From: Luca Barbato <lu_zero@gentoo.org>
Date: Sat, 29 Apr 2017 12:34:20 +0000 (+0000)
Subject: ppc: Add convolve8_horiz_avg
X-Git-Tag: v1.7.0~502^2~3
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=77772350f3082470e509ce47418e08af57d5fe28;p=libvpx

ppc: Add convolve8_horiz_avg

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

Change-Id: I54549b03ac6c7a4e3f485738b100c3cac7ac2e15
---

diff --git a/test/convolve_test.cc b/test/convolve_test.cc
index c0186d94b..9f9bf5d0c 100644
--- a/test/convolve_test.cc
+++ b/test/convolve_test.cc
@@ -1225,7 +1225,7 @@ INSTANTIATE_TEST_CASE_P(MSA, ConvolveTest,
 #if HAVE_VSX
 const ConvolveFunctions convolve8_vsx(
     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_avg_horiz_vsx, 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,
     vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
diff --git a/vpx_dsp/ppc/vpx_convolve_vsx.c b/vpx_dsp/ppc/vpx_convolve_vsx.c
index 7c02cfdea..af7032d9f 100644
--- a/vpx_dsp/ppc/vpx_convolve_vsx.c
+++ b/vpx_dsp/ppc/vpx_convolve_vsx.c
@@ -160,29 +160,53 @@ void vpx_convolve_avg_vsx(const uint8_t *src, ptrdiff_t src_stride,
   }
 }
 
+static inline void convolve_line(uint8_t *dst, const uint8_t *const src_x,
+                                 const int16_t *const x_filter) {
+  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);
+}
+
 // 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) {
+static inline 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) {
+      convolve_line(dst + x, &src[x_q4 >> SUBPEL_BITS],
+                    x_filters[x_q4 & SUBPEL_MASK]);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static inline void convolve_avg_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);
+      uint8_t v;
+      convolve_line(&v, &src[x_q4 >> SUBPEL_BITS],
+                    x_filters[x_q4 & SUBPEL_MASK]);
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] + v, 1);
       x_q4 += x_step_q4;
     }
     src += src_stride;
@@ -204,3 +228,18 @@ void vpx_convolve8_horiz_vsx(const uint8_t *src, ptrdiff_t src_stride,
   convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
                  w, h);
 }
+
+void vpx_convolve8_avg_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_avg_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                     x_step_q4, w, h);
+}
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 4c8fbef17..e7b28caba 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -350,7 +350,7 @@ add_proto qw/void vpx_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride,
 specialize qw/vpx_convolve8_avg sse2 ssse3 neon dspr2 msa/;
 
 add_proto qw/void vpx_convolve8_avg_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_avg_horiz sse2 ssse3 neon dspr2 msa/;
+specialize qw/vpx_convolve8_avg_horiz sse2 ssse3 neon dspr2 msa vsx/;
 
 add_proto qw/void vpx_convolve8_avg_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_avg_vert sse2 ssse3 neon dspr2 msa/;