Computing method and device with data sharing

A processor implemented method, the method comprising:

• loading, by a loader, input data of an input feature map stored in a memory in loading units according to a loading order;
• storing, by a buffer controller, the loaded input data in a reuse buffer of an address rotationally allocated according to the loading order;
• transmitting, by each of a plurality of senders, to an executer, respective input data corresponding to each output data of respective convolution operations among the input data stored in the reuse buffer;
• and
• flushing an address allocated to remaining input data, other than input data overlapping input data corresponding to output data on which a convolution operation is to be performed, among input data loaded in a previous clock,
• wherein portions of the transmitted respective input data overlap each other.

A processor-implemented method, the method comprising:

• loading, by a loader input data of an input feature map stored in a memory in loading units according to a loading order;
• storing, by a buffer controller, the loaded input data in a reuse buffer of an address rotationally allocated according to the loading order;
• and
• transmitting, by each of a plurality of senders, to an executer, respective input data corresponding to each output data of respective convolution operations among the input data stored in the reuse buffer,
• wherein portions of the transmitted respective input data overlap each other, and
• wherein the plurality of senders comprises a first sender and a second sender,
• the first sender corresponds to first output data,
• the second sender corresponds to second output data adjacent to the first output data,
• first input data, second input data, third input data, fourth input data, fifth input data, and sixth input data are stored in a first address, a second address, a third address, a fourth address, a fifth address, and a sixth address of the reuse buffer, respectively, and
• the transmitting comprises:
• transmitting, by the first sender, the first input data, the second input data, the third input data, and the fourth input data to the executer to perform a first convolution operation;
• and
• transmitting, by the second sender, the third input data, the fourth input data, the fifth input data, and the sixth input data to the executer to perform a second convolution operation.

A processor-implemented method of a fetcher that includes a first reuse buffer and a second reuse buffer, the method comprising:

• loading, by a loader of a fetcher, input data of an input feature map stored in a memory in loading units according to a loading order;
• allocating, by a buffer controller of the fetcher, an address of the first reuse buffer or the second reuse buffer rotationally to the loaded input data according to the loading order, and storing the loaded input data in the allocated address;
• and
• transmitting, by each of a plurality of senders, to an executer respective input data corresponding to output data of respective convolution operations among input data stored in the first reuse buffer or the second reuse buffer,
• wherein portions of the transmitted respective input data overlap, and
• wherein the rotational allocation comprises, with respect to previous input data and current input data included in the loaded input data:
• allocating a first address of the first reuse buffer to the current input data in response to an address allocated to the previous input data being a last address of the first reuse buffer, and allocating a first address of the second reuse buffer to the current input data in response to an address allocated to the previous input data being a last address of the second reuse buffer;
• or
• allocating the first address of the second reuse buffer to the current input data in response to the address allocated to the previous input data being the last address of the first reuse buffer, and allocating the first address of the first reuse buffer to the current input data in response to the address allocated to the previous input data being the last address of the second reuse buffer.

A computing device, the device comprises:

• a loader;
• a plurality of senders;
• a buffer controller;
• and
• a reuse buffer,
• wherein the loader is configured to load input data of an input feature map stored in a memory in loading units according to a loading order,
• the buffer controller is configured to store the loaded input data in the reuse buffer of an address rotationally allocated according to the loading order,
• each of the plurality of senders is configured to transmit, to an executer, respective input data corresponding to each output data of respective convolution operations among the input data stored in the reuse buffer, and
• portions of the transmitted respective input data overlap, and
• the rotational allocation comprises, with respect to previous input data and current input data included in the loaded input data, an allocation of a first address of the reuse buffer to the current input data in response to an address allocated to the previous input data being a last address of the reuse buffe.

A computing device, the device comprising:

• a loader;
• a plurality of senders;
• a buffer controller;
• and
• a plurality of reuse buffers,
• wherein the plurality of reuse buffers comprises a first reuse buffer and a second reuse buffer,
• the loader is configured to load input data of an input feature map stored in a memory in loading units according to a loading order,
• the buffer controller is configured to allocate an address of the first reuse buffer or the second reuse buffer rotationally to the loaded input data according to the loading order, and
• store the loaded input data in the allocated address,
• each of the plurality of senders is configured to transmit, to an executer, respective input data corresponding to output data of respective convolution operations among input data stored in the first reuse buffer or the second reuse buffer,
• portions of the transmitted respective input data overlap, and
• wherein the rotational allocation comprises, with respect to previous input data and current input data included in the loaded input data:
• an allocation of a first address of the first reuse buffer to the current input data in response to an address allocated to the previous input data being a last address of the first reuse buffer, and an allocation of a first address of the second reuse buffer to the current input data in response to an address allocated to the previous input data being a last address of the second reuse buffer;
• or
• an allocation of the first address of the second reuse buffer to the current input data in response to the address allocated to the previous input data being the last address of the first reuse buffer, and an allocation of the first address of the first reuse buffer to the current input data in response to the address allocated to the previous input data being the last address of the second reuse buffer.

A computing device, the device comprising:

• a memory;
• a fetcher;
• a buffer controller;
• an executer;
• and
• a storage,
• wherein the fetcher comprises:
• a loader;
• a plurality of senders;
• and
• a reuse buffer,
• the loader is configured to load input data of an input feature map stored in the memory in loading units according to a loading order,
• the buffer controller is configured to allocate an address of the reuse buffer to the loaded input data rotationally according to the loading order and store the loaded input data in the allocated address,
• each of the plurality of senders is configured to transmit, to the executer, respective input data corresponding to each output data of respective convolution operations among the input data stored in the reuse buffer,
• portions of transmitted respective input data overlap, and
• the rotational allocation comprises, with respect to previous input data and current input data included in the loaded input data, an allocation of a first address of the reuse buffer to the current input data in response to an address allocated to the previous input data being a last address of the reuse buffer.

A processor-implemented method of a computing device, the method comprising:

• loading, in loading units according to a loading order, input data of an input feature map stored in a memory exterior to the computing device;
• storing the loaded input data in a reuse buffer of the computing device according to addresses rotationally allocated according to the loading order;
• and
• transmitting to an executer respective input data, corresponding to each output data of respective convolution operations, among the input data stored in the reuse buffer,
• wherein a portion of the transmitted respective input data is spatially and/or temporally shared with respect to a single storing of the portion in the reuse memory, and
• wherein the rotational allocation comprises, with respect to previous input data and current input data included in the loaded input data, allocating a first address of the reuse buffer to the current input data in response to an address allocated to the previous input data being a last address of the reuse buffer.