Does TensorProductUniform4x1d not utilize the Tensor core of GPUs

[LHJ1098826475]

Hi

cuet.TensorProduct flattens, squeezes, and splits the descriptor so that it can use the TensorProductUniform4x1d. In fact, GPUs have Tensor core. Will processing data into 1d vectors reduce computational complexity, or may I have misunderstood? Please advise.

[mariogeiger]

Hi,

For now we have 3 kernels and the frontend, as you saw, tries to reduce the given STP to on of these kernels. We are currently working on two things related to that:

• improve the reduction of STP to the available kernels
• add more kernels.

Can you provide details about the specific STP you are executing?

[LHJ1098826475]

Hi,

We are mainly focusing on the TensorProductUniform4x1d interface. Can you provide more information?

In addition, the frontend of cuEqu shows four interfaces: TensorProductUniform3x1d, TensorProductUniform4x1d, FusedTensorProductOp3, and FusedTensorProductOp4, but you mentioned there are only three kernels. Is the backend kernels corresponding to these four interfaces?

Thanks.

[mariogeiger]

The third one is the symmetric contraction. You can see some explanations here.