Direct bindings to MLIR TOSA operations

[bryjen]

The frontend/backend should have direct bindings to the following MLIR TOSA ops. The parser and the emitter are straightforward; the semantic analyzer, however, should be able to assert that the passed arguments are valid for the operation (for example asserting that two tensors' shapes are valid for matrix multiplication) as well as perform inference on parameterized types (for example returning tensor size information after the multiplication of two tensors). Furthermore, these types should be easily translatable to MLIR TOSA types on the backend side.

Parser:

• tensors
• tosa.reshape
• tosa.transpose
• tosa.tile
• tosa.slice
• tosa.concat
• tosa.identity
• tosa.cast
• tosa.conv2d
• tosa.depthwise_conv2d
• tosa.matmul
• tosa.fully_connected
• tosa.avg_pool2d
• tosa.max_pool2d
• tosa.pad
• tosa.relu
• tosa.sigmoid
• tosa.tanh
• tosa.softmax

Semantics (type inference + argument validation):

• tensors
• tosa.reshape
• tosa.transpose
• tosa.tile
• tosa.slice
• tosa.concat
• tosa.identity
• tosa.cast
• tosa.conv2d
• tosa.depthwise_conv2d
• tosa.matmul
• tosa.fully_connected
• tosa.avg_pool2d
• tosa.max_pool2d
• tosa.pad
• tosa.relu
• tosa.sigmoid
• tosa.tanh
• tosa.softmax

Emitter:

• tensors
• tosa.reshape
• tosa.transpose
• tosa.tile
• tosa.slice
• tosa.concat
• tosa.identity
• tosa.cast
• tosa.conv2d
• tosa.depthwise_conv2d
• tosa.matmul
• tosa.fully_connected
• tosa.avg_pool2d
• tosa.max_pool2d
• tosa.pad
• tosa.relu
• tosa.sigmoid
• tosa.tanh
• tosa.softmax

[bryjen]

Auto-generated API reference for each of the listed MLIR TOSA ops listed above:

Operation	Operands	Attributes	Result Type	Notes
tosa.reshape	%input (tensor)	shape (array of int)	tensor<...>	New shape must have same number of elements
tosa.transpose	%input (tensor), %perm (1D tensor)	(none)	tensor<...>	Permutes dimensions according to %perm
tosa.tile	%input (tensor), %multiples (1D tensor)	(none)	tensor<...>	Repeats input along dimensions
tosa.slice	%input (tensor)	start, size (arrays)	tensor<...>	Extracts sub-tensor with bounds
tosa.concat	Variadic %inputs (tensors)	axis (int)	tensor<...>	Concatenates tensors along given axis
tosa.identity	%input (tensor)	(none)	same as input	No-op, used to forward or isolate values
tosa.cast	%input (tensor)	(none)	tensor<...>	Casts tensor to a different data type
tosa.conv2d	%input, %filter, %bias (tensors)	pad, stride, dilation (arrays)	tensor<...>	2D convolution operation
tosa.depthwise_conv2d	%input, %filter, %bias (tensors)	pad, stride, dilation (arrays)	tensor<...>	Depthwise 2D convolution operation
tosa.matmul	%a, %b (tensors)	(none)	tensor<...>	Matrix multiplication
tosa.fully_connected	%input, %weights, %bias (tensors)	(none)	tensor<...>	Fully connected layer operation
tosa.avg_pool2d	%input (tensor)	kernel, stride, pad (arrays)	tensor<...>	2D average pooling operation
tosa.max_pool2d	%input (tensor)	kernel, stride, pad (arrays)	tensor<...>	2D max pooling operation
tosa.pad	%input, %padding (tensor)	(none)	tensor<...>	Pads tensor with specified padding
tosa.relu	%input (tensor)	(none)	tensor<...>	Applies ReLU activation function
tosa.sigmoid	%input (tensor)	(none)	tensor<...>	Applies Sigmoid activation function
tosa.tanh	%input (tensor)	(none)	tensor<...>	Applies Tanh activation function
tosa.softmax	%input (tensor)	axis (int)	tensor<...>	Applies Softmax function along specified axis