Parallel architecture for matrix transposition
An extension to current multiple memory bank video processing architecture is presented. A more powerful memory controller is incorporated, allowing computation of multiple memory addresses at both the input and the output data paths making possible new combinations of reads and writes at the input and output ports. Matrix transposition computations required by the algorithms used in image and video processing are implemented in MAC modules and memory banks. The technique described here can be applied to other parallel processors including future VLIW DSP processors.
1. A matrix transposition accelerator comprising:
a plurality of N input memory banks;
a plurality of N input address registers, each input address register corresponding to one of said input memory banks;
a plurality of N multiply and add units, each multiply and add unit corresponding to one of said input memory banks;
a plurality of N output memory banks, each output memory bank corresponding to one of said multiply and add units;
a plurality of N output address registers, each output address unit corresponding to one of said output memory banks'
an input memory controller connected to said plurality of input address registers; and
an output memory controller connected to said plurality of output address registers;
said matrix transformation accelerator operating said input memory controller in cooperation with said output memory controller whereby said input memory controller supplies addresses to corresponding input address registers for recalling plural matrix elements from plural separate input memory banks and said output memory controller supplies addresses to corresponding output address register writing plural matrix elements into corresponding locations plural separate output memory banks.
2. The matrix transposition accelerator of claim 1 wherein:
said input memory controller operates to supply said addresses to said input address registers for recalling plural matrix elements from sequential memory locations within plural input memory banks; and
said output memory controller operates to supply said addresses to said output address registers to write plural matrix elements to scattered memory locations in plural output memory banks.
3. The matrix transposition accelerator of claim 2 , wherein:
said input memory controller and said output memory controller operate cooperatively according to the code sequence:
for (i=0; i<IH; i++)
for (j=0; j<IW; j+=N)
for (p=0; p<N; p++)
if (j+p < iW)
then dout[(j+p)*OL + I] and din[i*IL + j + p]
where: IH is an input matrix height; IW is an input matrix width; i, j and p are index variables; and the greatest common denominator between and output matrix offset OL and N is 1.
4. The matrix transposition accelerator of claim 1 wherein:
said input memory controller operates to supply said addresses to said input address registers for recalling plural matrix elements from scattered memory locations within plural input memory banks; and
said output memory controller operates to supply said addresses to said output address registers to write plural matrix elements to sequential memory locations in plural output memory banks.
5. The matrix transposition accelerator of claim 4 , wherein:
said input memory controller and said output memory controller operate cooperatively according to the code sequence:
for (j=0; j<IW j++),
for (i=0 i<IH; i+=N),
for (p=0; p<N; p++),
if (i+p < IH)
then dout{j*OL + i + p] and din[(i + p)*IL + j]
where: IH is an input matrix height; IW is an input matrix width; j, i and p are index variables; and the greatest common denominator between an input matrix offset IL and N is 1.
6. The matrix transposition accelerator of claim 1 wherein:
said input memory controller operates to supply said addresses to said input address registers for recalling plural matrix elements from scattered memory locations within plural input memory banks; and
said output memory controller operates to supply said addresses to said output address registers to write plural matrix elements to scattered memory locations in plural output memory banks.
7. The matrix transposition accelerator of claim 6 , wherein:
said input memory controller and said output memory controller operate cooperatively according to the code sequence:
for (i=0; i<IH; i+=N),
for (j=0; j<IW j+=N),
for (k=0; k<N; k++)
for (p=0; p<N; p++)
if ((i+p) < IH) AND (j+((k+p)%N) < IW))
then dout{(j++((k+p)%N))*OL + i + p) and
din[(i+p)*IL + j + (k+p)%N],
where: IH is an input matrix height; IW is an input matrix width; i, j, k and p are index variables; (X) % N is X modulo N; and the greatest common denominator between the sum of an input offset IL and 1 (IL+1), and N is 1 and the greatest common denominator between the sum of an output offset OL and 1 (OL+1), and N is 1.