Enable sse2 version of inverse wht for hbd build

Change-Id: If8f5efd701a11c8a7ad3078d10ec3cd0fe27667e

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 9abbe913e804e5913b38b513c3d713d37ba663e8..73726d217c5a36c6b5f52fbcb2e943fe4323a6ba 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -699,7 +699,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   specialize qw/vpx_iwht4x4_1_add/;
 
   add_proto qw/void vpx_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
-  specialize qw/vpx_iwht4x4_16_add/;
+  specialize qw/vpx_iwht4x4_16_add/, "$sse2_x86inc";
 
   add_proto qw/void vpx_highbd_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
   specialize qw/vpx_highbd_idct4x4_1_add/;

diff --git a/vpx_dsp/x86/inv_wht_sse2.asm b/vpx_dsp/x86/inv_wht_sse2.asm
index df6f4692b48dcc07f43e46330641f792625e12db..fbbcd76bd7b4eebfd139d0a381d884d5adfc754b 100644 (file)
--- a/vpx_dsp/x86/inv_wht_sse2.asm
+++ b/vpx_dsp/x86/inv_wht_sse2.asm
@@ -82,9 +82,15 @@ SECTION .text
 
 INIT_XMM sse2
 cglobal iwht4x4_16_add, 3, 3, 7, input, output, stride
+%if CONFIG_VP9_HIGHBITDEPTH
+  mova            m0,        [inputq +  0]
+  packssdw        m0,        [inputq + 16]
+  mova            m1,        [inputq + 32]
+  packssdw        m1,        [inputq + 48]
+%else
   mova            m0,        [inputq +  0]
   mova            m1,        [inputq + 16]
-
+%endif
   psraw           m0,        2
   psraw           m1,        2