vcs, palldium: support warmup, set difftest_perfCtrl_clean by DPI-C (#…

…506)

Previous for vcs/palladium, we use fixed default value for perfCtrl signal, which disable performance log report during simulation.

This change support passing warmup-instr from vcs to difftest, and difftest will use dpi-c to set perfCtrl_clean when commitIntrs greater than threshold, reseting perf counters after warm-up. 

After this change, user can get perf-log printed for vcs/palladium by setting log_begin/end (use +b/+e) correctly.

---------

Co-authored-by: klin02 <[email protected]>

.github/workflows/main.yml

@@ -450,3 +450,13 @@ jobs:
             make difftest_verilog PROFILE=../build/generated-src/difftest_profile.json NUMCORES=1 CONFIG=ZEL MFC=1
             make simv VCS=verilator WITH_CHISELDB=0 WITH_CONSTANTIN=0 IOTRACE_ZSTD=1
             ./build/simv +workload=../ready-to-run/microbench.bin +e=0 +diff=../ready-to-run/riscv64-nemu-interpreter-so +iotrace-name=../iotrace
+
+      - name: Verilator Build with VCS Top (with LogPerf and WarmUp)
+        run : |
+            cd $GITHUB_WORKSPACE/../xs-env
+            source ./env.sh
+            cd $GITHUB_WORKSPACE/../xs-env/NutShell
+            source ./env.sh
+            make clean
+            make simv VCS=verilator -j2
+            ./build/simv +workload=./ready-to-run/microbench.bin +b=0 +e=-1 +diff=./ready-to-run/riscv64-nemu-interpreter-so +max-instrs=5000 +warmup_instr=1000

src/test/csrc/vcs/vcs_main.cpp

@@ -26,11 +26,9 @@
 #ifndef CONFIG_NO_DIFFTEST
 #include "refproxy.h"
 #endif // CONFIG_NO_DIFFTEST
+#include "svdpi.h"
 #include <common.h>
 #include <locale.h>
-#ifdef CONFIG_DIFFTEST_DEFERRED_RESULT
-#include "svdpi.h"
-#endif // CONFIG_DIFFTEST_DEFERRED_RESULT
 #ifdef CONFIG_DIFFTEST_PERFCNT
 #include "perf.h"
 #endif // CONFIG_DIFFTEST_PERFCNT
@@ -44,6 +42,8 @@ static bool enable_difftest = true;
 static uint64_t max_instrs = 0;
 static char *workload_list = NULL;
 static uint32_t overwrite_nbytes = 0xe00;
+static uint64_t warmup_instr = 0;
+static svScope difftest_endpoint_scope;
 struct core_end_info_t {
   bool core_trap[NUM_CORES];
   double core_cpi[NUM_CORES];
@@ -84,6 +84,12 @@ extern "C" void set_overwrite_autoset() {
   fclose(fp);
 }
 
+// Support workload warms up and clean LogPerf after warmup instrs
+extern "C" void set_warmup_instr(uint64_t instrs) {
+  warmup_instr = instrs;
+  printf("Warmup instrs:%ld\n", instrs);
+}
+
 extern "C" void set_gcpt_bin(char *s) {
   gcpt_restore_bin = (char *)malloc(256);
   strcpy(gcpt_restore_bin, s);
@@ -94,6 +100,10 @@ extern "C" void set_max_instrs(uint64_t mc) {
   max_instrs = mc;
 }
 
+extern "C" void set_difftest_endpoint_scope() {
+  difftest_endpoint_scope = svGetScope();
+}
+
 extern "C" uint64_t get_stuck_limit() {
 #ifdef CONFIG_NO_DIFFTEST
   return 0;
@@ -161,6 +171,17 @@ extern "C" void set_simjtag() {
   enable_simjtag = true;
 }
 
+extern "C" void set_perfCtrl_clean();
+
+void difftest_perfCtrl_clean() {
+  if (difftest_endpoint_scope == NULL) {
+    printf("Error: Could not retrieve DifftestEndpoint scope, set first\n");
+    assert(difftest_endpoint_scope);
+  }
+  svSetScope(difftest_endpoint_scope);
+  set_perfCtrl_clean();
+}
+
 extern "C" uint8_t simv_init() {
   if (workload_list != NULL) {
     if (switch_workload())
@@ -239,11 +260,17 @@ extern "C" uint8_t simv_step() {
       return SIMV_FAIL;
   }
 
-  if (max_instrs != 0) { // 0 for no limit
-    for (int i = 0; i < NUM_CORES; i++) {
+  for (int i = 0; i < NUM_CORES; i++) {
+    auto trap = difftest[i]->get_trap_event();
+    if (warmup_instr != 0 && trap->instrCnt > warmup_instr) {
+      difftest_perfCtrl_clean();
+      warmup_instr = 0;
+      Info("Warmup finished. The performance counters will be reset.\n");
+    }
+
+    if (max_instrs != 0) { // 0 for no limit
       if (core_end_info.core_trap[i])
         continue;
-      auto trap = difftest[i]->get_trap_event();
       if (max_instrs < trap->instrCnt) {
         core_end_info.core_trap[i] = true;
         core_end_info.core_trap_num++;

src/test/vsrc/vcs/DifftestEndpoint.v

@@ -50,6 +50,8 @@ import "DPI-C" function void set_max_instrs(longint mc);
 import "DPI-C" function longint get_stuck_limit();
 import "DPI-C" function void set_overwrite_nbytes(longint len);
 import "DPI-C" function void set_overwrite_autoset();
+import "DPI-C" function void set_warmup_instr(longint instrs);
+import "DPI-C" context function void set_difftest_endpoint_scope();
 `ifdef WITH_DRAMSIM3
 import "DPI-C" function void simv_tick();
 `endif // WITH_DRAMSIM3
@@ -81,6 +83,7 @@ longint overwrite_nbytes;
 
 reg [63:0] max_instrs;
 reg [63:0] max_cycles;
+reg [63:0] warmup_instr;
 reg [63:0] stuck_limit;
 
 initial begin
@@ -100,12 +103,19 @@ initial begin
   end
   stuck_limit = 0;
 `ifndef TB_NO_DPIC
+  set_difftest_endpoint_scope();
   stuck_limit = get_stuck_limit();
   // workload: bin file
   if ($test$plusargs("workload")) begin
     $value$plusargs("workload=%s", bin_file);
     set_bin_file(bin_file);
   end
+  // warmup for instrs
+  warmup_instr = 0;
+  if ($test$plusargs("warmup_instr")) begin
+    $value$plusargs("warmup_instr=%d", warmup_instr);
+    set_warmup_instr(warmup_instr);
+  end
   // boot flash image: bin file
   if ($test$plusargs("flash")) begin
     $value$plusargs("flash=%s", flash_bin_file);
@@ -170,12 +180,27 @@ initial begin
   end
 end
 
+reg difftest_perfCtrl_clean_r;
 assign difftest_logCtrl_begin = difftest_logCtrl_begin_r;
 assign difftest_logCtrl_end = difftest_logCtrl_end_r;
 assign difftest_logCtrl_level = 0;
-assign difftest_perfCtrl_clean = 0;
+assign difftest_perfCtrl_clean = difftest_perfCtrl_clean_r;
 assign difftest_uart_in_ch = 8'hff;
 
+export "DPI-C" task set_perfCtrl_clean;
+task set_perfCtrl_clean();
+  difftest_perfCtrl_clean_r <= 1'b1;
+endtask
+
+always @(posedge clock) begin
+  if (reset) begin
+    difftest_perfCtrl_clean_r <= 1'b0;
+  end
+  else if (difftest_perfCtrl_clean_r) begin
+    difftest_perfCtrl_clean_r <= 1'b0;
+  end
+end
+
 always @(posedge clock) begin
   if (!reset && difftest_uart_out_valid) begin
     $fwrite(32'h8000_0001, "%c", difftest_uart_out_ch);