Optimize x86 intra_sa8d_x3_8x8

[x264] / common / x86 / x86inc.asm

diff --git a/common/x86/x86inc.asm b/common/x86/x86inc.asm
index dcc5e9f383a23446419c3cf808f6f53d2e0c6fee..9e90b1436be23b172d4810912bbdd61eaf1e94e3 100644 (file)
--- a/common/x86/x86inc.asm
+++ b/common/x86/x86inc.asm
@@ -297,6 +297,9 @@ DECLARE_REG 6, rax, eax, ax,  al,  [rsp + stack_offset + 56]
 
 %macro WIN64_SPILL_XMM 1
     %assign xmm_regs_used %1
+    %if mmsize == 8
+        %assign xmm_regs_used 0
+    %endif
     ASSERT xmm_regs_used <= 16
     %if xmm_regs_used > 6
         sub rsp, (xmm_regs_used-6)*16+16
@@ -641,7 +644,7 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
     INIT_CPUFLAGS %1
 %endmacro
 
-INIT_MMX
+INIT_XMM
 
 ; I often want to use macros that permute their arguments. e.g. there's no
 ; efficient way to implement butterfly or transpose or dct without swapping some
@@ -982,3 +985,22 @@ AVX_INSTR xorps, 1, 0
 AVX_INSTR pfadd, 1, 0
 AVX_INSTR pfsub, 1, 0
 AVX_INSTR pfmul, 1, 0
+
+
+; base-4 constants for shuffles
+%assign i 0
+%rep 256
+    %assign j ((i>>6)&3)*1000 + ((i>>4)&3)*100 + ((i>>2)&3)*10 + (i&3)
+    %if j < 10
+        CAT_XDEFINE q000, j, i
+    %elif j < 100
+        CAT_XDEFINE q00, j, i
+    %elif j < 1000
+        CAT_XDEFINE q0, j, i
+    %else
+        CAT_XDEFINE q, j, i
+    %endif
+%assign i i+1
+%endrep
+%undef i
+%undef j