yading@10: /* yading@10: * SPARC VIS optimized inverse DCT yading@10: * Copyright (c) 2007 Denes Balatoni < dbalatoni XatX interware XdotX hu > yading@10: * yading@10: * I did consult the following fine web page about dct yading@10: * http://www.geocities.com/ssavekar/dct.htm 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: #include yading@10: yading@10: #include "dsputil_vis.h" yading@10: #include "libavutil/mem.h" yading@10: yading@10: static const DECLARE_ALIGNED(8, int16_t, coeffs)[28] = { yading@10: - 1259,- 1259,- 1259,- 1259, yading@10: - 4989,- 4989,- 4989,- 4989, yading@10: -11045,-11045,-11045,-11045, yading@10: -19195,-19195,-19195,-19195, yading@10: -29126,-29126,-29126,-29126, yading@10: 25080, 25080, 25080, 25080, yading@10: 12785, 12785, 12785, 12785 yading@10: }; yading@10: static const DECLARE_ALIGNED(8, uint16_t, scale)[4] = { yading@10: 65536>>6, 65536>>6, 65536>>6, 65536>>6 yading@10: }; yading@10: static const DECLARE_ALIGNED(8, uint16_t, rounder)[4] = { yading@10: 1<<5, 1<<5, 1<<5, 1<<5 yading@10: }; yading@10: static const DECLARE_ALIGNED(8, uint16_t, expand)[4] = { yading@10: 1<<14, 1<<14, 1<<14, 1<<14 yading@10: }; yading@10: yading@10: #define INIT_IDCT \ yading@10: "ldd [%1], %%f32 \n\t"\ yading@10: "ldd [%1+8], %%f34 \n\t"\ yading@10: "ldd [%1+16], %%f36 \n\t"\ yading@10: "ldd [%1+24], %%f38 \n\t"\ yading@10: "ldd [%1+32], %%f40 \n\t"\ yading@10: "ldd [%1+40], %%f42 \n\t"\ yading@10: "ldd [%1+48], %%f44 \n\t"\ yading@10: "ldd [%0], %%f46 \n\t"\ yading@10: "fzero %%f62 \n\t"\ yading@10: yading@10: #define LOADSCALE(in) \ yading@10: "ldd [" in "], %%f0 \n\t"\ yading@10: "ldd [" in "+16], %%f2 \n\t"\ yading@10: "ldd [" in "+32], %%f4 \n\t"\ yading@10: "ldd [" in "+48], %%f6 \n\t"\ yading@10: "ldd [" in "+64], %%f8 \n\t"\ yading@10: "ldd [" in "+80], %%f10 \n\t"\ yading@10: "ldd [" in "+96], %%f12 \n\t"\ yading@10: "ldd [" in "+112], %%f14 \n\t"\ yading@10: "fpadd16 %%f0, %%f0, %%f0 \n\t"\ yading@10: "fpadd16 %%f2, %%f2, %%f2 \n\t"\ yading@10: "fpadd16 %%f4, %%f4, %%f4 \n\t"\ yading@10: "fpadd16 %%f6, %%f6, %%f6 \n\t"\ yading@10: "fpadd16 %%f8, %%f8, %%f8 \n\t"\ yading@10: "fpadd16 %%f10, %%f10, %%f10 \n\t"\ yading@10: "fpadd16 %%f12, %%f12, %%f12 \n\t"\ yading@10: "fpadd16 %%f14, %%f14, %%f14 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f0, %%f0, %%f0 \n\t"\ yading@10: "fpadd16 %%f2, %%f2, %%f2 \n\t"\ yading@10: "fpadd16 %%f4, %%f4, %%f4 \n\t"\ yading@10: "fpadd16 %%f6, %%f6, %%f6 \n\t"\ yading@10: "fpadd16 %%f8, %%f8, %%f8 \n\t"\ yading@10: "fpadd16 %%f10, %%f10, %%f10 \n\t"\ yading@10: "fpadd16 %%f12, %%f12, %%f12 \n\t"\ yading@10: "fpadd16 %%f14, %%f14, %%f14 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f0, %%f0, %%f0 \n\t"\ yading@10: "fpadd16 %%f2, %%f2, %%f2 \n\t"\ yading@10: "fpadd16 %%f4, %%f4, %%f4 \n\t"\ yading@10: "fpadd16 %%f6, %%f6, %%f6 \n\t"\ yading@10: "fpadd16 %%f8, %%f8, %%f8 \n\t"\ yading@10: "fpadd16 %%f10, %%f10, %%f10 \n\t"\ yading@10: "fpadd16 %%f12, %%f12, %%f12 \n\t"\ yading@10: "fpadd16 %%f14, %%f14, %%f14 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f0, %%f0, %%f0 \n\t"\ yading@10: "fpadd16 %%f2, %%f2, %%f2 \n\t"\ yading@10: "fpadd16 %%f4, %%f4, %%f4 \n\t"\ yading@10: "fpadd16 %%f6, %%f6, %%f6 \n\t"\ yading@10: "fpadd16 %%f8, %%f8, %%f8 \n\t"\ yading@10: "fpadd16 %%f10, %%f10, %%f10 \n\t"\ yading@10: "fpadd16 %%f12, %%f12, %%f12 \n\t"\ yading@10: "fpadd16 %%f14, %%f14, %%f14 \n\t"\ yading@10: yading@10: #define LOAD(in) \ yading@10: "ldd [" in "], %%f16 \n\t"\ yading@10: "ldd [" in "+8], %%f18 \n\t"\ yading@10: "ldd [" in "+16], %%f20 \n\t"\ yading@10: "ldd [" in "+24], %%f22 \n\t"\ yading@10: "ldd [" in "+32], %%f24 \n\t"\ yading@10: "ldd [" in "+40], %%f26 \n\t"\ yading@10: "ldd [" in "+48], %%f28 \n\t"\ yading@10: "ldd [" in "+56], %%f30 \n\t"\ yading@10: yading@10: #define TRANSPOSE \ yading@10: "fpmerge %%f16, %%f24, %%f0 \n\t"\ yading@10: "fpmerge %%f20, %%f28, %%f2 \n\t"\ yading@10: "fpmerge %%f17, %%f25, %%f4 \n\t"\ yading@10: "fpmerge %%f21, %%f29, %%f6 \n\t"\ yading@10: "fpmerge %%f18, %%f26, %%f8 \n\t"\ yading@10: "fpmerge %%f22, %%f30, %%f10 \n\t"\ yading@10: "fpmerge %%f19, %%f27, %%f12 \n\t"\ yading@10: "fpmerge %%f23, %%f31, %%f14 \n\t"\ yading@10: \ yading@10: "fpmerge %%f0, %%f2, %%f16 \n\t"\ yading@10: "fpmerge %%f1, %%f3, %%f18 \n\t"\ yading@10: "fpmerge %%f4, %%f6, %%f20 \n\t"\ yading@10: "fpmerge %%f5, %%f7, %%f22 \n\t"\ yading@10: "fpmerge %%f8, %%f10, %%f24 \n\t"\ yading@10: "fpmerge %%f9, %%f11, %%f26 \n\t"\ yading@10: "fpmerge %%f12, %%f14, %%f28 \n\t"\ yading@10: "fpmerge %%f13, %%f15, %%f30 \n\t"\ yading@10: \ yading@10: "fpmerge %%f16, %%f17, %%f0 \n\t"\ yading@10: "fpmerge %%f18, %%f19, %%f2 \n\t"\ yading@10: "fpmerge %%f20, %%f21, %%f4 \n\t"\ yading@10: "fpmerge %%f22, %%f23, %%f6 \n\t"\ yading@10: "fpmerge %%f24, %%f25, %%f8 \n\t"\ yading@10: "fpmerge %%f26, %%f27, %%f10 \n\t"\ yading@10: "fpmerge %%f28, %%f29, %%f12 \n\t"\ yading@10: "fpmerge %%f30, %%f31, %%f14 \n\t"\ yading@10: yading@10: #define IDCT4ROWS \ yading@10: /* 1. column */\ yading@10: "fmul8ulx16 %%f0, %%f38, %%f28 \n\t"\ yading@10: "for %%f4, %%f6, %%f60 \n\t"\ yading@10: "fmul8ulx16 %%f2, %%f32, %%f18 \n\t"\ yading@10: "fmul8ulx16 %%f2, %%f36, %%f22 \n\t"\ yading@10: "fmul8ulx16 %%f2, %%f40, %%f26 \n\t"\ yading@10: "fmul8ulx16 %%f2, %%f44, %%f30 \n\t"\ yading@10: \ yading@10: ADDROUNDER\ yading@10: \ yading@10: "fmul8sux16 %%f0, %%f38, %%f48 \n\t"\ yading@10: "fcmpd %%fcc0, %%f62, %%f60 \n\t"\ yading@10: "for %%f8, %%f10, %%f60 \n\t"\ yading@10: "fmul8sux16 %%f2, %%f32, %%f50 \n\t"\ yading@10: "fmul8sux16 %%f2, %%f36, %%f52 \n\t"\ yading@10: "fmul8sux16 %%f2, %%f40, %%f54 \n\t"\ yading@10: "fmul8sux16 %%f2, %%f44, %%f56 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f48, %%f28, %%f28 \n\t"\ yading@10: "fcmpd %%fcc1, %%f62, %%f60 \n\t"\ yading@10: "for %%f12, %%f14, %%f60 \n\t"\ yading@10: "fpadd16 %%f50, %%f18, %%f18 \n\t"\ yading@10: "fpadd16 %%f52, %%f22, %%f22 \n\t"\ yading@10: "fpadd16 %%f54, %%f26, %%f26 \n\t"\ yading@10: "fpadd16 %%f56, %%f30, %%f30 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f28, %%f0, %%f16 \n\t"\ yading@10: "fcmpd %%fcc2, %%f62, %%f60 \n\t"\ yading@10: "fpadd16 %%f28, %%f0, %%f20 \n\t"\ yading@10: "fpadd16 %%f28, %%f0, %%f24 \n\t"\ yading@10: "fpadd16 %%f28, %%f0, %%f28 \n\t"\ yading@10: "fpadd16 %%f18, %%f2, %%f18 \n\t"\ yading@10: "fpadd16 %%f22, %%f2, %%f22 \n\t"\ yading@10: /* 2. column */\ yading@10: "fbe %%fcc0, 3f \n\t"\ yading@10: "fpadd16 %%f26, %%f2, %%f26 \n\t"\ yading@10: "fmul8ulx16 %%f4, %%f34, %%f48 \n\t"\ yading@10: "fmul8ulx16 %%f4, %%f42, %%f50 \n\t"\ yading@10: "fmul8ulx16 %%f6, %%f36, %%f52 \n\t"\ yading@10: "fmul8ulx16 %%f6, %%f44, %%f54 \n\t"\ yading@10: "fmul8ulx16 %%f6, %%f32, %%f56 \n\t"\ yading@10: "fmul8ulx16 %%f6, %%f40, %%f58 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f16, %%f48, %%f16 \n\t"\ yading@10: "fpadd16 %%f20, %%f50, %%f20 \n\t"\ yading@10: "fpsub16 %%f24, %%f50, %%f24 \n\t"\ yading@10: "fpsub16 %%f28, %%f48, %%f28 \n\t"\ yading@10: "fpadd16 %%f18, %%f52, %%f18 \n\t"\ yading@10: "fpsub16 %%f22, %%f54, %%f22 \n\t"\ yading@10: "fpsub16 %%f26, %%f56, %%f26 \n\t"\ yading@10: "fpsub16 %%f30, %%f58, %%f30 \n\t"\ yading@10: \ yading@10: "fmul8sux16 %%f4, %%f34, %%f48 \n\t"\ yading@10: "fmul8sux16 %%f4, %%f42, %%f50 \n\t"\ yading@10: "fmul8sux16 %%f6, %%f36, %%f52 \n\t"\ yading@10: "fmul8sux16 %%f6, %%f44, %%f54 \n\t"\ yading@10: "fmul8sux16 %%f6, %%f32, %%f56 \n\t"\ yading@10: "fmul8sux16 %%f6, %%f40, %%f58 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f16, %%f48, %%f16 \n\t"\ yading@10: "fpadd16 %%f20, %%f50, %%f20 \n\t"\ yading@10: "fpsub16 %%f24, %%f50, %%f24 \n\t"\ yading@10: "fpsub16 %%f28, %%f48, %%f28 \n\t"\ yading@10: "fpadd16 %%f18, %%f52, %%f18 \n\t"\ yading@10: "fpsub16 %%f22, %%f54, %%f22 \n\t"\ yading@10: "fpsub16 %%f26, %%f56, %%f26 \n\t"\ yading@10: "fpsub16 %%f30, %%f58, %%f30 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f16, %%f4, %%f16 \n\t"\ yading@10: "fpsub16 %%f28, %%f4, %%f28 \n\t"\ yading@10: "fpadd16 %%f18, %%f6, %%f18 \n\t"\ yading@10: "fpsub16 %%f26, %%f6, %%f26 \n\t"\ yading@10: /* 3. column */\ yading@10: "3: \n\t"\ yading@10: "fbe %%fcc1, 4f \n\t"\ yading@10: "fpsub16 %%f30, %%f6, %%f30 \n\t"\ yading@10: "fmul8ulx16 %%f8, %%f38, %%f48 \n\t"\ yading@10: "fmul8ulx16 %%f10, %%f40, %%f50 \n\t"\ yading@10: "fmul8ulx16 %%f10, %%f32, %%f52 \n\t"\ yading@10: "fmul8ulx16 %%f10, %%f44, %%f54 \n\t"\ yading@10: "fmul8ulx16 %%f10, %%f36, %%f56 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f16, %%f48, %%f16 \n\t"\ yading@10: "fpsub16 %%f20, %%f48, %%f20 \n\t"\ yading@10: "fpsub16 %%f24, %%f48, %%f24 \n\t"\ yading@10: "fpadd16 %%f28, %%f48, %%f28 \n\t"\ yading@10: "fpadd16 %%f18, %%f50, %%f18 \n\t"\ yading@10: "fpsub16 %%f22, %%f52, %%f22 \n\t"\ yading@10: "fpadd16 %%f26, %%f54, %%f26 \n\t"\ yading@10: "fpadd16 %%f30, %%f56, %%f30 \n\t"\ yading@10: \ yading@10: "fmul8sux16 %%f8, %%f38, %%f48 \n\t"\ yading@10: "fmul8sux16 %%f10, %%f40, %%f50 \n\t"\ yading@10: "fmul8sux16 %%f10, %%f32, %%f52 \n\t"\ yading@10: "fmul8sux16 %%f10, %%f44, %%f54 \n\t"\ yading@10: "fmul8sux16 %%f10, %%f36, %%f56 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f16, %%f48, %%f16 \n\t"\ yading@10: "fpsub16 %%f20, %%f48, %%f20 \n\t"\ yading@10: "fpsub16 %%f24, %%f48, %%f24 \n\t"\ yading@10: "fpadd16 %%f28, %%f48, %%f28 \n\t"\ yading@10: "fpadd16 %%f18, %%f50, %%f18 \n\t"\ yading@10: "fpsub16 %%f22, %%f52, %%f22 \n\t"\ yading@10: "fpadd16 %%f26, %%f54, %%f26 \n\t"\ yading@10: "fpadd16 %%f30, %%f56, %%f30 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f16, %%f8, %%f16 \n\t"\ yading@10: "fpsub16 %%f20, %%f8, %%f20 \n\t"\ yading@10: "fpsub16 %%f24, %%f8, %%f24 \n\t"\ yading@10: "fpadd16 %%f28, %%f8, %%f28 \n\t"\ yading@10: "fpadd16 %%f18, %%f10, %%f18 \n\t"\ yading@10: "fpsub16 %%f22, %%f10, %%f22 \n\t"\ yading@10: /* 4. column */\ yading@10: "4: \n\t"\ yading@10: "fbe %%fcc2, 5f \n\t"\ yading@10: "fpadd16 %%f30, %%f10, %%f30 \n\t"\ yading@10: "fmul8ulx16 %%f12, %%f42, %%f48 \n\t"\ yading@10: "fmul8ulx16 %%f12, %%f34, %%f50 \n\t"\ yading@10: "fmul8ulx16 %%f14, %%f44, %%f52 \n\t"\ yading@10: "fmul8ulx16 %%f14, %%f40, %%f54 \n\t"\ yading@10: "fmul8ulx16 %%f14, %%f36, %%f56 \n\t"\ yading@10: "fmul8ulx16 %%f14, %%f32, %%f58 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f16, %%f48, %%f16 \n\t"\ yading@10: "fpsub16 %%f20, %%f50, %%f20 \n\t"\ yading@10: "fpadd16 %%f24, %%f50, %%f24 \n\t"\ yading@10: "fpsub16 %%f28, %%f48, %%f28 \n\t"\ yading@10: "fpadd16 %%f18, %%f52, %%f18 \n\t"\ yading@10: "fpsub16 %%f22, %%f54, %%f22 \n\t"\ yading@10: "fpadd16 %%f26, %%f56, %%f26 \n\t"\ yading@10: "fpsub16 %%f30, %%f58, %%f30 \n\t"\ yading@10: \ yading@10: "fmul8sux16 %%f12, %%f42, %%f48 \n\t"\ yading@10: "fmul8sux16 %%f12, %%f34, %%f50 \n\t"\ yading@10: "fmul8sux16 %%f14, %%f44, %%f52 \n\t"\ yading@10: "fmul8sux16 %%f14, %%f40, %%f54 \n\t"\ yading@10: "fmul8sux16 %%f14, %%f36, %%f56 \n\t"\ yading@10: "fmul8sux16 %%f14, %%f32, %%f58 \n\t"\ yading@10: \ yading@10: "fpadd16 %%f16, %%f48, %%f16 \n\t"\ yading@10: "fpsub16 %%f20, %%f50, %%f20 \n\t"\ yading@10: "fpadd16 %%f24, %%f50, %%f24 \n\t"\ yading@10: "fpsub16 %%f28, %%f48, %%f28 \n\t"\ yading@10: "fpadd16 %%f18, %%f52, %%f18 \n\t"\ yading@10: "fpsub16 %%f22, %%f54, %%f22 \n\t"\ yading@10: "fpadd16 %%f26, %%f56, %%f26 \n\t"\ yading@10: "fpsub16 %%f30, %%f58, %%f30 \n\t"\ yading@10: \ yading@10: "fpsub16 %%f20, %%f12, %%f20 \n\t"\ yading@10: "fpadd16 %%f24, %%f12, %%f24 \n\t"\ yading@10: "fpsub16 %%f22, %%f14, %%f22 \n\t"\ yading@10: "fpadd16 %%f26, %%f14, %%f26 \n\t"\ yading@10: "fpsub16 %%f30, %%f14, %%f30 \n\t"\ yading@10: /* final butterfly */\ yading@10: "5: \n\t"\ yading@10: "fpsub16 %%f16, %%f18, %%f48 \n\t"\ yading@10: "fpsub16 %%f20, %%f22, %%f50 \n\t"\ yading@10: "fpsub16 %%f24, %%f26, %%f52 \n\t"\ yading@10: "fpsub16 %%f28, %%f30, %%f54 \n\t"\ yading@10: "fpadd16 %%f16, %%f18, %%f16 \n\t"\ yading@10: "fpadd16 %%f20, %%f22, %%f20 \n\t"\ yading@10: "fpadd16 %%f24, %%f26, %%f24 \n\t"\ yading@10: "fpadd16 %%f28, %%f30, %%f28 \n\t"\ yading@10: yading@10: #define STOREROWS(out) \ yading@10: "std %%f48, [" out "+112] \n\t"\ yading@10: "std %%f50, [" out "+96] \n\t"\ yading@10: "std %%f52, [" out "+80] \n\t"\ yading@10: "std %%f54, [" out "+64] \n\t"\ yading@10: "std %%f16, [" out "] \n\t"\ yading@10: "std %%f20, [" out "+16] \n\t"\ yading@10: "std %%f24, [" out "+32] \n\t"\ yading@10: "std %%f28, [" out "+48] \n\t"\ yading@10: yading@10: #define SCALEROWS \ yading@10: "fmul8sux16 %%f46, %%f48, %%f48 \n\t"\ yading@10: "fmul8sux16 %%f46, %%f50, %%f50 \n\t"\ yading@10: "fmul8sux16 %%f46, %%f52, %%f52 \n\t"\ yading@10: "fmul8sux16 %%f46, %%f54, %%f54 \n\t"\ yading@10: "fmul8sux16 %%f46, %%f16, %%f16 \n\t"\ yading@10: "fmul8sux16 %%f46, %%f20, %%f20 \n\t"\ yading@10: "fmul8sux16 %%f46, %%f24, %%f24 \n\t"\ yading@10: "fmul8sux16 %%f46, %%f28, %%f28 \n\t"\ yading@10: yading@10: #define PUTPIXELSCLAMPED(dest) \ yading@10: "fpack16 %%f48, %%f14 \n\t"\ yading@10: "fpack16 %%f50, %%f12 \n\t"\ yading@10: "fpack16 %%f16, %%f0 \n\t"\ yading@10: "fpack16 %%f20, %%f2 \n\t"\ yading@10: "fpack16 %%f24, %%f4 \n\t"\ yading@10: "fpack16 %%f28, %%f6 \n\t"\ yading@10: "fpack16 %%f54, %%f8 \n\t"\ yading@10: "fpack16 %%f52, %%f10 \n\t"\ yading@10: "st %%f0, [%3+" dest "] \n\t"\ yading@10: "st %%f2, [%5+" dest "] \n\t"\ yading@10: "st %%f4, [%6+" dest "] \n\t"\ yading@10: "st %%f6, [%7+" dest "] \n\t"\ yading@10: "st %%f8, [%8+" dest "] \n\t"\ yading@10: "st %%f10, [%9+" dest "] \n\t"\ yading@10: "st %%f12, [%10+" dest "] \n\t"\ yading@10: "st %%f14, [%11+" dest "] \n\t"\ yading@10: yading@10: #define ADDPIXELSCLAMPED(dest) \ yading@10: "ldd [%5], %%f18 \n\t"\ yading@10: "ld [%3+" dest"], %%f0 \n\t"\ yading@10: "ld [%6+" dest"], %%f2 \n\t"\ yading@10: "ld [%7+" dest"], %%f4 \n\t"\ yading@10: "ld [%8+" dest"], %%f6 \n\t"\ yading@10: "ld [%9+" dest"], %%f8 \n\t"\ yading@10: "ld [%10+" dest"], %%f10 \n\t"\ yading@10: "ld [%11+" dest"], %%f12 \n\t"\ yading@10: "ld [%12+" dest"], %%f14 \n\t"\ yading@10: "fmul8x16 %%f0, %%f18, %%f0 \n\t"\ yading@10: "fmul8x16 %%f2, %%f18, %%f2 \n\t"\ yading@10: "fmul8x16 %%f4, %%f18, %%f4 \n\t"\ yading@10: "fmul8x16 %%f6, %%f18, %%f6 \n\t"\ yading@10: "fmul8x16 %%f8, %%f18, %%f8 \n\t"\ yading@10: "fmul8x16 %%f10, %%f18, %%f10 \n\t"\ yading@10: "fmul8x16 %%f12, %%f18, %%f12 \n\t"\ yading@10: "fmul8x16 %%f14, %%f18, %%f14 \n\t"\ yading@10: "fpadd16 %%f0, %%f16, %%f0 \n\t"\ yading@10: "fpadd16 %%f2, %%f20, %%f2 \n\t"\ yading@10: "fpadd16 %%f4, %%f24, %%f4 \n\t"\ yading@10: "fpadd16 %%f6, %%f28, %%f6 \n\t"\ yading@10: "fpadd16 %%f8, %%f54, %%f8 \n\t"\ yading@10: "fpadd16 %%f10, %%f52, %%f10 \n\t"\ yading@10: "fpadd16 %%f12, %%f50, %%f12 \n\t"\ yading@10: "fpadd16 %%f14, %%f48, %%f14 \n\t"\ yading@10: "fpack16 %%f0, %%f0 \n\t"\ yading@10: "fpack16 %%f2, %%f2 \n\t"\ yading@10: "fpack16 %%f4, %%f4 \n\t"\ yading@10: "fpack16 %%f6, %%f6 \n\t"\ yading@10: "fpack16 %%f8, %%f8 \n\t"\ yading@10: "fpack16 %%f10, %%f10 \n\t"\ yading@10: "fpack16 %%f12, %%f12 \n\t"\ yading@10: "fpack16 %%f14, %%f14 \n\t"\ yading@10: "st %%f0, [%3+" dest "] \n\t"\ yading@10: "st %%f2, [%6+" dest "] \n\t"\ yading@10: "st %%f4, [%7+" dest "] \n\t"\ yading@10: "st %%f6, [%8+" dest "] \n\t"\ yading@10: "st %%f8, [%9+" dest "] \n\t"\ yading@10: "st %%f10, [%10+" dest "] \n\t"\ yading@10: "st %%f12, [%11+" dest "] \n\t"\ yading@10: "st %%f14, [%12+" dest "] \n\t"\ yading@10: yading@10: yading@10: void ff_simple_idct_vis(int16_t *data) { yading@10: int out1, out2, out3, out4; yading@10: DECLARE_ALIGNED(8, int16_t, temp)[8*8]; yading@10: yading@10: __asm__ volatile( yading@10: INIT_IDCT yading@10: yading@10: #define ADDROUNDER yading@10: yading@10: // shift right 16-4=12 yading@10: LOADSCALE("%2+8") yading@10: IDCT4ROWS yading@10: STOREROWS("%3+8") yading@10: LOADSCALE("%2+0") yading@10: IDCT4ROWS yading@10: "std %%f48, [%3+112] \n\t" yading@10: "std %%f50, [%3+96] \n\t" yading@10: "std %%f52, [%3+80] \n\t" yading@10: "std %%f54, [%3+64] \n\t" yading@10: yading@10: // shift right 16+4 yading@10: "ldd [%3+8], %%f18 \n\t" yading@10: "ldd [%3+24], %%f22 \n\t" yading@10: "ldd [%3+40], %%f26 \n\t" yading@10: "ldd [%3+56], %%f30 \n\t" yading@10: TRANSPOSE yading@10: IDCT4ROWS yading@10: SCALEROWS yading@10: STOREROWS("%2+0") yading@10: LOAD("%3+64") yading@10: TRANSPOSE yading@10: IDCT4ROWS yading@10: SCALEROWS yading@10: STOREROWS("%2+8") yading@10: yading@10: : "=r" (out1), "=r" (out2), "=r" (out3), "=r" (out4) yading@10: : "0" (scale), "1" (coeffs), "2" (data), "3" (temp) yading@10: ); yading@10: } yading@10: yading@10: void ff_simple_idct_put_vis(uint8_t *dest, int line_size, int16_t *data) { yading@10: int out1, out2, out3, out4, out5; yading@10: int r1, r2, r3, r4, r5, r6, r7; yading@10: yading@10: __asm__ volatile( yading@10: "wr %%g0, 0x8, %%gsr \n\t" yading@10: yading@10: INIT_IDCT yading@10: yading@10: "add %3, %4, %5 \n\t" yading@10: "add %5, %4, %6 \n\t" yading@10: "add %6, %4, %7 \n\t" yading@10: "add %7, %4, %8 \n\t" yading@10: "add %8, %4, %9 \n\t" yading@10: "add %9, %4, %10 \n\t" yading@10: "add %10, %4, %11 \n\t" yading@10: yading@10: // shift right 16-4=12 yading@10: LOADSCALE("%2+8") yading@10: IDCT4ROWS yading@10: STOREROWS("%2+8") yading@10: LOADSCALE("%2+0") yading@10: IDCT4ROWS yading@10: "std %%f48, [%2+112] \n\t" yading@10: "std %%f50, [%2+96] \n\t" yading@10: "std %%f52, [%2+80] \n\t" yading@10: "std %%f54, [%2+64] \n\t" yading@10: yading@10: #undef ADDROUNDER yading@10: #define ADDROUNDER "fpadd16 %%f28, %%f46, %%f28 \n\t" yading@10: yading@10: // shift right 16+4 yading@10: "ldd [%2+8], %%f18 \n\t" yading@10: "ldd [%2+24], %%f22 \n\t" yading@10: "ldd [%2+40], %%f26 \n\t" yading@10: "ldd [%2+56], %%f30 \n\t" yading@10: TRANSPOSE yading@10: IDCT4ROWS yading@10: PUTPIXELSCLAMPED("0") yading@10: LOAD("%2+64") yading@10: TRANSPOSE yading@10: IDCT4ROWS yading@10: PUTPIXELSCLAMPED("4") yading@10: yading@10: : "=r" (out1), "=r" (out2), "=r" (out3), "=r" (out4), "=r" (out5), yading@10: "=r" (r1), "=r" (r2), "=r" (r3), "=r" (r4), "=r" (r5), "=r" (r6), "=r" (r7) yading@10: : "0" (rounder), "1" (coeffs), "2" (data), "3" (dest), "4" (line_size) yading@10: ); yading@10: } yading@10: yading@10: void ff_simple_idct_add_vis(uint8_t *dest, int line_size, int16_t *data) { yading@10: int out1, out2, out3, out4, out5, out6; yading@10: int r1, r2, r3, r4, r5, r6, r7; yading@10: yading@10: __asm__ volatile( yading@10: "wr %%g0, 0x8, %%gsr \n\t" yading@10: yading@10: INIT_IDCT yading@10: yading@10: "add %3, %4, %6 \n\t" yading@10: "add %6, %4, %7 \n\t" yading@10: "add %7, %4, %8 \n\t" yading@10: "add %8, %4, %9 \n\t" yading@10: "add %9, %4, %10 \n\t" yading@10: "add %10, %4, %11 \n\t" yading@10: "add %11, %4, %12 \n\t" yading@10: yading@10: #undef ADDROUNDER yading@10: #define ADDROUNDER yading@10: yading@10: // shift right 16-4=12 yading@10: LOADSCALE("%2+8") yading@10: IDCT4ROWS yading@10: STOREROWS("%2+8") yading@10: LOADSCALE("%2+0") yading@10: IDCT4ROWS yading@10: "std %%f48, [%2+112] \n\t" yading@10: "std %%f50, [%2+96] \n\t" yading@10: "std %%f52, [%2+80] \n\t" yading@10: "std %%f54, [%2+64] \n\t" yading@10: yading@10: #undef ADDROUNDER yading@10: #define ADDROUNDER "fpadd16 %%f28, %%f46, %%f28 \n\t" yading@10: yading@10: // shift right 16+4 yading@10: "ldd [%2+8], %%f18 \n\t" yading@10: "ldd [%2+24], %%f22 \n\t" yading@10: "ldd [%2+40], %%f26 \n\t" yading@10: "ldd [%2+56], %%f30 \n\t" yading@10: TRANSPOSE yading@10: IDCT4ROWS yading@10: ADDPIXELSCLAMPED("0") yading@10: LOAD("%2+64") yading@10: TRANSPOSE yading@10: IDCT4ROWS yading@10: ADDPIXELSCLAMPED("4") yading@10: yading@10: : "=r" (out1), "=r" (out2), "=r" (out3), "=r" (out4), "=r" (out5), "=r" (out6), yading@10: "=r" (r1), "=r" (r2), "=r" (r3), "=r" (r4), "=r" (r5), "=r" (r6), "=r" (r7) yading@10: : "0" (rounder), "1" (coeffs), "2" (data), "3" (dest), "4" (line_size), "5" (expand) yading@10: ); yading@10: }