Patent Application
20240045691
The disclosure relates generally to computer processor architecture, and, more specifically, to systems and methods for performing 8-bit floating-point matrix dot product instructions.
Matrices are increasingly important in many computing tasks such as machine learning and other bulk data processing. Deep Learning is a class of machine learning algorithms. Deep learning architectures, such as deep neural networks, have been applied to fields including computer vision, speech recognition, natural language processing, audio recognition, social network filtering, machine translation, bioinformatics and drug design.
Inference and training, two tools used for deep learning, are tending towards low precision arithmetic. Maximizing throughput of deep learning algorithms and computations may assist in meeting the needs of deep learning processors, for example, those performing deep learning in a data center.
Matrix-matrix multiplication (a.k.a., GEMM or General Matrix Multiplication) is a common compute-heavy operation on modern processors. Special hardware for matrix multiplication (e.g., GEMM) is a good option for improving the peak compute (and energy efficiency) of certain applications, such as deep learning.
Some of these applications, including deep learning, can operate on input data elements with relatively few bits without losing accuracy, as long as the output elements have enough bits (i.e., more than the inputs).
The present disclosure is illustrated by way of example and not limitation in the figures of the accompanying drawings, in which like references indicate similar elements and in which:
In the following description, numerous specific details are set forth. However, it is understood that embodiments may be practiced without these specific details. In other instances, well-known circuits, structures and techniques have not been shown in detail in order not to obscure the understanding of this description.
References in the specification to “one embodiment,” “an embodiment,” “an example embodiment,” etc., indicate that the embodiment described may include a particular feature, structure, or characteristic, but every embodiment may not necessarily include the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the knowledge of one skilled in the art to affect such feature, structure, or characteristic in connection with other embodiments whether or not explicitly described.
In many mainstream processors, handling matrices is a difficult and/or instruction intensive task. For example, rows of a matrix could be put into a plurality of packed data (e.g., SIMD or vector) registers and then operated on individually. For example, an add two 8×2 matrices may require a load or gather into four packed data registers depending upon data sizes. Then a first add of packed data registers corresponding to a first row from each matrix is performed and a second add of packed data registers corresponding to a second row from each matrix is performed. Then the resulting packed data registers are scattered back to memory. While for small matrices this scenario may be acceptable, it is often not acceptable with larger matrices.
Described herein are mechanisms to support matrix operations in computer hardware such as central processing units (CPUs), graphic processing units (GPUs), and accelerators. The matrix operations utilize 2-dimensional (2-D) data structures representing one or more packed regions of memory such as registers. Throughout this description, these 2-D data structures are referred to as tiles. Note that a matrix may be smaller than a tile (use less than all of a tile) or utilize a plurality of tiles (the matrix is larger than the size of any one tile). Throughout the description, matrix (tile) language is used to indicate operations performed using tiles that impact a matrix; whether or not that matrix is larger than any one tile is not typically relevant.
Each tile may be acted upon by different operations such as those that are detailed herein and include, but are not limited to: matrix (tile) multiplication, tile add, tile subtract, tile diagonal, tile zero, tile transform, tile dot product, tile broadcast, tile row broadcast, tile column broadcast, tile multiplication, tile multiplication and accumulation, tile move, etc. Additionally, support for operators such as the use of a scale and/or bias may be used with these operations or in support of non-numeric applications in the future, for instance, OpenCL “local memory,” data compression/decompression, etc. Also described herein are instructions for performing matrix (tile) 8-bit floating-point tile dot product (TDPBF8PS) instructions.
Portions of storage (such as memory (non-volatile and volatile), registers, cache, etc.) are arranged into tiles of different horizontal and vertical dimensions. For example, a tile may have horizontal dimension of 4 (e.g., four rows of a matrix) and a vertical dimension of 8 (e.g., 8 columns of the matrix). Typically, the horizontal dimension is related to element sizes (e.g., 2-, 4-, 8-, 16-, 32-, 64-, 128-bit, etc.). Multiple datatypes (single precision floating-point, double precision floating-point, integer, etc.) may be supported.
In some embodiments, tile parameters can be configured. For example, a given tile may be configured to provide tile options. Exemplary tile options include but are not limited to: a number of rows of the tile, a number of columns of the tile, whether the tile is VALID, and/or whether the tile consists of a PAIR of equal-sized tiles.
In some embodiments, tile parameters are definable. For example, a “palette” is used to provide tile options. Exemplary options include, but are not limited to: the number of tile names, the number of bytes in a row of storage, the number of rows and columns in a tile, etc. For example, a maximum “height” (number of rows) of a tile may be defined as:
Tile Max Rows=Architected Storage/(The Number of Palette Names*The Number of Bytes per row).
As such, an application can be written such that a fixed usage of names will be able to take advantage of different storage sizes across implementations.
Configuration of tiles is done using a matrix (tile) configuration (“TILECONFIG”) instruction, where a particular tile usage is defined in a selected palette. This declaration includes the number of tile names to be used, the requested number of rows and columns per name (tile), and, in some embodiments, the requested datatype of each tile. In some embodiments, consistency checks are performed during the execution of a TILECONFIG instruction to determine that it matches the restrictions of the palette entry.
Tile loads from memory and stores to memory are typically strided accesses from the application memory to packed rows of data. Exemplary TILELOAD and TILESTORE instructions, or other instruction references to application memory as a TILE operand in load-op instructions, are, in some embodiments, restartable to handle (up to) 2*rows of page faults, unmasked floating-point exceptions, and/or interrupts per instruction.
In (B), a matrix is stored in a tile comprised of a plurality of registers such as packed data registers (single instruction, multiple data (SIMD) or vector registers). In this example, the tile is overlaid on three physical registers. Typically, consecutive registers are used, however, this need not be the case.
In (C), a matrix is stored in a tile in non-register storage accessible to a fused multiply accumulate (FMA) circuit used in tile operations. This storage may be inside of an FMA, or adjacent to it. Additionally, in some embodiments, discussed below, the storage may be for a data element and not an entire row or tile.
The supported parameters for the TMMA architecture are reported via CPUID. In some embodiments, the list of information includes a maximum height and a maximum SIMD dimension. Configuring the TMMA architecture requires specifying the dimensions for each tile, the element size for each tile and the palette identifier. This configuration is done by executing the TILECONFIG instruction.
Successful execution of a TILECONFIG instruction enables subsequent TILE operators. A TILERELEASEALL instruction clears the tile configuration and disables the TILE operations (until the next TILECONFIG instructions executes). In some embodiments, XSAVE, XSTORE, etc. are used in context switching using tiles. In some embodiments, 2 XCR0 bits are used in XSAVE, one for TILECONFIG metadata and one bit corresponding to actual tile payload data.
TILECONFIG not only configures the tile usage, but also sets a state variable indicating that the program is in a region of code with tiles configured. An implementation may enumerate restrictions on other instructions that can be used with a tile region such as no usage of an existing register set, etc.
Exiting a tile region is typically done with the TILERELEASEALL instruction. It takes no parameters and swiftly invalidates all tiles (indicating that the data no longer needs any saving or restoring) and clears the internal state corresponding to being in a tile region.
In some embodiments, tile operations will zero any rows and any columns beyond the dimensions specified by the tile configuration. For example, tile operations will zero the data beyond the configured number of columns (factoring in the size of the elements) as each row is written. For example, with 64-byte rows and a tile configured with 10 rows and 12 columns, an operation writing FP32 elements would write each of the first 10 rows with 12*4 bytes with output/result data and zero the remaining 4*4 bytes in each row. Tile operations also fully zero any rows after the first 10 configured rows. When using 1K tile with 64-byte rows, there would be 16 rows, so in this example, the last 6 rows would also be zeroed.
In some embodiments, a context restore instruction (e.g., XRSTOR), when loading data, enforces that the data beyond the configured rows for a tile will be maintained as zero. If there is no valid configuration, all rows are zeroed. XRSTOR of tile data can load garbage in the columns beyond those configured. It should not be possible for XRSTOR to clear beyond the number of columns configured because there is not an element width associated with the tile configuration.
Context save (e.g., XSAVE) exposes the entire TILE storage area when writing it to memory. If XRSTOR loaded garbage data into the rightmost part of a tile, that data will be saved by XSAVE. XSAVE will write zeros for rows beyond the number specified for each tile.
In some embodiments, tile instructions are restartable. The operations that access memory allow restart after page faults. The computational instructions that deal with floating-point operations also allow for unmasked floating-point exceptions, with the masking of the exceptions controlled by a control and/or status register.
To support restarting instructions after these events, the instructions store information in the start registers detailed below.
In this example, a coherent memory interface 303 is coupled to the host processor/processing system 301 and matrix operations accelerator 307 such that they can share memory.
In some embodiments, tiles are supported using an overlay over physical registers. For example, a tile may utilize 16 1,024-bit registers, 32 512-bit registers, etc. depending on the implementation. In some embodiments, the matrix operations utilize 2-dimensional (2-D) data structures representing one or more packed regions of memory such as registers. Throughout this description, these 2-D data structures are referred to as tiles or tile registers.
In some embodiments, the matrix operations accelerator 307 includes a plurality of FMAs 309 coupled to data buffers 305 (in some implementations, one or more of these buffers 305 are stored in the FMAs of the grid as shown). The data buffers 305 buffer tiles loaded from memory and/or tiles to be stored to memory (e.g., using a tileload or tilestore instruction). Data buffers may be, for example, a plurality of registers. Typically, these FMAs are arranged as a grid of chained FMAs 309 which are able to read and write tiles. In this example, the matrix operations accelerator 307 is to perform a matrix multiply operation using tiles T0, T1, and T2. At least one of tiles is housed in the FMA grid 309. In some embodiments, all tiles in an operation are stored in the FMA grid 309. In other embodiments, only a subset is stored in the FMA grid 309. As shown, T1 is housed and T0 and T2 are not. Note that A, B, and C refer to the matrices of these tiles which may or may not take up the entire space of the tile.
The number of rows in the matrix (TILE A 601) matches the number of serial (chained) FMAs comprising the computation's latency. An implementation is free to recirculate on a grid of smaller height, but the computation remains the same.
The source/destination vector comes from a tile of N rows (TILE C 605) and the grid of FMAs 611 performs N vector-matrix operations resulting in a complete instruction performing a matrix multiplication of tiles. Tile B 603 is the other vector source and supplies “broadcast” terms to the FMAs in each stage.
In operation, in some embodiments, the elements of matrix B (stored in a tile B 603) are spread across the rectangular grid of FMAs. Matrix B (stored in tile A 601) has its elements of a row transformed to match up with the columnar dimension of the rectangular grid of FMAs. At each FMA in the grid, an element of A and B are multiplied and added to the incoming summand (from above in the FIG.) and the outgoing sum is passed to the next row of FMAs (or the final output).
The latency of a single step is proportional to K (row height of matrix B) and dependent TMMAs typically have enough source-destination rows (either in a single tile or across tile) to hide that latency. An implementation may also split the SIMD (packed data element) dimension M (row height of matrix A) across time steps, but this simply changes the constant that K is multiplied by. When a program specifies a smaller K than the maximum enumerated by the TMACC, an implementation is free to implement this with “masking” or “early outs.”
The latency of an entire TMMA is proportional to N*K. The repeat rate is proportional to N. The number of MACs per TMMA instruction is N*K*M.
A first signed source (source 1 701) and a second signed source (source 2 703) each have four packed data elements. Each of these packed data elements stores signed data such as floating-point data. A third signed source (source 3 709) has two packed data elements, each of which stores signed data. The sizes of the first and second signed sources 701 and 703 are half that of the third signed source (initial value or previous result) 709. For example, the first and second signed sources 701 and 703 could have 32-bit packed data elements (e.g., single precision floating-point) while the third signed source 709 could have 64-bit packed data elements (e.g., double precision floating-point).
packed data elements are processed in pairs. For example, the data of the most significant packed data element positions of the first and second signed sources 701 and 703 are multiplied using a multiplier circuits 705, and the data from second most significant packed data element positions of the first and second signed sources 701 and 703 are multiplied using a multiplier circuits 705, etc. In some embodiments, these multiplier circuits 705 are reused for other packed data elements positions. In other embodiments, additional multiplier circuits are used so that the packed data elements are processed in parallel. In some contexts, parallel execution is done using lanes that are the size of the signed third source 709. The results of each of the multiplications are added using addition circuitry 711.
The result of the addition of the results of the multiplications is added to the data from most significant packed data element position of the signed source 3 709 (using a different adder 713 or the same adder 711).
Finally, the result of the second addition is either stored into the signed destination 715 in a packed data element position that corresponds to the packed data element position used from the signed third source 709 or passed on to the next iteration if there is one. In some embodiments, a writemask is applied to this storage such that if a corresponding writemask (bit) is set, the storage happens, and, if not set, the storage does not happen.
A first signed source (source 1 801) and a second signed source (source 2 803) each have four packed data elements. Each of these packed data elements stores signed data such as integer data. A third signed source (source 3 809) has two packed data elements, each of which stores signed data. The sizes of the first and second signed sources 801 and 803 are half that of the third signed source 809. For example, the first and second signed sources 801 and 803 could have 32-bit packed data elements (e.g., single precision floating-point) the third signed source 809 could have 64-bit packed data elements (e.g., double precision floating-point).
As illustrated, packed data elements are processed in pairs. For example, the data of the most significant packed data element positions of the first and second signed sources 801 and 803 are multiplied using a multiplier circuits 805, and the data from second most significant packed data element positions of the first and second signed sources 801 and 803 are multiplied using a multiplier circuits 805, etc. In some embodiments, multiplier circuits 805 perform the multiplications with infinite precision without saturation and use adder/saturation circuitry 813 to saturate the results of the accumulation to plus or minus infinity in case of an overflow and to zero in case of any underflow. In other embodiments, multiplier circuits 805 perform the saturation themselves. In some embodiments, these multiplier circuits 805 are reused for other packed data element positions. In other embodiments, additional multiplier circuits are used so that the packed data elements are processed in parallel. In some contexts, parallel execution is done using lanes that are the size of the signed third source (initial value or previous iteration result) 809. The results of each of the multiplications are added to the signed third source 809 using addition/saturation circuitry 813.
Addition/saturation (accumulator) circuitry 813 preserves a sign of an operand when the addition results in a value that is too big. In particular, saturation evaluation occurs on the infinite precision result between the multi-way-add and the write to the destination or next iteration. When the accumulator 813 is floating-point and the input terms are integer, the sum of products and the floating-point accumulator input value are turned into infinite precision values (fixed point numbers of hundreds of bits), the addition of the multiplication results and the third input is performed, and a single rounding to the actual accumulator type is performed.
Unsigned saturation means the output values are limited to a maximum unsigned number for that element width (all 1s). Signed saturation means a value is limited to the be in the range between a minimum negative number and a max positive number for that element width (for bytes for example, the range is from −128 (=−2{circumflex over ( )}7) to 127(=2{circumflex over ( )}7-1)).
The result of the addition and saturation check is stored into the signed result 815 in a packed data element position that corresponds to the packed data element position used from the signed third source 809 or passed on to the next iteration if there is one. In some embodiments, a writemask is applied to this storage such that if a corresponding writemask (bit) is set, the storage happens, and, if not set, the storage does not happen.
A first signed source (source 1 901) and a second unsigned source (source 2 903) each have four packed data elements. Each of these packed data elements has data such as floating-point or integer data. A third signed source (initial value or result 915) has a packed data element of which stores signed data. The sizes of the first and second sources 901 and 903 are a quarter of the third signed source 915. For example, the first and second sources 901 and 903 could have 16-bit packed data elements (e.g., word) and the third signed source 915 could have 64-bit packed data elements (e.g., double precision floating-point or 64-bit integer).
In this illustration, the four most significant packed data element positions of the first and second sources 901 and 903 and the most significant packed data element position of the third signed source 915 are shown. Of course, other packed data element positions would also be processed if there are any.
As illustrated, packed data elements are processed in quadruplets. For example, the data of the most significant packed data element positions of the first and second sources 901 and 903 are multiplied using a multiplier circuit 905, data from second most significant packed data element positions of the first and second sources 901 and 903 are multiplied using a multiplier circuit 907, data from third most significant packed data element positions of the first and second sources 901 and 903 are multiplied using a multiplier circuit 909, and data from the least significant packed data element positions of the first and second sources 901 and 903 are multiplied using a multiplier circuit 911. In some embodiments, the signed packed data elements of the first source 901 are sign extended and the unsigned packed data elements of the second source 903 are zero extended prior to the multiplications.
In some embodiments, these multiplier circuits 905-911 are reused for other packed data elements positions. In other embodiments, additional multiplier circuits are used so that the packed data elements are processed in parallel. In some contexts, parallel execution is done using lanes that are the size of the signed third source 915. The results of each of the multiplications are added using addition circuitry 913.
The result of the addition of the results of the multiplications is added to the data from most significant packed data element position of the signed source 3 915 (using a different adder 917 or the same adder 913).
Finally, the result 919 of the second addition is either stored into the signed destination in a packed data element position that corresponds to the packed data element position used from the signed third source 915 or passed to the next iteration. In some embodiments, a writemask is applied to this storage such that if a corresponding writemask (bit) is set, the storage happens, and, if not set, the storage does not happen.
A first signed source 1001 and a second unsigned source 1003 each have four packed data elements. Each of these packed data elements stores data such as floating-point or integer data. A third signed source 1015 (initial or previous result) has a packed data element of which stores signed data. The sizes of the first and second sources are a quarter of the third signed source 1015 (initial or previous result). For example, the first and second sources could have 16-bit packed data elements (e.g., word) and the third signed source 1015 (initial or previous result) could have 64-bit packed data elements (e.g., double precision floating-point or 64-bit integer).
In this illustration, the four most significant packed data element positions of the first signed source 1001 and the second unsigned source 1003 and the most significant packed data element position of the third signed source 1015 are shown. Of course, other packed data element positions would also be processed if there are any.
As illustrated, packed data elements are processed in quadruplets. For example, the data of the most significant packed data element positions of the first signed source 1001 and the second unsigned source 1003 are multiplied using a multiplier circuit 1005, data from second most significant packed data element positions of the first signed source 1001 and the second unsigned source 1003 are multiplied using a multiplier circuit 1007, data from third most significant packed data element positions of the first signed source 1001 and the second unsigned source 1003 are multiplied using a multiplier circuit 1009, and data from the least significant packed data element positions of the first signed source 1001 and the second unsigned source 1003 are multiplied using a multiplier circuit 1011. In some embodiments, the signed packed data elements of the first signed source 1001 are sign extended and the unsigned packed data elements of the second unsigned source 1003 are zero extended prior to the multiplications.
In some embodiments, these multiplier circuits 1005-1011 are reused for other packed data elements positions. In other embodiments, additional multiplier circuits are used so that the packed data elements are processed in parallel. In some contexts, parallel execution is done using lanes that are the size of third signed source 1015 (initial or previous result). The result of the addition of the results of the multiplications is added to the data from most significant packed data element position of third signed source 1015 (initial or previous result) using adder/saturation 1013 circuitry.
Addition/saturation (accumulator) circuitry 1013 preserves a sign of an operand when the addition results in a value that is too big or too small for signed saturation. In particular, saturation evaluation occurs on the infinite precision result between the multi-way-add and the write to the destination. When the accumulator 1013 is floating-point and the input terms are integer, the sum of products and the floating-point accumulator input value are turned into infinite precision values (fixed point numbers of hundreds of bits), the addition of the multiplication results and the third input is performed, and a single rounding to the actual accumulator type is performed.
The result 1019 of the addition and saturation check is stored into the signed destination in a packed data element position that corresponds to the packed data element position used from third signed source 1015 (initial or previous result) or passed to the next iteration. In some embodiments, a writemask is applied to this storage such that if a corresponding writemask (bit) is set, the storage happens, and, if not set, the storage does not happen.
For an accumulator having 4× input sizes (in other words, the accumulator input value is four times the size of the packed data element sizes of the sources), table 1103 illustrates different configurations. For byte sized sources, the accumulator uses 32-bit integer or single-precision floating-point (SPFP) values that are 32-bit in size. For word sized sources, the accumulator uses 64-bit integer or double-precision floating-point (DPFP) values that are 64-bit in size in some embodiments.
For an accumulator having 8× input sizes (in other words, the accumulator input value is eight times the size of the packed data element sizes of the sources), table 1105 illustrates a configuration. For byte sized sources, the accumulator uses 64-bit integer.
As hinted at earlier, matrix operations circuitry may be included in a core, or as an external accelerator.
A plurality of cores, core 0 1201, core 1 1203, core 2 1205, and core N 1207 provide non-tile-based instruction support. In some embodiments, matrix operations circuitry 1251 is provided in a core 1203, and in other embodiments matrix operations circuitry 1211 and 1213 are accessible on the ring interconnect 1245.
Additionally, one or more memory controllers 1223-1225 are provided to communicate with memory 1233 and 1231 on behalf of the cores and/or matrix operations circuitry.
The branch prediction and decode circuitry 1303 is coupled to allocate/rename 1307 circuitry which is coupled, in some embodiments, to scheduler circuitry 1309. In some embodiments, these circuits provide register renaming, register allocation, and/or scheduling functionality by performing one or more of: 1) renaming logical operand values to physical operand values (e.g., a register alias table in some embodiments), 2) allocating status bits and flags to the decoded instruction, and 3) scheduling the decoded instruction for execution on execution circuitry out of an instruction pool (e.g., using a reservation station in some embodiments).
The scheduler circuitry 1309 represents any number of different schedulers, including reservations stations, central instruction window, etc. The scheduler circuitry 1309 is coupled to, or includes, physical register file(s) 1315. Each of the physical register file(s) 1315 represents one or more physical register files, different ones of which store one or more different data types, such as scalar integer, scalar floating-point, packed integer, packed floating-point, vector integer, vector floating-point, status (e.g., an instruction pointer that is the address of the next instruction to be executed), tiles, etc. In one embodiment, the physical register file(s) 1315 comprises vector registers circuitry, write mask registers circuitry, and scalar registers circuitry. These register circuits may provide architectural vector registers, vector mask registers, and general-purpose registers. The physical register file(s) 1315 is overlapped by a retirement circuit 1317 to illustrate various ways in which register renaming and out-of-order execution may be implemented (e.g., using a reorder buffer(s) and a retirement register file(s); using a future file(s), a history buffer(s), and a retirement register file(s); using a register maps and a pool of registers; etc.). The retirement circuit 1317 and the physical register file(s) 1315 are coupled to the execution circuitry 1311.
While register renaming is described in the context of out-of-order execution, it should be understood that register renaming may be used in an in-order architecture. While the illustrated embodiment of the processor may also include separate instruction and data cache units and a shared L2 cache unit, alternative embodiments may have a single internal cache for both instructions and data, such as, for example, a Level 1 (L1) internal cache, or multiple levels of internal cache. In some embodiments, the system may include a combination of an internal cache and an external cache that is external to the core and/or the processor. Alternatively, all of the cache may be external to the core and/or the processor.
The execution circuitry 1311 is a set of one or more execution circuits, including scalar circuitry 1321, vector/SIMD circuitry 1323, and matrix operations circuitry 1327, as well as memory access circuitry 1325 to access cache 1313. The execution circuits perform various operations (e.g., shifts, addition, subtraction, multiplication) and on various types of data (e.g., scalar floating-point, packed integer, packed floating-point, vector integer, vector floating-point). While some embodiments may include a number of execution units dedicated to specific functions or sets of functions, other embodiments may include only one execution unit or multiple execution units that all perform all functions. The scalar circuitry 1321 performs scalar operations, the vector/SIMD circuitry 1323 performs vector/SIMD operations, and matrix operations circuitry 1327 performs matrix (tile) operations detailed herein.
By way of example, the exemplary register renaming, out-of-order issue/execution core architecture may implement a pipeline as follows: 1) an instruction fetch circuit performs fetch and length decoding stages; 2) the branch and decode circuitry 1303 performs a decode stage; 3) the allocate/rename 1307 circuitry performs an allocation stage and renaming stage; 4) the scheduler circuitry 1309 performs a schedule stage; 5) physical register file(s) (coupled to, or included in, the scheduler circuitry 1309 and allocate/rename 1307 circuitry and a memory unit perform a register read/memory read stage; the execution circuitry 1311 performs an execute stage; 6) a memory unit and the physical register file(s) unit(s) perform a write back/memory write stage; 7) various units may be involved in the exception handling stage; and 8) a retirement unit and the physical register file(s) unit(s) perform a commit stage.
The core may support one or more instructions sets (e.g., the x86 instruction set (with some extensions that have been added with newer versions); the MIPS instruction set of MIPS Technologies of Sunnyvale, CA; the ARM instruction set (with optional additional extensions such as NEON) of ARM Holdings of Sunnyvale, CA), including the instruction(s) described herein. In one embodiment, the core 1390 includes logic to support a packed data instruction set extension (e.g., AVX1, AVX2), thereby allowing the operations used by many multimedia applications to be performed using packed data.
It should be understood that the core may support multithreading (executing two or more parallel sets of operations or threads), and may do so in a variety of ways including time sliced multithreading, simultaneous multithreading (where a single physical core provides a logical core for each of the threads that physical core is simultaneously multithreading), or a combination thereof (e.g., time sliced fetching and decoding and simultaneous multithreading thereafter such as in the Intel® Hyperthreading technology).
The branch prediction and decode circuitry 1403 is coupled to allocate/rename 1407 circuitry which is coupled, in some embodiments, to scheduler circuitry 1409. In some embodiments, these circuits provide register renaming, register allocation, and/or scheduling functionality by performing one or more of: 1) renaming logical operand values to physical operand values (e.g., a register alias table in some embodiments), 2) allocating status bits and flags to the decoded instruction, and 3) scheduling the decoded instruction for execution on execution circuitry out of an instruction pool (e.g., using a reservation station in some embodiments).
The scheduler circuitry 1409 represents any number of different schedulers, including reservations stations, central instruction window, etc. The scheduler unit(s) scheduler circuitry 1409 is coupled to, or includes, physical register file(s) 1415. Each of the physical register file(s) 1415 represents one or more physical register files, different ones of which store one or more different data types, such as scalar integer, scalar floating-point, packed integer, packed floating-point, vector integer, vector floating-point, status (e.g., an instruction pointer that is the address of the next instruction to be executed), tiles, etc. In one embodiment, the physical register file(s) 1415 comprises vector registers circuitry, write mask registers circuitry, and scalar registers circuitry. These register circuits may provide architectural vector registers, vector mask registers, and general-purpose registers. The physical register file(s) 1415 is overlapped by a retirement circuit 1417 to illustrate various ways in which register renaming and out-of-order execution may be implemented (e.g., using a reorder buffer(s) and a retirement register file(s); using a future file(s), a history buffer(s), and a retirement register file(s); using a register maps and a pool of registers; etc.). The retirement circuit 1417 and the physical register file(s) 1415 are coupled to the execution circuitry 1411.
While register renaming is described in the context of out-of-order execution, it should be understood that register renaming may be used in an in-order architecture. While the illustrated embodiment of the processor may also include separate instruction and data cache units and a shared L2 cache unit, alternative embodiments may have a single internal cache for both instructions and data, such as, for example, a Level 1 (L1) internal cache, or multiple levels of internal cache. In some embodiments, the system may include a combination of an internal cache and an external cache that is external to the core and/or the processor. Alternatively, all of the cache may be external to the core and/or the processor.
The execution circuitry 1411 a set of one or more execution circuits 1427 and a set of one or more memory access circuits 1425 to access cache 1413. The execution circuits 1427 perform matrix (tile) operations detailed herein.
By way of example, the exemplary register renaming, out-of-order issue/execution core architecture may implement a pipeline as follows: 1) an instruction fetch circuit performs fetch and length decoding stages; 2) the branch and decode circuitry 1403 performs a decode stage; 3) the allocate/rename 1407 circuitry performs an allocation stage and renaming stage; 4) the scheduler circuitry 1409 performs a schedule stage; 5) physical register file(s) (coupled to, or included in, the scheduler circuitry 1409 and allocate/rename 1407 circuitry and a memory unit perform a register read/memory read stage; the execution circuitry 1411 performs an execute stage; 6) a memory unit and the physical register file(s) unit(s) perform a write back/memory write stage; 7) various units may be involved in the exception handling stage; and 8) a retirement unit and the physical register file(s) unit(s) perform a commit stage.
The core may support one or more instructions sets (e.g., the x86 instruction set (with some extensions that have been added with newer versions); the MIPS instruction set of MIPS Technologies of Sunnyvale, CA; the ARM instruction set (with optional additional extensions such as NEON) of ARM Holdings of Sunnyvale, CA), including the instruction(s) described herein. In one embodiment, the core 1490 includes logic to support a packed data instruction set extension (e.g., AVX1, AVX2), thereby allowing the operations used by many multimedia applications to be performed using packed data.
It should be understood that the core may support multithreading (executing two or more parallel sets of operations or threads), and may do so in a variety of ways including time sliced multithreading, simultaneous multithreading (where a single physical core provides a logical core for each of the threads that physical core is simultaneously multithreading), or a combination thereof (e.g., time sliced fetching and decoding and simultaneous multithreading thereafter such as in the Intel® Hyperthreading technology).
Throughout this description, data is expressed using row major data layout. Column major users should translate the terms according to their orientation.
In some embodiments, row-major semantics are utilized in hardware, and column major data is to swap the operand order with the result being transforms of matrix, but for subsequent column-major reads from memory it is the correct, non-transformed matrix.
For example, if there are two column-major matrices to multiply:
The input matrices would be stored in linear memory (column-major) as:
Reading those matrices as row-major with dimensions 2×3 and 3×2, they would appear as:
Swapping the order and matrix multiplying:
The transform matrix is out and can then be stored in in row-major order:
The exemplary code as shown includes the usage of a tile configuration instruction and is executed to configure tile usage, load tiles, a loop to process the tiles, store tiles to memory, and release tile usage.
As discussed above, tile usage typically needs to be configured prior to use. For example, full usage of all rows and columns may not be needed. Not only does not configuring these rows and columns save power in some embodiments, but the configuration may be used to determine if an operation will generate an error. For example, a matrix multiplication of the form (N×M)*(L×N) will typically not work if M and L are not the same.
Prior to using matrices using tiles, in some embodiments, tile support is to be configured. For example, how many rows and columns per tile, tiles that are to be used, etc. are configured. A TILECONFIG instruction is an improvement to a computer itself as it provides for support to configure the computer to use a matrix accelerator (either as a part of a processor core, or as an external device). In particular, an execution of the TILECONFIG instruction causes a configuration to be retrieved from memory and applied to matrix (tile) settings within a matrix accelerator.
Instruction execution resources 1811 of a processor/core 1805 stores aspects of a tile description 1803 into tile configurations 1817. The tile configurations 1817 include palette table 1813 to detail what tiles for a palette are configured (the number of rows and columns in each tile) and a marking that matrix support is in use. In particular, instruction execution resources 1811 are configured to use tiles as specified by the tile configurations 1817. The instruction execution resources 1811 may also include a machine specific register or configuration register to indicate tile usage. Additional values such as in-use and start values are also set. The tile configurations 1817 utilize register(s) 1819 to store tile usage and configuration information.
Byte 1 stores a value to be stored in a “startRow” register 1903 and byte 2 stores a value to be stored in a register, startP 1905. To support restarting instructions after these events, the instructions store information these registers. To support restarting instructions after break events such as those detailed above, the instructions store information in these registers. The startRow value indicates the row that should be used for restart. The startP value indicates the position within the row for store operations when pairs are used and, in some embodiments, indicates the lower half of the row (in the lower tile of a pair) or higher half of the row (in the higher tile of a pair). Generally, the position in the row (the column) is not needed.
With the exception of TILECONFIG and STTILECFG, successfully executing matrix (tile) instructions will set both startRow and startP to zero.
Any time an interrupted matrix (tile) instruction is not restarted, it is the responsibility of software to zero the startRow and startP values. For example, unmasked floating-point exception handlers might decide to finish the operation in software and change the program counter value to another instruction, usually the next instruction. In this case the software exception handler must zero the startRow and startP values in the exception presented to it by the operating system before resuming the program. The operating system will subsequently reload those values using a restore instruction.
Byte 3 stores an indication of pairs (1b per tile) of tiles 1907.
Bytes 16-17 store the number of rows 1913 and columns 1915 for tile 0, bytes 18-19 store the number of rows and columns for tile 1, etc. In other words, each 2-byte group specifies a number of rows and columns for a tile. If a group of 2 bytes is not used to specify tile parameters, they should have the value zero. Specifying tile parameters for more tiles than the implementation limit or the palette limit results in a fault. Unconfigured tiles are set to an initial state with 0 rows, 0 columns.
Finally, the configuration in memory typically ends with an ending delineation such as all zeros for several consecutive bytes.
Other combinations are contemplated such as combining the start registers into a single register where they are shown separately, etc.
In contrast to the IEEE 754-standardized 16-bit (FP16) variant, BF16 does not compromise on range when being compared to FP32. FP32 numbers have 8 bits of exponent and 24 bits of mantissa (including the one implicit). BF16 cuts 16 bits from the 24-bit FP32 mantissa to create a 16-bit floating point datatype. In contrast FP16, roughly halves the FP32 mantissa to 10 explicit bits and reduces the exponent to 5 bits to fit the 16-bit datatype envelope.
Although BF16 offers less precision than FP16, it is typically better suited to support deep learning tasks. FP16's range is not enough to accomplish deep learning training out-of-the-box due to its limited range. BF16 does not suffer from this issue and the limited precision may actually help to generalize the learned weights in the neural net training task. In other words, lower precision can be seen as offering a built-in regularization property.
In some examples, an 8-bit floating point format (FP8) provides some advantages over a larger floating point format. For example, an 8-bit floating point format may reduce pressure on memory and bandwidth used for machine learning (such as weights, activations, and gradient values used for training and/or inference of neural networks). As shown, the IEEE and BF16 formats have a fixed number of bits allocated to the fraction (or mantissa which is the fraction bits+1 bit) and exponent fields. Additionally, in some examples, a fixed exponent bias may be provided for a FP16 or BF16 number. As eight bits allows for a small number of mantissa and exponent bits than FP16 or BF16 it may be advantageous to have some variance in FP8 formats (e.g., ensure high accuracy and convergence when training machine learning models).
In machine learning, different parameters, namely weights, gradients and activations, have different precision and range requirements to achieve high training accuracy and/or convergence. This allows for different allocations of the number of exponent and fraction (mantissa bits) depending on the parameter being represented.
An example FP8 format is shown in 2107. In some examples, this is called a bfloat8-bit floating point (BF8) format. As shown, this format uses 1 bit for a sign, 5 bits for the exponent, and 2 bits for the fraction (or 1+2 bits for the mantissa). An example FP8 format is shown in 2109. In some examples, this is called a hybrid8-bit floating point (HF8) format. As shown, this format uses 1 bit for a sign, 4 bits for the exponent, and 3 bits for the fraction (or 1+3 bits for the mantissa).
Normalized numbers, subnormal (denormal) numbers, and zeroes are supported in both FP8 formats. In some examples, infinity and not-a-number (NaN) encodings are not supported, however, in some examples one or more are. In examples where infinities are not supported, a maximum exponent value is not reserved for encoding NaN and +/−infinity and just used to represent normalized floating-point numbers.
In examples where infinities and NaN are supported, the are mapped to In some examples, for a NaN on an overflow, the value may be upconverted to IEEE754 NaN. In some examples, infinities and NaN raise exceptions for a hardware status register to delineate NaN from overflow.
In some examples, a zero is represented by an encoding with all zeroes the exponent and the fraction. Encodings with an all zero exponent and non-zero fraction represent denormal numbers. In the HF8 format, an exponent=00002 and mantissa=0002 represents numerical value of zero, while exponent=00002 and mantissa=0012, 0102, 0112, 1002, 1012, 1102, and 1112 represent the denormal numbers. Similarly, in the BF8 format an exponent=000002 and mantissa=002 represents numerical value of zero, while exponent=000002 and mantissa=012, 102, and 112 represent the denormal numbers.
In some examples, the FP8 formats utilize a variable exponential bias (e.g., a 6-bit unsigned integer value used as a bias). A bias skews the range of representable values more on the smaller numeric values at the expense of larger numerical values. In these examples, a numerical value of a normalized floating point number is (1)sign×2exponent-bias×1. mantissa and the numerical value of a denormal floating point number is (1)sign×2exponent-bias×0. mantissa. In some examples, the bias is provided by one or more packed data registers (e.g., SIMD or vector) where each data element position of the one or more packed data registers is to provides a bias value for a corresponding data element position of a source and/or destination. In some examples, the bias is provided by one or more general purpose registers where each general purpose register provides a bias to be used for each data element of a particular source and/or destination. Note that in some examples, a single general purpose register is used for a plurality of sources and/or destination. In some examples, the maximum bias is 16 for BF8 and 8 for HF8.
In some examples, not-a-number (NANs) and infinities are defined similarly to other IEEE floating points format, using an all-ones exponents. However, it is also acceptable in some examples to define versions of instructions that support other formats where “negative zero” is used to denote NANs and infinities, and the all-ones exponent is used to encode normal floating point numbers.
In some examples, hardware support for FP8 supports one or more status (condition code) flags: invalid, denormal, overflow, and underflow. An arithmetic operation with a denormal operand will set the denormal exception flag, while an arithmetic operation with any NaN operand or no useful definable result will set the invalid exception flag. An arithmetic operation with a result that that overflows or underflows a destination will set the overflow and underflow exception flags respectively in some examples.
Recent work has also shown that 8-bit float point formats, such as BF8 (using a 1-5-2 format (1-bit sign, 5-bit exponent, and 2-bit fraction or a 1-4-3 format), are a viable option for input data for mixed precision computation such as fused multiply-add (FMA) with BF8 inputs and a FP32 accumulator. To prepare higher-precision outputs to be used as the next operation's inputs, in some embodiments, those outputs need to be converted/rounded to FP8 numbers. Using 8-bit floating-point format instead of single-precision in at least some matrix operations is expected to alleviate memory utilization and bandwidth issues while providing a non-trivial performance upside (e.g., on the order of 2×) even during the compute operation. Additionally, numerical accuracy studies have shown that the precision of the Deep Learning application is not compromised. However, extensive workload studies have shown, that from time to time its required to avoid classic round-to-nearest behavior during these down converts. Instead, a stochastic rounding operation is needed. Examples herein relate to conversion using a provided bias term, including variable in-place, 2nd source merging and/or saturating.
Current experiments show bandwidth issues on the various cache levels and DRAM. So, as matrix compute capabilities speed up significantly (2×), the memory sub-systems capabilities only increase modestly due to reduce memory footprint. However, it has been found important to achieve convergence that FMAs accumulate into single-precision, IEEE float32. That means it may be important down-convert a result to FP8 after the operation completes.
In some examples, BF8-based operations support round to nearest even (RNE) and stochastic rounding. In some examples, HF8-based operations support round to nearest even (RNE) and stochastic rounding. In some examples, hybrid operations using both HF8 and BF8 are supported.
As mentioned above, special hardware for General Matrix Multiplication (a.k.a., GEMM), is a good option for improving the peak compute performance (and energy efficiency) of certain applications, such as deep learning. Some of these applications, including deep learning, can operate on input data elements with relatively few bits without losing accuracy, as long as the output elements have enough bits (i.e., more than the inputs).
Accordingly, disclosed methods and systems perform an 8-bit floating-point matrix dot product operation (e.g., TDP[X]BF8PS, TDP[X]HF8PS, TDP[X]BF8HF8PS, or TDP[X]HF8BF8PS) that takes source matrices (e.g., tiles) having 8-bit floating-point elements, performs dot product multiplications, and accumulates the resulting products with a 32-bit single-precision destination.
In certain embodiments, the disclosed an 8-bit floating-point matrix dot product operation instruction is to be executed by a processor that includes fetch circuitry to fetch an instruction having fields to specify an opcode and locations of a M by N destination matrix (tile) having single-precision elements, a M by K first source matrix (tile), and a K by N second source matrix (tile), each element of the specified first and second source matrices including a quadruple of 8-bit floating point values including first (e.g., having a quad index of zero in the (0, 1, 2, 3) quads in each element of the sources in
In certain embodiments, the disclosed TDP[X]BF8PS, TDP[X]HF8PS, TDP[X]BF8HF8PS, or TDP[X]HF8BF8PS instruction is to be executed by a processor that includes fetch circuitry to fetch an instruction having fields to specify an opcode that indicates execution circuitry is to cause, for each element of the first source matrix and corresponding element of the second source matrix, a conversion of the 8-bit floating-point values to single-precision values, a multiplication of converted single-precision values from first values of the quads together to generate a first result, a multiplication of converted single-precision values from second values of the quads together to generate a second result, a multiplication of converted single-precision values from third values of the quads together to generate a third result, a multiplication of converted single-precision values from fourth values of the quads together to generate a fourth result, and an accumulation of the first, second, third, and fourth results with previous contents of a corresponding element of the destination matrix.
In certain embodiments, the processor will also include other supporting hardware, such as decode circuitry to decode the fetched instruction, and execution circuitry to respond to the decoded instruction as specified by the opcode, e.g., execution circuitry that causes a matrix operations accelerator (e.g., matrix operations accelerator 307 in
An embodiment of a format for the 8-bit floating-point matrix dot product operation is TDP[X]BF8PS TSRCDST, TSRC1, TSRC2. In some embodiments, TDP[X]BF8PS is the opcode mnemonic of the instruction and indicates a tile dot product using BF8 data elements from TSRC1 and TSRC2. TSRCDST is indicated by one or more fields for a tile operand. TSRC1 and TSRC2 are indicated by one or more fields for the tile sources. In some embodiments, the destination is encoded using one or more fields for ModRM:reg(w) and the source is encoded using one or more fields for ModRM:r/m(r). In some embodiments, the instruction uses prefix 2901(C). The [X] indicates that a bias may be applied. The bias may be dynamic and provided by a second source such as a second register.
In some examples, the opcode is provided by at least field 2903, TSRCDST field is provided by at least field 3044, the first source is provided by at least bits VVVV of one of 3305, BPJ17, or 3417, and the source is provided by at least 3046. In some examples, the opcode is provided by at least field 2903, TSRCDST field is provided by at least field 3044, the source is provided by at least bits VVVV of one of 3305, BPJ17, or 3417, and the second source is provided by at least 3046 and/or the SIB byte 3004. In some examples, a bias is provided by register or memory location (such as provided by at least 3046 and/or the SIB byte 3004).
An embodiment of a format for the 8-bit floating-point matrix dot product operation is TDP[X]HF8PS TSRCDST, TSRC1, TSRC2. In some embodiments, TDP[X]HF8PS is the opcode mnemonic of the instruction and indicates a tile dot product using HF8 data elements from TSRC1 and TSRC2. TSRCDST is indicated by one or more fields for a tile operand. TSRC1 and TSRC2 are indicated by one or more fields for the tile sources. In some embodiments, the destination is encoded using one or more fields for ModRM:reg(w) and the source is encoded using one or more fields for ModRM:r/m(r). In some embodiments, the instruction uses prefix 2901(C). The [X] indicates that a bias may be applied. The bias may be dynamic and provided by a second source such as a second register.
In some examples, the opcode is provided by at least field 2903, TSRCDST field is provided by at least field 3044, the first source is provided by at least bits VVVV of one of 3305, BPJ17, or 3417, and the source is provided by at least 3046. In some examples, the opcode is provided by at least field 2903, TSRCDST field is provided by at least field 3044, the source is provided by at least bits VVVV of one of 3305, BPJ17, or 3417, and the second source is provided by at least 3046 and/or the SIB byte 3004. In some examples, a bias is provided by register or memory location (such as provided by at least 3046 and/or the SIB byte 3004).
An embodiment of a format for the 8-bit floating-point matrix dot product operation is TDP[X]HF8BF8PS TSRCDST, TSRC1, TSRC2. In some embodiments, TDP[X]HF8BF8PS is the opcode mnemonic of the instruction and indicates a tile dot product using HF8 data elements from TSRC1 and BF8 from TSRC2. TSRCDST is indicated by one or more fields for a tile operand. TSRC1 and TSRC2 are indicated by one or more fields for the tile sources. In some embodiments, the destination is encoded using one or more fields for ModRM:reg(w) and the source is encoded using one or more fields for ModRM:r/m(r). In some embodiments, the instruction uses prefix 2901(C). The [X] indicates that a bias may be applied. The bias may be dynamic and provided by a second source such as a second register.
In some examples, the opcode is provided by at least field 2903, TSRCDST field is provided by at least field 3044, the first source is provided by at least bits VVVV of one of 3305, BPJ17, or 3417, and the source is provided by at least 3046. In some examples, the opcode is provided by at least field 2903, TSRCDST field is provided by at least field 3044, the source is provided by at least bits VVVV of one of 3305, BPJ17, or 3417, and the second source is provided by at least 3046 and/or the SIB byte 3004. In some examples, a bias is provided by register or memory location (such as provided by at least 3046 and/or the SIB byte 3004).
An embodiment of a format for the 8-bit floating-point matrix dot product operation is TDP[X]BF8HF8PS TSRCDST, TSRC1, TSRC2. In some embodiments, TDP[X]BF8HF8PS is the opcode mnemonic of the instruction and indicates a tile dot product using HF8 data elements from TSRC2 and BF8 from TSRC1. TSRCDST is indicated by one or more fields for a tile operand. TSRC1 and TSRC2 are indicated by one or more fields for the tile sources. In some embodiments, the destination is encoded using one or more fields for ModRM:reg(w) and the source is encoded using one or more fields for ModRM:r/m(r). In some embodiments, the instruction uses prefix 2901(C). The [X] indicates that a bias may be applied. The bias may be dynamic and provided by a second source such as a second register.
In some examples, the opcode is provided by at least field 2903, TSRCDST field is provided by at least field 3044, the first source is provided by at least bits VVVV of one of 3305, BPJ17, or 3417, and the source is provided by at least 3046. In some examples, the opcode is provided by at least field 2903, TSRCDST field is provided by at least field 3044, the source is provided by at least bits VVVV of one of 3305, BPJ17, or 3417, and the second source is provided by at least 3046 and/or the SIB byte 3004. In some examples, a bias is provided by register or memory location (such as provided by at least 3046 and/or the SIB byte 3004).
In some examples, an architectural model specific register (MSR) (e.g., MXCSR register storing control and/or status information) is read (e.g., as part of the execution of an instruction), e.g., to determine exception information. DAZ may refer to a “denormals-are-zero” control (e.g., in an MSR)). In certain embodiments, 8-bit precision (e.g., BF8) values are able to be processed having denormal/subnormal values.
In one embodiment, an architectural machine specific register (MSR) MXCSR register (e.g., MXCSR register storing control and/or status information for an SSE register) is not read (e.g., is not inspected and/or is not updated) (e.g., as part of the execution of an instruction). In certain embodiments, exception information for an instruction is implicit in the instruction, for example, DAZ=1 being implied for a FP8 operation (e.g., without consulting MXCSR) (e.g., for a TDPBF8PS instruction) and/or DAZ=0 being implied for a non-FP8 operation (e.g., without consulting MXCSR).
In operation, M, K, and N may be specified in one or more of several ways: as operands to the TDP[X]BF8PS, TDP[X]HF8PS, TDP[X]BF8HF8PS, or TDP[X]HF8BF8PS instruction (e.g., shown as tile registers “t” here), as suffixes or prefixes to the specified opcode (an asterisk is used herein as a shorthand to refer to those optional suffixes and prefixes), as part of an immediate provided with the instruction (e.g., K, M, and N each to be specified as a different (e.g., 8) bits of an (e.g., 32-bit) immediate), as part of control registers programmed by software (e.g., XTILECONFIG is a register loaded by either a matrix accelerator configuration instruction, such as TILECFG or a XRSTORE* instructions, and is stored by matrix save instruction, such as XSAVE*), or even as architectural default values.
Here, the specified first source matrix (e.g., tile) 2212A has dimensions of M=4 by K=3. The specified second source matrix (e.g., tile) 2212B has dimensions of K=3 by N=5. K, M, and N are shown as having different values for illustrative purposes, but in other embodiments, they can be equal.
In one embodiment of operation, processor, core, or accelerator 2200 is to respond to opcode 2202 (TDP[X]BF8PS, TDP[X]HF8PS, TDP[X]BF8HF8PS, or TDP[X]HF8BF8PS) by, for each element (M,N) of the specified destination matrix (e.g., tile) 2222, convert, using convert circuit 2216A, K quads of values from row M of the specified first source matrix (e.g., tile) 222A, and, using convert circuit 2116B, K quads of values from column N of the specified second source matrix (e.g., tile) 2112B to single-precision floating-point as specified by IEEE 754. Processor, core, or accelerator 2200 is then to multiply, using multiply circuit 2218, the K pairs of converted first quad values together, the K pairs of converted second quad values together, the K pairs of converted third quad values together, and the K pairs of converted fourth quad values together, and accumulate, using accumulate circuit 2220, the products (4*K number of products) with previous contents of the element (M,N).
Performance of the TDP[X]BF8PS, TDP[X]HF8PS, TDP[X]BF8HF8PS, or TDP[X]HF8BF8PS instruction is illustrated here for setting the destination element at matrix (e.g., tile) location, e.g., the index of (row, column) being (1, 0) in C1,0. In
In certain embodiments, processor, core, or accelerator 2200 is to use accumulate circuit 2220 to accumulate the K first products with the K second products, and separately accumulate the K third products with the K fourth products, then accumulate the two accumulated sums with previous contents of the element (M,N).
As shown, three arrows travel from each of the specified first and second source matrices (e.g., tiles), to indicate that the conversions and multiplications occur in parallel. In some embodiments, the processor responds to the decoded instruction by generating and storing results into every element of the specified destination matrix (e.g., tile) in parallel. In some embodiments, new values are generated and stored into the destination on a row-at-a-time or a column-at-a-time basis.
Disclosed embodiments improve upon alternative approaches by allowing software to perform a TDP[X]BF8PS, TDP[X]HF8PS, TDP[X]BF8HF8PS, or TDP[X]HF8BF8PS instruction with reduced source element sizes, which allows less memory space and less memory bandwidth to be used and improves the peak compute performance (and energy efficiency) of certain applications. Some applications, such as deep learning, can operate on input data elements with relatively few bits without losing accuracy, as long as the output elements have enough bits (e.g., more than the inputs).
In certain embodiments, a TDP[X]BF8PS, TDP[X]HF8PS, TDP[X]BF8HF8PS, or TDP[X]HF8BF8PS instruction is part of a tile (e.g., AMX) architecture extension to an ISA that includes two-dimensional (2D) registers (for example, with each tile register identified as a single “tile register” (e.g., a single pointer to a single tile register), e.g., in contrast to vector (e.g., ZMM, YMM, or XMM) registers), and the ISA may include separate instructions to load/store 2D blocks from memory (e.g., strided sets of contiguous locations), instructions to perform matrix-matrix multiplication on three registers (e.g., matrix Cupdated=matrix A×matrix B+matrix Cprevious), and/or instructions to perform elementwise arithmetic operations on two (or three) source tiles. In one embodiment, a source matrix or source matrices are first loaded (e.g., via a host processor) into a cache (e.g., a first level (L1) data cache) and are then loaded (e.g., via execution of a tile load instruction) into a tile register (e.g., of a matrix operations accelerator) from the cache, e.g., via coherent memory interface 303 in
Example pseudocode illustrating exemplary execution of a TDP[X]BF8PS, TDP[X]HF8PS, TDP[X]BF8HF8PS, or TDP[X]HF8BF8PS instructions is provided below:
Example pseudocode for conversion functionality are as follows:
Note that the conversions include a field for the FP8 format (it is provided in the code above) and fields for bias and NaN handling.
Example helper functions are as follows:
At 2301, an instance of a single instruction an instruction having fields to specify an opcode and locations of a m by n destination matrix having single-precision elements, an m by k first source matrix, and a k by n second source matrix, the source matrices having elements that each comprise a quadruple of 8-bit floating point values, the opcode to indicate execution circuitry is to cause, for each element of the first source matrix and corresponding element of the second source matrix, a conversion of the 8-bit floating-point values to single-precision values, a multiplication of converted single-precision values from first values of the quadruples together to generate a first result, a multiplication of converted single-precision values from second values of the quadruples together to generate a second result, a multiplication of converted single-precision values from third values of the quadruples together to generate a third result, a multiplication of converted single-precision values from fourth values of the quadruples together to generate a fourth result, and an accumulation of the first, second, third, and fourth results with previous contents of a corresponding element of the destination matrix is fetched.
In some embodiments, the fetched instruction of the first instruction set is translated into one or more instructions of a second, different instruction set at 2302. This translation may be done in hardware, software, or a combination thereof. Note the translation may be from a tile instruction to one or more vector (SIMD) instructions such that vector (SIMD) hardware performs the execution.
The instance of the single instruction, or the one or more translated instructions of the second, different instruction set, is/are decoded at 2303. In some embodiments, the translation and decoding are merged.
Data values associated with the source operands of the decoded instruction are retrieved and the instruction is scheduled at 2305. For example, when one or more of the source operands are memory operands, the data from the indicated memory location is retrieved. Note that in some embodiments, the scheduling includes sending a command to an accelerator to perform the operation(s) according to the opcode of the instance of the single instruction.
At 2307, the decoded instruction(s) is/are executed by execution circuitry (hardware) such as that detailed herein. The execution circuitry may be in a processor core or in an accelerator external to a processor core. In some embodiments, the execution circuitry is configured such one or more conditions of: denormals are treated as zero (DAZ), flush to zero (FTZ) is used (denormal results are set to zero), round to nearest even (RNE) rounding is used, and/or all exceptions are suppressed (SAE). Note the execution circuitry may be in an accelerator and a command sent to the accelerator based on the decoded instruction. In some embodiments, unused rows of the source/destination matrix (tile) operand are zeroed. Examples of conversion of FP8 have been discussed. In some examples, the first and second source use different types of FP8 values (e.g., one uses BF8 and the other uses HF8).
In some embodiments, the instruction is committed or retired at 2309.
The instructions detailed above may be used in a variety of computer architectures and environments, utilize one or more instruction formats, etc. Examples of architectures, formats, etc. that support these instructions are detailed below.
Detailed below are descriptions of example computer architectures. Other system designs and configurations known in the arts for laptop, desktop, and handheld personal computers (PC)s, personal digital assistants, engineering workstations, servers, disaggregated servers, network devices, network hubs, switches, routers, embedded processors, digital signal processors (DSPs), graphics devices, video game devices, set-top boxes, micro controllers, cell phones, portable media players, hand-held devices, and various other electronic devices, are also suitable. In general, a variety of systems or electronic devices capable of incorporating a processor and/or other execution logic as disclosed herein are generally suitable.
Processors 2470 and 2480 are shown including integrated memory controller (IMC) circuitry 2472 and 2482, respectively. Processor 2470 also includes interface circuits 2476 and 2478; similarly, second processor 2480 includes interface circuits 2486 and 2488. Processors 2470, 2480 may exchange information via the interface 2450 using interface circuits 2478, 2488. IMCs 2472 and 2482 couple the processors 2470, 2480 to respective memories, namely a memory 2432 and a memory 2434, which may be portions of main memory locally attached to the respective processors.
Processors 2470, 2480 may each exchange information with a network interface (NW I/F) 2490 via individual interfaces 2452, 2454 using interface circuits 2476, 2494, 2486, 2498. The network interface 2490 (e.g., one or more of an interconnect, bus, and/or fabric, and in some examples is a chipset) may optionally exchange information with a coprocessor 2438 via an interface circuit 2492. In some examples, the coprocessor 2438 is a special-purpose processor, such as, for example, a high-throughput processor, a network or communication processor, compression engine, graphics processor, general purpose graphics processing unit (GPGPU), neural-network processing unit (NPU), embedded processor, or the like.
A shared cache (not shown) may be included in either processor 2470, 2480 or outside of both processors, yet connected with the processors via an interface such as P-P interconnect, such that either or both processors' local cache information may be stored in the shared cache if a processor is placed into a low power mode.
Network interface 2490 may be coupled to a first interface 2416 via interface circuit 2496. In some examples, first interface 2416 may be an interface such as a Peripheral Component Interconnect (PCI) interconnect, a PCI Express interconnect or another I/O interconnect. In some examples, first interface 2416 is coupled to a power control unit (PCU) 2417, which may include circuitry, software, and/or firmware to perform power management operations with regard to the processors 2470, 2480 and/or co-processor 2438. PCU 2417 provides control information to a voltage regulator (not shown) to cause the voltage regulator to generate the appropriate regulated voltage. PCU 2417 also provides control information to control the operating voltage generated. In various examples, PCU 2417 may include a variety of power management logic units (circuitry) to perform hardware-based power management. Such power management may be wholly processor controlled (e.g., by various processor hardware, and which may be triggered by workload and/or power, thermal or other processor constraints) and/or the power management may be performed responsive to external sources (such as a platform or power management source or system software).
PCU 2417 is illustrated as being present as logic separate from the processor 2470 and/or processor 2480. In other cases, PCU 2417 may execute on a given one or more of cores (not shown) of processor 2470 or 2480. In some cases, PCU 2417 may be implemented as a microcontroller (dedicated or general-purpose) or other control logic configured to execute its own dedicated power management code, sometimes referred to as P-code. In yet other examples, power management operations to be performed by PCU 2417 may be implemented externally to a processor, such as by way of a separate power management integrated circuit (PMIC) or another component external to the processor. In yet other examples, power management operations to be performed by PCU 2417 may be implemented within BIOS or other system software.
Various I/O devices 2414 may be coupled to first interface 2416, along with a bus bridge 2418 which couples first interface 2416 to a second interface 2420. In some examples, one or more additional processor(s) 2415, such as coprocessors, high throughput many integrated core (MIC) processors, GPGPUs, accelerators (such as graphics accelerators or digital signal processing (DSP) units), field programmable gate arrays (FPGAs), or any other processor, are coupled to first interface 2416. In some examples, second interface 2420 may be a low pin count (LPC) interface. Various devices may be coupled to second interface 2420 including, for example, a keyboard and/or mouse 2422, communication devices 2427 and storage circuitry 2428. Storage circuitry 2428 may be one or more non-transitory machine-readable storage media as described below, such as a disk drive or other mass storage device which may include instructions/code and data 2430 and may implement the storage ISAB03 in some examples. Further, an audio I/O 2424 may be coupled to second interface 2420. Note that other architectures than the point-to-point architecture described above are possible. For example, instead of the point-to-point architecture, a system such as multiprocessor system 2400 may implement a multi-drop interface or other such architecture.
Processor cores may be implemented in different ways, for different purposes, and in different processors. For instance, implementations of such cores may include: 1) a general purpose in-order core intended for general-purpose computing; 2) a high-performance general purpose out-of-order core intended for general-purpose computing; 3) a special purpose core intended primarily for graphics and/or scientific (throughput) computing. Implementations of different processors may include: 1) a CPU including one or more general purpose in-order cores intended for general-purpose computing and/or one or more general purpose out-of-order cores intended for general-purpose computing; and 2) a coprocessor including one or more special purpose cores intended primarily for graphics and/or scientific (throughput) computing. Such different processors lead to different computer system architectures, which may include: 1) the coprocessor on a separate chip from the CPU; 2) the coprocessor on a separate die in the same package as a CPU; 3) the coprocessor on the same die as a CPU (in which case, such a coprocessor is sometimes referred to as special purpose logic, such as integrated graphics and/or scientific (throughput) logic, or as special purpose cores); and 4) a system on a chip (SoC) that may be included on the same die as the described CPU (sometimes referred to as the application core(s) or application processor(s)), the above described coprocessor, and additional functionality. Example core architectures are described next, followed by descriptions of example processors and computer architectures.
Thus, different implementations of the processor 2500 may include: 1) a CPU with the special purpose logic 2508 being integrated graphics and/or scientific (throughput) logic (which may include one or more cores, not shown), and the cores 2502(A)-(N) being one or more general purpose cores (e.g., general purpose in-order cores, general purpose out-of-order cores, or a combination of the two); 2) a coprocessor with the cores 2502(A)-(N) being a large number of special purpose cores intended primarily for graphics and/or scientific (throughput); and 3) a coprocessor with the cores 2502(A)-(N) being a large number of general purpose in-order cores. Thus, the processor 2500 may be a general-purpose processor, coprocessor or special-purpose processor, such as, for example, a network or communication processor, compression engine, graphics processor, GPGPU (general purpose graphics processing unit), a high throughput many integrated core (MIC) coprocessor (including 30 or more cores), embedded processor, or the like. The processor may be implemented on one or more chips. The processor 2500 may be a part of and/or may be implemented on one or more substrates using any of a number of process technologies, such as, for example, complementary metal oxide semiconductor (CMOS), bipolar CMOS (BiCMOS), P-type metal oxide semiconductor (PMOS), or N-type metal oxide semiconductor (NMOS).
A memory hierarchy includes one or more levels of cache unit(s) circuitry 2504(A)-(N) within the cores 2502(A)-(N), a set of one or more shared cache unit(s) circuitry 2506, and external memory (not shown) coupled to the set of integrated memory controller unit(s) circuitry 2514. The set of one or more shared cache unit(s) circuitry 2506 may include one or more mid-level caches, such as level 2 (L2), level 3 (L3), level 4 (L4), or other levels of cache, such as a last level cache (LLC), and/or combinations thereof. While in some examples interface network circuitry 2512 (e.g., a ring interconnect) interfaces the special purpose logic 2508 (e.g., integrated graphics logic), the set of shared cache unit(s) circuitry 2506, and the system agent unit circuitry 2510, alternative examples use any number of well-known techniques for interfacing such units. In some examples, coherency is maintained between one or more of the shared cache unit(s) circuitry 2506 and cores 2502(A)-(N). In some examples, interface controller units circuitry 2516 couple the cores 2502 to one or more other devices 2518 such as one or more I/O devices, storage, one or more communication devices (e.g., wireless networking, wired networking, etc.), etc.
In some examples, one or more of the cores 2502(A)-(N) are capable of multithreading. The system agent unit circuitry 2510 includes those components coordinating and operating cores 2502(A)-(N). The system agent unit circuitry 2510 may include, for example, power control unit (PCU) circuitry and/or display unit circuitry (not shown). The PCU may be or may include logic and components needed for regulating the power state of the cores 2502(A)-(N) and/or the special purpose logic 2508 (e.g., integrated graphics logic). The display unit circuitry is for driving one or more externally connected displays.
The cores 2502(A)-(N) may be homogenous in terms of instruction set architecture (ISA). Alternatively, the cores 2502(A)-(N) may be heterogeneous in terms of ISA; that is, a subset of the cores 2502(A)-(N) may be capable of executing an ISA, while other cores may be capable of executing only a subset of that ISA or another ISA.
In
By way of example, the example register renaming, out-of-order issue/execution architecture core of
The front-end unit circuitry 2630 may include branch prediction circuitry 2632 coupled to instruction cache circuitry 2634, which is coupled to an instruction translation lookaside buffer (TLB) 2636, which is coupled to instruction fetch circuitry 2638, which is coupled to decode circuitry 2640. In one example, the instruction cache circuitry 2634 is included in the memory unit circuitry 2670 rather than the front-end circuitry 2630. The decode circuitry 2640 (or decoder) may decode instructions, and generate as an output one or more micro-operations, micro-code entry points, microinstructions, other instructions, or other control signals, which are decoded from, or which otherwise reflect, or are derived from, the original instructions. The decode circuitry 2640 may further include address generation unit (AGU, not shown) circuitry. In one example, the AGU generates an LSU address using forwarded register ports, and may further perform branch forwarding (e.g., immediate offset branch forwarding, LR register branch forwarding, etc.). The decode circuitry 2640 may be implemented using various different mechanisms. Examples of suitable mechanisms include, but are not limited to, look-up tables, hardware implementations, programmable logic arrays (PLAs), microcode read only memories (ROMs), etc. In one example, the core 2690 includes a microcode ROM (not shown) or other medium that stores microcode for certain macroinstructions (e.g., in decode circuitry 2640 or otherwise within the front-end circuitry 2630). In one example, the decode circuitry 2640 includes a micro-operation (micro-op) or operation cache (not shown) to hold/cache decoded operations, micro-tags, or micro-operations generated during the decode or other stages of the processor pipeline 2600. The decode circuitry 2640 may be coupled to rename/allocator unit circuitry 2652 in the execution engine circuitry 2650.
The execution engine circuitry 2650 includes the rename/allocator unit circuitry 2652 coupled to retirement unit circuitry 2654 and a set of one or more scheduler(s) circuitry 2656. The scheduler(s) circuitry 2656 represents any number of different schedulers, including reservations stations, central instruction window, etc. In some examples, the scheduler(s) circuitry 2656 can include arithmetic logic unit (ALU) scheduler/scheduling circuitry, ALU queues, address generation unit (AGU) scheduler/scheduling circuitry, AGU queues, etc. The scheduler(s) circuitry 2656 is coupled to the physical register file(s) circuitry 2658. Each of the physical register file(s) circuitry 2658 represents one or more physical register files, different ones of which store one or more different data types, such as scalar integer, scalar floating-point, packed integer, packed floating-point, vector integer, vector floating-point, status (e.g., an instruction pointer that is the address of the next instruction to be executed), etc. In one example, the physical register file(s) circuitry 2658 includes vector registers unit circuitry, writemask registers unit circuitry, and scalar register unit circuitry. These register units may provide architectural vector registers, vector mask registers, general-purpose registers, etc. The physical register file(s) circuitry 2658 is coupled to the retirement unit circuitry 2654 (also known as a retire queue or a retirement queue) to illustrate various ways in which register renaming and out-of-order execution may be implemented (e.g., using a reorder buffer(s) (ROB(s)) and a retirement register file(s); using a future file(s), a history buffer(s), and a retirement register file(s); using a register maps and a pool of registers; etc.). The retirement unit circuitry 2654 and the physical register file(s) circuitry 2658 are coupled to the execution cluster(s) 2660. The execution cluster(s) 2660 includes a set of one or more execution unit(s) circuitry 2662 and a set of one or more memory access circuitry 2664. The execution unit(s) circuitry 2662 may perform various arithmetic, logic, floating-point or other types of operations (e.g., shifts, addition, subtraction, multiplication) and on various types of data (e.g., scalar integer, scalar floating-point, packed integer, packed floating-point, vector integer, vector floating-point). While some examples may include a number of execution units or execution unit circuitry dedicated to specific functions or sets of functions, other examples may include only one execution unit circuitry or multiple execution units/execution unit circuitry that all perform all functions. The scheduler(s) circuitry 2656, physical register file(s) circuitry 2658, and execution cluster(s) 2660 are shown as being possibly plural because certain examples create separate pipelines for certain types of data/operations (e.g., a scalar integer pipeline, a scalar floating-point/packed integer/packed floating-point/vector integer/vector floating-point pipeline, and/or a memory access pipeline that each have their own scheduler circuitry, physical register file(s) circuitry, and/or execution cluster—and in the case of a separate memory access pipeline, certain examples are implemented in which only the execution cluster of this pipeline has the memory access unit(s) circuitry 2664). It should also be understood that where separate pipelines are used, one or more of these pipelines may be out-of-order issue/execution and the rest in-order.
In some examples, the execution engine unit circuitry 2650 may perform load store unit (LSU) address/data pipelining to an Advanced Microcontroller Bus (AMB) interface (not shown), and address phase and writeback, data phase load, store, and branches.
The set of memory access circuitry 2664 is coupled to the memory unit circuitry 2670, which includes data TLB circuitry 2672 coupled to data cache circuitry 2674 coupled to level 2 (L2) cache circuitry 2676. In one example, the memory access circuitry 2664 may include load unit circuitry, store address unit circuitry, and store data unit circuitry, each of which is coupled to the data TLB circuitry 2672 in the memory unit circuitry 2670. The instruction cache circuitry 2634 is further coupled to the level 2 (L2) cache circuitry 2676 in the memory unit circuitry 2670. In one example, the instruction cache 2634 and the data cache 2674 are combined into a single instruction and data cache (not shown) in L2 cache circuitry 2676, level 3 (L3) cache circuitry (not shown), and/or main memory. The L2 cache circuitry 2676 is coupled to one or more other levels of cache and eventually to a main memory.
The core 2690 may support one or more instructions sets (e.g., the x86 instruction set architecture (optionally with some extensions that have been added with newer versions); the MIPS instruction set architecture; the ARM instruction set architecture (optionally with optional additional extensions such as NEON)), including the instruction(s) described herein. In one example, the core 2690 includes logic to support a packed data instruction set architecture extension (e.g., AVX1, AVX2), thereby allowing the operations used by many multimedia applications to be performed using packed data.
In some examples, the register architecture 2800 includes writemask/predicate registers 2815. For example, in some examples, there are 8 writemask/predicate registers (sometimes called k0 through k7) that are each 16-bit, 32-bit, 64-bit, or 128-bit in size. Writemask/predicate registers 2815 may allow for merging (e.g., allowing any set of elements in the destination to be protected from updates during the execution of any operation) and/or zeroing (e.g., zeroing vector masks allow any set of elements in the destination to be zeroed during the execution of any operation). In some examples, each data element position in a given writemask/predicate register 2815 corresponds to a data element position of the destination. In other examples, the writemask/predicate registers 2815 are scalable and consists of a set number of enable bits for a given vector element (e.g., 8 enable bits per 64-bit vector element).
The register architecture 2800 includes a plurality of general-purpose registers 2825. These registers may be 16-bit, 32-bit, 64-bit, etc. and can be used for scalar operations. In some examples, these registers are referenced by the names RAX, RBX, RCX, RDX, RBP, RSI, RDI, RSP, and R8 through R15.
In some examples, the register architecture 2800 includes scalar floating-point (FP) register file 2845 which is used for scalar floating-point operations on 32/64/80-bit floating-point data using the x87 instruction set architecture extension or as MMX registers to perform operations on 64-bit packed integer data, as well as to hold operands for some operations performed between the MMX and XMM registers.
One or more flag registers 2840 (e.g., EFLAGS, RFLAGS, etc.) store status and control information for arithmetic, compare, and system operations. For example, the one or more flag registers 2840 may store condition code information such as carry, parity, auxiliary carry, zero, sign, and overflow. In some examples, the one or more flag registers 2840 are called program status and control registers.
Segment registers 2820 contain segment points for use in accessing memory. In some examples, these registers are referenced by the names CS, DS, SS, ES, FS, and GS.
Model specific registers or machine specific registers (MSRs) 2835 control and report on processor performance. Most MSRs 2835 handle system-related functions and are not accessible to an application program. For example, MSRs may provide control for one or more of: performance-monitoring counters, debug extensions, memory type range registers, thermal and power management, instruction-specific support, and/or processor feature/mode support. Machine check registers 2860 consist of control, status, and error reporting MSRs that are used to detect and report on hardware errors. Control register(s) 2855 (e.g., CR0-CR4) determine the operating mode of a processor (e.g., processor 2470, 2480, 2438, 2415, and/or 2500) and the characteristics of a currently executing task. In some examples, MSRs 2835 are a subset of control registers 2855.
One or more instruction pointer register(s) 2830 store an instruction pointer value. Debug registers 2850 control and allow for the monitoring of a processor or core's debugging operations.
Memory (mem) management registers 2865 specify the locations of data structures used in protected mode memory management. These registers may include a global descriptor table register (GDTR), interrupt descriptor table register (IDTR), task register, and a local descriptor table register (LDTR) register.
Alternative examples may use wider or narrower registers. Additionally, alternative examples may use more, less, or different register files and registers. The register architecture 2800 may, for example, be used in register file/memory 'ISAB08, or physical register file(s) circuitry 2658.
Instruction Set Architectures.
An instruction set architecture (ISA) may include one or more instruction formats. A given instruction format may define various fields (e.g., number of bits, location of bits) to specify, among other things, the operation to be performed (e.g., opcode) and the operand(s) on which that operation is to be performed and/or other data field(s) (e.g., mask). Some instruction formats are further broken down through the definition of instruction templates (or sub-formats). For example, the instruction templates of a given instruction format may be defined to have different subsets of the instruction format's fields (the included fields are typically in the same order, but at least some have different bit positions because there are less fields included) and/or defined to have a given field interpreted differently. Thus, each instruction of an ISA is expressed using a given instruction format (and, if defined, in a given one of the instruction templates of that instruction format) and includes fields for specifying the operation and the operands. For example, an example ADD instruction has a specific opcode and an instruction format that includes an opcode field to specify that opcode and operand fields to select operands (source1/destination and source2); and an occurrence of this ADD instruction in an instruction stream will have specific contents in the operand fields that select specific operands. In addition, though the description below is made in the context of x86 ISA, it is within the knowledge of one skilled in the art to apply the teachings of the present disclosure in another ISA.
Examples of the instruction(s) described herein may be embodied in different formats. Additionally, example systems, architectures, and pipelines are detailed below. Examples of the instruction(s) may be executed on such systems, architectures, and pipelines, but are not limited to those detailed.
The prefix(es) field(s) 2901, when used, modifies an instruction. In some examples, one or more prefixes are used to repeat string instructions (e.g., 0×F0, 0xF2, 0xF3, etc.), to provide section overrides (e.g., 0x2E, 0x36, 0x3E, 0x26, 0x64, 0x65, 0x2E, 0x3E, etc.), to perform bus lock operations, and/or to change operand (e.g., 0x66) and address sizes (e.g., 0x67). Certain instructions require a mandatory prefix (e.g., 0x66, 0xF2, 0xF3, etc.). Certain of these prefixes may be considered “legacy” prefixes. Other prefixes, one or more examples of which are detailed herein, indicate, and/or provide further capability, such as specifying particular registers, etc. The other prefixes typically follow the “legacy” prefixes.
The opcode field 2903 is used to at least partially define the operation to be performed upon a decoding of the instruction. In some examples, a primary opcode encoded in the opcode field 2903 is one, two, or three bytes in length. In other examples, a primary opcode can be a different length. An additional 3-bit opcode field is sometimes encoded in another field.
The addressing information field 2905 is used to address one or more operands of the instruction, such as a location in memory or one or more registers.
The content of the MOD field 3042 distinguishes between memory access and non-memory access modes. In some examples, when the MOD field 3042 has a binary value of 11 (11b), a register-direct addressing mode is utilized, and otherwise a register-indirect addressing mode is used.
The register field 3044 may encode either the destination register operand or a source register operand or may encode an opcode extension and not be used to encode any instruction operand. The content of register field 3044, directly or through address generation, specifies the locations of a source or destination operand (either in a register or in memory). In some examples, the register field 3044 is supplemented with an additional bit from a prefix (e.g., prefix 2901) to allow for greater addressing.
The R/M field 3046 may be used to encode an instruction operand that references a memory address or may be used to encode either the destination register operand or a source register operand. Note the R/M field 3046 may be combined with the MOD field 3042 to dictate an addressing mode in some examples.
The SIB byte 3004 includes a scale field 3052, an index field 3054, and a base field 3056 to be used in the generation of an address. The scale field 3052 indicates a scaling factor. The index field 3054 specifies an index register to use. In some examples, the index field 3054 is supplemented with an additional bit from a prefix (e.g., prefix 2901) to allow for greater addressing. The base field 3056 specifies a base register to use. In some examples, the base field 3056 is supplemented with an additional bit from a prefix (e.g., prefix 2901) to allow for greater addressing. In practice, the content of the scale field 3052 allows for the scaling of the content of the index field 3054 for memory address generation (e.g., for address generation that uses 2scale*index+base).
Some addressing forms utilize a displacement value to generate a memory address. For example, a memory address may be generated according to 2scale*index+base+displacement, index*scale+displacement, r/m+displacement, instruction pointer (RIP/EIP)+displacement, register+displacement, etc. The displacement may be a 1-byte, 2-byte, 4-byte, etc. value. In some examples, the displacement field 2907 provides this value. Additionally, in some examples, a displacement factor usage is encoded in the MOD field of the addressing information field 2905 that indicates a compressed displacement scheme for which a displacement value is calculated and stored in the displacement field 2907.
In some examples, the immediate value field 2909 specifies an immediate value for the instruction. An immediate value may be encoded as a 1-byte value, a 2-byte value, a 4-byte value, etc.
Instructions using the first prefix 2901(A) may specify up to three registers using 3-bit fields depending on the format: 1) using the reg field 3044 and the R/M field 3046 of the MOD R/M byte 3002; 2) using the MOD R/M byte 3002 with the SIB byte 3004 including using the reg field 3044 and the base field 3056 and index field 3054; or 3) using the register field of an opcode.
In the first prefix 2901(A), bit positions 7:4 are set as 0100. Bit position 3 (W) can be used to determine the operand size but may not solely determine operand width. As such, when W=0, the operand size is determined by a code segment descriptor (CS.D) and when W=1, the operand size is 64-bit.
Note that the addition of another bit allows for 16 (24) registers to be addressed, whereas the MOD R/M reg field 3044 and MOD R/M R/M field 3046 alone can each only address 8 registers.
In the first prefix 2901(A), bit position 2 (R) may be an extension of the MOD R/M reg field 3044 and may be used to modify the MOD R/M reg field 3044 when that field encodes a general-purpose register, a 64-bit packed data register (e.g., a SSE register), or a control or debug register. R is ignored when MOD R/M byte 3002 specifies other registers or defines an extended opcode.
Bit position 1 (X) may modify the SIB byte index field 3054.
Bit position 0 (B) may modify the base in the MOD R/M R/M field 3046 or the SIB byte base field 3056; or it may modify the opcode register field used for accessing general purpose registers (e.g., general purpose registers 2825).
In some examples, the second prefix 2901(B) comes in two forms—a two-byte form and a three-byte form. The two-byte second prefix 2901(B) is used mainly for 128-bit, scalar, and some 256-bit instructions; while the three-byte second prefix 2901(B) provides a compact replacement of the first prefix 2901(A) and 3-byte opcode instructions.
Instructions that use this prefix may use the MOD R/M R/M field 3046 to encode the instruction operand that references a memory address or encode either the destination register operand or a source register operand.
Instructions that use this prefix may use the MOD R/M reg field 3044 to encode either the destination register operand or a source register operand, or to be treated as an opcode extension and not used to encode any instruction operand.
For instruction syntax that support four operands, vvvv, the MOD R/M R/M field 3046 and the MOD R/M reg field 3044 encode three of the four operands. Bits[7:4] of the immediate value field 2909 are then used to encode the third source register operand.
Bit[7] of byte 2 3317 is used similar to W of the first prefix 2901(A) including helping to determine promotable operand sizes. Bit[2] is used to dictate the length (L) of the vector (where a value of 0 is a scalar or 128-bit vector and a value of 1 is a 256-bit vector). Bits[1:0] provide opcode extensionality equivalent to some legacy prefixes (e.g., 00=no prefix, 01=66H, 10=F3H, and 11=F2H). Bits[6:3], shown as vvvv, may be used to: 1) encode the first source register operand, specified in inverted (1s complement) form and valid for instructions with 2 or more source operands; 2) encode the destination register operand, specified in 1s complement form for certain vector shifts; or 3) not encode any operand, the field is reserved and should contain a certain value, such as 1111b.
Instructions that use this prefix may use the MOD R/M R/M field 3046 to encode the instruction operand that references a memory address or encode either the destination register operand or a source register operand.
Instructions that use this prefix may use the MOD R/M reg field 3044 to encode either the destination register operand or a source register operand, or to be treated as an opcode extension and not used to encode any instruction operand.
For instruction syntax that support four operands, vvvv, the MOD R/M R/M field 3046, and the MOD R/M reg field 3044 encode three of the four operands. Bits[7:4] of the immediate value field 2909 are then used to encode the third source register operand.
The third prefix 2901(C) can encode 32 vector registers (e.g., 128-bit, 256-bit, and 512-bit registers) in 64-bit mode. In some examples, instructions that utilize a writemask/opmask (see discussion of registers in a previous figure, such as
The third prefix 2901(C) may encode functionality that is specific to instruction classes (e.g., a packed instruction with “load+op” semantic can support embedded broadcast functionality, a floating-point instruction with rounding semantic can support static rounding functionality, a floating-point instruction with non-rounding arithmetic semantic can support “suppress all exceptions” functionality, etc.).
The first byte of the third prefix 2901(C) is a format field 3411 that has a value, in one example, of 62H. Subsequent bytes are referred to as payload bytes 3415-3419 and collectively form a 24-bit value of P[23:0] providing specific capability in the form of one or more fields (detailed herein).
In some examples, P[1:0] of payload byte 3419 are identical to the low two mm bits. P[3:2] are reserved in some examples. Bit P[4] (R′) allows access to the high 16 vector register set when combined with P[7] and the MOD R/M reg field 3044. P[6] can also provide access to a high 16 vector register when SIB-type addressing is not needed. P[7:5] consist of R, X, and B which are operand specifier modifier bits for vector register, general purpose register, memory addressing and allow access to the next set of 8 registers beyond the low 8 registers when combined with the MOD R/M register field 3044 and MOD R/M R/M field 3046. P[9:8] provide opcode extensionality equivalent to some legacy prefixes (e.g., 00=no prefix, 01=66H, 10=F3H, and 11=F2H). P[10] in some examples is a fixed value of 1. P[14:11], shown as vvvv, may be used to: 1) encode the first source register operand, specified in inverted (1s complement) form and valid for instructions with 2 or more source operands; 2) encode the destination register operand, specified in 1s complement form for certain vector shifts; or 3) not encode any operand, the field is reserved and should contain a certain value, such as 1111b.
P[15] is similar to W of the first prefix 2901(A) and second prefix 2911(B) and may serve as an opcode extension bit or operand size promotion.
P[18:16] specify the index of a register in the opmask (writemask) registers (e.g., writemask/predicate registers 2815). In one example, the specific value aaa=000 has a special behavior implying no opmask is used for the particular instruction (this may be implemented in a variety of ways including the use of a opmask hardwired to all ones or hardware that bypasses the masking hardware). When merging, vector masks allow any set of elements in the destination to be protected from updates during the execution of any operation (specified by the base operation and the augmentation operation); in other one example, preserving the old value of each element of the destination where the corresponding mask bit has a 0. In contrast, when zeroing vector masks allow any set of elements in the destination to be zeroed during the execution of any operation (specified by the base operation and the augmentation operation); in one example, an element of the destination is set to 0 when the corresponding mask bit has a 0 value. A subset of this functionality is the ability to control the vector length of the operation being performed (that is, the span of elements being modified, from the first to the last one); however, it is not necessary that the elements that are modified be consecutive. Thus, the opmask field allows for partial vector operations, including loads, stores, arithmetic, logical, etc. While examples are described in which the opmask field's content selects one of a number of opmask registers that contains the opmask to be used (and thus the opmask field's content indirectly identifies that masking to be performed), alternative examples instead or additional allow the mask write field's content to directly specify the masking to be performed.
P[19] can be combined with P[14:11] to encode a second source vector register in a non-destructive source syntax which can access an upper 16 vector registers using P[19]. P[20] encodes multiple functionalities, which differs across different classes of instructions and can affect the meaning of the vector length/rounding control specifier field (P[22:21]). P[23] indicates support for merging-writemasking (e.g., when set to 0) or support for zeroing and merging-writemasking (e.g., when set to 1).
Example examples of encoding of registers in instructions using the third prefix 2901(C) are detailed in the following tables.
Program code may be applied to input information to perform the functions described herein and generate output information. The output information may be applied to one or more output devices, in known fashion. For purposes of this application, a processing system includes any system that has a processor, such as, for example, a digital signal processor (DSP), a microcontroller, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a microprocessor, or any combination thereof.
The program code may be implemented in a high-level procedural or object-oriented programming language to communicate with a processing system. The program code may also be implemented in assembly or machine language, if desired. In fact, the mechanisms described herein are not limited in scope to any particular programming language. In any case, the language may be a compiled or interpreted language.
Examples of the mechanisms disclosed herein may be implemented in hardware, software, firmware, or a combination of such implementation approaches. Examples may be implemented as computer programs or program code executing on programmable systems comprising at least one processor, a storage system (including volatile and non-volatile memory and/or storage elements), at least one input device, and at least one output device.
One or more aspects of at least one example may be implemented by representative instructions stored on a machine-readable medium which represents various logic within the processor, which when read by a machine causes the machine to fabricate logic to perform the techniques described herein. Such representations, known as “intellectual property (IP) cores” may be stored on a tangible, machine readable medium and supplied to various customers or manufacturing facilities to load into the fabrication machines that make the logic or processor.
Such machine-readable storage media may include, without limitation, non-transitory, tangible arrangements of articles manufactured or formed by a machine or device, including storage media such as hard disks, any other type of disk including floppy disks, optical disks, compact disk read-only memories (CD-ROMs), compact disk rewritables (CD-RWs), and magneto-optical disks, semiconductor devices such as read-only memories (ROMs), random access memories (RAMs) such as dynamic random access memories (DRAMs), static random access memories (SRAMs), erasable programmable read-only memories (EPROMs), flash memories, electrically erasable programmable read-only memories (EEPROMs), phase change memory (PCM), magnetic or optical cards, or any other type of media suitable for storing electronic instructions.
Accordingly, examples also include non-transitory, tangible machine-readable media containing instructions or containing design data, such as Hardware Description Language (HDL), which defines structures, circuits, apparatuses, processors and/or system features described herein. Such examples may also be referred to as program products.
Emulation (Including Binary Translation, Code Morphing, Etc.).
In some cases, an instruction converter may be used to convert an instruction from a source instruction set architecture to a target instruction set architecture. For example, the instruction converter may translate (e.g., using static binary translation, dynamic binary translation including dynamic compilation), morph, emulate, or otherwise convert an instruction to one or more other instructions to be processed by the core. The instruction converter may be implemented in software, hardware, firmware, or a combination thereof. The instruction converter may be on processor, off processor, or part on and part off processor.
Example embodiments include, but are not limited to:
References to “one example,” “an example,” etc., indicate that the example described may include a particular feature, structure, or characteristic, but every example may not necessarily include the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same example. Further, when a particular feature, structure, or characteristic is described in connection with an example, it is submitted that it is within the knowledge of one skilled in the art to affect such feature, structure, or characteristic in connection with other examples whether or not explicitly described.
Moreover, in the various examples described above, unless specifically noted otherwise, disjunctive language such as the phrase “at least one of A, B, or C” or “A, B, and/or C” is intended to be understood to mean either A, B, or C, or any combination thereof (i.e. A and B, A and C, B and C, and A, B and C).
The specification and drawings are, accordingly, to be regarded in an illustrative rather than a restrictive sense. It will, however, be evident that various modifications and changes may be made thereunto without departing from the broader spirit and scope of the disclosure as set forth in the claims.
| Number | Date | Country | Kind |
|---|---|---|---|
| 202241044453 | Aug 2022 | IN | national |
