CUDA Fortran不必要數據拷貝

原創 wcj0626 2018-09-03 18:48

首先請看下面的代碼：

module add
	use cudafor
	contains
	attributes(global) subroutine call_global(a)
		integer :: a(:)
		integer :: i
		i=threadIdx%x+(blockIdx%x-1)*blockDim%x
		a(i)=i
	end subroutine call_global
end module

program test
	use add
	implicit none
	integer,device :: d_a(256)
	integer :: a(256)
	integer :: i	

	do i=1,10
		call call_global<<<2,128>>>(d_a)
	enddo
	a=d_a
	print*,"a(1)=",a(1)

end program
把COMPUTE_PROFILE設置爲1，profile的分析結果： 

# CUDA_PROFILE_LOG_VERSION 2.0
# CUDA_DEVICE 0 GeForce GTX 750
# CUDA_CONTEXT 1
# TIMESTAMPFACTOR 13859dd8cdd69c55
method,gputime,cputime,occupancy
method=[ memcpyHtoD ] gputime=[ 0.960 ] cputime=[ 6.234 ] 
method=[ add_call_global_ ] gputime=[ 3.360 ] cputime=[ 18.620 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.864 ] cputime=[ 4.235 ] 
method=[ add_call_global_ ] gputime=[ 1.536 ] cputime=[ 5.075 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 3.925 ] 
method=[ add_call_global_ ] gputime=[ 1.536 ] cputime=[ 4.782 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 3.893 ] 
method=[ add_call_global_ ] gputime=[ 1.568 ] cputime=[ 4.502 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 3.729 ] 
method=[ add_call_global_ ] gputime=[ 1.568 ] cputime=[ 4.471 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 3.835 ] 
method=[ add_call_global_ ] gputime=[ 1.568 ] cputime=[ 4.420 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.800 ] cputime=[ 3.744 ] 
method=[ add_call_global_ ] gputime=[ 1.568 ] cputime=[ 4.660 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.864 ] cputime=[ 3.731 ] 
method=[ add_call_global_ ] gputime=[ 1.568 ] cputime=[ 4.437 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 3.689 ] 
method=[ add_call_global_ ] gputime=[ 1.600 ] cputime=[ 4.476 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 3.696 ] 
method=[ add_call_global_ ] gputime=[ 1.568 ] cputime=[ 4.375 ] occupancy=[ 1.000 ] 
method=[ memcpyDtoH ] gputime=[ 2.080 ] cputime=[ 14.631 ]

什麼情況？每次循環之前都有數據拷貝！！然後給出PGI的解釋： 

The array is declared assumed-shape in the global subroutine, so we have to allocate, initialize, and pass the descriptor every time.

PGI給出兩個解決辦法：

1、參數以reference形式傳遞

  修改代碼如下：

module add
	use cudafor
	contains
	attributes(global) subroutine call_global(a)
		integer :: a(*)
		integer :: i
		i=threadIdx%x+(blockIdx%x-1)*blockDim%x
		a(i)=i
	end subroutine call_global
end module

program test
	use add
	implicit none
	integer,device :: d_a(256)
	integer :: a(256)
	integer :: i	

	do i=1,10
		call call_global<<<2,128>>>(d_a)
	enddo
	a=d_a
	print*,"a(1)=",a(1)

end program
    如果不注意看的話，估計沒看出啥區別，請注意代碼第5行。看懂了吧？

   profile的分析結果如下：

# CUDA_PROFILE_LOG_VERSION 2.0
# CUDA_DEVICE 0 GeForce GTX 750
# CUDA_CONTEXT 1
# TIMESTAMPFACTOR 13859dd8cdbb1678
method,gputime,cputime,occupancy
method=[ add_call_global_ ] gputime=[ 3.168 ] cputime=[ 16.332 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.408 ] cputime=[ 3.883 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.376 ] cputime=[ 3.722 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.344 ] cputime=[ 3.291 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.408 ] cputime=[ 3.566 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.408 ] cputime=[ 3.438 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.408 ] cputime=[ 3.333 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.376 ] cputime=[ 3.304 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.408 ] cputime=[ 3.311 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.408 ] cputime=[ 3.226 ] occupancy=[ 1.000 ] 
method=[ memcpyDtoH ] gputime=[ 2.080 ] cputime=[ 12.897 ]
   木有了每次循環之前都有數據拷貝的過程。
2、把變量放置到moule中(測試有問題)

我們繼續修改代碼如下：

module add
	use cudafor
	integer,device :: d_a(256)
	contains
	attributes(global) subroutine call_global(a)
		integer :: a(:)
		integer :: i
		i=threadIdx%x+(blockIdx%x-1)*blockDim%x
		a(i)=i
	end subroutine call_global
end module

program test
	use add
	implicit none
	
	integer :: a(256)
	integer :: i	

	
	do i=1,10
		call call_global<<<2,128>>>(d_a)
	enddo
	a=d_a
	print*,"a(1)=",a(1)

end program
profile的分析結果如下： 

# CUDA_PROFILE_LOG_VERSION 2.0
# CUDA_DEVICE 0 GeForce GTX 750
# CUDA_CONTEXT 1
# TIMESTAMPFACTOR 13859fff450cdbde
method,gputime,cputime,occupancy
method=[ memcpyHtoD ] gputime=[ 0.864 ] cputime=[ 21.286 ] 
method=[ add_call_global_ ] gputime=[ 2.112 ] cputime=[ 7.290 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 69.106 ] 
method=[ add_call_global_ ] gputime=[ 1.504 ] cputime=[ 6.707 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 54.235 ] 
method=[ add_call_global_ ] gputime=[ 1.504 ] cputime=[ 6.123 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 75.521 ] 
method=[ add_call_global_ ] gputime=[ 1.504 ] cputime=[ 6.707 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 78.729 ] 
method=[ add_call_global_ ] gputime=[ 1.504 ] cputime=[ 6.123 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.864 ] cputime=[ 67.357 ] 
method=[ add_call_global_ ] gputime=[ 1.504 ] cputime=[ 5.832 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 53.944 ] 
method=[ add_call_global_ ] gputime=[ 1.536 ] cputime=[ 5.832 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 53.652 ] 
method=[ add_call_global_ ] gputime=[ 1.536 ] cputime=[ 5.540 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 76.105 ] 
method=[ add_call_global_ ] gputime=[ 1.472 ] cputime=[ 6.123 ] occupancy=[ 1.000 ] 
method=[ memcpyHtoD ] gputime=[ 0.832 ] cputime=[ 69.690 ] 
method=[ add_call_global_ ] gputime=[ 1.536 ] cputime=[ 5.540 ] occupancy=[ 1.000 ] 
method=[ memcpyDtoH ] gputime=[ 2.688 ] cputime=[ 87.477 ] 

問題還是依舊。等待繼續反饋看結果。

7-31號補充：

把數據放置到module中用錯了，正確的方法如下：

module add
	use cudafor
	integer,device :: d_a(256,16)
	contains
	attributes(global) subroutine call_global()
		integer :: i,j
		i=threadIdx%x+(blockIdx%x-1)*blockDim%x
		j=threadIdx%y+(blockIdx%y-1)*blockDim%y	

		d_a(i,j)=i+j
	end subroutine call_global
end module

program test
	use add
	implicit none
	
	integer :: a(256,16)
	integer :: i	
	type(dim3) :: tblock,tgrid

	tblock=dim3(128,2,1)
	tgrid=dim3(2,8,1)
	do i=1,10
		call call_global<<<tgrid,tblock>>>()
	enddo
	a=d_a
	print*,"a(1)=",a(1,1)

end program

profile分析結果： 

# CUDA_PROFILE_LOG_VERSION 2.0
# CUDA_DEVICE 0 GeForce GTX 750
# CUDA_CONTEXT 1
# TIMESTAMPFACTOR 1385ef0816adc600
method,gputime,cputime,occupancy
method=[ add_call_global_ ] gputime=[ 1.888 ] cputime=[ 8.164 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.408 ] cputime=[ 5.249 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.408 ] cputime=[ 4.957 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.472 ] cputime=[ 4.957 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.440 ] cputime=[ 4.957 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.408 ] cputime=[ 4.665 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.440 ] cputime=[ 4.374 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.440 ] cputime=[ 4.957 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.408 ] cputime=[ 4.665 ] occupancy=[ 1.000 ] 
method=[ add_call_global_ ] gputime=[ 1.408 ] cputime=[ 4.374 ] occupancy=[ 1.000 ] 
method=[ memcpyDtoH ] gputime=[ 4.768 ] cputime=[ 167.663 ]
解決了不必要的H2D的數據拷貝。

發表評論
所有評論
還沒有人評論，想成為第一個評論的人麼? 請在上方評論欄輸入並且點擊發布.
相關文章

CUDA Fortran for Scientists and Engineers第二版翻譯

wcj0626 2020-02-20 15:39:21

cuda fotran 生成隨機數

想不出名字啦 2018-08-26 06:34:28