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"
29 av_cold int ff_init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
30 int16_t *filter, int32_t *filterPos,
34 x86_reg imm8OfPShufW1A;
35 x86_reg imm8OfPShufW2A;
36 x86_reg fragmentLengthA;
38 x86_reg imm8OfPShufW1B;
39 x86_reg imm8OfPShufW2B;
40 x86_reg fragmentLengthB;
45 // create an optimized horizontal scaling routine
46 /* This scaler is made of runtime-generated MMXEXT code using specially tuned
47 * pshufw instructions. For every four output pixels, if four input pixels
48 * are enough for the fast bilinear scaling, then a chunk of fragmentB is
49 * used. If five input pixels are needed, then a chunk of fragmentA is used.
58 "movq (%%"FF_REG_d", %%"FF_REG_a"), %%mm3 \n\t"
59 "movd (%%"FF_REG_c", %%"FF_REG_S"), %%mm0 \n\t"
60 "movd 1(%%"FF_REG_c", %%"FF_REG_S"), %%mm1 \n\t"
61 "punpcklbw %%mm7, %%mm1 \n\t"
62 "punpcklbw %%mm7, %%mm0 \n\t"
63 "pshufw $0xFF, %%mm1, %%mm1 \n\t"
65 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
67 "psubw %%mm1, %%mm0 \n\t"
68 "movl 8(%%"FF_REG_b", %%"FF_REG_a"), %%esi \n\t"
69 "pmullw %%mm3, %%mm0 \n\t"
70 "psllw $7, %%mm1 \n\t"
71 "paddw %%mm1, %%mm0 \n\t"
73 "movq %%mm0, (%%"FF_REG_D", %%"FF_REG_a") \n\t"
75 "add $8, %%"FF_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 (%%"FF_REG_d", %%"FF_REG_a"), %%mm3 \n\t"
98 "movd (%%"FF_REG_c", %%"FF_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(%%"FF_REG_b", %%"FF_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, (%%"FF_REG_D", %%"FF_REG_a") \n\t"
112 "add $8, %%"FF_REG_a" \n\t"
115 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
116 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
117 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
122 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
126 : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
127 "=r" (fragmentLengthB)
130 xpos = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
133 for (i = 0; i < dstW / numSplits; i++) {
138 int b = ((xpos + xInc) >> 16) - xx;
139 int c = ((xpos + xInc * 2) >> 16) - xx;
140 int d = ((xpos + xInc * 3) >> 16) - xx;
141 int inc = (d + 1 < 4);
142 uint8_t *fragment = inc ? fragmentB : fragmentA;
143 x86_reg imm8OfPShufW1 = inc ? imm8OfPShufW1B : imm8OfPShufW1A;
144 x86_reg imm8OfPShufW2 = inc ? imm8OfPShufW2B : imm8OfPShufW2A;
145 x86_reg fragmentLength = inc ? fragmentLengthB : fragmentLengthA;
146 int maxShift = 3 - (d + inc);
150 filter[i] = ((xpos & 0xFFFF) ^ 0xFFFF) >> 9;
151 filter[i + 1] = (((xpos + xInc) & 0xFFFF) ^ 0xFFFF) >> 9;
152 filter[i + 2] = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
153 filter[i + 3] = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
154 filterPos[i / 2] = xx;
156 memcpy(filterCode + fragmentPos, fragment, fragmentLength);
158 filterCode[fragmentPos + imm8OfPShufW1] = (a + inc) |
162 filterCode[fragmentPos + imm8OfPShufW2] = a | (b << 2) |
166 if (i + 4 - inc >= dstW)
167 shift = maxShift; // avoid overread
168 else if ((filterPos[i / 2] & 3) <= maxShift)
169 shift = filterPos[i / 2] & 3; // align
171 if (shift && i >= shift) {
172 filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
173 filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
174 filterPos[i / 2] -= shift;
178 fragmentPos += fragmentLength;
181 filterCode[fragmentPos] = RET;
186 filterPos[((i / 2) + 1) & (~1)] = xpos >> 16; // needed to jump to the next part
188 return fragmentPos + 1;
191 void ff_hyscale_fast_mmxext(SwsContext *c, int16_t *dst,
192 int dstWidth, const uint8_t *src,
195 int32_t *filterPos = c->hLumFilterPos;
196 int16_t *filter = c->hLumFilter;
197 void *mmxextFilterCode = c->lumMmxextFilterCode;
202 #if !HAVE_EBX_AVAILABLE
209 "mov -8(%%rsp), %%"FF_REG_a" \n\t"
210 "mov %%"FF_REG_a", %5 \n\t" // retsave
212 #if !HAVE_EBX_AVAILABLE
213 "mov %%"FF_REG_b", %5 \n\t" // ebxsave
216 "pxor %%mm7, %%mm7 \n\t"
217 "mov %0, %%"FF_REG_c" \n\t"
218 "mov %1, %%"FF_REG_D" \n\t"
219 "mov %2, %%"FF_REG_d" \n\t"
220 "mov %3, %%"FF_REG_b" \n\t"
221 "xor %%"FF_REG_a", %%"FF_REG_a" \n\t" // i
222 PREFETCH" (%%"FF_REG_c") \n\t"
223 PREFETCH" 32(%%"FF_REG_c") \n\t"
224 PREFETCH" 64(%%"FF_REG_c") \n\t"
227 #define CALL_MMXEXT_FILTER_CODE \
228 "movl (%%"FF_REG_b"), %%esi \n\t"\
230 "movl (%%"FF_REG_b", %%"FF_REG_a"), %%esi \n\t"\
231 "add %%"FF_REG_S", %%"FF_REG_c" \n\t"\
232 "add %%"FF_REG_a", %%"FF_REG_D" \n\t"\
233 "xor %%"FF_REG_a", %%"FF_REG_a" \n\t"\
236 #define CALL_MMXEXT_FILTER_CODE \
237 "movl (%%"FF_REG_b"), %%esi \n\t"\
239 "addl (%%"FF_REG_b", %%"FF_REG_a"), %%"FF_REG_c" \n\t"\
240 "add %%"FF_REG_a", %%"FF_REG_D" \n\t"\
241 "xor %%"FF_REG_a", %%"FF_REG_a" \n\t"\
243 #endif /* ARCH_X86_64 */
245 CALL_MMXEXT_FILTER_CODE
246 CALL_MMXEXT_FILTER_CODE
247 CALL_MMXEXT_FILTER_CODE
248 CALL_MMXEXT_FILTER_CODE
249 CALL_MMXEXT_FILTER_CODE
250 CALL_MMXEXT_FILTER_CODE
251 CALL_MMXEXT_FILTER_CODE
252 CALL_MMXEXT_FILTER_CODE
255 "mov %5, %%"FF_REG_a" \n\t"
256 "mov %%"FF_REG_a", -8(%%rsp) \n\t"
258 #if !HAVE_EBX_AVAILABLE
259 "mov %5, %%"FF_REG_b" \n\t"
262 :: "m" (src), "m" (dst), "m" (filter), "m" (filterPos),
263 "m" (mmxextFilterCode)
267 #if !HAVE_EBX_AVAILABLE
271 : "%"FF_REG_a, "%"FF_REG_c, "%"FF_REG_d, "%"FF_REG_S, "%"FF_REG_D
272 #if ARCH_X86_64 || HAVE_EBX_AVAILABLE
277 for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--)
278 dst[i] = src[srcW-1]*128;
281 void ff_hcscale_fast_mmxext(SwsContext *c, int16_t *dst1, int16_t *dst2,
282 int dstWidth, const uint8_t *src1,
283 const uint8_t *src2, int srcW, int xInc)
285 int32_t *filterPos = c->hChrFilterPos;
286 int16_t *filter = c->hChrFilter;
287 void *mmxextFilterCode = c->chrMmxextFilterCode;
290 DECLARE_ALIGNED(8, uint64_t, retsave);
292 #if !HAVE_EBX_AVAILABLE
293 DECLARE_ALIGNED(8, uint64_t, ebxsave);
298 "mov -8(%%rsp), %%"FF_REG_a" \n\t"
299 "mov %%"FF_REG_a", %7 \n\t" // retsave
301 #if !HAVE_EBX_AVAILABLE
302 "mov %%"FF_REG_b", %7 \n\t" // ebxsave
305 "pxor %%mm7, %%mm7 \n\t"
306 "mov %0, %%"FF_REG_c" \n\t"
307 "mov %1, %%"FF_REG_D" \n\t"
308 "mov %2, %%"FF_REG_d" \n\t"
309 "mov %3, %%"FF_REG_b" \n\t"
310 "xor %%"FF_REG_a", %%"FF_REG_a" \n\t" // i
311 PREFETCH" (%%"FF_REG_c") \n\t"
312 PREFETCH" 32(%%"FF_REG_c") \n\t"
313 PREFETCH" 64(%%"FF_REG_c") \n\t"
315 CALL_MMXEXT_FILTER_CODE
316 CALL_MMXEXT_FILTER_CODE
317 CALL_MMXEXT_FILTER_CODE
318 CALL_MMXEXT_FILTER_CODE
319 "xor %%"FF_REG_a", %%"FF_REG_a" \n\t" // i
320 "mov %5, %%"FF_REG_c" \n\t" // src2
321 "mov %6, %%"FF_REG_D" \n\t" // dst2
322 PREFETCH" (%%"FF_REG_c") \n\t"
323 PREFETCH" 32(%%"FF_REG_c") \n\t"
324 PREFETCH" 64(%%"FF_REG_c") \n\t"
326 CALL_MMXEXT_FILTER_CODE
327 CALL_MMXEXT_FILTER_CODE
328 CALL_MMXEXT_FILTER_CODE
329 CALL_MMXEXT_FILTER_CODE
332 "mov %7, %%"FF_REG_a" \n\t"
333 "mov %%"FF_REG_a", -8(%%rsp) \n\t"
335 #if !HAVE_EBX_AVAILABLE
336 "mov %7, %%"FF_REG_b" \n\t"
339 :: "m" (src1), "m" (dst1), "m" (filter), "m" (filterPos),
340 "m" (mmxextFilterCode), "m" (src2), "m"(dst2)
344 #if !HAVE_EBX_AVAILABLE
348 : "%"FF_REG_a, "%"FF_REG_c, "%"FF_REG_d, "%"FF_REG_S, "%"FF_REG_D
349 #if ARCH_X86_64 || HAVE_EBX_AVAILABLE
354 for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--) {
355 dst1[i] = src1[srcW-1]*128;
356 dst2[i] = src2[srcW-1]*128;
359 #endif //HAVE_INLINE_ASM