Modify sub-pixel filters to eliminate unnecessary calculations

In sub-pixel calculation, xoffset and yoffset mostly take some
specific values. Modified sub-pixel filter functions according to
these possible values to improve performance.

Change-Id: I83083570af8b00ff65093467914fbb97a4e9ea21

2 files changed, 254 insertions, 82 deletions

diff --git a/vp8/encoder/x86/variance_impl_sse2.asm b/vp8/encoder/x86/variance_impl_sse2.asm
index cefa0a956..7178e7e31 100644
--- a/vp8/encoder/x86/variance_impl_sse2.asm
+++ b/vp8/encoder/x86/variance_impl_sse2.asm
@@ -493,8 +493,8 @@ sym(vp8_get8x8var_sse2):
 ;    unsigned char *src_ptr,
 ;    int src_pixels_per_line,
 ;    unsigned int Height,
-;    unsigned short *HFilter,
-;    unsigned short *VFilter,
+;    int  xoffset,
+;    int  yoffset,
 ;    int *sum,
 ;    unsigned int *sumsquared;;
 ;
@@ -504,68 +504,80 @@ sym(vp8_filter_block2d_bil_var_sse2):
     push        rbp
     mov         rbp, rsp
     SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM
     GET_GOT     rbx
     push rsi
     push rdi
-    sub         rsp, 16
+    push rbx
     ; end prolog
 
         pxor            xmm6,           xmm6                 ;
         pxor            xmm7,           xmm7                 ;
-        mov             rax,            arg(5) ;HFilter             ;
 
-        mov             rdx,            arg(6) ;VFilter             ;
-        mov             rsi,            arg(0) ;ref_ptr              ;
+        lea             rsi,            [GLOBAL(xmm_bi_rd)]  ; rounding
+        movdqa          xmm4,           XMMWORD PTR [rsi]
 
-        mov             rdi,            arg(2) ;src_ptr              ;
-        movsxd          rcx,            dword ptr arg(4) ;Height              ;
+        lea             rcx,            [GLOBAL(vp8_bilinear_filters_sse2)]
+        movsxd          rax,            dword ptr arg(5)     ; xoffset
+
+        cmp             rax,            0                    ; skip first_pass filter if xoffset=0
+        je              filter_block2d_bil_var_sse2_sp_only
+
+        shl             rax,            5                    ; point to filter coeff with xoffset
+        lea             rax,            [rax + rcx]          ; HFilter
+
+        movsxd          rdx,            dword ptr arg(6)     ; yoffset
+
+        cmp             rdx,            0                    ; skip second_pass filter if yoffset=0
+        je              filter_block2d_bil_var_sse2_fp_only
+
+        shl             rdx,            5
+        lea             rdx,            [rdx + rcx]          ; VFilter
+
+        mov             rsi,            arg(0)               ;ref_ptr
+        mov             rdi,            arg(2)               ;src_ptr
+        movsxd          rcx,            dword ptr arg(4)     ;Height
 
         pxor            xmm0,           xmm0                 ;
-        movq            xmm1,           QWORD PTR [rsi]               ;
+        movq            xmm1,           QWORD PTR [rsi]      ;
+        movq            xmm3,           QWORD PTR [rsi+1]    ;
 
-        movq            xmm3,           QWORD PTR [rsi+1]        ;
         punpcklbw       xmm1,           xmm0                 ;
-
-        pmullw          xmm1,           [rax]               ;
+        pmullw          xmm1,           [rax]                ;
         punpcklbw       xmm3,           xmm0
-            ;
         pmullw          xmm3,           [rax+16]             ;
-        paddw           xmm1,           xmm3                 ;
-
-        paddw           xmm1,           [GLOBAL(xmm_bi_rd)]  ;
-        psraw           xmm1,           xmm_filter_shift    ;
 
+        paddw           xmm1,           xmm3                 ;
+        paddw           xmm1,           xmm4                 ;
+        psraw           xmm1,           xmm_filter_shift     ;
         movdqa          xmm5,           xmm1
-%if ABI_IS_32BIT
-        add             rsi,            dword ptr arg(1) ;ref_pixels_per_line    ;
-%else
-        movsxd          r8,             dword ptr arg(1) ;ref_pixels_per_line    ;
-        add             rsi,            r8
+
+        movsxd          rbx,            dword ptr arg(1) ;ref_pixels_per_line
+        lea             rsi,            [rsi + rbx]
+%if ABI_IS_32BIT=0
+        movsxd          r9,             dword ptr arg(3) ;src_pixels_per_line
 %endif
-filter_block2d_bil_var_sse2_loop:
 
+filter_block2d_bil_var_sse2_loop:
         movq            xmm1,           QWORD PTR [rsi]               ;
         movq            xmm3,           QWORD PTR [rsi+1]             ;
 
         punpcklbw       xmm1,           xmm0                 ;
         pmullw          xmm1,           [rax]               ;
-
         punpcklbw       xmm3,           xmm0                 ;
         pmullw          xmm3,           [rax+16]             ;
 
         paddw           xmm1,           xmm3                 ;
-        paddw           xmm1,           [GLOBAL(xmm_bi_rd)]  ;
-
+        paddw           xmm1,           xmm4               ;
         psraw           xmm1,           xmm_filter_shift    ;
-        movdqa          xmm3,           xmm5                 ;
 
+        movdqa          xmm3,           xmm5                 ;
         movdqa          xmm5,           xmm1                 ;
-        pmullw          xmm3,           [rdx]               ;
 
+        pmullw          xmm3,           [rdx]               ;
         pmullw          xmm1,           [rdx+16]             ;
         paddw           xmm1,           xmm3                 ;
-
-        paddw           xmm1,           [GLOBAL(xmm_bi_rd)]  ;
+        paddw           xmm1,           xmm4                 ;
         psraw           xmm1,           xmm_filter_shift    ;
 
         movq            xmm3,           QWORD PTR [rdi]               ;
@@ -577,20 +589,103 @@ filter_block2d_bil_var_sse2_loop:
         pmaddwd         xmm1,           xmm1                 ;
         paddd           xmm7,           xmm1                 ;
 
+        lea             rsi,            [rsi + rbx]          ;ref_pixels_per_line
 %if ABI_IS_32BIT
-        add             rsi,            dword ptr arg(1) ;ref_pixels_per_line    ;
-        add             rdi,            dword ptr arg(3) ;src_pixels_per_line    ;
+        add             rdi,            dword ptr arg(3)     ;src_pixels_per_line
 %else
-        movsxd          r8,             dword ptr arg(1) ;ref_pixels_per_line    ;
-        movsxd          r9,             dword ptr arg(3) ;src_pixels_per_line    ;
-        add             rsi,            r8
-        add             rdi,            r9
+        lea             rdi,            [rdi + r9]
 %endif
 
         sub             rcx,            1                   ;
         jnz             filter_block2d_bil_var_sse2_loop       ;
 
+        jmp             filter_block2d_bil_variance
+
+filter_block2d_bil_var_sse2_sp_only:
+        movsxd          rdx,            dword ptr arg(6)     ; yoffset
+        shl             rdx,            5
+        lea             rdx,            [rdx + rcx]          ; VFilter
+
+        mov             rsi,            arg(0)               ;ref_ptr
+        mov             rdi,            arg(2)               ;src_ptr
+        movsxd          rcx,            dword ptr arg(4)     ;Height
+        movsxd          rax,            dword ptr arg(1)     ;ref_pixels_per_line
+
+        pxor            xmm0,           xmm0                 ;
+        movq            xmm1,           QWORD PTR [rsi]      ;
+        punpcklbw       xmm1,           xmm0                 ;
+
+        movsxd          rbx,            dword ptr arg(3)     ;src_pixels_per_line
+        lea             rsi,            [rsi + rax]
+
+filter_block2d_bil_sp_only_loop:
+        movq            xmm3,           QWORD PTR [rsi]             ;
+        punpcklbw       xmm3,           xmm0                 ;
+        movdqa          xmm5,           xmm3
+
+        pmullw          xmm1,           [rdx]               ;
+        pmullw          xmm3,           [rdx+16]             ;
+        paddw           xmm1,           xmm3                 ;
+        paddw           xmm1,           xmm4                 ;
+        psraw           xmm1,           xmm_filter_shift    ;
+
+        movq            xmm3,           QWORD PTR [rdi]               ;
+        punpcklbw       xmm3,           xmm0                 ;
+
+        psubw           xmm1,           xmm3                 ;
+        paddw           xmm6,           xmm1                 ;
+
+        pmaddwd         xmm1,           xmm1                 ;
+        paddd           xmm7,           xmm1                 ;
+
+        movdqa          xmm1,           xmm5                 ;
+        lea             rsi,            [rsi + rax]          ;ref_pixels_per_line
+        lea             rdi,            [rdi + rbx]          ;src_pixels_per_line
+
+        sub             rcx,            1                   ;
+        jnz             filter_block2d_bil_sp_only_loop       ;
+
+        jmp             filter_block2d_bil_variance
+
+filter_block2d_bil_var_sse2_fp_only:
+        mov             rsi,            arg(0)               ;ref_ptr
+        mov             rdi,            arg(2)               ;src_ptr
+        movsxd          rcx,            dword ptr arg(4)     ;Height
+        movsxd          rdx,            dword ptr arg(1)     ;ref_pixels_per_line
+
+        pxor            xmm0,           xmm0                 ;
+        movsxd          rbx,            dword ptr arg(3)     ;src_pixels_per_line
+
+filter_block2d_bil_fp_only_loop:
+        movq            xmm1,           QWORD PTR [rsi]       ;
+        movq            xmm3,           QWORD PTR [rsi+1]     ;
+
+        punpcklbw       xmm1,           xmm0                 ;
+        pmullw          xmm1,           [rax]               ;
+        punpcklbw       xmm3,           xmm0                 ;
+        pmullw          xmm3,           [rax+16]             ;
+
+        paddw           xmm1,           xmm3                 ;
+        paddw           xmm1,           xmm4  ;
+        psraw           xmm1,           xmm_filter_shift    ;
+
+        movq            xmm3,           QWORD PTR [rdi]     ;
+        punpcklbw       xmm3,           xmm0                 ;
+
+        psubw           xmm1,           xmm3                 ;
+        paddw           xmm6,           xmm1                 ;
+
+        pmaddwd         xmm1,           xmm1                 ;
+        paddd           xmm7,           xmm1                 ;
+        lea             rsi,            [rsi + rdx]
+        lea             rdi,            [rdi + rbx]          ;src_pixels_per_line
+
+        sub             rcx,            1                   ;
+        jnz             filter_block2d_bil_fp_only_loop       ;
+
+        jmp             filter_block2d_bil_variance
 
+filter_block2d_bil_variance:
         movdq2q         mm6,            xmm6                ;
         movdq2q         mm7,            xmm7                ;
 
@@ -627,12 +722,12 @@ filter_block2d_bil_var_sse2_loop:
         movd            [rsi],          mm2    ; xsum
         movd            [rdi],          mm4    ; xxsum
 
-
     ; begin epilog
-    add rsp, 16
+    pop rbx
     pop rdi
     pop rsi
     RESTORE_GOT
+    RESTORE_XMM
     UNSHADOW_ARGS
     pop         rbp
     ret
@@ -974,3 +1069,13 @@ SECTION_RODATA
 align 16
 xmm_bi_rd:
     times 8 dw 64
+align 16
+vp8_bilinear_filters_sse2:
+    dw 128, 128, 128, 128, 128, 128, 128, 128,  0,  0,  0,  0,  0,  0,  0,  0
+    dw 112, 112, 112, 112, 112, 112, 112, 112, 16, 16, 16, 16, 16, 16, 16, 16
+    dw 96, 96, 96, 96, 96, 96, 96, 96, 32, 32, 32, 32, 32, 32, 32, 32
+    dw 80, 80, 80, 80, 80, 80, 80, 80, 48, 48, 48, 48, 48, 48, 48, 48
+    dw 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64
+    dw 48, 48, 48, 48, 48, 48, 48, 48, 80, 80, 80, 80, 80, 80, 80, 80
+    dw 32, 32, 32, 32, 32, 32, 32, 32, 96, 96, 96, 96, 96, 96, 96, 96
+    dw 16, 16, 16, 16, 16, 16, 16, 16, 112, 112, 112, 112, 112, 112, 112, 112
diff --git a/vp8/encoder/x86/variance_sse2.c b/vp8/encoder/x86/variance_sse2.c
index 006e0a24a..6f79f0d23 100644
--- a/vp8/encoder/x86/variance_sse2.c
+++ b/vp8/encoder/x86/variance_sse2.c
@@ -76,8 +76,8 @@ void vp8_filter_block2d_bil_var_sse2
     const unsigned char *src_ptr,
     int src_pixels_per_line,
     unsigned int Height,
-    const short *HFilter,
-    const short *VFilter,
+    int  xoffset,
+    int  yoffset,
     int *sum,
     unsigned int *sumsquared
 );
@@ -222,21 +222,6 @@ unsigned int vp8_variance8x16_wmt
 
 }
 
-///////////////////////////////////////////////////////////////////////////
-// the mmx function that does the bilinear filtering and var calculation //
-// int one pass                                                          //
-///////////////////////////////////////////////////////////////////////////
-DECLARE_ALIGNED(16, const short, vp8_bilinear_filters_xmm[8][16]) =
-{
-    { 128, 128, 128, 128, 128, 128, 128, 128,  0,  0,  0,  0,  0,  0,  0,  0 },
-    { 112, 112, 112, 112, 112, 112, 112, 112, 16, 16, 16, 16, 16, 16, 16, 16 },
-    {  96, 96, 96, 96, 96, 96, 96, 96, 32, 32, 32, 32, 32, 32, 32, 32 },
-    {  80, 80, 80, 80, 80, 80, 80, 80, 48, 48, 48, 48, 48, 48, 48, 48 },
-    {  64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64 },
-    {  48, 48, 48, 48, 48, 48, 48, 48, 80, 80, 80, 80, 80, 80, 80, 80 },
-    {  32, 32, 32, 32, 32, 32, 32, 32, 96, 96, 96, 96, 96, 96, 96, 96 },
-    {  16, 16, 16, 16, 16, 16, 16, 16, 112, 112, 112, 112, 112, 112, 112, 112 }
-};
 unsigned int vp8_sub_pixel_variance4x4_wmt
 (
     const unsigned char  *src_ptr,
@@ -272,15 +257,38 @@ unsigned int vp8_sub_pixel_variance8x8_wmt
     unsigned int *sse
 )
 {
-
     int xsum;
     unsigned int xxsum;
-    vp8_filter_block2d_bil_var_sse2(
-        src_ptr, src_pixels_per_line,
-        dst_ptr, dst_pixels_per_line, 8,
-        vp8_bilinear_filters_xmm[xoffset], vp8_bilinear_filters_xmm[yoffset],
-        &xsum, &xxsum
-    );
+
+    if (xoffset == 4 && yoffset == 0)
+    {
+        vp8_half_horiz_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            &xsum, &xxsum);
+    }
+    else if (xoffset == 0 && yoffset == 4)
+    {
+        vp8_half_vert_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            &xsum, &xxsum);
+    }
+    else if (xoffset == 4 && yoffset == 4)
+    {
+        vp8_half_horiz_vert_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            &xsum, &xxsum);
+    }
+    else
+    {
+        vp8_filter_block2d_bil_var_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            xoffset, yoffset,
+            &xsum, &xxsum);
+    }
 
     *sse = xxsum;
     return (xxsum - ((xsum * xsum) >> 6));
@@ -344,7 +352,7 @@ unsigned int vp8_sub_pixel_variance16x16_wmt
         vp8_filter_block2d_bil_var_sse2(
             src_ptr, src_pixels_per_line,
             dst_ptr, dst_pixels_per_line, 16,
-            vp8_bilinear_filters_xmm[xoffset], vp8_bilinear_filters_xmm[yoffset],
+            xoffset, yoffset,
             &xsum0, &xxsum0
         );
 
@@ -352,7 +360,7 @@ unsigned int vp8_sub_pixel_variance16x16_wmt
         vp8_filter_block2d_bil_var_sse2(
             src_ptr + 8, src_pixels_per_line,
             dst_ptr + 8, dst_pixels_per_line, 16,
-            vp8_bilinear_filters_xmm[xoffset], vp8_bilinear_filters_xmm[yoffset],
+            xoffset, yoffset,
             &xsum1, &xxsum1
         );
     }
@@ -392,21 +400,56 @@ unsigned int vp8_sub_pixel_variance16x8_wmt
     int xsum0, xsum1;
     unsigned int xxsum0, xxsum1;
 
+    if (xoffset == 4 && yoffset == 0)
+    {
+        vp8_half_horiz_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            &xsum0, &xxsum0);
 
-    vp8_filter_block2d_bil_var_sse2(
-        src_ptr, src_pixels_per_line,
-        dst_ptr, dst_pixels_per_line, 8,
-        vp8_bilinear_filters_xmm[xoffset], vp8_bilinear_filters_xmm[yoffset],
-        &xsum0, &xxsum0
-    );
+        vp8_half_horiz_variance16x_h_sse2(
+            src_ptr + 8, src_pixels_per_line,
+            dst_ptr + 8, dst_pixels_per_line, 8,
+            &xsum1, &xxsum1);
+    }
+    else if (xoffset == 0 && yoffset == 4)
+    {
+        vp8_half_vert_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            &xsum0, &xxsum0);
+
+        vp8_half_vert_variance16x_h_sse2(
+            src_ptr + 8, src_pixels_per_line,
+            dst_ptr + 8, dst_pixels_per_line, 8,
+            &xsum1, &xxsum1);
+    }
+    else if (xoffset == 4 && yoffset == 4)
+    {
+        vp8_half_horiz_vert_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            &xsum0, &xxsum0);
 
+        vp8_half_horiz_vert_variance16x_h_sse2(
+            src_ptr + 8, src_pixels_per_line,
+            dst_ptr + 8, dst_pixels_per_line, 8,
+            &xsum1, &xxsum1);
+    }
+    else
+    {
+        vp8_filter_block2d_bil_var_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 8,
+            xoffset, yoffset,
+            &xsum0, &xxsum0);
 
-    vp8_filter_block2d_bil_var_sse2(
-        src_ptr + 8, src_pixels_per_line,
-        dst_ptr + 8, dst_pixels_per_line, 8,
-        vp8_bilinear_filters_xmm[xoffset], vp8_bilinear_filters_xmm[yoffset],
-        &xsum1, &xxsum1
-    );
+        vp8_filter_block2d_bil_var_sse2(
+            src_ptr + 8, src_pixels_per_line,
+            dst_ptr + 8, dst_pixels_per_line, 8,
+            xoffset, yoffset,
+            &xsum1, &xxsum1);
+    }
 
     xsum0 += xsum1;
     xxsum0 += xxsum1;
@@ -428,12 +471,36 @@ unsigned int vp8_sub_pixel_variance8x16_wmt
 {
     int xsum;
     unsigned int xxsum;
-    vp8_filter_block2d_bil_var_sse2(
-        src_ptr, src_pixels_per_line,
-        dst_ptr, dst_pixels_per_line, 16,
-        vp8_bilinear_filters_xmm[xoffset], vp8_bilinear_filters_xmm[yoffset],
-        &xsum, &xxsum
-    );
+
+    if (xoffset == 4 && yoffset == 0)
+    {
+        vp8_half_horiz_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 16,
+            &xsum, &xxsum);
+    }
+    else if (xoffset == 0 && yoffset == 4)
+    {
+        vp8_half_vert_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 16,
+            &xsum, &xxsum);
+    }
+    else if (xoffset == 4 && yoffset == 4)
+    {
+        vp8_half_horiz_vert_variance16x_h_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 16,
+            &xsum, &xxsum);
+    }
+    else
+    {
+        vp8_filter_block2d_bil_var_sse2(
+            src_ptr, src_pixels_per_line,
+            dst_ptr, dst_pixels_per_line, 16,
+            xoffset, yoffset,
+            &xsum, &xxsum);
+    }
 
     *sse = xxsum;
     return (xxsum - ((xsum * xsum) >> 7));