mips msa vpx_dsp sad sad4d avgsad optimization

average improvement ~3x-5x

Change-Id: Ie30748cfbedebbd544b7ef4f286055ccb7f60306

5 files changed, 1008 insertions, 39 deletions

diff --git a/test/sad_test.cc b/test/sad_test.cc
index 001d21dad..1c17bfec7 100644
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -1114,4 +1114,98 @@ const SadMxNx4Param x4d_avx2_tests[] = {
 INSTANTIATE_TEST_CASE_P(AVX2, SADx4Test, ::testing::ValuesIn(x4d_avx2_tests));
 #endif  // HAVE_AVX2
 
+//------------------------------------------------------------------------------
+// MIPS functions
+#if HAVE_MSA
+const SadMxNFunc sad64x64_msa = vpx_sad64x64_msa;
+const SadMxNFunc sad64x32_msa = vpx_sad64x32_msa;
+const SadMxNFunc sad32x64_msa = vpx_sad32x64_msa;
+const SadMxNFunc sad32x32_msa = vpx_sad32x32_msa;
+const SadMxNFunc sad32x16_msa = vpx_sad32x16_msa;
+const SadMxNFunc sad16x32_msa = vpx_sad16x32_msa;
+const SadMxNFunc sad16x16_msa = vpx_sad16x16_msa;
+const SadMxNFunc sad16x8_msa = vpx_sad16x8_msa;
+const SadMxNFunc sad8x16_msa = vpx_sad8x16_msa;
+const SadMxNFunc sad8x8_msa = vpx_sad8x8_msa;
+const SadMxNFunc sad8x4_msa = vpx_sad8x4_msa;
+const SadMxNFunc sad4x8_msa = vpx_sad4x8_msa;
+const SadMxNFunc sad4x4_msa = vpx_sad4x4_msa;
+const SadMxNParam msa_tests[] = {
+  make_tuple(64, 64, sad64x64_msa, -1),
+  make_tuple(64, 32, sad64x32_msa, -1),
+  make_tuple(32, 64, sad32x64_msa, -1),
+  make_tuple(32, 32, sad32x32_msa, -1),
+  make_tuple(32, 16, sad32x16_msa, -1),
+  make_tuple(16, 32, sad16x32_msa, -1),
+  make_tuple(16, 16, sad16x16_msa, -1),
+  make_tuple(16, 8, sad16x8_msa, -1),
+  make_tuple(8, 16, sad8x16_msa, -1),
+  make_tuple(8, 8, sad8x8_msa, -1),
+  make_tuple(8, 4, sad8x4_msa, -1),
+  make_tuple(4, 8, sad4x8_msa, -1),
+  make_tuple(4, 4, sad4x4_msa, -1),
+};
+INSTANTIATE_TEST_CASE_P(MSA, SADTest, ::testing::ValuesIn(msa_tests));
+
+const SadMxNAvgFunc sad64x64_avg_msa = vpx_sad64x64_avg_msa;
+const SadMxNAvgFunc sad64x32_avg_msa = vpx_sad64x32_avg_msa;
+const SadMxNAvgFunc sad32x64_avg_msa = vpx_sad32x64_avg_msa;
+const SadMxNAvgFunc sad32x32_avg_msa = vpx_sad32x32_avg_msa;
+const SadMxNAvgFunc sad32x16_avg_msa = vpx_sad32x16_avg_msa;
+const SadMxNAvgFunc sad16x32_avg_msa = vpx_sad16x32_avg_msa;
+const SadMxNAvgFunc sad16x16_avg_msa = vpx_sad16x16_avg_msa;
+const SadMxNAvgFunc sad16x8_avg_msa = vpx_sad16x8_avg_msa;
+const SadMxNAvgFunc sad8x16_avg_msa = vpx_sad8x16_avg_msa;
+const SadMxNAvgFunc sad8x8_avg_msa = vpx_sad8x8_avg_msa;
+const SadMxNAvgFunc sad8x4_avg_msa = vpx_sad8x4_avg_msa;
+const SadMxNAvgFunc sad4x8_avg_msa = vpx_sad4x8_avg_msa;
+const SadMxNAvgFunc sad4x4_avg_msa = vpx_sad4x4_avg_msa;
+const SadMxNAvgParam avg_msa_tests[] = {
+  make_tuple(64, 64, sad64x64_avg_msa, -1),
+  make_tuple(64, 32, sad64x32_avg_msa, -1),
+  make_tuple(32, 64, sad32x64_avg_msa, -1),
+  make_tuple(32, 32, sad32x32_avg_msa, -1),
+  make_tuple(32, 16, sad32x16_avg_msa, -1),
+  make_tuple(16, 32, sad16x32_avg_msa, -1),
+  make_tuple(16, 16, sad16x16_avg_msa, -1),
+  make_tuple(16, 8, sad16x8_avg_msa, -1),
+  make_tuple(8, 16, sad8x16_avg_msa, -1),
+  make_tuple(8, 8, sad8x8_avg_msa, -1),
+  make_tuple(8, 4, sad8x4_avg_msa, -1),
+  make_tuple(4, 8, sad4x8_avg_msa, -1),
+  make_tuple(4, 4, sad4x4_avg_msa, -1),
+};
+INSTANTIATE_TEST_CASE_P(MSA, SADavgTest, ::testing::ValuesIn(avg_msa_tests));
+
+const SadMxNx4Func sad64x64x4d_msa = vpx_sad64x64x4d_msa;
+const SadMxNx4Func sad64x32x4d_msa = vpx_sad64x32x4d_msa;
+const SadMxNx4Func sad32x64x4d_msa = vpx_sad32x64x4d_msa;
+const SadMxNx4Func sad32x32x4d_msa = vpx_sad32x32x4d_msa;
+const SadMxNx4Func sad32x16x4d_msa = vpx_sad32x16x4d_msa;
+const SadMxNx4Func sad16x32x4d_msa = vpx_sad16x32x4d_msa;
+const SadMxNx4Func sad16x16x4d_msa = vpx_sad16x16x4d_msa;
+const SadMxNx4Func sad16x8x4d_msa = vpx_sad16x8x4d_msa;
+const SadMxNx4Func sad8x16x4d_msa = vpx_sad8x16x4d_msa;
+const SadMxNx4Func sad8x8x4d_msa = vpx_sad8x8x4d_msa;
+const SadMxNx4Func sad8x4x4d_msa = vpx_sad8x4x4d_msa;
+const SadMxNx4Func sad4x8x4d_msa = vpx_sad4x8x4d_msa;
+const SadMxNx4Func sad4x4x4d_msa = vpx_sad4x4x4d_msa;
+const SadMxNx4Param x4d_msa_tests[] = {
+  make_tuple(64, 64, sad64x64x4d_msa, -1),
+  make_tuple(64, 32, sad64x32x4d_msa, -1),
+  make_tuple(32, 64, sad32x64x4d_msa, -1),
+  make_tuple(32, 32, sad32x32x4d_msa, -1),
+  make_tuple(32, 16, sad32x16x4d_msa, -1),
+  make_tuple(16, 32, sad16x32x4d_msa, -1),
+  make_tuple(16, 16, sad16x16x4d_msa, -1),
+  make_tuple(16, 8, sad16x8x4d_msa, -1),
+  make_tuple(8, 16, sad8x16x4d_msa, -1),
+  make_tuple(8, 8, sad8x8x4d_msa, -1),
+  make_tuple(8, 4, sad8x4x4d_msa, -1),
+  make_tuple(4, 8, sad4x8x4d_msa, -1),
+  make_tuple(4, 4, sad4x4x4d_msa, -1),
+};
+INSTANTIATE_TEST_CASE_P(MSA, SADx4Test, ::testing::ValuesIn(x4d_msa_tests));
+#endif  // HAVE_MSA
+
 }  // namespace
diff --git a/vpx_dsp/mips/macros_msa.h b/vpx_dsp/mips/macros_msa.h
index e36e1b62c..30e152bc1 100644
--- a/vpx_dsp/mips/macros_msa.h
+++ b/vpx_dsp/mips/macros_msa.h
@@ -82,12 +82,24 @@
 }
 #define LD_UB2(...) LD_B2(v16u8, __VA_ARGS__)
 
+#define LD_B3(RTYPE, psrc, stride, out0, out1, out2) {  \
+  LD_B2(RTYPE, (psrc), stride, out0, out1);             \
+  out2 = LD_B(RTYPE, (psrc) + 2 * stride);              \
+}
+#define LD_UB3(...) LD_B3(v16u8, __VA_ARGS__)
+
 #define LD_B4(RTYPE, psrc, stride, out0, out1, out2, out3) {  \
   LD_B2(RTYPE, (psrc), stride, out0, out1);                   \
   LD_B2(RTYPE, (psrc) + 2 * stride , stride, out2, out3);     \
 }
 #define LD_UB4(...) LD_B4(v16u8, __VA_ARGS__)
 
+#define LD_B5(RTYPE, psrc, stride, out0, out1, out2, out3, out4) {  \
+  LD_B4(RTYPE, (psrc), stride, out0, out1, out2, out3);             \
+  out4 = LD_B(RTYPE, (psrc) + 4 * stride);                          \
+}
+#define LD_UB5(...) LD_B5(v16u8, __VA_ARGS__)
+
 /* Description : Load vectors with 8 halfword elements with stride
    Arguments   : Inputs  - psrc, stride
                  Outputs - out0, out1
@@ -105,6 +117,40 @@
 }
 #define LD_SH4(...) LD_H4(v8i16, __VA_ARGS__)
 
+/* Description : average with rounding (in0 + in1 + 1) / 2.
+   Arguments   : Inputs  - in0, in1, in2, in3,
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each unsigned byte element from 'in0' vector is added with
+                 each unsigned byte element from 'in1' vector. Then the average
+                 with rounding is calculated and written to 'out0'
+*/
+#define AVER_UB2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_aver_u_b((v16u8)in0, (v16u8)in1);    \
+  out1 = (RTYPE)__msa_aver_u_b((v16u8)in2, (v16u8)in3);    \
+}
+#define AVER_UB2_UB(...) AVER_UB2(v16u8, __VA_ARGS__)
+
+#define AVER_UB4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                 out0, out1, out2, out3) {                       \
+  AVER_UB2(RTYPE, in0, in1, in2, in3, out0, out1)                \
+  AVER_UB2(RTYPE, in4, in5, in6, in7, out2, out3)                \
+}
+#define AVER_UB4_UB(...) AVER_UB4(v16u8, __VA_ARGS__)
+
+/* Description : Immediate number of elements to slide
+   Arguments   : Inputs  - in0_0, in0_1, in1_0, in1_1, slide_val
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Byte elements from 'in0_0' vector are slid into 'in1_0' by
+                 value specified in the 'slide_val'
+*/
+#define SLDI_B2(RTYPE, in0_0, in0_1, in1_0, in1_1, out0, out1, slide_val) {  \
+  out0 = (RTYPE)__msa_sldi_b((v16i8)in0_0, (v16i8)in1_0, slide_val);         \
+  out1 = (RTYPE)__msa_sldi_b((v16i8)in0_1, (v16i8)in1_1, slide_val);         \
+}
+#define SLDI_B2_UB(...) SLDI_B2(v16u8, __VA_ARGS__)
+
 /* Description : Dot product & addition of halfword vector elements
    Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
                  Outputs - out0, out1
@@ -155,6 +201,26 @@
   sum_m;                                          \
 })
 
+/* Description : Horizontal addition of 8 unsigned halfword elements
+   Arguments   : Inputs  - in       (unsigned halfword vector)
+                 Outputs - sum_m    (u32 sum)
+                 Return Type - unsigned word
+   Details     : 8 unsigned halfword elements of input vector are added
+                 together and the resulting integer sum is returned
+*/
+#define HADD_UH_U32(in) ({                           \
+  v4u32 res_m;                                       \
+  v2u64 res0_m, res1_m;                              \
+  uint32_t sum_m;                                    \
+                                                     \
+  res_m = __msa_hadd_u_w((v8u16)in, (v8u16)in);      \
+  res0_m = __msa_hadd_u_d(res_m, res_m);             \
+  res1_m = (v2u64)__msa_splati_d((v2i64)res0_m, 1);  \
+  res0_m = res0_m + res1_m;                          \
+  sum_m = __msa_copy_u_w((v4i32)res0_m, 0);          \
+  sum_m;                                             \
+})
+
 /* Description : Horizontal subtraction of unsigned byte vector elements
    Arguments   : Inputs  - in0, in1
                  Outputs - out0, out1
@@ -169,6 +235,27 @@
 }
 #define HSUB_UB2_SH(...) HSUB_UB2(v8i16, __VA_ARGS__)
 
+/* Description : SAD (Sum of Absolute Difference)
+   Arguments   : Inputs  - in0, in1, ref0, ref1
+                 Outputs - sad_m                 (halfword vector)
+                 Return Type - unsigned halfword
+   Details     : Absolute difference of all the byte elements from 'in0' with
+                 'ref0' is calculated and preserved in 'diff0'. Then even-odd
+                 pairs are added together to generate 8 halfword results.
+*/
+#define SAD_UB2_UH(in0, in1, ref0, ref1) ({                 \
+  v16u8 diff0_m, diff1_m;                                   \
+  v8u16 sad_m = { 0 };                                      \
+                                                            \
+  diff0_m = __msa_asub_u_b((v16u8)in0, (v16u8)ref0);        \
+  diff1_m = __msa_asub_u_b((v16u8)in1, (v16u8)ref1);        \
+                                                            \
+  sad_m += __msa_hadd_u_h((v16u8)diff0_m, (v16u8)diff0_m);  \
+  sad_m += __msa_hadd_u_h((v16u8)diff1_m, (v16u8)diff1_m);  \
+                                                            \
+  sad_m;                                                    \
+})
+
 /* Description : Set element n input vector to GPR value
    Arguments   : Inputs - in0, in1, in2, in3
                  Output - out
diff --git a/vpx_dsp/mips/sad_msa.c b/vpx_dsp/mips/sad_msa.c
new file mode 100644
index 000000000..19ac47e03
--- /dev/null
+++ b/vpx_dsp/mips/sad_msa.c
@@ -0,0 +1,787 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/mips/macros_msa.h"
+
+static uint32_t sad_4width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v16u8 diff;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad += __msa_hadd_u_h(diff, diff);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_8width_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2,
+                src0, src1, ref0, ref1);
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_16width_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t *ref, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(src, src_stride, src0, src1);
+    src += (2 * src_stride);
+    LD_UB2(ref, ref_stride, ref0, ref1);
+    ref += (2 * ref_stride);
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, src_stride, src0, src1);
+    src += (2 * src_stride);
+    LD_UB2(ref, ref_stride, ref0, ref1);
+    ref += (2 * ref_stride);
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_32width_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t *ref, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_64width_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t *ref, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  uint32_t sad = 0;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+
+  for (ht_cnt = (height >> 1); ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad = HADD_UH_U32(sad0);
+  sad += HADD_UH_U32(sad1);
+
+  return sad;
+}
+
+static void sad_4width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t * const aref_ptr[],
+                               int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v16u8 diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    src_ptr += (4 * src_stride);
+
+    LW4(ref0_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref0_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    LW4(ref1_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref1_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    LW4(ref2_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref2_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+
+    LW4(ref3_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref3_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad3 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_8width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t * const aref_ptr[],
+                               int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
+  v16u8 ref8, ref9, ref10, ref11, ref12, ref13, ref14, ref15;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LD_UB4(ref0_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref0_ptr += (4 * ref_stride);
+    LD_UB4(ref1_ptr, ref_stride, ref4, ref5, ref6, ref7);
+    ref1_ptr += (4 * ref_stride);
+    LD_UB4(ref2_ptr, ref_stride, ref8, ref9, ref10, ref11);
+    ref2_ptr += (4 * ref_stride);
+    LD_UB4(ref3_ptr, ref_stride, ref12, ref13, ref14, ref15);
+    ref3_ptr += (4 * ref_stride);
+
+    PCKEV_D2_UB(src1, src0, src3, src2, src0, src1);
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    PCKEV_D2_UB(ref5, ref4, ref7, ref6, ref0, ref1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    PCKEV_D2_UB(ref9, ref8, ref11, ref10, ref0, ref1);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    PCKEV_D2_UB(ref13, ref12, ref15, ref14, ref0, ref1);
+    sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_16width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                const uint8_t * const aref_ptr[],
+                                int32_t ref_stride,
+                                int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  v16u8 src, ref0, ref1, ref2, ref3, diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = (height >> 1); ht_cnt--;) {
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref0 = LD_UB(ref0_ptr);
+    ref0_ptr += ref_stride;
+    ref1 = LD_UB(ref1_ptr);
+    ref1_ptr += ref_stride;
+    ref2 = LD_UB(ref2_ptr);
+    ref2_ptr += ref_stride;
+    ref3 = LD_UB(ref3_ptr);
+    ref3_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref1);
+    sad1 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref2);
+    sad2 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref3);
+    sad3 += __msa_hadd_u_h(diff, diff);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref0 = LD_UB(ref0_ptr);
+    ref0_ptr += ref_stride;
+    ref1 = LD_UB(ref1_ptr);
+    ref1_ptr += ref_stride;
+    ref2 = LD_UB(ref2_ptr);
+    ref2_ptr += ref_stride;
+    ref3 = LD_UB(ref3_ptr);
+    ref3_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref1);
+    sad1 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref2);
+    sad2 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref3);
+    sad3 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_32width_x4d_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t * const aref_ptr[],
+                                int32_t ref_stride,
+                                int32_t height, uint32_t *sad_array) {
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+
+    LD_UB2(ref0_ptr, 16, ref0, ref1);
+    ref0_ptr += ref_stride;
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(ref1_ptr, 16, ref0, ref1);
+    ref1_ptr += ref_stride;
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(ref2_ptr, 16, ref0, ref1);
+    ref2_ptr += ref_stride;
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(ref3_ptr, 16, ref0, ref1);
+    ref3_ptr += ref_stride;
+    sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_64width_x4d_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t * const aref_ptr[],
+                                int32_t ref_stride,
+                                int32_t height, uint32_t *sad_array) {
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8u16 sad0_0 = { 0 };
+  v8u16 sad0_1 = { 0 };
+  v8u16 sad1_0 = { 0 };
+  v8u16 sad1_1 = { 0 };
+  v8u16 sad2_0 = { 0 };
+  v8u16 sad2_1 = { 0 };
+  v8u16 sad3_0 = { 0 };
+  v8u16 sad3_1 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+
+    LD_UB4(ref0_ptr, 16, ref0, ref1, ref2, ref3);
+    ref0_ptr += ref_stride;
+    sad0_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad0_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(ref1_ptr, 16, ref0, ref1, ref2, ref3);
+    ref1_ptr += ref_stride;
+    sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(ref2_ptr, 16, ref0, ref1, ref2, ref3);
+    ref2_ptr += ref_stride;
+    sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(ref3_ptr, 16, ref0, ref1, ref2, ref3);
+    ref3_ptr += ref_stride;
+    sad3_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad3_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0_0);
+  sad_array[0] += HADD_UH_U32(sad0_1);
+  sad_array[1] = HADD_UH_U32(sad1_0);
+  sad_array[1] += HADD_UH_U32(sad1_1);
+  sad_array[2] = HADD_UH_U32(sad2_0);
+  sad_array[2] += HADD_UH_U32(sad2_1);
+  sad_array[3] = HADD_UH_U32(sad3_0);
+  sad_array[3] += HADD_UH_U32(sad3_1);
+}
+
+static uint32_t avgsad_4width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                  const uint8_t *ref_ptr, int32_t ref_stride,
+                                  int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v16u8 diff, pred, comp;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+
+    comp = __msa_aver_u_b(pred, ref);
+    diff = __msa_asub_u_b(src, comp);
+    sad += __msa_hadd_u_h(diff, diff);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_8width_msa(const uint8_t *src, int32_t src_stride,
+                                  const uint8_t *ref, int32_t ref_stride,
+                                  int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 diff0, diff1, pred0, pred1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2,
+                src0, src1, ref0, ref1);
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, diff0, diff1);
+    sad += SAD_UB2_UH(src0, src1, diff0, diff1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_16width_msa(const uint8_t *src, int32_t src_stride,
+                                   const uint8_t *ref, int32_t ref_stride,
+                                   int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, pred2, pred3, comp0, comp1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 3); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * 16);
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1);
+    sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+    AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1);
+    sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * 16);
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1);
+    sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+    AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1);
+    sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_32width_msa(const uint8_t *src, int32_t src_stride,
+                                   const uint8_t *ref, int32_t ref_stride,
+                                   int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
+  v16u8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7;
+  v16u8 comp0, comp1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src2, src4, src6);
+    LD_UB4(src + 16, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+
+    LD_UB4(ref, ref_stride, ref0, ref2, ref4, ref6);
+    LD_UB4(ref + 16, ref_stride, ref1, ref3, ref5, ref7);
+    ref += (4 * ref_stride);
+
+    LD_UB4(sec_pred, 32, pred0, pred2, pred4, pred6);
+    LD_UB4(sec_pred + 16, 32, pred1, pred3, pred5, pred7);
+    sec_pred += (4 * 32);
+
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1);
+    sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+    AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1);
+    sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+    AVER_UB2_UB(pred4, ref4, pred5, ref5, comp0, comp1);
+    sad += SAD_UB2_UH(src4, src5, comp0, comp1);
+    AVER_UB2_UB(pred6, ref6, pred7, ref7, comp0, comp1);
+    sad += SAD_UB2_UH(src6, src7, comp0, comp1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_64width_msa(const uint8_t *src, int32_t src_stride,
+                                   const uint8_t *ref, int32_t ref_stride,
+                                   int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 comp0, comp1, comp2, comp3;
+  v16u8 pred0, pred1, pred2, pred3;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v4u32 sad;
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3,
+                comp0, comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3,
+                comp0, comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3,
+                comp0, comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3,
+                comp0, comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+  }
+
+  sad = __msa_hadd_u_w(sad0, sad0);
+  sad += __msa_hadd_u_w(sad1, sad1);
+
+  return HADD_SW_S32(sad);
+}
+
+#define VPX_SAD_4xHEIGHT_MSA(height)                                        \
+uint32_t vpx_sad4x##height##_msa(const uint8_t *src, int32_t src_stride,    \
+                                 const uint8_t *ref, int32_t ref_stride) {  \
+  return sad_4width_msa(src, src_stride,  ref, ref_stride, height);         \
+}
+
+#define VPX_SAD_8xHEIGHT_MSA(height)                                        \
+uint32_t vpx_sad8x##height##_msa(const uint8_t *src, int32_t src_stride,    \
+                                 const uint8_t *ref, int32_t ref_stride) {  \
+  return sad_8width_msa(src, src_stride, ref, ref_stride, height);          \
+}
+
+#define VPX_SAD_16xHEIGHT_MSA(height)                                        \
+uint32_t vpx_sad16x##height##_msa(const uint8_t *src, int32_t src_stride,    \
+                                  const uint8_t *ref, int32_t ref_stride) {  \
+  return sad_16width_msa(src, src_stride, ref, ref_stride, height);          \
+}
+
+#define VPX_SAD_32xHEIGHT_MSA(height)                                        \
+uint32_t vpx_sad32x##height##_msa(const uint8_t *src, int32_t src_stride,    \
+                                  const uint8_t *ref, int32_t ref_stride) {  \
+  return sad_32width_msa(src, src_stride, ref, ref_stride, height);          \
+}
+
+#define VPX_SAD_64xHEIGHT_MSA(height)                                        \
+uint32_t vpx_sad64x##height##_msa(const uint8_t *src, int32_t src_stride,    \
+                                  const uint8_t *ref, int32_t ref_stride) {  \
+  return sad_64width_msa(src, src_stride, ref, ref_stride, height);          \
+}
+
+#define VPX_SAD_4xHEIGHTx4D_MSA(height)                                  \
+void vpx_sad4x##height##x4d_msa(const uint8_t *src, int32_t src_stride,  \
+                                const uint8_t *const refs[],             \
+                                int32_t ref_stride, uint32_t *sads) {    \
+  sad_4width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);   \
+}
+
+#define VPX_SAD_8xHEIGHTx4D_MSA(height)                                  \
+void vpx_sad8x##height##x4d_msa(const uint8_t *src, int32_t src_stride,  \
+                                const uint8_t *const refs[],             \
+                                int32_t ref_stride, uint32_t *sads) {    \
+  sad_8width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);   \
+}
+
+#define VPX_SAD_16xHEIGHTx4D_MSA(height)                                  \
+void vpx_sad16x##height##x4d_msa(const uint8_t *src, int32_t src_stride,  \
+                                 const uint8_t *const refs[],             \
+                                 int32_t ref_stride, uint32_t *sads) {    \
+  sad_16width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);   \
+}
+
+#define VPX_SAD_32xHEIGHTx4D_MSA(height)                                  \
+void vpx_sad32x##height##x4d_msa(const uint8_t *src, int32_t src_stride,  \
+                                 const uint8_t *const refs[],             \
+                                 int32_t ref_stride, uint32_t *sads) {    \
+  sad_32width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);   \
+}
+
+#define VPX_SAD_64xHEIGHTx4D_MSA(height)                                  \
+void vpx_sad64x##height##x4d_msa(const uint8_t *src, int32_t src_stride,  \
+                                 const uint8_t *const refs[],             \
+                                 int32_t ref_stride, uint32_t *sads) {    \
+  sad_64width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);   \
+}
+
+#define VPX_AVGSAD_4xHEIGHT_MSA(height)                                       \
+uint32_t vpx_sad4x##height##_avg_msa(const uint8_t *src, int32_t src_stride,  \
+                                     const uint8_t *ref, int32_t ref_stride,  \
+                                     const uint8_t *second_pred) {            \
+  return avgsad_4width_msa(src, src_stride, ref, ref_stride,                  \
+                           height, second_pred);                              \
+}
+
+#define VPX_AVGSAD_8xHEIGHT_MSA(height)                                       \
+uint32_t vpx_sad8x##height##_avg_msa(const uint8_t *src, int32_t src_stride,  \
+                                     const uint8_t *ref, int32_t ref_stride,  \
+                                     const uint8_t *second_pred) {            \
+  return avgsad_8width_msa(src, src_stride, ref, ref_stride,                  \
+                           height, second_pred);                              \
+}
+
+#define VPX_AVGSAD_16xHEIGHT_MSA(height)                                       \
+uint32_t vpx_sad16x##height##_avg_msa(const uint8_t *src, int32_t src_stride,  \
+                                      const uint8_t *ref, int32_t ref_stride,  \
+                                      const uint8_t *second_pred) {            \
+  return avgsad_16width_msa(src, src_stride, ref, ref_stride,                  \
+                            height, second_pred);                              \
+}
+
+#define VPX_AVGSAD_32xHEIGHT_MSA(height)                                       \
+uint32_t vpx_sad32x##height##_avg_msa(const uint8_t *src, int32_t src_stride,  \
+                                      const uint8_t *ref, int32_t ref_stride,  \
+                                      const uint8_t *second_pred) {            \
+  return avgsad_32width_msa(src, src_stride, ref, ref_stride,                  \
+                            height, second_pred);                              \
+}
+
+#define VPX_AVGSAD_64xHEIGHT_MSA(height)                                       \
+uint32_t vpx_sad64x##height##_avg_msa(const uint8_t *src, int32_t src_stride,  \
+                                      const uint8_t *ref, int32_t ref_stride,  \
+                                      const uint8_t *second_pred) {            \
+  return avgsad_64width_msa(src, src_stride, ref, ref_stride,                  \
+                            height, second_pred);                              \
+}
+
+// 64x64
+VPX_SAD_64xHEIGHT_MSA(64);
+VPX_SAD_64xHEIGHTx4D_MSA(64);
+VPX_AVGSAD_64xHEIGHT_MSA(64);
+
+// 64x32
+VPX_SAD_64xHEIGHT_MSA(32);
+VPX_SAD_64xHEIGHTx4D_MSA(32);
+VPX_AVGSAD_64xHEIGHT_MSA(32);
+
+// 32x64
+VPX_SAD_32xHEIGHT_MSA(64);
+VPX_SAD_32xHEIGHTx4D_MSA(64);
+VPX_AVGSAD_32xHEIGHT_MSA(64);
+
+// 32x32
+VPX_SAD_32xHEIGHT_MSA(32);
+VPX_SAD_32xHEIGHTx4D_MSA(32);
+VPX_AVGSAD_32xHEIGHT_MSA(32);
+
+// 32x16
+VPX_SAD_32xHEIGHT_MSA(16);
+VPX_SAD_32xHEIGHTx4D_MSA(16);
+VPX_AVGSAD_32xHEIGHT_MSA(16);
+
+// 16x32
+VPX_SAD_16xHEIGHT_MSA(32);
+VPX_SAD_16xHEIGHTx4D_MSA(32);
+VPX_AVGSAD_16xHEIGHT_MSA(32);
+
+// 16x16
+VPX_SAD_16xHEIGHT_MSA(16);
+VPX_SAD_16xHEIGHTx4D_MSA(16);
+VPX_AVGSAD_16xHEIGHT_MSA(16);
+
+// 16x8
+VPX_SAD_16xHEIGHT_MSA(8);
+VPX_SAD_16xHEIGHTx4D_MSA(8);
+VPX_AVGSAD_16xHEIGHT_MSA(8);
+
+// 8x16
+VPX_SAD_8xHEIGHT_MSA(16);
+VPX_SAD_8xHEIGHTx4D_MSA(16);
+VPX_AVGSAD_8xHEIGHT_MSA(16);
+
+// 8x8
+VPX_SAD_8xHEIGHT_MSA(8);
+VPX_SAD_8xHEIGHTx4D_MSA(8);
+VPX_AVGSAD_8xHEIGHT_MSA(8);
+
+// 8x4
+VPX_SAD_8xHEIGHT_MSA(4);
+VPX_SAD_8xHEIGHTx4D_MSA(4);
+VPX_AVGSAD_8xHEIGHT_MSA(4);
+
+// 4x8
+VPX_SAD_4xHEIGHT_MSA(8);
+VPX_SAD_4xHEIGHTx4D_MSA(8);
+VPX_AVGSAD_4xHEIGHT_MSA(8);
+
+// 4x4
+VPX_SAD_4xHEIGHT_MSA(4);
+VPX_SAD_4xHEIGHTx4D_MSA(4);
+VPX_AVGSAD_4xHEIGHT_MSA(4);
diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index 90b0f9c93..9987f15be 100644
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -28,6 +28,7 @@ DSP_SRCS-$(HAVE_AVX2)   += x86/sad4d_avx2.c
 DSP_SRCS-$(HAVE_AVX2)   += x86/sad_avx2.c
 
 DSP_SRCS-$(HAVE_MSA)    += mips/macros_msa.h
+DSP_SRCS-$(HAVE_MSA)    += mips/sad_msa.c
 
 ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
 DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad4d_sse2.asm
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 934520115..f12270c1c 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -40,85 +40,85 @@ if (vpx_config("CONFIG_ENCODERS") eq "yes") {
 # Single block SAD
 #
 add_proto qw/unsigned int vpx_sad64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad64x64 avx2 neon/, "$sse2_x86inc";
+specialize qw/vpx_sad64x64 avx2 neon msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad64x32 avx2/, "$sse2_x86inc";
+specialize qw/vpx_sad64x32 avx2 msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad32x64 avx2/, "$sse2_x86inc";
+specialize qw/vpx_sad32x64 avx2 msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad32x32 avx2 neon/, "$sse2_x86inc";
+specialize qw/vpx_sad32x32 avx2 neon msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad32x16 avx2/, "$sse2_x86inc";
+specialize qw/vpx_sad32x16 avx2 msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad16x32/, "$sse2_x86inc";
+specialize qw/vpx_sad16x32 msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad16x16 mmx media neon/, "$sse2_x86inc";
+specialize qw/vpx_sad16x16 mmx media neon msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad16x8 mmx neon/, "$sse2_x86inc";
+specialize qw/vpx_sad16x8 mmx neon msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad8x16 mmx neon/, "$sse2_x86inc";
+specialize qw/vpx_sad8x16 mmx neon msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad8x8 mmx neon/, "$sse2_x86inc";
+specialize qw/vpx_sad8x8 mmx neon msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad8x4/, "$sse2_x86inc";
+specialize qw/vpx_sad8x4 msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad4x8/, "$sse_x86inc";
+specialize qw/vpx_sad4x8 msa/, "$sse_x86inc";
 
 add_proto qw/unsigned int vpx_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad4x4 mmx neon/, "$sse_x86inc";
+specialize qw/vpx_sad4x4 mmx neon msa/, "$sse_x86inc";
 
 #
 # Avg
 #
 add_proto qw/unsigned int vpx_sad64x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad64x64_avg avx2/, "$sse2_x86inc";
+specialize qw/vpx_sad64x64_avg avx2 msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad64x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad64x32_avg avx2/, "$sse2_x86inc";
+specialize qw/vpx_sad64x32_avg avx2 msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad32x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad32x64_avg avx2/, "$sse2_x86inc";
+specialize qw/vpx_sad32x64_avg avx2 msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad32x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad32x32_avg avx2/, "$sse2_x86inc";
+specialize qw/vpx_sad32x32_avg avx2 msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad32x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad32x16_avg avx2/, "$sse2_x86inc";
+specialize qw/vpx_sad32x16_avg avx2 msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad16x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad16x32_avg/, "$sse2_x86inc";
+specialize qw/vpx_sad16x32_avg msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad16x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad16x16_avg/, "$sse2_x86inc";
+specialize qw/vpx_sad16x16_avg msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad16x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad16x8_avg/, "$sse2_x86inc";
+specialize qw/vpx_sad16x8_avg msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad8x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad8x16_avg/, "$sse2_x86inc";
+specialize qw/vpx_sad8x16_avg msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad8x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad8x8_avg/, "$sse2_x86inc";
+specialize qw/vpx_sad8x8_avg msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad8x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad8x4_avg/, "$sse2_x86inc";
+specialize qw/vpx_sad8x4_avg msa/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vpx_sad4x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad4x8_avg/, "$sse_x86inc";
+specialize qw/vpx_sad4x8_avg msa/, "$sse_x86inc";
 
 add_proto qw/unsigned int vpx_sad4x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad4x4_avg/, "$sse_x86inc";
+specialize qw/vpx_sad4x4_avg msa/, "$sse_x86inc";
 
 #
 # Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
@@ -171,43 +171,43 @@ specialize qw/vpx_sad4x4x8 sse4_1/;
 # 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/, "$sse2_x86inc";
+specialize qw/vpx_sad64x64x4d avx2 neon msa/, "$sse2_x86inc";
 
 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/, "$sse2_x86inc";
+specialize qw/vpx_sad64x32x4d msa/, "$sse2_x86inc";
 
 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/, "$sse2_x86inc";
+specialize qw/vpx_sad32x64x4d msa/, "$sse2_x86inc";
 
 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/, "$sse2_x86inc";
+specialize qw/vpx_sad32x32x4d avx2 neon msa/, "$sse2_x86inc";
 
 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/, "$sse2_x86inc";
+specialize qw/vpx_sad32x16x4d msa/, "$sse2_x86inc";
 
 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/, "$sse2_x86inc";
+specialize qw/vpx_sad16x32x4d msa/, "$sse2_x86inc";
 
 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/, "$sse2_x86inc";
+specialize qw/vpx_sad16x16x4d neon msa/, "$sse2_x86inc";
 
 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/, "$sse2_x86inc";
+specialize qw/vpx_sad16x8x4d msa/, "$sse2_x86inc";
 
 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/, "$sse2_x86inc";
+specialize qw/vpx_sad8x16x4d msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_sad8x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad8x8x4d/, "$sse2_x86inc";
+specialize qw/vpx_sad8x8x4d msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_sad8x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad8x4x4d/, "$sse2_x86inc";
+specialize qw/vpx_sad8x4x4d msa/, "$sse2_x86inc";
 
 add_proto qw/void vpx_sad4x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad4x8x4d/, "$sse_x86inc";
+specialize qw/vpx_sad4x8x4d msa/, "$sse_x86inc";
 
 add_proto qw/void vpx_sad4x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad4x4x4d/, "$sse_x86inc";
+specialize qw/vpx_sad4x4x4d msa/, "$sse_x86inc";
 
 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   #