Patent Application
20210089316
The field of invention relates generally to computer processor architecture, and, more specifically, to deep learning implementations using systolic arrays and fused operations.
Matrices are increasingly important in many computing tasks such as machine learning and other bulk data processing. 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.
Unfortunately, widespread adoption and application of deep learning based machine learning systems faces challenges relating to computational requirements, power-consumption, and memory bandwidth utilization. Deep learning neural network models, for example, can be many megabytes in size and require millions of arithmetic operations per second to efficiently process.
The present invention 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 effect 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) multi-operation (TileMultiOp) 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 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 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 multiple accumulate (FMA) circuit used in tile operations. This storage may be inside of or adjacent to an FMA. 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 instruction 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 in to 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 A (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 Figure) 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 TMMA, 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 1701) and a second signed source (source 2703) 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 3709) 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).
In this illustration, only the two most significant packed data element positions of the first and second signed sources 701 and 703 and the most significant packed data element position of the third signed source 709 are shown. Of course, the other packed data element positions would also be processed.
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 701 and 703 are multiplied using a multiplier circuit 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 circuit 707. In some embodiments, these multiplier circuits 705 and 707 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 3709 (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 1801) and a second signed source (source 2803) each have four packed data elements. Each of these packed data elements stores signed data such as integer data. A third signed source (source 3809) 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).
In this illustration, only the two most significant packed data element positions of the first and second signed sources 801 and 803 and the most significant packed data element position of the third signed source 809 are shown. Of course, the other packed data element positions would also be processed.
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 circuit 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 circuit 807. In some embodiments, these multiplier circuits 805 and 807 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 (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 1901) and a second unsigned source (source 2903) 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 3915 (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 01201, core 11203, core 21205, 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 circuitry 1307 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, Calif.; the ARM instruction set (with optional additional extensions such as NEON) of ARM Holdings of Sunnyvale, Calif.), 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 is 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, Calif.; the ARM instruction set (with optional additional extensions such as NEON) of ARM Holdings of Sunnyvale, Calif.), 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:
a c e b d f
and
g h i j k l.
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 row-major order:
and used in subsequent column major computations, it is the correct un-transformed matrix:
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 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.
A machine learning algorithm is an algorithm that can learn based on a set of data. Embodiments of machine learning algorithms can be designed to model high-level abstractions within a data set. For example, image recognition algorithms can be used to determine to which of several categories a given input belongs; regression algorithms can output a numeric value given an input; and pattern recognition algorithms can be used to generate translated text or perform text-to-speech and/or speech recognition.
An exemplary type of machine learning algorithm is a neural network. There are many types of neural networks; a simple type of neural network is a feedforward network. A feedforward network may be implemented as an acyclic graph in which the nodes are arranged in layers. Typically, a feedforward network topology includes an input layer and an output layer that are separated by at least one hidden layer. The hidden layer transforms input received by the input layer into a representation that is useful for generating output in the output layer. The network nodes are fully connected via edges to the nodes in adjacent layers, but there are no edges between nodes within each layer. Data received at the nodes of an input layer of a feedforward network are propagated (i.e., “fed forward”) to the nodes of the output layer via an activation function that calculates the states of the nodes of each successive layer in the network based on coefficients (“weights”) respectively associated with each of the edges connecting the layers. Depending on the specific model being represented by the algorithm being executed, the output from the neural network algorithm can take various forms.
Before a machine learning algorithm can be used to model a particular problem, the algorithm is trained using a training data set. Training a neural network involves selecting a network topology, using a set of training data representing a problem being modeled by the network, and adjusting the weights until the network model performs with minimal error for all instances of the training data set. For example, during a supervised learning training process for a neural network, the output produced by the network in response to the input representing an instance in a training data set is compared to the “correct” labeled output for that instance, an error signal representing the difference between the output and the “correct” labeled output is calculated, and the weights associated with the connections are adjusted to minimize that error as the error signal is backward propagated through the layers of the network. The network is considered “trained” when the errors for each of the outputs generated from the instances of the training data set are minimized.
The accuracy of a machine learning algorithm can be affected significantly by the quality of the data set used to train the algorithm. The training process can be computationally intensive and may require a significant amount of time on a conventional general-purpose processor. Accordingly, parallel processing hardware is used to train many types of machine learning algorithms. This is particularly useful for optimizing the training of neural networks, as the computations performed in adjusting the coefficients in neural networks lend themselves naturally to parallel implementations. Specifically, many machine learning algorithms and software applications have been adapted to make use of the parallel processing hardware within general-purpose graphics processing devices.
Hardware acceleration for the machine learning application 2102 can be enabled via a machine learning framework 2104. The machine learning framework 2104 can provide a library of machine learning primitives. Machine learning primitives are basic operations that are commonly performed by machine learning algorithms. Without the machine learning framework 2104, developers of machine learning algorithms would be required to create and optimize the main computational logic associated with the machine learning algorithm, then re-optimize the computational logic as new parallel processors are developed. Instead, the machine learning application 2102 can be configured to perform the necessary computations using the primitives provided by the machine learning framework 2104. Exemplary primitives include tensor convolutions, activation functions, and pooling, which are computational operations that are performed while training a convolutional neural network (CNN). The machine learning framework 2104 can also provide primitives to implement basic linear algebra subprograms performed by many machine-learning algorithms, such as matrix and vector operations.
The machine learning framework 2104 can process input data received from the machine learning application 2102 and generate the appropriate input to a compute framework 2106. Compute framework 2106 can abstract the underlying instructions provided to the GPGPU driver 2108 to enable the machine learning framework 2104 to take advantage of hardware acceleration via the GPGPU hardware 2110 without requiring the machine learning framework 2104 to have intimate knowledge of the architecture of the GPGPU hardware 2110. Additionally, the compute framework 2106 can enable hardware acceleration for the machine learning framework 2104 across a variety of types and generations of the GPGPU hardware 2110.
The computing architecture provided by embodiments described herein can be configured to perform the types of parallel processing that is particularly suited for training and deploying neural networks for machine learning. A neural network can be generalized as a network of functions having a graph relationship. There are a variety of types of neural network implementations used in machine learning. One exemplary type of neural network is the feedforward network, as previously described.
A second exemplary type of neural network is the convolutional neural network (CNN). A CNN is a specialized feedforward neural network for processing data having a known, grid-like topology, such as image data. Accordingly, CNNs are commonly used for image recognition applications, but they also may be used for other types of pattern recognition such as speech and language processing. The nodes in the CNN input layer are organized into a set of “filters” (feature detectors inspired by the receptive fields found in the retina), and the output of each set of filters is propagated to nodes in successive layers of the network. The computations for a CNN include applying the convolution mathematical operation to each filter to produce the output of that filter. Convolution is a specialized kind of mathematical operation performed by two functions to produce a third function that is a modified version of one of the two original functions. In convolutional network terminology, the first function to the convolution can be referred to as the input, while the second function can be referred to as the convolution kernel. The output may be referred to as the feature map. For example, the input to a convolution layer can be a multidimensional array of data that defines the various color components of an input image. The convolution kernel can be a multidimensional array of parameters, where the parameters are adapted by the training process for the neural network.
Recurrent neural networks (RNNs) are a family of feedforward neural networks that include feedback connections between layers. RNNs enable modeling of sequential data by sharing parameter data across different parts of the neural network. The architecture for a RNN includes cycles. The cycles represent the influence of a present value of a variable on its own value at a future time, as at least a portion of the output data from the RNN is used as feedback for processing a subsequent input in a sequence. This feature makes RNNs particularly useful for language processing due to the variable nature in which language data can be composed.
The exemplary neural networks described above can be used to perform deep learning. Deep learning is machine learning using deep neural networks. The deep neural networks used in deep learning are artificial neural networks composed of multiple hidden layers, as opposed to shallow neural networks that include only a single hidden layer. Deeper neural networks are generally more computationally intensive to train. However, the additional hidden layers of the network enable multistep pattern recognition that results in reduced output error relative to shallow machine learning techniques.
Deep neural networks used in deep learning typically include a front-end network to perform feature recognition coupled to a back-end network which represents a mathematical model that can perform operations (e.g., object classification, speech recognition, etc.) based on the feature representation provided to the model. Deep learning enables machine learning to be performed without requiring hand crafted feature engineering to be performed for the model. Instead, deep neural networks can learn features based on statistical structure or correlation within the input data. The learned features can be provided to a mathematical model that can map detected features to an output. The mathematical model used by the network is generally specialized for the specific task to be performed, and different models will be used to perform different task.
Once the neural network is structured, a learning model can be applied to the network to train the network to perform specific tasks. The learning model describes how to adjust the weights within the model to reduce the output error of the network. Backpropagation of errors is a common method used to train neural networks. An input vector is presented to the network for processing. The output of the network is compared to the desired output using a loss function and an error value is calculated for each of the neurons in the output layer. The error values are then propagated backwards until each neuron has an associated error value which roughly represents its contribution to the original output. The network can then learn from those errors using an algorithm, such as the stochastic gradient descent algorithm, to update the weights of the of the neural network.
The convolutional layers are sparsely connected, which differs from a traditional neural network configuration found in the fully connected layers 2118. Traditional neural network layers are fully connected, such that every output unit interacts with every input unit. However, the convolutional layers are sparsely connected because the output of the convolution of a field is input (instead of the respective state value of each of the nodes in the field) to the nodes of the subsequent layer, as illustrated. The kernels associated with the convolutional layers perform convolution operations, the output of which is sent to the next layer. The dimensionality reduction performed within the convolutional layers is one aspect that enables the CNN to scale to process large images.
In the convolution stage 2126, execution circuitry performs several convolutions in parallel to produce a set of linear activations. The convolution stage 2126 can include an affine transformation, which is any transformation that can be specified as a linear transformation plus a translation. Affine transformations include rotations, translations, scaling, and combinations of these transformations. The convolution stage computes the output of functions (e.g., neurons) that are connected to specific regions in the input, which can be determined as the local region associated with the neuron. The neurons compute a dot product between the weights of the neurons and the region in the local input to which the neurons are connected. The output from the convolution stage 2126 defines a set of linear activations that are processed by successive stages of the convolutional layer 2124.
The linear activations can be processed by a detector stage 2128. In the detector stage 2128, each linear activation is processed by a non-linear activation function. The non-linear activation function increases the nonlinear properties of the overall network without affecting the receptive fields of the convolution layer. Several types of non-linear activation functions may be used. One particular type is the rectified linear unit (ReLU), which uses an activation function defined as f(x)=max(0, x), such that the activation threshold is at zero. Other types of activation functions that can be used include the sigmoid function, the tanh function, and the soft max function,
The pooling stage 2130 uses a pooling function that replaces the output of the convolutional layer 2124 with a summary statistic of the nearby outputs. The pooling function can be used to introduce translation invariance into the neural network, such that small translations to the input do not change the pooled outputs. Invariance to local translation can be useful in scenarios where the presence of a feature in the input data is more important than the precise location of the feature. Various types of pooling functions can be used during the pooling stage 2130, including max pooling, average pooling, and 12-norm pooling. A numeric example of an embodiment of pooling is illustrated and described with respect to
The output from the convolutional layer 2124 can then be processed by the next layer 2132. The next layer 2132 can be an additional convolutional layer or one of the fully connected layers 2118. For example, the first convolutional layer 2114 of
As mentioned above, widespread adoption and application of deep learning based machine learning systems faces challenges relating to computational requirements, power-consumption, and memory bandwidth utilization. Disclosed herein are embodiments of an invention for fusing multiple machine learning primitives (such as convolution followed by activation) by using registers to pass the results of a first primitive to second and subsequent primitives. By fusing machine learning primitives, disclosed embodiments can significantly reduce memory bandwidth utilization and power expenditure by avoiding the need to write intermediate results to memory. For example, performing a convolution by convolving a 5×5 convolution filter over a 32×32 input matrix at a stride of 1 involves over 25,000 multiplications. Disclosed embodiments avoid the need to write those 25,000 intermediate results back to memory.
Alternative, inferior approaches implement these machine learning primitives serially, using separate, explicit instructions or sequences of instructions, writing intermediate results back to memory. Such approaches utilize more memory bandwidth and consume more power than disclosed embodiments.
Thus, in one embodiment, a processor implements machine learning primitives by fetching an instruction using fetch circuitry and decoding the instruction using decode circuitry. The instruction has fields to specify an opcode and locations of a destination and N source matrices. In one example, the opcode indicates the processor is to load the N source matrices from memory, perform a convolution on each of the N source matrices, and store results of the N convolutions in registers to be passed to an activation layer. The processor then generates N feature maps by using a systolic array of execution units to perform a convolution on each of the N source matrices. Advantageously, the processor can perform the convolutions and the activation with at most one memory load of each of the N source matrices. In other words, the processor stores intermediate results in registers, and avoids using up memory bandwidth and consuming power. The processor schedules, using scheduling circuitry, execution of the instruction, and executes the instruction as per the opcode.
In general, disclosed embodiments support fusing a first machine learning primitive with a second machine learning primitive, wherein the fusing means results of the first primitive are passed using registers, to serve as inputs to the second primitive.
In operation, multi-op execution circuitry is to receive three input matrixes, 2142, 2144, 2146. As per the opcode, the processor is to execute a convolution and a ReLU at 2148, and write the ReLU results to destination 2150.
TileMultiOp instruction 2400 further includes several optional parameters to control the processor's behavior, including M 2412, K 2414, and N 2416 (to set the matrix dimensions), element size 2418 (crumb, nibble, byte, word, doubleword, or quadword), element format 2420 (packed or scalar single- or double-precision floating-point data and packed or scalar integer data), and mask 2422 (multi-bit value with one bit per destination element, the bit to control whether the destination element is to be updated, or if it is to be masked from being updated. In some embodiments, mask 2422 identifies one of the architectural mask registers in the processor register file to serve as the mask. Masked destination elements are either to be zeroed or merged, as controlled by another instruction field, or by a control register programmed by software.
Opcode 2402 is shown including an asterisk, which is to convey that additional prefixes and/or suffixes may be added to specify instruction behavior. One or more of instruction modifiers 2412, 2414, 2416, 2418, 2420, and 2422 may be specified using prefixes or suffixes to opcode 2402.
In some embodiments, one or more of optional instructions modifiers 2412, 2414, 2416, 2418, 2420, and 2422 are encoded in an immediate field (not shown) optionally included with the instruction 2400. In some embodiments, one or more of optional instruction modifiers 2412, 2414, 2416, 2418, 2420, and 2422 is specified via a configuration/status register (e.g., XTILECONFIG). In other words, when any one or more of optional modifiers 2412, 2414, 2416, 2418, 2420, and 2422 are not specified by the instruction, they sometimes are specified with implicit parameters that are inherited from other parts of the tile architecture.
Detailed herein are examples of hardware, software, etc. to execute the above described instructions. For example, what is described below details aspects of instruction execution including various pipeline stages such as fetch, decode, schedule, execute, retire, etc.
An instruction set 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 though the definition of instruction templates (or subformats). 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 exemplary 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. A set of SIMD extensions referred to as the Advanced Vector Extensions (AVX) (AVX1 and AVX2) and using the Vector Extensions (VEX) coding scheme has been released and/or published (e.g., see Intel® 64 and IA-32 Architectures Software Developer's Manual, September 2014; and see Intel® Advanced Vector Extensions Programming Reference, October 2014).
Embodiments of the instruction(s) described herein may be embodied in different formats. Additionally, exemplary systems, architectures, and pipelines are detailed below. Embodiments of the instruction(s) may be executed on such systems, architectures, and pipelines, but are not limited to those detailed.
A vector friendly instruction format is an instruction format that is suited for vector instructions (e.g., there are certain fields specific to vector operations). While embodiments are described in which both vector and scalar operations are supported through the vector friendly instruction format, alternative embodiments use only vector operations the vector friendly instruction format.
While embodiments will be described in which the vector friendly instruction format supports the following: a 64 byte vector operand length (or size) with 32 bit (4 byte) or 64 bit (8 byte) data element widths (or sizes) (and thus, a 64 byte vector consists of either 16 doubleword-size elements or alternatively, 8 quadword-size elements); a 64 byte vector operand length (or size) with 16 bit (2 byte) or 8 bit (1 byte) data element widths (or sizes); a 32 byte vector operand length (or size) with 32 bit (4 byte), 64 bit (8 byte), 16 bit (2 byte), or 8 bit (1 byte) data element widths (or sizes); and a 16 byte vector operand length (or size) with 32 bit (4 byte), 64 bit (8 byte), 16 bit (2 byte), or 8 bit (1 byte) data element widths (or sizes); alternative embodiments may support more, less and/or different vector operand sizes (e.g., 256 byte vector operands) with more, less, or different data element widths (e.g., 128 bit (16 byte) data element widths).
The class A instruction templates in
The generic vector friendly instruction format 2500 includes the following fields listed below in the order illustrated in
Format field 2540—a specific value (an instruction format identifier value) in this field uniquely identifies the vector friendly instruction format, and thus occurrences of instructions in the vector friendly instruction format in instruction streams. As such, this field is optional in the sense that it is not needed for an instruction set that has only the generic vector friendly instruction format.
Base operation field 2542—its content distinguishes different base operations.
Register index field 2544—its content, directly or through address generation, specifies the locations of the source and destination operands, be they in registers or in memory. These include a sufficient number of bits to select N registers from a P×Q (e.g. 32×512, 16×128, 32×1024, 64×1024) register file. While in one embodiment N may be up to three sources and one destination register, alternative embodiments may support more or less sources and destination registers (e.g., may support up to two sources where one of these sources also acts as the destination, may support up to three sources where one of these sources also acts as the destination, may support up to two sources and one destination).
Modifier field 2546—its content distinguishes occurrences of instructions in the generic vector instruction format that specify memory access from those that do not; that is, between no memory access 2505 instruction templates and memory access 2520 instruction templates. Memory access operations read and/or write to the memory hierarchy (in some cases specifying the source and/or destination addresses using values in registers), while non-memory access operations do not (e.g., the source and destinations are registers). While in one embodiment this field also selects between three different ways to perform memory address calculations, alternative embodiments may support more, less, or different ways to perform memory address calculations.
Augmentation operation field 2550—its content distinguishes which one of a variety of different operations to be performed in addition to the base operation. This field is context specific. In one embodiment, this field is divided into a class field 2568, an alpha field 2552, and a beta field 2554. The augmentation operation field 2550 allows common groups of operations to be performed in a single instruction rather than 2, 3, or 4 instructions.
Scale field 2560—its content allows for the scaling of the index field's content for memory address generation (e.g., for address generation that uses 2scale*index+base).
Displacement Field 2562A—its content is used as part of memory address generation (e.g., for address generation that uses 2scale*index+base+displacement).
Displacement Factor Field 2562B (note that the juxtaposition of displacement field 2562A directly over displacement factor field 2562B indicates one or the other is used)—its content is used as part of address generation; it specifies a displacement factor that is to be scaled by the size of a memory access (N)—where N is the number of bytes in the memory access (e.g., for address generation that uses 2scale*index+base+scaled displacement). Redundant low-order bits are ignored and hence, the displacement factor field's content is multiplied by the memory operands total size (N) in order to generate the final displacement to be used in calculating an effective address. The value of N is determined by the processor hardware at runtime based on the full opcode field 2574 (described later herein) and the data manipulation field 2554C. The displacement field 2562A and the displacement factor field 2562B are optional in the sense that they are not used for the no memory access 2505 instruction templates and/or different embodiments may implement only one or none of the two.
Data element width field 2564—its content distinguishes which one of a number of data element widths is to be used (in some embodiments for all instructions; in other embodiments for only some of the instructions). This field is optional in the sense that it is not needed if only one data element width is supported and/or data element widths are supported using some aspect of the opcodes.
Write mask field 2570—its content controls, on a per data element position basis, whether that data element position in the destination vector operand reflects the result of the base operation and augmentation operation. Class A instruction templates support merging-writemasking, while class B instruction templates support both merging- and zeroing-writemasking. 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 one embodiment, 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 embodiment, 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 write mask field 2570 allows for partial vector operations, including loads, stores, arithmetic, logical, etc. While embodiments are described in which the write mask field's 2570 content selects one of a number of write mask registers that contains the write mask to be used (and thus the write mask field's 2570 content indirectly identifies that masking to be performed), alternative embodiments instead or additional allow the mask write field's 2570 content to directly specify the masking to be performed.
Immediate field 2572—its content allows for the specification of an immediate. This field is optional in the sense that is it not present in an implementation of the generic vector friendly format that does not support immediate and it is not present in instructions that do not use an immediate.
Class field 2568—its content distinguishes between different classes of instructions. With reference to
In the case of the non-memory access 2505 instruction templates of class A, the alpha field 2552 is interpreted as an RS field 2552A, whose content distinguishes which one of the different augmentation operation types are to be performed (e.g., round 2552A.1 and data transform 2552A.2 are respectively specified for the no memory access, round type operation 2510 and the no memory access, data transform type operation 2515 instruction templates), while the beta field 2554 distinguishes which of the operations of the specified type is to be performed. In the no memory access 2505 instruction templates, the scale field 2560, the displacement field 2562A, and the displacement factor field 2562B are not present.
In the no memory access full round control type operation 2510 instruction template, the beta field 2554 is interpreted as a round control field 2554A, whose content(s) provide static rounding. While in the described embodiments the round control field 2554A includes a suppress all floating-point exceptions (SAE) field 2556 and a round operation control field 2558, alternative embodiments may support may encode both these concepts into the same field or only have one or the other of these concepts/fields (e.g., may have only the round operation control field 2558).
SAE field 2556—its content distinguishes whether or not to disable the exception event reporting; when the SAE field's 2556 content indicates suppression is enabled, a given instruction does not report any kind of floating-point exception flag and does not raise any floating-point exception handler.
Round operation control field 2558—its content distinguishes which one of a group of rounding operations to perform (e.g., Round-up, Round-down, Round-towards-zero and Round-to-nearest). Thus, the round operation control field 2558 allows for the changing of the rounding mode on a per instruction basis. In one embodiment where a processor includes a control register for specifying rounding modes, the round operation control field's 2558 content overrides that register value.
In the no memory access data transform type operation 2515 instruction template, the beta field 2554 is interpreted as a data transform field 2554B, whose content distinguishes which one of a number of data transforms is to be performed (e.g., no data transform, swizzle, broadcast).
In the case of a memory access 2520 instruction template of class A, the alpha field 2552 is interpreted as an eviction hint field 2552B, whose content distinguishes which one of the eviction hints is to be used (in
Vector memory instructions perform vector loads from and vector stores to memory, with conversion support. As with regular vector instructions, vector memory instructions transfer data from/to memory in a data element-wise fashion, with the elements that are actually transferred is dictated by the contents of the vector mask that is selected as the write mask.
Temporal data is data likely to be reused soon enough to benefit from caching. This is, however, a hint, and different processors may implement it in different ways, including ignoring the hint entirely.
Non-temporal data is data unlikely to be reused soon enough to benefit from caching in the 1st-level cache and should be given priority for eviction. This is, however, a hint, and different processors may implement it in different ways, including ignoring the hint entirely.
In the case of the instruction templates of class B, the alpha field 2552 is interpreted as a write mask control (Z) field 2552C, whose content distinguishes whether the write masking controlled by the write mask field 2570 should be a merging or a zeroing.
In the case of the non-memory access 2505 instruction templates of class B, part of the beta field 2554 is interpreted as an RL field 2557A, whose content distinguishes which one of the different augmentation operation types are to be performed (e.g., round 2557A.1 and vector length (VSIZE) 2557A.2 are respectively specified for the no memory access, write mask control, partial round control type operation 2512 instruction template and the no memory access, write mask control, VSIZE type operation 2517 instruction template), while the rest of the beta field 2554 distinguishes which of the operations of the specified type is to be performed. In the no memory access 2505 instruction templates, the scale field 2560, the displacement field 2562A, and the displacement factor field 2562B are not present.
In the no memory access, write mask control, partial round control type operation 2510 instruction template, the rest of the beta field 2554 is interpreted as a round operation field 2559A and exception event reporting is disabled (a given instruction does not report any kind of floating-point exception flag and does not raise any floating-point exception handler).
Round operation control field 2559A—just as round operation control field 2558, its content distinguishes which one of a group of rounding operations to perform (e.g., Round-up, Round-down, Round-towards-zero and Round-to-nearest). Thus, the round operation control field 2559A allows for the changing of the rounding mode on a per instruction basis. In one embodiment where a processor includes a control register for specifying rounding modes, the round operation control field's 2550 content overrides that register value.
In the no memory access, write mask control, VSIZE type operation 2517 instruction template, the rest of the beta field 2554 is interpreted as a vector length field 2559B, whose content distinguishes which one of a number of data vector lengths is to be performed on (e.g., 128, 256, or 512 byte).
In the case of a memory access 2520 instruction template of class B, part of the beta field 2554 is interpreted as a broadcast field 2557B, whose content distinguishes whether or not the broadcast type data manipulation operation is to be performed, while the rest of the beta field 2554 is interpreted the vector length field 2559B. The memory access 2520 instruction templates include the scale field 2560, and optionally the displacement field 2562A or the displacement factor field 2562B.
With regard to the generic vector friendly instruction format 2500, a full opcode field 2574 is shown including the format field 2540, the base operation field 2542, and the data element width field 2564. While one embodiment is shown where the full opcode field 2574 includes all of these fields, the full opcode field 2574 includes less than all of these fields in embodiments that do not support all of them. The full opcode field 2574 provides the operation code (opcode).
The augmentation operation field 2550, the data element width field 2564, and the write mask field 2570 allow these features to be specified on a per instruction basis in the generic vector friendly instruction format.
The combination of write mask field and data element width field create typed instructions in that they allow the mask to be applied based on different data element widths.
The various instruction templates found within class A and class B are beneficial in different situations. In some embodiments, different processors or different cores within a processor may support only class A, only class B, or both classes. For instance, a high performance general purpose out-of-order core intended for general-purpose computing may support only class B, a core intended primarily for graphics and/or scientific (throughput) computing may support only class A, and a core intended for both may support both (of course, a core that has some mix of templates and instructions from both classes but not all templates and instructions from both classes is within the purview of the invention). Also, a single processor may include multiple cores, all of which support the same class or in which different cores support different class. For instance, in a processor with separate graphics and general-purpose cores, one of the graphics cores intended primarily for graphics and/or scientific computing may support only class A, while one or more of the general-purpose cores may be high performance general purpose cores with out of order execution and register renaming intended for general-purpose computing that support only class B. Another processor that does not have a separate graphics core may include one more general purpose in-order or out-of-order cores that support both class A and class B. Of course, features from one class may also be implement in the other class in different embodiments. Programs written in a high level language would be put (e.g., just in time compiled or statically compiled) into an variety of different executable forms, including: 1) a form having only instructions of the class(es) supported by the target processor for execution; or 2) a form having alternative routines written using different combinations of the instructions of all classes and having control flow code that selects the routines to execute based on the instructions supported by the processor which is currently executing the code.
It should be understood that, although embodiments are described with reference to the specific vector friendly instruction format 2600 in the context of the generic vector friendly instruction format 2500 for illustrative purposes, the invention is not limited to the specific vector friendly instruction format 2600 except where claimed. For example, the generic vector friendly instruction format 2500 contemplates a variety of possible sizes for the various fields, while the specific vector friendly instruction format 2600 is shown as having fields of specific sizes. By way of specific example, while the data element width field 2564 is illustrated as a one-bit field in the specific vector friendly instruction format 2600, the invention is not so limited (that is, the generic vector friendly instruction format 2500 contemplates other sizes of the data element width field 2564).
The specific vector friendly instruction format 2600 includes the following fields listed below in the order illustrated in
EVEX Prefix 2602 (Bytes 0-3)—is encoded in a four-byte form.
Format Field 2540 (EVEX Byte 0, bits [7:0])—the first byte (EVEX Byte 0) is the format field 2540 and it contains 0x62 (the unique value used for distinguishing the vector friendly instruction format in one embodiment).
The second-fourth bytes (EVEX Bytes 1-3) include a number of bit fields providing specific capability.
REX field 2605 (EVEX Byte 1, bits [7-5])—consists of an EVEX.R bit field (EVEX Byte 1, bit [7]—R), EVEX.X bit field (EVEX byte 1, bit [6]—X), and EVEX.B bit field (EVEX Byte 1, bit [5]—B). The EVEX.R, EVEX.X, and EVEX.B bit fields provide the same functionality as the corresponding VEX bit fields, and are encoded using 1s complement form, i.e. ZMM0 is encoded as 1111B, ZMM15 is encoded as 0000B. Other fields of the instructions encode the lower three bits of the register indexes as is known in the art (rrr, xxx, and bbb), so that Rrrr, Xxxx, and Bbbb may be formed by adding EVEX.R, EVEX.X, and EVEX.B.
REX′ field 2610A—this is the first part of the REX′ field 2610 and is the EVEX.R′ bit field (EVEX Byte 1, bit [4]—R′) that is used to encode either the upper 16 or lower 16 of the extended 32 register set. In one embodiment, this bit, along with others as indicated below, is stored in bit inverted format to distinguish (in the well-known x86 32-bit mode) from the BOUND instruction, whose real opcode byte is 62, but does not accept in the MOD R/M field (described below) the value of 11 in the MOD field; alternative embodiments do not store this and the other indicated bits below in the inverted format. A value of 1 is used to encode the lower 16 registers. In other words, R′Rrrr is formed by combining EVEX.R′, EVEX.R, and the other RRR from other fields.
Opcode map field 2615 (EVEX byte 1, bits [3:0]—mmmm)—its content encodes an implied leading opcode byte (0F, 0F 38, or 0F 3).
Data element width field 2564 (EVEX byte 2, bit [7]—W)—is represented by the notation EVEX.W. EVEX.W is used to define the granularity (size) of the datatype (either 32-bit data elements or 64-bit data elements).
EVEX.vvvv 2620 (EVEX Byte 2, bits [6:3]-vvvv)—the role of EVEX.vvvv may include the following: 1) EVEX.vvvv encodes the first source register operand, specified in inverted (1s complement) form and is valid for instructions with 2 or more source operands; 2) EVEX.vvvv encodes the destination register operand, specified in 1s complement form for certain vector shifts; or 3) EVEX.vvvv does not encode any operand, the field is reserved and should contain 1111b. Thus, EVEX.vvvv field 2620 encodes the 4 low-order bits of the first source register specifier stored in inverted (1s complement) form. Depending on the instruction, an extra different EVEX bit field is used to extend the specifier size to 32 registers.
EVEX.02568 Class field (EVEX byte 2, bit [2]-U)—If EVEX.0=0, it indicates class A or EVEX.U0; if EVEX.0=1, it indicates class B or EVEX.U1.
Prefix encoding field 2625 (EVEX byte 2, bits [1:0]-pp)—provides additional bits for the base operation field. In addition to providing support for the legacy SSE instructions in the EVEX prefix format, this also has the benefit of compacting the SIMD prefix (rather than requiring a byte to express the SIMD prefix, the EVEX prefix requires only 2 bits). In one embodiment, to support legacy SSE instructions that use a SIMD prefix (66H, F2H, F3H) in both the legacy format and in the EVEX prefix format, these legacy SIMD prefixes are encoded into the SIMD prefix encoding field; and at runtime are expanded into the legacy SIMD prefix prior to being provided to the decoder's PLA (so the PLA can execute both the legacy and EVEX format of these legacy instructions without modification). Although newer instructions could use the EVEX prefix encoding field's content directly as an opcode extension, certain embodiments expand in a similar fashion for consistency but allow for different meanings to be specified by these legacy SIMD prefixes. An alternative embodiment may redesign the PLA to support the 2-bit SIMD prefix encodings, and thus not require the expansion.
Alpha field 2552 (EVEX byte 3, bit [7]—EH; also known as EVEX.EH, EVEX.rs, EVEX.RL, EVEX.write mask control, and EVEX.N; also illustrated with a)—as previously described, this field is context specific.
Beta field 2554 (EVEX byte 3, bits [6:4]—SSS, also known as EVEX.s2-0, EVEX.r2-0, EVEX.rr1, EVEX.LL0, EVEX.LLB; also illustrated with BB(3)—as previously described, this field is context specific.
REX′ field 2610B—this is the remainder of the REX′ field 2610 and is the EVEX.V′ bit field (EVEX Byte 3, bit [3]—V′) that may be used to encode either the upper 16 or lower 16 of the extended 32 register set. This bit is stored in bit inverted format. A value of 1 is used to encode the lower 16 registers. In other words, V′VVVV is formed by combining EVEX.V′, EVEX.vvvv.
Write mask field 2570 (EVEX byte 3, bits [2:0]-kkk)—its content specifies the index of a register in the write mask registers as previously described. In one embodiment, the specific value EVEX.kkk=000 has a special behavior implying no write mask is used for the particular instruction (this may be implemented in a variety of ways including the use of a write mask hardwired to all ones or hardware that bypasses the masking hardware).
Real Opcode Field 2630 (Byte 4) is also known as the opcode byte. Part of the opcode is specified in this field.
MOD R/M Field 2640 (Byte 5) includes MOD field 2642, Reg field 2644, and R/M field 2646. As previously described, the MOD field's 2642 content distinguishes between memory access and non-memory access operations. The role of Reg field 2644 can be summarized to two situations: encoding either the destination register operand or a source register operand or be treated as an opcode extension and not used to encode any instruction operand. The role of R/M field 2646 may include the following: encoding the instruction operand that references a memory address or encoding either the destination register operand or a source register operand.
Scale, Index, Base (SIB) Byte (Byte 6)—as previously described, the content of SIB 2650, including SIB.ss 2652, SIB.xxx 2654, and SIB bbb 2656, is used for memory address generation. SIB.xxx 2654 and SIB.bbb 2656—the contents of these fields have been previously referred to with regard to the register indexes Xxxx and Bbbb.
Displacement field 2562A (Bytes 7-10)—when MOD field 2642 contains 10, bytes 7-10 are the displacement field 2562A, and it works the same as the legacy 32-bit displacement (disp32) and works at byte granularity.
Displacement factor field 2562B (Byte 7)—when MOD field 2642 contains 01, byte 7 is the displacement factor field 2562B. The location of this field is that same as that of the legacy x86 instruction set 8-bit displacement (disp8), which works at byte granularity. Since disp8 is sign extended, it can only address between −128 and 127 bytes offsets; in terms of 64-byte cache lines, disp8 uses 8 bits that can be set to only four really useful values −128, −64, 0, and 64; since a greater range is often needed, disp32 is used; however, disp32 requires 4 bytes. In contrast to disp8 and disp32, the displacement factor field 2562B is a reinterpretation of disp8; when using displacement factor field 2562B, the actual displacement is determined by the content of the displacement factor field multiplied by the size of the memory operand access (N). This type of displacement is referred to as disp8*N. This reduces the average instruction length (a single byte of used for the displacement but with a much greater range). Such compressed displacement assumes that the effective displacement is multiple of the granularity of the memory access, and hence, the redundant low-order bits of the address offset do not need to be encoded. In other words, the displacement factor field 2562B substitutes the legacy x86 instruction set 8-bit displacement. Thus, the displacement factor field 2562B is encoded the same way as an x86 instruction set 8-bit displacement (so no changes in the ModRM/SIB encoding rules) with the only exception that disp8 is overloaded to disp8*N. In other words, there are no changes in the encoding rules or encoding lengths but only in the interpretation of the displacement value by hardware (which needs to scale the displacement by the size of the memory operand to obtain a byte-wise address offset). Immediate field 2572 operates as previously described.
When U=1, the alpha field 2552 (EVEX byte 3, bit [7]—EH) is interpreted as the write mask control (Z) field 2552C. When U=1 and the MOD field 2642 contains 11 (signifying a no memory access operation), part of the beta field 2554 (EVEX byte 3, bit [4]—S0) is interpreted as the RL field 2557A; when it contains a 1 (round 2557A.1) the rest of the beta field 2554 (EVEX byte 3, bit [6-5]—S2-1) is interpreted as the round operation field 2559A, while when the RL field 2557A contains a 0 (VSIZE 2557A.2) the rest of the beta field 2554 (EVEX byte 3, bit [6-5]—S2-1) is interpreted as the vector length field 2559B (EVEX byte 3, bit [6-5]—L1-0). When U=1 and the MOD field 2642 contains 00, 01, or 10 (signifying a memory access operation), the beta field 2554 (EVEX byte 3, bits [6:4]—SSS) is interpreted as the vector length field 2559B (EVEX byte 3, bit [6-5]—L1-0) and the broadcast field 2557B (EVEX byte 3, bit [4]—B).
In other words, the vector length field 2559B selects between a maximum length and one or more other shorter lengths, where each such shorter length is half the length of the preceding length; and instructions templates without the vector length field 2559B operate on the maximum vector length. Further, in one embodiment, the class B instruction templates of the specific vector friendly instruction format 2600 operate on packed or scalar single/double-precision floating-point data and packed or scalar integer data. Scalar operations are operations performed on the lowest order data element position in a zmm/ymm/xmm register; the higher order data element positions are either left the same as they were prior to the instruction or zeroed depending on the embodiment.
Write mask registers 2715—in the embodiment illustrated, there are 8 write mask registers (k0 through k7), each 64 bits in size. In an alternate embodiment, the write mask registers 2715 are 16 bits in size. As previously described, in one embodiment, the vector mask register k0 cannot be used as a write mask; when the encoding that would normally indicate k0 is used for a write mask, it selects a hardwired write mask of 0xFFFF, effectively disabling write masking for that instruction.
General-purpose registers 2725—in the embodiment illustrated, there are sixteen 64-bit general-purpose registers that are used along with the existing x86 addressing modes to address memory operands. These registers are referenced by the names RAX, RBX, RCX, RDX, RBP, RSI, RDI, RSP, and R8 through R15.
Scalar floating-point stack register file (x87 stack) 2745, on which is aliased the MMX packed integer flat register file 2750—in the embodiment illustrated, the x87 stack is an eight-element stack used to perform scalar floating-point operations on 32/64/80-bit floating-point data using the x87 instruction set extension; while the MMX registers are used 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.
Alternative embodiments may use wider or narrower registers. Additionally, alternative embodiments may use more, less, or different register files and registers.
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). 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 that may include on the same die the described CPU (sometimes referred to as the application core(s) or application processor(s)), the above described coprocessor, and additional functionality. Exemplary core architectures are described next, followed by descriptions of exemplary processors and computer architectures.
In
The front-end unit 2830 includes a branch prediction unit 2832 coupled to an instruction cache unit 2834, which is coupled to an instruction translation lookaside buffer (TLB) 2836, which is coupled to an instruction fetch unit 2838, which is coupled to a decode unit 2840. The decode unit 2840 (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 unit 2840 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 embodiment, the core 2890 includes a microcode ROM or other medium that stores microcode for certain macroinstructions (e.g., in decode unit 2840 or otherwise within the front-end unit 2830). The decode unit 2840 is coupled to a rename/allocator unit 2852 in the execution engine unit 2850.
The execution engine unit 2850 includes the rename/allocator unit 2852 coupled to a retirement unit 2854 and a set of one or more scheduler unit(s) 2856. The scheduler unit(s) 2856 represents any number of different schedulers, including reservations stations, central instruction window, etc. The scheduler unit(s) 2856 is coupled to the physical register file(s) unit(s) 2858. Each of the physical register file(s) units 2858 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 embodiment, the physical register file(s) unit 2858 comprises a vector registers unit, a write mask registers unit, and a scalar registers unit. These register units may provide architectural vector registers, vector mask registers, and general-purpose registers. The physical register file(s) unit(s) 2858 is overlapped by the retirement unit 2854 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 unit 2854 and the physical register file(s) unit(s) 2858 are coupled to the execution cluster(s) 2860. The execution cluster(s) 2860 includes a set of one or more execution units 2862 and a set of one or more memory access units 2864. The execution units 2862 may 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 scheduler unit(s) 2856, physical register file(s) unit(s) 2858, and execution cluster(s) 2860 are shown as being possibly plural because certain embodiments 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 unit, physical register file(s) unit, and/or execution cluster—and in the case of a separate memory access pipeline, certain embodiments are implemented in which only the execution cluster of this pipeline has the memory access unit(s) 2864). 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.
The set of memory access units 2864 is coupled to the memory unit 2870, which includes a data TLB unit 2872 coupled to a data cache unit 2874 coupled to a level 2 (L2) cache unit 2876. In one exemplary embodiment, the memory access units 2864 may include a load unit, a store address unit, and a store data unit, each of which is coupled to the data TLB unit 2872 in the memory unit 2870. The instruction cache unit 2834 is further coupled to a level 2 (L2) cache unit 2876 in the memory unit 2870. The L2 cache unit 2876 is coupled to one or more other levels of cache and eventually to a main memory.
By way of example, the exemplary register renaming, out-of-order issue/execution core architecture may implement the pipeline 2800 as follows: 1) the instruction fetch 2838 performs the fetch and length decoding stages 2802 and 2804; 2) the decode unit 2840 performs the decode stage 2806; 3) the rename/allocator unit 2852 performs the allocation stage 2808 and renaming stage 2810; 4) the scheduler unit(s) 2856 performs the schedule stage 2812; 5) the physical register file(s) unit(s) 2858 and the memory unit 2870 perform the register read/memory read stage 2814; the execution cluster 2860 perform the execute stage 2816; 6) the memory unit 2870 and the physical register file(s) unit(s) 2858 perform the write back/memory write stage 2818; 7) various units may be involved in the exception handling stage 2822; and 8) the retirement unit 2854 and the physical register file(s) unit(s) 2858 perform the commit stage 2824.
The core 2890 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, Calif.; the ARM instruction set (with optional additional extensions such as NEON) of ARM Holdings of Sunnyvale, Calif.), including the instruction(s) described herein. In one embodiment, the core 2890 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).
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 also includes separate instruction and data cache units 2834/2874 and a shared L2 cache unit 2876, 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.
Specific Exemplary in-Order Core Architecture
The local subset of the L2 cache 2904 is part of a global L2 cache that is divided into separate local subsets, one per processor core. Each processor core has a direct access path to its own local subset of the L2 cache 2904. Data read by a processor core is stored in its L2 cache subset 2904 and can be accessed quickly, in parallel with other processor cores accessing their own local L2 cache subsets. Data written by a processor core is stored in its own L2 cache subset 2904 and is flushed from other subsets, if necessary. The ring network ensures coherency for shared data. The ring network is bi-directional to allow agents such as processor cores, L2 caches and other logic blocks to communicate with each other within the chip. Each ring data-path is 1012-bits wide per direction.
Thus, different implementations of the processor 3000 may include: 1) a CPU with the special purpose logic 3008 being integrated graphics and/or scientific (throughput) logic (which may include one or more cores), and the cores 3002A-N being one or more general purpose cores (e.g., general purpose in-order cores, general purpose out-of-order cores, a combination of the two); 2) a coprocessor with the cores 3002A-N being a large number of special purpose cores intended primarily for graphics and/or scientific (throughput); and 3) a coprocessor with the cores 3002A-N being a large number of general purpose in-order cores. Thus, the processor 3000 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 3000 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, BiCMOS, CMOS, or NMOS.
The memory hierarchy includes one or more levels of cache within the cores, a set or one or more shared cache units 3006, and external memory (not shown) coupled to the set of integrated memory controller units 3014. The set of shared cache units 3006 may include one or more mid-level caches, such as level 2 (L2), level 3 (L3), level 4 (L4), or other levels of cache, a last level cache (LLC), and/or combinations thereof. While in one embodiment a ring based interconnect unit 3012 interconnects the special purpose logic 3008 (integrated graphics logic is an example of and is also referred to herein as special purpose logic), the set of shared cache units 3006, and the system agent unit 3010/integrated memory controller unit(s) 3014, alternative embodiments may use any number of well-known techniques for interconnecting such units. In one embodiment, coherency is maintained between one or more cache units 3006 and cores 3002A-N.
In some embodiments, one or more of the cores 3002A-N are capable of multithreading. The system agent 3010 includes those components coordinating and operating cores 3002A-N. The system agent unit 3010 may include for example a power control unit (PCU) and a display unit. The PCU may be or include logic and components needed for regulating the power state of the cores 3002A-N and the special purpose logic 3008. The display unit is for driving one or more externally connected displays.
The cores 3002A-N may be homogenous or heterogeneous in terms of architecture instruction set; that is, two or more of the cores 3002A-N may be capable of execution the same instruction set, while others may be capable of executing only a subset of that instruction set or a different instruction set.
Referring now to
The optional nature of additional processors 3115 is denoted in
The memory 3140 may be, for example, dynamic random-access memory (DRAM), phase change memory (PCM), or a combination of the two. For at least one embodiment, the controller hub 3120 communicates with the processor(s) 3110, 3115 via a multi-drop bus, such as a frontside bus (FSB), point-to-point interface such as QuickPath Interconnect (QPI), or similar connection 3195.
In one embodiment, the coprocessor 3145 is a special-purpose processor, such as, for example, a high-throughput MIC processor, a network or communication processor, compression engine, graphics processor, GPGPU, embedded processor, or the like. In one embodiment, controller hub 3120 may include an integrated graphics accelerator.
There can be a variety of differences between the physical resources 3110, 3115 in terms of a spectrum of metrics of merit including architectural, microarchitectural, thermal, power consumption characteristics, and the like.
In one embodiment, the processor 3110 executes instructions that control data processing operations of a general type. Embedded within the instructions may be coprocessor instructions. The processor 3110 recognizes these coprocessor instructions as being of a type that should be executed by the attached coprocessor 3145. Accordingly, the processor 3110 issues these coprocessor instructions (or control signals representing coprocessor instructions) on a coprocessor bus or other interconnect, to coprocessor 3145. Coprocessor(s) 3145 accept and execute the received coprocessor instructions.
Referring now to
Processors 3270 and 3280 are shown including integrated memory controller (IMC) units 3272 and 3282, respectively. Processor 3270 also includes as part of its bus controller units point-to-point (P-P) interfaces 3276 and 3278; similarly, second processor 3280 includes P-P interfaces 3286 and 3288. Processors 3270, 3280 may exchange information via a point-to-point (P-P) interface 3250 using P-P interface circuits 3278, 3288. As shown in
Processors 3270, 3280 may each exchange information with a chipset 3290 via individual P-P interfaces 3252, 3254 using point to point interface circuits 3276, 3294, 3286, 3298. Chipset 3290 may optionally exchange information with the coprocessor 3238 via a high-performance interface 3292. In one embodiment, the coprocessor 3238 is a special-purpose processor, such as, for example, a high-throughput MIC processor, a network or communication processor, compression engine, graphics processor, GPGPU, embedded processor, or the like.
A shared cache (not shown) may be included in either processor or outside of both processors yet connected with the processors via 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.
Chipset 3290 may be coupled to a first bus 3216 via an interface 3296. In one embodiment, first bus 3216 may be a Peripheral Component Interconnect (PCI) bus, or a bus such as a PCI Express bus or another third generation I/O interconnect bus, although the scope of the present invention is not so limited.
As shown in
Referring now to
Referring now to
Embodiments of the mechanisms disclosed herein may be implemented in hardware, software, firmware, or a combination of such implementation approaches. Embodiments 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.
Program code, such as code 3230 illustrated in
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.
One or more aspects of at least one embodiment 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 “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 actually 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 rewritable's (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, embodiments 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 embodiments may also be referred to as program products.
In some cases, an instruction converter may be used to convert an instruction from a source instruction set to a target instruction set. 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 1 provides an exemplary processor including: fetch circuitry to fetch an instruction, decode circuitry to decode the instruction having fields to specify an opcode and locations of a destination and N source matrices, the opcode indicating the processor is to load the N source matrices from memory, perform N convolutions on the N source matrices to generate N feature maps, and store results of the N convolutions in registers to be passed to an activation layer, wherein the processor is to perform the N convolutions and the activation layer with at most one memory load of each of the N source matrices, scheduling circuitry to schedule execution of the instruction, and execution circuitry to execute the instruction as per the opcode.
Example 2 includes the substance of the exemplary processor of Example 1, wherein the processor is to perform each of the N convolutions by convolving a feature identifier over a source matrix one element at a time, each time generating products of each element of the feature identifier and a corresponding element of a receptive field of the source matrix, and storing a sum of the products into a corresponding element of the feature map.
Example 3 includes the substance of the exemplary processor of Example 1, wherein the activation layer includes one of Rectified Linear Unit (ReLU), tanh, sigmoid, and soft max.
Example 4 includes the substance of the exemplary processor of Example 1, wherein N equals three, and elements of the N source matrices include red, green, and blue pixel values.
Example 5 includes the substance of the exemplary processor of Example 1, wherein the opcode further calls for the processor to perform pooling on the results of the N convolutions in order to down-sample each of the N feature maps.
Example 6 includes the substance of the exemplary processor of Example 1, wherein the opcode further calls for the processor to perform a fully connected layer.
Example 7 includes the substance of the exemplary processor of Example 1, wherein the opcode further calls for the processor to pad each of the N source matrices with zeroes, such that each of the N feature maps has the same dimensions as a corresponding one of the N source matrices.
Example 8 includes the substance of the exemplary processor of Example 1, wherein the opcode further calls for the processor to generate N additional feature maps by performing an additional convolution on each of the N feature maps, results of the N additional convolutions to be stored in registers to be passed to the activation layer, wherein the processor is to perform the N convolutions, the N additional convolutions, and the activation layer with at most one memory load of each of the N source matrices.
Example 9 provides an exemplary processor-implemented method including: fetching an instruction using fetch circuitry, decoding, using decode circuitry, the instruction having fields to specify an opcode and locations of a destination and N source matrices, the opcode indicating the processor is to load the N source matrices from memory, perform N convolutions on the N source matrices to generate N feature maps, and store results of the N convolutions in registers to be passed to an activation layer, wherein the processor is to perform the N convolutions and the activation layer with at most one memory load of each of the N source matrices, scheduling, using scheduling circuitry, execution of the instruction, and executing, using execution circuitry, the instruction as per the opcode.
Example 10 includes the substance of the exemplary processor-implemented method of Example 9, wherein the processor is to perform each of the N convolutions by convolving a feature identifier over a source matrix one element at a time, each time generating products of each element of the feature identifier and a corresponding element of a receptive field of the source matrix, and storing a sum of the products into a corresponding element of the feature map.
Example 11 includes the substance of the exemplary processor-implemented method of Example 9, wherein the activation layer includes one of Rectified Linear Unit (ReLU), tanh, sigmoid, and soft max.
Example 12 includes the substance of the exemplary processor-implemented method of Example 9, wherein N equals three, and elements of the N source matrices include red, green, and blue pixel values.
Example 13 includes the substance of the exemplary processor-implemented method of Example 9, wherein the opcode further calls for the processor to perform pooling on the results of the N convolutions in order to down-sample each of the N feature maps.
Example 14 includes the substance of the exemplary processor-implemented method of Example 9, wherein the opcode further calls for the processor to perform a fully connected layer.
Example 15 includes the substance of the exemplary processor-implemented method of Example 9, wherein the opcode further calls for the processor to pad each of the N source matrices with zeroes, such that each of the N feature maps has the same dimensions as a corresponding one of the N source matrices.
Example 16 includes the substance of the exemplary processor-implemented method of Example 9, wherein the opcode further calls for the processor to generate N additional feature maps by performing an additional convolution on each of the N feature maps, results of the N additional convolutions to be stored in registers to be passed to the activation layer, wherein the processor is to perform the N convolutions, the N additional convolutions, and the activation layer with at most one memory load of each of the N source matrices.
Example 17 provides an exemplary non-transitory machine-readable medium containing instructions to which a processor is to respond by: fetching an instruction using fetch circuitry, decoding, using decode circuitry, the instruction having fields to specify an opcode and locations of a destination and N source matrices, the opcode indicating the processor is to load the N source matrices from memory, perform N convolutions on the N source matrices to generate N feature maps, and store results of the N convolutions in registers to be passed to an activation layer, wherein the processor is to perform the N convolutions and the activation layer with at most one memory load of each of the N source matrices, scheduling, using scheduling circuitry, execution of the instruction, and executing, using execution circuitry, the instruction as per the opcode.
Example 18 includes the substance of the exemplary non-transitory machine-readable medium of Example 17, wherein the processor is to perform each of the N convolutions by convolving a feature identifier over a source matrix one element at a time, each time generating products of each element of the feature identifier and a corresponding element of a receptive field of the source matrix, and storing a sum of the products into a corresponding element of the feature map.
Example 19 includes the substance of the exemplary non-transitory machine-readable medium of Example 17, wherein the activation layer includes one of Rectified Linear Unit (ReLU), tanh, sigmoid, and soft max.
Example 20 includes the substance of the exemplary non-transitory machine-readable medium of Example 17, wherein N equals three, and elements of the N source matrices include red, green, and blue pixel values.
