Rewrite vp8_short_walsh4x4_sse2()

This rewriting reflects changes made in commit "Improve the
accuracy of forward walsh-hadamard transform". Since this function
is not called much, only a small encoder performance gain (~0.5% )
is seen.

Change-Id: Ie9df58a43028a11fd5b115c4bbe3141f7596578b

diff --git a/vp8/encoder/x86/fwalsh_sse2.asm b/vp8/encoder/x86/fwalsh_sse2.asm
index 38812c8..ffc9b3d 100644 (file)
--- a/vp8/encoder/x86/fwalsh_sse2.asm
+++ b/vp8/encoder/x86/fwalsh_sse2.asm
@@ -21,94 +21,122 @@ sym(vp8_short_walsh4x4_sse2):
     push        rdi
     ; end prolog
 
-    mov     rsi, arg(0)
-    mov     rdi, arg(1)
-
-    movdqu    xmm4, [rsi + 0]       ;ip[4] ip[0]
-    movdqu    xmm0, [rsi + 16]      ;ip[12] ip[8]
-
-    pxor  xmm7, xmm7
-    ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-    ; 13 12 11 10 03 02 01 00
-    ;
-    ; 33 32 31 30 23 22 21 20
-    ;
-    movdqa    xmm3, xmm4          ; 13 12 11 10 03 02 01 00
-    punpcklwd xmm4, xmm0          ; 23 03 22 02 21 01 20 00
-    punpckhwd xmm3, xmm0          ; 33 13 32 12 31 11 30 10
-    movdqa    xmm1, xmm4          ; 23 03 22 02 21 01 20 00
-    punpcklwd xmm4, xmm3          ; 31 21 11 01 30 20 10 00
-    punpckhwd xmm1, xmm3          ; 33 23 13 03 32 22 12 02
-    ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-    pshufd    xmm2, xmm1, 4eh       ;ip[8] ip[12]
-    movdqa    xmm3, xmm4          ;ip[4] ip[0]
-
-    paddw   xmm4, xmm2          ;ip[4]+ip[8] ip[0]+ip[12] aka b1 a1
-    psubw   xmm3, xmm2          ;ip[4]-ip[8] ip[0]-ip[12] aka c1 d1
-
+    mov     rsi, arg(0)           ; input
+    mov     rdi, arg(1)           ; output
+    movsxd  rdx, dword ptr arg(2) ; pitch
+
+    ; first for loop
+    movq    xmm0, MMWORD PTR [rsi]           ; load input
+    movq    xmm1, MMWORD PTR [rsi + rdx]
+    lea     rsi,  [rsi + rdx*2]
+    movq    xmm2, MMWORD PTR [rsi]
+    movq    xmm3, MMWORD PTR [rsi + rdx]
+
+    punpcklwd xmm0,  xmm1
+    punpcklwd xmm2,  xmm3
+
+    movdqa    xmm1, xmm0
+    punpckldq xmm0, xmm2           ; ip[1] ip[0]
+    punpckhdq xmm1, xmm2           ; ip[3] ip[2]
+
+    movdqa    xmm2, xmm0
+    paddw     xmm0, xmm1
+    psubw     xmm2, xmm1
+
+    psllw     xmm0, 2              ; d1  a1
+    psllw     xmm2, 2              ; c1  b1
+
+    movdqa    xmm1, xmm0
+    punpcklqdq xmm0, xmm2          ; b1  a1
+    punpckhqdq xmm1, xmm2          ; c1  d1
+
+    pxor      xmm6, xmm6
+    movq      xmm6, xmm0
+    pxor      xmm7, xmm7
+    pcmpeqw   xmm7, xmm6
+    paddw     xmm7, [GLOBAL(c1)]
+
+    movdqa    xmm2, xmm0
+    paddw     xmm0, xmm1           ; b1+c1  a1+d1
+    psubw     xmm2, xmm1           ; b1-c1  a1-d1
+    paddw     xmm0, xmm7           ; b1+c1  a1+d1+(a1!=0)
+
+    ; second for loop
+    ; input: 13  9  5  1 12  8  4  0 (xmm0)
+    ;        14 10  6  2 15 11  7  3 (xmm2)
+    ; after shuffle:
+    ;        13  5  9  1 12  4  8  0 (xmm0)
+    ;        14  6 10  2 15  7 11  3 (xmm1)
+    pshuflw   xmm3, xmm0, 0xd8
+    pshufhw   xmm0, xmm3, 0xd8
+    pshuflw   xmm3, xmm2, 0xd8
+    pshufhw   xmm1, xmm3, 0xd8
+
+    movdqa    xmm2, xmm0
+    pmaddwd   xmm0, [GLOBAL(c1)]    ; d11 a11 d10 a10
+    pmaddwd   xmm2, [GLOBAL(cn1)]   ; c11 b11 c10 b10
+    movdqa    xmm3, xmm1
+    pmaddwd   xmm1, [GLOBAL(c1)]    ; d12 a12 d13 a13
+    pmaddwd   xmm3, [GLOBAL(cn1)]   ; c12 b12 c13 b13
+
+    pshufd    xmm4, xmm0, 0xd8      ; d11 d10 a11 a10
+    pshufd    xmm5, xmm2, 0xd8      ; c11 c10 b11 b10
+    pshufd    xmm6, xmm1, 0x72      ; d13 d12 a13 a12
+    pshufd    xmm7, xmm3, 0x72      ; c13 c12 b13 b12
+
+    movdqa    xmm0, xmm4
+    punpcklqdq xmm0, xmm5           ; b11 b10 a11 a10
+    punpckhqdq xmm4, xmm5           ; c11 c10 d11 d10
+    movdqa    xmm1, xmm6
+    punpcklqdq xmm1, xmm7           ; b13 b12 a13 a12
+    punpckhqdq xmm6, xmm7           ; c13 c12 d13 d12
+
+    movdqa    xmm2, xmm0
+    paddd     xmm0, xmm4            ; b21 b20 a21 a20
+    psubd     xmm2, xmm4            ; c21 c20 d21 d20
+    movdqa    xmm3, xmm1
+    paddd     xmm1, xmm6            ; b23 b22 a23 a22
+    psubd     xmm3, xmm6            ; c23 c22 d23 d22
+
+    pxor      xmm4, xmm4
     movdqa    xmm5, xmm4
-    punpcklqdq  xmm4, xmm3          ;d1 a1
-    punpckhqdq  xmm5, xmm3          ;c1 b1
-
-    movdqa    xmm1, xmm5          ;c1 b1
-    paddw   xmm5, xmm4          ;dl+cl a1+b1 aka op[4] op[0]
-    psubw   xmm4, xmm1          ;d1-c1 a1-b1 aka op[12] op[8]
-    ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-    ; 13 12 11 10 03 02 01 00
-    ;
-    ; 33 32 31 30 23 22 21 20
-    ;
-    movdqa    xmm0, xmm5          ; 13 12 11 10 03 02 01 00
-    punpcklwd xmm5, xmm4          ; 23 03 22 02 21 01 20 00
-    punpckhwd xmm0, xmm4          ; 33 13 32 12 31 11 30 10
-    movdqa    xmm1, xmm5          ; 23 03 22 02 21 01 20 00
-    punpcklwd xmm5, xmm0          ; 31 21 11 01 30 20 10 00
-    punpckhwd xmm1, xmm0          ; 33 23 13 03 32 22 12 02
-    ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-    pshufd    xmm2, xmm1, 4eh       ;ip[8] ip[12]
-    movdqa    xmm3, xmm5          ;ip[4] ip[0]
-
-    paddw   xmm5, xmm2          ;ip[4]+ip[8] ip[0]+ip[12] aka b1 a1
-    psubw   xmm3, xmm2          ;ip[4]-ip[8] ip[0]-ip[12] aka c1 d1
-
-    movdqa    xmm6, xmm5
-    punpcklqdq  xmm5, xmm3          ;d1 a1
-    punpckhqdq  xmm6, xmm3          ;c1 b1
-
-    movdqa    xmm1, xmm6          ;c1 b1
-    paddw   xmm6, xmm5          ;dl+cl a1+b1 aka op[4] op[0]
-    psubw   xmm5, xmm1          ;d1-c1 a1-b1 aka op[12] op[8]
-
-    movdqa    xmm0, xmm6          ;aka b2 a2
-    movdqa    xmm1, xmm5          ;aka d2 c2
-
-    pcmpgtw   xmm0, xmm7
-    pcmpgtw   xmm1, xmm7
-
-    psrlw   xmm0, 15
-    psrlw   xmm1, 15
-
-    paddw   xmm6, xmm0
-    paddw   xmm5, xmm1
-
-    psraw   xmm6, 1
-    psraw   xmm5, 1
-
-    ;   a2 = a1 + b1;
-    ;   b2 = c1 + d1;
-    ;   c2 = a1 - b1;
-    ;   d2 = d1 - c1;
-    ;        a2 += (a2>0);
-    ;        b2 += (b2>0);
-    ;        c2 += (c2>0);
-    ;        d2 += (d2>0);
-    ;   op[0] = (a2)>>1;
-    ;   op[4] = (b2)>>1;
-    ;   op[8] = (c2)>>1;
-    ;   op[12]= (d2)>>1;
-
-    movdqu  [rdi + 0], xmm6
-    movdqu  [rdi + 16], xmm5
+    pcmpgtd   xmm4, xmm0
+    pcmpgtd   xmm5, xmm2
+    pand      xmm4, [GLOBAL(cd1)]
+    pand      xmm5, [GLOBAL(cd1)]
+
+    pxor      xmm6, xmm6
+    movdqa    xmm7, xmm6
+    pcmpgtd   xmm6, xmm1
+    pcmpgtd   xmm7, xmm3
+    pand      xmm6, [GLOBAL(cd1)]
+    pand      xmm7, [GLOBAL(cd1)]
+
+    paddd     xmm0, xmm4
+    paddd     xmm2, xmm5
+    paddd     xmm0, [GLOBAL(cd3)]
+    paddd     xmm2, [GLOBAL(cd3)]
+    paddd     xmm1, xmm6
+    paddd     xmm3, xmm7
+    paddd     xmm1, [GLOBAL(cd3)]
+    paddd     xmm3, [GLOBAL(cd3)]
+
+    psrad     xmm0, 3
+    psrad     xmm1, 3
+    psrad     xmm2, 3
+    psrad     xmm3, 3
+    movdqa    xmm4, xmm0
+    punpcklqdq xmm0, xmm1           ; a23 a22 a21 a20
+    punpckhqdq xmm4, xmm1           ; b23 b22 b21 b20
+    movdqa    xmm5, xmm2
+    punpckhqdq xmm2, xmm3           ; c23 c22 c21 c20
+    punpcklqdq xmm5, xmm3           ; d23 d22 d21 d20
+
+    packssdw  xmm0, xmm4            ; b23 b22 b21 b20 a23 a22 a21 a20
+    packssdw  xmm2, xmm5            ; d23 d22 d21 d20 c23 c22 c21 c20
+
+    movdqa  XMMWORD PTR [rdi], xmm0
+    movdqa  XMMWORD PTR [rdi + 16], xmm2
 
     ; begin epilog
     pop rdi
@@ -116,3 +144,17 @@ sym(vp8_short_walsh4x4_sse2):
     UNSHADOW_ARGS
     pop         rbp
     ret
+
+SECTION_RODATA
+align 16
+c1:
+    dw 0x0001, 0x0001, 0x0001, 0x0001, 0x0001, 0x0001, 0x0001, 0x0001
+align 16
+cn1:
+    dw 0x0001, 0xffff, 0x0001, 0xffff, 0x0001, 0xffff, 0x0001, 0xffff
+align 16
+cd1:
+    dd 0x00000001, 0x00000001, 0x00000001, 0x00000001
+align 16
+cd3:
+    dd 0x00000003, 0x00000003, 0x00000003, 0x00000003

diff --git a/vp8/encoder/x86/x86_csystemdependent.c b/vp8/encoder/x86/x86_csystemdependent.c
index 9b753bf..b4035ff 100644 (file)
--- a/vp8/encoder/x86/x86_csystemdependent.c
+++ b/vp8/encoder/x86/x86_csystemdependent.c
@@ -289,7 +289,7 @@ void vp8_arch_x86_encoder_init(VP8_COMP *cpi)
         cpi->rtcd.fdct.fast4x4                   = vp8_short_fdct4x4_sse2;
         cpi->rtcd.fdct.fast8x4                   = vp8_short_fdct8x4_sse2;
 
-        cpi->rtcd.fdct.walsh_short4x4            = vp8_short_walsh4x4_c ;
+        cpi->rtcd.fdct.walsh_short4x4            = vp8_short_walsh4x4_sse2 ;
 
         cpi->rtcd.encodemb.berr                  = vp8_block_error_xmm;
         cpi->rtcd.encodemb.mberr                 = vp8_mbblock_error_xmm;