;*****************************************************************************
;* pixel.asm: x86 pixel metrics
;*****************************************************************************
-;* Copyright (C) 2003-2011 x264 project
+;* Copyright (C) 2003-2013 x264 project
;*
;* Authors: Loren Merritt <lorenm@u.washington.edu>
;* Holger Lubitz <holger@lubitz.org>
times 4 db 1, -1
mask_10: times 4 dw 0, -1
mask_1100: times 2 dd 0, -1
+pb_pppm: times 4 db 1,1,1,-1
deinterleave_shuf: db 0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15
-
+intrax3_shuf: db 7,6,7,6,5,4,5,4,3,2,3,2,1,0,1,0
+
+intrax9a_ddlr1: db 6, 7, 8, 9, 7, 8, 9,10, 4, 5, 6, 7, 3, 4, 5, 6
+intrax9a_ddlr2: db 8, 9,10,11, 9,10,11,12, 2, 3, 4, 5, 1, 2, 3, 4
+intrax9a_hdu1: db 15, 4, 5, 6,14, 3,15, 4,14, 2,13, 1,13, 1,12, 0
+intrax9a_hdu2: db 13, 2,14, 3,12, 1,13, 2,12, 0,11,11,11,11,11,11
+intrax9a_vrl1: db 10,11,12,13, 3, 4, 5, 6,11,12,13,14, 5, 6, 7, 8
+intrax9a_vrl2: db 2,10,11,12, 1, 3, 4, 5,12,13,14,15, 6, 7, 8, 9
+intrax9a_vh1: db 6, 7, 8, 9, 6, 7, 8, 9, 4, 4, 4, 4, 3, 3, 3, 3
+intrax9a_vh2: db 6, 7, 8, 9, 6, 7, 8, 9, 2, 2, 2, 2, 1, 1, 1, 1
+intrax9a_dc: db 1, 2, 3, 4, 6, 7, 8, 9,-1,-1,-1,-1,-1,-1,-1,-1
+intrax9a_lut: db 0x60,0x68,0x80,0x00,0x08,0x20,0x40,0x28,0x48,0,0,0,0,0,0,0
+pw_s01234567: dw 0x8000,0x8001,0x8002,0x8003,0x8004,0x8005,0x8006,0x8007
+pw_s01234657: dw 0x8000,0x8001,0x8002,0x8003,0x8004,0x8006,0x8005,0x8007
+intrax9_edge: db 0, 0, 1, 2, 3, 7, 8, 9,10,11,12,13,14,15,15,15
+
+intrax9b_ddlr1: db 6, 7, 8, 9, 4, 5, 6, 7, 7, 8, 9,10, 3, 4, 5, 6
+intrax9b_ddlr2: db 8, 9,10,11, 2, 3, 4, 5, 9,10,11,12, 1, 2, 3, 4
+intrax9b_hdu1: db 15, 4, 5, 6,14, 2,13, 1,14, 3,15, 4,13, 1,12, 0
+intrax9b_hdu2: db 13, 2,14, 3,12, 0,11,11,12, 1,13, 2,11,11,11,11
+intrax9b_vrl1: db 10,11,12,13,11,12,13,14, 3, 4, 5, 6, 5, 6, 7, 8
+intrax9b_vrl2: db 2,10,11,12,12,13,14,15, 1, 3, 4, 5, 6, 7, 8, 9
+intrax9b_vh1: db 6, 7, 8, 9, 4, 4, 4, 4, 6, 7, 8, 9, 3, 3, 3, 3
+intrax9b_vh2: db 6, 7, 8, 9, 2, 2, 2, 2, 6, 7, 8, 9, 1, 1, 1, 1
+intrax9b_edge2: db 6, 7, 8, 9, 6, 7, 8, 9, 4, 3, 2, 1, 4, 3, 2, 1
+intrax9b_v1: db 0, 1,-1,-1,-1,-1,-1,-1, 4, 5,-1,-1,-1,-1,-1,-1
+intrax9b_v2: db 2, 3,-1,-1,-1,-1,-1,-1, 6, 7,-1,-1,-1,-1,-1,-1
+intrax9b_lut: db 0x60,0x64,0x80,0x00,0x04,0x20,0x40,0x24,0x44,0,0,0,0,0,0,0
+
+intra8x9_h1: db 7, 7, 7, 7, 7, 7, 7, 7, 5, 5, 5, 5, 5, 5, 5, 5
+intra8x9_h2: db 6, 6, 6, 6, 6, 6, 6, 6, 4, 4, 4, 4, 4, 4, 4, 4
+intra8x9_h3: db 3, 3, 3, 3, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1
+intra8x9_h4: db 2, 2, 2, 2, 2, 2, 2, 2, 0, 0, 0, 0, 0, 0, 0, 0
+intra8x9_ddl1: db 1, 2, 3, 4, 5, 6, 7, 8, 3, 4, 5, 6, 7, 8, 9,10
+intra8x9_ddl2: db 2, 3, 4, 5, 6, 7, 8, 9, 4, 5, 6, 7, 8, 9,10,11
+intra8x9_ddl3: db 5, 6, 7, 8, 9,10,11,12, 7, 8, 9,10,11,12,13,14
+intra8x9_ddl4: db 6, 7, 8, 9,10,11,12,13, 8, 9,10,11,12,13,14,15
+intra8x9_vl1: db 0, 1, 2, 3, 4, 5, 6, 7, 1, 2, 3, 4, 5, 6, 7, 8
+intra8x9_vl2: db 1, 2, 3, 4, 5, 6, 7, 8, 2, 3, 4, 5, 6, 7, 8, 9
+intra8x9_vl3: db 2, 3, 4, 5, 6, 7, 8, 9, 3, 4, 5, 6, 7, 8, 9,10
+intra8x9_vl4: db 3, 4, 5, 6, 7, 8, 9,10, 4, 5, 6, 7, 8, 9,10,11
+intra8x9_ddr1: db 8, 9,10,11,12,13,14,15, 6, 7, 8, 9,10,11,12,13
+intra8x9_ddr2: db 7, 8, 9,10,11,12,13,14, 5, 6, 7, 8, 9,10,11,12
+intra8x9_ddr3: db 4, 5, 6, 7, 8, 9,10,11, 2, 3, 4, 5, 6, 7, 8, 9
+intra8x9_ddr4: db 3, 4, 5, 6, 7, 8, 9,10, 1, 2, 3, 4, 5, 6, 7, 8
+intra8x9_vr1: db 8, 9,10,11,12,13,14,15, 7, 8, 9,10,11,12,13,14
+intra8x9_vr2: db 8, 9,10,11,12,13,14,15, 6, 8, 9,10,11,12,13,14
+intra8x9_vr3: db 5, 7, 8, 9,10,11,12,13, 3, 5, 7, 8, 9,10,11,12
+intra8x9_vr4: db 4, 6, 8, 9,10,11,12,13, 2, 4, 6, 8, 9,10,11,12
+intra8x9_hd1: db 3, 8, 9,10,11,12,13,14, 1, 6, 2, 7, 3, 8, 9,10
+intra8x9_hd2: db 2, 7, 3, 8, 9,10,11,12, 0, 5, 1, 6, 2, 7, 3, 8
+intra8x9_hd3: db 7, 8, 9,10,11,12,13,14, 3, 4, 5, 6, 7, 8, 9,10
+intra8x9_hd4: db 5, 6, 7, 8, 9,10,11,12, 1, 2, 3, 4, 5, 6, 7, 8
+intra8x9_hu1: db 13,12,11,10, 9, 8, 7, 6, 9, 8, 7, 6, 5, 4, 3, 2
+intra8x9_hu2: db 11,10, 9, 8, 7, 6, 5, 4, 7, 6, 5, 4, 3, 2, 1, 0
+intra8x9_hu3: db 5, 4, 3, 2, 1, 0,15,15, 1, 0,15,15,15,15,15,15
+intra8x9_hu4: db 3, 2, 1, 0,15,15,15,15,15,15,15,15,15,15,15,15
+pw_s00112233: dw 0x8000,0x8000,0x8001,0x8001,0x8002,0x8002,0x8003,0x8003
+pw_s00001111: dw 0x8000,0x8000,0x8000,0x8000,0x8001,0x8001,0x8001,0x8001
+
+transd_shuf1: SHUFFLE_MASK_W 0, 8, 2, 10, 4, 12, 6, 14
+transd_shuf2: SHUFFLE_MASK_W 1, 9, 3, 11, 5, 13, 7, 15
+
+sw_f0: dq 0xfff0, 0
+sq_0f: dq 0xffffffff, 0
pd_f0: times 4 dd 0xffff0000
-sq_0f: times 1 dq 0xffffffff
SECTION .text
+cextern pb_0
+cextern pb_1
cextern pw_1
+cextern pw_8
+cextern pw_16
+cextern pw_32
cextern pw_00ff
-
+cextern pw_ppppmmmm
+cextern pw_ppmmppmm
+cextern pw_pmpmpmpm
+cextern pw_pmmpzzzz
cextern hsub_mul
;=============================================================================
; SSD
;=============================================================================
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
;-----------------------------------------------------------------------------
-; int pixel_ssd_MxN( uint16_t *, int, uint16_t *, int )
+; int pixel_ssd_MxN( uint16_t *, intptr_t, uint16_t *, intptr_t )
;-----------------------------------------------------------------------------
-%macro SSD_ONE 3
-cglobal pixel_ssd_%1x%2_%3, 4,5,6*(mmsize/16)
+%macro SSD_ONE 2
+cglobal pixel_ssd_%1x%2, 4,5,6
mov r4, %1*%2/mmsize
pxor m0, m0
.loop
%define offset mmsize
%define num_rows 1
%endif
+ lea r0, [r0+r1*2*num_rows]
psubw m1, [r2]
psubw m3, [r2+offset]
+ lea r2, [r2+r3*2*num_rows]
pmaddwd m1, m1
pmaddwd m3, m3
- dec r4
- lea r0, [r0+r1*2*num_rows]
- lea r2, [r2+r3*2*num_rows]
paddd m0, m1
paddd m0, m3
+ dec r4
jg .loop
HADDD m0, m5
movd eax, m0
%endmacro
%macro SSD_16_MMX 2
-cglobal pixel_ssd_%1x%2_mmxext, 4,5
+cglobal pixel_ssd_%1x%2, 4,5
mov r4, %1*%2/mmsize/2
pxor m0, m0
.loop
psubw m7, m2
pmaddwd m3, m3
pmaddwd m5, m5
- dec r4
lea r0, [r0+r1*2]
lea r2, [r2+r3*2]
pmaddwd m7, m7
paddd m5, m7
paddd m0, m1
paddd m0, m5
+ dec r4
jg .loop
HADDD m0, m7
movd eax, m0
RET
%endmacro
-INIT_MMX
-SSD_ONE 4, 4, mmxext
-SSD_ONE 4, 8, mmxext
-SSD_ONE 8, 4, mmxext
-SSD_ONE 8, 8, mmxext
-SSD_ONE 8, 16, mmxext
+INIT_MMX mmx2
+SSD_ONE 4, 4
+SSD_ONE 4, 8
+SSD_ONE 4, 16
+SSD_ONE 8, 4
+SSD_ONE 8, 8
+SSD_ONE 8, 16
SSD_16_MMX 16, 8
SSD_16_MMX 16, 16
-INIT_XMM
-SSD_ONE 8, 4, sse2
-SSD_ONE 8, 8, sse2
-SSD_ONE 8, 16, sse2
-SSD_ONE 16, 8, sse2
-SSD_ONE 16, 16, sse2
+INIT_XMM sse2
+SSD_ONE 8, 4
+SSD_ONE 8, 8
+SSD_ONE 8, 16
+SSD_ONE 16, 8
+SSD_ONE 16, 16
%endif ; HIGH_BIT_DEPTH
-%ifndef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH == 0
%macro SSD_LOAD_FULL 5
mova m1, [t0+%1]
mova m2, [t2+%2]
psubusb m%3, m%7
por m%1, m%2
por m%3, m%4
- mova m%2, m%1
- mova m%4, m%3
+ punpcklbw m%2, m%1, m%5
punpckhbw m%1, m%5
- punpcklbw m%2, m%5
+ punpcklbw m%4, m%3, m%5
punpckhbw m%3, m%5
- punpcklbw m%4, m%5
%endif
pmaddwd m%1, m%1
pmaddwd m%2, m%2
%macro SSD_CORE_SSSE3 7-8
%ifidn %8, FULL
- mova m%6, m%1
- mova m%7, m%3
+ punpckhbw m%6, m%1, m%2
+ punpckhbw m%7, m%3, m%4
punpcklbw m%1, m%2
punpcklbw m%3, m%4
- punpckhbw m%6, m%2
- punpckhbw m%7, m%4
SWAP %6, %2, %3
SWAP %7, %4
%endif
%endmacro
;-----------------------------------------------------------------------------
-; int pixel_ssd_16x16( uint8_t *, int, uint8_t *, int )
+; int pixel_ssd_16x16( uint8_t *, intptr_t, uint8_t *, intptr_t )
;-----------------------------------------------------------------------------
-%macro SSD 3-4 0
+%macro SSD 2
%if %1 != %2
%assign function_align 8
%else
%assign function_align 16
%endif
-cglobal pixel_ssd_%1x%2_%3, 0,0,0
+cglobal pixel_ssd_%1x%2, 0,0,0
mov al, %1*%2/mmsize/2
%if %1 != %2
- jmp mangle(x264_pixel_ssd_%1x%1_%3.startloop)
+ jmp mangle(x264_pixel_ssd_%1x%1 %+ SUFFIX %+ .startloop)
%else
.startloop:
-%ifdef ARCH_X86_64
+%if ARCH_X86_64
DECLARE_REG_TMP 0,1,2,3
-%ifnidn %3, mmx
PROLOGUE 0,0,8
-%endif
%else
PROLOGUE 0,5
DECLARE_REG_TMP 1,2,3,4
mov t3, r3m
%endif
-%ifidn %3, ssse3
+%if cpuflag(ssse3)
mova m7, [hsub_mul]
-%elifidn %3, sse2
+%elifidn cpuname, sse2
mova m7, [pw_00ff]
%elif %1 >= mmsize
pxor m7, m7
%endif
%endmacro
-INIT_MMX
-SSD 16, 16, mmx
-SSD 16, 8, mmx
-SSD 8, 8, mmx
-SSD 8, 16, mmx
-SSD 4, 4, mmx
-SSD 8, 4, mmx
-SSD 4, 8, mmx
-INIT_XMM
-SSD 16, 16, sse2slow, 8
-SSD 8, 8, sse2slow, 8
-SSD 16, 8, sse2slow, 8
-SSD 8, 16, sse2slow, 8
-SSD 8, 4, sse2slow, 8
+INIT_MMX mmx
+SSD 16, 16
+SSD 16, 8
+SSD 8, 8
+SSD 8, 16
+SSD 4, 4
+SSD 8, 4
+SSD 4, 8
+SSD 4, 16
+INIT_XMM sse2slow
+SSD 16, 16
+SSD 8, 8
+SSD 16, 8
+SSD 8, 16
+SSD 8, 4
+INIT_XMM sse2
%define SSD_CORE SSD_CORE_SSE2
%define JOIN JOIN_SSE2
-SSD 16, 16, sse2, 8
-SSD 8, 8, sse2, 8
-SSD 16, 8, sse2, 8
-SSD 8, 16, sse2, 8
-SSD 8, 4, sse2, 8
+SSD 16, 16
+SSD 8, 8
+SSD 16, 8
+SSD 8, 16
+SSD 8, 4
+INIT_XMM ssse3
%define SSD_CORE SSD_CORE_SSSE3
%define JOIN JOIN_SSSE3
-SSD 16, 16, ssse3, 8
-SSD 8, 8, ssse3, 8
-SSD 16, 8, ssse3, 8
-SSD 8, 16, ssse3, 8
-SSD 8, 4, ssse3, 8
-INIT_MMX
-SSD 4, 4, ssse3
-SSD 4, 8, ssse3
+SSD 16, 16
+SSD 8, 8
+SSD 16, 8
+SSD 8, 16
+SSD 8, 4
+INIT_XMM avx
+SSD 16, 16
+SSD 8, 8
+SSD 16, 8
+SSD 8, 16
+SSD 8, 4
+INIT_MMX ssse3
+SSD 4, 4
+SSD 4, 8
+SSD 4, 16
+INIT_XMM xop
+SSD 16, 16
+SSD 8, 8
+SSD 16, 8
+SSD 8, 16
+SSD 8, 4
%assign function_align 16
%endif ; !HIGH_BIT_DEPTH
;-----------------------------------------------------------------------------
-; void pixel_ssd_nv12_core( uint16_t *pixuv1, int stride1, uint16_t *pixuv2, int stride2,
+; void pixel_ssd_nv12_core( uint16_t *pixuv1, intptr_t stride1, uint16_t *pixuv2, intptr_t stride2,
; int width, int height, uint64_t *ssd_u, uint64_t *ssd_v )
;
; The maximum width this function can handle without risk of overflow is given
; For 10-bit MMX this means width >= 16416 and for XMM >= 32832. At sane
; distortion levels it will take much more than that though.
;-----------------------------------------------------------------------------
-%ifdef HIGH_BIT_DEPTH
-%macro SSD_NV12 1-2 0
-cglobal pixel_ssd_nv12_core_%1, 6,7,7*(mmsize/16)
+%if HIGH_BIT_DEPTH
+%macro SSD_NV12 0
+cglobal pixel_ssd_nv12_core, 6,7,7
shl r4d, 2
FIX_STRIDES r1, r3
add r0, r4
mova m1, [r0+r6+mmsize]
psubw m0, [r2+r6]
psubw m1, [r2+r6+mmsize]
-%if mmsize==8
- pshufw m0, m0, 11011000b
- pshufw m1, m1, 11011000b
-%else
- pshuflw m0, m0, 11011000b
- pshuflw m1, m1, 11011000b
- pshufhw m0, m0, 11011000b
- pshufhw m1, m1, 11011000b
+ PSHUFLW m0, m0, q3120
+ PSHUFLW m1, m1, q3120
+%if mmsize==16
+ pshufhw m0, m0, q3120
+ pshufhw m1, m1, q3120
%endif
pmaddwd m0, m0
pmaddwd m1, m1
jl .loopx
%if mmsize==16 ; using HADDD would remove the mmsize/32 part from the
; equation above, putting the width limit at 8208
- mova m0, m2
- mova m1, m3
+ punpckhdq m0, m2, m6
+ punpckhdq m1, m3, m6
punpckldq m2, m6
punpckldq m3, m6
- punpckhdq m0, m6
- punpckhdq m1, m6
paddq m3, m2
paddq m1, m0
paddq m4, m3
paddq m4, m1
%else ; unfortunately paddq is sse2
- ; emulate 48 bit precision for mmxext instead
+ ; emulate 48 bit precision for mmx2 instead
mova m0, m2
mova m1, m3
punpcklwd m2, m6
%if mmsize==16
movq [r3], m4
movhps [r4], m4
-%else ; fixup for mmxext
+%else ; fixup for mmx2
SBUTTERFLY dq, 4, 5, 0
mova m0, m4
psrld m4, 16
%endmacro ; SSD_NV12
%endif ; HIGH_BIT_DEPTH
-%ifndef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH == 0
;-----------------------------------------------------------------------------
-; void pixel_ssd_nv12_core( uint8_t *pixuv1, int stride1, uint8_t *pixuv2, int stride2,
+; void pixel_ssd_nv12_core( uint8_t *pixuv1, intptr_t stride1, uint8_t *pixuv2, intptr_t stride2,
; int width, int height, uint64_t *ssd_u, uint64_t *ssd_v )
;
; This implementation can potentially overflow on image widths >= 11008 (or
; 6604 if interlaced), since it is called on blocks of height up to 12 (resp
; 20). At sane distortion levels it will take much more than that though.
;-----------------------------------------------------------------------------
-%macro SSD_NV12 1-2 0
-cglobal pixel_ssd_nv12_core_%1, 6,7
+%macro SSD_NV12 0
+cglobal pixel_ssd_nv12_core, 6,7
shl r4d, 1
add r0, r4
add r2, r4
psubusb m0, m1
psubusb m1, [r0+r6]
por m0, m1
- mova m2, m0
+ psrlw m2, m0, 8
pand m0, m5
- psrlw m2, 8
- pmaddwd m0, m0
pmaddwd m2, m2
+ pmaddwd m0, m0
paddd m3, m0
paddd m4, m2
add r6, mmsize
movq [r4], m4
RET
%endmacro ; SSD_NV12
-%endif ; !X264_HIGHT_BIT_DEPTH
+%endif ; !HIGH_BIT_DEPTH
-INIT_MMX
-SSD_NV12 mmxext
-INIT_XMM
-SSD_NV12 sse2
+INIT_MMX mmx2
+SSD_NV12
+INIT_XMM sse2
+SSD_NV12
+INIT_XMM avx
+SSD_NV12
;=============================================================================
; variance
%macro VAR_START 1
pxor m5, m5 ; sum
pxor m6, m6 ; sum squared
-%ifndef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH == 0
%if %1
mova m7, [pw_00ff]
%else
%endmacro
%macro VAR_END 2
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
%if mmsize == 8 && %1*%2 == 256
HADDUW m5, m2
%else
movd eax, m5
HADDD m6, m1
movd edx, m6
-%ifdef ARCH_X86_64
+%if ARCH_X86_64
shl rdx, 32
add rax, rdx
%endif
%macro VAR_2ROW 2
mov r2d, %2
.loop:
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
mova m0, [r0]
mova m1, [r0+mmsize]
mova m3, [r0+%1]
mova m4, [r0+%1+mmsize]
%else ; !HIGH_BIT_DEPTH
mova m0, [r0]
- mova m1, m0
+ punpckhbw m1, m0, m7
mova m3, [r0+%1]
mova m4, m3
punpcklbw m0, m7
- punpckhbw m1, m7
%endif ; HIGH_BIT_DEPTH
%ifidn %1, r1
lea r0, [r0+%1*2]
%else
add r0, r1
%endif
-%ifndef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH == 0
punpcklbw m3, m7
punpckhbw m4, m7
%endif ; !HIGH_BIT_DEPTH
- dec r2d
VAR_CORE
+ dec r2d
jg .loop
%endmacro
;-----------------------------------------------------------------------------
-; int pixel_var_wxh( uint8_t *, int )
+; int pixel_var_wxh( uint8_t *, intptr_t )
;-----------------------------------------------------------------------------
-INIT_MMX
-cglobal pixel_var_16x16_mmxext, 2,3
+INIT_MMX mmx2
+cglobal pixel_var_16x16, 2,3
FIX_STRIDES r1
VAR_START 0
VAR_2ROW 8*SIZEOF_PIXEL, 16
VAR_END 16, 16
-cglobal pixel_var_8x8_mmxext, 2,3
+cglobal pixel_var_8x16, 2,3
+ FIX_STRIDES r1
+ VAR_START 0
+ VAR_2ROW r1, 8
+ VAR_END 8, 16
+
+cglobal pixel_var_8x8, 2,3
FIX_STRIDES r1
VAR_START 0
VAR_2ROW r1, 4
VAR_END 8, 8
-INIT_XMM
-%ifdef HIGH_BIT_DEPTH
-cglobal pixel_var_16x16_sse2, 2,3,8
+%if HIGH_BIT_DEPTH
+%macro VAR 0
+cglobal pixel_var_16x16, 2,3,8
FIX_STRIDES r1
VAR_START 0
VAR_2ROW r1, 8
VAR_END 16, 16
-cglobal pixel_var_8x8_sse2, 2,3,8
+cglobal pixel_var_8x8, 2,3,8
lea r2, [r1*3]
VAR_START 0
mova m0, [r0]
mova m4, [r0+r2*2]
VAR_CORE
VAR_END 8, 8
+%endmacro ; VAR
+
+INIT_XMM sse2
+VAR
+INIT_XMM avx
+VAR
+INIT_XMM xop
+VAR
%endif ; HIGH_BIT_DEPTH
-%ifndef HIGH_BIT_DEPTH
-cglobal pixel_var_16x16_sse2, 2,3,8
+%if HIGH_BIT_DEPTH == 0
+%macro VAR 0
+cglobal pixel_var_16x16, 2,3,8
VAR_START 1
mov r2d, 8
.loop:
jg .loop
VAR_END 16, 16
-cglobal pixel_var_8x8_sse2, 2,4,8
+cglobal pixel_var_8x8, 2,4,8
VAR_START 1
mov r2d, 2
lea r3, [r1*3]
dec r2d
jg .loop
VAR_END 8, 8
+
+cglobal pixel_var_8x16, 2,4,8
+ VAR_START 1
+ mov r2d, 4
+ lea r3, [r1*3]
+.loop:
+ movh m0, [r0]
+ movh m3, [r0+r1]
+ movhps m0, [r0+r1*2]
+ movhps m3, [r0+r3]
+ DEINTB 1, 0, 4, 3, 7
+ lea r0, [r0+r1*4]
+ VAR_CORE
+ dec r2d
+ jg .loop
+ VAR_END 8, 16
+%endmacro ; VAR
+
+INIT_XMM sse2
+VAR
+INIT_XMM avx
+VAR
+INIT_XMM xop
+VAR
%endif ; !HIGH_BIT_DEPTH
-%macro VAR2_END 0
+%macro VAR2_END 1
HADDW m5, m7
movd r1d, m5
imul r1d, r1d
HADDD m6, m1
- shr r1d, 6
+ shr r1d, %1
movd eax, m6
mov [r4], eax
sub eax, r1d ; sqr - (sum * sum >> shift)
%endmacro
;-----------------------------------------------------------------------------
-; int pixel_var2_8x8( pixel *, int, pixel *, int, int * )
+; int pixel_var2_8x8( pixel *, intptr_t, pixel *, intptr_t, int * )
;-----------------------------------------------------------------------------
-INIT_MMX
-cglobal pixel_var2_8x8_mmxext, 5,6
+%macro VAR2_8x8_MMX 2
+cglobal pixel_var2_8x%1, 5,6
FIX_STRIDES r1, r3
VAR_START 0
- mov r5d, 8
+ mov r5d, %1
.loop:
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
mova m0, [r0]
mova m1, [r0+mmsize]
psubw m0, [r2]
add r2, r3
dec r5d
jg .loop
- VAR2_END
- RET
+ VAR2_END %2
+%endmacro
-INIT_XMM
-cglobal pixel_var2_8x8_sse2, 5,6,8
+%if ARCH_X86_64 == 0
+INIT_MMX mmx2
+VAR2_8x8_MMX 8, 6
+VAR2_8x8_MMX 16, 7
+%endif
+
+%macro VAR2_8x8_SSE2 2
+cglobal pixel_var2_8x%1, 5,6,8
VAR_START 1
- mov r5d, 4
+ mov r5d, %1/2
.loop:
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
mova m0, [r0]
mova m1, [r0+r1*2]
mova m2, [r2]
lea r2, [r2+r3*2*SIZEOF_PIXEL]
dec r5d
jg .loop
- VAR2_END
- RET
+ VAR2_END %2
+%endmacro
+
+INIT_XMM sse2
+VAR2_8x8_SSE2 8, 6
+VAR2_8x8_SSE2 16, 7
-%ifndef HIGH_BIT_DEPTH
-cglobal pixel_var2_8x8_ssse3, 5,6,8
+%if HIGH_BIT_DEPTH == 0
+%macro VAR2_8x8_SSSE3 2
+cglobal pixel_var2_8x%1, 5,6,8
pxor m5, m5 ; sum
pxor m6, m6 ; sum squared
mova m7, [hsub_mul]
- mov r5d, 2
+ mov r5d, %1/4
.loop:
movq m0, [r0]
movq m2, [r2]
lea r2, [r2+r3*2]
dec r5d
jg .loop
- VAR2_END
- RET
+ VAR2_END %2
+%endmacro
+
+INIT_XMM ssse3
+VAR2_8x8_SSSE3 8, 6
+VAR2_8x8_SSSE3 16, 7
+INIT_XMM xop
+VAR2_8x8_SSSE3 8, 6
+VAR2_8x8_SSSE3 16, 7
+
%endif ; !HIGH_BIT_DEPTH
;=============================================================================
; SATD
;=============================================================================
-%define TRANS TRANS_SSE2
-
-%macro JDUP_SSE2 2
- punpckldq %1, %2
- ; doesn't need to dup. sse2 does things by zero extending to words and full h_2d
-%endmacro
-
-%macro JDUP_CONROE 2
+%macro JDUP 2
+%if cpuflag(sse4)
+ ; just use shufps on anything post conroe
+ shufps %1, %2, 0
+%elif cpuflag(ssse3)
; join 2x 32 bit and duplicate them
; emulating shufps is faster on conroe
punpcklqdq %1, %2
movsldup %1, %1
-%endmacro
-
-%macro JDUP_PENRYN 2
- ; just use shufps on anything post conroe
- shufps %1, %2, 0
+%else
+ ; doesn't need to dup. sse2 does things by zero extending to words and full h_2d
+ punpckldq %1, %2
+%endif
%endmacro
%macro HSUMSUB 5
DEINTB %1, %2, %3, %4, %5
psubw m%1, m%3
psubw m%2, m%4
- SUMSUB_BA w, m%1, m%2, m%3
+ SUMSUB_BA w, %1, %2, %3
%endmacro
%macro LOAD_SUMSUB_16x4P 10-13 r0, r2, none
%ifidn %1, sse2
HADAMARD4_2D_SSE %2, %3, %4, %5, %6, amax
%else
- HADAMARD4_V m%2, m%3, m%4, m%5, m%6
+ HADAMARD4_V %2, %3, %4, %5, %6
; doing the abs first is a slight advantage
- ABS4 m%2, m%4, m%3, m%5, m%6, m%7
+ ABSW2 m%2, m%4, m%2, m%4, m%6, m%7
+ ABSW2 m%3, m%5, m%3, m%5, m%6, m%7
HADAMARD 1, max, %2, %4, %6, %7
%endif
%ifnidn %9, swap
%endmacro
%macro SATD_END_MMX 0
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
HADDUW m0, m1
movd eax, m0
%else ; !HIGH_BIT_DEPTH
- pshufw m1, m0, 01001110b
+ pshufw m1, m0, q1032
paddw m0, m1
- pshufw m1, m0, 10110001b
+ pshufw m1, m0, q2301
paddw m0, m1
movd eax, m0
and eax, 0xffff
; for small blocks on x86_32, modify pixel pointer instead.
;-----------------------------------------------------------------------------
-; int pixel_satd_16x16( uint8_t *, int, uint8_t *, int )
+; int pixel_satd_16x16( uint8_t *, intptr_t, uint8_t *, intptr_t )
;-----------------------------------------------------------------------------
-INIT_MMX
-cglobal pixel_satd_16x4_internal_mmxext
+INIT_MMX mmx2
+cglobal pixel_satd_16x4_internal
SATD_4x4_MMX m2, 0, 0
SATD_4x4_MMX m1, 4, 0
paddw m0, m2
paddw m0, m1
ret
-cglobal pixel_satd_8x8_internal_mmxext
+cglobal pixel_satd_8x8_internal
SATD_4x4_MMX m2, 0, 0
SATD_4x4_MMX m1, 4, 1
paddw m0, m2
paddw m0, m1
-pixel_satd_8x4_internal_mmxext:
+pixel_satd_8x4_internal_mmx2:
SATD_4x4_MMX m2, 0, 0
SATD_4x4_MMX m1, 4, 0
paddw m0, m2
paddw m0, m1
ret
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
%macro SATD_MxN_MMX 3
-cglobal pixel_satd_%1x%2_mmxext, 4,7
+cglobal pixel_satd_%1x%2, 4,7
SATD_START_MMX
pxor m0, m0
- call pixel_satd_%1x%3_internal_mmxext
+ call pixel_satd_%1x%3_internal_mmx2
HADDUW m0, m1
movd r6d, m0
%rep %2/%3-1
pxor m0, m0
lea r0, [r0+4*r1]
lea r2, [r2+4*r3]
- call pixel_satd_%1x%3_internal_mmxext
+ call pixel_satd_%1x%3_internal_mmx2
movd m2, r4
HADDUW m0, m1
movd r4, m0
SATD_MxN_MMX 8, 16, 8
%endif ; HIGH_BIT_DEPTH
-%ifndef HIGH_BIT_DEPTH
-cglobal pixel_satd_16x16_mmxext, 4,6
+%if HIGH_BIT_DEPTH == 0
+cglobal pixel_satd_16x16, 4,6
SATD_START_MMX
pxor m0, m0
%rep 3
- call pixel_satd_16x4_internal_mmxext
+ call pixel_satd_16x4_internal_mmx2
lea r0, [r0+4*r1]
lea r2, [r2+4*r3]
%endrep
- call pixel_satd_16x4_internal_mmxext
+ call pixel_satd_16x4_internal_mmx2
HADDUW m0, m1
movd eax, m0
RET
-cglobal pixel_satd_16x8_mmxext, 4,6
+cglobal pixel_satd_16x8, 4,6
SATD_START_MMX
pxor m0, m0
- call pixel_satd_16x4_internal_mmxext
+ call pixel_satd_16x4_internal_mmx2
lea r0, [r0+4*r1]
lea r2, [r2+4*r3]
- call pixel_satd_16x4_internal_mmxext
+ call pixel_satd_16x4_internal_mmx2
SATD_END_MMX
-cglobal pixel_satd_8x16_mmxext, 4,6
+cglobal pixel_satd_8x16, 4,6
SATD_START_MMX
pxor m0, m0
- call pixel_satd_8x8_internal_mmxext
+ call pixel_satd_8x8_internal_mmx2
lea r0, [r0+4*r1]
lea r2, [r2+4*r3]
- call pixel_satd_8x8_internal_mmxext
+ call pixel_satd_8x8_internal_mmx2
SATD_END_MMX
%endif ; !HIGH_BIT_DEPTH
-cglobal pixel_satd_8x8_mmxext, 4,6
+cglobal pixel_satd_8x8, 4,6
SATD_START_MMX
pxor m0, m0
- call pixel_satd_8x8_internal_mmxext
+ call pixel_satd_8x8_internal_mmx2
SATD_END_MMX
-cglobal pixel_satd_8x4_mmxext, 4,6
+cglobal pixel_satd_8x4, 4,6
SATD_START_MMX
pxor m0, m0
- call pixel_satd_8x4_internal_mmxext
+ call pixel_satd_8x4_internal_mmx2
+ SATD_END_MMX
+
+cglobal pixel_satd_4x16, 4,6
+ SATD_START_MMX
+ SATD_4x4_MMX m0, 0, 1
+ SATD_4x4_MMX m1, 0, 1
+ paddw m0, m1
+ SATD_4x4_MMX m1, 0, 1
+ paddw m0, m1
+ SATD_4x4_MMX m1, 0, 0
+ paddw m0, m1
SATD_END_MMX
-cglobal pixel_satd_4x8_mmxext, 4,6
+cglobal pixel_satd_4x8, 4,6
SATD_START_MMX
SATD_4x4_MMX m0, 0, 1
SATD_4x4_MMX m1, 0, 0
paddw m0, m1
SATD_END_MMX
-cglobal pixel_satd_4x4_mmxext, 4,6
+cglobal pixel_satd_4x4, 4,6
SATD_START_MMX
SATD_4x4_MMX m0, 0, 0
SATD_END_MMX
-%macro SATD_START_SSE2 3
-%ifnidn %1, sse2
- mova %3, [hmul_8p]
+%macro SATD_START_SSE2 2
+%if cpuflag(ssse3)
+ mova %2, [hmul_8p]
%endif
lea r4, [3*r1]
lea r5, [3*r3]
- pxor %2, %2
+ pxor %1, %1
%endmacro
-%macro SATD_END_SSE2 2
- HADDW %2, m7
- movd eax, %2
+%macro SATD_END_SSE2 1
+ HADDW %1, m7
+ movd eax, %1
RET
%endmacro
%macro BACKUP_POINTERS 0
-%ifdef ARCH_X86_64
- mov r10, r0
- mov r11, r2
+%if ARCH_X86_64
+%if WIN64
+ PUSH r7
+%endif
+ mov r6, r0
+ mov r7, r2
%endif
%endmacro
%macro RESTORE_AND_INC_POINTERS 0
-%ifdef ARCH_X86_64
- lea r0, [r10+8]
- lea r2, [r11+8]
+%if ARCH_X86_64
+ lea r0, [r6+8]
+ lea r2, [r7+8]
+%if WIN64
+ POP r7
+%endif
%else
mov r0, r0mp
mov r2, r2mp
%endif
%endmacro
-;-----------------------------------------------------------------------------
-; int pixel_satd_8x4( uint8_t *, int, uint8_t *, int )
-;-----------------------------------------------------------------------------
-%macro SATDS_SSE2 1
-INIT_XMM
-%ifnidn %1, sse2
-cglobal pixel_satd_4x4_%1, 4, 6, 6
- SATD_START_MMX
- mova m4, [hmul_4p]
- LOAD_DUP_2x4P m2, m5, [r2], [r2+r3]
- LOAD_DUP_2x4P m3, m5, [r2+2*r3], [r2+r5]
- LOAD_DUP_2x4P m0, m5, [r0], [r0+r1]
- LOAD_DUP_2x4P m1, m5, [r0+2*r1], [r0+r4]
- DIFF_SUMSUB_SSSE3 0, 2, 1, 3, 4
- HADAMARD 0, sumsub, 0, 1, 2, 3
- HADAMARD 4, sumsub, 0, 1, 2, 3
- HADAMARD 1, amax, 0, 1, 2, 3
- HADDW m0, m1
- movd eax, m0
- RET
-%endif
-
-cglobal pixel_satd_4x8_%1, 4, 6, 8
- SATD_START_MMX
-%ifnidn %1, sse2
- mova m7, [hmul_4p]
-%endif
+%macro SATD_4x8_SSE 2
movd m4, [r2]
movd m5, [r2+r3]
movd m6, [r2+2*r3]
JDUP m5, m3
movd m3, [r0+2*r1]
JDUP m1, m3
+%if cpuflag(ssse3) && %1==1
+ mova m3, [hmul_4p]
+ DIFFOP 0, 4, 1, 5, 3
+%else
DIFFOP 0, 4, 1, 5, 7
+%endif
movd m5, [r2]
add r2, r5
movd m3, [r0]
JDUP m5, m4
movd m4, [r0+r1]
JDUP m3, m4
+%if cpuflag(ssse3) && %1==1
+ mova m4, [hmul_4p]
+ DIFFOP 2, 6, 3, 5, 4
+%else
DIFFOP 2, 6, 3, 5, 7
- SATD_8x4_SSE %1, 0, 1, 2, 3, 4, 5, 6, swap
- HADDW m6, m1
- movd eax, m6
+%endif
+ SATD_8x4_SSE cpuname, 0, 1, 2, 3, 4, 5, 7, %2
+%endmacro
+
+;-----------------------------------------------------------------------------
+; int pixel_satd_8x4( uint8_t *, intptr_t, uint8_t *, intptr_t )
+;-----------------------------------------------------------------------------
+%macro SATDS_SSE2 0
+%if cpuflag(ssse3)
+cglobal pixel_satd_4x4, 4, 6, 6
+ SATD_START_MMX
+ mova m4, [hmul_4p]
+ LOAD_DUP_2x4P m2, m5, [r2], [r2+r3]
+ LOAD_DUP_2x4P m3, m5, [r2+2*r3], [r2+r5]
+ LOAD_DUP_2x4P m0, m5, [r0], [r0+r1]
+ LOAD_DUP_2x4P m1, m5, [r0+2*r1], [r0+r4]
+ DIFF_SUMSUB_SSSE3 0, 2, 1, 3, 4
+ HADAMARD 0, sumsub, 0, 1, 2, 3
+ HADAMARD 4, sumsub, 0, 1, 2, 3
+ HADAMARD 1, amax, 0, 1, 2, 3
+ HADDW m0, m1
+ movd eax, m0
+ RET
+%endif
+
+cglobal pixel_satd_4x8, 4, 6, 8
+ SATD_START_MMX
+%if cpuflag(ssse3)
+ mova m7, [hmul_4p]
+%endif
+ SATD_4x8_SSE 0, swap
+ HADDW m7, m1
+ movd eax, m7
+ RET
+
+cglobal pixel_satd_4x16, 4, 6, 8
+ SATD_START_MMX
+%if cpuflag(ssse3)
+ mova m7, [hmul_4p]
+%endif
+ SATD_4x8_SSE 0, swap
+ lea r0, [r0+r1*2]
+ lea r2, [r2+r3*2]
+ SATD_4x8_SSE 1, add
+ HADDW m7, m1
+ movd eax, m7
RET
-cglobal pixel_satd_8x8_internal_%1
+cglobal pixel_satd_8x8_internal
LOAD_SUMSUB_8x4P 0, 1, 2, 3, 4, 5, 7, r0, r2, 1
- SATD_8x4_SSE %1, 0, 1, 2, 3, 4, 5, 6
-pixel_satd_8x4_internal_%1:
+ SATD_8x4_SSE cpuname, 0, 1, 2, 3, 4, 5, 6
+%%pixel_satd_8x4_internal:
LOAD_SUMSUB_8x4P 0, 1, 2, 3, 4, 5, 7, r0, r2, 1
- SATD_8x4_SSE %1, 0, 1, 2, 3, 4, 5, 6
+ SATD_8x4_SSE cpuname, 0, 1, 2, 3, 4, 5, 6
ret
-%ifdef UNIX64 ; 16x8 regresses on phenom win64, 16x16 is almost the same
-cglobal pixel_satd_16x4_internal_%1
+%if UNIX64 ; 16x8 regresses on phenom win64, 16x16 is almost the same
+cglobal pixel_satd_16x4_internal
LOAD_SUMSUB_16x4P 0, 1, 2, 3, 4, 8, 5, 9, 6, 7, r0, r2, 11
lea r2, [r2+4*r3]
lea r0, [r0+4*r1]
+ ; FIXME: this doesn't really mean ssse3, but rather selects between two different behaviors implemented with sse2?
SATD_8x4_SSE ssse3, 0, 1, 2, 3, 6, 11, 10
SATD_8x4_SSE ssse3, 4, 8, 5, 9, 6, 3, 10
ret
-cglobal pixel_satd_16x8_%1, 4,6,12
- SATD_START_SSE2 %1, m10, m7
-%ifidn %1, sse2
+cglobal pixel_satd_16x8, 4,6,12
+ SATD_START_SSE2 m10, m7
+%if notcpuflag(ssse3)
mova m7, [pw_00ff]
%endif
- jmp pixel_satd_16x8_internal_%1
+ jmp %%pixel_satd_16x8_internal
-cglobal pixel_satd_16x16_%1, 4,6,12
- SATD_START_SSE2 %1, m10, m7
-%ifidn %1, sse2
+cglobal pixel_satd_16x16, 4,6,12
+ SATD_START_SSE2 m10, m7
+%if notcpuflag(ssse3)
mova m7, [pw_00ff]
%endif
- call pixel_satd_16x4_internal_%1
- call pixel_satd_16x4_internal_%1
-pixel_satd_16x8_internal_%1:
- call pixel_satd_16x4_internal_%1
- call pixel_satd_16x4_internal_%1
- SATD_END_SSE2 %1, m10
+ call pixel_satd_16x4_internal
+ call pixel_satd_16x4_internal
+%%pixel_satd_16x8_internal:
+ call pixel_satd_16x4_internal
+ call pixel_satd_16x4_internal
+ SATD_END_SSE2 m10
%else
-cglobal pixel_satd_16x8_%1, 4,6,8
- SATD_START_SSE2 %1, m6, m7
+cglobal pixel_satd_16x8, 4,6,8
+ SATD_START_SSE2 m6, m7
BACKUP_POINTERS
- call pixel_satd_8x8_internal_%1
+ call pixel_satd_8x8_internal
RESTORE_AND_INC_POINTERS
- call pixel_satd_8x8_internal_%1
- SATD_END_SSE2 %1, m6
+ call pixel_satd_8x8_internal
+ SATD_END_SSE2 m6
-cglobal pixel_satd_16x16_%1, 4,6,8
- SATD_START_SSE2 %1, m6, m7
+cglobal pixel_satd_16x16, 4,6,8
+ SATD_START_SSE2 m6, m7
BACKUP_POINTERS
- call pixel_satd_8x8_internal_%1
- call pixel_satd_8x8_internal_%1
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
RESTORE_AND_INC_POINTERS
- call pixel_satd_8x8_internal_%1
- call pixel_satd_8x8_internal_%1
- SATD_END_SSE2 %1, m6
-%endif
-
-cglobal pixel_satd_8x16_%1, 4,6,8
- SATD_START_SSE2 %1, m6, m7
- call pixel_satd_8x8_internal_%1
- call pixel_satd_8x8_internal_%1
- SATD_END_SSE2 %1, m6
-
-cglobal pixel_satd_8x8_%1, 4,6,8
- SATD_START_SSE2 %1, m6, m7
- call pixel_satd_8x8_internal_%1
- SATD_END_SSE2 %1, m6
-
-cglobal pixel_satd_8x4_%1, 4,6,8
- SATD_START_SSE2 %1, m6, m7
- call pixel_satd_8x4_internal_%1
- SATD_END_SSE2 %1, m6
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_END_SSE2 m6
+%endif
+
+cglobal pixel_satd_8x16, 4,6,8
+ SATD_START_SSE2 m6, m7
+ call pixel_satd_8x8_internal
+ call pixel_satd_8x8_internal
+ SATD_END_SSE2 m6
+
+cglobal pixel_satd_8x8, 4,6,8
+ SATD_START_SSE2 m6, m7
+ call pixel_satd_8x8_internal
+ SATD_END_SSE2 m6
+
+cglobal pixel_satd_8x4, 4,6,8
+ SATD_START_SSE2 m6, m7
+ call %%pixel_satd_8x4_internal
+ SATD_END_SSE2 m6
%endmacro ; SATDS_SSE2
%macro SA8D_INTER 0
-%ifdef ARCH_X86_64
+%if ARCH_X86_64
%define lh m10
%define rh m0
%else
%define lh m0
%define rh [esp+48]
%endif
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
HADDUW m0, m1
paddd lh, rh
%else
%endif ; HIGH_BIT_DEPTH
%endmacro
-%macro SA8D 1
-%ifdef HIGH_BIT_DEPTH
- %define vertical 1
-%elifidn %1, sse2 ; sse2 doesn't seem to like the horizontal way of doing things
+%macro SA8D 0
+%if HIGH_BIT_DEPTH
%define vertical 1
-%else
- %define vertical 0
+%else ; sse2 doesn't seem to like the horizontal way of doing things
+ %define vertical (cpuflags == cpuflags_sse2)
%endif
-%ifdef ARCH_X86_64
+%if ARCH_X86_64
;-----------------------------------------------------------------------------
-; int pixel_sa8d_8x8( uint8_t *, int, uint8_t *, int )
+; int pixel_sa8d_8x8( uint8_t *, intptr_t, uint8_t *, intptr_t )
;-----------------------------------------------------------------------------
-cglobal pixel_sa8d_8x8_internal_%1
- lea r10, [r0+4*r1]
- lea r11, [r2+4*r3]
+cglobal pixel_sa8d_8x8_internal
+ lea r6, [r0+4*r1]
+ lea r7, [r2+4*r3]
LOAD_SUMSUB_8x4P 0, 1, 2, 8, 5, 6, 7, r0, r2
- LOAD_SUMSUB_8x4P 4, 5, 3, 9, 11, 6, 7, r10, r11
+ LOAD_SUMSUB_8x4P 4, 5, 3, 9, 11, 6, 7, r6, r7
%if vertical
HADAMARD8_2D 0, 1, 2, 8, 4, 5, 3, 9, 6, amax
%else ; non-sse2
- HADAMARD4_V m0, m1, m2, m8, m6
- HADAMARD4_V m4, m5, m3, m9, m6
- SUMSUB_BADC w, m0, m4, m1, m5, m6
- HADAMARD 2, sumsub, 0, 4, 6, 11
- HADAMARD 2, sumsub, 1, 5, 6, 11
- SUMSUB_BADC w, m2, m3, m8, m9, m6
- HADAMARD 2, sumsub, 2, 3, 6, 11
- HADAMARD 2, sumsub, 8, 9, 6, 11
- HADAMARD 1, amax, 0, 4, 6, 11
- HADAMARD 1, amax, 1, 5, 6, 4
- HADAMARD 1, amax, 2, 3, 6, 4
- HADAMARD 1, amax, 8, 9, 6, 4
+ HADAMARD8_2D_HMUL 0, 1, 2, 8, 4, 5, 3, 9, 6, 11
%endif
paddw m0, m1
paddw m0, m2
paddw m0, m8
- SAVE_MM_PERMUTATION pixel_sa8d_8x8_internal_%1
+ SAVE_MM_PERMUTATION
ret
-cglobal pixel_sa8d_8x8_%1, 4,6,12
+cglobal pixel_sa8d_8x8, 4,8,12
FIX_STRIDES r1, r3
lea r4, [3*r1]
lea r5, [3*r3]
%if vertical == 0
mova m7, [hmul_8p]
%endif
- call pixel_sa8d_8x8_internal_%1
-%ifdef HIGH_BIT_DEPTH
+ call pixel_sa8d_8x8_internal
+%if HIGH_BIT_DEPTH
HADDUW m0, m1
%else
HADDW m0, m1
shr eax, 1
RET
-cglobal pixel_sa8d_16x16_%1, 4,6,12
+cglobal pixel_sa8d_16x16, 4,8,12
FIX_STRIDES r1, r3
lea r4, [3*r1]
lea r5, [3*r3]
%if vertical == 0
mova m7, [hmul_8p]
%endif
- call pixel_sa8d_8x8_internal_%1 ; pix[0]
+ call pixel_sa8d_8x8_internal ; pix[0]
add r2, 8*SIZEOF_PIXEL
add r0, 8*SIZEOF_PIXEL
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
HADDUW m0, m1
%endif
mova m10, m0
- call pixel_sa8d_8x8_internal_%1 ; pix[8]
+ call pixel_sa8d_8x8_internal ; pix[8]
lea r2, [r2+8*r3]
lea r0, [r0+8*r1]
SA8D_INTER
- call pixel_sa8d_8x8_internal_%1 ; pix[8*stride+8]
+ call pixel_sa8d_8x8_internal ; pix[8*stride+8]
sub r2, 8*SIZEOF_PIXEL
sub r0, 8*SIZEOF_PIXEL
SA8D_INTER
- call pixel_sa8d_8x8_internal_%1 ; pix[8*stride]
+ call pixel_sa8d_8x8_internal ; pix[8*stride]
SA8D_INTER
- SWAP m0, m10
-%ifndef HIGH_BIT_DEPTH
+ SWAP 0, 10
+%if HIGH_BIT_DEPTH == 0
HADDUW m0, m1
%endif
movd eax, m0
RET
%else ; ARCH_X86_32
-%ifnidn %1, mmxext
-cglobal pixel_sa8d_8x8_internal_%1
+%if mmsize == 16
+cglobal pixel_sa8d_8x8_internal
%define spill0 [esp+4]
%define spill1 [esp+20]
%define spill2 [esp+36]
movdqa m3, spill0
paddw m0, m1
HADAMARD2_2D 2, 6, 3, 7, 5, qdq, amax
-%else ; non-sse2
+%else ; mmsize == 8
mova m7, [hmul_8p]
LOAD_SUMSUB_8x4P 0, 1, 2, 3, 5, 6, 7, r0, r2, 1
; could do first HADAMARD4_V here to save spilling later
mova spill2, m1
SWAP 1, 7
LOAD_SUMSUB_8x4P 4, 5, 6, 7, 2, 3, 1, r0, r2, 1
- HADAMARD4_V m4, m5, m6, m7, m3
+ HADAMARD4_V 4, 5, 6, 7, 3
mova m1, spill2
mova m2, spill0
mova m3, spill1
mova spill0, m6
mova spill1, m7
- HADAMARD4_V m0, m1, m2, m3, m7
- SUMSUB_BADC w, m0, m4, m1, m5, m7
+ HADAMARD4_V 0, 1, 2, 3, 7
+ SUMSUB_BADC w, 0, 4, 1, 5, 7
HADAMARD 2, sumsub, 0, 4, 7, 6
HADAMARD 2, sumsub, 1, 5, 7, 6
HADAMARD 1, amax, 0, 4, 7, 6
mova m6, spill0
mova m7, spill1
paddw m0, m1
- SUMSUB_BADC w, m2, m6, m3, m7, m4
+ SUMSUB_BADC w, 2, 6, 3, 7, 4
HADAMARD 2, sumsub, 2, 6, 4, 5
HADAMARD 2, sumsub, 3, 7, 4, 5
HADAMARD 1, amax, 2, 6, 4, 5
%endif ; sse2/non-sse2
paddw m0, m2
paddw m0, m3
+ SAVE_MM_PERMUTATION
ret
-%endif ; ifndef mmxext
+%endif ; ifndef mmx2
-cglobal pixel_sa8d_8x8_%1, 4,7
+cglobal pixel_sa8d_8x8, 4,7
FIX_STRIDES r1, r3
mov r6, esp
and esp, ~15
sub esp, 48
lea r4, [3*r1]
lea r5, [3*r3]
- call pixel_sa8d_8x8_internal_%1
-%ifdef HIGH_BIT_DEPTH
+ call pixel_sa8d_8x8_internal
+%if HIGH_BIT_DEPTH
HADDUW m0, m1
%else
HADDW m0, m1
mov esp, r6
RET
-cglobal pixel_sa8d_16x16_%1, 4,7
+cglobal pixel_sa8d_16x16, 4,7
FIX_STRIDES r1, r3
mov r6, esp
and esp, ~15
sub esp, 64
lea r4, [3*r1]
lea r5, [3*r3]
- call pixel_sa8d_8x8_internal_%1
-%ifidn %1, mmxext
+ call pixel_sa8d_8x8_internal
+%if mmsize == 8
lea r0, [r0+4*r1]
lea r2, [r2+4*r3]
%endif
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
HADDUW m0, m1
%endif
mova [esp+48], m0
- call pixel_sa8d_8x8_internal_%1
+ call pixel_sa8d_8x8_internal
mov r0, [r6+20]
mov r2, [r6+28]
add r0, 8*SIZEOF_PIXEL
add r2, 8*SIZEOF_PIXEL
SA8D_INTER
mova [esp+48], m0
- call pixel_sa8d_8x8_internal_%1
-%ifidn %1, mmxext
+ call pixel_sa8d_8x8_internal
+%if mmsize == 8
lea r0, [r0+4*r1]
lea r2, [r2+4*r3]
-%endif
-%if mmsize == 16
+%else
SA8D_INTER
%endif
mova [esp+64-mmsize], m0
- call pixel_sa8d_8x8_internal_%1
-%ifdef HIGH_BIT_DEPTH
+ call pixel_sa8d_8x8_internal
+%if HIGH_BIT_DEPTH
SA8D_INTER
%else ; !HIGH_BIT_DEPTH
paddusw m0, [esp+64-mmsize]
; INTRA SATD
;=============================================================================
-%macro INTRA_SA8D_SSE2 1
-%ifdef ARCH_X86_64
-INIT_XMM
+%macro HSUMSUB2 8
+ pshufd %4, %2, %7
+ pshufd %5, %3, %7
+ %1 %2, %8
+ %1 %6, %8
+ paddw %2, %4
+ paddw %3, %5
+%endmacro
+
+; intra_sa8d_x3_8x8 and intra_satd_x3_4x4 are obsoleted by x9 on ssse3+,
+; and are only retained for old cpus.
+%macro INTRA_SA8D_SSE2 0
+%if ARCH_X86_64
;-----------------------------------------------------------------------------
-; void intra_sa8d_x3_8x8_core( uint8_t *fenc, int16_t edges[2][8], int *res )
+; void intra_sa8d_x3_8x8( uint8_t *fenc, uint8_t edge[36], int *res )
;-----------------------------------------------------------------------------
-cglobal intra_sa8d_x3_8x8_core_%1, 3,3,16
+cglobal intra_sa8d_x3_8x8, 3,3,14
; 8x8 hadamard
pxor m8, m8
movq m0, [r0+0*FENC_STRIDE]
punpcklbw m6, m8
punpcklbw m7, m8
- HADAMARD8_2D 0, 1, 2, 3, 4, 5, 6, 7, 8
+ HADAMARD8_2D 0, 1, 2, 3, 4, 5, 6, 7, 8
- ; dc
- movzx r0d, word [r1+0]
- add r0w, word [r1+16]
- add r0d, 8
- and r0d, -16
- shl r0d, 2
-
- pxor m15, m15
- movdqa m8, m2
- movdqa m9, m3
- movdqa m10, m4
- movdqa m11, m5
- ABS4 m8, m9, m10, m11, m12, m13
+ ABSW2 m8, m9, m2, m3, m2, m3
+ ABSW2 m10, m11, m4, m5, m4, m5
paddusw m8, m10
paddusw m9, m11
-%ifidn %1, ssse3
- pabsw m10, m6
- pabsw m11, m7
- pabsw m15, m1
-%else
- movdqa m10, m6
- movdqa m11, m7
- movdqa m15, m1
- ABS2 m10, m11, m13, m14
- ABS1 m15, m13
-%endif
+ ABSW2 m10, m11, m6, m7, m6, m7
+ ABSW m13, m1, m1
paddusw m10, m11
paddusw m8, m9
- paddusw m15, m10
- paddusw m15, m8
- movdqa m14, m15 ; 7x8 sum
-
- movdqa m8, [r1+0] ; left edge
- movd m9, r0d
- psllw m8, 3
+ paddusw m13, m10
+ paddusw m13, m8
+
+ ; 1D hadamard of edges
+ movq m8, [r1+7]
+ movq m9, [r1+16]
+ pxor m10, m10
+ punpcklbw m8, m10
+ punpcklbw m9, m10
+ HSUMSUB2 pmullw, m8, m9, m10, m11, m11, q1032, [pw_ppppmmmm]
+ HSUMSUB2 pmullw, m8, m9, m10, m11, m11, q2301, [pw_ppmmppmm]
+ pshuflw m10, m8, q2301
+ pshuflw m11, m9, q2301
+ pshufhw m10, m10, q2301
+ pshufhw m11, m11, q2301
+ pmullw m8, [pw_pmpmpmpm]
+ pmullw m11, [pw_pmpmpmpm]
+ paddw m8, m10
+ paddw m9, m11
+
+ ; differences
+ paddw m10, m8, m9
+ paddw m10, [pw_8]
+ pand m10, [sw_f0]
+ psllw m10, 2 ; dc
+
+ psllw m8, 3 ; left edge
psubw m8, m0
- psubw m9, m0
- ABS1 m8, m10
- ABS1 m9, m11 ; 1x8 sum
- paddusw m14, m8
- paddusw m15, m9
+ psubw m10, m0
+ ABSW2 m8, m10, m8, m10, m11, m12 ; 1x8 sum
+ paddusw m8, m13
+ paddusw m13, m10
punpcklwd m0, m1
punpcklwd m2, m3
punpcklwd m4, m5
punpckldq m0, m2
punpckldq m4, m6
punpcklqdq m0, m4 ; transpose
- movdqa m1, [r1+16] ; top edge
- movdqa m2, m15
- psllw m1, 3
- psrldq m2, 2 ; 8x7 sum
- psubw m0, m1 ; 8x1 sum
- ABS1 m0, m1
+ psllw m9, 3 ; top edge
+ psrldq m2, m13, 2 ; 8x7 sum
+ psubw m0, m9 ; 8x1 sum
+ ABSW m0, m0, m9
paddusw m2, m0
; 3x HADDW
movdqa m7, [pw_1]
pmaddwd m2, m7
- pmaddwd m14, m7
- pmaddwd m15, m7
- movdqa m3, m2
- punpckldq m2, m14
- punpckhdq m3, m14
- pshufd m5, m15, 0xf5
+ pmaddwd m8, m7
+ pmaddwd m13, m7
+ punpckhdq m3, m2, m8
+ punpckldq m2, m8
+ pshufd m5, m13, q3311
paddd m2, m3
- paddd m5, m15
- movdqa m3, m2
+ paddd m5, m13
+ punpckhqdq m0, m2, m5
punpcklqdq m2, m5
- punpckhqdq m3, m5
- pavgw m3, m2
- pxor m0, m0
- pavgw m3, m0
- movq [r2], m3 ; i8x8_v, i8x8_h
- psrldq m3, 8
- movd [r2+8], m3 ; i8x8_dc
+ pavgw m0, m2
+ pxor m1, m1
+ pavgw m0, m1
+ movq [r2], m0 ; i8x8_v, i8x8_h
+ psrldq m0, 8
+ movd [r2+8], m0 ; i8x8_dc
RET
%endif ; ARCH_X86_64
%endmacro ; INTRA_SA8D_SSE2
INIT_MMX
cglobal hadamard_load
; not really a global, but otherwise cycles get attributed to the wrong function in profiling
+%if HIGH_BIT_DEPTH
+ mova m0, [r0+0*FENC_STRIDEB]
+ mova m1, [r0+1*FENC_STRIDEB]
+ mova m2, [r0+2*FENC_STRIDEB]
+ mova m3, [r0+3*FENC_STRIDEB]
+%else
pxor m7, m7
movd m0, [r0+0*FENC_STRIDE]
movd m1, [r0+1*FENC_STRIDE]
punpcklbw m1, m7
punpcklbw m2, m7
punpcklbw m3, m7
+%endif
HADAMARD4_2D 0, 1, 2, 3, 4
- SAVE_MM_PERMUTATION hadamard_load
+ SAVE_MM_PERMUTATION
ret
-%macro SCALAR_SUMSUB 4
- add %1, %2
- add %3, %4
- add %2, %2
- add %4, %4
- sub %2, %1
- sub %4, %3
-%endmacro
-
-%macro SCALAR_HADAMARD_LEFT 5 ; y, 4x tmp
-%ifnidn %1, 0
- shl %1d, 5 ; log(FDEC_STRIDE)
-%endif
- movzx %2d, byte [r1+%1-1+0*FDEC_STRIDE]
- movzx %3d, byte [r1+%1-1+1*FDEC_STRIDE]
- movzx %4d, byte [r1+%1-1+2*FDEC_STRIDE]
- movzx %5d, byte [r1+%1-1+3*FDEC_STRIDE]
-%ifnidn %1, 0
- shr %1d, 5
-%endif
- SCALAR_SUMSUB %2d, %3d, %4d, %5d
- SCALAR_SUMSUB %2d, %4d, %3d, %5d
- mov [left_1d+2*%1+0], %2w
- mov [left_1d+2*%1+2], %3w
- mov [left_1d+2*%1+4], %4w
- mov [left_1d+2*%1+6], %5w
-%endmacro
-
-%macro SCALAR_HADAMARD_TOP 5 ; x, 4x tmp
- movzx %2d, byte [r1+%1-FDEC_STRIDE+0]
- movzx %3d, byte [r1+%1-FDEC_STRIDE+1]
- movzx %4d, byte [r1+%1-FDEC_STRIDE+2]
- movzx %5d, byte [r1+%1-FDEC_STRIDE+3]
- SCALAR_SUMSUB %2d, %3d, %4d, %5d
- SCALAR_SUMSUB %2d, %4d, %3d, %5d
- mov [top_1d+2*%1+0], %2w
- mov [top_1d+2*%1+2], %3w
- mov [top_1d+2*%1+4], %4w
- mov [top_1d+2*%1+6], %5w
+%macro SCALAR_HADAMARD 4-5 ; direction, offset, 3x tmp
+%ifidn %1, top
+%if HIGH_BIT_DEPTH
+ mova %3, [r1+%2*SIZEOF_PIXEL-FDEC_STRIDEB]
+%else
+ movd %3, [r1+%2*SIZEOF_PIXEL-FDEC_STRIDEB]
+ pxor %5, %5
+ punpcklbw %3, %5
+%endif
+%else ; left
+%ifnidn %2, 0
+ shl %2d, 5 ; log(FDEC_STRIDEB)
+%endif
+ movd %3, [r1+%2*SIZEOF_PIXEL-4+1*FDEC_STRIDEB]
+ pinsrw %3, [r1+%2*SIZEOF_PIXEL-2+0*FDEC_STRIDEB], 0
+ pinsrw %3, [r1+%2*SIZEOF_PIXEL-2+2*FDEC_STRIDEB], 2
+ pinsrw %3, [r1+%2*SIZEOF_PIXEL-2+3*FDEC_STRIDEB], 3
+%if HIGH_BIT_DEPTH == 0
+ psrlw %3, 8
+%endif
+%ifnidn %2, 0
+ shr %2d, 5
+%endif
+%endif ; direction
+%if cpuflag(ssse3)
+ %define %%sign psignw
+%else
+ %define %%sign pmullw
+%endif
+ pshufw %4, %3, q1032
+ %%sign %4, [pw_ppmmppmm]
+ paddw %3, %4
+ pshufw %4, %3, q2301
+ %%sign %4, [pw_pmpmpmpm]
+ paddw %3, %4
+ psllw %3, 2
+ mova [%1_1d+2*%2], %3
%endmacro
%macro SUM_MM_X3 8 ; 3x sum, 4x tmp, op
pxor %7, %7
- pshufw %4, %1, 01001110b
- pshufw %5, %2, 01001110b
- pshufw %6, %3, 01001110b
+ pshufw %4, %1, q1032
+ pshufw %5, %2, q1032
+ pshufw %6, %3, q1032
paddw %1, %4
paddw %2, %5
paddw %3, %6
punpcklwd %1, %7
punpcklwd %2, %7
punpcklwd %3, %7
- pshufw %4, %1, 01001110b
- pshufw %5, %2, 01001110b
- pshufw %6, %3, 01001110b
+ pshufw %4, %1, q1032
+ pshufw %5, %2, q1032
+ pshufw %6, %3, q1032
%8 %1, %4
%8 %2, %5
%8 %3, %6
%endmacro
-%macro CLEAR_SUMS 0
-%ifdef ARCH_X86_64
- mov qword [sums+0], 0
- mov qword [sums+8], 0
- mov qword [sums+16], 0
-%else
- pxor m7, m7
- movq [sums+0], m7
- movq [sums+8], m7
- movq [sums+16], m7
-%endif
-%endmacro
-
; in: m1..m3
; out: m7
; clobber: m4..m6
-%macro SUM3x4 1
-%ifidn %1, ssse3
- pabsw m4, m1
- pabsw m5, m2
- pabsw m7, m3
+%macro SUM3x4 0
+ ABSW2 m4, m5, m1, m2, m1, m2
+ ABSW m7, m3, m3
paddw m4, m5
-%else
- movq m4, m1
- movq m5, m2
- ABS2 m4, m5, m6, m7
- movq m7, m3
- paddw m4, m5
- ABS1 m7, m6
-%endif
paddw m7, m4
%endmacro
-; in: m0..m3 (4x4), m7 (3x4)
+; in: m0..m3 (4x4)
; out: m0 v, m4 h, m5 dc
-; clobber: m6
+; clobber: m1..m3
%macro SUM4x3 3 ; dc, left, top
movq m4, %2
+%ifid %1
+ movq m5, %1
+%else
movd m5, %1
- psllw m4, 2
+%endif
psubw m4, m0
psubw m5, m0
punpcklwd m0, m1
punpcklwd m2, m3
punpckldq m0, m2 ; transpose
- movq m1, %3
- psllw m1, 2
- psubw m0, m1
- ABS2 m4, m5, m2, m3 ; 1x4 sum
- ABS1 m0, m1 ; 4x1 sum
+ psubw m0, %3
+ ABSW2 m4, m5, m4, m5, m2, m3 ; 1x4 sum
+ ABSW m0, m0, m1 ; 4x1 sum
%endmacro
-%macro INTRA_SATDS_MMX 1
-INIT_MMX
+%macro INTRA_X3_MMX 0
;-----------------------------------------------------------------------------
; void intra_satd_x3_4x4( uint8_t *fenc, uint8_t *fdec, int *res )
;-----------------------------------------------------------------------------
-cglobal intra_satd_x3_4x4_%1, 2,6
-%ifdef ARCH_X86_64
+cglobal intra_satd_x3_4x4, 3,3
+%if UNIX64
; stack is 16 byte aligned because abi says so
%define top_1d rsp-8 ; size 8
%define left_1d rsp-16 ; size 8
- %define t0 r10
%else
- ; stack is 16 byte aligned at least in gcc, and we've pushed 3 regs + return address, so it's still aligned
- SUB esp, 16
- %define top_1d esp+8
- %define left_1d esp
- %define t0 r2
+ ; WIN64: stack is 16 byte aligned because abi says so
+ ; X86_32: stack is 16 byte aligned at least in gcc, and we've pushed 3 regs + return address, so it's still aligned
+ SUB rsp, 16
+ %define top_1d rsp+8
+ %define left_1d rsp
%endif
call hadamard_load
- SCALAR_HADAMARD_LEFT 0, r0, r3, r4, r5
- mov t0d, r0d
- SCALAR_HADAMARD_TOP 0, r0, r3, r4, r5
- lea t0d, [t0d + r0d + 4]
- and t0d, -8
- shl t0d, 1 ; dc
-
- SUM3x4 %1
- SUM4x3 t0d, [left_1d], [top_1d]
+ SCALAR_HADAMARD left, 0, m4, m5
+ SCALAR_HADAMARD top, 0, m6, m5, m7
+ paddw m6, m4
+ pavgw m6, [pw_16]
+ pand m6, [sw_f0] ; dc
+
+ SUM3x4
+ SUM4x3 m6, [left_1d], [top_1d]
paddw m4, m7
paddw m5, m7
movq m1, m5
paddw m0, m1
SUM_MM_X3 m0, m4, m5, m1, m2, m3, m6, pavgw
-%ifndef ARCH_X86_64
- mov r2, r2mp
-%endif
movd [r2+0], m0 ; i4x4_v satd
movd [r2+4], m4 ; i4x4_h satd
movd [r2+8], m5 ; i4x4_dc satd
-%ifndef ARCH_X86_64
- ADD esp, 16
+%if UNIX64 == 0
+ ADD rsp, 16
%endif
RET
-%ifdef ARCH_X86_64
- %define t0 r10
- %define t2 r11
-%else
- %define t0 r0
- %define t2 r2
-%endif
-
;-----------------------------------------------------------------------------
; void intra_satd_x3_16x16( uint8_t *fenc, uint8_t *fdec, int *res )
;-----------------------------------------------------------------------------
-cglobal intra_satd_x3_16x16_%1, 0,7
-%ifdef ARCH_X86_64
- %assign stack_pad 88
-%else
- %assign stack_pad 88 + ((stack_offset+88+4)&15)
-%endif
+cglobal intra_satd_x3_16x16, 0,5
+ %assign stack_pad 120 + ((stack_offset+120+gprsize)&15)
; not really needed on x86_64, just shuts up valgrind about storing data below the stack across a function call
SUB rsp, stack_pad
-%define sums rsp+64 ; size 24
+%define sums rsp+64 ; size 56
%define top_1d rsp+32 ; size 32
%define left_1d rsp ; size 32
movifnidn r1, r1mp
- CLEAR_SUMS
+
+ pxor m7, m7
+ mova [sums+ 0], m7
+ mova [sums+ 8], m7
+ mova [sums+16], m7
+%if HIGH_BIT_DEPTH
+ mova [sums+24], m7
+ mova [sums+32], m7
+ mova [sums+40], m7
+ mova [sums+48], m7
+%endif
; 1D hadamards
- xor t2d, t2d
- mov t0d, 12
+ mov r3d, 12
+ movd m6, [pw_32]
.loop_edge:
- SCALAR_HADAMARD_LEFT t0, r3, r4, r5, r6
- add t2d, r3d
- SCALAR_HADAMARD_TOP t0, r3, r4, r5, r6
- add t2d, r3d
- sub t0d, 4
+ SCALAR_HADAMARD left, r3, m0, m1
+ SCALAR_HADAMARD top, r3, m1, m2, m3
+ pavgw m0, m1
+ paddw m6, m0
+ sub r3d, 4
jge .loop_edge
- shr t2d, 1
- add t2d, 8
- and t2d, -16 ; dc
+ psrlw m6, 2
+ pand m6, [sw_f0] ; dc
; 2D hadamards
- movifnidn r0, r0mp
- xor r3d, r3d
+ movifnidn r0, r0mp
+ mov r3, -4
.loop_y:
- xor r4d, r4d
+ mov r4, -4
.loop_x:
call hadamard_load
- SUM3x4 %1
- SUM4x3 t2d, [left_1d+8*r3], [top_1d+8*r4]
+ SUM3x4
+ SUM4x3 m6, [left_1d+8*(r3+4)], [top_1d+8*(r4+4)]
pavgw m4, m7
pavgw m5, m7
- paddw m0, [sums+0] ; i16x16_v satd
- paddw m4, [sums+8] ; i16x16_h satd
+ paddw m0, [sums+ 0] ; i16x16_v satd
+ paddw m4, [sums+ 8] ; i16x16_h satd
paddw m5, [sums+16] ; i16x16_dc satd
- movq [sums+0], m0
- movq [sums+8], m4
- movq [sums+16], m5
+ mova [sums+ 0], m0
+ mova [sums+ 8], m4
+ mova [sums+16], m5
- add r0, 4
- inc r4d
- cmp r4d, 4
+ add r0, 4*SIZEOF_PIXEL
+ inc r4
jl .loop_x
- add r0, 4*FENC_STRIDE-16
- inc r3d
- cmp r3d, 4
+%if HIGH_BIT_DEPTH
+ psrld m7, m4, 16
+ pslld m4, 16
+ psrld m4, 16
+ paddd m4, m7
+ psrld m7, m0, 16
+ pslld m0, 16
+ psrld m0, 16
+ paddd m0, m7
+ paddd m4, [sums+32]
+ paddd m0, [sums+24]
+ mova [sums+32], m4
+ mova [sums+24], m0
+ pxor m7, m7
+ punpckhwd m3, m5, m7
+ punpcklwd m5, m7
+ paddd m3, [sums+48]
+ paddd m5, [sums+40]
+ mova [sums+48], m3
+ mova [sums+40], m5
+ mova [sums+ 0], m7
+ mova [sums+ 8], m7
+ mova [sums+16], m7
+%endif
+ add r0, 4*FENC_STRIDEB-16*SIZEOF_PIXEL
+ inc r3
jl .loop_y
; horizontal sum
movifnidn r2, r2mp
- movq m2, [sums+16]
- movq m1, [sums+8]
- movq m0, [sums+0]
- movq m7, m2
- SUM_MM_X3 m0, m1, m2, m3, m4, m5, m6, paddd
+%if HIGH_BIT_DEPTH
+ mova m1, m5
+ paddd m5, m3
+ HADDD m5, m7 ; DC satd
+ HADDD m4, m7 ; H satd
+ HADDD m0, m7 ; the part of V satd that doesn't overlap with DC
+ psrld m0, 1
+ psrlq m1, 32 ; DC[1]
+ paddd m0, m3 ; DC[2]
+ psrlq m3, 32 ; DC[3]
+ paddd m0, m1
+ paddd m0, m3
+%else
+ mova m7, m5
+ SUM_MM_X3 m0, m4, m5, m3, m1, m2, m6, paddd
psrld m0, 1
pslld m7, 16
psrld m7, 16
- paddd m0, m2
+ paddd m0, m5
psubd m0, m7
- movd [r2+8], m2 ; i16x16_dc satd
- movd [r2+4], m1 ; i16x16_h satd
- movd [r2+0], m0 ; i16x16_v satd
- ADD rsp, stack_pad
+%endif
+ movd [r2+8], m5 ; i16x16_dc satd
+ movd [r2+4], m4 ; i16x16_h satd
+ movd [r2+0], m0 ; i16x16_v satd
+ ADD rsp, stack_pad
RET
+%if ARCH_X86_64
+ %define t0 r6
+%else
+ %define t0 r2
+%endif
+
;-----------------------------------------------------------------------------
; void intra_satd_x3_8x8c( uint8_t *fenc, uint8_t *fdec, int *res )
;-----------------------------------------------------------------------------
-cglobal intra_satd_x3_8x8c_%1, 0,6
+cglobal intra_satd_x3_8x8c, 0,6
; not really needed on x86_64, just shuts up valgrind about storing data below the stack across a function call
SUB rsp, 72
%define sums rsp+48 ; size 24
%define top_1d rsp+16 ; size 16
%define left_1d rsp ; size 16
movifnidn r1, r1mp
- CLEAR_SUMS
+ pxor m7, m7
+ mova [sums+ 0], m7
+ mova [sums+ 8], m7
+ mova [sums+16], m7
; 1D hadamards
- mov t0d, 4
+ mov r3d, 4
.loop_edge:
- SCALAR_HADAMARD_LEFT t0, t2, r3, r4, r5
- SCALAR_HADAMARD_TOP t0, t2, r3, r4, r5
- sub t0d, 4
+ SCALAR_HADAMARD left, r3, m0, m1
+ SCALAR_HADAMARD top, r3, m0, m1, m2
+ sub r3d, 4
jge .loop_edge
; dc
- movzx t2d, word [left_1d+0]
+ movzx t0d, word [left_1d+0]
movzx r3d, word [top_1d+0]
movzx r4d, word [left_1d+8]
movzx r5d, word [top_1d+8]
- add t2d, r3d
- lea r3, [r4 + r5]
- lea t2, [2*t2 + 8]
- lea r3, [2*r3 + 8]
- lea r4, [4*r4 + 8]
- lea r5, [4*r5 + 8]
- and t2d, -16 ; tl
+ lea t0d, [t0 + r3 + 16]
+ lea r3d, [r4 + r5 + 16]
+ shr t0d, 1
+ shr r3d, 1
+ add r4d, 8
+ add r5d, 8
+ and t0d, -16 ; tl
and r3d, -16 ; br
and r4d, -16 ; bl
and r5d, -16 ; tr
- mov [dc_1d+ 0], t2d ; tl
+ mov [dc_1d+ 0], t0d ; tl
mov [dc_1d+ 4], r5d ; tr
mov [dc_1d+ 8], r4d ; bl
mov [dc_1d+12], r3d ; br
; 2D hadamards
movifnidn r0, r0mp
movifnidn r2, r2mp
- xor r3d, r3d
+ mov r3, -2
.loop_y:
- xor r4d, r4d
+ mov r4, -2
.loop_x:
call hadamard_load
- SUM3x4 %1
- SUM4x3 [r5+4*r4], [left_1d+8*r3], [top_1d+8*r4]
+ SUM3x4
+ SUM4x3 [r5+4*(r4+2)], [left_1d+8*(r3+2)], [top_1d+8*(r4+2)]
pavgw m4, m7
pavgw m5, m7
paddw m0, [sums+16] ; i4x4_v satd
movq [sums+8], m4
movq [sums+0], m5
- add r0, 4
- inc r4d
- cmp r4d, 2
+ add r0, 4*SIZEOF_PIXEL
+ inc r4
jl .loop_x
- add r0, 4*FENC_STRIDE-8
+ add r0, 4*FENC_STRIDEB-8*SIZEOF_PIXEL
add r5, 8
- inc r3d
- cmp r3d, 2
+ inc r3
jl .loop_y
; horizontal sum
movq m1, [sums+8]
movq m2, [sums+16]
movq m7, m0
+%if HIGH_BIT_DEPTH
+ psrlq m7, 16
+ HADDW m7, m3
+ SUM_MM_X3 m0, m1, m2, m3, m4, m5, m6, paddd
+ psrld m2, 1
+ paddd m2, m7
+%else
psrlq m7, 15
paddw m2, m7
SUM_MM_X3 m0, m1, m2, m3, m4, m5, m6, paddd
psrld m2, 1
+%endif
movd [r2+0], m0 ; i8x8c_dc satd
movd [r2+4], m1 ; i8x8c_h satd
movd [r2+8], m2 ; i8x8c_v satd
ADD rsp, 72
RET
-%endmacro ; INTRA_SATDS_MMX
+%endmacro ; INTRA_X3_MMX
-%macro ABS_MOV_SSSE3 2
- pabsw %1, %2
-%endmacro
-%macro ABS_MOV_MMX 2
- pxor %1, %1
- psubw %1, %2
- pmaxsw %1, %2
+%macro PRED4x4_LOWPASS 5
+%ifid %5
+ pavgb %5, %2, %3
+ pxor %3, %2
+ pand %3, [pb_1]
+ psubusb %5, %3
+ pavgb %1, %4, %5
+%else
+ mova %5, %2
+ pavgb %2, %3
+ pxor %3, %5
+ pand %3, [pb_1]
+ psubusb %2, %3
+ pavgb %1, %4, %2
+%endif
%endmacro
-%define ABS_MOV ABS_MOV_MMX
+%macro INTRA_X9_PRED 2
+%if cpuflag(sse4)
+ movu m1, [r1-1*FDEC_STRIDE-8]
+ pinsrb m1, [r1+3*FDEC_STRIDE-1], 0
+ pinsrb m1, [r1+2*FDEC_STRIDE-1], 1
+ pinsrb m1, [r1+1*FDEC_STRIDE-1], 2
+ pinsrb m1, [r1+0*FDEC_STRIDE-1], 3
+%else
+ movd mm0, [r1+3*FDEC_STRIDE-4]
+ punpcklbw mm0, [r1+2*FDEC_STRIDE-4]
+ movd mm1, [r1+1*FDEC_STRIDE-4]
+ punpcklbw mm1, [r1+0*FDEC_STRIDE-4]
+ punpckhwd mm0, mm1
+ psrlq mm0, 32
+ movq2dq m0, mm0
+ movu m1, [r1-1*FDEC_STRIDE-8]
+ movss m1, m0 ; l3 l2 l1 l0 __ __ __ lt t0 t1 t2 t3 t4 t5 t6 t7
+%endif ; cpuflag
+ pshufb m1, [intrax9_edge] ; l3 l3 l2 l1 l0 lt t0 t1 t2 t3 t4 t5 t6 t7 t7 __
+ psrldq m0, m1, 1 ; l3 l2 l1 l0 lt t0 t1 t2 t3 t4 t5 t6 t7 t7 __ __
+ psrldq m2, m1, 2 ; l2 l1 l0 lt t0 t1 t2 t3 t4 t5 t6 t7 t7 __ __ __
+ pavgb m5, m0, m1 ; Gl3 Gl2 Gl1 Gl0 Glt Gt0 Gt1 Gt2 Gt3 Gt4 Gt5 __ __ __ __ __
+ mova %2, m1
+ PRED4x4_LOWPASS m0, m1, m2, m0, m4 ; Fl3 Fl2 Fl1 Fl0 Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5 Ft6 Ft7 __ __ __
+ ; ddl ddr
+ ; Ft1 Ft2 Ft3 Ft4 Flt Ft0 Ft1 Ft2
+ ; Ft2 Ft3 Ft4 Ft5 Fl0 Flt Ft0 Ft1
+ ; Ft3 Ft4 Ft5 Ft6 Fl1 Fl0 Flt Ft0
+ ; Ft4 Ft5 Ft6 Ft7 Fl2 Fl1 Fl0 Flt
+ pshufb m2, m0, [%1_ddlr1] ; a: ddl row0, ddl row1, ddr row0, ddr row1 / b: ddl row0, ddr row0, ddl row1, ddr row1
+ pshufb m3, m0, [%1_ddlr2] ; rows 2,3
+ ; hd hu
+ ; Glt Flt Ft0 Ft1 Gl0 Fl1 Gl1 Fl2
+ ; Gl0 Fl0 Glt Flt Gl1 Fl2 Gl2 Fl3
+ ; Gl1 Fl1 Gl0 Fl0 Gl2 Fl3 Gl3 Gl3
+ ; Gl2 Fl2 Gl1 Fl1 Gl3 Gl3 Gl3 Gl3
+ pslldq m0, 5 ; ___ ___ ___ ___ ___ Fl3 Fl2 Fl1 Fl0 Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5
+ palignr m7, m5, m0, 5 ; Fl3 Fl2 Fl1 Fl0 Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5 Gl3 Gl2 Gl1 Gl0 Glt
+ pshufb m6, m7, [%1_hdu1]
+ pshufb m7, m7, [%1_hdu2]
+ ; vr vl
+ ; Gt0 Gt1 Gt2 Gt3 Gt1 Gt2 Gt3 Gt4
+ ; Flt Ft0 Ft1 Ft2 Ft1 Ft2 Ft3 Ft4
+ ; Fl0 Gt0 Gt1 Gt2 Gt2 Gt3 Gt4 Gt5
+ ; Fl1 Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5
+ psrldq m5, 5 ; Gt0 Gt1 Gt2 Gt3 Gt4 Gt5 ...
+ palignr m5, m0, 6 ; ___ Fl1 Fl0 Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5 Gt0 Gt1 Gt2 Gt3 Gt4 Gt5
+ pshufb m4, m5, [%1_vrl1]
+ pshufb m5, m5, [%1_vrl2]
+%endmacro ; INTRA_X9_PRED
+
+%macro INTRA_X9_VHDC 5 ; edge, fenc01, fenc23, tmp, tmp
+ pshufb m2, m%1, [intrax9b_vh1]
+ pshufb m3, m%1, [intrax9b_vh2]
+ mova [pred_buf+0x60], m2
+ mova [pred_buf+0x70], m3
+ pshufb m%1, [intrax9b_edge2] ; t0 t1 t2 t3 t0 t1 t2 t3 l0 l1 l2 l3 l0 l1 l2 l3
+ pmaddubsw m%1, [hmul_4p]
+ pshufhw m0, m%1, q2301
+ pshuflw m0, m0, q2301
+ psignw m%1, [pw_pmpmpmpm]
+ paddw m0, m%1
+ psllw m0, 2 ; hadamard(top), hadamard(left)
+ movhlps m3, m0
+ pshufb m1, m0, [intrax9b_v1]
+ pshufb m2, m0, [intrax9b_v2]
+ paddw m0, m3
+ psignw m3, [pw_pmmpzzzz] ; FIXME could this be eliminated?
+ pavgw m0, [pw_16]
+ pand m0, [sw_f0] ; dc
+ ; This (as well as one of the steps in intra_satd_x9_4x4.satd_8x4) could be
+ ; changed from a wd transpose to a qdq, with appropriate rearrangement of inputs.
+ ; Which would be faster on conroe, but slower on penryn and sandybridge, and too invasive to ifdef.
+ HADAMARD 0, sumsub, %2, %3, %4, %5
+ HADAMARD 1, sumsub, %2, %3, %4, %5
+ movd r3d, m0
+ shr r3d, 4
+ imul r3d, 0x01010101
+ mov [pred_buf+0x80], r3d
+ mov [pred_buf+0x88], r3d
+ mov [pred_buf+0x90], r3d
+ mov [pred_buf+0x98], r3d
+ psubw m3, m%2
+ psubw m0, m%2
+ psubw m1, m%2
+ psubw m2, m%3
+ pabsw m%3, m%3
+ pabsw m3, m3
+ pabsw m0, m0
+ pabsw m1, m1
+ pabsw m2, m2
+ pavgw m3, m%3
+ pavgw m0, m%3
+ pavgw m1, m2
+%if cpuflag(sse4)
+ phaddw m3, m0
+%else
+ SBUTTERFLY qdq, 3, 0, 2
+ paddw m3, m0
+%endif
+ movhlps m2, m1
+ paddw m1, m2
+%if cpuflag(xop)
+ vphaddwq m3, m3
+ vphaddwq m1, m1
+ packssdw m1, m3
+%else
+ phaddw m1, m3
+ pmaddwd m1, [pw_1] ; v, _, h, dc
+%endif
+%endmacro ; INTRA_X9_VHDC
+
+%macro INTRA_X9_END 2
+%if cpuflag(sse4)
+ phminposuw m0, m0 ; h,dc,ddl,ddr,vr,hd,vl,hu
+ movd eax, m0
+ add eax, 1<<16
+ cmp ax, r3w
+ cmovge eax, r3d
+%else
+%if %1
+ ; 4x4 sad is up to 12 bits; +bitcosts -> 13 bits; pack with 3 bit index
+ psllw m0, 3
+ paddw m0, [pw_s01234567] ; h,dc,ddl,ddr,vr,hd,vl,hu
+%else
+ ; 4x4 satd is up to 13 bits; +bitcosts and saturate -> 13 bits; pack with 3 bit index
+ psllw m0, 2
+ paddusw m0, m0
+ paddw m0, [pw_s01234657] ; h,dc,ddl,ddr,vr,vl,hd,hu
+%endif
+ movhlps m1, m0
+ pminsw m0, m1
+ pshuflw m1, m0, q0032
+ pminsw m0, m1
+ pshuflw m1, m0, q0001
+ pminsw m0, m1
+ movd eax, m0
+ movsx r2d, ax
+ and eax, 7
+ sar r2d, 3
+ shl eax, 16
+ ; 1<<16: increment index to match intra4x4_pred_e. couldn't do this before because it had to fit in 3 bits
+ ; 1<<12: undo sign manipulation
+ lea eax, [rax+r2+(1<<16)+(1<<12)]
+ cmp ax, r3w
+ cmovge eax, r3d
+%endif ; cpuflag
+
+ ; output the predicted samples
+ mov r3d, eax
+ shr r3d, 16
+%ifdef PIC
+ lea r2, [%2_lut]
+ movzx r2d, byte [r2+r3]
+%else
+ movzx r2d, byte [%2_lut+r3]
+%endif
+%if %1 ; sad
+ movq mm0, [pred_buf+r2]
+ movq mm1, [pred_buf+r2+16]
+ movd [r1+0*FDEC_STRIDE], mm0
+ movd [r1+2*FDEC_STRIDE], mm1
+ psrlq mm0, 32
+ psrlq mm1, 32
+ movd [r1+1*FDEC_STRIDE], mm0
+ movd [r1+3*FDEC_STRIDE], mm1
+%else ; satd
+%assign i 0
+%rep 4
+ mov r3d, [pred_buf+r2+8*i]
+ mov [r1+i*FDEC_STRIDE], r3d
+%assign i i+1
+%endrep
+%endif
+%endmacro ; INTRA_X9_END
+
+%macro INTRA_X9 0
+;-----------------------------------------------------------------------------
+; int intra_sad_x9_4x4( uint8_t *fenc, uint8_t *fdec, uint16_t *bitcosts )
+;-----------------------------------------------------------------------------
+%if notcpuflag(xop)
+cglobal intra_sad_x9_4x4, 3,4,9
+ %assign pad 0xc0-gprsize-(stack_offset&15)
+ %define pred_buf rsp
+ sub rsp, pad
+%if ARCH_X86_64
+ INTRA_X9_PRED intrax9a, m8
+%else
+ INTRA_X9_PRED intrax9a, [rsp+0xa0]
+%endif
+ mova [rsp+0x00], m2
+ mova [rsp+0x10], m3
+ mova [rsp+0x20], m4
+ mova [rsp+0x30], m5
+ mova [rsp+0x40], m6
+ mova [rsp+0x50], m7
+%if cpuflag(sse4)
+ movd m0, [r0+0*FENC_STRIDE]
+ pinsrd m0, [r0+1*FENC_STRIDE], 1
+ movd m1, [r0+2*FENC_STRIDE]
+ pinsrd m1, [r0+3*FENC_STRIDE], 1
+%else
+ movd mm0, [r0+0*FENC_STRIDE]
+ punpckldq mm0, [r0+1*FENC_STRIDE]
+ movd mm1, [r0+2*FENC_STRIDE]
+ punpckldq mm1, [r0+3*FENC_STRIDE]
+ movq2dq m0, mm0
+ movq2dq m1, mm1
+%endif
+ punpcklqdq m0, m0
+ punpcklqdq m1, m1
+ psadbw m2, m0
+ psadbw m3, m1
+ psadbw m4, m0
+ psadbw m5, m1
+ psadbw m6, m0
+ psadbw m7, m1
+ paddd m2, m3
+ paddd m4, m5
+ paddd m6, m7
+%if ARCH_X86_64
+ SWAP 7, 8
+ pxor m8, m8
+ %define %%zero m8
+%else
+ mova m7, [rsp+0xa0]
+ %define %%zero [pb_0]
+%endif
+ pshufb m3, m7, [intrax9a_vh1]
+ pshufb m5, m7, [intrax9a_vh2]
+ pshufb m7, [intrax9a_dc]
+ psadbw m7, %%zero
+ psrlw m7, 2
+ mova [rsp+0x60], m3
+ mova [rsp+0x70], m5
+ psadbw m3, m0
+ pavgw m7, %%zero
+ pshufb m7, %%zero
+ psadbw m5, m1
+ movq [rsp+0x80], m7
+ movq [rsp+0x90], m7
+ psadbw m0, m7
+ paddd m3, m5
+ psadbw m1, m7
+ paddd m0, m1
+ movzx r3d, word [r2]
+ movd r0d, m3 ; v
+ add r3d, r0d
+ punpckhqdq m3, m0 ; h, dc
+ shufps m3, m2, q2020
+ psllq m6, 32
+ por m4, m6
+ movu m0, [r2+2]
+ packssdw m3, m4
+ paddw m0, m3
+ INTRA_X9_END 1, intrax9a
+ add rsp, pad
+ RET
+%endif ; cpuflag
+
+%if ARCH_X86_64
+;-----------------------------------------------------------------------------
+; int intra_satd_x9_4x4( uint8_t *fenc, uint8_t *fdec, uint16_t *bitcosts )
+;-----------------------------------------------------------------------------
+cglobal intra_satd_x9_4x4, 3,4,16
+ %assign pad 0xb0-gprsize-(stack_offset&15)
+ %define pred_buf rsp
+ sub rsp, pad
+ INTRA_X9_PRED intrax9b, m15
+ mova [rsp+0x00], m2
+ mova [rsp+0x10], m3
+ mova [rsp+0x20], m4
+ mova [rsp+0x30], m5
+ mova [rsp+0x40], m6
+ mova [rsp+0x50], m7
+ movd m8, [r0+0*FENC_STRIDE]
+ movd m9, [r0+1*FENC_STRIDE]
+ movd m10, [r0+2*FENC_STRIDE]
+ movd m11, [r0+3*FENC_STRIDE]
+ mova m12, [hmul_8p]
+ pshufd m8, m8, 0
+ pshufd m9, m9, 0
+ pshufd m10, m10, 0
+ pshufd m11, m11, 0
+ pmaddubsw m8, m12
+ pmaddubsw m9, m12
+ pmaddubsw m10, m12
+ pmaddubsw m11, m12
+ movddup m0, m2
+ pshufd m1, m2, q3232
+ movddup m2, m3
+ movhlps m3, m3
+ call .satd_8x4 ; ddr, ddl
+ movddup m2, m5
+ pshufd m3, m5, q3232
+ mova m5, m0
+ movddup m0, m4
+ pshufd m1, m4, q3232
+ call .satd_8x4 ; vr, vl
+ movddup m2, m7
+ pshufd m3, m7, q3232
+ mova m4, m0
+ movddup m0, m6
+ pshufd m1, m6, q3232
+ call .satd_8x4 ; hd, hu
+%if cpuflag(sse4)
+ punpckldq m4, m0
+%else
+ punpcklqdq m4, m0 ; conroe dislikes punpckldq, and ssse3 INTRA_X9_END can handle arbitrary orders whereas phminposuw can't
+%endif
+ mova m1, [pw_ppmmppmm]
+ psignw m8, m1
+ psignw m10, m1
+ paddw m8, m9
+ paddw m10, m11
+ INTRA_X9_VHDC 15, 8, 10, 6, 7
+ ; find minimum
+ movu m0, [r2+2]
+ movd r3d, m1
+ palignr m5, m1, 8
+%if notcpuflag(sse4)
+ pshufhw m0, m0, q3120 ; compensate for different order in unpack
+%endif
+ packssdw m5, m4
+ paddw m0, m5
+ movzx r0d, word [r2]
+ add r3d, r0d
+ INTRA_X9_END 0, intrax9b
+ add rsp, pad
+ RET
+RESET_MM_PERMUTATION
+ALIGN 16
+.satd_8x4:
+ pmaddubsw m0, m12
+ pmaddubsw m1, m12
+ pmaddubsw m2, m12
+ pmaddubsw m3, m12
+ psubw m0, m8
+ psubw m1, m9
+ psubw m2, m10
+ psubw m3, m11
+ SATD_8x4_SSE cpuname, 0, 1, 2, 3, 13, 14, 0, swap
+ pmaddwd m0, [pw_1]
+%if cpuflag(sse4)
+ pshufd m1, m0, q0032
+%else
+ movhlps m1, m0
+%endif
+ paddd xmm0, m0, m1 ; consistent location of return value. only the avx version of hadamard permutes m0, so 3arg is free
+ ret
+
+%else ; !ARCH_X86_64
+cglobal intra_satd_x9_4x4, 3,4,8
+ %assign pad 0x120-gprsize-(stack_offset&15)
+ %define fenc_buf rsp
+ %define pred_buf rsp+0x40
+ %define spill rsp+0xe0
+ sub rsp, pad
+ INTRA_X9_PRED intrax9b, [spill+0x20]
+ mova [pred_buf+0x00], m2
+ mova [pred_buf+0x10], m3
+ mova [pred_buf+0x20], m4
+ mova [pred_buf+0x30], m5
+ mova [pred_buf+0x40], m6
+ mova [pred_buf+0x50], m7
+ movd m4, [r0+0*FENC_STRIDE]
+ movd m5, [r0+1*FENC_STRIDE]
+ movd m6, [r0+2*FENC_STRIDE]
+ movd m0, [r0+3*FENC_STRIDE]
+ mova m7, [hmul_8p]
+ pshufd m4, m4, 0
+ pshufd m5, m5, 0
+ pshufd m6, m6, 0
+ pshufd m0, m0, 0
+ pmaddubsw m4, m7
+ pmaddubsw m5, m7
+ pmaddubsw m6, m7
+ pmaddubsw m0, m7
+ mova [fenc_buf+0x00], m4
+ mova [fenc_buf+0x10], m5
+ mova [fenc_buf+0x20], m6
+ mova [fenc_buf+0x30], m0
+ movddup m0, m2
+ pshufd m1, m2, q3232
+ movddup m2, m3
+ movhlps m3, m3
+ pmaddubsw m0, m7
+ pmaddubsw m1, m7
+ pmaddubsw m2, m7
+ pmaddubsw m3, m7
+ psubw m0, m4
+ psubw m1, m5
+ psubw m2, m6
+ call .satd_8x4b ; ddr, ddl
+ mova m3, [pred_buf+0x30]
+ mova m1, [pred_buf+0x20]
+ movddup m2, m3
+ movhlps m3, m3
+ movq [spill+0x08], m0
+ movddup m0, m1
+ movhlps m1, m1
+ call .satd_8x4 ; vr, vl
+ mova m3, [pred_buf+0x50]
+ mova m1, [pred_buf+0x40]
+ movddup m2, m3
+ movhlps m3, m3
+ movq [spill+0x10], m0
+ movddup m0, m1
+ movhlps m1, m1
+ call .satd_8x4 ; hd, hu
+ movq [spill+0x18], m0
+ mova m1, [spill+0x20]
+ mova m4, [fenc_buf+0x00]
+ mova m5, [fenc_buf+0x20]
+ mova m2, [pw_ppmmppmm]
+ psignw m4, m2
+ psignw m5, m2
+ paddw m4, [fenc_buf+0x10]
+ paddw m5, [fenc_buf+0x30]
+ INTRA_X9_VHDC 1, 4, 5, 6, 7
+ ; find minimum
+ movu m0, [r2+2]
+ movd r3d, m1
+ punpckhqdq m1, [spill+0x00]
+ packssdw m1, [spill+0x10]
+%if cpuflag(sse4)
+ pshufhw m1, m1, q3120
+%else
+ pshufhw m0, m0, q3120
+%endif
+ paddw m0, m1
+ movzx r0d, word [r2]
+ add r3d, r0d
+ INTRA_X9_END 0, intrax9b
+ add rsp, pad
+ RET
+RESET_MM_PERMUTATION
+ALIGN 16
+.satd_8x4:
+ pmaddubsw m0, m7
+ pmaddubsw m1, m7
+ pmaddubsw m2, m7
+ pmaddubsw m3, m7
+ %xdefine fenc_buf fenc_buf+gprsize
+ psubw m0, [fenc_buf+0x00]
+ psubw m1, [fenc_buf+0x10]
+ psubw m2, [fenc_buf+0x20]
+.satd_8x4b:
+ psubw m3, [fenc_buf+0x30]
+ SATD_8x4_SSE cpuname, 0, 1, 2, 3, 4, 5, 0, swap
+ pmaddwd m0, [pw_1]
+%if cpuflag(sse4)
+ pshufd m1, m0, q0032
+%else
+ movhlps m1, m0
+%endif
+ paddd xmm0, m0, m1
+ ret
+%endif ; ARCH
+%endmacro ; INTRA_X9
+
+
+
+%macro INTRA8_X9 0
+;-----------------------------------------------------------------------------
+; int intra_sad_x9_8x8( uint8_t *fenc, uint8_t *fdec, uint8_t edge[36], uint16_t *bitcosts, uint16_t *satds )
+;-----------------------------------------------------------------------------
+cglobal intra_sad_x9_8x8, 5,6,9
+ %define fenc02 m4
+ %define fenc13 m5
+ %define fenc46 m6
+ %define fenc57 m7
+%if ARCH_X86_64
+ %define tmp m8
+ %assign padbase 0x0
+%else
+ %define tmp [rsp]
+ %assign padbase 0x10
+%endif
+ %assign pad 0x240+0x10+padbase-gprsize-(stack_offset&15)
+ %define pred(i,j) [rsp+i*0x40+j*0x10+padbase]
+
+ SUB rsp, pad
+ movq fenc02, [r0+FENC_STRIDE* 0]
+ movq fenc13, [r0+FENC_STRIDE* 1]
+ movq fenc46, [r0+FENC_STRIDE* 4]
+ movq fenc57, [r0+FENC_STRIDE* 5]
+ movhps fenc02, [r0+FENC_STRIDE* 2]
+ movhps fenc13, [r0+FENC_STRIDE* 3]
+ movhps fenc46, [r0+FENC_STRIDE* 6]
+ movhps fenc57, [r0+FENC_STRIDE* 7]
+
+ ; save instruction size: avoid 4-byte memory offsets
+ lea r0, [intra8x9_h1+128]
+ %define off(m) (r0+m-(intra8x9_h1+128))
+
+; v
+ movddup m0, [r2+16]
+ mova pred(0,0), m0
+ psadbw m1, m0, fenc02
+ mova pred(0,1), m0
+ psadbw m2, m0, fenc13
+ mova pred(0,2), m0
+ psadbw m3, m0, fenc46
+ mova pred(0,3), m0
+ psadbw m0, m0, fenc57
+ paddw m1, m2
+ paddw m0, m3
+ paddw m0, m1
+ movhlps m1, m0
+ paddw m0, m1
+ movd [r4+0], m0
+
+; h
+ movq m0, [r2+7]
+ pshufb m1, m0, [off(intra8x9_h1)]
+ pshufb m2, m0, [off(intra8x9_h2)]
+ mova pred(1,0), m1
+ psadbw m1, fenc02
+ mova pred(1,1), m2
+ psadbw m2, fenc13
+ paddw m1, m2
+ pshufb m3, m0, [off(intra8x9_h3)]
+ pshufb m2, m0, [off(intra8x9_h4)]
+ mova pred(1,2), m3
+ psadbw m3, fenc46
+ mova pred(1,3), m2
+ psadbw m2, fenc57
+ paddw m1, m3
+ paddw m1, m2
+ movhlps m2, m1
+ paddw m1, m2
+ movd [r4+2], m1
+
+ lea r5, [rsp+padbase+0x100]
+ %define pred(i,j) [r5+i*0x40+j*0x10-0x100]
+
+; dc
+ movhps m0, [r2+16]
+ pxor m2, m2
+ psadbw m0, m2
+ movhlps m1, m0
+ paddw m0, m1
+ psrlw m0, 3
+ pavgw m0, m2
+ pshufb m0, m2
+ mova pred(2,0), m0
+ psadbw m1, m0, fenc02
+ mova pred(2,1), m0
+ psadbw m2, m0, fenc13
+ mova pred(2,2), m0
+ psadbw m3, m0, fenc46
+ mova pred(2,3), m0
+ psadbw m0, m0, fenc57
+ paddw m1, m2
+ paddw m0, m3
+ paddw m0, m1
+ movhlps m1, m0
+ paddw m0, m1
+ movd [r4+4], m0
+
+; ddl
+; Ft1 Ft2 Ft3 Ft4 Ft5 Ft6 Ft7 Ft8
+; Ft2 Ft3 Ft4 Ft5 Ft6 Ft7 Ft8 Ft9
+; Ft3 Ft4 Ft5 Ft6 Ft7 Ft8 Ft9 FtA
+; Ft4 Ft5 Ft6 Ft7 Ft8 Ft9 FtA FtB
+; Ft5 Ft6 Ft7 Ft8 Ft9 FtA FtB FtC
+; Ft6 Ft7 Ft8 Ft9 FtA FtB FtC FtD
+; Ft7 Ft8 Ft9 FtA FtB FtC FtD FtE
+; Ft8 Ft9 FtA FtB FtC FtD FtE FtF
+ mova m0, [r2+16]
+ movu m2, [r2+17]
+ pslldq m1, m0, 1
+ pavgb m3, m0, m2 ; Gt1 Gt2 Gt3 Gt4 Gt5 Gt6 Gt7 Gt8 Gt9 GtA GtB ___ ___ ___ ___ ___
+ PRED4x4_LOWPASS m0, m1, m2, m0, tmp ; ___ Ft1 Ft2 Ft3 Ft4 Ft5 Ft6 Ft7 Ft8 Ft9 FtA FtB FtC FtD FtE FtF
+ pshufb m1, m0, [off(intra8x9_ddl1)]
+ pshufb m2, m0, [off(intra8x9_ddl2)]
+ mova pred(3,0), m1
+ psadbw m1, fenc02
+ mova pred(3,1), m2
+ psadbw m2, fenc13
+ paddw m1, m2
+ pshufb m2, m0, [off(intra8x9_ddl3)]
+ mova pred(3,2), m2
+ psadbw m2, fenc46
+ paddw m1, m2
+ pshufb m2, m0, [off(intra8x9_ddl4)]
+ mova pred(3,3), m2
+ psadbw m2, fenc57
+ paddw m1, m2
+ movhlps m2, m1
+ paddw m1, m2
+ movd [r4+6], m1
+
+; vl
+; Gt1 Gt2 Gt3 Gt4 Gt5 Gt6 Gt7 Gt8
+; Ft1 Ft2 Ft3 Ft4 Ft5 Ft6 Ft7 Ft8
+; Gt2 Gt3 Gt4 Gt5 Gt6 Gt7 Gt8 Gt9
+; Ft2 Ft3 Ft4 Ft5 Ft6 Ft7 Ft8 Ft9
+; Gt3 Gt4 Gt5 Gt6 Gt7 Gt8 Gt9 GtA
+; Ft3 Ft4 Ft5 Ft6 Ft7 Ft8 Ft9 FtA
+; Gt4 Gt5 Gt6 Gt7 Gt8 Gt9 GtA GtB
+; Ft4 Ft5 Ft6 Ft7 Ft8 Ft9 FtA FtB
+ pshufb m1, m3, [off(intra8x9_vl1)]
+ pshufb m2, m0, [off(intra8x9_vl2)]
+ pshufb m3, m3, [off(intra8x9_vl3)]
+ pshufb m0, m0, [off(intra8x9_vl4)]
+ mova pred(7,0), m1
+ psadbw m1, fenc02
+ mova pred(7,1), m2
+ psadbw m2, fenc13
+ mova pred(7,2), m3
+ psadbw m3, fenc46
+ mova pred(7,3), m0
+ psadbw m0, fenc57
+ paddw m1, m2
+ paddw m0, m3
+ paddw m0, m1
+ movhlps m1, m0
+ paddw m0, m1
+%if cpuflag(sse4)
+ pextrw [r4+14], m0, 0
+%else
+ movd r5d, m0
+ mov [r4+14], r5w
+ lea r5, [rsp+padbase+0x100]
+%endif
+
+; ddr
+; Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5 Ft6
+; Fl0 Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5
+; Fl1 Fl0 Flt Ft0 Ft1 Ft2 Ft3 Ft4
+; Fl2 Fl1 Fl0 Flt Ft0 Ft1 Ft2 Ft3
+; Fl3 Fl2 Fl1 Fl0 Flt Ft0 Ft1 Ft2
+; Fl4 Fl3 Fl2 Fl1 Fl0 Flt Ft0 Ft1
+; Fl5 Fl4 Fl3 Fl2 Fl1 Fl0 Flt Ft0
+; Fl6 Fl5 Fl4 Fl3 Fl2 Fl1 Fl0 Flt
+ movu m2, [r2+8]
+ movu m0, [r2+7]
+ movu m1, [r2+6]
+ pavgb m3, m2, m0 ; Gl6 Gl5 Gl4 Gl3 Gl2 Gl1 Gl0 Glt Gt0 Gt1 Gt2 Gt3 Gt4 Gt5 Gt6 Gt7
+ PRED4x4_LOWPASS m0, m1, m2, m0, tmp ; Fl7 Fl6 Fl5 Fl4 Fl3 Fl2 Fl1 Fl0 Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5 Ft6
+ pshufb m1, m0, [off(intra8x9_ddr1)]
+ pshufb m2, m0, [off(intra8x9_ddr2)]
+ mova pred(4,0), m1
+ psadbw m1, fenc02
+ mova pred(4,1), m2
+ psadbw m2, fenc13
+ paddw m1, m2
+ pshufb m2, m0, [off(intra8x9_ddr3)]
+ mova pred(4,2), m2
+ psadbw m2, fenc46
+ paddw m1, m2
+ pshufb m2, m0, [off(intra8x9_ddr4)]
+ mova pred(4,3), m2
+ psadbw m2, fenc57
+ paddw m1, m2
+ movhlps m2, m1
+ paddw m1, m2
+ movd [r4+8], m1
+
+ add r0, 256
+ add r5, 0xC0
+ %define off(m) (r0+m-(intra8x9_h1+256+128))
+ %define pred(i,j) [r5+i*0x40+j*0x10-0x1C0]
+
+; vr
+; Gt0 Gt1 Gt2 Gt3 Gt4 Gt5 Gt6 Gt7
+; Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5 Ft6
+; Fl0 Gt0 Gt1 Gt2 Gt3 Gt4 Gt5 Gt6
+; Fl1 Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5
+; Fl2 Fl0 Gt0 Gt1 Gt2 Gt3 Gt4 Gt5
+; Fl3 Fl1 Flt Ft0 Ft1 Ft2 Ft3 Ft4
+; Fl4 Fl2 Fl0 Gt0 Gt1 Gt2 Gt3 Gt4
+; Fl5 Fl3 Fl1 Flt Ft0 Ft1 Ft2 Ft3
+ movsd m2, m3, m0 ; Fl7 Fl6 Fl5 Fl4 Fl3 Fl2 Fl1 Fl0 Gt0 Gt1 Gt2 Gt3 Gt4 Gt5 Gt6 Gt7
+ pshufb m1, m2, [off(intra8x9_vr1)]
+ pshufb m2, m2, [off(intra8x9_vr3)]
+ mova pred(5,0), m1
+ psadbw m1, fenc02
+ mova pred(5,2), m2
+ psadbw m2, fenc46
+ paddw m1, m2
+ pshufb m2, m0, [off(intra8x9_vr2)]
+ mova pred(5,1), m2
+ psadbw m2, fenc13
+ paddw m1, m2
+ pshufb m2, m0, [off(intra8x9_vr4)]
+ mova pred(5,3), m2
+ psadbw m2, fenc57
+ paddw m1, m2
+ movhlps m2, m1
+ paddw m1, m2
+ movd [r4+10], m1
+
+; hd
+; Glt Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5
+; Gl0 Fl0 Glt Flt Ft0 Ft1 Ft2 Ft3
+; Gl1 Fl1 Gl0 Fl0 Glt Flt Ft0 Ft1
+; Gl2 Fl2 Gl1 Fl1 Gl0 Fl0 Glt Flt
+; Gl3 Fl3 Gl2 Fl2 Gl1 Fl1 Gl0 Fl0
+; Gl4 Fl4 Gl3 Fl3 Gl2 Fl2 Gl1 Fl1
+; Gl5 Fl5 Gl4 Fl4 Gl3 Fl3 Gl2 Fl2
+; Gl6 Fl6 Gl5 Fl5 Gl4 Fl4 Gl3 Fl3
+ pshufd m2, m3, q0001
+%if cpuflag(sse4)
+ pblendw m2, m0, q3330 ; Gl2 Gl1 Gl0 Glt ___ Fl2 Fl1 Fl0 Flt Ft0 Ft1 Ft2 Ft3 Ft4 Ft5 ___
+%else
+ movss m1, m0, m2
+ SWAP 1, 2
+%endif
+ punpcklbw m0, m3 ; Fl7 Gl6 Fl6 Gl5 Fl5 Gl4 Fl4 Gl3 Fl3 Gl2 Fl2 Gl1 Fl1 Gl0 Fl0 ___
+ pshufb m1, m2, [off(intra8x9_hd1)]
+ pshufb m2, m2, [off(intra8x9_hd2)]
+ mova pred(6,0), m1
+ psadbw m1, fenc02
+ mova pred(6,1), m2
+ psadbw m2, fenc13
+ paddw m1, m2
+ pshufb m2, m0, [off(intra8x9_hd3)]
+ pshufb m3, m0, [off(intra8x9_hd4)]
+ mova pred(6,2), m2
+ psadbw m2, fenc46
+ mova pred(6,3), m3
+ psadbw m3, fenc57
+ paddw m1, m2
+ paddw m1, m3
+ movhlps m2, m1
+ paddw m1, m2
+ ; don't just store to [r4+12]. this is too close to the load of dqword [r4] and would cause a forwarding stall
+ pslldq m1, 12
+ SWAP 3, 1
+
+; hu
+; Gl0 Fl1 Gl1 Fl2 Gl2 Fl3 Gl3 Fl4
+; Gl1 Fl2 Gl2 Fl3 Gl3 Fl4 Gl4 Fl5
+; Gl2 Fl3 Gl3 Gl3 Gl4 Fl5 Gl5 Fl6
+; Gl3 Gl3 Gl4 Fl5 Gl5 Fl6 Gl6 Fl7
+; Gl4 Fl5 Gl5 Fl6 Gl6 Fl7 Gl7 Gl7
+; Gl5 Fl6 Gl6 Fl7 Gl7 Gl7 Gl7 Gl7
+; Gl6 Fl7 Gl7 Gl7 Gl7 Gl7 Gl7 Gl7
+; Gl7 Gl7 Gl7 Gl7 Gl7 Gl7 Gl7 Gl7
+%if cpuflag(sse4)
+ pinsrb m0, [r2+7], 15 ; Gl7
+%else
+ movd m1, [r2+7]
+ pslldq m0, 1
+ palignr m1, m0, 1
+ SWAP 0, 1
+%endif
+ pshufb m1, m0, [off(intra8x9_hu1)]
+ pshufb m2, m0, [off(intra8x9_hu2)]
+ mova pred(8,0), m1
+ psadbw m1, fenc02
+ mova pred(8,1), m2
+ psadbw m2, fenc13
+ paddw m1, m2
+ pshufb m2, m0, [off(intra8x9_hu3)]
+ pshufb m0, m0, [off(intra8x9_hu4)]
+ mova pred(8,2), m2
+ psadbw m2, fenc46
+ mova pred(8,3), m0
+ psadbw m0, fenc57
+ paddw m1, m2
+ paddw m1, m0
+ movhlps m2, m1
+ paddw m1, m2
+ movd r2d, m1
+
+ movu m0, [r3]
+ por m3, [r4]
+ paddw m0, m3
+ mova [r4], m0
+ movzx r5d, word [r3+16]
+ add r2d, r5d
+ mov [r4+16], r2w
+
+%if cpuflag(sse4)
+ phminposuw m0, m0 ; v,h,dc,ddl,ddr,vr,hd,vl
+ movd eax, m0
+%else
+ ; 8x8 sad is up to 14 bits; +bitcosts and saturate -> 14 bits; pack with 2 bit index
+ paddusw m0, m0
+ paddusw m0, m0
+ paddw m0, [off(pw_s00112233)]
+ movhlps m1, m0
+ pminsw m0, m1
+ pshuflw m1, m0, q0032
+ pminsw m0, m1
+ movd eax, m0
+ ; repack with 3 bit index
+ xor eax, 0x80008000
+ movzx r3d, ax
+ shr eax, 15
+ add r3d, r3d
+ or eax, 1
+ cmp eax, r3d
+ cmovg eax, r3d
+ ; reverse to phminposuw order
+ mov r3d, eax
+ and eax, 7
+ shr r3d, 3
+ shl eax, 16
+ or eax, r3d
+%endif
+ add r2d, 8<<16
+ cmp ax, r2w
+ cmovg eax, r2d
+
+ mov r2d, eax
+ shr r2d, 16
+ shl r2d, 6
+ add r1, 4*FDEC_STRIDE
+ mova m0, [rsp+padbase+r2+0x00]
+ mova m1, [rsp+padbase+r2+0x10]
+ mova m2, [rsp+padbase+r2+0x20]
+ mova m3, [rsp+padbase+r2+0x30]
+ movq [r1+FDEC_STRIDE*-4], m0
+ movhps [r1+FDEC_STRIDE*-2], m0
+ movq [r1+FDEC_STRIDE*-3], m1
+ movhps [r1+FDEC_STRIDE*-1], m1
+ movq [r1+FDEC_STRIDE* 0], m2
+ movhps [r1+FDEC_STRIDE* 2], m2
+ movq [r1+FDEC_STRIDE* 1], m3
+ movhps [r1+FDEC_STRIDE* 3], m3
+ ADD rsp, pad
+ RET
+
+%if ARCH_X86_64
+;-----------------------------------------------------------------------------
+; int intra_sa8d_x9_8x8( uint8_t *fenc, uint8_t *fdec, uint8_t edge[36], uint16_t *bitcosts, uint16_t *satds )
+;-----------------------------------------------------------------------------
+cglobal intra_sa8d_x9_8x8, 5,6,16
+ %assign pad 0x2c0+0x10-gprsize-(stack_offset&15)
+ %define fenc_buf rsp
+ %define pred_buf rsp+0x80
+ SUB rsp, pad
+ mova m15, [hmul_8p]
+ pxor m8, m8
+%assign %%i 0
+%rep 8
+ movddup m %+ %%i, [r0+%%i*FENC_STRIDE]
+ pmaddubsw m9, m %+ %%i, m15
+ punpcklbw m %+ %%i, m8
+ mova [fenc_buf+%%i*0x10], m9
+%assign %%i %%i+1
+%endrep
+
+ ; save instruction size: avoid 4-byte memory offsets
+ lea r0, [intra8x9_h1+0x80]
+ %define off(m) (r0+m-(intra8x9_h1+0x80))
+ lea r5, [pred_buf+0x80]
+
+; v, h, dc
+ HADAMARD8_2D 0, 1, 2, 3, 4, 5, 6, 7, 8
+ pabsw m11, m1
+%assign %%i 2
+%rep 6
+ pabsw m8, m %+ %%i
+ paddw m11, m8
+%assign %%i %%i+1
+%endrep
+
+ ; 1D hadamard of edges
+ movq m8, [r2+7]
+ movddup m9, [r2+16]
+ mova [r5-0x80], m9
+ mova [r5-0x70], m9
+ mova [r5-0x60], m9
+ mova [r5-0x50], m9
+ punpcklwd m8, m8
+ pshufb m9, [intrax3_shuf]
+ pmaddubsw m8, [pb_pppm]
+ pmaddubsw m9, [pb_pppm]
+ HSUMSUB2 psignw, m8, m9, m12, m13, m9, q1032, [pw_ppppmmmm]
+ HSUMSUB2 psignw, m8, m9, m12, m13, m9, q2301, [pw_ppmmppmm]
+
+ ; dc
+ paddw m10, m8, m9
+ paddw m10, [pw_8]
+ pand m10, [sw_f0]
+ psrlw m12, m10, 4
+ psllw m10, 2
+ pxor m13, m13
+ pshufb m12, m13
+ mova [r5+0x00], m12
+ mova [r5+0x10], m12
+ mova [r5+0x20], m12
+ mova [r5+0x30], m12
+
+ ; differences
+ psllw m8, 3 ; left edge
+ psubw m8, m0
+ psubw m10, m0
+ pabsw m8, m8 ; 1x8 sum
+ pabsw m10, m10
+ paddw m8, m11
+ paddw m11, m10
+ punpcklwd m0, m1
+ punpcklwd m2, m3
+ punpcklwd m4, m5
+ punpcklwd m6, m7
+ punpckldq m0, m2
+ punpckldq m4, m6
+ punpcklqdq m0, m4 ; transpose
+ psllw m9, 3 ; top edge
+ psrldq m10, m11, 2 ; 8x7 sum
+ psubw m0, m9 ; 8x1 sum
+ pabsw m0, m0
+ paddw m10, m0
+
+ phaddd m10, m8 ; logically phaddw, but this is faster and it won't overflow
+ psrlw m11, 1
+ psrlw m10, 1
+
+; store h
+ movq m3, [r2+7]
+ pshufb m0, m3, [off(intra8x9_h1)]
+ pshufb m1, m3, [off(intra8x9_h2)]
+ pshufb m2, m3, [off(intra8x9_h3)]
+ pshufb m3, m3, [off(intra8x9_h4)]
+ mova [r5-0x40], m0
+ mova [r5-0x30], m1
+ mova [r5-0x20], m2
+ mova [r5-0x10], m3
+
+; ddl
+ mova m8, [r2+16]
+ movu m2, [r2+17]
+ pslldq m1, m8, 1
+ pavgb m9, m8, m2
+ PRED4x4_LOWPASS m8, m1, m2, m8, m3
+ pshufb m0, m8, [off(intra8x9_ddl1)]
+ pshufb m1, m8, [off(intra8x9_ddl2)]
+ pshufb m2, m8, [off(intra8x9_ddl3)]
+ pshufb m3, m8, [off(intra8x9_ddl4)]
+ add r5, 0x40
+ call .sa8d
+ phaddd m11, m0
+
+; vl
+ pshufb m0, m9, [off(intra8x9_vl1)]
+ pshufb m1, m8, [off(intra8x9_vl2)]
+ pshufb m2, m9, [off(intra8x9_vl3)]
+ pshufb m3, m8, [off(intra8x9_vl4)]
+ add r5, 0x100
+ call .sa8d
+ phaddd m10, m11
+ mova m12, m0
+
+; ddr
+ movu m2, [r2+8]
+ movu m8, [r2+7]
+ movu m1, [r2+6]
+ pavgb m9, m2, m8
+ PRED4x4_LOWPASS m8, m1, m2, m8, m3
+ pshufb m0, m8, [off(intra8x9_ddr1)]
+ pshufb m1, m8, [off(intra8x9_ddr2)]
+ pshufb m2, m8, [off(intra8x9_ddr3)]
+ pshufb m3, m8, [off(intra8x9_ddr4)]
+ sub r5, 0xc0
+ call .sa8d
+ mova m11, m0
+
+ add r0, 0x100
+ %define off(m) (r0+m-(intra8x9_h1+0x180))
+
+; vr
+ movsd m2, m9, m8
+ pshufb m0, m2, [off(intra8x9_vr1)]
+ pshufb m1, m8, [off(intra8x9_vr2)]
+ pshufb m2, m2, [off(intra8x9_vr3)]
+ pshufb m3, m8, [off(intra8x9_vr4)]
+ add r5, 0x40
+ call .sa8d
+ phaddd m11, m0
+
+; hd
+%if cpuflag(sse4)
+ pshufd m1, m9, q0001
+ pblendw m1, m8, q3330
+%else
+ pshufd m2, m9, q0001
+ movss m1, m8, m2
+%endif
+ punpcklbw m8, m9
+ pshufb m0, m1, [off(intra8x9_hd1)]
+ pshufb m1, m1, [off(intra8x9_hd2)]
+ pshufb m2, m8, [off(intra8x9_hd3)]
+ pshufb m3, m8, [off(intra8x9_hd4)]
+ add r5, 0x40
+ call .sa8d
+ phaddd m0, m12
+ phaddd m11, m0
+
+; hu
+%if cpuflag(sse4)
+ pinsrb m8, [r2+7], 15
+%else
+ movd m9, [r2+7]
+ pslldq m8, 1
+ palignr m9, m8, 1
+ SWAP 8, 9
+%endif
+ pshufb m0, m8, [off(intra8x9_hu1)]
+ pshufb m1, m8, [off(intra8x9_hu2)]
+ pshufb m2, m8, [off(intra8x9_hu3)]
+ pshufb m3, m8, [off(intra8x9_hu4)]
+ add r5, 0x80
+ call .sa8d
+
+ pmaddwd m0, [pw_1]
+ phaddw m10, m11
+ movhlps m1, m0
+ paddw m0, m1
+ pshuflw m1, m0, q0032
+ pavgw m0, m1
+ pxor m2, m2
+ pavgw m10, m2
+ movd r2d, m0
+
+ movu m0, [r3]
+ paddw m0, m10
+ mova [r4], m0
+ movzx r5d, word [r3+16]
+ add r2d, r5d
+ mov [r4+16], r2w
+
+%if cpuflag(sse4)
+ phminposuw m0, m0
+ movd eax, m0
+%else
+ ; 8x8 sa8d is up to 15 bits; +bitcosts and saturate -> 15 bits; pack with 1 bit index
+ paddusw m0, m0
+ paddw m0, [off(pw_s00001111)]
+ movhlps m1, m0
+ pminsw m0, m1
+ pshuflw m1, m0, q0032
+ mova m2, m0
+ pminsw m0, m1
+ pcmpgtw m2, m1 ; 2nd index bit
+ movd r3d, m0
+ movd r4d, m2
+ ; repack with 3 bit index
+ xor r3d, 0x80008000
+ and r4d, 0x00020002
+ movzx eax, r3w
+ movzx r5d, r4w
+ shr r3d, 16
+ shr r4d, 16
+ lea eax, [rax*4+r5]
+ lea r3d, [ r3*4+r4+1]
+ cmp eax, r3d
+ cmovg eax, r3d
+ ; reverse to phminposuw order
+ mov r3d, eax
+ and eax, 7
+ shr r3d, 3
+ shl eax, 16
+ or eax, r3d
+%endif
+ add r2d, 8<<16
+ cmp ax, r2w
+ cmovg eax, r2d
+
+ mov r2d, eax
+ shr r2d, 16
+ shl r2d, 6
+ add r1, 4*FDEC_STRIDE
+ mova m0, [pred_buf+r2+0x00]
+ mova m1, [pred_buf+r2+0x10]
+ mova m2, [pred_buf+r2+0x20]
+ mova m3, [pred_buf+r2+0x30]
+ movq [r1+FDEC_STRIDE*-4], m0
+ movhps [r1+FDEC_STRIDE*-2], m0
+ movq [r1+FDEC_STRIDE*-3], m1
+ movhps [r1+FDEC_STRIDE*-1], m1
+ movq [r1+FDEC_STRIDE* 0], m2
+ movhps [r1+FDEC_STRIDE* 2], m2
+ movq [r1+FDEC_STRIDE* 1], m3
+ movhps [r1+FDEC_STRIDE* 3], m3
+ ADD rsp, pad
+ RET
+
+ALIGN 16
+.sa8d:
+ %xdefine mret m0
+ %xdefine fenc_buf fenc_buf+gprsize
+ mova [r5+0x00], m0
+ mova [r5+0x10], m1
+ mova [r5+0x20], m2
+ mova [r5+0x30], m3
+ movddup m4, m0
+ movddup m5, m1
+ movddup m6, m2
+ movddup m7, m3
+ punpckhqdq m0, m0
+ punpckhqdq m1, m1
+ punpckhqdq m2, m2
+ punpckhqdq m3, m3
+ PERMUTE 0,4, 1,5, 2,0, 3,1, 4,6, 5,7, 6,2, 7,3
+ pmaddubsw m0, m15
+ pmaddubsw m1, m15
+ psubw m0, [fenc_buf+0x00]
+ psubw m1, [fenc_buf+0x10]
+ pmaddubsw m2, m15
+ pmaddubsw m3, m15
+ psubw m2, [fenc_buf+0x20]
+ psubw m3, [fenc_buf+0x30]
+ pmaddubsw m4, m15
+ pmaddubsw m5, m15
+ psubw m4, [fenc_buf+0x40]
+ psubw m5, [fenc_buf+0x50]
+ pmaddubsw m6, m15
+ pmaddubsw m7, m15
+ psubw m6, [fenc_buf+0x60]
+ psubw m7, [fenc_buf+0x70]
+ HADAMARD8_2D_HMUL 0, 1, 2, 3, 4, 5, 6, 7, 13, 14
+ paddw m0, m1
+ paddw m0, m2
+ paddw mret, m0, m3
+ ret
+%endif ; ARCH_X86_64
+%endmacro ; INTRA8_X9
; in: r0=pix, r1=stride, r2=stride*3, r3=tmp, m6=mask_ac4, m7=0
; out: [tmp]=hadamard4, m0=satd
-cglobal hadamard_ac_4x4_mmxext
-%ifdef HIGH_BIT_DEPTH
+INIT_MMX mmx2
+cglobal hadamard_ac_4x4
+%if HIGH_BIT_DEPTH
mova m0, [r0]
mova m1, [r0+r1]
mova m2, [r0+r1*2]
mova [r3+8], m1
mova [r3+16], m2
mova [r3+24], m3
- ABS1 m0, m4
- ABS1 m1, m4
+ ABSW m0, m0, m4
+ ABSW m1, m1, m4
pand m0, m6
- ABS1 m2, m4
- ABS1 m3, m4
+ ABSW m2, m2, m4
+ ABSW m3, m3, m4
paddw m0, m1
paddw m2, m3
paddw m0, m2
- SAVE_MM_PERMUTATION hadamard_ac_4x4_mmxext
+ SAVE_MM_PERMUTATION
ret
-cglobal hadamard_ac_2x2max_mmxext
+cglobal hadamard_ac_2x2max
mova m0, [r3+0x00]
mova m1, [r3+0x20]
mova m2, [r3+0x40]
mova m3, [r3+0x60]
sub r3, 8
- SUMSUB_BADC w, m0, m1, m2, m3, m4
- ABS4 m0, m2, m1, m3, m4, m5
+ SUMSUB_BADC w, 0, 1, 2, 3, 4
+ ABSW2 m0, m2, m0, m2, m4, m5
+ ABSW2 m1, m3, m1, m3, m4, m5
HADAMARD 0, max, 0, 2, 4, 5
HADAMARD 0, max, 1, 3, 4, 5
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
pmaddwd m0, m7
pmaddwd m1, m7
paddd m6, m0
paddw m7, m0
paddw m7, m1
%endif ; HIGH_BIT_DEPTH
- SAVE_MM_PERMUTATION hadamard_ac_2x2max_mmxext
+ SAVE_MM_PERMUTATION
ret
%macro AC_PREP 2
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
pmaddwd %1, %2
%endif
%endmacro
%macro AC_PADD 3
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
AC_PREP %2, %3
paddd %1, %2
%else
%endif ; HIGH_BIT_DEPTH
%endmacro
-cglobal hadamard_ac_8x8_mmxext
+cglobal hadamard_ac_8x8
mova m6, [mask_ac4]
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
mova m7, [pw_1]
%else
pxor m7, m7
%endif ; HIGH_BIT_DEPTH
- call hadamard_ac_4x4_mmxext
+ call hadamard_ac_4x4_mmx2
add r0, 4*SIZEOF_PIXEL
add r3, 32
mova m5, m0
AC_PREP m5, m7
- call hadamard_ac_4x4_mmxext
+ call hadamard_ac_4x4_mmx2
lea r0, [r0+4*r1]
add r3, 64
AC_PADD m5, m0, m7
- call hadamard_ac_4x4_mmxext
+ call hadamard_ac_4x4_mmx2
sub r0, 4*SIZEOF_PIXEL
sub r3, 32
AC_PADD m5, m0, m7
- call hadamard_ac_4x4_mmxext
+ call hadamard_ac_4x4_mmx2
AC_PADD m5, m0, m7
sub r3, 40
mova [rsp+gprsize+8], m5 ; save satd
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
pxor m6, m6
%endif
%rep 3
- call hadamard_ac_2x2max_mmxext
+ call hadamard_ac_2x2max_mmx2
%endrep
mova m0, [r3+0x00]
mova m1, [r3+0x20]
mova m2, [r3+0x40]
mova m3, [r3+0x60]
- SUMSUB_BADC w, m0, m1, m2, m3, m4
+ SUMSUB_BADC w, 0, 1, 2, 3, 4
HADAMARD 0, sumsub, 0, 2, 4, 5
- ABS4 m1, m3, m0, m2, m4, m5
+ ABSW2 m1, m3, m1, m3, m4, m5
+ ABSW2 m0, m2, m0, m2, m4, m5
HADAMARD 0, max, 1, 3, 4, 5
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
pand m0, [mask_ac4]
pmaddwd m1, m7
pmaddwd m0, m7
paddd m0, m2
paddd m6, m6
paddd m0, m6
- SWAP m0, m6
+ SWAP 0, 6
%else ; !HIGH_BIT_DEPTH
pand m6, m0
paddw m7, m1
paddw m6, m7
%endif ; HIGH_BIT_DEPTH
mova [rsp+gprsize], m6 ; save sa8d
- SWAP m0, m6
- SAVE_MM_PERMUTATION hadamard_ac_8x8_mmxext
+ SWAP 0, 6
+ SAVE_MM_PERMUTATION
ret
-%macro HADAMARD_AC_WXH_SUM_MMXEXT 2
+%macro HADAMARD_AC_WXH_SUM_MMX 2
mova m1, [rsp+1*mmsize]
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
%if %1*%2 >= 128
paddd m0, [rsp+2*mmsize]
paddd m1, [rsp+3*mmsize]
%endmacro
%macro HADAMARD_AC_WXH_MMX 2
-cglobal pixel_hadamard_ac_%1x%2_mmxext, 2,4
+cglobal pixel_hadamard_ac_%1x%2, 2,4
%assign pad 16-gprsize-(stack_offset&15)
%define ysub r1
FIX_STRIDES r1
sub rsp, 16+128+pad
lea r2, [r1*3]
lea r3, [rsp+16]
- call hadamard_ac_8x8_mmxext
+ call hadamard_ac_8x8_mmx2
%if %2==16
%define ysub r2
lea r0, [r0+r1*4]
sub rsp, 16
- call hadamard_ac_8x8_mmxext
+ call hadamard_ac_8x8_mmx2
%endif
%if %1==16
neg ysub
sub rsp, 16
lea r0, [r0+ysub*4+8*SIZEOF_PIXEL]
neg ysub
- call hadamard_ac_8x8_mmxext
+ call hadamard_ac_8x8_mmx2
%if %2==16
lea r0, [r0+r1*4]
sub rsp, 16
- call hadamard_ac_8x8_mmxext
+ call hadamard_ac_8x8_mmx2
%endif
%endif
- HADAMARD_AC_WXH_SUM_MMXEXT %1, %2
+ HADAMARD_AC_WXH_SUM_MMX %1, %2
movd edx, m0
movd eax, m1
shr edx, 1
-%ifdef ARCH_X86_64
+%if ARCH_X86_64
shl rdx, 32
add rax, rdx
%endif
HADAMARD_AC_WXH_MMX 8, 8
%macro LOAD_INC_8x4W_SSE2 5
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
movu m%1, [r0]
movu m%2, [r0+r1]
movu m%3, [r0+r1*2]
HSUMSUB %1, %2, %3, %4, %5
%endmacro
-%macro HADAMARD_AC_SSE2 1
-INIT_XMM
+%macro HADAMARD_AC_SSE2 0
; in: r0=pix, r1=stride, r2=stride*3
; out: [esp+16]=sa8d, [esp+32]=satd, r0+=stride*4
-cglobal hadamard_ac_8x8_%1
-%ifdef ARCH_X86_64
+cglobal hadamard_ac_8x8
+%if ARCH_X86_64
%define spill0 m8
%define spill1 m9
%define spill2 m10
%define spill1 [rsp+gprsize+16]
%define spill2 [rsp+gprsize+32]
%endif
-%ifdef HIGH_BIT_DEPTH
- %define vertical 1
-%elifidn %1, sse2
+%if HIGH_BIT_DEPTH
%define vertical 1
- ;LOAD_INC only unpacks to words
- pxor m7, m7
-%else
+%elif cpuflag(ssse3)
%define vertical 0
;LOAD_INC loads sumsubs
mova m7, [hmul_8p]
+%else
+ %define vertical 1
+ ;LOAD_INC only unpacks to words
+ pxor m7, m7
%endif
LOAD_INC_8x4W 0, 1, 2, 3, 7
%if vertical
HADAMARD4_2D_SSE 0, 1, 2, 3, 4
%else
- HADAMARD4_V m0, m1, m2, m3, m4
+ HADAMARD4_V 0, 1, 2, 3, 4
%endif
mova spill0, m1
SWAP 1, 7
%if vertical
HADAMARD4_2D_SSE 4, 5, 6, 7, 1
%else
- HADAMARD4_V m4, m5, m6, m7, m1
+ HADAMARD4_V 4, 5, 6, 7, 1
+ ; FIXME SWAP
mova m1, spill0
mova spill0, m6
mova spill1, m7
%endif
mova spill1, m2
mova spill2, m3
- ABS_MOV m1, m0
- ABS_MOV m2, m4
- ABS_MOV m3, m5
+ ABSW m1, m0, m0
+ ABSW m2, m4, m4
+ ABSW m3, m5, m5
paddw m1, m2
- SUMSUB_BA w, m0, m4
+ SUMSUB_BA w, 0, 4
%if vertical
pand m1, [mask_ac4]
%else
pand m1, [mask_ac4b]
%endif
AC_PREP m1, [pw_1]
- ABS_MOV m2, spill0
+ ABSW m2, spill0
AC_PADD m1, m3, [pw_1]
- ABS_MOV m3, spill1
+ ABSW m3, spill1
AC_PADD m1, m2, [pw_1]
- ABS_MOV m2, spill2
+ ABSW m2, spill2
AC_PADD m1, m3, [pw_1]
- ABS_MOV m3, m6
+ ABSW m3, m6, m6
AC_PADD m1, m2, [pw_1]
- ABS_MOV m2, m7
+ ABSW m2, m7, m7
AC_PADD m1, m3, [pw_1]
mova m3, m7
AC_PADD m1, m2, [pw_1]
AC_PREP m2, [pw_1]
AC_PADD m2, m3, [pw_1]
AC_PADD m2, m1, [pw_1]
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
paddd m2, m2
%else
paddw m2, m2
%endif ; HIGH_BIT_DEPTH
- ABS1 m4, m7
+ ABSW m4, m4, m7
pand m0, [mask_ac8]
- ABS1 m0, m7
+ ABSW m0, m0, m7
AC_PADD m2, m4, [pw_1]
AC_PADD m2, m0, [pw_1]
mova [rsp+gprsize+16], m2 ; save sa8d
- SWAP m0, m2
- SAVE_MM_PERMUTATION hadamard_ac_8x8_%1
+ SWAP 0, 2
+ SAVE_MM_PERMUTATION
ret
-HADAMARD_AC_WXH_SSE2 16, 16, %1
-HADAMARD_AC_WXH_SSE2 8, 16, %1
-HADAMARD_AC_WXH_SSE2 16, 8, %1
-HADAMARD_AC_WXH_SSE2 8, 8, %1
+HADAMARD_AC_WXH_SSE2 16, 16
+HADAMARD_AC_WXH_SSE2 8, 16
+HADAMARD_AC_WXH_SSE2 16, 8
+HADAMARD_AC_WXH_SSE2 8, 8
%endmacro ; HADAMARD_AC_SSE2
%macro HADAMARD_AC_WXH_SUM_SSE2 2
mova m1, [rsp+2*mmsize]
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
%if %1*%2 >= 128
paddd m0, [rsp+3*mmsize]
paddd m1, [rsp+4*mmsize]
%endmacro
; struct { int satd, int sa8d; } pixel_hadamard_ac_16x16( uint8_t *pix, int stride )
-%macro HADAMARD_AC_WXH_SSE2 3
-cglobal pixel_hadamard_ac_%1x%2_%3, 2,3,11
+%macro HADAMARD_AC_WXH_SSE2 2
+cglobal pixel_hadamard_ac_%1x%2, 2,3,11
%assign pad 16-gprsize-(stack_offset&15)
%define ysub r1
FIX_STRIDES r1
sub rsp, 48+pad
lea r2, [r1*3]
- call hadamard_ac_8x8_%3
+ call hadamard_ac_8x8
%if %2==16
%define ysub r2
lea r0, [r0+r1*4]
sub rsp, 32
- call hadamard_ac_8x8_%3
+ call hadamard_ac_8x8
%endif
%if %1==16
neg ysub
sub rsp, 32
lea r0, [r0+ysub*4+8*SIZEOF_PIXEL]
neg ysub
- call hadamard_ac_8x8_%3
+ call hadamard_ac_8x8
%if %2==16
lea r0, [r0+r1*4]
sub rsp, 32
- call hadamard_ac_8x8_%3
+ call hadamard_ac_8x8
%endif
%endif
HADAMARD_AC_WXH_SUM_SSE2 %1, %2
movd eax, m1
shr edx, 2 - (%1*%2 >> 8)
shr eax, 1
-%ifdef ARCH_X86_64
+%if ARCH_X86_64
shl rdx, 32
add rax, rdx
%endif
; instantiate satds
-%ifndef ARCH_X86_64
-cextern pixel_sa8d_8x8_internal_mmxext
-SA8D mmxext
+%if ARCH_X86_64 == 0
+cextern pixel_sa8d_8x8_internal_mmx2
+INIT_MMX mmx2
+SA8D
%endif
%define TRANS TRANS_SSE2
-%define ABS1 ABS1_MMX
-%define ABS2 ABS2_MMX
%define DIFFOP DIFF_UNPACK_SSE2
-%define JDUP JDUP_SSE2
%define LOAD_INC_8x4W LOAD_INC_8x4W_SSE2
%define LOAD_SUMSUB_8x4P LOAD_DIFF_8x4P
%define LOAD_SUMSUB_16P LOAD_SUMSUB_16P_SSE2
%define movdqa movaps ; doesn't hurt pre-nehalem, might as well save size
%define movdqu movups
%define punpcklqdq movlhps
-INIT_XMM
-SA8D sse2
-SATDS_SSE2 sse2
-INTRA_SA8D_SSE2 sse2
-%ifndef HIGH_BIT_DEPTH
-INTRA_SATDS_MMX mmxext
-%endif
-HADAMARD_AC_SSE2 sse2
-
-%define ABS1 ABS1_SSSE3
-%define ABS2 ABS2_SSSE3
-%define ABS_MOV ABS_MOV_SSSE3
+INIT_XMM sse2
+SA8D
+SATDS_SSE2
+%if HIGH_BIT_DEPTH == 0
+INTRA_SA8D_SSE2
+%endif
+INIT_MMX mmx2
+INTRA_X3_MMX
+INIT_XMM sse2
+HADAMARD_AC_SSE2
+
%define DIFFOP DIFF_SUMSUB_SSSE3
-%define JDUP JDUP_CONROE
%define LOAD_DUP_4x8P LOAD_DUP_4x8P_CONROE
-%ifndef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH == 0
%define LOAD_INC_8x4W LOAD_INC_8x4W_SSSE3
%define LOAD_SUMSUB_8x4P LOAD_SUMSUB_8x4P_SSSE3
%define LOAD_SUMSUB_16P LOAD_SUMSUB_16P_SSSE3
%endif
-SATDS_SSE2 ssse3
-SA8D ssse3
-HADAMARD_AC_SSE2 ssse3
+INIT_XMM ssse3
+SATDS_SSE2
+SA8D
+HADAMARD_AC_SSE2
+%if HIGH_BIT_DEPTH == 0
+INTRA_X9
+INTRA8_X9
+%endif
%undef movdqa ; nehalem doesn't like movaps
%undef movdqu ; movups
%undef punpcklqdq ; or movlhps
-INTRA_SA8D_SSE2 ssse3
-INTRA_SATDS_MMX ssse3
+%if HIGH_BIT_DEPTH == 0
+INIT_MMX ssse3
+INTRA_X3_MMX
+%endif
%define TRANS TRANS_SSE4
-%define JDUP JDUP_PENRYN
%define LOAD_DUP_4x8P LOAD_DUP_4x8P_PENRYN
-SATDS_SSE2 sse4
-SA8D sse4
-HADAMARD_AC_SSE2 sse4
+INIT_XMM sse4
+SATDS_SSE2
+SA8D
+HADAMARD_AC_SSE2
+%if HIGH_BIT_DEPTH == 0
+INTRA_X9
+INTRA8_X9
+%endif
+
+INIT_XMM avx
+SATDS_SSE2
+SA8D
+%if HIGH_BIT_DEPTH == 0
+INTRA_X9
+INTRA8_X9
+%endif
+HADAMARD_AC_SSE2
+
+%define TRANS TRANS_XOP
+INIT_XMM xop
+SATDS_SSE2
+SA8D
+%if HIGH_BIT_DEPTH == 0
+INTRA_X9
+; no xop INTRA8_X9. it's slower than avx on bulldozer. dunno why.
+%endif
+HADAMARD_AC_SSE2
;=============================================================================
; SSIM
;=============================================================================
;-----------------------------------------------------------------------------
-; void pixel_ssim_4x4x2_core( const uint8_t *pix1, int stride1,
-; const uint8_t *pix2, int stride2, int sums[2][4] )
+; void pixel_ssim_4x4x2_core( const uint8_t *pix1, intptr_t stride1,
+; const uint8_t *pix2, intptr_t stride2, int sums[2][4] )
;-----------------------------------------------------------------------------
-
%macro SSIM_ITER 1
-%ifdef HIGH_BIT_DEPTH
+%if HIGH_BIT_DEPTH
movdqu m5, [r0+(%1&1)*r1]
movdqu m6, [r2+(%1&1)*r3]
%else
paddw m1, m5
paddw m2, m6
%endif
- movdqa m7, m5
+ pmaddwd m7, m5, m6
pmaddwd m5, m5
- pmaddwd m7, m6
pmaddwd m6, m6
-%if %1==0
- SWAP m3, m5
- SWAP m4, m7
-%else
- paddd m3, m5
- paddd m4, m7
-%endif
+ ACCUM paddd, 3, 5, %1
+ ACCUM paddd, 4, 7, %1
paddd m3, m6
%endmacro
-cglobal pixel_ssim_4x4x2_core_sse2, 4,4,8
+%macro SSIM 0
+cglobal pixel_ssim_4x4x2_core, 4,4,8
FIX_STRIDES r1, r3
pxor m0, m0
SSIM_ITER 0
; PHADDW m1, m2
; PHADDD m3, m4
movdqa m7, [pw_1]
- pshufd m5, m3, 0xb1
+ pshufd m5, m3, q2301
pmaddwd m1, m7
pmaddwd m2, m7
- pshufd m6, m4, 0xb1
+ pshufd m6, m4, q2301
packssdw m1, m2
paddd m3, m5
- pshufd m1, m1, 0xd8
+ pshufd m1, m1, q3120
paddd m4, m6
pmaddwd m1, m7
- movdqa m5, m3
+ punpckhdq m5, m3, m4
punpckldq m3, m4
- punpckhdq m5, m4
-%ifdef UNIX64
+%if UNIX64
%define t0 r4
%else
%define t0 rax
;-----------------------------------------------------------------------------
; float pixel_ssim_end( int sum0[5][4], int sum1[5][4], int width )
;-----------------------------------------------------------------------------
-cglobal pixel_ssim_end4_sse2, 3,3,7
+cglobal pixel_ssim_end4, 3,3,7
movdqa m0, [r0+ 0]
movdqa m1, [r0+16]
movdqa m2, [r0+32]
addps m2, m6 ; vars + ssim_c2
addps m3, m6 ; covar*2 + ssim_c2
%else
- movdqa m4, m1
+ pmaddwd m4, m1, m0 ; s1*s2
pslld m1, 16
- pmaddwd m4, m0 ; s1*s2
por m0, m1
pmaddwd m0, m0 ; s1*s1 + s2*s2
pslld m4, 1
.skip:
movhlps m0, m4
addps m0, m4
- pshuflw m4, m0, 0xE
+ pshuflw m4, m0, q0032
addss m0, m4
-%ifndef ARCH_X86_64
+%if ARCH_X86_64 == 0
movd r0m, m0
fld dword r0m
%endif
RET
+%endmacro ; SSIM
+
+INIT_XMM sse2
+SSIM
+INIT_XMM avx
+SSIM
+;-----------------------------------------------------------------------------
+; int pixel_asd8( pixel *pix1, intptr_t stride1, pixel *pix2, intptr_t stride2, int height );
+;-----------------------------------------------------------------------------
+%macro ASD8 0
+cglobal pixel_asd8, 5,5
+ pxor m0, m0
+ pxor m1, m1
+.loop:
+%if HIGH_BIT_DEPTH
+ paddw m0, [r0]
+ paddw m1, [r2]
+ paddw m0, [r0+2*r1]
+ paddw m1, [r2+2*r3]
+ lea r0, [r0+4*r1]
+ paddw m0, [r0]
+ paddw m1, [r2+4*r3]
+ lea r2, [r2+4*r3]
+ paddw m0, [r0+2*r1]
+ paddw m1, [r2+2*r3]
+ lea r0, [r0+4*r1]
+ lea r2, [r2+4*r3]
+%else
+ movq m2, [r0]
+ movq m3, [r2]
+ movhps m2, [r0+r1]
+ movhps m3, [r2+r3]
+ lea r0, [r0+2*r1]
+ psadbw m2, m1
+ psadbw m3, m1
+ movq m4, [r0]
+ movq m5, [r2+2*r3]
+ lea r2, [r2+2*r3]
+ movhps m4, [r0+r1]
+ movhps m5, [r2+r3]
+ lea r0, [r0+2*r1]
+ paddw m0, m2
+ psubw m0, m3
+ psadbw m4, m1
+ psadbw m5, m1
+ lea r2, [r2+2*r3]
+ paddw m0, m4
+ psubw m0, m5
+%endif
+ sub r4d, 4
+ jg .loop
+%if HIGH_BIT_DEPTH
+ psubw m0, m1
+ HADDW m0, m1
+ ABSD m1, m0
+%else
+ movhlps m1, m0
+ paddw m0, m1
+ ABSW m1, m0
+%endif
+ movd eax, m1
+ RET
+%endmacro
+INIT_XMM sse2
+ASD8
+INIT_XMM ssse3
+ASD8
+%if HIGH_BIT_DEPTH
+INIT_XMM xop
+ASD8
+%endif
;=============================================================================
; Successive Elimination ADS
;=============================================================================
%macro ADS_START 0
-%ifdef WIN64
+%if UNIX64
movsxd r5, r5d
+%else
+ mov r5d, r5m
%endif
mov r0d, r5d
lea r6, [r4+r5+15]
jmp ads_mvs
%endmacro
-%define ABS1 ABS1_MMX
-
;-----------------------------------------------------------------------------
; int pixel_ads4( int enc_dc[4], uint16_t *sums, int delta,
; uint16_t *cost_mvx, int16_t *mvs, int width, int thresh )
;-----------------------------------------------------------------------------
-cglobal pixel_ads4_mmxext, 6,7
+INIT_MMX mmx2
+cglobal pixel_ads4, 5,7
movq mm6, [r0]
movq mm4, [r0+8]
pshufw mm7, mm6, 0
- pshufw mm6, mm6, 0xAA
+ pshufw mm6, mm6, q2222
pshufw mm5, mm4, 0
- pshufw mm4, mm4, 0xAA
+ pshufw mm4, mm4, q2222
ADS_START
.loop:
movq mm0, [r1]
movq mm1, [r1+16]
psubw mm0, mm7
psubw mm1, mm6
- ABS1 mm0, mm2
- ABS1 mm1, mm3
+ ABSW mm0, mm0, mm2
+ ABSW mm1, mm1, mm3
movq mm2, [r1+r2]
movq mm3, [r1+r2+16]
psubw mm2, mm5
psubw mm3, mm4
paddw mm0, mm1
- ABS1 mm2, mm1
- ABS1 mm3, mm1
+ ABSW mm2, mm2, mm1
+ ABSW mm3, mm3, mm1
paddw mm0, mm2
paddw mm0, mm3
pshufw mm1, r6m, 0
movd [r6], mm1
ADS_END 1
-cglobal pixel_ads2_mmxext, 6,7
+cglobal pixel_ads2, 5,7
movq mm6, [r0]
pshufw mm5, r6m, 0
pshufw mm7, mm6, 0
- pshufw mm6, mm6, 0xAA
+ pshufw mm6, mm6, q2222
ADS_START
.loop:
movq mm0, [r1]
movq mm1, [r1+r2]
psubw mm0, mm7
psubw mm1, mm6
- ABS1 mm0, mm2
- ABS1 mm1, mm3
+ ABSW mm0, mm0, mm2
+ ABSW mm1, mm1, mm3
paddw mm0, mm1
paddusw mm0, [r3]
movq mm4, mm5
movd [r6], mm4
ADS_END 1
-cglobal pixel_ads1_mmxext, 6,7
+cglobal pixel_ads1, 5,7
pshufw mm7, [r0], 0
pshufw mm6, r6m, 0
ADS_START
movq mm1, [r1+8]
psubw mm0, mm7
psubw mm1, mm7
- ABS1 mm0, mm2
- ABS1 mm1, mm3
+ ABSW mm0, mm0, mm2
+ ABSW mm1, mm1, mm3
paddusw mm0, [r3]
paddusw mm1, [r3+8]
movq mm4, mm6
movq [r6], mm4
ADS_END 2
-%macro ADS_SSE2 1
-cglobal pixel_ads4_%1, 6,7,12
+%macro ADS_XMM 0
+cglobal pixel_ads4, 5,7,12
movdqa xmm4, [r0]
pshuflw xmm7, xmm4, 0
- pshuflw xmm6, xmm4, 0xAA
+ pshuflw xmm6, xmm4, q2222
pshufhw xmm5, xmm4, 0
- pshufhw xmm4, xmm4, 0xAA
+ pshufhw xmm4, xmm4, q2222
punpcklqdq xmm7, xmm7
punpcklqdq xmm6, xmm6
punpckhqdq xmm5, xmm5
punpckhqdq xmm4, xmm4
-%ifdef ARCH_X86_64
+%if ARCH_X86_64
pshuflw xmm8, r6m, 0
punpcklqdq xmm8, xmm8
ADS_START
movdqu xmm10, [r1]
movdqu xmm11, [r1+r2]
.loop:
- movdqa xmm0, xmm10
- movdqu xmm1, [r1+16]
- movdqa xmm10, xmm1
- psubw xmm0, xmm7
- psubw xmm1, xmm6
- ABS1 xmm0, xmm2
- ABS1 xmm1, xmm3
- movdqa xmm2, xmm11
- movdqu xmm3, [r1+r2+16]
- movdqa xmm11, xmm3
- psubw xmm2, xmm5
- psubw xmm3, xmm4
+ psubw xmm0, xmm10, xmm7
+ movdqu xmm10, [r1+16]
+ psubw xmm1, xmm10, xmm6
+ ABSW xmm0, xmm0, xmm2
+ ABSW xmm1, xmm1, xmm3
+ psubw xmm2, xmm11, xmm5
+ movdqu xmm11, [r1+r2+16]
paddw xmm0, xmm1
+ psubw xmm3, xmm11, xmm4
movdqu xmm9, [r3]
- ABS1 xmm2, xmm1
- ABS1 xmm3, xmm1
+ ABSW xmm2, xmm2, xmm1
+ ABSW xmm3, xmm3, xmm1
paddw xmm0, xmm2
paddw xmm0, xmm3
paddusw xmm0, xmm9
- movdqa xmm1, xmm8
- psubusw xmm1, xmm0
+ psubusw xmm1, xmm8, xmm0
packsswb xmm1, xmm1
movq [r6], xmm1
%else
movdqu xmm1, [r1+16]
psubw xmm0, xmm7
psubw xmm1, xmm6
- ABS1 xmm0, xmm2
- ABS1 xmm1, xmm3
+ ABSW xmm0, xmm0, xmm2
+ ABSW xmm1, xmm1, xmm3
movdqu xmm2, [r1+r2]
movdqu xmm3, [r1+r2+16]
psubw xmm2, xmm5
psubw xmm3, xmm4
paddw xmm0, xmm1
- ABS1 xmm2, xmm1
- ABS1 xmm3, xmm1
+ ABSW xmm2, xmm2, xmm1
+ ABSW xmm3, xmm3, xmm1
paddw xmm0, xmm2
paddw xmm0, xmm3
movd xmm1, r6m
%endif ; ARCH
ADS_END 2
-cglobal pixel_ads2_%1, 6,7,8
+cglobal pixel_ads2, 5,7,8
movq xmm6, [r0]
movd xmm5, r6m
pshuflw xmm7, xmm6, 0
- pshuflw xmm6, xmm6, 0xAA
+ pshuflw xmm6, xmm6, q2222
pshuflw xmm5, xmm5, 0
punpcklqdq xmm7, xmm7
punpcklqdq xmm6, xmm6
psubw xmm0, xmm7
psubw xmm1, xmm6
movdqu xmm4, [r3]
- ABS1 xmm0, xmm2
- ABS1 xmm1, xmm3
+ ABSW xmm0, xmm0, xmm2
+ ABSW xmm1, xmm1, xmm3
paddw xmm0, xmm1
paddusw xmm0, xmm4
- movdqa xmm1, xmm5
- psubusw xmm1, xmm0
+ psubusw xmm1, xmm5, xmm0
packsswb xmm1, xmm1
movq [r6], xmm1
ADS_END 2
-cglobal pixel_ads1_%1, 6,7,8
+cglobal pixel_ads1, 5,7,8
movd xmm7, [r0]
movd xmm6, r6m
pshuflw xmm7, xmm7, 0
psubw xmm1, xmm7
movdqu xmm2, [r3]
movdqu xmm3, [r3+16]
- ABS1 xmm0, xmm4
- ABS1 xmm1, xmm5
+ ABSW xmm0, xmm0, xmm4
+ ABSW xmm1, xmm1, xmm5
paddusw xmm0, xmm2
paddusw xmm1, xmm3
- movdqa xmm4, xmm6
- movdqa xmm5, xmm6
- psubusw xmm4, xmm0
- psubusw xmm5, xmm1
+ psubusw xmm4, xmm6, xmm0
+ psubusw xmm5, xmm6, xmm1
packsswb xmm4, xmm5
movdqa [r6], xmm4
ADS_END 4
%endmacro
-ADS_SSE2 sse2
-%define ABS1 ABS1_SSSE3
-ADS_SSE2 ssse3
+INIT_XMM sse2
+ADS_XMM
+INIT_XMM ssse3
+ADS_XMM
+INIT_XMM avx
+ADS_XMM
; int pixel_ads_mvs( int16_t *mvs, uint8_t *masks, int width )
; {
inc r1d
%endmacro
+INIT_MMX
cglobal pixel_ads_mvs, 0,7,0
ads_mvs:
lea r6, [r4+r5+15]
jge .end
.loopi:
mov r2, [r6+r1]
-%ifdef ARCH_X86_64
+%if ARCH_X86_64
test r2, r2
%else
mov r3, r2
- or r3d, [r6+r1+4]
+ add r3d, [r6+r1+4]
%endif
jz .loopi0
xor r3d, r3d
TEST 1
TEST 2
TEST 3
-%ifdef ARCH_X86_64
+%if ARCH_X86_64
shr r2, 32
%else
mov r2d, [r6+r1]
projects / x264 / blobdiff