mirror of
https://github.com/RRZE-HPC/OSACA.git
synced 2026-01-07 03:30:06 +01:00
removed old testcases
This commit is contained in:
Jan Laukemann
@@ -1,100 +0,0 @@
|
||||
#define INSTR add
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR edx, eax
|
||||
INSTR r9d, ebx
|
||||
INSTR r10d, ecx
|
||||
INSTR edx, eax
|
||||
INSTR r9d, ebx
|
||||
INSTR r10d, ecx
|
||||
INSTR r11d, eax
|
||||
INSTR r12d, ebx
|
||||
INSTR r13d, ecx
|
||||
INSTR r14d, eax
|
||||
INSTR r15d, ebx
|
||||
INSTR eax, ecx
|
||||
INSTR ebx, eax
|
||||
INSTR ecx, ebx
|
||||
INSTR edx, ecx
|
||||
INSTR r9d, eax
|
||||
INSTR r10d, ebx
|
||||
INSTR r11d, ecx
|
||||
INSTR r12d, eax
|
||||
INSTR r13d, ebx
|
||||
INSTR r14d, ecx
|
||||
INSTR r15d, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, ecx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR add
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
INSTR eax, ebx
|
||||
INSTR ebx, eax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR cmp
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rdx, rax
|
||||
INSTR r9, rbx
|
||||
INSTR r10, rcx
|
||||
INSTR rdx, rax
|
||||
INSTR r9, rbx
|
||||
INSTR r10, rcx
|
||||
INSTR r11, rax
|
||||
INSTR r12, rbx
|
||||
INSTR r13, rcx
|
||||
INSTR r14, rax
|
||||
INSTR r15, rbx
|
||||
INSTR rax, rcx
|
||||
INSTR rbx, rax
|
||||
INSTR rcx, rbx
|
||||
INSTR rdx, rcx
|
||||
INSTR r9, rax
|
||||
INSTR r10, rbx
|
||||
INSTR r11, rcx
|
||||
INSTR r12, rax
|
||||
INSTR r13, rbx
|
||||
INSTR r14, rcx
|
||||
INSTR r15, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rcx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR cmp
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR dec
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR edx
|
||||
INSTR r9d
|
||||
INSTR r10d
|
||||
INSTR edx
|
||||
INSTR r9d
|
||||
INSTR r10d
|
||||
INSTR r11d
|
||||
INSTR r12d
|
||||
INSTR r13d
|
||||
INSTR r14d
|
||||
INSTR r15d
|
||||
INSTR eax
|
||||
INSTR ebx
|
||||
INSTR ecx
|
||||
INSTR edx
|
||||
INSTR r9d
|
||||
INSTR r10d
|
||||
INSTR r11d
|
||||
INSTR r12d
|
||||
INSTR r13d
|
||||
INSTR r14d
|
||||
INSTR r15d
|
||||
INSTR eax
|
||||
INSTR ebx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR dec
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR inc
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rdx
|
||||
INSTR r9
|
||||
INSTR r10
|
||||
INSTR rdx
|
||||
INSTR r9
|
||||
INSTR r10
|
||||
INSTR r11
|
||||
INSTR r12
|
||||
INSTR r13
|
||||
INSTR r14
|
||||
INSTR r15
|
||||
INSTR rax
|
||||
INSTR rbx
|
||||
INSTR rcx
|
||||
INSTR rdx
|
||||
INSTR r9
|
||||
INSTR r10
|
||||
INSTR r11
|
||||
INSTR r12
|
||||
INSTR r13
|
||||
INSTR r14
|
||||
INSTR r15
|
||||
INSTR rax
|
||||
INSTR rbx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR inc
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,82 +0,0 @@
|
||||
#define INSTR janadd
|
||||
#define NINST 6
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rdx, eax
|
||||
INSTR r9, ebx
|
||||
INSTR r10, ecx
|
||||
INSTR rdx, eax
|
||||
INSTR r9, ebx
|
||||
INSTR r10, ecx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,82 +0,0 @@
|
||||
#define INSTR janadd
|
||||
#define NINST 6
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,82 +0,0 @@
|
||||
#define INSTR janadd
|
||||
#define NINST 6
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rdx, eax
|
||||
INSTR r9, ebx
|
||||
INSTR r10, ecx
|
||||
INSTR rdx, eax
|
||||
INSTR r9, ebx
|
||||
INSTR r10, ecx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,82 +0,0 @@
|
||||
#define INSTR janadd
|
||||
#define NINST 6
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR mov
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rdx, rax
|
||||
INSTR r9, rbx
|
||||
INSTR r10, rcx
|
||||
INSTR rdx, rax
|
||||
INSTR r9, rbx
|
||||
INSTR r10, rcx
|
||||
INSTR r11, rax
|
||||
INSTR r12, rbx
|
||||
INSTR r13, rcx
|
||||
INSTR r14, rax
|
||||
INSTR r15, rbx
|
||||
INSTR rax, rcx
|
||||
INSTR rbx, rax
|
||||
INSTR rcx, rbx
|
||||
INSTR rdx, rcx
|
||||
INSTR r9, rax
|
||||
INSTR r10, rbx
|
||||
INSTR r11, rcx
|
||||
INSTR r12, rax
|
||||
INSTR r13, rbx
|
||||
INSTR r14, rcx
|
||||
INSTR r15, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rcx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR mov
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR movslq
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rdx, eax
|
||||
INSTR r9, ebx
|
||||
INSTR r10, ecx
|
||||
INSTR rdx, eax
|
||||
INSTR r9, ebx
|
||||
INSTR r10, ecx
|
||||
INSTR r11, eax
|
||||
INSTR r12, ebx
|
||||
INSTR r13, ecx
|
||||
INSTR r14, eax
|
||||
INSTR r15, ebx
|
||||
INSTR rax, ecx
|
||||
INSTR rbx, eax
|
||||
INSTR rcx, ebx
|
||||
INSTR rdx, ecx
|
||||
INSTR r9, eax
|
||||
INSTR r10, ebx
|
||||
INSTR r11, ecx
|
||||
INSTR r12, eax
|
||||
INSTR r13, ebx
|
||||
INSTR r14, ecx
|
||||
INSTR r15, eax
|
||||
INSTR rax, ebx
|
||||
INSTR rbx, ecx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR movslq
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
INSTR rax, eax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR movzbl
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR edx, al
|
||||
INSTR r9d, bl
|
||||
INSTR r10d, cl
|
||||
INSTR edx, al
|
||||
INSTR r9d, bl
|
||||
INSTR r10d, cl
|
||||
INSTR r11d, al
|
||||
INSTR r12d, bl
|
||||
INSTR r13d, cl
|
||||
INSTR r14d, al
|
||||
INSTR r15d, bl
|
||||
INSTR eax, cl
|
||||
INSTR ebx, al
|
||||
INSTR ecx, bl
|
||||
INSTR edx, cl
|
||||
INSTR r9d, al
|
||||
INSTR r10d, bl
|
||||
INSTR r11d, cl
|
||||
INSTR r12d, al
|
||||
INSTR r13d, bl
|
||||
INSTR r14d, cl
|
||||
INSTR r15d, al
|
||||
INSTR eax, bl
|
||||
INSTR ebx, cl
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR movzbl
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
INSTR eax, al
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR neg
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR edx
|
||||
INSTR r9d
|
||||
INSTR r10d
|
||||
INSTR edx
|
||||
INSTR r9d
|
||||
INSTR r10d
|
||||
INSTR r11d
|
||||
INSTR r12d
|
||||
INSTR r13d
|
||||
INSTR r14d
|
||||
INSTR r15d
|
||||
INSTR eax
|
||||
INSTR ebx
|
||||
INSTR ecx
|
||||
INSTR edx
|
||||
INSTR r9d
|
||||
INSTR r10d
|
||||
INSTR r11d
|
||||
INSTR r12d
|
||||
INSTR r13d
|
||||
INSTR r14d
|
||||
INSTR r15d
|
||||
INSTR eax
|
||||
INSTR ebx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR neg
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
INSTR eax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR pop
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rdx
|
||||
INSTR r9
|
||||
INSTR r10
|
||||
INSTR rdx
|
||||
INSTR r9
|
||||
INSTR r10
|
||||
INSTR r11
|
||||
INSTR r12
|
||||
INSTR r13
|
||||
INSTR r14
|
||||
INSTR r15
|
||||
INSTR rax
|
||||
INSTR rbx
|
||||
INSTR rcx
|
||||
INSTR rdx
|
||||
INSTR r9
|
||||
INSTR r10
|
||||
INSTR r11
|
||||
INSTR r12
|
||||
INSTR r13
|
||||
INSTR r14
|
||||
INSTR r15
|
||||
INSTR rax
|
||||
INSTR rbx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR pop
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
INSTR rax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR sub
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rdx, rax
|
||||
INSTR r9, rbx
|
||||
INSTR r10, rcx
|
||||
INSTR rdx, rax
|
||||
INSTR r9, rbx
|
||||
INSTR r10, rcx
|
||||
INSTR r11, rax
|
||||
INSTR r12, rbx
|
||||
INSTR r13, rcx
|
||||
INSTR r14, rax
|
||||
INSTR r15, rbx
|
||||
INSTR rax, rcx
|
||||
INSTR rbx, rax
|
||||
INSTR rcx, rbx
|
||||
INSTR rdx, rcx
|
||||
INSTR r9, rax
|
||||
INSTR r10, rbx
|
||||
INSTR r11, rcx
|
||||
INSTR r12, rax
|
||||
INSTR r13, rbx
|
||||
INSTR r14, rcx
|
||||
INSTR r15, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rcx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR sub
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR test
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rdx, rax
|
||||
INSTR r9, rbx
|
||||
INSTR r10, rcx
|
||||
INSTR rdx, rax
|
||||
INSTR r9, rbx
|
||||
INSTR r10, rcx
|
||||
INSTR r11, rax
|
||||
INSTR r12, rbx
|
||||
INSTR r13, rcx
|
||||
INSTR r14, rax
|
||||
INSTR r15, rbx
|
||||
INSTR rax, rcx
|
||||
INSTR rbx, rax
|
||||
INSTR rcx, rbx
|
||||
INSTR rdx, rcx
|
||||
INSTR r9, rax
|
||||
INSTR r10, rbx
|
||||
INSTR r11, rcx
|
||||
INSTR r12, rax
|
||||
INSTR r13, rbx
|
||||
INSTR r14, rcx
|
||||
INSTR r15, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rcx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR test
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
INSTR rax, rbx
|
||||
INSTR rbx, rax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,67 +0,0 @@
|
||||
#define INSTR vaddpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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, ymm0, ymm0
|
||||
INSTR ymm4, ymm1, ymm1
|
||||
INSTR ymm5, ymm2, ymm2
|
||||
INSTR ymm3, ymm0, ymm0
|
||||
INSTR ymm4, ymm1, ymm1
|
||||
INSTR ymm5, ymm2, ymm2
|
||||
INSTR ymm6, ymm0, ymm0
|
||||
INSTR ymm7, ymm1, ymm1
|
||||
INSTR ymm8, ymm2, ymm2
|
||||
INSTR ymm9, ymm0, ymm0
|
||||
INSTR ymm10, ymm1, ymm1
|
||||
INSTR ymm11, ymm2, ymm2
|
||||
INSTR ymm12, ymm0, ymm0
|
||||
INSTR ymm13, ymm1, ymm1
|
||||
INSTR ymm14, ymm2, ymm2
|
||||
INSTR ymm15, ymm0, ymm0
|
||||
INSTR ymm16, ymm1, ymm1
|
||||
INSTR ymm17, ymm2, ymm2
|
||||
INSTR ymm18, ymm0, ymm0
|
||||
INSTR ymm19, ymm1, ymm1
|
||||
INSTR ymm20, ymm2, ymm2
|
||||
INSTR ymm21, ymm0, ymm0
|
||||
INSTR ymm22, ymm1, ymm1
|
||||
INSTR ymm23, ymm2, ymm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,67 +0,0 @@
|
||||
#define INSTR vaddpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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 ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vaddpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm6, xmm0, xmm0
|
||||
INSTR xmm7, xmm1, xmm1
|
||||
INSTR xmm8, xmm2, xmm2
|
||||
INSTR xmm9, xmm0, xmm0
|
||||
INSTR xmm10, xmm1, xmm1
|
||||
INSTR xmm11, xmm2, xmm2
|
||||
INSTR xmm12, xmm0, xmm0
|
||||
INSTR xmm13, xmm1, xmm1
|
||||
INSTR xmm14, xmm2, xmm2
|
||||
INSTR xmm15, xmm0, xmm0
|
||||
INSTR xmm16, xmm1, xmm1
|
||||
INSTR xmm17, xmm2, xmm2
|
||||
INSTR xmm18, xmm0, xmm0
|
||||
INSTR xmm19, xmm1, xmm1
|
||||
INSTR xmm20, xmm2, xmm2
|
||||
INSTR xmm21, xmm0, xmm0
|
||||
INSTR xmm22, xmm1, xmm1
|
||||
INSTR xmm23, xmm2, xmm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vaddpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vaddsd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm6, xmm0, xmm0
|
||||
INSTR xmm7, xmm1, xmm1
|
||||
INSTR xmm8, xmm2, xmm2
|
||||
INSTR xmm9, xmm0, xmm0
|
||||
INSTR xmm10, xmm1, xmm1
|
||||
INSTR xmm11, xmm2, xmm2
|
||||
INSTR xmm12, xmm0, xmm0
|
||||
INSTR xmm13, xmm1, xmm1
|
||||
INSTR xmm14, xmm2, xmm2
|
||||
INSTR xmm15, xmm0, xmm0
|
||||
INSTR xmm16, xmm1, xmm1
|
||||
INSTR xmm17, xmm2, xmm2
|
||||
INSTR xmm18, xmm0, xmm0
|
||||
INSTR xmm19, xmm1, xmm1
|
||||
INSTR xmm20, xmm2, xmm2
|
||||
INSTR xmm21, xmm0, xmm0
|
||||
INSTR xmm22, xmm1, xmm1
|
||||
INSTR xmm23, xmm2, xmm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vaddsd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,67 +0,0 @@
|
||||
#define INSTR vmovapd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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, ymm0
|
||||
INSTR ymm4, ymm1
|
||||
INSTR ymm5, ymm2
|
||||
INSTR ymm3, ymm0
|
||||
INSTR ymm4, ymm1
|
||||
INSTR ymm5, ymm2
|
||||
INSTR ymm6, ymm0
|
||||
INSTR ymm7, ymm1
|
||||
INSTR ymm8, ymm2
|
||||
INSTR ymm9, ymm0
|
||||
INSTR ymm10, ymm1
|
||||
INSTR ymm11, ymm2
|
||||
INSTR ymm12, ymm0
|
||||
INSTR ymm13, ymm1
|
||||
INSTR ymm14, ymm2
|
||||
INSTR ymm15, ymm0
|
||||
INSTR ymm16, ymm1
|
||||
INSTR ymm17, ymm2
|
||||
INSTR ymm18, ymm0
|
||||
INSTR ymm19, ymm1
|
||||
INSTR ymm20, ymm2
|
||||
INSTR ymm21, ymm0
|
||||
INSTR ymm22, ymm1
|
||||
INSTR ymm23, ymm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,67 +0,0 @@
|
||||
#define INSTR vmovapd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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 ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
INSTR ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
INSTR ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
INSTR ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
INSTR ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
INSTR ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
INSTR ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
INSTR ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
INSTR ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
INSTR ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
INSTR ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
INSTR ymm0, ymm1
|
||||
INSTR ymm1, ymm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vmovapd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm3, xmm0
|
||||
INSTR xmm4, xmm1
|
||||
INSTR xmm5, xmm2
|
||||
INSTR xmm3, xmm0
|
||||
INSTR xmm4, xmm1
|
||||
INSTR xmm5, xmm2
|
||||
INSTR xmm6, xmm0
|
||||
INSTR xmm7, xmm1
|
||||
INSTR xmm8, xmm2
|
||||
INSTR xmm9, xmm0
|
||||
INSTR xmm10, xmm1
|
||||
INSTR xmm11, xmm2
|
||||
INSTR xmm12, xmm0
|
||||
INSTR xmm13, xmm1
|
||||
INSTR xmm14, xmm2
|
||||
INSTR xmm15, xmm0
|
||||
INSTR xmm16, xmm1
|
||||
INSTR xmm17, xmm2
|
||||
INSTR xmm18, xmm0
|
||||
INSTR xmm19, xmm1
|
||||
INSTR xmm20, xmm2
|
||||
INSTR xmm21, xmm0
|
||||
INSTR xmm22, xmm1
|
||||
INSTR xmm23, xmm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vmovapd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vmovaps
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm3, xmm0
|
||||
INSTR xmm4, xmm1
|
||||
INSTR xmm5, xmm2
|
||||
INSTR xmm3, xmm0
|
||||
INSTR xmm4, xmm1
|
||||
INSTR xmm5, xmm2
|
||||
INSTR xmm6, xmm0
|
||||
INSTR xmm7, xmm1
|
||||
INSTR xmm8, xmm2
|
||||
INSTR xmm9, xmm0
|
||||
INSTR xmm10, xmm1
|
||||
INSTR xmm11, xmm2
|
||||
INSTR xmm12, xmm0
|
||||
INSTR xmm13, xmm1
|
||||
INSTR xmm14, xmm2
|
||||
INSTR xmm15, xmm0
|
||||
INSTR xmm16, xmm1
|
||||
INSTR xmm17, xmm2
|
||||
INSTR xmm18, xmm0
|
||||
INSTR xmm19, xmm1
|
||||
INSTR xmm20, xmm2
|
||||
INSTR xmm21, xmm0
|
||||
INSTR xmm22, xmm1
|
||||
INSTR xmm23, xmm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vmovaps
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
INSTR xmm0, xmm1
|
||||
INSTR xmm1, xmm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,98 +0,0 @@
|
||||
#define INSTR vmovq
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR rdx, xmm0
|
||||
INSTR r9, xmm1
|
||||
INSTR r10, xmm2
|
||||
INSTR rdx, xmm0
|
||||
INSTR r9, xmm1
|
||||
INSTR r10, xmm2
|
||||
INSTR r11, xmm0
|
||||
INSTR r12, xmm1
|
||||
INSTR r13, xmm2
|
||||
INSTR r14, xmm0
|
||||
INSTR r15, xmm1
|
||||
INSTR rax, xmm2
|
||||
INSTR rbx, xmm0
|
||||
INSTR rcx, xmm1
|
||||
INSTR rdx, xmm2
|
||||
INSTR r9, xmm0
|
||||
INSTR r10, xmm1
|
||||
INSTR r11, xmm2
|
||||
INSTR r12, xmm0
|
||||
INSTR r13, xmm1
|
||||
INSTR r14, xmm2
|
||||
INSTR r15, xmm0
|
||||
INSTR rax, xmm1
|
||||
INSTR rbx, xmm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,98 +0,0 @@
|
||||
#define INSTR vmovq
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
INSTR rax, xmm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR vmovq
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm3, rax
|
||||
INSTR xmm4, rbx
|
||||
INSTR xmm5, rcx
|
||||
INSTR xmm3, rax
|
||||
INSTR xmm4, rbx
|
||||
INSTR xmm5, rcx
|
||||
INSTR xmm6, rax
|
||||
INSTR xmm7, rbx
|
||||
INSTR xmm8, rcx
|
||||
INSTR xmm9, rax
|
||||
INSTR xmm10, rbx
|
||||
INSTR xmm11, rcx
|
||||
INSTR xmm12, rax
|
||||
INSTR xmm13, rbx
|
||||
INSTR xmm14, rcx
|
||||
INSTR xmm15, rax
|
||||
INSTR xmm16, rbx
|
||||
INSTR xmm17, rcx
|
||||
INSTR xmm18, rax
|
||||
INSTR xmm19, rbx
|
||||
INSTR xmm20, rcx
|
||||
INSTR xmm21, rax
|
||||
INSTR xmm22, rbx
|
||||
INSTR xmm23, rcx
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR vmovq
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
INSTR xmm0, rax
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vmovsd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm6, xmm0, xmm0
|
||||
INSTR xmm7, xmm1, xmm1
|
||||
INSTR xmm8, xmm2, xmm2
|
||||
INSTR xmm9, xmm0, xmm0
|
||||
INSTR xmm10, xmm1, xmm1
|
||||
INSTR xmm11, xmm2, xmm2
|
||||
INSTR xmm12, xmm0, xmm0
|
||||
INSTR xmm13, xmm1, xmm1
|
||||
INSTR xmm14, xmm2, xmm2
|
||||
INSTR xmm15, xmm0, xmm0
|
||||
INSTR xmm16, xmm1, xmm1
|
||||
INSTR xmm17, xmm2, xmm2
|
||||
INSTR xmm18, xmm0, xmm0
|
||||
INSTR xmm19, xmm1, xmm1
|
||||
INSTR xmm20, xmm2, xmm2
|
||||
INSTR xmm21, xmm0, xmm0
|
||||
INSTR xmm22, xmm1, xmm1
|
||||
INSTR xmm23, xmm2, xmm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vmovsd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,67 +0,0 @@
|
||||
#define INSTR vmulpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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, ymm0, ymm0
|
||||
INSTR ymm4, ymm1, ymm1
|
||||
INSTR ymm5, ymm2, ymm2
|
||||
INSTR ymm3, ymm0, ymm0
|
||||
INSTR ymm4, ymm1, ymm1
|
||||
INSTR ymm5, ymm2, ymm2
|
||||
INSTR ymm6, ymm0, ymm0
|
||||
INSTR ymm7, ymm1, ymm1
|
||||
INSTR ymm8, ymm2, ymm2
|
||||
INSTR ymm9, ymm0, ymm0
|
||||
INSTR ymm10, ymm1, ymm1
|
||||
INSTR ymm11, ymm2, ymm2
|
||||
INSTR ymm12, ymm0, ymm0
|
||||
INSTR ymm13, ymm1, ymm1
|
||||
INSTR ymm14, ymm2, ymm2
|
||||
INSTR ymm15, ymm0, ymm0
|
||||
INSTR ymm16, ymm1, ymm1
|
||||
INSTR ymm17, ymm2, ymm2
|
||||
INSTR ymm18, ymm0, ymm0
|
||||
INSTR ymm19, ymm1, ymm1
|
||||
INSTR ymm20, ymm2, ymm2
|
||||
INSTR ymm21, ymm0, ymm0
|
||||
INSTR ymm22, ymm1, ymm1
|
||||
INSTR ymm23, ymm2, ymm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,67 +0,0 @@
|
||||
#define INSTR vmulpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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 ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vmulsd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm6, xmm0, xmm0
|
||||
INSTR xmm7, xmm1, xmm1
|
||||
INSTR xmm8, xmm2, xmm2
|
||||
INSTR xmm9, xmm0, xmm0
|
||||
INSTR xmm10, xmm1, xmm1
|
||||
INSTR xmm11, xmm2, xmm2
|
||||
INSTR xmm12, xmm0, xmm0
|
||||
INSTR xmm13, xmm1, xmm1
|
||||
INSTR xmm14, xmm2, xmm2
|
||||
INSTR xmm15, xmm0, xmm0
|
||||
INSTR xmm16, xmm1, xmm1
|
||||
INSTR xmm17, xmm2, xmm2
|
||||
INSTR xmm18, xmm0, xmm0
|
||||
INSTR xmm19, xmm1, xmm1
|
||||
INSTR xmm20, xmm2, xmm2
|
||||
INSTR xmm21, xmm0, xmm0
|
||||
INSTR xmm22, xmm1, xmm1
|
||||
INSTR xmm23, xmm2, xmm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vmulsd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,67 +0,0 @@
|
||||
#define INSTR vsubpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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, ymm0, ymm0
|
||||
INSTR ymm4, ymm1, ymm1
|
||||
INSTR ymm5, ymm2, ymm2
|
||||
INSTR ymm3, ymm0, ymm0
|
||||
INSTR ymm4, ymm1, ymm1
|
||||
INSTR ymm5, ymm2, ymm2
|
||||
INSTR ymm6, ymm0, ymm0
|
||||
INSTR ymm7, ymm1, ymm1
|
||||
INSTR ymm8, ymm2, ymm2
|
||||
INSTR ymm9, ymm0, ymm0
|
||||
INSTR ymm10, ymm1, ymm1
|
||||
INSTR ymm11, ymm2, ymm2
|
||||
INSTR ymm12, ymm0, ymm0
|
||||
INSTR ymm13, ymm1, ymm1
|
||||
INSTR ymm14, ymm2, ymm2
|
||||
INSTR ymm15, ymm0, ymm0
|
||||
INSTR ymm16, ymm1, ymm1
|
||||
INSTR ymm17, ymm2, ymm2
|
||||
INSTR ymm18, ymm0, ymm0
|
||||
INSTR ymm19, ymm1, ymm1
|
||||
INSTR ymm20, ymm2, ymm2
|
||||
INSTR ymm21, ymm0, ymm0
|
||||
INSTR ymm22, ymm1, ymm1
|
||||
INSTR ymm23, ymm2, ymm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,67 +0,0 @@
|
||||
#define INSTR vsubpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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 ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vsubsd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm6, xmm0, xmm0
|
||||
INSTR xmm7, xmm1, xmm1
|
||||
INSTR xmm8, xmm2, xmm2
|
||||
INSTR xmm9, xmm0, xmm0
|
||||
INSTR xmm10, xmm1, xmm1
|
||||
INSTR xmm11, xmm2, xmm2
|
||||
INSTR xmm12, xmm0, xmm0
|
||||
INSTR xmm13, xmm1, xmm1
|
||||
INSTR xmm14, xmm2, xmm2
|
||||
INSTR xmm15, xmm0, xmm0
|
||||
INSTR xmm16, xmm1, xmm1
|
||||
INSTR xmm17, xmm2, xmm2
|
||||
INSTR xmm18, xmm0, xmm0
|
||||
INSTR xmm19, xmm1, xmm1
|
||||
INSTR xmm20, xmm2, xmm2
|
||||
INSTR xmm21, xmm0, xmm0
|
||||
INSTR xmm22, xmm1, xmm1
|
||||
INSTR xmm23, xmm2, xmm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vsubsd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vunpckhpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm6, xmm0, xmm0
|
||||
INSTR xmm7, xmm1, xmm1
|
||||
INSTR xmm8, xmm2, xmm2
|
||||
INSTR xmm9, xmm0, xmm0
|
||||
INSTR xmm10, xmm1, xmm1
|
||||
INSTR xmm11, xmm2, xmm2
|
||||
INSTR xmm12, xmm0, xmm0
|
||||
INSTR xmm13, xmm1, xmm1
|
||||
INSTR xmm14, xmm2, xmm2
|
||||
INSTR xmm15, xmm0, xmm0
|
||||
INSTR xmm16, xmm1, xmm1
|
||||
INSTR xmm17, xmm2, xmm2
|
||||
INSTR xmm18, xmm0, xmm0
|
||||
INSTR xmm19, xmm1, xmm1
|
||||
INSTR xmm20, xmm2, xmm2
|
||||
INSTR xmm21, xmm0, xmm0
|
||||
INSTR xmm22, xmm1, xmm1
|
||||
INSTR xmm23, xmm2, xmm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vunpckhpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,67 +0,0 @@
|
||||
#define INSTR vxorpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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, ymm0, ymm0
|
||||
INSTR ymm4, ymm1, ymm1
|
||||
INSTR ymm5, ymm2, ymm2
|
||||
INSTR ymm3, ymm0, ymm0
|
||||
INSTR ymm4, ymm1, ymm1
|
||||
INSTR ymm5, ymm2, ymm2
|
||||
INSTR ymm6, ymm0, ymm0
|
||||
INSTR ymm7, ymm1, ymm1
|
||||
INSTR ymm8, ymm2, ymm2
|
||||
INSTR ymm9, ymm0, ymm0
|
||||
INSTR ymm10, ymm1, ymm1
|
||||
INSTR ymm11, ymm2, ymm2
|
||||
INSTR ymm12, ymm0, ymm0
|
||||
INSTR ymm13, ymm1, ymm1
|
||||
INSTR ymm14, ymm2, ymm2
|
||||
INSTR ymm15, ymm0, ymm0
|
||||
INSTR ymm16, ymm1, ymm1
|
||||
INSTR ymm17, ymm2, ymm2
|
||||
INSTR ymm18, ymm0, ymm0
|
||||
INSTR ymm19, ymm1, ymm1
|
||||
INSTR ymm20, ymm2, ymm2
|
||||
INSTR ymm21, ymm0, ymm0
|
||||
INSTR ymm22, ymm1, ymm1
|
||||
INSTR ymm23, ymm2, ymm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,67 +0,0 @@
|
||||
#define INSTR vxorpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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 ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
INSTR ymm0, ymm1, ymm0
|
||||
INSTR ymm1, ymm0, ymm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vxorpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm6, xmm0, xmm0
|
||||
INSTR xmm7, xmm1, xmm1
|
||||
INSTR xmm8, xmm2, xmm2
|
||||
INSTR xmm9, xmm0, xmm0
|
||||
INSTR xmm10, xmm1, xmm1
|
||||
INSTR xmm11, xmm2, xmm2
|
||||
INSTR xmm12, xmm0, xmm0
|
||||
INSTR xmm13, xmm1, xmm1
|
||||
INSTR xmm14, xmm2, xmm2
|
||||
INSTR xmm15, xmm0, xmm0
|
||||
INSTR xmm16, xmm1, xmm1
|
||||
INSTR xmm17, xmm2, xmm2
|
||||
INSTR xmm18, xmm0, xmm0
|
||||
INSTR xmm19, xmm1, xmm1
|
||||
INSTR xmm20, xmm2, xmm2
|
||||
INSTR xmm21, xmm0, xmm0
|
||||
INSTR xmm22, xmm1, xmm1
|
||||
INSTR xmm23, xmm2, xmm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vxorpd
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vxorps
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm3, xmm0, xmm0
|
||||
INSTR xmm4, xmm1, xmm1
|
||||
INSTR xmm5, xmm2, xmm2
|
||||
INSTR xmm6, xmm0, xmm0
|
||||
INSTR xmm7, xmm1, xmm1
|
||||
INSTR xmm8, xmm2, xmm2
|
||||
INSTR xmm9, xmm0, xmm0
|
||||
INSTR xmm10, xmm1, xmm1
|
||||
INSTR xmm11, xmm2, xmm2
|
||||
INSTR xmm12, xmm0, xmm0
|
||||
INSTR xmm13, xmm1, xmm1
|
||||
INSTR xmm14, xmm2, xmm2
|
||||
INSTR xmm15, xmm0, xmm0
|
||||
INSTR xmm16, xmm1, xmm1
|
||||
INSTR xmm17, xmm2, xmm2
|
||||
INSTR xmm18, xmm0, xmm0
|
||||
INSTR xmm19, xmm1, xmm1
|
||||
INSTR xmm20, xmm2, xmm2
|
||||
INSTR xmm21, xmm0, xmm0
|
||||
INSTR xmm22, xmm1, xmm1
|
||||
INSTR xmm23, xmm2, xmm2
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,65 +0,0 @@
|
||||
#define INSTR vxorps
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
# copy DP 1.0
|
||||
vmovaps xmm0, xmm0
|
||||
vmovaps xmm1, xmm0
|
||||
# Create DP 2.0
|
||||
vaddpd xmm1, xmm1, xmm1
|
||||
# Create DP 0.5
|
||||
vdivpd xmm2, xmm0, xmm1
|
||||
loop:
|
||||
inc i
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
INSTR xmm0, xmm1, xmm0
|
||||
INSTR xmm1, xmm0, xmm0
|
||||
cmp i, N
|
||||
jl loop
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR xor
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR dl, al
|
||||
INSTR r9l, bl
|
||||
INSTR r10l, cl
|
||||
INSTR dl, al
|
||||
INSTR r9l, bl
|
||||
INSTR r10l, cl
|
||||
INSTR r11l, al
|
||||
INSTR r12l, bl
|
||||
INSTR r13l, cl
|
||||
INSTR r14l, al
|
||||
INSTR r15l, bl
|
||||
INSTR al, cl
|
||||
INSTR bl, al
|
||||
INSTR cl, bl
|
||||
INSTR dl, cl
|
||||
INSTR r9l, al
|
||||
INSTR r10l, bl
|
||||
INSTR r11l, cl
|
||||
INSTR r12l, al
|
||||
INSTR r13l, bl
|
||||
INSTR r14l, cl
|
||||
INSTR r15l, al
|
||||
INSTR al, bl
|
||||
INSTR bl, cl
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
@@ -1,100 +0,0 @@
|
||||
#define INSTR xor
|
||||
#define NINST 24
|
||||
#define N edi
|
||||
#define i r8d
|
||||
|
||||
|
||||
.intel_syntax noprefix
|
||||
.globl ninst
|
||||
.data
|
||||
ninst:
|
||||
.long NINST
|
||||
.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
|
||||
push rax
|
||||
push rbx
|
||||
push rcx
|
||||
push rdx
|
||||
push r9
|
||||
push r10
|
||||
push r11
|
||||
push r12
|
||||
push r13
|
||||
push r14
|
||||
push r15
|
||||
xor rax, rax
|
||||
xor rbx, rbx
|
||||
xor rcx, rcx
|
||||
xor rdx, rdx
|
||||
xor r9, r9
|
||||
xor r10, r10
|
||||
xor r11, r11
|
||||
xor r12, r12
|
||||
xor r13, r13
|
||||
xor r14, r14
|
||||
xor r15, r15
|
||||
# copy DP 1.0
|
||||
vmovq rax, xmm0
|
||||
vmovq rbx, xmm0
|
||||
# Create DP 2.0
|
||||
add rbx, rax
|
||||
# Create DP 0.5
|
||||
div rax
|
||||
movq rcx, rax
|
||||
vmovq rax, xmm0
|
||||
loop:
|
||||
inc i
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
INSTR al, bl
|
||||
INSTR bl, al
|
||||
cmp i, N
|
||||
jl loop
|
||||
pop r15
|
||||
pop r14
|
||||
pop r13
|
||||
pop r12
|
||||
pop r11
|
||||
pop r10
|
||||
pop r9
|
||||
pop rdx
|
||||
pop rcx
|
||||
pop rbx
|
||||
pop rax
|
||||
done:
|
||||
mov rsp, rbp
|
||||
pop rbp
|
||||
ret
|
||||
.size latency, .-latency
|
||||
Reference in New Issue
Block a user