simdnotes.txt

different ways of doing the same thing

	mov ecx, 0xF0F0F0F0
	shrd rcx, rcx, 32 ; fill up the rest

	mov rcx, 0xF0F0F0F0F0F0F0F0

	pcmpeqw mm0, mm0 ; set mm0 to all 0xFFF...
	psllw mm0, 8
	movq rax, mm0

	mov eax, 0xF0F0F0F0
	movd mm0, eax
	punpcklbw mm0, mm0

	mov al, 0xF0
	movd mm0, eax
	pxor mm1, mm1
	pshufb mm0, mm1


vector idioms:
set a vector register to all 0's
	pxor xmm0, xmm0
set a vector register to all 1's
	pcmpeq[bwd] xmm0, xmm0
extend low {byte,word,dword,qword}-wise pattern to high
	punpckl{bw,wd,dq,qdq} xmm0, xmm0
set each {byte,word,dword} to 1
	pxor xmm0, xmm0
	pcmpeq xmm1, xmm1
	psub xmm0, xmm1
set each word to 2^n - 1 (similar version for dwords too)
	pcmpeq xmm0, xmm0
	psrlw xmm0, 16-n
set each word to -2^n
	pcmpeq xmm0, xmm0
	psllw xmm0, n
look in Intel Opt. Man. pg 280 ("building blocks") for more
also, see Agner Vol 2, section 13.4

repeatedly doing pshufb on itself to a register effectively amounts to
permuting it, based on itself. This could be interesting if there are
multiple permutation cycles (with different periods) inside the vector
register
in general, pshufb can compose permutations

to broadcast the lowest byte of xmm0, do
	punpcklbw xmm0, xmm0
	punpcklbw xmm0, xmm0 ; punpcklwd would work too, but prob be slower
	punpcklbw xmm0, xmm0
	punpcklbw xmm0, xmm0
or if a scratch register is available:
	pxor xmm1, xmm1
	pshufb xmm0, xmm1
to propagate the lowest word, do
	punpcklwd xmm0, xmm0
	punpcklwd xmm0, xmm0
	punpcklwd xmm0, xmm0




not really simd, but worth stashing somewhere
use cqo to sign extend rax to rax:rdx useful to create a mask of the sign
bit of rax