#define INSTR vinsertf128
#define NINST 64
#define N edi
#define i r8d


.intel_syntax noprefix
.globl ninst
.data
ninst:
.long NINST
.align 32
PI:
.long  0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9, 0xf01b866e, 0x400921f9
.text
.globl latency
.type latency, @function
.align 32
latency:
		push      rbp
		mov       rbp, rsp
		xor       i, i
		test      N, N
		jle       done
		# create DP 1.0
		vpcmpeqw xmm0, xmm0, xmm0		# all ones
		vpsllq xmm0, xmm0, 54			# logical left shift: 11111110..0 (54=64-(10-1))
		vpsrlq xmm0, xmm0, 2			# logical right shift: 1 bit for sign; leading      mantissa bit is zero
		# expand from SSE to AVX
		vinsertf128 ymm0, ymm0, xmm0, 0x1
		# copy DP 1.0
		vmovaps ymm0, ymm0
		vmovaps ymm1, ymm0
		# Create DP 2.0
		vaddpd ymm1, ymm1, ymm1
		# Create DP 0.5
		vdivpd ymm2, ymm0, ymm1
loop:
		inc      i
		INSTR    ymm3
		INSTR    ymm4
		INSTR    ymm5
		INSTR    ymm6
		INSTR    ymm7
		INSTR    ymm8
		INSTR    ymm9
		INSTR    ymm10
		INSTR    ymm11
		INSTR    ymm12
		INSTR    ymm13
		INSTR    ymm14
		INSTR    ymm15
		INSTR    ymm3
		INSTR    ymm4
		INSTR    ymm5
		INSTR    ymm6
		INSTR    ymm7
		INSTR    ymm8
		INSTR    ymm9
		INSTR    ymm10
		INSTR    ymm11
		INSTR    ymm12
		INSTR    ymm13
		INSTR    ymm14
		INSTR    ymm15
		INSTR    ymm3
		INSTR    ymm4
		INSTR    ymm5
		INSTR    ymm6
		INSTR    ymm7
		INSTR    ymm8
		INSTR    ymm9
		INSTR    ymm10
		INSTR    ymm11
		INSTR    ymm12
		INSTR    ymm13
		INSTR    ymm14
		INSTR    ymm15
		INSTR    ymm3
		INSTR    ymm4
		INSTR    ymm5
		INSTR    ymm6
		INSTR    ymm7
		INSTR    ymm8
		INSTR    ymm9
		INSTR    ymm10
		INSTR    ymm11
		INSTR    ymm12
		INSTR    ymm13
		INSTR    ymm14
		INSTR    ymm15
		INSTR    ymm3
		INSTR    ymm4
		INSTR    ymm5
		INSTR    ymm6
		INSTR    ymm7
		INSTR    ymm8
		INSTR    ymm9
		INSTR    ymm10
		INSTR    ymm11
		INSTR    ymm12
		INSTR    ymm13
		INSTR    ymm14
		cmp      i, N
		jl       loop
done:
		mov	rsp, rbp
		pop	rbp
		ret
.size latency, .-latency