libavcodec/x86/idct_sse2_xvid.c File Reference

SSE2 idct compatible with xvidmmx. More...

#include "libavcodec/dsputil.h"
#include "libavutil/x86_cpu.h"
#include "idct_xvid.h"
#include "dsputil_mmx.h"

Go to the source code of this file.

Defines

#define X8(x)   x,x,x,x,x,x,x,x

#define ROW_SHIFT   11

#define COL_SHIFT   6

#define ROW1   "%%xmm6"

#define ROW3   "%%xmm4"

#define ROW5   "%%xmm5"

#define ROW7   "%%xmm7"

#define CLEAR_ODD(r)   "pxor "r","r" \n\t"

#define PUT_ODD(dst)   "pshufhw $0x1B, %%xmm2, "dst" \n\t"

#define ROW0   "(%0)"

#define REG0   "%%xmm4"

#define ROW2   "2*16(%0)"

#define REG2   "%%xmm4"

#define ROW4   "4*16(%0)"

#define REG4   "%%xmm6"

#define ROW6   "6*16(%0)"

#define REG6   "%%xmm6"

#define CLEAR_EVEN(r)

#define PUT_EVEN(dst)

#define XMMS   "%%xmm2"

#define MOV_32_ONLY   "movdqa "

#define SREG2   "%%xmm7"

#define TAN3   "%%xmm0"

#define TAN1   "%%xmm2"

#define ROUND(x)   "paddd "MANGLE(x)

#define JZ(reg, to)

#define JNZ(reg, to)

#define TEST_ONE_ROW(src, reg, clear)

#define TEST_TWO_ROWS(row1, row2, reg1, reg2, clear1, clear2)

#define iMTX_MULT(src, table, rounder, put)

IDCT pass on rows.

#define iLLM_HEAD

#define iLLM_PASS(dct)

IDCT pass on columns.

#define iLLM_PASS_SPARSE(dct)

IDCT pass on columns, assuming rows 4-7 are zero.

Functions

DECLARE_ASM_CONST (16, int16_t, tan1)[]

DECLARE_ASM_CONST (16, int16_t, tan2)[]

DECLARE_ASM_CONST (16, int16_t, tan3)[]

DECLARE_ASM_CONST (16, int16_t, sqrt2)[]

DECLARE_ASM_CONST (8, uint8_t, m127)[]

DECLARE_ASM_CONST (16, int16_t, iTab1)[]

DECLARE_ASM_CONST (16, int16_t, iTab2)[]

DECLARE_ASM_CONST (16, int16_t, iTab3)[]

DECLARE_ASM_CONST (16, int16_t, iTab4)[]

DECLARE_ASM_CONST (16, int32_t, walkenIdctRounders)[]

void ff_idct_xvid_sse2 (short *block)

void ff_idct_xvid_sse2_put (uint8_t *dest, int line_size, short *block)

void ff_idct_xvid_sse2_add (uint8_t *dest, int line_size, short *block)

Detailed Description

SSE2 idct compatible with xvidmmx.

Definition in file idct_sse2_xvid.c.

Define Documentation

#define CLEAR_EVEN ( r )

Definition at line 136 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define CLEAR_ODD ( r ) "pxor "r","r" \n\t"

Definition at line 105 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define COL_SHIFT 6

Definition at line 54 of file idct_sse2_xvid.c.

#define iLLM_HEAD

Value:

"movdqa   "MANGLE(tan3)", "TAN3"  \n\t" \
    "movdqa   "MANGLE(tan1)", "TAN1"  \n\t" \

Definition at line 201 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define iLLM_PASS ( dct )

IDCT pass on columns.

Definition at line 206 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define iLLM_PASS_SPARSE ( dct )

IDCT pass on columns, assuming rows 4-7 are zero.

Definition at line 284 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define iMTX_MULT	(	src,
		table,
		rounder,
		put	)

Value:

"movdqa        "src", %%xmm3      \n\t" \
    "movdqa       %%xmm3, %%xmm0      \n\t" \
    "pshufd   $0x11, %%xmm3, %%xmm1   \n\t" /* 4602 */ \
    "punpcklqdq   %%xmm0, %%xmm0      \n\t" /* 0246 */ \
    "pmaddwd     "table", %%xmm0      \n\t" \
    "pmaddwd  16+"table", %%xmm1      \n\t" \
    "pshufd   $0xBB, %%xmm3, %%xmm2   \n\t" /* 5713 */ \
    "punpckhqdq   %%xmm3, %%xmm3      \n\t" /* 1357 */ \
    "pmaddwd  32+"table", %%xmm2      \n\t" \
    "pmaddwd  48+"table", %%xmm3      \n\t" \
    "paddd        %%xmm1, %%xmm0      \n\t" \
    "paddd        %%xmm3, %%xmm2      \n\t" \
    rounder",     %%xmm0              \n\t" \
    "movdqa       %%xmm2, %%xmm3      \n\t" \
    "paddd        %%xmm0, %%xmm2      \n\t" \
    "psubd        %%xmm3, %%xmm0      \n\t" \
    "psrad           $11, %%xmm2      \n\t" \
    "psrad           $11, %%xmm0      \n\t" \
    "packssdw     %%xmm0, %%xmm2      \n\t" \
    put                                     \
    "1:                               \n\t"

IDCT pass on rows.

Definition at line 178 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define JNZ	(	reg,
		to		)

Value:

"testl     "reg","reg"            \n\t" \
    "jnz       "to"                   \n\t"

Definition at line 154 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define JZ	(	reg,
		to		)

Value:

"testl     "reg","reg"            \n\t" \
    "jz        "to"                   \n\t"

Definition at line 150 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define MOV_32_ONLY "movdqa "

Definition at line 141 of file idct_sse2_xvid.c.

#define PUT_EVEN ( dst )

Value:

"pshufhw  $0x1B, %%xmm2, %%xmm2   \n\t" \
    "movdqa          %%xmm2, "dst"    \n\t"

Definition at line 137 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define PUT_ODD ( dst ) "pshufhw $0x1B, %%xmm2, "dst" \n\t"

Definition at line 106 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define REG0 "%%xmm4"

Definition at line 129 of file idct_sse2_xvid.c.

#define REG2 "%%xmm4"

Definition at line 131 of file idct_sse2_xvid.c.

#define REG4 "%%xmm6"

Definition at line 133 of file idct_sse2_xvid.c.

#define REG6 "%%xmm6"

Definition at line 135 of file idct_sse2_xvid.c.

#define ROUND ( x ) "paddd "MANGLE(x)

Definition at line 148 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define ROW0 "(%0)"

Definition at line 128 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define ROW1 "%%xmm6"

Definition at line 100 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define ROW2 "2*16(%0)"

Definition at line 130 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define ROW3 "%%xmm4"

Definition at line 101 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define ROW4 "4*16(%0)"

Definition at line 132 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define ROW5 "%%xmm5"

Definition at line 102 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define ROW6 "6*16(%0)"

Definition at line 134 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define ROW7 "%%xmm7"

Definition at line 103 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define ROW_SHIFT 11

Definition at line 53 of file idct_sse2_xvid.c.

#define SREG2 "%%xmm7"

Definition at line 142 of file idct_sse2_xvid.c.

#define TAN1 "%%xmm2"

Definition at line 144 of file idct_sse2_xvid.c.

#define TAN3 "%%xmm0"

Definition at line 143 of file idct_sse2_xvid.c.

#define TEST_ONE_ROW	(	src,
		reg,
		clear	)

Value:

clear                                   \
    "movq     "src", %%mm1            \n\t" \
    "por    8+"src", %%mm1            \n\t" \
    "paddusb  %%mm0, %%mm1            \n\t" \
    "pmovmskb %%mm1, "reg"            \n\t"

Definition at line 158 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define TEST_TWO_ROWS	(	row1,
		row2,
		reg1,
		reg2,
		clear1,
		clear2	)

Value:

clear1                                  \
    clear2                                  \
    "movq     "row1", %%mm1           \n\t" \
    "por    8+"row1", %%mm1           \n\t" \
    "movq     "row2", %%mm2           \n\t" \
    "por    8+"row2", %%mm2           \n\t" \
    "paddusb   %%mm0, %%mm1           \n\t" \
    "paddusb   %%mm0, %%mm2           \n\t" \
    "pmovmskb  %%mm1, "reg1"          \n\t" \
    "pmovmskb  %%mm2, "reg2"          \n\t"

Definition at line 165 of file idct_sse2_xvid.c.

Referenced by ff_idct_xvid_sse2().

#define X8 ( x ) x,x,x,x,x,x,x,x

Definition at line 51 of file idct_sse2_xvid.c.

#define XMMS "%%xmm2"

Definition at line 140 of file idct_sse2_xvid.c.

Function Documentation

DECLARE_ASM_CONST	(	16	,
		int32_t	,
		walkenIdctRounders
	)

DECLARE_ASM_CONST	(	16	,
		int16_t	,
		iTab4
	)

DECLARE_ASM_CONST	(	16	,
		int16_t	,
		iTab3
	)

DECLARE_ASM_CONST	(	16	,
		int16_t	,
		iTab2
	)

DECLARE_ASM_CONST	(	16	,
		int16_t	,
		iTab1
	)

DECLARE_ASM_CONST	(	8	,
		uint8_t	,
		m127
	)

DECLARE_ASM_CONST	(	16	,
		int16_t	,
		sqrt2
	)

DECLARE_ASM_CONST	(	16	,
		int16_t	,
		tan3
	)

DECLARE_ASM_CONST	(	16	,
		int16_t	,
		tan2
	)

DECLARE_ASM_CONST	(	16	,
		int16_t	,
		tan1
	)

void ff_idct_xvid_sse2 ( short * block ) [inline]

Definition at line 344 of file idct_sse2_xvid.c.

Referenced by dsputil_init_mmx(), ff_idct_xvid_sse2_add(), and ff_idct_xvid_sse2_put().

void ff_idct_xvid_sse2_add	(	uint8_t *	dest,
		int	line_size,
		short *	block
	)

Definition at line 399 of file idct_sse2_xvid.c.

Referenced by dsputil_init_mmx().

void ff_idct_xvid_sse2_put	(	uint8_t *	dest,
		int	line_size,
		short *	block
	)

Definition at line 393 of file idct_sse2_xvid.c.

Referenced by dsputil_init_mmx().


Defines
#define	X8(x) x,x,x,x,x,x,x,x
#define	ROW_SHIFT 11
#define	COL_SHIFT 6
#define	ROW1 "%%xmm6"
#define	ROW3 "%%xmm4"
#define	ROW5 "%%xmm5"
#define	ROW7 "%%xmm7"
#define	CLEAR_ODD(r) "pxor "r","r" \n\t"
#define	PUT_ODD(dst) "pshufhw $0x1B, %%xmm2, "dst" \n\t"
#define	ROW0 "(%0)"
#define	REG0 "%%xmm4"
#define	ROW2 "2*16(%0)"
#define	REG2 "%%xmm4"
#define	ROW4 "4*16(%0)"
#define	REG4 "%%xmm6"
#define	ROW6 "6*16(%0)"
#define	REG6 "%%xmm6"
#define	CLEAR_EVEN(r)
#define	PUT_EVEN(dst)
#define	XMMS "%%xmm2"
#define	MOV_32_ONLY "movdqa "
#define	SREG2 "%%xmm7"
#define	TAN3 "%%xmm0"
#define	TAN1 "%%xmm2"
#define	ROUND(x) "paddd "MANGLE(x)
#define	JZ(reg, to)
#define	JNZ(reg, to)
#define	TEST_ONE_ROW(src, reg, clear)
#define	TEST_TWO_ROWS(row1, row2, reg1, reg2, clear1, clear2)
#define	iMTX_MULT(src, table, rounder, put)
	IDCT pass on rows.
#define	iLLM_HEAD
#define	iLLM_PASS(dct)
	IDCT pass on columns.
#define	iLLM_PASS_SPARSE(dct)
	IDCT pass on columns, assuming rows 4-7 are zero.
Functions
	DECLARE_ASM_CONST (16, int16_t, tan1)[]
	DECLARE_ASM_CONST (16, int16_t, tan2)[]
	DECLARE_ASM_CONST (16, int16_t, tan3)[]
	DECLARE_ASM_CONST (16, int16_t, sqrt2)[]
	DECLARE_ASM_CONST (8, uint8_t, m127)[]
	DECLARE_ASM_CONST (16, int16_t, iTab1)[]
	DECLARE_ASM_CONST (16, int16_t, iTab2)[]
	DECLARE_ASM_CONST (16, int16_t, iTab3)[]
	DECLARE_ASM_CONST (16, int16_t, iTab4)[]
	DECLARE_ASM_CONST (16, int32_t, walkenIdctRounders)[]
void	ff_idct_xvid_sse2 (short *block)
void	ff_idct_xvid_sse2_put (uint8_t dest, int line_size, short block)
void	ff_idct_xvid_sse2_add (uint8_t dest, int line_size, short block)