Merge "mips msa vpx_dsp sad sad4d avgsad optimization"

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

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

vpx_dsp/mips/sad_msa.c

+/*
+ *  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 };