39 #if COMPILE_TEMPLATE_SSE2
40 #define MMREG_WIDTH "16"
44 "pshuflw $0, "a", "a" \n\t"\
45 "punpcklwd "a", "a" \n\t"
46 #define PMAXW(a,b) "pmaxsw "a", "b" \n\t"
48 "movhlps "a", "b" \n\t"\
50 "pshuflw $0x0E, "a", "b" \n\t"\
52 "pshuflw $0x01, "a", "b" \n\t"\
55 #define MMREG_WIDTH "8"
58 #if COMPILE_TEMPLATE_MMXEXT
59 #define SPREADW(a) "pshufw $0, "a", "a" \n\t"
60 #define PMAXW(a,b) "pmaxsw "a", "b" \n\t"
62 "pshufw $0x0E, "a", "b" \n\t"\
64 "pshufw $0x01, "a", "b" \n\t"\
68 "punpcklwd "a", "a" \n\t"\
69 "punpcklwd "a", "a" \n\t"
71 "psubusw "a", "b" \n\t"\
75 "psrlq $32, "a" \n\t"\
78 "psrlq $16, "a" \n\t"\
84 #if COMPILE_TEMPLATE_SSSE3
85 #define SAVE_SIGN(a,b) \
86 "movdqa "b", "a" \n\t"\
88 #define RESTORE_SIGN(a,b) \
89 "psignw "a", "b" \n\t"
91 #define SAVE_SIGN(a,b) \
93 "pcmpgtw "b", "a" \n\t" \
96 #define RESTORE_SIGN(a,b) \
98 "psubw "a", "b" \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
102 int16_t *
block,
int n,
103 int qscale,
int *overflow)
107 const uint16_t *qmat, *bias;
110 assert((7&(
int)(&temp_block[0])) == 0);
113 RENAME_FDCT(ff_fdct)(
block);
116 s->denoise_dct(s,
block);
128 :
"=d" (
level),
"=a"(dummy)
133 level = (
block[0] + 4)>>3;
137 last_non_zero_p1 = 1;
138 bias = s->q_intra_matrix16[qscale][1];
139 qmat = s->q_intra_matrix16[qscale][0];
141 last_non_zero_p1 = 0;
142 bias = s->q_inter_matrix16[qscale][1];
143 qmat = s->q_inter_matrix16[qscale][0];
146 if((s->out_format ==
FMT_H263 || s->out_format ==
FMT_H261) && s->mpeg_quant==0){
149 "movd %%"REG_a
", "MM"3 \n\t"
151 "pxor "MM"7, "MM"7 \n\t"
152 "pxor "MM"4, "MM"4 \n\t"
154 "pxor "MM"6, "MM"6 \n\t"
155 "psubw (%3), "MM"6 \n\t"
156 "mov $-128, %%"REG_a
" \n\t"
159 MOVQ" (%1, %%"REG_a
"), "MM"0 \n\t"
161 "psubusw "MM"6, "MM"0 \n\t"
162 "pmulhw "MM"5, "MM"0 \n\t"
163 "por "MM"0, "MM"4 \n\t"
165 MOVQ" "MM"0, (%5, %%"REG_a
") \n\t"
166 "pcmpeqw "MM"7, "MM"0 \n\t"
167 MOVQ" (%4, %%"REG_a
"), "MM"1 \n\t"
168 MOVQ" "MM"7, (%1, %%"REG_a
") \n\t"
169 "pandn "MM"1, "MM"0 \n\t"
174 "movd "MM"3, %%"REG_a
" \n\t"
175 "movzb %%al, %%"REG_a
" \n\t"
176 :
"+a" (last_non_zero_p1)
177 :
"r" (
block+64),
"r" (qmat),
"r" (bias),
180 "%xmm4",
"%xmm5",
"%xmm6",
"%xmm7")
184 "movd %%"REG_a
", "MM"3 \n\t"
186 "pxor "MM"7, "MM"7 \n\t"
187 "pxor "MM"4, "MM"4 \n\t"
188 "mov $-128, %%"REG_a
" \n\t"
191 MOVQ" (%1, %%"REG_a
"), "MM"0 \n\t"
193 MOVQ" (%3, %%"REG_a
"), "MM"6 \n\t"
194 "paddusw "MM"6, "MM"0 \n\t"
195 MOVQ" (%2, %%"REG_a
"), "MM"5 \n\t"
196 "pmulhw "MM"5, "MM"0 \n\t"
197 "por "MM"0, "MM"4 \n\t"
199 MOVQ" "MM"0, (%5, %%"REG_a
") \n\t"
200 "pcmpeqw "MM"7, "MM"0 \n\t"
201 MOVQ" (%4, %%"REG_a
"), "MM"1 \n\t"
202 MOVQ" "MM"7, (%1, %%"REG_a
") \n\t"
203 "pandn "MM"1, "MM"0 \n\t"
208 "movd "MM"3, %%"REG_a
" \n\t"
209 "movzb %%al, %%"REG_a
" \n\t"
210 :
"+a" (last_non_zero_p1)
211 :
"r" (
block+64),
"r" (qmat+64),
"r" (bias+64),
214 "%xmm4",
"%xmm5",
"%xmm6",
"%xmm7")
218 "movd %1, "MM"1 \n\t"
220 "psubusw "MM"1, "MM"4 \n\t"
221 "packuswb "MM"4, "MM"4 \n\t"
222 #if COMPILE_TEMPLATE_SSE2
223 "packuswb "MM"4, "MM"4 \n\t"
225 "movd "MM"4, %0 \n\t"
227 :
"g" (s->max_qcoeff)
231 else block[0]= temp_block[0];
234 if(last_non_zero_p1 <= 1)
goto end;
235 block[0x08] = temp_block[0x01];
block[0x10] = temp_block[0x08];
236 block[0x20] = temp_block[0x10];
237 if(last_non_zero_p1 <= 4)
goto end;
238 block[0x18] = temp_block[0x09];
block[0x04] = temp_block[0x02];
239 block[0x09] = temp_block[0x03];
240 if(last_non_zero_p1 <= 7)
goto end;
241 block[0x14] = temp_block[0x0A];
block[0x28] = temp_block[0x11];
242 block[0x12] = temp_block[0x18];
block[0x02] = temp_block[0x20];
243 if(last_non_zero_p1 <= 11)
goto end;
244 block[0x1A] = temp_block[0x19];
block[0x24] = temp_block[0x12];
245 block[0x19] = temp_block[0x0B];
block[0x01] = temp_block[0x04];
246 block[0x0C] = temp_block[0x05];
247 if(last_non_zero_p1 <= 16)
goto end;
248 block[0x11] = temp_block[0x0C];
block[0x29] = temp_block[0x13];
249 block[0x16] = temp_block[0x1A];
block[0x0A] = temp_block[0x21];
250 block[0x30] = temp_block[0x28];
block[0x22] = temp_block[0x30];
251 block[0x38] = temp_block[0x29];
block[0x06] = temp_block[0x22];
252 if(last_non_zero_p1 <= 24)
goto end;
253 block[0x1B] = temp_block[0x1B];
block[0x21] = temp_block[0x14];
254 block[0x1C] = temp_block[0x0D];
block[0x05] = temp_block[0x06];
255 block[0x0D] = temp_block[0x07];
block[0x15] = temp_block[0x0E];
256 block[0x2C] = temp_block[0x15];
block[0x13] = temp_block[0x1C];
257 if(last_non_zero_p1 <= 32)
goto end;
258 block[0x0B] = temp_block[0x23];
block[0x34] = temp_block[0x2A];
259 block[0x2A] = temp_block[0x31];
block[0x32] = temp_block[0x38];
260 block[0x3A] = temp_block[0x39];
block[0x26] = temp_block[0x32];
261 block[0x39] = temp_block[0x2B];
block[0x03] = temp_block[0x24];
262 if(last_non_zero_p1 <= 40)
goto end;
263 block[0x1E] = temp_block[0x1D];
block[0x25] = temp_block[0x16];
264 block[0x1D] = temp_block[0x0F];
block[0x2D] = temp_block[0x17];
265 block[0x17] = temp_block[0x1E];
block[0x0E] = temp_block[0x25];
266 block[0x31] = temp_block[0x2C];
block[0x2B] = temp_block[0x33];
267 if(last_non_zero_p1 <= 48)
goto end;
268 block[0x36] = temp_block[0x3A];
block[0x3B] = temp_block[0x3B];
269 block[0x23] = temp_block[0x34];
block[0x3C] = temp_block[0x2D];
270 block[0x07] = temp_block[0x26];
block[0x1F] = temp_block[0x1F];
271 block[0x0F] = temp_block[0x27];
block[0x35] = temp_block[0x2E];
272 if(last_non_zero_p1 <= 56)
goto end;
273 block[0x2E] = temp_block[0x35];
block[0x33] = temp_block[0x3C];
274 block[0x3E] = temp_block[0x3D];
block[0x27] = temp_block[0x36];
275 block[0x3D] = temp_block[0x2F];
block[0x2F] = temp_block[0x37];
276 block[0x37] = temp_block[0x3E];
block[0x3F] = temp_block[0x3F];
278 if(last_non_zero_p1 <= 1)
goto end;
279 block[0x01] = temp_block[0x01];
280 block[0x08] = temp_block[0x08];
block[0x10] = temp_block[0x10];
281 if(last_non_zero_p1 <= 4)
goto end;
282 block[0x09] = temp_block[0x09];
block[0x02] = temp_block[0x02];
283 block[0x03] = temp_block[0x03];
284 if(last_non_zero_p1 <= 7)
goto end;
285 block[0x0A] = temp_block[0x0A];
block[0x11] = temp_block[0x11];
286 block[0x18] = temp_block[0x18];
block[0x20] = temp_block[0x20];
287 if(last_non_zero_p1 <= 11)
goto end;
288 block[0x19] = temp_block[0x19];
289 block[0x12] = temp_block[0x12];
block[0x0B] = temp_block[0x0B];
290 block[0x04] = temp_block[0x04];
block[0x05] = temp_block[0x05];
291 if(last_non_zero_p1 <= 16)
goto end;
292 block[0x0C] = temp_block[0x0C];
block[0x13] = temp_block[0x13];
293 block[0x1A] = temp_block[0x1A];
block[0x21] = temp_block[0x21];
294 block[0x28] = temp_block[0x28];
block[0x30] = temp_block[0x30];
295 block[0x29] = temp_block[0x29];
block[0x22] = temp_block[0x22];
296 if(last_non_zero_p1 <= 24)
goto end;
297 block[0x1B] = temp_block[0x1B];
block[0x14] = temp_block[0x14];
298 block[0x0D] = temp_block[0x0D];
block[0x06] = temp_block[0x06];
299 block[0x07] = temp_block[0x07];
block[0x0E] = temp_block[0x0E];
300 block[0x15] = temp_block[0x15];
block[0x1C] = temp_block[0x1C];
301 if(last_non_zero_p1 <= 32)
goto end;
302 block[0x23] = temp_block[0x23];
block[0x2A] = temp_block[0x2A];
303 block[0x31] = temp_block[0x31];
block[0x38] = temp_block[0x38];
304 block[0x39] = temp_block[0x39];
block[0x32] = temp_block[0x32];
305 block[0x2B] = temp_block[0x2B];
block[0x24] = temp_block[0x24];
306 if(last_non_zero_p1 <= 40)
goto end;
307 block[0x1D] = temp_block[0x1D];
block[0x16] = temp_block[0x16];
308 block[0x0F] = temp_block[0x0F];
block[0x17] = temp_block[0x17];
309 block[0x1E] = temp_block[0x1E];
block[0x25] = temp_block[0x25];
310 block[0x2C] = temp_block[0x2C];
block[0x33] = temp_block[0x33];
311 if(last_non_zero_p1 <= 48)
goto end;
312 block[0x3A] = temp_block[0x3A];
block[0x3B] = temp_block[0x3B];
313 block[0x34] = temp_block[0x34];
block[0x2D] = temp_block[0x2D];
314 block[0x26] = temp_block[0x26];
block[0x1F] = temp_block[0x1F];
315 block[0x27] = temp_block[0x27];
block[0x2E] = temp_block[0x2E];
316 if(last_non_zero_p1 <= 56)
goto end;
317 block[0x35] = temp_block[0x35];
block[0x3C] = temp_block[0x3C];
318 block[0x3D] = temp_block[0x3D];
block[0x36] = temp_block[0x36];
319 block[0x2F] = temp_block[0x2F];
block[0x37] = temp_block[0x37];
320 block[0x3E] = temp_block[0x3E];
block[0x3F] = temp_block[0x3F];
323 return last_non_zero_p1 - 1;
static int RENAME() dct_quantize(MpegEncContext *s, int16_t *block, int n, int qscale, int *overflow)
static uint16_t inv_zigzag_direct16[64]
const uint32_t ff_inverse[257]
#define XMM_CLOBBERS_ONLY(...)
#define RESTORE_SIGN(a, b)
common internal API header
#define LOCAL_ALIGNED_16(t, v,...)