yading@10: /* yading@10: * MMX optimized DSP utils yading@10: * Copyright (c) 2000, 2001 Fabrice Bellard yading@10: * Copyright (c) 2002-2004 Michael Niedermayer yading@10: * yading@10: * This file is part of FFmpeg. yading@10: * yading@10: * FFmpeg is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * FFmpeg is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with FFmpeg; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: * yading@10: * MMX optimization by Nick Kurshev yading@10: */ yading@10: yading@10: #include "libavutil/attributes.h" yading@10: #include "libavutil/cpu.h" yading@10: #include "libavutil/x86/asm.h" yading@10: #include "libavcodec/dsputil.h" yading@10: #include "libavcodec/h264dsp.h" yading@10: #include "libavcodec/mpegvideo.h" yading@10: #include "libavcodec/simple_idct.h" yading@10: #include "libavcodec/videodsp.h" yading@10: #include "dsputil_mmx.h" yading@10: #include "idct_xvid.h" yading@10: #include "diracdsp_mmx.h" yading@10: yading@10: //#undef NDEBUG yading@10: //#include yading@10: yading@10: /* pixel operations */ yading@10: DECLARE_ALIGNED(8, const uint64_t, ff_pw_15) = 0x000F000F000F000FULL; yading@10: DECLARE_ALIGNED(16, const xmm_reg, ff_pw_17) = { 0x0011001100110011ULL, 0x0011001100110011ULL }; yading@10: DECLARE_ALIGNED(8, const uint64_t, ff_pw_20) = 0x0014001400140014ULL; yading@10: DECLARE_ALIGNED(8, const uint64_t, ff_pw_42) = 0x002A002A002A002AULL; yading@10: DECLARE_ALIGNED(8, const uint64_t, ff_pw_53) = 0x0035003500350035ULL; yading@10: DECLARE_ALIGNED(8, const uint64_t, ff_pw_96) = 0x0060006000600060ULL; yading@10: DECLARE_ALIGNED(8, const uint64_t, ff_pw_128) = 0x0080008000800080ULL; yading@10: DECLARE_ALIGNED(8, const uint64_t, ff_pw_255) = 0x00ff00ff00ff00ffULL; yading@10: DECLARE_ALIGNED(16, const xmm_reg, ff_pw_512) = { 0x0200020002000200ULL, 0x0200020002000200ULL }; yading@10: DECLARE_ALIGNED(16, const xmm_reg, ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL }; yading@10: yading@10: DECLARE_ALIGNED(8, const uint64_t, ff_pb_3F) = 0x3F3F3F3F3F3F3F3FULL; yading@10: DECLARE_ALIGNED(8, const uint64_t, ff_pb_FC) = 0xFCFCFCFCFCFCFCFCULL; yading@10: yading@10: DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 }; yading@10: DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 }; yading@10: yading@10: yading@10: #if HAVE_YASM yading@10: void ff_put_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2, yading@10: int dstStride, int src1Stride, int h); yading@10: void ff_put_no_rnd_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, yading@10: uint8_t *src2, int dstStride, yading@10: int src1Stride, int h); yading@10: void ff_avg_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2, yading@10: int dstStride, int src1Stride, int h); yading@10: void ff_put_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2, yading@10: int dstStride, int src1Stride, int h); yading@10: void ff_avg_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2, yading@10: int dstStride, int src1Stride, int h); yading@10: void ff_put_no_rnd_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2, yading@10: int dstStride, int src1Stride, int h); yading@10: void ff_avg_pixels8_mmxext(uint8_t *block, const uint8_t *pixels, yading@10: ptrdiff_t line_size, int h); yading@10: yading@10: static void ff_put_pixels16_mmxext(uint8_t *block, const uint8_t *pixels, yading@10: ptrdiff_t line_size, int h) yading@10: { yading@10: ff_put_pixels8_mmxext(block, pixels, line_size, h); yading@10: ff_put_pixels8_mmxext(block + 8, pixels + 8, line_size, h); yading@10: } yading@10: yading@10: void ff_put_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride, int h); yading@10: void ff_avg_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride, int h); yading@10: void ff_put_no_rnd_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride, yading@10: int h); yading@10: void ff_put_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride, int h); yading@10: void ff_avg_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride, int h); yading@10: void ff_put_no_rnd_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride, yading@10: int h); yading@10: void ff_put_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride); yading@10: void ff_avg_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride); yading@10: void ff_put_no_rnd_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride); yading@10: void ff_put_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride); yading@10: void ff_avg_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride); yading@10: void ff_put_no_rnd_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src, yading@10: int dstStride, int srcStride); yading@10: #define ff_put_no_rnd_pixels16_mmxext ff_put_pixels16_mmxext yading@10: #define ff_put_no_rnd_pixels8_mmxext ff_put_pixels8_mmxext yading@10: #endif /* HAVE_YASM */ yading@10: yading@10: yading@10: #if HAVE_INLINE_ASM yading@10: yading@10: #define JUMPALIGN() __asm__ volatile (".p2align 3"::) yading@10: #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::) yading@10: yading@10: #define MOVQ_BFE(regd) \ yading@10: __asm__ volatile ( \ yading@10: "pcmpeqd %%"#regd", %%"#regd" \n\t" \ yading@10: "paddb %%"#regd", %%"#regd" \n\t" ::) yading@10: yading@10: #ifndef PIC yading@10: #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone)) yading@10: #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo)) yading@10: #else yading@10: // for shared library it's better to use this way for accessing constants yading@10: // pcmpeqd -> -1 yading@10: #define MOVQ_BONE(regd) \ yading@10: __asm__ volatile ( \ yading@10: "pcmpeqd %%"#regd", %%"#regd" \n\t" \ yading@10: "psrlw $15, %%"#regd" \n\t" \ yading@10: "packuswb %%"#regd", %%"#regd" \n\t" ::) yading@10: yading@10: #define MOVQ_WTWO(regd) \ yading@10: __asm__ volatile ( \ yading@10: "pcmpeqd %%"#regd", %%"#regd" \n\t" \ yading@10: "psrlw $15, %%"#regd" \n\t" \ yading@10: "psllw $1, %%"#regd" \n\t"::) yading@10: yading@10: #endif yading@10: yading@10: // using regr as temporary and for the output result yading@10: // first argument is unmodifed and second is trashed yading@10: // regfe is supposed to contain 0xfefefefefefefefe yading@10: #define PAVGB_MMX(rega, regb, regr, regfe) \ yading@10: "movq "#rega", "#regr" \n\t" \ yading@10: "por "#regb", "#regr" \n\t" \ yading@10: "pxor "#rega", "#regb" \n\t" \ yading@10: "pand "#regfe", "#regb" \n\t" \ yading@10: "psrlq $1, "#regb" \n\t" \ yading@10: "psubb "#regb", "#regr" \n\t" yading@10: yading@10: // mm6 is supposed to contain 0xfefefefefefefefe yading@10: #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp) \ yading@10: "movq "#rega", "#regr" \n\t" \ yading@10: "movq "#regc", "#regp" \n\t" \ yading@10: "por "#regb", "#regr" \n\t" \ yading@10: "por "#regd", "#regp" \n\t" \ yading@10: "pxor "#rega", "#regb" \n\t" \ yading@10: "pxor "#regc", "#regd" \n\t" \ yading@10: "pand %%mm6, "#regb" \n\t" \ yading@10: "pand %%mm6, "#regd" \n\t" \ yading@10: "psrlq $1, "#regd" \n\t" \ yading@10: "psrlq $1, "#regb" \n\t" \ yading@10: "psubb "#regb", "#regr" \n\t" \ yading@10: "psubb "#regd", "#regp" \n\t" yading@10: yading@10: /***********************************/ yading@10: /* MMX rounding */ yading@10: yading@10: #define DEF(x, y) x ## _ ## y ## _mmx yading@10: #define SET_RND MOVQ_WTWO yading@10: #define PAVGBP(a, b, c, d, e, f) PAVGBP_MMX(a, b, c, d, e, f) yading@10: #define PAVGB(a, b, c, e) PAVGB_MMX(a, b, c, e) yading@10: #define OP_AVG(a, b, c, e) PAVGB_MMX(a, b, c, e) yading@10: yading@10: #include "dsputil_rnd_template.c" yading@10: yading@10: #undef DEF yading@10: #undef SET_RND yading@10: #undef PAVGBP yading@10: #undef PAVGB yading@10: #undef OP_AVG yading@10: yading@10: #endif /* HAVE_INLINE_ASM */ yading@10: yading@10: yading@10: #if HAVE_YASM yading@10: yading@10: /***********************************/ yading@10: /* MMXEXT specific */ yading@10: yading@10: //FIXME the following could be optimized too ... yading@10: static void ff_avg_pixels16_mmxext(uint8_t *block, const uint8_t *pixels, yading@10: int line_size, int h) yading@10: { yading@10: ff_avg_pixels8_mmxext(block, pixels, line_size, h); yading@10: ff_avg_pixels8_mmxext(block + 8, pixels + 8, line_size, h); yading@10: } yading@10: yading@10: #endif /* HAVE_YASM */ yading@10: yading@10: yading@10: #if HAVE_INLINE_ASM yading@10: /***********************************/ yading@10: /* standard MMX */ yading@10: yading@10: void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels, yading@10: int line_size) yading@10: { yading@10: const int16_t *p; yading@10: uint8_t *pix; yading@10: yading@10: /* read the pixels */ yading@10: p = block; yading@10: pix = pixels; yading@10: /* unrolled loop */ yading@10: __asm__ volatile ( yading@10: "movq (%3), %%mm0 \n\t" yading@10: "movq 8(%3), %%mm1 \n\t" yading@10: "movq 16(%3), %%mm2 \n\t" yading@10: "movq 24(%3), %%mm3 \n\t" yading@10: "movq 32(%3), %%mm4 \n\t" yading@10: "movq 40(%3), %%mm5 \n\t" yading@10: "movq 48(%3), %%mm6 \n\t" yading@10: "movq 56(%3), %%mm7 \n\t" yading@10: "packuswb %%mm1, %%mm0 \n\t" yading@10: "packuswb %%mm3, %%mm2 \n\t" yading@10: "packuswb %%mm5, %%mm4 \n\t" yading@10: "packuswb %%mm7, %%mm6 \n\t" yading@10: "movq %%mm0, (%0) \n\t" yading@10: "movq %%mm2, (%0, %1) \n\t" yading@10: "movq %%mm4, (%0, %1, 2) \n\t" yading@10: "movq %%mm6, (%0, %2) \n\t" yading@10: :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), yading@10: "r"(p) yading@10: : "memory"); yading@10: pix += line_size * 4; yading@10: p += 32; yading@10: yading@10: // if here would be an exact copy of the code above yading@10: // compiler would generate some very strange code yading@10: // thus using "r" yading@10: __asm__ volatile ( yading@10: "movq (%3), %%mm0 \n\t" yading@10: "movq 8(%3), %%mm1 \n\t" yading@10: "movq 16(%3), %%mm2 \n\t" yading@10: "movq 24(%3), %%mm3 \n\t" yading@10: "movq 32(%3), %%mm4 \n\t" yading@10: "movq 40(%3), %%mm5 \n\t" yading@10: "movq 48(%3), %%mm6 \n\t" yading@10: "movq 56(%3), %%mm7 \n\t" yading@10: "packuswb %%mm1, %%mm0 \n\t" yading@10: "packuswb %%mm3, %%mm2 \n\t" yading@10: "packuswb %%mm5, %%mm4 \n\t" yading@10: "packuswb %%mm7, %%mm6 \n\t" yading@10: "movq %%mm0, (%0) \n\t" yading@10: "movq %%mm2, (%0, %1) \n\t" yading@10: "movq %%mm4, (%0, %1, 2) \n\t" yading@10: "movq %%mm6, (%0, %2) \n\t" yading@10: :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p) yading@10: : "memory"); yading@10: } yading@10: yading@10: #define put_signed_pixels_clamped_mmx_half(off) \ yading@10: "movq "#off"(%2), %%mm1 \n\t" \ yading@10: "movq 16 + "#off"(%2), %%mm2 \n\t" \ yading@10: "movq 32 + "#off"(%2), %%mm3 \n\t" \ yading@10: "movq 48 + "#off"(%2), %%mm4 \n\t" \ yading@10: "packsswb 8 + "#off"(%2), %%mm1 \n\t" \ yading@10: "packsswb 24 + "#off"(%2), %%mm2 \n\t" \ yading@10: "packsswb 40 + "#off"(%2), %%mm3 \n\t" \ yading@10: "packsswb 56 + "#off"(%2), %%mm4 \n\t" \ yading@10: "paddb %%mm0, %%mm1 \n\t" \ yading@10: "paddb %%mm0, %%mm2 \n\t" \ yading@10: "paddb %%mm0, %%mm3 \n\t" \ yading@10: "paddb %%mm0, %%mm4 \n\t" \ yading@10: "movq %%mm1, (%0) \n\t" \ yading@10: "movq %%mm2, (%0, %3) \n\t" \ yading@10: "movq %%mm3, (%0, %3, 2) \n\t" \ yading@10: "movq %%mm4, (%0, %1) \n\t" yading@10: yading@10: void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels, yading@10: int line_size) yading@10: { yading@10: x86_reg line_skip = line_size; yading@10: x86_reg line_skip3; yading@10: yading@10: __asm__ volatile ( yading@10: "movq "MANGLE(ff_pb_80)", %%mm0 \n\t" yading@10: "lea (%3, %3, 2), %1 \n\t" yading@10: put_signed_pixels_clamped_mmx_half(0) yading@10: "lea (%0, %3, 4), %0 \n\t" yading@10: put_signed_pixels_clamped_mmx_half(64) yading@10: : "+&r"(pixels), "=&r"(line_skip3) yading@10: : "r"(block), "r"(line_skip) yading@10: : "memory"); yading@10: } yading@10: yading@10: void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels, yading@10: int line_size) yading@10: { yading@10: const int16_t *p; yading@10: uint8_t *pix; yading@10: int i; yading@10: yading@10: /* read the pixels */ yading@10: p = block; yading@10: pix = pixels; yading@10: MOVQ_ZERO(mm7); yading@10: i = 4; yading@10: do { yading@10: __asm__ volatile ( yading@10: "movq (%2), %%mm0 \n\t" yading@10: "movq 8(%2), %%mm1 \n\t" yading@10: "movq 16(%2), %%mm2 \n\t" yading@10: "movq 24(%2), %%mm3 \n\t" yading@10: "movq %0, %%mm4 \n\t" yading@10: "movq %1, %%mm6 \n\t" yading@10: "movq %%mm4, %%mm5 \n\t" yading@10: "punpcklbw %%mm7, %%mm4 \n\t" yading@10: "punpckhbw %%mm7, %%mm5 \n\t" yading@10: "paddsw %%mm4, %%mm0 \n\t" yading@10: "paddsw %%mm5, %%mm1 \n\t" yading@10: "movq %%mm6, %%mm5 \n\t" yading@10: "punpcklbw %%mm7, %%mm6 \n\t" yading@10: "punpckhbw %%mm7, %%mm5 \n\t" yading@10: "paddsw %%mm6, %%mm2 \n\t" yading@10: "paddsw %%mm5, %%mm3 \n\t" yading@10: "packuswb %%mm1, %%mm0 \n\t" yading@10: "packuswb %%mm3, %%mm2 \n\t" yading@10: "movq %%mm0, %0 \n\t" yading@10: "movq %%mm2, %1 \n\t" yading@10: : "+m"(*pix), "+m"(*(pix + line_size)) yading@10: : "r"(p) yading@10: : "memory"); yading@10: pix += line_size * 2; yading@10: p += 16; yading@10: } while (--i); yading@10: } yading@10: yading@10: static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels, yading@10: ptrdiff_t line_size, int h) yading@10: { yading@10: __asm__ volatile ( yading@10: "lea (%3, %3), %%"REG_a" \n\t" yading@10: ".p2align 3 \n\t" yading@10: "1: \n\t" yading@10: "movq (%1 ), %%mm0 \n\t" yading@10: "movq (%1, %3), %%mm1 \n\t" yading@10: "movq %%mm0, (%2) \n\t" yading@10: "movq %%mm1, (%2, %3) \n\t" yading@10: "add %%"REG_a", %1 \n\t" yading@10: "add %%"REG_a", %2 \n\t" yading@10: "movq (%1 ), %%mm0 \n\t" yading@10: "movq (%1, %3), %%mm1 \n\t" yading@10: "movq %%mm0, (%2) \n\t" yading@10: "movq %%mm1, (%2, %3) \n\t" yading@10: "add %%"REG_a", %1 \n\t" yading@10: "add %%"REG_a", %2 \n\t" yading@10: "subl $4, %0 \n\t" yading@10: "jnz 1b \n\t" yading@10: : "+g"(h), "+r"(pixels), "+r"(block) yading@10: : "r"((x86_reg)line_size) yading@10: : "%"REG_a, "memory" yading@10: ); yading@10: } yading@10: yading@10: static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels, yading@10: ptrdiff_t line_size, int h) yading@10: { yading@10: __asm__ volatile ( yading@10: "lea (%3, %3), %%"REG_a" \n\t" yading@10: ".p2align 3 \n\t" yading@10: "1: \n\t" yading@10: "movq (%1 ), %%mm0 \n\t" yading@10: "movq 8(%1 ), %%mm4 \n\t" yading@10: "movq (%1, %3), %%mm1 \n\t" yading@10: "movq 8(%1, %3), %%mm5 \n\t" yading@10: "movq %%mm0, (%2) \n\t" yading@10: "movq %%mm4, 8(%2) \n\t" yading@10: "movq %%mm1, (%2, %3) \n\t" yading@10: "movq %%mm5, 8(%2, %3) \n\t" yading@10: "add %%"REG_a", %1 \n\t" yading@10: "add %%"REG_a", %2 \n\t" yading@10: "movq (%1 ), %%mm0 \n\t" yading@10: "movq 8(%1 ), %%mm4 \n\t" yading@10: "movq (%1, %3), %%mm1 \n\t" yading@10: "movq 8(%1, %3), %%mm5 \n\t" yading@10: "movq %%mm0, (%2) \n\t" yading@10: "movq %%mm4, 8(%2) \n\t" yading@10: "movq %%mm1, (%2, %3) \n\t" yading@10: "movq %%mm5, 8(%2, %3) \n\t" yading@10: "add %%"REG_a", %1 \n\t" yading@10: "add %%"REG_a", %2 \n\t" yading@10: "subl $4, %0 \n\t" yading@10: "jnz 1b \n\t" yading@10: : "+g"(h), "+r"(pixels), "+r"(block) yading@10: : "r"((x86_reg)line_size) yading@10: : "%"REG_a, "memory" yading@10: ); yading@10: } yading@10: yading@10: #define CLEAR_BLOCKS(name, n) \ yading@10: static void name(int16_t *blocks) \ yading@10: { \ yading@10: __asm__ volatile ( \ yading@10: "pxor %%mm7, %%mm7 \n\t" \ yading@10: "mov %1, %%"REG_a" \n\t" \ yading@10: "1: \n\t" \ yading@10: "movq %%mm7, (%0, %%"REG_a") \n\t" \ yading@10: "movq %%mm7, 8(%0, %%"REG_a") \n\t" \ yading@10: "movq %%mm7, 16(%0, %%"REG_a") \n\t" \ yading@10: "movq %%mm7, 24(%0, %%"REG_a") \n\t" \ yading@10: "add $32, %%"REG_a" \n\t" \ yading@10: "js 1b \n\t" \ yading@10: :: "r"(((uint8_t *)blocks) + 128 * n), \ yading@10: "i"(-128 * n) \ yading@10: : "%"REG_a \ yading@10: ); \ yading@10: } yading@10: CLEAR_BLOCKS(clear_blocks_mmx, 6) yading@10: CLEAR_BLOCKS(clear_block_mmx, 1) yading@10: yading@10: static void clear_block_sse(int16_t *block) yading@10: { yading@10: __asm__ volatile ( yading@10: "xorps %%xmm0, %%xmm0 \n" yading@10: "movaps %%xmm0, (%0) \n" yading@10: "movaps %%xmm0, 16(%0) \n" yading@10: "movaps %%xmm0, 32(%0) \n" yading@10: "movaps %%xmm0, 48(%0) \n" yading@10: "movaps %%xmm0, 64(%0) \n" yading@10: "movaps %%xmm0, 80(%0) \n" yading@10: "movaps %%xmm0, 96(%0) \n" yading@10: "movaps %%xmm0, 112(%0) \n" yading@10: :: "r"(block) yading@10: : "memory" yading@10: ); yading@10: } yading@10: yading@10: static void clear_blocks_sse(int16_t *blocks) yading@10: { yading@10: __asm__ volatile ( yading@10: "xorps %%xmm0, %%xmm0 \n" yading@10: "mov %1, %%"REG_a" \n" yading@10: "1: \n" yading@10: "movaps %%xmm0, (%0, %%"REG_a") \n" yading@10: "movaps %%xmm0, 16(%0, %%"REG_a") \n" yading@10: "movaps %%xmm0, 32(%0, %%"REG_a") \n" yading@10: "movaps %%xmm0, 48(%0, %%"REG_a") \n" yading@10: "movaps %%xmm0, 64(%0, %%"REG_a") \n" yading@10: "movaps %%xmm0, 80(%0, %%"REG_a") \n" yading@10: "movaps %%xmm0, 96(%0, %%"REG_a") \n" yading@10: "movaps %%xmm0, 112(%0, %%"REG_a") \n" yading@10: "add $128, %%"REG_a" \n" yading@10: "js 1b \n" yading@10: :: "r"(((uint8_t *)blocks) + 128 * 6), yading@10: "i"(-128 * 6) yading@10: : "%"REG_a yading@10: ); yading@10: } yading@10: yading@10: static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w) yading@10: { yading@10: x86_reg i = 0; yading@10: __asm__ volatile ( yading@10: "jmp 2f \n\t" yading@10: "1: \n\t" yading@10: "movq (%1, %0), %%mm0 \n\t" yading@10: "movq (%2, %0), %%mm1 \n\t" yading@10: "paddb %%mm0, %%mm1 \n\t" yading@10: "movq %%mm1, (%2, %0) \n\t" yading@10: "movq 8(%1, %0), %%mm0 \n\t" yading@10: "movq 8(%2, %0), %%mm1 \n\t" yading@10: "paddb %%mm0, %%mm1 \n\t" yading@10: "movq %%mm1, 8(%2, %0) \n\t" yading@10: "add $16, %0 \n\t" yading@10: "2: \n\t" yading@10: "cmp %3, %0 \n\t" yading@10: "js 1b \n\t" yading@10: : "+r"(i) yading@10: : "r"(src), "r"(dst), "r"((x86_reg)w - 15) yading@10: ); yading@10: for ( ; i < w; i++) yading@10: dst[i + 0] += src[i + 0]; yading@10: } yading@10: yading@10: #if HAVE_7REGS yading@10: static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top, yading@10: const uint8_t *diff, int w, yading@10: int *left, int *left_top) yading@10: { yading@10: x86_reg w2 = -w; yading@10: x86_reg x; yading@10: int l = *left & 0xff; yading@10: int tl = *left_top & 0xff; yading@10: int t; yading@10: __asm__ volatile ( yading@10: "mov %7, %3 \n" yading@10: "1: \n" yading@10: "movzbl (%3, %4), %2 \n" yading@10: "mov %2, %k3 \n" yading@10: "sub %b1, %b3 \n" yading@10: "add %b0, %b3 \n" yading@10: "mov %2, %1 \n" yading@10: "cmp %0, %2 \n" yading@10: "cmovg %0, %2 \n" yading@10: "cmovg %1, %0 \n" yading@10: "cmp %k3, %0 \n" yading@10: "cmovg %k3, %0 \n" yading@10: "mov %7, %3 \n" yading@10: "cmp %2, %0 \n" yading@10: "cmovl %2, %0 \n" yading@10: "add (%6, %4), %b0 \n" yading@10: "mov %b0, (%5, %4) \n" yading@10: "inc %4 \n" yading@10: "jl 1b \n" yading@10: : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2) yading@10: : "r"(dst + w), "r"(diff + w), "rm"(top + w) yading@10: ); yading@10: *left = l; yading@10: *left_top = tl; yading@10: } yading@10: #endif yading@10: yading@10: /* Draw the edges of width 'w' of an image of size width, height yading@10: * this MMX version can only handle w == 8 || w == 16. */ yading@10: static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, yading@10: int w, int h, int sides) yading@10: { yading@10: uint8_t *ptr, *last_line; yading@10: int i; yading@10: yading@10: last_line = buf + (height - 1) * wrap; yading@10: /* left and right */ yading@10: ptr = buf; yading@10: if (w == 8) { yading@10: __asm__ volatile ( yading@10: "1: \n\t" yading@10: "movd (%0), %%mm0 \n\t" yading@10: "punpcklbw %%mm0, %%mm0 \n\t" yading@10: "punpcklwd %%mm0, %%mm0 \n\t" yading@10: "punpckldq %%mm0, %%mm0 \n\t" yading@10: "movq %%mm0, -8(%0) \n\t" yading@10: "movq -8(%0, %2), %%mm1 \n\t" yading@10: "punpckhbw %%mm1, %%mm1 \n\t" yading@10: "punpckhwd %%mm1, %%mm1 \n\t" yading@10: "punpckhdq %%mm1, %%mm1 \n\t" yading@10: "movq %%mm1, (%0, %2) \n\t" yading@10: "add %1, %0 \n\t" yading@10: "cmp %3, %0 \n\t" yading@10: "jb 1b \n\t" yading@10: : "+r"(ptr) yading@10: : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height) yading@10: ); yading@10: } else if(w==16){ yading@10: __asm__ volatile ( yading@10: "1: \n\t" yading@10: "movd (%0), %%mm0 \n\t" yading@10: "punpcklbw %%mm0, %%mm0 \n\t" yading@10: "punpcklwd %%mm0, %%mm0 \n\t" yading@10: "punpckldq %%mm0, %%mm0 \n\t" yading@10: "movq %%mm0, -8(%0) \n\t" yading@10: "movq %%mm0, -16(%0) \n\t" yading@10: "movq -8(%0, %2), %%mm1 \n\t" yading@10: "punpckhbw %%mm1, %%mm1 \n\t" yading@10: "punpckhwd %%mm1, %%mm1 \n\t" yading@10: "punpckhdq %%mm1, %%mm1 \n\t" yading@10: "movq %%mm1, (%0, %2) \n\t" yading@10: "movq %%mm1, 8(%0, %2) \n\t" yading@10: "add %1, %0 \n\t" yading@10: "cmp %3, %0 \n\t" yading@10: "jb 1b \n\t" yading@10: : "+r"(ptr) yading@10: : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height) yading@10: ); yading@10: } else { yading@10: av_assert1(w == 4); yading@10: __asm__ volatile ( yading@10: "1: \n\t" yading@10: "movd (%0), %%mm0 \n\t" yading@10: "punpcklbw %%mm0, %%mm0 \n\t" yading@10: "punpcklwd %%mm0, %%mm0 \n\t" yading@10: "movd %%mm0, -4(%0) \n\t" yading@10: "movd -4(%0, %2), %%mm1 \n\t" yading@10: "punpcklbw %%mm1, %%mm1 \n\t" yading@10: "punpckhwd %%mm1, %%mm1 \n\t" yading@10: "punpckhdq %%mm1, %%mm1 \n\t" yading@10: "movd %%mm1, (%0, %2) \n\t" yading@10: "add %1, %0 \n\t" yading@10: "cmp %3, %0 \n\t" yading@10: "jb 1b \n\t" yading@10: : "+r"(ptr) yading@10: : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height) yading@10: ); yading@10: } yading@10: yading@10: /* top and bottom (and hopefully also the corners) */ yading@10: if (sides & EDGE_TOP) { yading@10: for (i = 0; i < h; i += 4) { yading@10: ptr = buf - (i + 1) * wrap - w; yading@10: __asm__ volatile ( yading@10: "1: \n\t" yading@10: "movq (%1, %0), %%mm0 \n\t" yading@10: "movq %%mm0, (%0) \n\t" yading@10: "movq %%mm0, (%0, %2) \n\t" yading@10: "movq %%mm0, (%0, %2, 2) \n\t" yading@10: "movq %%mm0, (%0, %3) \n\t" yading@10: "add $8, %0 \n\t" yading@10: "cmp %4, %0 \n\t" yading@10: "jb 1b \n\t" yading@10: : "+r"(ptr) yading@10: : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap), yading@10: "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w) yading@10: ); yading@10: } yading@10: } yading@10: yading@10: if (sides & EDGE_BOTTOM) { yading@10: for (i = 0; i < h; i += 4) { yading@10: ptr = last_line + (i + 1) * wrap - w; yading@10: __asm__ volatile ( yading@10: "1: \n\t" yading@10: "movq (%1, %0), %%mm0 \n\t" yading@10: "movq %%mm0, (%0) \n\t" yading@10: "movq %%mm0, (%0, %2) \n\t" yading@10: "movq %%mm0, (%0, %2, 2) \n\t" yading@10: "movq %%mm0, (%0, %3) \n\t" yading@10: "add $8, %0 \n\t" yading@10: "cmp %4, %0 \n\t" yading@10: "jb 1b \n\t" yading@10: : "+r"(ptr) yading@10: : "r"((x86_reg)last_line - (x86_reg)ptr - w), yading@10: "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3), yading@10: "r"(ptr + width + 2 * w) yading@10: ); yading@10: } yading@10: } yading@10: } yading@10: #endif /* HAVE_INLINE_ASM */ yading@10: yading@10: yading@10: #if HAVE_YASM yading@10: #define QPEL_OP(OPNAME, ROUNDER, RND, MMX) \ yading@10: static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: ff_ ## OPNAME ## pixels8_ ## MMX(dst, src, stride, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t temp[8]; \ yading@10: uint8_t * const half = (uint8_t*)temp; \ yading@10: ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, \ yading@10: stride, 8); \ yading@10: ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half, \ yading@10: stride, stride, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: ff_ ## OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride, \ yading@10: stride, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t temp[8]; \ yading@10: uint8_t * const half = (uint8_t*)temp; \ yading@10: ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8, \ yading@10: stride, 8); \ yading@10: ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride, \ yading@10: stride, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t temp[8]; \ yading@10: uint8_t * const half = (uint8_t*)temp; \ yading@10: ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, \ yading@10: 8, stride); \ yading@10: ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half, \ yading@10: stride, stride, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src, \ yading@10: stride, stride); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t temp[8]; \ yading@10: uint8_t * const half = (uint8_t*)temp; \ yading@10: ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src, \ yading@10: 8, stride); \ yading@10: ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,\ yading@10: stride, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[8 + 9]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 64; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ yading@10: ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, \ yading@10: stride, 8, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[8 + 9]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 64; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ yading@10: ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, \ yading@10: stride, 8, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[8 + 9]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 64; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ yading@10: ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, \ yading@10: stride, 8, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[8 + 9]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 64; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ yading@10: ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, \ yading@10: stride, 8, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[8 + 9]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 64; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ yading@10: ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV, \ yading@10: stride, 8, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[8 + 9]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 64; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\ yading@10: ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV, \ yading@10: stride, 8, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[8 + 9]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, \ yading@10: 8, stride, 9); \ yading@10: ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, \ yading@10: stride, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[8 + 9]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, \ yading@10: stride, 9); \ yading@10: ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8, \ yading@10: stride, 9); \ yading@10: ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, \ yading@10: stride, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[9]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8, \ yading@10: stride, 9); \ yading@10: ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH, \ yading@10: stride, 8); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: ff_ ## OPNAME ## pixels16_ ## MMX(dst, src, stride, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t temp[32]; \ yading@10: uint8_t * const half = (uint8_t*)temp; \ yading@10: ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, \ yading@10: stride, 16); \ yading@10: ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, \ yading@10: stride, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: ff_ ## OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src, \ yading@10: stride, stride, 16);\ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t temp[32]; \ yading@10: uint8_t * const half = (uint8_t*)temp; \ yading@10: ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16, \ yading@10: stride, 16); \ yading@10: ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half, \ yading@10: stride, stride, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t temp[32]; \ yading@10: uint8_t * const half = (uint8_t*)temp; \ yading@10: ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, \ yading@10: stride); \ yading@10: ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride, \ yading@10: stride, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src, \ yading@10: stride, stride); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t temp[32]; \ yading@10: uint8_t * const half = (uint8_t*)temp; \ yading@10: ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16, \ yading@10: stride); \ yading@10: ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half, \ yading@10: stride, stride, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[16 * 2 + 17 * 2]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 256; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, \ yading@10: 16, 16); \ yading@10: ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, \ yading@10: stride, 16, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[16 * 2 + 17 * 2]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 256; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, \ yading@10: 16, 16); \ yading@10: ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, \ yading@10: stride, 16, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[16 * 2 + 17 * 2]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 256; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, \ yading@10: 16, 16); \ yading@10: ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, \ yading@10: stride, 16, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[16 * 2 + 17 * 2]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 256; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, \ yading@10: 16, 16); \ yading@10: ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, \ yading@10: stride, 16, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[16 * 2 + 17 * 2]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 256; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, \ yading@10: 16, 16); \ yading@10: ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV, \ yading@10: stride, 16, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[16 * 2 + 17 * 2]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half) + 256; \ yading@10: uint8_t * const halfHV = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH, \ yading@10: 16, 16); \ yading@10: ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV, \ yading@10: stride, 16, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[17 * 2]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16, \ yading@10: stride, 17); \ yading@10: ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, \ yading@10: stride, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[17 * 2]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, \ yading@10: stride, 17); \ yading@10: ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16, \ yading@10: stride, 17); \ yading@10: ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, \ yading@10: stride, 16); \ yading@10: } \ yading@10: \ yading@10: static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: uint64_t half[17 * 2]; \ yading@10: uint8_t * const halfH = ((uint8_t*)half); \ yading@10: ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16, \ yading@10: stride, 17); \ yading@10: ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH, \ yading@10: stride, 16); \ yading@10: } yading@10: yading@10: QPEL_OP(put_, ff_pw_16, _, mmxext) yading@10: QPEL_OP(avg_, ff_pw_16, _, mmxext) yading@10: QPEL_OP(put_no_rnd_, ff_pw_15, _no_rnd_, mmxext) yading@10: #endif /* HAVE_YASM */ yading@10: yading@10: yading@10: #if HAVE_INLINE_ASM yading@10: void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride) yading@10: { yading@10: put_pixels8_xy2_mmx(dst, src, stride, 8); yading@10: } yading@10: void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride) yading@10: { yading@10: put_pixels16_xy2_mmx(dst, src, stride, 16); yading@10: } yading@10: void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride) yading@10: { yading@10: avg_pixels8_xy2_mmx(dst, src, stride, 8); yading@10: } yading@10: void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t stride) yading@10: { yading@10: avg_pixels16_xy2_mmx(dst, src, stride, 16); yading@10: } yading@10: yading@10: typedef void emulated_edge_mc_func(uint8_t *dst, const uint8_t *src, yading@10: ptrdiff_t linesize, int block_w, int block_h, yading@10: int src_x, int src_y, int w, int h); yading@10: yading@10: static av_always_inline void gmc(uint8_t *dst, uint8_t *src, yading@10: int stride, int h, int ox, int oy, yading@10: int dxx, int dxy, int dyx, int dyy, yading@10: int shift, int r, int width, int height, yading@10: emulated_edge_mc_func *emu_edge_fn) yading@10: { yading@10: const int w = 8; yading@10: const int ix = ox >> (16 + shift); yading@10: const int iy = oy >> (16 + shift); yading@10: const int oxs = ox >> 4; yading@10: const int oys = oy >> 4; yading@10: const int dxxs = dxx >> 4; yading@10: const int dxys = dxy >> 4; yading@10: const int dyxs = dyx >> 4; yading@10: const int dyys = dyy >> 4; yading@10: const uint16_t r4[4] = { r, r, r, r }; yading@10: const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys }; yading@10: const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys }; yading@10: const uint64_t shift2 = 2 * shift; yading@10: #define MAX_STRIDE 4096U yading@10: #define MAX_H 8U yading@10: uint8_t edge_buf[(MAX_H + 1) * MAX_STRIDE]; yading@10: int x, y; yading@10: yading@10: const int dxw = (dxx - (1 << (16 + shift))) * (w - 1); yading@10: const int dyh = (dyy - (1 << (16 + shift))) * (h - 1); yading@10: const int dxh = dxy * (h - 1); yading@10: const int dyw = dyx * (w - 1); yading@10: int need_emu = (unsigned)ix >= width - w || yading@10: (unsigned)iy >= height - h; yading@10: yading@10: if ( // non-constant fullpel offset (3% of blocks) yading@10: ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) | yading@10: (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift) yading@10: // uses more than 16 bits of subpel mv (only at huge resolution) yading@10: || (dxx | dxy | dyx | dyy) & 15 yading@10: || (need_emu && (h > MAX_H || stride > MAX_STRIDE))) { yading@10: // FIXME could still use mmx for some of the rows yading@10: ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy, yading@10: shift, r, width, height); yading@10: return; yading@10: } yading@10: yading@10: src += ix + iy * stride; yading@10: if (need_emu) { yading@10: emu_edge_fn(edge_buf, src, stride, w + 1, h + 1, ix, iy, width, height); yading@10: src = edge_buf; yading@10: } yading@10: yading@10: __asm__ volatile ( yading@10: "movd %0, %%mm6 \n\t" yading@10: "pxor %%mm7, %%mm7 \n\t" yading@10: "punpcklwd %%mm6, %%mm6 \n\t" yading@10: "punpcklwd %%mm6, %%mm6 \n\t" yading@10: :: "r"(1<PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \ yading@10: c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \ yading@10: } while (0) yading@10: yading@10: static av_cold void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx, yading@10: int mm_flags) yading@10: { yading@10: const int high_bit_depth = avctx->bits_per_raw_sample > 8; yading@10: yading@10: #if HAVE_INLINE_ASM yading@10: c->put_pixels_clamped = ff_put_pixels_clamped_mmx; yading@10: c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx; yading@10: c->add_pixels_clamped = ff_add_pixels_clamped_mmx; yading@10: yading@10: if (!high_bit_depth) { yading@10: c->clear_block = clear_block_mmx; yading@10: c->clear_blocks = clear_blocks_mmx; yading@10: c->draw_edges = draw_edges_mmx; yading@10: } yading@10: yading@10: #if CONFIG_VIDEODSP && (ARCH_X86_32 || !HAVE_YASM) yading@10: c->gmc = gmc_mmx; yading@10: #endif yading@10: yading@10: c->add_bytes = add_bytes_mmx; yading@10: #endif /* HAVE_INLINE_ASM */ yading@10: yading@10: #if HAVE_YASM yading@10: if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) { yading@10: c->h263_v_loop_filter = ff_h263_v_loop_filter_mmx; yading@10: c->h263_h_loop_filter = ff_h263_h_loop_filter_mmx; yading@10: } yading@10: yading@10: c->vector_clip_int32 = ff_vector_clip_int32_mmx; yading@10: #endif yading@10: yading@10: } yading@10: yading@10: static av_cold void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx, yading@10: int mm_flags) yading@10: { yading@10: yading@10: #if HAVE_YASM yading@10: SET_QPEL_FUNCS(avg_qpel, 0, 16, mmxext, ); yading@10: SET_QPEL_FUNCS(avg_qpel, 1, 8, mmxext, ); yading@10: yading@10: SET_QPEL_FUNCS(put_qpel, 0, 16, mmxext, ); yading@10: SET_QPEL_FUNCS(put_qpel, 1, 8, mmxext, ); yading@10: SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, ); yading@10: SET_QPEL_FUNCS(put_no_rnd_qpel, 1, 8, mmxext, ); yading@10: #endif /* HAVE_YASM */ yading@10: yading@10: #if HAVE_MMXEXT_EXTERNAL yading@10: /* slower than cmov version on AMD */ yading@10: if (!(mm_flags & AV_CPU_FLAG_3DNOW)) yading@10: c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmxext; yading@10: yading@10: c->scalarproduct_int16 = ff_scalarproduct_int16_mmxext; yading@10: c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext; yading@10: yading@10: if (avctx->flags & CODEC_FLAG_BITEXACT) { yading@10: c->apply_window_int16 = ff_apply_window_int16_mmxext; yading@10: } else { yading@10: c->apply_window_int16 = ff_apply_window_int16_round_mmxext; yading@10: } yading@10: #endif /* HAVE_MMXEXT_EXTERNAL */ yading@10: } yading@10: yading@10: static av_cold void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx, yading@10: int mm_flags) yading@10: { yading@10: #if HAVE_INLINE_ASM yading@10: const int high_bit_depth = avctx->bits_per_raw_sample > 8; yading@10: yading@10: if (!high_bit_depth) { yading@10: if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) { yading@10: /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */ yading@10: c->clear_block = clear_block_sse; yading@10: c->clear_blocks = clear_blocks_sse; yading@10: } yading@10: } yading@10: yading@10: c->vector_clipf = vector_clipf_sse; yading@10: #endif /* HAVE_INLINE_ASM */ yading@10: yading@10: #if HAVE_YASM yading@10: #if HAVE_INLINE_ASM && CONFIG_VIDEODSP yading@10: c->gmc = gmc_sse; yading@10: #endif yading@10: #endif /* HAVE_YASM */ yading@10: } yading@10: yading@10: static av_cold void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx, yading@10: int mm_flags) yading@10: { yading@10: const int high_bit_depth = avctx->bits_per_raw_sample > 8; yading@10: yading@10: #if HAVE_SSE2_INLINE yading@10: if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) { yading@10: c->idct_put = ff_idct_xvid_sse2_put; yading@10: c->idct_add = ff_idct_xvid_sse2_add; yading@10: c->idct = ff_idct_xvid_sse2; yading@10: c->idct_permutation_type = FF_SSE2_IDCT_PERM; yading@10: } yading@10: #endif /* HAVE_SSE2_INLINE */ yading@10: yading@10: #if HAVE_SSE2_EXTERNAL yading@10: c->scalarproduct_int16 = ff_scalarproduct_int16_sse2; yading@10: c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2; yading@10: if (mm_flags & AV_CPU_FLAG_ATOM) { yading@10: c->vector_clip_int32 = ff_vector_clip_int32_int_sse2; yading@10: } else { yading@10: c->vector_clip_int32 = ff_vector_clip_int32_sse2; yading@10: } yading@10: if (avctx->flags & CODEC_FLAG_BITEXACT) { yading@10: c->apply_window_int16 = ff_apply_window_int16_sse2; yading@10: } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) { yading@10: c->apply_window_int16 = ff_apply_window_int16_round_sse2; yading@10: } yading@10: c->bswap_buf = ff_bswap32_buf_sse2; yading@10: #endif /* HAVE_SSE2_EXTERNAL */ yading@10: } yading@10: yading@10: static av_cold void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx, yading@10: int mm_flags) yading@10: { yading@10: #if HAVE_SSSE3_EXTERNAL yading@10: c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3; yading@10: if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe yading@10: c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4; yading@10: yading@10: if (mm_flags & AV_CPU_FLAG_ATOM) yading@10: c->apply_window_int16 = ff_apply_window_int16_ssse3_atom; yading@10: else yading@10: c->apply_window_int16 = ff_apply_window_int16_ssse3; yading@10: if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit yading@10: c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3; yading@10: c->bswap_buf = ff_bswap32_buf_ssse3; yading@10: #endif /* HAVE_SSSE3_EXTERNAL */ yading@10: } yading@10: yading@10: static av_cold void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx, yading@10: int mm_flags) yading@10: { yading@10: #if HAVE_SSE4_EXTERNAL yading@10: c->vector_clip_int32 = ff_vector_clip_int32_sse4; yading@10: #endif /* HAVE_SSE4_EXTERNAL */ yading@10: } yading@10: yading@10: av_cold void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx) yading@10: { yading@10: int mm_flags = av_get_cpu_flags(); yading@10: yading@10: #if HAVE_7REGS && HAVE_INLINE_ASM yading@10: if (mm_flags & AV_CPU_FLAG_CMOV) yading@10: c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov; yading@10: #endif yading@10: yading@10: if (mm_flags & AV_CPU_FLAG_MMX) { yading@10: #if HAVE_INLINE_ASM yading@10: const int idct_algo = avctx->idct_algo; yading@10: yading@10: if (avctx->lowres == 0 && avctx->bits_per_raw_sample <= 8) { yading@10: if (idct_algo == FF_IDCT_AUTO || idct_algo == FF_IDCT_SIMPLEMMX) { yading@10: c->idct_put = ff_simple_idct_put_mmx; yading@10: c->idct_add = ff_simple_idct_add_mmx; yading@10: c->idct = ff_simple_idct_mmx; yading@10: c->idct_permutation_type = FF_SIMPLE_IDCT_PERM; yading@10: } else if (idct_algo == FF_IDCT_XVIDMMX) { yading@10: if (mm_flags & AV_CPU_FLAG_SSE2) { yading@10: c->idct_put = ff_idct_xvid_sse2_put; yading@10: c->idct_add = ff_idct_xvid_sse2_add; yading@10: c->idct = ff_idct_xvid_sse2; yading@10: c->idct_permutation_type = FF_SSE2_IDCT_PERM; yading@10: } else if (mm_flags & AV_CPU_FLAG_MMXEXT) { yading@10: c->idct_put = ff_idct_xvid_mmxext_put; yading@10: c->idct_add = ff_idct_xvid_mmxext_add; yading@10: c->idct = ff_idct_xvid_mmxext; yading@10: } else { yading@10: c->idct_put = ff_idct_xvid_mmx_put; yading@10: c->idct_add = ff_idct_xvid_mmx_add; yading@10: c->idct = ff_idct_xvid_mmx; yading@10: } yading@10: } yading@10: } yading@10: #endif /* HAVE_INLINE_ASM */ yading@10: yading@10: dsputil_init_mmx(c, avctx, mm_flags); yading@10: } yading@10: yading@10: if (mm_flags & AV_CPU_FLAG_MMXEXT) yading@10: dsputil_init_mmxext(c, avctx, mm_flags); yading@10: yading@10: if (mm_flags & AV_CPU_FLAG_SSE) yading@10: dsputil_init_sse(c, avctx, mm_flags); yading@10: yading@10: if (mm_flags & AV_CPU_FLAG_SSE2) yading@10: dsputil_init_sse2(c, avctx, mm_flags); yading@10: yading@10: if (mm_flags & AV_CPU_FLAG_SSSE3) yading@10: dsputil_init_ssse3(c, avctx, mm_flags); yading@10: yading@10: if (mm_flags & AV_CPU_FLAG_SSE4) yading@10: dsputil_init_sse4(c, avctx, mm_flags); yading@10: yading@10: if (CONFIG_ENCODERS) yading@10: ff_dsputilenc_init_mmx(c, avctx); yading@10: }