Abstract: An embodiment of an integrated circuit comprises circuitry to generate a cache tag for data to be stored in a cache memory, store a first portion of the cache tag in a primary tag memory, and store a second portion of the cache tag in a secondary tag memory, wherein a size of the first portion is smaller than a size of the second portion. Other embodiments are disclosed and claimed.

Abstract: Examples include techniques for near data acceleration for a multi-core architecture. A near data processor included in a memory controller of a processor may access data maintained in a memory device coupled with the near data processor via one or more memory channels responsive to a work request to execute a kernel, an application or a loop routine using the accessed data to generate values. The near data processor provides an indication to the requestor of the work request that values have been generated.

Abstract: An DNN accelerator can perform fixed-point emulation of floating-point computation. In a multiplication operation on two floating-point matrices, the DNN accelerator determines an extreme exponent for a row in the first floating-point matrix and determines another extreme exponent for a column in the second floating-point matrix. The row and column can be converted to fixed-point vectors based on the extreme exponents. The two fixed-point vectors are fed into a PE array in the DNN accelerator. The PE array performs a multiplication operation on the two fixed-point vectors and generates a fixed-point inner product. The fixed-point inner product can be converted back to a floating-point inner product based on the extreme exponents. The floating-point inner product is an element in the matrix resulted from the multiplication operation on the two floating-point matrices. The matrix can be accumulated with another matrix resulted from a fixed-point emulation of a floating-point matrix multiplication.

Abstract: An embodiment of an integrated circuit comprises circuitry to generate a cache tag for data to be stored in a cache memory, store a first portion of the cache tag in a primary tag memory, and store a second portion of the cache tag in a secondary tag memory, wherein a size of the first portion is smaller than a size of the second portion. Other embodiments are disclosed and claimed.

Abstract: Methods, apparatus, systems, and articles of manufacture are disclosed for improving data transfer for heterogeneous programs. An example apparatus includes instructions in the apparatus, and processor circuitry to at least one of execute or instantiate the instructions to determine a runtime associated with executing a code object by a heterogeneous electronic device based on at least one of a location of a memory object or a data transfer penalty, the data transfer penalty associated with access of the memory object in response to execution of the code object, identify a memory operation for the memory object based on the runtime, and generate an executable file based on the memory operation, the executable file, when executed, to cause execution of the code object by at least one of first hardware or second hardware of the heterogeneous electronic device based on the memory operation.

Abstract: Systems, apparatuses and methods may provide for technology that determines that a control loop is to be executed for an unspecified number of iterations and automatically forces the control loop to be executed for a fixed number of iterations in addition to the unspecified number of iterations, where execution of the control loop for the fixed number of iterations is conducted in parallel. In one example, the technology also removes one or more dataflow tokens associated with the execution of the control loop for the fixed number of iterations.

Abstract: Systems, apparatuses and methods may provide for technology that determines a plurality of memory operations associated with a data-flow graph that represents a computer code, where a spatial architecture executes the data-flow graph and the spatial architecture includes a plurality of memory controllers, randomly assigns one or more of the plurality of memory operations to one or more of the plurality of memory controllers to generate a first allocation of the plurality of memory operations to the memory controllers, and determines that the first allocation is to be stored as a permanent memory allocation based on a first performance metric associated with the first allocation.

Abstract: Systems, methods, and apparatuses relating to arbitration among a plurality of memory interface circuits in a configurable spatial accelerator are described. In one embodiment, a configurable spatial accelerator (CSA) includes a plurality of processing elements; a plurality of request address file (RAF) circuits, and a circuit switched interconnect network between the plurality of processing elements and the RAF circuits. As a dataflow architecture, embodiments of CSA have a unique memory architecture where memory accesses are decoupled into an explicit request and response phase allowing pipelining through memory. Certain embodiments herein provide for improved memory sub-system design via arbitration and the improvements to arbitration discussed herein.

Abstract: Systems, apparatuses and methods may provide for technology that determines that a control loop is to be executed for an unspecified number of iterations and automatically forces the control loop to be executed for a fixed number of iterations in addition to the unspecified number of iterations, where execution of the control loop for the fixed number of iterations is conducted in parallel. In one example, the technology also removes one or more dataflow tokens associated with the execution of the control loop for the fixed number of iterations.

Abstract: Systems, methods, and apparatuses relating to memory interface circuit allocation in a configurable spatial accelerator are described. In one embodiment, a configurable spatial accelerator (CSA) includes a plurality of processing elements; a plurality of request address file (RAF) circuits, and a circuit switched interconnect network between the plurality of processing elements and the RAF circuits. As a dataflow architecture, embodiments of CSA have a unique memory architecture where memory accesses are decoupled into an explicit request and response phase allowing pipelining through memory. Certain embodiments herein provide for an improved memory sub-system design via the improvements to allocation discussed herein.

Abstract: Systems, methods, and apparatuses relating to arbitration among a plurality of memory interface circuits in a configurable spatial accelerator are described. In one embodiment, a configurable spatial accelerator (CSA) includes a plurality of processing elements; a plurality of request address file (RAF) circuits, and a circuit switched interconnect network between the plurality of processing elements and the RAF circuits. As a dataflow architecture, embodiments of CSA have a unique memory architecture where memory accesses are decoupled into an explicit request and response phase allowing pipelining through memory. Certain embodiments herein provide for improved memory sub-system design via arbitration and the improvements to arbitration discussed herein.

Abstract: Systems, methods, and apparatuses relating to time-multiplexing circuitry in a configurable spatial accelerator are described. In one embodiment, a configurable spatial accelerator (CSA) includes a plurality of processing elements; and a time-multiplexed, circuit switched interconnect network between the plurality of processing elements. In another embodiment, a configurable spatial accelerator (CSA) includes a plurality of time-multiplexed processing elements; and a time-multiplexed, circuit switched interconnect network between the plurality of time-multiplexed processing elements.

Abstract: Systems, apparatuses and methods may provide for technology that determines that a control loop is to be executed for an unspecified number of iterations and automatically forces the control loop to be executed for a fixed number of iterations in addition to the unspecified number of iterations, where execution of the control loop for the fixed number of iterations is conducted in parallel. In one example, the technology also removes one or more dataflow tokens associated with the execution of the control loop for the fixed number of iterations.

Abstract: Systems, methods, and apparatuses relating to swizzle operations and disable operations in a configurable spatial accelerator (CSA) are described. Certain embodiments herein provide for an encoding system for a specific set of swizzle primitives across a plurality of packed data elements in a CSA.

Abstract: Systems, methods, and apparatuses relating to memory interface circuit allocation in a configurable spatial accelerator are described. In one embodiment, a configurable spatial accelerator (CSA) includes a plurality of processing elements; a plurality of request address file (RAF) circuits, and a circuit switched interconnect network between the plurality of processing elements and the RAF circuits. As a dataflow architecture, embodiments of CSA have a unique memory architecture where memory accesses are decoupled into an explicit request and response phase allowing pipelining through memory. Certain embodiments herein provide for an improved memory sub-system design via the improvements to allocation discussed herein.

Abstract: Systems, methods, and apparatuses relating to swizzle operations and disable operations in a configurable spatial accelerator (CSA) are described. Certain embodiments herein provide for an encoding system for a specific set of swizzle primitives across a plurality of packed data elements in a CSA.

Abstract: Systems, methods, and apparatuses relating to in-network storage for a configurable spatial accelerator are described.

Abstract: Systems, apparatuses and methods may provide for technology that determines a plurality of memory operations associated with a data-flow graph that represents a computer code, where a spatial architecture executes the data-flow graph and the spatial architecture includes a plurality of memory controllers, randomly assigns one or more of the plurality of memory operations to one or more of the plurality of memory controllers to generate a first allocation of the plurality of memory operations to the memory controllers, and determines that the first allocation is to be stored as a permanent memory allocation based on a first performance metric associated with the first allocation.

Abstract: Systems, methods, and apparatuses relating to in-network storage for a configurable spatial accelerator are described.

Abstract: Systems, apparatuses and methods may provide for technology that determines that a control loop is to be executed for an unspecified number of iterations and automatically forces the control loop to be executed for a fixed number of iterations in addition to the unspecified number of iterations, where execution of the control loop for the fixed number of iterations is conducted in parallel. In one example, the technology also removes one or more dataflow tokens associated with the execution of the control loop for the fixed number of iterations.