libvpx: Pull from upstream

source/libvpx/vp9/common/vp9_rtcd_defs.pl

Index: source/libvpx/vp9/common/vp9_rtcd_defs.pl
diff --git a/source/libvpx/vp9/common/vp9_rtcd_defs.pl b/source/libvpx/vp9/common/vp9_rtcd_defs.pl
index bbe200d0d6327fc7dafc62e85d5b04b1aba0ddb9..22a5efdd5b2a288c5165ea1fb9e3bbe423e31499 100644
--- a/source/libvpx/vp9/common/vp9_rtcd_defs.pl
+++ b/source/libvpx/vp9/common/vp9_rtcd_defs.pl
@@ -60,52 +60,52 @@ add_proto qw/void vp9_d207_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vp9_d207_predictor_4x4/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d45_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_d45_predictor_4x4/, "$ssse3_x86inc";
+specialize qw/vp9_d45_predictor_4x4 neon/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d63_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d63_predictor_4x4/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_h_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_h_predictor_4x4 neon dspr2/, "$ssse3_x86inc";
+specialize qw/vp9_h_predictor_4x4 neon dspr2 msa/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d117_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d117_predictor_4x4/;
 
 add_proto qw/void vp9_d135_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_d135_predictor_4x4/;
+specialize qw/vp9_d135_predictor_4x4 neon/;
 
 add_proto qw/void vp9_d153_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d153_predictor_4x4/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_v_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_v_predictor_4x4 neon/, "$sse_x86inc";
+specialize qw/vp9_v_predictor_4x4 neon msa/, "$sse_x86inc";
 
 add_proto qw/void vp9_tm_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_tm_predictor_4x4 neon dspr2/, "$sse_x86inc";
+specialize qw/vp9_tm_predictor_4x4 neon dspr2 msa/, "$sse_x86inc";
 
 add_proto qw/void vp9_dc_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_predictor_4x4 dspr2/, "$sse_x86inc";
+specialize qw/vp9_dc_predictor_4x4 dspr2 msa neon/, "$sse_x86inc";
 
 add_proto qw/void vp9_dc_top_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_top_predictor_4x4/, "$sse_x86inc";
+specialize qw/vp9_dc_top_predictor_4x4 msa neon/, "$sse_x86inc";
 
 add_proto qw/void vp9_dc_left_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_left_predictor_4x4/, "$sse_x86inc";
+specialize qw/vp9_dc_left_predictor_4x4 msa neon/, "$sse_x86inc";
 
 add_proto qw/void vp9_dc_128_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_128_predictor_4x4/, "$sse_x86inc";
+specialize qw/vp9_dc_128_predictor_4x4 msa neon/, "$sse_x86inc";
 
 add_proto qw/void vp9_d207_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d207_predictor_8x8/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d45_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_d45_predictor_8x8/, "$ssse3_x86inc";
+specialize qw/vp9_d45_predictor_8x8 neon/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d63_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d63_predictor_8x8/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_h_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_h_predictor_8x8 neon dspr2/, "$ssse3_x86inc";
+specialize qw/vp9_h_predictor_8x8 neon dspr2 msa/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d117_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d117_predictor_8x8/;
@@ -117,34 +117,34 @@ add_proto qw/void vp9_d153_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vp9_d153_predictor_8x8/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_v_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_v_predictor_8x8 neon/, "$sse_x86inc";
+specialize qw/vp9_v_predictor_8x8 neon msa/, "$sse_x86inc";
 
 add_proto qw/void vp9_tm_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_tm_predictor_8x8 neon dspr2/, "$sse2_x86inc";
+specialize qw/vp9_tm_predictor_8x8 neon dspr2 msa/, "$sse2_x86inc";
 
 add_proto qw/void vp9_dc_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_predictor_8x8 dspr2 neon/, "$sse_x86inc";
+specialize qw/vp9_dc_predictor_8x8 dspr2 neon msa/, "$sse_x86inc";
 
 add_proto qw/void vp9_dc_top_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_top_predictor_8x8 neon/, "$sse_x86inc";
+specialize qw/vp9_dc_top_predictor_8x8 neon msa/, "$sse_x86inc";
 
 add_proto qw/void vp9_dc_left_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_left_predictor_8x8 neon/, "$sse_x86inc";
+specialize qw/vp9_dc_left_predictor_8x8 neon msa/, "$sse_x86inc";
 
 add_proto qw/void vp9_dc_128_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_128_predictor_8x8 neon/, "$sse_x86inc";
+specialize qw/vp9_dc_128_predictor_8x8 neon msa/, "$sse_x86inc";
 
 add_proto qw/void vp9_d207_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d207_predictor_16x16/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d45_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_d45_predictor_16x16/, "$ssse3_x86inc";
+specialize qw/vp9_d45_predictor_16x16 neon/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d63_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d63_predictor_16x16/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_h_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_h_predictor_16x16 neon dspr2/, "$ssse3_x86inc";
+specialize qw/vp9_h_predictor_16x16 neon dspr2 msa/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d117_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d117_predictor_16x16/;
@@ -156,22 +156,22 @@ add_proto qw/void vp9_d153_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride,
 specialize qw/vp9_d153_predictor_16x16/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_v_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_v_predictor_16x16 neon/, "$sse2_x86inc";
+specialize qw/vp9_v_predictor_16x16 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vp9_tm_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_tm_predictor_16x16 neon/, "$sse2_x86inc";
+specialize qw/vp9_tm_predictor_16x16 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vp9_dc_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_predictor_16x16 dspr2 neon/, "$sse2_x86inc";
+specialize qw/vp9_dc_predictor_16x16 dspr2 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vp9_dc_top_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_top_predictor_16x16 neon/, "$sse2_x86inc";
+specialize qw/vp9_dc_top_predictor_16x16 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vp9_dc_left_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_left_predictor_16x16 neon/, "$sse2_x86inc";
+specialize qw/vp9_dc_left_predictor_16x16 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vp9_dc_128_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_128_predictor_16x16 neon/, "$sse2_x86inc";
+specialize qw/vp9_dc_128_predictor_16x16 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vp9_d207_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d207_predictor_32x32/, "$ssse3_x86inc";
@@ -183,7 +183,7 @@ add_proto qw/void vp9_d63_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, c
 specialize qw/vp9_d63_predictor_32x32/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_h_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_h_predictor_32x32 neon/, "$ssse3_x86inc";
+specialize qw/vp9_h_predictor_32x32 neon msa/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d117_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d117_predictor_32x32/;
@@ -192,25 +192,25 @@ add_proto qw/void vp9_d135_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride,
 specialize qw/vp9_d135_predictor_32x32/;
 
 add_proto qw/void vp9_d153_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_d153_predictor_32x32/;
+specialize qw/vp9_d153_predictor_32x32/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_v_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_v_predictor_32x32 neon/, "$sse2_x86inc";
+specialize qw/vp9_v_predictor_32x32 neon msa/, "$sse2_x86inc";
 
 add_proto qw/void vp9_tm_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_tm_predictor_32x32 neon/, "$sse2_x86_64";
+specialize qw/vp9_tm_predictor_32x32 neon msa/, "$sse2_x86_64";
 
 add_proto qw/void vp9_dc_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_predictor_32x32/, "$sse2_x86inc";
+specialize qw/vp9_dc_predictor_32x32 msa neon/, "$sse2_x86inc";
 
 add_proto qw/void vp9_dc_top_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_top_predictor_32x32/, "$sse2_x86inc";
+specialize qw/vp9_dc_top_predictor_32x32 msa neon/, "$sse2_x86inc";
 
 add_proto qw/void vp9_dc_left_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_left_predictor_32x32/, "$sse2_x86inc";
+specialize qw/vp9_dc_left_predictor_32x32 msa neon/, "$sse2_x86inc";
 
 add_proto qw/void vp9_dc_128_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_128_predictor_32x32/, "$sse2_x86inc";
+specialize qw/vp9_dc_128_predictor_32x32 msa neon/, "$sse2_x86inc";
 
 #
 # Loopfilter
@@ -276,10 +276,10 @@ specialize qw/vp9_plane_add_noise sse2/;
 $vp9_plane_add_noise_sse2=vp9_plane_add_noise_wmt;
 
 add_proto qw/void vp9_filter_by_weight16x16/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight";
-specialize qw/vp9_filter_by_weight16x16 sse2/;
+specialize qw/vp9_filter_by_weight16x16 sse2 msa/;
 
 add_proto qw/void vp9_filter_by_weight8x8/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight";
-specialize qw/vp9_filter_by_weight8x8 sse2/;
+specialize qw/vp9_filter_by_weight8x8 sse2 msa/;
 }
 
 #
@@ -878,10 +878,10 @@ add_proto qw/unsigned int vp9_sub_pixel_avg_variance4x4/, "const uint8_t *src_pt
 specialize qw/vp9_sub_pixel_avg_variance4x4/, "$sse_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_avg_8x8/, "const uint8_t *, int p";
-specialize qw/vp9_avg_8x8 sse2 neon/;
+specialize qw/vp9_avg_8x8 sse2 neon msa/;
 
 add_proto qw/unsigned int vp9_avg_4x4/, "const uint8_t *, int p";
-specialize qw/vp9_avg_4x4 sse2/;
+specialize qw/vp9_avg_4x4 sse2 msa/;
 
 add_proto qw/void vp9_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
 specialize qw/vp9_minmax_8x8 sse2/;
@@ -916,7 +916,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 # ENCODEMB INVOKE
 
 add_proto qw/void vp9_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
-specialize qw/vp9_subtract_block neon/, "$sse2_x86inc";
+specialize qw/vp9_subtract_block neon msa/, "$sse2_x86inc";
 
 #
 # Denoiser
@@ -948,7 +948,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vp9_fdct8x8_quant/;
 } else {
   add_proto qw/int64_t vp9_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
-  specialize qw/vp9_block_error avx2/, "$sse2_x86inc";
+  specialize qw/vp9_block_error avx2 msa/, "$sse2_x86inc";
 
   add_proto qw/int64_t vp9_block_error_fp/, "const int16_t *coeff, const int16_t *dqcoeff, int block_size";
   specialize qw/vp9_block_error_fp sse2/;
@@ -1023,43 +1023,43 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vp9_fdct32x32_rd sse2/;
 } else {
   add_proto qw/void vp9_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
-  specialize qw/vp9_fht4x4 sse2/;
+  specialize qw/vp9_fht4x4 sse2 msa/;
 
   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/;
+  specialize qw/vp9_fht16x16 sse2 msa/;
 
   add_proto qw/void vp9_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vp9_fwht4x4/, "$mmx_x86inc";
+  specialize qw/vp9_fwht4x4 msa/, "$mmx_x86inc";
 
   add_proto qw/void vp9_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
   specialize qw/vp9_fdct4x4_1 sse2/;
 
   add_proto qw/void vp9_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vp9_fdct4x4 sse2/;
+  specialize qw/vp9_fdct4x4 sse2 msa/;
 
   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/;
+  specialize qw/vp9_fdct16x16_1 sse2 msa/;
 
   add_proto qw/void vp9_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vp9_fdct16x16 sse2/;
+  specialize qw/vp9_fdct16x16 sse2 msa/;
 
   add_proto qw/void vp9_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vp9_fdct32x32_1 sse2/;
+  specialize qw/vp9_fdct32x32_1 sse2 msa/;
 
   add_proto qw/void vp9_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vp9_fdct32x32 sse2 avx2/;
+  specialize qw/vp9_fdct32x32 sse2 avx2 msa/;
 
   add_proto qw/void vp9_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
-  specialize qw/vp9_fdct32x32_rd sse2 avx2/;
+  specialize qw/vp9_fdct32x32_rd sse2 avx2 msa/;
 }
 
 #
@@ -1077,7 +1077,7 @@ add_proto qw/int vp9_full_range_search/, "const struct macroblock *x, const stru
 specialize qw/vp9_full_range_search/;
 
 add_proto qw/void vp9_temporal_filter_apply/, "uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count";
-specialize qw/vp9_temporal_filter_apply sse2/;
+specialize qw/vp9_temporal_filter_apply sse2 msa/;
 
 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {