# US 2018/0293691 A1 · Machine learning sparse computation mechanism

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2018/0293691 A1
- Document ID: US20180293691A1
- Status: Published application
- Kind: A1
- Application number: 15/482,791
- Filed: 2017-04-09
- Published: 2018-10-11
- Earliest claimed: 2017-04-09
- Assignees: Intel Corporation
- Applicants: Intel Corporation
- Inventors: Eriko Nurvitadhi; Balaji Vembu; Tsung-Han Lin; Kamal Sinha; Rajkishore Barik; Nicolas C. Galoppo Von Borries
- Web page: https://patentgrep.com/patent/US20180293691A1

## CPC

- G06T 1/20: Processor architectures; Processor configuration, e.g. pipelining
- G06F 9/3001: {Arithmetic instructions}
- G06F 9/3885: {using a plurality of independent parallel functional units}
- G06F 9/4881: {Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues}
- G06F 12/0207: {with multidimensional access, e.g. row/column, matrix}
- G06F 12/0811: with multilevel cache hierarchies
- G06F 12/0815: Cache consistency protocols
- G06F 12/0831: using a bus scheme, e.g. with bus monitoring or watching means
- G06F 12/0888: using selective caching, e.g. bypass
- G06F 17/16: Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)}
- G06F 18/2136: based on sparsity criteria, e.g. with an overcomplete basis
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/048: Activation functions
- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- G06N 3/063: using electronic means
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/09: Supervised learning
- G06N 20/00: Machine learning
- G06T 1/60: Memory management
- G06T 15/005: {General purpose rendering architectures}
- H03M 7/30: Compression (speech analysis-synthesis for redundancy reduction G10L19/00; for image communication H04N); Expansion; Suppression of unnecessary data, e.g. redundancy reduction
- G06F 2212/1024: Latency reduction
- G06F 2212/302: In image processor or graphics adapter
- G06F 2212/401: Compressed data
- G06F 2212/621: Coherency control relating to peripheral accessing, e.g. from DMA or I/O device
- G06T 2200/28: involving image processing hardware
- Y02D 10/00: Energy efficient computing, e.g. low power processors, power management or thermal management
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- H03M: CODING; DECODING; CODE CONVERSION IN GENERAL (using fluidic means F15C4/00; optical analogue/digital converters G02F7/00; coding, decoding or code conversion, specially adapted for particular applications, see the relevant subclasses, e.g. G01D, G01R, G06F, G06T, G09G, G10L, G11B, G11C, H04B, H04L, H04M, H04N; ciphering or deciphering for cryptography or other purposes involving the need for secrecy G09C)
- Y02D: CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE

## Abstract

An apparatus to facilitate processing of a sparse matrix is disclosed. The apparatus includes a plurality of processing units each comprising one or more processing elements, including logic to read operands, a multiplication unit to multiply two or more operands and a scheduler to identify operands having a zero value and prevent scheduling of the operands having the zero value at the multiplication unit.

## Claims

### Claim 1

An apparatus to facilitate processing a sparse matrix, comprising:
a graphics processing unit, including:
a plurality of processing units each comprising one or more processing elements to perform neural network matrix multiplication operations, including:
pattern tracking logic to detect one or more segments of sparse data in a stored block of matrix data;
logic to read operands;
a multiplication unit to multiply two or more operands; and
a scheduler to identify operands having a zero value and prevent scheduling of the operands having the zero value at the multiplication unit.

### Claim 2 (depends on 1)

The apparatus of claim 1, wherein the scheduler schedules non-zero operands at the multiplication unit.

### Claim 3 (depends on 2)

The apparatus of claim 2, wherein the scheduler retrieves stored operand values upon the operands being received at the logic to read operands.

### Claim 4 (depends on 3)

The apparatus of claim 3, further comprising write result logic to write a result of the multiplication of the two or more operands at the multiplication unit, wherein the scheduler writes a zero value to the write result logic for operands having a zero value.

### Claim 5

(canceled)

### Claim 6 (depends on 1)

The apparatus of claim 1, wherein the pattern tracking logic further comprises pattern recognition logic to perform a bounding box operation on the block of data to determine a similarity of the data.

### Claim 7 (depends on 6)

The apparatus of claim 6, wherein the pattern recognition logic tracks data stored at a memory device.

### Claim 8 (depends on 6)

The apparatus of claim 6, wherein the pattern recognition logic tracks data stored at a cache memory device.

### Claim 9 (depends on 6)

The apparatus of claim 6, wherein the pattern recognition logic tracks data stored at a page table.

### Claim 10 (depends on 1)

The apparatus of claim 1, wherein the pattern tracking logic further comprises segment logic to record an address location for each detected segment of sparse data.

### Claim 11 (depends on 1)

The apparatus of claim 1, further comprising:
logic to compress a sparse matrix; and
a sparse compressed buffer to store the compressed sparse matrix.

### Claim 12 (depends on 11)

The apparatus of claim 11, wherein the compressed sparse matrix is dynamically generated based on a sparse index.

### Claim 13 (depends on 12)

The apparatus of claim 12, wherein the compressed sparse matrix comprises a sparse matrix frequently processed at the one or more processing units.

### Claim at position 14

14-16. (canceled)

### Claim 17

A graphics processing unit comprising:
a plurality of processing units each comprising one or more processing elements to perform neural network matrix multiplication operations, including:
pattern recognition logic to detect one or more segments of sparse data in a stored block of data;
compression logic to generate a compressed sparse matrix of sparse data upon a determination that the segment of sparse data meets a sparsity predetermined threshold; and
a storage device to store the compressed sparse matrix.

### Claim 18 (depends on 17)

The method of claim 17, wherein the compressed sparse matrix is dynamically generated based on a sparse index.

### Claim 19 (depends on 17)

The method of claim 17, wherein the compressed sparse matrix comprises a sparse matrix frequently processed at the one or more processing units.

### Claim 20 (depends on 17)

The method of claim 17, wherein the pattern recognition logic detects the one or more segments of sparse data in the stored block of data by performing a bounding box operation on the block of data to determine a similarity of the data.

### Claim 21 (depends on 19)

The method of claim 19, wherein the pattern recognition logic records an address location for each detected segment of sparse data.
