Flow control for reconfigurable processors
The technology disclosed relates to storing a dataflow graph with a plurality of compute nodes that transmit data along data connections, and controlling data transmission between compute nodes in the plurality of compute nodes along the data connections by using control connections to control writing of data.
1. A computer-implemented method, including:
executing a dataflow graph on a processing system having a plurality of compute nodes, each having a ready-to-read credit counter and a write credit counter, that transmit data along data connections; and
controlling data transmission between compute nodes in the plurality of compute nodes along the data connections by using dedicated control connections between the compute nodes to selectively control writing of data based on both the ready-to-read credit counter and the write credit counter in a particular compute node of the plurality of compute nodes that provides the data for a particular data transmission, wherein the control connections are distinct from the data connections and are configured to transmit control signals that manage flow of the data by selectively incrementing the ready-to-read credit counter and the write credit counter in the particular compute node.
2. The computer-implemented method of claim 1 , further including initializing the ready-to-read credit counter of the particular compute node with as many read credits as a buffer depth of a corresponding compute node of the plurality of compute nodes that reads data from the particular compute node.
3. The computer-implemented method of claim 2 , further including decrementing the ready-to-read credit counter in the particular compute node when the particular compute node begins writing a buffer data unit into the corresponding compute node along a data connection of the data connections.
4. The computer-implemented method of claim 3 , further including incrementing the ready-to-read credit counter in the particular compute node when the particular compute node receives from the corresponding compute node a read ready token along a control connection of the control connections, wherein the read ready token indicates to the particular compute node that the corresponding compute node has freed a buffer data unit and is ready to receive an additional buffer data unit.
5. The computer-implemented method of claim 4 , wherein the particular compute node stops writing data into the corresponding compute node when the ready-to-read credit counter in the particular compute node has zero read credits.
6. The computer-implemented method of claim 5 , wherein the particular compute node resumes writing data into the corresponding compute node when the particular compute node receives the read ready token from the corresponding compute node.
7. The computer-implemented method of claim 1 , further including initializing the write credit counter in the particular compute node with one or more write credits.
8. The computer-implemented method of claim 7 , further including decrementing the write credit counter in the particular compute node when the particular compute node begins writing a buffer data unit of the data into a corresponding compute node of the plurality of compute nodes along a data connection of the data connections.
9. The computer-implemented method of claim 8 , further including incrementing the write credit counter in the particular compute node when the particular compute node receives from the corresponding compute node a write done token along a control connection of the control connections, wherein the write done token indicates to the particular compute node that the writing of the buffer data unit into the corresponding compute node has completed.
10. The computer-implemented method of claim 9 , wherein the particular compute node stops writing data into the corresponding compute node when the write credit counter in the particular compute node has zero write credits.
11. The computer-implemented method of claim 10 , wherein the particular compute node resumes writing data into the corresponding compute node when the particular compute node receives the write done token from the corresponding compute node.
12. A computer-implemented method comprising:
executing a dataflow graph on a processing system having a plurality of compute nodes coupled by data connections and control connections, including a first node coupled to a second node by a first data connection and a first control connection, the first node including a write credit counter and a ready-to-read credit counter;
sending data from the first node to the second node over the first data connection while both the write credit counter and the ready-to-read credit counter are greater than zero while pausing transmission of the data from the first node to the second node over the first data connection while either the write credit counter or the ready-to-read credit counter are equal to zero;
decrementing both the write credit counter and the ready-to-read credit counter upon sending a buffer unit of the data from the first node to the second node over the first data connection;
incrementing the write credit counter upon receipt of a write done token from the second node over the first control connection; and
incrementing the ready-to-read credit counter upon receipt of a read ready token from the second node over the first control connection.
13. The computer-implemented method of claim 12 , further comprising:
initializing the write credit counter to a first predetermined value; and
initializing the ready-to-read credit counter to a second predetermined value that is based on a size of a buffer in the second node.
14. The computer-implemented method of claim 12 , further including:
sending the write done token from the second node to the first node over the first control connection in response to storing the buffer unit of the data received from the first node over the first data connection into a buffer of the second node.
15. The computer-implemented method of claim 12 , further including:
sending the read ready token from the second node to the first node over the first control connection in response to removing the buffer unit of the data from a buffer of the second node.
16. A processing system comprising:
a plurality of compute nodes coupled by data connections and control connections, including a first node coupled to a second node by a first data connection and a first control connection, wherein the first node is configured to send data to the second node over the first data connection;
a write credit counter in the first node that is incremented in response to receipt of a write done token from the second node over the first control connection and decremented in response to receipt of a write done token from the second node over the first control connection;
a ready-to-read credit counter in the first node that is incremented in response to receipt of a read ready token from the second node over the first control connection and decremented in response to receipt of a write done token from the second node over the first control connection; and
dataflow control circuitry in the first node to pause transmission of the data from the first node to the second node over the first data connection while either the write credit counter or the ready-to-read credit counter are equal to zero.
17. The processing system of claim 16 , further comprising a buffer in the second node, wherein the ready-to-read credit counter is initialized to a predetermined value that is based on a size of the buffer in the second node.
18. The processing system of claim 17 , further comprising buffer management circuitry configured to:
send the write done token from the second node to the first node over the first control connection in response to storing a buffer unit of the data received from the first node over the first data connection into the buffer of the second node; and
send the read ready token from the second node to the first node over the first control connection in response to removing the buffer unit of the data from the buffer of the second node.