Intel Corporation

Roger Gramunt

Jesus Corbal

Roger Espasa

Benjamin C Chaffin

Ramon Matas

Julio Gago

Federico Ardanaz

Rolf Kassa

Alexey Suprun

Rammohan Padmanabhan

Edward Thomas Grochowski

Alexander F Heinecke

Joel S Emer

Dipankar Das

Naveen K Mellempudi

Mikhail Smelyanskiy

Jonathan Cannon Hall

Dheevatsa Mudigere

Geoff Lowney

Chung-Lun Chan

Santiago Galan

Toni Juan

Dennis Ryan Bradford

Aditya Kesiraju

Christopher B Wilkerson

Jayesh Gaur

Isaac Hernandez

Neal S Moyer

Matthew G Smith

Jonathan M Eastep

Seth H Pugsley

Ed Grochowski

Taraneh Bahrami

Rohan Sharma

Jesus Corbal San Adrian

Prabhat Jain

Thomas B Maciukenas

Vikram S Sundaram

Gerardo A Fernandez

Michael S Yang

Sufiyan Syed

Priyank Deshpande

Described herein is a graphics processor including a processing resource including a multiplier configured to multiply input associated with the instruction at one of a first plurality of bit widths, an adder configured to add a product output from the multiplier with an accumulator value at one of a second plurality of bit widths, and circuitry to select a first bit width of the first plurality of bit widths for the multiplier and a second bit width of the second plurality of bit widths for the adder.

Optimized compute hardware for machine learning operations

A processing cluster of a processing cluster array comprises a plurality of registers to store input values of vector input operands, the input values of at least some of the vector input operands having different bit lengths than those of other input values of other vector input operands, and a compute unit to execute a dot-product instruction with the vector input operands to perform a number of parallel multiply operations and an accumulate operation per 32-bit lane based on a bit length of the smallest-sized input value of a first vector input operand relative to the 32-bit lane.

One embodiment provides for a compute apparatus to perform machine learning operations, the compute apparatus comprising a fetch unit to fetch a single instruction having multiple input operands, wherein the multiple input operands have an unequal bit-length, a first input operand having a first bit-length and a second input operand having a second bit-length; a decode unit to decode the single instruction into a decoded instruction; an operand length unit to determine a smaller bit-length of the first bit-length and the second bit-length; and a compute unit to perform a matrix operation on the multiple input operands to generate an output value having a bit length of the smaller bit length.

An apparatus is provided which comprises: a component; a voltage generator to supply load current to the component; first one or more circuitries to predict that the load current is to increase from a first time; and second one or more circuitries to, in anticipation of the increase in the load current from the first time, cause the component to execute first instructions during a time period that occurs prior to the first time.

Power noise injection to control rate of change of current

A method, system, and apparatus may initialize a fixed plurality of page table entries for a fixed plurality of pages in memory, each page having a first size, wherein a linear address for each page table entry corresponds to a physical address and the fixed plurality of pages are aligned. A bit in each of the page table entries for the aligned pages may be set to indicate whether or not the fixed plurality of pages is to be treated as one combined page having a second page size larger than the first page size. Other embodiments are described and claimed.

Method, system, and apparatus for page sizing extension

A processor includes a logic for stateless capture of data linear addresses (DLA) during precise event based sampling (PEBS) for an out-of-order execution engine. The engine may include a PEBS unit with logic to increment a counter each time an instance of a designated micro-op is retired a reorder buffer, capture output DLA referenced by an instance of the micro-op that executes after the counter overflows, set a captured bit associated with a reorder buffer identifier for the instance of the micro-op, and store a PEBS record in a debug storage when the instance of the micro-op is retired from the reorder buffer. The designated micro-op references a DLA of a memory accessible to the processor.

Stateless capture of data linear addresses during precise event based sampling

A processor includes a set associative cache and a cache controller. The cache controller makes an initial association between first and second groups of sampled sets in the cache and first and second cache replacement policies. Follower sets in the cache are initially associated with the more conservative of the two policies. Following cache line insertions in a first epoch, the associations between the groups of sampled sets and cache replacement policies are swapped for the next epoch. If the less conservative policy outperforms the more conservative policy during two consecutive epochs, the follower sets are associated with the less conservative policy for the next epoch. Subsequently, if the more conservative policy outperforms the less conservative policy during any epoch, the follower sets are again associated with the more conservative policy. Performance may be measured based the number of cache misses associated with each policy.

System and method for cache replacement using conservative set dueling

An apparatus and method for performing an efficient scatter operation. For example, one embodiment of a processor comprises: an allocator unit to receive a scatter operation comprising a number of data elements and responsively allocate resources to execute the scatter operation; a memory execution cluster comprising at least a portion of the resources to execute the scatter operation, the resources including one or more store data buffers and one or more store address buffers; and a senior store pipeline to transfer store data elements from the store data buffers to system memory using addresses from the store address buffers prior to retirement of the scatter operation.

Method and apparatus for performing an efficient scatter

In one embodiment, a processor includes a frontend unit having an instruction decoder to receive and to decode instructions of a plurality of threads, an execution unit coupled to the instruction decoder to receive and execute the decoded instructions, and an instruction retirement unit having a retirement logic to receive the instructions from the execution unit and to retire the instructions associated with one or more of the threads that have an instruction or an event pending to be retired. The instruction retirement unit includes a thread arbitration logic to select one of the threads at a time and to dispatch the selected thread to the retirement logic for retirement processing.

Scalable event handling in multi-threaded processor cores

An apparatus and method for efficiently managing the architectural state of a processor. For example, one embodiment of a processor comprises: a source mask register to be logically subdivided into at least a first portion to store a usable portion of a mask value and a second portion to store an indication of whether the usable portion of the mask value has been updated; a control register to store an unusable portion of the mask value; architectural state management logic to read the indication to determine whether the mask value has been updated prior to performing a store operation, wherein if the mask value has been updated, then the architectural state management logic is to read the usable portion of the mask value from the first portion of the source mask register and zero out bits of the unusable portion of the mask value to generate a final mask value to be saved to memory, and wherein if the mask value has not been updated, then the architectural state management logic is to concatenate the usable portion of the mask value with the unusable portion of the mask value read from the control register to generate a final mask value to be saved to memory.

Method and apparatus for efficiently managing architectural register state of a processor

Exemplary aspects are directed toward resolving fault suppression in hardware, which at the same time does not incur a performance hit. For example, when multiple instructions are executing simultaneously, a mask can specify which elements need not be executed. If the mask is disabled, those elements do not need to be executed. A determination is then made as to whether a fault happens in one of the elements that have been disabled. If there is a fault in one of the elements that has been disabled, a state machine re-fetches the instructions in a special mode. More specifically, the state machine determines if the fault is on a disabled element, and if the fault is on a disabled element, then the state machine specifies that the fault should be ignored. If during the first execution there was no mask, if there is an error present during execution, then the element is re-run with the mask to see if the error is a 'real' fault.

Memory fault suppression via re-execution and hardware FSM

A technique to generate a vector mask. In particular, at least one embodiment of the invention matches at least two instructions used in generating a vector mask and prevents at least one of the two instructions from executing if the correlation is found.

Technique for setting a vector mask

Growing community of inventors

Portland, OR, United States of America

Roger Gramunt