US 4,881,168 AGranted

Vector processor with vector data compression/expansion capability

AssigneeHitachi, Ltd.

InventorsYasuhiro Inagami, Takayuki Nakagawa, Yoshiko Tamaki, Shigeo Nagashima

Life of this patent

Estimating the term…

Loading patent…

Abstract

A vector processor has a memory for storing vector data, a plurality of vector registers each capable of reading or writing plural (m) vector elements in parallel, at least one mask vector register capable of storing m mask bits in parallel, and a transfer section connected to the memory, the plurality of vector registers and the mask vector register and responsive to a store compression instruction or a load expansion instruction for transferring vector elements to or from regularly spaced address locations within the memory from or to selected storage locations of a selected vector register corresponding to valid mask bits. The transfer section includes at least one count unit connected to the mask vector register for counting a total number of valid mask bits within all of the already read out mask bits, and plural (m) access units operable concurrently and connected to the count unit and the mask vector register. Each access unit is responsive to the validity of a corresponding one within currently read out m mask bits, to a total number of valid mask bits or invalid mask bits included within the currently read m mask bits and having preceding sequential numbers of elements to that of the corresponding mask bit, and to the counted total number, and operates to generate an address of a location within the memory which holds a vector element to be transferred to a storage location corresponding to the corresponding mask bit within the selected vector register or which should receive a vector element read out from the storage location.

Claims6

Dependent claims

A vector processor, comprising:

(a) a memory for holding data;
(b) at least one vector register including means for holding a set of vector elements and means for reading out said vector elements sequentially in groups in response to a store compression instruction, wherein each group comprises m (m>1) vector elements which are read out in parallel;
(c) at least one vector mask register including means for holding a set of mask bits each corresponding to one vector element of the set of vector elements, each mask bit comprising a valid mask bit or an invalid mask bit, and means for reading out the set of mask bits sequentially in groups in response to said store compression instruction, wherein each group comprises m mask bits corresponding to a respective group of m vector elements, wherein a respective group of m mask bits are read out in parallel and in synchronism with read out of a corresponding group of vector elements;
(d) store means connected to said memory, said vector register and said vector mask register and responsive to said store compression instruction for selectively writing vector elements, within a set of vector elements currently read out of said vector register and corresponding to valid mask bits within the corresponding set of mask bits, into regularly spaced storage locations of said memory;
(d1) wherein said store means comprises first to m-th parallelly operable access means, wherein an i-th (1.ltoreq.i.ltoreq.m) access means includes means for writing into said memory an i-th vector element of a respective group of vector elements currently read out from said vector register, in response to an i-th mask bit of a group of corresponding mask bits currently read out from the mask register when said i-th mask bit is a valid mask bit;
(d2) wherein said first access means includes means for generating an address of a storage location of said memory into which a first vector element of the read out group of vector elements is to be written when a first mask bit of the read out mask bits is a valid mask bit included in all groups of m mask bits read out before the currently read out group of m mask bits;
(d3) wherein each j-th (2.ltoreq.j.ltoreq.m) access means includes means for generating an address of a storage location of said memory into which a j-th vector element of the currently read out group of vector elements is to be written when a j-th mask bit of the currently read out group of m mask bits is a valid mask bit, depending upon said total number of valid mask bits in all previously read out groups of mask bits and upon a total number of valid or invalid mask bits included in the first to (j-1)th mask bits included in the currently read out m mask bits.

4 more elements

A vector processor, comprising:

(a) a memory for holding a set of vector elements;
(b) at least one vector mask register including means for holding a set of mask bits and means for sequentially reading out said mask bits in first groups in response to a load expansion instruction, wherein each first group comprises m (m>1) mask bits which are read out in parallel, wherein each mask bit comprises either a valid mask bit or an invalid mask bit;
(c) at least one vector register comprised of a set of storage locations divided into a plurality of second groups each having m storage locations, wherein each respective second group comprises storage locations corresponding to the respective m mask bits belonging to one of the first groups, wherein the different second groups of m storage locations are accessible sequentially for write-in of vector elements and respective storage locations of each second group are accessible in parallel for write-in of vector elements;
(d) load means connected to said memory, said vector register and said vector mask register and responsive to said load expansion instruction for transferring vector elements corresponding to valid mask bits from said memory to storage locations within said vector register;
(d1) wherein said load means includes first to m-th parallelly operable access means, wherein an i-th (1.ltoreq.i.ltoreq.m) access means includes means, responsive to an i-th mask bit of a group of m mask bits currently read out from the vector mask register when said i-th mask bit is a valid mask bit, for reading from said memory one of the set of vector elements to be written into an i-th storage location of a group of m storage locations corresponding to the currently read out group of m mask bits;
(d2) wherein said first access means includes means for generating an address of a storage location of said memory from which a vector element to be written into a first storage location of the corresponding group of m storage locations is to be read when a first mask bit of the currently read out group of mask bits is a valid mask bit, said address being determined depending upon a total number of valid mask bits included in all groups of mask bits read out before the currently read out group of m mask bits;
(d3) wherein each j-th (2.ltoreq.j.ltoreq.m) access means includes means for generating an address of a storage location of said memory from which a vector element to be written into a j-th storage location of the corresponding group of m storage locations is to be read when a j-th mask bit of the read out group of m mask bits is a valid mask bit, said address being determined depending upon said total number of said valid mask bits in all previous groups of read out mask bits and upon a total number of valid or invalid mask bits included in the first to (j-1)-th mask bits of the currently read out group of m mask bits.

4 more elements

Dependencies are inferred from the claim text.