Hello,
Noticed in the Tosa specification that some operators like ReductionSum and Negate have an acc_t field in the Supported Data Types table, but no argument that corresponds with that field, while other operators like the Convolutions have a matching acc_type attribute.
I think an attribute is needed for these operations to convey the type of the accumulator.
Similarly is the case for Resize resize_t field.
Hi,
We don’t have an attribute for these operators, as the current specification doesn’t allow different accumulator types for these operations. For the convolutions, the choice of accumulator type is selectable, so you could have a fp16 convolution with either an fp16 accumulator or an fp32 accumulator. For reduce_sum, a fp16 input always uses an fp16 accumulator.
It’s an attempt to balance broad functionality with controlling the scope that needs to be tested, so it would be interesting to understand areas where it would be valuable to support multiple accumulator types for these operators.
Eric
