"Method and apparatus for obtaining a scalar value directly from a vector register
Abstract
A method and apparatus for obtaining a scalar value from a vector register for use in a mixed vector and scalar instruction, including providing a vector in a vector register file, and embedding a location identifier of the scalar value within the vector in the bits defining the mixed vector and scalar instruction. The scalar value can be used directly from the vector register without the need to load the scalar to a scalar register prior to executing the instruction. The scalar location identifier may be embedded in the secondary op code of the instruction, or the instruction may have dedicated bits for providing the location of the scalar within the vector.
FIELD OF THE INVENTION
This invention relates to information processors, such as microprocessors, and, more particularly, to a method and apparatus which improves the operational efficiency of information processors having a vector processing unit by enabling a scalar value to be directly selected from a vector register for use, for example, in a mixed vector and scalar operation.
BACKGROUND OF THE INVENTION
The electronic industry is in a state of evolution spurred by the seemingly unquenchable desire of the consumer for better, faster, smaller, cheaper and more functional electronic devices. In their attempt to satisfy these demands, the electronic industry must constantly strive to increase the speed at which functions are performed by data processors. Videogame consoles are one primary example of an electronic device that constantly demands greater speed and reduced cost. These consoles must be high in performance and low in cost to satisfy the ever increasing demands associated therewith. The instant invention is directed to increasing the speed at which a vector processing units of information processors can perform mathematical operations when a scalar is needed from a vector register to perform the operation.
Microprocessors typically have a number of execution units for performing mathematical operations. One example of an execution unit commonly found on microprocessors is a fixed point unit (FXU), also known as an integer unit, designed to execute integer (whole number) data manipulation instructions using general purpose registers (GPRs) which provide the source operands and the destination results for the instructions. Integer load instructions move data from memory to GPRs and store instructions move data from GPRs to memory. An exemplary GPR file may have 32 registers, wherein each register has 32 bits. These registers are used to hold and store integer data needed by the integer unit to execute integer instructions, such as an integer add instruction, which, for example, adds an integer in a first GPR to an integer in a second GPR and then places the result thereof back into the first GPR or into another GPR in the general purpose register file.
Another type of execution unit found on most microprocessors is a floating point unit (FPU), which is used to execute floating point instructions involving non-integers or floating point numbers. Floating point numbers are represented in the form of a mantissa and an exponent, such as 6.02.times.10.sup.3 A floating point register file containing floating point registers (FPRs) is used in a similar manner as the GPRs are used in connection with the fixed point execution unit, as explained above. In other words, the FPRs provide source operands and destination results for floating point instructions. Floating point load instructions move data from memory to FPRs and store instructions move data from FPRs to memory. An exemplary FPR file may have 32 registers, wherein each register has 64 bits. These registers are used to hold and store floating point data needed by the floating point execution unit (FPU) to execute floating point instructions, such as a floating point add instruction, which, for example, adds a floating point number in a first FPR to a floating point number in a second FPR and then places the result thereof back into the first FPR or into another FPR in the floating point register file.
Microprocessor having floating point execution units typically enable data movement and arithmetic operations on two floating point formats: double precision and single precision. In the example of the floating point register file described above having 64 bits per register, a double precision floating point number is represented using all 64 bits of the FPR, while a single precision number only uses 32 of the 64 available bits in each FPR. Generally, microprocessors having single precision capabilities have single precision instructions that use a double precision format.
For applications that perform low precision vector and matrix arithmetic, a third floating point format is sometimes provided which is known as paired singles. The paired singles capability can improve performance of an application by enabling two single precision floating point values to be moved and processed in parallel, thereby substantially doubling the speed of certain operations performed on single precision values. The term "paired singles" means that the floating point register is logically divided in half so that each register contains two single precision values. In the example 64-bit FPR described above, a pair of single precision floating point numbers comprising 32 bits each can be stored in each 64 bit FPR. Special instructions are then provided in the instruction set of the microprocessor to enable paired single operations which process each 32-bit portion of the 64 bit register in parallel. The paired singles format basically converts the floating point register file to a vector register file, wherein each vector has a dimension of two. As a result, part of the floating point execution unit becomes a vector processing unit (paired singles unit) in order to execute the paired singles instructions.
Some information processors, from microprocessors to supercomputers, have vector processing units specifically designed to process vectors. Vectors are basically an array or set of values. In contrast, a scalar includes only one value, such as a single number (integer or non-integer). A vector may have any number of elements ranging from 2 to 256 or more. Supercomputers typically provide large dimension vector processing capabilities. On the other hand, the paired singles unit on the microprocessor described above involves vectors with a dimension of only 2. In either case, in order to store vectors for use by the vector processing unit, vector registers are provided which are similar to those of the GPR and FPR register files as described above, except that the register size corresponds to the dimension of the vector on which the vector processing unit operates. For example, if the vector includes 64 values (such as integers or floating point numbers) each of which require 32 bits, then each vector register will have 2048 bits which are logically divided into 64 32-bit sections. Thus, in this example, each vector register is capable of storing a vector having a dimension of 64. FIG. 2 shows: an exemplary vector register file 2 storing four 64 dimension vectors A, B, C and D.
A primary advantage of a vector processing unit with vector register as compared to a scalar processing unit with scalar registers is demonstrated with the following example: Assume vectors A and B are defined to have a dimension of 64, i.e. A=(A.sub.0 . . . A.sub.63) and B=(B.sub.0 . . . B.sub.63). In order to perform a common mathematical operation such as an add operation using the values in vectors A and B, a scalar processor would have to execute 64 scalar addition instructions so that the resulting vector would be R=((A.sub.1 +B.sub.1). (A.sub.63 +B.sub.63)). Similarly, in order to perform a common operation known as Dot_Product, wherein each corresponding value in vectors A and B are multiplied together and then each element in the resulting vector are added together to provide a resultant scalar, 128 scalar instructions would have to be performed (64 multiplication and 64 addition). In contrast, in vector processing a single vector addition instruction and a single vector Dot_Product instruction can achieve the same result. Moreover, each of the corresponding elements in the vectors can be processed in parallel when executing the instruction. Thus, vector processing is very advantageous in many information processing applications.
One problem, however, that is encountered in vector processing, is that sometimes it is desired to perform an operation using a scalar value contained within a vector register. For example, some applications may require mixed vector and scalar calculations, wherein the scalar needed (e.g. C.sub.10) to perform the calculation is a single element within a particular vector (e.g. C) stored in a vector register. In other words, while a vector processing unit may easily execute a vector instruction which adds vector A to B and places the result in vector C (i.e. C=A+B), the vector processing unit cannot directly perform a mixed vector and scalar operation when the desired scalar is an element in a vector register (i.e. D=C.sub.10 +A). The primary reason for this limitation is that mixed scalar and vector instructions require that the scalar used in the operation be stored is a scalar register. In other words, such instructions do not have the ability to select a particular scalar element, such as C.sub.10, from a vector register. FIG. 1 shows an exemplary format of prior art instructions for mixed scalar and vector instructions.
As can be seen in FIG. 1, the typical format for a mixed scalar and vector instruction 3 includes a primary op-code 4, a scalar register address 5, a vector register address 6 and a destination register address 7. The primary op-code identifies the particular type of instruction, such as vector-scalar multiplication, and may, for example, comprise the most significant 6 bits (bits 0-5) of the instruction. The scalar register address 5 provides the particular address of the register in the GPR file that contains the scalar value needed to execute the instruction. The vector register address 6 provides the particular address of the vector register in the vector register file which contains the vector needed to execute the instruction. The destination register address 7 provides the location for the result of the operation. It is noted that the instruction format 3 of FIG. 1 is only exemplary and that prior art instructions may have other formats and/or include other parts, such as a secondary op-code, status bits, etc., as one skilled in the art will readily understand. However, as explained above, regardless of the particular format of the instruction, the instruction still requires that a scalar register be used to store the scalar value needed to execute the instruction.
As a result, if the required scalar is a particular element of a vector register (e.g. C.sub.10), the entire vector register must first be copied to memory in order to enable the desired scalar (C.sub.10) to be loaded into a scalar register. In other words, the prior art provides no suitable mechanism for enabling a scalar to be used from a vector register. Thus, while such mixed scalar and vector instructions can be performed, they require significant overhead in terms of time required to store the vector to memory and load the scalar from memory to a scalar register, so that the scalar register contains the required scalar value to execute the instruction. Even assuming that the required vector is in a cache (high speed on-chip memory), thereby eliminating the need to access external memory, significant overhead still exists. For example, a typical cache may require approximately 30-50 CPU clock cycles (a time unit by which the central processing unit (CPU) operates) to load data from a 64-bit 128 dimension vector. Moreover, if cache is not available or if a cache miss occurs, the overhead would be approximately an order of magnitude higher to load or access the vector in an external memory as compared to a cache. Thus, large CPU cycle overhead is required to execute an instruction that, without the above limitations, could execute in for example, as fast as 10 clock cycles, i.e. 40 to 100s of clock cycle overhead for a 10 cycle instruction.
Accordingly, a need exists for reducing the large overhead associated with such mixed scalar and vector instructions, so that the operations associated therewith can be performed faster and so that application performance can be improved.
SUMMARY OF THE INVENTION
The instant invention provides a mechanism and a method for enabling mixed scalar and vector instructions to run more efficiently and with less CPU cycle overhead by eliminating the need to load a value from a vector register into a scalar register in order to be used during execution of the instruction. The invention provides an improved instruction format which may be used in connection with any suitable type of data processor, from microprocessors to supercomputers, having a vector processing unit in order to improve the operational efficiency thereof.
In accordance with the invention, the improved instruction format has an embedded bit or a plurality of embedded bits that identify a particular element in a vector to be used as a scalar during execution of the instruction. In this way, a mixed scalar and vector instruction can be executed without the need to load the scalar operand into a scalar or general purpose register. By identifying, in the instruction, the location of the scalar in the vector, the scalar can be directly used from the vector register file for execution of the instruction.
In accordance with a preferred embodiment of the invention, the instruction format for mixed scalar and vector operations includes a primary op code, a first source vector register address, a second source vector register address, a destination register vector address, and at least one position bit which indicates the location of a desired scalar in one of the vector registers needed to execute the instruction. The number of bits needed to indicate the position of the desired scalar within a vector depends on the particular dimension of the vector involved. For example, if the vector has a dimension of 64, then six bits are needed to provide a unique identifier for the particular scalar within the vector. In other words, if the dimension of the vector is 2.sup.n, then n bits are needed, in this embodiment, to indicate the location of any scalar within the vector.
In another embodiment of the invention, the location of the scalar within the vector is determined based on the value of a secondary op code in the instruction. It is noted, however, that the invention is not limited to any particular implementation of the scalar position indicator in the instruction. Instead, the invention covers any suitable way in which the location of a scalar within the vector can be represented or embedded in the bit format comprising the instruction.
In a preferred embodiment, the invention is implemented on a microprocessor, such as the microprocessors in IBM's PowerPC (IBM Trademark) family of microprocessors (hereafter "PowerPC"), wherein the microprocessor has been modified or redesigned to include a vector processing unit, such as a paired singles unit. For more information on the PowerPC microprocessors see PowerPC 740 and PowerPC 750 RISC Microprocessor Family User Manual, IBM 1998 and PowerPC Microprocessor Family: The Programming Environments, Motorola Inc. 1994, both of which are hereby incorporated by reference in their entirety.
In the modified PowerPC example described above, the paired singles operation may be selectively enabled by, for example, providing a hardware implementation specific special purpose register (e.g. HID2) having a bit (e.g. 3.sup.rd bit) which controls whether paired single instructions can be executed. Other bits in the special purpose register can be used, for example, to control other enhancement options that may be available on the microprocessor.
The invention also provides specific instruction definitions for mixed vector and scalar operations. The invention is also directed to a decoder, such as a microprocessor or a virtual machine (e.g. software implemented hardware emulator), which is capable of decoding any of all of these particular instructions disclosed herein. The invention further relates to a storage medium which stores any or all of the particular instructions disclosed herein.
