shithub: libvpx

Info  •  Files  •  Log  •  Branches

[VSX] Add support to Power9-only vec_absd

~5% gain for SAD.

Change-Id: Ief7d7691f837474f5b6b582129628276fdcce319

--- a/vpx_dsp/ppc/deblock_vsx.c

+++ b/vpx_dsp/ppc/deblock_vsx.c

@@ -23,17 +23,13 @@

                                      0x06, 0x07, 0x18, 0x19, 0x1A, 0x1B,

                                      0x1C, 0x1D, 0x1E, 0x1F };

-static INLINE uint8x16_t vec_abd_s8(uint8x16_t a, uint8x16_t b) {

-  return vec_sub(vec_max(a, b), vec_min(a, b));

-}

-

 static INLINE uint8x16_t apply_filter(uint8x16_t ctx[4], uint8x16_t v,

                                       uint8x16_t filter) {

   const uint8x16_t k1 = vec_avg(ctx[0], ctx[1]);

   const uint8x16_t k2 = vec_avg(ctx[3], ctx[2]);

   const uint8x16_t k3 = vec_avg(k1, k2);

-  const uint8x16_t f_a = vec_max(vec_abd_s8(v, ctx[0]), vec_abd_s8(v, ctx[1]));

-  const uint8x16_t f_b = vec_max(vec_abd_s8(v, ctx[2]), vec_abd_s8(v, ctx[3]));

+  const uint8x16_t f_a = vec_max(vec_absd(v, ctx[0]), vec_absd(v, ctx[1]));

+  const uint8x16_t f_b = vec_max(vec_absd(v, ctx[2]), vec_absd(v, ctx[3]));

   const bool8x16_t mask = vec_cmplt(vec_max(f_a, f_b), filter);

   return vec_sel(v, vec_avg(k3, v), mask);

 }

--- a/vpx_dsp/ppc/sad_vsx.c

+++ b/vpx_dsp/ppc/sad_vsx.c

@@ -17,10 +17,10 @@

 #include "vpx/vpx_integer.h"

 #include "vpx_ports/mem.h"

-#define PROCESS16(offset)                                \

-  v_a = vec_vsx_ld(offset, a);                           \

-  v_b = vec_vsx_ld(offset, b);                           \

-  v_abs = vec_sub(vec_max(v_a, v_b), vec_min(v_a, v_b)); \

+#define PROCESS16(offset)      \

+  v_a = vec_vsx_ld(offset, a); \

+  v_b = vec_vsx_ld(offset, b); \

+  v_abs = vec_absd(v_a, v_b);  \

   v_sad = vec_sum4s(v_abs, v_sad);

 #define SAD8(height)                                                     \

--- a/vpx_dsp/ppc/types_vsx.h

+++ b/vpx_dsp/ppc/types_vsx.h

@@ -75,6 +75,10 @@

   return (uint8x16_t)vec_mergeh(a0, a1);

 }

+#ifndef __POWER9_VECTOR__

+#define vec_absd(a, b) vec_sub(vec_max(a, b), vec_min(a, b))

+#endif

+

 static const uint8x16_t vec_zeros_u8 = { 0, 0, 0, 0, 0, 0, 0, 0,

                                          0, 0, 0, 0, 0, 0, 0, 0 };

 static const int16x8_t vec_zeros_s16 = { 0, 0, 0, 0, 0, 0, 0, 0 };