Scientific processor vector file organization

Abstract

Abstract An improved memory organization for a pipelined scien. tific vector processor is disclosed. It provides an efficient vectorfile organization with a data transfer capability which is adequate to support multiple program execution pipelines with the processor, It enables the multiple pipelines to simultaneously access various blocks of the overall vector file, The invention provides a novel manner of segmenting the file storage and of addressing the various elements of the segments. it also provides a unique method of granting and maintaining memory access to the vector files. The vector files themselves are programmable registers. There are sixteen (16) files of vector registers and each file has space for sixty-four 36-bit words. Each file may be formatted as single or double word elements. A preferred embodiment of the scientific processor using this vector file organization is shown. it has six independent instruction execution pipelines. They all access the vector files for either source (read) operands or destination (write) operands or both. Any combination of execution pipes can be active at any given time. They are initiated into execution in a serial manner depending on the instructions encountered in a program. The Vector Files are organized into eight independent blocks. Each block has a random access memory (RAM), an address register and a write data register. Each of the blocks can do a read or write operation each clock cycle. Each block holds certain pairs of words of each of the 16 files. The first block has words 0, 1, 16, 17, 32, 33, 48 and 49. The next block has the succeeding word pairs (each eighth pair). The pattern continues for the remaining 6 blocks. Only one copy of the vector files is usually shown, however, two are actually provided. These copies, while termed the primary and secondary vector files, are equal in capability, The purpose of the two copies is to allow reading pairs of operands, as required by most instructions. For example, during a vector multiply instruction (executed by multiply pipeline) the multiplicand is read from the primary copy and the multiplier is read from the secondary copy. The read registers are associated with a pipeline rather than with a vector file block. Writing of the corresponding multiply product element into both copies occurs nine clock cycles later. A time slot management mechanism included in the Vector File organization consists of 8 registers and performs the following functions. It allocates and reserves access to the Vector Files by the execution pipe for each individual instruction execution. It maintains the reservation and its purpose (read or write) until the instruction is completed. It forms or selects Vector File read or write address for both primary and secondary copies. The address consists of the appropriate 4-bit file number plus certain other fields. Once formed, an address is used to reference all 8 blocks, in what is termed a pass. It selectively initiates one of 5 counters used to select read data from the 8 RAMS. Finally, it selectively initiates one of 6 counters used to route write data to 1 of the 8 write data registers. The 8 time slot registers (0-7) are connected together in a serial loop. All the reservation data is advanced for register to register each clock cycle. As each reservation is in register 0 it is used to form or select a Vector File address. Typically a reservation is maintained for 4 passes or 32 clock cycles and is released by being cleared upon completion of the instruction.

Inventors

• 👤 Archie E. Lahti

Bibliographic Details