Revision as of 16:34, 2 May 2019

This page describes the L1 cache controller (CC) allocated in the nu+ core and directly connected to the LDST unit, Core Interface (CI), Thread Controller (TC), Memory Controller (MeC) and Instruction Cache (IC). The Single Core CC handles requests from the core (load/store miss, instruction miss, flush, evict, data invalidate) and serializes them. Pending requests are scheduled with a fixed priority.

TODO: magari un disegno di tutti i componenti collegati al CC

This lightweight CC handles a transaction at time, hence there is no need for a Miss Status Holding Register (MSHR).

Interface

This section shows the interface of the CC to/from all other units.

To/from Core interface

CI buffers memory requests from LDST unit, providing a separate queue for each type of request. This interface is directly forwarded to the CC, which schedules a request at time. Regards to this component, it can be possible to decouple a service speed of CC and a service speed of LDST units [[MIRKO] non ho capito questo periodo]. In fact, the cache controller executes one request at a time, although LDST unit might issue multiple request concurrently, up to the current number of threads allocated into the core. Those requests are scheduled by a fixed priority arbiter, which selects a pending request and dequeues it once the CC forwards a memory transaction to the memory.

Following lines of code define interface to/from core interface:

~
output logic                                       cc_dequeue_store_request,
~
input  logic                                       ci_store_request_valid,
input  thread_id_t                                 ci_store_request_thread_id,
input  dcache_address_t                            ci_store_request_address,
input  logic                                       ci_store_request_coherent,
~

To/from LDST

As described in load/store unit, Load/Store unit does not store any information about the coherency protocol used, although it keeps track of information regarding privileges on all cache addresses. Each cache line stored, in fact, has two privileges: can read and can write that are used to determine cache miss/hit and are updated by the Cache Controller.

Memory requests from LDST unit:

LOAD
STORE
REPLACEMENT
FLUSH
DINV
INSTR
IOM

All the listed operation above are handled be the control FSM in the CC. The IDLE state selects a pending operation performing a fixed-priority scheduling, . On the coherence protocol side, as said before, the LDST has an abstract view of the implemented protocol, and the CC grants it READ or WRITE privileges, while the current coherence protocol is transparent to the LDST unit.

Following lines of code define interface to/from load/store unit:

~
output logic                                       cc_update_ldst_valid,
output dcache_way_idx_t                            cc_update_ldst_way,
output dcache_address_t                            cc_update_ldst_address,
output dcache_privileges_t                         cc_update_ldst_privileges,
output dcache_line_t                               cc_update_ldst_store_value,
output cc_command_t                                cc_update_ldst_command,
output logic                                       cc_wakeup,
output thread_id_t                                 cc_wakeup_thread_id,
~

To/from Memory controller

TODO

To/from Instruction cache

TODO

To/from Thread controller

TODO

Implementation

In this section is described how to is implemented CC.

FSM

The behaviour is implemented by a finite state machine (FSM). The states are three:

• idle
• send request
• wait response

Below is represented the graph of FSM of CC.

The FSM is implemented dividing sequential and combinatorial behaviour.

Sequential section

First of all, some signals are initialized.

In the IDLE state, preparation of request is performed through setting some bits. Preparation depends on the type of request. Below there is an example of LOAD request:

~
if (grants[LOAD]) begin
 granted_read          <= 1'b1;
 granted_write         <= 1'b0;
 granted_need_snoop    <= 1'b1;
 granted_need_hit_miss <= 1'b0;
 granted_wakeup        <= 1'b1;
 granted_thread_id     <= ci_load_request_thread_id;
 granted_address       <= ci_load_request_address;
~

Some information is reported such as thread ID that it sent a request or address of LOAD request or type of request. If there is at least one request, then CC move into SEND REQ state.

The SEND REQ state implements a logic which allows CC to send a request to the memory. If the request is executable and the memory is available, then CC sends the request by loading the address and kind (READ or WRITE) of request to the memory interface. If the request is READ, then CC waits for a response from the memory. else if kind of request is WRITE, then CC computes a DINV request by executing following lines of code and it comes back in the IDLE state [[MIRKO] non credo sia così, se esegui una DINV and ogni write hai cotinui miss]. When the memory is not available, then CC waits until the availability bit of the memory is asserted. When the request is not executable, then CC jumps back into the IDLE state.

~
if (grants_reg[DINV]) begin
 cc_update_ldst_valid       <= 1'b1;
 cc_update_ldst_way         <= way_matched_reg;
 cc_update_ldst_address     <= granted_address;
 cc_update_ldst_privileges  <= dcache_privileges_t'(0);
 cc_update_ldst_command     <= CC_UPDATE_INFO_DATA;
end
~

In these lines of code, CC send to LDST the response of line cache invalidation request without any privileges. So, no thread can read or write that cache line until next LOAD/STORE request.

In the WAIT RESP state, CC waits for a response from the memory. If memory response is available, then CC comes back in the IDLE state. If kind of request is INSTR, then CC return to TC the contents of memory. Else if kind of request is LOAD or STORE, then CC performs a request by executing below lines of code:

~
end else if (grants_reg[LOAD] | grants_reg[STORE]) begin
 cc_update_ldst_valid       <= 1'b1;
 cc_update_ldst_way         <= counter_way[granted_address.index];
 cc_update_ldst_address     <= granted_address;
 cc_update_ldst_privileges  <= dcache_privileges_t'(2'b11);
 cc_update_ldst_store_value <= m2n_response_data_swap;
 cc_update_ldst_command     <= ways_full ? CC_REPLACEMENT : CC_UPDATE_INFO_DATA;
~

CC returns to LDST the way that a request has to execute, READ and WRITE privileges, data from memory that has to write if the type of request is STORE. So, all threads can read or can write a line of a cache of request.

Combinatorial section

First of all, some signals are initialized.

In the IDLE state, the address of cache (tag and set), that a WRITE request has to compute, is defined.

In the SEND REQ state, there is a logic for dequeue request. If the request is executable and the memory is available, then dequeue signal of REPLACEMENT, FLUSH, DINV and I/O write request is asserted. Else if the request isn't executable, then dequeue signal of LOAD, STORE, FLUSH and DINV is asserted.

In the WAIT RESP, once a response from memory is ready, dequeue signal of LOAD, STORE, I/O read and INSTR request is asserted.

IO, Instruction and Core Interface requests buffering

Every request has a valid bit. If this bit is asserted means that CC receives a request from CI. Also, there are two bits that mean there is IO or INSTR pending request. All these bits are buffered into a vector. Regards to this vector, it can be possible to schedule requests. There is a component (described [here]) that allow rounding robin schedule.

Following lines of code defined when a request is executable and when a request is completed:

~
assign req_completed = dequeue_instr | dequeue_iom | cc_dequeue_store_request | cc_dequeue_load_request | cc_dequeue_replacement_request | cc_dequeue_flush_request | cc_dequeue_dinv_request;
assign execute_req = ~granted_need_snoop | (granted_need_snoop & ((snoop_hit & granted_need_hit_miss) | (~snoop_hit & ~granted_need_hit_miss)));
~

A request is completed when relative dequeue signal is asserted, after FSM execution.

A request is executable if it doesn't need for a snoop operation or it needs for a snoop operation and it needs for a snoop hit and FSM needs for a hit (such as FLUSH request) or it needs for a snoop miss and FSM needs for a miss (such as LOAD operation).

IO and INSTR requests are managed in a different way.

IO Map request

There are two queues of size 8:

• IO FIFO REQUEST
• IO FIFO RESPONSE

The format of the queues is defined by the following lines of code:

~
typedef struct packed {
 thread_id_t thread;
 io_operation_t operation;
 dcache_address_t address;
 register_t data;
} io_fifo_t;
~

Requests are buffered into IO FIFO REQUEST queue. When the IO FIFO REQUEST is full, the CC refuses further IO Map requests. Once that processing of a request is terminated, the request is dequeue from IO FIFO REQUEST and the relative response is queued into IO FIFO RESPONSE queue. Once the LDST unit consumed the response, it asserts ldst_io_resp_consumed signal, then the response is dequeued from IO FIFO RESPONSE queue.

It needs two queues because LDST is implemented by FSM, so LDST can be busy sometimes.

Instruction miss request

INSTR requests are buffered in INSTR FIFO queue of size 2. Data that are stored into the INSTR FIFO queue is the address of the request. The relative response of a request is sent to the TC during the execution of FSM, like below lines of code:

~
WAIT_RESP: begin
 if (m2n_response_valid) begin
  state <= IDLE;
  // handle req
  if (grants_reg[INSTR]) begin
   mem_instr_request_data_in <= m2n_response_data_swap;
   mem_instr_request_valid   <= 1'b1;
~

m2n_response_data_swap register contains data from memory that are transformed into big-endian if needed.

It needs one queue because MeC is implemented by a pipeline architecture, so response time is known.

Snoop managing

Snoop managing checks whether a cache hit occurs or not after a snoop request, through following lines of code:

~
generate
 for ( dcache_way = 0; dcache_way < `DCACHE_WAY; dcache_way++ ) begin
  assign way_busy_oh[dcache_way]    = ( ldst_snoop_privileges[dcache_way].can_read | ldst_snoop_privileges[dcache_way].can_write );
  assign way_matched_oh[dcache_way] = ( ldst_snoop_tag[dcache_way] == granted_address.tag ) & way_busy_oh[dcache_way];
 end
endgenerate
assign snoop_hit = |way_matched_oh;
assign ways_full = &way_busy_oh;
~

A cache hit is asserted if a thread has read or write privileges on such address and if the tag of the requested address is equal to an element present in the tag array.

snoop_hit is asserted if there is at least one cache hit. ways_full is asserted if all ways are busy.

way_matched_oh array is the index of the matched way that is one-hot encoded. There is a component that converts one-hot string to a natural number, it is described [here]. The output is put in way_matched_idx register.

For each clock period, there is a matched way used by CC, thanks to these lines of code:

~
always_ff @( posedge clk, posedge reset ) begin
 if ( reset ) begin
  way_matched_reg <= dcache_way_idx_t'(0);
  end else if (snoop_hit) begin
  way_matched_reg <= way_matched_idx;
 end
end
~

Some requests need for a snoop, such as LOAD/STORE, FLUSH and DINV request because they need to access to L1 cache for hitting a specific line of cache, so CC has to evaluate a cache hit. All other types of request don't need for a snoop, like REPLACEMENT because this type of request has to replace a line of cache when the relative cache is full and INSTR and IO request because these types of request have to communicate directly with memory.

CC checks a cache hit occurs every request that needs for a snoop operation because if CC has some request on the same address, at the first request CC regains the data from memory and refresh data and privileges of the line of the cache. From the second request onwards, CC checks a cache hit but it doesn't have to access to the memory. This optimizes performances of nu+ because access to memory is an onerous operation.

Memory swap

nu+ is a parametrizable little-endian manycore, so it can be embedded in architecture with various memory with different endianness. So, CC has a piece of code that is used to convert data from little-endian to big-endian and vice versa, if it needed.