volatile data

Accelerated Computing HPC Compilers Legacy PGI Compilers
#1

I realized that if I declare a variable as ‘volatile’, the newly assigned value is not seen in the next statement. Is this a bug in compiler PGI 12.2?

attributes(global) subroutine test()
INTEGER :: lcc_ct, ryr_ct

ii = (blockIdx%x-1) * blockDim%x + threadIdx%x

IF (ii .LE. 20000) THEN
ryr_ct = 1000
compPdt_dev(ii,1) = ryr_ct + 10 ! seen as 1010
ENDIF

ENDIF



attributes(global) subroutine test()
INTEGER,VOLATILE :: lcc_ct, ryr_ct

ii = (blockIdx%x-1) * blockDim%x + threadIdx%x

IF (ii .LE. 20000) THEN
ryr_ct = 1000
compPdt_dev(ii,1) = ryr_ct + 10 ! seen as 10
ENDIF

ENDIF

Thanks,
Tuan

#2

Hi Tuan,

The CUDA C we generate are identical between the two examples, with the exception of volatile attribute (See the output from -Mcuda=keepgpu). There may be a problem with the NVIDIA back-end compiler or an issue with your code. Though, we would need to see a full reproducing example to determine which. Can you put one together?

Thanks,
Mat

#3

Hi Mat,
I sent a working sample that reproduce the error yesterday to trs@pgroup.com Could you please take a look at that.

Thanks,
Tuan

#4

Thanks Tuan. Turns out that it is our error but only occurs when optimisation is applied. When I went back and recompiled your small example with “-O2”, then I saw the bug.

The problem is that the compiler is performing constant propagation and replaces the “ryr_ct” variable with the constant value “1010”. When “volatile” is used, the reference to “ryr_ct” in “compPdt_dev(ii,1) = ryr_ct + 10” is correctly preserved, however, the assignment “ryr_ct = 1000” is still removed. I have sent a report to our engineers (TPR#18530) to be fixed.

  • Mat