Vector-matrix multiplication

• US 20040243657A1
• Filed: 03/03/2004
• Published: 12/02/2004
• Est. Priority Date: 09/03/2001
• Status: Active Grant

• Current Assignee: Lenslet Ltd.
• Original Assignee: Lenslet Ltd.
• Inventors: Moses, Eyal, Levit, Shimon, Asaf, Yosefa, Sariel, Aviram, Liberman, Sergio, Machal, Vered, Goren, Avner, Hefetz, Yaron, Sender, Benzion, Tzelnick, Tzvi
• Granted Patent: US 7,536,431 B2
• Time in Patent Office: Days
• US Class Current: 708/607
• CPC Class Codes: G06E 1/045 Matrix or vector computation
• CPC Class Codes: G06T 1/0007 Image acquisition

Abstract

An integrated VMM (vector-matrix multiplier) module, including an electro-optical VMM component that multiplies an input vector by a matrix to produce an output vector; and an electronic VPU (vector processing unit) that processes at least one of the input and output vectors. Various error reducing mechanisms are also discussed.

Claims

1. An integrated VMM (vector-matrix multiplier) module, comprising:

an electro-optical VMM component that multiplies an input vector by a matrix to produce an output vector; and

an electronic VPU (vector processing unit) that processes at least one of the input and output vectors.

2. A module according to claim 1, comprising a DSP (digital signal processor) that processes at least one of the input vector, the output vector or the matrix.

3. A module according to claim 1, comprising a memory that stores at least one of matrix replacement values, at least one previous output vector and instructions for a component of the module.

4. A module according to claim 2, wherein at least one of said DSP and VPU are programmed to calculate an update value for at least part of said matrix.

5. A module according to claim 1, wherein said VMM component includes a local memory buffer for update values of said matrix.

6. A module according to claim 1, comprising a register file adapted for exchanging information between said VMM and said VPU.

7. A module according to claim 6, wherein said register file includes a register copy ability for transferring information between registers.

8. A module according to claim 1, comprising a parameter extractor which extracts at least one parameter from at least one of said vectors.

9. A module according to claim 8, wherein said parameter comprises an extreme value element.

10. A module according to claim 1, wherein said VMM module comprises a pre-processor which preprocesses said input vector, to improve a quality of its processing by a matrix component of said VMM.

11. A module according to claim 1, wherein said VMM module comprises a pre-processor which preprocesses said input vector, to correct for artifacts caused by processing by a matrix component of said VMM.

12. A module according to claim 1, comprising a vector buffer for buffered input from an external circuit.

13. A module according to claim 12, wherein said buffer receives 8 bit data in parallel.

69. A module according to claim 1, comprising a controller which controls the operations of the VMM component and the electronic VPU.

70. A module according to claim 69, wherein the VMM component and the electronic VPU have operations cycles and the controller controls the operations of the VMM component and the electronic VPU in each operation cycle based on commands it receives.

71. A module according to claim 70, comprising controller memory which stores the commands of the controller.

72. A module according to claim 70, wherein the commands received by the controller include a field for each of the VMM component and the electronic VPU.

73. A module according to claim 70, wherein the commands received by the controller include a field for program flow commands.

74. A module according to claim 70, wherein the commands indicate for the VPU one or more of a precision to be used by the VPU, a method of overflow treatment and a parameter of truncation.

75. A module according to claim 70, wherein the commands are generated by compiling high language directives.

14. An integrated VMM (vector-matrix multiplier) module, comprising:

a electro-optical VMM component that multiplies an input vector by a matrix to produce an output vector, and

a controller, wherein said controller is operative to replace values in only a part of said matrix.

15. A VMM (vector-matrix multiplier) component, comprising:

a plurality of input elements that represent an input vector;

a plurality of electro-optical matrix elements that represent a transformation matrix; and

a plurality of detector elements that detect signals from said input elements after they are modulated by said matrix elements such that said detected signals represent result vector of a vector matrix multiplication of said input vector,

wherein said component comprises at least one redundant element in at least one of said input, matrix and output elements.

16. A component according to claim 15, comprising a fan element for at least one of input fanning and output fanning, wherein said fan is programmable to selectively utilize said at least one redundant element.

17. A component according to claim 16, comprising a controller which manages said at least one redundant element.

18-32. -32. (Cancelled)

33. A method of improving signal detection in an electro-optical VMM, comprising:

receiving an input vector and a matrix to be processed by said VMM; and

rearranging said input vector on an input of said VMM and said matrix in a matrix portion of said VMM, in a manner that improves signal detection.

34. A method according to claim 33, wherein rearranging comprises spatially separating vector elements to reduce cross-talk.

35. A method according to claim 33, wherein rearranging comprises duplicating at least some vector elements.

36. A method according to claim 35, wherein rearranging comprises duplicating an entire vector.

37. A method according to claim 35, wherein rearranging comprises rearranging said matrix.

38. A method according to claim 35, wherein rearranging comprises rearranging said input vector so that at least some light sources of said VMM can be extinguished.

39. A method of improving signal detection in an electro-optical VMM, comprising:

receiving an input vector and a matrix to be processed by said VMM; and

adapting values of at least one of said input vector on an input of said VMM and said matrix in a matrix portion of said VMM, in a manner that improves signal detection.

40. A method according to claim 39, wherein adapting comprising negating values of at least some vector elements.

41. A method according to claim 39, wherein adapting comprises shifting a baseline value to be non-zero, such that light sources of the VMM are not extinguished to achieve this base line value.

42. A method according to claim 39, wherein adapting comprises amplifying or reducing input value to make use of an available dynamic range of said VMM.

43. A method according to claim 39, wherein adapting comprises shifting an input value base line to make use of an available dynamic range of said VMM.

44. A method according to claim 39, wherein adapting comprises applying a linearity correction.

45. A method according to claim 39, wherein adapting comprises weighting vector elements with weights that correspond to a number of zero values in a corresponding matrix column.

46. A method according to claim 39, wherein adapting comprises weighting at least one of vector elements and matrix elements with weights that correspond to an average of values in corresponding matrix columns.

47. A method of improving signal detection in an electro-optical VMM, comprising:

receiving an input vector and a matrix to be processed by said VMM;

processing said vector by said VMM to produce an output vector; and

adapting values of said output vector, by applying a history correction which corrects for residual a effects of a previous computation performed by said VMM.

48. A method according to claim 47, wherein said adapting comprises applying a temperature correction.

49. A method according to claim 47, wherein said adapting comprises applying a correction for an adaptation made to said input vector.

50-68. -68. (Cancelled)