shithub: libvpx

Info  •  Files  •  Log  •  Branches

ppc: Add convolve8_horiz

The 8x8 and larger blocks cases can be sped up further.

Change-Id: I89b635d6b01c59f523f2d54b1284ed32916c5046

--- a/test/convolve_test.cc

+++ b/test/convolve_test.cc

@@ -1224,7 +1224,7 @@

 #if HAVE_VSX

 const ConvolveFunctions convolve8_vsx(

-    vpx_convolve_copy_vsx, vpx_convolve_avg_vsx, vpx_convolve8_horiz_c,

+    vpx_convolve_copy_vsx, vpx_convolve_avg_vsx, vpx_convolve8_horiz_vsx,

     vpx_convolve8_avg_horiz_c, vpx_convolve8_vert_c, vpx_convolve8_avg_vert_c,

     vpx_convolve8_c, vpx_convolve8_avg_c, vpx_scaled_horiz_c,

     vpx_scaled_avg_horiz_c, vpx_scaled_vert_c, vpx_scaled_avg_vert_c,

--- a/vpx_dsp/ppc/vpx_convolve_vsx.c

+++ b/vpx_dsp/ppc/vpx_convolve_vsx.c

@@ -9,6 +9,7 @@

  */

 #include <string.h>

 #include "./vpx_dsp_rtcd.h"

+#include "vpx_dsp/vpx_filter.h"

 #include "vpx_dsp/ppc/types_vsx.h"

 // TODO(lu_zero): unroll

@@ -157,4 +158,49 @@

       break;

     }

   }

+}

+

+// TODO(lu_zero): Implement 8x8 and bigger block special cases

+static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,

+                           uint8_t *dst, ptrdiff_t dst_stride,

+                           const InterpKernel *x_filters, int x0_q4,

+                           int x_step_q4, int w, int h) {

+  int x, y;

+  src -= SUBPEL_TAPS / 2 - 1;

+

+  for (y = 0; y < h; ++y) {

+    int x_q4 = x0_q4;

+    for (x = 0; x < w; ++x) {

+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];

+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];

+      const int16x8_t s = unpack_to_s16_h(vec_vsx_ld(0, src_x));

+      const int16x8_t f = vec_vsx_ld(0, x_filter);

+      const int32x4_t sum = vec_msum(s, f, vec_splat_s32(0));

+      const int32x4_t bias =

+          vec_sl(vec_splat_s32(1), vec_splat_u32(FILTER_BITS - 1));

+      const int32x4_t avg =

+          vec_sr(vec_sums(sum, bias), vec_splat_u32(FILTER_BITS));

+      const uint8x16_t v = vec_splat(

+          vec_packsu(vec_pack(avg, vec_splat_s32(0)), vec_splat_s16(0)), 3);

+      vec_ste(v, 0, dst + x);

+      x_q4 += x_step_q4;

+    }

+    src += src_stride;

+    dst += dst_stride;

+  }

+}

+

+void vpx_convolve8_horiz_vsx(const uint8_t *src, ptrdiff_t src_stride,

+                             uint8_t *dst, ptrdiff_t dst_stride,

+                             const int16_t *filter_x, int x_step_q4,

+                             const int16_t *filter_y, int y_step_q4, int w,

+                             int h) {

+  const InterpKernel *const filters_x = get_filter_base(filter_x);

+  const int x0_q4 = get_filter_offset(filter_x, filters_x);

+

+  (void)filter_y;

+  (void)y_step_q4;

+

+  convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,

+                 w, h);

 }

--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl

+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -341,7 +341,7 @@

 specialize qw/vpx_convolve8 sse2 ssse3 avx2 neon dspr2 msa/;

 add_proto qw/void vpx_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";

-specialize qw/vpx_convolve8_horiz sse2 ssse3 avx2 neon dspr2 msa/;

+specialize qw/vpx_convolve8_horiz sse2 ssse3 avx2 neon dspr2 msa vsx/;

 add_proto qw/void vpx_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";

 specialize qw/vpx_convolve8_vert sse2 ssse3 avx2 neon dspr2 msa/;