Global Sync across different blocks in IR builder

Questions
#1

@were @tqchen Is there a way to do global synchronization in IR builder or the hybrid IR builder on the GPU? If I write gpu code, I would separate the two parts into two kernel functions. But I don’t know how to do that in TVM. Currently I’m experiencing bug after transposing data. Thread 0 reads different values from say temp[1000] compared with thread 1000 if the block size if 256.

1 Like
#2

You can launch multiple kernels in tvm. Just make sure you build two blocks of code, both of them bind to a different thread block as outer most

#3

I see. In order to call barrier(CLK_GLOBAL_MEM_FENCE) in IR builder what shall I do?

#4

In theory you do not need to call global fence, because what will happen is that two kernels will be launched

#5

I meant if I don’t want to separate the kernel into two kernel functions.

#6

@were does hybrid interface already expose barrier/ syncthreads apis to python users?

#7

The global barrier do not work well in opencl due to language limitations, so best way would indeed be use two kernels

#8

For one of my use cases, the results seem acceptable. I experimented it first using manually changed opencl code.

#9

CUDA9 cooperative groups allows global sync across work-groups. Does TVM support this feature?

#10

@milindn TVM hasn’t yet take benefit from global sync in CUDA9, contribution is welcomed

1 Like