为CPU提供足够的,稳定的指令流和数据流是计算机体系结构设计中两个永恒的话题。为了给CPU提供指令流,需要设计分支预测机构,为了给CPU提供数据流,就需要设计cache了。其实,无论是insn还是data,都需要访问存储器,所以从这个角度来说,cache需要承担更重要的角色。
本小节我们就分析一下or1200的cache部分的实现。
还是那句话,研究一个东西,首先要了解其来龙去脉,cache也不例外。
cache的出现是为了解决memory wall问题。由于cpu的频率越来越高,处理能力越来越大,但存储系统虽有一定发展,但还是和CPU的距离越来越大。这样就会出现“茶壶里倒饺子”的情况,就是所谓的存储墙问题。cache,正是为了解决这个问题而出现的。
关于cache,我们需要先了解cache的映射方式,写策略,替换策略,cache的优化技术,等等相关内容。这些内容,我们之前都已介绍过了,这里不再赘述,如有疑问,请参考:http://blog.csdn.net/rill_zhen/article/details/9491095
在分析or1200的cache的具体实现之前,我们有必要先了解cache的一般工作机制。为了清晰的展示这个过程,我假设了一个例子,这个例子是MMU模块分析时,那个例子的延伸。
在分析or1200的MMU时,我们假设了一个例子,那个示例中,MMU将变量test的虚拟地址(0x2008),转换成了物理地址(0x1006008)。
cpu访问内存,虚实地址转换,是其中的第一步,在完成虚实转换之后,并不是直接用这个地址访问外部的SDRAM,而是MMU先将物理地址发送到cache,如果cache hit则直接ack cpu,如果cache miss则才需要访问下一级cache或外部SDRAM。
上面我们介绍了cache的大致工作流程,但是,cache的具体工作细节是怎样的呢?
得到test的物理地址之后是如何运作的呢,下面,我们就以直接映射的,大小为8K,line数目为512,line宽度为16-Bytes的一个cache,来说明,如下图所示:
通过这幅图,我们可以很清楚的看到其工作细节。
说明:
a,这个cache的映射方式是direct mapped。
b,cache的总容量是8K,也正好就是一个内存页。
c,整个cache有512个cache line,或者叫cache entry。
d,每个cache line缓存16个字节的数据。
e,由于是直接映射,所以不存在什么替换算法,哪个line出现cache miss就替换哪个。
f,写策略,write through和write back两种。
g,由于cache一般是对软件编程模型透明的,所以很少需要和软件交互,只需要最基本的控制,比如,需要把那个通道lock啊,cache flush啊,如果采用LRU替换算法,及时更新LRU值啊,等等。这一点和MMU大不相同,MMU需要软件的大量的干预和控制。
h,简单介绍一下工作机制:
首先,cache将虚拟地址的index域进行取模运算(%),具体和那个值取模,就看cache line的数量和缓存的数据大小。本例子中cacheline数量是512,缓存数量是16B,所以,需要将index分成cache line index(定位到哪一行),和行内偏移(定位到这一行的哪一个字节)。
cache根据cache line index定位到cache的具体一行,判断这一行的valid标志,如果有效,在将这一行的tag和MMU产生的PPN进行比较(因为一个cache line可能会对应多个内存地址)。如果tag和PPN匹配,那么说明cache hit,如果两个判断条件有一个不满足,说明cache miss,这时,cache会burst access(突发访问,本例子是叠4,每次4B,正好16B),更新这一个cache line。
i,cache的操作
刷新:cache将valid置0即可。
锁定:加入有某个程序运行时间很长,为了防止其他程序在出现cache miss时将这个程序的cache line刷新,可以将这个程序使用的cache line 锁定。具体锁定方式可以是通道锁定,也可以是某一行锁定(将整个cache分成若干组,每个组有若干行,一个组就叫一个通道(way))。
上面我们介绍了直接映射cache的工作机制,其他两种映射方式的cache也大体相同,不同的地方是cache line搜索方法,替换策略,写策略不同。
全相连映射cache的工作机制,如下图所示:
介于直接映射和全相连映射之间,不再赘述。
了解了cache的工作机制之后,再分析or1200的cache的具体实现就相对容易一些,由于cache只是内存的一个子集,没有独立的编程空间,所以与软件的交互比较少,分析起来就更简单一些。
or1200的cache采用直接映射方式,大小是8K,共512个entry,每个line缓存16个字节,每个line由1-bit标志位,19-bit tag和16*8-bit数据组成。
上面我们已经详细说明了这种cache的工作机制,or1200的cache也不例外。
or1200的cache,由qmem模块组成一级cache,dcache/icache组成二级cache,sb模块组成数据的三级cache。
下面是整个ordb2a开饭板的存储系统的框图,从中,我们可以清晰的看出整个系统的存储子系统的数据通路。
qmem模块的实质是一块小的RAM,在or1200_define.v中,对qmem有如下描述,从中我们可以知道qmem的作用,意义,容量等信息。
///////////////////////////////////////////////// // // Quick Embedded Memory (QMEM) // // // Quick Embedded Memory // // Instantiation of dedicated insn/data memory (RAM or ROM). // Insn fetch has effective throughput 1insn / clock cycle. // Data load takes two clock cycles / access, data store // takes 1 clock cycle / access (if there is no insn fetch)). // Memory instantiation is shared between insn and data, // meaning if insn fetch are performed, data load/store // performance will be lower. // // Main reason for QMEM is to put some time critical functions // into this memory and to have predictable and fast access // to these functions. (soft fpu, context switch, exception // handlers, stack, etc) // // It makes design a bit bigger and slower. QMEM sits behind // IMMU/DMMU so all addresses are physical (so the MMUs can be // used with QMEM and QMEM is seen by the CPU just like any other // memory in the system). IC/DC are sitting behind QMEM so the // whole design timing might be worse with QMEM implemented. // //`define OR1200_QMEM_IMPLEMENTED // // Base address and mask of QMEM // // Base address defines first address of QMEM. Mask defines // QMEM range in address space. Actual size of QMEM is however // determined with instantiated RAM/ROM. However bigger // mask will reserve more address space for QMEM, but also // make design faster, while more tight mask will take // less address space but also make design slower. If // instantiated RAM/ROM is smaller than space reserved with // the mask, instatiated RAM/ROM will also be shadowed // at higher addresses in reserved space. // `define OR1200_QMEM_IADDR 32'h0080_0000 `define OR1200_QMEM_IMASK 32'hfff0_0000 // Max QMEM size 1MB `define OR1200_QMEM_DADDR 32'h0080_0000 `define OR1200_QMEM_DMASK 32'hfff0_0000 // Max QMEM size 1MB // // QMEM interface byte-select capability // // To enable qmem_sel* ports, define this macro. // //`define OR1200_QMEM_BSEL // // QMEM interface acknowledge // // To enable qmem_ack port, define this macro. // //`define OR1200_QMEM_ACK
从上面我们可以看出,qmem是总线上的一个buffer,但对于cpu内核来说,qmem不是透明的,是可见的。
当cpu从读cache的时候,地址线对于cache和qmem都是有效的,但是qmem要比cache响应快。
当cpu写cache的时候,如果在qmem的地址范围内,qmem也会更新,所以不会出错。
咱们举一个例子。
假如你是qmem,我是cache,还有路人甲是cpu。
你和我负责提供旅游路线咨询,咱们使用同一个电话号码的两个分机,如果有人打电话,咱们的分机会同时响铃。
你手上只有北京的地图,我手上有全国地图,路人甲如果想去一个地方旅游,打了咱们的那个电话(read cache),咱们都会拿起电话,如果路人甲想去北京旅游,由于你的地图小,反应快,这样,你就会把路线信息告诉他(qmem hit)。如果路人甲想去上海旅游,上海不在你的地图上(qmem miss),那么就由我来回答,但时间会长一点(而不是你再问我,还由你来回答)。
如果路人甲从北京旅游回来打咱们的电话说,颐和园正在维修,不对外开放了(write cache),你和我都会听到这个信息,所以你的北京地图和我的全国地图都会更新。
所以,如果有路人乙再去北京旅游,直接从你那得到消息是不会出错的。
既然qmem对cpu是可见的,那么qmem就和外部的SDRAM在逻辑上是平等的。所不同的是qmem容量更小,访问速度更快(只需一个clock),还有一个重要的区别是qmem miss不会引起异常,但是如果SDRAM miss就会引起异常。
既然qmem对cpu是可见的,那么qmem的使用是由软件程序员控制的。软件程序员必须知道qmem的存在。
既然qmem也是编程空间的一部分,那么qmem就和一般的SDRAM一样使用了,所以,也就没有reload操作了。
关于qmem,我们需要注意一下几点:
a,qmem的地址空间有多大,如下代码所示:
// // Base address and mask of QMEM // // Base address defines first address of QMEM. Mask defines // QMEM range in address space. Actual size of QMEM is however // determined with instantiated RAM/ROM. However bigger // mask will reserve more address space for QMEM, but also // make design faster, while more tight mask will take // less address space but also make design slower. If // instantiated RAM/ROM is smaller than space reserved with // the mask, instatiated RAM/ROM will also be shadowed // at higher addresses in reserved space. // `define OR1200_QMEM_IADDR 32'h0080_0000 `define OR1200_QMEM_IMASK 32'hfff0_0000 // Max QMEM size 1MB `define OR1200_QMEM_DADDR 32'h0080_0000 `define OR1200_QMEM_DMASK 32'hfff0_0000 // Max QMEM size 1MB
b,qmem的hit判断条件
如何判断cpu的访问地址落在qmem范围内呢?代码如下:
// // Address comparison whether QMEM was hit // assign iaddr_qmem_hit = (qmemimmu_adr_i & `OR1200_QMEM_IMASK) == `OR1200_QMEM_IADDR; assign daddr_qmem_hit = (qmemdmmu_adr_i & `OR1200_QMEM_DMASK) == `OR1200_QMEM_DADDR;
c,qmem的实际例化RAM大小
qmem的地址空间是1MB,但实际上qmem模块内部的RAM并没有那么大,这一点,从上面qmem的整体分析部分的注释,可以很清楚的看出。那么qmem的RAM到底有多大呢?代码如下:
// // Instantiation of embedded memory // or1200_spram_2048x32 or1200_qmem_ram( .clk(clk), .rst(rst), .addr(qmem_addr[12:2]), .ce(qmem_en), .we(qmem_we), .oe(1'b1), .di(qmem_di), .doq(qmem_do) );
从中可以看出,qmem实际上只例化了8KB。这就有一个疑问了,qmem的地址空间是1MB,但实际上只例化了8KB的RAM,那么有可能cpu的访问地址在这1MB内,不在8KB内,那qmem如何响应cpu呢?这就涉及到8K的使用问题了,分析如下:
由于qmem实际只有8KB,但其地址空间是1MB,在例化spram_1024x32模块时,地址忽略了高位,只使用了12~2位,所以对于0x80_0000和0x88_0000两个地址会对应相同的qmem的地址,所以这就需要软件的编写人员自己处理类似cache line替换的工作。
If // instantiated RAM/ROM is smaller than space reserved with // the mask, instatiated RAM/ROM will also be shadowed // at higher addresses in reserved space.
qmem模块只有一个RTL文件,就是or1200_qmem_top.v,代码分析,不是代码的复制,粘贴之后加点注释那么简单。为了突出重点,在了解了qmem的大体功能之后,我们需要了解其核心代码,下面,我们分析一下qmem模块的核心,也就是其FSM,如下所示:
`define OR1200_QMEMFSM_IDLE 3'd0 `define OR1200_QMEMFSM_STORE 3'd1 `define OR1200_QMEMFSM_LOAD 3'd2 `define OR1200_QMEMFSM_FETCH 3'd3 // // QMEM control FSM // always @(`OR1200_RST_EVENT rst or posedge clk) if (rst == `OR1200_RST_VALUE) begin state <= `OR1200_QMEMFSM_IDLE; qmem_dack <= 1'b0; qmem_iack <= 1'b0; end else case (state) // synopsys parallel_case `OR1200_QMEMFSM_IDLE: begin if (qmemdmmu_cycstb_i & daddr_qmem_hit & qmemdcpu_we_i & qmem_ack) begin state <= `OR1200_QMEMFSM_STORE; qmem_dack <= 1'b1; qmem_iack <= 1'b0; end else if (qmemdmmu_cycstb_i & daddr_qmem_hit & qmem_ack) begin state <= `OR1200_QMEMFSM_LOAD; qmem_dack <= 1'b1; qmem_iack <= 1'b0; end else if (qmemimmu_cycstb_i & iaddr_qmem_hit & qmem_ack) begin state <= `OR1200_QMEMFSM_FETCH; qmem_iack <= 1'b1; qmem_dack <= 1'b0; end end `OR1200_QMEMFSM_STORE: begin if (qmemdmmu_cycstb_i & daddr_qmem_hit & qmemdcpu_we_i & qmem_ack) begin state <= `OR1200_QMEMFSM_STORE; qmem_dack <= 1'b1; qmem_iack <= 1'b0; end else if (qmemdmmu_cycstb_i & daddr_qmem_hit & qmem_ack) begin state <= `OR1200_QMEMFSM_LOAD; qmem_dack <= 1'b1; qmem_iack <= 1'b0; end else if (qmemimmu_cycstb_i & iaddr_qmem_hit & qmem_ack) begin state <= `OR1200_QMEMFSM_FETCH; qmem_iack <= 1'b1; qmem_dack <= 1'b0; end else begin state <= `OR1200_QMEMFSM_IDLE; qmem_dack <= 1'b0; qmem_iack <= 1'b0; end end `OR1200_QMEMFSM_LOAD: begin if (qmemdmmu_cycstb_i & daddr_qmem_hit & qmemdcpu_we_i & qmem_ack) begin state <= `OR1200_QMEMFSM_STORE; qmem_dack <= 1'b1; qmem_iack <= 1'b0; end else if (qmemdmmu_cycstb_i & daddr_qmem_hit & qmem_ack) begin state <= `OR1200_QMEMFSM_LOAD; qmem_dack <= 1'b1; qmem_iack <= 1'b0; end else if (qmemimmu_cycstb_i & iaddr_qmem_hit & qmem_ack) begin state <= `OR1200_QMEMFSM_FETCH; qmem_iack <= 1'b1; qmem_dack <= 1'b0; end else begin state <= `OR1200_QMEMFSM_IDLE; qmem_dack <= 1'b0; qmem_iack <= 1'b0; end end `OR1200_QMEMFSM_FETCH: begin if (qmemdmmu_cycstb_i & daddr_qmem_hit & qmemdcpu_we_i & qmem_ack) begin state <= `OR1200_QMEMFSM_STORE; qmem_dack <= 1'b1; qmem_iack <= 1'b0; end else if (qmemdmmu_cycstb_i & daddr_qmem_hit & qmem_ack) begin state <= `OR1200_QMEMFSM_LOAD; qmem_dack <= 1'b1; qmem_iack <= 1'b0; end else if (qmemimmu_cycstb_i & iaddr_qmem_hit & qmem_ack) begin state <= `OR1200_QMEMFSM_FETCH; qmem_iack <= 1'b1; qmem_dack <= 1'b0; end else begin state <= `OR1200_QMEMFSM_IDLE; qmem_dack <= 1'b0; qmem_iack <= 1'b0; end end default: begin state <= `OR1200_QMEMFSM_IDLE; qmem_dack <= 1'b0; qmem_iack <= 1'b0; end endcase
可以看出qmem共有4个状态,为了便于查看,我画出了qmem的状态图,如下所示,有状态和状态转移条件,一目了然,不再赘述。
data cache和instruction cache机制相似,这里只分析data cache。
data cache是外部内存的一个子集,其作用也是一般意义上的cache的作用。
这里只说明一下几点:
a,cache的预取,在cache空闲的时候,可以事先将内存中的部分数据填充到cache里,降低cache miss概率。
b,cache的无效控制,如果有些cache line有特殊要求,软件可以设置这些line为无效。
c,cache的锁定,本小节开始部分已经介绍了。
dcache由四个文件组成,分别是:or1200_dc_top.v,or1200_dc_fsm.v,or1200_dc_tag.v,or1200_dc_ram.v。这里只介绍其核心部分,也就是or1200_dc_fsm.v中的FSM,代码如下所示:
`define OR1200_DCFSM_IDLE 3'd0 `define OR1200_DCFSM_CLOADSTORE 3'd1 `define OR1200_DCFSM_LOOP2 3'd2 `define OR1200_DCFSM_LOOP3 3'd3 `define OR1200_DCFSM_LOOP4 3'd4 `define OR1200_DCFSM_FLUSH5 3'd5 `define OR1200_DCFSM_INV6 3'd6 //invalidate `define OR1200_DCFSM_WAITSPRCS7 3'd7 // // Main DC FSM // always @(posedge clk or `OR1200_RST_EVENT rst) begin if (rst == `OR1200_RST_VALUE) begin state <= `OR1200_DCFSM_IDLE; addr_r <= 32'd0; hitmiss_eval <= 1'b0; store <= 1'b0; load <= 1'b0; cnt <= `OR1200_DCLS'd0; cache_miss <= 1'b0; cache_dirty_needs_writeback <= 1'b0; cache_inhibit <= 1'b0; did_early_load_ack <= 1'b0; cache_spr_block_flush <= 1'b0; cache_spr_block_writeback <= 1'b0; end else case (state) // synopsys parallel_case `OR1200_DCFSM_IDLE : begin if (dc_en & (dc_block_flush | dc_block_writeback)) begin cache_spr_block_flush <= dc_block_flush; cache_spr_block_writeback <= dc_block_writeback; hitmiss_eval <= 1'b1; state <= `OR1200_DCFSM_FLUSH5; addr_r <= spr_dat_i; end else if (dc_en & dcqmem_cycstb_i) begin state <= `OR1200_DCFSM_CLOADSTORE; hitmiss_eval <= 1'b1; store <= dcqmem_we_i; load <= !dcqmem_we_i; end end // case: `OR1200_DCFSM_IDLE `OR1200_DCFSM_CLOADSTORE: begin hitmiss_eval <= 1'b0; if (hitmiss_eval) begin cache_inhibit <= dcqmem_ci_i; // Check for cache inhibit here cache_miss <= tagcomp_miss; cache_dirty_needs_writeback <= dirty; addr_r <= lsu_addr; end // Evaluate any cache line load/stores in first cycle: if (hitmiss_eval & tagcomp_miss & !(store & writethrough) & !dcqmem_ci_i) begin // Miss - first either: // 1) write back dirty line if (dirty) begin // Address for writeback addr_r <= {tag, lsu_addr[`OR1200_DCINDXH:2],2'd0}; load <= 1'b0; store <= 1'b1; `ifdef OR1200_VERBOSE $display("%t: dcache miss and dirty", $time); `endif end // 2) load requested line else begin addr_r <= lsu_addr; load <= 1'b1; store <= 1'b0; end // else: !if(dirty) state <= `OR1200_DCFSM_LOOP2; // Set the counter for the burst accesses cnt <= ((1 << `OR1200_DCLS) - 4); end else if (// Strobe goes low !dcqmem_cycstb_i | // Cycle finishes (!hitmiss_eval & (biudata_valid | biudata_error)) | // Cache hit in first cycle.... (hitmiss_eval & !tagcomp_miss & !dcqmem_ci_i & // .. and you're not doing a writethrough store.. !(store & writethrough))) begin state <= `OR1200_DCFSM_IDLE; load <= 1'b0; store <= 1'b0; cache_inhibit <= 1'b0; cache_dirty_needs_writeback <= 1'b0; end end // case: `OR1200_DCFSM_CLOADSTORE `OR1200_DCFSM_LOOP2 : begin // loop/abort if (!dc_en| biudata_error) begin state <= `OR1200_DCFSM_IDLE; load <= 1'b0; store <= 1'b0; cnt <= `OR1200_DCLS'd0; end if (biudata_valid & (|cnt)) begin cnt <= cnt - 4; addr_r[`OR1200_DCLS-1:2] <= addr_r[`OR1200_DCLS-1:2] + 1; end else if (biudata_valid & !(|cnt)) begin state <= `OR1200_DCFSM_LOOP3; addr_r <= lsu_addr; load <= 1'b0; store <= 1'b0; end // Track if we did an early ack during a load if (load_miss_ack) did_early_load_ack <= 1'b1; end // case: `OR1200_DCFSM_LOOP2 `OR1200_DCFSM_LOOP3: begin // figure out next step if (cache_dirty_needs_writeback) begin // Just did store of the dirty line so now load new one load <= 1'b1; // Set the counter for the burst accesses cnt <= ((1 << `OR1200_DCLS) - 4); // Address of line to be loaded addr_r <= lsu_addr; cache_dirty_needs_writeback <= 1'b0; state <= `OR1200_DCFSM_LOOP2; end // if (cache_dirty_needs_writeback) else if (cache_spr_block_flush | cache_spr_block_writeback) begin // Just wrote back the line to memory, we're finished. cache_spr_block_flush <= 1'b0; cache_spr_block_writeback <= 1'b0; state <= `OR1200_DCFSM_WAITSPRCS7; end else begin // Just loaded a new line, finish up did_early_load_ack <= 1'b0; state <= `OR1200_DCFSM_LOOP4; end end // case: `OR1200_DCFSM_LOOP3 `OR1200_DCFSM_LOOP4: begin state <= `OR1200_DCFSM_IDLE; end `OR1200_DCFSM_FLUSH5: begin hitmiss_eval <= 1'b0; if (hitmiss_eval & !tag_v) begin // Not even cached, just ignore cache_spr_block_flush <= 1'b0; cache_spr_block_writeback <= 1'b0; state <= `OR1200_DCFSM_WAITSPRCS7; end else if (hitmiss_eval & tag_v) begin // Tag is valid - what do we do? if ((cache_spr_block_flush | cache_spr_block_writeback) & dirty) begin // Need to writeback // Address for writeback (spr_dat_i has already changed so // use line number from addr_r) addr_r <= {tag, addr_r[`OR1200_DCINDXH:2],2'd0}; load <= 1'b0; store <= 1'b1; `ifdef OR1200_VERBOSE $display("%t: block flush: dirty block", $time); `endif state <= `OR1200_DCFSM_LOOP2; // Set the counter for the burst accesses cnt <= ((1 << `OR1200_DCLS) - 4); end else if (cache_spr_block_flush & !dirty) begin // Line not dirty, just need to invalidate state <= `OR1200_DCFSM_INV6; end // else: !if(dirty) else if (cache_spr_block_writeback & !dirty) begin // Nothing to do - line is valid but not dirty cache_spr_block_writeback <= 1'b0; state <= `OR1200_DCFSM_WAITSPRCS7; end end // if (hitmiss_eval & tag_v) end `OR1200_DCFSM_INV6: begin cache_spr_block_flush <= 1'b0; // Wait until SPR CS goes low before going back to idle if (!spr_cswe) state <= `OR1200_DCFSM_IDLE; end `OR1200_DCFSM_WAITSPRCS7: begin // Wait until SPR CS goes low before going back to idle if (!spr_cswe) state <= `OR1200_DCFSM_IDLE; end endcase // case (state) end // always @ (posedge clk or `OR1200_RST_EVENT rst)
为了便于理解,我画出了其状态图,如下所示:
store buffer,其本质是一个FIFO,相当于一个write back的cache,其功能和相关分析,之前已经做过,请参考:http://blog.csdn.net/rill_zhen/article/details/9491095 中的第2.1章节。
关于这个FIFO的depth和width,or1200-define.v中有如下定义:
// // Number of store buffer entries // // Verified number of entries are 4 and 8 entries // (2 and 3 for OR1200_SB_LOG). OR1200_SB_ENTRIES must // always match 2**OR1200_SB_LOG. // To disable store buffer, undefine // OR1200_SB_IMPLEMENTED. // `define OR1200_SB_LOG 2 // 2 or 3 `define OR1200_SB_ENTRIES 4 // 4 or 8
sb模块包含两个文件,or1200_sb.v和or1200_sb_fifo.v,第二个从文件名就可以看出是一个FIFO,其物理结构是一个双口的RAM,这里只分析第一个,主要代码如下所示:
代码很少,只有150多行。
module or1200_sb( // RISC clock, reset clk, rst, // Internal RISC bus (SB) sb_en, // Internal RISC bus (DC<->SB) dcsb_dat_i, dcsb_adr_i, dcsb_cyc_i, dcsb_stb_i, dcsb_we_i, dcsb_sel_i, dcsb_cab_i, dcsb_dat_o, dcsb_ack_o, dcsb_err_o, // BIU bus sbbiu_dat_o, sbbiu_adr_o, sbbiu_cyc_o, sbbiu_stb_o, sbbiu_we_o, sbbiu_sel_o, sbbiu_cab_o, sbbiu_dat_i, sbbiu_ack_i, sbbiu_err_i ); parameter dw = `OR1200_OPERAND_WIDTH; parameter aw = `OR1200_OPERAND_WIDTH; // // RISC clock, reset // input clk; // RISC clock input rst; // RISC reset // // Internal RISC bus (SB) // input sb_en; // SB enable // // Internal RISC bus (DC<->SB) // input [dw-1:0] dcsb_dat_i; // input data bus input [aw-1:0] dcsb_adr_i; // address bus input dcsb_cyc_i; // WB cycle input dcsb_stb_i; // WB strobe input dcsb_we_i; // WB write enable input dcsb_cab_i; // CAB input input [3:0] dcsb_sel_i; // byte selects output [dw-1:0] dcsb_dat_o; // output data bus output dcsb_ack_o; // ack output output dcsb_err_o; // err output // // BIU bus // output [dw-1:0] sbbiu_dat_o; // output data bus output [aw-1:0] sbbiu_adr_o; // address bus output sbbiu_cyc_o; // WB cycle output sbbiu_stb_o; // WB strobe output sbbiu_we_o; // WB write enable output sbbiu_cab_o; // CAB input output [3:0] sbbiu_sel_o; // byte selects input [dw-1:0] sbbiu_dat_i; // input data bus input sbbiu_ack_i; // ack output input sbbiu_err_i; // err output `ifdef OR1200_SB_IMPLEMENTED // // Internal wires and regs // wire [4+dw+aw-1:0] fifo_dat_i; // FIFO data in wire [4+dw+aw-1:0] fifo_dat_o; // FIFO data out wire fifo_wr; wire fifo_rd; wire fifo_full; wire fifo_empty; wire sel_sb; reg sb_en_reg; reg outstanding_store; reg fifo_wr_ack; // // FIFO data in/out // assign fifo_dat_i = {dcsb_sel_i, dcsb_dat_i, dcsb_adr_i}; assign {sbbiu_sel_o, sbbiu_dat_o, sbbiu_adr_o} = sel_sb ? fifo_dat_o : {dcsb_sel_i, dcsb_dat_i, dcsb_adr_i}; // // Control // assign fifo_wr = dcsb_cyc_i & dcsb_stb_i & dcsb_we_i & ~fifo_full & ~fifo_wr_ack; assign fifo_rd = ~outstanding_store; assign dcsb_dat_o = sbbiu_dat_i; assign dcsb_ack_o = sel_sb ? fifo_wr_ack : sbbiu_ack_i; assign dcsb_err_o = sel_sb ? 1'b0 : sbbiu_err_i; // SB never returns error assign sbbiu_cyc_o = sel_sb ? outstanding_store : dcsb_cyc_i; assign sbbiu_stb_o = sel_sb ? outstanding_store : dcsb_stb_i; assign sbbiu_we_o = sel_sb ? 1'b1 : dcsb_we_i; assign sbbiu_cab_o = sel_sb ? 1'b0 : dcsb_cab_i; assign sel_sb = sb_en_reg & (~fifo_empty | (fifo_empty & outstanding_store)); // // SB enable // always @(posedge clk or `OR1200_RST_EVENT rst) if (rst == `OR1200_RST_VALUE) sb_en_reg <= 1'b0; else if (sb_en & ~dcsb_cyc_i) sb_en_reg <= 1'b1; // enable SB when there is no dcsb transfer in progress else if (~sb_en & (~fifo_empty | (fifo_empty & outstanding_store))) sb_en_reg <= 1'b0; // disable SB when there is no pending transfers from SB // // Store buffer FIFO instantiation // or1200_sb_fifo or1200_sb_fifo ( .clk_i(clk), .rst_i(rst), .dat_i(fifo_dat_i), .wr_i(fifo_wr), .rd_i(fifo_rd), .dat_o(fifo_dat_o), .full_o(fifo_full), .empty_o(fifo_empty) ); // // fifo_rd // always @(posedge clk or `OR1200_RST_EVENT rst) if (rst == `OR1200_RST_VALUE) outstanding_store <= 1'b0; else if (sbbiu_ack_i) outstanding_store <= 1'b0; else if (sel_sb | fifo_wr) outstanding_store <= 1'b1; // // fifo_wr_ack // always @(posedge clk or `OR1200_RST_EVENT rst) if (rst == `OR1200_RST_VALUE) fifo_wr_ack <= 1'b0; else if (fifo_wr) fifo_wr_ack <= 1'b1; else fifo_wr_ack <= 1'b0; `else // !OR1200_SB_IMPLEMENTED assign sbbiu_dat_o = dcsb_dat_i; assign sbbiu_adr_o = dcsb_adr_i; assign sbbiu_cyc_o = dcsb_cyc_i; assign sbbiu_stb_o = dcsb_stb_i; assign sbbiu_we_o = dcsb_we_i; assign sbbiu_cab_o = dcsb_cab_i; assign sbbiu_sel_o = dcsb_sel_i; assign dcsb_dat_o = sbbiu_dat_i; assign dcsb_ack_o = sbbiu_ack_i; assign dcsb_err_o = sbbiu_err_i; `endif endmodule
biu(bus ingerface unit)模块,是or1200_top和外界进行数据交换的窗口,对于or1200,例化了两个,分别是dbiu和ibiu。biu模块除了和外界交换数据外,还有判断字节对齐等功能。
这个模块主要是一个wishbone协议的slave和master的一个wrapper,如果你对wishbone总线protocol比较熟悉的话,这个模块看起来就简单多了,我之前也写过wishbone的相关的内容,请参考:http://blog.csdn.net/rill_zhen/article/details/8659788
biu模块包含一个文件,or1200_wb_biu.v,主要是wishbone协议的时序产生逻辑,这里不做细说,为了保持本文的完整性,其主要代码,如下所示:
module or1200_wb_biu( // RISC clock, reset and clock control clk, rst, clmode, // WISHBONE interface wb_clk_i, wb_rst_i, wb_ack_i, wb_err_i, wb_rty_i, wb_dat_i, wb_cyc_o, wb_adr_o, wb_stb_o, wb_we_o, wb_sel_o, wb_dat_o, `ifdef OR1200_WB_CAB wb_cab_o, `endif `ifdef OR1200_WB_B3 wb_cti_o, wb_bte_o, `endif // Internal RISC bus biu_dat_i, biu_adr_i, biu_cyc_i, biu_stb_i, biu_we_i, biu_sel_i, biu_cab_i, biu_dat_o, biu_ack_o, biu_err_o ); parameter dw = `OR1200_OPERAND_WIDTH; parameter aw = `OR1200_OPERAND_WIDTH; parameter bl = 4; /* Can currently be either 4 or 8 - the two optional line sizes for the OR1200. */ // // RISC clock, reset and clock control // input clk; // RISC clock input rst; // RISC reset input [1:0] clmode; // 00 WB=RISC, 01 WB=RISC/2, 10 N/A, 11 WB=RISC/4 // // WISHBONE interface // input wb_clk_i; // clock input input wb_rst_i; // reset input input wb_ack_i; // normal termination input wb_err_i; // termination w/ error input wb_rty_i; // termination w/ retry input [dw-1:0] wb_dat_i; // input data bus output wb_cyc_o; // cycle valid output output [aw-1:0] wb_adr_o; // address bus outputs output wb_stb_o; // strobe output output wb_we_o; // indicates write transfer output [3:0] wb_sel_o; // byte select outputs output [dw-1:0] wb_dat_o; // output data bus `ifdef OR1200_WB_CAB output wb_cab_o; // consecutive address burst `endif `ifdef OR1200_WB_B3 output [2:0] wb_cti_o; // cycle type identifier output [1:0] wb_bte_o; // burst type extension `endif // // Internal RISC interface // input [dw-1:0] biu_dat_i; // input data bus input [aw-1:0] biu_adr_i; // address bus input biu_cyc_i; // WB cycle input biu_stb_i; // WB strobe input biu_we_i; // WB write enable input biu_cab_i; // CAB input input [3:0] biu_sel_i; // byte selects output [31:0] biu_dat_o; // output data bus output biu_ack_o; // ack output output biu_err_o; // err output // // Registers // wire wb_ack; // normal termination reg [aw-1:0] wb_adr_o; // address bus outputs reg wb_cyc_o; // cycle output reg wb_stb_o; // strobe output reg wb_we_o; // indicates write transfer reg [3:0] wb_sel_o; // byte select outputs `ifdef OR1200_WB_CAB reg wb_cab_o; // CAB output `endif `ifdef OR1200_WB_B3 reg [2:0] wb_cti_o; // cycle type identifier reg [1:0] wb_bte_o; // burst type extension `endif `ifdef OR1200_NO_DC reg [dw-1:0] wb_dat_o; // output data bus `else assign wb_dat_o = biu_dat_i; // No register on this - straight from DCRAM `endif `ifdef OR1200_WB_RETRY reg [`OR1200_WB_RETRY-1:0] retry_cnt; // Retry counter `else wire retry_cnt; assign retry_cnt = 1'b0; `endif `ifdef OR1200_WB_B3 reg [3:0] burst_len; // burst counter `endif reg biu_stb_reg; // WB strobe wire biu_stb; // WB strobe reg wb_cyc_nxt; // next WB cycle value reg wb_stb_nxt; // next WB strobe value reg [2:0] wb_cti_nxt; // next cycle type identifier value reg wb_ack_cnt; // WB ack toggle counter reg wb_err_cnt; // WB err toggle counter reg wb_rty_cnt; // WB rty toggle counter reg biu_ack_cnt; // BIU ack toggle counter reg biu_err_cnt; // BIU err toggle counter reg biu_rty_cnt; // BIU rty toggle counter wire biu_rty; // BIU rty indicator reg [1:0] wb_fsm_state_cur; // WB FSM - surrent state reg [1:0] wb_fsm_state_nxt; // WB FSM - next state wire [1:0] wb_fsm_idle = 2'h0; // WB FSM state - IDLE wire [1:0] wb_fsm_trans = 2'h1; // WB FSM state - normal TRANSFER wire [1:0] wb_fsm_last = 2'h2; // EB FSM state - LAST transfer // // WISHBONE I/F <-> Internal RISC I/F conversion // //assign wb_ack = wb_ack_i; assign wb_ack = wb_ack_i & !wb_err_i & !wb_rty_i; // // WB FSM - register part // always @(posedge wb_clk_i or `OR1200_RST_EVENT wb_rst_i) begin if (wb_rst_i == `OR1200_RST_VALUE) wb_fsm_state_cur <= wb_fsm_idle; else wb_fsm_state_cur <= wb_fsm_state_nxt; end // // WB burst tength counter // always @(posedge wb_clk_i or `OR1200_RST_EVENT wb_rst_i) begin if (wb_rst_i == `OR1200_RST_VALUE) begin burst_len <= 0; end else begin // burst counter if (wb_fsm_state_cur == wb_fsm_idle) burst_len <= bl[3:0] - 2; else if (wb_stb_o & wb_ack) burst_len <= burst_len - 1; end end // // WB FSM - combinatorial part // always @(wb_fsm_state_cur or burst_len or wb_err_i or wb_rty_i or wb_ack or wb_cti_o or wb_sel_o or wb_stb_o or wb_we_o or biu_cyc_i or biu_stb or biu_cab_i or biu_sel_i or biu_we_i) begin // States of WISHBONE Finite State Machine case(wb_fsm_state_cur) // IDLE wb_fsm_idle : begin wb_cyc_nxt = biu_cyc_i & biu_stb; wb_stb_nxt = biu_cyc_i & biu_stb; wb_cti_nxt = {!biu_cab_i, 1'b1, !biu_cab_i}; if (biu_cyc_i & biu_stb) wb_fsm_state_nxt = wb_fsm_trans; else wb_fsm_state_nxt = wb_fsm_idle; end // normal TRANSFER wb_fsm_trans : begin wb_cyc_nxt = !wb_stb_o | !wb_err_i & !wb_rty_i & !(wb_ack & wb_cti_o == 3'b111); wb_stb_nxt = !wb_stb_o | !wb_err_i & !wb_rty_i & !wb_ack | !wb_err_i & !wb_rty_i & wb_cti_o == 3'b010 ; wb_cti_nxt[2] = wb_stb_o & wb_ack & burst_len == 'h0 | wb_cti_o[2]; wb_cti_nxt[1] = 1'b1 ; wb_cti_nxt[0] = wb_stb_o & wb_ack & burst_len == 'h0 | wb_cti_o[0]; if ((!biu_cyc_i | !biu_stb | !biu_cab_i | biu_sel_i != wb_sel_o | biu_we_i != wb_we_o) & wb_cti_o == 3'b010) wb_fsm_state_nxt = wb_fsm_last; else if ((wb_err_i | wb_rty_i | wb_ack & wb_cti_o==3'b111) & wb_stb_o) wb_fsm_state_nxt = wb_fsm_idle; else wb_fsm_state_nxt = wb_fsm_trans; end // LAST transfer wb_fsm_last : begin wb_cyc_nxt = !wb_stb_o | !wb_err_i & !wb_rty_i & !(wb_ack & wb_cti_o == 3'b111); wb_stb_nxt = !wb_stb_o | !wb_err_i & !wb_rty_i & !(wb_ack & wb_cti_o == 3'b111); wb_cti_nxt[2] = wb_ack & wb_stb_o | wb_cti_o[2]; wb_cti_nxt[1] = 1'b1 ; wb_cti_nxt[0] = wb_ack & wb_stb_o | wb_cti_o[0]; if ((wb_err_i | wb_rty_i | wb_ack & wb_cti_o == 3'b111) & wb_stb_o) wb_fsm_state_nxt = wb_fsm_idle; else wb_fsm_state_nxt = wb_fsm_last; end // default state default:begin wb_cyc_nxt = 1'bx; wb_stb_nxt = 1'bx; wb_cti_nxt = 3'bxxx; wb_fsm_state_nxt = 2'bxx; end endcase end // // WB FSM - output signals // always @(posedge wb_clk_i or `OR1200_RST_EVENT wb_rst_i) begin if (wb_rst_i == `OR1200_RST_VALUE) begin wb_cyc_o <= 1'b0; wb_stb_o <= 1'b0; wb_cti_o <= 3'b111; wb_bte_o <= (bl==8) ? 2'b10 : (bl==4) ? 2'b01 : 2'b00; `ifdef OR1200_WB_CAB wb_cab_o <= 1'b0; `endif wb_we_o <= 1'b0; wb_sel_o <= 4'hf; wb_adr_o <= {aw{1'b0}}; `ifdef OR1200_NO_DC wb_dat_o <= {dw{1'b0}}; `endif end else begin wb_cyc_o <= wb_cyc_nxt; if (wb_ack & wb_cti_o == 3'b111) wb_stb_o <= 1'b0; else wb_stb_o <= wb_stb_nxt; `ifndef OR1200_NO_BURSTS wb_cti_o <= wb_cti_nxt; `endif wb_bte_o <= (bl==8) ? 2'b10 : (bl==4) ? 2'b01 : 2'b00; `ifdef OR1200_WB_CAB wb_cab_o <= biu_cab_i; `endif // we and sel - set at beginning of access if (wb_fsm_state_cur == wb_fsm_idle) begin wb_we_o <= biu_we_i; wb_sel_o <= biu_sel_i; end // adr - set at beginning of access and changed at every termination if (wb_fsm_state_cur == wb_fsm_idle) begin wb_adr_o <= biu_adr_i; end else if (wb_stb_o & wb_ack) begin if (bl==4) begin wb_adr_o[3:2] <= wb_adr_o[3:2] + 1; end if (bl==8) begin wb_adr_o[4:2] <= wb_adr_o[4:2] + 1; end end `ifdef OR1200_NO_DC // dat - write data changed after avery subsequent write access if (!wb_stb_o) begin wb_dat_o <= biu_dat_i; end `endif end end // // WB & BIU termination toggle counters // always @(posedge wb_clk_i or `OR1200_RST_EVENT wb_rst_i) begin if (wb_rst_i == `OR1200_RST_VALUE) begin wb_ack_cnt <= 1'b0; wb_err_cnt <= 1'b0; wb_rty_cnt <= 1'b0; end else begin // WB ack toggle counter if (wb_fsm_state_cur == wb_fsm_idle | !(|clmode)) wb_ack_cnt <= 1'b0; else if (wb_stb_o & wb_ack) wb_ack_cnt <= !wb_ack_cnt; // WB err toggle counter if (wb_fsm_state_cur == wb_fsm_idle | !(|clmode)) wb_err_cnt <= 1'b0; else if (wb_stb_o & wb_err_i) wb_err_cnt <= !wb_err_cnt; // WB rty toggle counter if (wb_fsm_state_cur == wb_fsm_idle | !(|clmode)) wb_rty_cnt <= 1'b0; else if (wb_stb_o & wb_rty_i) wb_rty_cnt <= !wb_rty_cnt; end end always @(posedge clk or `OR1200_RST_EVENT rst) begin if (rst == `OR1200_RST_VALUE) begin biu_stb_reg <= 1'b0; biu_ack_cnt <= 1'b0; biu_err_cnt <= 1'b0; biu_rty_cnt <= 1'b0; `ifdef OR1200_WB_RETRY retry_cnt <= {`OR1200_WB_RETRY{1'b0}}; `endif end else begin // BIU strobe if (biu_stb_i & !biu_cab_i & biu_ack_o) biu_stb_reg <= 1'b0; else biu_stb_reg <= biu_stb_i; // BIU ack toggle counter if (wb_fsm_state_cur == wb_fsm_idle | !(|clmode)) biu_ack_cnt <= 1'b0 ; else if (biu_ack_o) biu_ack_cnt <= !biu_ack_cnt ; // BIU err toggle counter if (wb_fsm_state_cur == wb_fsm_idle | !(|clmode)) biu_err_cnt <= 1'b0 ; else if (wb_err_i & biu_err_o) biu_err_cnt <= !biu_err_cnt ; // BIU rty toggle counter if (wb_fsm_state_cur == wb_fsm_idle | !(|clmode)) biu_rty_cnt <= 1'b0 ; else if (biu_rty) biu_rty_cnt <= !biu_rty_cnt ; `ifdef OR1200_WB_RETRY if (biu_ack_o | biu_err_o) retry_cnt <= {`OR1200_WB_RETRY{1'b0}}; else if (biu_rty) retry_cnt <= retry_cnt + 1'b1; `endif end end assign biu_stb = biu_stb_i & biu_stb_reg; // // Input BIU data bus // assign biu_dat_o = wb_dat_i; // // Input BIU termination signals // assign biu_rty = (wb_fsm_state_cur == wb_fsm_trans) & wb_rty_i & wb_stb_o & (wb_rty_cnt ~^ biu_rty_cnt); assign biu_ack_o = (wb_fsm_state_cur == wb_fsm_trans) & wb_ack & wb_stb_o & (wb_ack_cnt ~^ biu_ack_cnt); assign biu_err_o = (wb_fsm_state_cur == wb_fsm_trans) & wb_err_i & wb_stb_o & (wb_err_cnt ~^ biu_err_cnt) `ifdef OR1200_WB_RETRY | biu_rty & retry_cnt[`OR1200_WB_RETRY-1]; `else ; `endif endmodule
终于可以告一段落了,下面弄个小问题放松一下。
很多人可能曾经遇到过这样一个软件方面的笔试题,题目是,下面两段程序,一般情况下,哪个的执行时间短(假设cache大小为8K)?
自此,我们完成了对OpenRISC的MMU,cache系统的分析,对计算机体系结构中很重要的部分--存储器组织有了一个完整,清晰,透彻的了解了。
1,ORPSoC RTL code
2,OpenRISC arch-manual
// // Instantiation of embedded memory // or1200_spram_2048x32 or1200_qmem_ram( .clk(clk), .rst(rst), .addr(qmem_addr[12:2]), .ce(qmem_en), .we(qmem_we), .oe(1'b1), .di(qmem_di), .doq(qmem_do) );