659F45 */
34 #define WIND_4_SUBBANDS_0_2 \
35 (int32_t)0x115B1ED2 /* gas32CoeffFor4SBs[16] = -gas32CoeffFor4SBs[24] = \
36 0x115B1ED2 */
37 #define WIND_4_SUBBANDS_1_0 \
38 (int32_t)0x001194E6 /* gas32CoeffFor4SBs[1 et 39] = 0x001194E6 */
39 #define WIND_4_SUBBANDS_1_1 \
40 (int32_t)0x029DBAA3 /* gas32CoeffFor4SBs[9 et 31] = 0x029DBAA3 */
41 #define WIND_4_SUBBANDS_1_2 \
42 (int32_t)0x18F55C90 /* gas32CoeffFor4SBs[17 et 23] = 0x18F55C90 */
43 #define WIND_4_SUBBANDS_1_3 \
44 (int32_t)0xF60FAF37 /* gas32CoeffFor4SBs[15 et 25] = 0xF60FAF37 */
45 #define WIND_4_SUBBANDS_1_4 \
46 (int32_t)0xFF9BB9D5 /* gas32CoeffFor4SBs[7 et 33] = 0xFF9BB9D5 */
47 #define WIND_4_SUBBANDS_2_0 \
48 (int32_t)0x0030E2D3 /* gas32CoeffFor4SBs[2 et 38] = 0x0030E2D3 */
49 #define WIND_4_SUBBANDS_2_1 \
50 (int32_t)0x03B23341 /* gas32CoeffFor4SBs[10 et 30] = 0x03B23341 */
51 #define WIND_4_SUBBANDS_2_2 \
52 (int32_t)0x1F91CA46 /* gas32CoeffFor4SBs[18 et 22] = 0x1F91CA46 */
53 #define WIND_4_SUBBANDS_2_3 \
54 (int32_t)0xFC4F91D4 /* gas32CoeffFor4SBs[14 et 26] = 0xFC4F91D4 */
55 #define WIND_4_SUBBANDS_2_4 \
56 (int32_t)0x003D239B /* gas32CoeffFor4SBs[6 et 34] = 0x003D239B */
57 #define WIND_4_SUBBANDS_3_0 \
58 (int32_t)0x00599403 /* gas32CoeffFor4SBs[3 et 37] = 0x00599403 */
59 #define WIND_4_SUBBANDS_3_1 \
60 (int32_t)0x041EEE40 /* gas32CoeffFor4SBs[11 et 29] = 0x041EEE40 */
61 #define WIND_4_SUBBANDS_3_2 \
62 (int32_t)0x2412F251 /* gas32CoeffFor4SBs[19 et 21] = 0x2412F251 */
63 #define WIND_4_SUBBANDS_3_3 \
64 (int32_t)0x00C8F2BC /* gas32CoeffFor4SBs[13 et 27] = 0x00C8F2BC */
65 #define WIND_4_SUBBANDS_3_4 \
66 (int32_t)0x007F88E4 /* gas32CoeffFor4SBs[5 et 35] = 0x007F88E4 */
67 #define WIND_4_SUBBANDS_4_0 \
68 (int32_t)0x007DBCC8 /* gas32CoeffFor4SBs[4 et 36] = 0x007DBCC8 */
69 #define WIND_4_SUBBANDS_4_1 \
70 (int32_t)0x034FEE2C /* gas32CoeffFor4SBs[12 et 28] = 0x034FEE2C */
71 #define WIND_4_SUBBANDS_4_2 \
72 (int32_t)0x25AC1FF2 /* gas32CoeffFor4SBs[20] = 0x25AC1FF2 */
73
74 #define WIND_8_SUBBANDS_0_1 (int32_t)0x00B97348 /* 16 0x00B97348 */
75 #define WIND_8_SUBBANDS_0_2 (int32_t)0x08B4307A /* 32 0x08B4307A */
76 #define WIND_8_SUBBANDS_1_0 (int32_t)0x00052173 /* 1 et 79 = 0x00052173 */
77 #define WIND_8_SUBBANDS_1_1 (int32_t)0x01071B96 /* 17 et 63 = 0x01071B96 */
78 #define WIND_8_SUBBANDS_1_2 (int32_t)0x0A9F3E9A /* 33 et 47 = 0x0A9F3E9A*/
79 #define WIND_8_SUBBANDS_1_3 (int32_t)0xF9312891 /* 31 et 49 = 0xF9312891 */
80 #define WIND_8_SUBBANDS_1_4 (int32_t)0xFF8D6793 /* 15 et 65 = 0xFF8D6793 */
81 #define WIND_8_SUBBANDS_2_0 (int32_t)0x000B3F71 /* 2 et 78 = 0x000B3F71 */
82 #define WIND_8_SUBBANDS_2_1 (int32_t)0x0156B3CA /* 18 et 62 = 0x0156B3CA */
83 #define WIND_8_SUBBANDS_2_2 (int32_t)0x0C7D59B6 /* 34 et 46 = 0x0C7D59B6 */
84 #define WIND_8_SUBBANDS_2_3 (int32_t)0xFAFF95FC /* 30 et 50 = 0xFAFF95FC */
85 #define WIND_8_SUBBANDS_2_4 (int32_t)0xFFC9F10E /* 14 et 66 = 0xFFC9F10E */
86 #define WIND_8_SUBBANDS_3_0 (int32_t)0x00122C7D /* 3 et 77 = 0x00122C7D*/
87 #define WIND_8_SUBBANDS_3_1 (int32_t)0x01A1B38B /* 19 et 61 = 0x01A1B38B */
88 #define WIND_8_SUBBANDS_3_2 (int32_t)0x0E3BB16F /* 35 et 45 = 0x0E3BB16F */
89 #define WIND_8_SUBBANDS_3_3 (int32_t)0xFCA86E7E /* 29 et 51 = 0xFCA86E7E */
90 #define WIND_8_SUBBANDS_3_4 (int32_t)0xFFFA2413 /* 13 et 67 = 0xFFFA2413 */
91 #define WIND_8_SUBBANDS_4_0 (int32_t)0x001AFF89 /* 4 et 66 = 0x001AFF89 */
92 #define WIND_8_SUBBANDS_4_1 (int32_t)0x01E0224C /* 20 et 60 = 0x01E0224C */
93 #define WIND_8_SUBBANDS_4_2 (int32_t)0x0FC721F9 /* 36 et 44 = 0x0FC721F9 */
94 #define WIND_8_SUBBANDS_4_3 (int32_t)0xFE20435D /* 28 et 52 = 0xFE20435D */
95 #define WIND_8_SUBBANDS_4_4 (int32_t)0x001D8FD2 /* 12 et 68 = 0x001D8FD2 */
96 #define WIND_8_SUBBANDS_5_0 (int32_t)0x00255A62 /* 5 et 75 = 0x00255A62 */
97 #define WIND_8_SUBBANDS_5_1 (int32_t)0x0209291F /* 21 et 59 = 0x0209291F */
98 #define WIND_8_SUBBANDS_5_2 (int32_t)0x110ECEF0 /* 37 et 43 = 0x110ECEF0 */
99 #define WIND_8_SUBBANDS_5_3 (int32_t)0xFF5EEB73 /* 27 et 53 = 0xFF5EEB73 */
100 #define WIND_8_SUBBANDS_5_4 (int32_t)0x0034F8B6 /* 11 et 69 = 0x0034F8B6 */
101 #define WIND_8_SUBBANDS_6_0 (int32_t)0x003060F4 /* 6 et 74 = 0x003060F4 */
102 #define WIND_8_SUBBANDS_6_1 (int32_t)0x02138653 /* 22 et 58 = 0x02138653 */
103 #define WIND_8_SUBBANDS_6_2 (int32_t)0x120435FA /* 38 et 42 = 0x120435FA */
104 #define WIND_8_SUBBANDS_6_3 (int32_t)0x005FD0FF /* 26 et 54 = 0x005FD0FF */
105 #define WIND_8_SUBBANDS_6_4 (int32_t)0x00415B75 /* 10 et 70 = 0x00415B75 */
106 #define WIND_8_SUBBANDS_7_0 (int32_t)0x003A72E7 /* 7 et 73 = 0x003A72E7 */
107 #define WIND_8_SUBBANDS_7_1 (int32_t)0x01F5F424 /* 23 et 57 = 0x01F5F424 */
108 #define WIND_8_SUBBANDS_7_2 (int32_t)0x129C226F /* 39 et 41 = 0x129C226F */
109 #define WIND_8_SUBBANDS_7_3 (int32_t)0x01223EBA /* 25 et 55 = 0x01223EBA */
110 #define WIND_8_SUBBANDS_7_4 (int32_t)0x0044EF48 /* 9 et 71 = 0x0044EF48 */
111 #define WIND_8_SUBBANDS_8_0 (int32_t)0x0041EC6A /* 8 et 72 = 0x0041EC6A */
112 #define WIND_8_SUBBANDS_8_1 (int32_t)0x01A7ECEF /* 24 et 56 = 0x01A7ECEF */
113 #define WIND_8_SUBBANDS_8_2 (int32_t)0x12CF6C75 /* 40 = 0x12CF6C75 */
114 #else
115 #define WIND_4_SUBBANDS_0_1 \
116 (int16_t)0x0166 /* gas32CoeffFor4SBs[8] = -gas32CoeffFor4SBs[32] = \
117 0x01659F45 */
118 #define WIND_4_SUBBANDS_0_2 \
119 (int16_t)0x115B /* gas32CoeffFor4SBs[16] = -gas32CoeffFor4SBs[24] = \
120 0x115B1ED2 */
121 #define WIND_4_SUBBANDS_1_0 \
122 (int16_t)0x0012 /* gas32CoeffFor4SBs[1 et 39] = 0x001194E6 */
123 #define WIND_4_SUBBANDS_1_1 \
124 (int16_t)0x029E /* gas32CoeffFor4SBs[9 et 31] = 0x029DBAA3 */
125 #define WIND_4_SUBBANDS_1_2 \
126 (int16_t)0x18F5 /* gas32CoeffFor4SBs[17 et 23] = 0x18F55C90 */
127 #define WIND_4_SUBBANDS_1_3 \
128 (int16_t)0xF610 /* gas32CoeffFor4SBs[15 et 25] = 0xF60FAF37 */
129 #define WIND_4_SUBBANDS_1_4 \
130 (int16_t)0xFF9C /* gas32CoeffFor4SBs[7 et 33] = 0xFF9BB9D5 */
131 #define WIND_4_SUBBANDS_2_0 \
132 (int16_t)0x0031 /* gas32CoeffFor4SBs[2 et 38] = 0x0030E2D3 */
133 #define WIND_4_SUBBANDS_2_1 \
134 (int16_t)0x03B2 /* gas32CoeffFor4SBs[10 et 30] = 0x03B23341 */
135 #define WIND_4_SUBBANDS_2_2 \
136 (int16_t)0x1F91 /* gas32CoeffFor4SBs[18 et 22] = 0x1F91CA46 */
137 #define WIND_4_SUBBANDS_2_3 \
138 (int16_t)0xFC50 /* gas32CoeffFor4SBs[14 et 26] = 0xFC4F91D4 */
139 #define WIND_4_SUBBANDS_2_4 \
140 (int16_t)0x003D /* gas32CoeffFor4SBs[6 et 34] = 0x003D239B */
141 #define WIND_4_SUBBANDS_3_0 \
142 (int16_t)0x005A /* gas32CoeffFor4SBs[3 et 37] = 0x00599403 */
143 #define WIND_4_SUBBANDS_3_1 \
144 (int16_t)0x041F /* gas32CoeffFor4SBs[11 et 29] = 0x041EEE40 */
145 #define WIND_4_SUBBANDS_3_2 \
146 (int16_t)0x2413 /* gas32CoeffFor4SBs[19 et 21] = 0x2412F251 */
147 #define WIND_4_SUBBANDS_3_3 \
148 (int16_t)0x00C9 /* gas32CoeffFor4SBs[13 et 27] = 0x00C8F2BC */
149 #define WIND_4_SUBBANDS_3_4 \
150 (int16_t)0x0080 /* gas32CoeffFor4SBs[5 et 35] = 0x007F88E4 */
151 #define WIND_4_SUBBANDS_4_0 \
152 (int16_t)0x007E /* gas32CoeffFor4SBs[4 et 36] = 0x007DBCC8 */
153 #define WIND_4_SUBBANDS_4_1 \
154 (int16_t)0x0350 /* gas32CoeffFor4SBs[12 et 28] = 0x034FEE2C */
155 #define WIND_4_SUBBANDS_4_2 \
156 (int16_t)0x25AC /* gas32CoeffFor4SBs[20] = 25AC1FF2 */
157
158 #define WIND_8_SUBBANDS_0_1 (int16_t)0x00B9 /* 16 0x12CF6C75 */
159 #define WIND_8_SUBBANDS_0_2 (int16_t)0x08B4 /* 32 0x08B4307A */
160 #define WIND_8_SUBBANDS_1_0 (int16_t)0x0005 /* 1 et 79 = 0x00052173 */
161 #define WIND_8_SUBBANDS_1_1 (int16_t)0x0107 /* 17 et 63 = 0x01071B96 */
162 #define WIND_8_SUBBANDS_1_2 (int16_t)0x0A9F /* 33 et 47 = 0x0A9F3E9A*/
163 #define WIND_8_SUBBANDS_1_3 (int16_t)0xF931 /* 31 et 49 = 0xF9312891 */
164 #define WIND_8_SUBBANDS_1_4 (int16_t)0xFF8D /* 15 et 65 = 0xFF8D6793 */
165 #define WIND_8_SUBBANDS_2_0 (int16_t)0x000B /* 2 et 78 = 0x000B3F71 */
166 #define WIND_8_SUBBANDS_2_1 (int16_t)0x0157 /* 18 et 62 = 0x0156B3CA */
167 #define WIND_8_SUBBANDS_2_2 (int16_t)0x0C7D /* 34 et 46 = 0x0C7D59B6 */
168 #define WIND_8_SUBBANDS_2_3 (int16_t)0xFB00 /* 30 et 50 = 0xFAFF95FC */
169 #define WIND_8_SUBBANDS_2_4 (int16_t)0xFFCA /* 14 et 66 = 0xFFC9F10E */
170 #define WIND_8_SUBBANDS_3_0 (int16_t)0x0012 /* 3 et 77 = 0x00122C7D*/
171 #define WIND_8_SUBBANDS_3_1 (int16_t)0x01A2 /* 19 et 61 = 0x01A1B38B */
172 #define WIND_8_SUBBANDS_3_2 (int16_t)0x0E3C /* 35 et 45 = 0x0E3BB16F */
173 #define WIND_8_SUBBANDS_3_3 (int16_t)0xFCA8 /* 29 et 51 = 0xFCA86E7E */
174 #define WIND_8_SUBBANDS_3_4 (int16_t)0xFFFA /* 13 et 67 = 0xFFFA2413 */
175 #define WIND_8_SUBBANDS_4_0 (int16_t)0x001B /* 4 et 66 = 0x001AFF89 */
176 #define WIND_8_SUBBANDS_4_1 (int16_t)0x01E0 /* 20 et 60 = 0x01E0224C */
177 #define WIND_8_SUBBANDS_4_2 (int16_t)0x0FC7 /* 36 et 44 = 0x0FC721F9 */
178 #define WIND_8_SUBBANDS_4_3 (int16_t)0xFE20 /* 28 et 52 = 0xFE20435D */
179 #define WIND_8_SUBBANDS_4_4 (int16_t)0x001E /* 12 et 68 = 0x001D8FD2 */
180 #define WIND_8_SUBBANDS_5_0 (int16_t)0x0025 /* 5 et 75 = 0x00255A62 */
181 #define WIND_8_SUBBANDS_5_1 (int16_t)0x0209 /* 21 et 59 = 0x0209291F */
182 #define WIND_8_SUBBANDS_5_2 (int16_t)0x110F /* 37 et 43 = 0x110ECEF0 */
183 #define WIND_8_SUBBANDS_5_3 (int16_t)0xFF5F /* 27 et 53 = 0xFF5EEB73 */
184 #define WIND_8_SUBBANDS_5_4 (int16_t)0x0035 /* 11 et 69 = 0x0034F8B6 */
185 #define WIND_8_SUBBANDS_6_0 (int16_t)0x0030 /* 6 et 74 = 0x003060F4 */
186 #define WIND_8_SUBBANDS_6_1 (int16_t)0x0214 /* 22 et 58 = 0x02138653 */
187 #define WIND_8_SUBBANDS_6_2 (int16_t)0x1204 /* 38 et 42 = 0x120435FA */
188 #define WIND_8_SUBBANDS_6_3 (int16_t)0x0060 /* 26 et 54 = 0x005FD0FF */
189 #define WIND_8_SUBBANDS_6_4 (int16_t)0x0041 /* 10 et 70 = 0x00415B75 */
190 #define WIND_8_SUBBANDS_7_0 (int16_t)0x003A /* 7 et 73 = 0x003A72E7 */
191 #define WIND_8_SUBBANDS_7_1 (int16_t)0x01F6 /* 23 et 57 = 0x01F5F424 */
192 #define WIND_8_SUBBANDS_7_2 (int16_t)0x129C /* 39 et 41 = 0x129C226F */
193 #define WIND_8_SUBBANDS_7_3 (int16_t)0x0122 /* 25 et 55 = 0x01223EBA */
194 #define WIND_8_SUBBANDS_7_4 (int16_t)0x0045 /* 9 et 71 = 0x0044EF48 */
195 #define WIND_8_SUBBANDS_8_0 (int16_t)0x0042 /* 8 et 72 = 0x0041EC6A */
196 #define WIND_8_SUBBANDS_8_1 (int16_t)0x01A8 /* 24 et 56 = 0x01A7ECEF */
197 #define WIND_8_SUBBANDS_8_2 (int16_t)0x12CF /* 40 = 0x12CF6C75 */
198 #endif
199
200 #if (SBC_USE_ARM_PRAGMA == TRUE)
201 #pragma arm section zidata = "sbc_s32_analysis_section"
202 #endif
203 static int32_t s32DCTY[16] = {0};
204 static int32_t s32X[ENC_VX_BUFFER_SIZE / 2];
205 static int16_t* s16X =
206 (int16_t*)s32X; /* s16X must be 32 bits aligned cf SHIFTUP_X8_2*/
207 #if (SBC_USE_ARM_PRAGMA == TRUE)
208 #pragma arm section zidata
209 #endif
210
211 /* This macro is for 4 subbands */
212 #define SHIFTUP_X4 \
213 { \
214 ps32X = (int32_t*)(s16X + EncMaxShiftCounter + 38); \
215 for (i = 0; i < 9; i++) { \
216 *ps32X = *(ps32X - 2 - (ShiftCounter >> 1)); \
217 ps32X--; \
218 *ps32X = *(ps32X - 2 - (ShiftCounter >> 1)); \
219 ps32X--; \
220 } \
221 }
222 #define SHIFTUP_X4_2 \
223 { \
224 ps32X = (int32_t*)(s16X + EncMaxShiftCounter + 38); \
225 ps32X2 = (int32_t*)(s16X + (EncMaxShiftCounter << 1) + 78); \
226 for (i = 0; i < 9; i++) { \
227 *ps32X = *(ps32X - 2 - (ShiftCounter >> 1)); \
228 *(ps32X2) = *(ps32X2 - 2 - (ShiftCounter >> 1)); \
229 ps32X--; \
230 ps32X2--; \
231 *ps32X = *(ps32X - 2 - (ShiftCounter >> 1)); \
232 *(ps32X2) = *(ps32X2 - 2 - (ShiftCounter >> 1)); \
233 ps32X--; \
234 ps32X2--; \
235 } \
236 }
237
238 /* This macro is for 8 subbands */
239 #define SHIFTUP_X8 \
240 { \
241 ps32X = (int32_t*)(s16X + EncMaxShiftCounter + 78); \
242 for (i = 0; i < 9; i++) { \
243 *ps32X = *(ps32X - 4 - (ShiftCounter >> 1)); \
244 ps32X--; \
245 *ps32X = *(ps32X - 4 - (ShiftCounter >> 1)); \
246 ps32X--; \
247 *ps32X = *(ps32X - 4 - (ShiftCounter >> 1)); \
248 ps32X--; \
249 *ps32X = *(ps32X - 4 - (ShiftCounter >> 1)); \
250 ps32X--; \
251 } \
252 }
253 #define SHIFTUP_X8_2 \
254 { \
255 ps32X = (int32_t*)(s16X + EncMaxShiftCounter + 78); \
256 ps32X2 = (int32_t*)(s16X + (EncMaxShiftCounter << 1) + 158); \
257 for (i = 0; i < 9; i++) { \
258 *ps32X = *(ps32X - 4 - (ShiftCounter >> 1)); \
259 *(ps32X2) = *(ps32X2 - 4 - (ShiftCounter >> 1)); \
260 ps32X--; \
261 ps32X2--; \
262 *ps32X = *(ps32X - 4 - (ShiftCounter >> 1)); \
263 *(ps32X2) = *(ps32X2 - 4 - (ShiftCounter >> 1)); \
264 ps32X--; \
265 ps32X2--; \
266 *ps32X = *(ps32X - 4 - (ShiftCounter >> 1)); \
267 *(ps32X2) = *(ps32X2 - 4 - (ShiftCounter >> 1)); \
268 ps32X--; \
269 ps32X2--; \
270 *ps32X = *(ps32X - 4 - (ShiftCounter >> 1)); \
271 *(ps32X2) = *(ps32X2 - 4 - (ShiftCounter >> 1)); \
272 ps32X--; \
273 ps32X2--; \
274 } \
275 }
276
277 #if (SBC_ARM_ASM_OPT == TRUE)
278 #define WINDOW_ACCU_8_0 \
279 { \
280 __asm {\
281 MUL s32Hi,WIND_8_SUBBANDS_0_1,(s16X[ChOffset+16]-s16X[ChOffset+64]);\
282 MLA s32Hi,WIND_8_SUBBANDS_0_2,(s16X[ChOffset+32]-s16X[ChOffset+48]),s32Hi;\
283 MOV s32DCTY[0],s32Hi; \
284 } \
285 }
286 #define WINDOW_ACCU_8_1_15 \
287 { \
288 __asm {\
289 MUL s32Hi,WIND_8_SUBBANDS_1_0,s16X[ChOffset+1];\
290 MUL s32Hi2,WIND_8_SUBBANDS_1_0,s16X[ChOffset+64+15];\
291 MLA s32Hi,WIND_8_SUBBANDS_1_1,s16X[ChOffset+16+1],s32Hi;\
292 MLA s32Hi2,WIND_8_SUBBANDS_1_1,s16X[ChOffset+48+15],s32Hi2;\
293 MLA s32Hi,WIND_8_SUBBANDS_1_2,s16X[ChOffset+32+1],s32Hi;\
294 MLA s32Hi2,WIND_8_SUBBANDS_1_2,s16X[ChOffset+32+15],s32Hi2;\
295 MLA s32Hi,WIND_8_SUBBANDS_1_3,s16X[ChOffset+48+1],s32Hi;\
296 MLA s32Hi2,WIND_8_SUBBANDS_1_3,s16X[ChOffset+16+15],s32Hi2;\
297 MLA s32Hi,WIND_8_SUBBANDS_1_4,s16X[ChOffset+64+1],s32Hi;\
298 MLA s32Hi2,WIND_8_SUBBANDS_1_4,s16X[ChOffset+15],s32Hi2;\
299 MOV s32DCTY[1],s32Hi;\
300 MOV s32DCTY[15],s32Hi2; \
301 } \
302 }
303 #define WINDOW_ACCU_8_2_14 \
304 { \
305 __asm {\
306 MUL s32Hi,WIND_8_SUBBANDS_2_0,s16X[ChOffset+2];\
307 MUL s32Hi2,WIND_8_SUBBANDS_2_0,s16X[ChOffset+64+14];\
308 MLA s32Hi,WIND_8_SUBBANDS_2_1,s16X[ChOffset+16+2],s32Hi;\
309 MLA s32Hi2,WIND_8_SUBBANDS_2_1,s16X[ChOffset+48+14],s32Hi2;\
310 MLA s32Hi,WIND_8_SUBBANDS_2_2,s16X[ChOffset+32+2],s32Hi;\
311 MLA s32Hi2,WIND_8_SUBBANDS_2_2,s16X[ChOffset+32+14],s32Hi2;\
312 MLA s32Hi,WIND_8_SUBBANDS_2_3,s16X[ChOffset+48+2],s32Hi;\
313 MLA s32Hi2,WIND_8_SUBBANDS_2_3,s16X[ChOffset+16+14],s32Hi2;\
314 MLA s32Hi,WIND_8_SUBBANDS_2_4,s16X[ChOffset+64+2],s32Hi;\
315 MLA s32Hi2,WIND_8_SUBBANDS_2_4,s16X[ChOffset+14],s32Hi2;\
316 MOV s32DCTY[2],s32Hi;\
317 MOV s32DCTY[14],s32Hi2; \
318 } \
319 }
320 #define WINDOW_ACCU_8_3_13 \
321 { \
322 __asm {\
323 MUL s32Hi,WIND_8_SUBBANDS_3_0,s16X[ChOffset+3];\
324 MUL s32Hi2,WIND_8_SUBBANDS_3_0,s16X[ChOffset+64+13];\
325 MLA s32Hi,WIND_8_SUBBANDS_3_1,s16X[ChOffset+16+3],s32Hi;\
326 MLA s32Hi2,WIND_8_SUBBANDS_3_1,s16X[ChOffset+48+13],s32Hi2;\
327 MLA s32Hi,WIND_8_SUBBANDS_3_2,s16X[ChOffset+32+3],s32Hi;\
328 MLA s32Hi2,WIND_8_SUBBANDS_3_2,s16X[ChOffset+32+13],s32Hi2;\
329 MLA s32Hi,WIND_8_SUBBANDS_3_3,s16X[ChOffset+48+3],s32Hi;\
330 MLA s32Hi2,WIND_8_SUBBANDS_3_3,s16X[ChOffset+16+13],s32Hi2;\
331 MLA s32Hi,WIND_8_SUBBANDS_3_4,s16X[ChOffset+64+3],s32Hi;\
332 MLA s32Hi2,WIND_8_SUBBANDS_3_4,s16X[ChOffset+13],s32Hi2;\
333 MOV s32DCTY[3],s32Hi;\
334 MOV s32DCTY[13],s32Hi2; \
335 } \
336 }
337 #define WINDOW_ACCU_8_4_12 \
338 { \
339 __asm {\
340 MUL s32Hi,WIND_8_SUBBANDS_4_0,s16X[ChOffset+4];\
341 MUL s32Hi2,WIND_8_SUBBANDS_4_0,s16X[ChOffset+64+12];\
342 MLA s32Hi,WIND_8_SUBBANDS_4_1,s16X[ChOffset+16+4],s32Hi;\
343 MLA s32Hi2,WIND_8_SUBBANDS_4_1,s16X[ChOffset+48+12],s32Hi2;\
344 MLA s32Hi,WIND_8_SUBBANDS_4_2,s16X[ChOffset+32+4],s32Hi;\
345 MLA s32Hi2,WIND_8_SUBBANDS_4_2,s16X[ChOffset+32+12],s32Hi2;\
346 MLA s32Hi,WIND_8_SUBBANDS_4_3,s16X[ChOffset+48+4],s32Hi;\
347 MLA s32Hi2,WIND_8_SUBBANDS_4_3,s16X[ChOffset+16+12],s32Hi2;\
348 MLA s32Hi,WIND_8_SUBBANDS_4_4,s16X[ChOffset+64+4],s32Hi;\
349 MLA s32Hi2,WIND_8_SUBBANDS_4_4,s16X[ChOffset+12],s32Hi2;\
350 MOV s32DCTY[4],s32Hi;\
351 MOV s32DCTY[12],s32Hi2; \
352 } \
353 }
354 #define WINDOW_ACCU_8_5_11 \
355 { \
356 __asm {\
357 MUL s32Hi,WIND_8_SUBBANDS_5_0,s16X[ChOffset+5];\
358 MUL s32Hi2,WIND_8_SUBBANDS_5_0,s16X[ChOffset+64+11];\
359 MLA s32Hi,WIND_8_SUBBANDS_5_1,s16X[ChOffset+16+5],s32Hi;\
360 MLA s32Hi2,WIND_8_SUBBANDS_5_1,s16X[ChOffset+48+11],s32Hi2;\
361 MLA s32Hi,WIND_8_SUBBANDS_5_2,s16X[ChOffset+32+5],s32Hi;\
362 MLA s32Hi2,WIND_8_SUBBANDS_5_2,s16X[ChOffset+32+11],s32Hi2;\
363 MLA s32Hi,WIND_8_SUBBANDS_5_3,s16X[ChOffset+48+5],s32Hi;\
364 MLA s32Hi2,WIND_8_SUBBANDS_5_3,s16X[ChOffset+16+11],s32Hi2;\
365 MLA s32Hi,WIND_8_SUBBANDS_5_4,s16X[ChOffset+64+5],s32Hi;\
366 MLA s32Hi2,WIND_8_SUBBANDS_5_4,s16X[ChOffset+11],s32Hi2;\
367 MOV s32DCTY[5],s32Hi;\
368 MOV s32DCTY[11],s32Hi2; \
369 } \
370 }
371 #define WINDOW_ACCU_8_6_10 \
372 { \
373 __asm {\
374 MUL s32Hi,WIND_8_SUBBANDS_6_0,s16X[ChOffset+6];\
375 MUL s32Hi2,WIND_8_SUBBANDS_6_0,s16X[ChOffset+64+10];\
376 MLA s32Hi,WIND_8_SUBBANDS_6_1,s16X[ChOffset+16+6],s32Hi;\
377 MLA s32Hi2,WIND_8_SUBBANDS_6_1,s16X[ChOffset+48+10],s32Hi2;\
378 MLA s32Hi,WIND_8_SUBBANDS_6_2,s16X[ChOffset+32+6],s32Hi;\
379 MLA s32Hi2,WIND_8_SUBBANDS_6_2,s16X[ChOffset+32+10],s32Hi2;\
380 MLA s32Hi,WIND_8_SUBBANDS_6_3,s16X[ChOffset+48+6],s32Hi;\
381 MLA s32Hi2,WIND_8_SUBBANDS_6_3,s16X[ChOffset+16+10],s32Hi2;\
382 MLA s32Hi,WIND_8_SUBBANDS_6_4,s16X[ChOffset+64+6],s32Hi;\
383 MLA s32Hi2,WIND_8_SUBBANDS_6_4,s16X[ChOffset+10],s32Hi2;\
384 MOV s32DCTY[6],s32Hi;\
385 MOV s32DCTY[10],s32Hi2; \
386 } \
387 }
388 #define WINDOW_ACCU_8_7_9 \
389 { \
390 __asm {\
391 MUL s32Hi,WIND_8_SUBBANDS_7_0,s16X[ChOffset+7];\
392 MUL s32Hi2,WIND_8_SUBBANDS_7_0,s16X[ChOffset+64+9];\
393 MLA s32Hi,WIND_8_SUBBANDS_7_1,s16X[ChOffset+16+7],s32Hi;\
394 MLA s32Hi2,WIND_8_SUBBANDS_7_1,s16X[ChOffset+48+9],s32Hi2;\
395 MLA s32Hi,WIND_8_SUBBANDS_7_2,s16X[ChOffset+32+7],s32Hi;\
396 MLA s32Hi2,WIND_8_SUBBANDS_7_2,s16X[ChOffset+32+9],s32Hi2;\
397 MLA s32Hi,WIND_8_SUBBANDS_7_3,s16X[ChOffset+48+7],s32Hi;\
398 MLA s32Hi2,WIND_8_SUBBANDS_7_3,s16X[ChOffset+16+9],s32Hi2;\
399 MLA s32Hi,WIND_8_SUBBANDS_7_4,s16X[ChOffset+64+7],s32Hi;\
400 MLA s32Hi2,WIND_8_SUBBANDS_7_4,s16X[ChOffset+9],s32Hi2;\
401 MOV s32DCTY[7],s32Hi;\
402 MOV s32DCTY[9],s32Hi2; \
403 } \
404 }
405 #define WINDOW_ACCU_8_8 \
406 { \
407 __asm {\
408 MUL s32Hi,WIND_8_SUBBANDS_8_0,(s16X[ChOffset+8]+s16X[ChOffset+8+64]);\
409 MLA s32Hi,WIND_8_SUBBANDS_8_1,(s16X[ChOffset+8+16]+s16X[ChOffset+8+64]),s32Hi;\
410 MLA s32Hi,WIND_8_SUBBANDS_8_2,s16X[ChOffset+8+32],s32Hi;\
411 MOV s32DCTY[8],s32Hi; \
412 } \
413 }
414 #define WINDOW_ACCU_4_0 \
415 { \
416 __asm {\
417 MUL s32Hi,WIND_4_SUBBANDS_0_1,(s16X[ChOffset+8]-s16X[ChOffset+32]);\
418 MLA s32Hi,WIND_4_SUBBANDS_0_2,(s16X[ChOffset+16]-s16X[ChOffset+24]),s32Hi;\
419 MOV s32DCTY[0],s32Hi; \
420 } \
421 }
422 #define WINDOW_ACCU_4_1_7 \
423 { \
424 __asm {\
425 MUL s32Hi,WIND_4_SUBBANDS_1_0,s16X[ChOffset+1];\
426 MUL s32Hi2,WIND_4_SUBBANDS_1_0,s16X[ChOffset+32+7];\
427 MLA s32Hi,WIND_4_SUBBANDS_1_1,s16X[ChOffset+8+1],s32Hi;\
428 MLA s32Hi2,WIND_4_SUBBANDS_1_1,s16X[ChOffset+24+7],s32Hi2;\
429 MLA s32Hi,WIND_4_SUBBANDS_1_2,s16X[ChOffset+16+1],s32Hi;\
430 MLA s32Hi2,WIND_4_SUBBANDS_1_2,s16X[ChOffset+16+7],s32Hi2;\
431 MLA s32Hi,WIND_4_SUBBANDS_1_3,s16X[ChOffset+24+1],s32Hi;\
432 MLA s32Hi2,WIND_4_SUBBANDS_1_3,s16X[ChOffset+8+7],s32Hi2;\
433 MLA s32Hi,WIND_4_SUBBANDS_1_4,s16X[ChOffset+32+1],s32Hi;\
434 MLA s32Hi2,WIND_4_SUBBANDS_1_4,s16X[ChOffset+7],s32Hi2;\
435 MOV s32DCTY[1],s32Hi;\
436 MOV s32DCTY[7],s32Hi2; \
437 } \
438 }
439 #define WINDOW_ACCU_4_2_6 \
440 { \
441 __asm {\
442 MUL s32Hi,WIND_4_SUBBANDS_2_0,s16X[ChOffset+2];\
443 MUL s32Hi2,WIND_4_SUBBANDS_2_0,s16X[ChOffset+32+6];\
444 MLA s32Hi,WIND_4_SUBBANDS_2_1,s16X[ChOffset+8+2],s32Hi;\
445 MLA s32Hi2,WIND_4_SUBBANDS_2_1,s16X[ChOffset+24+6],s32Hi2;\
446 MLA s32Hi,WIND_4_SUBBANDS_2_2,s16X[ChOffset+16+2],s32Hi;\
447 MLA s32Hi2,WIND_4_SUBBANDS_2_2,s16X[ChOffset+16+6],s32Hi2;\
448 MLA s32Hi,WIND_4_SUBBANDS_2_3,s16X[ChOffset+24+2],s32Hi;\
449 MLA s32Hi2,WIND_4_SUBBANDS_2_3,s16X[ChOffset+8+6],s32Hi2;\
450 MLA s32Hi,WIND_4_SUBBANDS_2_4,s16X[ChOffset+32+2],s32Hi;\
451 MLA s32Hi2,WIND_4_SUBBANDS_2_4,s16X[ChOffset+6],s32Hi2;\
452 MOV s32DCTY[2],s32Hi;\
453 MOV s32DCTY[6],s32Hi2; \
454 } \
455 }
456 #define WINDOW_ACCU_4_3_5 \
457 { \
458 __asm {\
459 MUL s32Hi,WIND_4_SUBBANDS_3_0,s16X[ChOffset+3];\
460 MUL s32Hi2,WIND_4_SUBBANDS_3_0,s16X[ChOffset+32+5];\
461 MLA s32Hi,WIND_4_SUBBANDS_3_1,s16X[ChOffset+8+3],s32Hi;\
462 MLA s32Hi2,WIND_4_SUBBANDS_3_1,s16X[ChOffset+24+5],s32Hi2;\
463 MLA s32Hi,WIND_4_SUBBANDS_3_2,s16X[ChOffset+16+3],s32Hi;\
464 MLA s32Hi2,WIND_4_SUBBANDS_3_2,s16X[ChOffset+16+5],s32Hi2;\
465 MLA s32Hi,WIND_4_SUBBANDS_3_3,s16X[ChOffset+24+3],s32Hi;\
466 MLA s32Hi2,WIND_4_SUBBANDS_3_3,s16X[ChOffset+8+5],s32Hi2;\
467 MLA s32Hi,WIND_4_SUBBANDS_3_4,s16X[ChOffset+32+3],s32Hi;\
468 MLA s32Hi2,WIND_4_SUBBANDS_3_4,s16X[ChOffset+5],s32Hi2;\
469 MOV s32DCTY[3],s32Hi;\
470 MOV s32DCTY[5],s32Hi2; \
471 } \
472 }
473 #define WINDOW_ACCU_4_4 \
474 { \
475 __asm {\
476 MUL s32Hi,WIND_4_SUBBANDS_4_0,(s16X[ChOffset+4]+s16X[ChOffset+4+32]);\
477 MLA s32Hi,WIND_4_SUBBANDS_4_1,(s16X[ChOffset+4+8]+s16X[ChOffset+4+24]),s32Hi;\
478 MLA s32Hi,WIND_4_SUBBANDS_4_2,s16X[ChOffset+4+16],s32Hi;\
479 MOV s32DCTY[4],s32Hi; \
480 } \
481 }
482
483 #define WINDOW_PARTIAL_4 \
484 { \
485 WINDOW_ACCU_4_0; \
486 WINDOW_ACCU_4_1_7; \
487 WINDOW_ACCU_4_2_6; \
488 WINDOW_ACCU_4_3_5; \
489 WINDOW_ACCU_4_4; \
490 }
491
492 #define WINDOW_PARTIAL_8 \
493 { \
494 WINDOW_ACCU_8_0; \
495 WINDOW_ACCU_8_1_15; \
496 WINDOW_ACCU_8_2_14; \
497 WINDOW_ACCU_8_3_13; \
498 WINDOW_ACCU_8_4_12; \
499 WINDOW_ACCU_8_5_11; \
500 WINDOW_ACCU_8_6_10; \
501 WINDOW_ACCU_8_7_9; \
502 WINDOW_ACCU_8_8; \
503 }
504
505 #else
506 #if (SBC_IPAQ_OPT == TRUE)
507
508 #if (SBC_IS_64_MULT_IN_WINDOW_ACCU == TRUE)
509 #define WINDOW_ACCU_8_0 \
510 { \
511 s64Temp = (int64_t)WIND_8_SUBBANDS_0_1 * \
512 (int64_t)(s16X[ChOffset + 16] - s16X[ChOffset + 64]); \
513 s64Temp += (int64_t)WIND_8_SUBBANDS_0_2 * \
514 (int64_t)(s16X[ChOffset + 32] - s16X[ChOffset + 48]); \
515 s32DCTY[0] = (int32_t)(s64Temp >> 16); \
516 }
517 #define WINDOW_ACCU_8_1_15 \
518 { \
519 s64Temp = (int64_t)WIND_8_SUBBANDS_1_0 * (int64_t)s16X[ChOffset + 1]; \
520 s64Temp2 = \
521 (int64_t)WIND_8_SUBBANDS_1_0 * (int64_t)s16X[ChOffset + 64 + 15]; \
522 s64Temp += \
523 (int64_t)WIND_8_SUBBANDS_1_1 * (int64_t)s16X[ChOffset + 16 + 1]; \
524 s64Temp2 += \
525 (int64_t)WIND_8_SUBBANDS_1_1 * (int64_t)s16X[ChOffset + 48 + 15]; \
526 s64Temp += \
527 (int64_t)WIND_8_SUBBANDS_1_2 * (int64_t)s16X[ChOffset + 32 + 1]; \
528 s64Temp2 += \
529 (int64_t)WIND_8_SUBBANDS_1_2 * (int64_t)s16X[ChOffset + 32 + 15]; \
530 s64Temp += \
531 (int64_t)WIND_8_SUBBANDS_1_3 * (int64_t)s16X[ChOffset + 48 + 1]; \
532 s64Temp2 += \
533 (int64_t)WIND_8_SUBBANDS_1_3 * (int64_t)s16X[ChOffset + 16 + 15]; \
534 s64Temp += \
535 (int64_t)WIND_8_SUBBANDS_1_4 * (int64_t)s16X[ChOffset + 64 + 1]; \
536 s64Temp2 += (int64_t)WIND_8_SUBBANDS_1_4 * (int64_t)s16X[ChOffset + 15]; \
537 s32DCTY[1] = (int32_t)(s64Temp >> 16); \
538 s32DCTY[15] = (int32_t)(s64Temp2 >> 16); \
539 }
540 #define WINDOW_ACCU_8_2_14 \
541 { \
542 s64Temp = (int64_t)WIND_8_SUBBANDS_2_0 * (int64_t)s16X[ChOffset + 2]; \
543 s64Temp2 = \
544 (int64_t)WIND_8_SUBBANDS_2_0 * (int64_t)s16X[ChOffset + 64 + 14]; \
545 s64Temp += \
546 (int64_t)WIND_8_SUBBANDS_2_1 * (int64_t)s16X[ChOffset + 16 + 2]; \
547 s64Temp2 += \
548 (int64_t)WIND_8_SUBBANDS_2_1 * (int64_t)s16X[ChOffset + 48 + 14]; \
549 s64Temp += \
550 (int64_t)WIND_8_SUBBANDS_2_2 * (int64_t)s16X[ChOffset + 32 + 2]; \
551 s64Temp2 += \
552 (int64_t)WIND_8_SUBBANDS_2_2 * (int64_t)s16X[ChOffset + 32 + 14]; \
553 s64Temp += \
554 (int64_t)WIND_8_SUBBANDS_2_3 * (int64_t)s16X[ChOffset + 48 + 2]; \
555 s64Temp2 += \
556 (int64_t)WIND_8_SUBBANDS_2_3 * (int64_t)s16X[ChOffset + 16 + 14]; \
557 s64Temp += \
558 (int64_t)WIND_8_SUBBANDS_2_4 * (int64_t)s16X[ChOffset + 64 + 2]; \
559 s64Temp2 += (int64_t)WIND_8_SUBBANDS_2_4 * (int64_t)s16X[ChOffset + 14]; \
560 s32DCTY[2] = (int32_t)(s64Temp >> 16); \
561 s32DCTY[14] = (int32_t)(s64Temp2 >> 16); \
562 }
563 #define WINDOW_ACCU_8_3_13 \
564 { \
565 s64Temp = (int64_t)WIND_8_SUBBANDS_3_0 * (int64_t)s16X[ChOffset + 3]; \
566 s64Temp2 = \
567 (int64_t)WIND_8_SUBBANDS_3_0 * (int64_t)s16X[ChOffset + 64 + 13]; \
568 s64Temp += \
569 (int64_t)WIND_8_SUBBANDS_3_1 * (int64_t)s16X[ChOffset + 16 + 3]; \
570 s64Temp2 += \
571 (int64_t)WIND_8_SUBBANDS_3_1 * (int64_t)s16X[ChOffset + 48 + 13]; \
572 s64Temp += \
573 (int64_t)WIND_8_SUBBANDS_3_2 * (int64_t)s16X[ChOffset + 32 + 3]; \
574 s64Temp2 += \
575 (int64_t)WIND_8_SUBBANDS_3_2 * (int64_t)s16X[ChOffset + 32 + 13]; \
576 s64Temp += \
577 (int64_t)WIND_8_SUBBANDS_3_3 * (int64_t)s16X[ChOffset + 48 + 3]; \
578 s64Temp2 += \
579 (int64_t)WIND_8_SUBBANDS_3_3 * (int64_t)s16X[ChOffset + 16 + 13]; \
580 s64Temp += \
581 (int64_t)WIND_8_SUBBANDS_3_4 * (int64_t)s16X[ChOffset + 64 + 3]; \
582 s64Temp2 += (int64_t)WIND_8_SUBBANDS_3_4 * (int64_t)s16X[ChOffset + 13]; \
583 s32DCTY[3] = (int32_t)(s64Temp >> 16); \
584 s32DCTY[13] = (int32_t)(s64Temp2 >> 16); \
585 }
586 #define WINDOW_ACCU_8_4_12 \
587 { \
588 s64Temp = (int64_t)WIND_8_SUBBANDS_4_0 * (int64_t)s16X[ChOffset + 4]; \
589 s64Temp2 = \
590 (int64_t)WIND_8_SUBBANDS_4_0 * (int64_t)s16X[ChOffset + 64 + 12]; \
591 s64Temp += \
592 (int64_t)WIND_8_SUBBANDS_4_1 * (int64_t)s16X[ChOffset + 16 + 4]; \
593 s64Temp2 += \
594 (int64_t)WIND_8_SUBBANDS_4_1 * (int64_t)s16X[ChOffset + 48 + 12]; \
595 s64Temp += \
596 (int64_t)WIND_8_SUBBANDS_4_2 * (int64_t)s16X[ChOffset + 32 + 4]; \
597 s64Temp2 += \
598 (int64_t)WIND_8_SUBBANDS_4_2 * (int64_t)s16X[ChOffset + 32 + 12]; \
599 s64Temp += \
600 (int64_t)WIND_8_SUBBANDS_4_3 * (int64_t)s16X[ChOffset + 48 + 4]; \
601 s64Temp2 += \
602 (int64_t)WIND_8_SUBBANDS_4_3 * (int64_t)s16X[ChOffset + 16 + 12]; \
603 s64Temp += \
604 (int64_t)WIND_8_SUBBANDS_4_4 * (int64_t)s16X[ChOffset + 64 + 4]; \
605 s64Temp2 += (int64_t)WIND_8_SUBBANDS_4_4 * (int64_t)s16X[ChOffset + 12]; \
606 s32DCTY[4] = (int32_t)(s64Temp >> 16); \
607 s32DCTY[12] = (int32_t)(s64Temp2 >> 16); \
608 }
609 #define WINDOW_ACCU_8_5_11 \
610 { \
611 s64Temp = (int64_t)WIND_8_SUBBANDS_5_0 * (int64_t)s16X[ChOffset + 5]; \
612 s64Temp2 = \
613 (int64_t)WIND_8_SUBBANDS_5_0 * (int64_t)s16X[ChOffset + 64 + 11]; \
614 s64Temp += \
615 (int64_t)WIND_8_SUBBANDS_5_1 * (int64_t)s16X[ChOffset + 16 + 5]; \
616 s64Temp2 += \
617 (int64_t)WIND_8_SUBBANDS_5_1 * (int64_t)s16X[ChOffset + 48 + 11]; \
618 s64Temp += \
619 (int64_t)WIND_8_SUBBANDS_5_2 * (int64_t)s16X[ChOffset + 32 + 5]; \
620 s64Temp2 += \
621 (int64_t)WIND_8_SUBBANDS_5_2 * (int64_t)s16X[ChOffset + 32 + 11]; \
622 s64Temp += \
623 (int64_t)WIND_8_SUBBANDS_5_3 * (int64_t)s16X[ChOffset + 48 + 5]; \
624 s64Temp2 += \
625 (int64_t)WIND_8_SUBBANDS_5_3 * (int64_t)s16X[ChOffset + 16 + 11]; \
626 s64Temp += \
627 (int64_t)WIND_8_SUBBANDS_5_4 * (int64_t)s16X[ChOffset + 64 + 5]; \
628 s64Temp2 += (int64_t)WIND_8_SUBBANDS_5_4 * (int64_t)s16X[ChOffset + 11]; \
629 s32DCTY[5] = (int32_t)(s64Temp >> 16); \
630 s32DCTY[11] = (int32_t)(s64Temp2 >> 16); \
631 }
632 #define WINDOW_ACCU_8_6_10 \
633 { \
634 s64Temp = (int64_t)WIND_8_SUBBANDS_6_0 * (int64_t)s16X[ChOffset + 6]; \
635 s64Temp2 = \
636 (int64_t)WIND_8_SUBBANDS_6_0 * (int64_t)s16X[ChOffset + 64 + 10]; \
637 s64Temp += \
638 (int64_t)WIND_8_SUBBANDS_6_1 * (int64_t)s16X[ChOffset + 16 + 6]; \
639 s64Temp2 += \
640 (int64_t)WIND_8_SUBBANDS_6_1 * (int64_t)s16X[ChOffset + 48 + 10]; \
641 s64Temp += \
642 (int64_t)WIND_8_SUBBANDS_6_2 * (int64_t)s16X[ChOffset + 32 + 6]; \
643 s64Temp2 += \
644 (int64_t)WIND_8_SUBBANDS_6_2 * (int64_t)s16X[ChOffset + 32 + 10]; \
645 s64Temp += \
646 (int64_t)WIND_8_SUBBANDS_6_3 * (int64_t)s16X[ChOffset + 48 + 6]; \
647 s64Temp2 += \
648 (int64_t)WIND_8_SUBBANDS_6_3 * (int64_t)s16X[ChOffset + 16 + 10]; \
649 s64Temp += \
650 (int64_t)WIND_8_SUBBANDS_6_4 * (int64_t)s16X[ChOffset + 64 + 6]; \
651 s64Temp2 += (int64_t)WIND_8_SUBBANDS_6_4 * (int64_t)s16X[ChOffset + 10]; \
652 s32DCTY[6] = (int32_t)(s64Temp >> 16); \
653 s32DCTY[10] = (int32_t)(s64Temp2 >> 16); \
654 }
655 #define WINDOW_ACCU_8_7_9 \
656 { \
657 s64Temp = (int64_t)WIND_8_SUBBANDS_7_0 * (int64_t)s16X[ChOffset + 7]; \
658 s64Temp2 = \
659 (int64_t)WIND_8_SUBBANDS_7_0 * (int64_t)s16X[ChOffset + 64 + 9]; \
660 s64Temp += \
661 (int64_t)WIND_8_SUBBANDS_7_1 * (int64_t)s16X[ChOffset + 16 + 7]; \
662 s64Temp2 += \
663 (int64_t)WIND_8_SUBBANDS_7_1 * (int64_t)s16X[ChOffset + 48 + 9]; \
664 s64Temp += \
665 (int64_t)WIND_8_SUBBANDS_7_2 * (int64_t)s16X[ChOffset + 32 + 7]; \
666 s64Temp2 += \
667 (int64_t)WIND_8_SUBBANDS_7_2 * (int64_t)s16X[ChOffset + 32 + 9]; \
668 s64Temp += \
669 (int64_t)WIND_8_SUBBANDS_7_3 * (int64_t)s16X[ChOffset + 48 + 7]; \
670 s64Temp2 += \
671 (int64_t)WIND_8_SUBBANDS_7_3 * (int64_t)s16X[ChOffset + 16 + 9]; \
672 s64Temp += \
673 (int64_t)WIND_8_SUBBANDS_7_4 * (int64_t)s16X[ChOffset + 64 + 7]; \
674 s64Temp2 += (int64_t)WIND_8_SUBBANDS_7_4 * (int64_t)s16X[ChOffset + 9]; \
675 s32DCTY[7] = (int32_t)(s64Temp >> 16); \
676 s32DCTY[9] = (int32_t)(s64Temp2 >> 16); \
677 }
678 #define WINDOW_ACCU_8_8 \
679 { \
680 s64Temp = (int64_t)WIND_8_SUBBANDS_8_0 * \
681 (int64_t)(s16X[ChOffset + 8] + s16X[ChOffset + 64 + 8]); \
682 s64Temp += (int64_t)WIND_8_SUBBANDS_8_1 * \
683 (int64_t)(s16X[ChOffset + 16 + 8] + s16X[ChOffset + 48 + 8]); \
684 s64Temp += \
685 (int64_t)WIND_8_SUBBANDS_8_2 * (int64_t)s16X[ChOffset + 32 + 8]; \
686 s32DCTY[8] = (int32_t)(s64Temp >> 16); \
687 }
688 #define WINDOW_ACCU_4_0 \
689 { \
690 s64Temp = (int64_t)WIND_4_SUBBANDS_0_1 * \
691 (int64_t)(s16X[ChOffset + 8] - s16X[ChOffset + 32]); \
692 s64Temp += (int64_t)WIND_4_SUBBANDS_0_2 * \
693 (int64_t)(s16X[ChOffset + 16] - s16X[ChOffset + 24]); \
694 s32DCTY[0] = (int32_t)(s64Temp >> 16); \
695 }
696 #define WINDOW_ACCU_4_1_7 \
697 { \
698 s64Temp = (int64_t)WIND_4_SUBBANDS_1_0 * (int64_t)s16X[ChOffset + 1]; \
699 s64Temp2 = \
700 (int64_t)WIND_4_SUBBANDS_1_0 * (int64_t)s16X[ChOffset + 32 + 7]; \
701 s64Temp += (int64_t)WIND_4_SUBBANDS_1_1 * (int64_t)s16X[ChOffset + 8 + 1]; \
702 s64Temp2 += \
703 (int64_t)WIND_4_SUBBANDS_1_1 * (int64_t)s16X[ChOffset + 24 + 7]; \
704 s64Temp += \
705 (int64_t)WIND_4_SUBBANDS_1_2 * (int64_t)s16X[ChOffset + 16 + 1]; \
706 s64Temp2 += \
707 (int64_t)WIND_4_SUBBANDS_1_2 * (int64_t)s16X[ChOffset + 16 + 7]; \
708 s64Temp += \
709 (int64_t)WIND_4_SUBBANDS_1_3 * (int64_t)s16X[ChOffset + 24 + 1]; \
710 s64Temp2 += \
711 (int64_t)WIND_4_SUBBANDS_1_3 * (int64_t)s16X[ChOffset + 8 + 7]; \
712 s64Temp += \
713 (int64_t)WIND_4_SUBBANDS_1_4 * (int64_t)s16X[ChOffset + 32 + 1]; \
714 s64Temp2 += (int64_t)WIND_4_SUBBANDS_1_4 * (int64_t)s16X[ChOffset + 7]; \
715 s32DCTY[1] = (int32_t)(s64Temp >> 16); \
716 s32DCTY[7] = (int32_t)(s64Temp2 >> 16); \
717 }
718 #define WINDOW_ACCU_4_2_6 \
719 { \
720 s64Temp = (int64_t)WIND_4_SUBBANDS_2_0 * (int64_t)s16X[ChOffset + 2]; \
721 s64Temp2 = \
722 (int64_t)WIND_4_SUBBANDS_2_0 * (int64_t)s16X[ChOffset + 32 + 6]; \
723 s64Temp += (int64_t)WIND_4_SUBBANDS_2_1 * (int64_t)s16X[ChOffset + 8 + 2]; \
724 s64Temp2 += \
725 (int64_t)WIND_4_SUBBANDS_2_1 * (int64_t)s16X[ChOffset + 24 + 6]; \
726 s64Temp += \
727 (int64_t)WIND_4_SUBBANDS_2_2 * (int64_t)s16X[ChOffset + 16 + 2]; \
728 s64Temp2 += \
729 (int64_t)WIND_4_SUBBANDS_2_2 * (int64_t)s16X[ChOffset + 16 + 6]; \
730 s64Temp += \
731 (int64_t)WIND_4_SUBBANDS_2_3 * (int64_t)s16X[ChOffset + 24 + 2]; \
732 s64Temp2 += \
733 (int64_t)WIND_4_SUBBANDS_2_3 * (int64_t)s16X[ChOffset + 8 + 6]; \
734 s64Temp += \
735 (int64_t)WIND_4_SUBBANDS_2_4 * (int64_t)s16X[ChOffset + 32 + 2]; \
736 s64Temp2 += (int64_t)WIND_4_SUBBANDS_2_4 * (int64_t)s16X[ChOffset + 6]; \
737 s32DCTY[2] = (int32_t)(s64Temp >> 16); \
738 s32DCTY[6] = (int32_t)(s64Temp2 >> 16); \
739 }
740 #define WINDOW_ACCU_4_3_5 \
741 { \
742 s64Temp = (int64_t)WIND_4_SUBBANDS_3_0 * (int64_t)s16X[ChOffset + 3]; \
743 s64Temp2 = \
744 (int64_t)WIND_4_SUBBANDS_3_0 * (int64_t)s16X[ChOffset + 32 + 5]; \
745 s64Temp += (int64_t)WIND_4_SUBBANDS_3_1 * (int64_t)s16X[ChOffset + 8 + 3]; \
746 s64Temp2 += \
747 (int64_t)WIND_4_SUBBANDS_3_1 * (int64_t)s16X[ChOffset + 24 + 5]; \
748 s64Temp += \
749 (int64_t)WIND_4_SUBBANDS_3_2 * (int64_t)s16X[ChOffset + 16 + 3]; \
750 s64Temp2 += \
751 (int64_t)WIND_4_SUBBANDS_3_2 * (int64_t)s16X[ChOffset + 16 + 5]; \
752 s64Temp += \
753 (int64_t)WIND_4_SUBBANDS_3_3 * (int64_t)s16X[ChOffset + 24 + 3]; \
754 s64Temp2 += \
755 (int64_t)WIND_4_SUBBANDS_3_3 * (int64_t)s16X[ChOffset + 8 + 5]; \
756 s64Temp += \
757 (int64_t)WIND_4_SUBBANDS_3_4 * (int64_t)s16X[ChOffset + 32 + 3]; \
758 s64Temp2 += (int64_t)WIND_4_SUBBANDS_3_4 * (int64_t)s16X[ChOffset + 5]; \
759 s32DCTY[3] = (int32_t)(s64Temp >> 16); \
760 s32DCTY[5] = (int32_t)(s64Temp2 >> 16); \
761 }
762
763 #define WINDOW_ACCU_4_4 \
764 { \
765 s64Temp = (int64_t)WIND_4_SUBBANDS_4_0 * \
766 (int64_t)(s16X[ChOffset + 4] + s16X[ChOffset + 4 + 32]); \
767 s64Temp += (int64_t)WIND_4_SUBBANDS_4_1 * \
768 (int64_t)(s16X[ChOffset + 4 + 8] + s16X[ChOffset + 4 + 24]); \
769 s64Temp += \
770 (int64_t)WIND_4_SUBBANDS_4_2 * (int64_t)s16X[ChOffset + 4 + 16]; \
771 s32DCTY[4] = (int32_t)(s64Temp >> 16); \
772 }
773 #else /* SBC_IS_64_MULT_IN_WINDOW_ACCU == FALSE */
774 #define WINDOW_ACCU_8_0 \
775 { \
776 s32Temp = (int32_t)WIND_8_SUBBANDS_0_1 * \
777 (int32_t)(s16X[ChOffset + 16] - s16X[ChOffset + 64]); \
778 s32Temp += (int32_t)WIND_8_SUBBANDS_0_2 * \
779 (int32_t)(s16X[ChOffset + 32] - s16X[ChOffset + 48]); \
780 s32DCTY[0] = (int32_t)s32Temp; \
781 }
782 #define WINDOW_ACCU_8_1_15 \
783 { \
784 s32Temp = (int32_t)WIND_8_SUBBANDS_1_0 * (int32_t)s16X[ChOffset + 1]; \
785 s32Temp2 = \
786 (int32_t)WIND_8_SUBBANDS_1_0 * (int32_t)s16X[ChOffset + 64 + 15]; \
787 s32Temp += \
788 (int32_t)WIND_8_SUBBANDS_1_1 * (int32_t)s16X[ChOffset + 16 + 1]; \
789 s32Temp2 += \
790 (int32_t)WIND_8_SUBBANDS_1_1 * (int32_t)s16X[ChOffset + 48 + 15]; \
791 s32Temp += \
792 (int32_t)WIND_8_SUBBANDS_1_2 * (int32_t)s16X[ChOffset + 32 + 1]; \
793 s32Temp2 += \
794 (int32_t)WIND_8_SUBBANDS_1_2 * (int32_t)s16X[ChOffset + 32 + 15]; \
795 s32Temp += \
796 (int32_t)WIND_8_SUBBANDS_1_3 * (int32_t)s16X[ChOffset + 48 + 1]; \
797 s32Temp2 += \
798 (int32_t)WIND_8_SUBBANDS_1_3 * (int32_t)s16X[ChOffset + 16 + 15]; \
799 s32Temp += \
800 (int32_t)WIND_8_SUBBANDS_1_4 * (int32_t)s16X[ChOffset + 64 + 1]; \
801 s32Temp2 += (int32_t)WIND_8_SUBBANDS_1_4 * (int32_t)s16X[ChOffset + 15]; \
802 s32DCTY[1] = (int32_t)s32Temp; \
803 s32DCTY[15] = (int32_t)s32Temp2; \
804 }
805 #define WINDOW_ACCU_8_2_14 \
806 { \
807 s32Temp = (int32_t)WIND_8_SUBBANDS_2_0 * (int32_t)s16X[ChOffset + 2]; \
808 s32Temp2 = \
809 (int32_t)WIND_8_SUBBANDS_2_0 * (int32_t)s16X[ChOffset + 64 + 14]; \
810 s32Temp += \
811 (int32_t)WIND_8_SUBBANDS_2_1 * (int32_t)s16X[ChOffset + 16 + 2]; \
812 s32Temp2 += \
813 (int32_t)WIND_8_SUBBANDS_2_1 * (int32_t)s16X[ChOffset + 48 + 14]; \
814 s32Temp += \
815 (int32_t)WIND_8_SUBBANDS_2_2 * (int32_t)s16X[ChOffset + 32 + 2]; \
816 s32Temp2 += \
817 (int32_t)WIND_8_SUBBANDS_2_2 * (int32_t)s16X[ChOffset + 32 + 14]; \
818 s32Temp += \
819 (int32_t)WIND_8_SUBBANDS_2_3 * (int32_t)s16X[ChOffset + 48 + 2]; \
820 s32Temp2 += \
821 (int32_t)WIND_8_SUBBANDS_2_3 * (int32_t)s16X[ChOffset + 16 + 14]; \
822 s32Temp += \
823 (int32_t)WIND_8_SUBBANDS_2_4 * (int32_t)s16X[ChOffset + 64 + 2]; \
824 s32Temp2 += (int32_t)WIND_8_SUBBANDS_2_4 * (int32_t)s16X[ChOffset + 14]; \
825 s32DCTY[2] = (int32_t)s32Temp; \
826 s32DCTY[14] = (int32_t)s32Temp2; \
827 }
828 #define WINDOW_ACCU_8_3_13 \
829 { \
830 s32Temp = (int32_t)WIND_8_SUBBANDS_3_0 * (int32_t)s16X[ChOffset + 3]; \
831 s32Temp2 = \
832 (int32_t)WIND_8_SUBBANDS_3_0 * (int32_t)s16X[ChOffset + 64 + 13]; \
833 s32Temp += \
834 (int32_t)WIND_8_SUBBANDS_3_1 * (int32_t)s16X[ChOffset + 16 + 3]; \
835 s32Temp2 += \
836 (int32_t)WIND_8_SUBBANDS_3_1 * (int32_t)s16X[ChOffset + 48 + 13]; \
837 s32Temp += \
838 (int32_t)WIND_8_SUBBANDS_3_2 * (int32_t)s16X[ChOffset + 32 + 3]; \
839 s32Temp2 += \
840 (int32_t)WIND_8_SUBBANDS_3_2 * (int32_t)s16X[ChOffset + 32 + 13]; \
841 s32Temp += \
842 (int32_t)WIND_8_SUBBANDS_3_3 * (int32_t)s16X[ChOffset + 48 + 3]; \
843 s32Temp2 += \
844 (int32_t)WIND_8_SUBBANDS_3_3 * (int32_t)s16X[ChOffset + 16 + 13]; \
845 s32Temp += \
846 (int32_t)WIND_8_SUBBANDS_3_4 * (int32_t)s16X[ChOffset + 64 + 3]; \
847 s32Temp2 += (int32_t)WIND_8_SUBBANDS_3_4 * (int32_t)s16X[ChOffset + 13]; \
848 s32DCTY[3] = (int32_t)s32Temp; \
849 s32DCTY[13] = (int32_t)s32Temp2; \
850 }
851 #define WINDOW_ACCU_8_4_12 \
852 { \
853 s32Temp = (int32_t)WIND_8_SUBBANDS_4_0 * (int32_t)s16X[ChOffset + 4]; \
854 s32Temp2 = \
855 (int32_t)WIND_8_SUBBANDS_4_0 * (int32_t)s16X[ChOffset + 64 + 12]; \
856 s32Temp += \
857 (int32_t)WIND_8_SUBBANDS_4_1 * (int32_t)s16X[ChOffset + 16 + 4]; \
858 s32Temp2 += \
859 (int32_t)WIND_8_SUBBANDS_4_1 * (int32_t)s16X[ChOffset + 48 + 12]; \
860 s32Temp += \
861 (int32_t)WIND_8_SUBBANDS_4_2 * (int32_t)s16X[ChOffset + 32 + 4]; \
862 s32Temp2 += \
863 (int32_t)WIND_8_SUBBANDS_4_2 * (int32_t)s16X[ChOffset + 32 + 12]; \
864 s32Temp += \
865 (int32_t)WIND_8_SUBBANDS_4_3 * (int32_t)s16X[ChOffset + 48 + 4]; \
866 s32Temp2 += \
867 (int32_t)WIND_8_SUBBANDS_4_3 * (int32_t)s16X[ChOffset + 16 + 12]; \
868 s32Temp += \
869 (int32_t)WIND_8_SUBBANDS_4_4 * (int32_t)s16X[ChOffset + 64 + 4]; \
870 s32Temp2 += (int32_t)WIND_8_SUBBANDS_4_4 * (int32_t)s16X[ChOffset + 12]; \
871 s32DCTY[4] = (int32_t)s32Temp; \
872 s32DCTY[12] = (int32_t)s32Temp2; \
873 }
874 #define WINDOW_ACCU_8_5_11 \
875 { \
876 s32Temp = (int32_t)WIND_8_SUBBANDS_5_0 * (int32_t)s16X[ChOffset + 5]; \
877 s32Temp2 = \
878 (int32_t)WIND_8_SUBBANDS_5_0 * (int32_t)s16X[ChOffset + 64 + 11]; \
879 s32Temp += \
880 (int32_t)WIND_8_SUBBANDS_5_1 * (int32_t)s16X[ChOffset + 16 + 5]; \
881 s32Temp2 += \
882 (int32_t)WIND_8_SUBBANDS_5_1 * (int32_t)s16X[ChOffset + 48 + 11]; \
883 s32Temp += \
884 (int32_t)WIND_8_SUBBANDS_5_2 * (int32_t)s16X[ChOffset + 32 + 5]; \
885 s32Temp2 += \
886 (int32_t)WIND_8_SUBBANDS_5_2 * (int32_t)s16X[ChOffset + 32 + 11]; \
887 s32Temp += \
888 (int32_t)WIND_8_SUBBANDS_5_3 * (int32_t)s16X[ChOffset + 48 + 5]; \
889 s32Temp2 += \
890 (int32_t)WIND_8_SUBBANDS_5_3 * (int32_t)s16X[ChOffset + 16 + 11]; \
891 s32Temp += \
892 (int32_t)WIND_8_SUBBANDS_5_4 * (int32_t)s16X[ChOffset + 64 + 5]; \
893 s32Temp2 += (int32_t)WIND_8_SUBBANDS_5_4 * (int32_t)s16X[ChOffset + 11]; \
894 s32DCTY[5] = (int32_t)s32Temp; \
895 s32DCTY[11] = (int32_t)s32Temp2; \
896 }
897 #define WINDOW_ACCU_8_6_10 \
898 { \
899 s32Temp = (int32_t)WIND_8_SUBBANDS_6_0 * (int32_t)s16X[ChOffset + 6]; \
900 s32Temp2 = \
901 (int32_t)WIND_8_SUBBANDS_6_0 * (int32_t)s16X[ChOffset + 64 + 10]; \
902 s32Temp += \
903 (int32_t)WIND_8_SUBBANDS_6_1 * (int32_t)s16X[ChOffset + 16 + 6]; \
904 s32Temp2 += \
905 (int32_t)WIND_8_SUBBANDS_6_1 * (int32_t)s16X[ChOffset + 48 + 10]; \
906 s32Temp += \
907 (int32_t)WIND_8_SUBBANDS_6_2 * (int32_t)s16X[ChOffset + 32 + 6]; \
908 s32Temp2 += \
909 (int32_t)WIND_8_SUBBANDS_6_2 * (int32_t)s16X[ChOffset + 32 + 10]; \
910 s32Temp += \
911 (int32_t)WIND_8_SUBBANDS_6_3 * (int32_t)s16X[ChOffset + 48 + 6]; \
912 s32Temp2 += \
913 (int32_t)WIND_8_SUBBANDS_6_3 * (int32_t)s16X[ChOffset + 16 + 10]; \
914 s32Temp += \
915 (int32_t)WIND_8_SUBBANDS_6_4 * (int32_t)s16X[ChOffset + 64 + 6]; \
916 s32Temp2 += (int32_t)WIND_8_SUBBANDS_6_4 * (int32_t)s16X[ChOffset + 10]; \
917 s32DCTY[6] = (int32_t)s32Temp; \
918 s32DCTY[10] = (int32_t)s32Temp2; \
919 }
920 #define WINDOW_ACCU_8_7_9 \
921 { \
922 s32Temp = (int32_t)WIND_8_SUBBANDS_7_0 * (int32_t)s16X[ChOffset + 7]; \
923 s32Temp2 = \
924 (int32_t)WIND_8_SUBBANDS_7_0 * (int32_t)s16X[ChOffset + 64 + 9]; \
925 s32Temp += \
926 (int32_t)WIND_8_SUBBANDS_7_1 * (int32_t)s16X[ChOffset + 16 + 7]; \
927 s32Temp2 += \
928 (int32_t)WIND_8_SUBBANDS_7_1 * (int32_t)s16X[ChOffset + 48 + 9]; \
929 s32Temp += \
930 (int32_t)WIND_8_SUBBANDS_7_2 * (int32_t)s16X[ChOffset + 32 + 7]; \
931 s32Temp2 += \
932 (int32_t)WIND_8_SUBBANDS_7_2 * (int32_t)s16X[ChOffset + 32 + 9]; \
933 s32Temp += \
934 (int32_t)WIND_8_SUBBANDS_7_3 * (int32_t)s16X[ChOffset + 48 + 7]; \
935 s32Temp2 += \
936 (int32_t)WIND_8_SUBBANDS_7_3 * (int32_t)s16X[ChOffset + 16 + 9]; \
937 s32Temp += \
938 (int32_t)WIND_8_SUBBANDS_7_4 * (int32_t)s16X[ChOffset + 64 + 7]; \
939 s32Temp2 += (int32_t)WIND_8_SUBBANDS_7_4 * (int32_t)s16X[ChOffset + 9]; \
940 s32DCTY[7] = (int32_t)s32Temp; \
941 s32DCTY[9] = (int32_t)s32Temp2; \
942 }
943 #define WINDOW_ACCU_8_8 \
944 { \
945 s32Temp = (int32_t)WIND_8_SUBBANDS_8_0 * \
946 (int32_t)(s16X[ChOffset + 8] + s16X[ChOffset + 64 + 8]); \
947 s32Temp += (int32_t)WIND_8_SUBBANDS_8_1 * \
948 (int32_t)(s16X[ChOffset + 16 + 8] + s16X[ChOffset + 48 + 8]); \
949 s32Temp += \
950 (int32_t)WIND_8_SUBBANDS_8_2 * (int32_t)s16X[ChOffset + 32 + 8]; \
951 s32DCTY[8] = (int32_t)s32Temp; \
952 }
953 #define WINDOW_ACCU_4_0 \
954 { \
955 s32Temp = (int32_t)WIND_4_SUBBANDS_0_1 * \
956 (int32_t)(s16X[ChOffset + 8] - s16X[ChOffset + 32]); \
957 s32Temp += (int32_t)WIND_4_SUBBANDS_0_2 * \
958 (int32_t)(s16X[ChOffset + 16] - s16X[ChOffset + 24]); \
959 s32DCTY[0] = (int32_t)(s32Temp); \
960 }
961 #define WINDOW_ACCU_4_1_7 \
962 { \
963 s32Temp = (int32_t)WIND_4_SUBBANDS_1_0 * (int32_t)s16X[ChOffset + 1]; \
964 s32Temp2 = \
965 (int32_t)WIND_4_SUBBANDS_1_0 * (int32_t)s16X[ChOffset + 32 + 7]; \
966 s32Temp += (int32_t)WIND_4_SUBBANDS_1_1 * (int32_t)s16X[ChOffset + 8 + 1]; \
967 s32Temp2 += \
968 (int32_t)WIND_4_SUBBANDS_1_1 * (int32_t)s16X[ChOffset + 24 + 7]; \
969 s32Temp += \
970 (int32_t)WIND_4_SUBBANDS_1_2 * (int32_t)s16X[ChOffset + 16 + 1]; \
971 s32Temp2 += \
972 (int32_t)WIND_4_SUBBANDS_1_2 * (int32_t)s16X[ChOffset + 16 + 7]; \
973 s32Temp += \
974 (int32_t)WIND_4_SUBBANDS_1_3 * (int32_t)s16X[ChOffset + 24 + 1]; \
975 s32Temp2 += \
976 (int32_t)WIND_4_SUBBANDS_1_3 * (int32_t)s16X[ChOffset + 8 + 7]; \
977 s32Temp += \
978 (int32_t)WIND_4_SUBBANDS_1_4 * (int32_t)s16X[ChOffset + 32 + 1]; \
979 s32Temp2 += (int32_t)WIND_4_SUBBANDS_1_4 * (int32_t)s16X[ChOffset + 7]; \
980 s32DCTY[1] = (int32_t)(s32Temp); \
981 s32DCTY[7] = (int32_t)(s32Temp2); \
982 }
983 #define WINDOW_ACCU_4_2_6 \
984 { \
985 s32Temp = (int32_t)WIND_4_SUBBANDS_2_0 * (int32_t)s16X[ChOffset + 2]; \
986 s32Temp2 = \
987 (int32_t)WIND_4_SUBBANDS_2_0 * (int32_t)s16X[ChOffset + 32 + 6]; \
988 s32Temp += (int32_t)WIND_4_SUBBANDS_2_1 * (int32_t)s16X[ChOffset + 8 + 2]; \
989 s32Temp2 += \
990 (int32_t)WIND_4_SUBBANDS_2_1 * (int32_t)s16X[ChOffset + 24 + 6]; \
991 s32Temp += \
992 (int32_t)WIND_4_SUBBANDS_2_2 * (int32_t)s16X[ChOffset + 16 + 2]; \
993 s32Temp2 += \
994 (int32_t)WIND_4_SUBBANDS_2_2 * (int32_t)s16X[ChOffset + 16 + 6]; \
995 s32Temp += \
996 (int32_t)WIND_4_SUBBANDS_2_3 * (int32_t)s16X[ChOffset + 24 + 2]; \
997 s32Temp2 += \
998 (int32_t)WIND_4_SUBBANDS_2_3 * (int32_t)s16X[ChOffset + 8 + 6]; \
999 s32Temp += \
1000 (int32_t)WIND_4_SUBBANDS_2_4 * (int32_t)s16X[ChOffset + 32 + 2]; \
1001 s32Temp2 += (int32_t)WIND_4_SUBBANDS_2_4 * (int32_t)s16X[ChOffset + 6]; \
1002 s32DCTY[2] = (int32_t)(s32Temp); \
1003 s32DCTY[6] = (int32_t)(s32Temp2); \
1004 }
1005 #define WINDOW_ACCU_4_3_5 \
1006 { \
1007 s32Temp = (int32_t)WIND_4_SUBBANDS_3_0 * (int32_t)s16X[ChOffset + 3]; \
1008 s32Temp2 = \
1009 (int32_t)WIND_4_SUBBANDS_3_0 * (int32_t)s16X[ChOffset + 32 + 5]; \
1010 s32Temp += (int32_t)WIND_4_SUBBANDS_3_1 * (int32_t)s16X[ChOffset + 8 + 3]; \
1011 s32Temp2 += \
1012 (int32_t)WIND_4_SUBBANDS_3_1 * (int32_t)s16X[ChOffset + 24 + 5]; \
1013 s32Temp += \
1014 (int32_t)WIND_4_SUBBANDS_3_2 * (int32_t)s16X[ChOffset + 16 + 3]; \
1015 s32Temp2 += \
1016 (int32_t)WIND_4_SUBBANDS_3_2 * (int32_t)s16X[ChOffset + 16 + 5]; \
1017 s32Temp += \
1018 (int32_t)WIND_4_SUBBANDS_3_3 * (int32_t)s16X[ChOffset + 24 + 3]; \
1019 s32Temp2 += \
1020 (int32_t)WIND_4_SUBBANDS_3_3 * (int32_t)s16X[ChOffset + 8 + 5]; \
1021 s32Temp += \
1022 (int32_t)WIND_4_SUBBANDS_3_4 * (int32_t)s16X[ChOffset + 32 + 3]; \
1023 s32Temp2 += (int32_t)WIND_4_SUBBANDS_3_4 * (int32_t)s16X[ChOffset + 5]; \
1024 s32DCTY[3] = (int32_t)(s32Temp); \
1025 s32DCTY[5] = (int32_t)(s32Temp2); \
1026 }
1027
1028 #define WINDOW_ACCU_4_4 \
1029 { \
1030 s32Temp = (int32_t)WIND_4_SUBBANDS_4_0 * \
1031 (int32_t)(s16X[ChOffset + 4] + s16X[ChOffset + 4 + 32]); \
1032 s32Temp += (int32_t)WIND_4_SUBBANDS_4_1 * \
1033 (int32_t)(s16X[ChOffset + 4 + 8] + s16X[ChOffset + 4 + 24]); \
1034 s32Temp += \
1035 (int32_t)WIND_4_SUBBANDS_4_2 * (int32_t)s16X[ChOffset + 4 + 16]; \
1036 s32DCTY[4] = (int32_t)(s32Temp); \
1037 }
1038 #endif
1039 #define WINDOW_PARTIAL_4 \
1040 { \
1041 WINDOW_ACCU_4_0; \
1042 WINDOW_ACCU_4_1_7; \
1043 WINDOW_ACCU_4_2_6; \
1044 WINDOW_ACCU_4_3_5; \
1045 WINDOW_ACCU_4_4; \
1046 }
1047
1048 #define WINDOW_PARTIAL_8 \
1049 { \
1050 WINDOW_ACCU_8_0; \
1051 WINDOW_ACCU_8_1_15; \
1052 WINDOW_ACCU_8_2_14; \
1053 WINDOW_ACCU_8_3_13; \
1054 WINDOW_ACCU_8_4_12; \
1055 WINDOW_ACCU_8_5_11; \
1056 WINDOW_ACCU_8_6_10; \
1057 WINDOW_ACCU_8_7_9; \
1058 WINDOW_ACCU_8_8; \
1059 }
1060 #else
1061 #if (SBC_IS_64_MULT_IN_WINDOW_ACCU == TRUE)
1062 #define WINDOW_ACCU_4(i) \
1063 { \
1064 s64Temp = ((int64_t)gas32CoeffFor4SBs[i] * (int64_t)s16X[ChOffset + i]); \
1065 s64Temp += ((int64_t)gas32CoeffFor4SBs[(i + 8)] * \
1066 (int64_t)s16X[ChOffset + i + 8]); \
1067 s64Temp += ((int64_t)gas32CoeffFor4SBs[(i + 16)] * \
1068 (int64_t)s16X[ChOffset + i + 16]); \
1069 s64Temp += ((int64_t)gas32CoeffFor4SBs[(i + 24)] * \
1070 (int64_t)s16X[ChOffset + i + 24]); \
1071 s64Temp += ((int64_t)gas32CoeffFor4SBs[(i + 32)] * \
1072 (int64_t)s16X[ChOffset + i + 32]); \
1073 s32DCTY[i] = (int32_t)(s64Temp >> 16); \
1074 /*printf("s32DCTY4: 0x%x \n", s32DCTY[i]);*/ \
1075 }
1076 #else
1077 #define WINDOW_ACCU_4(i) \
1078 { \
1079 s32DCTY[i] = (gas32CoeffFor4SBs[i * 2] * s16X[ChOffset + i]) + \
1080 (((int32_t)(uint16_t)(gas32CoeffFor4SBs[(i * 2) + 1]) * \
1081 s16X[ChOffset + i]) >> \
1082 16); \
1083 s32DCTY[i] += \
1084 (gas32CoeffFor4SBs[(i + 8) * 2] * s16X[ChOffset + i + 8]) + \
1085 (((int32_t)(uint16_t)(gas32CoeffFor4SBs[((i + 8) * 2) + 1]) * \
1086 s16X[ChOffset + i + 8]) >> \
1087 16); \
1088 s32DCTY[i] += \
1089 (gas32CoeffFor4SBs[(i + 16) * 2] * s16X[ChOffset + i + 16]) + \
1090 (((int32_t)(uint16_t)(gas32CoeffFor4SBs[((i + 16) * 2) + 1]) * \
1091 s16X[ChOffset + i + 16]) >> \
1092 16); \
1093 s32DCTY[i] += \
1094 (gas32CoeffFor4SBs[(i + 24) * 2] * s16X[ChOffset + i + 24]) + \
1095 (((int32_t)(uint16_t)(gas32CoeffFor4SBs[((i + 24) * 2) + 1]) * \
1096 s16X[ChOffset + i + 24]) >> \
1097 16); \
1098 s32DCTY[i] += \
1099 (gas32CoeffFor4SBs[(i + 32) * 2] * s16X[ChOffset + i + 32]) + \
1100 (((int32_t)(uint16_t)(gas32CoeffFor4SBs[((i + 32) * 2) + 1]) * \
1101 s16X[ChOffset + i + 32]) >> \
1102 16); \
1103 }
1104 #endif
1105 #define WINDOW_PARTIAL_4 \
1106 { \
1107 WINDOW_ACCU_4(0); \
1108 WINDOW_ACCU_4(1); \
1109 WINDOW_ACCU_4(2); \
1110 WINDOW_ACCU_4(3); \
1111 WINDOW_ACCU_4(4); \
1112 WINDOW_ACCU_4(5); \
1113 WINDOW_ACCU_4(6); \
1114 WINDOW_ACCU_4(7); \
1115 }
1116
1117 #if (SBC_IS_64_MULT_IN_WINDOW_ACCU == TRUE)
1118 #define WINDOW_ACCU_8(i) \
1119 { \
1120 s64Temp = \
1121 ((((int64_t)gas32CoeffFor8SBs[i] * (int64_t)s16X[ChOffset + i]))); \
1122 s64Temp += ((((int64_t)gas32CoeffFor8SBs[(i + 16)] * \
1123 (int64_t)s16X[ChOffset + i + 16]))); \
1124 s64Temp += ((((int64_t)gas32CoeffFor8SBs[(i + 32)] * \
1125 (int64_t)s16X[ChOffset + i + 32]))); \
1126 s64Temp += ((((int64_t)gas32CoeffFor8SBs[(i + 48)] * \
1127 (int64_t)s16X[ChOffset + i + 48]))); \
1128 s64Temp += ((((int64_t)gas32CoeffFor8SBs[(i + 64)] * \
1129 (int64_t)s16X[ChOffset + i + 64]))); \
1130 /*printf("s32DCTY8: %d= 0x%x * %d\n", s32DCTY[i], gas32CoeffFor8SBs[i], \
1131 * s16X[ChOffset+i]);*/ \
1132 s32DCTY[i] = (int32_t)(s64Temp >> 16); \
1133 }
1134 #else
1135 #define WINDOW_ACCU_8(i) \
1136 { \
1137 s32DCTY[i] = (gas32CoeffFor8SBs[i * 2] * s16X[ChOffset + i]) + \
1138 (((int32_t)(uint16_t)(gas32CoeffFor8SBs[(i * 2) + 1]) * \
1139 s16X[ChOffset + i]) >> \
1140 16); \
1141 s32DCTY[i] += \
1142 (gas32CoeffFor8SBs[(i + 16) * 2] * s16X[ChOffset + i + 16]) + \
1143 (((int32_t)(uint16_t)(gas32CoeffFor8SBs[((i + 16) * 2) + 1]) * \
1144 s16X[ChOffset + i + 16]) >> \
1145 16); \
1146 s32DCTY[i] += \
1147 (gas32CoeffFor8SBs[(i + 32) * 2] * s16X[ChOffset + i + 32]) + \
1148 (((int32_t)(uint16_t)(gas32CoeffFor8SBs[((i + 32) * 2) + 1]) * \
1149 s16X[ChOffset + i + 32]) >> \
1150 16); \
1151 s32DCTY[i] += \
1152 (gas32CoeffFor8SBs[(i + 48) * 2] * s16X[ChOffset + i + 48]) + \
1153 (((int32_t)(uint16_t)(gas32CoeffFor8SBs[((i + 48) * 2) + 1]) * \
1154 s16X[ChOffset + i + 48]) >> \
1155 16); \
1156 s32DCTY[i] += \
1157 (gas32CoeffFor8SBs[(i + 64) * 2] * s16X[ChOffset + i + 64]) + \
1158 (((int32_t)(uint16_t)(gas32CoeffFor8SBs[((i + 64) * 2) + 1]) * \
1159 s16X[ChOffset + i + 64]) >> \
1160 16); \
1161 /*printf("s32DCTY8: %d = 0x%4x%4x * %d\n", s32DCTY[i], gas32CoeffFor8SBs[i \
1162 * * 2], (gas32CoeffFor8SBs[(i * 2) + 1]), s16X[ChOffset+i]);*/ \
1163 /*s32DCTY[i]=(int32_t)(s64Temp>>16);*/ \
1164 }
1165 #endif
1166 #define WINDOW_PARTIAL_8 \
1167 { \
1168 WINDOW_ACCU_8(0); \
1169 WINDOW_ACCU_8(1); \
1170 WINDOW_ACCU_8(2); \
1171 WINDOW_ACCU_8(3); \
1172 WINDOW_ACCU_8(4); \
1173 WINDOW_ACCU_8(5); \
1174 WINDOW_ACCU_8(6); \
1175 WINDOW_ACCU_8(7); \
1176 WINDOW_ACCU_8(8); \
1177 WINDOW_ACCU_8(9); \
1178 WINDOW_ACCU_8(10); \
1179 WINDOW_ACCU_8(11); \
1180 WINDOW_ACCU_8(12); \
1181 WINDOW_ACCU_8(13); \
1182 WINDOW_ACCU_8(14); \
1183 WINDOW_ACCU_8(15); \
1184 }
1185 #endif
1186 #endif
1187
1188 static int16_t ShiftCounter = 0;
1189 extern int16_t EncMaxShiftCounter;
1190 /****************************************************************************
1191 * SbcAnalysisFilter - performs Analysis of the input audio stream
1192 *
1193 * RETURNS : N/A
1194 */
SbcAnalysisFilter4(SBC_ENC_PARAMS * pstrEncParams,int16_t * input)1195 void SbcAnalysisFilter4(SBC_ENC_PARAMS* pstrEncParams, int16_t* input) {
1196 int16_t* ps16PcmBuf;
1197 int32_t* ps32SbBuf;
1198 int32_t s32Blk, s32Ch;
1199 int32_t s32NumOfChannels, s32NumOfBlocks;
1200 int32_t i, *ps32X, *ps32X2;
1201 int32_t Offset, Offset2, ChOffset;
1202 #if (SBC_ARM_ASM_OPT == TRUE)
1203 register int32_t s32Hi, s32Hi2;
1204 #else
1205 #if (SBC_IPAQ_OPT == TRUE)
1206 #if (SBC_IS_64_MULT_IN_WINDOW_ACCU == TRUE)
1207 register int64_t s64Temp, s64Temp2;
1208 #else
1209 register int32_t s32Temp, s32Temp2;
1210 #endif
1211 #else
1212
1213 #if (SBC_IS_64_MULT_IN_WINDOW_ACCU == TRUE)
1214 int64_t s64Temp;
1215 #endif
1216
1217 #endif
1218 #endif
1219
1220 s32NumOfChannels = pstrEncParams->s16NumOfChannels;
1221 s32NumOfBlocks = pstrEncParams->s16NumOfBlocks;
1222
1223 ps16PcmBuf = input;
1224
1225 ps32SbBuf = pstrEncParams->s32SbBuffer;
1226 Offset2 = (int32_t)(EncMaxShiftCounter + 40);
1227 for (s32Blk = 0; s32Blk < s32NumOfBlocks; s32Blk++) {
1228 Offset = (int32_t)(EncMaxShiftCounter - ShiftCounter);
1229 /* Store new samples */
1230 if (s32NumOfChannels == 1) {
1231 s16X[3 + Offset] = *ps16PcmBuf;
1232 ps16PcmBuf++;
1233 s16X[2 + Offset] = *ps16PcmBuf;
1234 ps16PcmBuf++;
1235 s16X[1 + Offset] = *ps16PcmBuf;
1236 ps16PcmBuf++;
1237 s16X[0 + Offset] = *ps16PcmBuf;
1238 ps16PcmBuf++;
1239 } else {
1240 s16X[3 + Offset] = *ps16PcmBuf;
1241 ps16PcmBuf++;
1242 s16X[Offset2 + 3 + Offset] = *ps16PcmBuf;
1243 ps16PcmBuf++;
1244 s16X[2 + Offset] = *ps16PcmBuf;
1245 ps16PcmBuf++;
1246 s16X[Offset2 + 2 + Offset] = *ps16PcmBuf;
1247 ps16PcmBuf++;
1248 s16X[1 + Offset] = *ps16PcmBuf;
1249 ps16PcmBuf++;
1250 s16X[Offset2 + 1 + Offset] = *ps16PcmBuf;
1251 ps16PcmBuf++;
1252 s16X[0 + Offset] = *ps16PcmBuf;
1253 ps16PcmBuf++;
1254 s16X[Offset2 + 0 + Offset] = *ps16PcmBuf;
1255 ps16PcmBuf++;
1256 }
1257 for (s32Ch = 0; s32Ch < s32NumOfChannels; s32Ch++) {
1258 ChOffset = s32Ch * Offset2 + Offset;
1259
1260 WINDOW_PARTIAL_4
1261
1262 SBC_FastIDCT4(s32DCTY, ps32SbBuf);
1263
1264 ps32SbBuf += SUB_BANDS_4;
1265 }
1266 if (s32NumOfChannels == 1) {
1267 if (ShiftCounter >= EncMaxShiftCounter) {
1268 SHIFTUP_X4;
1269 ShiftCounter = 0;
1270 } else {
1271 ShiftCounter += SUB_BANDS_4;
1272 }
1273 } else {
1274 if (ShiftCounter >= EncMaxShiftCounter) {
1275 SHIFTUP_X4_2;
1276 ShiftCounter = 0;
1277 } else {
1278 ShiftCounter += SUB_BANDS_4;
1279 }
1280 }
1281 }
1282 }
1283
1284 /* ////////////////////////////////////////////////////////////////////////// */
SbcAnalysisFilter8(SBC_ENC_PARAMS * pstrEncParams,int16_t * input)1285 void SbcAnalysisFilter8(SBC_ENC_PARAMS* pstrEncParams, int16_t* input) {
1286 int16_t* ps16PcmBuf;
1287 int32_t* ps32SbBuf;
1288 int32_t s32Blk, s32Ch; /* counter for block*/
1289 int32_t Offset, Offset2;
1290 int32_t s32NumOfChannels, s32NumOfBlocks;
1291 int32_t i, *ps32X, *ps32X2;
1292 int32_t ChOffset;
1293 #if (SBC_ARM_ASM_OPT == TRUE)
1294 register int32_t s32Hi, s32Hi2;
1295 #else
1296 #if (SBC_IPAQ_OPT == TRUE)
1297 #if (SBC_IS_64_MULT_IN_WINDOW_ACCU == TRUE)
1298 register int64_t s64Temp, s64Temp2;
1299 #else
1300 register int32_t s32Temp, s32Temp2;
1301 #endif
1302 #else
1303 #if (SBC_IS_64_MULT_IN_WINDOW_ACCU == TRUE)
1304 int64_t s64Temp;
1305 #endif
1306 #endif
1307 #endif
1308
1309 s32NumOfChannels = pstrEncParams->s16NumOfChannels;
1310 s32NumOfBlocks = pstrEncParams->s16NumOfBlocks;
1311
1312 ps16PcmBuf = input;
1313
1314 ps32SbBuf = pstrEncParams->s32SbBuffer;
1315 Offset2 = (int32_t)(EncMaxShiftCounter + 80);
1316 for (s32Blk = 0; s32Blk < s32NumOfBlocks; s32Blk++) {
1317 Offset = (int32_t)(EncMaxShiftCounter - ShiftCounter);
1318 /* Store new samples */
1319 if (s32NumOfChannels == 1) {
1320 s16X[7 + Offset] = *ps16PcmBuf;
1321 ps16PcmBuf++;
1322 s16X[6 + Offset] = *ps16PcmBuf;
1323 ps16PcmBuf++;
1324 s16X[5 + Offset] = *ps16PcmBuf;
1325 ps16PcmBuf++;
1326 s16X[4 + Offset] = *ps16PcmBuf;
1327 ps16PcmBuf++;
1328 s16X[3 + Offset] = *ps16PcmBuf;
1329 ps16PcmBuf++;
1330 s16X[2 + Offset] = *ps16PcmBuf;
1331 ps16PcmBuf++;
1332 s16X[1 + Offset] = *ps16PcmBuf;
1333 ps16PcmBuf++;
1334 s16X[0 + Offset] = *ps16PcmBuf;
1335 ps16PcmBuf++;
1336 } else {
1337 s16X[7 + Offset] = *ps16PcmBuf;
1338 ps16PcmBuf++;
1339 s16X[Offset2 + 7 + Offset] = *ps16PcmBuf;
1340 ps16PcmBuf++;
1341 s16X[6 + Offset] = *ps16PcmBuf;
1342 ps16PcmBuf++;
1343 s16X[Offset2 + 6 + Offset] = *ps16PcmBuf;
1344 ps16PcmBuf++;
1345 s16X[5 + Offset] = *ps16PcmBuf;
1346 ps16PcmBuf++;
1347 s16X[Offset2 + 5 + Offset] = *ps16PcmBuf;
1348 ps16PcmBuf++;
1349 s16X[4 + Offset] = *ps16PcmBuf;
1350 ps16PcmBuf++;
1351 s16X[Offset2 + 4 + Offset] = *ps16PcmBuf;
1352 ps16PcmBuf++;
1353 s16X[3 + Offset] = *ps16PcmBuf;
1354 ps16PcmBuf++;
1355 s16X[Offset2 + 3 + Offset] = *ps16PcmBuf;
1356 ps16PcmBuf++;
1357 s16X[2 + Offset] = *ps16PcmBuf;
1358 ps16PcmBuf++;
1359 s16X[Offset2 + 2 + Offset] = *ps16PcmBuf;
1360 ps16PcmBuf++;
1361 s16X[1 + Offset] = *ps16PcmBuf;
1362 ps16PcmBuf++;
1363 s16X[Offset2 + 1 + Offset] = *ps16PcmBuf;
1364 ps16PcmBuf++;
1365 s16X[0 + Offset] = *ps16PcmBuf;
1366 ps16PcmBuf++;
1367 s16X[Offset2 + 0 + Offset] = *ps16PcmBuf;
1368 ps16PcmBuf++;
1369 }
1370 for (s32Ch = 0; s32Ch < s32NumOfChannels; s32Ch++) {
1371 ChOffset = s32Ch * Offset2 + Offset;
1372
1373 WINDOW_PARTIAL_8
1374
1375 SBC_FastIDCT8(s32DCTY, ps32SbBuf);
1376
1377 ps32SbBuf += SUB_BANDS_8;
1378 }
1379 if (s32NumOfChannels == 1) {
1380 if (ShiftCounter >= EncMaxShiftCounter) {
1381 SHIFTUP_X8;
1382 ShiftCounter = 0;
1383 } else {
1384 ShiftCounter += SUB_BANDS_8;
1385 }
1386 } else {
1387 if (ShiftCounter >= EncMaxShiftCounter) {
1388 SHIFTUP_X8_2;
1389 ShiftCounter = 0;
1390 } else {
1391 ShiftCounter += SUB_BANDS_8;
1392 }
1393 }
1394 }
1395 }
1396
SbcAnalysisInit(void)1397 void SbcAnalysisInit(void) {
1398 memset(s16X, 0, ENC_VX_BUFFER_SIZE * sizeof(int16_t));
1399 ShiftCounter = 0;
1400 }
1401