Simple H/V loopfilter for VP8 in MMX, MMX2 and SSE2 (yay for yasm macros).

Originally committed as revision 24029 to svn://svn.ffmpeg.org/ffmpeg/trunk

4 files changed, 334 insertions, 0 deletions

diff --git a/libavcodec/x86/dsputil_mmx.c b/libavcodec/x86/dsputil_mmx.c
index 4e1efb1887..174e893803 100644
--- a/libavcodec/x86/dsputil_mmx.c
+++ b/libavcodec/x86/dsputil_mmx.c
@@ -63,12 +63,16 @@ DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255) = 0x00ff00ff00ff00ffULL;
 
 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1  ) = 0x0101010101010101ULL;
 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_3  ) = {0x0303030303030303ULL, 0x0303030303030303ULL};
+DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_4  ) = {0x0404040404040404ULL, 0x0404040404040404ULL};
 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7  ) = 0x0707070707070707ULL;
 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F ) = 0x1F1F1F1F1F1F1F1FULL;
 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F ) = 0x3F3F3F3F3F3F3F3FULL;
+DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_80 ) = {0x8080808080808080ULL, 0x8080808080808080ULL};
 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81 ) = 0x8181818181818181ULL;
 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_A1 ) = 0xA1A1A1A1A1A1A1A1ULL;
+DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_F8 ) = {0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL};
 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC ) = 0xFCFCFCFCFCFCFCFCULL;
+DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_FE ) = {0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL};
 
 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
diff --git a/libavcodec/x86/vp8dsp-init.c b/libavcodec/x86/vp8dsp-init.c
index 698d394278..9cd72b6955 100644
--- a/libavcodec/x86/vp8dsp-init.c
+++ b/libavcodec/x86/vp8dsp-init.c
@@ -222,6 +222,13 @@ extern void ff_vp8_idct_dc_add_mmx(uint8_t *dst, DCTELEM block[16], int stride);
 extern void ff_vp8_idct_dc_add_sse4(uint8_t *dst, DCTELEM block[16], int stride);
 extern void ff_vp8_luma_dc_wht_mmxext(DCTELEM block[4][4][16], DCTELEM dc[16]);
 extern void ff_vp8_idct_add_mmx(uint8_t *dst, DCTELEM block[16], int stride);
+
+extern void ff_vp8_v_loop_filter_simple_mmx   (uint8_t *dst, int stride, int flim);
+extern void ff_vp8_v_loop_filter_simple_mmxext(uint8_t *dst, int stride, int flim);
+extern void ff_vp8_v_loop_filter_simple_sse2  (uint8_t *dst, int stride, int flim);
+extern void ff_vp8_h_loop_filter_simple_mmx   (uint8_t *dst, int stride, int flim);
+extern void ff_vp8_h_loop_filter_simple_mmxext(uint8_t *dst, int stride, int flim);
+extern void ff_vp8_h_loop_filter_simple_sse2  (uint8_t *dst, int stride, int flim);
 #endif
 
 #define VP8_LUMA_MC_FUNC(IDX, SIZE, OPT) \
@@ -260,6 +267,9 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
         c->put_vp8_bilinear_pixels_tab[0][0][0] = ff_put_vp8_pixels16_mmx;
         c->put_vp8_epel_pixels_tab[1][0][0]     =
         c->put_vp8_bilinear_pixels_tab[1][0][0] = ff_put_vp8_pixels8_mmx;
+
+        c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_mmx;
+        c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_mmx;
     }
 
     /* note that 4-tap width=16 functions are missing because w=16
@@ -272,6 +282,9 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
         VP8_BILINEAR_MC_FUNC(0, 16, mmxext);
         VP8_BILINEAR_MC_FUNC(1, 8, mmxext);
         VP8_BILINEAR_MC_FUNC(2, 4, mmxext);
+
+        c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_mmxext;
+        c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_mmxext;
     }
 
     if (mm_flags & FF_MM_SSE) {
@@ -284,6 +297,9 @@ av_cold void ff_vp8dsp_init_x86(VP8DSPContext* c)
         VP8_MC_FUNC(1, 8, sse2);
         VP8_BILINEAR_MC_FUNC(0, 16, sse2);
         VP8_BILINEAR_MC_FUNC(1, 8, sse2);
+
+        c->vp8_v_loop_filter_simple = ff_vp8_v_loop_filter_simple_sse2;
+        c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_sse2;
     }
 
     if (mm_flags & FF_MM_SSSE3) {
diff --git a/libavcodec/x86/vp8dsp.asm b/libavcodec/x86/vp8dsp.asm
index 2c3eee4009..aedd09e5ac 100644
--- a/libavcodec/x86/vp8dsp.asm
+++ b/libavcodec/x86/vp8dsp.asm
@@ -146,8 +146,13 @@ pw_20091: times 4 dw 20091
 pw_17734: times 4 dw 17734
 
 cextern pw_3
+cextern pb_3
 cextern pw_4
+cextern pb_4
 cextern pw_64
+cextern pb_80
+cextern pb_F8
+cextern pb_FE
 
 SECTION .text
 
@@ -1063,3 +1068,304 @@ cglobal vp8_luma_dc_wht_mmxext, 2,3
     add           r0, 2*16*4
     SCATTER_WHT   3
     RET
+
+;-----------------------------------------------------------------------------
+; void vp8_h/v_loop_filter_simple_<opt>(uint8_t *dst, int stride, int flim);
+;-----------------------------------------------------------------------------
+
+; macro called with 7 mm register indexes as argument, and 4 regular registers
+;
+; first 4 mm registers will carry the transposed pixel data
+; the other three are scratchspace (one would be sufficient, but this allows
+; for more spreading/pipelining and thus faster execution on OOE CPUs)
+;
+; first two regular registers are buf+4*stride and buf+5*stride
+; third is -stride, fourth is +stride
+%macro READ_8x4_INTERLEAVED 11
+    ; interleave 8 (A-H) rows of 4 pixels each
+    movd          m%1, [%8+%10*4]   ; A0-3
+    movd          m%5, [%9+%10*4]   ; B0-3
+    movd          m%2, [%8+%10*2]   ; C0-3
+    movd          m%6, [%8+%10]     ; D0-3
+    movd          m%3, [%8]         ; E0-3
+    movd          m%7, [%9]         ; F0-3
+    movd          m%4, [%9+%11]     ; G0-3
+    punpcklbw     m%1, m%5          ; A/B interleaved
+    movd          m%5, [%9+%11*2]   ; H0-3
+    punpcklbw     m%2, m%6          ; C/D interleaved
+    punpcklbw     m%3, m%7          ; E/F interleaved
+    punpcklbw     m%4, m%5          ; G/H interleaved
+%endmacro
+
+; macro called with 7 mm register indexes as argument, and 5 regular registers
+; first 11 mean the same as READ_8x4_TRANSPOSED above
+; fifth regular register is scratchspace to reach the bottom 8 rows, it
+; will be set to second regular register + 8*stride at the end
+%macro READ_16x4_INTERLEAVED 12
+    ; transpose 16 (A-P) rows of 4 pixels each
+    lea           %12, [r0+8*r2]
+
+    ; read (and interleave) those addressable by %8 (=r0), A/C/D/E/I/K/L/M
+    movd          m%1, [%8+%10*4]   ; A0-3
+    movd          m%3, [%12+%10*4]  ; I0-3
+    movd          m%2, [%8+%10*2]   ; C0-3
+    movd          m%4, [%12+%10*2]  ; K0-3
+    movd          m%6, [%8+%10]     ; D0-3
+    movd          m%5, [%12+%10]    ; L0-3
+    movd          m%7, [%12]        ; M0-3
+    add           %12, %11
+    punpcklbw     m%1, m%3          ; A/I
+    movd          m%3, [%8]         ; E0-3
+    punpcklbw     m%2, m%4          ; C/K
+    punpcklbw     m%6, m%5          ; D/L
+    punpcklbw     m%3, m%7          ; E/M
+    punpcklbw     m%2, m%6          ; C/D/K/L interleaved
+
+    ; read (and interleave) those addressable by %9 (=r4), B/F/G/H/J/N/O/P
+    movd         m%5, [%9+%10*4]   ; B0-3
+    movd         m%4, [%12+%10*4]  ; J0-3
+    movd         m%7, [%9]         ; F0-3
+    movd         m%6, [%12]        ; N0-3
+    punpcklbw    m%5, m%4          ; B/J
+    punpcklbw    m%7, m%6          ; F/N
+    punpcklbw    m%1, m%5          ; A/B/I/J interleaved
+    punpcklbw    m%3, m%7          ; E/F/M/N interleaved
+    movd         m%4, [%9+%11]     ; G0-3
+    movd         m%6, [%12+%11]    ; O0-3
+    movd         m%5, [%9+%11*2]   ; H0-3
+    movd         m%7, [%12+%11*2]  ; P0-3
+    punpcklbw    m%4, m%6          ; G/O
+    punpcklbw    m%5, m%7          ; H/P
+    punpcklbw    m%4, m%5          ; G/H/O/P interleaved
+%endmacro
+
+; write 4 mm registers of 2 dwords each
+; first four arguments are mm register indexes containing source data
+; last four are registers containing buf+4*stride, buf+5*stride,
+; -stride and +stride
+%macro WRITE_4x2D 8
+    ; write out (2 dwords per register)
+    movd    [%5+%7*4], m%1
+    movd    [%5+%7*2], m%2
+    movd         [%5], m%3
+    movd      [%6+%8], m%4
+    punpckhdq     m%1, m%1
+    punpckhdq     m%2, m%2
+    punpckhdq     m%3, m%3
+    punpckhdq     m%4, m%4
+    movd    [%6+%7*4], m%1
+    movd      [%5+%7], m%2
+    movd         [%6], m%3
+    movd    [%6+%8*2], m%4
+%endmacro
+
+; write 4 xmm registers of 4 dwords each
+; arguments same as WRITE_2x4D, but with an extra register, so that the 5 regular
+; registers contain buf+4*stride, buf+5*stride, buf+12*stride, -stride and +stride
+; we add 1*stride to the third regular registry in the process
+%macro WRITE_4x4D 9
+    ; write out (4 dwords per register), start with dwords zero
+    movd    [%5+%8*4], m%1
+    movd         [%5], m%2
+    movd    [%5+%9*4], m%3
+    movd    [%5+%9*8], m%4
+
+    ; store dwords 1
+    psrldq        m%1, 4
+    psrldq        m%2, 4
+    psrldq        m%3, 4
+    psrldq        m%4, 4
+    movd    [%6+%8*4], m%1
+    movd         [%6], m%2
+    movd    [%6+%9*4], m%3
+    movd    [%6+%9*8], m%4
+
+    ; write dwords 2
+    psrldq        m%1, 4
+    psrldq        m%2, 4
+    psrldq        m%3, 4
+    psrldq        m%4, 4
+    movd    [%5+%8*2], m%1
+    movd      [%6+%9], m%2
+    movd    [%7+%8*2], m%3
+    movd    [%7+%9*2], m%4
+    add            %7, %9
+
+    ; store dwords 3
+    psrldq        m%1, 4
+    psrldq        m%2, 4
+    psrldq        m%3, 4
+    psrldq        m%4, 4
+    movd      [%5+%8], m%1
+    movd    [%6+%9*2], m%2
+    movd    [%7+%8*2], m%3
+    movd    [%7+%9*2], m%4
+%endmacro
+
+%macro SIMPLE_LOOPFILTER 3
+cglobal vp8_%2_loop_filter_simple_%1, 3, %3
+%ifidn %2, h
+    mov            r5, rsp          ; backup stack pointer
+    and           rsp, ~(mmsize-1)  ; align stack
+%endif
+%if mmsize == 8 ; mmx/mmxext
+    mov            r3, 2
+%endif
+
+    ; splat register with "flim"
+    movd           m7, r2
+    punpcklbw      m7, m7
+%if mmsize == 16 ; sse2
+    punpcklwd      m7, m7
+    pshufd         m7, m7, 0x0
+%elifidn %1, mmx
+    punpcklwd      m7, m7
+    punpckldq      m7, m7
+%else ; mmxext
+    pshufw         m7, m7, 0x0
+%endif
+
+    ; set up indexes to address 4 rows
+    mov            r2, r1
+    neg            r1
+%ifidn %2, h
+    lea            r0, [r0+4*r2-2]
+    sub           rsp, mmsize*2     ; (aligned) storage space for saving p1/q1
+%endif
+
+%if mmsize == 8 ; mmx / mmxext
+.next8px
+%endif
+%ifidn %2, v
+    ; read 4 half/full rows of pixels
+    mova           m0, [r0+r1*2]    ; p1
+    mova           m1, [r0+r1]      ; p0
+    mova           m2, [r0]         ; q0
+    mova           m3, [r0+r2]      ; q1
+%else ; h
+    lea            r4, [r0+r2]
+
+%if mmsize == 8 ; mmx/mmxext
+    READ_8x4_INTERLEAVED  0, 1, 2, 3, 4, 5, 6, r0, r4, r1, r2
+%else ; sse2
+    READ_16x4_INTERLEAVED 0, 1, 2, 3, 4, 5, 6, r0, r4, r1, r2, r3
+%endif
+    TRANSPOSE4x4W         0, 1, 2, 3, 4
+
+    mova        [rsp], m0           ; store p1
+    mova [rsp+mmsize], m3           ; store q1
+%endif
+
+    ; simple_limit
+    mova           m5, m2           ; m5=backup of q0
+    mova           m6, m1           ; m6=backup of p0
+    psubusb        m1, m2           ; p0-q0
+    psubusb        m2, m6           ; q0-p0
+    por            m1, m2           ; FFABS(p0-q0)
+    paddusb        m1, m1           ; m1=FFABS(p0-q0)*2
+
+    mova           m4, m3
+    mova           m2, m0
+    psubusb        m3, m0           ; q1-p1
+    psubusb        m0, m4           ; p1-q1
+    por            m3, m0           ; FFABS(p1-q1)
+    mova           m0, [pb_80]
+    pxor           m2, m0
+    pxor           m4, m0
+    psubsb         m2, m4           ; m2=p1-q1 (signed) backup for below
+    pand           m3, [pb_FE]
+    psrlq          m3, 1            ; m3=FFABS(p1-q1)/2, this can be used signed
+    paddusb        m3, m1
+    psubusb        m3, m7
+    pxor           m1, m1
+    pcmpeqb        m3, m1           ; abs(p0-q0)*2+abs(p1-q1)/2<=flim mask(0xff/0x0)
+
+    ; filter_common (use m2/p1-q1, m4=q0, m6=p0, m5/q0-p0 and m3/mask)
+    mova           m4, m5
+    pxor           m5, m0
+    pxor           m0, m6
+    psubsb         m5, m0           ; q0-p0 (signed)
+    paddsb         m2, m5
+    paddsb         m2, m5
+    paddsb         m2, m5           ; a=(p1-q1) + 3*(q0-p0)
+    pand           m2, m3           ; apply filter mask (m3)
+
+    mova           m3, [pb_F8]
+    mova           m1, m2
+    paddsb         m2, [pb_4]       ; f1<<3=a+4
+    paddsb         m1, [pb_3]       ; f2<<3=a+3
+    pand           m2, m3
+    pand           m1, m3           ; cache f2<<3
+
+    pxor           m0, m0
+    pxor           m3, m3
+    pcmpgtb        m0, m2           ; which values are <0?
+    psubb          m3, m2           ; -f1<<3
+    psrlq          m2, 3            ; +f1
+    psrlq          m3, 3            ; -f1
+    pand           m3, m0
+    pandn          m0, m2
+    psubusb        m4, m0
+    paddusb        m4, m3           ; q0-f1
+
+    pxor           m0, m0
+    pxor           m3, m3
+    pcmpgtb        m0, m1           ; which values are <0?
+    psubb          m3, m1           ; -f2<<3
+    psrlq          m1, 3            ; +f2
+    psrlq          m3, 3            ; -f2
+    pand           m3, m0
+    pandn          m0, m1
+    paddusb        m6, m0
+    psubusb        m6, m3           ; p0+f2
+
+    ; store
+%ifidn %2, v
+    mova         [r0], m4
+    mova      [r0+r1], m6
+%else ; h
+    mova           m0, [rsp]        ; p1
+    SWAP            2, 4            ; p0
+    SWAP            1, 6            ; q0
+    mova           m3, [rsp+mmsize] ; q1
+
+    TRANSPOSE4x4B  0, 1, 2, 3, 4
+%if mmsize == 16 ; sse2
+    add            r3, r1           ; change from r4*8*stride to r0+8*stride
+    WRITE_4x4D 0, 1, 2, 3, r0, r4, r3, r1, r2
+%else ; mmx/mmxext
+    WRITE_4x2D 0, 1, 2, 3, r0, r4, r1, r2
+%endif
+%endif
+
+%if mmsize == 8 ; mmx/mmxext
+    ; next 8 pixels
+%ifidn %2, v
+    add            r0, 8            ; advance 8 cols = pixels
+%else ; h
+    lea            r0, [r0+r2*8]    ; advance 8 rows = lines
+%endif
+    dec            r3
+    jg .next8px
+%ifidn %2, v
+    REP_RET
+%else ; h
+    mov           rsp, r5           ; restore stack pointer
+    RET
+%endif
+%else ; sse2
+%ifidn %2, h
+    mov           rsp, r5           ; restore stack pointer
+%endif
+    RET
+%endif
+%endmacro
+
+INIT_MMX
+SIMPLE_LOOPFILTER mmx,    v, 4
+SIMPLE_LOOPFILTER mmx,    h, 6
+SIMPLE_LOOPFILTER mmxext, v, 4
+SIMPLE_LOOPFILTER mmxext, h, 6
+INIT_XMM
+SIMPLE_LOOPFILTER sse2,   v, 3
+SIMPLE_LOOPFILTER sse2,   h, 6
diff --git a/libavcodec/x86/x86util.asm b/libavcodec/x86/x86util.asm
index 3051a7f147..f99a277d74 100644
--- a/libavcodec/x86/x86util.asm
+++ b/libavcodec/x86/x86util.asm
@@ -37,6 +37,14 @@
     SWAP %2, %4, %3
 %endmacro
 
+%macro TRANSPOSE4x4B 5
+    SBUTTERFLY bw, %1, %2, %5
+    SBUTTERFLY bw, %3, %4, %5
+    SBUTTERFLY wd, %1, %3, %5
+    SBUTTERFLY wd, %2, %4, %5
+    SWAP %2, %3
+%endmacro
+
 %macro TRANSPOSE4x4W 5
     SBUTTERFLY wd, %1, %2, %5
     SBUTTERFLY wd, %3, %4, %5