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"
24 #include "libavutil/mem_internal.h"
26 #define RET 0xC3 // near return opcode for x86
27 #define PREFETCH "prefetchnta"
30 av_cold int ff_init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
31 int16_t *filter, int32_t *filterPos,
35 x86_reg imm8OfPShufW1A;
36 x86_reg imm8OfPShufW2A;
37 x86_reg fragmentLengthA;
39 x86_reg imm8OfPShufW1B;
40 x86_reg imm8OfPShufW2B;
41 x86_reg fragmentLengthB;
46 // create an optimized horizontal scaling routine
47 /* This scaler is made of runtime-generated MMXEXT code using specially tuned
48 * pshufw instructions. For every four output pixels, if four input pixels
49 * are enough for the fast bilinear scaling, then a chunk of fragmentB is
50 * used. If five input pixels are needed, then a chunk of fragmentA is used.
59 "movq (%%"FF_REG_d", %%"FF_REG_a"), %%mm3 \n\t"
60 "movd (%%"FF_REG_c", %%"FF_REG_S"), %%mm0 \n\t"
61 "movd 1(%%"FF_REG_c", %%"FF_REG_S"), %%mm1 \n\t"
62 "punpcklbw %%mm7, %%mm1 \n\t"
63 "punpcklbw %%mm7, %%mm0 \n\t"
64 "pshufw $0xFF, %%mm1, %%mm1 \n\t"
66 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
68 "psubw %%mm1, %%mm0 \n\t"
69 "movl 8(%%"FF_REG_b", %%"FF_REG_a"), %%esi \n\t"
70 "pmullw %%mm3, %%mm0 \n\t"
71 "psllw $7, %%mm1 \n\t"
72 "paddw %%mm1, %%mm0 \n\t"
74 "movq %%mm0, (%%"FF_REG_D", %%"FF_REG_a") \n\t"
76 "add $8, %%"FF_REG_a" \n\t"
79 "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
80 "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
81 "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
86 "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
90 : "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
91 "=r" (fragmentLengthA)
98 "movq (%%"FF_REG_d", %%"FF_REG_a"), %%mm3 \n\t"
99 "movd (%%"FF_REG_c", %%"FF_REG_S"), %%mm0 \n\t"
100 "punpcklbw %%mm7, %%mm0 \n\t"
101 "pshufw $0xFF, %%mm0, %%mm1 \n\t"
103 "pshufw $0xFF, %%mm0, %%mm0 \n\t"
105 "psubw %%mm1, %%mm0 \n\t"
106 "movl 8(%%"FF_REG_b", %%"FF_REG_a"), %%esi \n\t"
107 "pmullw %%mm3, %%mm0 \n\t"
108 "psllw $7, %%mm1 \n\t"
109 "paddw %%mm1, %%mm0 \n\t"
111 "movq %%mm0, (%%"FF_REG_D", %%"FF_REG_a") \n\t"
113 "add $8, %%"FF_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;
203 #if !HAVE_EBX_AVAILABLE
210 "mov -8(%%rsp), %%"FF_REG_a" \n\t"
211 "mov %%"FF_REG_a", %5 \n\t" // retsave
213 #if !HAVE_EBX_AVAILABLE
214 "mov %%"FF_REG_b", %5 \n\t" // ebxsave
217 "pxor %%mm7, %%mm7 \n\t"
218 "mov %0, %%"FF_REG_c" \n\t"
219 "mov %1, %%"FF_REG_D" \n\t"
220 "mov %2, %%"FF_REG_d" \n\t"
221 "mov %3, %%"FF_REG_b" \n\t"
222 "xor %%"FF_REG_a", %%"FF_REG_a" \n\t" // i
223 PREFETCH" (%%"FF_REG_c") \n\t"
224 PREFETCH" 32(%%"FF_REG_c") \n\t"
225 PREFETCH" 64(%%"FF_REG_c") \n\t"
228 #define CALL_MMXEXT_FILTER_CODE \
229 "movl (%%"FF_REG_b"), %%esi \n\t"\
231 "movl (%%"FF_REG_b", %%"FF_REG_a"), %%esi \n\t"\
232 "add %%"FF_REG_S", %%"FF_REG_c" \n\t"\
233 "add %%"FF_REG_a", %%"FF_REG_D" \n\t"\
234 "xor %%"FF_REG_a", %%"FF_REG_a" \n\t"\
237 #define CALL_MMXEXT_FILTER_CODE \
238 "movl (%%"FF_REG_b"), %%esi \n\t"\
240 "addl (%%"FF_REG_b", %%"FF_REG_a"), %%"FF_REG_c" \n\t"\
241 "add %%"FF_REG_a", %%"FF_REG_D" \n\t"\
242 "xor %%"FF_REG_a", %%"FF_REG_a" \n\t"\
244 #endif /* ARCH_X86_64 */
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
253 CALL_MMXEXT_FILTER_CODE
256 "mov %5, %%"FF_REG_a" \n\t"
257 "mov %%"FF_REG_a", -8(%%rsp) \n\t"
259 #if !HAVE_EBX_AVAILABLE
260 "mov %5, %%"FF_REG_b" \n\t"
263 :: "m" (src), "m" (dst), "m" (filter), "m" (filterPos),
264 "m" (mmxextFilterCode)
268 #if !HAVE_EBX_AVAILABLE
272 : "%"FF_REG_a, "%"FF_REG_c, "%"FF_REG_d, "%"FF_REG_S, "%"FF_REG_D
273 #if ARCH_X86_64 || HAVE_EBX_AVAILABLE
278 for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--)
279 dst[i] = src[srcW-1]*128;
282 void ff_hcscale_fast_mmxext(SwsContext *c, int16_t *dst1, int16_t *dst2,
283 int dstWidth, const uint8_t *src1,
284 const uint8_t *src2, int srcW, int xInc)
286 int32_t *filterPos = c->hChrFilterPos;
287 int16_t *filter = c->hChrFilter;
288 void *mmxextFilterCode = c->chrMmxextFilterCode;
291 DECLARE_ALIGNED(8, uint64_t, retsave);
293 #if !HAVE_EBX_AVAILABLE
294 DECLARE_ALIGNED(8, uint64_t, ebxsave);
299 "mov -8(%%rsp), %%"FF_REG_a" \n\t"
300 "mov %%"FF_REG_a", %7 \n\t" // retsave
302 #if !HAVE_EBX_AVAILABLE
303 "mov %%"FF_REG_b", %7 \n\t" // ebxsave
306 "pxor %%mm7, %%mm7 \n\t"
307 "mov %0, %%"FF_REG_c" \n\t"
308 "mov %1, %%"FF_REG_D" \n\t"
309 "mov %2, %%"FF_REG_d" \n\t"
310 "mov %3, %%"FF_REG_b" \n\t"
311 "xor %%"FF_REG_a", %%"FF_REG_a" \n\t" // i
312 PREFETCH" (%%"FF_REG_c") \n\t"
313 PREFETCH" 32(%%"FF_REG_c") \n\t"
314 PREFETCH" 64(%%"FF_REG_c") \n\t"
316 CALL_MMXEXT_FILTER_CODE
317 CALL_MMXEXT_FILTER_CODE
318 CALL_MMXEXT_FILTER_CODE
319 CALL_MMXEXT_FILTER_CODE
320 "xor %%"FF_REG_a", %%"FF_REG_a" \n\t" // i
321 "mov %5, %%"FF_REG_c" \n\t" // src2
322 "mov %6, %%"FF_REG_D" \n\t" // dst2
323 PREFETCH" (%%"FF_REG_c") \n\t"
324 PREFETCH" 32(%%"FF_REG_c") \n\t"
325 PREFETCH" 64(%%"FF_REG_c") \n\t"
327 CALL_MMXEXT_FILTER_CODE
328 CALL_MMXEXT_FILTER_CODE
329 CALL_MMXEXT_FILTER_CODE
330 CALL_MMXEXT_FILTER_CODE
333 "mov %7, %%"FF_REG_a" \n\t"
334 "mov %%"FF_REG_a", -8(%%rsp) \n\t"
336 #if !HAVE_EBX_AVAILABLE
337 "mov %7, %%"FF_REG_b" \n\t"
340 :: "m" (src1), "m" (dst1), "m" (filter), "m" (filterPos),
341 "m" (mmxextFilterCode), "m" (src2), "m"(dst2)
345 #if !HAVE_EBX_AVAILABLE
349 : "%"FF_REG_a, "%"FF_REG_c, "%"FF_REG_d, "%"FF_REG_S, "%"FF_REG_D
350 #if ARCH_X86_64 || HAVE_EBX_AVAILABLE
355 for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--) {
356 dst1[i] = src1[srcW-1]*128;
357 dst2[i] = src2[srcW-1]*128;
360 #endif //HAVE_INLINE_ASM