* License along with this library; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*/
+
+/**
+ * @file dsputil.h
+ * DSP utils.
+ * note, many functions in here may use MMX which trashes the FPU state, it is
+ * absolutely necessary to call emms_c() between dsp & float/double code
+ */
+
#ifndef DSPUTIL_H
#define DSPUTIL_H
#include "common.h"
#include "avcodec.h"
+
//#define DEBUG
/* dct code */
typedef short DCTELEM;
+//typedef int DCTELEM;
void fdct_ifast (DCTELEM *data);
+void ff_jpeg_fdct_islow (DCTELEM *data);
void j_rev_dct (DCTELEM *data);
-void fdct_mmx(DCTELEM *block);
-
-extern void (*av_fdct)(DCTELEM *block);
+void ff_fdct_mmx(DCTELEM *block);
/* encoding scans */
-extern UINT8 ff_alternate_horizontal_scan[64];
-extern UINT8 ff_alternate_vertical_scan[64];
-extern UINT8 zigzag_direct[64];
-
-/* permutation table */
-extern UINT8 permutation[64];
+extern const uint8_t ff_alternate_horizontal_scan[64];
+extern const uint8_t ff_alternate_vertical_scan[64];
+extern const uint8_t ff_zigzag_direct[64];
/* pixel operations */
#define MAX_NEG_CROP 384
/* temporary */
-extern UINT32 squareTbl[512];
-extern UINT8 cropTbl[256 + 2 * MAX_NEG_CROP];
+extern uint32_t squareTbl[512];
+extern uint8_t cropTbl[256 + 2 * MAX_NEG_CROP];
-void dsputil_init(void);
-/* pixel ops : interface with DCT */
+/* minimum alignment rules ;)
+if u notice errors in the align stuff, need more alignment for some asm code for some cpu
+or need to use a function with less aligned data then send a mail to the ffmpeg-dev list, ...
-extern void (*ff_idct)(DCTELEM *block);
-extern void (*ff_idct_put)(UINT8 *dest, int line_size, DCTELEM *block);
-extern void (*ff_idct_add)(UINT8 *dest, int line_size, DCTELEM *block);
-extern void (*get_pixels)(DCTELEM *block, const UINT8 *pixels, int line_size);
-extern void (*diff_pixels)(DCTELEM *block, const UINT8 *s1, const UINT8 *s2, int stride);
-extern void (*put_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size);
-extern void (*add_pixels_clamped)(const DCTELEM *block, UINT8 *pixels, int line_size);
-extern void (*gmc1)(UINT8 *dst, UINT8 *src, int srcStride, int h, int x16, int y16, int rounder);
-extern void (*clear_blocks)(DCTELEM *blocks);
-extern int (*pix_sum)(UINT8 * pix, int line_size);
-extern int (*pix_norm1)(UINT8 * pix, int line_size);
+!warning these alignments might not match reallity, (missing attribute((align)) stuff somewhere possible)
+i (michael) didnt check them, these are just the alignents which i think could be reached easily ...
+!future video codecs might need functions with less strict alignment
+*/
-
-void get_pixels_c(DCTELEM *block, const UINT8 *pixels, int line_size);
-void diff_pixels_c(DCTELEM *block, const UINT8 *s1, const UINT8 *s2, int stride);
-void put_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size);
-void add_pixels_clamped_c(const DCTELEM *block, UINT8 *pixels, int line_size);
+/*
+void get_pixels_c(DCTELEM *block, const uint8_t *pixels, int line_size);
+void diff_pixels_c(DCTELEM *block, const uint8_t *s1, const uint8_t *s2, int stride);
+void put_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
+void add_pixels_clamped_c(const DCTELEM *block, uint8_t *pixels, int line_size);
void clear_blocks_c(DCTELEM *blocks);
+*/
/* add and put pixel (decoding) */
-typedef void (*op_pixels_func)(UINT8 *block, const UINT8 *pixels, int line_size, int h);
-typedef void (*qpel_mc_func)(UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my);
-
-extern op_pixels_func put_pixels_tab[4];
-extern op_pixels_func avg_pixels_tab[4];
-extern op_pixels_func put_no_rnd_pixels_tab[4];
-extern op_pixels_func avg_no_rnd_pixels_tab[4];
-extern qpel_mc_func qpel_mc_rnd_tab[16];
-extern qpel_mc_func qpel_mc_no_rnd_tab[16];
+// blocksizes for op_pixels_func are 8x4,8x8 16x8 16x16
+typedef void (*op_pixels_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int h);
+typedef void (*tpel_mc_func)(uint8_t *block/*align width (8 or 16)*/, const uint8_t *pixels/*align 1*/, int line_size, int w, int h);
+typedef void (*qpel_mc_func)(uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
+typedef void (*h264_chroma_mc_func)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x, int y);
+
+#define DEF_OLD_QPEL(name)\
+void ff_put_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
+void ff_put_no_rnd_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);\
+void ff_avg_ ## name (uint8_t *dst/*align width (8 or 16)*/, uint8_t *src/*align 1*/, int stride);
+
+DEF_OLD_QPEL(qpel16_mc11_old_c)
+DEF_OLD_QPEL(qpel16_mc31_old_c)
+DEF_OLD_QPEL(qpel16_mc12_old_c)
+DEF_OLD_QPEL(qpel16_mc32_old_c)
+DEF_OLD_QPEL(qpel16_mc13_old_c)
+DEF_OLD_QPEL(qpel16_mc33_old_c)
+DEF_OLD_QPEL(qpel8_mc11_old_c)
+DEF_OLD_QPEL(qpel8_mc31_old_c)
+DEF_OLD_QPEL(qpel8_mc12_old_c)
+DEF_OLD_QPEL(qpel8_mc32_old_c)
+DEF_OLD_QPEL(qpel8_mc13_old_c)
+DEF_OLD_QPEL(qpel8_mc33_old_c)
+
+#define CALL_2X_PIXELS(a, b, n)\
+static void a(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
+ b(block , pixels , line_size, h);\
+ b(block+n, pixels+n, line_size, h);\
+}
/* motion estimation */
-typedef int (*op_pixels_abs_func)(UINT8 *blk1, UINT8 *blk2, int line_size);
+typedef int (*op_pixels_abs_func)(uint8_t *blk1/*align width (8 or 16)*/, uint8_t *blk2/*align 1*/, int line_size)/* __attribute__ ((const))*/;
+
+typedef int (*me_cmp_func)(void /*MpegEncContext*/ *s, uint8_t *blk1/*align width (8 or 16)*/, uint8_t *blk2/*align 1*/, int line_size)/* __attribute__ ((const))*/;
+
-extern op_pixels_abs_func pix_abs16x16;
-extern op_pixels_abs_func pix_abs16x16_x2;
-extern op_pixels_abs_func pix_abs16x16_y2;
-extern op_pixels_abs_func pix_abs16x16_xy2;
-extern op_pixels_abs_func pix_abs8x8;
-extern op_pixels_abs_func pix_abs8x8_x2;
-extern op_pixels_abs_func pix_abs8x8_y2;
-extern op_pixels_abs_func pix_abs8x8_xy2;
+/**
+ * DSPContext.
+ */
+typedef struct DSPContext {
+ /* pixel ops : interface with DCT */
+ void (*get_pixels)(DCTELEM *block/*align 16*/, const uint8_t *pixels/*align 8*/, int line_size);
+ void (*diff_pixels)(DCTELEM *block/*align 16*/, const uint8_t *s1/*align 8*/, const uint8_t *s2/*align 8*/, int stride);
+ void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
+ void (*add_pixels_clamped)(const DCTELEM *block/*align 16*/, uint8_t *pixels/*align 8*/, int line_size);
+ /**
+ * translational global motion compensation.
+ */
+ void (*gmc1)(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int srcStride, int h, int x16, int y16, int rounder);
+ /**
+ * global motion compensation.
+ */
+ void (*gmc )(uint8_t *dst/*align 8*/, uint8_t *src/*align 1*/, int stride, int h, int ox, int oy,
+ int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height);
+ void (*clear_blocks)(DCTELEM *blocks/*align 16*/);
+ int (*pix_sum)(uint8_t * pix, int line_size);
+ int (*pix_norm1)(uint8_t * pix, int line_size);
+ me_cmp_func sad[2]; /* identical to pix_absAxA except additional void * */
+ me_cmp_func sse[2];
+ me_cmp_func hadamard8_diff[2];
+ me_cmp_func dct_sad[2];
+ me_cmp_func quant_psnr[2];
+ me_cmp_func bit[2];
+ me_cmp_func rd[2];
+ int (*hadamard8_abs )(uint8_t *src, int stride, int mean);
+
+ me_cmp_func me_pre_cmp[11];
+ me_cmp_func me_cmp[11];
+ me_cmp_func me_sub_cmp[11];
+ me_cmp_func mb_cmp[11];
+
+ /* maybe create an array for 16/8/4/2 functions */
+ /**
+ * Halfpel motion compensation with rounding (a+b+1)>>1.
+ * this is an array[4][4] of motion compensation funcions for 4
+ * horizontal blocksizes (2,4,8,16) and the 4 halfpel positions<br>
+ * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
+ * @param block destination where the result is stored
+ * @param pixels source
+ * @param line_size number of bytes in a horizontal line of block
+ * @param h height
+ */
+ op_pixels_func put_pixels_tab[4][4];
+
+ /**
+ * Halfpel motion compensation with rounding (a+b+1)>>1.
+ * this is an array[2][4] of motion compensation funcions for 2
+ * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
+ * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
+ * @param block destination into which the result is averaged (a+b+1)>>1
+ * @param pixels source
+ * @param line_size number of bytes in a horizontal line of block
+ * @param h height
+ */
+ op_pixels_func avg_pixels_tab[2][4];
+
+ /**
+ * Halfpel motion compensation with no rounding (a+b)>>1.
+ * this is an array[2][4] of motion compensation funcions for 2
+ * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
+ * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
+ * @param block destination where the result is stored
+ * @param pixels source
+ * @param line_size number of bytes in a horizontal line of block
+ * @param h height
+ */
+ op_pixels_func put_no_rnd_pixels_tab[2][4];
+
+ /**
+ * Halfpel motion compensation with no rounding (a+b)>>1.
+ * this is an array[2][4] of motion compensation funcions for 2
+ * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
+ * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
+ * @param block destination into which the result is averaged (a+b)>>1
+ * @param pixels source
+ * @param line_size number of bytes in a horizontal line of block
+ * @param h height
+ */
+ op_pixels_func avg_no_rnd_pixels_tab[2][4];
+
+ /**
+ * Thirdpel motion compensation with rounding (a+b+1)>>1.
+ * this is an array[12] of motion compensation funcions for the 9 thirdpel positions<br>
+ * *pixels_tab[ xthirdpel + 4*ythirdpel ]
+ * @param block destination where the result is stored
+ * @param pixels source
+ * @param line_size number of bytes in a horizontal line of block
+ * @param h height
+ */
+ tpel_mc_func put_tpel_pixels_tab[11]; //FIXME individual func ptr per width?
+
+ qpel_mc_func put_qpel_pixels_tab[2][16];
+ qpel_mc_func avg_qpel_pixels_tab[2][16];
+ qpel_mc_func put_no_rnd_qpel_pixels_tab[2][16];
+ qpel_mc_func avg_no_rnd_qpel_pixels_tab[2][16];
+ qpel_mc_func put_mspel_pixels_tab[8];
+
+ /**
+ * h264 Chram MC
+ */
+ h264_chroma_mc_func put_h264_chroma_pixels_tab[3];
+ h264_chroma_mc_func avg_h264_chroma_pixels_tab[3];
+
+ qpel_mc_func put_h264_qpel_pixels_tab[3][16];
+ qpel_mc_func avg_h264_qpel_pixels_tab[3][16];
+
+ op_pixels_abs_func pix_abs16x16;
+ op_pixels_abs_func pix_abs16x16_x2;
+ op_pixels_abs_func pix_abs16x16_y2;
+ op_pixels_abs_func pix_abs16x16_xy2;
+ op_pixels_abs_func pix_abs8x8;
+ op_pixels_abs_func pix_abs8x8_x2;
+ op_pixels_abs_func pix_abs8x8_y2;
+ op_pixels_abs_func pix_abs8x8_xy2;
+
+ /* huffyuv specific */
+ void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
+ void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
+ void (*bswap_buf)(uint32_t *dst, uint32_t *src, int w);
+
+ /* (I)DCT */
+ void (*fdct)(DCTELEM *block/* align 16*/);
+
+ /**
+ * block -> idct -> clip to unsigned 8 bit -> dest.
+ * (-1392, 0, 0, ...) -> idct -> (-174, -174, ...) -> put -> (0, 0, ...)
+ * @param line_size size in bytes of a horizotal line of dest
+ */
+ void (*idct_put)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
+
+ /**
+ * block -> idct -> add dest -> clip to unsigned 8 bit -> dest.
+ * @param line_size size in bytes of a horizotal line of dest
+ */
+ void (*idct_add)(uint8_t *dest/*align 8*/, int line_size, DCTELEM *block/*align 16*/);
+
+ /**
+ * idct input permutation.
+ * several optimized IDCTs need a permutated input (relative to the normal order of the reference
+ * IDCT)
+ * this permutation must be performed before the idct_put/add, note, normally this can be merged
+ * with the zigzag/alternate scan<br>
+ * an example to avoid confusion:
+ * - (->decode coeffs -> zigzag reorder -> dequant -> reference idct ->...)
+ * - (x -> referece dct -> reference idct -> x)
+ * - (x -> referece dct -> simple_mmx_perm = idct_permutation -> simple_idct_mmx -> x)
+ * - (->decode coeffs -> zigzag reorder -> simple_mmx_perm -> dequant -> simple_idct_mmx ->...)
+ */
+ uint8_t idct_permutation[64];
+ int idct_permutation_type;
+#define FF_NO_IDCT_PERM 1
+#define FF_LIBMPEG2_IDCT_PERM 2
+#define FF_SIMPLE_IDCT_PERM 3
+#define FF_TRANSPOSE_IDCT_PERM 4
+
+} DSPContext;
+
+void dsputil_static_init(void);
+void dsputil_init(DSPContext* p, AVCodecContext *avctx);
+
+/**
+ * permute block according to permuatation.
+ * @param last last non zero element in scantable order
+ */
+void ff_block_permute(DCTELEM *block, uint8_t *permutation, const uint8_t *scantable, int last);
-int pix_abs16x16_c(UINT8 *blk1, UINT8 *blk2, int lx);
-int pix_abs16x16_x2_c(UINT8 *blk1, UINT8 *blk2, int lx);
-int pix_abs16x16_y2_c(UINT8 *blk1, UINT8 *blk2, int lx);
-int pix_abs16x16_xy2_c(UINT8 *blk1, UINT8 *blk2, int lx);
+#define BYTE_VEC32(c) ((c)*0x01010101UL)
-static inline int block_permute_op(int j)
+static inline uint32_t rnd_avg32(uint32_t a, uint32_t b)
{
- return permutation[j];
+ return (a | b) - (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
}
-void block_permute(INT16 *block);
+static inline uint32_t no_rnd_avg32(uint32_t a, uint32_t b)
+{
+ return (a & b) + (((a ^ b) & ~BYTE_VEC32(0x01)) >> 1);
+}
+
+/**
+ * Empty mmx state.
+ * this must be called between any dsp function and float/double code.
+ * for example sin(); dsp->idct_put(); emms_c(); cos()
+ */
+#define emms_c()
+
+/* should be defined by architectures supporting
+ one or more MultiMedia extension */
+int mm_support(void);
#if defined(HAVE_MMX)
+#undef emms_c
+
#define MM_MMX 0x0001 /* standard MMX */
#define MM_3DNOW 0x0004 /* AMD 3DNOW */
#define MM_MMXEXT 0x0002 /* SSE integer functions or AMD MMX ext */
extern int mm_flags;
-int mm_support(void);
+void add_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
+void put_pixels_clamped_mmx(const DCTELEM *block, uint8_t *pixels, int line_size);
static inline void emms(void)
{
__asm __volatile ("emms;":::"memory");
}
+
#define emms_c() \
{\
if (mm_flags & MM_MMX)\
#define __align8 __attribute__ ((aligned (8)))
-void dsputil_init_mmx(void);
-void dsputil_set_bit_exact_mmx(void);
+void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx);
+void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx);
#elif defined(ARCH_ARMV4L)
-#define emms_c()
-
/* This is to use 4 bytes read to the IDCT pointers for some 'zero'
line ptimizations */
#define __align8 __attribute__ ((aligned (4)))
-void dsputil_init_armv4l(void);
+void dsputil_init_armv4l(DSPContext* c, AVCodecContext *avctx);
#elif defined(HAVE_MLIB)
-
-#define emms_c()
/* SPARC/VIS IDCT needs 8-byte aligned DCT blocks */
#define __align8 __attribute__ ((aligned (8)))
-void dsputil_init_mlib(void);
+void dsputil_init_mlib(DSPContext* c, AVCodecContext *avctx);
#elif defined(ARCH_ALPHA)
-#define emms_c()
#define __align8 __attribute__ ((aligned (8)))
-void dsputil_init_alpha(void);
+void dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
-#else
+#elif defined(ARCH_POWERPC)
-#define emms_c()
+#define MM_ALTIVEC 0x0001 /* standard AltiVec */
+
+extern int mm_flags;
+
+#if defined(HAVE_ALTIVEC) && !defined(CONFIG_DARWIN)
+#include <altivec.h>
+#endif
+
+#define __align8 __attribute__ ((aligned (16)))
+
+void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx);
+
+#elif defined(HAVE_MMI)
+
+#define __align8 __attribute__ ((aligned (16)))
+
+void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx);
+
+#elif defined(ARCH_SH4)
+
+#define __align8 __attribute__ ((aligned (8)))
+
+void dsputil_init_sh4(DSPContext* c, AVCodecContext *avctx);
+
+#else
#define __align8
struct unaligned_64 { uint64_t l; } __attribute__((packed));
struct unaligned_32 { uint32_t l; } __attribute__((packed));
+struct unaligned_16 { uint16_t l; } __attribute__((packed));
+#define LD16(a) (((const struct unaligned_16 *) (a))->l)
#define LD32(a) (((const struct unaligned_32 *) (a))->l)
#define LD64(a) (((const struct unaligned_64 *) (a))->l)
#else /* __GNUC__ */
+#define LD16(a) (*((uint16_t*)(a)))
#define LD32(a) (*((uint32_t*)(a)))
#define LD64(a) (*((uint64_t*)(a)))
#endif /* !__GNUC__ */
/* PSNR */
-void get_psnr(UINT8 *orig_image[3], UINT8 *coded_image[3],
+void get_psnr(uint8_t *orig_image[3], uint8_t *coded_image[3],
int orig_linesize[3], int coded_linesize,
AVCodecContext *avctx);
-
+
+/* FFT computation */
+
+/* NOTE: soon integer code will be added, so you must use the
+ FFTSample type */
+typedef float FFTSample;
+
+typedef struct FFTComplex {
+ FFTSample re, im;
+} FFTComplex;
+
+typedef struct FFTContext {
+ int nbits;
+ int inverse;
+ uint16_t *revtab;
+ FFTComplex *exptab;
+ FFTComplex *exptab1; /* only used by SSE code */
+ void (*fft_calc)(struct FFTContext *s, FFTComplex *z);
+} FFTContext;
+
+int fft_init(FFTContext *s, int nbits, int inverse);
+void fft_permute(FFTContext *s, FFTComplex *z);
+void fft_calc_c(FFTContext *s, FFTComplex *z);
+void fft_calc_sse(FFTContext *s, FFTComplex *z);
+void fft_calc_altivec(FFTContext *s, FFTComplex *z);
+
+static inline void fft_calc(FFTContext *s, FFTComplex *z)
+{
+ s->fft_calc(s, z);
+}
+void fft_end(FFTContext *s);
+
+/* MDCT computation */
+
+typedef struct MDCTContext {
+ int n; /* size of MDCT (i.e. number of input data * 2) */
+ int nbits; /* n = 2^nbits */
+ /* pre/post rotation tables */
+ FFTSample *tcos;
+ FFTSample *tsin;
+ FFTContext fft;
+} MDCTContext;
+
+int ff_mdct_init(MDCTContext *s, int nbits, int inverse);
+void ff_imdct_calc(MDCTContext *s, FFTSample *output,
+ const FFTSample *input, FFTSample *tmp);
+void ff_mdct_calc(MDCTContext *s, FFTSample *out,
+ const FFTSample *input, FFTSample *tmp);
+void ff_mdct_end(MDCTContext *s);
+
+#define WARPER88_1616(name8, name16)\
+static int name16(void /*MpegEncContext*/ *s, uint8_t *dst, uint8_t *src, int stride){\
+ return name8(s, dst , src , stride)\
+ +name8(s, dst+8 , src+8 , stride)\
+ +name8(s, dst +8*stride, src +8*stride, stride)\
+ +name8(s, dst+8+8*stride, src+8+8*stride, stride);\
+}
+
+#ifndef HAVE_LRINTF
+/* XXX: add ISOC specific test to avoid specific BSD testing. */
+/* better than nothing implementation. */
+/* btw, rintf() is existing on fbsd too -- alex */
+static inline long int lrintf(float x)
+{
+#ifdef CONFIG_WIN32
+ /* XXX: incorrect, but make it compile */
+ return (int)(x);
+#else
+ return (int)(rint(x));
+#endif
+}
+#endif
+
#endif