Programmable data movement processor for collective communication offload

A data movement processor, comprising:

• a microcode fetch circuit configured to receive a microcode instruction from a client system, wherein the microcode instruction specifies data transfers of a collective communication operation;
• a plurality of compute unit circuits;
• and
• a microcode dispatch circuit coupled to the plurality of compute unit circuits and the microcode fetch circuit, wherein the microcode dispatch circuit is configured to distribute slices of the microcode instruction to different ones of the plurality of compute unit circuits to effectuate the collective communication operation;
• wherein a selected compute unit circuit of the plurality of compute unit circuits, in response to the microcode instruction, configures a network-on-chip to establish data paths therein that communicatively link different ones of a plurality of circuits for performing the data transfers of the collective communication operation.

A method, comprising:

• receiving, within a data movement processor disposed in an integrated circuit, a microcode instruction specifying a collective communication operation;
• in response to the microcode instruction, establishing data paths within a network-on-chip disposed in the integrated circuit, wherein the data paths communicatively link a plurality of circuits involved in data transfers that form the collective communication operation as specified by the microcode instruction;
• and
• initiating, by the data movement processor, the data transfers between the plurality of circuits through the data paths of the network-on-chip by providing commands to a data mover circuit coupled to the network-on-chip.