shithub: dav1d

Download patch

ref: a91a03b0e143f77bd647b481aff034eb3315ee7b
parent: 713aa34c421726719a2e9676ba2d0d5af99512bf
author: Victorien Le Couviour--Tuffet <[email protected]>
date: Wed Sep 18 05:25:57 EDT 2019

x86: add warp_affine SSE4 and SSSE3 asm

------------------------------------------
x86_64: warp_8x8_8bpc_c: 1773.4
x86_32: warp_8x8_8bpc_c: 1740.4
----------
x86_64: warp_8x8_8bpc_ssse3: 317.5
x86_32: warp_8x8_8bpc_ssse3: 378.4
----------
x86_64: warp_8x8_8bpc_sse4: 303.7
x86_32: warp_8x8_8bpc_sse4: 367.7
----------
x86_64: warp_8x8_8bpc_avx2: 224.9
---------------------
---------------------
x86_64: warp_8x8t_8bpc_c: 1664.6
x86_32: warp_8x8t_8bpc_c: 1674.0
----------
x86_64: warp_8x8t_8bpc_ssse3: 320.7
x86_32: warp_8x8t_8bpc_ssse3: 379.5
----------
x86_64: warp_8x8t_8bpc_sse4: 304.8
x86_32: warp_8x8t_8bpc_sse4: 369.8
----------
x86_64: warp_8x8t_8bpc_avx2: 228.5
------------------------------------------

--- a/src/x86/mc_init_tmpl.c
+++ b/src/x86/mc_init_tmpl.c
@@ -88,7 +88,11 @@
 decl_blend_dir_fn(dav1d_blend_h_ssse3);
 
 decl_warp8x8_fn(dav1d_warp_affine_8x8_avx2);
+decl_warp8x8_fn(dav1d_warp_affine_8x8_sse4);
+decl_warp8x8_fn(dav1d_warp_affine_8x8_ssse3);
 decl_warp8x8t_fn(dav1d_warp_affine_8x8t_avx2);
+decl_warp8x8t_fn(dav1d_warp_affine_8x8t_sse4);
+decl_warp8x8t_fn(dav1d_warp_affine_8x8t_ssse3);
 
 decl_emu_edge_fn(dav1d_emu_edge_avx2);
 decl_emu_edge_fn(dav1d_emu_edge_ssse3);
@@ -134,7 +138,19 @@
     c->blend = dav1d_blend_ssse3;
     c->blend_v = dav1d_blend_v_ssse3;
     c->blend_h = dav1d_blend_h_ssse3;
+
+    c->warp8x8  = dav1d_warp_affine_8x8_ssse3;
+    c->warp8x8t = dav1d_warp_affine_8x8t_ssse3;
+
     c->emu_edge = dav1d_emu_edge_ssse3;
+#endif
+
+    if(!(flags & DAV1D_X86_CPU_FLAG_SSE41))
+        return;
+
+#if BITDEPTH == 8
+    c->warp8x8  = dav1d_warp_affine_8x8_sse4;
+    c->warp8x8t = dav1d_warp_affine_8x8t_sse4;
 #endif
 
     if (!(flags & DAV1D_X86_CPU_FLAG_AVX2))
--- a/src/x86/mc_ssse3.asm
+++ b/src/x86/mc_ssse3.asm
@@ -44,6 +44,10 @@
             db 45, 19, 47, 17, 48, 16, 50, 14, 51, 13, 52, 12, 53, 11, 55,  9
             db 56,  8, 57,  7, 58,  6, 59,  5, 60,  4, 60,  4, 61,  3, 62,  2
 
+warp_8x8_shufA: db 0,  2,  4,  6,  1,  3,  5,  7,  1,  3,  5,  7,  2,  4,  6,  8
+warp_8x8_shufB: db 4,  6,  8, 10,  5,  7,  9, 11,  5,  7,  9, 11,  6,  8, 10, 12
+warp_8x8_shufC: db 2,  4,  6,  8,  3,  5,  7,  9,  3,  5,  7,  9,  4,  6,  8, 10
+warp_8x8_shufD: db 6,  8, 10, 12,  7,  9, 11, 13,  7,  9, 11, 13,  8, 10, 12, 14
 blend_shuf:     db 0,  1,  0,  1,  0,  1,  0,  1,  2,  3,  2,  3,  2,  3,  2,  3
 subpel_h_shuf4: db 0,  1,  2,  3,  1,  2,  3,  4,  8,  9, 10, 11,  9, 10, 11, 12
                 db 2,  3,  4,  5,  3,  4,  5,  6, 10, 11, 12, 13, 11, 12, 13, 14
@@ -53,17 +57,18 @@
 bilin_h_shuf4:  db 1,  0,  2,  1,  3,  2,  4,  3,  9,  8, 10,  9, 11, 10, 12, 11
 bilin_h_shuf8:  db 1,  0,  2,  1,  3,  2,  4,  3,  5,  4,  6,  5,  7,  6,  8,  7
 
-pb_64:   times 16 db 64
-pw_8:    times 8 dw 8
-pw_26:   times 8 dw 26
-pw_34:   times 8 dw 34
-pw_512:  times 8 dw 512
-pw_1024: times 8 dw 1024
-pw_2048: times 8 dw 2048
-pw_6903: times 8 dw 6903
-pw_8192: times 8 dw 8192
-pd_32:   times 4 dd 32
-pd_512:  times 4 dd 512
+pb_64:    times 16 db 64
+pw_8:     times 8 dw 8
+pw_26:    times 8 dw 26
+pw_34:    times 8 dw 34
+pw_512:   times 8 dw 512
+pw_1024:  times 8 dw 1024
+pw_2048:  times 8 dw 2048
+pw_6903:  times 8 dw 6903
+pw_8192:  times 8 dw 8192
+pd_32:    times 4 dd 32
+pd_512:   times 4 dd 512
+pd_32768: times 4 dd 32768
 
 pw_258:  times 2 dw 258
 
@@ -146,6 +151,8 @@
 
 %define table_offset(type, fn) type %+ fn %+ SUFFIX %+ _table - type %+ SUFFIX
 
+cextern mc_warp_filter
+
 SECTION .text
 
 INIT_XMM ssse3
@@ -3301,6 +3308,580 @@
     sub                 r5d, 1<<16
     jg .hv_w8_loop0
     RET
+
+%if ARCH_X86_32
+ %macro SAVE_ALPHA_BETA 0
+    mov              alpham, alphad
+    mov               betam, betad
+ %endmacro
+
+ %macro SAVE_DELTA_GAMMA 0
+    mov              deltam, deltad
+    mov              gammam, gammad
+ %endmacro
+
+ %macro LOAD_ALPHA_BETA_MX 0
+    mov                 mym, myd
+    mov              alphad, alpham
+    mov               betad, betam
+    mov                 mxd, mxm
+ %endmacro
+
+ %macro LOAD_DELTA_GAMMA_MY 0
+    mov                 mxm, mxd
+    mov              deltad, deltam
+    mov              gammad, gammam
+    mov                 myd, mym
+ %endmacro
+
+ %define PIC_reg r2
+ %define PIC_base_offset $$
+ %define PIC_sym(sym) (PIC_reg+(sym)-PIC_base_offset)
+%else
+ %define SAVE_ALPHA_BETA
+ %define SAVE_DELTA_GAMMA
+ %define PIC_sym(sym) sym
+%endif
+
+%if ARCH_X86_32
+ %if STACK_ALIGNMENT < required_stack_alignment
+  %assign copy_args 8*4
+ %else
+  %assign copy_args 0
+ %endif
+%endif
+
+%macro RELOC_ARGS 0
+ %if copy_args
+    mov                  r0, r0m
+    mov                  r1, r1m
+    mov                  r2, r2m
+    mov                  r3, r3m
+    mov                  r5, r5m
+    mov                dstm, r0
+    mov                 dsm, r1
+    mov                srcm, r2
+    mov                 ssm, r3
+    mov                 mxm, r5
+    mov                  r0, r6m
+    mov                 mym, r0
+ %endif
+%endmacro
+
+%macro BLENDHWDW 2 ; blend high words from dwords, src1, src2
+ %if cpuflag(sse4)
+    pblendw              %1, %2, 0xAA
+ %else
+    pand                 %2, m10
+    por                  %1, %2
+ %endif
+%endmacro
+
+%macro WARP_V 10 ; dst0, dst1, 0, 2, 4, 6, 1, 3, 5, 7
+    ; Can be done using gathers, but that's terribly slow on many CPU:s
+ %if ARCH_X86_32
+  %define m8  m4
+  %define m9  m5
+  %define m14 m6
+  %define m15 m7
+  %define m11 m7
+    pxor                m11, m11
+ %endif
+    lea               tmp1d, [myq+deltaq*4]
+    lea               tmp2d, [myq+deltaq*1]
+    shr                 myd, 10
+    shr               tmp1d, 10
+    movq                 m2, [filterq+myq  *8] ; a
+    movq                 m8, [filterq+tmp1q*8] ; e
+    lea               tmp1d, [tmp2q+deltaq*4]
+    lea                 myd, [tmp2q+deltaq*1]
+    shr               tmp2d, 10
+    shr               tmp1d, 10
+    movq                 m3, [filterq+tmp2q*8] ; b
+    movq                 m0, [filterq+tmp1q*8] ; f
+    punpcklwd            m2, m3
+    punpcklwd            m8, m0
+    lea               tmp1d, [myq+deltaq*4]
+    lea               tmp2d, [myq+deltaq*1]
+    shr                 myd, 10
+    shr               tmp1d, 10
+    movq                 m0, [filterq+myq  *8] ; c
+    movq                 m9, [filterq+tmp1q*8] ; g
+    lea               tmp1d, [tmp2q+deltaq*4]
+    lea                 myd, [tmp2q+gammaq]       ; my += gamma
+    shr               tmp2d, 10
+    shr               tmp1d, 10
+    movq                 m3, [filterq+tmp2q*8] ; d
+    movq                 m1, [filterq+tmp1q*8] ; h
+    punpcklwd            m0, m3
+    punpcklwd            m9, m1
+    punpckldq            m1, m2, m0
+    punpckhdq            m2, m0
+    punpcklbw            m0, m11, m1 ; a0 a2 b0 b2 c0 c2 d0 d2 << 8
+    punpckhbw            m3, m11, m1 ; a4 a6 b4 b6 c4 c6 d4 d6 << 8
+    punpcklbw            m1, m11, m2 ; a1 a3 b1 b3 c1 c3 d1 d3 << 8
+    punpckhbw           m14, m11, m2 ; a5 a7 b5 b7 c5 c7 d5 d7 << 8
+    pmaddwd              m0, %3
+    pmaddwd              m3, %5
+    pmaddwd              m1, %7
+    pmaddwd             m14, %9
+    paddd                m0, m3
+    paddd                m1, m14
+    paddd                m0, m1
+    mova                 %1, m0
+ %if ARCH_X86_64
+    SWAP                 m3, m14
+ %endif
+    punpckldq            m0, m8, m9
+    punpckhdq            m8, m9
+    punpcklbw            m1, m11, m0 ; e0 e2 f0 f2 g0 g2 h0 h2 << 8
+    punpckhbw           m14, m11, m0 ; e4 e6 f4 f6 g4 g6 h4 h6 << 8
+    punpcklbw            m2, m11, m8 ; e1 e3 f1 f3 g1 g3 h1 h3 << 8
+    punpckhbw           m15, m11, m8 ; e5 e7 f5 f7 g5 g7 h5 h7 << 8
+    pmaddwd              m1, %4
+    pmaddwd             m14, %6
+    pmaddwd              m2, %8
+    pmaddwd             m15, %10
+    paddd                m1, m14
+    paddd                m2, m15
+    paddd                m1, m2
+    mova                 %2, m1
+ %if ARCH_X86_64
+    SWAP                m14, m3
+ %endif
+%endmacro
+
+%if ARCH_X86_64
+ %define counterd r4d
+%else
+ %if copy_args == 0
+  %define counterd dword r4m
+ %else
+  %define counterd dword [esp+stack_size-4*7]
+ %endif
+%endif
+
+%macro WARP_AFFINE_8X8T 0
+%if ARCH_X86_64
+cglobal warp_affine_8x8t, 6, 14, 16, 0x90, tmp, ts
+%else
+cglobal warp_affine_8x8t, 0, 7, 16, -0x130-copy_args, tmp, ts
+ %if copy_args
+  %define tmpm [esp+stack_size-4*1]
+  %define tsm  [esp+stack_size-4*2]
+ %endif
+%endif
+    call mangle(private_prefix %+ _warp_affine_8x8_%+cpuname).main
+.loop:
+%if ARCH_X86_32
+ %define m12 m4
+ %define m13 m5
+ %define m14 m6
+ %define m15 m7
+    mova                m12, [esp+0xC0]
+    mova                m13, [esp+0xD0]
+    mova                m14, [esp+0xE0]
+    mova                m15, [esp+0xF0]
+%endif
+    psrad               m12, 13
+    psrad               m13, 13
+    psrad               m14, 13
+    psrad               m15, 13
+    packssdw            m12, m13
+    packssdw            m14, m15
+    mova                m13, [PIC_sym(pw_8192)]
+    pmulhrsw            m12, m13 ; (x + (1 << 6)) >> 7
+    pmulhrsw            m14, m13
+    mova       [tmpq+tsq*0], m12
+    mova       [tmpq+tsq*2], m14
+    dec            counterd
+    jz   mangle(private_prefix %+ _warp_affine_8x8_%+cpuname).end
+%if ARCH_X86_32
+    mov                tmpm, tmpd
+    mov                  r0, [esp+0x100]
+    mov                  r1, [esp+0x104]
+%endif
+    call mangle(private_prefix %+ _warp_affine_8x8_%+cpuname).main2
+    lea                tmpq, [tmpq+tsq*4]
+    jmp .loop
+%endmacro
+
+%macro WARP_AFFINE_8X8 0
+%if ARCH_X86_64
+cglobal warp_affine_8x8, 6, 14, 16, 0x90, \
+                         dst, ds, src, ss, abcd, mx, tmp2, alpha, beta, \
+                         filter, tmp1, delta, my, gamma
+%else
+cglobal warp_affine_8x8, 0, 7, 16, -0x130-copy_args, \
+                         dst, ds, src, ss, abcd, mx, tmp2, alpha, beta, \
+                         filter, tmp1, delta, my, gamma
+ %define alphaq     r0
+ %define alphad     r0
+ %define alpham     [esp+gprsize+0x100]
+ %define betaq      r1
+ %define betad      r1
+ %define betam      [esp+gprsize+0x104]
+ %define deltaq     r0
+ %define deltad     r0
+ %define deltam     [esp+gprsize+0x108]
+ %define gammaq     r1
+ %define gammad     r1
+ %define gammam     [esp+gprsize+0x10C]
+ %define filterq    r3
+ %define tmp1q      r4
+ %define tmp1d      r4
+ %define tmp1m      [esp+gprsize+0x110]
+ %define myq        r5
+ %define myd        r5
+ %define mym        r6m
+ %if copy_args
+  %define dstm [esp+stack_size-4*1]
+  %define dsm  [esp+stack_size-4*2]
+  %define srcm [esp+stack_size-4*3]
+  %define ssm  [esp+stack_size-4*4]
+  %define mxm  [esp+stack_size-4*5]
+  %define mym  [esp+stack_size-4*6]
+ %endif
+%endif
+    call .main
+    jmp .start
+.loop:
+%if ARCH_X86_32
+    mov                dstm, dstd
+    mov              alphad, [esp+0x100]
+    mov               betad, [esp+0x104]
+%endif
+    call .main2
+    lea                dstq, [dstq+dsq*2]
+.start:
+%if cpuflag(ssse3)
+ %if ARCH_X86_64
+    mova                m10, [PIC_sym(pw_8192)]
+ %else
+  %define m10 [PIC_sym(pw_8192)]
+ %endif
+%endif
+%if ARCH_X86_32
+ %define m12 m5
+ %define m13 m6
+    mova                m12, [esp+0xC0]
+    mova                m13, [esp+0xD0]
+%endif
+%if cpuflag(sse4)
+ %if ARCH_X86_32
+  %define m11 m4
+    pxor                m11, m11
+ %endif
+    psrad               m12, 18
+    psrad               m13, 18
+    packusdw            m12, m13
+    pavgw               m12, m11 ; (x + (1 << 10)) >> 11
+%else
+    psrad               m12, 17
+    psrad               m13, 17
+    packssdw            m12, m13
+    pmulhrsw            m12, m10 ; (x + (1 << 10)) >> 11
+%endif
+%if ARCH_X86_32
+ %define m14 m6
+ %define m15 m7
+    mova                m14, [esp+0xE0]
+    mova                m15, [esp+0xF0]
+%endif
+%if cpuflag(sse4)
+    psrad               m14, 18
+    psrad               m15, 18
+    packusdw            m14, m15
+    pavgw               m14, m11 ; (x + (1 << 10)) >> 11
+%else
+    psrad               m14, 17
+    psrad               m15, 17
+    packssdw            m14, m15
+    pmulhrsw            m14, m10 ; (x + (1 << 10)) >> 11
+%endif
+    packuswb            m12, m14
+    movq       [dstq+dsq*0], m12
+    movhps     [dstq+dsq*1], m12
+    dec            counterd
+    jg .loop
+.end:
+    RET
+ALIGN function_align
+.main:
+%assign stack_offset stack_offset+gprsize
+%if ARCH_X86_32
+ %assign stack_size stack_size+4
+ %if copy_args
+  %assign stack_offset stack_offset-4
+ %endif
+    RELOC_ARGS
+    LEA             PIC_reg, $$
+ %define PIC_mem [esp+gprsize+0x114]
+    mov               abcdd, abcdm
+ %if copy_args == 0
+    mov                 ssd, ssm
+    mov                 mxd, mxm
+ %endif
+    mov             PIC_mem, PIC_reg
+    mov                srcd, srcm
+%endif
+    movsx            deltad, word [abcdq+2*2]
+    movsx            gammad, word [abcdq+2*3]
+    lea               tmp1d, [deltaq*3]
+    sub              gammad, tmp1d    ; gamma -= delta*3
+    SAVE_DELTA_GAMMA
+%if ARCH_X86_32
+    mov               abcdd, abcdm
+%endif
+    movsx            alphad, word [abcdq+2*0]
+    movsx             betad, word [abcdq+2*1]
+    lea               tmp1q, [ssq*3+3]
+    add                 mxd, 512+(64<<10)
+    lea               tmp2d, [alphaq*3]
+    sub                srcq, tmp1q    ; src -= src_stride*3 + 3
+%if ARCH_X86_32
+    mov                srcm, srcd
+    mov             PIC_reg, PIC_mem
+%endif
+    sub               betad, tmp2d    ; beta -= alpha*3
+    lea             filterq, [PIC_sym(mc_warp_filter)]
+%if ARCH_X86_64
+    mov                 myd, r6m
+    pxor                m11, m11
+%endif
+    call .h
+    psrld                m2, m0, 16
+    psrld                m3, m1, 16
+%if ARCH_X86_32
+    mova [esp+gprsize+0x10], m3
+%endif
+    call .h
+    psrld                m4, m0, 16
+    psrld                m5, m1, 16
+%if ARCH_X86_32
+    mova [esp+gprsize+0x20], m4
+    mova [esp+gprsize+0x30], m5
+%endif
+    call .h
+%if ARCH_X86_64
+ %define blendmask [rsp+gprsize+0x80]
+%else
+    mova                 m3, [esp+gprsize+0x10]
+ %define blendmask [esp+gprsize+0x120]
+ %define m10 m7
+%endif
+    pcmpeqd             m10, m10
+    pslld               m10, 16
+    mova          blendmask, m10
+    BLENDHWDW            m2, m0 ; 0
+    BLENDHWDW            m3, m1 ; 2
+    mova [rsp+gprsize+0x00], m2
+    mova [rsp+gprsize+0x10], m3
+    call .h
+%if ARCH_X86_32
+    mova                 m4, [esp+gprsize+0x20]
+    mova                 m5, [esp+gprsize+0x30]
+%endif
+    mova                m10, blendmask
+    BLENDHWDW            m4, m0 ; 1
+    BLENDHWDW            m5, m1 ; 3
+    mova [rsp+gprsize+0x20], m4
+    mova [rsp+gprsize+0x30], m5
+    call .h
+%if ARCH_X86_32
+    mova                 m3, [esp+gprsize+0x10]
+ %define m10 m5
+%endif
+    psrld                m6, m2, 16
+    psrld                m7, m3, 16
+    mova                m10, blendmask
+    BLENDHWDW            m6, m0 ; 2
+    BLENDHWDW            m7, m1 ; 4
+    mova [rsp+gprsize+0x40], m6
+    mova [rsp+gprsize+0x50], m7
+    call .h
+%if ARCH_X86_32
+    mova                m4, [esp+gprsize+0x20]
+    mova                m5, [esp+gprsize+0x30]
+%endif
+    psrld               m2, m4, 16
+    psrld               m3, m5, 16
+    mova                m10, blendmask
+    BLENDHWDW           m2, m0 ; 3
+    BLENDHWDW           m3, m1 ; 5
+    mova [rsp+gprsize+0x60], m2
+    mova [rsp+gprsize+0x70], m3
+    call .h
+%if ARCH_X86_32
+    mova                 m6, [esp+gprsize+0x40]
+    mova                 m7, [esp+gprsize+0x50]
+ %define m10 m7
+%endif
+    psrld                m4, m6, 16
+    psrld                m5, m7, 16
+    mova                m10, blendmask
+    BLENDHWDW            m4, m0 ; 4
+    BLENDHWDW            m5, m1 ; 6
+%if ARCH_X86_64
+    add                 myd, 512+(64<<10)
+    mova                 m6, m2
+    mova                 m7, m3
+%else
+    mova [esp+gprsize+0x80], m4
+    mova [esp+gprsize+0x90], m5
+    add           dword mym, 512+(64<<10)
+%endif
+    mov            counterd, 4
+    SAVE_ALPHA_BETA
+.main2:
+    call .h
+%if ARCH_X86_32
+    mova                 m6, [esp+gprsize+0x60]
+    mova                 m7, [esp+gprsize+0x70]
+ %define m10 m5
+%endif
+    psrld                m6, 16
+    psrld                m7, 16
+    mova                m10, blendmask
+    BLENDHWDW            m6, m0 ; 5
+    BLENDHWDW            m7, m1 ; 7
+%if ARCH_X86_64
+    WARP_V              m12, m13, [rsp+gprsize+0x00], [rsp+gprsize+0x10], \
+                                  m4, m5, \
+                                  [rsp+gprsize+0x20], [rsp+gprsize+0x30], \
+                                  m6, m7
+%else
+    mova [esp+gprsize+0xA0], m6
+    mova [esp+gprsize+0xB0], m7
+    LOAD_DELTA_GAMMA_MY
+    WARP_V [esp+gprsize+0xC0], [esp+gprsize+0xD0], \
+           [esp+gprsize+0x00], [esp+gprsize+0x10], \
+           [esp+gprsize+0x80], [esp+gprsize+0x90], \
+           [esp+gprsize+0x20], [esp+gprsize+0x30], \
+           [esp+gprsize+0xA0], [esp+gprsize+0xB0]
+    LOAD_ALPHA_BETA_MX
+%endif
+    call .h
+    mova                 m2, [rsp+gprsize+0x40]
+    mova                 m3, [rsp+gprsize+0x50]
+%if ARCH_X86_32
+    mova                 m4, [rsp+gprsize+0x80]
+    mova                 m5, [rsp+gprsize+0x90]
+ %define m10 m7
+%endif
+    mova [rsp+gprsize+0x00], m2
+    mova [rsp+gprsize+0x10], m3
+    mova [rsp+gprsize+0x40], m4
+    mova [rsp+gprsize+0x50], m5
+    psrld                m4, 16
+    psrld                m5, 16
+    mova                m10, blendmask
+    BLENDHWDW            m4, m0 ; 6
+    BLENDHWDW            m5, m1 ; 8
+%if ARCH_X86_64
+    WARP_V              m14, m15, [rsp+gprsize+0x20], [rsp+gprsize+0x30], \
+                                  m6, m7, \
+                                  [rsp+gprsize+0x00], [rsp+gprsize+0x10], \
+                                  m4, m5
+%else
+    mova [esp+gprsize+0x80], m4
+    mova [esp+gprsize+0x90], m5
+    LOAD_DELTA_GAMMA_MY
+    WARP_V [esp+gprsize+0xE0], [esp+gprsize+0xF0], \
+           [esp+gprsize+0x20], [esp+gprsize+0x30], \
+           [esp+gprsize+0xA0], [esp+gprsize+0xB0], \
+           [esp+gprsize+0x00], [esp+gprsize+0x10], \
+           [esp+gprsize+0x80], [esp+gprsize+0x90]
+    mov                 mym, myd
+    mov                dstd, dstm
+    mov                 dsd, dsm
+    mov                 mxd, mxm
+%endif
+    mova                 m2, [rsp+gprsize+0x60]
+    mova                 m3, [rsp+gprsize+0x70]
+%if ARCH_X86_32
+    mova                 m6, [esp+gprsize+0xA0]
+    mova                 m7, [esp+gprsize+0xB0]
+%endif
+    mova [rsp+gprsize+0x20], m2
+    mova [rsp+gprsize+0x30], m3
+    mova [rsp+gprsize+0x60], m6
+    mova [rsp+gprsize+0x70], m7
+    ret
+ALIGN function_align
+.h:
+%if ARCH_X86_32
+ %define m8  m3
+ %define m9  m4
+ %define m10 m5
+ %define m14 m6
+ %define m15 m7
+%endif
+    lea               tmp1d, [mxq+alphaq*4]
+    lea               tmp2d, [mxq+alphaq*1]
+%if ARCH_X86_32
+ %assign stack_offset stack_offset+4
+ %assign stack_size stack_size+4
+ %define PIC_mem [esp+gprsize*2+0x114]
+    mov             PIC_mem, PIC_reg
+    mov                srcd, srcm
+%endif
+    movu                m10, [srcq]
+%if ARCH_X86_32
+    add                srcd, ssm
+    mov                srcm, srcd
+    mov             PIC_reg, PIC_mem
+%else
+    add                srcq, ssq
+%endif
+    shr                 mxd, 10
+    shr               tmp1d, 10
+    movq                 m1, [filterq+mxq  *8]  ; 0 X
+    movq                 m8, [filterq+tmp1q*8]  ; 4 X
+    lea               tmp1d, [tmp2q+alphaq*4]
+    lea                 mxd, [tmp2q+alphaq*1]
+    shr               tmp2d, 10
+    shr               tmp1d, 10
+    movhps               m1, [filterq+tmp2q*8]  ; 0 1
+    movhps               m8, [filterq+tmp1q*8]  ; 4 5
+    lea               tmp1d, [mxq+alphaq*4]
+    lea               tmp2d, [mxq+alphaq*1]
+    shr                 mxd, 10
+    shr               tmp1d, 10
+    movq                m14, [filterq+mxq  *8]  ; 2 X
+    movq                 m9, [filterq+tmp1q*8]  ; 6 X
+    lea               tmp1d, [tmp2q+alphaq*4]
+    lea                 mxd, [tmp2q+betaq]  ; mx += beta
+    shr               tmp2d, 10
+    shr               tmp1d, 10
+    movhps              m14, [filterq+tmp2q*8]  ; 2 3
+    movhps               m9, [filterq+tmp1q*8]  ; 6 7
+    pshufb               m0, m10, [PIC_sym(warp_8x8_shufA)]
+    pmaddubsw            m0, m1
+    pshufb               m1, m10, [PIC_sym(warp_8x8_shufB)]
+    pmaddubsw            m1, m8
+    pshufb              m15, m10, [PIC_sym(warp_8x8_shufC)]
+    pmaddubsw           m15, m14
+    pshufb              m10, m10, [PIC_sym(warp_8x8_shufD)]
+    pmaddubsw           m10, m9
+    mova                m14, [PIC_sym(pw_8192)]
+    mova                 m9, [PIC_sym(pd_32768)]
+    phaddw               m0, m15
+    phaddw               m1, m10
+    pmaddwd              m0, m14 ; 17-bit intermediate, upshifted by 13
+    pmaddwd              m1, m14
+    paddd                m0, m9  ; rounded 14-bit result in upper 16 bits of dword
+    paddd                m1, m9
+    ret
+%endmacro
+
+INIT_XMM sse4
+WARP_AFFINE_8X8
+WARP_AFFINE_8X8T
+
+INIT_XMM ssse3
+WARP_AFFINE_8X8
+WARP_AFFINE_8X8T
 
 %if WIN64
 DECLARE_REG_TMP 6, 4