Skip to content
Snippets Groups Projects
Commit 2f7f2e4b authored by Loren Merritt's avatar Loren Merritt Committed by Ronald S. Bultje
Browse files

Update x86inc.asm to latest x264 version, and add AVX symmetry. 

We keep INIT_AVX (for backwards compatibility). 3arg AVX ops with
a memory arg can only have it in src2, whereas SSE emulation of
3arg prefers to have it in src1 (i.e. the mov). So, if the op is
symmetric and the wrong one is memory, swap them.
parent 18ae3626
Branches
Tags
No related merge requests found
Hide whitespace changes
Inline Side-by-side
libavutil/x86/x86inc.asm
+ 340
− 183
View file @ 2f7f2e4b
;***************************************************************************** ;*****************************************************************************
;* x86inc.asm ;* x86inc.asm: x264asm abstraction layer
;***************************************************************************** ;*****************************************************************************
;* Copyright (C) 2005-2011 x264 project ;* Copyright (C) 2005-2011 x264 project
;* ;*
...@@ -112,7 +112,7 @@ ...@@ -112,7 +112,7 @@
; we need more flexible macro. ; we need more flexible macro.
; RET: ; RET:
; Pops anything that was pushed by PROLOGUE ; Pops anything that was pushed by PROLOGUE, and returns.
; REP_RET: ; REP_RET:
; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons ; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons
...@@ -297,6 +297,9 @@ DECLARE_REG 6, rax, eax, ax, al, [rsp + stack_offset + 56] ...@@ -297,6 +297,9 @@ DECLARE_REG 6, rax, eax, ax, al, [rsp + stack_offset + 56]
%macro WIN64_SPILL_XMM 1 %macro WIN64_SPILL_XMM 1
%assign xmm_regs_used %1 %assign xmm_regs_used %1
%if mmsize == 8
%assign xmm_regs_used 0
%endif
ASSERT xmm_regs_used <= 16 ASSERT xmm_regs_used <= 16
%if xmm_regs_used > 6 %if xmm_regs_used > 6
sub rsp, (xmm_regs_used-6)*16+16 sub rsp, (xmm_regs_used-6)*16+16
...@@ -459,10 +462,24 @@ DECLARE_REG 6, ebp, ebp, bp, null, [esp + stack_offset + 28] ...@@ -459,10 +462,24 @@ DECLARE_REG 6, ebp, ebp, bp, null, [esp + stack_offset + 28]
%assign function_align 16 %assign function_align 16
; Symbol prefix for C linkage ; Begin a function.
%macro cglobal 1-2+ ; Applies any symbol mangling needed for C linkage, and sets up a define such that
%xdefine %1 mangle(program_name %+ _ %+ %1) ; subsequent uses of the function name automatically refer to the mangled version.
%xdefine %1.skip_prologue %1 %+ .skip_prologue ; Appends cpuflags to the function name if cpuflags has been specified.
%macro cglobal 1-2+ ; name, [PROLOGUE args]
%if %0 == 1
cglobal_internal %1 %+ SUFFIX
%else
cglobal_internal %1 %+ SUFFIX, %2
%endif
%endmacro
%macro cglobal_internal 1-2+
%ifndef cglobaled_%1
%xdefine %1 mangle(program_name %+ _ %+ %1)
%xdefine %1.skip_prologue %1 %+ .skip_prologue
CAT_XDEFINE cglobaled_, %1, 1
%endif
%xdefine current_function %1
%ifidn __OUTPUT_FORMAT__,elf %ifidn __OUTPUT_FORMAT__,elf
global %1:function hidden global %1:function hidden
%else %else
...@@ -479,12 +496,14 @@ DECLARE_REG 6, ebp, ebp, bp, null, [esp + stack_offset + 28] ...@@ -479,12 +496,14 @@ DECLARE_REG 6, ebp, ebp, bp, null, [esp + stack_offset + 28]
%macro cextern 1 %macro cextern 1
%xdefine %1 mangle(program_name %+ _ %+ %1) %xdefine %1 mangle(program_name %+ _ %+ %1)
CAT_XDEFINE cglobaled_, %1, 1
extern %1 extern %1
%endmacro %endmacro
;like cextern, but without the prefix ; like cextern, but without the prefix
%macro cextern_naked 1 %macro cextern_naked 1
%xdefine %1 mangle(%1) %xdefine %1 mangle(%1)
CAT_XDEFINE cglobaled_, %1, 1
extern %1 extern %1
%endmacro %endmacro
...@@ -500,6 +519,61 @@ DECLARE_REG 6, ebp, ebp, bp, null, [esp + stack_offset + 28] ...@@ -500,6 +519,61 @@ DECLARE_REG 6, ebp, ebp, bp, null, [esp + stack_offset + 28]
SECTION .note.GNU-stack noalloc noexec nowrite progbits SECTION .note.GNU-stack noalloc noexec nowrite progbits
%endif %endif
; cpuflags
%assign cpuflags_mmx (1<<0)
%assign cpuflags_mmx2 (1<<1) | cpuflags_mmx
%assign cpuflags_3dnow (1<<2) | cpuflags_mmx
%assign cpuflags_3dnow2 (1<<3) | cpuflags_3dnow
%assign cpuflags_sse (1<<4) | cpuflags_mmx2
%assign cpuflags_sse2 (1<<5) | cpuflags_sse
%assign cpuflags_sse2slow (1<<6) | cpuflags_sse2
%assign cpuflags_sse3 (1<<7) | cpuflags_sse2
%assign cpuflags_ssse3 (1<<8) | cpuflags_sse3
%assign cpuflags_sse4 (1<<9) | cpuflags_ssse3
%assign cpuflags_sse42 (1<<10)| cpuflags_sse4
%assign cpuflags_avx (1<<11)| cpuflags_sse42
%assign cpuflags_xop (1<<12)| cpuflags_avx
%assign cpuflags_fma4 (1<<13)| cpuflags_avx
%assign cpuflags_cache32 (1<<16)
%assign cpuflags_cache64 (1<<17)
%assign cpuflags_slowctz (1<<18)
%assign cpuflags_lzcnt (1<<19)
%assign cpuflags_misalign (1<<20)
%assign cpuflags_aligned (1<<21) ; not a cpu feature, but a function variant
%assign cpuflags_atom (1<<22)
%define cpuflag(x) ((cpuflags & (cpuflags_ %+ x)) == (cpuflags_ %+ x))
%define notcpuflag(x) ((cpuflags & (cpuflags_ %+ x)) != (cpuflags_ %+ x))
; Takes up to 2 cpuflags from the above list.
; All subsequent functions (up to the next INIT_CPUFLAGS) is built for the specified cpu.
; You shouldn't need to invoke this macro directly, it's a subroutine for INIT_MMX &co.
%macro INIT_CPUFLAGS 0-2
%if %0 >= 1
%xdefine cpuname %1
%assign cpuflags cpuflags_%1
%if %0 >= 2
%xdefine cpuname %1_%2
%assign cpuflags cpuflags | cpuflags_%2
%endif
%xdefine SUFFIX _ %+ cpuname
%if cpuflag(avx)
%assign avx_enabled 1
%endif
%if cpuflag(aligned)
%define movu mova
%elifidn %1, sse3
%define movu lddqu
%endif
%else
%xdefine SUFFIX
%undef cpuname
%undef cpuflags
%endif
%endmacro
; merge mmx and sse* ; merge mmx and sse*
%macro CAT_XDEFINE 3 %macro CAT_XDEFINE 3
...@@ -510,9 +584,9 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits ...@@ -510,9 +584,9 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
%undef %1%2 %undef %1%2
%endmacro %endmacro
%macro INIT_MMX 0 %macro INIT_MMX 0-1+
%assign avx_enabled 0 %assign avx_enabled 0
%define RESET_MM_PERMUTATION INIT_MMX %define RESET_MM_PERMUTATION INIT_MMX %1
%define mmsize 8 %define mmsize 8
%define num_mmregs 8 %define num_mmregs 8
%define mova movq %define mova movq
...@@ -530,11 +604,12 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits ...@@ -530,11 +604,12 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
CAT_UNDEF nmm, %%i CAT_UNDEF nmm, %%i
%assign %%i %%i+1 %assign %%i %%i+1
%endrep %endrep
INIT_CPUFLAGS %1
%endmacro %endmacro
%macro INIT_XMM 0 %macro INIT_XMM 0-1+
%assign avx_enabled 0 %assign avx_enabled 0
%define RESET_MM_PERMUTATION INIT_XMM %define RESET_MM_PERMUTATION INIT_XMM %1
%define mmsize 16 %define mmsize 16
%define num_mmregs 8 %define num_mmregs 8
%ifdef ARCH_X86_64 %ifdef ARCH_X86_64
...@@ -550,8 +625,10 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits ...@@ -550,8 +625,10 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
CAT_XDEFINE nxmm, %%i, %%i CAT_XDEFINE nxmm, %%i, %%i
%assign %%i %%i+1 %assign %%i %%i+1
%endrep %endrep
INIT_CPUFLAGS %1
%endmacro %endmacro
; FIXME: INIT_AVX can be replaced by INIT_XMM avx
%macro INIT_AVX 0 %macro INIT_AVX 0
INIT_XMM INIT_XMM
%assign avx_enabled 1 %assign avx_enabled 1
...@@ -559,9 +636,9 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits ...@@ -559,9 +636,9 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
%define RESET_MM_PERMUTATION INIT_AVX %define RESET_MM_PERMUTATION INIT_AVX
%endmacro %endmacro
%macro INIT_YMM 0 %macro INIT_YMM 0-1+
%assign avx_enabled 1 %assign avx_enabled 1
%define RESET_MM_PERMUTATION INIT_YMM %define RESET_MM_PERMUTATION INIT_YMM %1
%define mmsize 32 %define mmsize 32
%define num_mmregs 8 %define num_mmregs 8
%ifdef ARCH_X86_64 %ifdef ARCH_X86_64
...@@ -569,15 +646,18 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits ...@@ -569,15 +646,18 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
%endif %endif
%define mova vmovaps %define mova vmovaps
%define movu vmovups %define movu vmovups
%undef movh
%define movnta vmovntps
%assign %%i 0 %assign %%i 0
%rep num_mmregs %rep num_mmregs
CAT_XDEFINE m, %%i, ymm %+ %%i CAT_XDEFINE m, %%i, ymm %+ %%i
CAT_XDEFINE nymm, %%i, %%i CAT_XDEFINE nymm, %%i, %%i
%assign %%i %%i+1 %assign %%i %%i+1
%endrep %endrep
INIT_CPUFLAGS %1
%endmacro %endmacro
INIT_MMX INIT_XMM
; I often want to use macros that permute their arguments. e.g. there's no ; 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 ; efficient way to implement butterfly or transpose or dct without swapping some
...@@ -633,31 +713,46 @@ INIT_MMX ...@@ -633,31 +713,46 @@ INIT_MMX
%endrep %endrep
%endmacro %endmacro
; If SAVE_MM_PERMUTATION is placed at the end of a function and given the ; If SAVE_MM_PERMUTATION is placed at the end of a function, then any later
; function name, then any later calls to that function will automatically ; calls to that function will automatically load the permutation, so values can
; load the permutation, so values can be returned in mmregs. ; be returned in mmregs.
%macro SAVE_MM_PERMUTATION 1 ; name to save as %macro SAVE_MM_PERMUTATION 0-1
%if %0
%xdefine %%f %1_m
%else
%xdefine %%f current_function %+ _m
%endif
%assign %%i 0 %assign %%i 0
%rep num_mmregs %rep num_mmregs
CAT_XDEFINE %1_m, %%i, m %+ %%i CAT_XDEFINE %%f, %%i, m %+ %%i
%assign %%i %%i+1 %assign %%i %%i+1
%endrep %endrep
%endmacro %endmacro
%macro LOAD_MM_PERMUTATION 1 ; name to load from %macro LOAD_MM_PERMUTATION 1 ; name to load from
%assign %%i 0 %ifdef %1_m0
%rep num_mmregs %assign %%i 0
CAT_XDEFINE m, %%i, %1_m %+ %%i %rep num_mmregs
CAT_XDEFINE n, m %+ %%i, %%i CAT_XDEFINE m, %%i, %1_m %+ %%i
%assign %%i %%i+1 CAT_XDEFINE n, m %+ %%i, %%i
%endrep %assign %%i %%i+1
%endrep
%endif
%endmacro %endmacro
; Append cpuflags to the callee's name iff the appended name is known and the plain name isn't
%macro call 1 %macro call 1
call %1 call_internal %1, %1 %+ SUFFIX
%ifdef %1_m0 %endmacro
LOAD_MM_PERMUTATION %1 %macro call_internal 2
%xdefine %%i %1
%ifndef cglobaled_%1
%ifdef cglobaled_%2
%xdefine %%i %2
%endif
%endif %endif
call %%i
LOAD_MM_PERMUTATION %%i
%endmacro %endmacro
; Substitutions that reduce instruction size but are functionally equivalent ; Substitutions that reduce instruction size but are functionally equivalent
...@@ -702,14 +797,19 @@ INIT_MMX ...@@ -702,14 +797,19 @@ INIT_MMX
;%1 == instruction ;%1 == instruction
;%2 == 1 if float, 0 if int ;%2 == 1 if float, 0 if int
;%3 == 0 if 3-operand (xmm, xmm, xmm), 1 if 4-operand (xmm, xmm, xmm, imm) ;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 3-operand (xmm, xmm, xmm)
;%4 == number of operands given ;%4 == number of operands given
;%5+: operands ;%5+: operands
%macro RUN_AVX_INSTR 6-7+ %macro RUN_AVX_INSTR 6-7+
%if sizeof%5==32 %ifid %5
%define %%size sizeof%5
%else
%define %%size mmsize
%endif
%if %%size==32
v%1 %5, %6, %7 v%1 %5, %6, %7
%else %else
%if sizeof%5==8 %if %%size==8
%define %%regmov movq %define %%regmov movq
%elif %2 %elif %2
%define %%regmov movaps %define %%regmov movaps
...@@ -736,15 +836,37 @@ INIT_MMX ...@@ -736,15 +836,37 @@ INIT_MMX
%endif %endif
%endmacro %endmacro
; 3arg AVX ops with a memory arg can only have it in src2,
; whereas SSE emulation of 3arg prefers to have it in src1 (i.e. the mov).
; So, if the op is symmetric and the wrong one is memory, swap them.
%macro RUN_AVX_INSTR1 8
%assign %%swap 0
%if avx_enabled
%ifnid %6
%assign %%swap 1
%endif
%elifnidn %5, %6
%ifnid %7
%assign %%swap 1
%endif
%endif
%if %%swap && %3 == 0 && %8 == 1
RUN_AVX_INSTR %1, %2, %3, %4, %5, %7, %6
%else
RUN_AVX_INSTR %1, %2, %3, %4, %5, %6, %7
%endif
%endmacro
;%1 == instruction ;%1 == instruction
;%2 == 1 if float, 0 if int ;%2 == 1 if float, 0 if int
;%3 == 0 if 3-operand (xmm, xmm, xmm), 1 if 4-operand (xmm, xmm, xmm, imm) ;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 3-operand (xmm, xmm, xmm)
%macro AVX_INSTR 3 ;%4 == 1 if symmetric (i.e. doesn't matter which src arg is which), 0 if not
%macro %1 2-8 fnord, fnord, fnord, %1, %2, %3 %macro AVX_INSTR 4
%macro %1 2-9 fnord, fnord, fnord, %1, %2, %3, %4
%ifidn %3, fnord %ifidn %3, fnord
RUN_AVX_INSTR %6, %7, %8, 2, %1, %2 RUN_AVX_INSTR %6, %7, %8, 2, %1, %2
%elifidn %4, fnord %elifidn %4, fnord
RUN_AVX_INSTR %6, %7, %8, 3, %1, %2, %3 RUN_AVX_INSTR1 %6, %7, %8, 3, %1, %2, %3, %9
%elifidn %5, fnord %elifidn %5, fnord
RUN_AVX_INSTR %6, %7, %8, 4, %1, %2, %3, %4 RUN_AVX_INSTR %6, %7, %8, 4, %1, %2, %3, %4
%else %else
...@@ -753,153 +875,188 @@ INIT_MMX ...@@ -753,153 +875,188 @@ INIT_MMX
%endmacro %endmacro
%endmacro %endmacro
AVX_INSTR addpd, 1, 0 AVX_INSTR addpd, 1, 0, 1
AVX_INSTR addps, 1, 0 AVX_INSTR addps, 1, 0, 1
AVX_INSTR addsd, 1, 0 AVX_INSTR addsd, 1, 0, 1
AVX_INSTR addss, 1, 0 AVX_INSTR addss, 1, 0, 1
AVX_INSTR addsubpd, 1, 0 AVX_INSTR addsubpd, 1, 0, 0
AVX_INSTR addsubps, 1, 0 AVX_INSTR addsubps, 1, 0, 0
AVX_INSTR andpd, 1, 0 AVX_INSTR andpd, 1, 0, 1
AVX_INSTR andps, 1, 0 AVX_INSTR andps, 1, 0, 1
AVX_INSTR andnpd, 1, 0 AVX_INSTR andnpd, 1, 0, 0
AVX_INSTR andnps, 1, 0 AVX_INSTR andnps, 1, 0, 0
AVX_INSTR blendpd, 1, 0 AVX_INSTR blendpd, 1, 0, 0
AVX_INSTR blendps, 1, 0 AVX_INSTR blendps, 1, 0, 0
AVX_INSTR blendvpd, 1, 0 AVX_INSTR blendvpd, 1, 0, 0
AVX_INSTR blendvps, 1, 0 AVX_INSTR blendvps, 1, 0, 0
AVX_INSTR cmppd, 1, 0 AVX_INSTR cmppd, 1, 0, 0
AVX_INSTR cmpps, 1, 0 AVX_INSTR cmpps, 1, 0, 0
AVX_INSTR cmpsd, 1, 0 AVX_INSTR cmpsd, 1, 0, 0
AVX_INSTR cmpss, 1, 0 AVX_INSTR cmpss, 1, 0, 0
AVX_INSTR divpd, 1, 0 AVX_INSTR divpd, 1, 0, 0
AVX_INSTR divps, 1, 0 AVX_INSTR divps, 1, 0, 0
AVX_INSTR divsd, 1, 0 AVX_INSTR divsd, 1, 0, 0
AVX_INSTR divss, 1, 0 AVX_INSTR divss, 1, 0, 0
AVX_INSTR dppd, 1, 0 AVX_INSTR dppd, 1, 1, 0
AVX_INSTR dpps, 1, 0 AVX_INSTR dpps, 1, 1, 0
AVX_INSTR haddpd, 1, 0 AVX_INSTR haddpd, 1, 0, 0
AVX_INSTR haddps, 1, 0 AVX_INSTR haddps, 1, 0, 0
AVX_INSTR hsubpd, 1, 0 AVX_INSTR hsubpd, 1, 0, 0
AVX_INSTR hsubps, 1, 0 AVX_INSTR hsubps, 1, 0, 0
AVX_INSTR maxpd, 1, 0 AVX_INSTR maxpd, 1, 0, 1
AVX_INSTR maxps, 1, 0 AVX_INSTR maxps, 1, 0, 1
AVX_INSTR maxsd, 1, 0 AVX_INSTR maxsd, 1, 0, 1
AVX_INSTR maxss, 1, 0 AVX_INSTR maxss, 1, 0, 1
AVX_INSTR minpd, 1, 0 AVX_INSTR minpd, 1, 0, 1
AVX_INSTR minps, 1, 0 AVX_INSTR minps, 1, 0, 1
AVX_INSTR minsd, 1, 0 AVX_INSTR minsd, 1, 0, 1
AVX_INSTR minss, 1, 0 AVX_INSTR minss, 1, 0, 1
AVX_INSTR mpsadbw, 0, 1 AVX_INSTR movsd, 1, 0, 0
AVX_INSTR mulpd, 1, 0 AVX_INSTR movss, 1, 0, 0
AVX_INSTR mulps, 1, 0 AVX_INSTR mpsadbw, 0, 1, 0
AVX_INSTR mulsd, 1, 0 AVX_INSTR mulpd, 1, 0, 1
AVX_INSTR mulss, 1, 0 AVX_INSTR mulps, 1, 0, 1
AVX_INSTR orpd, 1, 0 AVX_INSTR mulsd, 1, 0, 1
AVX_INSTR orps, 1, 0 AVX_INSTR mulss, 1, 0, 1
AVX_INSTR packsswb, 0, 0 AVX_INSTR orpd, 1, 0, 1
AVX_INSTR packssdw, 0, 0 AVX_INSTR orps, 1, 0, 1
AVX_INSTR packuswb, 0, 0 AVX_INSTR packsswb, 0, 0, 0
AVX_INSTR packusdw, 0, 0 AVX_INSTR packssdw, 0, 0, 0
AVX_INSTR paddb, 0, 0 AVX_INSTR packuswb, 0, 0, 0
AVX_INSTR paddw, 0, 0 AVX_INSTR packusdw, 0, 0, 0
AVX_INSTR paddd, 0, 0 AVX_INSTR paddb, 0, 0, 1
AVX_INSTR paddq, 0, 0 AVX_INSTR paddw, 0, 0, 1
AVX_INSTR paddsb, 0, 0 AVX_INSTR paddd, 0, 0, 1
AVX_INSTR paddsw, 0, 0 AVX_INSTR paddq, 0, 0, 1
AVX_INSTR paddusb, 0, 0 AVX_INSTR paddsb, 0, 0, 1
AVX_INSTR paddusw, 0, 0 AVX_INSTR paddsw, 0, 0, 1
AVX_INSTR palignr, 0, 1 AVX_INSTR paddusb, 0, 0, 1
AVX_INSTR pand, 0, 0 AVX_INSTR paddusw, 0, 0, 1
AVX_INSTR pandn, 0, 0 AVX_INSTR palignr, 0, 1, 0
AVX_INSTR pavgb, 0, 0 AVX_INSTR pand, 0, 0, 1
AVX_INSTR pavgw, 0, 0 AVX_INSTR pandn, 0, 0, 0
AVX_INSTR pblendvb, 0, 0 AVX_INSTR pavgb, 0, 0, 1
AVX_INSTR pblendw, 0, 1 AVX_INSTR pavgw, 0, 0, 1
AVX_INSTR pcmpestri, 0, 0 AVX_INSTR pblendvb, 0, 0, 0
AVX_INSTR pcmpestrm, 0, 0 AVX_INSTR pblendw, 0, 1, 0
AVX_INSTR pcmpistri, 0, 0 AVX_INSTR pcmpestri, 0, 0, 0
AVX_INSTR pcmpistrm, 0, 0 AVX_INSTR pcmpestrm, 0, 0, 0
AVX_INSTR pcmpeqb, 0, 0 AVX_INSTR pcmpistri, 0, 0, 0
AVX_INSTR pcmpeqw, 0, 0 AVX_INSTR pcmpistrm, 0, 0, 0
AVX_INSTR pcmpeqd, 0, 0 AVX_INSTR pcmpeqb, 0, 0, 1
AVX_INSTR pcmpeqq, 0, 0 AVX_INSTR pcmpeqw, 0, 0, 1
AVX_INSTR pcmpgtb, 0, 0 AVX_INSTR pcmpeqd, 0, 0, 1
AVX_INSTR pcmpgtw, 0, 0 AVX_INSTR pcmpeqq, 0, 0, 1
AVX_INSTR pcmpgtd, 0, 0 AVX_INSTR pcmpgtb, 0, 0, 0
AVX_INSTR pcmpgtq, 0, 0 AVX_INSTR pcmpgtw, 0, 0, 0
AVX_INSTR phaddw, 0, 0 AVX_INSTR pcmpgtd, 0, 0, 0
AVX_INSTR phaddd, 0, 0 AVX_INSTR pcmpgtq, 0, 0, 0
AVX_INSTR phaddsw, 0, 0 AVX_INSTR phaddw, 0, 0, 0
AVX_INSTR phsubw, 0, 0 AVX_INSTR phaddd, 0, 0, 0
AVX_INSTR phsubd, 0, 0 AVX_INSTR phaddsw, 0, 0, 0
AVX_INSTR phsubsw, 0, 0 AVX_INSTR phsubw, 0, 0, 0
AVX_INSTR pmaddwd, 0, 0 AVX_INSTR phsubd, 0, 0, 0
AVX_INSTR pmaddubsw, 0, 0 AVX_INSTR phsubsw, 0, 0, 0
AVX_INSTR pmaxsb, 0, 0 AVX_INSTR pmaddwd, 0, 0, 1
AVX_INSTR pmaxsw, 0, 0 AVX_INSTR pmaddubsw, 0, 0, 0
AVX_INSTR pmaxsd, 0, 0 AVX_INSTR pmaxsb, 0, 0, 1
AVX_INSTR pmaxub, 0, 0 AVX_INSTR pmaxsw, 0, 0, 1
AVX_INSTR pmaxuw, 0, 0 AVX_INSTR pmaxsd, 0, 0, 1
AVX_INSTR pmaxud, 0, 0 AVX_INSTR pmaxub, 0, 0, 1
AVX_INSTR pminsb, 0, 0 AVX_INSTR pmaxuw, 0, 0, 1
AVX_INSTR pminsw, 0, 0 AVX_INSTR pmaxud, 0, 0, 1
AVX_INSTR pminsd, 0, 0 AVX_INSTR pminsb, 0, 0, 1
AVX_INSTR pminub, 0, 0 AVX_INSTR pminsw, 0, 0, 1
AVX_INSTR pminuw, 0, 0 AVX_INSTR pminsd, 0, 0, 1
AVX_INSTR pminud, 0, 0 AVX_INSTR pminub, 0, 0, 1
AVX_INSTR pmulhuw, 0, 0 AVX_INSTR pminuw, 0, 0, 1
AVX_INSTR pmulhrsw, 0, 0 AVX_INSTR pminud, 0, 0, 1
AVX_INSTR pmulhw, 0, 0 AVX_INSTR pmulhuw, 0, 0, 1
AVX_INSTR pmullw, 0, 0 AVX_INSTR pmulhrsw, 0, 0, 1
AVX_INSTR pmulld, 0, 0 AVX_INSTR pmulhw, 0, 0, 1
AVX_INSTR pmuludq, 0, 0 AVX_INSTR pmullw, 0, 0, 1
AVX_INSTR pmuldq, 0, 0 AVX_INSTR pmulld, 0, 0, 1
AVX_INSTR por, 0, 0 AVX_INSTR pmuludq, 0, 0, 1
AVX_INSTR psadbw, 0, 0 AVX_INSTR pmuldq, 0, 0, 1
AVX_INSTR pshufb, 0, 0 AVX_INSTR por, 0, 0, 1
AVX_INSTR psignb, 0, 0 AVX_INSTR psadbw, 0, 0, 1
AVX_INSTR psignw, 0, 0 AVX_INSTR pshufb, 0, 0, 0
AVX_INSTR psignd, 0, 0 AVX_INSTR psignb, 0, 0, 0
AVX_INSTR psllw, 0, 0 AVX_INSTR psignw, 0, 0, 0
AVX_INSTR pslld, 0, 0 AVX_INSTR psignd, 0, 0, 0
AVX_INSTR psllq, 0, 0 AVX_INSTR psllw, 0, 0, 0
AVX_INSTR pslldq, 0, 0 AVX_INSTR pslld, 0, 0, 0
AVX_INSTR psraw, 0, 0 AVX_INSTR psllq, 0, 0, 0
AVX_INSTR psrad, 0, 0 AVX_INSTR pslldq, 0, 0, 0
AVX_INSTR psrlw, 0, 0 AVX_INSTR psraw, 0, 0, 0
AVX_INSTR psrld, 0, 0 AVX_INSTR psrad, 0, 0, 0
AVX_INSTR psrlq, 0, 0 AVX_INSTR psrlw, 0, 0, 0
AVX_INSTR psrldq, 0, 0 AVX_INSTR psrld, 0, 0, 0
AVX_INSTR psubb, 0, 0 AVX_INSTR psrlq, 0, 0, 0
AVX_INSTR psubw, 0, 0 AVX_INSTR psrldq, 0, 0, 0
AVX_INSTR psubd, 0, 0 AVX_INSTR psubb, 0, 0, 0
AVX_INSTR psubq, 0, 0 AVX_INSTR psubw, 0, 0, 0
AVX_INSTR psubsb, 0, 0 AVX_INSTR psubd, 0, 0, 0
AVX_INSTR psubsw, 0, 0 AVX_INSTR psubq, 0, 0, 0
AVX_INSTR psubusb, 0, 0 AVX_INSTR psubsb, 0, 0, 0
AVX_INSTR psubusw, 0, 0 AVX_INSTR psubsw, 0, 0, 0
AVX_INSTR punpckhbw, 0, 0 AVX_INSTR psubusb, 0, 0, 0
AVX_INSTR punpckhwd, 0, 0 AVX_INSTR psubusw, 0, 0, 0
AVX_INSTR punpckhdq, 0, 0 AVX_INSTR punpckhbw, 0, 0, 0
AVX_INSTR punpckhqdq, 0, 0 AVX_INSTR punpckhwd, 0, 0, 0
AVX_INSTR punpcklbw, 0, 0 AVX_INSTR punpckhdq, 0, 0, 0
AVX_INSTR punpcklwd, 0, 0 AVX_INSTR punpckhqdq, 0, 0, 0
AVX_INSTR punpckldq, 0, 0 AVX_INSTR punpcklbw, 0, 0, 0
AVX_INSTR punpcklqdq, 0, 0 AVX_INSTR punpcklwd, 0, 0, 0
AVX_INSTR pxor, 0, 0 AVX_INSTR punpckldq, 0, 0, 0
AVX_INSTR shufps, 0, 1 AVX_INSTR punpcklqdq, 0, 0, 0
AVX_INSTR subpd, 1, 0 AVX_INSTR pxor, 0, 0, 1
AVX_INSTR subps, 1, 0 AVX_INSTR shufps, 0, 1, 0
AVX_INSTR subsd, 1, 0 AVX_INSTR subpd, 1, 0, 0
AVX_INSTR subss, 1, 0 AVX_INSTR subps, 1, 0, 0
AVX_INSTR unpckhpd, 1, 0 AVX_INSTR subsd, 1, 0, 0
AVX_INSTR unpckhps, 1, 0 AVX_INSTR subss, 1, 0, 0
AVX_INSTR unpcklpd, 1, 0 AVX_INSTR unpckhpd, 1, 0, 0
AVX_INSTR unpcklps, 1, 0 AVX_INSTR unpckhps, 1, 0, 0
AVX_INSTR xorpd, 1, 0 AVX_INSTR unpcklpd, 1, 0, 0
AVX_INSTR xorps, 1, 0 AVX_INSTR unpcklps, 1, 0, 0
AVX_INSTR xorpd, 1, 0, 1
AVX_INSTR xorps, 1, 0, 1
; 3DNow instructions, for sharing code between AVX, SSE and 3DN ; 3DNow instructions, for sharing code between AVX, SSE and 3DN
AVX_INSTR pfadd, 1, 0 AVX_INSTR pfadd, 1, 0, 1
AVX_INSTR pfsub, 1, 0 AVX_INSTR pfsub, 1, 0, 0
AVX_INSTR pfmul, 1, 0 AVX_INSTR pfmul, 1, 0, 1
; 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
%macro FMA_INSTR 3
%macro %1 4-7 %1, %2, %3
%if cpuflag(xop)
v%5 %1, %2, %3, %4
%else
%6 %1, %2, %3
%7 %1, %4
%endif
%endmacro
%endmacro
FMA_INSTR pmacsdd, pmulld, paddd
FMA_INSTR pmacsww, pmullw, paddw
FMA_INSTR pmadcswd, pmaddwd, paddd
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment