为什么使用浮点类型时,O3优化不提高性能? [英] Why O3 optimization does not improve the performance when using float type?
问题描述
我编译相应的 C
实施两个浮动
和 INT
当我编译它们 O2
几乎每一件事情都是一样的,但是当我使用 O3
标志使用自动向量化能力两者产生变异的速度提升。我看到大会放出来,发现差异,但我不知道为什么编译 GCC
这样吗?之间是什么原因和差异浮动
键入和 INT
键入<?/ p>
乘法之前,我调换的,因为某些原因,第二个矩阵。矩阵的尺寸是128×128和 O2
标INT执行速度可达5.4在同一实现时启用 O3
标记和浮动
实施加速是咬人几乎 0.94
。结果
诠释组件从放:
.L2:
vmovdqa 448(%RDI),%ymm0
MOVL $ c_tra,EAX%
MOVQ%R8,RDX%
vmovdqa(%RDI),%ymm15
vmovdqa%ymm0,-48(%RSP)
vmovdqa 480(%RDI),%ymm0
vmovdqa 32(%RDI),%ymm14
vmovdqa 64(%RDI),%ymm13
vmovdqa 96(%RDI),%ymm12
vmovdqa 128(%RDI),%ymm11
vmovdqa 160(%RDI),%ymm10
vmovdqa 192(%RDI),%ymm9
vmovdqa 224(%RDI),%ymm8
vmovdqa 256(%RDI),%ymm7
vmovdqa 288(%RDI),%ymm6
vmovdqa 320(%RDI),%ymm5
vmovdqa 352(%RDI),%ymm4
vmovdqa 384(%RDI),%ymm3
vmovdqa 416(%RDI),%ymm2
vmovdqa%ymm0,-80(%RSP)
.p2align 4日,10
.p2align 3
.L5:
vpmulld 32(%RAX),%ymm14,%ymm0
vpmulld(RAX%),%ymm15,%ymm1
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 64(%RAX),%ymm13,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 96(%RAX),%ymm12,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 128(RAX%),%ymm11,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 160(RAX%),%ymm10,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 192(RAX%),%ymm9,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 224(RAX%),%ymm8,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 256(RAX%),%ymm7,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 288(RAX%),%ymm6,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 320(RAX%),%ymm5,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 352(RAX%),%ymm4,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 384(RAX%),%ymm3,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vpmulld 416(RAX%),%ymm2,%ymm0
vpaddd%ymm0,%ymm1,%ymm1
vmovdqa -48(%RSP),%ymm0
addq $ 512%RAX
addq $ 4%的RDX
vpmulld -64(RAX%),%ymm0,%ymm0
vpaddd%ymm0,%ymm1,%ymm0
vmovdqa -80(%RSP),%ymm1
vpmulld -32(RAX%),%ymm1,%ymm1
vpaddd%ymm0,%ymm1,%ymm1
vmovdqa%将xmm1,%XMM0
vextracti128 $为0x1,%ymm1,xmm1中的%
vpextrd $ 1,%XMM0,ESI%
vpextrd $ 0%XMM0,ECX%
ADDL%ESI,ECX%
vpextrd $ 2%XMM0,ESI%
ADDL%ESI,ECX%
vpextrd $ 3%XMM0,ESI%
ADDL%ESI,ECX%
vpextrd $ 0%将xmm1,ESI%
ADDL%ESI,ECX%
vpextrd $ 1,%将xmm1,ESI%
ADDL%ESI,ECX%
vpextrd $ 2%将xmm1,ESI%
ADDL%ESI,ECX%
vpextrd $ 3%将xmm1,ESI%
ADDL%ESI,ECX%
MOVL%ECX,-4(%RDX)
cmpq $ c_tra + 65536%RAX
JNE .L5
addq $ 512%R8
addq $ 512%RDI
cmpq $ c_result + 65536%R8
JNE .L2
浮子组件出放:
.L2:
xorl%ESI,ESI%
.p2align 4日,10
.p2align 3
.L7:
MOVQ%RDI,RSI%
xorl%EAX,EAX%
xorl%EDX,EDX%
salq $ 5%RSI
.p2align 4日,10
.p2align 3
.L5:
vcvtsi2ss%EDX,%XMM0,%XMM0
vmovss一个(RCX%,%RAX),%XMM2
vfmadd231ss c_tra(RSI%,%RAX),%XMM2,%XMM0
addq $ 4%RAX
vcvttss2si%XMM0,EDX%
cmpq $ 128,RAX%
JNE .L5
vcvtsi2ss%EDX,%XMM0,%XMM0
vmovss%XMM0,c_result(RCX%,%RDI)
addq $ 4%RDI
cmpq $ 128,%RDI
JNE .L7
这可以看出, vcvttss2si
和 vcvtsi2ss
限制自动向量化我这种转换和经销商的矢量化标记 Ofast
矢量程序自动改变了一些变量prevent。因此,答案是自动矢量具有与转换问题。
I compiled the corresponding C
implementation of two float
and int
matrix multiplication program when I compile them in O2
almost every thing is the same but when I use O3
flag to use auto vectorization capability both of them yield variant speedups. I see the assembly out put and found out the differences but I don't know why GCC
compiled like this? what is the reason and differences between float
type and int
type ?
Before the multiplication I transposed the second matrix because of some reasons. size of the matrices are 128x128 and the speed up of O2
scalar int implementation is 5.4 over the same implementation when I enable O3
flag and for float
implementation speedup is a bite worse almost 0.94
.
Int assembly out put:
.L2:
vmovdqa 448(%rdi), %ymm0
movl $c_tra, %eax
movq %r8, %rdx
vmovdqa (%rdi), %ymm15
vmovdqa %ymm0, -48(%rsp)
vmovdqa 480(%rdi), %ymm0
vmovdqa 32(%rdi), %ymm14
vmovdqa 64(%rdi), %ymm13
vmovdqa 96(%rdi), %ymm12
vmovdqa 128(%rdi), %ymm11
vmovdqa 160(%rdi), %ymm10
vmovdqa 192(%rdi), %ymm9
vmovdqa 224(%rdi), %ymm8
vmovdqa 256(%rdi), %ymm7
vmovdqa 288(%rdi), %ymm6
vmovdqa 320(%rdi), %ymm5
vmovdqa 352(%rdi), %ymm4
vmovdqa 384(%rdi), %ymm3
vmovdqa 416(%rdi), %ymm2
vmovdqa %ymm0, -80(%rsp)
.p2align 4,,10
.p2align 3
.L5:
vpmulld 32(%rax), %ymm14, %ymm0
vpmulld (%rax), %ymm15, %ymm1
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 64(%rax), %ymm13, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 96(%rax), %ymm12, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 128(%rax), %ymm11, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 160(%rax), %ymm10, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 192(%rax), %ymm9, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 224(%rax), %ymm8, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 256(%rax), %ymm7, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 288(%rax), %ymm6, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 320(%rax), %ymm5, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 352(%rax), %ymm4, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 384(%rax), %ymm3, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vpmulld 416(%rax), %ymm2, %ymm0
vpaddd %ymm0, %ymm1, %ymm1
vmovdqa -48(%rsp), %ymm0
addq $512, %rax
addq $4, %rdx
vpmulld -64(%rax), %ymm0, %ymm0
vpaddd %ymm0, %ymm1, %ymm0
vmovdqa -80(%rsp), %ymm1
vpmulld -32(%rax), %ymm1, %ymm1
vpaddd %ymm0, %ymm1, %ymm1
vmovdqa %xmm1, %xmm0
vextracti128 $0x1, %ymm1, %xmm1
vpextrd $1, %xmm0, %esi
vpextrd $0, %xmm0, %ecx
addl %esi, %ecx
vpextrd $2, %xmm0, %esi
addl %esi, %ecx
vpextrd $3, %xmm0, %esi
addl %esi, %ecx
vpextrd $0, %xmm1, %esi
addl %esi, %ecx
vpextrd $1, %xmm1, %esi
addl %esi, %ecx
vpextrd $2, %xmm1, %esi
addl %esi, %ecx
vpextrd $3, %xmm1, %esi
addl %esi, %ecx
movl %ecx, -4(%rdx)
cmpq $c_tra+65536, %rax
jne .L5
addq $512, %r8
addq $512, %rdi
cmpq $c_result+65536, %r8
jne .L2
Float assembly out put:
.L2:
xorl %esi, %esi
.p2align 4,,10
.p2align 3
.L7:
movq %rdi, %rsi
xorl %eax, %eax
xorl %edx, %edx
salq $5, %rsi
.p2align 4,,10
.p2align 3
.L5:
vcvtsi2ss %edx, %xmm0, %xmm0
vmovss a(%rcx,%rax), %xmm2
vfmadd231ss c_tra(%rsi,%rax), %xmm2, %xmm0
addq $4, %rax
vcvttss2si %xmm0, %edx
cmpq $128, %rax
jne .L5
vcvtsi2ss %edx, %xmm0, %xmm0
vmovss %xmm0, c_result(%rcx,%rdi)
addq $4, %rdi
cmpq $128, %rdi
jne .L7
It could be seen that vcvttss2si
and vcvtsi2ss
restrict the auto vectorization I changed some variable to prevent this conversion and auto vectorization flag Ofast
vectorized the program automatically. So the answer is auto-vectorization has a problem with the conversions.
这篇关于为什么使用浮点类型时,O3优化不提高性能?的文章就介绍到这了,希望我们推荐的答案对大家有所帮助,也希望大家多多支持IT屋!