avfilter/vf_ssim: improve precision

Use doubles for accumulating floats.

libavfilter/ssim.h

@@ -28,7 +28,7 @@ typedef struct SSIMDSPContext {
     void (*ssim_4x4_line)(const uint8_t *buf, ptrdiff_t buf_stride,
                           const uint8_t *ref, ptrdiff_t ref_stride,
                           int (*sums)[4], int w);
-    float (*ssim_end_line)(const int (*sum0)[4], const int (*sum1)[4], int w);
+    double (*ssim_end_line)(const int (*sum0)[4], const int (*sum1)[4], int w);
 } SSIMDSPContext;
 
 void ff_ssim_init_x86(SSIMDSPContext *dsp);

libavfilter/vf_ssim.c

@@ -55,13 +55,13 @@ typedef struct SSIMContext {
     uint64_t nb_frames;
     double ssim[4], ssim_total;
     char comps[4];
-    float coefs[4];
+    double coefs[4];
     uint8_t rgba_map[4];
     int planewidth[4];
     int planeheight[4];
     int *temp;
     int is_rgb;
-    float (*ssim_plane)(SSIMDSPContext *dsp,
+    double (*ssim_plane)(SSIMDSPContext *dsp,
                         uint8_t *main, int main_stride,
                         uint8_t *ref, int ref_stride,
                         int width, int height, void *temp,
@@ -206,9 +206,9 @@ static float ssim_endn_16bit(const int64_t (*sum0)[4], const int64_t (*sum1)[4],
     return ssim;
 }
 
-static float ssim_endn_8bit(const int (*sum0)[4], const int (*sum1)[4], int width)
+static double ssim_endn_8bit(const int (*sum0)[4], const int (*sum1)[4], int width)
 {
-    float ssim = 0.0;
+    double ssim = 0.0;
     int i;
 
     for (i = 0; i < width; i++)
@@ -221,14 +221,14 @@ static float ssim_endn_8bit(const int (*sum0)[4], const int (*sum1)[4], int widt
 
 #define SUM_LEN(w) (((w) >> 2) + 3)
 
-static float ssim_plane_16bit(SSIMDSPContext *dsp,
+static double ssim_plane_16bit(SSIMDSPContext *dsp,
                               uint8_t *main, int main_stride,
                               uint8_t *ref, int ref_stride,
                               int width, int height, void *temp,
                               int max)
 {
     int z = 0, y;
-    float ssim = 0.0;
+    double ssim = 0.0;
     int64_t (*sum0)[4] = temp;
     int64_t (*sum1)[4] = sum0 + SUM_LEN(width);
 
@@ -249,14 +249,14 @@ static float ssim_plane_16bit(SSIMDSPContext *dsp,
     return ssim / ((height - 1) * (width - 1));
 }
 
-static float ssim_plane(SSIMDSPContext *dsp,
+static double ssim_plane(SSIMDSPContext *dsp,
                         uint8_t *main, int main_stride,
                         uint8_t *ref, int ref_stride,
                         int width, int height, void *temp,
                         int max)
 {
     int z = 0, y;
-    float ssim = 0.0;
+    double ssim = 0.0;
     int (*sum0)[4] = temp;
     int (*sum1)[4] = sum0 + SUM_LEN(width);
 
@@ -279,7 +279,7 @@ static float ssim_plane(SSIMDSPContext *dsp,
 
 static double ssim_db(double ssim, double weight)
 {
-    return 10 * log10(weight / (weight - ssim));
+    return (fabs(weight - ssim) > 1e-9) ? 10.0 * log10(weight / (weight - ssim)) : INFINITY;
 }
 
 static int do_ssim(FFFrameSync *fs)
@@ -288,7 +288,7 @@ static int do_ssim(FFFrameSync *fs)
     SSIMContext *s = ctx->priv;
     AVFrame *master, *ref;
     AVDictionary **metadata;
-    float c[4], ssimv = 0.0;
+    double c[4] = { 0 }, ssimv = 0.0;
     int ret, i;
 
     ret = ff_framesync_dualinput_get(fs, &master, &ref);

libavfilter/x86/vf_ssim.asm

@@ -169,8 +169,9 @@ SSIM_4X4_LINE 8
 %endif
 
 INIT_XMM sse4
-cglobal ssim_end_line, 3, 3, 6, sum0, sum1, w
+cglobal ssim_end_line, 3, 3, 7, sum0, sum1, w
     pxor              m0, m0
+    pxor              m6, m6
 .loop:
     mova              m1, [sum0q+mmsize*0]
     mova              m2, [sum0q+mmsize*1]
@@ -214,34 +215,46 @@ cglobal ssim_end_line, 3, 3, 6, sum0, sum1, w
     mulps             m4, m5
     mulps             m3, m1
     divps             m4, m3                    ; ssim_endl
-    addps             m0, m4                    ; ssim
+    mova              m5, m4
+    cvtps2pd          m3, m5
+    movhlps           m5, m5
+    cvtps2pd          m5, m5
+    addpd             m0, m3                    ; ssim
+    addpd             m6, m5                    ; ssim
     add            sum0q, mmsize*4
     add            sum1q, mmsize*4
     sub               wd, 4
     jg .loop
 
-    ; subps the ones we added too much
+    ; subpd the ones we added too much
     test              wd, wd
     jz .end
     add               wd, 4
+    test              wd, 3
+    jz .skip3
     test              wd, 2
     jz .skip2
-    psrldq            m4, 8
-.skip2:
     test              wd, 1
     jz .skip1
-    psrldq            m4, 4
+.skip3:
+    psrldq            m5, 8
+    subpd             m6, m5
+    jmp .end
+.skip2:
+    psrldq            m5, 8
+    subpd             m6, m5
+    subpd             m0, m3
+    jmp .end
 .skip1:
-    subps             m0, m4
+    psrldq            m3, 16
+    subpd             m6, m5
 
 .end:
+    addpd             m0, m6
     movhlps           m4, m0
-    addps             m0, m4
-    movss             m4, m0
-    shufps            m0, m0, 1
-    addss             m0, m4
+    addpd             m0, m4
 %if ARCH_X86_32
-    movss            r0m, m0
+    movsd            r0m, m0
     fld             r0mp
 %endif
     RET

libavfilter/x86/vf_ssim_init.c

@@ -28,7 +28,7 @@ void ff_ssim_4x4_line_ssse3(const uint8_t *buf, ptrdiff_t buf_stride,
 void ff_ssim_4x4_line_xop  (const uint8_t *buf, ptrdiff_t buf_stride,
                             const uint8_t *ref, ptrdiff_t ref_stride,
                             int (*sums)[4], int w);
-float ff_ssim_end_line_sse4(const int (*sum0)[4], const int (*sum1)[4], int w);
+double ff_ssim_end_line_sse4(const int (*sum0)[4], const int (*sum1)[4], int w);
 
 void ff_ssim_init_x86(SSIMDSPContext *dsp)
 {