From ed208f7d2f07a0a6c6c64e5e9fe25d370761b97e Mon Sep 17 00:00:00 2001
From: Jim Bankoski <jimbankoski@google.com>
Date: Tue, 17 Jan 2012 17:27:39 -0800
Subject: [PATCH] vp8d - valgrind warnings in mb post processor

Solved by extending the border in the postproc buffer as necessary

Change-Id: Ic3f61397fe5bc8e4db6fc78050b0b160bd0aee86
---
 vp8/common/postproc.c            | 30 ++++++++++++++++--
 vp8/common/x86/postproc_mmx.asm  | 45 +++++++++++++++++++++++++++
 vp8/common/x86/postproc_sse2.asm | 66 +++++++++++++++++++++++++++++++++++++++-
 3 files changed, 138 insertions(+), 3 deletions(-)

diff --git a/vp8/common/postproc.c b/vp8/common/postproc.c
index eb17a7c..b657675 100644
--- a/vp8/common/postproc.c
+++ b/vp8/common/postproc.c
@@ -178,6 +178,12 @@ void vp8_post_proc_down_and_across_c
         p_src = dst_ptr;
         p_dst = dst_ptr;
 
+        for (i = -8; i<0; i++)
+          p_src[i]=p_src[0];
+
+        for (i = cols; i<cols+8; i++)
+          p_src[i]=p_src[cols-1];
+
         for (i = 0; i < 8; i++)
             d[i] = p_src[i];
 
@@ -228,12 +234,19 @@ void vp8_mbpost_proc_across_ip_c(unsigned char *src, int pitch, int rows, int co
     unsigned char *s = src;
     unsigned char d[16];
 
-
     for (r = 0; r < rows; r++)
     {
         int sumsq = 0;
         int sum   = 0;
 
+        for (i = -8; i<0; i++)
+          s[i]=s[0];
+
+        // 17 avoids valgrind warning - we buffer values in c in d
+        // and only write them when we've read 8 ahead...
+        for (i = cols; i<cols+17; i++)
+          s[i]=s[cols-1];
+
         for (i = -8; i <= 6; i++)
         {
             sumsq += s[i] * s[i];
@@ -272,7 +285,7 @@ void vp8_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols, i
     int r, c, i;
     const short *rv3 = &vp8_rv[63&rand()];
 
-    for (c = 0; c < cols; c++)
+    for (c = 0; c < cols; c++ )
     {
         unsigned char *s = &dst[c];
         int sumsq = 0;
@@ -280,6 +293,14 @@ void vp8_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols, i
         unsigned char d[16];
         const short *rv2 = rv3 + ((c * 17) & 127);
 
+        for (i = -8; i < 0; i++)
+          s[i*pitch]=s[0];
+
+        // 17 avoids valgrind warning - we buffer values in c in d
+        // and only write them when we've read 8 ahead...
+        for (i = rows; i < rows+17; i++)
+          s[i*pitch]=s[(rows-1)*pitch];
+
         for (i = -8; i <= 6; i++)
         {
             sumsq += s[i*pitch] * s[i*pitch];
@@ -320,6 +341,7 @@ static void vp8_deblock_and_de_macro_block(YV12_BUFFER_CONFIG         *source,
     POSTPROC_INVOKE(rtcd, across)(post->y_buffer, post->y_stride, post->y_height, post->y_width, q2mbl(q));
     POSTPROC_INVOKE(rtcd, down)(post->y_buffer, post->y_stride, post->y_height, post->y_width, q2mbl(q));
 
+
     POSTPROC_INVOKE(rtcd, downacross)(source->u_buffer, post->u_buffer, source->uv_stride, post->uv_stride, source->uv_height, source->uv_width, ppl);
     POSTPROC_INVOKE(rtcd, downacross)(source->v_buffer, post->v_buffer, source->uv_stride, post->uv_stride, source->uv_height, source->uv_width, ppl);
 
@@ -935,6 +957,10 @@ int vp8_post_proc_frame(VP8_COMMON *oci, YV12_BUFFER_CONFIG *dest, vp8_ppflags_t
             {
                 oci->post_proc_buffer_int_used = 1;
             }
+            // insure that postproc is set to all 0's so that post proc
+            // doesn't pull random data in from edge
+            vpx_memset((&oci->post_proc_buffer_int)->buffer_alloc,126,(&oci->post_proc_buffer)->frame_size);
+
         }
     }
 
diff --git a/vp8/common/x86/postproc_mmx.asm b/vp8/common/x86/postproc_mmx.asm
index 8112218..d24f740 100644
--- a/vp8/common/x86/postproc_mmx.asm
+++ b/vp8/common/x86/postproc_mmx.asm
@@ -151,6 +151,23 @@ sym(vp8_post_proc_down_and_across_mmx):
         sub         rsi, rdx
         sub         rdi, rdx
 
+        ; dup the first byte into the left border 8 times
+        movq        mm1,   [rdi]
+        punpcklbw   mm1,   mm1
+        punpcklwd   mm1,   mm1
+        punpckldq   mm1,   mm1
+
+        mov         rdx,    -8
+        movq        [rdi+rdx], mm1
+
+        ; dup the last byte into the right border
+        movsxd      rdx,    dword arg(5)
+        movq        mm1,   [rdi + rdx + -1]
+        punpcklbw   mm1,   mm1
+        punpcklwd   mm1,   mm1
+        punpckldq   mm1,   mm1
+        movq        [rdi+rdx], mm1
+
 
         push        rax
         xor         rdx,    rdx
@@ -298,8 +315,36 @@ sym(vp8_mbpost_proc_down_mmx):
             pxor        mm0,        mm0     ;
 
             movsxd      rax,        dword ptr arg(1) ;pitch       ;
+
+            ; this copies the last row down into the border 8 rows
+            mov         rdi,        rsi
+            mov         rdx,        arg(2)
+            sub         rdx,        9
+            imul        rdx,        rax
+            lea         rdi,        [rdi+rdx]
+            movq        mm1,        QWORD ptr[rdi]              ; first row
+            mov         rcx,        8
+.init_borderd                                                    ; initialize borders
+            lea         rdi,        [rdi + rax]
+            movq        [rdi],      xmm1
+
+            dec         rcx
+            jne         .init_borderd
+
             neg         rax                                     ; rax = -pitch
 
+            ; this copies the first row up into the border 8 rows
+            mov         rdi,        rsi
+            movq        mm1,        QWORD ptr[rdi]              ; first row
+            mov         rcx,        8
+.init_border                                                    ; initialize borders
+            lea         rdi,        [rdi + rax]
+            movq        [rdi],      mm1
+
+            dec         rcx
+            jne         .init_border
+
+
             lea         rsi,        [rsi + rax*8];              ; rdi = s[-pitch*8]
             neg         rax
 
diff --git a/vp8/common/x86/postproc_sse2.asm b/vp8/common/x86/postproc_sse2.asm
index 1f219ca..966aafd 100644
--- a/vp8/common/x86/postproc_sse2.asm
+++ b/vp8/common/x86/postproc_sse2.asm
@@ -139,6 +139,24 @@ sym(vp8_post_proc_down_and_across_xmm):
         sub         rsi,        rdx
         sub         rdi,        rdx
 
+
+        ; dup the first byte into the left border 8 times
+        movq        mm1,   [rdi]
+        punpcklbw   mm1,   mm1
+        punpcklwd   mm1,   mm1
+        punpckldq   mm1,   mm1
+
+        mov         rdx,    -8
+        movq        [rdi+rdx], mm1
+
+        ; dup the last byte into the right border
+        movsxd      rdx,    dword arg(5)
+        movq        mm1,   [rdi + rdx + -1]
+        punpcklbw   mm1,   mm1
+        punpcklwd   mm1,   mm1
+        punpckldq   mm1,   mm1
+        movq        [rdi+rdx], mm1
+
         xor         rdx,        rdx
         movq        mm0,        QWORD PTR [rdi-8];
 
@@ -287,12 +305,40 @@ sym(vp8_mbpost_proc_down_xmm):
             pxor        xmm0,       xmm0        ;
 
             movsxd      rax,        dword ptr arg(1) ;pitch       ;
+
+            ; this copies the last row down into the border 8 rows
+            mov         rdi,        rsi
+            mov         rdx,        arg(2)
+            sub         rdx,        9
+            imul        rdx,        rax
+            lea         rdi,        [rdi+rdx]
+            movq        xmm1,       QWORD ptr[rdi]              ; first row
+            mov         rcx,        8
+.init_borderd                                                    ; initialize borders
+            lea         rdi,        [rdi + rax]
+            movq        [rdi],      xmm1
+
+            dec         rcx
+            jne         .init_borderd
+
             neg         rax                                     ; rax = -pitch
 
+            ; this copies the first row up into the border 8 rows
+            mov         rdi,        rsi
+            movq        xmm1,       QWORD ptr[rdi]              ; first row
+            mov         rcx,        8
+.init_border                                                    ; initialize borders
+            lea         rdi,        [rdi + rax]
+            movq        [rdi],      xmm1
+
+            dec         rcx
+            jne         .init_border
+
+
+
             lea         rsi,        [rsi + rax*8];              ; rdi = s[-pitch*8]
             neg         rax
 
-
             pxor        xmm5,       xmm5
             pxor        xmm6,       xmm6        ;
 
@@ -480,7 +526,25 @@ sym(vp8_mbpost_proc_across_ip_xmm):
         xor         rdx,    rdx ;sumsq=0;
         xor         rcx,    rcx ;sum=0;
         mov         rsi,    arg(0); s
+
+
+        ; dup the first byte into the left border 8 times
+        movq        mm1,   [rsi]
+        punpcklbw   mm1,   mm1
+        punpcklwd   mm1,   mm1
+        punpckldq   mm1,   mm1
+
         mov         rdi,    -8
+        movq        [rsi+rdi], mm1
+
+        ; dup the last byte into the right border
+        movsxd      rdx,    dword arg(3)
+        movq        mm1,   [rsi + rdx + -1]
+        punpcklbw   mm1,   mm1
+        punpcklwd   mm1,   mm1
+        punpckldq   mm1,   mm1
+        movq        [rsi+rdx], mm1
+
 .ip_var_loop:
         ;for(i=-8;i<=6;i++)
         ;{
-- 
2.7.4