vp8dsp_altivec.c
Go to the documentation of this file.
1 
23 #include "libavutil/cpu.h"
24 #include "libavcodec/vp8dsp.h"
25 #include "dsputil_altivec.h"
26 #include "types_altivec.h"
27 #include "util_altivec.h"
28 
29 #define REPT4(...) { __VA_ARGS__, __VA_ARGS__, __VA_ARGS__, __VA_ARGS__ }
30 
31 // h subpel filter uses msum to multiply+add 4 pixel taps at once
32 static const vec_s8 h_subpel_filters_inner[7] =
33 {
34  REPT4( -6, 123, 12, -1),
35  REPT4(-11, 108, 36, -8),
36  REPT4( -9, 93, 50, -6),
37  REPT4(-16, 77, 77, -16),
38  REPT4( -6, 50, 93, -9),
39  REPT4( -8, 36, 108, -11),
40  REPT4( -1, 12, 123, -6),
41 };
42 
43 // for 6tap filters, these are the outer two taps
44 // The zeros mask off pixels 4-7 when filtering 0-3
45 // and vice-versa
46 static const vec_s8 h_subpel_filters_outer[3] =
47 {
48  REPT4(0, 0, 2, 1),
49  REPT4(0, 0, 3, 3),
50  REPT4(0, 0, 1, 2),
51 };
52 
53 #define LOAD_H_SUBPEL_FILTER(i) \
54  vec_s8 filter_inner = h_subpel_filters_inner[i]; \
55  vec_s8 filter_outerh = h_subpel_filters_outer[(i)>>1]; \
56  vec_s8 filter_outerl = vec_sld(filter_outerh, filter_outerh, 2)
57 
58 #define FILTER_H(dstv, off) \
59  a = vec_ld((off)-is6tap-1, src); \
60  b = vec_ld((off)-is6tap-1+15, src); \
61 \
62  pixh = vec_perm(a, b, permh##off); \
63  pixl = vec_perm(a, b, perml##off); \
64  filth = vec_msum(filter_inner, pixh, c64); \
65  filtl = vec_msum(filter_inner, pixl, c64); \
66 \
67  if (is6tap) { \
68  outer = vec_perm(a, b, perm_6tap##off); \
69  filth = vec_msum(filter_outerh, outer, filth); \
70  filtl = vec_msum(filter_outerl, outer, filtl); \
71  } \
72  if (w == 4) \
73  filtl = filth; /* discard pixels 4-7 */ \
74  dstv = vec_packs(filth, filtl); \
75  dstv = vec_sra(dstv, c7)
76 
77 static av_always_inline
78 void put_vp8_epel_h_altivec_core(uint8_t *dst, int dst_stride,
79  uint8_t *src, int src_stride,
80  int h, int mx, int w, int is6tap)
81 {
83  vec_u8 align_vec0, align_vec8, permh0, permh8, filt;
84  vec_u8 perm_6tap0, perm_6tap8, perml0, perml8;
85  vec_u8 a, b, pixh, pixl, outer;
86  vec_s16 f16h, f16l;
87  vec_s32 filth, filtl;
88 
89  vec_u8 perm_inner6 = { 1,2,3,4, 2,3,4,5, 3,4,5,6, 4,5,6,7 };
90  vec_u8 perm_inner4 = { 0,1,2,3, 1,2,3,4, 2,3,4,5, 3,4,5,6 };
91  vec_u8 perm_inner = is6tap ? perm_inner6 : perm_inner4;
92  vec_u8 perm_outer = { 4,9, 0,5, 5,10, 1,6, 6,11, 2,7, 7,12, 3,8 };
93  vec_s32 c64 = vec_sl(vec_splat_s32(1), vec_splat_u32(6));
94  vec_u16 c7 = vec_splat_u16(7);
95 
96  align_vec0 = vec_lvsl( -is6tap-1, src);
97  align_vec8 = vec_lvsl(8-is6tap-1, src);
98 
99  permh0 = vec_perm(align_vec0, align_vec0, perm_inner);
100  permh8 = vec_perm(align_vec8, align_vec8, perm_inner);
101  perm_inner = vec_add(perm_inner, vec_splat_u8(4));
102  perml0 = vec_perm(align_vec0, align_vec0, perm_inner);
103  perml8 = vec_perm(align_vec8, align_vec8, perm_inner);
104  perm_6tap0 = vec_perm(align_vec0, align_vec0, perm_outer);
105  perm_6tap8 = vec_perm(align_vec8, align_vec8, perm_outer);
106 
107  while (h --> 0) {
108  FILTER_H(f16h, 0);
109 
110  if (w == 16) {
111  FILTER_H(f16l, 8);
112  filt = vec_packsu(f16h, f16l);
113  vec_st(filt, 0, dst);
114  } else {
115  filt = vec_packsu(f16h, f16h);
116  vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
117  if (w == 8)
118  vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
119  }
120  src += src_stride;
121  dst += dst_stride;
122  }
123 }
124 
125 // v subpel filter does a simple vertical multiply + add
126 static const vec_u8 v_subpel_filters[7] =
127 {
128  { 0, 6, 123, 12, 1, 0 },
129  { 2, 11, 108, 36, 8, 1 },
130  { 0, 9, 93, 50, 6, 0 },
131  { 3, 16, 77, 77, 16, 3 },
132  { 0, 6, 50, 93, 9, 0 },
133  { 1, 8, 36, 108, 11, 2 },
134  { 0, 1, 12, 123, 6, 0 },
135 };
136 
137 #define LOAD_V_SUBPEL_FILTER(i) \
138  vec_u8 subpel_filter = v_subpel_filters[i]; \
139  vec_u8 f0 = vec_splat(subpel_filter, 0); \
140  vec_u8 f1 = vec_splat(subpel_filter, 1); \
141  vec_u8 f2 = vec_splat(subpel_filter, 2); \
142  vec_u8 f3 = vec_splat(subpel_filter, 3); \
143  vec_u8 f4 = vec_splat(subpel_filter, 4); \
144  vec_u8 f5 = vec_splat(subpel_filter, 5)
145 
146 #define FILTER_V(dstv, vec_mul) \
147  s1f = (vec_s16)vec_mul(s1, f1); \
148  s2f = (vec_s16)vec_mul(s2, f2); \
149  s3f = (vec_s16)vec_mul(s3, f3); \
150  s4f = (vec_s16)vec_mul(s4, f4); \
151  s2f = vec_subs(s2f, s1f); \
152  s3f = vec_subs(s3f, s4f); \
153  if (is6tap) { \
154  s0f = (vec_s16)vec_mul(s0, f0); \
155  s5f = (vec_s16)vec_mul(s5, f5); \
156  s2f = vec_adds(s2f, s0f); \
157  s3f = vec_adds(s3f, s5f); \
158  } \
159  dstv = vec_adds(s2f, s3f); \
160  dstv = vec_adds(dstv, c64); \
161  dstv = vec_sra(dstv, c7)
162 
163 static av_always_inline
164 void put_vp8_epel_v_altivec_core(uint8_t *dst, int dst_stride,
165  uint8_t *src, int src_stride,
166  int h, int my, int w, int is6tap)
167 {
168  LOAD_V_SUBPEL_FILTER(my-1);
169  vec_u8 s0, s1, s2, s3, s4, s5, filt, align_vech, perm_vec, align_vecl;
170  vec_s16 s0f, s1f, s2f, s3f, s4f, s5f, f16h, f16l;
171  vec_s16 c64 = vec_sl(vec_splat_s16(1), vec_splat_u16(6));
172  vec_u16 c7 = vec_splat_u16(7);
173 
174  // we want pixels 0-7 to be in the even positions and 8-15 in the odd,
175  // so combine this permute with the alignment permute vector
176  align_vech = vec_lvsl(0, src);
177  align_vecl = vec_sld(align_vech, align_vech, 8);
178  if (w ==16)
179  perm_vec = vec_mergeh(align_vech, align_vecl);
180  else
181  perm_vec = vec_mergeh(align_vech, align_vech);
182 
183  if (is6tap)
184  s0 = load_with_perm_vec(-2*src_stride, src, perm_vec);
185  s1 = load_with_perm_vec(-1*src_stride, src, perm_vec);
186  s2 = load_with_perm_vec( 0*src_stride, src, perm_vec);
187  s3 = load_with_perm_vec( 1*src_stride, src, perm_vec);
188  if (is6tap)
189  s4 = load_with_perm_vec( 2*src_stride, src, perm_vec);
190 
191  src += (2+is6tap)*src_stride;
192 
193  while (h --> 0) {
194  if (is6tap)
195  s5 = load_with_perm_vec(0, src, perm_vec);
196  else
197  s4 = load_with_perm_vec(0, src, perm_vec);
198 
199  FILTER_V(f16h, vec_mule);
200 
201  if (w == 16) {
202  FILTER_V(f16l, vec_mulo);
203  filt = vec_packsu(f16h, f16l);
204  vec_st(filt, 0, dst);
205  } else {
206  filt = vec_packsu(f16h, f16h);
207  if (w == 4)
208  filt = (vec_u8)vec_splat((vec_u32)filt, 0);
209  else
210  vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
211  vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
212  }
213 
214  if (is6tap)
215  s0 = s1;
216  s1 = s2;
217  s2 = s3;
218  s3 = s4;
219  if (is6tap)
220  s4 = s5;
221 
222  dst += dst_stride;
223  src += src_stride;
224  }
225 }
226 
227 #define EPEL_FUNCS(WIDTH, TAPS) \
228 static av_noinline \
229 void put_vp8_epel ## WIDTH ## _h ## TAPS ## _altivec(uint8_t *dst, int dst_stride, uint8_t *src, int src_stride, int h, int mx, int my) \
230 { \
231  put_vp8_epel_h_altivec_core(dst, dst_stride, src, src_stride, h, mx, WIDTH, TAPS == 6); \
232 } \
233 \
234 static av_noinline \
235 void put_vp8_epel ## WIDTH ## _v ## TAPS ## _altivec(uint8_t *dst, int dst_stride, uint8_t *src, int src_stride, int h, int mx, int my) \
236 { \
237  put_vp8_epel_v_altivec_core(dst, dst_stride, src, src_stride, h, my, WIDTH, TAPS == 6); \
238 }
239 
240 #define EPEL_HV(WIDTH, HTAPS, VTAPS) \
241 static void put_vp8_epel ## WIDTH ## _h ## HTAPS ## v ## VTAPS ## _altivec(uint8_t *dst, int stride, uint8_t *src, int s, int h, int mx, int my) \
242 { \
243  DECLARE_ALIGNED(16, uint8_t, tmp)[(2*WIDTH+5)*16]; \
244  if (VTAPS == 6) { \
245  put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16, src-2*stride, stride, h+5, mx, my); \
246  put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, stride, tmp+2*16, 16, h, mx, my); \
247  } else { \
248  put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16, src-stride, stride, h+4, mx, my); \
249  put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, stride, tmp+16, 16, h, mx, my); \
250  } \
251 }
252 
253 EPEL_FUNCS(16,6)
254 EPEL_FUNCS(8, 6)
255 EPEL_FUNCS(8, 4)
256 EPEL_FUNCS(4, 6)
257 EPEL_FUNCS(4, 4)
258 
259 EPEL_HV(16, 6,6)
260 EPEL_HV(8, 6,6)
261 EPEL_HV(8, 4,6)
262 EPEL_HV(8, 6,4)
263 EPEL_HV(8, 4,4)
264 EPEL_HV(4, 6,6)
265 EPEL_HV(4, 4,6)
266 EPEL_HV(4, 6,4)
267 EPEL_HV(4, 4,4)
268 
269 static void put_vp8_pixels16_altivec(uint8_t *dst, int stride, uint8_t *src, int s, int h, int mx, int my)
270 {
271  put_pixels16_altivec(dst, src, stride, h);
272 }
273 
275 {
277  return;
278 
280  c->put_vp8_epel_pixels_tab[0][0][2] = put_vp8_epel16_h6_altivec;
281  c->put_vp8_epel_pixels_tab[0][2][0] = put_vp8_epel16_v6_altivec;
282  c->put_vp8_epel_pixels_tab[0][2][2] = put_vp8_epel16_h6v6_altivec;
283 
284  c->put_vp8_epel_pixels_tab[1][0][2] = put_vp8_epel8_h6_altivec;
285  c->put_vp8_epel_pixels_tab[1][2][0] = put_vp8_epel8_v6_altivec;
286  c->put_vp8_epel_pixels_tab[1][0][1] = put_vp8_epel8_h4_altivec;
287  c->put_vp8_epel_pixels_tab[1][1][0] = put_vp8_epel8_v4_altivec;
288 
289  c->put_vp8_epel_pixels_tab[1][2][2] = put_vp8_epel8_h6v6_altivec;
290  c->put_vp8_epel_pixels_tab[1][1][1] = put_vp8_epel8_h4v4_altivec;
291  c->put_vp8_epel_pixels_tab[1][1][2] = put_vp8_epel8_h6v4_altivec;
292  c->put_vp8_epel_pixels_tab[1][2][1] = put_vp8_epel8_h4v6_altivec;
293 
294  c->put_vp8_epel_pixels_tab[2][0][2] = put_vp8_epel4_h6_altivec;
295  c->put_vp8_epel_pixels_tab[2][2][0] = put_vp8_epel4_v6_altivec;
296  c->put_vp8_epel_pixels_tab[2][0][1] = put_vp8_epel4_h4_altivec;
297  c->put_vp8_epel_pixels_tab[2][1][0] = put_vp8_epel4_v4_altivec;
298 
299  c->put_vp8_epel_pixels_tab[2][2][2] = put_vp8_epel4_h6v6_altivec;
300  c->put_vp8_epel_pixels_tab[2][1][1] = put_vp8_epel4_h4v4_altivec;
301  c->put_vp8_epel_pixels_tab[2][1][2] = put_vp8_epel4_h6v4_altivec;
302  c->put_vp8_epel_pixels_tab[2][2][1] = put_vp8_epel4_h4v6_altivec;
303 }