WO2009101119A1

WO2009101119A1 - A processor

Info

Publication number: WO2009101119A1
Application number: PCT/EP2009/051598
Authority: WO
Inventors: David Maloney
Original assignee: Linear Algebra Technologies
Priority date: 2008-02-11
Filing date: 2009-02-11
Publication date: 2009-08-20
Also published as: EP2260380A1; CN101971139A; US9858073B2; GB0802484D0; US20110047360A1; US20180189066A1; JP2011511986A; GB2457303A; KR20100122493A; US10719318B2

Abstract

The present application provides a method of randomly accessing a compressed structure in memory without the need for retrieving and decompressing the entire compressed structure.

Description

Title

A processor Field of the Application

The present application relates to the field of processors and to methods of transferring data between memory and the processor. More particularly, the present application is directed to a method of accessing an individual value, or contiguous or non-contiguous group of values within a compressed data-structure in memory without the need for retrieving and decompressing the entire compressed structure.

Background Of The Application

There is a fundamental problem in the design of computing systems, namely that of minimising the time cost of memory accesses.

This is a fundamental limitation on the design of computer systems as no matter what memory technology is used to support computation and no matter what technology is used to connect that memory to the processor, there is a maximum limitation on how much information can be transferred between processor and memory in a given time, this is the available memory bandwidth and the limitation of computer power by available memory bandwidth is often referred to as the "memory- wall".

The present application seeks to increase the effective memory bandwidth and thus minimise the limitation of the "memory-wall" through the use of data compression.

It is known to employ data compression to reduce the effects of the "memory wall". However, a problem for programmers using compressed memory sub-systems is that data has to be decompressed before it can be operated upon as shown in the system of Figure 1. This usually involves reading the compressed data from one part of memory 10 into the register files of the processor 16, decompressing it using program code retrieved from program memory 18 and storing the decompressed data in another uncompressed portion of memory 12.

However this solution has the disadvantage that additional memory bandwidth is required to read compressed data, store it in uncompressed form, and read it back into the processor to be operated upon. Additional memory capacity is also required to hold the uncompressed data and the decompression process will increase pressure on the processors register-files. Clearly this is a sub-optimal solution which it is suggested explains why such compressed memory sub-systems have remained an academic curiosity rather than entering the mainstream microprocessor industry. EP-0240032-A2 discloses a vector processor comprises a memory for storing and retrieving vector data. The vector processor comprises a plurality of vector registers each capable of reading or writing plural (m) vector elements in parallel, at least one mask vector register capable of m mask bits in parallel, transfer portion connected to the memory, the plurality of vector registers and the mask vector register and responsive to an instruction for transferring vector elements from regularly spaced address locations within the memory to selected storage locations of a selected vector register corresponding to valid mask bits. Whilst this approach is useful, it is limited in that the storage\retheval of vector data is limited to an entire register. Register-blocking is a useful technique for accelerating matrix algebra (particularly Finite-Element), however it has the disadvantage in that for many matrices (ex. As used in search engines such as GOOGLETM) zero fill has to be added decreasing effective FLOPS (Floating Point Operations Per Second), and increasing memory bandwidth requirements, both of which are commodities which are in short supply in modern computing systems.

In fact the growing gap between processing capabilities and memory bandwidth which are increasing at highly disparate rates of 50% and 7% per annum respectively is referred to, as mentioned above, as the "Memory Wall". There have been many claims of "breaking" the memory wall and they usually consist of using a cache to reduce the probability of having to go off-chip, and/or using multi-threading so that the latency and penalties associated with going off-chip can be mitigated.

These approaches merely hide the problem of limited external memory bandwidth rather than solving it and generally rely on the data-set exhibiting sufficient data locality, and/or the program exhibiting sufficient Thread-Level Parallelism (TLP) in order to be effective at all, and this may not be true of all problems, and is certainly not always known a priori. In fact many of the larger and more interesting problems exhibit neither sufficient data-locality, nor sufficient TLP and the throughput of the whole system degenerates to the point where it is limited by external memory bandwidth, and the extra hardware which has been added on-chip is of no use. For this reason it is not uncommon to see large engineering applications pulling down processor performance to 1 % or less of the manufacturers quoted peak performance specification.

Summary

The present application seeks to increase the effective memory bandwidth and minimise the limitation of the "memory-wall" on computation by storing data in a compressed format. This is achieved by providing a means of compression and decompression which is suitable for block-structured data used in many applications including, for example, computer graphics, rigid-body dynamics, finite-element analysis and other scientific and engineering applications, which operate on large data sets which must be stored in memory. In order to further reduce the effect of the "memory-wall", the processor pipeline is also modified in such a way as to take advantage of compression, increasing the processing rate beyond what can be achieved by operating on compressed data alone. More particularly, the present application allows random-access to individual values within a compressed structure stored in memory through the use of a value map and associated field bits specified as part of an instruction.

Accordingly, a first embodiment of the application provides a processor as detailed in claims 1 and 38. The application also provides a method as detailed in claims 21 and 56. The application further provides a data processing machine as detailed in claim 35 and 69. Advantageous embodiments are provided in the dependent claims.

Brief Description Of The Drawings

The present application will now be described with reference to the accompanying drawings in which:

Figure 1 is a processor arrangement known in the art,

Figure 2 is a series of transformations for which the processor of the present application is suitable for processing,

Figure 3 is an exemplary processor according to an embodiment of this application, Figure 4 illustrates a exemplary compression method that may be employed by the processor of Figure 3,

Figure 5 illustrates how a memory structure can be compressed to fit in 32 bit RAM by the processor of Figure 3,

Figure 6 illustrates the format of exemplary instruction codes that may be employed by the processor of Figure 3,

Figure 7 illustrates how the same memory structure of Figure 5 can be compressed to fit in 64 bit RAM by the processor of Figure 3, Figure 8 is an exemplary circuit for generating addresses which might be employed by the processor of Figure 3, Figure 9 demonstrates the operation of

Figure 8 with exemplary values,

Figure 10 demonstrates exemplary load\store circuit that might be employed in the processor of Figure 3, Figure 11 is an illustration of address sequence generation for use in

Figure 10, Figure 12 is an exemplary timing diagram illustrating the mode of operation of the circuit of Figure 10

Figure 13 is a more detailed block diagram of the processor of Figure 3, Figure 14 is an instruction format,

Figure 15 is a table showing various opcode instructions available in an exemplary processor including instructions for loading\stohng compressed data,

Figure 16 is a circuit for address generation for a non compression based instruction, Figure 17 is circuit for generating a bit as employed by the compression method of the processor, Figure 18 illustrates the computational savings that may provided by the exemplary processor of Figure 3 for different types of matrix structure,

Figure 19 is a circuit suitable for the processor of Figure 2 for loading\storing compressed data,

Figure 20 illustrates the effects of port width on memory organisation,

Figure 21 illustrates the resulting effect on address generation for the addresses shown in Figure 20, and

Figure 22 is exemplary timing diagrams illustrating the effects of Figure 20.

Detailed Description Of The Drawings

A typical example of the desirability of compression is the use of matrix representations and linear algebra operators to simulate reality on a 2-dimensional screen in computer graphics and related applications. Operations on a source data matrix often consist of rotations and other transformations, and often sequences of them, of the type shown in Figure 2. As can be seen these matrices contain a large number of zeroes which require a large amount of memory bandwidth to load/store while not contributing to the result of the calculations performed using the whole matrix, i.e. they are trivial values from the point of view of arithmetic.

The present application, shown in Figure 3, provides a modified processor which is capable of operating directly upon compressed data in memory without the requirement for decompression processing, thus eliminating the requirement for an additional uncompressed data buffer and additional processor and memory bandwidth required to handle decompression of compressed data into a buffer area for further processing. Instead, the compression and decompression processes are integrated in the memory access structures of the processor. The application allows compressed structures of arbitrary size and complexity, consisting, for example but not limited to, any combination of the following: floatingpoint matrices, floating-point vectors, floating-point scalars, signed integers, unsigned integers, packed characters and address-pointers. Suitably, the compression and decompression logic is embedded within the processor load/store ports, making the programmer's job easier in terms of not having to worry about the details of how data is read/written from/to the compressed memory subsystem on a transaction by transaction basis.

The solution described herein employs a value map to code for zero and nonzero entries in dense sub-matrix in the manner shown in Figure 4. The use of such a value map has been described previously in related applications (US60/911 ,273 filed 11 th April 2007, US60/911 ,274 filed 11 th April 2007 and UK 0704999.2 and UK0704976.0 and corresponding PCT applications claiming priority therefrom) the entire contents of which are incorporated herein by reference. In the example shown a compression of 37.5% is achieved. The proposed solution has the advantage that each zero entry in the matrix to be stored is represented by a 1 -bit rather than a 32- bit single15 precision, 64-bit double-precision, etc. number. Overall compression is achieved as long as 1 or more entries per dense-matrix are zero in a 32-bit memory system storing a sparse matrix with 32-bit single-precision entries.

While the basic compression method outlined above is interesting and offers advantages, in most practical applications the programmer desires to store more than just sparse matrices and often needs to store and retrieve data stored in complex data-structures. The challenge is how to integrate a compression scheme so that it can provide similar performance and memory compression benefits to those outlined previously when handling mixed compressed/uncompressed data- structures without overly burdening the programmer. The present application provides a means of random accessing data within a larger structure which may be compressed in the manner described previously, but may also contain uncompressed floating-point or integer data. The scheme allows a basic structure containing up to thirty-two 32-bit entries to be represented by a 32-bit integer (value map). In principle, the scheme is not limited to 32 but values or structures with 32 entries and the size of the value map and the number of entries may be selected to meet the requirements of the processor design. The presence of a 1 -bit in the value map indicates the presence of a non-trivial floating-point or integer value at that position in memory as before. As each value is represented by a single bit, the value map may be referred to also as a bit map. An example of how a mixed (compressed/uncompressed) structure would be stored in memory is shown in Figure 5.

The basic principle of operation of the proposed compressed structure decompression logic is that a value map is used to represent the presence of trivial values. For the purposes of this application a trivial value is a 0.0 single precision (32-bit) floating-point value (Hex 00000000), however the same scheme may easily be extended to double-precision floating-point or beyond. Alternately if there were a preponderance of some other value than 0.0 in a particular data-set the scheme outlined here could easily be extended to allow a value map 0 represent that value rather than 0.0, for instance some data sets might contain many 1.0 values.

As can be seen from the RAM addresses, for an exemplary arrangement of data, in Figure 5 it can be seen that for each non-zero entry (sx, sy, sz, xo -X3, yo - V3, zo - Z3) in the 32-bit value map an offset of 1 is added to the base address. In practice both the value map and the base address of the next structure in memory may be embedded in the current structure and retrieved by the programmer.

To make use of compressed structures it is beneficial to reduce the overhead for the programmer in storing and retrieving the data stored in compressed form. In the present application, the programming overhead is minimised by hiding the complication of storing and retrieving compressed data inside the Load-Store Unit (LSU) address-generators of the processor which may for example be a RISC processor. The address generators allow a programmer to specify which elements of data are to be retrieved from a compressed structure. In turn, the LSU is able to retrieve the requested elements from the compressed structure in memory without decompressing the entire compressed structure.

Thus in the exemplary arrangement described herein, the address generators allow the programmer to specify which of 8 possible 4-entry fields (vectors) from a 32-entry compressed structure are to be loaded/stored using a simple 8-bit field mask which is included in the load/store opcode as shown in Figure 6. In the exemplary arrangement described herein, the instruction for loading a compressed vector is LDCV, whereas the instruction for storing a compressed vector is STCV. This exemplary arrangement allows a programmer to specify the 4-entry vector register file (VRF) register to which data will be loaded/stored from/to. The base register in the integer register file (IRF) to which the addressgenerator will add the compressed-structure offset and the value map register also in the IRF. The opcode then contains an immediate 8-bit value f[7:0] which is used to select which of the eight four-entry compressed vectors actually get loaded/stored from/to the VRF.

As the LSU port-width is generally limited, in the case of exemplary arrangement in this application to 64-bits, it will be appreciated that the LDCV and STCV instructions necessarily take multiple clock-cycles. As a result the value map is decoded in an iterative fashion in the order of the f-bits causing a 64-bit compressed entry to be written to or read from the correct VRF register and stored to or read from the correct memory (RAM) address as seen from the LSU port. The layout in RAM of a typical compressed structure in a 64-bit memory subsystem is as shown in Figure 7.

It will be appreciated that the method and arrangement is not limited to 32 bit words or 64 bit wide memories and that similar layouts are possible for other memory widths including 96 and 128-bit wide memories etc, however for simplicity of explanation such arrangements organisations will not be discussed in detail here. The VRF may be used as a staging area for the complete compressed structure before parts of it are dispersed to other register files internal to the processor in the case of a compressed load, or alternately for the storing of a compressed structure to external RAM attached to the LSU port. An exemplary logic for the LSU will now be discussed with reference to Figure 8, which shows the logic for the LSU compressed-structure address generator for loading values from within a compressed structure at a given base address (base_addr). A first row of adders 50 (compressors) is provided to compress the value map structure of 32 bits arranged in eight 4-bit value map fields (bm7-bmθ) into eight 3-bit numbers, each of which represents the number of ones in a given 4- bit value map. A second row of adders 52 is provided to add the results of the row of compressors to generate the start address of each value map so that each value map field may be addressed independently if required by applying the appropriate f- bits in the compressed load or store opcode. A row of AND gates 54 is provided to mask out unrequired fields depending on the state of the f (field) bits set in the compressed load/store opcode. The outputs of these AND gates are then summed by a further set of adders 56 to generate the total number of 32-bit words (word_cnt) which must be loaded/stored from/to memory based on the opcode f-bits. A final 8:1 multiplexer stage 58 is controlled by an 8 field-enable which allows the required value map addresses selected to be applied sequentially to a final adder 60 in order to generate the required RAM addresses.

The values of the various signals for the example compressed data structure are shown in Figure 9. The addresses shown are calculated statically from the value map in the selected IRF (Integer Register File) register and are then applied sequentially depending on the stage of the f-bits specified in the compressed load or store instructions.

The complete address-generation logic for the compressed load/store operations including the counter and finite-state machine to generate the required control signals is shown in Figure 10.

The fen[2:0] signal which selects the sequence of addresses to be applied to the address selection multiplexer operates according to the truth-table shown in Figure 11.

As may be seen in Figure 12 not only are the source/destination addresses required for the compressed load/store operations generated by the LSU address-generator, but the source/destination addresses for the source/destination VRF rows relative to that specified in the compressed load/store operation are also generated. In the case of a 64-bit LSU bus up to 2 VRF rows may be addressed in a compressed load/store instruction. Although in the present exemplary arrangement, the VRF was chosen as the source/destination for data loaded or stored in compressed load/store operations, in principle the Integer Register File (IRF) or scalar register file (SRF) could equally have been chosen.

The exemplary timing diagram of Figure 12 shows clearly that the compressed load/store operation takes multiple cycles to execute and all of the required control signals are generated from knowledge of the IRF value map descriptor for the compressed structure coupled with the f-bits specified in the compressed load/store control word. The sequence of control signals is derived using the f-bits and value map in conjunction with a counter which is driven by the clock input to the address- generator.

The diagram shows behaviour for both the load and store operations, the only difference between the 2 operations would be:

1. In the case of a compressed load a read signal is generated to read from the

RAM attached to the LSU port and a write signal is generated for the appropriate VRF row which acts as a destination for the RAM contents. In fact a series of RAM reads and VRF writes may be generated depending on the state of the value map in the IRF and the f-bits set in the compressed load opcode. 2. Similarly in the case of a compressed store a series of VRF reads and RAM writes may be generated depending on the state of the IRF value map and the f-bits set in the compressed store opcode.

A block-diagram of the SVU architecture, of which the Load Store Unit (LSU) with compressed load/store operations forms a part, is shown in Figure 13. As can be seen the principal external interfaces are to the SVU instruction memory, and the LSU ports which connect to the data memory-banks as well as the debug control unit (DCU) which allows SVU registers and diagnostic information to be read/written via the processor bus.

The instruction word has 7 slots which allow it to control up to 7 of the 9 functional units per cycle as shown in Figure 14.

In the present exemplary embodiment, the leftmost bit of the 128-bit instructions is reserved to provide for future features and the remaining 127 bits are split into fields, each of which controls one of up to 9 functional units contained in the preferred embodiment of processor architecture:

• 1x Predicated Execution Unit (PEU) • 1x Branch/Repeat Unit (BRU)

• 2x Load-Store Units (LSU)

• 1x Vector Arithmetic Unit (VAU) • 1x Scalar Arithmetic Unit (SAU)

• 2x Integer Arithmetic Units (IAU)

• 1x Compare/Move Unit (CMU) A final functional unit not under the direct control of the instruction-word is the Debug Control-Unit (DCU) which allows the processor to monitor and control program execution, especially in the event that exceptional conditions and/or interrupts occur. The exemplary LSU allows vector data to be loaded from memory into VRF registers by the programmer. The LSU is controlled by a 24-bit opcode and can perform one of 8 operations:

• NOP i.e. not load any data from memory (clocks gated off for power- saving) • LDVX load 4-element uncompressed vector from memory

• STVX store 4-element uncompressed vector to memory

• LDCV load 4-element compressed vector from memory

• STCV store 4-element compressed vector to memory • LDLI load 16-bit immediate into lower 16-bits of specified IRF register • LDHI load 16-bit immediate into upper 16-bits of specified IRF register

• CPXI copy size/address of field 1 -8 in IRF value map to IRF base register

An exemplary opcode table for the LSU is shown in Figure 15. A block diagram of the LSU in uncompressed load/store mode is shown in Figure 16. In this mode, the bru_hold signal prevents the program-counter (PC) in the BRU from advancing until the load/store operation has been completed. This is important where loads and stores do not complete in a single cycle because the 4 vector elements must be transferred in 1 s or 2s rather than 4 elements in a single cycle (128-bits).

The block-diagram of the value map generation logic and load/store path to the VRF is shown in Figure 17 for Non-Zero Value map Generation. Whenever an uncompressed 4-element vector is loaded from memory into the VRF an additional 4-entry non-zero value map is generated using a row of 4 comparators (4x 32-input NOR gates) and inserted on the corresponding VRF row as the data fetched from memory. This 4-bit value map entry may be used later to store data back to memory in compressed format and to minimise power dissipation by eliminating trivial operations on zero values.

A unique capability of the processor architecture described herein is its support for the loading and storing of compressed data in a random access manner without requiring the entire compressed structure to be retrieved. The reason for this support is that many interesting applications operate on large bodies of sparse data i.e. the data contains large numbers of zeroes which have to be fetched from memory and when operated upon generate zero results.

The main assumptions in the architecture are that all data to be handled by the processor is packed into 4-element vectors to be stored in memory, whether this data is compressed or uncompressed, and whether the data consists of matrices, vectors and packed scalars or integers. Based on this assumption the programmer may organise matrix/vector/scalar/integer data into 32-entry compressed blocks where only those entries which are non-zero are explicitly stored. All zero values and non-zero values are represented by a 1-bit entry in a non-zero value map which can then be used to com press/uncompress the 32-entries into 32 or fewer, 32-bit words in RAM.

The transformation matrices used in 3D graphics and game physics (rigid body dynamics) applications shown in Figure are a good example. The 32-bit value maps corresponding to the 8 matrices are shown in detail in Figure 2.

As can be seen in Figure 18 graphics transformation matrices contain a large percentage of trivial (zero values) allowing over 50% data-compression to be achieved.

The compression value maps are used to allow random-access into programmer defined complex data-structures using the hardware shown in Figure 19. The compressed load operation allows the programmer to gather a series of up to 8 x A- element values specified by an 8-bit field in the opcode from addresses beginning at the base address specified in the IRF. The hardware automatically generates the required addresses and read signals as well as generating a hold signal to hold the BRU in the case of multi-cycle loads being required, for example if more than 4 non- zeros must be loaded from memory. In contrast, in the prior art this type of complex scatter gather functionality is typically complex to achieve with traditional processor architectures requiring multiple registers to be set up with base addresses and offsets to allow the indexing of complex data-structures. Furthermore such traditional processor architectures make no provision for the compression of trivial data-values. Suitably, the LSU port widths are configurable independently via the SVU_LCR register detailed in section 6. The port width parameter allows top level SoC interconnect to be traded off against the throughput of the SVU by allowing A- element vectors to be loaded/stored in multiple transactions over a /2/4 element wide data bus attached to the SVU LSU ports as shown in Figure 20.

The generation of the additional address lsbs to support the port-width parametehsation is performed using a counter which appends 0, 1 or 2 address lsbs to the address generated by the LSU as shown in Figure 21. The effect of the port-width parameter on a 4-element store to SVU RAM is shown in Figure 22. As can be seen in Figure 18 graphics transformation matrices contain a large percentage of trivial (zero values) allowing over 50% data-compression to be achieved. The compression value maps are used to allow random-access into programmer defined complex data-structures using the hardware shown in Figure 19. The compressed load operation allows the programmer to gather a series of up to 8 x A- element values specified by an 8-bit field in the opcode from addresses beginning at the base address specified in the IRF. The hardware automatically generates the required addresses and read signals as well as generating a hold signal to hold the BRU in the case of multi-cycle loads being required, for example if more than 4 non- zeros must be loaded from memory. In contrast, in the prior art this type of complex scatter gather functionality is typically complex to achieve with traditional processor architectures requiring multiple registers to be set up with base addresses and offsets to allow the indexing of complex data-structures. Furthermore such traditional processor architectures make no provision for the compression of trivial data-values.

Suitably, the LSU port widths are configurable independently, for example, by means of a parameter provided as an entry in a register. The port width parameter allows top-level SoC interconnect to be traded off against the throughput of the SVU by allowing 4-element vectors to be loaded/stored in multiple transactions over a 1/2/4 element wide data bus attached to the SVU LSU ports as shown in Figure 20. The generation of the additional address lsbs to support the port-width parameterisation is performed using a counter which appends 0, 1 or 2 address lsbs to the address generated by the LSU as shown in Figure 21.

The effect of the port-width parameter on a 4-element store to SVU RAM is shown in Figure 22. As can be seen the number of cycles taken to load or store a 128-bit vector takes 4 cycles when port width is 1 (32-bits), 2 cycles when the port width is 2 (64-bits) and 1 cycle when the port width is 4 (128-bits).

Although the present application has been described with respect to a processor with an instruction set having an identified opcode for performing the data transfer between the memory and the processor, it will be appreciated that the function may be performed where there is no explicit instruction-set in program storage, but rather loads and stores are performed under the control of a state-machine or other means. A significant advantage is that one or more items of data may be retrieved from a compressed structure in memory without having to retrieve or decompress the entire compressed structure. Similarly, a single item of data may be updated in a compressed structure stored in memory without having to retrieve or decompress the entire compressed structure. The words comprises/comprising when used in this specification are to specify the presence of stated features, integers, steps or components but does not preclude the presence or addition of one or more other features, integers , steps, components or groups thereof.

Claims

1. A processor having an instruction set architecture, the instruction set architecture comprising at least one load instruction for retrieving a data value from a compressed structure stored in random access memory, the at least one load instruction comprising an opcode and at least one operand identifying the position of the data value in a non-compressed form of the structure, the processor being adapted to convert the position of the data value in non-compressed form into a memory location for the data value within the compressed structure and being further adapted to retrieve the data value from the memory location.

2. A processor according to claim 1 , wherein the non-compressed form of the structure comprises trivial and non-trivial values and the compressed structure only comprises non-trivial data values.

3. A processor according to claim 1 or claim 2, wherein the compressed data structure comprises a map representing the locations of non-trivial data values in the non-compressed form of the structure.

4. A processor according to claim 3, wherein the processor is configured to obtain the map of the compressed data structure prior to transferring the requested value.

5. A processor according to claim 3 or 4, where the non-trivial data values are non- zero values.

6. A processor according to any one of claims 3 to 5, wherein the processor determines from the map whether the requested value is a trivial or non-trivial value. 7. A processor according to claim 6, wherein the at least one operand includes a base address for the compressed structure in memory.

8. A processor according to claim 7, wherein in the event that the requested value is a non-trivial value, the processor is configured to count from the map the number of non-trivial values and add it as an offset to the base-address in advance of the requested value and to set the random access memory access address accordingly.

9. A processor according to claim 8, wherein said count is implemented using a plurality of adders.

10. A processor according to claim 8, wherein in the event that the requested value is a trivial value, the processor is configured to return the trivial value without accessing the memory. 11. A processor according to any preceding claim wherein the processor is adapted to load a plurality of data values in response to the instruction.

12. A processor according to claim 1 , wherein the data values are floating point numbers.

13. A processor according to claim 12, wherein the data values are single precision floating point numbers.

14. A processor according to claim 12, wherein the data values are double precision floating point numbers. 15. A processor according to claim 12, wherein the data values are extended precision floating point numbers.

16. A processor according to claim 12, wherein the data values are 128-bit precision floating point numbers.

17. A processor according to claim 1 , wherein the data values are integers.

18. A processor according to claim 1 , wherein the data values are a combination of floating-point, floating-point matrices, floating-point vectors and fixed-point integers.

20. A method of loading at least one data value from within a data structure into a processor, wherein the data structure is stored in memory in a compressed form in which only non-trivial data values are present, the compressed form further comprising a map specifying the locations of the non-trivial data values in a non- compressed structure, the method comprising the steps of: receiving at least one position for the at least one data value to be loaded in the non- compressed structure, using the map to identify the at least one stored position of the at least one data value within the compressed structure, and using said at least one stored position to load the at least one data value from the memory into the processor.

21. A method according claim 20, wherein the method only loads data values for the identified at least one position.

22. A method according to claim 21 , wherein the method comprises the step of obtaining the map of the compressed data structure prior to retrieving the at least one data value.

23. A method according to claim 22, wherein the step of obtaining the map comprises retrieving the map from the memory.

24. A method according to anyone of claims 21 to 23, where the non-trivial data values are non-zero values. 25. A method according to any one of claims 21 to 24, wherein in the event that the requested data value is a non-trivial value, the method is configured to count from the map the number of non-trivial values in advance of the requested data value and to set the random access memory access address accordingly.

26. A method according to any one of claims 21 to 25, wherein the method transfers a multiple of data values at a time between the processor and memory.

27. A method according to any one of claims 21 to 24, wherein in the event that the received at least one position corresponds to the location of a trivial value in the non- compressed structure, the method is configured to provide the trivial value directly to the processor.

28. A method according to any one of claim 21 to 27, wherein the data values are floating point numbers.

29. A method according to claim 28, wherein the data values are single precision floating point numbers.

30. A method according to claim 28, wherein the data values are double precision floating point numbers.

31. A method according to claim 28, wherein the data values are extended precision floating point numbers.

32. A method according to claim 28, wherein the data values are 128-bit precision floating point numbers.

33. A method according to any one of claims 21 to 27, wherein the data values are integers.

34. A method according to claim 21 , wherein the data values are a combination of floating-point, floating-point matrices, floating-point vectors and fixed-point integers. 35. A data processing machine having a data transfer feature for loading at least one data value from a compressed structure in memory, the compressed structure comprising a plurality of data values in which only non-trivial data values are present and a map identifying the locations of the non-trivial data values in an equivalent non-compressed structure, the data transfer feature being responsive to an instruction to load the data value, the instruction identifying the position of the data value in the equivalent non-compressed structure, the data transfer function being adapted to generate a memory offset from said identified position and to use said offset to transfer the data value from memory. 36. A data processing machine according to claim 35 wherein the data processing machine is a processor.

37. A data processing machine according to claim 35, wherein the data machine is a state engine.

38. A processor having an instruction set architecture, the instruction set architecture comprising at least one store instruction for storing at least one data value from a non-compressed data structure into a corresponding compressed data structure in random access memory, the at least one store instruction comprising an opcode and at least one operand identifying the position of the at least one data value in the non- compressed data structure, the processor being adapted to convert the position of the at least one data value in non-compressed form into a memory location for the data value within the compressed structure using a map showing the locations of data values within the compressed structure and being further adapted to store the at least one data value at this memory location.

39. A processor according to claim 38, wherein the non-compressed structure comprises trivial and non-trivial values and the compressed structure only comprises non-trivial data values.

40. A processor according to claim 38 or claim 39, wherein the map is part of the compressed data structure and represents the locations of non-trivial data values in the non compressed data structure. 41. A processor according to claim 40, wherein the processor is configured to obtain the map of the compressed data structure from memory prior to storing the requested value.

42. A processor according to claim 40 or 41 , where the non-trivial data values are non-zero values.

43. A processor according to any one of claims 38 to 42, wherein the processor is configured to count from the map the number of non-trivial values in advance of the data value and to set the random access memory access address accordingly.

44. A processor according to claim 43, wherein said count is implemented using a series of adders. 45. A processor according to any one of claims 38 to 44 wherein the processor is responsive to the opcode to store a plurality of data values at a time.

46. A processor according to any one of claims 38 to 45, wherein the at least one operand includes the base address of the compressed data structure in RAM.

47. A processor according to any one of claims 38 to 47, wherein a plurality of data values in the compressed data-structure may be stored with the one instruction.

48. A processor according to claim 38, wherein the data values are floating point numbers.

49. A processor according to claim 48, wherein the data values are single precision floating point numbers. 50. A processor according to claim 48, wherein the data values are double precision floating point numbers.

51. A processor according to claim 48, wherein the data values are extended precision floating point numbers.

52. A processor according to claim 48, wherein the data values are 128-bit precision floating point numbers.

53. A processor according to claim 38, wherein the data values are integers.

54. A processor according to claim 38, wherein the data values are a combination of floating-point, floating-point matrices, floating-point vectors and fixed-point integers.

55. A method of inserting a data value from a processor into a compressed structure in a memory in which only non-trivial data values are present, the compressed structure comprising a map identifying the locations of the non-trivial data values in an equivalent non-compressed structure, the method comprising the steps of: receiving the position in the non-compressed structure of the data value to be inserted, using the map to identify from the received position the equivalent position of the data value in the compressed structure, and using said equivalent position to store the data value in the memory.

56. A method according to claim 55, wherein the method comprises the step of obtaining the map of the compressed data structure prior to storing the data value.

57. A method according to claim 56, wherein the step of obtaining the map comprises retrieving the map from the memory.

58. A method according to anyone of claims 55 to 57, where the non-trivial data values are non-zero values.

59. A method according to any one of claims 55 to 58, wherein in the event that the requested data value is a non-trivial value, the method is configured to count from the map the number of non-trivial values in advance of the requested data value and to set the random access memory access address accordingly.

60. A method according to any one of claims 55 to 59, wherein the method transfers a plurality of two data values at a time between the processor and memory.

61. A method according to any one of claim 55 to 60, wherein the data values are floating point numbers.

62. A method according to claim 61 , wherein the data values are single precision floating point numbers.

63. A method according to claim 61 , wherein the data values are double precision floating point numbers.

64. A method according to claim 61 , wherein the data values are extended precision floating point numbers.

65. A method according to claim 61 , wherein the data values are 128-bit precision floating point numbers.

66. A method according to any one of claims 55 to 60, wherein the data values are integers.

67. A method according to claim 55, wherein the data values are a combination of floating-point, floating-point matrices, floating-point vectors and fixed-point integers.

68. A data processing machine having a data transfer feature for storing a data value directly into a compressed structure in memory, the compressed structure comprising a plurality of data values in which only non-trivial data values are present and a map identifying the locations of the non-trivial data values in an equivalent non-compressed structure, the data transfer feature being responsive to an instruction to store the data value, the instruction identifying the position of the data value in the equivalent non-compressed structure, the data transfer function being adapted to generate a memory offset from said identified position and to use said offset to access the memory to store the data value into memory.

69. A data processing machine according to claim 68 wherein the data processing machine is a processor. 70. A data processing machine according to claim 68, wherein the data machine is a state engine.