Systolic array image processing system and method

1. A system for performing a convolution of pixel data with kernel data to process an image represented by the pixel data, the system comprising:

a systolic array of processing elements, wherein each of said processing elements comprises:

a pixel data port for receiving an element of the pixel data,

a kernel data port for receiving an element of the kernel data,

a multiplier coupled to the pixel data port and the kernel data port for multiplying the pixel data element by the kernel data element received by said ports to form a product of said elements, and

an accumulator coupled to the multiplier for adding the product to the contents of the accumulator,

said systolic array being arranged in a plurality of rows and columns, each row having an equal number of processing elements and each column having an equal number of processing elements,

wherein the processing elements pixel data ports in each row are coupled in series such that a pixel data element entered into the input end processing element of the row is transmitted to the other processing element pixel data ports in the row, and

wherein the processing element kernel data ports in each column are coupled in series such that a kernel data element entered into the input end processing element of the columns is transmitted to the other processing element kernel data ports in the column;

outputs bus means comprising a plurality of output channels, each of the output channels being coupled to the accumulators of the processing elements in one of the rows and receiving the contents of said accumulators, said plurality being equal in number to the number of rows;

kernel input means comprising:

a plurality of kernel channels, each of the kernel channels being coupled to the kernel data port in the input end processing element of one column and supplying kernel data elements to said input end processing element, said plurality being equal in number to the number of columns, and

kernel storage means coupled to the kernel channels for storing the kernel data;

multiplexor means coupled to the pixel data ports in the input end processing elements in each row, and supplying pixel data elements to said input end processing elements;

a first pixel input and a second pixel input coupled to said multiplexor means for supplying pixel data elements to said multiplexor means, said first and second pixel inputs each providing independent data streams of the entire pixel data, such that said multiplexor means selects each pixel data element for each input end processing element from one or the other of said data streams; and

clock means coupled to the processing elements of the systolic array, the output bus means, the kernel input means, the multiplexor means, and the first and second pixel inputs for providing timing signals that synchronize the operation of the system by a sequence of clock cycles,

wherein said pixel data ports in the input end processing elements of the columns are all coupled to the first pixel input, said input end processing elements being the first row of the systolic array,

wherein the multiplexor means comprises a plurality of multiplexors, each of the multiplexors generating one output pixel data element selected from one of two input pixel data elements, the output of each multiplex or being coupled to the pixel data port in the input end processing element of one row of the systolic array other than the first row, one input of the multiplexor being coupled to receive pixel data from another row of the systolic array, and the other input of the multiplexor being coupled to receive pixel data from the second pixel input, and

wherein the number of said kernel data elements is the square of some integer L, and wherein each of said multiplexors operates in the sequence of said clock cycles to apply data input to the row corresponding to said multiplexor from said second pixel input each one clock cycle out of L clock cycles, and from the multiplexor output from another row each L-1 clock cycles out of L clock cycles.

2. A system for performing a convolution of pixel data with kernel data to process an image represented by the pixel data, the system comprising:

a systolic array of processing elements, wherein each of said processing elements comprises:

a pixel data port for receiving an element of the pixel data,

a kernel data port for receiving an element of the kernel data,

a multiplier coupled to the pixel data port and the kernel data port for multiplying the pixel data element by the kernel data element received by said ports to form a product of said elements, and

an accumulator coupled to the multiplier for adding the product to the contents of the accumulator,

said systolic array being arranged in a plurality of rows and columns, each row having an equal number of processing elements and each column having an equal number of processing elements,

wherein the processing elements pixel data ports in each row are coupled in series such that a pixel data element entered into the input end processing element of the row is transmitted to the other processing element pixel data ports in the row, and

wherein the processing element kernel data ports in each column are coupled in series such that a kernel data element entered into the input end processing element of the column is transmitted to the other processing element kernel data ports in the columns;

output bus means comprising a plurality of output channels, each of the output channels being coupled to the accumulators of the processing elements in one of the rows and receiving the contents of said accumulators, said plurality being equal in number to the number of rows;

kernel input means comprising:

a plurality of kernel channels, each of the kernel channels being coupled to the kernel data port in the input end processing element of one column and supplying kernel data elements to said input end processing element, said plurality being equal in number to the number of columns, and

kernel storage means coupled to the kernel channels for storing the kernel data;

multiplexor means coupled to the pixel data ports in the input end processing elements in each row, and supplying pixel data elements to said input end processing elements;

a first pixel input and a second pixel input coupled to said multiplexor means for supplying pixel data elements to said multiplexor means, said first and second pixel inputs each providing independent data streams of the entire pixel data, such that said multiplexor means selects each pixel data element for each input end processing elements from one or the other of said data streams; and

clock means coupled to the processing elements of the systolic array, the output bus means, the kernel input means, the multiplexor means, and the first and second pixel inputs for providing timing signals that synchronize the operation of the system by a sequence of clock cycles,

wherein said pixel data ports in the input end processing elements of the columns are all coupled to the first pixel input, said input end processing elements being the first row of the systolic array,

wherein the multiplexor means comprises a plurality of multiplexors, each of the multiplexors generating one output pixel data element selected from one of two input pixel data elements, the output of each multiplexor being coupled to the pixel data port in the input end processing element of one row of the systolic array other than the first row, one input of the multiplexor being coupled to receive pixel data from another row of the systolic array, and the other input of the multiplexor being coupled to receive pixel data from the second pixel input,

wherein the number of said kernel data elements is the square of some integer L, and wherein each of said multiplexors operates in the sequence of said clock cycles to apply data input to the row corresponding to said multiplexor from said second pixel input each one clock cycle out of L clock cycles, and from the multiplexor output from another row each L-1 clock cycles out of L clock cycles, and

wherein said multiplexors, said kernel input means, said first pixel input, and said second pixel input are all programmable to enable the system to compute the convolution for any value of L.

3. A method for performing a two-dimensional image processing convolution of an M.times.M pixel data array with an L.times.L kernel data array in an N.times.N systolic array of processing elements, wherein in each clock cycle of the systolic array each processing element in row i and column j of the systolic array, PE(i,j), receives a pixel data element and a kernel data element, forms the product of the elements and adds the product to the contents of an accumulator, comprising the steps of:

scanning the L.times.L kernel data array for kernel data values and transmitting these values to the processing elements such that the value received by PE(i,j) is also received by PE(i+1,j) delayed by one clock cycle, and also received by PE(i,j+1) delayed by L+1 clock cycles;

scanning the M.times.M pixel data array for pixel data values and transmitting these values to the processing elements such that the value received by PE(i,j) is also received by PE(i,j+1) delayed by one clock cycle;

scanning the M.times.M pixel data array to form a first pixel data stream A and a second pixel data stream B, wherein each stream contains an entire scan of the pixel data array;

transmitting data stream A to the first row of processing elements PE(i,j);

transmitting pixel data from pixel data streams A and B to the remaining rows of processing elements such that during each sequence of L clock cycles the pixel data values received by PE(i+1,j) are the same as the values received by PE(i,j) for L-1 of said cycles, and for the remaining 1 cycle the pixel data value received by PE(i+1,j) is obtained from said second data stream B; and

after each scan of the pixel data array and kernel data array, repeating each such scan until the entire convolution has been computed.

4. The method according to claim 3, wherein said first and second data streams are such formed by scanning successively each block of L rows and M columns of said M.times.M pixel data array, and wherein in each of said blocks the streams are formed by scanning successively each column of L rows of said pixel data array values.

5. The method according to claim 4, wherein during the period that said first data stream is formed by scanning a block of L rows and M columns of pixel data, said second data stream B is formed by scanning another of said blocks of L rows and M columns, delayed by L clock cycles relative to the stream A scan.

6. The method according to claim 3, wherein the L.times.L kernel data array is scanned by scanning successively each column of kernel data values.

7. The method according to claim 3, wherein at the outset of each clock cycle, the kernel data value in each processing element PE(i,j) having i less than N is transmitted to the processing element PE(i+1,j).

8. The method according to claim 3, wherein at the outset of each clock cycle, the pixel data value in each processing element PE(i,j) having j less than N is transmitted to the processing element PE(i,j+1).

9. The method according to claim 3, further comprising the step of, at the beginning of each successive scan of the L.times.L kernel data array by each processing element in row i, PE(i,j):

transferring the contents of the accumulator of PE(i,j) to an output bus; and

resetting said accumulator to zero.

10. A method for performing a two-dimensional image processing convolution of an M.times.M pixel data array with an L.times.L kernel data array in an N.times.N systolic array of processing elements, wherein in each clock cycle of the systolic array each processing element in row i and column j of the systolic array, PE(i,j), receives a pixel data element and a kernel data element, forms the product of the elements and adds the product to the contents of an accumulator, comprising the steps of:

scanning the L.times.L kernel data array for kernel data values and transmitting these values to the processing elements such that the value received by PE(i,j) is also received by PE(i+1,j) delayed by one clock cycle, and also received by PE(i,j+1) delayed by L+1 clock cycles;

scanning the M.times.M pixel data array for pixel data values and transmitting these values to the processing elements such that the value received by PE(i,j) is also received by PE(i,j+1) simultaneously;

scanning the M.times.M pixel data array to form a first pixel data stream A and a second pixel data stream B, wherein each stream contains an entire scan of the pixel data array;

transmitting data stream A to the first row of processing elements PE(i,j);

transmitting pixel data from pixel data streams A and B to the remaining rows of processing elements such that during each sequence of L clock cycles the pixel data values received by PE(i+1, j) are the same as the values received by PE(i,j) for L-1 of said cycles, and for the remaining 1 cycle the pixel data value received by PE(i+1,j) is obtained from said second data stream B; and

after each scan of the pixel data array and kernel data array, repeating each such scan until the entire convolution has been computed.

11. The method according to claim 10, wherein said first and second data systems are each formed by scanning successively each block of L columns and M rows of said M.times.M pixel data array, and wherein in each of said blocks the streams are formed by scanning successively each row of L columns of said pixel data array values.

12. The method according to claim 11, wherein during the period that said first data stream is formed by scanning a block of L columns and M rows of pixel data, said second data stream B is formed by scanning another of said blocks of L columns and M rows, delayed by L clock cycles relative to the stream A scan.

13. The method according to claim 10, wherein the L.times.L kernel data array is scanned by scanning successively each row of kernel data values.

14. The method according to claim 10, wherein at the outset of each clock cycle, the kernel data value in each processing element PE(i,j) having i less than N is transmitted to the processing element PE(i+1,j).

15. The method according to claim 10, further comprising the step of, at the beginning of each successive scan of the L.times.L kernel data array by each processing element in row i, PE(i,j):

transferring the contents of the accumulator of PE(i,j) to an output bus; and

resetting said accumulator to zero.