What is the meaning of the gpuparams (64,256,1,1,1,1) ?

[Jiede1]

Hello，i am studying the codes on the test examples beneath the gpuenabler dirctory，and I can't understand the meaning of the number like (64,256,1,...) ? Could you please explain this?

[josiahsams]

The register CUDA kernel will be invoked twice (2 stages) and we are providing the GPU dimensions for each stage (64, 256, 1, 1, 1, 1).

In this case, it means for the first stage, schedule the kernel with a 256: threadsPerBlock.x and 64: blockDim.x and set the blockDim.y,blockDim.z,threadsPerBlock.y,threadsPerBlock.z to 1.

That particular kernel is expected to be executed in two stages Ref: https://github.com/IBMSparkGPU/GPUEnabler/blob/master/examples/src/main/resources/SparkGPUExamples.cu#L29

So reduce operations are done in 2 stages, Block-vise reduce and 1 whole reduce with just one block & one thread.

Hope it helps.

[Jiede1]

Thanks, but I have another question now ...if 64 means blockDim.x ,then the number of the threads of a block are 6411 =64 ,and threadsPerBlock.x is 256 , not 64 , it doesn't make sense...

[josiahsams]

To be clear, Number of Blocks required is 64(GridDim.x) and each block is of size 256 (blockDim.x).

Refer to this PR for more clarity, #60

[Jiede1]

thanks a lot ,it helps.