alternative LGPL-licensed, MMX-optimized YUV to RGB conversion routines
authorDiego Biurrun <diego@biurrun.de>
Tue, 4 May 2010 09:11:01 +0000 (09:11 +0000)
committerDiego Biurrun <diego@biurrun.de>
Tue, 4 May 2010 09:11:01 +0000 (09:11 +0000)
written by Kostya Shishkov

Originally committed as revision 31135 to svn://svn.mplayerhq.hu/mplayer/trunk/libswscale

libswscale/Makefile
libswscale/x86/yuv2rgb_mmx.c
libswscale/x86/yuv2rgb_template2.c [new file with mode: 0644]
libswscale/yuv2rgb.c

index 9ae28e0..6976079 100644 (file)
@@ -12,10 +12,9 @@ OBJS-$(ARCH_BFIN)          +=  bfin/internal_bfin.o     \
                                bfin/yuv2rgb_bfin.o
 OBJS-$(CONFIG_MLIB)        +=  mlib/yuv2rgb_mlib.o
 OBJS-$(HAVE_ALTIVEC)       +=  ppc/yuv2rgb_altivec.o
+OBJS-$(HAVE_MMX)           +=  x86/yuv2rgb_mmx.o
 OBJS-$(HAVE_VIS)           +=  sparc/yuv2rgb_vis.o
 
-MMX-OBJS-$(CONFIG_GPL)     +=  x86/yuv2rgb_mmx.o        \
-
 TESTPROGS = colorspace swscale
 
 DIRS = bfin mlib ppc sparc x86
index 1e69df1..d042421 100644 (file)
@@ -49,14 +49,22 @@ DECLARE_ASM_CONST(8, uint64_t, mmx_grnmask) = 0xfcfcfcfcfcfcfcfcULL;
 #define HAVE_MMX2 0
 #define HAVE_AMD3DNOW 0
 #define RENAME(a) a ## _MMX
+#if CONFIG_GPL
 #include "yuv2rgb_template.c"
+#else
+#include "yuv2rgb_template2.c"
+#endif
 
 //MMX2 versions
 #undef RENAME
 #undef HAVE_MMX2
 #define HAVE_MMX2 1
 #define RENAME(a) a ## _MMX2
+#if CONFIG_GPL
 #include "yuv2rgb_template.c"
+#else
+#include "yuv2rgb_template2.c"
+#endif
 
 SwsFunc ff_yuv2rgb_init_mmx(SwsContext *c)
 {
diff --git a/libswscale/x86/yuv2rgb_template2.c b/libswscale/x86/yuv2rgb_template2.c
new file mode 100644 (file)
index 0000000..f391cde
--- /dev/null
@@ -0,0 +1,459 @@
+/*
+ * software YUV to RGB converter
+ *
+ * Copyright (C) 2001-2007 Michael Niedermayer
+ *           (c) 2010 Konstantin Shishkov
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#undef MOVNTQ
+#undef EMMS
+#undef SFENCE
+
+#if HAVE_AMD3DNOW
+/* On K6 femms is faster than emms. On K7 femms is directly mapped to emms. */
+#define EMMS   "femms"
+#else
+#define EMMS   "emms"
+#endif
+
+#if HAVE_MMX2
+#define MOVNTQ "movntq"
+#define SFENCE "sfence"
+#else
+#define MOVNTQ "movq"
+#define SFENCE " # nop"
+#endif
+
+#define REG_BLUE  "0"
+#define REG_RED   "1"
+#define REG_GREEN "2"
+#define REG_ALPHA "3"
+
+#define YUV2RGB_LOOP(depth)                                          \
+    h_size = (c->dstW + 7) & ~7;                                     \
+    if (h_size * depth > FFABS(dstStride[0]))                        \
+        h_size -= 8;                                                 \
+                                                                     \
+    if (c->srcFormat == PIX_FMT_YUV422P) {                           \
+        srcStride[1] *= 2;                                           \
+        srcStride[2] *= 2;                                           \
+    }                                                                \
+                                                                     \
+    __asm__ volatile ("pxor %mm4, %mm4\n\t");                        \
+    for (y = 0; y < srcSliceH; y++) {                                \
+        uint8_t *image    = dst[0] + (y + srcSliceY) * dstStride[0]; \
+        const uint8_t *py = src[0] +               y * srcStride[0]; \
+        const uint8_t *pu = src[1] +        (y >> 1) * srcStride[1]; \
+        const uint8_t *pv = src[2] +        (y >> 1) * srcStride[2]; \
+        x86_reg index = -h_size / 2;                                 \
+
+#define YUV2RGB_INITIAL_LOAD          \
+    __asm__ volatile (                \
+        "movq (%5, %0, 2), %%mm6\n\t" \
+        "movd    (%2, %0), %%mm0\n\t" \
+        "movd    (%3, %0), %%mm1\n\t" \
+        "1: \n\t"                     \
+
+/* YUV2RGB core
+ * Conversion is performed in usual way:
+ * R = Y' * Ycoef + Vred * V'
+ * G = Y' * Ycoef + Vgreen * V' + Ugreen * U'
+ * B = Y' * Ycoef               + Ublue * U'
+ *
+ * where X' = X * 8 - Xoffset (multiplication is performed to increase
+ * precision a bit).
+ * Since it operates in YUV420 colorspace, Y component is additionally
+ * split into Y1 and Y2 for even and odd pixels.
+ *
+ * Input:
+ * mm0 - U (4 elems), mm1 - V (4 elems), mm6 - Y (8 elems), mm4 - zero register
+ * Output:
+ * mm1 - R, mm2 - G, mm0 - B
+ */
+#define YUV2RGB                                  \
+    /* convert Y, U, V into Y1', Y2', U', V' */  \
+    "movq      %%mm6, %%mm7\n\t"                 \
+    "punpcklbw %%mm4, %%mm0\n\t"                 \
+    "punpcklbw %%mm4, %%mm1\n\t"                 \
+    "pand     "MANGLE(mmx_00ffw)", %%mm6\n\t"    \
+    "psrlw     $8,    %%mm7\n\t"                 \
+    "psllw     $3,    %%mm0\n\t"                 \
+    "psllw     $3,    %%mm1\n\t"                 \
+    "psllw     $3,    %%mm6\n\t"                 \
+    "psllw     $3,    %%mm7\n\t"                 \
+    "psubsw   "U_OFFSET"(%4), %%mm0\n\t"         \
+    "psubsw   "V_OFFSET"(%4), %%mm1\n\t"         \
+    "psubw    "Y_OFFSET"(%4), %%mm6\n\t"         \
+    "psubw    "Y_OFFSET"(%4), %%mm7\n\t"         \
+\
+     /* multiply by coefficients */              \
+    "movq      %%mm0, %%mm2\n\t"                 \
+    "movq      %%mm1, %%mm3\n\t"                 \
+    "pmulhw   "UG_COEFF"(%4), %%mm2\n\t"         \
+    "pmulhw   "VG_COEFF"(%4), %%mm3\n\t"         \
+    "pmulhw   "Y_COEFF" (%4), %%mm6\n\t"         \
+    "pmulhw   "Y_COEFF" (%4), %%mm7\n\t"         \
+    "pmulhw   "UB_COEFF"(%4), %%mm0\n\t"         \
+    "pmulhw   "VR_COEFF"(%4), %%mm1\n\t"         \
+    "paddsw    %%mm3, %%mm2\n\t"                 \
+    /* now: mm0 = UB, mm1 = VR, mm2 = CG */      \
+    /*      mm6 = Y1, mm7 = Y2 */                \
+\
+    /* produce RGB */                            \
+    "movq      %%mm7, %%mm3\n\t"                 \
+    "movq      %%mm7, %%mm5\n\t"                 \
+    "paddsw    %%mm0, %%mm3\n\t"                 \
+    "paddsw    %%mm1, %%mm5\n\t"                 \
+    "paddsw    %%mm2, %%mm7\n\t"                 \
+    "paddsw    %%mm6, %%mm0\n\t"                 \
+    "paddsw    %%mm6, %%mm1\n\t"                 \
+    "paddsw    %%mm6, %%mm2\n\t"                 \
+\
+    /* pack and interleave even/odd pixels */    \
+    "packuswb  %%mm0, %%mm0\n\t"                 \
+    "packuswb  %%mm1, %%mm1\n\t"                 \
+    "packuswb  %%mm2, %%mm2\n\t"                 \
+    "packuswb  %%mm3, %%mm3\n\t"                 \
+    "packuswb  %%mm5, %%mm5\n\t"                 \
+    "packuswb  %%mm7, %%mm7\n\t"                 \
+    "punpcklbw %%mm3, %%mm0\n\t"                 \
+    "punpcklbw %%mm5, %%mm1\n\t"                 \
+    "punpcklbw %%mm7, %%mm2\n\t"                 \
+
+#define YUV2RGB_ENDLOOP(depth)                   \
+    "movq 8 (%5, %0, 2), %%mm6\n\t"              \
+    "movd 4 (%3, %0),    %%mm1\n\t"              \
+    "movd 4 (%2, %0),    %%mm0\n\t"              \
+    "add $"AV_STRINGIFY(depth * 8)", %1\n\t"     \
+    "add  $4, %0\n\t"                            \
+    "js   1b\n\t"                                \
+
+#define YUV2RGB_OPERANDS                                          \
+        : "+r" (index), "+r" (image)                              \
+        : "r" (pu - index), "r" (pv - index), "r"(&c->redDither), \
+          "r" (py - 2*index)                                      \
+        );                                                        \
+    }                                                             \
+
+#define YUV2RGB_OPERANDS_ALPHA                                    \
+        : "+r" (index), "+r" (image)                              \
+        : "r" (pu - index), "r" (pv - index), "r"(&c->redDither), \
+          "r" (py - 2*index), "r" (pa - 2*index)                  \
+        );                                                        \
+    }                                                             \
+
+#define YUV2RGB_ENDFUNC                          \
+    __asm__ volatile (SFENCE"\n\t"EMMS);         \
+    return srcSliceH;                            \
+
+
+#define RGB_PACK16(gmask, gshift, rshift)        \
+    "pand      "MANGLE(mmx_redmask)", %%mm0\n\t" \
+    "pand      "MANGLE(mmx_redmask)", %%mm1\n\t" \
+    "psrlw     $3,        %%mm0\n\t"             \
+    "pand      "MANGLE(gmask)",       %%mm2\n\t" \
+    "movq      %%mm0,     %%mm5\n\t"             \
+    "movq      %%mm1,     %%mm6\n\t"             \
+    "movq      %%mm2,     %%mm7\n\t"             \
+    "punpcklbw %%mm4,     %%mm0\n\t"             \
+    "punpcklbw %%mm4,     %%mm1\n\t"             \
+    "punpcklbw %%mm4,     %%mm2\n\t"             \
+    "punpckhbw %%mm4,     %%mm5\n\t"             \
+    "punpckhbw %%mm4,     %%mm6\n\t"             \
+    "punpckhbw %%mm4,     %%mm7\n\t"             \
+    "psllw     $"rshift", %%mm1\n\t"             \
+    "psllw     $"rshift", %%mm6\n\t"             \
+    "psllw     $"gshift", %%mm2\n\t"             \
+    "psllw     $"gshift", %%mm7\n\t"             \
+    "por       %%mm1,     %%mm0\n\t"             \
+    "por       %%mm6,     %%mm5\n\t"             \
+    "por       %%mm2,     %%mm0\n\t"             \
+    "por       %%mm7,     %%mm5\n\t"             \
+    MOVNTQ "   %%mm0,      (%1)\n\t"             \
+    MOVNTQ "   %%mm5,     8(%1)\n\t"             \
+
+#define DITHER_RGB                               \
+    "paddusb "BLUE_DITHER"(%4),  %%mm0\n\t"      \
+    "paddusb "GREEN_DITHER"(%4), %%mm2\n\t"      \
+    "paddusb "RED_DITHER"(%4),   %%mm1\n\t"      \
+
+static inline int RENAME(yuv420_rgb15)(SwsContext *c, const uint8_t *src[],
+                                       int srcStride[],
+                                       int srcSliceY, int srcSliceH,
+                                       uint8_t *dst[], int dstStride[])
+{
+    int y, h_size;
+
+    YUV2RGB_LOOP(2)
+
+#ifdef DITHER1XBPP
+        c->blueDither  = ff_dither8[y       & 1];
+        c->greenDither = ff_dither8[y       & 1];
+        c->redDither   = ff_dither8[(y + 1) & 1];
+#endif
+
+        YUV2RGB_INITIAL_LOAD
+        YUV2RGB
+#ifdef DITHER1XBPP
+        DITHER_RGB
+#endif
+        RGB_PACK16(mmx_redmask, "2", "7")
+
+    YUV2RGB_ENDLOOP(2)
+    YUV2RGB_OPERANDS
+    YUV2RGB_ENDFUNC
+}
+
+static inline int RENAME(yuv420_rgb16)(SwsContext *c, const uint8_t *src[],
+                                       int srcStride[],
+                                       int srcSliceY, int srcSliceH,
+                                       uint8_t *dst[], int dstStride[])
+{
+    int y, h_size;
+
+    YUV2RGB_LOOP(2)
+
+#ifdef DITHER1XBPP
+        c->blueDither  = ff_dither8[y       & 1];
+        c->greenDither = ff_dither4[y       & 1];
+        c->redDither   = ff_dither8[(y + 1) & 1];
+#endif
+
+        YUV2RGB_INITIAL_LOAD
+        YUV2RGB
+#ifdef DITHER1XBPP
+        DITHER_RGB
+#endif
+        RGB_PACK16(mmx_grnmask, "3", "8")
+
+    YUV2RGB_ENDLOOP(2)
+    YUV2RGB_OPERANDS
+    YUV2RGB_ENDFUNC
+}
+
+
+#define RGB_PACK24(red, blue)              \
+    /* generate first packed RGB octet */  \
+    "movq      %%mm2,      %%mm5\n\t"      \
+    "movq      %%mm"blue", %%mm6\n\t"      \
+    "movq      %%mm"red",  %%mm7\n\t"      \
+    "punpcklbw %%mm5,      %%mm6\n\t"      \
+    "punpcklbw %%mm4,      %%mm7\n\t"      \
+    "movq      %%mm6,      %%mm3\n\t"      \
+    "punpcklwd %%mm7,      %%mm6\n\t"      \
+    "psrlq     $32,        %%mm3\n\t"      \
+    "movq      %%mm6,      %%mm5\n\t"      \
+    "psllq     $40,        %%mm6\n\t"      \
+    "psllq     $48,        %%mm3\n\t"      \
+    "psrlq     $32,        %%mm5\n\t"      \
+    "psrlq     $40,        %%mm6\n\t"      \
+    "psllq     $24,        %%mm5\n\t"      \
+    "por       %%mm3,      %%mm6\n\t"      \
+    "por       %%mm5,      %%mm6\n\t"      \
+    MOVNTQ "   %%mm6,      (%1)\n\t"       \
+\
+    /* generate second packed RGB octet */ \
+    "movq      %%mm"red",  %%mm7\n\t"      \
+    "movq      %%mm2,      %%mm5\n\t"      \
+    "movq      %%mm"blue", %%mm6\n\t"      \
+    "punpcklbw %%mm4,      %%mm7\n\t"      \
+    "punpcklbw %%mm5,      %%mm6\n\t"      \
+    "movq      %%mm7,      %%mm3\n\t"      \
+    "punpckhwd %%mm7,      %%mm6\n\t"      \
+    "psllq     $16,        %%mm3\n\t"      \
+    "psrlq     $32,        %%mm6\n\t"      \
+    "psrlq     $48,        %%mm3\n\t"      \
+    "psllq     $8,         %%mm6\n\t"      \
+    "movq      %%mm"red",  %%mm7\n\t"      \
+    "por       %%mm6,      %%mm3\n\t"      \
+    "movq      %%mm"blue", %%mm6\n\t"      \
+    "movq      %%mm2,      %%mm5\n\t"      \
+    "punpckhbw %%mm4,      %%mm7\n\t"      \
+    "punpckhbw %%mm5,      %%mm6\n\t"      \
+    "movq      %%mm6,      %%mm5\n\t"      \
+    "punpcklwd %%mm7,      %%mm6\n\t"      \
+    "psrlq     $16,        %%mm5\n\t"      \
+    "psllq     $56,        %%mm5\n\t"      \
+    "por       %%mm5,      %%mm3\n\t"      \
+    "psllq     $32,        %%mm6\n\t"      \
+    "por       %%mm6,      %%mm3\n\t"      \
+    MOVNTQ "   %%mm3,      8(%1)\n\t"      \
+\
+    /* generate third packed RGB octet */  \
+    "movq      %%mm"red",  %%mm7\n\t"      \
+    "movq      %%mm2,      %%mm5\n\t"      \
+    "movq      %%mm2,      %%mm3\n\t"      \
+    "movq      %%mm"blue", %%mm6\n\t"      \
+    "punpckhbw %%mm"red",  %%mm3\n\t"      \
+    "punpckhbw %%mm4,      %%mm7\n\t"      \
+    "psllq     $32,        %%mm3\n\t"      \
+    "punpckhbw %%mm5,      %%mm6\n\t"      \
+    "psrlq     $48,        %%mm3\n\t"      \
+    "punpckhwd %%mm7,      %%mm6\n\t"      \
+    "movq      %%mm6,      %%mm7\n\t"      \
+    "psrlq     $32,        %%mm6\n\t"      \
+    "psllq     $32,        %%mm7\n\t"      \
+    "psllq     $40,        %%mm6\n\t"      \
+    "psrlq     $16,        %%mm7\n\t"      \
+    "por       %%mm6,      %%mm3\n\t"      \
+    "por       %%mm7,      %%mm3\n\t"      \
+    MOVNTQ "   %%mm3,      16(%1)\n\t"     \
+
+static inline int RENAME(yuv420_rgb24)(SwsContext *c, const uint8_t *src[],
+                                       int srcStride[],
+                                       int srcSliceY, int srcSliceH,
+                                       uint8_t *dst[], int dstStride[])
+{
+    int y, h_size;
+
+    YUV2RGB_LOOP(3)
+
+        YUV2RGB_INITIAL_LOAD
+        YUV2RGB
+        RGB_PACK24(REG_BLUE, REG_RED)
+
+    YUV2RGB_ENDLOOP(3)
+    YUV2RGB_OPERANDS
+    YUV2RGB_ENDFUNC
+}
+
+static inline int RENAME(yuv420_bgr24)(SwsContext *c, const uint8_t *src[],
+                                       int srcStride[],
+                                       int srcSliceY, int srcSliceH,
+                                       uint8_t *dst[], int dstStride[])
+{
+    int y, h_size;
+
+    YUV2RGB_LOOP(3)
+
+        YUV2RGB_INITIAL_LOAD
+        YUV2RGB
+        RGB_PACK24(REG_RED, REG_BLUE)
+
+    YUV2RGB_ENDLOOP(3)
+    YUV2RGB_OPERANDS
+    YUV2RGB_ENDFUNC
+}
+
+
+#define SET_EMPTY_ALPHA                                                      \
+    "pcmpeqd   %%mm"REG_ALPHA", %%mm"REG_ALPHA"\n\t" /* set alpha to 0xFF */ \
+
+#define LOAD_ALPHA                                   \
+    "movq      (%6, %0, 2),     %%mm"REG_ALPHA"\n\t" \
+
+#define RGB_PACK32(red, green, blue, alpha)  \
+    "movq      %%mm"blue",  %%mm5\n\t"       \
+    "movq      %%mm"red",   %%mm6\n\t"       \
+    "punpckhbw %%mm"green", %%mm5\n\t"       \
+    "punpcklbw %%mm"green", %%mm"blue"\n\t"  \
+    "punpckhbw %%mm"alpha", %%mm6\n\t"       \
+    "punpcklbw %%mm"alpha", %%mm"red"\n\t"   \
+    "movq      %%mm"blue",  %%mm"green"\n\t" \
+    "movq      %%mm5,       %%mm"alpha"\n\t" \
+    "punpcklwd %%mm"red",   %%mm"blue"\n\t"  \
+    "punpckhwd %%mm"red",   %%mm"green"\n\t" \
+    "punpcklwd %%mm6,       %%mm5\n\t"       \
+    "punpckhwd %%mm6,       %%mm"alpha"\n\t" \
+    MOVNTQ "   %%mm"blue",   0(%1)\n\t"      \
+    MOVNTQ "   %%mm"green",  8(%1)\n\t"      \
+    MOVNTQ "   %%mm5,       16(%1)\n\t"      \
+    MOVNTQ "   %%mm"alpha", 24(%1)\n\t"      \
+
+static inline int RENAME(yuv420_rgb32)(SwsContext *c, const uint8_t *src[],
+                                       int srcStride[],
+                                       int srcSliceY, int srcSliceH,
+                                       uint8_t *dst[], int dstStride[])
+{
+    int y, h_size;
+
+    YUV2RGB_LOOP(4)
+
+        YUV2RGB_INITIAL_LOAD
+        YUV2RGB
+        SET_EMPTY_ALPHA
+        RGB_PACK32(REG_RED, REG_GREEN, REG_BLUE, REG_ALPHA)
+
+    YUV2RGB_ENDLOOP(4)
+    YUV2RGB_OPERANDS
+    YUV2RGB_ENDFUNC
+}
+
+static inline int RENAME(yuva420_rgb32)(SwsContext *c, const uint8_t *src[],
+                                        int srcStride[],
+                                        int srcSliceY, int srcSliceH,
+                                        uint8_t *dst[], int dstStride[])
+{
+#if HAVE_7REGS
+    int y, h_size;
+
+    YUV2RGB_LOOP(4)
+
+        const uint8_t *pa = src[3] + y * srcStride[3];
+        YUV2RGB_INITIAL_LOAD
+        YUV2RGB
+        LOAD_ALPHA
+        RGB_PACK32(REG_RED, REG_GREEN, REG_BLUE, REG_ALPHA)
+
+    YUV2RGB_ENDLOOP(4)
+    YUV2RGB_OPERANDS_ALPHA
+    YUV2RGB_ENDFUNC
+#endif
+}
+
+static inline int RENAME(yuv420_bgr32)(SwsContext *c, const uint8_t *src[],
+                                       int srcStride[],
+                                       int srcSliceY, int srcSliceH,
+                                       uint8_t *dst[], int dstStride[])
+{
+    int y, h_size;
+
+    YUV2RGB_LOOP(4)
+
+        YUV2RGB_INITIAL_LOAD
+        YUV2RGB
+        SET_EMPTY_ALPHA
+        RGB_PACK32(REG_BLUE, REG_GREEN, REG_RED, REG_ALPHA)
+
+    YUV2RGB_ENDLOOP(4)
+    YUV2RGB_OPERANDS
+    YUV2RGB_ENDFUNC
+}
+
+static inline int RENAME(yuva420_bgr32)(SwsContext *c, const uint8_t *src[],
+                                        int srcStride[],
+                                        int srcSliceY, int srcSliceH,
+                                        uint8_t *dst[], int dstStride[])
+{
+#if HAVE_7REGS
+    int y, h_size;
+
+    YUV2RGB_LOOP(4)
+
+        const uint8_t *pa = src[3] + y * srcStride[3];
+        YUV2RGB_INITIAL_LOAD
+        YUV2RGB
+        LOAD_ALPHA
+        RGB_PACK32(REG_BLUE, REG_GREEN, REG_RED, REG_ALPHA)
+
+    YUV2RGB_ENDLOOP(4)
+    YUV2RGB_OPERANDS_ALPHA
+    YUV2RGB_ENDFUNC
+#endif
+}
index 60764ae..f47b15d 100644 (file)
@@ -543,7 +543,7 @@ CLOSEYUV2RGBFUNC(1)
 SwsFunc ff_yuv2rgb_get_func_ptr(SwsContext *c)
 {
     SwsFunc t = NULL;
-#if HAVE_MMX && CONFIG_GPL
+#if HAVE_MMX
      t = ff_yuv2rgb_init_mmx(c);
 #endif
 #if HAVE_VIS