 |
Description  |
|
|
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to the fields of microprocessor and embedded DRAM architectures. More particularly, the invention pertains to a split processor architecture whereby a CPU portion performs standard processing and control functions,
an embedded DRAM portion performs memory-intensive manipulations, and the CPU and embedded DRAM portions function in concert to execute a single program.
2. Description of the Prior Art
Microprocessor technology continues to evolve rapidly. Every few years processor circuit speeds double, and the amount of logic that can be implemented on a single chip increases similarly. In addition, RISC, superscalar, very long instruction
word (VLIW), and other architectural advances enable the processor to perform more useful work per clock cycle. Meanwhile, the number of DRAM cells per chip doubles and the required refresh rate halves every few years. The fact that DRAM access times
do not double every few years results in a processor-DRAM speed mismatch. If the processor is to execute a program and manipulate data stored in a DRAM, it will have to insert wait states into its bus cycles to work with the slower DRAM. To combat
this, hierarchical cache structures or large on-board SRAM banks are used so that on average, much less time is spent waiting for the large but slower DRAM.
Real-time multimedia capabilities are becoming increasingly important in microcomputer systems. Especially with video and image data, it is not practical to build caches large enough to hold the requisite data structures while they are being
processed. This gives rise to large amounts of data traffic between the memory and the processor and decreases cache efficiency. For example, the Intel Pentium processors employ MMX technology, which essentially provides a vector processor subsystem
that can process multiple pixels in parallel. However, even with faster synchronous DRAM, the problem remains that performance is limited by the DRAM access time needed to transfer data to and from the processor.
Other applications where external DRAM presents a system bottleneck are database applications. Database processing involves such algorithms as searching, sorting, and list processing in general. A key identifying requirement is the frequent use
of memory indirect addressing. In memory indirect addressing, a pointer is stored in memory. The pointer must be retrieved from memory and then used to determine the address of another pointer located in memory. This addressing mode is used
extensively in linked list searching and in dealing with recursive data structures such as trees and heaps. In these situations, cache performance diminishes as the processor is burdened with having to manipulate large data structures distributed across
large areas in memory. In many cases, these memory accesses are interleaved with disk accesses, further reducing system performance.
Several prior art approaches have been used to increase processing speed in microsystems involving a fast processor and a slower DRAM. Many of these techniques, especially cache oriented solutions, are detailed in "Computer Architecture: A
Quantitative Approach, 2nd Ed.," by John Hennessy and David Patterson (Morgan Kaufmann Publishers, 1996). This reference also discusses pipelined processing architectures together with instruction-level parallel processing techniques, as embodied in
superscalar and VLIW architectures. These concepts are extended herein to provide improved performance by providing split caching and instruction-level parallel processing structures and methods that employ a CPU core and embedded DRAM logic.
The concept of using a coprocessor to extend a processor architecture is known in the art. Floating point coprocessors, such as the Intel 80.times.87 family, monitor the instruction stream from the memory into the processor, and, when certain
coprocessor instructions are detected, the coprocessor latches and executes the coprocessor instructions. Upon completion, the coprocessor presents the results to the processor. In such systems, the processor is aware of the presence of the
coprocessor, and the two work together to accelerate processing. However, the coprocessor is external from the memory, and no increase in effective memory bandwidth is realized. Rather, this solution speeds up computation by employing a faster
arithmetic processor than could be integrated onto a single die at the time. Also, this solution does not provide for the important situation when the CPU involves a cache. In such situations, the coprocessor instructions cannot be intercepted, for
example, when the CPU executes looped floating point code from cache. Another deficiency with this prior art is its inability to provide a solution for situations where the processor is not aware of the presence of the coprocessor. Such a situation
becomes desirable in light of the present invention, whereby a standard DRAM may be replaced by an embedded DRAM to accelerate processing without modification of preexisting application software.
Motorola employed a different coprocessor interface for the MC68020 and MC68030 processors. In this protocol, when the processor executes a coprocessor instruction, a specialized sequence of bus cycles is initiated to pass the coprocessor
instruction and any required operands across the coprocessor interface. If, for example, the coprocessor is a floating point processor, then the combination of the processor and the coprocessor appears as an extended processor with floating point
capabilities. This interface serves as a good starting point, but does not define a protocol to fork execution threads or to jointly execute instructions on both sides of the interface. Furthermore, it does not define a protocol to allow the
coprocessor to interact with the instruction sequence before it arrives at the processor. Moreover, the interface requires the processor to wait while a sequence of slow bus transactions are performed. This interface concept is not sufficient to
support the features and required performance needed of the embedded DRAM coprocessors.
U.S. Pat. No. 5,485,624 discloses a coprocessor architecture for CPUs that are unaware of the presence of a coprocessor. In this architecture, the coprocessor monitors addresses generated by the CPU while fetching instructions, and when
certain addresses are detected, interprets an opcode field not used by the CPU as a coprocessor instruction. In this system, the coprocessor then performs DMA transfers between memory and an interface card. This system does not involve an embedded DRAM
that can speed processing by minimizing the bottleneck between the CPU and DRAM. Moreover, the coprocessor interface is designed to monitor the address bus and to respond only to specific preprogrammed addresses. When one of these addresses is
identified, then an unused portion of an opcode is needed in which to insert coprocessor instructions. This system is thus not suited to systems that use large numbers of coprocessor instructions as in the split processor architecture of the present
invention. A very large content addressable memory (CAM) would be required to handle all the coprocessor instruction addresses, and this CAM would need to be flushed and loaded on each task switch. The need for a large CAM eliminates the DRAM area
advantage associated with an embedded DRAM solution. Moreover, introduction of a large task switching overhead eliminates the acceleration advantages. Finally, this technique involves a CPU unaware of the coprocessor but having opcodes that include
unused fields that can be used by the coprocessor. A more powerful and general solution is needed.
The concept of memory based processors is also known in the art. The term "intelligent memories" is often used to describe such systems. For example, U.S. Pat. No. 5,396,641 discloses a memory based processor that is designed increase
processor-memory bandwidth. In this system, a set of bit serial processor elements function as a single instruction, multiple data (SIMD) parallel machine. Data is accessed in the memory based processor using normal row address and column address
strobe oriented bus protocols. SIMD instructions are additionally latched in along with row addresses to control the operation of the SIMD machine under control by a host CPU. Hence, the description in U.S. Pat. No. 5,396,641 views the intelligent
memory as a separate parallel processor controlled via write operations from the CPU. While this system may be useful as an attached vector processor, it does not serve to accelerate the normal software executed on a host processor. This architecture
requires the CPU to execute instructions to explicitly control and route data to and from the memory based coprocessor. This architecture does not provide a tightly coupled acceleration unit that can accelerate performance with specialized instruction
set extensions, and it cannot be used to accelerate existing applications software unaware of the existence of the embedded DRAM coprocessor. This architecture requires a very specialized form of programming where SIMD parallelism is expressly
identified and coded into the application program.
It would be desirable to have an architecture that could accelerate the manipulation of data stored in a slower DRAM. It would also be desirable to be able to program such a system in a high level language programming model whereby the
acceleration means are transparent to the programmer. It would also be desirable to maintain the processing features and capabilities of current microprocessors, to include caching systems, instruction pipelining, superscalar or VLIW operation, and the
like. It would also be desirable to have a general purpose processor core that could implement operating system and applications programs so that this core could be mixed with different embedded DRAM coprocessors to accelerate the memory intensive
processing of, for example, digital signal processing, multimedia or database algorithms. Finally, it would be desirable if a standard DRAM module could be replaced by an embedded DRAM module with processor architectural extensions, whereby existing
software would be accelerated by the embedded DRAM extension.
SUMMARY OF THE INVENTION
One aspect of the present invention is a processor whose architecture is partitioned into a CPU core portion and an embedded DRAM portion. The CPU core portion handles the main processing and control functions, while the embedded DRAM portion
performs memory-intensive data manipulations. In the architecture, instructions execute either on the CPU core portion of the processor, the embedded DRAM portion of the processor, or across both portions of the processor.
In another aspect of the present invention, the CPU portion is able to effectively cache instructions and data while still sharing the instruction stream with the embedded DRAM portion of the processor implemented in the embedded DRAM. A
separate caching structure is employed for a different program space on the embedded DRAM. Using this system, the separation of the CPU and embedded DRAM portions of the architecture is transparent to the programmer, allowing standard high level
language software to run. In one embodiment, a special compiler is used to segment the code into a plurality of instruction types. The processor architecture takes advantage of the embedded DRAM, advantageously employing multiple address spaces that
are transparent to the first portion of the processor, and that minimize data bussing traffic between the processors.
Another aspect of the present invention is an apparatus and method to execute standard available software on a split architecture. For example, in the personal computer and workstation markets there are already multi-billion dollar investments
in preexisting software. In this aspect of the invention, an embedded DRAM module may be inserted into an existing single in line memory module (SIMM) slot. Thus, an accelerator may be added without needing to modify existing application software, and
the upgrade can be performed effortlessly in the field. This functionality is enabled by allowing the embedded DRAM coprocessor to monitor the instruction stream and to replace certain instruction sequences with read and write commands. In one
embodiment a profiler analyzes uniprocessor execution either statistically or dynamically and then constructs modification tables to reassign certain code segments to the embedded DRAM coprocessor. In another embodiment, the embedded DRAM performs the
analysis in real-time. In still another embodiment, the embedded DRAM is exercised by standard software through the use of preloaded driver programs accessed via operating system calls.
Another aspect of the present invention is a computer system which comprises a central processing unit and an external memory coupled to the central processor. The external memory comprises one or more dynamic random access memory (DRAM) arrays,
a set of local functional units, a local program prefetch unit, and a monitor/modify unit. The monitor/modify unit is operative to evaluate each instruction opcode as it is fetched from the DRAM array, and, in response to the opcode, to perform one of
the following actions: (i) sending the opcode to the central processing unit; (ii) sending the opcode to the set of local functional units; and (iii) sending the opcode to the local program prefetch unit to fork a separate execution thread for execution
by the set of local functional units.
Preferably, in response to the opcode, the monitor/modify unit also performs the actions of sending the opcode to the set of local functional units, substituting at least one different opcode for the opcode, and sending the at least one different
opcode to the central processing unit. Also preferably, the at least one different opcode instructs the central processing unit to read values from the external memory representative of the register contents that would have been present in the central
processing unit had the central processing unit executed the original instruction stream.
Another aspect of the present invention is an embedded dynamic random access memory (DRAM) coprocessor designed to be coupled to a central processing unit. The embedded DRAM coprocessor comprises one or more DRAM arrays. An external memory
interface is responsive to address and control signals generated from an external source to transfer data between the DRAM arrays and the external source. A set of local functional units execute program instructions. A local program prefetch unit
fetches program instructions. A monitor/modify unit evaluates each instruction opcode as it is fetched under control of the external source from the DRAM array, and, in response to the opcode, performs one of the following actions: (i) sending the
opcode to the external source; (ii) sending the opcode to the set of local functional units; and (iii) sending the opcode to the local program prefetch unit to fork a separate execution thread for execution by the set of local functional units.
Preferably, in response to the opcode, the monitor/modify unit also performs the actions of sending the opcode to the set of local functional units, substituting one or more different opcodes for the opcode, and sending the one or more different
opcodes to the external source.
Another aspect of the present invention is a computer system which comprises a central processing unit coupled to an external memory. The central processor unit comprises a first set of functional units responsive to program instructions. A
first program cache memory has at least one level of caching and provides high speed access to the program instructions. A first prefetch unit controls the fetching of a sequence of instructions to be executed by the first set of functional units. The
instructions are fetched from the external memory unless the program instructions are found in the first program cache memory; in which case, the program instructions are fetched from the first program cache memory. The external memory comprises one or
more dynamic random access memory (DRAM) arrays, a second set of local functional units, a second program prefetch unit, and a second program cache memory. The first program cache memory only caches instructions executed by the functional units on the
central processing unit, and the second program cache memory only caches instructions executed by the second set of functional units on the external memory device. Preferably, the first program cache memory is a unified cache which also serves as a data
cache. Also preferably, the central processing unit sends one or more attribute signals to identify certain memory read signals to be instruction fetch cycles. The attribute signals are decoded by logic embedded in the external memory so that the
second program cache memory can identify opcode fetch cycles. In particular embodiments, the external memory further includes a monitor/modify unit which intercepts opcodes fetched by the first prefetch unit and passes the opcodes to the second prefetch
unit to cause the second prefetch unit to fetch a sequence of program instructions for execution. The opcodes of the sequence of program instructions are fetched from the one or more DRAM arrays unless they are found to reside in the second program
cache.
Another aspect of the present invention is an embedded dynamic random access memory (DRAM) coprocessor which comprises an external memory interface for transferring instructions and data in response to address and control signals received from an
external bus master. The coprocessor also comprises one or more DRAM arrays, a set of local functional units, a program prefetch unit, and a program cache memory. The program cache memory only caches instructions executed by the functional units on the
external memory device. Preferably, the external memory interface receives one or more attribute signals to identify certain memory read signals to be instruction fetch cycles. The attribute signals are decoded by logic embedded in the external memory
so that the program cache can identify externally generated opcode fetch cycles. The coprocessor preferably includes a monitor/modify unit which intercepts opcodes in instructions transferred over the external memory interface and which passes the
opcodes to the program prefetch unit to cause the program prefetch unit to fetch a sequence of program instructions for execution. The opcodes of the sequence of program instructions are fetched from the one or more DRAM arrays unless the opcodes of the
sequence of program instructions are found to reside in the program cache.
Another aspect of the present invention is a computer system which comprises a central processing unit coupled to an external memory. The central processing unit comprises a first set of functional units responsive to program instructions. A
first prefetch unit controls the fetching of a sequence of instructions from the external memory to be executed by the first set of functional units. The external memory comprises one or more dynamic random access memory (DRAM) arrays, a second set of
local functional units, one or more external interface busses, and a second program prefetch unit. The central processing unit and the external program memory jointly execute a single program which is segmented into first and second program spaces. The
first program space comprises type I, type II and optionally type III instructions. The second program space comprises type II and type III instructions. The type I instructions always execute on the first set of functional units. The type II
instructions generate interface control exchanges between the central processing unit and the external memory. The type II instructions selectively are split into portions executed on the central processing unit and portions executed on the external
memory. The type III instructions always execute on the second set of functional units. Preferably, the central processing unit has a first program cache, and the external memory has a second program cache. The first cache only caches the type I and
the type II instructions accessed in the first program space. The second program cache only caches type II and type III instructions accessed in the second program space. Preferably, upon the execution of the type II instruction on the central
processing unit, a logical address is transferred over one of the external interface busses to the external memory. The external memory passes the logical address to the second prefetch unit, which, in turn, fetches a sequence of instructions from the
second program space. The sequence of instructions is executed by a second set of functional units in the external memory. Preferably, the type II instructions comprise first and second opcodes. The first opcode executes on the central processing
unit, and the second opcode executes on the external memory. The first opcode comprises instruction type identifier information, opcode information to direct execution of a one of the first set of functional units, and an address field to be transferred
over one of the external interface busses to reference instructions in the second program space. The second opcode comprises instruction type identifier information and opcode information to direct execution of a one of the second set of functional
units. Preferably, the second opcode further comprises signaling information to be passed across one of the external interface busses to the central processing unit. A stop field indicates to the second prefetch unit to stop fetching instructions from
the second program space. Preferably, the type II instruction is a split branch to subroutine instruction, and upon execution of the split branch to subroutine instruction, a subroutine branch address is passed across one of the external interface
busses to activate a subroutine stored in the second program space. Preferably, the type II instruction involves a first operand stored in memory and a second operand stored in a register located on the central processing unit. The type II instruction
is split into a first portion and a second portion. The first portion executes on the external memory to access the first operand and to place it on one of the external interface busses. The second portion executes on the central processing unit which
reads the first operand from one of the external interface busses and computes a result of the type II instruction.
Another aspect of the present invention is an embedded dynamic random access memory (DRAM) coprocessor which jointly executes a program with an external central processing unit. The embedded DRAM coprocessor comprises a DRAM array which
comprises one or more DRAM banks. Each bank has an associated row pointer. Each row pointer is operative to precharge and activate a row in the respective DRAM bank. A first synchronous external memory interface accepts address and control information
used to access memory locations in the DRAM array. A second synchronous external memory interface receives type II instruction information from an external source. A prefetch unit is responsive to the received type II information to execute one or more
instructions referenced by the received type II information. A set of one or more functional units is responsive to instructions fetched by the prefetch unit. Preferably, the first and the second synchronous interfaces share a common bus. Also
preferably, the embedded DRAM coprocessor further comprises a program cache which caches program instructions fetched under the control of the prefetch unit from the DRAM array. The embedded DRAM coprocessor may also further comprise a register file
coupled to the DRAM array and to the functional units, wherein at least a subset of the register file contains a mirror image of a register set contained on the external central processing unit. In certain embodiments, at least a subset of the set of
one or more functional units includes a replica of at least one functional unit contained on the external central processing unit. In preferred embodiments, the register file may further include a set of multimedia extension (MMX) registers, and the at
least one functional unit may include at least one MMX functional unit.
Another aspect of the present invention is a computer system which comprises a central processing unit which includes at least one level of program cache memory. An embedded dynamic random access memory (DRAM) coprocessor is coupled to the
central processing unit. The embedded DRAM coprocessor comprises a DRAM array which comprises one or more DRAM banks, each with an associated row pointer. The row pointer precharges and activates a row in the DRAM bank. A first synchronous external
memory interface accepts address and control information used to access memory locations in the DRAM array. A second synchronous external memory interface receives from the central processing unit addresses into the second program space. A prefetch
unit is responsive to the received addresses and prefetches one or more instructions referenced by the received addresses. A set of one or more functional units is responsive to instructions fetched by the prefetch unit. When the central processor
executes specified instructions in an instruction stream read from a first program memory space in the embedded DRAM coprocessor, the central processor sends address information to the embedded DRAM coprocessor which references instructions in a second
program memory space located in the embedded DRAM coprocessor. As a result, the central processing unit and the embedded DRAM coprocessor jointly execute a program. Preferably, the embedded DRAM coprocessor further includes a register file coupled to
the DRAM array and the functional units. At least a subset of the register file contains a mirror image of a register set contained on the external central processing unit. At least a subset of the set of the one or more functional units is capable of
executing a subset of the instruction set executed on the central processing unit. Also preferably, the register file further includes a set of multimedia extension (MX) registers., and the functional units include one or more MMX functional units.
Another aspect of the present invention is a central processing unit cooperative to jointly execute programs fetched from an embedded dynamic random access memory (DRAM) coprocessor. The central processing unit comprises a prefetch unit which
fetches instructions to be executed by the central processing unit, set of internal registers, a set of one or more functional units which executes instructions, an optional program cache, a first external memory interface which transfers addresses,
control signals and data to and from external memory and input/output (I/O) devices, and a second external memory interface which transfers synchronization signals and address information between the central processing unit and the embedded DRAM
coprocessor. The central processing unit and the embedded DRAM coprocessor jointly execute a single program that is partitioned into first and second memory spaces. The instructions in the first memory space are executed by the central processing unit. The instructions in the second memory space are executed by the embedded DRAM coprocessor. The instructions in the first memory space include a first type of instruction and a second type of instruction. The first type of instruction is executed wholly
on the central processing unit. Upon execution, the second type of instruction sends address information which references instructions in the second program space to the embedded DRAM coprocessor. Upon execution of the second type of instruction, the
central processing unit directs the embedded DRAM coprocessor to perform at least one of the following operations: (i) fork a separate execution thread to execute a sequence of instructions stored in the second program space; (ii) execute a fixed number
of instructions and then stop; and (iii) execute a fixed number of instructions and supply one or more results over one of the first external memory interface and the second external memory interface in alignment with a clock edge, a fixed number of
clock cycles later.
Another aspect of the present invention is a central processing unit cooperative to jointly execute programs fetched from an embedded dynamic random access memory (DRAM) coprocessor. The central processing unit comprises a prefetch unit
operative to fetch instructions to be executed by the central processing unit, a set of internal registers, a set of one or more functional units operative to execute instructions, a program cache, a first external memory interface operative to transfer
addresses, control signals, and data to and from external memory and input/output (I/O) devices, and a second external memory interface operative to transfer synchronization signals and optionally address information between the central processing unit
and the embedded DRAM coprocessor. The central processing unit and the embedded DRAM coprocessor jointly execute a single program which is partitioned into first and second memory spaces. The instructions in the first memory space are executed by the
central processing unit, and the instructions in the second memory space are executed by the embedded DRAM coprocessor. The instructions in the first memory space include a first type of instruction which is executed wholly on the central processing
unit and a second type of instruction which, upon execution, sends address information which references instructions in the second program space to the embedded DRAM coprocessor. The central processor unit and the embedded DRAM coprocessor have
overlapping architectures which include mirror image subsets of registers and mirror image subsets of functionality of the functional units. The central processing unit and the embedded DRAM coprocessor execute an overlapping instruction set.
Another aspect of the present invention is a method to jointly execute programs on a central processing unit coupled to an embedded dynamic random access memory (DRAM) coprocessor. The method comprises the steps of replicating a portion of a
register set of the central processing unit on the embedded DRAM coprocessor, and replicating a portion of the functionality of functional units of the central processing unit to support the replicating of a portion of the instruction set of the central
processing unit on the embedded DRAM coprocessor. A program is jointly executed on the central processing unit and the embedded DRAM coprocessor by partitioning computationally intensive portions of the code to run on the central processing unit and by
partitioning memory intensive code segments to run on the embedded DRAM coprocessor. The contents of selected ones of the replicated register subsets are transferred between the central processing unit and the embedded DRAM coprocessor in order to
maintain program level synchronization between the central processing unit and the embedded DRAM coprocessor. Preferably, the method further includes the step of adding an architectural extension on the embedded DRAM coprocessor. The architectural
extension comprises an additional set of registers beyond those contained on the central processing unit. The architectural extension also comprises additional instructions beyond those processed by the central processing unit. The method also
preferably includes the step of partitioning code segments which reference the additional registers and code segments which use the additional instructions to be executed on the embedded DRAM coprocessor.
Another aspect of the present invention is a split very long instruction word (VLIW) processing apparatus which comprises a VLIW central processor. The VLIW central processor comprises a set of functional units which receive a plurality of
instructions for execution in parallel and a first VLIW program cache which holds a collection of very long instruction words. Each very long instruction word comprises a set of instruction fields. Each instruction field comprises an instruction to be
executed by a functional unit. The central processor further includes a dispatch unit which scans bit fields within the instruction fields to decide how many instructions to dispatch in parallel and to which functional unit to direct each instruction.
One or more register files are coupled to the functional units. An external memory interface carries instructions and data from an external source. An on-board data memory is coupled to the functional units, the register files, and the external memory
interface. At least one of the functional units includes a branch processing unit which processes branch instructions. The branch processing unit is coupled to a prefetch unit used to sequence the VLIW control words from the VLIW program cache or
external memory. The branch processing unit is also coupled to an external interface for transferring branch related information. The processing apparatus also includes a VLIW extension processor which cooperates with the VLIW central processor to
jointly execute a single VLIW program. The VLIW extension processor comprises a set of at least one functional unit which receives one or more instructions for execution in a given clock cycle. A second VLIW program cache holds a collection of very
long instruction words, whereby each very long instruction word comprises one or more instruction fields, wherein each instruction field comprises an instruction to be executed by a functional unit. A second dispatch unit scans bit fields within the
instruction fields to decide how many instructions to dispatch in parallel and to which functional unit to direct each instruction. At least one of the functional units includes a second branch processing unit which processes branch instructions. The
branch processing unit is coupled to a prefetch unit which sequences VLIW control words from the second VLIW program cache. The branch processing unit is also coupled to a second external interface which transfers branch related information.
Preferably, the VLIW processing apparatus further includes an on-board dynamic random access memory (DRAM) coupled to the functional units. The on-board DRAM is coupled to an externally controllable synchronous memory interface. Also preferably, the
DRAM is buffered via one of an active row sense amp techniques, an SRAM cache, or a data register file. Also preferably, the first and second branch processing units simultaneously execute control dependent branches, as encountered in looping, without
transferring branch related information across the second external interface. Preferably, only one of the first and second branch processing units simultaneously executes data dependent branches, and the branch target address is passed to or from the
central VLIW from or to the VLIW extension processor. Preferably, the number of the data dependent branches requiring synchronization is minimized using conditional execution of instructions on the central VLIW processor and the VLIW extension
processor. Preferably, the first and second prefetch units fetch a single very long instruction word which comprises a first portion stored in the first VLIW program cache which controls the first prefetch unit and a second portion stored in the second
VLIW program cache which controls the second prefetch unit. Information encoded into the first portion of the very long instruction word indicates whether the first prefetch unit is to dispatch a set of instructions each clock cycle or to insert delays
between the dispatching of certain groups of instructions. Information encoded into the second portion of the very long instruction word indicates whether the second prefetch unit is to dispatch a set of instructions each clock cycle or to insert delays
between the dispatching of certain groups of instructions. The first and second prefetch units operate together in response to the first and second portions of the very long instruction word so as synchronize the dispatching of instructions in the
central VLIW processor and the extension VLIW processor. This conserves program memory space when the processing loads of the VLIW central processor and the VLIW extension processor are uneven. Also preferably, the first branch processing unit sends a
fork branch address to the second branch processing unit. The second branch processing unit responds to the fork branch address by forking an execution thread to a specified address. As a result, the central VLIW processor and the VLIW extension
processor execute separate instruction sequences in a decoupled manner. Upon completion of the separate instruction sequences, the first and second branch processing units resynchronize by executing join instructions which cause the central VLIW
processor and the extension VLIW processor to cease fetching instructions from a specified address until a synchronization signal has been received over the second external interface. Preferably, the processing apparatus also includes a sequential
access memory interface to the DRAM array so that the VLIW extension processor can directly control a frame buffer.
Another aspect of the present invention is a method to accelerate application programs written without knowledge of an embedded dynamic random access memory (DRAM) coprocessor. The application programs are written to execute on a central
processing unit. The method comprises the step of implementing a set of operating system level application program interface routines (APIs) called by the application program and the step of writing the APIs to implement a specified functionality of the
APIs by executing a first portion of a driver program on the central processing unit and by executing a second portion of the driver program on the embedded DRAM coprocessor. The driver program further executes instructions which generate interface
transactions between the central processor unit and the embedded DRAM coprocessor. Preferably, the method includes the step of evaluating specified ones of calls to the APIs and the step of determining whether to generate an operating system message or
to call the driver routine directly to bypass the normally associated operating system message delay.
Another aspect of the present invention is a method to accelerate application programs written without knowledge of an embedded dynamic random access memory (DRAM) coprocessor, wherein the application programs are written to execute on a central
processing unit having a level one (L1) cache and having a level two (L2) cache. The method comprises the step of executing an execution profiler routine which monitors system parameters such as memory waiting time, loop indices, and cache miss rate,
and the step of identifying program segments which generate specified memory traffic patterns which generate excessive numbers of L1 or L2 cache misses. The method includes the further step of constructing a modification table to reallocate the
identified program segments to execute in the embedded DRAM coprocessor. The modification table includes at least references to instructions used to communicate parameters and results between the central processor unit and the embedded DRAM coprocessor. The method further includes the step of loading the program to run with a loader program. The loader program also evaluates the modification tables in order to insert communication instructions where needed, to load instructions to be executed by the
central processing unit into a first program space, and to load instructions to be executed by the embedded DRAM coprocessor into a second program space. Preferably, the execution profiler further comprises a first communicating component which runs on
the central processing unit and which monitors processor and L1 cache performance. A second communicating component runs on the embedded DRAM coprocessor and monitors DRAM traffic. Also preferably, the second communicating component further monitors
the L2 cache miss rate.
Another aspect of the present invention is a method to accelerate application programs written without knowledge of an embedded dynamic random access memory (DRAM) coprocessor, wherein the application programs are written to execute on a central
processing unit. The method comprises the step of parsing a machine code program with a translator program knowledgeable of the machine language. The translator identifies program segments which contain specific opcode types and identifies program
segments which contain loop constructs whose loop counters are initialized with numbers beyond prespecified thresholds. The method also includes the step of constructing a modification table to reallocate the identified program segments to execute in
the embedded DRAM coprocessor. The modification table includes at least references to instructions used to communicate parameters and results between the central processor unit and the embedded DRAM coprocessor. The method also includes the step of
loading the program to run with a loader program. The loader program also evaluates the modification tables in order to insert the parameter and result communication instructions where needed, to load instructions to be executed by the central
processing unit into a first program space, and to load instructions to be executed by the embedded DRAM coprocessor into a second program space. Preferably, the program segment boundaries are aligned with subroutine call and return instructions. Also
preferably, the specific opcode types are multimedia extension (MMX) instructions. Preferably, at least one of the specific opcode types is not executable on the central processor unit, but is executable on the embedded DRAM coprocessor. At least one
of specific opcode types may be a multimedia extension (MMX) instruction.
Another aspect of the present invention is an embedded dynamic random access memory (DRAM) coprocessor implemented as individual bit slice units which are equipped with standard single in-line memory module (SIMM) interface connectors so they can
be interchanged with standard DRAM SIMMs found on computer boards in personal computers, workstations, and other forms of electronic equipment. Preferably, the coprocessor further comprises an additional interface connector which connects the embedded
DRAM bit slice processors, together via a separate backplane not found on the computer board into which the embedded DRAM coprocessors are plugged. Also preferably, the bit slice width of the SIMM is equal to the bus word width of the processor to which
the memory modules are attached.
Another aspect of the present invention is a processing architecture which includes a first CPU core portion coupled to a second embedded dynamic random access memory (DRAM) portion. These architectural components jointly implement a single
processor and instruction set. Advantageously, the embedded logic on the DRAM chip implements the memory intensive processing tasks, thus reducing the amount of traffic that needs to be bussed back and forth between the CPU core and the embedded DRAM
chips. The embedded DRAM logic monitors and manipulates the instruction stream into the CPU core. The architecture of the instruction set, data paths, addressing, control, caching, and interfaces are developed to allow the system to operate using a
standard programming model. Specialized video and graphics processing systems are developed. Also, an extended very long instruction word (VLIW) architecture implemented as a primary VLIW processor coupled to an embedded DRAM VLIW extension processor
efficiently deals with memory intensive tasks. In different embodiments, standard software can be accelerated either with or without the express knowledge of the processor.
BRIEF DESCRIPTION OF THE DRAWINGS
The various novel features of the invention are illustrated in the figures listed below and described in the detailed description which follows.
FIG. 1 is a high level bl | | |
