cextern pd_1
pd_151: times 4 dd 151
-; used in ff_apply_window_int16()
-pb_revwords: SHUFFLE_MASK_W 7, 6, 5, 4, 3, 2, 1, 0
-pd_16384: times 4 dd 16384
-
SECTION .text
;-----------------------------------------------------------------------------
%endif
%undef LOOP_ALIGN
-;-----------------------------------------------------------------------------
-; int ff_ac3_max_msb_abs_int16(const int16_t *src, int len)
-;
-; This function uses 2 different methods to calculate a valid result.
-; 1) logical 'or' of abs of each element
-; This is used for ssse3 because of the pabsw instruction.
-; It is also used for mmx because of the lack of min/max instructions.
-; 2) calculate min/max for the array, then or(abs(min),abs(max))
-; This is used for mmxext and sse2 because they have pminsw/pmaxsw.
-;-----------------------------------------------------------------------------
-
-; logical 'or' of 4 or 8 words in an mmx or xmm register into the low word
-%macro OR_WORDS_HORIZ 2 ; src, tmp
-%if cpuflag(sse2)
- movhlps %2, %1
- por %1, %2
- pshuflw %2, %1, q0032
- por %1, %2
- pshuflw %2, %1, q0001
- por %1, %2
-%elif cpuflag(mmxext)
- pshufw %2, %1, q0032
- por %1, %2
- pshufw %2, %1, q0001
- por %1, %2
-%else ; mmx
- movq %2, %1
- psrlq %2, 32
- por %1, %2
- movq %2, %1
- psrlq %2, 16
- por %1, %2
-%endif
-%endmacro
-
-%macro AC3_MAX_MSB_ABS_INT16 1
-cglobal ac3_max_msb_abs_int16, 2,2,5, src, len
- pxor m2, m2
- pxor m3, m3
-.loop:
-%ifidn %1, min_max
- mova m0, [srcq]
- mova m1, [srcq+mmsize]
- pminsw m2, m0
- pminsw m2, m1
- pmaxsw m3, m0
- pmaxsw m3, m1
-%else ; or_abs
-%if notcpuflag(ssse3)
- mova m0, [srcq]
- mova m1, [srcq+mmsize]
- ABS2 m0, m1, m3, m4
-%else ; ssse3
- ; using memory args is faster for ssse3
- pabsw m0, [srcq]
- pabsw m1, [srcq+mmsize]
-%endif
- por m2, m0
- por m2, m1
-%endif
- add srcq, mmsize*2
- sub lend, mmsize
- ja .loop
-%ifidn %1, min_max
- ABS2 m2, m3, m0, m1
- por m2, m3
-%endif
- OR_WORDS_HORIZ m2, m0
- movd eax, m2
- and eax, 0xFFFF
- RET
-%endmacro
-
-INIT_MMX mmx
-AC3_MAX_MSB_ABS_INT16 or_abs
-INIT_MMX mmxext
-AC3_MAX_MSB_ABS_INT16 min_max
-INIT_XMM sse2
-AC3_MAX_MSB_ABS_INT16 min_max
-INIT_XMM ssse3
-AC3_MAX_MSB_ABS_INT16 or_abs
-
-;-----------------------------------------------------------------------------
-; macro used for ff_ac3_lshift_int16() and ff_ac3_rshift_int32()
-;-----------------------------------------------------------------------------
-
-%macro AC3_SHIFT 3 ; l/r, 16/32, shift instruction, instruction set
-cglobal ac3_%1shift_int%2, 3, 3, 5, src, len, shift
- movd m0, shiftd
-.loop:
- mova m1, [srcq ]
- mova m2, [srcq+mmsize ]
- mova m3, [srcq+mmsize*2]
- mova m4, [srcq+mmsize*3]
- %3 m1, m0
- %3 m2, m0
- %3 m3, m0
- %3 m4, m0
- mova [srcq ], m1
- mova [srcq+mmsize ], m2
- mova [srcq+mmsize*2], m3
- mova [srcq+mmsize*3], m4
- add srcq, mmsize*4
- sub lend, mmsize*32/%2
- ja .loop
-.end:
- REP_RET
-%endmacro
-
-;-----------------------------------------------------------------------------
-; void ff_ac3_lshift_int16(int16_t *src, unsigned int len, unsigned int shift)
-;-----------------------------------------------------------------------------
-
-INIT_MMX mmx
-AC3_SHIFT l, 16, psllw
-INIT_XMM sse2
-AC3_SHIFT l, 16, psllw
-
-;-----------------------------------------------------------------------------
-; void ff_ac3_rshift_int32(int32_t *src, unsigned int len, unsigned int shift)
-;-----------------------------------------------------------------------------
-
-INIT_MMX mmx
-AC3_SHIFT r, 32, psrad
-INIT_XMM sse2
-AC3_SHIFT r, 32, psrad
-
;-----------------------------------------------------------------------------
; void ff_float_to_fixed24(int32_t *dst, const float *src, unsigned int len)
;-----------------------------------------------------------------------------
INIT_XMM ssse3
AC3_EXTRACT_EXPONENTS
%endif
-
-;-----------------------------------------------------------------------------
-; void ff_apply_window_int16(int16_t *output, const int16_t *input,
-; const int16_t *window, unsigned int len)
-;-----------------------------------------------------------------------------
-
-%macro REVERSE_WORDS 1-2
-%if cpuflag(ssse3) && notcpuflag(atom)
- pshufb %1, %2
-%elif cpuflag(sse2)
- pshuflw %1, %1, 0x1B
- pshufhw %1, %1, 0x1B
- pshufd %1, %1, 0x4E
-%elif cpuflag(mmxext)
- pshufw %1, %1, 0x1B
-%endif
-%endmacro
-
-%macro MUL16FIXED 3
-%if cpuflag(ssse3) ; dst, src, unused
-; dst = ((dst * src) + (1<<14)) >> 15
- pmulhrsw %1, %2
-%elif cpuflag(mmxext) ; dst, src, temp
-; dst = (dst * src) >> 15
-; pmulhw cuts off the bottom bit, so we have to lshift by 1 and add it back
-; in from the pmullw result.
- mova %3, %1
- pmulhw %1, %2
- pmullw %3, %2
- psrlw %3, 15
- psllw %1, 1
- por %1, %3
-%endif
-%endmacro
-
-%macro APPLY_WINDOW_INT16 1 ; %1 bitexact version
-%if %1
-cglobal apply_window_int16, 4,5,6, output, input, window, offset, offset2
-%else
-cglobal apply_window_int16_round, 4,5,6, output, input, window, offset, offset2
-%endif
- lea offset2q, [offsetq-mmsize]
-%if cpuflag(ssse3) && notcpuflag(atom)
- mova m5, [pb_revwords]
- ALIGN 16
-%elif %1
- mova m5, [pd_16384]
-%endif
-.loop:
-%if cpuflag(ssse3)
- ; This version does the 16x16->16 multiplication in-place without expanding
- ; to 32-bit. The ssse3 version is bit-identical.
- mova m0, [windowq+offset2q]
- mova m1, [ inputq+offset2q]
- pmulhrsw m1, m0
- REVERSE_WORDS m0, m5
- pmulhrsw m0, [ inputq+offsetq ]
- mova [outputq+offset2q], m1
- mova [outputq+offsetq ], m0
-%elif %1
- ; This version expands 16-bit to 32-bit, multiplies by the window,
- ; adds 16384 for rounding, right shifts 15, then repacks back to words to
- ; save to the output. The window is reversed for the second half.
- mova m3, [windowq+offset2q]
- mova m4, [ inputq+offset2q]
- pxor m0, m0
- punpcklwd m0, m3
- punpcklwd m1, m4
- pmaddwd m0, m1
- paddd m0, m5
- psrad m0, 15
- pxor m2, m2
- punpckhwd m2, m3
- punpckhwd m1, m4
- pmaddwd m2, m1
- paddd m2, m5
- psrad m2, 15
- packssdw m0, m2
- mova [outputq+offset2q], m0
- REVERSE_WORDS m3
- mova m4, [ inputq+offsetq]
- pxor m0, m0
- punpcklwd m0, m3
- punpcklwd m1, m4
- pmaddwd m0, m1
- paddd m0, m5
- psrad m0, 15
- pxor m2, m2
- punpckhwd m2, m3
- punpckhwd m1, m4
- pmaddwd m2, m1
- paddd m2, m5
- psrad m2, 15
- packssdw m0, m2
- mova [outputq+offsetq], m0
-%else
- ; This version does the 16x16->16 multiplication in-place without expanding
- ; to 32-bit. The mmxext and sse2 versions do not use rounding, and
- ; therefore are not bit-identical to the C version.
- mova m0, [windowq+offset2q]
- mova m1, [ inputq+offset2q]
- mova m2, [ inputq+offsetq ]
- MUL16FIXED m1, m0, m3
- REVERSE_WORDS m0
- MUL16FIXED m2, m0, m3
- mova [outputq+offset2q], m1
- mova [outputq+offsetq ], m2
-%endif
- add offsetd, mmsize
- sub offset2d, mmsize
- jae .loop
- REP_RET
-%endmacro
-
-INIT_MMX mmxext
-APPLY_WINDOW_INT16 0
-INIT_XMM sse2
-APPLY_WINDOW_INT16 0
-
-INIT_MMX mmxext
-APPLY_WINDOW_INT16 1
-INIT_XMM sse2
-APPLY_WINDOW_INT16 1
-INIT_XMM ssse3
-APPLY_WINDOW_INT16 1
-INIT_XMM ssse3, atom
-APPLY_WINDOW_INT16 1
projects / ffmpeg / blobdiff