Merge "mips msa vp9 fdct 8x8 optimization"

5 files changed, 324 insertions, 3 deletions

diff --git a/vp9/common/mips/msa/vp9_macros_msa.h b/vp9/common/mips/msa/vp9_macros_msa.h
index 863b2dc15..23b281f1f 100644
--- a/vp9/common/mips/msa/vp9_macros_msa.h
+++ b/vp9/common/mips/msa/vp9_macros_msa.h
@@ -856,6 +856,15 @@
 }
 #define INSERT_W2_SB(...) INSERT_W2(v16i8, __VA_ARGS__)
 
+#define INSERT_W4(RTYPE, in0, in1, in2, in3, out) {  \
+  out = (RTYPE)__msa_insert_w((v4i32)out, 0, in0);   \
+  out = (RTYPE)__msa_insert_w((v4i32)out, 1, in1);   \
+  out = (RTYPE)__msa_insert_w((v4i32)out, 2, in2);   \
+  out = (RTYPE)__msa_insert_w((v4i32)out, 3, in3);   \
+}
+#define INSERT_W4_UB(...) INSERT_W4(v16u8, __VA_ARGS__)
+#define INSERT_W4_SB(...) INSERT_W4(v16i8, __VA_ARGS__)
+
 /* Description : Insert specified double word elements from input vectors to 1
                  destination vector
    Arguments   : Inputs  - in0, in1      (2 input vectors)
@@ -902,6 +911,19 @@
 #define ILVEV_H2_SH(...) ILVEV_H2(v8i16, __VA_ARGS__)
 #define ILVEV_H2_SW(...) ILVEV_H2(v4i32, __VA_ARGS__)
 
+/* Description : Interleave even word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even word elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_W2(RTYPE, in0, in1, in2, in3, out0, out1) {  \
+  out0 = (RTYPE)__msa_ilvev_w((v4i32)in1, (v4i32)in0);     \
+  out1 = (RTYPE)__msa_ilvev_w((v4i32)in3, (v4i32)in2);     \
+}
+#define ILVEV_W2_SB(...) ILVEV_W2(v16i8, __VA_ARGS__)
+
 /* Description : Interleave even double word elements from vectors
    Arguments   : Inputs  - in0, in1, in2, in3
                  Outputs - out0, out1
@@ -1296,6 +1318,24 @@
 }
 #define XORI_B7_128_SB(...) XORI_B7_128(v16i8, __VA_ARGS__)
 
+/* Description : Average of signed halfword elements -> (a + b) / 2
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3
+                 Return Type - as per RTYPE
+   Details     : Each signed halfword element from 'in0' is added to each
+                 signed halfword element of 'in1' with full precision resulting
+                 in one extra bit in the result. The result is then divided by
+                 2 and written to 'out0'
+*/
+#define AVE_SH4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7,  \
+                out0, out1, out2, out3) {                       \
+  out0 = (RTYPE)__msa_ave_s_h((v8i16)in0, (v8i16)in1);          \
+  out1 = (RTYPE)__msa_ave_s_h((v8i16)in2, (v8i16)in3);          \
+  out2 = (RTYPE)__msa_ave_s_h((v8i16)in4, (v8i16)in5);          \
+  out3 = (RTYPE)__msa_ave_s_h((v8i16)in6, (v8i16)in7);          \
+}
+#define AVE_SH4_SH(...) AVE_SH4(v8i16, __VA_ARGS__)
+
 /* Description : Addition of signed halfword elements and signed saturation
    Arguments   : Inputs  - in0, in1, in2, in3
                  Outputs - out0, out1
@@ -1350,6 +1390,27 @@
   in3 = in3 >> shift;                        \
 }
 
+/* Description : Shift right arithmetic rounded words
+   Arguments   : Inputs  - in0, in1, shift
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each element of vector 'in0' is shifted right arithmetically by
+                 the number of bits in the corresponding element in the vector
+                 'shift'. The last discarded bit is added to shifted value for
+                 rounding and the result is written in-place.
+                 'shift' is a vector.
+*/
+#define SRAR_W2(RTYPE, in0, in1, shift) {               \
+  in0 = (RTYPE)__msa_srar_w((v4i32)in0, (v4i32)shift);  \
+  in1 = (RTYPE)__msa_srar_w((v4i32)in1, (v4i32)shift);  \
+}
+
+#define SRAR_W4(RTYPE, in0, in1, in2, in3, shift) {  \
+  SRAR_W2(RTYPE, in0, in1, shift)                    \
+  SRAR_W2(RTYPE, in2, in3, shift)                    \
+}
+#define SRAR_W4_SW(...) SRAR_W4(v4i32, __VA_ARGS__)
+
 /* Description : Shift right arithmetic rounded (immediate)
    Arguments   : Inputs  - in0, in1, in2, in3, shift
                  Outputs - in0, in1, in2, in3 (in place)
@@ -1396,6 +1457,21 @@
 }
 #define SRARI_W4_SW(...) SRARI_W4(v4i32, __VA_ARGS__)
 
+/* Description : Logical shift right all elements of vector (immediate)
+   Arguments   : Inputs  - in0, in1, in2, in3, shift
+                 Outputs - out0, out1, out2, out3
+                 Return Type - as per RTYPE
+   Details     : Each element of vector 'in0' is right shifted by 'shift' and
+                 the result is written in-place. 'shift' is an immediate value.
+*/
+#define SRLI_H4(RTYPE, in0, in1, in2, in3, out0, out1, out2, out3, shift) {  \
+  out0 = (RTYPE)__msa_srli_h((v8i16)in0, shift);                             \
+  out1 = (RTYPE)__msa_srli_h((v8i16)in1, shift);                             \
+  out2 = (RTYPE)__msa_srli_h((v8i16)in2, shift);                             \
+  out3 = (RTYPE)__msa_srli_h((v8i16)in3, shift);                             \
+}
+#define SRLI_H4_SH(...) SRLI_H4(v8i16, __VA_ARGS__)
+
 /* Description : Addition of 2 pairs of vectors
    Arguments   : Inputs  - in0, in1, in2, in3
                  Outputs - out0, out1
diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 27cd3d010..dac142397 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -1026,7 +1026,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vp9_fht4x4 sse2/;
 
   add_proto qw/void vp9_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
-  specialize qw/vp9_fht8x8 sse2/;
+  specialize qw/vp9_fht8x8 sse2 msa/;
 
   add_proto qw/void vp9_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
   specialize qw/vp9_fht16x16 sse2 msa/;
@@ -1041,10 +1041,10 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vp9_fdct4x4 sse2/;
 
   add_proto qw/void vp9_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vp9_fdct8x8_1 sse2 neon/;
+  specialize qw/vp9_fdct8x8_1 sse2 neon msa/;
 
   add_proto qw/void vp9_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vp9_fdct8x8 sse2 neon/, "$ssse3_x86_64";
+  specialize qw/vp9_fdct8x8 sse2 neon msa/, "$ssse3_x86_64";
 
   add_proto qw/void vp9_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
   specialize qw/vp9_fdct16x16_1 sse2 msa/;
diff --git a/vp9/encoder/mips/msa/vp9_fdct8x8_msa.c b/vp9/encoder/mips/msa/vp9_fdct8x8_msa.c
new file mode 100644
index 000000000..68e65723a
--- /dev/null
+++ b/vp9/encoder/mips/msa/vp9_fdct8x8_msa.c
@@ -0,0 +1,90 @@
+/*
+ *  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 <assert.h>
+
+#include "./vp9_rtcd.h"
+#include "vp9/encoder/mips/msa/vp9_fdct_msa.h"
+
+void vp9_fdct8x8_msa(const int16_t *input, int16_t *output,
+                     int32_t src_stride) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+
+  LD_SH8(input, src_stride, in0, in1, in2, in3, in4, in5, in6, in7);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+  VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+            in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+            in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  VP9_SRLI_AVE_S_4V_H(in0, in1, in2, in3, in4, in5, in6, in7);
+  ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output, 8);
+}
+
+void vp9_fdct8x8_1_msa(const int16_t *input, int16_t *out, int32_t stride) {
+  out[0] = VP9_LD_HADD(input, stride);
+  out[1] = 0;
+}
+
+void vp9_fht8x8_msa(const int16_t *input, int16_t *output, int32_t stride,
+                    int32_t tx_type) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+
+  LD_SH8(input, stride, in0, in1, in2, in3, in4, in5, in6, in7);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    case ADST_DCT:
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    case DCT_ADST:
+      VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    case ADST_ADST:
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                         in0, in1, in2, in3, in4, in5, in6, in7);
+      VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,
+                in0, in1, in2, in3, in4, in5, in6, in7);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7,
+                     in0, in1, in2, in3, in4, in5, in6, in7);
+  VP9_SRLI_AVE_S_4V_H(in0, in1, in2, in3, in4, in5, in6, in7);
+  ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output, 8);
+}
diff --git a/vp9/encoder/mips/msa/vp9_fdct_msa.h b/vp9/encoder/mips/msa/vp9_fdct_msa.h
index 99d299a92..3523a3d51 100644
--- a/vp9/encoder/mips/msa/vp9_fdct_msa.h
+++ b/vp9/encoder/mips/msa/vp9_fdct_msa.h
@@ -33,6 +33,23 @@
   out1 = __msa_pckev_h((v8i16)s0_m, (v8i16)s1_m);                    \
 }
 
+#define VP9_DOT_ADD_SUB_SRARI_PCK(in0, in1, in2, in3, in4, in5, in6, in7,  \
+                                  dst0, dst1, dst2, dst3) {                \
+  v4i32 tp0_m, tp1_m, tp2_m, tp3_m, tp4_m;                                 \
+  v4i32 tp5_m, tp6_m, tp7_m, tp8_m, tp9_m;                                 \
+                                                                           \
+  DOTP_SH4_SW(in0, in1, in0, in1, in4, in4, in5, in5,                      \
+              tp0_m, tp2_m, tp3_m, tp4_m);                                 \
+  DOTP_SH4_SW(in2, in3, in2, in3, in6, in6, in7, in7,                      \
+              tp5_m, tp6_m, tp7_m, tp8_m);                                 \
+  BUTTERFLY_4(tp0_m, tp3_m, tp7_m, tp5_m, tp1_m, tp9_m, tp7_m, tp5_m);     \
+  BUTTERFLY_4(tp2_m, tp4_m, tp8_m, tp6_m, tp3_m, tp0_m, tp4_m, tp2_m);     \
+  SRARI_W4_SW(tp1_m, tp9_m, tp7_m, tp5_m, DCT_CONST_BITS);                 \
+  SRARI_W4_SW(tp3_m, tp0_m, tp4_m, tp2_m, DCT_CONST_BITS);                 \
+  PCKEV_H4_SH(tp1_m, tp3_m, tp9_m, tp0_m, tp7_m, tp4_m, tp5_m, tp2_m,      \
+              dst0, dst1, dst2, dst3);                                     \
+}
+
 #define VP9_DOT_SHIFT_RIGHT_PCK_H(in0, in1, in2) ({   \
   v8i16 dst_m;                                        \
   v4i32 tp0_m, tp1_m;                                 \
@@ -44,6 +61,72 @@
   dst_m;                                              \
 })
 
+#define VP9_ADST8(in0, in1, in2, in3, in4, in5, in6, in7,                   \
+                  out0, out1, out2, out3, out4, out5, out6, out7) {         \
+  v8i16 cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst4_m;                        \
+  v8i16 vec0_m, vec1_m, vec2_m, vec3_m, s0_m, s1_m;                         \
+  v8i16 coeff0_m = { cospi_2_64, cospi_6_64, cospi_10_64, cospi_14_64,      \
+                     cospi_18_64, cospi_22_64, cospi_26_64, cospi_30_64 };  \
+  v8i16 coeff1_m = { cospi_8_64, -cospi_8_64, cospi_16_64, -cospi_16_64,    \
+                     cospi_24_64, -cospi_24_64, 0, 0 };                     \
+                                                                            \
+  SPLATI_H2_SH(coeff0_m, 0, 7, cnst0_m, cnst1_m);                           \
+  cnst2_m = -cnst0_m;                                                       \
+  ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);        \
+  SPLATI_H2_SH(coeff0_m, 4, 3, cnst2_m, cnst3_m);                           \
+  cnst4_m = -cnst2_m;                                                       \
+  ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);        \
+                                                                            \
+  ILVRL_H2_SH(in0, in7, vec1_m, vec0_m);                                    \
+  ILVRL_H2_SH(in4, in3, vec3_m, vec2_m);                                    \
+  VP9_DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m,        \
+                            cnst1_m, cnst2_m, cnst3_m, in7, in0,            \
+                            in4, in3);                                      \
+                                                                            \
+  SPLATI_H2_SH(coeff0_m, 2, 5, cnst0_m, cnst1_m);                           \
+  cnst2_m = -cnst0_m;                                                       \
+  ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);        \
+  SPLATI_H2_SH(coeff0_m, 6, 1, cnst2_m, cnst3_m);                           \
+  cnst4_m = -cnst2_m;                                                       \
+  ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);        \
+                                                                            \
+  ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                    \
+  ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                    \
+                                                                            \
+  VP9_DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m,        \
+                            cnst1_m, cnst2_m, cnst3_m, in5, in2,            \
+                            in6, in1);                                      \
+  BUTTERFLY_4(in7, in0, in2, in5, s1_m, s0_m, in2, in5);                    \
+  out7 = -s0_m;                                                             \
+  out0 = s1_m;                                                              \
+                                                                            \
+  SPLATI_H4_SH(coeff1_m, 0, 4, 1, 5, cnst0_m, cnst1_m, cnst2_m, cnst3_m);   \
+                                                                            \
+  ILVEV_H2_SH(cnst3_m, cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst2_m);        \
+  cnst0_m = __msa_ilvev_h(cnst1_m, cnst0_m);                                \
+  cnst1_m = cnst0_m;                                                        \
+                                                                            \
+  ILVRL_H2_SH(in4, in3, vec1_m, vec0_m);                                    \
+  ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                    \
+  VP9_DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m,        \
+                            cnst2_m, cnst3_m, cnst1_m, out1, out6,          \
+                            s0_m, s1_m);                                    \
+                                                                            \
+  SPLATI_H2_SH(coeff1_m, 2, 3, cnst0_m, cnst1_m);                           \
+  cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                                \
+                                                                            \
+  ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                    \
+  ILVRL_H2_SH(s0_m, s1_m, vec3_m, vec2_m);                                  \
+  out3 = VP9_DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                \
+  out4 = VP9_DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst1_m);                \
+  out2 = VP9_DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst0_m);                \
+  out5 = VP9_DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst1_m);                \
+                                                                            \
+  out1 = -out1;                                                             \
+  out3 = -out3;                                                             \
+  out5 = -out5;                                                             \
+}
+
 #define VP9_MADD_SHORT(m0, m1, c0, c1, res0, res1) {                \
   v4i32 madd0_m, madd1_m, madd2_m, madd3_m;                         \
   v8i16 madd_s0_m, madd_s1_m;                                       \
@@ -107,6 +190,77 @@
   vec1 >>= 2;                                     \
 }
 
+#define VP9_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7,            \
+                  out0, out1, out2, out3, out4, out5, out6, out7) {  \
+  v8i16 s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m;                    \
+  v8i16 s7_m, x0_m, x1_m, x2_m, x3_m;                                \
+  v8i16 coeff_m = { cospi_16_64, -cospi_16_64, cospi_8_64,           \
+                    cospi_24_64, cospi_4_64, cospi_28_64,            \
+                    cospi_12_64, cospi_20_64 };                      \
+                                                                     \
+  /* FDCT stage1 */                                                  \
+  BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7,                \
+              s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m, s7_m);       \
+  BUTTERFLY_4(s0_m, s1_m, s2_m, s3_m, x0_m, x1_m, x2_m, x3_m);       \
+  ILVL_H2_SH(x1_m, x0_m, x3_m, x2_m, s0_m, s2_m);                    \
+  ILVR_H2_SH(x1_m, x0_m, x3_m, x2_m, s1_m, s3_m);                    \
+  SPLATI_H2_SH(coeff_m, 0, 1, x0_m, x1_m);                           \
+  x1_m = __msa_ilvev_h(x1_m, x0_m);                                  \
+  out4 = VP9_DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x1_m);                \
+                                                                     \
+  SPLATI_H2_SH(coeff_m, 2, 3, x2_m, x3_m);                           \
+  x2_m = -x2_m;                                                      \
+  x2_m = __msa_ilvev_h(x3_m, x2_m);                                  \
+  out6 = VP9_DOT_SHIFT_RIGHT_PCK_H(s2_m, s3_m, x2_m);                \
+                                                                     \
+  out0 = VP9_DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x0_m);                \
+  x2_m = __msa_splati_h(coeff_m, 2);                                 \
+  x2_m = __msa_ilvev_h(x2_m, x3_m);                                  \
+  out2 = VP9_DOT_SHIFT_RIGHT_PCK_H(s2_m, s3_m, x2_m);                \
+                                                                     \
+  /* stage2 */                                                       \
+  ILVRL_H2_SH(s5_m, s6_m, s1_m, s0_m);                               \
+                                                                     \
+  s6_m = VP9_DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x0_m);                \
+  s5_m = VP9_DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x1_m);                \
+                                                                     \
+  /* stage3 */                                                       \
+  BUTTERFLY_4(s4_m, s7_m, s6_m, s5_m, x0_m, x3_m, x2_m, x1_m);       \
+                                                                     \
+  /* stage4 */                                                       \
+  ILVL_H2_SH(x3_m, x0_m, x2_m, x1_m, s4_m, s6_m);                    \
+  ILVR_H2_SH(x3_m, x0_m, x2_m, x1_m, s5_m, s7_m);                    \
+                                                                     \
+  SPLATI_H2_SH(coeff_m, 4, 5, x0_m, x1_m);                           \
+  x1_m = __msa_ilvev_h(x0_m, x1_m);                                  \
+  out1 = VP9_DOT_SHIFT_RIGHT_PCK_H(s4_m, s5_m, x1_m);                \
+                                                                     \
+  SPLATI_H2_SH(coeff_m, 6, 7, x2_m, x3_m);                           \
+  x2_m = __msa_ilvev_h(x3_m, x2_m);                                  \
+  out5 = VP9_DOT_SHIFT_RIGHT_PCK_H(s6_m, s7_m, x2_m);                \
+                                                                     \
+  x1_m = __msa_splati_h(coeff_m, 5);                                 \
+  x0_m = -x0_m;                                                      \
+  x0_m = __msa_ilvev_h(x1_m, x0_m);                                  \
+  out7 = VP9_DOT_SHIFT_RIGHT_PCK_H(s4_m, s5_m, x0_m);                \
+                                                                     \
+  x2_m = __msa_splati_h(coeff_m, 6);                                 \
+  x3_m = -x3_m;                                                      \
+  x2_m = __msa_ilvev_h(x2_m, x3_m);                                  \
+  out3 = VP9_DOT_SHIFT_RIGHT_PCK_H(s6_m, s7_m, x2_m);                \
+}
+
+#define VP9_SRLI_AVE_S_4V_H(in0, in1, in2, in3, in4, in5, in6, in7) {    \
+  v8i16 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;  \
+                                                                         \
+  SRLI_H4_SH(in0, in1, in2, in3, vec0_m, vec1_m, vec2_m, vec3_m, 15);    \
+  SRLI_H4_SH(in4, in5, in6, in7, vec4_m, vec5_m, vec6_m, vec7_m, 15);    \
+  AVE_SH4_SH(vec0_m, in0, vec1_m, in1, vec2_m, in2, vec3_m, in3,         \
+             in0, in1, in2, in3);                                        \
+  AVE_SH4_SH(vec4_m, in4, vec5_m, in5, vec6_m, in6, vec7_m, in7,         \
+             in4, in5, in6, in7);                                        \
+}
+
 #define VP9_FDCT8x16_EVEN(in0, in1, in2, in3, in4, in5, in6, in7,            \
                           out0, out1, out2, out3, out4, out5, out6, out7) {  \
   v8i16 s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m, s7_m;                      \
diff --git a/vp9/vp9cx.mk b/vp9/vp9cx.mk
index 6074da21a..4743c7c70 100644
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -152,6 +152,7 @@ VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_quantize_neon.c
 VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_subtract_neon.c
 VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_variance_neon.c
 
+VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct8x8_msa.c
 VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct16x16_msa.c
 VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct32x32_msa.c
 VP9_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/vp9_fdct_msa.h