diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index ecfd332ef..24e5b8b91 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -563,234 +563,162 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 #
 # Inverse transform
 if (vpx_config("CONFIG_VP9") eq "yes") {
+
+add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+
+if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") {
+  specialize qw/vpx_idct4x4_16_add neon sse2/;
+
+  specialize qw/vpx_idct4x4_1_add neon sse2/;
+
+  specialize qw/vpx_idct8x8_64_add neon sse2 ssse3/;
+
+  specialize qw/vpx_idct8x8_12_add neon sse2 ssse3/;
+
+  specialize qw/vpx_idct8x8_1_add neon sse2/;
+
+  specialize qw/vpx_idct16x16_256_add neon sse2/;
+
+  specialize qw/vpx_idct16x16_38_add neon sse2/;
+  $vpx_idct16x16_38_add_sse2=vpx_idct16x16_256_add_sse2;
+
+  specialize qw/vpx_idct16x16_10_add neon sse2/;
+
+  specialize qw/vpx_idct16x16_1_add neon sse2/;
+
+  specialize qw/vpx_idct32x32_1024_add neon sse2 ssse3/;
+
+  specialize qw/vpx_idct32x32_135_add neon sse2 ssse3/;
+  $vpx_idct32x32_135_add_sse2=vpx_idct32x32_1024_add_sse2;
+
+  specialize qw/vpx_idct32x32_34_add neon sse2 ssse3/;
+
+  specialize qw/vpx_idct32x32_1_add neon sse2/;
+}  # !CONFIG_EMULATE_HARDWARE
+
 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   # Note as optimized versions of these functions are added we need to add a check to ensure
   # that when CONFIG_EMULATE_HARDWARE is on, it defaults to the C versions only.
-  add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-  add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride";
   specialize qw/vpx_iwht4x4_16_add sse2/;
 
+  add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
+
   add_proto qw/void vpx_highbd_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
   specialize qw/vpx_highbd_idct4x4_1_add neon/;
 
+  add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
+
+  add_proto qw/void vpx_highbd_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
+
   add_proto qw/void vpx_highbd_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
   specialize qw/vpx_highbd_idct8x8_1_add neon/;
 
+  add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
+
+  add_proto qw/void vpx_highbd_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
+
+  add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
+
   add_proto qw/void vpx_highbd_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
   specialize qw/vpx_highbd_idct16x16_1_add neon/;
 
   add_proto qw/void vpx_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
 
+  add_proto qw/void vpx_highbd_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
+
   add_proto qw/void vpx_highbd_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
 
   add_proto qw/void vpx_highbd_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
   specialize qw/vpx_highbd_idct32x32_1_add neon sse2/;
 
-  add_proto qw/void vpx_highbd_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
-
   add_proto qw/void vpx_highbd_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
 
-  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
-  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
-    add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+  add_proto qw/void vpx_highbd_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
 
-    add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
-
-    add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
-
-    add_proto qw/void vpx_highbd_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
-
-    add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
-
-    add_proto qw/void vpx_highbd_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
-
-    add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
-
-    add_proto qw/void vpx_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
-
-    add_proto qw/void vpx_highbd_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
-
-    add_proto qw/void vpx_highbd_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
-  } else {
-    add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct4x4_16_add neon sse2/;
-
-    add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct4x4_1_add neon sse2/;
-
-    add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct8x8_64_add neon sse2 ssse3/;
-
-    add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct8x8_12_add neon sse2 ssse3/;
-
-    add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct8x8_1_add neon sse2/;
-
-    add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct16x16_256_add neon sse2/;
-
-    add_proto qw/void vpx_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct16x16_38_add neon sse2/;
-    $vpx_idct16x16_38_add_sse2=vpx_idct16x16_256_add_sse2;
-
-    add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct16x16_10_add neon sse2/;
-
-    add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct16x16_1_add neon sse2/;
-
-    add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct32x32_1024_add neon sse2 ssse3/;
-
-    add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct32x32_135_add neon sse2 ssse3/;
-    # Need to add 135 eob idct32x32 implementations.
-    $vpx_idct32x32_135_add_sse2=vpx_idct32x32_1024_add_sse2;
-
-    add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct32x32_34_add neon sse2 ssse3/;
-
-    add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct32x32_1_add neon sse2/;
-
-    add_proto qw/void vpx_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
+  if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") {
     specialize qw/vpx_highbd_idct4x4_16_add neon sse2/;
 
-    add_proto qw/void vpx_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
     specialize qw/vpx_highbd_idct8x8_64_add neon sse2/;
 
-    add_proto qw/void vpx_highbd_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
     specialize qw/vpx_highbd_idct8x8_12_add neon sse2/;
 
-    add_proto qw/void vpx_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
     specialize qw/vpx_highbd_idct16x16_256_add neon sse2/;
 
-    add_proto qw/void vpx_highbd_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
     specialize qw/vpx_highbd_idct16x16_38_add neon sse2/;
     $vpx_highbd_idct16x16_38_add_sse2=vpx_highbd_idct16x16_256_add_sse2;
 
-    add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
     specialize qw/vpx_highbd_idct16x16_10_add neon sse2/;
 
-    add_proto qw/void vpx_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
     specialize qw/vpx_highbd_idct32x32_1024_add neon/;
 
-    add_proto qw/void vpx_highbd_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
     specialize qw/vpx_highbd_idct32x32_135_add neon/;
 
-    add_proto qw/void vpx_highbd_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
     specialize qw/vpx_highbd_idct32x32_34_add neon/;
-  }  # CONFIG_EMULATE_HARDWARE
+  }  # !CONFIG_EMULATE_HARDWARE
 } else {
-  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
-  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
-    add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+  if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") {
+    specialize qw/vpx_idct4x4_16_add dspr2 msa/;
 
-    add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+    specialize qw/vpx_idct4x4_1_add dspr2 msa/;
 
-    add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+    specialize qw/vpx_idct8x8_64_add dspr2 msa/;
 
-    add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+    specialize qw/vpx_idct8x8_12_add dspr2 msa/;
 
-    add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+    specialize qw/vpx_idct8x8_1_add dspr2 msa/;
 
-    add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
+    specialize qw/vpx_idct16x16_256_add dspr2 msa/;
 
-    add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-
-    add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-  } else {
-    add_proto qw/void vpx_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct4x4_1_add sse2 neon dspr2 msa/;
-
-    add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct4x4_16_add sse2 neon dspr2 msa/;
-
-    add_proto qw/void vpx_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct8x8_1_add sse2 neon dspr2 msa/;
-
-    add_proto qw/void vpx_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct8x8_64_add sse2 ssse3 neon dspr2 msa/;
-
-    add_proto qw/void vpx_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct8x8_12_add sse2 ssse3 neon dspr2 msa/;
-
-    add_proto qw/void vpx_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct16x16_1_add sse2 neon dspr2 msa/;
-
-    add_proto qw/void vpx_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct16x16_256_add sse2 neon dspr2 msa/;
-
-    add_proto qw/void vpx_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct16x16_38_add sse2 neon dspr2 msa/;
-    $vpx_idct16x16_38_add_sse2=vpx_idct16x16_256_add_sse2;
+    specialize qw/vpx_idct16x16_38_add dspr2 msa/;
     $vpx_idct16x16_38_add_dspr2=vpx_idct16x16_256_add_dspr2;
     $vpx_idct16x16_38_add_msa=vpx_idct16x16_256_add_msa;
 
-    add_proto qw/void vpx_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct16x16_10_add sse2 neon dspr2 msa/;
+    specialize qw/vpx_idct16x16_10_add dspr2 msa/;
 
-    add_proto qw/void vpx_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct32x32_1024_add sse2 ssse3 neon dspr2 msa/;
+    specialize qw/vpx_idct16x16_1_add dspr2 msa/;
 
-    add_proto qw/void vpx_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct32x32_135_add sse2 ssse3 neon dspr2 msa/;
-    $vpx_idct32x32_135_add_sse2=vpx_idct32x32_1024_add_sse2;
+    specialize qw/vpx_idct32x32_1024_add dspr2 msa/;
+
+    specialize qw/vpx_idct32x32_135_add dspr2 msa/;
     $vpx_idct32x32_135_add_dspr2=vpx_idct32x32_1024_add_dspr2;
     $vpx_idct32x32_135_add_msa=vpx_idct32x32_1024_add_msa;
 
-    add_proto qw/void vpx_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct32x32_34_add sse2 ssse3 neon dspr2 msa/;
+    specialize qw/vpx_idct32x32_34_add dspr2 msa/;
 
-    add_proto qw/void vpx_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_idct32x32_1_add sse2 neon dspr2 msa/;
+    specialize qw/vpx_idct32x32_1_add dspr2 msa/;
 
-    add_proto qw/void vpx_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int stride";
-    specialize qw/vpx_iwht4x4_1_add msa/;
-
-    add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride";
     specialize qw/vpx_iwht4x4_16_add msa sse2/;
-  }  # CONFIG_EMULATE_HARDWARE
+
+    specialize qw/vpx_iwht4x4_1_add msa/;
+  }  # !CONFIG_EMULATE_HARDWARE
 }  # CONFIG_VP9_HIGHBITDEPTH
 }  # CONFIG_VP9