System for Performing Fast Data Accessing in Multiply/Accumulate Operations While Using a VRAM

This invention relates to computer memory access systems generally, and particularly to systems which provide for fast sequential data access, useful in sequential computation operations such as multiply/accumulate operations required, for example, in neural network calculations.

In performing neural network calculations, a large number of multiply/accumulate operations are performed. Typically, a DSP (digital signal processor) is used to perform these calculations because it can complete a multiply and accumulate operation in a single memory access cycle (generally, within 75-125 nanoseconds). However, to provide a low cost, high performance neural network operation, a large sequentially accessed table of static weights (data values) must be available to the DSP. If the data values from this table, which is sequentially accessed, cannot be supplied within the access cycle time (i.e., within 75-125 nanoseconds), the DSP is obliged to wait for the data; in which event, the full speed of the DSP would not be utilized.

An SRAM (fast static random access memory) is capable of providing a DSP with data at a sufficiently fast rate (i.e., within 75-125 nanoseconds). However, an SRAM carries with it the disadvantages of high cost and high power dissipation.

What is needed and would be useful, therefore, is a lower cost system which could provide both fast memory access and fast multiply/accumulate operation, as would be required for example in neural network calculations.

A VRAM (video dynamic RAM) with a serial access port is disclosed for use with a DSP (Digital Signal Processor) for providing fast sequential access to stored data such as neural network weights.

According to one aspect of the invention, a data access system is provided which includes a memory for storing data, a processor, and a data bus.

The memory has a parallel access port and a serial access port, and the processor has a first input-output port and a second input-output port. The data bus couples the second input-output port of the processor to the parallel and serial access ports of the memory. The processor is responsive to data applied via the first input-output port for transferring the applied data, via the second input-output port, the data bus, and the parallel access port, to the memory for storage. The stored data is then accessed by the processor as a serial sequence via the serial access port.

According to a second aspect of the invention, a method is provided for sequentially accessing data stored in a memory using a shift register. The method includes the steps of specifying an address of a group of the data for access, transferring the group of data in parallel to the shift register, specifying a count corresponding to the number of shifts of the shift register, and outputting the group of data as a serial sequence in accordance with the count.

FIG. 1 is a block diagram of a system of the present invention, including a DSP (digital signal processor), a data/address/control bus, and a VRAM (video DRAM);

FIG. 2 is a block diagram of the system of FIG. 1 in greater detail, including a memory access controller;

FIG. 3 is a block diagram of the memory access controller of FIG. 2 in greater detail;

FIG. 4 is a block diagram of the VRAM employed in the system of FIG. 1;

FIG. 5 is a waveform diagram showing timing relationships between clock signals and data and memory address values transferred between the DSP and VRAM of the system of FIG. 1; and

FIG. 6 is a graphic representation of a neural network, showing input values (X.sub.i), weight values (W.sub.ij), and output values (Y.sub.j).

Referring now to FIG. 1, there is shown a system 11 which embodies the present invention. The system includes a processor 13, a data/address/control bus 15, and a memory 17.

The processor 13 has an input/output port 19 for receiving data (e.g., data such as input variables or weight values) from an external source such as a host computer (not shown), and for outputting information to the host computer. The processor 13 also has a second input/output port 21 for accessing (reading from and writing to) the memory 17. The memory has a parallel port 23 for receiving data from or transmitting data to the processor 13. To access data stored in the memory via this parallel port 23, the address of each stored data value (element) is supplied, and the memory is randomly accessed, by the processor 13. The memory 17 also has a serial port 25 for receiving data from and transmitting data to the processor 13. Data stored in the memory are generally arranged in matrix form. A matrix location is accessible by specifying individual row and column addresses of the location.

As shown in FIG. 2, the memory 17 includes a VRAM memory chip 37 with an internal shift register 41, and a VRAM timing controller 39.

Data is transferred via serial port 25 to or from the shift register. This internal shift register (shown in greater detail in FIG. 4) can be initialized (loaded) with the contents of any row of the VRAM's memory matrix or can have its contents transferred (unloaded) to any row of the VRAM's memory matrix. The advantage of serial port 25 is the fast sequential access that it provides to data stored in the VRAM. Fast sequential access can be made without specification of the individual data addresses.

As shown in FIG. 2, processor 13 includes a dedicated microprocessor chip 33 (e.g., a model ADSP-2101 processor from Analog Devices Corporation), and a memory access controller 35 capable of providing fast serial access to the memory 17 in addition to normal parallel access. As indicated above, the memory 17 includes the VRAM chip 37 and VRAM timing controller 39. VRAM chip 37 may be a model TI44c251 VRAM from Texas Instruments Corporation, or a model MT42C4064 VRAM from Micron Technologies Corporation, for example, and controller 39 may be, for example, a model 74F1764 from Signetics Corporation.

As shown in FIGS. 2 and 3, memory access controller 35 includes a row address register 43 and a column address register 45 for specifying individual addresses of data values (e.g., a row address and a column address of a data value in an array of data values) stored in VRAM 37. Memory access controller 35 also includes an access counter 47 and a transfer counter 49. The access counter 47 and the transfer counter 49 are used to denote the current address of the VRAM location accessed by the processor 13 via the serial access port 25. The access counter contains the least significant bits (e.g., nine bits) of the address, and the transfer counter 49 contains the most significant bits (e.g., nine leading bits) of the address. The memory access controller 35 further includes an address-decode and control circuit 51 which maps addresses in the microprocessor's memory space to the various registers of the controller. For example, the address-decode logic of control circuit 51 maps the address space of the microprocessor 33 to the following registers and functions of the control circuit 51.

Generally, the memory space of the microprocessor 33 is less than the memory space available in the memory 17, and data is stored in the memory 37 In order to allow the microprocessor 33 to access the full address space available in the memory 17, the microprocessor first loads the row address register 43 with the row address of a location in memory 17 by writing to address xx0, then loads the column address register 45 with the column address of that location by writing to address xx1. The microprocessor then accesses address xx2. This causes the memory access controller 35 to apply output-enable (OE) signals 44, 46 (FIG. 3) to the row and column address registers 43, 45, to pass the row and column addresses to the VRAM timing controller 39, and to assert the appropriate RQ (request) signal 52. The VRAM timing controller 39 then completes the memory access by asserting RAS/CAS signals 18, 20 (FIG. 2), and supplying the appropriate address values on the address bus 28 to the VRAM 37 at predetermined VRAM-access times. In this way the microprocessor 33 reads data from and writes data to selected locations (addresses) in the VRAM 37. Data may be written in any sequence.

After the processor 13 loads the VRAM 37 with data (i.e., lists the data in a sequence suitable for later sequential access), the fast sequential access feature of the system is initialized, as follows. First, the transfer counter 49 (FIG. 3) is loaded with the row address of the first data word required (i.e., the first data word to be accessed), then the access counter 47 is loaded with the column address of that first data word. Next, the mode register 52 is set to serial read mode (for serial data transfer), and a transfer cycle is initiated by accessing address xx7. This access operation results in the memory access controller 35 applying output-enable (OE) signals 48, 50 to the access counter 47 and to the transfer counter 49, respectively. This allows the row/column addresses in the counters 47, 49 to be passed to the VRAM timing controller 39, and allows the serial request signal (RQ) 56 to be applied to controller 39. In response to the serial request signal 56, the VRAM timing controller completes the transfer cycle by applying the RAS/CAS signals 18, 20 and the transfer signal 58 (FIG. 4) to the VRAM while supplying the row/column addresses, on the address bus 27. This action of completing the transfer cycle results in the transfer of the data values in the memory locations having the row address to the SAM or shift register 41. The associated column address in the access counter 47 indicates which data value in the SAM is the first to be shifted out.

The processor then accesses the data list sequentially, by accessing the SAM ACCESS address (xx6). Each time the processor asserts (specifies) the SAM ACCESS address, the following events occur. First a serial clock signal 30 (FIGS. 2 and 3) is asserted (applied to VRAM 37), resulting in the next data word being clocked out of the SAM 41 and transmitted to the processor via the serial access port 25 (FIG. 1). Then, the access counter 47 is decremented and, when the access counter value (i.e., the contents of the access counter) equals zero, the transfer counter 49 is decremented. After which, a transfer cycle is initiated (forced) by the memory access controller 35 (FIG. 2). Next, the row address is obtained from the transfer counter 49, and the SAM's start address is obtained from the access counter 47 (the contents of which would now read 0). In this way, the processor sequentially accesses the data in the VRAM without supplying any intermediate addresses. The processor is required to supply the initial (starting) address only.

Use of VRAMs which have split SAMs (each SAM having one half the width of a row in the VRAM matrix) allows one half of the SAM to be involved in the transfer of data, while the other half is involved in the shifting of data. This type of VRAM architecture allows the processor to sequentially access data from the VRAM without requiring the processor to be involved with VRAM's transfer cycles. The transfer cycles are transparent to the processor.

A time relationship between clock signal, address access, and serial data output is shown in FIG. 5.

As indicated previously, the ability of the system to perform fast sequential data access operations is important for performance of compute-intensive sequential operations (e.g., fast multiply/accumulate operations that are common to neural net computations. An example of a neural net configuration is shown in FIG. 6. An associated neural net computation is shown by equation (1) below for one layer of the neural net.

In preparation for performing a neural net computation, the system 11 (FIG. 1), upon receipt of weight data (W.sub.ij) from the host computer (not shown), develops (builds) a table of the weight data in memory 17 by loading the weight data via parallel port 23 into the memory in a predetermined sequence (as shown in Table I below) suitable for sequential access subsequently. The table may be represented as an array of weight values (e.g., a 512.times.512 matrix of memory locations/addresses, each location having a four-bit weight value).

The above table represents a sequentially arranged three-layer list (data list) for performing n.times.m computations, where n represents the number of inputs, and m represents the number of outputs. The output of one layer becomes the input to the next layer. For example, for the first layer there are n=10 inputs (columns) and m=5 outputs (rows). For the second layer there are n=5 inputs and m=3 outputs, and for the third layer there are n=3 inputs and m=2 outputs.

As an alternative to loading the VRAM 37 via parallel port 23, loading may be achieved via serial port 25 and SAM 41.

After the weight values (W.sub.ij) are stored in the memory 17, the input values (X.sub.i) from the host computer are stored locally in the processor 13. (Typically, the quantity (n) of input values number about one hundred). The DSP then performs the multiply-accumulate operation shown in equation (1) below for each output (Y.sub.j): ##EQU1## The output value Y.sub.j may be adjusted to ensure that it falls within a predefined range of values or meets a predefined threshold. When so adjusted the adjusted value (Y') may be expressed as a function of Y.sub.j as follows: ##EQU2##

where n represents the number of input (X) values.

In the case where m represents the number of output (Y) values, the total number of outputs (Z) per layer may be expressed as: ##EQU3##

When n=100 and m=30, for example, the DSP 13 performs 3000 MAC (multiply/accumulate) operations for each layer of the neural net.

In the event there are k layers, with each layer having Z outputs, the total number of outputs (O.sub.T) may be expressed as: ##EQU4## and the total number of multiply/accumulate operations (O.sub.P) may be expressed as follows: ##EQU5##

where, Z.sub.i represents the initial number of inputs (X) to the neural net.

To accommodate such computer-intensive multiply-accumulate operations, the system configuration (architecture) enables the processor 13 to access (via the serial access port of the memory 17) the weight values in the memory at a rate of up to 40.times.10.sup.6 words (memory locations) per second, significantly faster than normal (parallel) DRAM access rates of about 5.times.10.sup.6 words (memory locations) per second.

While the present invention is susceptible of embodiment in various forms, there is shown in the drawings and described in the specification a presently preferred embodiment, with the understanding that the present disclosure is to be considered as an exemplification of the invention, and is not intended to limit the invention to the specific embodiment illustrated.