shithub: libvpx

Info  •  Files  •  Log  •  Branches

loopfilter_sse2: call unsuffixed lpf functions

this allows calls to use better versions (e.g., avx2) if available. in
most other cases the function pointer will be defined to the sse2
variant if another isn't available. this improves performance at 1080P
by ~2% on a Xeon E5-2690.

Change-Id: Ie9da3a567021f8416651a29b8c9ab9238dc4bdf1

--- a/vpx_dsp/x86/loopfilter_sse2.c

+++ b/vpx_dsp/x86/loopfilter_sse2.c

@@ -1674,8 +1674,8 @@

   transpose8x16(s - 4, s - 4 + pitch * 8, pitch, t_dst, 16);

   // Loop filtering

-  vpx_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,

-                                 blimit1, limit1, thresh1);

+  vpx_lpf_horizontal_4_dual(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,

+                            blimit1, limit1, thresh1);

   src[0] = t_dst;

   src[1] = t_dst + 8;

   dst[0] = s - 4;

@@ -1700,7 +1700,7 @@

   transpose(src, pitch, dst, 8, 1);

   // Loop filtering

-  vpx_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh);

+  vpx_lpf_horizontal_8(t_dst + 4 * 8, 8, blimit, limit, thresh);

   src[0] = t_dst;

   dst[0] = s - 4;

@@ -1721,8 +1721,8 @@

   transpose8x16(s - 4, s - 4 + pitch * 8, pitch, t_dst, 16);

   // Loop filtering

-  vpx_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,

-                                 blimit1, limit1, thresh1);

+  vpx_lpf_horizontal_8_dual(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,

+                            blimit1, limit1, thresh1);

   src[0] = t_dst;

   src[1] = t_dst + 8;

@@ -1750,7 +1750,7 @@

   transpose(src, pitch, dst, 8, 2);

   // Loop filtering

-  vpx_lpf_horizontal_16_sse2(t_dst + 8 * 8, 8, blimit, limit, thresh);

+  vpx_lpf_horizontal_16(t_dst + 8 * 8, 8, blimit, limit, thresh);

   src[0] = t_dst;

   src[1] = t_dst + 8 * 8;

@@ -1771,7 +1771,7 @@

   transpose8x16(s, s + 8 * pitch, pitch, t_dst + 8 * 16, 16);

   // Loop filtering

-  vpx_lpf_horizontal_16_dual_sse2(t_dst + 8 * 16, 16, blimit, limit, thresh);

+  vpx_lpf_horizontal_16_dual(t_dst + 8 * 16, 16, blimit, limit, thresh);

   // Transpose back

   transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, pitch);