1 /*
2  *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
3  *
4  *  Use of this source code is governed by a BSD-style license
5  *  that can be found in the LICENSE file in the root of the source
6  *  tree. An additional intellectual property rights grant can be found
7  *  in the file PATENTS.  All contributing project authors may
8  *  be found in the AUTHORS file in the root of the source tree.
9  */
10 
11 #include "vp8_rtcd.h"
12 
13 #if HAVE_DSPR2
14 #define CROP_WIDTH 256
15 
16 /******************************************************************************
17  * Notes:
18  *
19  * This implementation makes use of 16 bit fixed point version of two multiply
20  * constants:
21  *         1.   sqrt(2) * cos (pi/8)
22  *         2.   sqrt(2) * sin (pi/8)
23  * Since the first constant is bigger than 1, to maintain the same 16 bit
24  * fixed point precision as the second one, we use a trick of
25  *         x * a = x + x*(a-1)
26  * so
27  *         x * sqrt(2) * cos (pi/8) = x + x * (sqrt(2) *cos(pi/8)-1).
28  ****************************************************************************/
29 extern unsigned char ff_cropTbl[256 + 2 * CROP_WIDTH];
30 static const int cospi8sqrt2minus1 = 20091;
31 static const int sinpi8sqrt2      = 35468;
32 
prefetch_load_short(short * src)33 inline void prefetch_load_short(short *src)
34 {
35     __asm__ __volatile__ (
36         "pref   0,  0(%[src])   \n\t"
37         :
38         : [src] "r" (src)
39     );
40 }
41 
vp8_short_idct4x4llm_dspr2(short * input,unsigned char * pred_ptr,int pred_stride,unsigned char * dst_ptr,int dst_stride)42 void vp8_short_idct4x4llm_dspr2(short *input, unsigned char *pred_ptr,
43                                 int pred_stride, unsigned char *dst_ptr,
44                                 int dst_stride)
45 {
46     int r, c;
47     int a1, b1, c1, d1;
48     short output[16];
49     short *ip = input;
50     short *op = output;
51     int temp1, temp2;
52     int shortpitch = 4;
53 
54     int c2, d2;
55     int temp3, temp4;
56     unsigned char *cm = ff_cropTbl + CROP_WIDTH;
57 
58     /* prepare data for load */
59     prefetch_load_short(ip + 8);
60 
61     /* first loop is unrolled */
62     a1 = ip[0] + ip[8];
63     b1 = ip[0] - ip[8];
64 
65     temp1 = (ip[4] * sinpi8sqrt2) >> 16;
66     temp2 = ip[12] + ((ip[12] * cospi8sqrt2minus1) >> 16);
67     c1 = temp1 - temp2;
68 
69     temp1 = ip[4] + ((ip[4] * cospi8sqrt2minus1) >> 16);
70     temp2 = (ip[12] * sinpi8sqrt2) >> 16;
71     d1 = temp1 + temp2;
72 
73     temp3 = (ip[5] * sinpi8sqrt2) >> 16;
74     temp4 = ip[13] + ((ip[13] * cospi8sqrt2minus1) >> 16);
75     c2 = temp3 - temp4;
76 
77     temp3 = ip[5] + ((ip[5] * cospi8sqrt2minus1) >> 16);
78     temp4 = (ip[13] * sinpi8sqrt2) >> 16;
79     d2 = temp3 + temp4;
80 
81     op[0] = a1 + d1;
82     op[12] = a1 - d1;
83     op[4] = b1 + c1;
84     op[8] = b1 - c1;
85 
86     a1 = ip[1] + ip[9];
87     b1 = ip[1] - ip[9];
88 
89     op[1] = a1 + d2;
90     op[13] = a1 - d2;
91     op[5] = b1 + c2;
92     op[9] = b1 - c2;
93 
94     a1 = ip[2] + ip[10];
95     b1 = ip[2] - ip[10];
96 
97     temp1 = (ip[6] * sinpi8sqrt2) >> 16;
98     temp2 = ip[14] + ((ip[14] * cospi8sqrt2minus1) >> 16);
99     c1 = temp1 - temp2;
100 
101     temp1 = ip[6] + ((ip[6] * cospi8sqrt2minus1) >> 16);
102     temp2 = (ip[14] * sinpi8sqrt2) >> 16;
103     d1 = temp1 + temp2;
104 
105     temp3 = (ip[7] * sinpi8sqrt2) >> 16;
106     temp4 = ip[15] + ((ip[15] * cospi8sqrt2minus1) >> 16);
107     c2 = temp3 - temp4;
108 
109     temp3 = ip[7] + ((ip[7] * cospi8sqrt2minus1) >> 16);
110     temp4 = (ip[15] * sinpi8sqrt2) >> 16;
111     d2 = temp3 + temp4;
112 
113     op[2] = a1 + d1;
114     op[14] = a1 - d1;
115     op[6] = b1 + c1;
116     op[10] = b1 - c1;
117 
118     a1 = ip[3] + ip[11];
119     b1 = ip[3] - ip[11];
120 
121     op[3] = a1 + d2;
122     op[15] = a1 - d2;
123     op[7] = b1 + c2;
124     op[11] = b1 - c2;
125 
126     ip = output;
127 
128     /* prepare data for load */
129     prefetch_load_short(ip + shortpitch);
130 
131     /* second loop is unrolled */
132     a1 = ip[0] + ip[2];
133     b1 = ip[0] - ip[2];
134 
135     temp1 = (ip[1] * sinpi8sqrt2) >> 16;
136     temp2 = ip[3] + ((ip[3] * cospi8sqrt2minus1) >> 16);
137     c1 = temp1 - temp2;
138 
139     temp1 = ip[1] + ((ip[1] * cospi8sqrt2minus1) >> 16);
140     temp2 = (ip[3] * sinpi8sqrt2) >> 16;
141     d1 = temp1 + temp2;
142 
143     temp3 = (ip[5] * sinpi8sqrt2) >> 16;
144     temp4 = ip[7] + ((ip[7] * cospi8sqrt2minus1) >> 16);
145     c2 = temp3 - temp4;
146 
147     temp3 = ip[5] + ((ip[5] * cospi8sqrt2minus1) >> 16);
148     temp4 = (ip[7] * sinpi8sqrt2) >> 16;
149     d2 = temp3 + temp4;
150 
151     op[0] = (a1 + d1 + 4) >> 3;
152     op[3] = (a1 - d1 + 4) >> 3;
153     op[1] = (b1 + c1 + 4) >> 3;
154     op[2] = (b1 - c1 + 4) >> 3;
155 
156     a1 = ip[4] + ip[6];
157     b1 = ip[4] - ip[6];
158 
159     op[4] = (a1 + d2 + 4) >> 3;
160     op[7] = (a1 - d2 + 4) >> 3;
161     op[5] = (b1 + c2 + 4) >> 3;
162     op[6] = (b1 - c2 + 4) >> 3;
163 
164     a1 = ip[8] + ip[10];
165     b1 = ip[8] - ip[10];
166 
167     temp1 = (ip[9] * sinpi8sqrt2) >> 16;
168     temp2 = ip[11] + ((ip[11] * cospi8sqrt2minus1) >> 16);
169     c1 = temp1 - temp2;
170 
171     temp1 = ip[9] + ((ip[9] * cospi8sqrt2minus1) >> 16);
172     temp2 = (ip[11] * sinpi8sqrt2) >> 16;
173     d1 = temp1 + temp2;
174 
175     temp3 = (ip[13] * sinpi8sqrt2) >> 16;
176     temp4 = ip[15] + ((ip[15] * cospi8sqrt2minus1) >> 16);
177     c2 = temp3 - temp4;
178 
179     temp3 = ip[13] + ((ip[13] * cospi8sqrt2minus1) >> 16);
180     temp4 = (ip[15] * sinpi8sqrt2) >> 16;
181     d2 = temp3 + temp4;
182 
183     op[8] = (a1 + d1 + 4) >> 3;
184     op[11] = (a1 - d1 + 4) >> 3;
185     op[9] = (b1 + c1 + 4) >> 3;
186     op[10] = (b1 - c1 + 4) >> 3;
187 
188     a1 = ip[12] + ip[14];
189     b1 = ip[12] - ip[14];
190 
191     op[12] = (a1 + d2 + 4) >> 3;
192     op[15] = (a1 - d2 + 4) >> 3;
193     op[13] = (b1 + c2 + 4) >> 3;
194     op[14] = (b1 - c2 + 4) >> 3;
195 
196     ip = output;
197 
198     for (r = 0; r < 4; r++)
199     {
200         for (c = 0; c < 4; c++)
201         {
202             short a = ip[c] + pred_ptr[c] ;
203             dst_ptr[c] = cm[a] ;
204         }
205 
206         ip += 4;
207         dst_ptr += dst_stride;
208         pred_ptr += pred_stride;
209     }
210 }
211 
vp8_dc_only_idct_add_dspr2(short input_dc,unsigned char * pred_ptr,int pred_stride,unsigned char * dst_ptr,int dst_stride)212 void vp8_dc_only_idct_add_dspr2(short input_dc, unsigned char *pred_ptr, int pred_stride, unsigned char *dst_ptr, int dst_stride)
213 {
214     int a1;
215     int i, absa1;
216     int t2, vector_a1, vector_a;
217 
218     /* a1 = ((input_dc + 4) >> 3); */
219     __asm__ __volatile__ (
220         "addi  %[a1], %[input_dc], 4   \n\t"
221         "sra   %[a1], %[a1],       3   \n\t"
222         : [a1] "=r" (a1)
223         : [input_dc] "r" (input_dc)
224     );
225 
226     if (a1 < 0)
227     {
228         /* use quad-byte
229          * input and output memory are four byte aligned
230          */
231         __asm__ __volatile__ (
232             "abs        %[absa1],     %[a1]         \n\t"
233             "replv.qb   %[vector_a1], %[absa1]      \n\t"
234             : [absa1] "=r" (absa1), [vector_a1] "=r" (vector_a1)
235             : [a1] "r" (a1)
236         );
237 
238         /* use (a1 - predptr[c]) instead a1 + predptr[c] */
239         for (i = 4; i--;)
240         {
241             __asm__ __volatile__ (
242                 "lw             %[t2],       0(%[pred_ptr])                     \n\t"
243                 "add            %[pred_ptr], %[pred_ptr],    %[pred_stride]     \n\t"
244                 "subu_s.qb      %[vector_a], %[t2],          %[vector_a1]       \n\t"
245                 "sw             %[vector_a], 0(%[dst_ptr])                      \n\t"
246                 "add            %[dst_ptr],  %[dst_ptr],     %[dst_stride]      \n\t"
247                 : [t2] "=&r" (t2), [vector_a] "=&r" (vector_a),
248                   [dst_ptr] "+&r" (dst_ptr), [pred_ptr] "+&r" (pred_ptr)
249                 : [dst_stride] "r" (dst_stride), [pred_stride] "r" (pred_stride), [vector_a1] "r" (vector_a1)
250             );
251         }
252     }
253     else
254     {
255         /* use quad-byte
256          * input and output memory are four byte aligned
257          */
258         __asm__ __volatile__ (
259             "replv.qb       %[vector_a1], %[a1]     \n\t"
260             : [vector_a1] "=r" (vector_a1)
261             : [a1] "r" (a1)
262         );
263 
264         for (i = 4; i--;)
265         {
266             __asm__ __volatile__ (
267                 "lw             %[t2],       0(%[pred_ptr])                 \n\t"
268                 "add            %[pred_ptr], %[pred_ptr],    %[pred_stride] \n\t"
269                 "addu_s.qb      %[vector_a], %[vector_a1],   %[t2]          \n\t"
270                 "sw             %[vector_a], 0(%[dst_ptr])                  \n\t"
271                 "add            %[dst_ptr],  %[dst_ptr],     %[dst_stride]  \n\t"
272                 : [t2] "=&r" (t2), [vector_a] "=&r" (vector_a),
273                   [dst_ptr] "+&r" (dst_ptr), [pred_ptr] "+&r" (pred_ptr)
274                 : [dst_stride] "r" (dst_stride), [pred_stride] "r" (pred_stride), [vector_a1] "r" (vector_a1)
275             );
276         }
277     }
278 
279 }
280 
vp8_short_inv_walsh4x4_dspr2(short * input,short * mb_dqcoeff)281 void vp8_short_inv_walsh4x4_dspr2(short *input, short *mb_dqcoeff)
282 {
283     short output[16];
284     int i;
285     int a1, b1, c1, d1;
286     int a2, b2, c2, d2;
287     short *ip = input;
288     short *op = output;
289 
290     prefetch_load_short(ip);
291 
292     for (i = 4; i--;)
293     {
294         a1 = ip[0] + ip[12];
295         b1 = ip[4] + ip[8];
296         c1 = ip[4] - ip[8];
297         d1 = ip[0] - ip[12];
298 
299         op[0] = a1 + b1;
300         op[4] = c1 + d1;
301         op[8] = a1 - b1;
302         op[12] = d1 - c1;
303 
304         ip++;
305         op++;
306     }
307 
308     ip = output;
309     op = output;
310 
311     prefetch_load_short(ip);
312 
313     for (i = 4; i--;)
314     {
315         a1 = ip[0] + ip[3] + 3;
316         b1 = ip[1] + ip[2];
317         c1 = ip[1] - ip[2];
318         d1 = ip[0] - ip[3] + 3;
319 
320         a2 = a1 + b1;
321         b2 = d1 + c1;
322         c2 = a1 - b1;
323         d2 = d1 - c1;
324 
325         op[0] = a2 >> 3;
326         op[1] = b2 >> 3;
327         op[2] = c2 >> 3;
328         op[3] = d2 >> 3;
329 
330         ip += 4;
331         op += 4;
332     }
333 
334     for (i = 0; i < 16; i++)
335     {
336         mb_dqcoeff[i * 16] = output[i];
337     }
338 }
339 
vp8_short_inv_walsh4x4_1_dspr2(short * input,short * mb_dqcoeff)340 void vp8_short_inv_walsh4x4_1_dspr2(short *input, short *mb_dqcoeff)
341 {
342     int a1;
343 
344     a1 = ((input[0] + 3) >> 3);
345 
346     __asm__ __volatile__ (
347         "sh             %[a1], 0(%[mb_dqcoeff])                    \n\t"
348         "sh             %[a1], 32(%[mb_dqcoeff])                   \n\t"
349         "sh             %[a1], 64(%[mb_dqcoeff])                   \n\t"
350         "sh             %[a1], 96(%[mb_dqcoeff])                   \n\t"
351         "sh             %[a1], 128(%[mb_dqcoeff])                  \n\t"
352         "sh             %[a1], 160(%[mb_dqcoeff])                  \n\t"
353         "sh             %[a1], 192(%[mb_dqcoeff])                  \n\t"
354         "sh             %[a1], 224(%[mb_dqcoeff])                  \n\t"
355         "sh             %[a1], 256(%[mb_dqcoeff])                  \n\t"
356         "sh             %[a1], 288(%[mb_dqcoeff])                  \n\t"
357         "sh             %[a1], 320(%[mb_dqcoeff])                  \n\t"
358         "sh             %[a1], 352(%[mb_dqcoeff])                  \n\t"
359         "sh             %[a1], 384(%[mb_dqcoeff])                  \n\t"
360         "sh             %[a1], 416(%[mb_dqcoeff])                  \n\t"
361         "sh             %[a1], 448(%[mb_dqcoeff])                  \n\t"
362         "sh             %[a1], 480(%[mb_dqcoeff])                  \n\t"
363 
364         :
365         : [a1] "r" (a1), [mb_dqcoeff] "r" (mb_dqcoeff)
366     );
367 }
368 
369 #endif
370