US 11,106,968 B1Granted

Circuit arrangements and methods for traversing input feature maps

AssigneeXilinx, Inc.

InventorsEhsan Ghasemi, Elliott Delaye, Ashish Sirasao

Life of this patent

Estimating the term…

Loading patent…

Abstract

A circuit arrangement includes a buffer, a height traversal circuit configured to generate a sequence of IFM height values in response to first control signals, a width traversal circuit configured to generate a sequence of IFM width values in response to second control signals, a control circuit, and an address generation circuit. The control circuit is configured to input an OFM height, an OFM width, a kernel height, and a kernel width; generate the first control signals at times based on the OFM height and the kernel height; and generate the second control signals at times based on the OFM width and the kernel width. The address generation circuit is configured to generate a sequence of addresses based on the sequences of IFM height values and IFM width values, provide the sequence of addresses to the buffer, and enable reading from the buffer.

Claims20

Dependent claims

A circuit arrangement, comprising:

a buffer circuit configured for storage of elements i and i+1, for a plurality of sequential values of i, of an input feature map (IFM) at a plurality of corresponding sequential addresses of the buffer circuit, wherein elements i and i+1 are in row-major or column major order;
a height traversal circuit responsive to first control signals and including a first adder circuit configured to increment an output feature map (OFM) height counter value, a second adder circuit configured to increment a height offset value, and a third adder circuit configured to generate a sequence of height traversal values based on the OFM height counter value and the height offset value;
a width traversal circuit responsive to second control signals and including a fourth adder circuit configured to increment an OFM width counter value, a fifth adder circuit configured to increment a width offset value, and a sixth adder circuit configured to generate a sequence of width traversal values based on the OFM width counter value and the width offset value;
and
a control circuit configured to:
input an OFM height value, an OFM width value, a kernel height value, and a kernel width value;
generate the first control signals at times based on the OFM height value and the kernel height value;
generate the second control signals at times based on the OFM width value and the kernel width value;
and
an address generation circuit including a multiplier circuit and a seventh adder circuit and coupled to the buffer circuit, the height traversal circuit, and the width traversal circuit, the address generation circuit configured to:
generate a sequence of addresses based on the multiplier circuit multiplying the height traversal values by an IFM width value and the seventh adder circuit adding output of the multiplier circuit to the width traversal values;
wherein the sequence of addresses is out of sequential order;
provide the sequence of addresses to the buffer circuit;
and
enable reading from the buffer circuit.

12 more elements

A circuit arrangement, comprising:

a buffer circuit configured for storage of elements i and i+1, for a plurality of sequential values of i, an unrolled input feature map (IFM) in row major order as a one-dimensional array at a plurality of corresponding sequential addresses and addressed sequentially in the buffer circuit;
a first circuit configured to:
traverse by a first adder circuit, a first range of values associated with a height of an output feature map (OFM) in response to a first counter control signal;
shift a value of the first range by a stride;
traverse by a second adder circuit, a second range of values associated with a height of a kernel in response to a second counter control signal;
and
sum the shifted value of the first range and a value of the second range by a third adder circuit to generate a height traversal location;
a second circuit configured to:
traverse by a fourth adder circuit, a third range of values associated with a width of the OFM in response to a third counter control signal;
shift a value of the third range by the stride;
traverse by a fifth adder circuit, a fourth range of values associated with a width of the kernel in response to a fourth counter control signal;
and
sum the shifted value of the third range and a value of the fourth range by a sixth adder circuit to generate a width traversal location;
and
a third circuit configured to:
sum by a seventh adder circuit, a product generated by a multiplier circuit from the height traversal location and a width of an input feature map (IFM), and the width traversal location to generate an address of data of the IFM for a convolution operation on the data and the kernel;
and
output the address in response to:
the height traversal location being within a height of the IFM;
and
the width traversal location being within the width of the IFM;
and
wherein the third circuit generates a sequence of addresses that is out of sequential order.

20 more elements

A method, comprising:

inputting, to a control circuit, an output feature map (OFM) height value, an OFM width value, a kernel height value, and a kernel width value;
generating, by the control circuit, first control signals at times based on the OFM height value and the kernel height value;
generating, by the control circuit, second control signals at times based on the OFM width value and the kernel width value;
generating, by a height traversal circuit, a sequence of height traversal values in response to the first control signals, the generating the sequence of height traversal values including incrementing an OFM height counter value by a first adder circuit, incrementing a height offset value by a second adder circuit, and generating a sequence of height traversal values based on the OFM height counter value and the height offset value by a third adder circuit;
generating, by a width traversal circuit, a sequence of width traversal values in response to the second control signals, the generating the sequence of width traversal values including increment an OFM width counter value by a fourth adder circuit, incrementing a width offset value by a fifth adder circuit, and generating a sequence of width traversal values based on the OFM width counter value and the width offset value by a sixth adder circuit;
generating, by an address generation circuit, a sequence of addresses based on a multiplier circuit multiplying the sequence of height traversal values by an IFM width value and an adder circuit adding output of the multiplier circuit to the width traversal values;
providing, by the address generation circuit, the sequence of addresses to a buffer circuit configured for storage of elements i and i+1, for a plurality of sequential values of i, of an unrolled IFM as a one-dimensional array in row-major or column major order at a plurality of corresponding sequential addresses and addressed sequentially in the buffer circuit, wherein the sequence of addresses is out of sequential order;
and
enabling, by the address generation circuit, reading from the buffer circuit.

6 more elements

Dependencies are inferred from the claim text.