yading@11: /* yading@11: * AltiVec-enhanced yuv-to-yuv conversion routines. yading@11: * yading@11: * Copyright (C) 2004 Romain Dolbeau yading@11: * based on the equivalent C code in swscale.c 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: yading@11: #include "config.h" yading@11: #include "libswscale/swscale.h" yading@11: #include "libswscale/swscale_internal.h" yading@11: #include "libavutil/cpu.h" yading@11: yading@11: static int yv12toyuy2_unscaled_altivec(SwsContext *c, const uint8_t *src[], yading@11: int srcStride[], int srcSliceY, yading@11: int srcSliceH, uint8_t *dstParam[], yading@11: int dstStride_a[]) yading@11: { yading@11: uint8_t *dst = dstParam[0] + dstStride_a[0] * srcSliceY; yading@11: // yv12toyuy2(src[0], src[1], src[2], dst, c->srcW, srcSliceH, yading@11: // srcStride[0], srcStride[1], dstStride[0]); yading@11: const uint8_t *ysrc = src[0]; yading@11: const uint8_t *usrc = src[1]; yading@11: const uint8_t *vsrc = src[2]; yading@11: const int width = c->srcW; yading@11: const int height = srcSliceH; yading@11: const int lumStride = srcStride[0]; yading@11: const int chromStride = srcStride[1]; yading@11: const int dstStride = dstStride_a[0]; yading@11: const vector unsigned char yperm = vec_lvsl(0, ysrc); yading@11: const int vertLumPerChroma = 2; yading@11: register unsigned int y; yading@11: yading@11: /* This code assumes: yading@11: * yading@11: * 1) dst is 16 bytes-aligned yading@11: * 2) dstStride is a multiple of 16 yading@11: * 3) width is a multiple of 16 yading@11: * 4) lum & chrom stride are multiples of 8 yading@11: */ yading@11: yading@11: for (y = 0; y < height; y++) { yading@11: int i; yading@11: for (i = 0; i < width - 31; i += 32) { yading@11: const unsigned int j = i >> 1; yading@11: vector unsigned char v_yA = vec_ld(i, ysrc); yading@11: vector unsigned char v_yB = vec_ld(i + 16, ysrc); yading@11: vector unsigned char v_yC = vec_ld(i + 32, ysrc); yading@11: vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm); yading@11: vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm); yading@11: vector unsigned char v_uA = vec_ld(j, usrc); yading@11: vector unsigned char v_uB = vec_ld(j + 16, usrc); yading@11: vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc)); yading@11: vector unsigned char v_vA = vec_ld(j, vsrc); yading@11: vector unsigned char v_vB = vec_ld(j + 16, vsrc); yading@11: vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc)); yading@11: vector unsigned char v_uv_a = vec_mergeh(v_u, v_v); yading@11: vector unsigned char v_uv_b = vec_mergel(v_u, v_v); yading@11: vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a); yading@11: vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a); yading@11: vector unsigned char v_yuy2_2 = vec_mergeh(v_y2, v_uv_b); yading@11: vector unsigned char v_yuy2_3 = vec_mergel(v_y2, v_uv_b); yading@11: vec_st(v_yuy2_0, (i << 1), dst); yading@11: vec_st(v_yuy2_1, (i << 1) + 16, dst); yading@11: vec_st(v_yuy2_2, (i << 1) + 32, dst); yading@11: vec_st(v_yuy2_3, (i << 1) + 48, dst); yading@11: } yading@11: if (i < width) { yading@11: const unsigned int j = i >> 1; yading@11: vector unsigned char v_y1 = vec_ld(i, ysrc); yading@11: vector unsigned char v_u = vec_ld(j, usrc); yading@11: vector unsigned char v_v = vec_ld(j, vsrc); yading@11: vector unsigned char v_uv_a = vec_mergeh(v_u, v_v); yading@11: vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a); yading@11: vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a); yading@11: vec_st(v_yuy2_0, (i << 1), dst); yading@11: vec_st(v_yuy2_1, (i << 1) + 16, dst); yading@11: } yading@11: if ((y & (vertLumPerChroma - 1)) == vertLumPerChroma - 1) { yading@11: usrc += chromStride; yading@11: vsrc += chromStride; yading@11: } yading@11: ysrc += lumStride; yading@11: dst += dstStride; yading@11: } yading@11: yading@11: return srcSliceH; yading@11: } yading@11: yading@11: static int yv12touyvy_unscaled_altivec(SwsContext *c, const uint8_t *src[], yading@11: int srcStride[], int srcSliceY, yading@11: int srcSliceH, uint8_t *dstParam[], yading@11: int dstStride_a[]) yading@11: { yading@11: uint8_t *dst = dstParam[0] + dstStride_a[0] * srcSliceY; yading@11: // yv12toyuy2(src[0], src[1], src[2], dst, c->srcW, srcSliceH, yading@11: // srcStride[0], srcStride[1], dstStride[0]); yading@11: const uint8_t *ysrc = src[0]; yading@11: const uint8_t *usrc = src[1]; yading@11: const uint8_t *vsrc = src[2]; yading@11: const int width = c->srcW; yading@11: const int height = srcSliceH; yading@11: const int lumStride = srcStride[0]; yading@11: const int chromStride = srcStride[1]; yading@11: const int dstStride = dstStride_a[0]; yading@11: const int vertLumPerChroma = 2; yading@11: const vector unsigned char yperm = vec_lvsl(0, ysrc); yading@11: register unsigned int y; yading@11: yading@11: /* This code assumes: yading@11: * yading@11: * 1) dst is 16 bytes-aligned yading@11: * 2) dstStride is a multiple of 16 yading@11: * 3) width is a multiple of 16 yading@11: * 4) lum & chrom stride are multiples of 8 yading@11: */ yading@11: yading@11: for (y = 0; y < height; y++) { yading@11: int i; yading@11: for (i = 0; i < width - 31; i += 32) { yading@11: const unsigned int j = i >> 1; yading@11: vector unsigned char v_yA = vec_ld(i, ysrc); yading@11: vector unsigned char v_yB = vec_ld(i + 16, ysrc); yading@11: vector unsigned char v_yC = vec_ld(i + 32, ysrc); yading@11: vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm); yading@11: vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm); yading@11: vector unsigned char v_uA = vec_ld(j, usrc); yading@11: vector unsigned char v_uB = vec_ld(j + 16, usrc); yading@11: vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc)); yading@11: vector unsigned char v_vA = vec_ld(j, vsrc); yading@11: vector unsigned char v_vB = vec_ld(j + 16, vsrc); yading@11: vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc)); yading@11: vector unsigned char v_uv_a = vec_mergeh(v_u, v_v); yading@11: vector unsigned char v_uv_b = vec_mergel(v_u, v_v); yading@11: vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1); yading@11: vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1); yading@11: vector unsigned char v_uyvy_2 = vec_mergeh(v_uv_b, v_y2); yading@11: vector unsigned char v_uyvy_3 = vec_mergel(v_uv_b, v_y2); yading@11: vec_st(v_uyvy_0, (i << 1), dst); yading@11: vec_st(v_uyvy_1, (i << 1) + 16, dst); yading@11: vec_st(v_uyvy_2, (i << 1) + 32, dst); yading@11: vec_st(v_uyvy_3, (i << 1) + 48, dst); yading@11: } yading@11: if (i < width) { yading@11: const unsigned int j = i >> 1; yading@11: vector unsigned char v_y1 = vec_ld(i, ysrc); yading@11: vector unsigned char v_u = vec_ld(j, usrc); yading@11: vector unsigned char v_v = vec_ld(j, vsrc); yading@11: vector unsigned char v_uv_a = vec_mergeh(v_u, v_v); yading@11: vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1); yading@11: vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1); yading@11: vec_st(v_uyvy_0, (i << 1), dst); yading@11: vec_st(v_uyvy_1, (i << 1) + 16, dst); yading@11: } yading@11: if ((y & (vertLumPerChroma - 1)) == vertLumPerChroma - 1) { yading@11: usrc += chromStride; yading@11: vsrc += chromStride; yading@11: } yading@11: ysrc += lumStride; yading@11: dst += dstStride; yading@11: } yading@11: return srcSliceH; yading@11: } yading@11: yading@11: void ff_swscale_get_unscaled_altivec(SwsContext *c) yading@11: { yading@11: if ((av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC) && !(c->srcW & 15) && yading@11: !(c->flags & SWS_BITEXACT) && c->srcFormat == AV_PIX_FMT_YUV420P) { yading@11: enum AVPixelFormat dstFormat = c->dstFormat; yading@11: yading@11: // unscaled YV12 -> packed YUV, we want speed yading@11: if (dstFormat == AV_PIX_FMT_YUYV422) yading@11: c->swScale = yv12toyuy2_unscaled_altivec; yading@11: else if (dstFormat == AV_PIX_FMT_UYVY422) yading@11: c->swScale = yv12touyvy_unscaled_altivec; yading@11: } yading@11: }