yading@10: /* yading@10: * RV40 decoder motion compensation functions x86-optimised yading@10: * Copyright (c) 2008 Konstantin Shishkov 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: yading@10: /** yading@10: * @file yading@10: * RV40 decoder motion compensation functions x86-optimised yading@10: * 2,0 and 0,2 have h264 equivalents. yading@10: * 3,3 is bugged in the rv40 format and maps to _xy2 version yading@10: */ yading@10: yading@10: #include "libavcodec/rv34dsp.h" yading@10: #include "libavutil/attributes.h" yading@10: #include "libavutil/mem.h" yading@10: #include "libavutil/x86/cpu.h" yading@10: #include "dsputil_mmx.h" yading@10: yading@10: #if HAVE_YASM yading@10: void ff_put_rv40_chroma_mc8_mmx (uint8_t *dst, uint8_t *src, yading@10: int stride, int h, int x, int y); yading@10: void ff_avg_rv40_chroma_mc8_mmxext(uint8_t *dst, uint8_t *src, yading@10: int stride, int h, int x, int y); yading@10: void ff_avg_rv40_chroma_mc8_3dnow(uint8_t *dst, uint8_t *src, yading@10: int stride, int h, int x, int y); yading@10: yading@10: void ff_put_rv40_chroma_mc4_mmx (uint8_t *dst, uint8_t *src, yading@10: int stride, int h, int x, int y); yading@10: void ff_avg_rv40_chroma_mc4_mmxext(uint8_t *dst, uint8_t *src, yading@10: int stride, int h, int x, int y); yading@10: void ff_avg_rv40_chroma_mc4_3dnow(uint8_t *dst, uint8_t *src, yading@10: int stride, int h, int x, int y); yading@10: yading@10: #define DECLARE_WEIGHT(opt) \ yading@10: void ff_rv40_weight_func_rnd_16_##opt(uint8_t *dst, uint8_t *src1, uint8_t *src2, \ yading@10: int w1, int w2, ptrdiff_t stride); \ yading@10: void ff_rv40_weight_func_rnd_8_##opt (uint8_t *dst, uint8_t *src1, uint8_t *src2, \ yading@10: int w1, int w2, ptrdiff_t stride); \ yading@10: void ff_rv40_weight_func_nornd_16_##opt(uint8_t *dst, uint8_t *src1, uint8_t *src2, \ yading@10: int w1, int w2, ptrdiff_t stride); \ yading@10: void ff_rv40_weight_func_nornd_8_##opt (uint8_t *dst, uint8_t *src1, uint8_t *src2, \ yading@10: int w1, int w2, ptrdiff_t stride); yading@10: DECLARE_WEIGHT(mmxext) yading@10: DECLARE_WEIGHT(sse2) yading@10: DECLARE_WEIGHT(ssse3) yading@10: yading@10: /** @{ */ yading@10: /** yading@10: * Define one qpel function. yading@10: * LOOPSIZE must be already set to the number of pixels processed per yading@10: * iteration in the inner loop of the called functions. yading@10: * COFF(x) must be already defined so as to provide the offset into any yading@10: * array of coeffs used by the called function for the qpel position x. yading@10: */ yading@10: #define QPEL_FUNC_DECL(OP, SIZE, PH, PV, OPT) \ yading@10: static void OP ## rv40_qpel ##SIZE ##_mc ##PH ##PV ##OPT(uint8_t *dst, \ yading@10: uint8_t *src, \ yading@10: ptrdiff_t stride) \ yading@10: { \ yading@10: int i; \ yading@10: if (PH && PV) { \ yading@10: DECLARE_ALIGNED(16, uint8_t, tmp)[SIZE * (SIZE + 5)]; \ yading@10: uint8_t *tmpptr = tmp + SIZE * 2; \ yading@10: src -= stride * 2; \ yading@10: \ yading@10: for (i = 0; i < SIZE; i += LOOPSIZE) \ yading@10: ff_put_rv40_qpel_h ##OPT(tmp + i, SIZE, src + i, stride, \ yading@10: SIZE + 5, HCOFF(PH)); \ yading@10: for (i = 0; i < SIZE; i += LOOPSIZE) \ yading@10: ff_ ##OP ##rv40_qpel_v ##OPT(dst + i, stride, tmpptr + i, \ yading@10: SIZE, SIZE, VCOFF(PV)); \ yading@10: } else if (PV) { \ yading@10: for (i = 0; i < SIZE; i += LOOPSIZE) \ yading@10: ff_ ##OP ##rv40_qpel_v ## OPT(dst + i, stride, src + i, \ yading@10: stride, SIZE, VCOFF(PV)); \ yading@10: } else { \ yading@10: for (i = 0; i < SIZE; i += LOOPSIZE) \ yading@10: ff_ ##OP ##rv40_qpel_h ## OPT(dst + i, stride, src + i, \ yading@10: stride, SIZE, HCOFF(PH)); \ yading@10: } \ yading@10: }; yading@10: yading@10: /** Declare functions for sizes 8 and 16 and given operations yading@10: * and qpel position. */ yading@10: #define QPEL_FUNCS_DECL(OP, PH, PV, OPT) \ yading@10: QPEL_FUNC_DECL(OP, 8, PH, PV, OPT) \ yading@10: QPEL_FUNC_DECL(OP, 16, PH, PV, OPT) yading@10: yading@10: /** Declare all functions for all sizes and qpel positions */ yading@10: #define QPEL_MC_DECL(OP, OPT) \ yading@10: void ff_ ##OP ##rv40_qpel_h ##OPT(uint8_t *dst, ptrdiff_t dstStride, \ yading@10: const uint8_t *src, \ yading@10: ptrdiff_t srcStride, \ yading@10: int len, int m); \ yading@10: void ff_ ##OP ##rv40_qpel_v ##OPT(uint8_t *dst, ptrdiff_t dstStride, \ yading@10: const uint8_t *src, \ yading@10: ptrdiff_t srcStride, \ yading@10: int len, int m); \ yading@10: QPEL_FUNCS_DECL(OP, 0, 1, OPT) \ yading@10: QPEL_FUNCS_DECL(OP, 0, 3, OPT) \ yading@10: QPEL_FUNCS_DECL(OP, 1, 0, OPT) \ yading@10: QPEL_FUNCS_DECL(OP, 1, 1, OPT) \ yading@10: QPEL_FUNCS_DECL(OP, 1, 2, OPT) \ yading@10: QPEL_FUNCS_DECL(OP, 1, 3, OPT) \ yading@10: QPEL_FUNCS_DECL(OP, 2, 1, OPT) \ yading@10: QPEL_FUNCS_DECL(OP, 2, 2, OPT) \ yading@10: QPEL_FUNCS_DECL(OP, 2, 3, OPT) \ yading@10: QPEL_FUNCS_DECL(OP, 3, 0, OPT) \ yading@10: QPEL_FUNCS_DECL(OP, 3, 1, OPT) \ yading@10: QPEL_FUNCS_DECL(OP, 3, 2, OPT) yading@10: /** @} */ yading@10: yading@10: #define LOOPSIZE 8 yading@10: #define HCOFF(x) (32 * (x - 1)) yading@10: #define VCOFF(x) (32 * (x - 1)) yading@10: QPEL_MC_DECL(put_, _ssse3) yading@10: QPEL_MC_DECL(avg_, _ssse3) yading@10: yading@10: #undef LOOPSIZE yading@10: #undef HCOFF yading@10: #undef VCOFF yading@10: #define LOOPSIZE 8 yading@10: #define HCOFF(x) (64 * (x - 1)) yading@10: #define VCOFF(x) (64 * (x - 1)) yading@10: QPEL_MC_DECL(put_, _sse2) yading@10: QPEL_MC_DECL(avg_, _sse2) yading@10: yading@10: #if ARCH_X86_32 yading@10: #undef LOOPSIZE yading@10: #undef HCOFF yading@10: #undef VCOFF yading@10: #define LOOPSIZE 4 yading@10: #define HCOFF(x) (64 * (x - 1)) yading@10: #define VCOFF(x) (64 * (x - 1)) yading@10: yading@10: QPEL_MC_DECL(put_, _mmx) yading@10: yading@10: #define ff_put_rv40_qpel_h_mmxext ff_put_rv40_qpel_h_mmx yading@10: #define ff_put_rv40_qpel_v_mmxext ff_put_rv40_qpel_v_mmx yading@10: QPEL_MC_DECL(avg_, _mmxext) yading@10: yading@10: #define ff_put_rv40_qpel_h_3dnow ff_put_rv40_qpel_h_mmx yading@10: #define ff_put_rv40_qpel_v_3dnow ff_put_rv40_qpel_v_mmx yading@10: QPEL_MC_DECL(avg_, _3dnow) yading@10: #endif yading@10: yading@10: /** @{ */ yading@10: /** Set one function */ yading@10: #define QPEL_FUNC_SET(OP, SIZE, PH, PV, OPT) \ yading@10: c-> OP ## pixels_tab[2 - SIZE / 8][4 * PV + PH] = OP ## rv40_qpel ##SIZE ## _mc ##PH ##PV ##OPT; yading@10: yading@10: /** Set functions put and avg for sizes 8 and 16 and a given qpel position */ yading@10: #define QPEL_FUNCS_SET(OP, PH, PV, OPT) \ yading@10: QPEL_FUNC_SET(OP, 8, PH, PV, OPT) \ yading@10: QPEL_FUNC_SET(OP, 16, PH, PV, OPT) yading@10: yading@10: /** Set all functions for all sizes and qpel positions */ yading@10: #define QPEL_MC_SET(OP, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 0, 1, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 0, 3, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 1, 0, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 1, 1, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 1, 2, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 1, 3, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 2, 1, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 2, 2, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 2, 3, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 3, 0, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 3, 1, OPT) \ yading@10: QPEL_FUNCS_SET (OP, 3, 2, OPT) yading@10: /** @} */ yading@10: yading@10: #endif /* HAVE_YASM */ yading@10: yading@10: av_cold void ff_rv40dsp_init_x86(RV34DSPContext *c) yading@10: { yading@10: #if HAVE_YASM yading@10: int mm_flags = av_get_cpu_flags(); yading@10: yading@10: if (EXTERNAL_MMX(mm_flags)) { yading@10: c->put_chroma_pixels_tab[0] = ff_put_rv40_chroma_mc8_mmx; yading@10: c->put_chroma_pixels_tab[1] = ff_put_rv40_chroma_mc4_mmx; yading@10: #if HAVE_MMX_INLINE yading@10: c->put_pixels_tab[0][15] = ff_put_rv40_qpel16_mc33_mmx; yading@10: c->put_pixels_tab[1][15] = ff_put_rv40_qpel8_mc33_mmx; yading@10: c->avg_pixels_tab[0][15] = ff_avg_rv40_qpel16_mc33_mmx; yading@10: c->avg_pixels_tab[1][15] = ff_avg_rv40_qpel8_mc33_mmx; yading@10: #endif /* HAVE_MMX_INLINE */ yading@10: #if ARCH_X86_32 yading@10: QPEL_MC_SET(put_, _mmx) yading@10: #endif yading@10: } yading@10: if (EXTERNAL_MMXEXT(mm_flags)) { yading@10: c->avg_chroma_pixels_tab[0] = ff_avg_rv40_chroma_mc8_mmxext; yading@10: c->avg_chroma_pixels_tab[1] = ff_avg_rv40_chroma_mc4_mmxext; yading@10: c->rv40_weight_pixels_tab[0][0] = ff_rv40_weight_func_rnd_16_mmxext; yading@10: c->rv40_weight_pixels_tab[0][1] = ff_rv40_weight_func_rnd_8_mmxext; yading@10: c->rv40_weight_pixels_tab[1][0] = ff_rv40_weight_func_nornd_16_mmxext; yading@10: c->rv40_weight_pixels_tab[1][1] = ff_rv40_weight_func_nornd_8_mmxext; yading@10: #if ARCH_X86_32 yading@10: QPEL_MC_SET(avg_, _mmxext) yading@10: #endif yading@10: } else if (EXTERNAL_AMD3DNOW(mm_flags)) { yading@10: c->avg_chroma_pixels_tab[0] = ff_avg_rv40_chroma_mc8_3dnow; yading@10: c->avg_chroma_pixels_tab[1] = ff_avg_rv40_chroma_mc4_3dnow; yading@10: #if ARCH_X86_32 yading@10: QPEL_MC_SET(avg_, _3dnow) yading@10: #endif yading@10: } yading@10: if (EXTERNAL_SSE2(mm_flags)) { yading@10: c->rv40_weight_pixels_tab[0][0] = ff_rv40_weight_func_rnd_16_sse2; yading@10: c->rv40_weight_pixels_tab[0][1] = ff_rv40_weight_func_rnd_8_sse2; yading@10: c->rv40_weight_pixels_tab[1][0] = ff_rv40_weight_func_nornd_16_sse2; yading@10: c->rv40_weight_pixels_tab[1][1] = ff_rv40_weight_func_nornd_8_sse2; yading@10: QPEL_MC_SET(put_, _sse2) yading@10: QPEL_MC_SET(avg_, _sse2) yading@10: } yading@10: if (EXTERNAL_SSSE3(mm_flags)) { yading@10: c->rv40_weight_pixels_tab[0][0] = ff_rv40_weight_func_rnd_16_ssse3; yading@10: c->rv40_weight_pixels_tab[0][1] = ff_rv40_weight_func_rnd_8_ssse3; yading@10: c->rv40_weight_pixels_tab[1][0] = ff_rv40_weight_func_nornd_16_ssse3; yading@10: c->rv40_weight_pixels_tab[1][1] = ff_rv40_weight_func_nornd_8_ssse3; yading@10: QPEL_MC_SET(put_, _ssse3) yading@10: QPEL_MC_SET(avg_, _ssse3) yading@10: } yading@10: #endif /* HAVE_YASM */ yading@10: }