34 |
const uint32_t rounding); |
const uint32_t rounding); |
35 |
typedef INTERPOLATE8X8 *INTERPOLATE8X8_PTR; |
typedef INTERPOLATE8X8 *INTERPOLATE8X8_PTR; |
36 |
|
|
37 |
|
typedef void (INTERPOLATE8X4) (uint8_t * const dst, |
38 |
|
const uint8_t * const src, |
39 |
|
const uint32_t stride, |
40 |
|
const uint32_t rounding); |
41 |
|
typedef INTERPOLATE8X4 *INTERPOLATE8X4_PTR; |
42 |
|
|
43 |
typedef void (INTERPOLATE8X8_AVG2) (uint8_t *dst, |
typedef void (INTERPOLATE8X8_AVG2) (uint8_t *dst, |
44 |
const uint8_t *src1, |
const uint8_t *src1, |
45 |
const uint8_t *src2, |
const uint8_t *src2, |
84 |
extern INTERPOLATE8X8_PTR interpolate8x8_halfpel_v; |
extern INTERPOLATE8X8_PTR interpolate8x8_halfpel_v; |
85 |
extern INTERPOLATE8X8_PTR interpolate8x8_halfpel_hv; |
extern INTERPOLATE8X8_PTR interpolate8x8_halfpel_hv; |
86 |
|
|
87 |
|
extern INTERPOLATE8X4_PTR interpolate8x4_halfpel_h; |
88 |
|
extern INTERPOLATE8X4_PTR interpolate8x4_halfpel_v; |
89 |
|
extern INTERPOLATE8X4_PTR interpolate8x4_halfpel_hv; |
90 |
|
|
91 |
/* These functions do: dst = (dst+interpolate(src) + 1)/2 |
/* These functions do: dst = (dst+interpolate(src) + 1)/2 |
92 |
* Suitable for direct/interpolated bvop prediction block |
* Suitable for direct/interpolated bvop prediction block |
93 |
* building w/o the need for intermediate interpolated result |
* building w/o the need for intermediate interpolated result |
117 |
INTERPOLATE8X8 interpolate8x8_halfpel_h_c; |
INTERPOLATE8X8 interpolate8x8_halfpel_h_c; |
118 |
INTERPOLATE8X8 interpolate8x8_halfpel_v_c; |
INTERPOLATE8X8 interpolate8x8_halfpel_v_c; |
119 |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_c; |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_c; |
120 |
|
|
121 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_h_c; |
122 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_v_c; |
123 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_hv_c; |
124 |
|
|
125 |
INTERPOLATE8X8 interpolate8x8_halfpel_add_c; |
INTERPOLATE8X8 interpolate8x8_halfpel_add_c; |
126 |
INTERPOLATE8X8 interpolate8x8_halfpel_h_add_c; |
INTERPOLATE8X8 interpolate8x8_halfpel_h_add_c; |
127 |
INTERPOLATE8X8 interpolate8x8_halfpel_v_add_c; |
INTERPOLATE8X8 interpolate8x8_halfpel_v_add_c; |
128 |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_add_c; |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_add_c; |
129 |
|
|
130 |
#ifdef ARCH_IS_IA32 |
#if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) |
131 |
INTERPOLATE8X8 interpolate8x8_halfpel_h_mmx; |
INTERPOLATE8X8 interpolate8x8_halfpel_h_mmx; |
132 |
INTERPOLATE8X8 interpolate8x8_halfpel_v_mmx; |
INTERPOLATE8X8 interpolate8x8_halfpel_v_mmx; |
133 |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_mmx; |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_mmx; |
134 |
|
|
135 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_h_mmx; |
136 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_v_mmx; |
137 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_hv_mmx; |
138 |
|
|
139 |
INTERPOLATE8X8 interpolate8x8_halfpel_add_mmx; |
INTERPOLATE8X8 interpolate8x8_halfpel_add_mmx; |
140 |
INTERPOLATE8X8 interpolate8x8_halfpel_h_add_mmx; |
INTERPOLATE8X8 interpolate8x8_halfpel_h_add_mmx; |
141 |
INTERPOLATE8X8 interpolate8x8_halfpel_v_add_mmx; |
INTERPOLATE8X8 interpolate8x8_halfpel_v_add_mmx; |
145 |
INTERPOLATE8X8 interpolate8x8_halfpel_v_xmm; |
INTERPOLATE8X8 interpolate8x8_halfpel_v_xmm; |
146 |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_xmm; |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_xmm; |
147 |
|
|
148 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_h_xmm; |
149 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_v_xmm; |
150 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_hv_xmm; |
151 |
|
|
152 |
INTERPOLATE8X8 interpolate8x8_halfpel_add_xmm; |
INTERPOLATE8X8 interpolate8x8_halfpel_add_xmm; |
153 |
INTERPOLATE8X8 interpolate8x8_halfpel_h_add_xmm; |
INTERPOLATE8X8 interpolate8x8_halfpel_h_add_xmm; |
154 |
INTERPOLATE8X8 interpolate8x8_halfpel_v_add_xmm; |
INTERPOLATE8X8 interpolate8x8_halfpel_v_add_xmm; |
158 |
INTERPOLATE8X8 interpolate8x8_halfpel_v_3dn; |
INTERPOLATE8X8 interpolate8x8_halfpel_v_3dn; |
159 |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_3dn; |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_3dn; |
160 |
|
|
161 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_h_3dn; |
162 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_v_3dn; |
163 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_hv_3dn; |
164 |
|
|
165 |
INTERPOLATE8X8 interpolate8x8_halfpel_h_3dne; |
INTERPOLATE8X8 interpolate8x8_halfpel_h_3dne; |
166 |
INTERPOLATE8X8 interpolate8x8_halfpel_v_3dne; |
INTERPOLATE8X8 interpolate8x8_halfpel_v_3dne; |
167 |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_3dne; |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_3dne; |
168 |
|
|
169 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_h_3dne; |
170 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_v_3dne; |
171 |
|
INTERPOLATE8X4 interpolate8x4_halfpel_hv_3dne; |
172 |
#endif |
#endif |
173 |
|
|
174 |
#ifdef ARCH_IS_IA64 |
#ifdef ARCH_IS_IA64 |
188 |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_add_altivec_c; |
INTERPOLATE8X8 interpolate8x8_halfpel_hv_add_altivec_c; |
189 |
#endif |
#endif |
190 |
|
|
|
#ifdef ARCH_IS_X86_64 |
|
|
INTERPOLATE8X8 interpolate8x8_halfpel_h_x86_64; |
|
|
INTERPOLATE8X8 interpolate8x8_halfpel_v_x86_64; |
|
|
INTERPOLATE8X8 interpolate8x8_halfpel_hv_x86_64; |
|
|
|
|
|
INTERPOLATE8X8 interpolate8x8_halfpel_add_x86_64; |
|
|
INTERPOLATE8X8 interpolate8x8_halfpel_h_add_x86_64; |
|
|
INTERPOLATE8X8 interpolate8x8_halfpel_v_add_x86_64; |
|
|
INTERPOLATE8X8 interpolate8x8_halfpel_hv_add_x86_64; |
|
|
#endif |
|
|
|
|
191 |
INTERPOLATE8X8_AVG2 interpolate8x8_avg2_c; |
INTERPOLATE8X8_AVG2 interpolate8x8_avg2_c; |
192 |
INTERPOLATE8X8_AVG4 interpolate8x8_avg4_c; |
INTERPOLATE8X8_AVG4 interpolate8x8_avg4_c; |
193 |
|
|
194 |
#ifdef ARCH_IS_IA32 |
#if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) |
195 |
INTERPOLATE8X8_AVG2 interpolate8x8_avg2_mmx; |
INTERPOLATE8X8_AVG2 interpolate8x8_avg2_mmx; |
196 |
INTERPOLATE8X8_AVG4 interpolate8x8_avg4_mmx; |
INTERPOLATE8X8_AVG4 interpolate8x8_avg4_mmx; |
197 |
#endif |
#endif |
201 |
INTERPOLATE8X8_AVG4 interpolate8x8_avg4_altivec_c; |
INTERPOLATE8X8_AVG4 interpolate8x8_avg4_altivec_c; |
202 |
#endif |
#endif |
203 |
|
|
|
#ifdef ARCH_IS_X86_64 |
|
|
INTERPOLATE8X8_AVG2 interpolate8x8_avg2_x86_64; |
|
|
INTERPOLATE8X8_AVG4 interpolate8x8_avg4_x86_64; |
|
|
#endif |
|
|
|
|
204 |
INTERPOLATE_LOWPASS interpolate8x8_lowpass_h_c; |
INTERPOLATE_LOWPASS interpolate8x8_lowpass_h_c; |
205 |
INTERPOLATE_LOWPASS interpolate8x8_lowpass_v_c; |
INTERPOLATE_LOWPASS interpolate8x8_lowpass_v_c; |
206 |
|
|
213 |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_h_c; |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_h_c; |
214 |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_v_c; |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_v_c; |
215 |
|
|
216 |
#ifdef ARCH_IS_IA32 |
#if defined(ARCH_IS_IA32) || defined(ARCH_IS_X86_64) |
217 |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_h_mmx; |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_h_mmx; |
218 |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_v_mmx; |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_v_mmx; |
219 |
#endif |
#endif |
222 |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_h_altivec_c; |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_h_altivec_c; |
223 |
#endif |
#endif |
224 |
|
|
225 |
#ifdef ARCH_IS_X86_64 |
static __inline void |
226 |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_h_x86_64; |
interpolate8x4_switch(uint8_t * const cur, |
227 |
INTERPOLATE8X8_6TAP_LOWPASS interpolate8x8_6tap_lowpass_v_x86_64; |
const uint8_t * const refn, |
228 |
#endif |
const uint32_t x, |
229 |
|
const uint32_t y, |
230 |
|
const int32_t dx, |
231 |
|
const int dy, |
232 |
|
const uint32_t stride, |
233 |
|
const uint32_t rounding) |
234 |
|
{ |
235 |
|
|
236 |
|
const uint8_t * const src = refn + (int)((y + (dy>>1)) * stride + x + (dx>>1)); |
237 |
|
uint8_t * const dst = cur + (int)(y * stride + x); |
238 |
|
|
239 |
|
switch (((dx & 1) << 1) + (dy & 1)) |
240 |
|
{ /* ((dx%2)?2:0)+((dy%2)?1:0) */ |
241 |
|
case 0: |
242 |
|
transfer8x4_copy(dst, src, stride); |
243 |
|
break; |
244 |
|
case 1: |
245 |
|
interpolate8x4_halfpel_v(dst, src, stride, rounding); |
246 |
|
break; |
247 |
|
case 2: |
248 |
|
interpolate8x4_halfpel_h(dst, src, stride, rounding); |
249 |
|
break; |
250 |
|
default: |
251 |
|
interpolate8x4_halfpel_hv(dst, src, stride, rounding); |
252 |
|
break; |
253 |
|
} |
254 |
|
} |
255 |
|
|
256 |
static __inline void |
static __inline void |
257 |
interpolate8x8_switch(uint8_t * const cur, |
interpolate8x8_switch(uint8_t * const cur, |