ppc: Add get_mb_ss_vsx

Change-Id: I1b54a7a5bb642e4b836d786ea1ae506eed025e3f

diff --git a/test/variance_test.cc b/test/variance_test.cc
index e1c549aa8b12682e3697aa01a286cc49ce1c9dda..4fc5cf5d65293232d8212e2a9ee0e3030b1f8197 100644 (file)
--- a/test/variance_test.cc
+++ b/test/variance_test.cc
@@ -1340,6 +1340,9 @@ INSTANTIATE_TEST_CASE_P(
 #endif  // HAVE_MSA
 
 #if HAVE_VSX
+INSTANTIATE_TEST_CASE_P(VSX, SumOfSquaresTest,
+                        ::testing::Values(vpx_get_mb_ss_vsx));
+
 INSTANTIATE_TEST_CASE_P(VSX, VpxSseTest,
                         ::testing::Values(SseParams(2, 2,
                                                     &vpx_get4x4sse_cs_vsx)));

diff --git a/vpx_dsp/ppc/variance_vsx.c b/vpx_dsp/ppc/variance_vsx.c
index 8eb6f1a129f706dfb950d9498bfc093fc3408e2b..1efe2f00569f5498f3b6cfdc1ad35640263ca638 100644 (file)
--- a/vpx_dsp/ppc/variance_vsx.c
+++ b/vpx_dsp/ppc/variance_vsx.c
@@ -38,6 +38,23 @@ uint32_t vpx_get4x4sse_cs_vsx(const uint8_t *a, int a_stride, const uint8_t *b,
   return distortion;
 }
 
+// TODO(lu_zero): Unroll
+uint32_t vpx_get_mb_ss_vsx(const int16_t *a) {
+  unsigned int i, sum = 0;
+  int32x4_t s = vec_splat_s32(0);
+
+  for (i = 0; i < 256; i += 8) {
+    const int16x8_t v = vec_vsx_ld(0, a + i);
+    s = vec_msum(v, v, s);
+  }
+
+  s = vec_splat(vec_sums(s, vec_splat_s32(0)), 3);
+
+  vec_ste((uint32x4_t)s, 0, &sum);
+
+  return sum;
+}
+
 void vpx_comp_avg_pred_vsx(uint8_t *comp_pred, const uint8_t *pred, int width,
                            int height, const uint8_t *ref, int ref_stride) {
   int i, j;

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 7368d90ea827bdfdceb55a2626a3822aea7665f7..56fe66542e638af70825b29c6c41acf3cd0a2d72 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -1171,7 +1171,7 @@ add_proto qw/unsigned int vpx_mse8x8/, "const uint8_t *src_ptr, int  source_stri
   specialize qw/vpx_mse8x8 sse2 msa/;
 
 add_proto qw/unsigned int vpx_get_mb_ss/, "const int16_t *";
-  specialize qw/vpx_get_mb_ss sse2 msa/;
+  specialize qw/vpx_get_mb_ss sse2 msa vsx/;
 
 add_proto qw/unsigned int vpx_get4x4sse_cs/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride";
   specialize qw/vpx_get4x4sse_cs neon msa vsx/;