Add MSA optimized Interpolate/MergeUV/Misc functions

BUG=libyuv:634

Change-Id: If8d60bd57f01fe95bc2fd26196466574195cc126

Performance Gain (vs C auto-vectorized)
InterpolateRow_MSA      - ~3.3x
InterpolateRow_Any_MSA  - ~2.5x
ARGBSetRow_MSA          - ~1.0x
ARGBSetRow_Any_MSA      - ~1.0x
ARGBToRGB24Row_MSA      - ~1.9x
ARGBToRGB24Row_Any_MSA  - ~1.6x
MergeUVRow_MSA          - ~1.6x
MergeUVRow_Any_MSA      - ~1.2x

Performance Gain (vs C non-vectorized)
InterpolateRow_MSA      - ~11.3x
InterpolateRow_Any_MSA  - ~ 7.9x
ARGBSetRow_MSA          - ~ 6.2x
ARGBSetRow_Any_MSA      - ~ 4.0x
ARGBToRGB24Row_MSA      - ~ 9.9x
ARGBToRGB24Row_Any_MSA  - ~ 8.4x
MergeUVRow_MSA          - ~12.7x
MergeUVRow_Any_MSA      - ~ 8.0x

Change-Id: If8d60bd57f01fe95bc2fd26196466574195cc126
Reviewed-on: https://chromium-review.googlesource.com/445817Reviewed-by: Frank Barchard <fbarchard@google.com>
Commit-Queue: Frank Barchard <fbarchard@google.com>

include/libyuv/row.h

@@ -447,6 +447,10 @@ extern "C" {
 #define HAS_J400TOARGBROW_MSA
 #define HAS_YUY2TOARGBROW_MSA
 #define HAS_UYVYTOARGBROW_MSA
+#define HAS_INTERPOLATEROW_MSA
+#define HAS_ARGBSETROW_MSA
+#define HAS_RAWTORGB24ROW_MSA
+#define HAS_MERGEUVROW_MSA
 #endif
 
 #if defined(_MSC_VER) && !defined(__CLR_VER) && !defined(__clang__)
@@ -1412,6 +1416,10 @@ void MergeUVRow_NEON(const uint8* src_u,
                      const uint8* src_v,
                      uint8* dst_uv,
                      int width);
+void MergeUVRow_MSA(const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* dst_uv,
+                    int width);
 void MergeUVRow_Any_SSE2(const uint8* src_u,
                          const uint8* src_v,
                          uint8* dst_uv,
@@ -1424,6 +1432,10 @@ void MergeUVRow_Any_NEON(const uint8* src_u,
                          const uint8* src_v,
                          uint8* dst_uv,
                          int width);
+void MergeUVRow_Any_MSA(const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uv,
+                        int width);
 
 void CopyRow_SSE2(const uint8* src, uint8* dst, int count);
 void CopyRow_AVX(const uint8* src, uint8* dst, int count);
@@ -1482,6 +1494,8 @@ void ARGBSetRow_C(uint8* dst_argb, uint32 v32, int count);
 void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int count);
 void ARGBSetRow_NEON(uint8* dst_argb, uint32 v32, int count);
 void ARGBSetRow_Any_NEON(uint8* dst_argb, uint32 v32, int count);
+void ARGBSetRow_MSA(uint8* dst_argb, uint32 v32, int count);
+void ARGBSetRow_Any_MSA(uint8* dst_argb, uint32 v32, int count);
 
 // ARGBShufflers for BGRAToARGB etc.
 void ARGBShuffleRow_C(const uint8* src_argb,
@@ -1552,6 +1566,7 @@ void RGB24ToARGBRow_MSA(const uint8* src_rgb24, uint8* dst_argb, int width);
 void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int width);
 void RAWToARGBRow_MSA(const uint8* src_raw, uint8* dst_argb, int width);
 void RAWToRGB24Row_NEON(const uint8* src_raw, uint8* dst_rgb24, int width);
+void RAWToRGB24Row_MSA(const uint8* src_raw, uint8* dst_rgb24, int width);
 void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int width);
 void RGB565ToARGBRow_MSA(const uint8* src_rgb565, uint8* dst_argb, int width);
 void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555,
@@ -1613,6 +1628,7 @@ void RGB24ToARGBRow_Any_MSA(const uint8* src_rgb24, uint8* dst_argb, int width);
 void RAWToARGBRow_Any_NEON(const uint8* src_raw, uint8* dst_argb, int width);
 void RAWToARGBRow_Any_MSA(const uint8* src_raw, uint8* dst_argb, int width);
 void RAWToRGB24Row_Any_NEON(const uint8* src_raw, uint8* dst_rgb24, int width);
+void RAWToRGB24Row_Any_MSA(const uint8* src_raw, uint8* dst_rgb24, int width);
 void RGB565ToARGBRow_Any_NEON(const uint8* src_rgb565,
                               uint8* dst_argb,
                               int width);
@@ -2930,6 +2946,11 @@ void InterpolateRow_DSPR2(uint8* dst_ptr,
                           ptrdiff_t src_stride_ptr,
                           int width,
                           int source_y_fraction);
+void InterpolateRow_MSA(uint8* dst_ptr,
+                        const uint8* src_ptr,
+                        ptrdiff_t src_stride_ptr,
+                        int width,
+                        int source_y_fraction);
 void InterpolateRow_Any_NEON(uint8* dst_ptr,
                              const uint8* src_ptr,
                              ptrdiff_t src_stride_ptr,
@@ -2950,6 +2971,11 @@ void InterpolateRow_Any_DSPR2(uint8* dst_ptr,
                               ptrdiff_t src_stride_ptr,
                               int width,
                               int source_y_fraction);
+void InterpolateRow_Any_MSA(uint8* dst_ptr,
+                            const uint8* src_ptr,
+                            ptrdiff_t src_stride_ptr,
+                            int width,
+                            int source_y_fraction);
 
 void InterpolateRow_16_C(uint16* dst_ptr,
                          const uint16* src_ptr,

source/convert_from_argb.cc

@@ -359,6 +359,14 @@ int ARGBToNV12(const uint8* src_argb,
       ARGBToUVRow = ARGBToUVRow_DSPR2;
     }
   }
+#endif
+#if defined(HAS_MERGEUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MergeUVRow_ = MergeUVRow_Any_MSA;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_MSA;
+    }
+  }
 #endif
   {
     // Allocate a rows of uv.
@@ -502,6 +510,14 @@ int ARGBToNV21(const uint8* src_argb,
       ARGBToUVRow = ARGBToUVRow_DSPR2;
     }
   }
+#endif
+#if defined(HAS_MERGEUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MergeUVRow_ = MergeUVRow_Any_MSA;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_MSA;
+    }
+  }
 #endif
   {
     // Allocate a rows of uv.

source/planar_functions.cc

@@ -381,6 +381,14 @@ void MergeUVPlane(const uint8* src_u,
     }
   }
 #endif
+#if defined(HAS_MERGEUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MergeUVRow = MergeUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      MergeUVRow = MergeUVRow_MSA;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     // Merge a row of U and V into a row of UV.
@@ -1509,6 +1517,14 @@ int RAWToRGB24(const uint8* src_raw,
     }
   }
 #endif
+#if defined(HAS_RAWTORGB24ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RAWToRGB24Row = RAWToRGB24Row_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToRGB24Row = RAWToRGB24Row_MSA;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RAWToRGB24Row(src_raw, dst_rgb24, width);
@@ -1638,6 +1654,14 @@ int ARGBRect(uint8* dst_argb,
     ARGBSetRow = ARGBSetRow_X86;
   }
 #endif
+#if defined(HAS_ARGBSETROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBSetRow = ARGBSetRow_Any_MSA;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSetRow = ARGBSetRow_MSA;
+    }
+  }
+#endif
 
   // Set plane
   for (y = 0; y < height; ++y) {
@@ -2374,6 +2398,14 @@ int InterpolatePlane(const uint8* src0,
     InterpolateRow = InterpolateRow_DSPR2;
   }
 #endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     InterpolateRow(dst, src0, src1 - src0, width, interpolation);
@@ -3152,6 +3184,14 @@ int YUY2ToNV12(const uint8* src_yuy2,
     }
   }
 #endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
 
   {
     int awidth = halfwidth * 2;
@@ -3252,6 +3292,14 @@ int UYVYToNV12(const uint8* src_uyvy,
     }
   }
 #endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
 
   {
     int awidth = halfwidth * 2;

source/row_any.cc

@@ -212,6 +212,9 @@ ANY21(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, 0, 1, 1, 2, 31)
 #ifdef HAS_MERGEUVROW_NEON
 ANY21(MergeUVRow_Any_NEON, MergeUVRow_NEON, 0, 1, 1, 2, 15)
 #endif
+#ifdef HAS_MERGEUVROW_MSA
+ANY21(MergeUVRow_Any_MSA, MergeUVRow_MSA, 0, 1, 1, 2, 15)
+#endif
 
 // Math functions.
 #ifdef HAS_ARGBMULTIPLYROW_SSE2
@@ -429,6 +432,9 @@ ANY11(I400ToARGBRow_Any_MSA, I400ToARGBRow_MSA, 0, 1, 4, 15)
 #if defined(HAS_RAWTORGB24ROW_NEON)
 ANY11(RAWToRGB24Row_Any_NEON, RAWToRGB24Row_NEON, 0, 3, 3, 7)
 #endif
+#if defined(HAS_RAWTORGB24ROW_MSA)
+ANY11(RAWToRGB24Row_Any_MSA, RAWToRGB24Row_MSA, 0, 3, 3, 15)
+#endif
 #ifdef HAS_ARGBTOYROW_AVX2
 ANY11(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 0, 4, 1, 31)
 #endif
@@ -797,6 +803,9 @@ ANY11T(InterpolateRow_Any_NEON, InterpolateRow_NEON, 1, 1, 15)
 #ifdef HAS_INTERPOLATEROW_DSPR2
 ANY11T(InterpolateRow_Any_DSPR2, InterpolateRow_DSPR2, 1, 1, 3)
 #endif
+#ifdef HAS_INTERPOLATEROW_MSA
+ANY11T(InterpolateRow_Any_MSA, InterpolateRow_MSA, 1, 1, 31)
+#endif
 #undef ANY11T
 
 // Any 1 to 1 mirror.
@@ -862,6 +871,9 @@ ANY1(SetRow_Any_NEON, SetRow_NEON, uint8, 1, 15)
 #ifdef HAS_ARGBSETROW_NEON
 ANY1(ARGBSetRow_Any_NEON, ARGBSetRow_NEON, uint32, 4, 3)
 #endif
+#ifdef HAS_ARGBSETROW_MSA
+ANY1(ARGBSetRow_Any_MSA, ARGBSetRow_MSA, uint32, 4, 3)
+#endif
 #undef ANY1
 
 // Any 1 to 2.  Outputs UV planes.

source/row_msa.cc

@@ -8,6 +8,8 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
+#include <string.h>
+
 #include "libyuv/row.h"
 
 // This module is for GCC MSA
@@ -2849,6 +2851,124 @@ void UYVYToARGBRow_MSA(const uint8* src_uyvy,
   }
 }
 
+void InterpolateRow_MSA(uint8* dst_ptr,
+                        const uint8* src_ptr,
+                        ptrdiff_t src_stride,
+                        int width,
+                        int32 source_y_fraction) {
+  int32 y1_fraction = source_y_fraction;
+  int32 y0_fraction = 256 - y1_fraction;
+  uint16 y_fractions;
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  v16u8 src0, src1, src2, src3, dst0, dst1;
+  v8u16 vec0, vec1, vec2, vec3, y_frac;
+
+  if (0 == y1_fraction) {
+    memcpy(dst_ptr, src_ptr, width);
+    return;
+  }
+
+  if (128 == y1_fraction) {
+    for (x = 0; x < width; x += 32) {
+      src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+      src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+      src2 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+      src3 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+      dst0 = __msa_aver_u_b(src0, src2);
+      dst1 = __msa_aver_u_b(src1, src3);
+      ST_UB2(dst0, dst1, dst_ptr, 16);
+      s += 32;
+      t += 32;
+      dst_ptr += 32;
+    }
+    return;
+  }
+
+  y_fractions = (uint16)(y0_fraction + (y1_fraction << 8));
+  y_frac = (v8u16)__msa_fill_h(y_fractions);
+
+  for (x = 0; x < width; x += 32) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)s, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)t, 0);
+    src3 = (v16u8)__msa_ld_b((v16i8*)t, 16);
+    vec0 = (v8u16)__msa_ilvr_b((v16i8)src2, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b((v16i8)src2, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b((v16i8)src3, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b((v16i8)src3, (v16i8)src1);
+    vec0 = (v8u16)__msa_dotp_u_h((v16u8)vec0, (v16u8)y_frac);
+    vec1 = (v8u16)__msa_dotp_u_h((v16u8)vec1, (v16u8)y_frac);
+    vec2 = (v8u16)__msa_dotp_u_h((v16u8)vec2, (v16u8)y_frac);
+    vec3 = (v8u16)__msa_dotp_u_h((v16u8)vec3, (v16u8)y_frac);
+    vec0 = (v8u16)__msa_srari_h((v8i16)vec0, 8);
+    vec1 = (v8u16)__msa_srari_h((v8i16)vec1, 8);
+    vec2 = (v8u16)__msa_srari_h((v8i16)vec2, 8);
+    vec3 = (v8u16)__msa_srari_h((v8i16)vec3, 8);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    ST_UB2(dst0, dst1, dst_ptr, 16);
+    s += 32;
+    t += 32;
+    dst_ptr += 32;
+  }
+}
+
+void ARGBSetRow_MSA(uint8* dst_argb, uint32 v32, int width) {
+  int x;
+  v16u8 dst0 = (v16u8)__msa_fill_w(v32);
+
+  for (x = 0; x < width; x += 4) {
+    ST_UB(dst0, dst_argb);
+    dst_argb += 16;
+  }
+}
+
+void RAWToRGB24Row_MSA(const uint8* src_raw, uint8* dst_rgb24, int width) {
+  int x;
+  v16u8 src0, src1, src2, src3, src4, dst0, dst1, dst2;
+  v16i8 shuffler0 = {2, 1, 0, 5, 4, 3, 8, 7, 6, 11, 10, 9, 14, 13, 12, 17};
+  v16i8 shuffler1 = {8,  7,  12, 11, 10, 15, 14, 13,
+                     18, 17, 16, 21, 20, 19, 24, 23};
+  v16i8 shuffler2 = {14, 19, 18, 17, 22, 21, 20, 25,
+                     24, 23, 28, 27, 26, 31, 30, 29};
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_raw, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_raw, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_raw, 32);
+    src3 = (v16u8)__msa_sldi_b((v16i8)src1, (v16i8)src0, 8);
+    src4 = (v16u8)__msa_sldi_b((v16i8)src2, (v16i8)src1, 8);
+    dst0 = (v16u8)__msa_vshf_b(shuffler0, (v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_vshf_b(shuffler1, (v16i8)src4, (v16i8)src3);
+    dst2 = (v16u8)__msa_vshf_b(shuffler2, (v16i8)src2, (v16i8)src1);
+    ST_UB2(dst0, dst1, dst_rgb24, 16);
+    ST_UB(dst2, (dst_rgb24 + 32));
+    src_raw += 48;
+    dst_rgb24 += 48;
+  }
+}
+
+void MergeUVRow_MSA(const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* dst_uv,
+                    int width) {
+  int x;
+  v16u8 src0, src1, dst0, dst1;
+
+  for (x = 0; x < width; x += 16) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_u, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_v, 0);
+    dst0 = (v16u8)__msa_ilvr_b((v16i8)src1, (v16i8)src0);
+    dst1 = (v16u8)__msa_ilvl_b((v16i8)src1, (v16i8)src0);
+    ST_UB2(dst0, dst1, dst_uv, 16);
+    src_u += 16;
+    src_v += 16;
+    dst_uv += 32;
+  }
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv

source/scale.cc

@@ -45,10 +45,9 @@ static void ScalePlaneDown2(int src_width,
   int y;
   void (*ScaleRowDown2)(const uint8* src_ptr, ptrdiff_t src_stride,
                         uint8* dst_ptr, int dst_width) =
-      filtering == kFilterNone
-          ? ScaleRowDown2_C
-          : (filtering == kFilterLinear ? ScaleRowDown2Linear_C
-                                        : ScaleRowDown2Box_C);
+      filtering == kFilterNone ? ScaleRowDown2_C : (filtering == kFilterLinear
+                                                        ? ScaleRowDown2Linear_C
+                                                        : ScaleRowDown2Box_C);
   int row_stride = src_stride << 1;
   (void)src_width;
   (void)src_height;
@@ -1057,6 +1056,14 @@ void ScalePlaneBilinearDown(int src_width,
     }
   }
 #endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
 
 #if defined(HAS_SCALEFILTERCOLS_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {

source/scale_argb.cc

@@ -315,6 +315,14 @@ static void ScaleARGBBilinearDown(int src_width,
     }
   }
 #endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
 #if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
     ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
@@ -408,6 +416,14 @@ static void ScaleARGBBilinearUp(int src_width,
       IS_ALIGNED(dst_stride, 4)) {
     InterpolateRow = InterpolateRow_DSPR2;
   }
+#endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
 #endif
   if (src_width >= 32768) {
     ScaleARGBFilterCols =
@@ -598,6 +614,14 @@ static void ScaleYUVToARGBBilinearUp(int src_width,
     InterpolateRow = InterpolateRow_DSPR2;
   }
 #endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
 
   void (*ScaleARGBFilterCols)(uint8 * dst_argb, const uint8* src_argb,
                               int dst_width, int x, int dx) =

source/scale_common.cc

@@ -1072,6 +1072,14 @@ void ScalePlaneVertical(int src_height,
       InterpolateRow = InterpolateRow_DSPR2;
     }
   }
+#endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
 #endif
   for (j = 0; j < dst_height; ++j) {
     int yi;