fixed an overflow in ssim calculation
This commit fixed an overflow in ssim calculation, added register
save and restore to make sure assembly code working for x64 platform.
It also changed the sampling points to every 4x4 instead of 8x8 and
adjusted the constants in SSIM calculation to match the scale of
previous VPXSSIM.
Change-Id: Ia4dbb8c69eac55812f4662c88ab4653b6720537b
diff --git a/vp8/encoder/ssim.c b/vp8/encoder/ssim.c
index 64d67c6..c78be37 100644
--- a/vp8/encoder/ssim.c
+++ b/vp8/encoder/ssim.c
@@ -290,8 +290,8 @@
}
}
-const static long long c1 = 426148; // (256^2*(.01*255)^2
-const static long long c2 = 3835331; //(256^2*(.03*255)^2
+const static long long cc1 = 26634; // (64^2*(.01*255)^2
+const static long long cc2 = 239708; // (64^2*(.03*255)^2
static double similarity
(
@@ -303,10 +303,19 @@
int count
)
{
- long long ssim_n = (2*sum_s*sum_r+ c1)*(2*count*sum_sxr-2*sum_s*sum_r+c2);
+ long long ssim_n, ssim_d;
+ long long c1, c2;
- long long ssim_d = (sum_s*sum_s +sum_r*sum_r+c1)*
- (count*sum_sq_s-sum_s*sum_s + count*sum_sq_r-sum_r*sum_r +c2) ;
+ //scale the constants by number of pixels
+ c1 = (cc1*count*count)>>12;
+ c2 = (cc2*count*count)>>12;
+
+ ssim_n = (2*sum_s*sum_r+ c1)*((long long) 2*count*sum_sxr-
+ (long long) 2*sum_s*sum_r+c2);
+
+ ssim_d = (sum_s*sum_s +sum_r*sum_r+c1)*
+ ((long long)count*sum_sq_s-(long long)sum_s*sum_s +
+ (long long)count*sum_sq_r-(long long) sum_r*sum_r +c2) ;
return ssim_n * 1.0 / ssim_d;
}
@@ -332,18 +341,33 @@
const vp8_variance_rtcd_vtable_t *rtcd)
{
unsigned long sum_s=0,sum_r=0,sum_sq_s=0,sum_sq_r=0,sum_sxr=0;
- double ssim3;
- long long ssim_n;
- long long ssim_d;
+ long long ssim3;
+ long long ssim_n,ssim_n1,ssim_n2;
+ long long ssim_d,ssim_d1,ssim_d2;
+ long long ssim_t1,ssim_t2;
+ long long c1, c2;
+
+ // normalize by 256/64
+ c1 = cc1*16;
+ c2 = cc2*16;
rtcd->ssimpf(s, sp, r, rp, &sum_s, &sum_r, &sum_sq_s, &sum_sq_r, &sum_sxr);
- ssim_n = (2*sum_s*sum_r+ c1)*(2*256*sum_sxr-2*sum_s*sum_r+c2);
+ ssim_n1 = (2*sum_s*sum_r+ c1);
- ssim_d = (sum_s*sum_s +sum_r*sum_r+c1)*
- (256*sum_sq_s-sum_s*sum_s + 256*sum_sq_r-sum_r*sum_r +c2) ;
+ ssim_n2 =((long long) 2*256*sum_sxr-(long long) 2*sum_s*sum_r+c2);
- ssim3 = 256 * (ssim_d-ssim_n) / ssim_d;
- return (long)( 256*ssim3 * ssim3 );
+ ssim_d1 =((long long)sum_s*sum_s +(long long)sum_r*sum_r+c1);
+
+ ssim_d2 = (256 * (long long) sum_sq_s-(long long) sum_s*sum_s +
+ (long long) 256*sum_sq_r-(long long) sum_r*sum_r +c2) ;
+
+ ssim_t1 = 256 - 256 * ssim_n1 / ssim_d1;
+ ssim_t2 = 256 - 256 * ssim_n2 / ssim_d2;
+
+ ssim3 = 256 *ssim_t1 * ssim_t2;
+ if(ssim3 <0 )
+ ssim3=0;
+ return (long)( ssim3 );
}
// TODO: (jbb) this 8x8 window might be too big + we may want to pick pixels
// such that the window regions overlap block boundaries to penalize blocking
@@ -361,18 +385,20 @@
)
{
int i,j;
-
+ int samples =0;
double ssim_total=0;
- // we can sample points as frequently as we like start with 1 per 8x8
- for(i=0; i < height; i+=8, img1 += stride_img1*8, img2 += stride_img2*8)
+ // we can sample points as frequently as we like start with 1 per 4x4
+ for(i=0; i < height-8; i+=4, img1 += stride_img1*4, img2 += stride_img2*4)
{
- for(j=0; j < width; j+=8 )
+ for(j=0; j < width-8; j+=4 )
{
- ssim_total += ssim_8x8(img1, stride_img1, img2, stride_img2, rtcd);
+ double v = ssim_8x8(img1+j, stride_img1, img2+j, stride_img2, rtcd);
+ ssim_total += v;
+ samples++;
}
}
- ssim_total /= (width/8 * height /8);
+ ssim_total /= samples;
return ssim_total;
}
@@ -405,4 +431,4 @@
*weight = 1;
return ssimv;
-}
+}
\ No newline at end of file
diff --git a/vp8/encoder/x86/ssim_opt.asm b/vp8/encoder/x86/ssim_opt.asm
index c267cdb..d6cebf3 100644
--- a/vp8/encoder/x86/ssim_opt.asm
+++ b/vp8/encoder/x86/ssim_opt.asm
@@ -16,12 +16,12 @@
paddusw xmm14, xmm4 ; sum_r
movdqa xmm1, xmm3
pmaddwd xmm1, xmm1
- paddq xmm13, xmm1 ; sum_sq_s
+ paddd xmm13, xmm1 ; sum_sq_s
movdqa xmm2, xmm4
pmaddwd xmm2, xmm2
- paddq xmm12, xmm2 ; sum_sq_r
+ paddd xmm12, xmm2 ; sum_sq_r
pmaddwd xmm3, xmm4
- paddq xmm11, xmm3 ; sum_sxr
+ paddd xmm11, xmm3 ; sum_sxr
%endmacro
; Sum across the register %1 starting with q words
@@ -66,6 +66,7 @@
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 9
+ SAVE_XMM
push rsi
push rdi
; end prolog
@@ -115,19 +116,20 @@
SUM_ACROSS_Q xmm11
mov rdi,arg(4)
- movq [rdi], xmm15;
+ movd [rdi], xmm15;
mov rdi,arg(5)
- movq [rdi], xmm14;
+ movd [rdi], xmm14;
mov rdi,arg(6)
- movq [rdi], xmm13;
+ movd [rdi], xmm13;
mov rdi,arg(7)
- movq [rdi], xmm12;
+ movd [rdi], xmm12;
mov rdi,arg(8)
- movq [rdi], xmm11;
+ movd [rdi], xmm11;
; begin epilog
pop rdi
pop rsi
+ RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
@@ -154,6 +156,7 @@
push rbp
mov rbp, rsp
SHADOW_ARGS_TO_STACK 9
+ SAVE_XMM
push rsi
push rdi
; end prolog
@@ -174,11 +177,8 @@
NextRow2:
;grab source and reference pixels
- movq xmm5, [rsi]
- movq xmm6, [rdi]
-
- movdqa xmm3, xmm5
- movdqa xmm4, xmm6
+ movq xmm3, [rsi]
+ movq xmm4, [rdi]
punpcklbw xmm3, xmm0 ; low_s
punpcklbw xmm4, xmm0 ; low_r
@@ -197,19 +197,20 @@
SUM_ACROSS_Q xmm11
mov rdi,arg(4)
- movq [rdi], xmm15;
+ movd [rdi], xmm15;
mov rdi,arg(5)
- movq [rdi], xmm14;
+ movd [rdi], xmm14;
mov rdi,arg(6)
- movq [rdi], xmm13;
+ movd [rdi], xmm13;
mov rdi,arg(7)
- movq [rdi], xmm12;
+ movd [rdi], xmm12;
mov rdi,arg(8)
- movq [rdi], xmm11;
+ movd [rdi], xmm11;
; begin epilog
pop rdi
pop rsi
+ RESTORE_XMM
UNSHADOW_ARGS
pop rbp
ret
