Merge "ppc: Add vpx_sadnxmx4d_vsx for n,m = {8, 16, 32 ,64}"

diff --git a/test/sad_test.cc b/test/sad_test.cc
index 5d515b589987d246f709c0b9577e4b1cb46d040a..7bf6e13856cebc84326b23cefa4dcb24dd60ae4a 100644 (file)
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -946,5 +946,17 @@ const SadMxNAvgParam avg_vsx_tests[] = {
   SadMxNAvgParam(16, 8, &vpx_sad16x8_avg_vsx),
 };
 INSTANTIATE_TEST_CASE_P(VSX, SADavgTest, ::testing::ValuesIn(avg_vsx_tests));
+
+const SadMxNx4Param x4d_vsx_tests[] = {
+  SadMxNx4Param(64, 64, &vpx_sad64x64x4d_vsx),
+  SadMxNx4Param(64, 32, &vpx_sad64x32x4d_vsx),
+  SadMxNx4Param(32, 64, &vpx_sad32x64x4d_vsx),
+  SadMxNx4Param(32, 32, &vpx_sad32x32x4d_vsx),
+  SadMxNx4Param(32, 16, &vpx_sad32x16x4d_vsx),
+  SadMxNx4Param(16, 32, &vpx_sad16x32x4d_vsx),
+  SadMxNx4Param(16, 16, &vpx_sad16x16x4d_vsx),
+  SadMxNx4Param(16, 8, &vpx_sad16x8x4d_vsx),
+};
+INSTANTIATE_TEST_CASE_P(VSX, SADx4Test, ::testing::ValuesIn(x4d_vsx_tests));
 #endif  // HAVE_VSX
 }  // namespace

diff --git a/vpx_dsp/ppc/sad_vsx.c b/vpx_dsp/ppc/sad_vsx.c
index 1657df1e660344b907dd1fcdda7e1f4f38368ffe..bb49addae179585ee7d0b3476d913bc36e0c3271 100644 (file)
--- a/vpx_dsp/ppc/sad_vsx.c
+++ b/vpx_dsp/ppc/sad_vsx.c
@@ -133,7 +133,6 @@ SAD64(64);
     DECLARE_ALIGNED(64, uint8_t, comp_pred[64 * height]);                     \
     vpx_comp_avg_pred_vsx(comp_pred, second_pred, 64, height, ref,            \
                           ref_stride);                                        \
-                                                                              \
     return vpx_sad64x##height##_vsx(src, src_stride, comp_pred, 64);          \
   }
 
@@ -145,3 +144,111 @@ SAD32AVG(32);
 SAD32AVG(64);
 SAD64AVG(32);
 SAD64AVG(64);
+
+#define PROCESS16_4D(offset, ref, v_h, v_l) \
+  v_b = vec_vsx_ld(offset, ref);            \
+  v_bh = unpack_to_s16_h(v_b);              \
+  v_bl = unpack_to_s16_l(v_b);              \
+  v_subh = vec_sub(v_h, v_bh);              \
+  v_subl = vec_sub(v_l, v_bl);              \
+  v_absh = vec_abs(v_subh);                 \
+  v_absl = vec_abs(v_subl);                 \
+  v_sad = vec_sum4s(v_absh, v_sad);         \
+  v_sad = vec_sum4s(v_absl, v_sad);
+
+#define UNPACK_SRC(offset, srcv_h, srcv_l) \
+  v_a = vec_vsx_ld(offset, src);           \
+  srcv_h = unpack_to_s16_h(v_a);           \
+  srcv_l = unpack_to_s16_l(v_a);
+
+#define SAD16_4D(height)                                                  \
+  void vpx_sad16x##height##x4d_vsx(const uint8_t *src, int src_stride,    \
+                                   const uint8_t *const ref_array[],      \
+                                   int ref_stride, uint32_t *sad_array) { \
+    int i;                                                                \
+    int y;                                                                \
+    unsigned int sad[4];                                                  \
+    uint8x16_t v_a, v_b;                                                  \
+    int16x8_t v_ah, v_al, v_bh, v_bl, v_absh, v_absl, v_subh, v_subl;     \
+                                                                          \
+    for (i = 0; i < 4; i++) sad_array[i] = 0;                             \
+                                                                          \
+    for (y = 0; y < height; y++) {                                        \
+      UNPACK_SRC(y *src_stride, v_ah, v_al);                              \
+      for (i = 0; i < 4; i++) {                                           \
+        int32x4_t v_sad = vec_splat_s32(0);                               \
+        PROCESS16_4D(y *ref_stride, ref_array[i], v_ah, v_al);            \
+                                                                          \
+        vec_vsx_st((uint32x4_t)v_sad, 0, sad);                            \
+        sad_array[i] += (sad[3] + sad[2] + sad[1] + sad[0]);              \
+      }                                                                   \
+    }                                                                     \
+  }
+
+#define SAD32_4D(height)                                                  \
+  void vpx_sad32x##height##x4d_vsx(const uint8_t *src, int src_stride,    \
+                                   const uint8_t *const ref_array[],      \
+                                   int ref_stride, uint32_t *sad_array) { \
+    int i;                                                                \
+    int y;                                                                \
+    unsigned int sad[4];                                                  \
+    uint8x16_t v_a, v_b;                                                  \
+    int16x8_t v_ah1, v_al1, v_ah2, v_al2, v_bh, v_bl;                     \
+    int16x8_t v_absh, v_absl, v_subh, v_subl;                             \
+                                                                          \
+    for (i = 0; i < 4; i++) sad_array[i] = 0;                             \
+                                                                          \
+    for (y = 0; y < height; y++) {                                        \
+      UNPACK_SRC(y *src_stride, v_ah1, v_al1);                            \
+      UNPACK_SRC(y *src_stride + 16, v_ah2, v_al2);                       \
+      for (i = 0; i < 4; i++) {                                           \
+        int32x4_t v_sad = vec_splat_s32(0);                               \
+        PROCESS16_4D(y *ref_stride, ref_array[i], v_ah1, v_al1);          \
+        PROCESS16_4D(y *ref_stride + 16, ref_array[i], v_ah2, v_al2);     \
+                                                                          \
+        vec_vsx_st((uint32x4_t)v_sad, 0, sad);                            \
+        sad_array[i] += (sad[3] + sad[2] + sad[1] + sad[0]);              \
+      }                                                                   \
+    }                                                                     \
+  }
+
+#define SAD64_4D(height)                                                  \
+  void vpx_sad64x##height##x4d_vsx(const uint8_t *src, int src_stride,    \
+                                   const uint8_t *const ref_array[],      \
+                                   int ref_stride, uint32_t *sad_array) { \
+    int i;                                                                \
+    int y;                                                                \
+    unsigned int sad[4];                                                  \
+    uint8x16_t v_a, v_b;                                                  \
+    int16x8_t v_ah1, v_al1, v_ah2, v_al2, v_bh, v_bl;                     \
+    int16x8_t v_ah3, v_al3, v_ah4, v_al4;                                 \
+    int16x8_t v_absh, v_absl, v_subh, v_subl;                             \
+                                                                          \
+    for (i = 0; i < 4; i++) sad_array[i] = 0;                             \
+                                                                          \
+    for (y = 0; y < height; y++) {                                        \
+      UNPACK_SRC(y *src_stride, v_ah1, v_al1);                            \
+      UNPACK_SRC(y *src_stride + 16, v_ah2, v_al2);                       \
+      UNPACK_SRC(y *src_stride + 32, v_ah3, v_al3);                       \
+      UNPACK_SRC(y *src_stride + 48, v_ah4, v_al4);                       \
+      for (i = 0; i < 4; i++) {                                           \
+        int32x4_t v_sad = vec_splat_s32(0);                               \
+        PROCESS16_4D(y *ref_stride, ref_array[i], v_ah1, v_al1);          \
+        PROCESS16_4D(y *ref_stride + 16, ref_array[i], v_ah2, v_al2);     \
+        PROCESS16_4D(y *ref_stride + 32, ref_array[i], v_ah3, v_al3);     \
+        PROCESS16_4D(y *ref_stride + 48, ref_array[i], v_ah4, v_al4);     \
+                                                                          \
+        vec_vsx_st((uint32x4_t)v_sad, 0, sad);                            \
+        sad_array[i] += (sad[3] + sad[2] + sad[1] + sad[0]);              \
+      }                                                                   \
+    }                                                                     \
+  }
+
+SAD16_4D(8);
+SAD16_4D(16);
+SAD16_4D(32);
+SAD32_4D(16);
+SAD32_4D(32);
+SAD32_4D(64);
+SAD64_4D(32);
+SAD64_4D(64);

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 9f7f11285a5a2ecacad403fec3282de9d65edbc1..9f005b284058092a643f64ee400dbcffc520fbb3 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -873,28 +873,28 @@ specialize qw/vpx_sad4x4x8 sse4_1 msa/;
 # Multi-block SAD, comparing a reference to N independent blocks
 #
 add_proto qw/void vpx_sad64x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad64x64x4d avx2 neon msa sse2/;
+specialize qw/vpx_sad64x64x4d avx2 neon msa sse2 vsx/;
 
 add_proto qw/void vpx_sad64x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad64x32x4d msa sse2/;
+specialize qw/vpx_sad64x32x4d msa sse2 vsx/;
 
 add_proto qw/void vpx_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad32x64x4d msa sse2/;
+specialize qw/vpx_sad32x64x4d msa sse2 vsx/;
 
 add_proto qw/void vpx_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad32x32x4d avx2 neon msa sse2/;
+specialize qw/vpx_sad32x32x4d avx2 neon msa sse2 vsx/;
 
 add_proto qw/void vpx_sad32x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad32x16x4d msa sse2/;
+specialize qw/vpx_sad32x16x4d msa sse2 vsx/;
 
 add_proto qw/void vpx_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad16x32x4d msa sse2/;
+specialize qw/vpx_sad16x32x4d msa sse2 vsx/;
 
 add_proto qw/void vpx_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad16x16x4d neon msa sse2/;
+specialize qw/vpx_sad16x16x4d neon msa sse2 vsx/;
 
 add_proto qw/void vpx_sad16x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad16x8x4d msa sse2/;
+specialize qw/vpx_sad16x8x4d msa sse2 vsx/;
 
 add_proto qw/void vpx_sad8x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
 specialize qw/vpx_sad8x16x4d msa sse2/;