2 * Copyright (C) 2001-2003 Michael Niedermayer <michaelni@gmx.at>
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 #include "../swscale_internal.h"
22 #include "libavutil/x86/asm.h"
23 #include "libavutil/x86/cpu.h"
25 #define RET 0xC3 // near return opcode for x86
26 #define PREFETCH "prefetchnta"
28 av_cold int ff_init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
29 int16_t *filter, int32_t *filterPos,
33 x86_reg imm8OfPShufW1A;
34 x86_reg imm8OfPShufW2A;
35 x86_reg fragmentLengthA;
37 x86_reg imm8OfPShufW1B;
38 x86_reg imm8OfPShufW2B;
39 x86_reg fragmentLengthB;
44 // create an optimized horizontal scaling routine
45 /* This scaler is made of runtime-generated MMXEXT code using specially tuned
46 * pshufw instructions. For every four output pixels, if four input pixels
47 * are enough for the fast bilinear scaling, then a chunk of fragmentB is
48 * used. If five input pixels are needed, then a chunk of fragmentA is used.
57 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
58 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
59 "movd 1(%%"REG_c", %%"REG_S"), %%mm1 \n\t"
60 "punpcklbw %%mm7, %%mm1 \n\t"
61 "punpcklbw %%mm7, %%mm0 \n\t"
62 "pshufw $0xFF, %%mm1, %%mm1 \n\t"
64 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
66 "psubw %%mm1, %%mm0 \n\t"
67 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
68 "pmullw %%mm3, %%mm0 \n\t"
69 "psllw $7, %%mm1 \n\t"
70 "paddw %%mm1, %%mm0 \n\t"
72 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
74 "add $8, %%"REG_a" \n\t"
78 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
79 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
80 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
85 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
89 : "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
90 "=r" (fragmentLengthA)
97 "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
98 "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
99 "punpcklbw %%mm7, %%mm0 \n\t"
100 "pshufw $0xFF, %%mm0, %%mm1 \n\t"
102 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
104 "psubw %%mm1, %%mm0 \n\t"
105 "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
106 "pmullw %%mm3, %%mm0 \n\t"
107 "psllw $7, %%mm1 \n\t"
108 "paddw %%mm1, %%mm0 \n\t"
110 "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
112 "add $8, %%"REG_a" \n\t"
116 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
117 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
118 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
123 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
127 : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
128 "=r" (fragmentLengthB)
131 xpos = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
134 for (i = 0; i < dstW / numSplits; i++) {
139 int b = ((xpos + xInc) >> 16) - xx;
140 int c = ((xpos + xInc * 2) >> 16) - xx;
141 int d = ((xpos + xInc * 3) >> 16) - xx;
142 int inc = (d + 1 < 4);
143 uint8_t *fragment = inc ? fragmentB : fragmentA;
144 x86_reg imm8OfPShufW1 = inc ? imm8OfPShufW1B : imm8OfPShufW1A;
145 x86_reg imm8OfPShufW2 = inc ? imm8OfPShufW2B : imm8OfPShufW2A;
146 x86_reg fragmentLength = inc ? fragmentLengthB : fragmentLengthA;
147 int maxShift = 3 - (d + inc);
151 filter[i] = ((xpos & 0xFFFF) ^ 0xFFFF) >> 9;
152 filter[i + 1] = (((xpos + xInc) & 0xFFFF) ^ 0xFFFF) >> 9;
153 filter[i + 2] = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
154 filter[i + 3] = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
155 filterPos[i / 2] = xx;
157 memcpy(filterCode + fragmentPos, fragment, fragmentLength);
159 filterCode[fragmentPos + imm8OfPShufW1] = (a + inc) |
163 filterCode[fragmentPos + imm8OfPShufW2] = a | (b << 2) |
167 if (i + 4 - inc >= dstW)
168 shift = maxShift; // avoid overread
169 else if ((filterPos[i / 2] & 3) <= maxShift)
170 shift = filterPos[i / 2] & 3; // align
172 if (shift && i >= shift) {
173 filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
174 filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
175 filterPos[i / 2] -= shift;
179 fragmentPos += fragmentLength;
182 filterCode[fragmentPos] = RET;
187 filterPos[((i / 2) + 1) & (~1)] = xpos >> 16; // needed to jump to the next part
189 return fragmentPos + 1;
192 void ff_hyscale_fast_mmxext(SwsContext *c, int16_t *dst,
193 int dstWidth, const uint8_t *src,
196 int32_t *filterPos = c->hLumFilterPos;
197 int16_t *filter = c->hLumFilter;
198 void *mmxextFilterCode = c->lumMmxextFilterCode;
209 "mov %%"REG_b", %5 \n\t"
211 "mov -8(%%rsp), %%"REG_a" \n\t"
212 "mov %%"REG_a", %6 \n\t"
216 "mov -8(%%rsp), %%"REG_a" \n\t"
217 "mov %%"REG_a", %5 \n\t"
220 "pxor %%mm7, %%mm7 \n\t"
221 "mov %0, %%"REG_c" \n\t"
222 "mov %1, %%"REG_D" \n\t"
223 "mov %2, %%"REG_d" \n\t"
224 "mov %3, %%"REG_b" \n\t"
225 "xor %%"REG_a", %%"REG_a" \n\t" // i
226 PREFETCH" (%%"REG_c") \n\t"
227 PREFETCH" 32(%%"REG_c") \n\t"
228 PREFETCH" 64(%%"REG_c") \n\t"
231 #define CALL_MMXEXT_FILTER_CODE \
232 "movl (%%"REG_b"), %%esi \n\t"\
234 "movl (%%"REG_b", %%"REG_a"), %%esi \n\t"\
235 "add %%"REG_S", %%"REG_c" \n\t"\
236 "add %%"REG_a", %%"REG_D" \n\t"\
237 "xor %%"REG_a", %%"REG_a" \n\t"\
240 #define CALL_MMXEXT_FILTER_CODE \
241 "movl (%%"REG_b"), %%esi \n\t"\
243 "addl (%%"REG_b", %%"REG_a"), %%"REG_c" \n\t"\
244 "add %%"REG_a", %%"REG_D" \n\t"\
245 "xor %%"REG_a", %%"REG_a" \n\t"\
247 #endif /* ARCH_X86_64 */
249 CALL_MMXEXT_FILTER_CODE
250 CALL_MMXEXT_FILTER_CODE
251 CALL_MMXEXT_FILTER_CODE
252 CALL_MMXEXT_FILTER_CODE
253 CALL_MMXEXT_FILTER_CODE
254 CALL_MMXEXT_FILTER_CODE
255 CALL_MMXEXT_FILTER_CODE
256 CALL_MMXEXT_FILTER_CODE
259 "mov %5, %%"REG_b" \n\t"
261 "mov %6, %%"REG_a" \n\t"
262 "mov %%"REG_a", -8(%%rsp) \n\t"
266 "mov %5, %%"REG_a" \n\t"
267 "mov %%"REG_a", -8(%%rsp) \n\t"
270 :: "m" (src), "m" (dst), "m" (filter), "m" (filterPos),
271 "m" (mmxextFilterCode)
278 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
284 for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--)
285 dst[i] = src[srcW-1]*128;
288 void ff_hcscale_fast_mmxext(SwsContext *c, int16_t *dst1, int16_t *dst2,
289 int dstWidth, const uint8_t *src1,
290 const uint8_t *src2, int srcW, int xInc)
292 int32_t *filterPos = c->hChrFilterPos;
293 int16_t *filter = c->hChrFilter;
294 void *mmxextFilterCode = c->chrMmxextFilterCode;
297 DECLARE_ALIGNED(8, uint64_t, ebxsave);
300 DECLARE_ALIGNED(8, uint64_t, retsave);
305 "mov %%"REG_b", %7 \n\t"
307 "mov -8(%%rsp), %%"REG_a" \n\t"
308 "mov %%"REG_a", %8 \n\t"
312 "mov -8(%%rsp), %%"REG_a" \n\t"
313 "mov %%"REG_a", %7 \n\t"
316 "pxor %%mm7, %%mm7 \n\t"
317 "mov %0, %%"REG_c" \n\t"
318 "mov %1, %%"REG_D" \n\t"
319 "mov %2, %%"REG_d" \n\t"
320 "mov %3, %%"REG_b" \n\t"
321 "xor %%"REG_a", %%"REG_a" \n\t" // i
322 PREFETCH" (%%"REG_c") \n\t"
323 PREFETCH" 32(%%"REG_c") \n\t"
324 PREFETCH" 64(%%"REG_c") \n\t"
326 CALL_MMXEXT_FILTER_CODE
327 CALL_MMXEXT_FILTER_CODE
328 CALL_MMXEXT_FILTER_CODE
329 CALL_MMXEXT_FILTER_CODE
330 "xor %%"REG_a", %%"REG_a" \n\t" // i
331 "mov %5, %%"REG_c" \n\t" // src
332 "mov %6, %%"REG_D" \n\t" // buf2
333 PREFETCH" (%%"REG_c") \n\t"
334 PREFETCH" 32(%%"REG_c") \n\t"
335 PREFETCH" 64(%%"REG_c") \n\t"
337 CALL_MMXEXT_FILTER_CODE
338 CALL_MMXEXT_FILTER_CODE
339 CALL_MMXEXT_FILTER_CODE
340 CALL_MMXEXT_FILTER_CODE
343 "mov %7, %%"REG_b" \n\t"
345 "mov %8, %%"REG_a" \n\t"
346 "mov %%"REG_a", -8(%%rsp) \n\t"
350 "mov %7, %%"REG_a" \n\t"
351 "mov %%"REG_a", -8(%%rsp) \n\t"
354 :: "m" (src1), "m" (dst1), "m" (filter), "m" (filterPos),
355 "m" (mmxextFilterCode), "m" (src2), "m"(dst2)
362 : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
368 for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--) {
369 dst1[i] = src1[srcW-1]*128;
370 dst2[i] = src2[srcW-1]*128;