Merge "Fix mips msa fwd xform mismatch"

3 files changed, 45 insertions, 37 deletions

diff --git a/vpx_dsp/mips/fwd_dct32x32_msa.c b/vpx_dsp/mips/fwd_dct32x32_msa.c
index e41a90480..06fdc951e 100644
--- a/vpx_dsp/mips/fwd_dct32x32_msa.c
+++ b/vpx_dsp/mips/fwd_dct32x32_msa.c
@@ -927,21 +927,21 @@ void vpx_fdct32x32_rd_msa(const int16_t *input, int16_t *out,
 }
 
 void vpx_fdct32x32_1_msa(const int16_t *input, int16_t *out, int32_t stride) {
-  int sum = LD_HADD(input, stride);
-  sum += LD_HADD(input + 8, stride);
-  sum += LD_HADD(input + 16, stride);
-  sum += LD_HADD(input + 24, stride);
-  sum += LD_HADD(input + 32 * 8, stride);
-  sum += LD_HADD(input + 32 * 8 + 8, stride);
-  sum += LD_HADD(input + 32 * 8 + 16, stride);
-  sum += LD_HADD(input + 32 * 8 + 24, stride);
-  sum += LD_HADD(input + 32 * 16, stride);
-  sum += LD_HADD(input + 32 * 16 + 8, stride);
-  sum += LD_HADD(input + 32 * 16 + 16, stride);
-  sum += LD_HADD(input + 32 * 16 + 24, stride);
-  sum += LD_HADD(input + 32 * 24, stride);
-  sum += LD_HADD(input + 32 * 24 + 8, stride);
-  sum += LD_HADD(input + 32 * 24 + 16, stride);
-  sum += LD_HADD(input + 32 * 24 + 24, stride);
+  int sum, i;
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v4i32 vec_w = { 0 };
+
+  for (i = 0; i < 16; ++i) {
+    LD_SH4(input, 8, in0, in1, in2, in3);
+    input += stride;
+    LD_SH4(input, 8, in4, in5, in6, in7);
+    input += stride;
+    ADD4(in0, in1, in2, in3, in4, in5, in6, in7, in0, in2, in4, in6);
+    ADD2(in0, in2, in4, in6, in0, in4);
+    vec_w += __msa_hadd_s_w(in0, in0);
+    vec_w += __msa_hadd_s_w(in4, in4);
+  }
+
+  sum = HADD_SW_S32(vec_w);
   out[0] = (int16_t)(sum >> 3);
 }
diff --git a/vpx_dsp/mips/fwd_txfm_msa.c b/vpx_dsp/mips/fwd_txfm_msa.c
index fdead5050..f786664bb 100644
--- a/vpx_dsp/mips/fwd_txfm_msa.c
+++ b/vpx_dsp/mips/fwd_txfm_msa.c
@@ -216,7 +216,15 @@ void vpx_fdct8x8_msa(const int16_t *input, int16_t *output,
 }
 
 void vpx_fdct8x8_1_msa(const int16_t *input, int16_t *out, int32_t stride) {
-  out[0] = LD_HADD(input, stride);
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v4i32 vec_w;
+
+  LD_SH8(input, stride, in0, in1, in2, in3, in4, in5, in6, in7);
+  ADD4(in0, in1, in2, in3, in4, in5, in6, in7, in0, in2, in4, in6);
+  ADD2(in0, in2, in4, in6, in0, in4);
+  vec_w = __msa_hadd_s_w(in0, in0);
+  vec_w += __msa_hadd_s_w(in4, in4);
+  out[0] = HADD_SW_S32(vec_w);
   out[1] = 0;
 }
 
@@ -237,9 +245,25 @@ void vpx_fdct16x16_msa(const int16_t *input, int16_t *output,
 }
 
 void vpx_fdct16x16_1_msa(const int16_t *input, int16_t *out, int32_t stride) {
-  int sum = LD_HADD(input, stride);
-  sum += LD_HADD(input + 8, stride);
-  sum += LD_HADD(input + 16 * 8, stride);
-  sum += LD_HADD(input + 16 * 8 + 8, stride);
+  int sum, i;
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v4i32 vec_w = { 0 };
+
+  for (i = 0; i < 4; ++i) {
+    LD_SH2(input, 8, in0, in1);
+    input += stride;
+    LD_SH2(input, 8, in2, in3);
+    input += stride;
+    LD_SH2(input, 8, in4, in5);
+    input += stride;
+    LD_SH2(input, 8, in6, in7);
+    input += stride;
+    ADD4(in0, in1, in2, in3, in4, in5, in6, in7, in0, in2, in4, in6);
+    ADD2(in0, in2, in4, in6, in0, in4);
+    vec_w += __msa_hadd_s_w(in0, in0);
+    vec_w += __msa_hadd_s_w(in4, in4);
+  }
+
+  sum = HADD_SW_S32(vec_w);
   out[0] = (int16_t)(sum >> 1);
 }
diff --git a/vpx_dsp/mips/fwd_txfm_msa.h b/vpx_dsp/mips/fwd_txfm_msa.h
index db5e90e7b..fd589224d 100644
--- a/vpx_dsp/mips/fwd_txfm_msa.h
+++ b/vpx_dsp/mips/fwd_txfm_msa.h
@@ -14,22 +14,6 @@
 #include "vpx_dsp/mips/txfm_macros_msa.h"
 #include "vpx_dsp/txfm_common.h"
 
-#define LD_HADD(psrc, stride)                                                  \
-  ({                                                                           \
-    v8i16 in0_m, in1_m, in2_m, in3_m, in4_m, in5_m, in6_m, in7_m;              \
-    v4i32 vec_w_m;                                                             \
-                                                                               \
-    LD_SH4((psrc), stride, in0_m, in1_m, in2_m, in3_m);                        \
-    ADD2(in0_m, in1_m, in2_m, in3_m, in0_m, in2_m);                            \
-    LD_SH4(((psrc) + 4 * stride), stride, in4_m, in5_m, in6_m, in7_m);         \
-    ADD4(in4_m, in5_m, in6_m, in7_m, in0_m, in2_m, in4_m, in6_m, in4_m, in6_m, \
-         in0_m, in4_m);                                                        \
-    in0_m += in4_m;                                                            \
-                                                                               \
-    vec_w_m = __msa_hadd_s_w(in0_m, in0_m);                                    \
-    HADD_SW_S32(vec_w_m);                                                      \
-  })
-
 #define VP9_FDCT4(in0, in1, in2, in3, out0, out1, out2, out3)                  \
   {                                                                            \
     v8i16 cnst0_m, cnst1_m, cnst2_m, cnst3_m;                                  \