Backports highbitdepth accelerations into vp10

[libvpx] / vp10 / common / vp10_rtcd_defs.pl

diff --git a/vp10/common/vp10_rtcd_defs.pl b/vp10/common/vp10_rtcd_defs.pl
index 78cc2d042655ce5f37e0bb12f1178a00e8de536f..4404701878dab452c48c21986e2c0f096ba06c3f 100644 (file)
--- a/vp10/common/vp10_rtcd_defs.pl
+++ b/vp10/common/vp10_rtcd_defs.pl
@@ -1,12 +1,12 @@
 sub vp10_common_forward_decls() {
 print <<EOF
 /*
- * VP9
+ * VP10
  */
 
 #include "vpx/vpx_integer.h"
-#include "vp10/common/vp9_common.h"
-#include "vp10/common/vp9_enums.h"
+#include "vp10/common/common.h"
+#include "vp10/common/enums.h"
 
 struct macroblockd;
 
@@ -87,14 +87,127 @@ specialize qw/vp10_filter_by_weight8x8 sse2 msa/;
 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 vp10_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
-  specialize qw/vp10_iht4x4_16_add/;
+  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
+    add_proto qw/void vp10_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp10_iht4x4_16_add/;
+
+    add_proto qw/void vp10_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp10_iht8x8_64_add/;
+
+    add_proto qw/void vp10_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+    specialize qw/vp10_iht16x16_256_add/;
+
+    add_proto qw/void vp10_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct4x4/;
+
+    add_proto qw/void vp10_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct4x4_1/;
+
+    add_proto qw/void vp10_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct8x8/;
+
+    add_proto qw/void vp10_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct8x8_1/;
+
+    add_proto qw/void vp10_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct16x16/;
+
+    add_proto qw/void vp10_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct16x16_1/;
+
+    add_proto qw/void vp10_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32/;
+
+    add_proto qw/void vp10_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32_rd/;
+
+    add_proto qw/void vp10_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32_1/;
+
+    add_proto qw/void vp10_highbd_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct4x4/;
+
+    add_proto qw/void vp10_highbd_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct8x8/;
+
+    add_proto qw/void vp10_highbd_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct8x8_1/;
+
+    add_proto qw/void vp10_highbd_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct16x16/;
+
+    add_proto qw/void vp10_highbd_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct16x16_1/;
+
+    add_proto qw/void vp10_highbd_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct32x32/;
+
+    add_proto qw/void vp10_highbd_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct32x32_rd/;
+
+    add_proto qw/void vp10_highbd_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct32x32_1/;
+  } else {
+    add_proto qw/void vp10_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp10_iht4x4_16_add sse2/;
+
+    add_proto qw/void vp10_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/vp10_iht8x8_64_add sse2/;
+
+    add_proto qw/void vp10_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+    specialize qw/vp10_iht16x16_256_add/;
+
+    add_proto qw/void vp10_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct4x4 sse2/;
+
+    add_proto qw/void vp10_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct4x4_1 sse2/;
 
-  add_proto qw/void vp10_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
-  specialize qw/vp10_iht8x8_64_add/;
+    add_proto qw/void vp10_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct8x8 sse2/;
 
-  add_proto qw/void vp10_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
-  specialize qw/vp10_iht16x16_256_add/;
+    add_proto qw/void vp10_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct8x8_1 sse2/;
+
+    add_proto qw/void vp10_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct16x16 sse2/;
+
+    add_proto qw/void vp10_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct16x16_1 sse2/;
+
+    add_proto qw/void vp10_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32 sse2/;
+
+    add_proto qw/void vp10_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32_rd sse2/;
+
+    add_proto qw/void vp10_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32_1 sse2/;
+
+    add_proto qw/void vp10_highbd_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct4x4 sse2/;
+
+    add_proto qw/void vp10_highbd_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct8x8 sse2/;
+
+    add_proto qw/void vp10_highbd_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct8x8_1/;
+
+    add_proto qw/void vp10_highbd_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct16x16 sse2/;
+
+    add_proto qw/void vp10_highbd_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct16x16_1/;
+
+    add_proto qw/void vp10_highbd_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct32x32 sse2/;
+
+    add_proto qw/void vp10_highbd_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct32x32_rd sse2/;
+
+    add_proto qw/void vp10_highbd_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_highbd_fdct32x32_1/;
+  }
 } else {
   # Force C versions if CONFIG_EMULATE_HARDWARE is 1
   if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
@@ -106,6 +219,33 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 
     add_proto qw/void vp10_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
     specialize qw/vp10_iht16x16_256_add/;
+
+    add_proto qw/void vp10_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct4x4/;
+
+    add_proto qw/void vp10_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct4x4_1/;
+
+    add_proto qw/void vp10_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct8x8/;
+
+    add_proto qw/void vp10_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct8x8_1/;
+
+    add_proto qw/void vp10_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct16x16/;
+
+    add_proto qw/void vp10_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct16x16_1/;
+
+    add_proto qw/void vp10_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32/;
+
+    add_proto qw/void vp10_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32_rd/;
+
+    add_proto qw/void vp10_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32_1/;
   } else {
     add_proto qw/void vp10_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
     specialize qw/vp10_iht4x4_16_add sse2 neon dspr2 msa/;
@@ -115,6 +255,33 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 
     add_proto qw/void vp10_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
     specialize qw/vp10_iht16x16_256_add sse2 dspr2 msa/;
+
+    add_proto qw/void vp10_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct4x4 sse2/;
+
+    add_proto qw/void vp10_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct4x4_1 sse2/;
+
+    add_proto qw/void vp10_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct8x8 sse2/;
+
+    add_proto qw/void vp10_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct8x8_1 sse2/;
+
+    add_proto qw/void vp10_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct16x16 sse2/;
+
+    add_proto qw/void vp10_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct16x16_1 sse2/;
+
+    add_proto qw/void vp10_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32 sse2/;
+
+    add_proto qw/void vp10_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32_rd sse2/;
+
+    add_proto qw/void vp10_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/vp10_fdct32x32_1 sse2/;
   }
 }
 
@@ -261,17 +428,6 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vp10_fdct8x8_quant sse2 ssse3 neon/;
 }
 
-#
-# Structured Similarity (SSIM)
-#
-if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
-    add_proto qw/void vp10_ssim_parms_8x8/, "uint8_t *s, int sp, uint8_t *r, int rp, unsigned long *sum_s, unsigned long *sum_r, unsigned long *sum_sq_s, unsigned long *sum_sq_r, unsigned long *sum_sxr";
-    specialize qw/vp10_ssim_parms_8x8/, "$sse2_x86_64";
-
-    add_proto qw/void vp10_ssim_parms_16x16/, "uint8_t *s, int sp, uint8_t *r, int rp, unsigned long *sum_s, unsigned long *sum_r, unsigned long *sum_sq_s, unsigned long *sum_sq_r, unsigned long *sum_sxr";
-    specialize qw/vp10_ssim_parms_16x16/, "$sse2_x86_64";
-}
-
 # fdct functions
 
 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
@@ -300,6 +456,188 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vp10_fwht4x4 msa/, "$mmx_x86inc";
 }
 
+# Inverse transform
+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 vp10_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct4x4_1_add/;
+
+  add_proto qw/void vp10_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct4x4_16_add/;
+
+  add_proto qw/void vp10_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct8x8_1_add/;
+
+  add_proto qw/void vp10_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct8x8_64_add/;
+
+  add_proto qw/void vp10_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct8x8_12_add/;
+
+  add_proto qw/void vp10_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct16x16_1_add/;
+
+  add_proto qw/void vp10_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct16x16_256_add/;
+
+  add_proto qw/void vp10_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct16x16_10_add/;
+
+  add_proto qw/void vp10_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct32x32_1024_add/;
+
+  add_proto qw/void vp10_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct32x32_34_add/;
+
+  add_proto qw/void vp10_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_idct32x32_1_add/;
+
+  add_proto qw/void vp10_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_iwht4x4_1_add/;
+
+  add_proto qw/void vp10_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/vp10_iwht4x4_16_add/;
+
+  add_proto qw/void vp10_highbd_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct4x4_1_add/;
+
+  add_proto qw/void vp10_highbd_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct8x8_1_add/;
+
+  add_proto qw/void vp10_highbd_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct16x16_1_add/;
+
+  add_proto qw/void vp10_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct32x32_1024_add/;
+
+  add_proto qw/void vp10_highbd_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct32x32_34_add/;
+
+  add_proto qw/void vp10_highbd_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_idct32x32_1_add/;
+
+  add_proto qw/void vp10_highbd_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_iwht4x4_1_add/;
+
+  add_proto qw/void vp10_highbd_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+  specialize qw/vp10_highbd_iwht4x4_16_add/;
+
+  # Force C versions if CONFIG_EMULATE_HARDWARE is 1
+  if (vpx_config("CONFIG_EMULATE_HARDWARE") eq "yes") {
+    add_proto qw/void vp10_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct4x4_16_add/;
+
+    add_proto qw/void vp10_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct8x8_64_add/;
+
+    add_proto qw/void vp10_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct8x8_10_add/;
+
+    add_proto qw/void vp10_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct16x16_256_add/;
+
+    add_proto qw/void vp10_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct16x16_10_add/;
+  } else {
+    add_proto qw/void vp10_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct4x4_16_add sse2/;
+
+    add_proto qw/void vp10_highbd_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct8x8_64_add sse2/;
+
+    add_proto qw/void vp10_highbd_idct8x8_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct8x8_10_add sse2/;
+
+    add_proto qw/void vp10_highbd_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct16x16_256_add sse2/;
+
+    add_proto qw/void vp10_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/vp10_highbd_idct16x16_10_add sse2/;
+  }  # 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 vp10_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct4x4_1_add/;
+
+    add_proto qw/void vp10_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct4x4_16_add/;
+
+    add_proto qw/void vp10_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_1_add/;
+
+    add_proto qw/void vp10_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_64_add/;
+
+    add_proto qw/void vp10_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_12_add/;
+
+    add_proto qw/void vp10_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_1_add/;
+
+    add_proto qw/void vp10_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_256_add/;
+
+    add_proto qw/void vp10_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_10_add/;
+
+    add_proto qw/void vp10_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_1024_add/;
+
+    add_proto qw/void vp10_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_34_add/;
+
+    add_proto qw/void vp10_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_1_add/;
+
+    add_proto qw/void vp10_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_iwht4x4_1_add/;
+
+    add_proto qw/void vp10_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_iwht4x4_16_add/;
+  } else {
+    add_proto qw/void vp10_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct4x4_1_add sse2/;
+
+    add_proto qw/void vp10_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct4x4_16_add sse2/;
+
+    add_proto qw/void vp10_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_1_add sse2/;
+
+    add_proto qw/void vp10_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_64_add sse2/;
+
+    add_proto qw/void vp10_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct8x8_12_add sse2/;
+
+    add_proto qw/void vp10_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_1_add sse2/;
+
+    add_proto qw/void vp10_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_256_add sse2/;
+
+    add_proto qw/void vp10_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct16x16_10_add sse2/;
+
+    add_proto qw/void vp10_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_1024_add sse2/;
+
+    add_proto qw/void vp10_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_34_add sse2/;
+
+    add_proto qw/void vp10_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_idct32x32_1_add sse2/;
+
+    add_proto qw/void vp10_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_iwht4x4_1_add/;
+
+    add_proto qw/void vp10_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/vp10_iwht4x4_16_add/;
+  }  # CONFIG_EMULATE_HARDWARE
+}  # CONFIG_VP9_HIGHBITDEPTH
+
 #
 # Motion search
 #
@@ -330,14 +668,6 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vp10_highbd_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
   specialize qw/vp10_highbd_quantize_fp_32x32/;
 
-  #
-  # Structured Similarity (SSIM)
-  #
-  if (vpx_config("CONFIG_INTERNAL_STATS") eq "yes") {
-    add_proto qw/void vp10_highbd_ssim_parms_8x8/, "uint16_t *s, int sp, uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
-    specialize qw/vp10_highbd_ssim_parms_8x8/;
-  }
-
   # fdct functions
   add_proto qw/void vp10_highbd_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
   specialize qw/vp10_highbd_fht4x4/;