yading@11: /* yading@11: * VIS optimized software YUV to RGB converter yading@11: * Copyright (c) 2007 Denes Balatoni yading@11: * yading@11: * This file is part of FFmpeg. yading@11: * yading@11: * FFmpeg is free software; you can redistribute it and/or yading@11: * modify it under the terms of the GNU Lesser General Public yading@11: * License as published by the Free Software Foundation; either yading@11: * version 2.1 of the License, or (at your option) any later version. yading@11: * yading@11: * FFmpeg is distributed in the hope that it will be useful, yading@11: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@11: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@11: * Lesser General Public License for more details. yading@11: * yading@11: * You should have received a copy of the GNU Lesser General Public yading@11: * License along with FFmpeg; if not, write to the Free Software yading@11: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@11: */ yading@11: yading@11: #include yading@11: #include yading@11: yading@11: #include "libavutil/attributes.h" yading@11: #include "libswscale/swscale.h" yading@11: #include "libswscale/swscale_internal.h" yading@11: yading@11: #define YUV2RGB_INIT \ yading@11: "wr %%g0, 0x10, %%gsr \n\t" \ yading@11: "ldd [%5], %%f32 \n\t" \ yading@11: "ldd [%5 + 8], %%f34 \n\t" \ yading@11: "ldd [%5 + 16], %%f36 \n\t" \ yading@11: "ldd [%5 + 24], %%f38 \n\t" \ yading@11: "ldd [%5 + 32], %%f40 \n\t" \ yading@11: "ldd [%5 + 40], %%f42 \n\t" \ yading@11: "ldd [%5 + 48], %%f44 \n\t" \ yading@11: "ldd [%5 + 56], %%f46 \n\t" \ yading@11: "ldd [%5 + 64], %%f48 \n\t" \ yading@11: "ldd [%5 + 72], %%f50 \n\t" yading@11: yading@11: #define YUV2RGB_KERNEL \ yading@11: /* ^^^^ f0=Y f3=u f5=v */ \ yading@11: "fmul8x16 %%f3, %%f48, %%f6 \n\t" \ yading@11: "fmul8x16 %%f19, %%f48, %%f22 \n\t" \ yading@11: "fmul8x16 %%f5, %%f44, %%f8 \n\t" \ yading@11: "fmul8x16 %%f21, %%f44, %%f24 \n\t" \ yading@11: "fmul8x16 %%f0, %%f42, %%f0 \n\t" \ yading@11: "fmul8x16 %%f16, %%f42, %%f16 \n\t" \ yading@11: "fmul8x16 %%f3, %%f50, %%f2 \n\t" \ yading@11: "fmul8x16 %%f19, %%f50, %%f18 \n\t" \ yading@11: "fmul8x16 %%f5, %%f46, %%f4 \n\t" \ yading@11: "fmul8x16 %%f21, %%f46, %%f20 \n\t" \ yading@11: \ yading@11: "fpsub16 %%f6, %%f34, %%f6 \n\t" /* 1 */ \ yading@11: "fpsub16 %%f22, %%f34, %%f22 \n\t" /* 1 */ \ yading@11: "fpsub16 %%f8, %%f38, %%f8 \n\t" /* 3 */ \ yading@11: "fpsub16 %%f24, %%f38, %%f24 \n\t" /* 3 */ \ yading@11: "fpsub16 %%f0, %%f32, %%f0 \n\t" /* 0 */ \ yading@11: "fpsub16 %%f16, %%f32, %%f16 \n\t" /* 0 */ \ yading@11: "fpsub16 %%f2, %%f36, %%f2 \n\t" /* 2 */ \ yading@11: "fpsub16 %%f18, %%f36, %%f18 \n\t" /* 2 */ \ yading@11: "fpsub16 %%f4, %%f40, %%f4 \n\t" /* 4 */ \ yading@11: "fpsub16 %%f20, %%f40, %%f20 \n\t" /* 4 */ \ yading@11: \ yading@11: "fpadd16 %%f0, %%f8, %%f8 \n\t" /* Gt */ \ yading@11: "fpadd16 %%f16, %%f24, %%f24 \n\t" /* Gt */ \ yading@11: "fpadd16 %%f0, %%f4, %%f4 \n\t" /* R */ \ yading@11: "fpadd16 %%f16, %%f20, %%f20 \n\t" /* R */ \ yading@11: "fpadd16 %%f0, %%f6, %%f6 \n\t" /* B */ \ yading@11: "fpadd16 %%f16, %%f22, %%f22 \n\t" /* B */ \ yading@11: "fpadd16 %%f8, %%f2, %%f2 \n\t" /* G */ \ yading@11: "fpadd16 %%f24, %%f18, %%f18 \n\t" /* G */ \ yading@11: \ yading@11: "fpack16 %%f4, %%f4 \n\t" \ yading@11: "fpack16 %%f20, %%f20 \n\t" \ yading@11: "fpack16 %%f6, %%f6 \n\t" \ yading@11: "fpack16 %%f22, %%f22 \n\t" \ yading@11: "fpack16 %%f2, %%f2 \n\t" \ yading@11: "fpack16 %%f18, %%f18 \n\t" yading@11: yading@11: // FIXME: must be changed to set alpha to 255 instead of 0 yading@11: static int vis_420P_ARGB32(SwsContext *c, uint8_t *src[], int srcStride[], yading@11: int srcSliceY, int srcSliceH, yading@11: uint8_t *dst[], int dstStride[]) yading@11: { yading@11: int y, out1, out2, out3, out4, out5, out6; yading@11: yading@11: for (y = 0; y < srcSliceH; ++y) yading@11: __asm__ volatile ( yading@11: YUV2RGB_INIT yading@11: "wr %%g0, 0xd2, %%asi \n\t" /* ASI_FL16_P */ yading@11: "1: \n\t" yading@11: "ldda [%1] %%asi, %%f2 \n\t" yading@11: "ldda [%1 + 2] %%asi, %%f18 \n\t" yading@11: "ldda [%2] %%asi, %%f4 \n\t" yading@11: "ldda [%2 + 2] %%asi, %%f20 \n\t" yading@11: "ld [%0], %%f0 \n\t" yading@11: "ld [%0+4], %%f16 \n\t" yading@11: "fpmerge %%f3, %%f3, %%f2 \n\t" yading@11: "fpmerge %%f19, %%f19, %%f18 \n\t" yading@11: "fpmerge %%f5, %%f5, %%f4 \n\t" yading@11: "fpmerge %%f21, %%f21, %%f20 \n\t" yading@11: YUV2RGB_KERNEL yading@11: "fzero %%f0 \n\t" yading@11: "fpmerge %%f4, %%f6, %%f8 \n\t" // r, b, t1 yading@11: "fpmerge %%f20, %%f22, %%f24 \n\t" // r, b, t1 yading@11: "fpmerge %%f0, %%f2, %%f10 \n\t" // 0, g, t2 yading@11: "fpmerge %%f0, %%f18, %%f26 \n\t" // 0, g, t2 yading@11: "fpmerge %%f10, %%f8, %%f4 \n\t" // t2, t1, msb yading@11: "fpmerge %%f26, %%f24, %%f20 \n\t" // t2, t1, msb yading@11: "fpmerge %%f11, %%f9, %%f6 \n\t" // t2, t1, lsb yading@11: "fpmerge %%f27, %%f25, %%f22 \n\t" // t2, t1, lsb yading@11: "std %%f4, [%3] \n\t" yading@11: "std %%f20, [%3 + 16] \n\t" yading@11: "std %%f6, [%3 + 8] \n\t" yading@11: "std %%f22, [%3 + 24] \n\t" yading@11: yading@11: "add %0, 8, %0 \n\t" yading@11: "add %1, 4, %1 \n\t" yading@11: "add %2, 4, %2 \n\t" yading@11: "subcc %4, 8, %4 \n\t" yading@11: "bne 1b \n\t" yading@11: "add %3, 32, %3 \n\t" // delay slot yading@11: : "=r" (out1), "=r" (out2), "=r" (out3), "=r" (out4), "=r" (out5), "=r" (out6) yading@11: : "0" (src[0] + (y + srcSliceY) * srcStride[0]), "1" (src[1] + ((y + srcSliceY) >> 1) * srcStride[1]), yading@11: "2" (src[2] + ((y + srcSliceY) >> 1) * srcStride[2]), "3" (dst[0] + (y + srcSliceY) * dstStride[0]), yading@11: "4" (c->dstW), yading@11: "5" (c->sparc_coeffs) yading@11: ); yading@11: yading@11: return srcSliceH; yading@11: } yading@11: yading@11: // FIXME: must be changed to set alpha to 255 instead of 0 yading@11: static int vis_422P_ARGB32(SwsContext *c, uint8_t *src[], int srcStride[], yading@11: int srcSliceY, int srcSliceH, yading@11: uint8_t *dst[], int dstStride[]) yading@11: { yading@11: int y, out1, out2, out3, out4, out5, out6; yading@11: yading@11: for (y = 0; y < srcSliceH; ++y) yading@11: __asm__ volatile ( yading@11: YUV2RGB_INIT yading@11: "wr %%g0, 0xd2, %%asi \n\t" /* ASI_FL16_P */ yading@11: "1: \n\t" yading@11: "ldda [%1] %%asi, %%f2 \n\t" yading@11: "ldda [%1 + 2] %%asi, %%f18 \n\t" yading@11: "ldda [%2] %%asi, %%f4 \n\t" yading@11: "ldda [%2 + 2] %%asi, %%f20 \n\t" yading@11: "ld [%0], %%f0 \n\t" yading@11: "ld [%0 + 4], %%f16 \n\t" yading@11: "fpmerge %%f3, %%f3, %%f2 \n\t" yading@11: "fpmerge %%f19, %%f19, %%f18 \n\t" yading@11: "fpmerge %%f5, %%f5, %%f4 \n\t" yading@11: "fpmerge %%f21, %%f21, %%f20 \n\t" yading@11: YUV2RGB_KERNEL yading@11: "fzero %%f0 \n\t" yading@11: "fpmerge %%f4, %%f6, %%f8 \n\t" // r,b,t1 yading@11: "fpmerge %%f20, %%f22, %%f24 \n\t" // r,b,t1 yading@11: "fpmerge %%f0, %%f2, %%f10 \n\t" // 0,g,t2 yading@11: "fpmerge %%f0, %%f18, %%f26 \n\t" // 0,g,t2 yading@11: "fpmerge %%f10, %%f8, %%f4 \n\t" // t2,t1,msb yading@11: "fpmerge %%f26, %%f24, %%f20 \n\t" // t2,t1,msb yading@11: "fpmerge %%f11, %%f9, %%f6 \n\t" // t2,t1,lsb yading@11: "fpmerge %%f27, %%f25, %%f22 \n\t" // t2,t1,lsb yading@11: "std %%f4, [%3] \n\t" yading@11: "std %%f20, [%3 + 16] \n\t" yading@11: "std %%f6, [%3 + 8] \n\t" yading@11: "std %%f22, [%3 + 24] \n\t" yading@11: yading@11: "add %0, 8, %0 \n\t" yading@11: "add %1, 4, %1 \n\t" yading@11: "add %2, 4, %2 \n\t" yading@11: "subcc %4, 8, %4 \n\t" yading@11: "bne 1b \n\t" yading@11: "add %3, 32, %3 \n\t" //delay slot yading@11: : "=r" (out1), "=r" (out2), "=r" (out3), "=r" (out4), "=r" (out5), "=r" (out6) yading@11: : "0" (src[0] + (y + srcSliceY) * srcStride[0]), "1" (src[1] + (y + srcSliceY) * srcStride[1]), yading@11: "2" (src[2] + (y + srcSliceY) * srcStride[2]), "3" (dst[0] + (y + srcSliceY) * dstStride[0]), yading@11: "4" (c->dstW), yading@11: "5" (c->sparc_coeffs) yading@11: ); yading@11: yading@11: return srcSliceH; yading@11: } yading@11: yading@11: av_cold SwsFunc ff_yuv2rgb_init_vis(SwsContext *c) yading@11: { yading@11: c->sparc_coeffs[5] = c->yCoeff; yading@11: c->sparc_coeffs[6] = c->vgCoeff; yading@11: c->sparc_coeffs[7] = c->vrCoeff; yading@11: c->sparc_coeffs[8] = c->ubCoeff; yading@11: c->sparc_coeffs[9] = c->ugCoeff; yading@11: yading@11: c->sparc_coeffs[0] = (((int16_t)c->yOffset * (int16_t)c->yCoeff >> 11) & 0xffff) * 0x0001000100010001ULL; yading@11: c->sparc_coeffs[1] = (((int16_t)c->uOffset * (int16_t)c->ubCoeff >> 11) & 0xffff) * 0x0001000100010001ULL; yading@11: c->sparc_coeffs[2] = (((int16_t)c->uOffset * (int16_t)c->ugCoeff >> 11) & 0xffff) * 0x0001000100010001ULL; yading@11: c->sparc_coeffs[3] = (((int16_t)c->vOffset * (int16_t)c->vgCoeff >> 11) & 0xffff) * 0x0001000100010001ULL; yading@11: c->sparc_coeffs[4] = (((int16_t)c->vOffset * (int16_t)c->vrCoeff >> 11) & 0xffff) * 0x0001000100010001ULL; yading@11: yading@11: if (c->dstFormat == AV_PIX_FMT_RGB32 && c->srcFormat == AV_PIX_FMT_YUV422P && (c->dstW & 7) == 0) { yading@11: av_log(c, AV_LOG_INFO, yading@11: "SPARC VIS accelerated YUV422P -> RGB32 (WARNING: alpha value is wrong)\n"); yading@11: return vis_422P_ARGB32; yading@11: } else if (c->dstFormat == AV_PIX_FMT_RGB32 && c->srcFormat == AV_PIX_FMT_YUV420P && (c->dstW & 7) == 0) { yading@11: av_log(c, AV_LOG_INFO, yading@11: "SPARC VIS accelerated YUV420P -> RGB32 (WARNING: alpha value is wrong)\n"); yading@11: return vis_420P_ARGB32; yading@11: } yading@11: return NULL; yading@11: }