MEMORY ACCESS UNIT

Abstract

A memory access unit for handling transfers of samples in a d-dimensional array between a one of m data buses, where m≧1, and k*m memories, where k≧2, is disclosed. The memory access unit comprises k address calculators, each address calculator configured to receive a bus address to add a respective offset to generate a sample bus address and to generate, from the sample bus address according to an addressing scheme, a respective address in each of the d dimensions for access along one of the dimensions from the bus address according to an addressing scheme, for accessing a sample. The memory access unit comprises k sample collectors, each sample collector operable to generate a memory select for a one of the k*m memories so as to transfer the sample between a predetermined position in a bus data word and the respective one of the k*m memories. Each sample collector is configured to calculate a respective memory select in dependence upon the address in each of the d dimensions such that each sample collector selects a different one of the k*m memories so as to allow the sample collectors to access k of the k*m memories concurrently. A memory controller may comprise m memory access units for handling transfers of samples in a d-dimensional array between m data buses and k*m memories.

Description

FIELD OF THE INVENTION

The present invention relates to a memory access unit and to a memory controller comprising one or more memory access units.

BACKGROUND

Motor vehicles are increasing being equipped with different forms of sensors for detecting harmful situations. These sensors produce large amounts of data in multiple dimensions. Ideally, samples should be stored sequentially in memory to allow hardware accelerators to process them efficiently. However, there is often a need to process samples in more than one dimension.

One solution is to re-order samples between processing steps. However, this increases computation overheads and memory requirements.

Another solution is to employ memory which allows data to be read out in orthogonal dimensions. An example of such a memory is described in WO 2009/003115 A1.

SUMMARY

The present invention seeks to provide a memory access unit for handling transfers of data words of fixed size (herein referred to as “samples”).

According to a first aspect of the present invention there is provided a memory access unit for handling transfers of samples in a d-dimensional array between a one of m data buses, where m≧1, and k*m memories, where k≧2. The memory access unit comprises k address calculators and k sample collectors. Each address calculator is configured to receive a bus address, to add a respective offset to generate a bus address for a sample and to generate, from the bus address of the sample according to an addressing scheme, a respective address in each of the d dimensions for accessing a respective sample. Each sample collector is operable to generate a memory select for a one of the k*m memories so as to transfer the sample between a predetermined position in a bus data word and the one of the k*m memories. Each sample collector is configured to calculate a respective memory select in dependence upon the address in each of the d dimensions such that each sample collector selects a different one of the k memories so as to allow the sample collectors to access k of the k*m memories concurrently. For a single bus, m=1 and k*m=k. Each sample collector may be operable to transfer the respective sample to or from the respective one of the k*m memories.

Thus, samples can be accessed (read or write) along any of the dimensions without sample re-ordering in the order of the dimension.

A sample data width and a sample storage data width may be the same. The sample data width may be an integer multiple of (for example, double) the sample storage data width. The sample data width may be or may be settable to be, for example, four bytes, consisting of 32 bits, or eight bytes, consisting of 64 bits. The storage data width may be, for example, four bytes, consisting of 32 bits. However, other storage data widths may be used. The data bus width may be 16 bytes, consisting of 128 bits, or 32 bytes, consisting of 256 bits.

The memory access unit may further comprise a set of registers for changeably setting the numbers of samples in each of the d dimensions and/or sample data width.

The number d of dimensions may be or may be settable to be two or three.

Each sample collector may be configured to calculate the memory select in dependence upon a sum of the addresses in each of the d dimensions.

Each sample collector may be configured to calculate the memory select, CS, using:

CS=[(A_1+A_2+ . . . +A_d)]%k

Each address calculator may be configured to generate an index. This may be used when the sample data width is greater than the sample storage data width.

Each sample collector may be configured to calculate the memory select, CS, using:

CS=[I_S+(w_s/w_m)*(A_1+A_2+ . . . +A_d)]%k

where w_s is the sample data width and w_m is memory data width.

Each address calculator may be configured to calculate the respective addresses in dependence upon a linearly-increasing word address, sample data width and numbers of samples in each dimension.

Each address calculator may be configured to receive a bus address and to adjust the bus address based on a respective offset. Thus, the address calculators can generate respective sets of addresses and, thus, generate different memory selects.

The memory access unit may further comprise a bus interface coupled to the address calculators and the sample collectors. The bus interface may be configured to pass a bus address to each of the address calculators. 30o The memory access unit is preferably implemented in hardware logic.

The memory access unit may be suitable for handling transfers of samples in a d-dimensional array between a data bus and k*m memories, where the k*m memories are shared by m data buses.

According to a second aspect of the present invention there is provided a memory controller comprising at least one memory access unit.

The memory controller may comprise at least two memory access units. The at least two memory access units may access a common (or “global”) set of registers.

The memory controller may comprise m memory access units for handling transfers of samples in a d-dimensional array between m data buses and k*m memories. According to a third aspect of the present invention there is provided a memory system comprising a memory controller and k*m memories, each set of k*m memories operatively connected to the m memory access units.

There may be one data bus, i.e. m=1. There may be more than one data bus, i.e. m≧2. There may be between two and ten, or more data buses, i.e. 10≧m≧2 or m≧10.

According to a fourth aspect of the present invention there is provided an integrated circuit comprising a memory access unit or a memory controller.

The integrated circuit may be a microcontroller. The integrated circuit may be an application specific integrated circuit (ASIC). The integrated circuit may be a system on a chip (SoC). The integrated circuit may be a hardware accelerator. The integrated circuit may be a graphical processing unit (GPU). The integrated circuit may be a digital signal processor (DSP).

According to a fifth aspect of the present invention there is provided a motor vehicle comprising a computing device comprising a memory access unit or a memory controller.

The motor vehicle may be a motorcycle, an automobile (sometimes referred to as a “car”), a minibus, a bus, a truck or lorry. The motor vehicle may be powered by an internal combustion engine and/or one or more electric motors.

According to a sixth aspect of the present invention there is provided a method of transferring samples in a d-dimensional array between a one of m data buses and k*m memories. The method comprises, for each of k samples: receiving a bus address and adding a respective offset to generate a bus address of a sample and generating, from the sample bus address according to an addressing scheme, a respective address in each of the d dimensions for accessing samples along one of the dimensions and generating a memory select for a one of the k*m memories so as to transfer a sample between a predetermined position in a bus data word and the respective one of the k*m memories, wherein generating the memory select comprises calculating a memory select in dependence upon the addresses in each of the d dimensions so as to allow the k sample to be written to or read from k of the k*m memories concurrently.

The method can be performed by a unit (or module) such that it allows the unit and other units to access the k*m memories concurrently.

The method is preferably a hardware-implemented method. The unit (or module) may be a logic unit (or “logic module”).

According to a seventh aspect of the present invention there is provided a computer program comprising instructions which, when executed by one or more processors, causes the one or more processors to perform the method.

According to an eighth aspect of the present invention there is provided a computer readable medium (which may be non-transitory) carrying or storing the computer program.

BRIEF DESCRIPTION OF THE DRAWINGS

Certain embodiments of the present invention will now be described, by way of example, with reference to the accompanying drawings, in which:

FIG. 1 is a schematic block diagram of a memory system which includes memory which is useful for understanding the present invention;

FIG. 2 illustrates a simple two-dimensional array of samples which can be stored in the memory shown in FIG. 1;

FIG. 3 shows how samples shown in FIG. 2 are stored at linearly increasing addresses;

FIG. 4 illustrates how samples can be accessed along a first dimension using the full width of a data bus;

FIG. 5 illustrates how samples can be accessed along a second dimension which does not use the full width of the data bus;

FIG. 6 is schematic block diagram of a memory system which includes memory and a memory access unit in accordance with the present invention;

FIG. 6a illustrates configuration registers;

FIG. 7 illustrates a simple two-dimensional array of samples which can be stored in the memory shown in FIG. 6;

FIG. 8 illustrates how, using the memory access unit shown in FIG. 6, samples can be accessed along a first dimension using the full width of a data bus;

FIG. 9 illustrates how, using the memory access unit shown in FIG. 6, samples can be accessed along a second dimension using the full width of a data bus; Figure to is a process flow diagram of a method performed by the memory access unit shown in FIG. 6;

FIG. 11 illustrates a three-dimensional array of samples;

FIG. 12 is a schematic block diagram of a radar sample processing pipeline;

FIG. 13a shows a first set of parameters for a configuration register and a first set of addresses and chip selects in a memory system in which the sample data width is the same as the memory data width;

FIG. 13b shows contents of first, second, third and fourth sample storages;

FIGS. 13c, 13d and 13e illustrate operation of first, second, third and fourth sample collector when accessing first, second, third and fourth sample storages shown in FIG. 13b;

FIG. 14a shows a second set of parameters for a configuration register and a second set of addresses and chip selects in a memory system in which the sample data width is double the memory data width;

FIG. 14b shows contents of first, second, third and fourth sample storages;

FIGS. 14c, 14d and 14e illustrate operation of first, second, third and fourth sample collector when accessing first, second, third and fourth sample storages shown in FIG. 14b;

FIG. 15 shows a bus address;

FIG. 16 shows dimensional addresses for a range address map in a case where the sample width is the same as the memory data width;

FIG. 17 shows dimensional addresses for a range address map in a case where the sample data width is double as the memory data width;

FIG. 18 shows dimensional addresses for a pulse address map in a case where the sample data width is the same as the memory data width;

FIG. 19 shows dimensional addresses for a pulse address map in a case where the sample data width is double as the memory data width;

FIG. 20 shows dimensional addresses for a channel address map in a case where the sample data width is the same as the memory data width;

FIG. 21 shows dimensional addresses for a channel address map in a case where the sample data width is double as the memory data width;

FIG. 22 is schematic block diagram of a memory system which includes memory and an array of memory access units in accordance with the present invention;

FIG. 22a illustrates a configuration register;

FIG. 23 is a schematic block diagram of an address calculator;

FIG. 24 is a schematic block diagram of a sample collector;

FIG. 25 is a schematic block diagram of a sample storage module;

FIG. 26 is a schematic block diagram of a bus interface;

FIG. 27 is a bus timing diagram; and

FIG. 28 illustrates a motor vehicle comprising a memory system.

DETAILED DESCRIPTION OF CERTAIN EMBODIMENTS

In the following description, all constants, variables and registers are of type integer unless otherwise specified.

Memory System 1

FIG. 1 is a schematic block diagram of a memory system 1 which is useful for understanding the present invention.

Referring to FIG. 1, the memory system 1 includes memory 2 in the form of an array of n random access memory (RAM) modules 2₁, . . . , 2_n (where n is an integer greater than one). The memory modules 2₁, . . . , 2_n may take the form of memory macros. The memory system 1 also includes bus wiring 3 and a memory access circuit 4 interconnecting the memory 2 and a bus 5 which consists of an address bus 6 and a data bus 7.

The memory access circuit 4 includes an address calculator 8 (which may also be referred to as an “address decoder”), a data multiplexer 9 and a bus interface 10.

Each memory module 2₁, . . . , 2_n has a memory data width w_m. In this example, the memory data width w_m is 128 bits. In this example, the data bus 7 is 128 bits wide.

The address calculator 8 selects one RAM module 2₁, . . . , 2_n from n RAM modules 2₁, . . . , 2_n using a chip select signal CS and specifies a part of memory 2 to be accessed using a memory address A_M via chip select and memory address buses 11. The address calculator 8 calculates chip select CS and memory address A_M using a bus address A_B received on the address bus 6 via the bus interface 10.

In a write transfer, bus data D_B are placed on memory data bus 12₁, . . . , 12_k. The data are stored in one of the RAM modules 2₁, . . . , 2_n upon selection using chip select CS at a respective address A_M. The bus interface can request only part of the bus data D_B to be stored.

In a read transfer, a selected RAM module 2₁, . . . , 2_n can transfer memory data D_M to the bus interface 10. The bus interface to can request only part of memory data D_M be transferred.

Referring also to FIGS. 2 to 5, data is transferred in fixed-sized units 13 (herein referred to as “words” or “samples”). Each memory 2₁, . . . , 2_n has a width w_m which is an integer multiple of the sample size. In this case, width w_m is 128 bits and each sample 13 is four bytes long and contains 32 bits. Thus, if the full bus width is to be used, then samples 13 should be stored in (or written to) adjacent positions across memory.

FIG. 2 schematically shows a simple two-dimensional picture (or “image”) 14 which comprises 4×4 samples 13 arranged in a two-dimensional array having x and y dimensions.

FIG. 3 shows how the picture 14 (FIG. 2) can be stored in memory 2, i.e. in first and second memory modules 2₁, 2₂ (FIG. 2), addressable with chip selects CS equal to 0 and 1 respectively, at linearly-increasing addresses A_M.

The picture 14 (FIG. 2) can be processed in either dimension, that is, along the x dimension or along the y dimension. In this example, samples 13 can be transferred over a data bus 7 (FIG. 1) that can carry up to four samples 13 in each cycle.

FIGS. 4 and 5 illustrate bus transfer performance when data is transferred by accessing samples 13 along the x dimension and the y dimension respectively.

As shown in FIG. 4, if samples 13 are read out by accessing the picture 14 along the x dimension, then the full width of the data bus 7 can be utilized. However, as shown in FIG. 5, if samples 13 are read out by accessing the picture 14 along the y dimension, then only one sample 13 can be transferred at a time, i.e. per cycle.

Thus, when samples 13 are read out along the y dimension, performance is degraded by a factor of four.

Performance can be degraded even further if data transfer occurs in bursts. Bus systems can have long latencies. If an access request is made, then it may be necessary to wait a several cycles before data is transferred. Hence, modern bus systems transfer multiple data words in sequence (called “burst”) with each request. However, bursts require that the address for each data word increases linearly without jumps. Therefore, they cannot be used to transfer the samples in y direction, reducing the performance even more.

If more than one sample is transferred in a bus word, then the full width of a bus cannot be used unless the sample order matches the storage order. However, often there is a need to read out samples along more than one dimension, for example, in image processing and radar applications, and other applications which involve multi-dimensional data access or transpositions, such as in multi-dimensional fast Fourier transform (FFT) processing. Some applications, such as radar, are automotive applications. In accordance with the present invention, memory access units and memory access methods can allow samples to be read out along different dimensions without the need for sample re-ordering and/or which can help improve bus utilization.

Memory System 21

FIG. 6 is a schematic block diagram of a reconfigurable multi-dimensionally accessible memory system 21 in accordance with the present invention.

Referring to FIG. 6, the memory system 21 includes memory 22 in the form of an array of k sample storage modules 22₁, . . . , 22_k (where k is an integer greater than one). Each sample storage module 22₁, . . . , 22_k takes the form of a RAM macro or other RAM module. The memory system 21 includes bus wiring 23 and a memory access unit 24 interconnecting the memory 22 and a bus 25 which consists of an address bus 26 and a data bus 27.

The memory access unit 24 includes an array of k address calculators modules 28₁, . . . , 28_k, an array of k sample collector modules 29₁, . . . , 29_k, a common bus interface 30 and a set of configuration registers 31.

Each memory module 22₁, . . . , 22_n has a memory data width w_m. The memory data width w_m and the minimum sample width are the same. In this example, the data bus 27 is 128 bits wide. However, the data bus 27 can be narrower or wider, for example, 256 bits wide.

Referring also to FIG. 6a, the configuration registers 31 include a set of registers 31₁, 31₂, . . . , 31_d (where d is a positive integer more than one) specifying the number of samples S_i which can be stored in each dimension i, where i={1, 2, . . . , d}, and a register 31_s specifying sample data width w_s. The memory data width w_m is equal to or an integer multiple of the sample data width w_s. The configuration registers 31 are set by a user before the memory system 21 is used.

Referring again to FIG. 6, an address calculator module 28₁, . . . , 28_k receives the bus address A_B, adds a respective offset corresponding to the position Nr of the sample in the bus data word to obtain an bus address A_B′ for a respective sample and converts the sample bus address A_B′ into a set of sample addresses A_1, . . . , A_d and a sample index I_S according to an address decoding scheme. The bus address A_B′ for a sample is calculated using:

A_B′=A_B+(Nr−1)*w_m  (1)

where Nr=1, 2, . . . , k and memory data width w_m.

An example of an address decoding scheme for three-dimensional accesses in radar signal processing is described hereinafter. However, any suitable address decoding scheme can be used. The dimension address A_L is the linear address along the access dimension.

FIG. 7 schematically shows a simple two-dimensional picture (or “image”) 33 which comprises 4×4 samples 34 arranged in a two-dimensional array having x and y dimensions.

Referring to FIGS. 7 and 8, an i-th sample collector 29₁ (where i={1, 2, . . . , k}) is responsible for collecting a sample 34 at a fixed bit position [w_m*8*i−1:w_m*8*(i−1)] in the bus data word.

A sample collector 29₁, . . . , 29_k identifies a sample storage 22₁, . . . , 22_k in which a requested sample 33 is stored or is to be stored, selects the identified sample storage 22₁, . . . , 22_k using a select signal CS, requests transfer (read or write) of the sample 34 at the address A_M and fills or retrieves the sample 34 at the assigned position in the bus data word D_B.

A dedicated sample collector 29₁, . . . , 29_k exists for each sample 34 in the bus data word D_B. As will be explained in more detail later, the sample collectors 29₁, . . . , 29_k perform memory access in parallel using a chip select arrangement which ensures that there are no access conflicts in sample storage 22₁, . . . , 22_k. Thus, the full bus data width can be utilized without introducing wait states.

FIGS. 8 and 9 illustrate bus transfer performance when data is transferred by accessing samples 34 along x and y dimensions respectively.

As shown in FIG. 8, if samples 34 are read out by accessing a picture 33 along the x dimension, then the full width (in this case, 128 bits) of the data bus 27 can be utilized.

And, as shown in FIG. 9, if samples 34 are read out by accessing a picture 33 along the y dimension, then the full width of the data bus 27 can be employed.

Accesses along the x and y dimensions can utilize the full bus width, in this case, four samples per bus data word. Furthermore, bursts are possible since A_L linearly increases, without jumps. This can further improve performance. For example, if the bus 25 is an Advanced eXtensible Interface (AXI) bus with ten pipeline stages, a single access requires ten cycles per word. A 16-beat burst requires ten cycles for the initial word and fifteen cycles for the remaining words. On average, one word requires 25/16=1.6 cycles per word. This is another speed up by an additional factor of six.

Thus, compared with the memory system 1 shown in FIG. 1 in which samples are ordered along the memory address, the memory system 21 can transfer data along the y dimension at a rate which is 24 times faster.

Referring again to FIG. 6 and Figure to, when an access request is received (step S1), an address calculator module 28₁, . . . , 28_k converts the bus address A_B into a representation in the multi-dimensional data array (step S2).

Each sample 34 is addressed within the array by its dimension address A_1, A_2, . . . , A_d and sample index I_S. The sample index I_S is used if the sample width w_s is greater than the sample storage width w_m.

Since the bus data word D_B contains more than one sample, a bus access field indicates along which dimension access is requested. Several bus address calculation schemes are possible which identify encoding access direction, linear address A_L and index.

Each sample collector 29₁, . . . , 29_k computes (steps S3 to S6) using equations 2 to 5 below:

• • a sample address A_S using:

$\begin{array}{cc}\mathrm{A\_S}=\left(\mathrm{A\_d★}\left(\mathrm{S\_d}-1\mathrm{★S\_d}-2★\dots \mathrm{★S\_}1\right)+\mathrm{A\_d}-1★\left(\mathrm{S\_d}-2\mathrm{★S\_d}-3★\dots \mathrm{★S\_}1\right)+\dots +\mathrm{A\_}1\right)★\left(\mathrm{W\_s}/\mathrm{W\_m}\right)& \left(2\right)\end{array}$

• • a physical word address A_P using:

A_P=A_S+I_S  (3)

• • a selected memory CS using:

CS=[I_S+(w_s/w_m)*(A_1+A_2+ . . . +A_d)]%k  (4)

• • an address in memory A_M using:

A_M=A_P/k  (5)

where + is addition, − is subtraction, / is integer division (fractions are discarded), % is remainder of integer division (i.e. modulo) and * is integer multiplication.

Each sample collector 29₁, . . . , 29_k accesses a respective sample storage 22₁, . . . , 22_k (step S7). The bus interface answers the transfer request (step S8).

Selected memory CS is different for adjacent samples in every dimension. If accesses occur along dimension j (j={1, . . . , k}), only the sample address A_j changes in equation 4. The other sample addresses A_i (i={1, . . . , k}, i< >j) remain constant. However, for the term which changes, the term is different for adjacent addresses in the accessed dimension. Thus, parallel accesses to sample storages are possible.

There is, however, an exception to this arrangement, namely when a transaction crosses an access dimension boundary. In that case, more than one term changes in the selected memory CS. Depending on the chosen sizes, wait states may be necessary. No wait states occur if each dimension size S_1, S_2, . . . , S_d is a multiple of w_m*k/w_s. A memory access system which has multiple bus interfaces and which uses wait states is described hereinafter.

Application of the Memory System 21 in a Radar Application

Referring to FIG. 11, in radar applications, multiple time responses from an antenna vector (not shown) are stored in memory resulting in a three-dimensional array 35 (herein also referred to as a “data cube” or simply “cube”) of samples 36. The array 35 may have, for example, 512 samples along a range dimension, 64 samples along a Doppler dimension and 8 samples along a channel dimension.

In radar and other applications, a change of dimension requires re-ordering samples in memory (referred to as performing “corner turns”). A large amount of processing time can be spent on the corner turns.

Referring to FIG. 12, the radar pulse processing system 41 can be implemented using a pipeline in which samples 36 are stored, as a data cube 35 (FIG. 11), in memory 22. Samples 36 are fed into a pulse compression block 50 and the output is stored in memory 22. Samples 36 are then read out and fed into a Doppler filtering block 51 and the output is stored in memory 22. is the samples are read out and fed, together with weights from a previous data cube, into a space-time adaptive processing section 52 comprising a beamform block 53 and an adaptive weight computation block 54 which generates weights which are saved for the next cube, and are stored. The transformed samples 36 are read out and fed to a constant false alarm rate (CFAR) detection block 55 which outputs a detection report 56.

The multi-dimensional memory system 21 can be used to store and then read out the right concatenation of the samples 35 in the bus response for the given dimension. Thus, data reorganization can be avoided.

Simple Examples Showing Address Calculation and Sample Storage Selection

Referring to FIG. 13a, a set of parameters 61 for the configuration register and global assignment of address and chip selects 62 for a simple example are shown in which the sample data width w_s is the same as the memory data width w_m.

The maximum number of samples per bus word k is four, the total number of dimensions d is three, the memory data width w_m is four bytes and the sample data width w_s is four bytes. There are four, five and two samples in first, second and third dimensions respectively S_1, S_2, S_3.

As shown in FIG. 13a, values of dimension address A_1, A_2, A_3, sample address A_S, sample index I_S, physical word address A_P, sample storage select CS and address in memory A_M are tabulated.

Referring to FIG. 13b, contents of first, second, third and fourth sample storages 22₁, 22₂, 22₃, 22₄ are shown. Sample addresses in the array 35 are shown in round brackets and the sample indexes are shown in square brackets.

FIGS. 13c, 13d and 13e illustrate operation of the sample collector 29₁, 29₂, 29₃, 29₄ when accessing the sample storages 22₁, 22₂, 22₃, 22₄ along each of the three dimensions. FIGS. 13c, 13d and 13e show linear access address A_L, dimension address A_1, A_2, A_3, sample index I_S, physical address A_P, sample storage select CS and memory address A_M accesses for each sample collector 29₁, 29₂, 29₃, 29₄.

FIG. 13c illustrates accessing the sample storages 221, 22₂, 22₃, 22₄ along the first dimension, i.e. range samples. In this case, no wait states are needed because only one dimensional address changes at a time.

FIG. 13d illustrates accessing the sample storages 22₁, 22₂, 22₃, 22₄ along the second dimension, i.e. pulse samples. In this case, no wait states are needed because, even though more than one dimensional address changes at a time, there are no CS conflicts.

FIG. 13e illustrates accessing the sample storages 22₁, 22₂, 22₃, 22₄ along the third dimension, i.e. channel samples. In this case, a wait state is needed.

Referring to FIG. 14a, a set of parameters 61 for the configuration register and global assignment of address and chip selects 62 for a simple example are shown in which the sample data width w_s is double the memory data width w_m.

FIG. 14b shows the contents of first, second, third and fourth sample storages 22₁, 22₂, 22₃, 22₄ and FIGS. 14c, 14d and 14e illustrate operation of the sample collector 29₁, 29₂, 29₃, 29₄ when accessing the sample storages 22₁, 22₂, 22₃, 22₄ along each of the three dimensions.

Bus Address Decoding Scheme for Radar Devices

Referring again to FIG. 6, the address calculators 29₁, . . . 28_k employ an address decoding scheme to generate dimensional address A_1, . . . , A_d from bus address A_B.

An address decoding scheme will now be described with reference to FIG. 15.

FIG. 15 shows a bus address A_B′ 63 (i.e. the bus address A_B after position adjustment for a given sample) comprising a base field 64, a dimension access mode (DIM) field 65 and a linear address A_L 66. As fixed parameters, the memory data width w_m is four bytes and the number of dimensions is three.

Range Addressing Mode

Range addressing mode can be selected by setting DIM to 0, i.e. 2b00. The first, second and third dimensional addresses A_1, A_2 and A3 are calculated as follows:

A_1=(A_L/(w_s))%S_1  (6-R-1)

A_2=(A_L/(S_1*w_s))%S_2  (6-R-2)

A_3=(A_L/(S_2*S_1*w_s))%S_3  (6-R-3)

Range addressing mode can also be used to access the memory in a traditional way by setting the sample width to be the same as memory width, i.e. w_s=w_m, and the dimension size S_i, . . . , S_d to be the maximum physical memory size.

Pulse Addressing Mode

Pulse addressing mode can be selected by setting DIM to 1, i.e. 2b01. The first, second and third dimensional addresses A_1, A_2 and A3 are calculated as follows:

A_1=(A_L/(S_2*w_s)%S_1  (6-P-1)

A_2=(A_L/(w_s))%S_2  (6-P-2)

A_3=(A_L/(S_2*S_1*w_s))%S_3  (6-P-3)

Channel Addressing Mode

Channel addressing mode can be selected by setting DIM to 2, i.e. 2b10. The first, second and third dimensional addresses A_1, A_2 and A3 are calculated as follows:

A_1=(A_L/(S_3*w_s)%S_1  (6-C-1)

A_2=(A_L/(S_3*S_1*w_s))%S_2  (6-C-2)

A_3=(A_L/(w_s))%S_3  (6-C-3)

Sample Index

The sample index I_S is calculated using:

I_S=A_L/w_m%(w_s/w_m)  (7)

FIGS. 16 to 21 show examples of dimensional addresses A_1, A_2 and A_3 and physical address A_P calculated using equations 2, 3, 6-R-1, 6-R-2, 6-R-3, 6-P-1, 6-P-2, 6-P-3, 6-C-1, 6-C-2, 6-C-3 and 7.

Referring to FIGS. 16 and 17, dimensional addresses A_1, A_2 and A_3 are shown for a range address map calculated using equations 6-R-1, 6-R-2, 6-R-3 when S_1=4, S_2=2, S_3=3.

Referring to FIGS. 18 and 19, values of dimensional addresses A_1, A_2 and A_3 are shown for a pulse address map calculated using equations 6-P-1, 6-P-2, 6-P-3 when S_1=4, S_2=2, S_3=3.

Referring to FIGS. 20 and 21, values of dimensional addresses A_1, A_2 and A_3 are shown for a channel address map calculated using equations 6-C-1, 6-C-2, 6-C-3 when S_1=4, S_2=2, S_3=3

Memory System 71

Referring to FIG. 6, the memory system 21 includes an interface with only one bus. However, memory can be accessed by m buses (where m is a positive integer greater than one) by providing a ready (RDY) feedback from the sample storage to the sample collector. For example, there may be sufficient samples in memory to require a plurality of digital signal processors (DSPs) to process the samples. Memory may be divided into a plurality of portions, with each DSP processing a respective portion of memory. Each DSP uses a respective bus to access the memory. This allows the DSPs to process data in parallel. For example, eight DSPs can be used, thereby requiring eight buses (i.e. m=8). Each bus is provided with a bus master (not shown).

An RDY feedback can also be used to handle conflicts in CS selection. If more than one sample collector accesses the same sample storage, then conflict can be resolved by an arbiter.

To support m bus interfaces, there are an equivalent number of memory segments. In the best case, m masters can accesses each segment without conflicts.

Referring to FIG. 22, the memory system 71 includes memory 72 in the form of an array of (k*m) sample storage modules 72₁, . . . , 72_k*m (where k is an integer greater than two). The memory system 71 includes bus wiring 73 and m memory access units 74₁, . . . , 74_m interconnecting the memory 72 and m buses 75₁, . . . , 75_m each of which consists of an address bus 76, a data bus 77 and RDY signal 78.

Each memory access unit 74₁, . . . , 74_m includes an array of k address calculators modules 79₁, . . . , 79_k, an array of k sample collector modules 80₁, . . . , 80_k and a common bus interface 81. The memory system 71 is provided with a set of configuration registers 82.

Each memory module 72₁, . . . , 72_n has a memory data width w_m. In this example, memory data width w_m is 32 bits and each data bus 77 is 128 bits wide.

Referring also to FIG. 22a, the configuration registers 82 include a set of registers 82₁, 82₂, . . . , 81_d (where d is a positive integer more than one) specifying the number of samples S_i where i={1, 2, . . . , d} and a register 82_s specifying sample data width w_s. The sample data width w_s equals or is an integer multiple of memory data width w_m. The configuration registers 82 are set by a user before the memory system 71 is used.

Referring to FIG. 23, an address calculator 28₁, . . . 28_k, 79₁, . . . , 79_k is shown in more detail.

The address calculator 28₁, . . . 28_k, 79₁, . . . , 79_k comprises an adder unit 28a, 79a and address calculator arithmetic logic unit 28b, 79b.

The adder unit 28a, 79a adjusts the bus address A_B to match the position Nr=1, 2, . . . , k of the sample in the bus data word using equation (1) above.

The address calculator arithmetic logic unit 28b, 79b converts the sample bus address A_B′ into the dimensional address A_1, . . . , A_d and the sample index I_S. Different address decoding schemes are possible.

Referring to FIG. 24, a sample collector 29₁, . . . , 29_k, 80₁, . . . , 80_k is shown in more detail.

The sample collector 29₁, . . . , 29_k, 80₁, . . . , 80_k comprises sample calculator arithmetic logic unit 80a and a multiplexer 80b.

To support parallel accesses without conflicts from m bus interfaces, the sample storage select signal CS partitions memory into m segments, where m is a positive integer greater than one. Each segment contains S_M words.

Within the segment, sample collector 80₁, . . . , 80_k calculates sample storage select signal CS(MULTIPLE BUS) (hereinafter also referred to simply as CS) using:

CS(MULTIPLE BUS)=(A_P/S_M)*k+CS(SINGLE BUS)  (4′)

where CS(SINGLE) can be calculated using equation 4 above.

The sample collector 80₁, . . . , 80_k calculates address in memory A_M using equation 4′ below:

A_M=(A_P%S_M)/k  (5′)

D_B is bi-directional depending on transfer direction (read/write).

RDY indicates that a bus must wait where there are multiple masters and/or multiple sample collectors access the same RAM module.

Referring to FIG. 25, a sample storage module 22₁, . . . , 22_k*m, 72₁, . . . , 72_k*m is shown in more detail.

The sample storage module 22₁, . . . , 22_k*m, 72₁, . . . , 72_k*m comprises a RAM macro 22a, 72a, 35a multiplexer 22b, 72b, an arbiter 22c, 72c and comparators 22d(1), . . . , 22d(k*m), 72d(1), . . . , 72d(k*m).

Each sample storage 72₁, . . . , 72_k*m reacts only to CS request that match its number Nr, where Nr={1, . . . , k*m}. The arbiter 72c selects one request from all active CS requests by a suitable policy, for example round-robin. All other sample collectors 80₁, . . . , 80_k are paused using the RDY signal.

RAM 72a handles the data word transfer depending on the direction (read or write).

Referring to FIG. 26, a bus interface 81 is shown in more detail.

The bus interface 81 includes a concatenate block 81a and k-input AND gate 81b.

Each sample collector 80₁, . . . , 80_k handles a fixed sample position in the bus word. The bus interface 81 concatenate data D_M from each sample collector 80₁, . . . , 80_k for output as bus data, namely:

D_B={D_M(k), . . . ,D_M(1)}  (7)

The k-input AND gate 81b receives RDY(1), . . . , RDY(k) from the sample collectors 80₁, . . . , 80_k and outputs an RDY signal.

Referring to FIG. 27, when address calculations are timing-critical, they can be pipelined provided burst-property is maintained.

FIG. 27 shows halting an AHB bus only for one cycle at the beginning of a burst to compute C_S and A_M.

Referring to FIG. 28, a motor vehicle 91 is shown.

The motor vehicle 91 includes an advanced driver assistance system (ADAS) 92 which includes sensors (not shown) and one or more memory systems 21, 71.

It will be appreciated that many modifications may be made to the embodiments hereinbefore described.

Claims

1-15. (canceled)

16. A memory access unit for handling transfers of samples in a d-dimensional array between a one of m data buses, where m≧1, and k*m memories, where k≧2, the memory access unit comprising: k address calculators, each address calculator configured to receive a bus address, to add a respective offset to generate a sample bus address and to generate, from the sample bus address according to an addressing scheme, a respective address in each of the d dimensions for accessing a sample; andk sample collectors, each sample collector operable to generate a memory select for a one of the k*m memories so as to transfer the sample between a predetermined position in a bus data word and the respective one of the k*m memories,wherein each sample collector is configured to calculate a respective memory select in dependence upon the address in each of the d dimensions such that each sample collector selects a different one of the k*m memories so as to allow the sample collectors to access k of the k*m memories concurrently.

17. A memory access unit according to claim 16, further comprising: a set of registers for changeably setting the numbers of samples in each of the d dimensions and/or sample width.

18. A memory access unit according to claim 16, wherein the number d of dimensions is set or is settable to be two or three.

19. A memory access unit according to claim 16, wherein each sample collector is configured to calculate the memory select in dependence upon a sum of the addresses in each of the d dimensions.

20. A memory access unit according to claim 16, wherein each address calculator is configured to generate an index.

21. A memory access unit according to claim 16, wherein each address calculator is configured to calculate the respective addresses in dependence upon a linearly-increasing word address, sample size and numbers of samples in each dimension.

22. A memory access unit according to claim 16, further comprising: a bus interface coupled to the address calculators and the sample collectors.

23. A memory access unit according to claim 16, implemented in hardware logic.

24. A memory controller comprising: at least one memory access unit according to claim 16.

25. A memory controller according to claim 24, comprising: m memory access units for handling transfers of samples in a d-dimensional array between m data buses and k*m memories.

26. A memory system comprising: a memory controller according to claim 24; andk*m memories, each set of k*m memories operatively connected to the m memory access units.

27. An integrated circuit comprising: a memory access unit according to claim 16.

28. An integrated circuit according to claim 27, which is a microcontroller.

29. A motor vehicle comprising: a computing device comprising a memory access unit according to claim 16.

30. A method of transferring samples in a d-dimensional array between a one of m data buses, where m≧1, and k*m memories, where k≧2, the method comprising: for each of k samples:receiving a bus address and adding a respective offset to generate a sample bus address;generating, from a sample bus address according to an addressing scheme, a respective address in each of the d dimensions for accessing samples along one of the dimensions; andgenerating a memory select for a one of the k*m memories so as to transfer a sample between a predetermined position in a bus data word and the respective one of the k*m memories, wherein generating the memory select comprises calculating a memory select in dependence upon the addresses,

31. An integrated circuit comprising: a memory controller according to claim 24.

32. An integrated circuit according to claim 31, which is a microcontroller.

33. A motor vehicle comprising: a computing device comprising a memory controller according to claim 24.