1 // Copyright 2014 Google Inc. All Rights Reserved.
2 //
3 // Use of this source code is governed by a BSD-style license
4 // that can be found in the COPYING file in the root of the source
5 // tree. An additional intellectual property rights grant can be found
6 // in the file PATENTS. All contributing project authors may
7 // be found in the AUTHORS file in the root of the source tree.
8 // -----------------------------------------------------------------------------
9 //
10 // MIPS version of speed-critical encoding functions.
11 //
12 // Author(s): Djordje Pesut (djordje.pesut@imgtec.com)
13 // Jovan Zelincevic (jovan.zelincevic@imgtec.com)
14 // Slobodan Prijic (slobodan.prijic@imgtec.com)
15
16 #include "./dsp.h"
17
18 #if defined(WEBP_USE_MIPS32)
19
20 #include "../enc/vp8enci.h"
21 #include "../enc/cost.h"
22
23 #if defined(__GNUC__) && defined(__ANDROID__) && LOCAL_GCC_VERSION == 0x409
24 #define WORK_AROUND_GCC
25 #endif
26
27 static const int kC1 = 20091 + (1 << 16);
28 static const int kC2 = 35468;
29
30 // macro for one vertical pass in ITransformOne
31 // MUL macro inlined
32 // temp0..temp15 holds tmp[0]..tmp[15]
33 // A..D - offsets in bytes to load from in buffer
34 // TEMP0..TEMP3 - registers for corresponding tmp elements
35 // TEMP4..TEMP5 - temporary registers
36 #define VERTICAL_PASS(A, B, C, D, TEMP4, TEMP0, TEMP1, TEMP2, TEMP3) \
37 "lh %[temp16], "#A"(%[temp20]) \n\t" \
38 "lh %[temp18], "#B"(%[temp20]) \n\t" \
39 "lh %[temp17], "#C"(%[temp20]) \n\t" \
40 "lh %[temp19], "#D"(%[temp20]) \n\t" \
41 "addu %["#TEMP4"], %[temp16], %[temp18] \n\t" \
42 "subu %[temp16], %[temp16], %[temp18] \n\t" \
43 "mul %["#TEMP0"], %[temp17], %[kC2] \n\t" \
44 "mul %[temp18], %[temp19], %[kC1] \n\t" \
45 "mul %[temp17], %[temp17], %[kC1] \n\t" \
46 "mul %[temp19], %[temp19], %[kC2] \n\t" \
47 "sra %["#TEMP0"], %["#TEMP0"], 16 \n\n" \
48 "sra %[temp18], %[temp18], 16 \n\n" \
49 "sra %[temp17], %[temp17], 16 \n\n" \
50 "sra %[temp19], %[temp19], 16 \n\n" \
51 "subu %["#TEMP2"], %["#TEMP0"], %[temp18] \n\t" \
52 "addu %["#TEMP3"], %[temp17], %[temp19] \n\t" \
53 "addu %["#TEMP0"], %["#TEMP4"], %["#TEMP3"] \n\t" \
54 "addu %["#TEMP1"], %[temp16], %["#TEMP2"] \n\t" \
55 "subu %["#TEMP2"], %[temp16], %["#TEMP2"] \n\t" \
56 "subu %["#TEMP3"], %["#TEMP4"], %["#TEMP3"] \n\t"
57
58 // macro for one horizontal pass in ITransformOne
59 // MUL and STORE macros inlined
60 // a = clip_8b(a) is replaced with: a = max(a, 0); a = min(a, 255)
61 // temp0..temp15 holds tmp[0]..tmp[15]
62 // A..D - offsets in bytes to load from ref and store to dst buffer
63 // TEMP0, TEMP4, TEMP8 and TEMP12 - registers for corresponding tmp elements
64 #define HORIZONTAL_PASS(A, B, C, D, TEMP0, TEMP4, TEMP8, TEMP12) \
65 "addiu %["#TEMP0"], %["#TEMP0"], 4 \n\t" \
66 "addu %[temp16], %["#TEMP0"], %["#TEMP8"] \n\t" \
67 "subu %[temp17], %["#TEMP0"], %["#TEMP8"] \n\t" \
68 "mul %["#TEMP0"], %["#TEMP4"], %[kC2] \n\t" \
69 "mul %["#TEMP8"], %["#TEMP12"], %[kC1] \n\t" \
70 "mul %["#TEMP4"], %["#TEMP4"], %[kC1] \n\t" \
71 "mul %["#TEMP12"], %["#TEMP12"], %[kC2] \n\t" \
72 "sra %["#TEMP0"], %["#TEMP0"], 16 \n\t" \
73 "sra %["#TEMP8"], %["#TEMP8"], 16 \n\t" \
74 "sra %["#TEMP4"], %["#TEMP4"], 16 \n\t" \
75 "sra %["#TEMP12"], %["#TEMP12"], 16 \n\t" \
76 "subu %[temp18], %["#TEMP0"], %["#TEMP8"] \n\t" \
77 "addu %[temp19], %["#TEMP4"], %["#TEMP12"] \n\t" \
78 "addu %["#TEMP0"], %[temp16], %[temp19] \n\t" \
79 "addu %["#TEMP4"], %[temp17], %[temp18] \n\t" \
80 "subu %["#TEMP8"], %[temp17], %[temp18] \n\t" \
81 "subu %["#TEMP12"], %[temp16], %[temp19] \n\t" \
82 "lw %[temp20], 0(%[args]) \n\t" \
83 "sra %["#TEMP0"], %["#TEMP0"], 3 \n\t" \
84 "sra %["#TEMP4"], %["#TEMP4"], 3 \n\t" \
85 "sra %["#TEMP8"], %["#TEMP8"], 3 \n\t" \
86 "sra %["#TEMP12"], %["#TEMP12"], 3 \n\t" \
87 "lbu %[temp16], "#A"(%[temp20]) \n\t" \
88 "lbu %[temp17], "#B"(%[temp20]) \n\t" \
89 "lbu %[temp18], "#C"(%[temp20]) \n\t" \
90 "lbu %[temp19], "#D"(%[temp20]) \n\t" \
91 "addu %["#TEMP0"], %[temp16], %["#TEMP0"] \n\t" \
92 "addu %["#TEMP4"], %[temp17], %["#TEMP4"] \n\t" \
93 "addu %["#TEMP8"], %[temp18], %["#TEMP8"] \n\t" \
94 "addu %["#TEMP12"], %[temp19], %["#TEMP12"] \n\t" \
95 "slt %[temp16], %["#TEMP0"], $zero \n\t" \
96 "slt %[temp17], %["#TEMP4"], $zero \n\t" \
97 "slt %[temp18], %["#TEMP8"], $zero \n\t" \
98 "slt %[temp19], %["#TEMP12"], $zero \n\t" \
99 "movn %["#TEMP0"], $zero, %[temp16] \n\t" \
100 "movn %["#TEMP4"], $zero, %[temp17] \n\t" \
101 "movn %["#TEMP8"], $zero, %[temp18] \n\t" \
102 "movn %["#TEMP12"], $zero, %[temp19] \n\t" \
103 "addiu %[temp20], $zero, 255 \n\t" \
104 "slt %[temp16], %["#TEMP0"], %[temp20] \n\t" \
105 "slt %[temp17], %["#TEMP4"], %[temp20] \n\t" \
106 "slt %[temp18], %["#TEMP8"], %[temp20] \n\t" \
107 "slt %[temp19], %["#TEMP12"], %[temp20] \n\t" \
108 "movz %["#TEMP0"], %[temp20], %[temp16] \n\t" \
109 "movz %["#TEMP4"], %[temp20], %[temp17] \n\t" \
110 "lw %[temp16], 8(%[args]) \n\t" \
111 "movz %["#TEMP8"], %[temp20], %[temp18] \n\t" \
112 "movz %["#TEMP12"], %[temp20], %[temp19] \n\t" \
113 "sb %["#TEMP0"], "#A"(%[temp16]) \n\t" \
114 "sb %["#TEMP4"], "#B"(%[temp16]) \n\t" \
115 "sb %["#TEMP8"], "#C"(%[temp16]) \n\t" \
116 "sb %["#TEMP12"], "#D"(%[temp16]) \n\t"
117
118 // Does one or two inverse transforms.
ITransformOne(const uint8_t * ref,const int16_t * in,uint8_t * dst)119 static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
120 uint8_t* dst) {
121 int temp0, temp1, temp2, temp3, temp4, temp5, temp6;
122 int temp7, temp8, temp9, temp10, temp11, temp12, temp13;
123 int temp14, temp15, temp16, temp17, temp18, temp19, temp20;
124 const int* args[3] = {(const int*)ref, (const int*)in, (const int*)dst};
125
126 __asm__ volatile(
127 "lw %[temp20], 4(%[args]) \n\t"
128 VERTICAL_PASS(0, 16, 8, 24, temp4, temp0, temp1, temp2, temp3)
129 VERTICAL_PASS(2, 18, 10, 26, temp8, temp4, temp5, temp6, temp7)
130 VERTICAL_PASS(4, 20, 12, 28, temp12, temp8, temp9, temp10, temp11)
131 VERTICAL_PASS(6, 22, 14, 30, temp20, temp12, temp13, temp14, temp15)
132
133 HORIZONTAL_PASS( 0, 1, 2, 3, temp0, temp4, temp8, temp12)
134 HORIZONTAL_PASS(16, 17, 18, 19, temp1, temp5, temp9, temp13)
135 HORIZONTAL_PASS(32, 33, 34, 35, temp2, temp6, temp10, temp14)
136 HORIZONTAL_PASS(48, 49, 50, 51, temp3, temp7, temp11, temp15)
137
138 : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
139 [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
140 [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8),
141 [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11),
142 [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), [temp14]"=&r"(temp14),
143 [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), [temp17]"=&r"(temp17),
144 [temp18]"=&r"(temp18), [temp19]"=&r"(temp19), [temp20]"=&r"(temp20)
145 : [args]"r"(args), [kC1]"r"(kC1), [kC2]"r"(kC2)
146 : "memory", "hi", "lo"
147 );
148 }
149
ITransform(const uint8_t * ref,const int16_t * in,uint8_t * dst,int do_two)150 static void ITransform(const uint8_t* ref, const int16_t* in,
151 uint8_t* dst, int do_two) {
152 ITransformOne(ref, in, dst);
153 if (do_two) {
154 ITransformOne(ref + 4, in + 16, dst + 4);
155 }
156 }
157
158 #undef VERTICAL_PASS
159 #undef HORIZONTAL_PASS
160
161 // macro for one pass through for loop in QuantizeBlock
162 // QUANTDIV macro inlined
163 // J - offset in bytes (kZigzag[n] * 2)
164 // K - offset in bytes (kZigzag[n] * 4)
165 // N - offset in bytes (n * 2)
166 #define QUANTIZE_ONE(J, K, N) \
167 "lh %[temp0], "#J"(%[ppin]) \n\t" \
168 "lhu %[temp1], "#J"(%[ppsharpen]) \n\t" \
169 "lw %[temp2], "#K"(%[ppzthresh]) \n\t" \
170 "sra %[sign], %[temp0], 15 \n\t" \
171 "xor %[coeff], %[temp0], %[sign] \n\t" \
172 "subu %[coeff], %[coeff], %[sign] \n\t" \
173 "addu %[coeff], %[coeff], %[temp1] \n\t" \
174 "slt %[temp4], %[temp2], %[coeff] \n\t" \
175 "addiu %[temp5], $zero, 0 \n\t" \
176 "addiu %[level], $zero, 0 \n\t" \
177 "beqz %[temp4], 2f \n\t" \
178 "lhu %[temp1], "#J"(%[ppiq]) \n\t" \
179 "lw %[temp2], "#K"(%[ppbias]) \n\t" \
180 "lhu %[temp3], "#J"(%[ppq]) \n\t" \
181 "mul %[level], %[coeff], %[temp1] \n\t" \
182 "addu %[level], %[level], %[temp2] \n\t" \
183 "sra %[level], %[level], 17 \n\t" \
184 "slt %[temp4], %[max_level], %[level] \n\t" \
185 "movn %[level], %[max_level], %[temp4] \n\t" \
186 "xor %[level], %[level], %[sign] \n\t" \
187 "subu %[level], %[level], %[sign] \n\t" \
188 "mul %[temp5], %[level], %[temp3] \n\t" \
189 "2: \n\t" \
190 "sh %[temp5], "#J"(%[ppin]) \n\t" \
191 "sh %[level], "#N"(%[pout]) \n\t"
192
QuantizeBlock(int16_t in[16],int16_t out[16],const VP8Matrix * const mtx)193 static int QuantizeBlock(int16_t in[16], int16_t out[16],
194 const VP8Matrix* const mtx) {
195 int temp0, temp1, temp2, temp3, temp4, temp5;
196 int sign, coeff, level, i;
197 int max_level = MAX_LEVEL;
198
199 int16_t* ppin = &in[0];
200 int16_t* pout = &out[0];
201 const uint16_t* ppsharpen = &mtx->sharpen_[0];
202 const uint32_t* ppzthresh = &mtx->zthresh_[0];
203 const uint16_t* ppq = &mtx->q_[0];
204 const uint16_t* ppiq = &mtx->iq_[0];
205 const uint32_t* ppbias = &mtx->bias_[0];
206
207 __asm__ volatile(
208 QUANTIZE_ONE( 0, 0, 0)
209 QUANTIZE_ONE( 2, 4, 2)
210 QUANTIZE_ONE( 8, 16, 4)
211 QUANTIZE_ONE(16, 32, 6)
212 QUANTIZE_ONE(10, 20, 8)
213 QUANTIZE_ONE( 4, 8, 10)
214 QUANTIZE_ONE( 6, 12, 12)
215 QUANTIZE_ONE(12, 24, 14)
216 QUANTIZE_ONE(18, 36, 16)
217 QUANTIZE_ONE(24, 48, 18)
218 QUANTIZE_ONE(26, 52, 20)
219 QUANTIZE_ONE(20, 40, 22)
220 QUANTIZE_ONE(14, 28, 24)
221 QUANTIZE_ONE(22, 44, 26)
222 QUANTIZE_ONE(28, 56, 28)
223 QUANTIZE_ONE(30, 60, 30)
224
225 : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1),
226 [temp2]"=&r"(temp2), [temp3]"=&r"(temp3),
227 [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
228 [sign]"=&r"(sign), [coeff]"=&r"(coeff),
229 [level]"=&r"(level)
230 : [pout]"r"(pout), [ppin]"r"(ppin),
231 [ppiq]"r"(ppiq), [max_level]"r"(max_level),
232 [ppbias]"r"(ppbias), [ppzthresh]"r"(ppzthresh),
233 [ppsharpen]"r"(ppsharpen), [ppq]"r"(ppq)
234 : "memory", "hi", "lo"
235 );
236
237 // moved out from macro to increase possibility for earlier breaking
238 for (i = 15; i >= 0; i--) {
239 if (out[i]) return 1;
240 }
241 return 0;
242 }
243
244 #undef QUANTIZE_ONE
245
246 // macro for one horizontal pass in Disto4x4 (TTransform)
247 // two calls of function TTransform are merged into single one
248 // A..D - offsets in bytes to load from a and b buffers
249 // E..H - offsets in bytes to store first results to tmp buffer
250 // E1..H1 - offsets in bytes to store second results to tmp buffer
251 #define HORIZONTAL_PASS(A, B, C, D, E, F, G, H, E1, F1, G1, H1) \
252 "lbu %[temp0], "#A"(%[a]) \n\t" \
253 "lbu %[temp1], "#B"(%[a]) \n\t" \
254 "lbu %[temp2], "#C"(%[a]) \n\t" \
255 "lbu %[temp3], "#D"(%[a]) \n\t" \
256 "lbu %[temp4], "#A"(%[b]) \n\t" \
257 "lbu %[temp5], "#B"(%[b]) \n\t" \
258 "lbu %[temp6], "#C"(%[b]) \n\t" \
259 "lbu %[temp7], "#D"(%[b]) \n\t" \
260 "addu %[temp8], %[temp0], %[temp2] \n\t" \
261 "subu %[temp0], %[temp0], %[temp2] \n\t" \
262 "addu %[temp2], %[temp1], %[temp3] \n\t" \
263 "subu %[temp1], %[temp1], %[temp3] \n\t" \
264 "addu %[temp3], %[temp4], %[temp6] \n\t" \
265 "subu %[temp4], %[temp4], %[temp6] \n\t" \
266 "addu %[temp6], %[temp5], %[temp7] \n\t" \
267 "subu %[temp5], %[temp5], %[temp7] \n\t" \
268 "addu %[temp7], %[temp8], %[temp2] \n\t" \
269 "subu %[temp2], %[temp8], %[temp2] \n\t" \
270 "addu %[temp8], %[temp0], %[temp1] \n\t" \
271 "subu %[temp0], %[temp0], %[temp1] \n\t" \
272 "addu %[temp1], %[temp3], %[temp6] \n\t" \
273 "subu %[temp3], %[temp3], %[temp6] \n\t" \
274 "addu %[temp6], %[temp4], %[temp5] \n\t" \
275 "subu %[temp4], %[temp4], %[temp5] \n\t" \
276 "sw %[temp7], "#E"(%[tmp]) \n\t" \
277 "sw %[temp2], "#H"(%[tmp]) \n\t" \
278 "sw %[temp8], "#F"(%[tmp]) \n\t" \
279 "sw %[temp0], "#G"(%[tmp]) \n\t" \
280 "sw %[temp1], "#E1"(%[tmp]) \n\t" \
281 "sw %[temp3], "#H1"(%[tmp]) \n\t" \
282 "sw %[temp6], "#F1"(%[tmp]) \n\t" \
283 "sw %[temp4], "#G1"(%[tmp]) \n\t"
284
285 // macro for one vertical pass in Disto4x4 (TTransform)
286 // two calls of function TTransform are merged into single one
287 // since only one accu is available in mips32r1 instruction set
288 // first is done second call of function TTransform and after
289 // that first one.
290 // const int sum1 = TTransform(a, w);
291 // const int sum2 = TTransform(b, w);
292 // return abs(sum2 - sum1) >> 5;
293 // (sum2 - sum1) is calculated with madds (sub2) and msubs (sub1)
294 // A..D - offsets in bytes to load first results from tmp buffer
295 // A1..D1 - offsets in bytes to load second results from tmp buffer
296 // E..H - offsets in bytes to load from w buffer
297 #define VERTICAL_PASS(A, B, C, D, A1, B1, C1, D1, E, F, G, H) \
298 "lw %[temp0], "#A1"(%[tmp]) \n\t" \
299 "lw %[temp1], "#C1"(%[tmp]) \n\t" \
300 "lw %[temp2], "#B1"(%[tmp]) \n\t" \
301 "lw %[temp3], "#D1"(%[tmp]) \n\t" \
302 "addu %[temp8], %[temp0], %[temp1] \n\t" \
303 "subu %[temp0], %[temp0], %[temp1] \n\t" \
304 "addu %[temp1], %[temp2], %[temp3] \n\t" \
305 "subu %[temp2], %[temp2], %[temp3] \n\t" \
306 "addu %[temp3], %[temp8], %[temp1] \n\t" \
307 "subu %[temp8], %[temp8], %[temp1] \n\t" \
308 "addu %[temp1], %[temp0], %[temp2] \n\t" \
309 "subu %[temp0], %[temp0], %[temp2] \n\t" \
310 "sra %[temp4], %[temp3], 31 \n\t" \
311 "sra %[temp5], %[temp1], 31 \n\t" \
312 "sra %[temp6], %[temp0], 31 \n\t" \
313 "sra %[temp7], %[temp8], 31 \n\t" \
314 "xor %[temp3], %[temp3], %[temp4] \n\t" \
315 "xor %[temp1], %[temp1], %[temp5] \n\t" \
316 "xor %[temp0], %[temp0], %[temp6] \n\t" \
317 "xor %[temp8], %[temp8], %[temp7] \n\t" \
318 "subu %[temp3], %[temp3], %[temp4] \n\t" \
319 "subu %[temp1], %[temp1], %[temp5] \n\t" \
320 "subu %[temp0], %[temp0], %[temp6] \n\t" \
321 "subu %[temp8], %[temp8], %[temp7] \n\t" \
322 "lhu %[temp4], "#E"(%[w]) \n\t" \
323 "lhu %[temp5], "#F"(%[w]) \n\t" \
324 "lhu %[temp6], "#G"(%[w]) \n\t" \
325 "lhu %[temp7], "#H"(%[w]) \n\t" \
326 "madd %[temp4], %[temp3] \n\t" \
327 "madd %[temp5], %[temp1] \n\t" \
328 "madd %[temp6], %[temp0] \n\t" \
329 "madd %[temp7], %[temp8] \n\t" \
330 "lw %[temp0], "#A"(%[tmp]) \n\t" \
331 "lw %[temp1], "#C"(%[tmp]) \n\t" \
332 "lw %[temp2], "#B"(%[tmp]) \n\t" \
333 "lw %[temp3], "#D"(%[tmp]) \n\t" \
334 "addu %[temp8], %[temp0], %[temp1] \n\t" \
335 "subu %[temp0], %[temp0], %[temp1] \n\t" \
336 "addu %[temp1], %[temp2], %[temp3] \n\t" \
337 "subu %[temp2], %[temp2], %[temp3] \n\t" \
338 "addu %[temp3], %[temp8], %[temp1] \n\t" \
339 "subu %[temp1], %[temp8], %[temp1] \n\t" \
340 "addu %[temp8], %[temp0], %[temp2] \n\t" \
341 "subu %[temp0], %[temp0], %[temp2] \n\t" \
342 "sra %[temp2], %[temp3], 31 \n\t" \
343 "xor %[temp3], %[temp3], %[temp2] \n\t" \
344 "subu %[temp3], %[temp3], %[temp2] \n\t" \
345 "msub %[temp4], %[temp3] \n\t" \
346 "sra %[temp2], %[temp8], 31 \n\t" \
347 "sra %[temp3], %[temp0], 31 \n\t" \
348 "sra %[temp4], %[temp1], 31 \n\t" \
349 "xor %[temp8], %[temp8], %[temp2] \n\t" \
350 "xor %[temp0], %[temp0], %[temp3] \n\t" \
351 "xor %[temp1], %[temp1], %[temp4] \n\t" \
352 "subu %[temp8], %[temp8], %[temp2] \n\t" \
353 "subu %[temp0], %[temp0], %[temp3] \n\t" \
354 "subu %[temp1], %[temp1], %[temp4] \n\t" \
355 "msub %[temp5], %[temp8] \n\t" \
356 "msub %[temp6], %[temp0] \n\t" \
357 "msub %[temp7], %[temp1] \n\t"
358
Disto4x4(const uint8_t * const a,const uint8_t * const b,const uint16_t * const w)359 static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
360 const uint16_t* const w) {
361 int tmp[32];
362 int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8;
363
364 __asm__ volatile(
365 HORIZONTAL_PASS( 0, 1, 2, 3, 0, 4, 8, 12, 64, 68, 72, 76)
366 HORIZONTAL_PASS(16, 17, 18, 19, 16, 20, 24, 28, 80, 84, 88, 92)
367 HORIZONTAL_PASS(32, 33, 34, 35, 32, 36, 40, 44, 96, 100, 104, 108)
368 HORIZONTAL_PASS(48, 49, 50, 51, 48, 52, 56, 60, 112, 116, 120, 124)
369 "mthi $zero \n\t"
370 "mtlo $zero \n\t"
371 VERTICAL_PASS( 0, 16, 32, 48, 64, 80, 96, 112, 0, 8, 16, 24)
372 VERTICAL_PASS( 4, 20, 36, 52, 68, 84, 100, 116, 2, 10, 18, 26)
373 VERTICAL_PASS( 8, 24, 40, 56, 72, 88, 104, 120, 4, 12, 20, 28)
374 VERTICAL_PASS(12, 28, 44, 60, 76, 92, 108, 124, 6, 14, 22, 30)
375 "mflo %[temp0] \n\t"
376 "sra %[temp1], %[temp0], 31 \n\t"
377 "xor %[temp0], %[temp0], %[temp1] \n\t"
378 "subu %[temp0], %[temp0], %[temp1] \n\t"
379 "sra %[temp0], %[temp0], 5 \n\t"
380
381 : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
382 [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
383 [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8)
384 : [a]"r"(a), [b]"r"(b), [w]"r"(w), [tmp]"r"(tmp)
385 : "memory", "hi", "lo"
386 );
387
388 return temp0;
389 }
390
391 #undef VERTICAL_PASS
392 #undef HORIZONTAL_PASS
393
Disto16x16(const uint8_t * const a,const uint8_t * const b,const uint16_t * const w)394 static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
395 const uint16_t* const w) {
396 int D = 0;
397 int x, y;
398 for (y = 0; y < 16 * BPS; y += 4 * BPS) {
399 for (x = 0; x < 16; x += 4) {
400 D += Disto4x4(a + x + y, b + x + y, w);
401 }
402 }
403 return D;
404 }
405
406 // macro for one horizontal pass in FTransform
407 // temp0..temp15 holds tmp[0]..tmp[15]
408 // A..D - offsets in bytes to load from src and ref buffers
409 // TEMP0..TEMP3 - registers for corresponding tmp elements
410 #define HORIZONTAL_PASS(A, B, C, D, TEMP0, TEMP1, TEMP2, TEMP3) \
411 "lw %["#TEMP1"], 0(%[args]) \n\t" \
412 "lw %["#TEMP2"], 4(%[args]) \n\t" \
413 "lbu %[temp16], "#A"(%["#TEMP1"]) \n\t" \
414 "lbu %[temp17], "#A"(%["#TEMP2"]) \n\t" \
415 "lbu %[temp18], "#B"(%["#TEMP1"]) \n\t" \
416 "lbu %[temp19], "#B"(%["#TEMP2"]) \n\t" \
417 "subu %[temp20], %[temp16], %[temp17] \n\t" \
418 "lbu %[temp16], "#C"(%["#TEMP1"]) \n\t" \
419 "lbu %[temp17], "#C"(%["#TEMP2"]) \n\t" \
420 "subu %["#TEMP0"], %[temp18], %[temp19] \n\t" \
421 "lbu %[temp18], "#D"(%["#TEMP1"]) \n\t" \
422 "lbu %[temp19], "#D"(%["#TEMP2"]) \n\t" \
423 "subu %["#TEMP1"], %[temp16], %[temp17] \n\t" \
424 "subu %["#TEMP2"], %[temp18], %[temp19] \n\t" \
425 "addu %["#TEMP3"], %[temp20], %["#TEMP2"] \n\t" \
426 "subu %["#TEMP2"], %[temp20], %["#TEMP2"] \n\t" \
427 "addu %[temp20], %["#TEMP0"], %["#TEMP1"] \n\t" \
428 "subu %["#TEMP0"], %["#TEMP0"], %["#TEMP1"] \n\t" \
429 "mul %[temp16], %["#TEMP2"], %[c5352] \n\t" \
430 "mul %[temp17], %["#TEMP2"], %[c2217] \n\t" \
431 "mul %[temp18], %["#TEMP0"], %[c5352] \n\t" \
432 "mul %[temp19], %["#TEMP0"], %[c2217] \n\t" \
433 "addu %["#TEMP1"], %["#TEMP3"], %[temp20] \n\t" \
434 "subu %[temp20], %["#TEMP3"], %[temp20] \n\t" \
435 "sll %["#TEMP0"], %["#TEMP1"], 3 \n\t" \
436 "sll %["#TEMP2"], %[temp20], 3 \n\t" \
437 "addiu %[temp16], %[temp16], 1812 \n\t" \
438 "addiu %[temp17], %[temp17], 937 \n\t" \
439 "addu %[temp16], %[temp16], %[temp19] \n\t" \
440 "subu %[temp17], %[temp17], %[temp18] \n\t" \
441 "sra %["#TEMP1"], %[temp16], 9 \n\t" \
442 "sra %["#TEMP3"], %[temp17], 9 \n\t"
443
444 // macro for one vertical pass in FTransform
445 // temp0..temp15 holds tmp[0]..tmp[15]
446 // A..D - offsets in bytes to store to out buffer
447 // TEMP0, TEMP4, TEMP8 and TEMP12 - registers for corresponding tmp elements
448 #define VERTICAL_PASS(A, B, C, D, TEMP0, TEMP4, TEMP8, TEMP12) \
449 "addu %[temp16], %["#TEMP0"], %["#TEMP12"] \n\t" \
450 "subu %[temp19], %["#TEMP0"], %["#TEMP12"] \n\t" \
451 "addu %[temp17], %["#TEMP4"], %["#TEMP8"] \n\t" \
452 "subu %[temp18], %["#TEMP4"], %["#TEMP8"] \n\t" \
453 "mul %["#TEMP8"], %[temp19], %[c2217] \n\t" \
454 "mul %["#TEMP12"], %[temp18], %[c2217] \n\t" \
455 "mul %["#TEMP4"], %[temp19], %[c5352] \n\t" \
456 "mul %[temp18], %[temp18], %[c5352] \n\t" \
457 "addiu %[temp16], %[temp16], 7 \n\t" \
458 "addu %["#TEMP0"], %[temp16], %[temp17] \n\t" \
459 "sra %["#TEMP0"], %["#TEMP0"], 4 \n\t" \
460 "addu %["#TEMP12"], %["#TEMP12"], %["#TEMP4"] \n\t" \
461 "subu %["#TEMP4"], %[temp16], %[temp17] \n\t" \
462 "sra %["#TEMP4"], %["#TEMP4"], 4 \n\t" \
463 "addiu %["#TEMP8"], %["#TEMP8"], 30000 \n\t" \
464 "addiu %["#TEMP12"], %["#TEMP12"], 12000 \n\t" \
465 "addiu %["#TEMP8"], %["#TEMP8"], 21000 \n\t" \
466 "subu %["#TEMP8"], %["#TEMP8"], %[temp18] \n\t" \
467 "sra %["#TEMP12"], %["#TEMP12"], 16 \n\t" \
468 "sra %["#TEMP8"], %["#TEMP8"], 16 \n\t" \
469 "addiu %[temp16], %["#TEMP12"], 1 \n\t" \
470 "movn %["#TEMP12"], %[temp16], %[temp19] \n\t" \
471 "sh %["#TEMP0"], "#A"(%[temp20]) \n\t" \
472 "sh %["#TEMP4"], "#C"(%[temp20]) \n\t" \
473 "sh %["#TEMP8"], "#D"(%[temp20]) \n\t" \
474 "sh %["#TEMP12"], "#B"(%[temp20]) \n\t"
475
FTransform(const uint8_t * src,const uint8_t * ref,int16_t * out)476 static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
477 int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8;
478 int temp9, temp10, temp11, temp12, temp13, temp14, temp15, temp16;
479 int temp17, temp18, temp19, temp20;
480 const int c2217 = 2217;
481 const int c5352 = 5352;
482 const int* const args[3] =
483 { (const int*)src, (const int*)ref, (const int*)out };
484
485 __asm__ volatile(
486 HORIZONTAL_PASS( 0, 1, 2, 3, temp0, temp1, temp2, temp3)
487 HORIZONTAL_PASS(16, 17, 18, 19, temp4, temp5, temp6, temp7)
488 HORIZONTAL_PASS(32, 33, 34, 35, temp8, temp9, temp10, temp11)
489 HORIZONTAL_PASS(48, 49, 50, 51, temp12, temp13, temp14, temp15)
490 "lw %[temp20], 8(%[args]) \n\t"
491 VERTICAL_PASS(0, 8, 16, 24, temp0, temp4, temp8, temp12)
492 VERTICAL_PASS(2, 10, 18, 26, temp1, temp5, temp9, temp13)
493 VERTICAL_PASS(4, 12, 20, 28, temp2, temp6, temp10, temp14)
494 VERTICAL_PASS(6, 14, 22, 30, temp3, temp7, temp11, temp15)
495
496 : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
497 [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
498 [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8),
499 [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11),
500 [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), [temp14]"=&r"(temp14),
501 [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), [temp17]"=&r"(temp17),
502 [temp18]"=&r"(temp18), [temp19]"=&r"(temp19), [temp20]"=&r"(temp20)
503 : [args]"r"(args), [c2217]"r"(c2217), [c5352]"r"(c5352)
504 : "memory", "hi", "lo"
505 );
506 }
507
508 #undef VERTICAL_PASS
509 #undef HORIZONTAL_PASS
510
511 // Forward declaration.
512 extern int VP8GetResidualCostMIPS32(int ctx0, const VP8Residual* const res);
513
VP8GetResidualCostMIPS32(int ctx0,const VP8Residual * const res)514 int VP8GetResidualCostMIPS32(int ctx0, const VP8Residual* const res) {
515 int n = res->first;
516 // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
517 int p0 = res->prob[n][ctx0][0];
518 const uint16_t* t = res->cost[n][ctx0];
519 int cost;
520 const int const_2 = 2;
521 const int const_255 = 255;
522 const int const_max_level = MAX_VARIABLE_LEVEL;
523 int res_cost;
524 int res_prob;
525 int res_coeffs;
526 int res_last;
527 int v_reg;
528 int b_reg;
529 int ctx_reg;
530 int cost_add, temp_1, temp_2, temp_3;
531
532 if (res->last < 0) {
533 return VP8BitCost(0, p0);
534 }
535
536 cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0;
537
538 res_cost = (int)res->cost;
539 res_prob = (int)res->prob;
540 res_coeffs = (int)res->coeffs;
541 res_last = (int)res->last;
542
543 __asm__ volatile(
544 ".set push \n\t"
545 ".set noreorder \n\t"
546
547 "sll %[temp_1], %[n], 1 \n\t"
548 "addu %[res_coeffs], %[res_coeffs], %[temp_1] \n\t"
549 "slt %[temp_2], %[n], %[res_last] \n\t"
550 "bnez %[temp_2], 1f \n\t"
551 " li %[cost_add], 0 \n\t"
552 "b 2f \n\t"
553 " nop \n\t"
554 "1: \n\t"
555 "lh %[v_reg], 0(%[res_coeffs]) \n\t"
556 "addu %[b_reg], %[n], %[VP8EncBands] \n\t"
557 "move %[temp_1], %[const_max_level] \n\t"
558 "addu %[cost], %[cost], %[cost_add] \n\t"
559 "negu %[temp_2], %[v_reg] \n\t"
560 "slti %[temp_3], %[v_reg], 0 \n\t"
561 "movn %[v_reg], %[temp_2], %[temp_3] \n\t"
562 "lbu %[b_reg], 1(%[b_reg]) \n\t"
563 "li %[cost_add], 0 \n\t"
564
565 "sltiu %[temp_3], %[v_reg], 2 \n\t"
566 "move %[ctx_reg], %[v_reg] \n\t"
567 "movz %[ctx_reg], %[const_2], %[temp_3] \n\t"
568 // cost += VP8LevelCost(t, v);
569 "slt %[temp_3], %[v_reg], %[const_max_level] \n\t"
570 "movn %[temp_1], %[v_reg], %[temp_3] \n\t"
571 "sll %[temp_2], %[v_reg], 1 \n\t"
572 "addu %[temp_2], %[temp_2], %[VP8LevelFixedCosts] \n\t"
573 "lhu %[temp_2], 0(%[temp_2]) \n\t"
574 "sll %[temp_1], %[temp_1], 1 \n\t"
575 "addu %[temp_1], %[temp_1], %[t] \n\t"
576 "lhu %[temp_3], 0(%[temp_1]) \n\t"
577 "addu %[cost], %[cost], %[temp_2] \n\t"
578
579 // t = res->cost[b][ctx];
580 "sll %[temp_1], %[ctx_reg], 7 \n\t"
581 "sll %[temp_2], %[ctx_reg], 3 \n\t"
582 "addu %[cost], %[cost], %[temp_3] \n\t"
583 "addu %[temp_1], %[temp_1], %[temp_2] \n\t"
584 "sll %[temp_2], %[b_reg], 3 \n\t"
585 "sll %[temp_3], %[b_reg], 5 \n\t"
586 "sub %[temp_2], %[temp_3], %[temp_2] \n\t"
587 "sll %[temp_3], %[temp_2], 4 \n\t"
588 "addu %[temp_1], %[temp_1], %[temp_3] \n\t"
589 "addu %[temp_2], %[temp_2], %[res_cost] \n\t"
590 "addiu %[n], %[n], 1 \n\t"
591 "addu %[t], %[temp_1], %[temp_2] \n\t"
592 "slt %[temp_1], %[n], %[res_last] \n\t"
593 "bnez %[temp_1], 1b \n\t"
594 " addiu %[res_coeffs], %[res_coeffs], 2 \n\t"
595 "2: \n\t"
596
597 ".set pop \n\t"
598 : [cost]"+r"(cost), [t]"+r"(t), [n]"+r"(n), [v_reg]"=&r"(v_reg),
599 [ctx_reg]"=&r"(ctx_reg), [b_reg]"=&r"(b_reg), [cost_add]"=&r"(cost_add),
600 [temp_1]"=&r"(temp_1), [temp_2]"=&r"(temp_2), [temp_3]"=&r"(temp_3)
601 : [const_2]"r"(const_2), [const_255]"r"(const_255), [res_last]"r"(res_last),
602 [VP8EntropyCost]"r"(VP8EntropyCost), [VP8EncBands]"r"(VP8EncBands),
603 [const_max_level]"r"(const_max_level), [res_prob]"r"(res_prob),
604 [VP8LevelFixedCosts]"r"(VP8LevelFixedCosts), [res_coeffs]"r"(res_coeffs),
605 [res_cost]"r"(res_cost)
606 : "memory"
607 );
608
609 // Last coefficient is always non-zero
610 {
611 const int v = abs(res->coeffs[n]);
612 assert(v != 0);
613 cost += VP8LevelCost(t, v);
614 if (n < 15) {
615 const int b = VP8EncBands[n + 1];
616 const int ctx = (v == 1) ? 1 : 2;
617 const int last_p0 = res->prob[b][ctx][0];
618 cost += VP8BitCost(0, last_p0);
619 }
620 }
621 return cost;
622 }
623
624 #define GET_SSE_INNER(A, B, C, D) \
625 "lbu %[temp0], "#A"(%[a]) \n\t" \
626 "lbu %[temp1], "#A"(%[b]) \n\t" \
627 "lbu %[temp2], "#B"(%[a]) \n\t" \
628 "lbu %[temp3], "#B"(%[b]) \n\t" \
629 "lbu %[temp4], "#C"(%[a]) \n\t" \
630 "lbu %[temp5], "#C"(%[b]) \n\t" \
631 "lbu %[temp6], "#D"(%[a]) \n\t" \
632 "lbu %[temp7], "#D"(%[b]) \n\t" \
633 "subu %[temp0], %[temp0], %[temp1] \n\t" \
634 "subu %[temp2], %[temp2], %[temp3] \n\t" \
635 "subu %[temp4], %[temp4], %[temp5] \n\t" \
636 "subu %[temp6], %[temp6], %[temp7] \n\t" \
637 "madd %[temp0], %[temp0] \n\t" \
638 "madd %[temp2], %[temp2] \n\t" \
639 "madd %[temp4], %[temp4] \n\t" \
640 "madd %[temp6], %[temp6] \n\t"
641
642 #define GET_SSE(A, B, C, D) \
643 GET_SSE_INNER(A, A + 1, A + 2, A + 3) \
644 GET_SSE_INNER(B, B + 1, B + 2, B + 3) \
645 GET_SSE_INNER(C, C + 1, C + 2, C + 3) \
646 GET_SSE_INNER(D, D + 1, D + 2, D + 3)
647
648 #if !defined(WORK_AROUND_GCC)
SSE16x16(const uint8_t * a,const uint8_t * b)649 static int SSE16x16(const uint8_t* a, const uint8_t* b) {
650 int count;
651 int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
652
653 __asm__ volatile(
654 "mult $zero, $zero \n\t"
655
656 GET_SSE( 0, 4, 8, 12)
657 GET_SSE( 16, 20, 24, 28)
658 GET_SSE( 32, 36, 40, 44)
659 GET_SSE( 48, 52, 56, 60)
660 GET_SSE( 64, 68, 72, 76)
661 GET_SSE( 80, 84, 88, 92)
662 GET_SSE( 96, 100, 104, 108)
663 GET_SSE(112, 116, 120, 124)
664 GET_SSE(128, 132, 136, 140)
665 GET_SSE(144, 148, 152, 156)
666 GET_SSE(160, 164, 168, 172)
667 GET_SSE(176, 180, 184, 188)
668 GET_SSE(192, 196, 200, 204)
669 GET_SSE(208, 212, 216, 220)
670 GET_SSE(224, 228, 232, 236)
671 GET_SSE(240, 244, 248, 252)
672
673 "mflo %[count] \n\t"
674 : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
675 [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
676 [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count)
677 : [a]"r"(a), [b]"r"(b)
678 : "memory", "hi" , "lo"
679 );
680 return count;
681 }
682
SSE16x8(const uint8_t * a,const uint8_t * b)683 static int SSE16x8(const uint8_t* a, const uint8_t* b) {
684 int count;
685 int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
686
687 __asm__ volatile(
688 "mult $zero, $zero \n\t"
689
690 GET_SSE( 0, 4, 8, 12)
691 GET_SSE( 16, 20, 24, 28)
692 GET_SSE( 32, 36, 40, 44)
693 GET_SSE( 48, 52, 56, 60)
694 GET_SSE( 64, 68, 72, 76)
695 GET_SSE( 80, 84, 88, 92)
696 GET_SSE( 96, 100, 104, 108)
697 GET_SSE(112, 116, 120, 124)
698
699 "mflo %[count] \n\t"
700 : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
701 [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
702 [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count)
703 : [a]"r"(a), [b]"r"(b)
704 : "memory", "hi" , "lo"
705 );
706 return count;
707 }
708
SSE8x8(const uint8_t * a,const uint8_t * b)709 static int SSE8x8(const uint8_t* a, const uint8_t* b) {
710 int count;
711 int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
712
713 __asm__ volatile(
714 "mult $zero, $zero \n\t"
715
716 GET_SSE( 0, 4, 16, 20)
717 GET_SSE(32, 36, 48, 52)
718 GET_SSE(64, 68, 80, 84)
719 GET_SSE(96, 100, 112, 116)
720
721 "mflo %[count] \n\t"
722 : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
723 [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
724 [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count)
725 : [a]"r"(a), [b]"r"(b)
726 : "memory", "hi" , "lo"
727 );
728 return count;
729 }
730
SSE4x4(const uint8_t * a,const uint8_t * b)731 static int SSE4x4(const uint8_t* a, const uint8_t* b) {
732 int count;
733 int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
734
735 __asm__ volatile(
736 "mult $zero, $zero \n\t"
737
738 GET_SSE(0, 16, 32, 48)
739
740 "mflo %[count] \n\t"
741 : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
742 [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5),
743 [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count)
744 : [a]"r"(a), [b]"r"(b)
745 : "memory", "hi" , "lo"
746 );
747 return count;
748 }
749
750 #endif // WORK_AROUND_GCC
751
752 #undef GET_SSE_MIPS32
753 #undef GET_SSE_MIPS32_INNER
754
755 #endif // WEBP_USE_MIPS32
756
757 //------------------------------------------------------------------------------
758 // Entry point
759
760 extern void VP8EncDspInitMIPS32(void);
761
VP8EncDspInitMIPS32(void)762 void VP8EncDspInitMIPS32(void) {
763 #if defined(WEBP_USE_MIPS32)
764 VP8ITransform = ITransform;
765 VP8EncQuantizeBlock = QuantizeBlock;
766 VP8TDisto4x4 = Disto4x4;
767 VP8TDisto16x16 = Disto16x16;
768 VP8FTransform = FTransform;
769 #if !defined(WORK_AROUND_GCC)
770 VP8SSE16x16 = SSE16x16;
771 VP8SSE8x8 = SSE8x8;
772 VP8SSE16x8 = SSE16x8;
773 VP8SSE4x4 = SSE4x4;
774 #endif
775 #endif // WEBP_USE_MIPS32
776 }
777