Getting "TypeError: Unable to parse some datatype(s) found in args" while executing Application builder

[soumitraghosh388]

I am using 2 kernels : master and slave with n number of instances of slave kernels, where n depends on the input. I have defined graph and am trying to build the application.

Graph code below:

class Sample_app(AppBuilder):
    def __init__(self, kernel_1: Kernel, kernel_2: Kernel):
        self.master = kernel_1
        self.slave = kernel_2
        super().__init__()
    def callgraph(self, x_in_1, x_in_2, x_out):
            out_1, out_2 = self.master(x_in_1[0], x_in_2[0])
            slave_out = self.slave (out_1, out_2)
            x_out = [slave_out]

Creating application builder as mentioned below :

app_builder = Sample_app(kernel_1=master, kernel_2=slave)
x_in_1 = np.array([1028112570], dtype=np.int32)
x_in_2 = np.array([526679170], dtype=np.int32)
x_out = np.array([0], dtype=np.int32)
app_builder(x_in_1, x_in_2, x_out)

I am getting below error while executing application builder code :

TypeError: Unable to parse some datatype(s) found in args: (1028112570, 526679170) ... currently supporting [<class 'npu.build.port.BufferPort'>, <class 'npu.build.buffers.Buffer'>, <class 'numpy.ndarray'>, <class 'int'>] datatypes

[mariodruiz]

Hi @soumitraghosh388,

Would you be able to share the kernels source code?
For the app description, it looks like the master kernel takes two input arguments (both scalars) and the slave returns a scalar. Is this correct?

[mariodruiz]

Hi @soumitraghosh388, were you able to build this for the NPU?
Currently, we have a limitation of two inputs and one output or one input and two outputs from the IT to the CT.

[soumitraghosh388]

While building I am getting error (I am using 9 kernels in total, using MT as well for broadcast) :
ValueError: Unable to place kernel nw_stream_kernel_4 - no more free CT tiles

I could not understand : "Currently, we have a limitation of two inputs and one output or one input and two outputs from the IT to the CT." Are these input and output ports of kernels?
Also, I could not find document explaining Rialto framework in details. Can you refer me to those (if any)?

[mariodruiz]

While building I am getting error (I am using 9 kernels in total, using MT as well for broadcast) :
ValueError: Unable to place kernel nw_stream_kernel_4 - no more free CT tiles

We only support one kernel per compute tile at the moment. I should be writing a post to show how to do kernel fusion so you can map multiple kernels in one compute tile. In the meantime, you can check how this is done here: https://github.com/AMDResearch/Riallto/blob/main/tests/test_kernel_fusion.py

[mariodruiz]

"Currently, we have a limitation of two inputs and one output or one input and two outputs from the IT to the CT." Are these input and output ports of kernels?

No, this is a limitation of input and output buffers from system memory via to the interface tile to the compute tile. A kernel itself could have two inputs and two output buffers.

[mariodruiz]

Also, I could not find document explaining Rialto framework in details. Can you refer me to those (if any)?

Section 4 notebooks describes the framework in detail, starting from writing your own compute kernels to defining the dataflow graph.
Notebook 4_2 and 4_5 describe the two parts in detail.

You can find information about the package here: https://riallto.ai/modules.html