snooping response handling fix

driver/rtlsim/vortex.cpp

@@ -3,8 +3,7 @@
 #include <stdlib.h>
 #include <assert.h>
 #include <iostream>
-#include <thread>
+#include <future>
-#include <mutex>
 #include <chrono>
 
 #include <vortex.h>
@@ -59,20 +58,14 @@ private:
 
 class vx_device {    
 public:
-    vx_device() 
+    vx_device() {        
-        : is_done_(false) {
-        thread_ = new std::thread(__thread_proc__, this);        
         mem_allocation_ = vx_dev_caps(VX_CAPS_ALLOC_BASE_ADDR);
+        simulator_.attach_ram(&ram_);
     } 
 
     ~vx_device() {     
-        if (thread_) {
+        if (future_.valid()) {
-            mutex_.lock();
+            future_.wait();
-            is_done_ = true;
-            mutex_.unlock();
-
-            thread_->join();
-            delete thread_;
         }
     }
 
@@ -115,79 +108,48 @@ public:
         return 0;
     }
 
-    int flush_caches(size_t dev_maddr, size_t size) {
-        
-        mutex_.lock();     
-        simulator_.attach_ram(&ram_);
-        simulator_.flush_caches(dev_maddr, size);
-        simulator_.attach_ram(nullptr);
-        mutex_.unlock();
-
-        return 0;
-    }
-
     int start() {   
-
+        if (future_.valid()) {
-        mutex_.lock();     
+            future_.wait(); // ensure prior run completed
+        }
+        future_ = std::async(std::launch::async, [&]{             
             simulator_.reset();        
-        simulator_.attach_ram(&ram_);
+            while (simulator_.is_busy()) {
-        mutex_.unlock();       
+                simulator_.step();
-
+            }
+        });
         return 0;
     }
 
     int wait(long long timeout) {
+        if (!future_.valid())
+            return 0;
         auto timeout_sec = (timeout < 0) ? timeout : (timeout / 1000);
+        std::chrono::seconds wait_time(1);
         for (;;) {
-            mutex_.lock();
+            auto status = future_.wait_for(wait_time); // wait for 1 sec and check status
-            bool is_busy = simulator_.is_busy();
+            if (status == std::future_status::ready 
-            mutex_.unlock();
+             || 0 == timeout_sec--)
-
-            if (!is_busy || 0 == timeout_sec--) {
-                if (!is_busy) {
-                    mutex_.lock();
-                    simulator_.attach_ram(nullptr);
-                    mutex_.unlock();
-                }
                 break;
         }
-
-            std::this_thread::sleep_for(std::chrono::seconds(1));            
-        }
         return 0;
     }
 
+    int flush_caches(size_t dev_maddr, size_t size) {
+        if (future_.valid()) {
+            future_.wait(); // ensure prior run completed
+        }
+        simulator_.flush_caches(dev_maddr, size);
+        return 0;
+    }
+
+
 private:
 
-    void thread_proc() {
-        std::cout << "Device ready..." << std::endl;
-
-        for (;;) {
-            mutex_.lock();
-            bool is_done = is_done_;
-            mutex_.unlock();
-
-            if (is_done)
-                break;
-
-            mutex_.lock();
-            simulator_.step();
-            mutex_.unlock();
-        }
-
-        std::cout << "Device shutdown..." << std::endl;
-    }
-
-    static void __thread_proc__(vx_device* device) {
-        device->thread_proc();
-    }
-
-    bool is_done_; 
     size_t mem_allocation_;     
     RAM ram_;
     Simulator simulator_;
-    std::thread* thread_;   
+    std::future<void> future_;
-    std::mutex mutex_;
 };
 
 ///////////////////////////////////////////////////////////////////////////////

hw/rtl/VX_define.vh

@@ -29,11 +29,6 @@
         if (!(cond)) $error(msg);   \
     endgenerate
 
-`define UNUSED(x)               \
-    `IGNORE_WARNINGS_BEGIN      \
-        if (x != 0) begin end   \
-    `IGNORE_WARNINGS_END
-
 `define CLOG2(x)    $clog2(x)
 `define FLOG2(x)    ($clog2(x) - (((1 << $clog2(x)) > x) ? 1 : 0))
 `define LOG2UP(x)   ((x > 1) ? $clog2(x) : 1)

hw/rtl/VX_icache_stage.v

@@ -64,10 +64,10 @@ module VX_icache_stage #(
 
     /*always_comb begin
         if (1'($time & 1) && icache_req_if.core_req_ready && icache_req_if.core_req_valid) begin
-            $display("*** %t: I%01d$ req: pc=%0h, warp=%0d", $time, CORE_ID, fe_inst_meta_fi.inst_pc, fe_inst_meta_fi.warp_num);
+            $display("*** %t: I%01d$ req: tag=%0h, pc=%0h, warp=%0d", $time, CORE_ID, icache_req_if.core_req_tag, fe_inst_meta_fi.inst_pc, fe_inst_meta_fi.warp_num);
         end
         if (1'($time & 1) && icache_rsp_if.core_rsp_ready && icache_rsp_if.core_rsp_valid) begin
-            $display("*** %t: I%01d$ rsp: pc=%0h, warp=%0d, instr=%0h", $time, CORE_ID, fe_inst_meta_id.inst_pc, fe_inst_meta_id.warp_num, fe_inst_meta_id.instruction);
+            $display("*** %t: I%01d$ rsp: tag=%0h, pc=%0h, warp=%0d, instr=%0h", $time, CORE_ID, icache_rsp_if.core_rsp_tag, fe_inst_meta_id.inst_pc, fe_inst_meta_id.warp_num, fe_inst_meta_id.instruction);
         end
     end*/
 

hw/rtl/VX_lsu_unit.v

@@ -64,10 +64,10 @@ module VX_lsu_unit #(
 
     /*always_comb begin
         if (1'($time & 1) && dcache_req_if.core_req_ready && (| dcache_req_if.core_req_valid)) begin
-            $display("*** %t: D%01d$ req: valid=%b, addr=%0h, r=%0d, w=%0d, pc=%0h, rd=%0d, warp=%0d, data=%0h", $time, CORE_ID, use_valid, use_address, use_mem_read, use_mem_write, use_pc, use_rd, use_warp_num, use_store_data);
+            $display("*** %t: D%01d$ req: valid=%b, addr=%0h, tag=%0h, r=%0d, w=%0d, pc=%0h, rd=%0d, warp=%0d, data=%0h", $time, CORE_ID, use_valid, use_address, dcache_req_if.core_req_tag, use_mem_read, use_mem_write, use_pc, use_rd, use_warp_num, use_store_data);
         end
         if (1'($time & 1) && dcache_rsp_if.core_rsp_ready && (| dcache_rsp_if.core_rsp_valid)) begin
-            $display("*** %t: D%01d$ rsp: valid=%b, pc=%0h, rd=%0d, warp=%0d, data=%0h", $time, CORE_ID, mem_wb_if.valid, mem_wb_if.pc, mem_wb_if.rd, mem_wb_if.warp_num, mem_wb_if.data);
+            $display("*** %t: D%01d$ rsp: valid=%b, tag=%0h, pc=%0h, rd=%0d, warp=%0d, data=%0h", $time, CORE_ID, mem_wb_if.valid, dcache_rsp_if.core_rsp_tag, mem_wb_if.pc, mem_wb_if.rd, mem_wb_if.warp_num, mem_wb_if.data);
         end
     end*/
     

hw/rtl/Vortex_Socket.v

@@ -330,7 +330,7 @@ module Vortex_Socket (
 
     /*always_comb begin
         if (1'($time & 1) && (dram_req_read || dram_req_write) && dram_req_ready) begin
-            $display("*** %t: DRAM req: w=%b addr=%0h, tag=%0h, data=%0h", $time, dram_req_write, dram_req_addr, dram_req_tag, dram_req_data);
+            $display("*** %t: DRAM req: w=%b addr=%0h, tag=%0h, data=%0h", $time, dram_req_write, {dram_req_addr, `CLOG2(`GLOBAL_BLOCK_SIZE)'(0)}, dram_req_tag, dram_req_data);
         end
         if (1'($time & 1) && dram_rsp_valid && dram_rsp_ready) begin
             $display("*** %t: DRAM rsp: tag=%0h, data=%0h", $time, dram_rsp_tag, dram_rsp_data);

hw/rtl/cache/VX_bank.v

@@ -407,10 +407,9 @@ module VX_bank #(
         .out ({is_snp_st2 , snrq_tag_st2,                   fill_saw_dirty_st2 , is_fill_st2                   , valid_st2        , addr_st2                  , wsel_st2,                   writeword_st2                  , readword_st2 , readdata_st2 , readtag_st2 , miss_st2 , dirty_st2 , inst_meta_st2                  })
     );
 
-    wire should_flush;
-    wire dwbq_push;
-
     wire cwbq_full;
+    wire dwbq_push;
+    wire dwbq_empty;    
     wire dwbq_full;
     wire srpq_full;
     wire invalidate_fill;
@@ -420,7 +419,6 @@ module VX_bank #(
                    && !is_snp_st2 
                    && miss_st2 
                    && !mrvq_full 
-                   && !(should_flush && dwbq_push) 
                    && !((is_snp_st2 && valid_st2 && srpq_full) 
                      || ((valid_st2 && !miss_st2) && cwbq_full) 
                      || (((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) && dwbq_full) 
@@ -443,7 +441,7 @@ module VX_bank #(
         .clk                     (clk),
         .reset                   (reset),
         // Enqueue
-        .miss_add                (miss_add), // Need to do all 
+        .miss_add                (miss_add),
         .miss_add_addr           (miss_add_addr),
         .miss_add_wsel           (miss_add_wsel),
         .miss_add_data           (miss_add_data),
@@ -505,30 +503,16 @@ module VX_bank #(
         .full    (cwbq_full)
     );
 
-    assign should_flush = valid_st2 
-                       && (miss_add_mem_write != `BYTE_EN_NO) 
-                       && !is_snp_st2 
-                       && !is_fill_st2;
-
     // Enqueue to DWB Queue
-    assign dwbq_push = ((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2 || should_flush) 
+    assign dwbq_push = ((valid_st2 && miss_st2 && dirty_st2) || fill_saw_dirty_st2) 
                     && !dwbq_full 
                     && !((is_snp_st2 && valid_st2 && srpq_full) 
                       || ((valid_st2 && !miss_st2) && cwbq_full) 
                       || (valid_st2 && miss_st2 && mrvq_full) 
                       || (valid_st2 && miss_st2 && !invalidate_fill && ~dram_fill_req_ready));
 
-    wire[`LINE_ADDR_WIDTH-1:0] dwbq_req_addr;
+    wire [`BANK_LINE_WIDTH-1:0] dwbq_req_data = readdata_st2;
-    wire[`BANK_LINE_WIDTH-1:0] dwbq_req_data;
+    wire [`LINE_ADDR_WIDTH-1:0] dwbq_req_addr = {readtag_st2, addr_st2[`LINE_SELECT_BITS-1:0]};
-    wire       dwbq_empty;
-
-    if (SNOOP_FORWARDING) begin
-        assign dwbq_req_data = (should_flush && dwbq_push) ? writeword_st2 : readdata_st2;
-        assign dwbq_req_addr = (should_flush && dwbq_push) ? addr_st2 : {readtag_st2, addr_st2[`LINE_SELECT_BITS-1:0]};
-    end else begin
-        assign dwbq_req_data = readdata_st2;
-        assign dwbq_req_addr = {readtag_st2, addr_st2[`LINE_SELECT_BITS-1:0]};
-    end
 
     wire possible_fill = valid_st2 && miss_st2 && dram_fill_req_ready && ~is_snp_st2;
     wire [`LINE_ADDR_WIDTH-1:0] fill_invalidator_addr = addr_st2;

hw/rtl/cache/VX_tag_data_access.v

@@ -72,6 +72,9 @@ module VX_tag_data_access #(
                        && ((valid_req_st1e && !use_read_valid_st1e) 
                         || (valid_req_st1e && use_read_valid_st1e && !tags_match)); 
 
+    wire[`TAG_SELECT_BITS-1:0] writetag_st1e = writeaddr_st1e[`TAG_LINE_ADDR_RNG];
+    wire[`LINE_SELECT_BITS-1:0] writeladdr_st1e = writeaddr_st1e[`LINE_SELECT_BITS-1:0];
+
     VX_tag_data_structure #(
         .CACHE_SIZE             (CACHE_SIZE),
         .BANK_LINE_SIZE         (BANK_LINE_SIZE),
@@ -91,8 +94,8 @@ module VX_tag_data_access #(
         .invalidate  (invalidate_line),
         .write_enable(use_write_enable),
         .write_fill  (real_writefill),
-        .write_addr  (writeaddr_st1e[`LINE_SELECT_BITS-1:0]),
+        .write_addr  (writeladdr_st1e),
-        .tag_index   (writeaddr_st1e[`TAG_LINE_ADDR_RNG]),
+        .tag_index   (writetag_st1e),
         .write_data  (use_write_data),
         .fill_sent   (fill_sent)
     );
@@ -125,20 +128,18 @@ module VX_tag_data_access #(
 
     assign use_read_valid_st1e = read_valid_st1c[STAGE_1_CYCLES-1] || ~DRAM_ENABLE; // If shared memory, always valid
     assign use_read_dirty_st1e = read_dirty_st1c[STAGE_1_CYCLES-1] && DRAM_ENABLE; // Dirty only applies in Dcache
-    assign use_read_tag_st1e   = DRAM_ENABLE ? read_tag_st1c[STAGE_1_CYCLES-1] : writeaddr_st1e[`TAG_LINE_ADDR_RNG]; // Tag is always the same in SM
+    assign use_read_tag_st1e   = DRAM_ENABLE ? read_tag_st1c[STAGE_1_CYCLES-1] : writetag_st1e; // Tag is always the same in SM
-
+    assign use_read_data_st1e  = read_data_st1c[STAGE_1_CYCLES-1];
-    for (i = 0; i < `BANK_LINE_WORDS; i++) begin
-        assign use_read_data_st1e[i * `WORD_WIDTH +: `WORD_WIDTH]  = read_data_st1c[STAGE_1_CYCLES-1][i * `WORD_WIDTH +: `WORD_WIDTH];
-    end
 
     wire force_write  = real_writefill;
+    wire should_write;
 
     wire [`BANK_LINE_WORDS-1:0][3:0] we;
     wire [`BANK_LINE_WIDTH-1:0] data_write;
 
     if (WORD_SIZE == BANK_LINE_SIZE) begin
 
-        wire should_write = ((mem_write_st1e != `BYTE_EN_NO)) 
+        assign should_write = ((mem_write_st1e != `BYTE_EN_NO)) 
                            && valid_req_st1e 
                            && use_read_valid_st1e 
                            && !miss_st1e 
@@ -148,7 +149,7 @@ module VX_tag_data_access #(
             assign we[i] = (force_write || (should_write && !real_writefill)) ? 4'b1111 : 4'b0000;
         end    
 
-        assign readword_st1e = read_data_st1c[STAGE_1_CYCLES-1];
+        assign readword_st1e = use_read_data_st1e;
         assign data_write = force_write ? writedata_st1e : writeword_st1e;
 
     end else begin
@@ -174,13 +175,13 @@ module VX_tag_data_access #(
         wire [3:0] sb_mask = (b0 ? 4'b0001 : (b1 ? 4'b0010 : (b2 ? 4'b0100 : 4'b1000)));
         wire [3:0] sh_mask = (b0 ? 4'b0011 : 4'b1100);
 
-        wire should_write = (sw || sb || sh) 
+        assign should_write = (sw || sb || sh) 
                            && valid_req_st1e 
                            && use_read_valid_st1e 
                            && !miss_st1e 
                            && !is_snp_st1e;
 
-        wire[`WORD_WIDTH-1:0] data_unmod  = read_data_st1c[STAGE_1_CYCLES-1][block_offset * 32 +: 32];
+        wire[`WORD_WIDTH-1:0] data_unmod  = use_read_data_st1e[block_offset * 32 +: 32];
         wire[`WORD_WIDTH-1:0] data_unQual = (b0 || lw) ? (data_unmod)      :
                                                     b1 ? (data_unmod >> 8)  :
                                                     b2 ? (data_unmod >> 16) :
@@ -200,7 +201,7 @@ module VX_tag_data_access #(
         assign readword_st1e = data_Qual;
 
         for (i = 0; i < `BANK_LINE_WORDS; i++) begin
-            wire normal_write = (block_offset == i[`WORD_SELECT_BITS-1:0]) && should_write && !real_writefill;
+            wire normal_write = (block_offset == `WORD_SELECT_BITS'(i)) && should_write && !real_writefill;
 
             assign we[i] = (force_write)        ? 4'b1111 : 
                            (normal_write && sw) ? 4'b1111 :
@@ -226,7 +227,7 @@ module VX_tag_data_access #(
     assign use_write_data   = data_write;
 
     // use "case equality" to handle uninitialized tag when block entry is not valid
-    assign tags_match = ((writeaddr_st1e[`TAG_LINE_ADDR_RNG] == use_read_tag_st1e) === 1'b1);
+    assign tags_match = ((writetag_st1e == use_read_tag_st1e) === 1'b1);
 
     wire snoop_hit   = valid_req_st1e &&  is_snp_st1e && use_read_valid_st1e && tags_match && use_read_dirty_st1e;
     wire req_invalid = valid_req_st1e && !is_snp_st1e && !use_read_valid_st1e && !writefill_st1e;

hw/simulate/simulator.cpp

@@ -2,10 +2,10 @@
 #include <iostream>
 #include <iomanip>
 
-uint64_t time_stamp = 0;
+uint64_t timestamp = 0;
 
 double sc_time_stamp() { 
-  return time_stamp;
+  return timestamp;
 }
 
 Simulator::Simulator() {    
@@ -28,18 +28,23 @@ Simulator::~Simulator() {
 }
 
 void Simulator::attach_ram(RAM* ram) {
+#ifndef NDEBUG
+  std::cout << timestamp << ": [sim] attach_ram" << std::endl;
+#endif
   ram_ = ram;
   dram_rsp_vec_.clear();
 }
 
 void Simulator::print_stats(std::ostream& out) {
   out << std::left;
-  out << std::setw(24) << "# of total cycles:" << std::dec << time_stamp/2 << std::endl;
+  out << std::setw(24) << "# of total cycles:" << std::dec << timestamp/2 << std::endl;
 }
 
 void Simulator::dbus_driver() {
-  if (ram_ == nullptr)
+  if (ram_ == nullptr) {
+    vortex_->dram_req_ready = false;
     return;
+  }
 
   // handle DRAM response cycle
   int dequeue_index = -1;
@@ -70,7 +75,7 @@ void Simulator::dbus_driver() {
   // handle DRAM stalls
   bool dram_stalled = false;
 #ifdef ENABLE_DRAM_STALLS
-  if (0 == ((time_stamp/2) % DRAM_STALLS_MODULO)) { 
+  if (0 == ((timestamp/2) % DRAM_STALLS_MODULO)) { 
     dram_stalled = true;
   } else
   if (dram_rsp_vec_.size() >= DRAM_RQ_SIZE) {
@@ -114,12 +119,15 @@ void Simulator::io_driver() {
    && vortex_->io_req_addr == IO_BUS_ADDR_COUT) {
     uint32_t data_write = (uint32_t)vortex_->io_req_data;
     char c = (char)data_write;
-    std::cerr << c;      
+    std::cout << c;      
   }
   vortex_->io_req_ready = true;
 }
 
 void Simulator::reset() {     
+#ifndef NDEBUG
+  std::cout << timestamp << ": [sim] reset()" << std::endl;
+#endif 
   vortex_->reset = 1;
   this->step();  
   vortex_->reset = 0;
@@ -141,9 +149,9 @@ void Simulator::step() {
 void Simulator::eval() {
   vortex_->eval();
 #ifdef VCD_OUTPUT
-  trace_->dump(time_stamp);
+  trace_->dump(timestamp);
 #endif
-  ++time_stamp;
+  ++timestamp;
 }
 
 void Simulator::wait(uint32_t cycles) {
@@ -157,6 +165,9 @@ bool Simulator::is_busy() {
 }
 
 void Simulator::flush_caches(uint32_t mem_addr, uint32_t size) {  
+#ifndef NDEBUG
+  std::cout << timestamp << ": [sim] flush_caches()" << std::endl;
+#endif
   // align address to LLC block boundaries
   auto aligned_addr_start = mem_addr / GLOBAL_BLOCK_SIZE;
   auto aligned_addr_end = (mem_addr + size + GLOBAL_BLOCK_SIZE - 1) / GLOBAL_BLOCK_SIZE;
@@ -186,6 +197,10 @@ void Simulator::flush_caches(uint32_t mem_addr, uint32_t size) {
 }
 
 bool Simulator::run() {
+#ifndef NDEBUG
+  std::cout << timestamp << ": [sim] run()" << std::endl;
+#endif 
+
   // reset the device
   this->reset();
 

hw/simulate/simulator.h

@@ -52,6 +52,7 @@ private:
 
   RAM *ram_;
   VVortex_Socket *vortex_;
+  bool enable_;
 #ifdef VCD_OUTPUT
   VerilatedVcdC *trace_;
 #endif