Merge "mips msa optimize vpx_scaled_2d function"

test/convolve_test.cc

@@ -1359,7 +1359,7 @@ const ConvolveFunctions convolve8_msa(
     vpx_convolve8_avg_horiz_msa, vpx_convolve8_vert_msa,
     vpx_convolve8_avg_vert_msa, vpx_convolve8_msa, vpx_convolve8_avg_msa,
     vpx_scaled_horiz_c, vpx_scaled_avg_horiz_c, vpx_scaled_vert_c,
-    vpx_scaled_avg_vert_c, vpx_scaled_2d_c, vpx_scaled_avg_2d_c, 0);
+    vpx_scaled_avg_vert_c, vpx_scaled_2d_msa, vpx_scaled_avg_2d_c, 0);
 
 const ConvolveParam kArrayConvolve8_msa[] = { ALL_SIZES(convolve8_msa) };
 INSTANTIATE_TEST_CASE_P(MSA, ConvolveTest,

vpx_dsp/mips/macros_msa.h

@@ -555,6 +555,7 @@
 #define VSHF_B2_UB(...) VSHF_B2(v16u8, __VA_ARGS__)
 #define VSHF_B2_SB(...) VSHF_B2(v16i8, __VA_ARGS__)
 #define VSHF_B2_UH(...) VSHF_B2(v8u16, __VA_ARGS__)
+#define VSHF_B2_SH(...) VSHF_B2(v8i16, __VA_ARGS__)
 
 #define VSHF_B4(RTYPE, in0, in1, mask0, mask1, mask2, mask3, out0, out1, out2, \
                 out3)                                                          \
@@ -1182,6 +1183,7 @@
     out1 = (RTYPE)__msa_ilvl_w((v4i32)in0, (v4i32)in1); \
   }
 #define ILVRL_W2_UB(...) ILVRL_W2(v16u8, __VA_ARGS__)
+#define ILVRL_W2_SB(...) ILVRL_W2(v16i8, __VA_ARGS__)
 #define ILVRL_W2_SH(...) ILVRL_W2(v8i16, __VA_ARGS__)
 #define ILVRL_W2_SW(...) ILVRL_W2(v4i32, __VA_ARGS__)
 
@@ -1595,6 +1597,25 @@
     out = (v4i32)__msa_ilvr_h(sign_m, (v8i16)in); \
   }
 
+/* Description : Sign extend byte elements from input vector and return
+                 halfword results in pair of vectors
+   Arguments   : Input   - in           (byte vector)
+                 Outputs - out0, out1   (sign extended halfword vectors)
+                 Return Type - signed halfword
+   Details     : Sign bit of byte elements from input vector 'in' is
+                 extracted and interleaved right with same vector 'in0' to
+                 generate 8 signed halfword elements in 'out0'
+                 Then interleaved left with same vector 'in0' to
+                 generate 8 signed halfword elements in 'out1'
+*/
+#define UNPCK_SB_SH(in, out0, out1)       \
+  {                                       \
+    v16i8 tmp_m;                          \
+                                          \
+    tmp_m = __msa_clti_s_b((v16i8)in, 0); \
+    ILVRL_B2_SH(tmp_m, in, out0, out1);   \
+  }
+
 /* Description : Zero extend unsigned byte elements to halfword elements
    Arguments   : Input   - in          (unsigned byte vector)
                  Outputs - out0, out1  (unsigned  halfword vectors)

vpx_dsp/mips/vpx_convolve8_msa.c

@@ -629,3 +629,599 @@ void vpx_convolve8_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
     }
   }
 }
+
+static void filter_horiz_w4_msa(const uint8_t *src_x, ptrdiff_t src_pitch,
+                                uint8_t *dst, const int16_t *x_filter) {
+  uint64_t srcd0, srcd1, srcd2, srcd3;
+  uint32_t res;
+  v16u8 src0 = { 0 }, src1 = { 0 }, dst0;
+  v16i8 out0, out1;
+  v16i8 shf1 = { 0, 8, 16, 24, 4, 12, 20, 28, 1, 9, 17, 25, 5, 13, 21, 29 };
+  v16i8 shf2 = shf1 + 2;
+  v16i8 filt_shf0 = { 0, 1, 0, 1, 0, 1, 0, 1, 8, 9, 8, 9, 8, 9, 8, 9 };
+  v16i8 filt_shf1 = filt_shf0 + 2;
+  v16i8 filt_shf2 = filt_shf0 + 4;
+  v16i8 filt_shf3 = filt_shf0 + 6;
+  v8i16 filt, src0_h, src1_h, src2_h, src3_h, filt0, filt1, filt2, filt3;
+
+  LD4(src_x, src_pitch, srcd0, srcd1, srcd2, srcd3);
+  INSERT_D2_UB(srcd0, srcd1, src0);
+  INSERT_D2_UB(srcd2, srcd3, src1);
+  VSHF_B2_SB(src0, src1, src0, src1, shf1, shf2, out0, out1);
+  XORI_B2_128_SB(out0, out1);
+  UNPCK_SB_SH(out0, src0_h, src1_h);
+  UNPCK_SB_SH(out1, src2_h, src3_h);
+
+  filt = LD_SH(x_filter);
+  VSHF_B2_SH(filt, filt, filt, filt, filt_shf0, filt_shf1, filt0, filt1);
+  VSHF_B2_SH(filt, filt, filt, filt, filt_shf2, filt_shf3, filt2, filt3);
+
+  src0_h *= filt0;
+  src0_h += src1_h * filt1;
+  src0_h += src2_h * filt2;
+  src0_h += src3_h * filt3;
+
+  src1_h = (v8i16)__msa_sldi_b((v16i8)src0_h, (v16i8)src0_h, 8);
+
+  src0_h = __msa_adds_s_h(src0_h, src1_h);
+  src0_h = __msa_srari_h(src0_h, FILTER_BITS);
+  src0_h = __msa_sat_s_h(src0_h, 7);
+  dst0 = PCKEV_XORI128_UB(src0_h, src0_h);
+  res = __msa_copy_u_w((v4i32)dst0, 0);
+  SW(res, dst);
+}
+
+static void filter_horiz_w8_msa(const uint8_t *src_x, ptrdiff_t src_pitch,
+                                uint8_t *dst, const int16_t *x_filter) {
+  uint64_t srcd0, srcd1, srcd2, srcd3;
+  v16u8 src0 = { 0 }, src1 = { 0 }, src2 = { 0 }, src3 = { 0 };
+  v16u8 tmp0, tmp1, tmp2, tmp3, dst0;
+  v16i8 out0, out1, out2, out3;
+  v16i8 shf1 = { 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27 };
+  v16i8 shf2 = shf1 + 4;
+  v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h;
+  v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7;
+
+  LD4(src_x, src_pitch, srcd0, srcd1, srcd2, srcd3);
+  INSERT_D2_UB(srcd0, srcd1, src0);
+  INSERT_D2_UB(srcd2, srcd3, src1);
+  LD4(src_x + 4 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3);
+  INSERT_D2_UB(srcd0, srcd1, src2);
+  INSERT_D2_UB(srcd2, srcd3, src3);
+
+  filt = LD_SH(x_filter);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+  SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7);
+
+  // transpose
+  VSHF_B2_UB(src0, src1, src0, src1, shf1, shf2, tmp0, tmp1);
+  VSHF_B2_UB(src2, src3, src2, src3, shf1, shf2, tmp2, tmp3);
+  ILVRL_W2_SB(tmp2, tmp0, out0, out1);
+  ILVRL_W2_SB(tmp3, tmp1, out2, out3);
+
+  XORI_B4_128_SB(out0, out1, out2, out3);
+  UNPCK_SB_SH(out0, src0_h, src1_h);
+  UNPCK_SB_SH(out1, src2_h, src3_h);
+  UNPCK_SB_SH(out2, src4_h, src5_h);
+  UNPCK_SB_SH(out3, src6_h, src7_h);
+
+  src0_h *= filt0;
+  src4_h *= filt4;
+  src0_h += src1_h * filt1;
+  src4_h += src5_h * filt5;
+  src0_h += src2_h * filt2;
+  src4_h += src6_h * filt6;
+  src0_h += src3_h * filt3;
+  src4_h += src7_h * filt7;
+
+  src0_h = __msa_adds_s_h(src0_h, src4_h);
+  src0_h = __msa_srari_h(src0_h, FILTER_BITS);
+  src0_h = __msa_sat_s_h(src0_h, 7);
+  dst0 = PCKEV_XORI128_UB(src0_h, src0_h);
+  ST8x1_UB(dst0, dst);
+}
+
+static void filter_horiz_w16_msa(const uint8_t *src_x, ptrdiff_t src_pitch,
+                                 uint8_t *dst, const int16_t *x_filter) {
+  uint64_t srcd0, srcd1, srcd2, srcd3;
+  v16u8 src0 = { 0 }, src1 = { 0 }, src2 = { 0 }, src3 = { 0 };
+  v16u8 src4 = { 0 }, src5 = { 0 }, src6 = { 0 }, src7 = { 0 };
+  v16u8 tmp0, tmp1, tmp2, tmp3, dst0;
+  v16i8 out0, out1, out2, out3, out4, out5, out6, out7;
+  v16i8 shf1 = { 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27 };
+  v16i8 shf2 = shf1 + 4;
+  v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h;
+  v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7;
+  v8i16 dst0_h, dst1_h, dst2_h, dst3_h;
+
+  LD4(src_x, src_pitch, srcd0, srcd1, srcd2, srcd3);
+  INSERT_D2_UB(srcd0, srcd1, src0);
+  INSERT_D2_UB(srcd2, srcd3, src1);
+  LD4(src_x + 4 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3);
+  INSERT_D2_UB(srcd0, srcd1, src2);
+  INSERT_D2_UB(srcd2, srcd3, src3);
+  LD4(src_x + 8 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3);
+  INSERT_D2_UB(srcd0, srcd1, src4);
+  INSERT_D2_UB(srcd2, srcd3, src5);
+  LD4(src_x + 12 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3);
+  INSERT_D2_UB(srcd0, srcd1, src6);
+  INSERT_D2_UB(srcd2, srcd3, src7);
+
+  filt = LD_SH(x_filter);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+  SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7);
+
+  // transpose
+  VSHF_B2_UB(src0, src1, src0, src1, shf1, shf2, tmp0, tmp1);
+  VSHF_B2_UB(src2, src3, src2, src3, shf1, shf2, tmp2, tmp3);
+  ILVRL_W2_SB(tmp2, tmp0, out0, out1);
+  ILVRL_W2_SB(tmp3, tmp1, out2, out3);
+  XORI_B4_128_SB(out0, out1, out2, out3);
+
+  UNPCK_SB_SH(out0, src0_h, src1_h);
+  UNPCK_SB_SH(out1, src2_h, src3_h);
+  UNPCK_SB_SH(out2, src4_h, src5_h);
+  UNPCK_SB_SH(out3, src6_h, src7_h);
+
+  VSHF_B2_UB(src4, src5, src4, src5, shf1, shf2, tmp0, tmp1);
+  VSHF_B2_UB(src6, src7, src6, src7, shf1, shf2, tmp2, tmp3);
+  ILVRL_W2_SB(tmp2, tmp0, out4, out5);
+  ILVRL_W2_SB(tmp3, tmp1, out6, out7);
+  XORI_B4_128_SB(out4, out5, out6, out7);
+
+  dst0_h = src0_h * filt0;
+  dst1_h = src4_h * filt4;
+  dst0_h += src1_h * filt1;
+  dst1_h += src5_h * filt5;
+  dst0_h += src2_h * filt2;
+  dst1_h += src6_h * filt6;
+  dst0_h += src3_h * filt3;
+  dst1_h += src7_h * filt7;
+
+  UNPCK_SB_SH(out4, src0_h, src1_h);
+  UNPCK_SB_SH(out5, src2_h, src3_h);
+  UNPCK_SB_SH(out6, src4_h, src5_h);
+  UNPCK_SB_SH(out7, src6_h, src7_h);
+
+  dst2_h = src0_h * filt0;
+  dst3_h = src4_h * filt4;
+  dst2_h += src1_h * filt1;
+  dst3_h += src5_h * filt5;
+  dst2_h += src2_h * filt2;
+  dst3_h += src6_h * filt6;
+  dst2_h += src3_h * filt3;
+  dst3_h += src7_h * filt7;
+
+  ADDS_SH2_SH(dst0_h, dst1_h, dst2_h, dst3_h, dst0_h, dst2_h);
+  SRARI_H2_SH(dst0_h, dst2_h, FILTER_BITS);
+  SAT_SH2_SH(dst0_h, dst2_h, 7);
+  dst0 = PCKEV_XORI128_UB(dst0_h, dst2_h);
+  ST_UB(dst0, dst);
+}
+
+static void transpose4x4_to_dst(const uint8_t *src, uint8_t *dst,
+                                ptrdiff_t dst_stride) {
+  v16u8 in0;
+  v16i8 out0 = { 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15 };
+
+  in0 = LD_UB(src);
+  out0 = __msa_vshf_b(out0, (v16i8)in0, (v16i8)in0);
+  ST4x4_UB(out0, out0, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void transpose8x8_to_dst(const uint8_t *src, uint8_t *dst,
+                                ptrdiff_t dst_stride) {
+  v16u8 in0, in1, in2, in3, out0, out1, out2, out3, tmp0, tmp1, tmp2, tmp3;
+  v16i8 shf1 = { 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27 };
+  v16i8 shf2 = shf1 + 4;
+
+  LD_UB4(src, 16, in0, in1, in2, in3);
+  VSHF_B2_UB(in0, in1, in0, in1, shf1, shf2, tmp0, tmp1);
+  VSHF_B2_UB(in2, in3, in2, in3, shf1, shf2, tmp2, tmp3);
+  ILVRL_W2_UB(tmp2, tmp0, out0, out1);
+  ILVRL_W2_UB(tmp3, tmp1, out2, out3);
+  ST8x4_UB(out0, out1, dst, dst_stride);
+  ST8x4_UB(out2, out3, dst + 4 * dst_stride, dst_stride);
+}
+
+static void transpose16x16_to_dst(const uint8_t *src, uint8_t *dst,
+                                  ptrdiff_t dst_stride) {
+  v16u8 in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11, in12;
+  v16u8 in13, in14, in15, out0, out1, out2, out3, out4, out5, out6, out7, out8;
+  v16u8 out9, out10, out11, out12, out13, out14, out15;
+
+  LD_UB8(src, 16, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_UB8(src + 16 * 8, 16, in8, in9, in10, in11, in12, in13, in14, in15);
+
+  TRANSPOSE16x8_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10,
+                      in11, in12, in13, in14, in15, out0, out1, out2, out3,
+                      out4, out5, out6, out7);
+  ST_UB8(out0, out1, out2, out3, out4, out5, out6, out7, dst, dst_stride);
+  dst += 8 * dst_stride;
+
+  SLDI_B4_0_UB(in0, in1, in2, in3, in0, in1, in2, in3, 8);
+  SLDI_B4_0_UB(in4, in5, in6, in7, in4, in5, in6, in7, 8);
+  SLDI_B4_0_UB(in8, in9, in10, in11, in8, in9, in10, in11, 8);
+  SLDI_B4_0_UB(in12, in13, in14, in15, in12, in13, in14, in15, 8);
+
+  TRANSPOSE16x8_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10,
+                      in11, in12, in13, in14, in15, out8, out9, out10, out11,
+                      out12, out13, out14, out15);
+  ST_UB8(out8, out9, out10, out11, out12, out13, out14, out15, dst, dst_stride);
+}
+
+static void scaledconvolve_horiz_w4(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const InterpKernel *x_filters, int x0_q4,
+                                    int x_step_q4, int h) {
+  DECLARE_ALIGNED(16, uint8_t, temp[4 * 4]);
+  int y, z, i;
+  src -= SUBPEL_TAPS / 2 - 1;
+
+  for (y = 0; y < h; y += 4) {
+    int x_q4 = x0_q4;
+    for (z = 0; z < 4; ++z) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+
+      if (x_q4 & SUBPEL_MASK) {
+        filter_horiz_w4_msa(src_x, src_stride, temp + (z * 4), x_filter);
+      } else {
+        for (i = 0; i < 4; ++i) {
+          temp[z * 4 + i] = src_x[i * src_stride + 3];
+        }
+      }
+
+      x_q4 += x_step_q4;
+    }
+
+    transpose4x4_to_dst(temp, dst, dst_stride);
+
+    src += src_stride * 4;
+    dst += dst_stride * 4;
+  }
+}
+
+static void scaledconvolve_horiz_w8(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const InterpKernel *x_filters, int x0_q4,
+                                    int x_step_q4, int h) {
+  DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]);
+  int y, z, i;
+  src -= SUBPEL_TAPS / 2 - 1;
+
+  // This function processes 8x8 areas. The intermediate height is not always
+  // a multiple of 8, so force it to be a multiple of 8 here.
+  y = h + (8 - (h & 0x7));
+
+  do {
+    int x_q4 = x0_q4;
+    for (z = 0; z < 8; ++z) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+
+      if (x_q4 & SUBPEL_MASK) {
+        filter_horiz_w8_msa(src_x, src_stride, temp + (z * 8), x_filter);
+      } else {
+        for (i = 0; i < 8; ++i) {
+          temp[z * 8 + i] = src_x[3 + i * src_stride];
+        }
+      }
+
+      x_q4 += x_step_q4;
+    }
+
+    transpose8x8_to_dst(temp, dst, dst_stride);
+
+    src += src_stride * 8;
+    dst += dst_stride * 8;
+  } while (y -= 8);
+}
+
+static void scaledconvolve_horiz_mul16(const uint8_t *src, ptrdiff_t src_stride,
+                                       uint8_t *dst, ptrdiff_t dst_stride,
+                                       const InterpKernel *x_filters, int x0_q4,
+                                       int x_step_q4, int w, int h) {
+  DECLARE_ALIGNED(16, uint8_t, temp[16 * 16]);
+  int x, y, z, i;
+
+  src -= SUBPEL_TAPS / 2 - 1;
+
+  // This function processes 16x16 areas.  The intermediate height is not always
+  // a multiple of 16, so force it to be a multiple of 8 here.
+  y = h + (16 - (h & 0xF));
+
+  do {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; x += 16) {
+      for (z = 0; z < 16; ++z) {
+        const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+        const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+
+        if (x_q4 & SUBPEL_MASK) {
+          filter_horiz_w16_msa(src_x, src_stride, temp + (z * 16), x_filter);
+        } else {
+          for (i = 0; i < 16; ++i) {
+            temp[z * 16 + i] = src_x[3 + i * src_stride];
+          }
+        }
+
+        x_q4 += x_step_q4;
+      }
+
+      transpose16x16_to_dst(temp, dst + x, dst_stride);
+    }
+
+    src += src_stride * 16;
+    dst += dst_stride * 16;
+  } while (y -= 16);
+}
+
+static void filter_vert_w4_msa(const uint8_t *src_y, ptrdiff_t src_pitch,
+                               uint8_t *dst, const int16_t *y_filter) {
+  uint32_t srcw0, srcw1, srcw2, srcw3, srcw4, srcw5, srcw6, srcw7;
+  uint32_t res;
+  v16u8 src0 = { 0 }, src1 = { 0 }, dst0;
+  v16i8 out0, out1;
+  v16i8 shf1 = { 0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23 };
+  v16i8 shf2 = shf1 + 8;
+  v16i8 filt_shf0 = { 0, 1, 0, 1, 0, 1, 0, 1, 8, 9, 8, 9, 8, 9, 8, 9 };
+  v16i8 filt_shf1 = filt_shf0 + 2;
+  v16i8 filt_shf2 = filt_shf0 + 4;
+  v16i8 filt_shf3 = filt_shf0 + 6;
+  v8i16 filt, src0_h, src1_h, src2_h, src3_h;
+  v8i16 filt0, filt1, filt2, filt3;
+
+  LW4(src_y, src_pitch, srcw0, srcw1, srcw2, srcw3);
+  LW4(src_y + 4 * src_pitch, src_pitch, srcw4, srcw5, srcw6, srcw7);
+  INSERT_W4_UB(srcw0, srcw1, srcw2, srcw3, src0);
+  INSERT_W4_UB(srcw4, srcw5, srcw6, srcw7, src1);
+  VSHF_B2_SB(src0, src1, src0, src1, shf1, shf2, out0, out1);
+  XORI_B2_128_SB(out0, out1);
+  UNPCK_SB_SH(out0, src0_h, src1_h);
+  UNPCK_SB_SH(out1, src2_h, src3_h);
+
+  filt = LD_SH(y_filter);
+  VSHF_B2_SH(filt, filt, filt, filt, filt_shf0, filt_shf1, filt0, filt1);
+  VSHF_B2_SH(filt, filt, filt, filt, filt_shf2, filt_shf3, filt2, filt3);
+
+  src0_h *= filt0;
+  src0_h += src1_h * filt1;
+  src0_h += src2_h * filt2;
+  src0_h += src3_h * filt3;
+
+  src1_h = (v8i16)__msa_sldi_b((v16i8)src0_h, (v16i8)src0_h, 8);
+
+  src0_h = __msa_adds_s_h(src0_h, src1_h);
+  src0_h = __msa_srari_h(src0_h, FILTER_BITS);
+  src0_h = __msa_sat_s_h(src0_h, 7);
+  dst0 = PCKEV_XORI128_UB(src0_h, src0_h);
+  res = __msa_copy_u_w((v4i32)dst0, 0);
+  SW(res, dst);
+}
+
+static void filter_vert_w8_msa(const uint8_t *src_y, ptrdiff_t src_pitch,
+                               uint8_t *dst, const int16_t *y_filter) {
+  uint64_t srcd0, srcd1, srcd2, srcd3;
+  v16u8 dst0;
+  v16i8 src0 = { 0 }, src1 = { 0 }, src2 = { 0 }, src3 = { 0 };
+  v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h;
+  v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7;
+
+  LD4(src_y, src_pitch, srcd0, srcd1, srcd2, srcd3);
+  INSERT_D2_SB(srcd0, srcd1, src0);
+  INSERT_D2_SB(srcd2, srcd3, src1);
+  LD4(src_y + 4 * src_pitch, src_pitch, srcd0, srcd1, srcd2, srcd3);
+  INSERT_D2_SB(srcd0, srcd1, src2);
+  INSERT_D2_SB(srcd2, srcd3, src3);
+
+  filt = LD_SH(y_filter);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+  SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7);
+
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  UNPCK_SB_SH(src0, src0_h, src1_h);
+  UNPCK_SB_SH(src1, src2_h, src3_h);
+  UNPCK_SB_SH(src2, src4_h, src5_h);
+  UNPCK_SB_SH(src3, src6_h, src7_h);
+
+  src0_h *= filt0;
+  src4_h *= filt4;
+  src0_h += src1_h * filt1;
+  src4_h += src5_h * filt5;
+  src0_h += src2_h * filt2;
+  src4_h += src6_h * filt6;
+  src0_h += src3_h * filt3;
+  src4_h += src7_h * filt7;
+
+  src0_h = __msa_adds_s_h(src0_h, src4_h);
+  src0_h = __msa_srari_h(src0_h, FILTER_BITS);
+  src0_h = __msa_sat_s_h(src0_h, 7);
+  dst0 = PCKEV_XORI128_UB(src0_h, src0_h);
+  ST8x1_UB(dst0, dst);
+}
+
+static void filter_vert_mul_w16_msa(const uint8_t *src_y, ptrdiff_t src_pitch,
+                                    uint8_t *dst, const int16_t *y_filter,
+                                    int w) {
+  int x;
+  v16u8 dst0;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v8i16 filt, src0_h, src1_h, src2_h, src3_h, src4_h, src5_h, src6_h, src7_h;
+  v8i16 src8_h, src9_h, src10_h, src11_h, src12_h, src13_h, src14_h, src15_h;
+  v8i16 filt0, filt1, filt2, filt3, filt4, filt5, filt6, filt7;
+
+  filt = LD_SH(y_filter);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+  SPLATI_H4_SH(filt, 4, 5, 6, 7, filt4, filt5, filt6, filt7);
+
+  for (x = 0; x < w; x += 16) {
+    LD_SB8(src_y, src_pitch, src0, src1, src2, src3, src4, src5, src6, src7);
+    src_y += 16;
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    XORI_B4_128_SB(src4, src5, src6, src7);
+    UNPCK_SB_SH(src0, src0_h, src1_h);
+    UNPCK_SB_SH(src1, src2_h, src3_h);
+    UNPCK_SB_SH(src2, src4_h, src5_h);
+    UNPCK_SB_SH(src3, src6_h, src7_h);
+    UNPCK_SB_SH(src4, src8_h, src9_h);
+    UNPCK_SB_SH(src5, src10_h, src11_h);
+    UNPCK_SB_SH(src6, src12_h, src13_h);
+    UNPCK_SB_SH(src7, src14_h, src15_h);
+
+    src0_h *= filt0;
+    src1_h *= filt0;
+    src8_h *= filt4;
+    src9_h *= filt4;
+    src0_h += src2_h * filt1;
+    src1_h += src3_h * filt1;
+    src8_h += src10_h * filt5;
+    src9_h += src11_h * filt5;
+    src0_h += src4_h * filt2;
+    src1_h += src5_h * filt2;
+    src8_h += src12_h * filt6;
+    src9_h += src13_h * filt6;
+    src0_h += src6_h * filt3;
+    src1_h += src7_h * filt3;
+    src8_h += src14_h * filt7;
+    src9_h += src15_h * filt7;
+
+    ADDS_SH2_SH(src0_h, src8_h, src1_h, src9_h, src0_h, src1_h);
+    SRARI_H2_SH(src0_h, src1_h, FILTER_BITS);
+    SAT_SH2_SH(src0_h, src1_h, 7);
+    dst0 = PCKEV_XORI128_UB(src0_h, src1_h);
+    ST_UB(dst0, dst);
+    dst += 16;
+  }
+}
+
+static void scaledconvolve_vert_w4(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const InterpKernel *y_filters, int y0_q4,
+                                   int y_step_q4, int h) {
+  int y;
+  int y_q4 = y0_q4;
+
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (y = 0; y < h; ++y) {
+    const uint8_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+    const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+
+    if (y_q4 & SUBPEL_MASK) {
+      filter_vert_w4_msa(src_y, src_stride, &dst[y * dst_stride], y_filter);
+    } else {
+      uint32_t srcd = LW(src_y + 3 * src_stride);
+      SW(srcd, dst + y * dst_stride);
+    }
+
+    y_q4 += y_step_q4;
+  }
+}
+
+static void scaledconvolve_vert_w8(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const InterpKernel *y_filters, int y0_q4,
+                                   int y_step_q4, int h) {
+  int y;
+  int y_q4 = y0_q4;
+
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (y = 0; y < h; ++y) {
+    const uint8_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+    const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+
+    if (y_q4 & SUBPEL_MASK) {
+      filter_vert_w8_msa(src_y, src_stride, &dst[y * dst_stride], y_filter);
+    } else {
+      uint64_t srcd = LD(src_y + 3 * src_stride);
+      SD(srcd, dst + y * dst_stride);
+    }
+
+    y_q4 += y_step_q4;
+  }
+}