Add MSA optimized ScaleFilterCols, ScaleARGBCols, ScaleARGBFilterCols and ScaleRowDown34 functions

TBR=kjellander@chromium.org
R=fbarchard@google.com

Bug:libyuv:634
Change-Id: Ib139b9701fc67e24d27a6886377c0cb8b2773fda
Reviewed-on: https://chromium-review.googlesource.com/620791Reviewed-by: Frank Barchard <fbarchard@google.com>

include/libyuv/scale_row.h

@@ -114,9 +114,14 @@ extern "C" {
 #define HAS_SCALEROWDOWN2_MSA
 #define HAS_SCALEROWDOWN4_MSA
 #define HAS_SCALEADDROW_MSA
+#define HAS_SCALEARGBCOLS_MSA
+#define HAS_SCALEROWDOWN34_MSA
+
 #ifndef DISABLE_CLANG_MSA
 #define HAS_SCALEARGBROWDOWNEVEN_MSA
 #define HAS_SCALEROWDOWN38_MSA
+#define HAS_SCALEFILTERCOLS_MSA
+#define HAS_SCALEARGBFILTERCOLS_MSA
 #endif
 #endif
 
@@ -553,6 +558,26 @@ void ScaleARGBCols_Any_NEON(uint8* dst_argb,
                             int dst_width,
                             int x,
                             int dx);
+void ScaleARGBFilterCols_MSA(uint8* dst_argb,
+                             const uint8* src_argb,
+                             int dst_width,
+                             int x,
+                             int dx);
+void ScaleARGBCols_MSA(uint8* dst_argb,
+                       const uint8* src_argb,
+                       int dst_width,
+                       int x,
+                       int dx);
+void ScaleARGBFilterCols_Any_MSA(uint8* dst_argb,
+                                 const uint8* src_argb,
+                                 int dst_width,
+                                 int x,
+                                 int dx);
+void ScaleARGBCols_Any_MSA(uint8* dst_argb,
+                           const uint8* src_argb,
+                           int dst_width,
+                           int x,
+                           int dx);
 
 // ARGB Row functions
 void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
@@ -892,6 +917,24 @@ void ScaleRowDown38_3_Box_MSA(const uint8_t* src_ptr,
                               uint8_t* dst_ptr,
                               int dst_width);
 void ScaleAddRow_MSA(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width);
+void ScaleFilterCols_MSA(uint8* dst_ptr,
+                         const uint8* src_ptr,
+                         int dst_width,
+                         int x,
+                         int dx);
+void ScaleRowDown34_MSA(const uint8* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8* dst_ptr,
+                        int dst_width);
+void ScaleRowDown34_0_Box_MSA(const uint8* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8* dst_ptr,
+                              int dst_width);
+void ScaleRowDown34_1_Box_MSA(const uint8* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8* dst_ptr,
+                              int dst_width);
+
 void ScaleRowDown2_Any_MSA(const uint8_t* src_ptr,
                            ptrdiff_t src_stride,
                            uint8_t* dst,
@@ -927,6 +970,23 @@ void ScaleRowDown38_3_Box_Any_MSA(const uint8_t* src_ptr,
 void ScaleAddRow_Any_MSA(const uint8_t* src_ptr,
                          uint16_t* dst_ptr,
                          int src_width);
+void ScaleFilterCols_Any_MSA(uint8* dst_ptr,
+                             const uint8* src_ptr,
+                             int dst_width,
+                             int x,
+                             int dx);
+void ScaleRowDown34_Any_MSA(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst_ptr,
+                            int dst_width);
+void ScaleRowDown34_0_Box_Any_MSA(const uint8* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_ptr,
+                                  int dst_width);
+void ScaleRowDown34_1_Box_Any_MSA(const uint8* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_ptr,
+                                  int dst_width);
 
 #ifdef __cplusplus
 }  // extern "C"

source/scale.cc

@@ -371,6 +371,26 @@ static void ScalePlaneDown34(int src_width,
     }
   }
 #endif
+#if defined(HAS_SCALEROWDOWN34_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_Any_MSA;
+      ScaleRowDown34_1 = ScaleRowDown34_Any_MSA;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_MSA;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_MSA;
+    }
+    if (dst_width % 48 == 0) {
+      if (!filtering) {
+        ScaleRowDown34_0 = ScaleRowDown34_MSA;
+        ScaleRowDown34_1 = ScaleRowDown34_MSA;
+      } else {
+        ScaleRowDown34_0 = ScaleRowDown34_0_Box_MSA;
+        ScaleRowDown34_1 = ScaleRowDown34_1_Box_MSA;
+      }
+    }
+  }
+#endif
 #if defined(HAS_SCALEROWDOWN34_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     if (!filtering) {
@@ -1078,6 +1098,14 @@ void ScalePlaneBilinearDown(int src_width,
       ScaleFilterCols = ScaleFilterCols_NEON;
     }
   }
+#endif
+#if defined(HAS_SCALEFILTERCOLS_MSA)
+  if (TestCpuFlag(kCpuHasMSA) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleFilterCols = ScaleFilterCols_MSA;
+    }
+  }
 #endif
   if (y > max_y) {
     y = max_y;
@@ -1276,6 +1304,14 @@ void ScalePlaneBilinearUp(int src_width,
       ScaleFilterCols = ScaleFilterCols_NEON;
     }
   }
+#endif
+#if defined(HAS_SCALEFILTERCOLS_MSA)
+  if (filtering && TestCpuFlag(kCpuHasMSA) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleFilterCols = ScaleFilterCols_MSA;
+    }
+  }
 #endif
   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
     ScaleFilterCols = ScaleColsUp2_C;

source/scale_any.cc

@@ -33,9 +33,15 @@ extern "C" {
 #ifdef HAS_SCALEFILTERCOLS_NEON
 CANY(ScaleFilterCols_Any_NEON, ScaleFilterCols_NEON, ScaleFilterCols_C, 1, 7)
 #endif
+#ifdef HAS_SCALEFILTERCOLS_MSA
+CANY(ScaleFilterCols_Any_MSA, ScaleFilterCols_MSA, ScaleFilterCols_C, 1, 15)
+#endif
 #ifdef HAS_SCALEARGBCOLS_NEON
 CANY(ScaleARGBCols_Any_NEON, ScaleARGBCols_NEON, ScaleARGBCols_C, 4, 7)
 #endif
+#ifdef HAS_SCALEARGBCOLS_MSA
+CANY(ScaleARGBCols_Any_MSA, ScaleARGBCols_MSA, ScaleARGBCols_C, 4, 3)
+#endif
 #ifdef HAS_SCALEARGBFILTERCOLS_NEON
 CANY(ScaleARGBFilterCols_Any_NEON,
      ScaleARGBFilterCols_NEON,
@@ -43,6 +49,13 @@ CANY(ScaleARGBFilterCols_Any_NEON,
      4,
      3)
 #endif
+#ifdef HAS_SCALEARGBFILTERCOLS_MSA
+CANY(ScaleARGBFilterCols_Any_MSA,
+     ScaleARGBFilterCols_MSA,
+     ScaleARGBFilterCols_C,
+     4,
+     7)
+#endif
 #undef CANY
 
 // Fixed scale down.
@@ -228,6 +241,26 @@ SDANY(ScaleRowDown34_1_Box_Any_NEON,
       1,
       23)
 #endif
+#ifdef HAS_SCALEROWDOWN34_MSA
+SDANY(ScaleRowDown34_Any_MSA,
+      ScaleRowDown34_MSA,
+      ScaleRowDown34_C,
+      4 / 3,
+      1,
+      47)
+SDANY(ScaleRowDown34_0_Box_Any_MSA,
+      ScaleRowDown34_0_Box_MSA,
+      ScaleRowDown34_0_Box_C,
+      4 / 3,
+      1,
+      47)
+SDANY(ScaleRowDown34_1_Box_Any_MSA,
+      ScaleRowDown34_1_Box_MSA,
+      ScaleRowDown34_1_Box_C,
+      4 / 3,
+      1,
+      47)
+#endif
 #ifdef HAS_SCALEROWDOWN38_SSSE3
 SDANY(ScaleRowDown38_Any_SSSE3,
       ScaleRowDown38_SSSE3,

source/scale_argb.cc

@@ -335,6 +335,14 @@ static void ScaleARGBBilinearDown(int src_width,
       ScaleARGBFilterCols = ScaleARGBFilterCols_NEON;
     }
   }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_MSA;
+    }
+  }
 #endif
   // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
   // Allocate a row of ARGB.
@@ -442,6 +450,14 @@ static void ScaleARGBBilinearUp(int src_width,
     }
   }
 #endif
+#if defined(HAS_SCALEARGBFILTERCOLS_MSA)
+  if (filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_MSA;
+    }
+  }
+#endif
 #if defined(HAS_SCALEARGBCOLS_SSE2)
   if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
     ScaleARGBFilterCols = ScaleARGBCols_SSE2;
@@ -454,6 +470,14 @@ static void ScaleARGBBilinearUp(int src_width,
       ScaleARGBFilterCols = ScaleARGBCols_NEON;
     }
   }
+#endif
+#if defined(HAS_SCALEARGBCOLS_MSA)
+  if (!filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBCols_MSA;
+    }
+  }
 #endif
   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
     ScaleARGBFilterCols = ScaleARGBColsUp2_C;
@@ -643,6 +667,14 @@ static void ScaleYUVToARGBBilinearUp(int src_width,
     }
   }
 #endif
+#if defined(HAS_SCALEARGBFILTERCOLS_MSA)
+  if (filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_MSA;
+    }
+  }
+#endif
 #if defined(HAS_SCALEARGBCOLS_SSE2)
   if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
     ScaleARGBFilterCols = ScaleARGBCols_SSE2;
@@ -655,6 +687,14 @@ static void ScaleYUVToARGBBilinearUp(int src_width,
       ScaleARGBFilterCols = ScaleARGBCols_NEON;
     }
   }
+#endif
+#if defined(HAS_SCALEARGBCOLS_MSA)
+  if (!filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBCols_MSA;
+    }
+  }
 #endif
   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
     ScaleARGBFilterCols = ScaleARGBColsUp2_C;
@@ -778,6 +818,14 @@ static void ScaleARGBSimple(int src_width,
       ScaleARGBCols = ScaleARGBCols_NEON;
     }
   }
+#endif
+#if defined(HAS_SCALEARGBCOLS_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleARGBCols = ScaleARGBCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBCols = ScaleARGBCols_MSA;
+    }
+  }
 #endif
   if (src_width * 2 == dst_width && x < 0x8000) {
     ScaleARGBCols = ScaleARGBColsUp2_C;

source/scale_msa.cc

@@ -21,6 +21,14 @@ namespace libyuv {
 extern "C" {
 #endif
 
+#define LOAD_INDEXED_DATA(srcp, indx0, out0) \
+  {                                          \
+    out0[0] = srcp[indx0[0]];                \
+    out0[1] = srcp[indx0[1]];                \
+    out0[2] = srcp[indx0[2]];                \
+    out0[3] = srcp[indx0[3]];                \
+  }
+
 void ScaleARGBRowDown2_MSA(const uint8_t* src_argb,
                            ptrdiff_t src_stride,
                            uint8_t* dst_argb,
@@ -549,6 +557,398 @@ void ScaleAddRow_MSA(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width) {
   }
 }
 
+#ifndef DISABLE_CLANG_MSA
+void ScaleFilterCols_MSA(uint8* dst_ptr,
+                         const uint8* src_ptr,
+                         int dst_width,
+                         int x,
+                         int dx) {
+  int j;
+  v4i32 vec_x = __msa_fill_w(x);
+  v4i32 vec_dx = __msa_fill_w(dx);
+  v4i32 vec_const = {0, 1, 2, 3};
+  v4i32 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9;
+  v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8u16 reg0, reg1;
+  v16u8 dst0;
+  v4i32 const_0xFFFF = __msa_fill_w(0xFFFF);
+  v4i32 const_0x40 = __msa_fill_w(0x40);
+
+  vec0 = vec_dx * vec_const;
+  vec1 = vec_dx * 4;
+  vec_x += vec0;
+
+  for (j = 0; j < dst_width - 1; j += 16) {
+    vec2 = vec_x >> 16;
+    vec6 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec3 = vec_x >> 16;
+    vec7 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec4 = vec_x >> 16;
+    vec8 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec5 = vec_x >> 16;
+    vec9 = vec_x & const_0xFFFF;
+    vec_x += vec1;
+    vec6 >>= 9;
+    vec7 >>= 9;
+    vec8 >>= 9;
+    vec9 >>= 9;
+    LOAD_INDEXED_DATA(src_ptr, vec2, tmp0);
+    LOAD_INDEXED_DATA(src_ptr, vec3, tmp1);
+    LOAD_INDEXED_DATA(src_ptr, vec4, tmp2);
+    LOAD_INDEXED_DATA(src_ptr, vec5, tmp3);
+    vec2 += 1;
+    vec3 += 1;
+    vec4 += 1;
+    vec5 += 1;
+    LOAD_INDEXED_DATA(src_ptr, vec2, tmp4);
+    LOAD_INDEXED_DATA(src_ptr, vec3, tmp5);
+    LOAD_INDEXED_DATA(src_ptr, vec4, tmp6);
+    LOAD_INDEXED_DATA(src_ptr, vec5, tmp7);
+    tmp4 -= tmp0;
+    tmp5 -= tmp1;
+    tmp6 -= tmp2;
+    tmp7 -= tmp3;
+    tmp4 *= vec6;
+    tmp5 *= vec7;
+    tmp6 *= vec8;
+    tmp7 *= vec9;
+    tmp4 += const_0x40;
+    tmp5 += const_0x40;
+    tmp6 += const_0x40;
+    tmp7 += const_0x40;
+    tmp4 >>= 7;
+    tmp5 >>= 7;
+    tmp6 >>= 7;
+    tmp7 >>= 7;
+    tmp0 += tmp4;
+    tmp1 += tmp5;
+    tmp2 += tmp6;
+    tmp3 += tmp7;
+    reg0 = (v8u16)__msa_pckev_h((v8i16)tmp1, (v8i16)tmp0);
+    reg1 = (v8u16)__msa_pckev_h((v8i16)tmp3, (v8i16)tmp2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    __msa_st_b(dst0, dst_ptr, 0);
+    dst_ptr += 16;
+  }
+}
+#endif
+
+void ScaleARGBCols_MSA(uint8* dst_argb,
+                       const uint8* src_argb,
+                       int dst_width,
+                       int x,
+                       int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  v4i32 x_vec = __msa_fill_w(x);
+  v4i32 dx_vec = __msa_fill_w(dx);
+  v4i32 const_vec = {0, 1, 2, 3};
+  v4i32 vec0, vec1, vec2;
+  v4i32 dst0;
+
+  vec0 = dx_vec * const_vec;
+  vec1 = dx_vec * 4;
+  x_vec += vec0;
+
+  for (j = 0; j < dst_width; j += 4) {
+    vec2 = x_vec >> 16;
+    x_vec += vec1;
+    LOAD_INDEXED_DATA(src, vec2, dst0);
+    __msa_st_w(dst0, dst, 0);
+    dst += 4;
+  }
+}
+
+#ifndef DISABLE_CLANG_MSA
+void ScaleARGBFilterCols_MSA(uint8* dst_argb,
+                             const uint8* src_argb,
+                             int dst_width,
+                             int x,
+                             int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  int j;
+  v4u32 src0, src1, src2, src3;
+  v4u32 vec0, vec1, vec2, vec3;
+  v16u8 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v16u8 mult0, mult1, mult2, mult3;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v16u8 dst0, dst1;
+  v4u32 vec_x = (v4u32)__msa_fill_w(x);
+  v4u32 vec_dx = (v4u32)__msa_fill_w(dx);
+  v4u32 vec_const = {0, 1, 2, 3};
+  v16u8 const_0x7f = (v16u8)__msa_fill_b(0x7f);
+
+  vec0 = vec_dx * vec_const;
+  vec1 = vec_dx * 4;
+  vec_x += vec0;
+
+  for (j = 0; j < dst_width - 1; j += 8) {
+    vec2 = vec_x >> 16;
+    reg0 = (v16u8)(vec_x >> 9);
+    vec_x += vec1;
+    vec3 = vec_x >> 16;
+    reg1 = (v16u8)(vec_x >> 9);
+    vec_x += vec1;
+    reg0 = reg0 & const_0x7f;
+    reg1 = reg1 & const_0x7f;
+    reg0 = (v16u8)__msa_shf_b((v16i8)reg0, 0);
+    reg1 = (v16u8)__msa_shf_b((v16i8)reg1, 0);
+    reg2 = reg0 ^ const_0x7f;
+    reg3 = reg1 ^ const_0x7f;
+    mult0 = (v16u8)__msa_ilvr_b((v16i8)reg0, (v16i8)reg2);
+    mult1 = (v16u8)__msa_ilvl_b((v16i8)reg0, (v16i8)reg2);
+    mult2 = (v16u8)__msa_ilvr_b((v16i8)reg1, (v16i8)reg3);
+    mult3 = (v16u8)__msa_ilvl_b((v16i8)reg1, (v16i8)reg3);
+    LOAD_INDEXED_DATA(src, vec2, src0);
+    LOAD_INDEXED_DATA(src, vec3, src1);
+    vec2 += 1;
+    vec3 += 1;
+    LOAD_INDEXED_DATA(src, vec2, src2);
+    LOAD_INDEXED_DATA(src, vec3, src3);
+    reg4 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    reg5 = (v16u8)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    reg6 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    reg7 = (v16u8)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    tmp0 = __msa_dotp_u_h(reg4, mult0);
+    tmp1 = __msa_dotp_u_h(reg5, mult1);
+    tmp2 = __msa_dotp_u_h(reg6, mult2);
+    tmp3 = __msa_dotp_u_h(reg7, mult3);
+    tmp0 >>= 7;
+    tmp1 >>= 7;
+    tmp2 >>= 7;
+    tmp3 >>= 7;
+    dst0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+    __msa_st_b(dst0, dst_argb, 0);
+    __msa_st_b(dst1, dst_argb, 16);
+    dst_argb += 32;
+  }
+}
+#endif
+
+void ScaleRowDown34_MSA(const uint8* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8* dst,
+                        int dst_width) {
+  int x;
+  (void)src_stride;
+  v16u8 src0, src1, src2, src3;
+  v16u8 vec0, vec1, vec2;
+  v16i8 mask0 = {0, 1, 3, 4, 5, 7, 8, 9, 11, 12, 13, 15, 16, 17, 19, 20};
+  v16i8 mask1 = {5, 7, 8, 9, 11, 12, 13, 15, 16, 17, 19, 20, 21, 23, 24, 25};
+  v16i8 mask2 = {11, 12, 13, 15, 16, 17, 19, 20,
+                 21, 23, 24, 25, 27, 28, 29, 31};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48);
+    vec0 = (v16u8)__msa_vshf_b(mask0, (v16i8)src1, (v16i8)src0);
+    vec1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src2, (v16i8)src1);
+    vec2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src3, (v16i8)src2);
+    __msa_st_b((v16i8)vec0, dst, 0);
+    __msa_st_b((v16i8)vec1, dst, 16);
+    __msa_st_b((v16i8)vec2, dst, 32);
+    src_ptr += 64;
+    dst += 48;
+  }
+}
+
+void ScaleRowDown34_0_Box_MSA(const uint8* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8* d,
+                              int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0, dst1, dst2;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5;
+  v16u8 vec6, vec7, vec8, vec9, vec10, vec11;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5;
+  v8i16 reg6, reg7, reg8, reg9, reg10, reg11;
+  v16u8 const0 = {3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1};
+  v16u8 const1 = {1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1};
+  v16u8 const2 = {1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3};
+  v16i8 mask0 = {0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10};
+  v16i8 mask1 = {10, 11, 12, 13, 13, 14, 14, 15,
+                 16, 17, 17, 18, 18, 19, 20, 21};
+  v16i8 mask2 = {5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15};
+  v8i16 shft0 = {2, 1, 2, 2, 1, 2, 2, 1};
+  v8i16 shft1 = {2, 2, 1, 2, 2, 1, 2, 2};
+  v8i16 shft2 = {1, 2, 2, 1, 2, 2, 1, 2};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t, 48);
+    vec0 = (v16u8)__msa_vshf_b(mask0, (v16i8)src0, (v16i8)src0);
+    vec1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
+    vec2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src1, (v16i8)src1);
+    vec3 = (v16u8)__msa_vshf_b(mask0, (v16i8)src2, (v16i8)src2);
+    vec4 = (v16u8)__msa_vshf_b(mask1, (v16i8)src3, (v16i8)src2);
+    vec5 = (v16u8)__msa_vshf_b(mask2, (v16i8)src3, (v16i8)src3);
+    vec6 = (v16u8)__msa_vshf_b(mask0, (v16i8)src4, (v16i8)src4);
+    vec7 = (v16u8)__msa_vshf_b(mask1, (v16i8)src5, (v16i8)src4);
+    vec8 = (v16u8)__msa_vshf_b(mask2, (v16i8)src5, (v16i8)src5);
+    vec9 = (v16u8)__msa_vshf_b(mask0, (v16i8)src6, (v16i8)src6);
+    vec10 = (v16u8)__msa_vshf_b(mask1, (v16i8)src7, (v16i8)src6);
+    vec11 = (v16u8)__msa_vshf_b(mask2, (v16i8)src7, (v16i8)src7);
+    reg0 = (v8i16)__msa_dotp_u_h(vec0, const0);
+    reg1 = (v8i16)__msa_dotp_u_h(vec1, const1);
+    reg2 = (v8i16)__msa_dotp_u_h(vec2, const2);
+    reg3 = (v8i16)__msa_dotp_u_h(vec3, const0);
+    reg4 = (v8i16)__msa_dotp_u_h(vec4, const1);
+    reg5 = (v8i16)__msa_dotp_u_h(vec5, const2);
+    reg6 = (v8i16)__msa_dotp_u_h(vec6, const0);
+    reg7 = (v8i16)__msa_dotp_u_h(vec7, const1);
+    reg8 = (v8i16)__msa_dotp_u_h(vec8, const2);
+    reg9 = (v8i16)__msa_dotp_u_h(vec9, const0);
+    reg10 = (v8i16)__msa_dotp_u_h(vec10, const1);
+    reg11 = (v8i16)__msa_dotp_u_h(vec11, const2);
+    reg0 = __msa_srar_h(reg0, shft0);
+    reg1 = __msa_srar_h(reg1, shft1);
+    reg2 = __msa_srar_h(reg2, shft2);
+    reg3 = __msa_srar_h(reg3, shft0);
+    reg4 = __msa_srar_h(reg4, shft1);
+    reg5 = __msa_srar_h(reg5, shft2);
+    reg6 = __msa_srar_h(reg6, shft0);
+    reg7 = __msa_srar_h(reg7, shft1);
+    reg8 = __msa_srar_h(reg8, shft2);
+    reg9 = __msa_srar_h(reg9, shft0);
+    reg10 = __msa_srar_h(reg10, shft1);
+    reg11 = __msa_srar_h(reg11, shft2);
+    reg0 = reg0 * 3 + reg6;
+    reg1 = reg1 * 3 + reg7;
+    reg2 = reg2 * 3 + reg8;
+    reg3 = reg3 * 3 + reg9;
+    reg4 = reg4 * 3 + reg10;
+    reg5 = reg5 * 3 + reg11;
+    reg0 = __msa_srari_h(reg0, 2);
+    reg1 = __msa_srari_h(reg1, 2);
+    reg2 = __msa_srari_h(reg2, 2);
+    reg3 = __msa_srari_h(reg3, 2);
+    reg4 = __msa_srari_h(reg4, 2);
+    reg5 = __msa_srari_h(reg5, 2);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)reg3, (v16i8)reg2);
+    dst2 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4);
+    __msa_st_b((v16i8)dst0, d, 0);
+    __msa_st_b((v16i8)dst1, d, 16);
+    __msa_st_b((v16i8)dst2, d, 32);
+    s += 64;
+    t += 64;
+    d += 48;
+  }
+}
+
+void ScaleRowDown34_1_Box_MSA(const uint8* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8* d,
+                              int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0, dst1, dst2;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5;
+  v16u8 vec6, vec7, vec8, vec9, vec10, vec11;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5;
+  v8i16 reg6, reg7, reg8, reg9, reg10, reg11;
+  v16u8 const0 = {3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1};
+  v16u8 const1 = {1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1};
+  v16u8 const2 = {1, 1, 1, 3, 3, 1, 1, 1, 1, 3, 3, 1, 1, 1, 1, 3};
+  v16i8 mask0 = {0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10};
+  v16i8 mask1 = {10, 11, 12, 13, 13, 14, 14, 15,
+                 16, 17, 17, 18, 18, 19, 20, 21};
+  v16i8 mask2 = {5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15};
+  v8i16 shft0 = {2, 1, 2, 2, 1, 2, 2, 1};
+  v8i16 shft1 = {2, 2, 1, 2, 2, 1, 2, 2};
+  v8i16 shft2 = {1, 2, 2, 1, 2, 2, 1, 2};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)s, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)s, 48);
+    src4 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src5 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    src6 = (v16u8)__msa_ld_b((v16i8*)t, 32);
+    src7 = (v16u8)__msa_ld_b((v16i8*)t, 48);
+    vec0 = (v16u8)__msa_vshf_b(mask0, (v16i8)src0, (v16i8)src0);
+    vec1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
+    vec2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src1, (v16i8)src1);
+    vec3 = (v16u8)__msa_vshf_b(mask0, (v16i8)src2, (v16i8)src2);
+    vec4 = (v16u8)__msa_vshf_b(mask1, (v16i8)src3, (v16i8)src2);
+    vec5 = (v16u8)__msa_vshf_b(mask2, (v16i8)src3, (v16i8)src3);
+    vec6 = (v16u8)__msa_vshf_b(mask0, (v16i8)src4, (v16i8)src4);
+    vec7 = (v16u8)__msa_vshf_b(mask1, (v16i8)src5, (v16i8)src4);
+    vec8 = (v16u8)__msa_vshf_b(mask2, (v16i8)src5, (v16i8)src5);
+    vec9 = (v16u8)__msa_vshf_b(mask0, (v16i8)src6, (v16i8)src6);
+    vec10 = (v16u8)__msa_vshf_b(mask1, (v16i8)src7, (v16i8)src6);
+    vec11 = (v16u8)__msa_vshf_b(mask2, (v16i8)src7, (v16i8)src7);
+    reg0 = (v8i16)__msa_dotp_u_h(vec0, const0);
+    reg1 = (v8i16)__msa_dotp_u_h(vec1, const1);
+    reg2 = (v8i16)__msa_dotp_u_h(vec2, const2);
+    reg3 = (v8i16)__msa_dotp_u_h(vec3, const0);
+    reg4 = (v8i16)__msa_dotp_u_h(vec4, const1);
+    reg5 = (v8i16)__msa_dotp_u_h(vec5, const2);
+    reg6 = (v8i16)__msa_dotp_u_h(vec6, const0);
+    reg7 = (v8i16)__msa_dotp_u_h(vec7, const1);
+    reg8 = (v8i16)__msa_dotp_u_h(vec8, const2);
+    reg9 = (v8i16)__msa_dotp_u_h(vec9, const0);
+    reg10 = (v8i16)__msa_dotp_u_h(vec10, const1);
+    reg11 = (v8i16)__msa_dotp_u_h(vec11, const2);
+    reg0 = __msa_srar_h(reg0, shft0);
+    reg1 = __msa_srar_h(reg1, shft1);
+    reg2 = __msa_srar_h(reg2, shft2);
+    reg3 = __msa_srar_h(reg3, shft0);
+    reg4 = __msa_srar_h(reg4, shft1);
+    reg5 = __msa_srar_h(reg5, shft2);
+    reg6 = __msa_srar_h(reg6, shft0);
+    reg7 = __msa_srar_h(reg7, shft1);
+    reg8 = __msa_srar_h(reg8, shft2);
+    reg9 = __msa_srar_h(reg9, shft0);
+    reg10 = __msa_srar_h(reg10, shft1);
+    reg11 = __msa_srar_h(reg11, shft2);
+    reg0 += reg6;
+    reg1 += reg7;
+    reg2 += reg8;
+    reg3 += reg9;
+    reg4 += reg10;
+    reg5 += reg11;
+    reg0 = __msa_srari_h(reg0, 1);
+    reg1 = __msa_srari_h(reg1, 1);
+    reg2 = __msa_srari_h(reg2, 1);
+    reg3 = __msa_srari_h(reg3, 1);
+    reg4 = __msa_srari_h(reg4, 1);
+    reg5 = __msa_srari_h(reg5, 1);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)reg3, (v16i8)reg2);
+    dst2 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4);
+    __msa_st_b((v16i8)dst0, d, 0);
+    __msa_st_b((v16i8)dst1, d, 16);
+    __msa_st_b((v16i8)dst2, d, 32);
+    s += 64;
+    t += 64;
+    d += 48;
+  }
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv