JP2010176692A - Arithmetic processing device, information processing apparatus, and control method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To suppress wasteful consumption of a low hierarchy cache access pipeline by a hardware prefetch request issued from a cash memory of high hierarchy to a cache memory of low hierarchy in a cache memory having a plurality of hierarchies. <P>SOLUTION: A processor having prefetch function includes a cache memory of first hierarchy having a first line size; a cache memory of second hierarchy lower in hierarchy than the cash memory of first hierarchy and having a second line size different from the first line size, and a prefetch control part which issues, for each second line size, the prefetch request from the cache memory of first hierarchy to the cache memory of second hierarchy so as to prefetch a block corresponding to the first line size. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a processor having two or more layers of cache memories having different line sizes, and more particularly to a technique for a processor having a prefetch function for a cache memory.

  2. Description of the Related Art Conventionally, in a computer that performs continuous access to a memory used in science and technology calculations, such as HPC (High Performance Computing), a prefetch technique is applied to a cash register.

  Prefetching is a technique of predicting instructions or data that will be required in the near future in advance and reading them into the cache memory or the like, and can reduce cache memory cache misses.

  Patent Document 1 discloses a cache system having a prefetch function. In the system of Patent Document 1, in the case of continuous access to memory data, a predicted address of the line size destination to be accessed next in continuous access is registered in a queue due to a cache miss, and the access address is actually queued. Is hit and the prediction is successful, it is determined that the access is continuous, and a prefetch is issued to the address accessed next to the line size.

  In a multi-level cache memory, if the upper-level cache memory and the lower-level cache memory have different line sizes, the data size to be moved in due to a cache error in the lower-level hierarchy is changed from the upper-level line size to the lower-level cache memory. Any size up to the line size. In the case of continuous access in which hardware application fetch functions, the case of the line size of the lower hierarchy with the largest data size provides the highest performance. In this case, the data size to be moved in is the cache of the lower hierarchy. The line size is likely to be.

  For example, in the Columbia 2 memory system, the data size moved in due to a cache miss is the line size of the lower layer cache in the case of memory access, but the line size of the upper layer cache in the copy back case.

  Since an HPC JOB with many continuous accesses has a low copyback rate, the data size to be moved in is likely to be the cache line size of the lower hierarchy in the case of the above continuous access.

When prefetching is performed in a cache memory system in which the line size is different between an upper layer cache memory and a lower layer cache memory, the following problems occur.
When the data size moved in due to a cache miss in the lower hierarchy is the line size of the cache in the lower hierarchy, a hardware application fetch request issued from the upper hierarchy cache to the lower hierarchy cache (move-in to the lower hierarchy cache (Request) may be performed once for the line size of the cache in the lower hierarchy. However, in the conventional cache system, it is issued for each line size of the cache of the upper hierarchy, and a useless lower hierarchy cache access pipeline is consumed.

  If the data size moved in due to a cache error in the lower hierarchy is the line size of the lower hierarchy cache, a hardware application fetch request issued to the lower hierarchy cache can be issued once per line size of the lower hierarchy cache. Good. However, hardware application fetch sometimes loses the prefetch request due to implementation restrictions, and if the hardware application fetch is lost, if the prefetch request is issued only once, the cache of the lower layer No move-in request for memory data is issued.

  If the data size to be moved in due to a cache miss that occurred in the lower-level cache register is the line size of the lower-level cache memory, the prefetch request issued to the lower-level cache memory is the lower-level cache memory. Once per line size. Therefore, if the address that is the line size of the cache memory in the upper hierarchy is the initial value of the prefetch address of the prefetch request for the address missed in the cache register in the upper hierarchy, it may be the same line for the cache in the lower hierarchy Because of this, the lower-level cache access pipeline is consumed with a useless prefetch request.

  In the case of continuous access to memory for which prefetch functions, the data size moved into the lower-level cache memory is likely to be the line size of the lower-level cache memory. The data size may be different from the line size.

JP 2004-38345 A

  An object of the present invention is to provide a processor having a prefetch function that solves the above problems.

  In order to solve the above problems, a processor having a prefetch function according to the present invention includes a first hierarchy cache memory, a second hierarchy cache memory, and a prefetch control unit.

The cache memory of the first hierarchy has a first line size.
The cache memory of the second hierarchy is a lower hierarchy of the cache memory of the first hierarchy and has a second line size that is different from the first line size.

  The prefetch control unit issues a prefetch request for the cache of the second hierarchy from the cache memory of the first hierarchy so as to prefetch the block corresponding to the first line size for each of the second line sizes. To do.

With this configuration, it is possible to prevent an unnecessary prefetch request from being issued.
The prefetch control unit may issue the prefetch request once to a plurality of times for each second line size.

The prefetch control unit may be configured to issue the prefetch request so as to prefetch a block that is twice or more the first line size.
With this configuration, it is possible to cope with a case where a prefetch request is lost due to restrictions in implementation.

  Furthermore, the prefetch control unit may be configured such that a prefetch destination address for performing the prefetch request is an address ahead of the second line size from an address missed in the cache memory of the first hierarchy.

  The prefetch control unit issues the prefetch request for each of the first line sizes and issues the prefetch request for each of the second line sizes based on the size of the moved-in data. It is also possible to employ a configuration that further includes a switching unit that switches between the two.

With this configuration, it is possible to cope with a move-in other than the size of the second line such as copy back.
According to the present invention, the prefetch request is issued for each second line size that is the line size of the cache memory of the second hierarchy, not the cache memory of the first hierarchy. It is possible to suppress the consumption of the access pipeline of the cache memory in the hierarchy.

  Further, even if a prefetch request is lost due to implementation restrictions, performance can be improved by increasing the possibility that a memory data move-in request to the second-level cache memory will be issued.

  Further, for the address missed in the cache memory of the first hierarchy, the address of the line size ahead of the cache memory of the second hierarchy is used instead of the address ahead of the first line size, and the prefetch address of the hardware application fetch By using this initial value, it is possible to suppress consumption of the cache access pipeline of the second hierarchy due to useless requests.

  When the data size moved into the second-level cache memory differs from the line size of the second-level cache memory by issuing a hardware application fetch request according to the finally moved-in data size However, the request can be correctly issued without leaking the necessary request.

It is the schematic of the processor of the computer system in this embodiment, and its periphery structure. It is the figure drawn focusing on the memory management part of the processor in this embodiment. It is a figure which shows the structural example of the prefetch queue (PFQ) of 1st Embodiment. It is a figure which shows the relationship between the output address of an adder, and the output of a comparator. It is the figure which showed each state in procedure 8,9,10. It is a flowchart which shows operation | movement of the prefetch queue (PFQ) of 1st Embodiment. It is a figure which shows the structural example of the prefetch queue (PFQ) of 2nd Embodiment. It is a figure which shows the structural example of the prefetch queue (PFQ) of 3rd Embodiment. It is a figure which shows the structural example of the prefetch queue (PFQ) of 4th Embodiment.

An embodiment of the present invention will be described below with reference to the drawings.
FIG. 1 is a schematic diagram of a processor of a computer system and its peripheral configuration in the present embodiment.

The configuration of FIG. 1 includes a processor unit 1, a prefetch control device 2, a primary cache 3, a secondary cache 4, and a main storage device 5.
The processor unit 1 includes an ALU, a register, and the like, and is a part that manages actual calculation and data processing. Further, in the configuration of the figure, branch prediction and the like are also performed in the processor unit 1 and a request based on the prediction result is made to the primary cache 3. The prefetch control device 2 is a device that is responsible for overall control of prefetch processing, and requests prefetch to the secondary cache 4 while monitoring the request address from the processor unit 1 to the primary cache 3. The primary cache 3 is a primary cache system and includes a memory having a high access speed and a primary cache control device. The secondary cache 4 is a secondary cache system, and includes a memory having a higher access speed than the main storage device 5 and a larger capacity than the primary cache 3, and a secondary cache control device. In the present embodiment, the prefetched data is held in the secondary cache 4. The main storage device 5 is a memory constituted by a DRAM or the like.

  When the processor unit 1 accesses data on the main storage device 5, the request address is designated from the request address 6, the fetch data 7 is read at the time of reading, and is output to the primary cache 3 as the store data 8 at the time of writing.

  In response to a read request from the processor unit 1, the primary cache 3 outputs the data to the processor unit 1 as fetch data 7 if it holds the data at the request address. Requests the one-line data including the data from the request bus 11 to the secondary cache 4 and notifies the prefetch control device 2 as a cache miss 9. When the fetch data 12 is received, the data requested to the processor unit 1 is output as the fetch data 7. Further, when the cache data held by itself is updated, the primary cache 3 writes back the data from the data bus 13 to the secondary cache 4 at an appropriate timing.

  If the secondary cache 4 holds the data in response to the data request from the primary cache 3, it outputs the data for one line including the data to the primary cache 3 as fetch data 7. If not, the main memory 5 is requested from the request bus 14 for one line of data including the data. When the fetch data 15 is received, data for one line is output to the primary cache 3. Similarly to the primary cache 3, the secondary cache 4 writes back the data from the data bus 16 to the main storage device 5 at an appropriate timing when the cache data held by the secondary cache 4 is updated.

  When the processor unit 1 requests data from the primary cache 3, the address is specified by the address bus 6, and the prefetch control device 2 monitors this address value, and the prefetch address queue provided therein is assigned to this address. Search by value. If this address is in one block starting from the address existing in the prefetch address queue (hereinafter referred to as hit), the prefetch request address is output from the prefetch address bus 10 to the secondary cache 4 and the prefetch request is issued. At the same time, the address is registered in the prefetch address queue 25, and if it does not exist in the prefetch address queue, no prefetch is requested.

  In the present embodiment, the primary cache 3 and the secondary cache 4 are cache memories having different line sizes. In the following description, the line size of the primary cache 3 is 64 bytes (hereinafter referred to as B), It is assumed that the line size of the secondary cache 4 is 256B.

FIG. 2 is a diagram depicting the memory management portion of the processor in this embodiment.
In the figure, the processor includes a fetch port (FP) 21, a store port (SP) 22, a primary cache access pipeline 23, and a primary cache move-in buffer (L1 $ MIB) 24 as components for memory management. Provided in the primary cache 3, provided with a prefetch queue 25 in the prefetch control device 2, a secondary cache move import (L2 $ MIP) 26, a secondary cache prefetch port (L2 $ PFP), and a secondary cache access pipeline 28 And a secondary cache move-in buffer (L2 $ MIP) 29 is provided in the secondary cache 4, and a system controller move import (SCMIP) 30 is provided in the main storage device 5.

  The fetch port (FP) 21 is a port that receives a load instruction, a store instruction, and the like from the processor unit 1. The store port (SP) 22 is a port for the store instruction that has committed the store to write data to the cache. The secondary cache move-in buffer (L2 $ MIP) 29 and the system controller move import (SCMIP) 30 are ports for receiving move-in requests to the secondary cache 4 and the main storage device 5, respectively.

  The primary cache access pipeline 23 and the secondary cache access pipeline 28 are pipelines that accept access requests to the primary cache 3 and the secondary cache 4. The primary cache access pipeline 23 has five stages of P, T, M, B, and R. In the P stage, an address is selected and the address is transferred, and in the T stage, the primary cache is transferred with the transferred address. Referring to the tag and TLB (translation lookup table), the M stage compares (matches) the data obtained as the T stage reference result, and the B stage uses the comparison result to compare the data in the primary cache. In the R stage, a flag indicating whether the transferred data is valid or invalid for a primary cache miss or a TLB miss is calculated and sent. The secondary cache access pipeline 28 has stages PR1 and XP0-14. At each stage, port selection, L2 $ tag search, address comparison, L2 $ miss is registered in the L2 $ MIB, and L2 $ hits L2 $ Data is read, L2 $ data is transferred to L1 $ MIB, etc.

  The primary cache move-in buffer (L1 $ MIB) 24 and the secondary cache move-in buffer (L2 $ MIB) 29 buffer move-in instructions generated for the primary cache 3 and the secondary cache 4. .

  The prefetch queue (PFQ) 25 registers an address one line ahead of the address previously prefetched. When a cache miss occurs in the primary cache 3, the address at which the cache miss occurs and the prefetch queue (PFQ) 25, a prefetch request is issued to the secondary cache prefetch port (L2 $ PFP) 27 if an address matching the prefetch queue (PFQ) 25 is registered. The secondary cache prefetch port (L2 $ PFP) 27 receives a prefetch request from the prefetch queue (PFQ) 25.

The operation in the figure will be described below.
When the processor unit 1 decodes a load instruction or the like and issues a memory read request, the request is input from the fetch port (FP) 25 to the primary cache access pipeline 23. If the primary cache 2 hits the read request, the data is returned as it is from the fetch port (FP) 25 to the processor unit 1 that issued the request, and the data is written to the register 31.

  When the primary cache 2 misses, data must be brought from the secondary cache 3, so a request is entered in the primary cache move-in buffer (L1 $ MIB) 24. The primary cache move-in buffer (L1 $ MIB) 24 issues a read request to the secondary cache 3. This enters the secondary cache access pipeline 28 via the secondary cache move import (L2 $ MIP) 26 that receives the secondary cache 3 request.

  If this read request hits the secondary cache 3, the data is put into the primary cache move-in buffer (L1 $ MIB) 24, and the primary cache move-in buffer (L1 $ MIB) 24 accesses the primary cache line. A pipeline is acquired and data is written to the primary cache cache 2 (in the case of a primary cache miss and secondary cache hit).

Next, a case where hardware application fetch is performed will be described.
If there is a miss in the primary cache 2 and the address to be operated as the hardware application fetch is not registered in the prefetch queue (PFQ) 25, the address is once registered in the prefetch queue (PFQ) 27. At this time, as shown in Patent Document 1, when the address of 64B ahead is registered, and when accessing next 64 bytes ahead, the prefetch queue (PFQ) 25 hits at the same time as the primary cache misses. At this time, the prefetch queue (PFQ) 25 adds 64 B and issues a prefetch request with an address of +128 B to the prefetch port (L 2 $ PFP) 27.

  The primary cache miss is registered in the secondary cache move import (L2 $ MIP) 26 and the secondary cache prefetch port (L2 $ PFP) 27, and returns data if the secondary cache is accessed and hit. If there is a miss, it is registered in the secondary cache move-in buffer (L2 $ MIB) 29 and output to the system controller move import (SCMIP) 30 to issue a request to the main storage device 5. When the main memory 5 data is returned, it is written to the secondary cache 3 via the secondary cache access pipeline 28, and simultaneously returned to the primary cache access pipeline 23 by bypass, and this is stored in the primary cache 2. Write.

FIG. 3 is a diagram illustrating a configuration example of the prefetch queue (PFQ) 25 according to the first embodiment.
The prefetch 3, the selection circuit 44, the adder 45, the selection circuits 46 and 47, and the adder 48 shown in the figure are provided. Each entry 41-1 to 41-n is set with an address value or the like registered in the entry 41. A register 49 for comparing the address in the register 49 with the request address, and an AND circuit 51 for obtaining an AND between a comparison result of the comparator 50 and a valid bit in the register 49 described later.

In addition to the address value, the register 49 records a valid bit that serves as a status flag, a standby bit, and an L2 $ PFP registration permission flag.
The valid bit in the register 49 indicates whether the address value set in the register 48 is valid. The valid value is set when the address value is registered in the register 48, and the address value is read from the entry 41. Sometimes reset. The standby bit is set when the request address from the primary cache access pipeline 23 matches the address value registered in the register 48 in the entry 41 in which the valid bit is set. The prefetch address queue (PFQ) 25 determines entries 41-1 to 41-n to be read from the state of the standby bit. The L2 $ PFP registration permission flag indicates that the request address from the primary cache access pipeline 23 and the address registered in this entry 41 match (hit) the next 256B continuous address as the secondary cache prefetch port. (L2 $ PFP) 27 is used to determine whether or not to register, and if the L2 $ PFP registration permission flag is set to 1, registration to the secondary cache prefetch port (L2 $ PFP) is performed. If set, registration is not performed in the secondary cache prefetch port (L2 $ PFP).

  When a request address is input from the primary cache access pipeline 23 and is newly registered, the valid bit is 1, the standby bit is 0, the L2 $ PFP registration permission flag is 1, and the address is the request address. A value obtained by adding 64 to the adder 45 is input.

  When the registration address of the register 49 is updated, the comparison result of whether the bits [7: 6] of the output of the adder 45 by the comparator 42 are 0 or not is newly registered in the L2 $ PFP registration permission flag. The result obtained by the OR circuit 43 obtaining the OR of whether is input. The content of the L2 $ PFP registration permission flag is a selection signal for the selection circuits 46 and 47. When the L2 $ PFP registration permission flag is 1, 1 is output as the PFP request signal and the entry as the PFP request address. A value obtained by adding 256 B to the register value registered in the register 49 of 41 by the adder 48 is output. With these outputs, the PFP request address is registered in the secondary cache PFP (L2 $ PFP) 27.

FIG. 4 is a diagram showing the relationship between the output address of the adder 45 and the output of the comparator 42.
When the address value set in the register 49 is incremented by 64B by the adder 45 and registered in the register 49, the bits [7: 6] of the output address become 0 once every four times. 1 is output once every four times, and this is set to the L2 $ PFP registration permission flag via the OR circuit 43. If the registration in the register 49 is a new registration, 1 is set in the L2 $ PFP registration permission flag. Therefore, every time the address is updated four times from the new address registration, 1 is set in the L2 $ PFP registration permission flag, and the PFP request address is registered in the secondary cache PFP (L2 $ PFP) 27.

  The L2 $ PFP registration permission flag is set when new registration is performed in the prefetch queue (PFQ) 25 and when the head 64B address at the 256B boundary is updated and registered in the prefetch queue (PFQ) 25. It is reset when an address other than the head 64B of the 256B boundary is updated and registered in the PFQ.

  When a prefetch request address is input from the primary cache access pipeline 23, the prefetch queue (PFQ) 25 is compared with the address value in the register 49 by the comparator 50, and this comparison result is validated. The AND circuit 51 takes an AND with the bit, and outputs the result to the primary cache access pipeline 23 as a PFQ hit signal indicating whether or not the PFQ has been hit. Accordingly, when the request address matches the register 49 and the valid bit is 1, the PFQ hit signal becomes 1.

  Even when the line size of the primary cache 2 is 64B and the line size of the secondary cache 3 is 256B, and the line sizes of the upper and lower cache memories are different, the secondary cache prefetch port is used. Registration of the address value to (L2 $ PFP) 27 can be made once every four times (256B / 64B), and a prefetch request is made once per line size of the secondary cache 3. Therefore, it is possible to suppress unnecessary consumption of the lower hierarchy cache access pipeline and improve performance.

Next, using FIG. 2 and FIG. 3, the detailed procedure of the processing for the memory access instruction of the processor including the prefetch operation is shown below.
In the following description, addresses A, A + 8, A + 16,. . . , A + 56 load instruction is decoded by the processor unit as an example.
1: The load instruction acquires the primary cache access pipeline 23 via the fetch port (FP) 21.
2: In the primary cache access pipeline 23, the primary cache is accessed at address A.
3: In the primary cache access pipeline 23, a primary cache miss is detected as a result of 2.
4: Register a miss address in the primary cache move-in buffer (L1 $ MIB) 24.

  4.1: The primary cache move-in buffer (L1 $ MIB) 24 issues a move-in request from the secondary cache 3 to the primary cache 2 to the secondary cache move import (L2 $ MIP) 26.

4.2: Secondary cache move import (L2 $ MIP) 26 acquires the secondary cache access pipeline 28 and accesses the secondary cache at address A.
4.3: In the secondary cache access pipeline 28, a secondary cache miss is detected as a result of the procedure 4.2.

4.4: A miss address is registered in the secondary cache move-in buffer (L2 $ MIB) 29.
4.5: The secondary cache move-in buffer (L2 $ MIB) 29 issues a move-in request from the main storage 5 to the secondary cache 3 to the system controller move import (SCMIP) 30.

  4.6: The system controller move import (SCMIP) 30 takes 256 B of data from the miss address A from the main storage device 5 and moves it into the secondary cache move-in buffer (L 2 $ MIB) 29.

  4.7: The secondary cache move-in buffer (L2 $ MIB) 29 acquires the secondary cache access pipeline 28 and writes 256B of move-in data to the secondary cache 4.

4.8: The secondary cache move-in buffer (L2 $ MIB) 29 bypass-transfers 64B of move-in data to the primary cache move-in buffer (L1 $ MIB) 24.
4.9: A load instruction having a primary cache miss at address A acquires the primary cache access pipeline 23 and bypasses the move-in data transferred to the primary cache move-in buffer (L1 $ MIB) 24. Read and write data to the register 31 in the processor unit 1.

4.10: The primary cache move-in buffer (L1 $ MIB) 24 acquires the primary cache access pipeline 23 and writes 64 B of move-in data to the primary cache 2.
5: The prefetch queue (PFQ) 25 detects a miss.
6: The next continuous address (A + 64) is newly registered in the prefetch queue (PFQ) 25. Set L2 $ PFP registration permission flag in register 49.
7: Similarly, a load instruction that accesses consecutive addresses (A + 8, A + 16,..., A + 56) acquires the primary cache access pipeline 23.
8: If the move-in data from the secondary cache 3 has not arrived at this time, the primary cache MIB hit and data miss are detected, and the primary cache access pipeline 23 is aborted. The aborted request returns to the fetch port (FP) 21.
9: If move-in data from the secondary cache 3 has arrived and no data has been written to the primary cache 2, a primary cache MIB hit and data hit are detected, and the primary cache move-in buffer ( L1 $ MIB) 24 data is bypassed and read, and data is written to the register 31.
10: If the move-in data from the secondary cache 3 has arrived and data has been written to the primary cache, the primary cache hit is detected, the data is read from the primary cache 2, and the register 31 is read. Write data.

FIG. 5 shows each state in the procedures 8, 9, and 10.
In the state of procedure 8, since the address has arrived from the primary cache move-in buffer (L1 $ MIB) 24 but no data has arrived, and no data has been written to the primary cache 2, 1 The next cache access pipeline 23 is aborted.

  In the state of the procedure 9, the address and data have arrived from the primary cache move-in buffer (L1 $ MIB) 24, but no data has been written to the primary cache 2, so the primary cache move-in buffer Data is read from (L1 $ MIB) 24 and written to register 31.

  In the state of procedure 10, since the address and data have arrived from the primary cache move-in buffer (L1 $ MIB) 24 and the data has been written to the primary cache 2, the data is transferred from the primary cache 2. And data is written to the register 31.

The process for the load instruction for accessing the continuous address (A + 64) will be described below.
11: Similar to Procedure 1, the load instruction that accesses the continuous address (A + 64) acquires the primary cache access pipeline 23.
12: As a result of the procedure 11, a primary cache miss is detected.

12.1: Register a miss address in the primary cache move-in buffer (L1 $ MIB) 24 and access the secondary cache 3.
12.2: A secondary cache hit is detected, 64B data is read from the secondary cache, and transferred to the primary cache move-in buffer (L1 $ MIB) 24.

12.3: A load instruction with a primary cache miss at address (A + 64) reads data by bypassing data in the primary cache move-in buffer (L1 $ MIB) 24 and writes data to the register 31.
12.4: The primary cache move-in buffer (L1 $ MIB) 24 writes 64B data to the primary cache 2.
13: A hit in the prefetch queue (PFQ) 25 is detected. Set the wait bit in register 49.
14: The next continuous address (A + 128) is registered in the prefetch queue (PFQ) 25. Reset L2 $ PFP registration permission flag in register 49.
15: Since the L2 $ PFP registration permission flag was set until it was reset in step 14, the next 256B continuous address (A + 64 + 256) is registered in the secondary cache prefetch port (L2 $ PFP) 27.

  15.1: The secondary cache prefetch port (L2 $ PFP) 27 acquires the secondary cache access pipeline 28 and accesses the secondary cache 3 at the address (A + 64 + 256).

15.2: A secondary cache miss is detected as a result of the procedure 15.1.
15.3: A cache miss address is registered in the secondary cache move-in buffer (L2 $ MIB) 29.

  15.4: The secondary cache move-in buffer (L2 $ MIB) 29 issues a move-in request from the main storage 5 to the secondary cache 3 to the system controller move import (SCMIP) 30.

  15.5: The system controller move import (SCMIP) 30 takes 256 B of data from the miss address (A + 64 + 256) from the main storage device 5 and moves it into the secondary cache move-in buffer (L2 $ MIB) 29.

15.6: The secondary cache move-in buffer (L2 $ MIB) 29 acquires the secondary cache access pipeline 28 and writes 256B of move-in data to the secondary cache 3.
16: Similarly, a load instruction that accesses consecutive addresses (A + 64 + 8, A + 64 + 16,..., A + 64 + 56) acquires the primary cache access pipeline 23.
17: If the move-in data from the secondary cache 4 has not arrived, the primary cache MIB hit and data miss are detected and the primary cache access pipeline 23 is aborted. The aborted request returns to the fetch port (FP) 21.
18: If move-in data from the secondary cache 3 has arrived and no data has been written to the primary cache 2, a primary cache MIB hit and data hit are detected, and the primary cache move-in buffer (L1 $ MIB) 24 data is bypassed and read and written to the register 31.
19: If move-in data from the secondary cache 3 has arrived and data has been written to the primary cache 2, a primary cache hit is detected, data is read from the primary cache 2, and it is stored in the register Write to 41.

The process for the load instruction for accessing the continuous address (A + 128) will be described below.
20: Similar to procedures 1 and 11, the load instruction that accesses the continuous address (A + 128) acquires the primary cache access pipeline 23.
21: As a result of the procedure 20, a primary cache miss is detected.

21.1: A cache miss address is registered in the primary cache move-in buffer (L1 $ MIB) 24 and the secondary cache 3 is accessed.
21.2: As a result of the procedure 21.1, a secondary cache hit is detected, 64B data is read from the secondary cache 3, and transferred to the primary cache move-in buffer (L1 $ MIB) 24.

  21.3: A load instruction having a primary cache miss at address (A + 128) reads data from the primary cache move-in buffer (L1 $ MIB) 24, bypassing the data, and writes it to the register 31.

21.4: The primary cache move-in buffer (L1 $ MIB) 24 writes 64B data to the primary cache 2.
22: A hit of the prefetch queue (PFQ) 25 is detected. Set the wait bit in register 49.
23: The next continuous address (A + 192) is registered in the prefetch queue (PFQ) 25. Reset L2 $ PFP registration permission flag.
24: (L2 $ PFP registration permission flag was set until reset in step 23, so the next 256B consecutive address (A + 128 + 256) is not registered in the secondary cache prefetch port (L2 $ PFP) 27)
25: Similarly, a load instruction that accesses consecutive addresses (A + 128 + 8, A + 128 + 16,..., A + 128 + 56) acquires the primary cache access pipeline 23.
26: If the move-in data from the secondary cache 3 has not arrived, the primary cache move-in buffer (L1 $ MIB) 24 detects a data miss and aborts the primary cache access pipeline 23. The aborted request returns to the fetch port (FP) 21.
27: If move-in data from the secondary cache 3 has arrived and no data has been written to the primary cache, a primary cache MIB hit or data hit is detected, and the data in the primary cache MIB is bypassed. Read and write data to the register 31.
28: If move-in data from the secondary cache 3 has arrived and data has been written to the primary cache 2, a primary cache hit is detected, the data is read from the primary cache 2 and stored in the register 31. Write data.

The process for the load instruction for accessing the continuous address (A + 192) will be described below.
29: Similar to procedures 1, 11, and 20, the load instruction that accesses the continuous address (A + 192) acquires the primary cache access pipeline.
30: As a result of the procedure 29, a primary cache miss is detected.

30.1: A cache miss address is registered in the primary cache move-in buffer (L1 $ MIB) 24 and the secondary cache 3 is accessed.
30.2: A secondary cache hit is detected, 64B data is read from the secondary cache, and transferred to the primary cache move-in buffer (L1 $ MIB) 24.

  30.3: A load instruction with a primary cache miss at address (A + 192) reads data from the primary cache move-in buffer (L1 $ MIB) 24, bypassing the data, and writes it to the register 31.

30.4: The primary cache move-in buffer (L1 $ MIB) 24 writes 64B data to the primary cache 2.
31: A hit in the prefetch queue (PFQ) 25 is detected. Set the wait bit in register 49.
32: The next continuous address (A + 256) is registered in the prefetch queue (PFQ) 25. Set L2 $ PFP registration permission flag of register 49.
33: (Since the L2 $ PFP registration permission flag was reset until it was set in step 32, the next 256B continuous address (A + 192 + 256) is not registered in the secondary cache prefetch port (L2 $ PFP) 27)
34: Similarly, a load instruction that accesses consecutive addresses (A + 192 + 8, A + 192 + 16,..., A + 192 + 56) acquires the primary cache access pipeline 23.
35: If move-in data from the secondary cache 3 has not arrived, the primary cache move-in buffer (L1 $ MIB) 24 hit, data miss is detected, and the primary cache access pipeline 23 is aborted. The aborted request returns to the fetch port (FP) 21.
36: If move-in data from the secondary cache 3 has arrived and data has not been written to the primary cache 2, a primary cache move-in buffer (L1 $ MIB) 24 hit, data hit is detected, Data is read out from the primary cache move-in buffer (L1 $ MIB) 24, bypassed, and written into the register 31.
37: If move-in data from the secondary cache 3 has arrived and data has been written to the primary cache 2, a primary cache hit is detected, data is read from the primary cache 2, and this is stored in the register Write to 31.

Next, the processing for the load instruction that accesses the continuous address (A + 256) will be described 38: Similarly, the load instruction that accesses the continuous address (A + 256) acquires the primary cache access pipeline 23.
39: As a result of the procedure 38, a primary cache 2 miss is detected.

39.1: A miss address is registered in the primary cache move-in buffer (L1 $ MIB) 24 and the secondary cache 3 is accessed.
39.2: A secondary cache hit is detected, 64B data is read from the secondary cache 3 and transferred to the primary cache move-in buffer (L1 $ MIB) 24.

  39.3: A load instruction with a primary cache miss at address (A + 256) reads data from the primary cache move-in buffer (L1 $ MIB) 24, bypassing the data, and writes it to the register 31.

39.4: The primary cache MIB writes 64B data to the primary cache.
40: Detect 25 prefetch queue (PFQ) hits. Set the wait bit in register 49.
41: The next continuous address (A + 320) is registered in the prefetch queue (PFQ) 25. Reset L2 $ PFP registration permission flag in register 49.
42: Since the L2 $ PFP registration permission flag was set until it was reset in step 41, the next 256B continuous address (A + 256 + 256) is registered in the secondary cache prefetch port (L2 $ PFP) 27.

  42.1: The secondary cache prefetch port (L2 $ PFP) 27 acquires the secondary cache access pipeline 28 and accesses the secondary cache 3 at the address (A + 256 + 256).

42.2: As a result of the procedure 42.1, a secondary cache miss is detected.
42.3: A miss address is registered in the primary cache move-in buffer (L1 $ MIB) 24.

  42.4: The primary cache move-in buffer (L1 $ MIB) 24 issues a move-in request from the main storage 5 to the secondary cache 3 to the system controller move import (SCMIP) 30.

  42.5: The system controller move import (SCMIP) 30 extracts 256 B of data from the cache miss address (A + 256 + 256) from the main memory 5 and moves this data into the secondary cache move-in buffer (L2 $ MIB) 29. .

  42.6: The secondary cache move-in buffer (L2 $ MIB) 29 acquires the secondary cache access pipeline 28 and writes 256B of move-in data to the secondary cache 3.

Thereafter, the same processing is repeated for the load instruction for accessing the continuous addresses (A + 320), (A + 384),.
FIG. 6 is a flowchart showing the operation of the prefetch queue (PFQ) 25 of the first embodiment shown in FIG.

  In step S1, when the primary cache 2 is accessed at the address A, the primary cache 2 misses the cache (step S2, Y), and the prefetch queue (PFQ) 25 also misses (step S3, Y), step S4 Register the address (A + 64) one line ahead of the primary cache 2 in the prefetch queue (PFQ) 25, and set the L2 $ registration permission flag of the register 49 in the prefetch queue (PFQ) 25 to perform processing. Return to step S1.

In step S2, when the primary cache 2 hits (step S2, N) and the prefetch queue misses (step S3, Y), the process returns to step S1.
In step S2, the primary cache 2 misses (step S2, Y), the prefetch queue (PFQ) 25 hits (step S3, N), and the primary cache 2 hits (step S2, N). When the prefetch queue (PFQ) 25 is hit (step S3, N), the process moves to step S6 and prefetch is performed.

  In step S6, the standby bit in the register 49 of the prefetch queue (PFQ) 25 is set. If the L2 $ registration permission flag in the register 49 is set (step S7, Y), a request to prefetch with the PFP request address (A + 64) is registered in the secondary cache prefetch port (L2 $ PFP) 27, and the L2 $ is registered. If the registration permission flag is not set (step S7, N), the request is not registered in the secondary cache prefetch port (L2 $ PFP) 27.

  Next, after updating the registration address of the prefetch queue (PFQ) 25 to A + 64 in step S9, if the remainder obtained by dividing (A / 64 + 1) by 4 is 0, the L2 $ registration permission flag in the register 49 is set in step S11. Then, the process returns to step S1. If the remainder obtained by dividing A / 64 + 1) by 4 is not 0, the L2 $ registration permission flag is reset in step S12, and the process returns to step S1.

  As described above, in the first embodiment, a prefetch request can be issued for each line size of a lower-layer cache register, so that a useless prefetch request does not occupy an access pipeline, thereby improving performance. I can do it.

Next, a second configuration example of the prefetch queue (PFQ) 25 will be described.
When the prefetch queue (PFQ) 25 of the first embodiment has a size of one line of the lower layer cache n times the size of one line of the upper layer cache, the prefetch queue (PFQ) 25 can handle n consecutive accesses. The prefetch request has been registered to the secondary cache prefetch port (L2 $ PFP) 27 once, but the prefetch queue (PFQ) 25 of the second embodiment registers the prefetch request at least twice n times. To do.

  When the data size to be moved in due to a miss in the lower hierarchy cache is the line size of the lower hierarchy cache, the hardware application fetch request issued to the lower hierarchy cache is the prefetch queue (PFQ) of the first embodiment. ) 25, it may be performed once for the cache line size of the lower hierarchy.

  However, there are cases where prefetch requests are lost due to restrictions on hardware implementation. If the prefetch request is issued only once, the memory to the lower-level cache is lost when the hardware application fetch is lost. Data move-in requests are no longer issued. For example, when the secondary cache 3 misses, it is registered in the secondary cache move-in buffer (L2 $ MIB) 29, but the secondary cache move-in buffer (L2 $ MIB) 29 is full. The prefetch request may be lost without re-registration.

  In response to this point, the prefetch queue (PFQ) 25 of the second embodiment issues a prefetch request to the secondary cache prefetch port (L2 $ PFP) 27 a plurality of times for the line size of the cache in the lower hierarchy.

  FIG. 7 is a diagram illustrating a configuration example of the prefetch queue (PFQ) 25 according to the second embodiment. This figure is described in contrast to the prefetch queue (PFQ) 25 of the first embodiment shown in FIG.

  Comparing FIG. 7 with the prefetch queue (PFQ) 25 of the first embodiment of FIG. 3, the input of the comparator 42a is only bit [6] in the address output from the adder 45a. Therefore, in the first embodiment, when the address of the register 49 is updated four times, the L2 $ PFP registration permission flag is set to 1 once. However, in the prefetch queue (PFQ) 25 of the second embodiment, the address Is updated four times, 1 is set in the L2 $ PFP registration permission flag twice, and the request is registered in the secondary cache prefetch port (L2 $ PFP) 27.

Accordingly, in the second embodiment, even if one prefetch request is lost due to a hardware implementation problem, it can be dealt with.
Next, a third embodiment of the prefetch queue (PFQ) 25 will be described.

Similarly to the second embodiment, the prefetch queue (PFQ) 25 of the third embodiment deals with a case where a prefetch request is lost due to a hardware implementation problem.
In the third embodiment, a prefetch request is issued to the secondary cache prefetch port (L2 $ PFP) 27 so as to prefetch a block that is twice or more the line size of the cache of the upper hierarchy. As a result, the prefetch request is expanded twice or more by the secondary cache prefetch port (L2 $ PFP) 27, and a plurality of prefetch requests are issued.

  FIG. 8 is a diagram illustrating a configuration example of the prefetch queue (PFQ) 25 according to the third embodiment. This figure is also shown in contrast to the prefetch queue (PFQ) 25 of the first embodiment shown in FIG.

  When the configuration of the prefetch queue (PFQ) 25 of the third embodiment of FIG. 8 is compared with the configuration of the first embodiment of FIG. 3, the PFP request block output to the secondary cache prefetch port (L2 $ PFP) 27 The size 61 is 128 B, which is twice the line of the primary cache 2. Although not shown, in the first embodiment shown in FIG. 3, the PFP request block size is 64 B, which is the same as the line size of the primary cache 2.

  With this configuration, the prefetch queue (PFQ) 25 of the third embodiment designates a block size that is twice the line size of the primary cache 2 and sends a prefetch request to the secondary cache prefetch port (L2 $ PFP). 27, the secondary cache prefetch port (L2 $ PFP) 27 issues two prefetch requests.

As a result, even in the third embodiment, even if one prefetch request is lost due to a hardware implementation problem, it can be dealt with.
Next, a fourth embodiment of the prefetch queue (PFQ) 25 will be described.

  The prefetch queue (PFQ) 25 of the fourth embodiment has a move-in for each line size (256B) of the lower layer cache performed in the present embodiment and a line size of the upper layer cache performed by a conventional processor. The move-in for each (64B) can be switched.

Thereby, it is possible to cope with a move-in for each line size (64 B) of the upper layer cache executed at the time of copy back.
FIG. 9 is a diagram illustrating a configuration example of the prefetch queue (PFQ) 25 according to the fourth embodiment. This figure is also shown in contrast to the prefetch queue (PFQ) 25 of the first embodiment shown in FIG.

  Comparing the prefetch queue (PFQ) 25 of the fourth embodiment shown in the figure and the first embodiment of FIG. 3, in the configuration of FIG. 9, the move-in (MI) data size is stored in the register 49b. When the MI data size is set to 0, the prefetch queue (PFQ) 25 performs move-in for every 256B, and when 1 is set, it performs move-in for every 64B.

  This MI data size is set to 0 as an initial value. As a result of comparing the cache-in move-in address with the address set in the register 49b by the comparator 71, they match, and the move-in data size is the same. When it is 64B, 1 is set by the output of the AND circuit 72. The MI data size is input to the selection circuit 46b as a result of ORing with the L2 $ PFP registration permission flag by the OR circuit 73. Therefore, the size of the move-in can be switched such that when the MI data size is set to 0, 256B move-in is performed, and when 1 is set, 64B move-in is performed.

  As described above, according to the present embodiment, even if the line sizes of the upper layer cache register and the lower layer cache register are different, the prefetch request is issued for each line size of the lower layer cache register. Therefore, it is possible to suppress the consumption of the cache access pipeline due to a useless prefetch request and to improve the performance.

In addition, the prefetch request may be lost due to implementation restrictions.
Furthermore, it is possible to cope with a move-in of the line size of the upper layer cache executed at the time of copy back.

  In the above example, the case where the present invention is applied to the prefetch between the secondary cache memory and the main storage device is shown as an example, but the present invention is not limited to this, and the system is not limited to the cache having the tertiary cache or higher. When the memory is provided, the present invention can be applied between the secondary cache and the tertiary cache, between the tertiary cache and the main storage device, and the like.

  In the above example, the case where the prefetch continuous access direction is applied to the ascending order is shown as an example. However, the present invention is not limited to this, and the prefetch continuous access direction is the descending order. The case can also be applied.

Claims (9)

In an arithmetic processing unit connected to a storage device that holds data,
An arithmetic processing unit for requesting data to be used for an operation to perform processing;
When having a plurality of first cache lines of the first line size and not holding the data requested by the arithmetic processing unit in any of the plurality of first cache lines, sending a cache miss notification, A first cache memory for holding first data to be input in response to the received first prefetch request;
A second cache memory having a plurality of second cache lines of a second line size and holding second data input from the storage device in response to the received second prefetch request;
When the cache miss notification is received, a second prefetch request for prefetching the data amount of the second line size from the storage device to the second cache memory is issued for each second cache line. And a prefetch control unit that issues a first prefetch request for prefetching the data amount of the first line size from the storage device to the first cache memory for each first cache line. Arithmetic processing device characterized. In the arithmetic processing unit,
The prefetch control unit
The arithmetic processing unit according to claim 1, wherein the second prefetch request is issued prior to the first prefetch request. In the arithmetic processing unit,
The first prefetch request is:
3. The arithmetic processing unit according to claim 1, wherein the arithmetic processing unit is issued to an address that is ahead of the first line size from an address where the cache miss has occurred. In the arithmetic processing unit,
The second prefetch request is
3. The arithmetic processing unit according to claim 1, wherein the arithmetic processing unit is issued to an address ahead of the second line size from an address where the cache miss has occurred. A storage device for holding data;
An arithmetic processing unit for requesting data to be used for an operation to perform processing;
When having a plurality of first cache lines of the first line size and not holding the data requested by the arithmetic processing unit in any of the plurality of first cache lines, sending a cache miss notification, A first cache memory for holding first data to be input in response to the received first prefetch request;
A second cache memory having a plurality of second cache lines of a second line size and holding second data input from the storage device in response to the received second prefetch request;
When the cache miss notification is received, a second prefetch request for prefetching the data amount of the second line size from the storage device to the second cache memory is issued for each second cache line. And a prefetch control unit that issues a first prefetch request for prefetching the data amount of the first line size from the storage device to the first cache memory for each first cache line. A characteristic information processing apparatus. In the information processing apparatus,
The prefetch control unit
6. The information processing apparatus according to claim 5, wherein the second prefetch request is issued prior to the first prefetch request. In the information processing apparatus,
The first prefetch request is:
The information processing apparatus according to claim 5, wherein the information processing apparatus is issued to an address that is ahead of the first line size from an address where the cache miss has occurred. In the information processing apparatus,
The second prefetch request is
7. The information processing apparatus according to claim 5, wherein the information processing apparatus is issued to an address ahead of the second line size from an address where the cache miss has occurred. In a control method of an arithmetic processing unit connected to a storage device that holds data,
A step of requesting data to be used for a calculation performed by a calculation processing unit included in the calculation processing device;
When the first cache memory having a plurality of first cache lines of the first line size does not hold the data requested by the arithmetic processing unit in any of the plurality of first cache lines, a cache miss notification A step of sending
When the prefetch control unit included in the arithmetic processing unit receives the cache miss notification, for each second cache line, the data amount of the second line size is transferred from the storage unit to the second line size. Issuing a second prefetch request for prefetching to a second cache memory having a plurality of second cache lines;
The second cache memory holding second data input from the storage device in response to the received second prefetch request;
The prefetch control unit issuing a first prefetch request for prefetching the data amount of the first line size from the storage device to the first cache memory for each first cache line;
A control method comprising the step of: the first cache memory holding first data input from the storage device in response to the received first prefetch request.

Images