Patent Grant
12045617
This relates to using a streaming engine to insert null elements into a stream of matrix elements.
Digital signal processors (DSP) are optimized for processing streams of data that may be derived from various input signals, such as sensor data, a video stream, a voice channel, radar signals, biomedical signals, etc. Digital signal processors operating on real-time data typically receive an input data stream, perform a filter function on the data stream (such as encoding or decoding) and output a transformed data stream. The system is called real-time because the application fails if the transformed data stream is not available for output when scheduled. Typical video encoding requires a predictable but non-sequential input data pattern. A typical application requires memory access to load data registers in a data register file and then supply data from the data registers to functional units which perform the data processing.
One or more DSP processing cores can be combined with various peripheral circuits, blocks of memory, etc. on a single integrated circuit (IC) die to form a system on chip (SoC). These systems can include multiple interconnected processors that share the use of on-chip and off-chip memory. A processor can include some combination of instruction cache (ICache) and data cache (DCache) to improve processing. Furthermore, multiple processors with shared memory can be incorporated in a single embedded system. The processors can physically share the same memory without accessing data or executing code located in the same memory locations or can use some portion of the shared memory as common shared memory.
Methods and apparatus are provided for software instructions to be executed on a processor within a computer system to configure a steaming engine with stream parameters to define a multidimensional array. The stream parameters define a size for each dimension of the multidimensional array and a specified width for two selected dimensions of the array. Data is fetched from a memory coupled to the streaming engine responsive to the stream parameters. A stream of vectors is formed for the multidimensional array responsive to the stream parameters from the data fetched from memory. When the selected dimension in the stream of vectors exceeds the specified width, the streaming inserts null elements into each portion of a respective vector for the selected dimension that exceeds the specified width in the stream of vectors. Null elements are inserted into respective vectors for each of the two selected dimensions. Stream vectors that are completely null are formed by the streaming engine without accessing the system memory for respective data.
In the drawings, like elements are denoted by like reference numerals for consistency.
Digital signal processors (DSP) are optimized for processing streams of data that may be derived from various input signals, such as sensor data, a video stream, a voice channel, radar signals, biomedical signals, etc. Memory bandwidth and scheduling are concerns for digital signal processors operating on real-time data. An example DSP is described herein that includes a streaming engine to improve memory bandwidth and data scheduling.
One or more DSP can be combined with various peripheral circuits, blocks of memory, etc. on a single integrated circuit (IC) die to form a system on chip (SoC). See, for example, “66AK2Hx Multicore Keystone™ DSP+ARM® System-on-Chip,” 2013 which is incorporated by reference herein.
In the example processor described herein, an autonomous streaming engine (SE) is coupled to the DSP. In this example, the streaming engine includes two closely coupled streaming engines that can manage two data streams simultaneously. In another example, the streaming engine is capable of managing only a single stream, while in other examples the streaming engine is capable of handling more than two streams. In each case, for each stream, the streaming engine includes an address generation stage, a data formatting stage, and some storage for formatted data waiting for consumption by the processor. In the examples described herein, addresses are derived from algorithms that can involve multi-dimensional loops, each dimension maintaining an iteration count. In one example, the streaming engine supports six levels of nested iteration. In other examples, more or fewer levels of iteration are supported.
In the example processor described hereinbelow, control logic and an alignment network enable the streaming engine to form null vectors or padded vectors and to insert them into two dimensions of a stream without accessing the null data or pad data from system memory.
Several examples of forming null vectors and padded vectors are described in more detail with regards to
An example DSP processor is described in detail herein with reference to
Processing unit core 110 fetches instructions from L1I cache 121 as controlled by instruction fetch unit 111. Instruction fetch unit 111 determines the next instructions to be executed and recalls a fetch packet sized set of such instructions. The nature and size of fetch packets are further detailed below. Instructions are directly fetched from L1I cache 121 upon a cache hit if the instructions are stored in L1I cache 121. Upon a cache miss occurring when the specified instructions are not stored in L1I cache 121, the instructions are sought in L2 combined cache 130. In this example, the size of a cache line in L1I cache 121 equals the size of a fetch packet which is 512 bits. The memory locations of these instructions are either a hit in L2 combined cache 130 or a miss. A hit is serviced from L2 combined cache 130. A miss is serviced from a higher level of cache (not illustrated) or from main memory (not illustrated). In this example, the requested instruction is simultaneously supplied to both L1I cache 121 and processing unit core 110 to speed use.
In this example, processing unit core 110 includes multiple functional units to perform instruction specified data processing tasks. Instruction dispatch unit 112 determines the target functional unit of each fetched instruction. In this example, processing unit 110 operates as a very long instruction word (VLIW) processor capable of operating on multiple instructions in corresponding functional units simultaneously. A complier organizes instructions in execute packets that are executed together. Instruction dispatch unit 112 directs each instruction to its target functional unit. The functional unit assigned to an instruction is completely specified by the instruction produced by the compiler. The hardware of processing unit core 110 has no part in the functional unit assignment. In this example, instruction dispatch unit 112 operates on several instructions in parallel. The number of such parallel instructions is set by the size of the execute packet. This is further described herein.
One part of the dispatch task of instruction dispatch unit 112 is determining whether the instruction is to execute on a functional unit in scalar data path side A 115 or vector data path side B 116. An instruction bit within each instruction called the s bit determines which data path the instruction controls. This is further described herein.
Instruction decode unit 113 decodes each instruction in a current execute packet. Decoding includes identification of the functional unit performing the instruction, identification of registers used to supply data for the corresponding data processing operation from among possible register files, and identification of the register destination of the results of the corresponding data processing operation. As further explained below, instructions can include a constant field in place of one register number operand field. The result of this decoding are signals for control of the target functional unit to perform the data processing operation specified by the corresponding instruction on the specified data.
Processing unit core 110 includes control registers 114. Control registers 114 store information for control of the functional units in scalar data path side A 115 and vector data path side B 116. This information may include mode information or the like.
The decoded instructions from instruction decode 113 and information stored in control registers 114 are supplied to scalar data path side A 115 and vector data path side B 116. As a result, functional units within scalar data path side A 115 and vector data path side B 116 perform instruction specified data processing operations upon instruction specified data and store the results in an instruction specified data register or registers. Each of scalar data path side A 115 and vector data path side B 116 include multiple functional units that operate in parallel. These are further described below in conjunction with
Processing unit core 110 includes further non-instruction-based modules. Emulation unit 118 permits determination of the machine state of processing unit core 110 in response to instructions. This capability can be employed for algorithmic development. Interrupts/exceptions unit 119 enables processing unit core 110 to be responsive to external, asynchronous events (interrupts) and to respond to attempts to perform improper operations (exceptions).
Processor 100 includes streaming engine 125. Streaming engine 125 supplies two data streams from predetermined addresses cached in L2 combined cache 130 to register files of vector data path side B of processing unit core 110. This provides controlled data movement from memory (as cached in L2 combined cache 130) directly to functional unit operand inputs. This is further described herein.
L1D cache 123 exchanges data with register files in scalar data path side A 115 via bus 143. Bus 143 is a 64-bit bus in this example. L1D cache 123 exchanges data with register files in vector data path side B 116 via bus 144. Bus 144 is a 512-bit bus in this example. Busses 143 and 144 are illustrated as bidirectional supporting both processing unit core 110 data reads and data writes. L1D cache 123 exchanges data with L2 combined cache 130 via bus 145. Bus 145 is a 512-bit bus in this example. Bus 145 is illustrated as bidirectional supporting cache service for both processing unit core 110 data reads and data writes.
Processor data requests are directly fetched from L1D cache 123 upon a cache hit (if the requested data is stored in L1D cache 123). Upon a cache miss (the specified data is not stored in L1D cache 123), the data is sought in L2 combined cache 130. The memory locations of the requested data are either a hit in L2 combined cache 130 or a miss. A hit is serviced from L2 combined cache 130. A miss is serviced from another level of cache (not illustrated) or from main memory (not illustrated). The requested data may be simultaneously supplied to both L1D cache 123 and processing unit core 110 to speed use.
L2 combined cache 130 supplies data of a first data stream to streaming engine 125 via bus 146. Bus 146 is a 512-bit bus in this example. Streaming engine 125 supplies data of the first data stream to functional units of vector data path side B 116 via bus 147. Bus 147 is a 512-bit bus in this example. L2 combined cache 130 supplies data of a second data stream to streaming engine 125 via bus 148. Bus 148 is a 512-bit bus in this example. Streaming engine 125 supplies data of this second data stream to functional units of vector data path side B 116 via bus 149, which is a 512-bit bus in this example. Busses 146, 147, 148 and 149 are illustrated as unidirectional from L2 combined cache 130 to streaming engine 125 and to vector data path side B 116 in accordance with this example.
Streaming engine data requests are directly fetched from L2 combined cache 130 upon a cache hit (if the requested data is stored in L2 combined cache 130). Upon a cache miss (the specified data is not stored in L2 combined cache 130), the data is sought from another level of cache (not illustrated) or from main memory (not illustrated). It is technically feasible in some examples for L1D cache 123 to cache data not stored in L2 combined cache 130. If such operation is supported, then upon a streaming engine data request that is a miss in L2 combined cache 130, L2 combined cache 130 snoops L1D cache 123 for the streaming engine requested data. If L1D cache 123 stores the data, the snoop response includes the data, which is then supplied to service the streaming engine request. If L1D cache 123 does not store the data, the snoop response indicates this and L2 combined cache 130 services the streaming engine request from another level of cache (not illustrated) or from main memory (not illustrated).
In this example, both L1D cache 123 and L2 combined cache 130 can be configured as selected amounts of cache or directly addressable memory in accordance with U.S. Pat. No. 6,606,686 entitled UNIFIED MEMORY SYSTEM ARCHITECTURE INCLUDING CACHE AND DIRECTLY ADDRESSABLE STATIC RANDOM ACCESS MEMORY, which is incorporated by reference herein.
In this example, processor 100 is fabricated on an integrated chip (IC) that is mounted on a ball grid array (BGA) substrate. A BGA substrate and IC die together may be referred to as “BGA package,” “IC package,” “integrated circuit,” “IC,” “chip,” “microelectronic device,” or similar terminology. The BGA package may include encapsulation material to cover and protect the IC die from damage. In another example, other types of known or later developed packaging techniques may be used with processor 100.
Scalar data path side A 115 includes L1 unit 221. L1 unit 221 generally accepts two 64-bit operands and produces one 64-bit result. The two operands are each recalled from an instruction specified register in either global scalar register file 211 or L1/S1 local register file 212. L1 unit 221 performs the following instruction selected operations: 64-bit add/subtract operations; 32-bit min/max operations; 8-bit Single Instruction Multiple Data (SIMD) instructions such as sum of absolute value, minimum and maximum determinations; circular min/max operations; and various move operations between register files. The result is written into an instruction specified register of global scalar register file 211, L1/S1 local register file 212, M1/N1 local register file 213 or D1/D2 local register file 214.
Scalar data path side A 115 includes S1 unit 222. S1 unit 222 generally accepts two 64-bit operands and produces one 64-bit result. The two operands are each recalled from an instruction specified register in either global scalar register file 211 or L1/S1 local register file 212. In this example, S1 unit 222 performs the same type operations as L1 unit 221. In another example, there may be slight variations between the data processing operations supported by L1 unit 221 and S1 unit 222. The result is written into an instruction specified register of global scalar register file 211, L1/S1 local register file 212, M1/N1 local register file 213 or D1/D2 local register file 214.
Scalar data path side A 115 includes M1 unit 223. M1 unit 223 generally accepts two 64-bit operands and produces one 64-bit result. The two operands are each recalled from an instruction specified register in either global scalar register file 211 or M1/N1 local register file 213. In this example, M1 unit 223 performs the following instruction selected operations: 8-bit multiply operations; complex dot product operations; 32-bit bit count operations; complex conjugate multiply operations; and bit-wise Logical Operations, moves, adds and subtracts. The result is written into an instruction specified register of global scalar register file 211, L1/S1 local register file 212, M1/N1 local register file 213 or D1/D2 local register file 214.
Scalar data path side A 115 includes N1 unit 224. N1 unit 224 generally accepts two 64-bit operands and produces one 64-bit result. The two operands are each recalled from an instruction specified register in either global scalar register file 211 or M1/N1 local register file 213. In this example, N1 unit 224 performs the same type operations as M1 unit 223. There are also double operations (called dual issued instructions) that employ both the M1 unit 223 and the N1 unit 224 together. The result is written into an instruction specified register of global scalar register file 211, L1/S1 local register file 212, M1/N1 local register file 213 or D1/D2 local register file 214.
Scalar data path side A 115 includes D1 unit 225 and D2 unit 226. D1 unit 225 and D2 unit 226 generally each accept two 64-bit operands and each produce one 64-bit result. D1 unit 225 and D2 unit 226 generally perform address calculations and corresponding load and store operations. D1 unit 225 is used for scalar loads and stores of 64 bits. D2 unit 226 is used for vector loads and stores of 512 bits. In this example, D1 unit 225 and D2 unit 226 also perform: swapping, pack and unpack on the load and store data; 64-bit SIMD arithmetic operations; and 64-bit bit-wise logical operations. D1/D2 local register file 214 stores base and offset addresses used in address calculations for the corresponding loads and stores. The two operands are each recalled from an instruction specified register in either global scalar register file 211 or D1/D2 local register file 214. The calculated result is written into an instruction specified register of global scalar register file 211, L1/S1 local register file 212, M1/N1 local register file 213 or D1/D2 local register file 214.
Vector data path side B 116 includes L2 unit 241. L2 unit 241 generally accepts two 512-bit operands and produces one 512-bit result. The two operands are each recalled from an instruction specified register in either global vector register file 231, L2/S2 local register file 232 or predicate register file 234. In this example, L2 unit 241 performs instruction similar to L1 unit 221 except on wider 512-bit data. The result may be written into an instruction specified register of global vector register file 231, L2/S2 local register file 232, M2/N2/C local register file 233 or predicate register file 234.
Vector data path side B 116 includes S2 unit 242. S2 unit 242 generally accepts two 512-bit operands and produces one 512-bit result. The two operands are each recalled from an instruction specified register in either global vector register file 231, L2/S2 local register file 232 or predicate register file 234. In this example, S2 unit 242 performs instructions similar to S1 unit 222. The result is written into an instruction specified register of global vector register file 231, L2/S2 local register file 232, M2/N2/C local register file 233 or predicate register file 234.
Vector data path side B 116 includes M2 unit 243. M2 unit 243 generally accepts two 512-bit operands and produces one 512-bit result. The two operands are each recalled from an instruction specified register in either global vector register file 231 or M2/N2/C local register file 233. In this example, M2 unit 243 performs instructions similar to M1 unit 223 except on wider 512-bit data. The result is written into an instruction specified register of global vector register file 231, L2/S2 local register file 232 or M2/N2/C local register file 233.
Vector data path side B 116 includes N2 unit 244. N2 unit 244 generally accepts two 512-bit operands and produces one 512-bit result. The two operands are each recalled from an instruction specified register in either global vector register file 231 or M2/N2/C local register file 233. In this example, N2 unit 244 performs the same type operations as M2 unit 243. There are also double operations (called dual issued instructions) that employ both M2 unit 243 and the N2 unit 244 together. The result is written into an instruction specified register of global vector register file 231, L2/S2 local register file 232 or M2/N2/C local register file 233.
Vector data path side B 116 includes correlation (C) unit 245. C unit 245 generally accepts two 512-bit operands and produces one 512-bit result. The two operands are each recalled from an instruction specified register in either global vector register file 231 or M2/N2/C local register file 233. In this example, C unit 245 performs “Rake” and “Search” instructions that are used for WCDMA (wideband code division multiple access) encoding/decoding. In this example, C unit 245 can perform up to 512 multiples per clock cycle of a 2-bit PN (pseudorandom number) and 8-bit I/Q (complex number), 8-bit and 16-bit Sum-of-Absolute-Difference (SAD) calculations, up to 512 SADs per clock cycle, horizontal add and horizontal min/max instructions, and vector permutes instructions. C unit 245 also contains 4 vector control registers (CUCR0 to CUCR3) used to control certain operations of C unit 245 instructions. Control registers CUCR0 to CUCR3 are used as operands in certain C unit 245 operations. In some examples, control registers CUCR0 to CUCR3 are used in control of a general permutation instruction (VPERM), and as masks for SIMD multiple DOT product operations (DOTPM) and SIMD multiple Sum-of-Absolute-Difference (SAD) operations. In further examples, control register CUCR0 is used to store the polynomials for Galois Field Multiply operations (GFMPY) and control register CUCR1 is used to store the Galois field polynomial generator function.
Vector data path side B 116 includes P unit 246. Vector predicate (P) unit 246 performs basic logic operations on registers of local predicate register file 234. P unit 246 has direct access to read from and write to predication register file 234. The logic operations include single register unary operations such as NEG (negate) which inverts each bit of the single register, BITCNT (bit count) which returns a count of the number of bits in the single register having a predetermined digital state (1 or 0), RMBD (right most bit detect) which returns a number of bit positions from the least significant bit position (right most) to a first bit position having a predetermined digital state (1 or 0), DECIMATE which selects every instruction specified Nth (1, 2, 4, etc.) bit to output, and EXPAND which replicates each bit an instruction specified N times (2, 4, etc.). The logic operations also include two register binary operations such as AND which is a bitwise AND of data of the two registers, NAND which is a bitwise AND and negate of data of the two registers, OR which is a bitwise OR of data of the two registers, NOR which is a bitwise OR and negate of data of the two registers, and XOR which is exclusive OR of data of the two registers. The logic operations include transfer of data from a predicate register of predicate register file 234 to another specified predicate register or to a specified data register in global vector register file 231. One use of P unit 246 is manipulation of the SIMD vector comparison results for use in control of a further SIMD vector operation. The BITCNT instruction can be used to count the number of 1's in a predicate register to determine the number of valid data elements from a predicate register.
The provision of global register files accessible by all functional units of a side and local register files accessible by some of the functional units of a side is a design choice. In another example, a different accessibility provision could be made, such as employing one type of register file corresponding to the global register files described herein.
Cross path 117 permits limited exchange of data between scalar data path side A 115 and vector data path side B 116. During each operational cycle one 64-bit data word can be recalled from global scalar register file A 211 for use as an operand by one or more functional units of vector data path side B 116 and one 64-bit data word can be recalled from global vector register file 231 for use as an operand by one or more functional units of scalar data path side A 115. Any scalar data path side A 115 functional unit (L1 unit 221, S1 unit 222, M1 unit 223, N1 unit 224, D1 unit 225 and D2 unit 226) can read a 64-bit operand from global vector register file 231. This 64-bit operand is the least significant bits of the 512-bit data in the accessed register of global vector register file 231. Multiple scalar data path side A 115 functional units can employ the same 64-bit cross path data as an operand during the same operational cycle. However, a single 64-bit operand is transferred from vector data path side B 116 to scalar data path side A 115 in a single operational cycle. Any vector data path side B 116 functional unit (L2 unit 241, S2 unit 242, M2 unit 243, N2 unit 244, C unit 245 and P unit 246) can read a 64-bit operand from global scalar register file 211. If the corresponding instruction is a scalar instruction, the cross-path operand data is treated as a 64-bit operand. If the corresponding instruction is a vector instruction, the upper 448 bits of the operand are zero filled. Multiple vector data path side B 116 functional units can employ the same 64-bit cross path data as an operand during the same operational cycle. In one example, a single 64-bit operand is transferred from scalar data path side A 115 to vector data path side B 116 in a single operational cycle.
Streaming engine 125 (
Streaming engine 125 provides several benefits. For example, streaming engine 125 permits multi-dimensional memory accesses, increases the available bandwidth to the functional units minimizes the number of cache miss stalls since the stream buffer bypasses L1D cache 123, and reduces the number of scalar operations required to maintain a loop. Streaming engine 125 also manages address pointers and handles address generation which frees up the address generation instruction slots and D1 unit 225 and D2 unit 226 for other computations.
Processor core 110 (
Fetch phase 1110 includes program address generation (PG) stage 1111, program access (PA) stage 1112 and program receive (PR) stage 1113. During program address generation stage 1111, the program address is generated in the processor and the read request is sent to the memory controller for the L1I cache. During the program access stage 1112, the L1I cache processes the request, accesses the data in its memory and sends a fetch packet to the processor boundary. During the program receive stage 1113, the processor registers the fetch packet.
Instructions are fetched in a fetch packet that includes sixteen 32-bit wide words.
The execution of the individual instructions is partially controlled by a p bit in each instruction. In this example, the p bit is bit 0 of the 32-bit wide slot. The p bit determines whether an instruction executes in parallel with the next instruction. In this example, instructions are scanned from lower to higher address. If the p bit of an instruction is 1, then the next following instruction (higher memory address) is executed in parallel with (in the same cycle as) that instruction. If the p bit of an instruction is 0, then the next following instruction is executed in the cycle after the instruction.
Processor core 110 (
The instructions executing in parallel constitute an execute packet. In this example, an execute packet can contain up to sixteen 32-bit wide slots for sixteen instructions. No two instructions in an execute packet can use the same functional unit. A slot is one of five types: 1) a self-contained instruction executed on one of the functional units of processor core 110 (L1 unit 221, S1 unit 222, M1 unit 223, N1 unit 224, D1 unit 225, D2 unit 226, L2 unit 241, S2 unit 242, M2 unit 243, N2 unit 244, C unit 245 and P unit 246); 2) a unitless instruction such as a NOP (no operation) instruction or multiple NOP instructions; 3) a branch instruction; 4) a constant field extension; and 5) a conditional code extension. Some of these slot types are further explained herein.
Dispatch and decode phases 1120 (
Execution phase 1130 includes execution (E1 to E5) stages 1131 to 1135. Different types of instructions require different numbers of such stages to complete execution. The execution stages of the pipeline play an important role in understanding the device state at processor cycle boundaries.
During E1 stage 1131, the conditions for the instructions are evaluated and operands are operated on. As illustrated in
During E2 stage 1132, load instructions send the address to memory. Store instructions send the address and data to memory. Single-cycle instructions that saturate results set the SAT bit in the control status register (CSR) if saturation occurs. For 2-cycle instructions, results are written to a destination register file.
During E3 stage 1133, data memory accesses are performed. Any multiply instructions that saturate results set the SAT bit in the control status register (CSR) if saturation occurs. For 3-cycle instructions, results are written to a destination register file.
During E4 stage 1134, load instructions bring data to the processor boundary. For 4-cycle instructions, results are written to a destination register file.
During E5 stage 1135, load instructions write data into a register as illustrated schematically in
The creg field 1301 (bits 29 to 31) and the z bit 1302 (bit 28) are optional fields used in conditional instructions. The bits are used for conditional instructions to identify the predicate register and the condition. The z bit 1302 (bit 28) indicates whether the predication is based upon zero or not zero in the predicate register. If z=1, the test is for equality with zero. If z=0, the test is for nonzero. The case of creg=0 and z=0 is treated as true to allow unconditional instruction execution. The creg field 1301 and the z field 1302 are encoded in the instruction as shown in Table 1.
Execution of a conditional instruction is conditional upon the value stored in the specified data register. The data register is in the global scalar register file 211 for all functional units. Note that “z” in the z bit column refers to the zero/not zero comparison selection noted above and “x” is a don't care state. This coding specifies a subset of the sixteen global registers as predicate registers which preserves bits in the instruction coding. Note that unconditional instructions do not have the optional bits. For unconditional instructions, the bits in fields 1301 and 1302 (28 to 31) are used as additional opcode bits.
The dst field 1303 (bits 23 to 27) specifies a register in a corresponding register file as the destination of the instruction results.
The src2/cst field 1304 (bits 18 to 22) has several meanings depending on the instruction opcode field (bits 3 to 12 for all instructions and additionally bits 28 to 31 for unconditional instructions). One meaning specifies a register of a corresponding register file as the second operand. Another meaning is an immediate constant. Depending on the instruction type, the field 1304 is treated as an unsigned integer and zero extended to a specified data length or is treated as a signed integer and sign extended to the specified data length.
The src1 field 1305 (bits 13 to 17) specifies a register in a corresponding register file as the first source operand.
The opcode field 1306 (bits 3 to 12) for all instructions (and additionally bits 28 to 31 for unconditional instructions) specifies the type of instruction and designates appropriate instruction options including unambiguous designation of the functional unit used and operation performed. A detailed explanation of the opcode is beyond the scope of this description except for the instruction options described below.
The e bit 1307 (bit 2) is used for immediate constant instructions where the constant can be extended. If e=1, then the immediate constant is extended in a manner described below. If e=0, then the immediate constant is not extended and the immediate constant is specified by the src2/cst field 1304 (bits 18 to 22). Note that the e bit 1307 is used for some instructions. Accordingly, with proper coding, the e bit 1307 can be omitted from some instructions and the bit can be used as an additional opcode bit.
The s bit 1308 (bit 1) designates scalar data path side A 115 or vector data path side B 116. If s=0, then scalar data path side A 115 is selected which limits the functional unit to L1 unit 221, S1 unit 222, M1 unit 223, N1 unit 224, D1 unit 225 and D2 unit 226 and the corresponding register files illustrated in
The p bit 1309 (bit 0) marks the execute packets. The p-bit determines whether the instruction executes in parallel with the following instruction. The p-bits are scanned from lower to higher address. If p=1 for the current instruction, then the next instruction executes in parallel with the current instruction. If p=0 for the current instruction, then the next instruction executes in the cycle after the current instruction. All instructions executing in parallel constitute an execute packet. An execute packet can contain up to sixteen instructions. Each instruction in an execute packet uses a different functional unit.
There are two different condition code extension slots. Each execute packet can contain one each of these unique 32-bit condition code extension slots which contains the 4-bit creg/z fields for the instructions in the same execute packet.
Both condition code extension slot 0 and condition code extension slot 1 can include a p bit to define an execute packet as described above in conjunction with
There are two different 32-bit constant extension slots. Each execute packet can contain one each of the unique constant extension slots which contains 27 bits to be concatenated as high order bits with the 5-bit constant field 1305 to form a 32-bit constant. As noted in the instruction coding description above, some instructions define the src2/cst field 1304 as a constant rather than a source register identifier. At least some of such instructions can employ a constant extension slot to extend the constant to 32 bits.
Constant extension slot 0 and constant extension slot 1 are used as follows. The target instruction is of the type permitting constant specification. In this example, the extension is implemented by replacing one input operand register specification field with the least significant bits of the constant as described above with respect to scr2/cst field 1304. Instruction decoder 113 determines this case, known as an immediate field, from the instruction opcode bits. The target instruction also includes one constant extension bit (e bit 1307) dedicated to signaling whether the specified constant is not extended (constant extension bit=0) or extended (constant extension bit=1). If instruction decoder 113 detects a constant extension slot 0 or a constant extension slot 1, instruction decoder 113 further checks the other instructions within the execute packet for an instruction corresponding to the detected constant extension slot. A constant extension is made if one corresponding instruction has a constant extension bit (e bit 1307) equal to 1.
Sign extension unit 1703 forms the final operand value from the input from multiplexer 1703. Sign extension unit 1703 receives control inputs Scalar/Vector and Data Size. The Scalar/Vector input indicates whether the corresponding instruction is a scalar instruction or a vector instruction. The functional units of data path side A 115 (L1 unit 221, 51 unit 222, M1 unit 223, N1 unit 224, D1 unit 225 and D2 unit 226) perform scalar instructions. Any instruction directed to one of these functional units is a scalar instruction. Data path side B functional units L2 unit 241, S2 unit 242, M2 unit 243, N2 unit 244 and C unit 245 can perform scalar instructions or vector instructions. Instruction decoder 113 determines whether the instruction is a scalar instruction or a vector instruction from the opcode bits. P unit 246 may perform scalar instructions. The Data Size can be eight bits (byte B), sixteen bits (half-word H), 32 bits (word W), or 64 bits (double word D).
Table 2 lists the operation of sign extension unit 1703 for the various options.
Both constant extension slot 0 and constant extension slot 1 can include a p bit to define an execute packet as described above in conjunction with
An execute packet can include a constant extension slot 0 or 1 and more than one corresponding instruction marked constant extended (e bit=1). For such an occurrence, for constant extension slot 0, more than one of an L1 unit 221 instruction, data in a D1 unit 225 instruction, an S2 unit 242 instruction, an offset in a D2 unit 226 instruction, an M2 unit 243 instruction or an N2 unit 244 instruction in an execute packet can have an e bit of 1. For such an occurrence, for constant extension slot 1, more than one of an L2 unit 241 instruction, data in a D2 unit 226 instruction, an S1 unit 222 instruction, an offset in a D1 unit 225 instruction, an M1 unit 223 instruction or an N1 unit 224 instruction in an execute packet can have an e bit of 1. In one example, instruction decoder 113 determines that such an occurrence is an invalid operation and not supported. Alternately, the combination can be supported with extension bits of the constant extension slot applied to each corresponding functional unit instruction marked constant extended.
L1 unit 221, S1 unit 222, L2 unit 241, S2 unit 242 and C unit 245 often operate in a single instruction multiple data (SIMD) mode. In this SIMD mode, the same instruction is applied to packed data from the two operands. Each operand holds multiple data elements disposed in predetermined slots. SIMD operation is enabled by carry control at the data boundaries. Such carry control enables operations on varying data widths.
Operation on data sizes that are integral powers of 2 (2N) is common. However, the carry control technique is not limited to integral powers of 2 and can be applied to other data sizes and operand widths.
In this example, at least L2 unit 241 and S2 unit 242 employ two types of SIMD instructions using registers in predicate register file 234. In this example, the SIMD vector predicate instructions operate on an instruction specified data size. The data sizes include byte (8 bit) data, half word (16 bit) data, word (32 bit) data, double word (64 bit) data, quad word (128 bit) data and half vector (256 bit) data. In the first of these instruction types, the functional unit (L unit 241 or S unit 242) performs a SIMD comparison on packed data in two general data registers and supplies results to a predicate data register. The instruction specifies a data size, the two general data register operands, and the destination predicate register. In this example, each predicate data register includes one bit corresponding to each minimal data size portion of the general data registers. In the current example, the general data registers are 512 bits (64 bytes) and the predicate data registers are 64 bits (8 bytes). Each bit of a predicate data register corresponds to eight bits of a general data register. The comparison is performed on a specified data size (8, 16, 32, 64, 128 or 256 bits). If the comparison is true, then the functional unit supplies 1's to all predicate register bits corresponding the that data size portion. If the comparison is false, the functional unit supplies zeroes to the predicate register bits corresponding to that data size portion. In this example, the enabled comparison operations include: less than, greater than, and equal to.
In the second of the instruction types, the functional unit (L2 unit 241 or S2 unit 242) separately performs a first SIMD operation or a second SIMD operation on packed data in general data registers based upon the state of data in a predicate data register. The instruction specifies a data size, one or two general data register operands, a controlling predicate register, and a general data register destination. For example, a functional unit can select, for each data sized portion of two vector operands, a first data element of a first operand or a second data element of a second operand dependent upon the 1/0 state of corresponding bits in the predicate data register to store in the destination register. In another example, the data elements of a single vector operand can be saved to memory or not saved dependent upon the data of the corresponding bits of the predicate register.
The operations of P unit 245 permit a variety of compound vector SIMD operations based upon more than one vector comparison. For example, a range determination can be made using two comparisons. In a SIMD operation, a candidate vector is compared with a vector reference having the minimum of the range packed within a data register. The greater than result is scalar data with bits corresponding to the SIMD data width set to 0 or 1 depending upon the SIMD comparison and is stored in a predicate data register. Another SIMD comparison of the candidate vector is performed with another reference vector having the maximum of the range packed within a different data register produces another scalar with less than results stored in another predicate register. The P unit then ANDs the two predicate registers. The AND result indicates whether each SIMD data part of the candidate vector is within range or out of range. A P unit BITCNT instruction of the AND result can produce a count of the data elements within the comparison range. The P unit NEG function can be used to convert various expressions, such as: a less than comparison result to a greater than or equal comparison result; a greater than comparison result to a less than or equal to comparison result; or, an equal to comparison result to a not equal to comparison result.
Streaming Engine
Stream elements typically reside in system memory. The memory imposes no particular structure upon the stream. Programs define streams and thereby impose structure by specifying the stream attributes such as address of the first element of the stream, size and type of the elements in the stream, formatting for data in the stream, and the address sequence associated with the stream.
The streaming engine defines an address sequence for elements of the stream in terms of a pointer walking through memory. A multiple-level nested loop controls the path the pointer takes. An iteration count for a loop level indicates the number of times the level repeats. A dimension gives the distance between pointer positions of the loop level.
In a basic forward stream, the innermost loop consumes physically contiguous elements from memory as the implicit dimension of the innermost loop is one element. The pointer moves from element to element in consecutive, increasing order. In each level outside the inner loop, that loop moves the pointer to a new location based on the size of the dimension of the loop level.
This form of addressing allows programs to specify regular paths through memory using a small number of parameters. Table 4 lists the addressing parameters of a basic stream.
In this example, ELEM_BYTES ranges from 1 to 64 bytes as shown in Table 5.
The definition above maps consecutive elements of the stream to increasing addresses in memory which is appropriate for many algorithms. Some algorithms are better served by reading elements in decreasing memory address order or reverse stream addressing. For example, a discrete convolution computes vector dot-products, as illustrated by expression (1).
In expression (1), f[ ] and g[ ] represent arrays in memory. For each output, the algorithm reads f[ ] in the forward direction and reads g[ ] in the reverse direction. Practical filters limit the range of indices for [x] and [t−x] to a finite number of elements. To support this pattern, the streaming engine supports reading elements in decreasing address order.
Matrix multiplication presents a unique problem to the streaming engine. Each element in the matrix product is a vector dot product between a row from the first matrix and a column from the second. Programs typically store matrices in row-major or column-major order. Row-major order stores all the elements of a single row contiguously in memory. Column-major order stores all elements of a single column contiguously in memory. Matrices are typically stored in the same order as the default array order for the language. As a result, only one of the two matrices in a matrix multiplication map on to the 2-dimensional stream definition of the streaming engine. In a typical example, an index steps through columns on one array and rows of the other array. The streaming engine supports implicit matrix transposition with transposed streams. Transposed streams avoid the cost of explicitly transforming the data in memory. Instead of accessing data in strictly consecutive-element order, the streaming engine effectively interchanges the inner two loop dimensions of the traversal order, fetching elements along the second dimension into contiguous vector lanes.
This algorithm works but is impractical to implement for small element sizes. Some algorithms work on matrix tiles which are multiple columns and rows together. Therefore, the streaming engine defines a separate transposition granularity. The hardware imposes a minimum granularity. The transpose granularity needs to be at least as large as the element size. Transposition granularity causes the streaming engine to fetch one or more consecutive elements from dimension 0 before moving along dimension 1. When the granularity equals the element size, a single column from a row-major array is fetched. Otherwise, the granularity specifies fetching two, four or more columns at a time from a row-major array. This is also applicable for column-major layout by exchanging row and column in the description. A parameter GRANULE indicates the transposition granularity in bytes.
Another common matrix multiplication technique exchanges the innermost two loops of the matrix multiply. The resulting inner loop no longer reads down the column of one matrix while reading across the row of another. For example, the algorithm may hoist one term outside the inner loop, replacing it with the scalar value. The innermost loop can be implemented with a single scalar by vector multiply followed by a vector add. Or, the scalar value can be duplicated across the length of the vector and a vector by vector multiply used. The streaming engine of this example directly supports the latter case and related use models with an element duplication mode. In this mode, the streaming engine reads a granule smaller than the full vector size and replicates that granule to fill the next vector output.
The streaming engine treats each complex number as a single element with two sub-elements that give the real and imaginary (rectangular) or magnitude and angle (polar) portions of the complex number. Not all programs or peripherals agree what order these sub-elements should appear in memory. Therefore, the streaming engine offers the ability to swap the two sub-elements of a complex number with no cost. The feature swaps the halves of an element without interpreting the contents of the element and can be used to swap pairs of sub-elements of any type, not just complex numbers.
Algorithms generally prefer to work at high precision, but high precision values require more storage and bandwidth than lower precision values. Commonly, programs store data in memory at low precision, promote those values to a higher precision for calculation, and then demote the values to lower precision for storage. The streaming engine supports such operations directly by allowing algorithms to specify one level of type promotion. In this example, every sub-element can be promoted to a larger type size with either sign or zero extension for integer types. In some examples, the streaming engine supports floating point promotion, promoting 16-bit and 32-bit floating point values to 32-bit and 64-bit formats, respectively.
While the streaming engine defines a stream as a discrete sequence of data elements, the processing unit core 110 consumes data elements packed contiguously in vectors. The vectors resemble streams as the vectors contain multiple homogeneous elements with some implicit sequence. Because the streaming engine reads streams, but the processing unit core 110 consumes vectors, the streaming engine maps streams onto vectors in a consistent way.
Vectors include equal-sized lanes, each lane containing a sub-element. The processing unit core 110 designates the rightmost lane of the vector as lane 0, regardless of current endian mode. Lane numbers increase right-to-left. The actual number of lanes within a vector varies depending on the length of the vector and the data size of the sub-element.
Formatting section 2020 includes various formatting blocks. The formatting performed within formatter 1903 by the blocks is further described below. Complex swap block 2021 optionally swaps two sub-elements forming a complex number element. Type promotion block 2022 optionally promotes each data element into a larger data size. Promotion includes zero extension for unsigned integers and sign extension for signed integers. Decimation block 2023 optionally decimates the data elements. In this example, decimation can be 2:1 retaining every other data element or 4:1 retaining every fourth data element. Element duplication block 2024 optionally duplicates individual data elements. In this example, the data element duplication is an integer power of 2 (2N, where N is an integer) including 2×, 4×, 8×, 16×, 32× and 64×. In this example, data duplication can extend over multiple destination vectors. Vector length masking/group duplication block 2025 has two primary functions. An independently specified vector length VECLEN controls the data elements supplied to each output data vector. When group duplication is off, excess lanes in the output data vector are zero filled and these lanes are marked invalid. When group duplication is on, input data elements of the specified vector length are duplicated to fill the output data vector.
Output section 2030 holds the data for output to the corresponding functional units. Register and buffer for processor 2031 stores a formatted vector of data to be used as an operand by the functional units of processing unit core 110 (
The streaming engine maps the innermost stream dimension directly to vector lanes. The streaming engine maps earlier elements within the innermost stream dimension to lower lane numbers and later elements to higher lane numbers, regardless of whether the stream advances in increasing or decreasing address order. Whatever order the stream defines, the streaming engine deposits elements in vectors in increasing-lane order. For non-complex data, the streaming engine places the first element in lane 0 of the vector processing unit core 110 (
The streaming engine fills each vector processing unit core 110 fetches with as many elements as possible from the innermost stream dimension. If the innermost dimension is not a multiple of the vector length, the streaming engine zero pads the dimension to a multiple of the vector length. As noted below, the streaming engine also marks the lanes invalid. Thus, for higher-dimension streams, the first element from each iteration of an outer dimension arrives in lane 0 of a vector. The streaming engine maps the innermost dimension to consecutive lanes in a vector. For transposed streams, the innermost dimension includes groups of sub-elements along dimension 1, not dimension 0, as transposition exchanges these two dimensions.
Two-dimensional (2D) streams exhibit greater variety as compared to one-dimensional streams. A basic 2D stream extracts a smaller rectangle from a larger rectangle. A transposed 2D stream reads a rectangle column-wise instead of row-wise. A looping stream, where the second dimension overlaps first, executes a finite impulse response (FIR) filter taps which loops repeatedly over FIR filter samples providing a sliding window of input samples.
Thus, the iteration count in the 0-dimension 2321 is nine and the iteration count in the 1-dimension 2322 is thirteen. Note that the ELEM_BYTES scales the innermost dimension. The first dimension has ICNT0 elements of size ELEM_BYTES. The stream address generator does not scale the outer dimensions. Therefore, DIM1=88, which is eleven elements scaled by eight bytes per element.
Transposed streams are accessed along dimension 1 before dimension 0. The following examples illustrate transposed streams with varying transposition granularity.
The streams examined so far read each element from memory exactly once. A stream can read a given element from memory multiple times, in effect looping over a portion of memory. FIR filters exhibit two common looping patterns: re-reading the same filter taps for each output and reading input samples from a sliding window. Two consecutive outputs need inputs from two overlapping windows.
Each streaming engine 125 includes a respective dedicated 6-dimensional (6D) stream address generator 2811/2821 that can each generate one new non-aligned request per cycle. As is further described herein, address generators 2811/2821 output 512-bit aligned addresses that overlap the elements in the sequence defined by the stream parameters.
Each address generator 2811/2821 connects to a respective dedicated micro table look-aside buffer (μTLB) 2812/2822. The μTLB 2812/2822 converts a single 48-bit virtual address to a 44-bit physical address each cycle. Each μTLB 2812/2822 has 8 entries, covering a minimum of 32 kB with 4 kB pages or a maximum of 16 MB with 2 MB pages. Each address generator 2811/2821 generates 2 addresses per cycle. The μTLB 2812/2822 only translates one address per cycle. To maintain throughput, streaming engine 125 operates under the assumption that most stream references are within the same 4 kB page. Thus, the address translation does not modify bits 0 to 11 of the address. If aout0 and aout1 line in the same 4 kB page (aout0[47:12] are the same aout1[47:12]), then the μTLB 2812/2822 only translates aout0 and reuses the translation for the upper bits of both addresses.
Translated addresses are queued in respective command queue 2813/2823. These addresses are aligned with information from the respective corresponding Storage Allocation and Tracking block 2814/2824. Streaming engine 125 does not explicitly manage μTLB 2812/2822. The system memory management unit (MMU) invalidates μTLBs as necessary during context switches.
Storage Allocation and Tracking 2814/2824 manages the internal storage of the stream, discovering data reuse and tracking the lifetime of each piece of data. The block accepts two virtual addresses per cycle and binds those addresses to slots in the internal storage. The data store is organized as an array of slots. The streaming engine maintains following metadata to track the contents and lifetime of the data in each slot: 49-bit virtual address associated with the slot, valid bit indicating valid address, ready bit indicating data has arrived for the address, active bit indicating if there are any references outstanding to this data, and a last reference value indicating the most recent reference to this slot in the reference queue. The storage allocation and tracking are further described herein.
Respective reference queue 2815/2825 stores the sequence of references generated by the respective corresponding address generator 2811/2821. The reference sequence enables the data formatting network to present data to processing unit core 110 in the correct order. Each entry in respective reference queue 2815/2825 contains the information necessary to read data out of the data store and align the data for processing unit core 110. Respective reference queue 2815/2825 maintains the information listed in Table 6 in each slot.
Storage allocation and tracking 2814/2824 inserts references in reference queue 2815/2825 as address generator 2811/2821 generates new addresses. Storage allocation and tracking 2814/2824 removes references from reference queue 2815/2825 when the data becomes available and there is room in the stream head registers. As storage allocation and tracking 2814/2824 removes slot references from reference queue 2815/2825 and formats data, the references are checked for the last reference to the corresponding slots. Storage allocation and tracking 2814/2824 compares reference queue 2815/2825 removal pointer against the recorded last reference of the slot. If the pointer and the recorded last reference match, then storage allocation and tracking 2814/2824 marks the slot inactive once the data is no longer needed.
Streaming engine 125 has respective data storage 2816/2826 for a selected number of elements. Deep buffering allows the streaming engine to fetch far ahead in the stream, hiding memory system latency. Each data storage 2816/2826 accommodates two simultaneous read operations and two simultaneous write operations per cycle and each is therefore referred to a two-read, two-write (2r2w) data storage. In other examples, the amount of buffering can be different. In the current example, streaming engine 125 dedicates 32 slots to each stream with each slot tagged by a virtual address. Each slot holds 64 bytes of data in eight banks of eight bytes.
Data storage 2816/2826 and the respective storage allocation/tracking logic 2814/2824 and reference queues 2815/2825 implement the data FIFO 1902 discussed with reference to
Respective butterfly network 2817/2827 includes a seven-stage butterfly network that implements the formatter 1903 (
Streaming engine 125 attempts to fetch and format data ahead of processing unit core 110's demand in order to maintain full throughput. Respective stream head registers 2818/2828 provide a small amount of buffering so that the process remains fully pipelined. Respective stream head registers 2818/2828 are not directly architecturally visible. Each stream also has a respective stream valid register 2819/2829. Valid registers 2819/2829 indicate which elements in the corresponding stream head registers 2818/2828 are valid. The outputs of stream head registers 2818/2828 and valid registers 2819/2829 are provided to processing unit core 110 via buses 2840/2841.
The two streams 2810/2820 share a pair of independent L2 interfaces 2830: L2 Interface A (IFA) 2833 and L2 Interface B (IFB) 2834. Each L2 interface provides 512 bits/cycle throughput direct to the L2 controller 130 (
To maximize performance, in this example both streams can use both L2 interfaces, allowing a single stream to send a peak command rate of two requests per cycle. Each interface prefers one stream over the other, but this preference changes dynamically from request to request. IFA 2833 and IFB 2834 prefer opposite streams, when IFA 2833 prefers Stream 0, IFB 2834 prefers Stream 1 and vice versa.
Respective arbiter 2831/2832 ahead of each respective interface 2833/2834 applies the following basic protocol on every cycle having credits available. Arbiter 2831/2832 checks if the preferred stream has a command ready to send. If so, arbiter 2831/2832 chooses that command. Arbiter 2831/2832 next checks if an alternate stream has at least two requests ready to send, or one command and no credits. If so, arbiter 2831/2832 pulls a command from the alternate stream. If either interface issues a command, the notion of preferred and alternate streams swap for the next request. Using this algorithm, the two interfaces dispatch requests as quickly as possible while retaining fairness between the two streams. The first rule ensures that each stream can send a request on every cycle that has available credits. The second rule provides a mechanism for one stream to borrow the interface of the other when the second interface is idle. The third rule spreads the bandwidth demand for each stream across both interfaces, ensuring neither interface becomes a bottleneck.
Respective coarse grain rotator 2835/2836 enables streaming engine 125 to support a transposed matrix addressing mode. In this mode, streaming engine 125 interchanges the two innermost dimensions of the multidimensional loop to access an array column-wise rather than row-wise. Respective rotators 2835/2836 are not architecturally visible.
Loop 0 is the innermost loop and loop 5 is the outermost loop. In the current example, DIM0 is equal to ELEM_BYTES defining physically contiguous data. Thus, the stream template register 2900 does not define DIM0. Streaming engine 125 interprets iteration counts as unsigned integers and dimensions as unscaled signed integers. An iteration count of zero at any level (ICNT0, ICNT1, ICNT2, ICNT3, ICNT4 or ICNT5) indicates an empty stream. Each iteration count must be at least one to define a valid stream. The template above specifies the type of elements, length and dimensions of the stream. The stream instructions separately specify a start address, e.g., by specification of a scalar register in scalar register file 211 which stores the start address. Thus, a program can open multiple streams using the same template but different registers storing the start address.
The Element Type (ELTYPE) field 3001 defines the data type of the elements in the stream. The coding of the four bits of the ELTYPE field 3001 is defined as shown in Table 9.
Real/Complex Type determines whether the streaming engine treats each element as a real number or two parts (real/imaginary or magnitude/angle) of a complex number and also specifies whether to swap the two parts of complex numbers. Complex types have a total element size twice the sub-element size. Otherwise, the sub-element size equals the total element size.
Sub-Element Size determines the type for purposes of type promotion and vector lane width. For example, 16-bit sub-elements get promoted to 32-bit sub-elements or 64-bit sub-elements when a stream requests type promotion. The vector lane width matters when processing unit core 110 (
Total Element Size specifies the minimal granularity of the stream which determines the number of bytes the stream fetches for each iteration of the innermost loop. Streams read whole elements, either in increasing or decreasing order. Therefore, the innermost dimension of a stream spans ICNT0×total-element-size bytes.
The TRANSPOSE field 3002 determines whether the streaming engine accesses the stream in a transposed order. The transposed order exchanges the inner two addressing levels. The TRANSPOSE field 3002 also indicated the granularity for transposing the stream. The coding of the three bits of the TRANSPOSE field 3002 is defined as shown in Table 10 for normal 2D operations.
Streaming engine 125 can transpose data elements at a different granularity than the element size thus allowing programs to fetch multiple columns of elements from each row. The transpose granularity cannot be smaller than the element size. The TRANSPOSE field 3002 interacts with the DIMFMT field 3009 in a manner further described below.
The PROMOTE field 3003 controls whether the streaming engine promotes sub-elements in the stream and the type of promotion. When enabled, streaming engine 125 promotes types by powers-of-2 sizes. The coding of the three bits of the PROMOTE field 3003 is defined as shown in Table 11.
When PROMOTE is 000, corresponding to a 1× promotion, each sub-element is unchanged and occupies a vector lane equal in width to the size specified by ELTYPE. When PROMOTE is 001, corresponding to a 2× promotion and zero extend, each sub-element is treated as an unsigned integer and zero extended to a vector lane twice the width specified by ELTYPE. A 2× promotion is invalid for an initial sub-element size of 64 bits. When PROMOTE is 010, corresponding to a 4× promotion and zero extend, each sub-element is treated as an unsigned integer and zero extended to a vector lane four times the width specified by ELTYPE. A 4× promotion is invalid for an initial sub-element size of 32 or 64 bits. When PROMOTE is 011, corresponding to an 8× promotion and zero extend, each sub-element is treated as an unsigned integer and zero extended to a vector lane eight times the width specified by ELTYPE. An 8× promotion is invalid for an initial sub-element size of 16, 32 or 64 bits. When PROMOTE is 101, corresponding to a 2× promotion and sign extend, each sub-element is treated as a signed integer and sign extended to a vector lane twice the width specified by ELTYPE. A 2× promotion is invalid for an initial sub-element size of 64 bits. When PROMOTE is 110, corresponding to a 4× promotion and sign extend, each sub-element is treated as a signed integer and sign extended to a vector lane four times the width specified by ELTYPE. A 4× promotion is invalid for an initial sub-element size of 32 or 64 bits. When PROMOTE is 111, corresponding to an 8× promotion and zero extend, each sub-element is treated as a signed integer and sign extended to a vector lane eight times the width specified by ELTYPE. An 8× promotion is invalid for an initial sub-element size of 16, 32 or 64 bits.
The VECLEN field 3004 defines the stream vector length for the stream in bytes. Streaming engine 125 breaks the stream into groups of elements that are VECLEN bytes long. The coding of the three bits of the VECLEN field 3004 is defined as shown in Table 12.
VECLEN cannot be less than the product of the element size in bytes and the duplication factor. As shown in Table 11, the maximum VECLEN of 64 bytes equals the preferred vector size of vector data path side B 116. When VECLEN is shorter than the native vector width of processing unit core 110, streaming engine 125 pads the extra lanes in the vector provided to processing unit core 110. The GRDUP field 3006 determines the type of padding. The VECLEN field 3004 interacts with ELDUP field 3005 and GRDUP field 3006 in a manner detailed below.
The ELDUP field 3005 specifies the number of times to duplicate each element. The element size multiplied with the element duplication amount cannot exceed the 64 bytes. The coding of the three bits of the ELDUP field 3005 is defined as shown in Table 13.
The ELDUP field 3005 interacts with VECLEN field 3004 and GRDUP field 3006 in a manner detailed below. The nature of the relationship between the permitted element size, the element duplication factor, and the destination vector length requires that a duplicated element that overflows the first destination register fills an integer number of destination registers upon completion of duplication. The data of the additional destination registers eventually supplies the respective stream head register 2818/2828. Upon completion of duplication of a first data element, the next data element is rotated down to the least significant bits of source register 3100 discarding the first data element. The process then repeats for the new data element.
The GRDUP bit 3006 determines whether group duplication is enabled. If GRDUP bit 3006 is 0, then group duplication is disabled. If the GRDUP bit 3006 is 1, then group duplication is enabled. When enabled by GRDUP bit 3006, streaming engine 125 duplicates a group of elements to fill the vector width. VECLEN field 3004 defines the length of the group to replicate. When VECLEN field 3004 is less than the vector length of processing unit core 110 and GRDUP bit 3006 enables group duplication, streaming engine 125 fills the extra lanes (see
The GRDUP field 3006 specifies how streaming engine 125 pads stream vectors for bits following the VECLEN length to the vector length of processing unit core 110. When GRDUP bit 3006 is 0, streaming engine 125 fills the extra lanes with zeros and marks the extra vector lanes invalid. When GRDUP bit 3006 is 1, streaming engine 125 fills extra lanes with copies of the group of elements in each stream vector. Setting GRDUP bit 3006 to 1 has no effect when VECLEN is set to the native vector width of processing unit core 110. VECLEN must be at least as large as the product of ELEM_BYTES and the element duplication factor ELDUP. That is, an element or the duplication factor number of elements cannot be separated using VECLEN.
Group duplication operates to the destination vector size. Group duplication does not change the data supplied when the product of the element size ELEM_BYTES and element duplication factor ELDUP equals or exceeds the destination vector width. Under such conditions, the states of the GRDUP bit 3006 and the VECLEN field 3004 have no effect on the supplied data.
The set of examples below illustrate the interaction between VECLEN and GRDUP. Each of the following examples show how the streaming engine maps a stream onto vectors across different stream vector lengths and the vector size of vector data path side B 116. The stream of this example includes twenty-nine elements (E0 to E28) of 64 bits/8 bytes. The stream can be a linear stream of twenty-nine elements or an inner loop of 29 elements. The tables illustrate eight byte lanes such as shown in
Table 14 illustrates how the example stream maps onto bits within the 64-byte processor vectors when VECLEN is 64 bytes.
As shown in Table 14, the stream extends over four vectors. As previously described, the lanes within vector 4 that extend beyond the stream are zero filled. When VECLEN has a size equal to the native vector length, the value of GRDUP does not matter as no duplication can take place with such a VECLEN.
Table 15 shows the same parameters as shown in Table 14, except with VECLEN of 32 bytes. Group duplicate is disabled (GRDUP=0).
The twenty-nine elements of the stream are distributed over lanes 0 to 3 in eight vectors. Extra lanes 4 to 7 in vectors 1-7 are zero filled. In vector 8, lane 1 has a stream element (E28) and the other lanes are zero filled.
Table 16 shows the same parameters as shown in Table 14, except with VECLEN of sixteen bytes. Group duplicate is disabled (GRDUP=0).
The twenty-nine elements of the stream are distributed over lane 0 and lane 1 in fifteen vectors. Extra lanes 2 to 7 in vectors 1-14 are zero filled. In vector 15, lane 1 has a stream element (E28) and the other lanes are zero filled.
Table 17 shows the same parameters as shown in Table 14, except with VECLEN of eight bytes. Group duplicate is disabled (GRDUP=0).
The twenty-nine elements of the stream appear in lane 0 in twenty-nine vectors. Extra lanes 1-7 in vectors 1-29 are zero filled.
Table 18 shows the same parameters as shown in Table 15, except with VECLEN of thirty-two bytes and group duplicate is enabled (GRDUP=1).
The twenty-nine elements of the stream are distributed over lanes 0-7 in eight vectors. Each vector 1-7 includes four elements duplicated. The duplication factor (2) results because VECLEN (32 bytes) is half the native vector length of 64 bytes. In vector 8, lane 0 has a stream element (E28) and lanes 1-3 are zero filled. Lanes 4-7 of vector 9 duplicate this pattern.
Table 19 shows the same parameters as shown in Table 16, except with VECLEN of sixteen bytes. Group duplicate is enabled (GRDUP=1).
The twenty-nine elements of the stream are distributed over lanes 0-7 in fifteen vectors. Each vector 1-7 includes two elements duplicated four times. The duplication factor (4) results because VECLEN (16 bytes) is one quarter the native vector length of 64 bytes. In vector 15, lane 0 has a stream element (E28) and lane 1 is zero filled. This pattern is duplicated in lanes 2 and 3, lanes 4 and 5, and lanes 6 and 7 of vector 15.
Table 20 shows the same parameters as shown in Table 17, except with VECLEN of eight bytes. Group duplicate is enabled (GRDUP=1).
The twenty-nine elements of the stream all appear on lanes 0 to 7 in twenty-nine vectors. Each vector includes one element duplicated eight times. The duplication factor (8) results because VECLEN (8 bytes) is one eighth the native vector length of 64 bytes. Thus, each lane is the same in vectors 1-29.
Multiplexers 3101 to 3163 are controlled by multiplexer control encoder 3180. Multiplexer control encoder 3180 receives ELEM_BYTES, VECLEN and GRDUP input signals and generates respective control signals for multiplexers 3101 to 3163. ELEM_BYTES and ELDUP are supplied to multiplexer control encoder 3180 to check to see that VECLEN is at least as great as the product of ELEM_BYTES and ELDUP. In operation, multiplexer control encoder 3180 controls multiplexers 3101 to 3163 to transfer least significant bits equal in number to VECLEN from input register 3100 to output register 3170. If GRDUP=0 indicating group duplication disabled, then multiplexer control encoder 3180 controls the remaining multiplexers 3101 to 3163 to transfer zeros to all bits in the remaining most significant lanes of output register 3170. If GRDUP=1 indicating group duplication enabled, then multiplexer control encoder 3180 controls the remaining multiplexers 3101 to 3163 to duplicate the VECLEN number of least significant bits of input register 3100 into the most significant lanes of output register 3170. This control is similar to the element duplication control described above and fills the output register 3170 with the first vector. For the next vector, data within input register 3100 is rotated down by VECLEN, discarding the previous VECLEN least significant bits. The rate of data movement in formatter 1903 (
Element duplication (ELDUP) and group duplication (GRUDP) are independent. Note these features include independent specification and parameter setting. Thus, element duplication and group duplication can be used together or separately. Because of how these are specified, element duplication permits overflow to the next vector while group duplication does not.
Referring again to
If DECIM field 3007 equals 00, then no decimation occurs. The data elements are passed to the corresponding stream head registers 2818/2828 without change. If DECIM field 3007 equals 01, then 2:1 decimation occurs. Streaming engine 125 removes odd number elements from the data stream upon storage in the stream head registers 2818/2828. Limitations in the formatting network require 2:1 decimation to be employed with data promotion by at least 2× (PROMOTE cannot be 000), ICNT0 must be multiple of 2, and the total vector length (VECLEN) must be large enough to hold a single promoted, duplicated element. For transposed streams (TRANSPOSE≠0), the transpose granule must be at least twice the element size in bytes before promotion. If DECIM field 3007 equals 10, then 4:1 decimation occurs. Streaming engine 125 retains every fourth data element removing three elements from the data stream upon storage in the stream head registers 2818/2828. Limitations in the formatting network require 4:1 decimation to be employed with data promotion by at least 4× (PROMOTE cannot be 000, 001 or 101), ICNT0 must be a multiple of 4 and the total vector length (VECLEN) must be large enough to hold a single promoted, duplicated element. For transposed streams (TRANSPOSE≠0), in one example, decimation removes columns, and does not remove rows. Thus, in such cases, the transpose granule must be at least twice the element size in bytes before promotion for 2:1 decimation (GRANULE≥2×ELEM_BYTES) and at least four times the element size in bytes before promotion for 4:1 decimation (GRANULE≥4×ELEM_BYTES).
The THROTTLE field 3008 controls how aggressively the streaming engine fetches ahead of processing unit core 110. The coding of the two bits of this field is defined as shown in Table 22.
THROTTLE does not change the meaning of the stream and serves only as a hint. The streaming engine can ignore this field. Programs should not rely on the specific throttle behavior for program correctness, because the architecture does not specify the precise throttle behavior. THROTTLE allows programmers to provide hints to the hardware about the program behavior. By default, the streaming engine attempts to get as far ahead of processing unit core 110 as possible to hide as much latency as possible (equivalent to THROTTLE=11), while providing full stream throughput to processing unit core 110. While some applications need this level of throughput, such throughput can cause bad system level behavior for others. For example, the streaming engine discards all fetched data across context switches. Therefore, aggressive fetch-ahead can lead to wasted bandwidth in a system with large numbers of context switches.
The DIMFMT field 3009 defines which of the loop count fields ICNT0 2901, ICNT1 2902, ICNT2 2903, ICNT3 2804, ICNT4 2905 and ICNT5 2906, of the loop dimension fields DIM1 2911, DIM2 2912, DIM3 2913, DIM4 2914 and DIM5 2915 and of the addressing mode fields AM0 3013, AM1 3014, AM2 3015, AM3 3016, AM4 3017 and AM5 3018 (part of FLAGS field 2921) of the stream template register 2900 are active for the particular stream. Table 23 lists the active loops for various values of the DIMFMT field 3009. Each active loop count must be at least 1 and the outer active loop count must be greater than 1.
The DIR bit 3010 determines the direction of fetch of the inner loop (Loop0). If the DIR bit 3010 is 0, Loop0 fetches are in the forward direction toward increasing addresses. If the DIR bit 3010 is 1, Loop0 fetches are in the backward direction toward decreasing addresses. The fetch direction of other loops is determined by the sign of the corresponding loop dimension DIM1, DIM2, DIM3, DIM4 and DIM5.
The CBK0 field 3011 and the CBK1 field 3012 control the circular block size upon selection of circular addressing. The manner of determining the circular block size is described herein.
The AM0 field 3013, AM1 field 3014, AM2 field 3015, AM3 field 3016, AM4 field 3017 and AM5 field 3018 control the addressing mode of a corresponding loop, thus permitting the addressing mode to be independently specified for each loop. Each of AM0 field 3013, AM1 field 3014, AM2 field 3015, AM3 field 3016, AM4 field 3017 and AM5 field 3018 are three bits and are decoded as listed in Table 24.
In linear addressing, the address advances according to the address arithmetic whether forward or reverse. In circular addressing, the address remains within a defined address block. Upon reaching the end of the circular address block the address wraps around to the beginning limit of the block. Circular addressing blocks are limited to 2N addresses where N is an integer. Circular address arithmetic can operate by cutting the carry chain between bits and not allowing a selected number of most significant bits to change. Thus, arithmetic beyond the end of the circular block changes only the least significant bits. The block size is set as listed in Table 25.
In this example, the circular block size is set by the number encoded by CBK0 (first circular address mode 01) or the number encoded by CBK0+CBK1+1 (second circular address mode 10). For example, in the first circular address mode, the circular address block size can range from 512 bytes to 16 M bytes. For the second circular address mode, the circular address block size can range from 1 K bytes to 64 G bytes. Thus, the encoded block size is 2(B+9) bytes, where B is the encoded block number which is CBK0 for the first block size (AMx of 01) and CBK0+CBK1+1 for the second block size (AMx of 10).
The processing unit 110 (
The stream start address register is a register in general scalar register file 211 (
SECLOSE explicitly marks a stream inactive, flushing any outstanding activity. Any further references to the stream trigger exceptions. SECLOSE also allows a program to prematurely terminate one or both streams.
An SESAVE instruction saves the state of a stream by capturing sufficient state information of a specified stream to restart that stream in the future. An SERSTR instruction restores a previously saved stream. An SESAVE instruction saves the stream metadata and does not save any of the stream data. The stream re-fetches stream data in response to an SERSTR instruction.
Each stream can be in one of three states: inactive, active, or frozen after reset. Both streams begin in the inactive state. Opening a stream moves the stream to the active state. Closing the stream returns the stream to the inactive state. In the absence of interrupts and exceptions, streams ordinarily do not make other state transitions. To account for interrupts, the streaming engine adds a third state: frozen. The frozen state represents an interrupted active stream.
In this example, four bits, two bits per stream, define the state of both streams. One bit per stream resides within the streaming engine, and the other bit resides within the processor core 110. The streaming engine internally tracks whether each stream holds a parameter set associated with an active stream. This bit distinguishes an inactive stream from a not-inactive stream. The processor core 110 separately tracks the state of each stream with a dedicated bit per stream in the Task State Register (TSR): TSR.SE0 for stream 0, and TSR.SE1 for stream 1. These bits distinguish between active and inactive streams.
Opening a stream moves the stream to the active state. Closing a stream moves the stream to the inactive state. If a program opens a new stream over a frozen stream, the new stream replaces the old stream and the streaming engine discards the contents of the previous stream. The streaming engine supports opening a new stream on a currently active stream. The streaming engine discards the contents of the previous stream, flushes the pipeline, and starts fetching data for the new opened stream. Data to processor is asserted once the data has returned. If a program closes an already closed stream, nothing happens. If a program closes an open or frozen stream, the streaming engine discards all state related to the stream, clears the internal stream-active bit, and clears the counter, tag and address registers. Closing a stream serves two purposes. Closing an active stream allows a program to specifically state the stream and the resources associated with the stream are no longer needed. Closing a frozen stream also allows context switching code to clear the state of the frozen stream, so that other tasks do not see it.
As noted above, there are circumstances when some data within a stream holding register 2818 or 2828 is not valid. As described above, such a state can occur at the end of an inner loop when the number of stream elements is less than the respective stream holding register 2818/2828 size or at the end of an inner loop when the number of stream elements remaining is less than the lanes defined by VECLEN. For times not at the end of an inner loop, if VECLEN is less than the width of stream holding register 2818/2828 and GRDUP is disabled, then lanes in stream holding register 2818/2828 in excess of VECLEN are invalid.
Referring again to
In this example, upon reading a respective one of the stream head registers 2818/2828 and transferring of data to the requesting functional unit, the invalid/valid data in the respective valid register 2819/2829 is automatically transferred to a data register within predicate register file 234 (
The valid data stored in the predicate register file 234 can be used in a variety of ways. The functional unit can combine the vector stream data with another set of vectors and then store the combined data to memory using the valid data indications as a mask, thus enabling the same process to be used for the end of loop data as is used for cases where all the lanes are valid which avoids storing invalid data. The valid indication stored in predicate register file 234 can be used as a mask or an operand in other processes. P unit 246 (
A second input PROMOTE is supplied to decoder 3202. Decoder 3202 produces an output promotion factor corresponding to the PROMOTE input. The meaning of various codings of PROMOTE are shown in Table 28, which shows an example output of decoder 3202 in bytes for the various PROMOTE codings. The difference in extension type (zero extension or sign extension) is not relevant to decoder 3202.
The outputs of decoders 3201 and 3202 are supplied to multiplier 3203. The product produced by multiplier 3203 is the lane size corresponding to the TOTAL ELEMENT SIZE and the promotion factor. Because the promotion factor is an integral power of 2 (2N), the multiplication can be achieved by corresponding shifting of the TOTAL ELEMENT SIZE, e.g., no shift for a promotion factor of 1, a one-bit shift for a promotion factor of 2, a two-bit shift for a promotion factor of 4, and a three-bit shift for a promotion factor of 8.
NUMBER OF LANES unit 3204 receives the vector length VECLEN and the LANE SIZE and generates the NUMBER OF LANES. Table 29 shows an example decoding of the number of lanes for lane size in bytes and the vector length VECLEN.
As previously stated, VECLEN must be greater than or equal to the product of the element size and the duplication factor. As shown in Table 29, VECLEN must also be greater than or equal to the product of the element size and the promotion factor. This means that VECLEN must be large enough to guarantee that an element cannot be separated from its extension produced by type promotion block 2022 (
The NUMBER OF LANES output of unit 3204 serves as one input to LANE/REMAINING ELEMENTS CONTROL WORD unit 3211. A second input comes from multiplexer 3212. Multiplexer 3212 receives a Loop0 input and a Loop1 input. The Loop0 input and the Loop 1 input represent the number of remaining elements in the current iteration of the corresponding loop.
Comparator 3314 compares the count stored in Loop0 count register 3311 (after incrementing by adder 3313) with the value of ICNT0 2901 (
Circuits for the higher loops (Loop1, Loop2, Loop3, Loop4 and Loop5) are similar to that illustrated in
Referring again to
That is, the loop down count is the difference between the initial iteration count specified in the stream template register and the loop up count produced as illustrated in
LANE/REMAINING ELEMENTS CONTROL WORD unit 3211 (
LANE/REMAINING ELEMENTS CONTROL WORD unit 3211 generates control word 3213 as follows. Control word 3213 has a number of bits equal to the number of lanes from unit 3204. If the remaining count of elements of the selected loop is greater than or equal to the number of lanes, then all lanes are valid. For this case, control word 3213 is all ones, indicating that all lanes within the vector length VECLEN are valid. If the remaining count of elements of the selected loop is nonzero and less than the number of lanes, then some lanes are valid and some are invalid. According to the lane allocation described above in conjunction with
Control word expansion unit 3214 expands the control word 3213 based upon the magnitude of LANE SIZE. The expanded control word includes one bit for each minimally sized lane. In this example, the minimum stream element size, and thus the minimum lane size, is one byte (8 bits). In this example, the size of holding registers 2818/2828 equals the vector size of 64 bytes (512 bits). Thus, the expanded control word has 64 bits, one bit for each byte of stream holding registers 2818/2828. This expanded control word fills the least significant bits of the corresponding valid register 2819 and 2829 (
For the case when VECLEN equals the vector length, the description is complete. The expanded control word includes bits for all places within respective valid register 2819/2829. There are some additional considerations when VECLEN does not equal the vector length. When VECLEN does not equal the vector length, the expanded control word does not have enough bits to fill the corresponding valid register 2819/2829. As illustrated in
Referring still to
In another example, mux 3215 and group duplicate block 3216 are replaced with group duplicate logic that is similar to the group duplicate logic 2025 illustrated in
As previously described, if group duplication is enabled (GRDUP=1), then the excess lanes of stream head register 2818 (
There are two possible outcomes. In one outcome, in most cases, all the lanes within VECLEN are valid and the bits from control word expansion unit 3214 are all ones. This occurs for vectors 1-7 of the group duplication example shown in Table 18 and vectors 1-14 of the group duplication example shown in Table 19. Under these conditions, all bits of the expanded control word from control word expansion unit 3214 are one and all lanes of stream head register 2818 are valid. Group duplicate unit 3216 thus fills all the excess lanes with ones. In the other outcome, the number of remaining stream data elements is less than the number of lanes within VECLEN. This occurs for vector 8 in the group duplication example shown in Table 18 and vector 15 in the group duplication example shown in Table 19. Under these conditions, some lanes within VECLEN are valid and some are invalid. Group duplicate unit 3216 fills the excess lanes with bits having the same pattern as the expanded control word bits. In either case, the excess lanes are filled corresponding to the expanded control bits.
Referring still to
Instruction decoder 113 receives bits 13-17 of src1 field 1305 of an instruction. The opcode field (bits 3-12 for all instructions and additionally bits 28-31 for unconditional instructions) unambiguously specifies a corresponding functional unit 3420 and the function to be performed. In this example, functional unit 3420 can be L2 unit 241, S2 unit 242, M2 unit 243, N2 unit 244 or C unit 245. The relevant part of instruction decoder 113 illustrated in FIG. 34 decodes src1 bit field 1305. Sub-decoder 3411 determines whether src1 bit field 1305 is in the range from 00000 to 01111. If this is the case, sub-decoder 3411 supplies a corresponding register number to global vector register file 231. In this example, the register number is the four least significant bits of src1 bit field 1305. Global vector register file 231 recalls data stored in the register corresponding to the register number and supplies the data to the src1 input of functional unit 3420.
Sub-decoder 3412 determines whether src1 bit field 1305 is in the range from 10000 to 10111. If this is the case, sub-decoder 3412 supplies a corresponding register number to the corresponding local vector register file. If the instruction is directed to L2 unit 241 or S2 unit 242, the corresponding local vector register file is local vector register file 232. If the instruction is directed to M2 unit 243, N2 unit 244 or C unit 245, the corresponding local vector register file is local vector register file 233. In this example, the register number is the three least significant bits of src1 bit field 1305. The corresponding local vector register file 232/233 recalls data stored in the register corresponding to the register number and supplies the data to the src1 input of functional unit 3420.
Sub-decoder 3413 determines whether src1 bit field 1305 is 11100. If this is the case, sub-decoder 3413 supplies a stream 0 read signal to streaming engine 125. Streaming engine 125 then supplies stream 0 data stored in holding register 2818 to the src1 input of functional unit 3420.
Sub-decoder 3414 determines whether src1 bit field 1305 is 11101. If this is the case, sub-decoder 3414 supplies a stream 0 read signal to streaming engine 125. Streaming engine 125 then supplies stream 0 data stored in holding register 2818 to the src1 input of functional unit 3420. Sub-decoder 3414 also supplies an advance signal to stream 0. As previously described, streaming engine 125 advances to store the next sequential vector of data elements of stream 0 in holding register 2818.
Supply of a stream 0 read signal to streaming engine 125 by either sub-decoder 3413 or sub-decoder 3414 triggers another data movement. Upon such a stream 0 read signal, streaming engine 125 supplies the data stored in valid register 2819 to predicate register file 234 for storage. In accordance with this example, this is a predetermined data register within predicate register file 234. In this example, data register P0 corresponds to stream 0.
Sub-decoder 3415 determines whether src1 bit field 1305 is 11110. If this is the case, sub-decoder 3415 supplies a stream 1 read signal to streaming engine 125. Streaming engine 125 then supplies stream 1 data stored in holding register 2828 to the src1 input of functional unit 3420.
Sub-decoder 3416 determines whether src1 bit field 1305 is 11111. If this is the case, sub-decoder 3416 supplies a stream 1 read signal to streaming engine 125. Streaming engine 125 then supplies stream 1 data stored in holding register 2828 to the src1 input of functional unit 3420. Sub-decoder 3414 also supplies an advance signal to stream 1. As previously described, streaming engine 125 advances to store the next sequential vector of data elements of stream 1 in holding register 2828.
Supply of a stream 1 read signal to streaming engine 125 by either sub-decoder 3415 or sub-decoder 3416 triggers another data movement. Upon such a stream 1 read signal, streaming engine 125 supplies the data stored in valid register 2829 to predicate register file 234 for storage. In accordance with this example, this is a predetermined data register within predicate register file 234. In this example, data register P1 corresponds to stream 1.
Similar circuits are used to select data supplied to scr2 input of functional unit 3402 in response to the bit coding of src2/cst field 1304. The src2 input of functional unit 3420 can be supplied with a constant input in a manner described above. If instruction decoder 113 generates a read signal for stream 0 from either scr1 field 1305 or scr2/cst field 1304, streaming engine 125 supplies the data stored in valid register 2819 to predicate register P0 of predicate register file 234 for storage. If instruction decode 113 generates a read signal for stream 1 from either scr1 field 1305 or scr2/cst field 1304, streaming engine 125 supplies the data stored in valid register 2829 to predicate register P1 of predicate register file 234 for storage.
The exact number of instruction bits devoted to operand specification and the number of data registers and streams are design choices. In particular, the specification of a single global vector register file and omission of local vector register files is feasible. This example employs a bit coding of an input operand selection field to designate a stream read and another bit coding to designate a stream read and advancing the stream.
The process illustrated in
Instruction decoder 113 receives bits 13-17 of src1 field 1305 of an instruction. The opcode field opcode field (bits 3-12 for all instructions and additionally bits 28-31 for unconditional instructions) unambiguously specifies P unit 246 and the function to be performed. The relevant part of instruction decoder 113 illustrated in
Sub-decoder 3512 determines whether src1 bit field 1305 is in the range 10000 to 10111. If this is the case, sub-decoder 3512 supplies a decoded register number to the predicate register file 234. In this example, the register number is the three least significant bits of src1 bit field 1305. The predicate register file 234 recalls data stored in the register corresponding to the register number and supplies the data to the src1 input of predicate unit 246.
Sub-decoder 3513 determines whether src1 bit field 1305 is 11100. If this is the case, sub-decoder 3513 supplies a stream 0 valid read signal to streaming engine 125. Streaming engine 125 then supplies valid data stored in valid register 2819 to the src1 input of P unit 246.
Sub-decoder 3514 determines whether src1 bit field 1305 is 11101. If this is the case, sub-decoder 3514 supplies a stream 1 valid read signal to streaming engine 125. Streaming engine 125 then supplies stream 1 valid data stored in valid register 2829 to the src1 input of P unit 246.
The P unit 246 instruction employing the stream valid register 2819/2829 as an operand can be any P unit instruction previously described such as NEG, BITCNT, RMBD, DECIMATE, EXPAND, AND, NAND, OR, NOR, and XOR.
The special instructions noted above can be limited to P unit 242. Thus, the operations outlined in
Matrix Multiplaction Accelerator
Generally speaking, MMA 3640 supports matrix multiplication of two matrices. As shown in expression (3), where matrix A is an [n×m] matrix and matrix B is an [m×p] matrix, the matrix multiplication outputs matrix C.
where each i,j entry is given by multiplying the entry Aik (across row i of A) by the entries Bkj (down column j of B), for k=1, 2, . . . , m.
Referring back to
In an example, MMA 3640 is also equipped to multiply two 64×64 8-bit matrices by breaking each into four 32×32 sub-matrices, multiplying various combinations of the sub-matrices and then combining the results to produce a final 64×64 8-bit matrix result. Other examples may be implemented to support larger or smaller matrices having a larger or smaller precision than 8 or 16-bits.
In order for MMA 3640 to operate correctly on a matrix that is smaller than its native size, such as the 32×32 example mentioned above, all of the elements of each input matrix A and B should contain valid data. Unused elements should be set to zero, or some other agreed upon null value. Streaming engine 125 includes support for nulling unused elements of an array during a stream access without performing additional memory accesses, as described in more detail hereinbelow. In an example, streaming engine 125 includes support for inserting zeros or a selected value, such as a max value or a min value, during a stream access without performing memory accesses to provide the selected values.
Streaming Engine Support for MMA
Referring still to
As a result, only one of the two matrices in a matrix multiplication map on to the streaming engine's 2-dimensional stream definition. This problem is not unique to the streaming engine. In fact, matrix multiplication's access pattern fits poorly with most general-purpose memory hierarchies. Some software libraries attack this problem by directly transposing one of the two matrices, so that both get accessed row-wise (or column-wise) during multiplication.
Transpose Mode
With the streaming engine, programs need not resort to that extreme. The streaming engine supports implicit matrix transposition with a notion of transposed streams. Transposed streams avoid the cost of explicitly transforming the data in memory. Instead of accessing data in strictly consecutive-element order, the streaming engine effectively interchanges the inner two loop dimensions in its traversal order, fetching elements along the second dimension into contiguous vector lanes.
Transpose mode interchanges the two innermost loop levels. That is, in transpose mode, the two innermost loops ICNT0 and ICNT1 are interchanged. ICNT1 determines the number of rows in each column. A column is defined as a GRANULE size. ICNT0 is the second dimension in a transpose stream and defines the horizontal width (which may or may not be a multiple of the GRANULE). In this example streaming engine, the maximum row height, ICNT1, must be at least 1 and less than or equal 16. There are no restrictions on the ICNT0 in transpose. However, if the ICNT0 is not a multiple of the GRANULE size, the streaming engine will pad zeros in the missing elements of each GRANULE.
Coarse rotator 2835 and data storage 2816 of stream 0 engine 2810 are illustrated in
In this example, data storage 2816 is organized as a register file 3810 with 32 slots of 64 bytes (512 bits). Other examples may provide a larger or a smaller amount of storage without changing the semantics of a stream. Data storage 2816 is organized as eight independent banks that are each eight bytes (64 bits) wide. Each bank includes two write ports and two read ports. In this example, each bank also includes two bits/line for parity protection.
In Transpose mode, the SE organizes the internal storage into sector tiles, and the number of sector tiles depends on what the current vertical count (ICNT1) is set to. This allows the SE to fetch as many rows and columns as possible and organizes and rotates the data coming back from L2 into the sectors. This allows the SE to use both read and write ports per bank when reading and writing the data in transpose mode, so that the data can be rotated and ordered according to its sector.
In this example, coarse rotator 2835 includes a set of sixteen multiplexors, represented by multiplexors 3806, 3807, 3808. Each multiplexor, such as multiplexor 3806, has sixteen inputs that are each four bytes (32 bits) wide and are connected to receive all 512 bits provided by the L2 interface 2833 on bus 3802. A four-byte output of each multiplexor is coupled to provide data to one half of a respective bank of register file 3810. Each bank of register file 3810 is coupled to receive data from two multiplexors in parallel, such as 3806, 3807, such that data received from the L2 interface via bus 3802 may be manipulated in four-byte elements.
Reference queue 2815 receives storage allocation and tracking meta-data from storage and allocation logic 2814 (
Alignment network 3820 and 3821 are each similar to coarse rotator 2835. In this example, alignment network 3820 includes a set of sixteen multiplexors, represented by multiplexors 3822, 3823. Each multiplexor, such as multiplexor 3822, has thirty-two inputs that are each four bytes (32 bits) wide and are connected to receive all 512 bits provided by each of the two read ports of the register file 3810. A four-byte output of each multiplexor, such as multiplexor 3822, is coupled to provide data to a respective input of butterfly network 2817. In this manner, multiplexors 3822, 3823 can select sixteen four-byte data elements from register file 3810 to form a 64-byte line of data to provide to butterfly network 2817.
Similarly, in this example, alignment network 3821 includes a set of sixteen multiplexors, represented by multiplexors 3824, 3825. Each multiplexor, such as multiplexor 3824, has thirty-two inputs that are each four bytes (32 bits) wide and are connected to receive all 512 bits provided by each of the two read ports of the register file 3810. A four-byte output of each multiplexor, such as multiplexor 3824, is coupled to provide data to a respective input of butterfly network 2817. In this manner, multiplexors 3824, 3825 can select sixteen four-byte data elements from register file 3810 to form a 64-byte line of data to provide to butterfly network 2817.
Control logic 3814 generates control signals to independently control each of the sixteen multiplexors 3822, 3823 in alignment network 3820 and each of the sixteen multiplexors 3824, 3825 in alignment network 3821 such that any four-byte data element retrieved from register file 3810 may be aligned to any four-byte location within two 64-byte output lines provided to butterfly network 2817 based on the meta-data provided by reference queue 2815.
Butterfly network 2817 is controlled by stream 0 engine 2810 to further format data retrieved from data storage 2816 prior to sending the formatted data to processing unit core 110, as described in more detail with regard to
In this example, control logic 3804 for coarse rotator 2835 and control logic 3814 for data storage 2816 are implemented as asynchronous Boolean logic that is capable of generating control signals for each of the multiplexors and register file 3810 in parallel based on the contents of the meta-data provided by reference queue 2815.
In Transpose mode, SE0 2810 organizes the internal storage 2816 into sector tiles, and the number of sector tiles depends on what the current vertical count (ICNT1) is set too. This allows the SE0 2810 to fetch as many rows and columns as possible and organizes and rotates the data coming back from L2 130 (
The coarse rotator is controlled by control logic 3804 based on meta-data that is queued up in reference queue 2815 (
Inserting Null Vectors into a Stream
Referring to
In another example, control logic 3814 can control each of the multiplexers on alignment networks 3820, 3821 to force a selected value onto selected elements of a stream vector based on the meta-data provided by reference queue 2815. In this manner, an entire stream vector can be set to a selected value, such as a min value or a max value without fetching any data from system memory.
In this example, control logic in the address generator 2811/2821 (see
Metadata produced by the address generator is pushed into the stream reference queue 2815/2825 (see
Data Strip Mining Support
Complex signal processing applications such as various vision algorithms, dense linear algebra, FFT operations and high-level apps including convolutional neural networks (CNNs), structure from motion (SFM), and Radar may require numeric manipulation of a complex multidimensional data structure. Boundaries of the data structure need to be clean so that anomalous data is not included in computations done at the boundary. Generally, this requires a programmer to allocate memory space at the boundary of a data structure that can be preset to a known value, such as zero, or a maximum or minimum data value. Presetting the boundary data then takes additional processing cycles to access and set the boundary data values.
An example steaming engine 125 (
In an example, the DECDIM feature uses two sets of flags that includes a first decrement flag (DECDIMx), where “x” is “1” or “2” in this example, a secondary decrement flag (DECDIMxSD), and a respective width parameter (DECDIMx_WIDTH). This allows a programmer to define four mask values on selected dimensions of the nested loops in order to mask or zero out parts of the stream data. Table 30 defines an example set of DECDIM flags that apply to an example of streaming engine 125. In this example, three bits are used to define the DECDIMx flag, while only two bits are used to define the DECDIMxSD flag. The DECDIMxSD flag is therefore limited to DIM1-DIM3 in this example. Various combinations of settings for DECDIM are supported, as described in more detail hereinbelow.
Referring to
The DECDIMx flag field and DECDIMx_WIDTH supports a data strip mining feature that allows a programmer to define a “total actual width” size of an image using the DECDIM_WIDTH count to provide this maximum actual width. In this mode, the DECDIM_WIDTH is decremented by the selected dimension (DIM1-DIM5), as shown in Table 30, when the address generator 2811/2821 (
The DECDIM_WIDTH count value will only reload to the programmed value again when the selected DECDIM dimension loop count (ICNT) expires. In the above example, the width would reload when dimension 2 (i.e., ICNT2) expires, and dimension 3 is entered. In other words, the width reloads when we enter any dimension higher than the selected DECDIM dimension. Thus, it is possible to program the width smaller than the loop iterations count for the selected dimension, which could cause the width count to underflow. This is explained further hereinbelow.
As mentioned above, there is also a secondary DECDIMxSD flag. The secondary flag allows a “secondary decrement” count mask on top of its respective DECDIMx flag for the respective DECDIM_WIDTH. In other words, the respective DECDIM_WIDTH is decremented or reloaded using the settings for both of the respective DECDIMx and DECDIMxSD when the respective selected dimension is entered or ends.
In an example, the DECDIM selection can be set for any DIM1-DIM5 while operating in linear mode. However, in transpose mode, DIM1 selection is not supported and only DIM2-DIM5 is supported. In this example, the selected DIMx value must be an unsigned value (i.e., DIMx bit 31 is zero). All other dimensions can be unsigned or signed.
For example, if DECDIM=010b, then DIM2 must be an unsigned value. DIM1, DIM3, DIM4, and DIM5 can be unsigned or signed. In example hardware, the innermost loop count is scaled in terms of total number of bytes. In other words, the innermost loop ICNT value is equal to (ICNT0×ELEM_BYTES), and in the case of DECDIM mode, the DECDIM_WIDTH count is (DECDIM_WIDTH×ELEM_BYTES). During processing unit core 110 fetches, the innermost loop count expires when all elements have been consumed before accounting for any data formatting (element duplication, promotion, decimation). Thus, with data formatting enabled, multiple processing unit core 110 fetches could occur before the innermost loop count expires.
In an example, the value of DECDIM_WIDTH count will be saturated at zero in the case where the loop iteration count parameters for the selected DECDIM (i.e., ICNT1-ICNT5) are incorrectly programmed causing DECDIM_WIDTH count to underflow when decremented with the selected DIMx value. When saturation occurs, streaming engine 125 will hold DECDIM_WIDTH count at zero and subsequent data phases to processing unit core 110 will also be zero.
The following tables and
SE 125 will not send any μTLB 2812/2822 (
Transpose data strip mining is similar to linear data strip mining except the data moves in a transpose pattern. DECDIM flags provides the dimension for when the DECDIM_WIDTH count decrement occurs. This allows the data strip mask to apply to the remaining elements in the GRANULE. In DECDIM mode and transpose, the setting for DECDIM=DIM1 is not supported as mentioned in previous tables. Note that for normal transpose patterns, when the ICNT0 is not a multiple of the GRANULE, gappy data occurs and is filled with zeros. Similarly, gappy data occurs in DECDIM transpose mode when the ICNT0 is not a multiple of the GRANULE, in addition to the data strip masking that is applied using DECDIM_WIDTH.
After ICNT0 expires, the next loop level DIM1 is entered, as indicated at 3911. In this example, DIM1 specifies a distance of 128 bytes; therefore fetch 5 3926 begins at address 0x128 and the inner loop is repeated to produce four stream vectors loaded with 56 bytes of element data. This repeats seven times since ICNT1=7 in this example.
After ICNT1 expires, the next loop level DIM2 is entered, as indicated at 3912. In this case, DIM2 specifies a distance of 80 bytes; therefore fetch 29 3927 begins at address 0x080 and the inner loop is repeated to produce four stream vectors loaded with 56 bytes of element data. Similar to 3925, eight bytes of element data is masked as indicated at 3928. This pattern repeats seven times since ICNT1=7 in this example.
Since ICNT0 expires on the first fetch to system memory, the next loop level DIM1 is entered, as indicated at 4011. In this example, DIM1 specifies a distance of 128 bytes; therefore fetch 2 4022 begins at address 0x128. ICNT1=7, therefore this loop is repeated seven times, such that fetch 2 4022, up to fetch 7 4023 access system memory 130 to form six additional stream vectors to provide to processing unit core 110.
After ICNT1 expires, the next loop level DIM2 is entered, as indicated at 4012. In this case, DIM2 specifies a distance of 80 bytes; therefore fetch 8 4024 begins at address 0x080 and the inner loop is repeated to produce a single stream vector loaded with 16 bytes of element data. This loop repeats up to fetch 14 4025, when ICNT 1 expires again and the next loop level DIM2 is entered again.
After ICNT1 expires again, the next loop level DIM2 is entered, as indicated at 4026. Again, DIM2 specifies a distance of 80 bytes; therefore fetch 15 4026 begins at address 0x160 and the inner loop is repeated to produce a single stream vector loaded with 16 bytes of element data. This loop repeats up to fetch 21 4027, when ICNT 1 expires again and the next loop level DIM2 is entered again.
After ICNT1 expires again, the next loop level DIM2 is entered, as indicated at 4026. Again, DIM2 specifies a distance of 80 bytes; therefore fetch 22 4028 begins at address 0x240 and the inner loop is repeated to produce a single stream vector loaded with 16 bytes of element data. However, in this case, the DECDIM_WIDTH parameter 4014 is set to be 248 elements, which is also 248 bytes in this example. Therefore, the array data elements in the border region indicated at 4029 are not to be used in the signal processing application that is consuming this stream. Metadata produced by address generator 2811/2821 (
This loop repeats up to fetch 28 4029, when ICNT 1 expires again. In this example, ICNT also expires after these four repetitions and the stream is complete.
However, in this example, ICNT2=5. Therefore, streaming engine 125 prepares to generate addresses for another repetition of the DIM2 loop. However, DECDIM_WIDTH 4112 has now underflowed, indicating that the remining stream vectors for border region 4130 are to be set to null vectors. The underflow condition is referred to as being “saturated.” In this example, null vectors are set to “0”. When the DECDIM_WIDTH count is saturated, there is no need to access system memory for array element data, since the stream vectors are set to null values. Control logic within address generator 2811/2821 suppresses address generation in response to the DECDIM_WIDTH being saturated so that no accesses are made to the pTLB 2812/2822 (
In this manner, a stream of array elements is provided to processing unit core 110 that is responsive to the stream parameters, but fetch cycles to system memory and the μTLB are avoided for the null vectors.
At fetch 4 4223, DECDIM_WIDTH count reaches zero and data in border region 4224 is masked using alignment network 3820/3821 as described in more detail hereinabove.
On the next iteration, DECDIM_WIDTH count is saturated and fetches to system memory and to the μTLB are suppressed as described hereinabove in more detail. In this manner, stream vectors 4225, 4226, and 4227 are formed by streaming engine 125 without accessing system memory or the μTLB.
After ICNT1 expires, the next loop level DIM2 is entered, as indicated at 4212. In this case, DIM2 specifies a distance of 80 bytes; therefore fetch 8 4228 begins at address 0x080 and the inner loop is repeated to produce a single stream vector loaded with 16 bytes of element data. This loop repeats up to fetch 11 4029, where a portion of the stream vector is masked in response to DECDIM_WIDTH count 4214 and the remaining three null stream vectors are formed by SE 125 without accessing system memory.
In this example, ICNT2=4, so this same loop of system access and formation of null vectors without system access is repeated two more times, as indicated at 4230.
The DIM1 level loop is performed in a similar manner to stream 4000 until DECDIM1_WIDTH 4312 saturates. In this example, ICNT0 4310=16 and ICNT1=7. As described hereinabove in more detail, once the DECDIM1_WIDTH 4312 count saturates then no further system memory accesses are needed for that loop. In this case, null stream vectors 4320 and 4321 are formed by streaming engine 125 without accessing system memory 130.
After ICNT1 expires, the next loop level DIM2 is entered, as indicated at 4311. The DECDIM1_WIDTH counter is reloaded with the count value from the stream template register 2900. In this case, DIM2 specifies a distance of 80 bytes; therefore fetch 8 4322 begins at address 0x080 and the inner loop is repeated to produce a single stream vector loaded with 16 bytes of element data. This loop repeats until the DECDIM1_WIDTH 4312 count again saturates, then no further system memory accesses are needed for that loop. In this case, null stream vectors 4323 and 4324 are formed by streaming engine 125 without accessing system memory 130.
In this example, ICNT2=4, so this same loop of system access and formation of null vectors without system access is repeated two more times.
However, in this case, the DECDIM2_WIDTH parameter 4313 is set to be 248 elements, which is also 248 bytes in this example. Therefore, the array data elements in the border region indicated at 4325 are not to be used in the signal processing application that is consuming this stream. Metadata produced by address generator 2811/2821 (
This loop repeats until the DECDIM1_WIDTH 4312 count again saturates, then no further system memory accesses are needed for that loop. In this case, null stream vectors 4326 and 4327 are formed by streaming engine 125 without accessing system memory 130.
In this example, ICNT3=2 and DIM3=1500, so this entire sequence is repeated again beginning at address 0x1500 as indicated at 4341.
In this example, a matrix 4400 is located in L2 system memory 130 and has nineteen columns of 32-bit elements and eighteen rows of 32-bit elements, where each element is indicated by “d” in
Using the stream parameters listed in Table 36, streaming engine 125 first fetches sub-array 4401 to provide to MMA 3640. Streaming engine 125 then fetches sub-array 4402. Sub-array 4402 includes a border region 4405 in which all the data elements provided to MMA 3640 need to be set to zero in order for a matrix multiplication operation to performed by MMA 3640 to operate correctly. As described hereinabove in more detail, streaming engine 125 uses DECDIM1_WIDTH 4410 to define the extent of matrix 440, which in this example is nineteen elements.
In this case, the DECDIM1_WIDTH parameter 4410 is set to be nineteen elements, which is 76 bytes in this example. Therefore, the array data elements in the border region indicated at 4405 are not to be used in the signal matrix multiplication application that is consuming this stream. Metadata produced by address generator 2811/2821 (
Streaming engine 125 then begins fetching sub-array 4403. At fetch 4412, DECDIM2_WIDTH count reaches zero and is saturated and fetches to system memory and to the μTLB are suppressed as described hereinabove in more detail. In this manner, stream vectors for border region 4406 are formed by streaming engine 125 without accessing system memory or the μTLB.
Streaming engine 125 then begins fetching sub-array 4404. Fetches to system memory are performed at 4413 and 4414 and the border region 4405 is again masked by alignment network 2820/2821 as described hereinabove in response to DECDEM1_WIDTH saturating. On the next cycle, DECDIM2_WIDTH count reaches zero and is saturated and fetches to system memory and to the μTLB are suppressed as described hereinabove in more detail. In this manner, stream vectors for border region 4406 in sub-array 4404 are formed by streaming engine 125 without accessing system memory or the μTLB.
Null Stream Vectors by Count
In the examples described hereinabove, several ways of padding stream vectors with constant values and several ways of forming null stream vectors without accessing system memory were described that are based on a specified value for a DECDIMx_WIDTH count. In another example, it may be useful to specify a number of null vectors to be inserted in a stream by the streaming engine.
Creation of a Convolution neural network (CNN) Toeplitz style matrix on the fly requires zero or constant values fed into the Matrix Multiplication Accelerator after the last feature map in the CNN layer. In linear algebra, a Toeplitz matrix or diagonal-constant matrix in which each descending diagonal from left to right is constant. As described with regard to
Assuring that all unused data elements are set to zero or to a selected null value using software requires allocation of memory and execution of instructions to write out the null values.
In an example, a loop end zero (LEZR) feature is implemented in the streaming engine 125 (
In this example, an eight-bit LEZR_CNT (LEZR count) field 2933 is provided in stream template 2900 (
Referring to
In another example, control logic 3814 can control each of the multiplexers on alignment networks 3820, 3821 to force a selected value onto selected elements of a stream vector based on the meta-data provided by reference queue 2815. In this manner, an entire stream vector can be set to a selected value, such as a min value or a max value without fetching any data from system memory.
In this example, control logic in the address generator 2811/2821 (see
Metadata produced by the address generator is pushed into the stream reference queue 2815/2825 (see
In this example, the MMA width can be selected to be 16, 32, or 64 elements, depending on the size of the elements, as described with respect to
Pad Value
In some applications, such as a pooling operation in a convolution neural network for maximum (max), minimum (min), and average pooling, the end of a column of image pixels needs to be either a max or a min value. Similarly, analysis tools for various deep learning neural networks, such as TensorFlow, Caffe networks, etc. require arbitrary zero, max, or min padding for the pooling operations.
In this example, a three-bit pad value (PADVAL) flag 3023 is included in the flag field 2921 of stream template 2900. An example definition of the PADVAL flags is included in Table 38. Pad size is done on a full element, as defined by the ELTYPE flags 3001 (
In an example, a stream vector is formed using the specified pad value as the last element of a stream vector of the stream. In another example, another flag field (such as LEZR flags 3023, Table 37) may be used to specify a selected dimension on which to include a pad value.
Referring to
In an example, control logic 3814 can control each of the multiplexers on alignment networks 3820, 3821 to force a selected value onto selected elements of a stream vector based on the meta-data provided by reference queue 2815. In this manner, an entire stream vector can be set to a selected value, such as a min value or a max value without fetching any data from system memory.
In this example, control logic in the address generator 2811/2821 (see
Metadata produced by the address generator is pushed into the stream reference queue 2815/2825 (see
In order to achieve this in software code requires performing a permute operation. In an example simulation, performance of the hardware based PADVAL solution was 4× better than the software solution for image width>64 and 2× better performance for an image with <64; furthermore, the software solution required three different loops.
At 4601, an address stream is generated according to the stream parameters stored in the stream template. Metadata is saved that indicates loop counts, end of loop, remaining width count, etc.
At 4602 a line of matrix data is fetched by the streaming engine from system memory using the sequence of addresses generated by the address generator.
At 4603, a check is made to see if a width count parameter is associated with the current dimension of the multi-dimension stream. If not, a normal stream vector is formed at 4605.
If a width count parameter is associated with the current dimension, then at 4604, a check is made to determine if the width count is depleted. If not, a normal stream vector is formed at 4605.
If the width count is depleted, then at 4606 a check is made to determine the width count is saturated. If not, then a stream vector is formed at 4607 by masking a portion of the vector that exceeds the width count. In some examples, the mask may insert a value of zero. In another example, a predefined value may be inserted by the mask, such as a min value, a max value, etc.
If the width count is saturated, then at 4608 a null vector is formed without accessing system memory. In an example, the null vector may be set to all zeros. In another example, the null vector may be set to a predetermined pad value, such as a min value, a max value, etc. A predetermined pad value may be designated by a flag field, such a PADVAL flag 3023 (
At 4609, if the current dimension is not complete the loop repeats at 4602. If the current dimension is complete, then the width count is decremented at 4610 if it is associated with the current dimension. Otherwise, the width count is not decremented.
At 4611 a check is made to determine if the stream is complete. If not, the process repeats to access more matrix data from the system and to form stream vectors.
At 4612, once the entire matrix has been accessed from memory, the data stream is closed.
At 4701, an address stream is generated according to the stream parameters stored in the stream template. Metadata is saved that indicates loop counts, end of loop, remaining width count, etc.
At 4702 a line of matrix data is fetched by the streaming engine from system memory using the sequence of addresses generated by the address generator.
At 4703 a normal stream vector is formed using the data elements fetched from system memory.
At 4704, if the current dimension is not complete the loop repeats at 4602. If the current dimension is complete.
At 4705, a check is made to determine if the null count applies to the current dimension. If so, a number n of null vectors equal to the null count value in the stream template are formed by the streaming engine without accessing null data from system memory. In this example, the data in the null vectors are set to zero. In another example, a preselected null value may be used to form the null vector, such as a min value, a max value, etc.
At 4707, a check is made to determine if the stream is complete. If not, the process repeats with the next loop at 4708 to access more matrix data from the system and to form stream vectors.
At 4709, once the entire matrix has been accessed from memory and the specified null vectors formed without accessing data from system memory, the data stream is closed.
In this example, processor A, processor B, L3 memory 4850 are all included in a SoC 4800 that may be encapsulated to form a package that may be mounted on a substrate such as a printed circuit board (PCB) using known or later developed packaging techniques. For example, SoC 4800 may be encapsulated in a ball grid array (BGA) package. In this example, external memory interface (EMI) 4852 allows additional external bulk memory 4854 to be accessed by processor A and/or processor B.
In this example, processor B is an ARM® processor that may be used for scalar processing and control functions. In other examples, various types of known or later developed processors may be combined with DSP 100. While two processors are illustrated in this example, in another example, multiple copies of DSP 100 and/or multiple copies of processor B may be included within an SoC and make use of the techniques for forming masked and null vectors without accessing system memory provided by streaming engine 125 that are described herein in more detail.
In a described example, a streaming engine includes two closely coupled streaming engines that can manage two data streams simultaneously. In another example, the streaming engine maybe capable of managing only a single stream, while in other examples the streaming engine is capable of handling more than two streams. In each case, for each stream, the streaming engine includes an address generation stage, a data formatting stage, and some storage for formatted data waiting for consumption by the processor.
In a described example, addresses are derived from algorithms that can involve multi-dimensional loops, each dimension maintaining an iteration count. In one example, the streaming engine supports six levels of nested iteration. In other examples, more or fewer levels of iteration are supported.
In a described example, one-dimensional zero padding of stream vectors is provided. In another example, two-dimensional zero padding of stream vectors is provided. In yet another example, more than two dimensions of zero padding may be provided.
In described examples, a complex DSP processor with multiple function units and dual data paths is described. In another example, a simpler DSP that is coupled to a stream processor may be used. In another example, other types of known or later developed processors may be coupled to a stream processor, such as a reduced instruction set computer (RISC), a traditional microprocessor, etc.
In a described example, the MMA supports 32×32 16-bit matrix multiplication, and 64×64 8-bit matrix multiplication, and the streaming engine is configured to provide 64-byte stream vectors. In another example, the MMA may be configured to support large or smaller matrix sizes. An associated streaming engine may be configured to provide stream vectors that have a size that is larger or smaller than 64 bytes.
In described examples, a processor that consumes a stream of data and a streaming engine that retrieves the stream of data from system memory are all included within a single integrated circuit (IC) as a system on a chip. In another example, the processor that consumes the stream of data may be packaged in a first IC and the streaming engine may be packaged in a second separate IC that is coupled to the first IC by a known or later developed communication channel or bus.
In this description, the term “couple” and derivatives thereof mean an indirect, direct, optical, and/or wireless electrical connection. Thus, if a first device couples to a second device, that connection may be through a direct electrical connection, through an indirect electrical connection via other devices and connections, through an optical electrical connection, and/or through a wireless electrical connection.
Modifications are possible in the described examples, and other examples are possible, within the scope of the claims.
This application is a continuation of U.S. patent application Ser. No. 16/420,457, filed on May 23, 2019, which is a continuation-in-part of U.S. patent application Ser. No. 16/227,238, filed on Dec. 20, 2018, now U.S. Pat. No. 11,036,648, issued Jun. 15, 2021, which is a continuation of U.S. patent application Ser. No. 15/429,205, filed on Feb. 10, 2017, now U.S. Pat. No. 10,162,641, issued on Dec. 25, 2018, which is a division of U.S. patent application Ser. No. 14/331,986, filed on Jul. 15, 2014, now U.S. Pat. No. 9,606,803, issued on Mar. 28, 2017, which claims the benefit of U.S. Provisional Patent Application No. 61/846,148, filed on Jul. 15, 2013, each of which is incorporated by reference herein in its entirety.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5502829 | Sachs | Mar 1996 | A |
| 5909573 | Sheaffer | Jun 1999 | A |
| 6606686 | Agarwala | Aug 2003 | B1 |
| 7750916 | Whitaker | Jul 2010 | B2 |
| 8537169 | Bolz | Sep 2013 | B1 |
| 8676874 | Fossum | Mar 2014 | B2 |
| 9606803 | Anderson et al. | Mar 2017 | B2 |
| 9667276 | Tunali | May 2017 | B1 |
| 9904645 | Thompson et al. | Feb 2018 | B2 |
| 10162641 | Anderson | Dec 2018 | B2 |
| 10203958 | Zbiciak | Feb 2019 | B2 |
| 10942741 | Tran | Mar 2021 | B2 |
| 10949206 | Tran | Mar 2021 | B2 |
| 11036648 | Anderson | Jun 2021 | B2 |
| 11113062 | Bhardwaj | Sep 2021 | B2 |
| 11231929 | Tran | Jan 2022 | B2 |
| 11249759 | Leven | Feb 2022 | B2 |
| 11256508 | Bhardwaj | Feb 2022 | B2 |
| 11269638 | Zbiciak | Mar 2022 | B2 |
| 11507520 | Bui | Nov 2022 | B2 |
| 11748270 | Bui | Sep 2023 | B2 |
| 20020026569 | Liao | Feb 2002 | A1 |
| 20030151774 | Washio | Aug 2003 | A1 |
| 20050273450 | McMillen | Dec 2005 | A1 |
| 20060179257 | Chu | Aug 2006 | A1 |
| 20070019005 | Van Baarsen | Jan 2007 | A1 |
| 20070189363 | Eriksson et al. | Aug 2007 | A1 |
| 20080183792 | Inoue | Jul 2008 | A1 |
| 20100054626 | Shi | Mar 2010 | A1 |
| 20100241808 | Toub | Sep 2010 | A1 |
| 20100303182 | Daneshrad et al. | Dec 2010 | A1 |
| 20100306300 | Lu et al. | Dec 2010 | A1 |
| 20120254192 | Gelbard | Oct 2012 | A1 |
| 20120281747 | Abrishamkar et al. | Nov 2012 | A1 |
| 20120324319 | Grinchuk | Dec 2012 | A1 |
| 20130339621 | Steely, Jr. et al. | Dec 2013 | A1 |
| 20140019635 | Reznik | Jan 2014 | A1 |
| 20150019840 | Anderson | Jan 2015 | A1 |
| 20150262332 | Park | Sep 2015 | A1 |
| 20190278595 | Tran et al. | Sep 2019 | A1 |
| 20190278596 | Leven | Sep 2019 | A1 |
| 20190278597 | Bhardwaj et al. | Sep 2019 | A1 |
| 20190278598 | Bhardwaj et al. | Sep 2019 | A1 |
| 20220164188 | Leven | May 2022 | A1 |
| Entry |
|---|
| Herruzo, E. et al., Using Padding to Optimize Locality in Scientific Application,2008, Springer-Verlag, pp. 863-872). (Year: 2008). |
| Hong, C. et al., Effective Padding of Multidimensional Arrays to Avoid Cache Conflict Misses, 2016, ACM, pp. 129-144. (Year: 2016). |
| Yeom, Jae-Seung et al., A Runtime Framework for Optimizing Multi-Dimensional Array Accesses on Multi-Core Processors, 2009, ResearchGate, 9 pages. (Year: 2009). |
| “66AK2Hx Keystone Multicore DSP+ARM System-on-Chips”, SPRT651A, Texas Instruments, Inc., 2013, pp. 1-3. |
| Number | Date | Country | |
|---|---|---|---|
| 20220164188 A1 | May 2022 | US |
| Number | Date | Country | |
|---|---|---|---|
| 61846148 | Jul 2013 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 14331986 | Jul 2014 | US |
| Child | 15429205 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 16420457 | May 2019 | US |
| Child | 17670611 | US | |
| Parent | 15429205 | Feb 2017 | US |
| Child | 16227238 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 16227238 | Dec 2018 | US |
| Child | 16420457 | US |
