ports/multimedia/dav1d/patches/patch-src_x86_cdef_avx2_asm

99 lines
2.3 KiB
Text

Index: src/x86/cdef_avx2.asm
--- src/x86/cdef_avx2.asm.orig
+++ src/x86/cdef_avx2.asm
@@ -572,6 +572,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
RET
.d0k0:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
vpbroadcastq m6, [dstq+strideq*1-1]
@@ -632,6 +633,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
.d1k0:
.d2k0:
.d3k0:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
movq xm6, [dstq+strideq*0-1]
@@ -683,6 +685,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
%endif
ret
.d4k0:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
vpbroadcastq m10, [dstq+strideq*1-1]
@@ -743,6 +746,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
.d5k0:
.d6k0:
.d7k0:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
movd xm6, [topq+strideq*1 ]
@@ -783,6 +787,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
%endif
ret
.d0k1:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
movd xm6, [dstq+strideq*2-2]
@@ -834,6 +839,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
%endif
ret
.d1k1:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
vpbroadcastq m6, [dstq+strideq*1-2]
@@ -891,6 +897,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
%endif
ret
.d2k1:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
movq xm11, [leftq]
@@ -935,6 +942,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
%endif
ret
.d3k1:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
vpbroadcastq m11, [dstq+strideq*1-2]
@@ -988,6 +996,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
%endif
ret
.d4k1:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
vinserti128 m6, [dstq+strideq*0-2], 1
@@ -1039,6 +1048,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
%endif
ret
.d5k1:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
movd xm6, [topq+strideq*0-1]
@@ -1097,6 +1107,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
%endif
ret
.d6k1:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
movd xm6, [topq+strideq*0]
@@ -1139,6 +1150,7 @@ cglobal cdef_filter_%1x%2_8bpc, 5, 11, 0, dst, stride,
%endif
ret
.d7k1:
+ _CET_ENDBR
%if %1 == 4
%if %2 == 4
movd xm5, [dstq+strideq*2-1]