Python square brackets between function name and arguments: func[…](…)
Question:
I was learning how to accelerate python computations on GPU from this notebook, where one line confuses me:
mandel_kernel[griddim, blockdim](-2.0, 1.0, -1.0, 1.0, d_image, 20)
Here, mandel_kernel is a decorated (by cuda.jit) function, griddim and blockdim are tuples of length 2: griddim=(32,16), blockdim=(32,8).
Is this square brackets in between function name and argument list part of python syntax, or something specific to the cuda.jit decoration?
Answers:
This is valid python syntax, i’ll try to break it down for you:
mandel_kernel is a dict, whose keys are 2-tuples (griddim, blockdim) and values are method (which is valid since methods are objects in python)
mandel_kernel[griddim, blockdim] therefore ‘returns’ (or evaluates to) a method
mandel_kernel[griddim, blockdim](-2.0, 1.0, -1.0, 1.0, d_image, 20) therefore calls that method with whatever arguments are inside the parenthesis.
This one line could be rewritten in three lines like so:
key = tuple(griddim, blockdim)
method = mandel_kernel[key]
method(-2.0, 1.0, -1.0, 1.0, d_image, 20)
As iCart stated, @cuda.jit overloads the getitem method which changes the behavior of the brackets to instead of acting like a dict, the brackets are used to pass a tuple in a manner spiritually similar to the CUDA C/C++ runtime API syntax that looks like:
cudaKernel<<<numBlocks, threadsPerBlock>>>(parameter1, parameter2, …, parameterN);
It’s cute, but it’s anything but clear at first glance IMO. It’s doubly confusing when the tuple is computed by a lambda on the fly as it is for the triton library examples. I’m gathering they got this idea from numba.
I was learning how to accelerate python computations on GPU from this notebook, where one line confuses me:
mandel_kernel[griddim, blockdim](-2.0, 1.0, -1.0, 1.0, d_image, 20)
Here, mandel_kernel is a decorated (by cuda.jit) function, griddim and blockdim are tuples of length 2: griddim=(32,16), blockdim=(32,8).
Is this square brackets in between function name and argument list part of python syntax, or something specific to the cuda.jit decoration?
This is valid python syntax, i’ll try to break it down for you:
mandel_kernel is a dict, whose keys are 2-tuples (griddim, blockdim) and values are method (which is valid since methods are objects in python)
mandel_kernel[griddim, blockdim] therefore ‘returns’ (or evaluates to) a method
mandel_kernel[griddim, blockdim](-2.0, 1.0, -1.0, 1.0, d_image, 20) therefore calls that method with whatever arguments are inside the parenthesis.
This one line could be rewritten in three lines like so:
key = tuple(griddim, blockdim)
method = mandel_kernel[key]
method(-2.0, 1.0, -1.0, 1.0, d_image, 20)
As iCart stated, @cuda.jit overloads the getitem method which changes the behavior of the brackets to instead of acting like a dict, the brackets are used to pass a tuple in a manner spiritually similar to the CUDA C/C++ runtime API syntax that looks like:
cudaKernel<<<numBlocks, threadsPerBlock>>>(parameter1, parameter2, …, parameterN);
It’s cute, but it’s anything but clear at first glance IMO. It’s doubly confusing when the tuple is computed by a lambda on the fly as it is for the triton library examples. I’m gathering they got this idea from numba.