diff --git a/.github/workflows/main.yml b/.github/workflows/main.yml
index c9fcb6cc8..b5b15ff89 100644
--- a/.github/workflows/main.yml
+++ b/.github/workflows/main.yml
@@ -57,6 +57,7 @@ jobs:
             make clean
             make emu NO_DIFF=1 EMU_TRACE=1
             ./build/emu -b 0 -e 100 -i ./ready-to-run/microbench.bin --dump-wave -C 10000
+            ./build/emu -b 0 -e 100 -i ./ready-to-run/microbench.bin --dump-wave-full -C 10000
 
       - name: Simulation with No Diff
         run: |
@@ -95,6 +96,7 @@ jobs:
             make clean
             make emu EMU_TRACE=1
             ./build/emu -b 10 -e 12 -i ./ready-to-run/microbench.bin --dump-wave --diff ./ready-to-run/riscv64-nemu-interpreter-so
+            ./build/emu -b 10 -e 12 -i ./ready-to-run/microbench.bin --dump-wave-full --diff ./ready-to-run/riscv64-nemu-interpreter-so
 
       - name: Difftest with two threads
         run: |
diff --git a/src/test/csrc/verilator/emu.cpp b/src/test/csrc/verilator/emu.cpp
index 089912b97..51b4b0c96 100644
--- a/src/test/csrc/verilator/emu.cpp
+++ b/src/test/csrc/verilator/emu.cpp
@@ -87,6 +87,7 @@ static inline void print_help(const char *file) {
   printf("      --load-snapshot=PATH   load snapshot from PATH\n");
   printf("      --no-snapshot          disable saving snapshots\n");
   printf("      --dump-wave            dump waveform when log is enabled\n");
+  printf("      --dump-wave-full       dump full waveform when log is enabled\n");
   printf("      --dump-ref-trace       dump REF trace when log is enabled\n");
   printf("      --dump-commit-trace    dump commit trace when log is enabled\n");
 #ifdef ENABLE_CHISEL_DB
@@ -143,6 +144,7 @@ inline EmuArgs parse_args(int argc, const char *argv[]) {
     { "dump-footprints",   1, NULL,  0  },
     { "as-footprints",     0, NULL,  0  },
     { "dump-linearized",   1, NULL,  0  },
+    { "dump-wave-full",    0, NULL,  0  },
     { "seed",              1, NULL, 's' },
     { "max-cycles",        1, NULL, 'C' },
     { "fork-interval",     1, NULL, 'X' },
@@ -213,6 +215,7 @@ inline EmuArgs parse_args(int argc, const char *argv[]) {
           case 19: args.footprints_name = optarg; continue;
           case 20: args.image_as_footprints = true; continue;
           case 21: args.linearized_name = optarg; continue;
+          case 22: args.enable_waveform = true; args.enable_waveform_full = true; continue;
         }
         // fall through
       default:
@@ -264,6 +267,8 @@ inline EmuArgs parse_args(int argc, const char *argv[]) {
     exit(0);
   }
 
+  args.enable_waveform = args.enable_waveform && !args.enable_fork;
+
 #ifdef ENABLE_IPC
   char *ipc_image = (char *)malloc(255);
   char *ipc_file = (char *)malloc(255);
@@ -317,6 +322,25 @@ Emulator::Emulator(int argc, const char *argv[]):
   // init flash
   init_flash(args.flash_bin);
 
+#if VM_TRACE == 1
+  if (args.enable_waveform) {
+    Verilated::traceEverOn(true);	// Verilator must compute traced signals
+#ifdef ENABLE_FST
+    tfp = new VerilatedFstC;
+#else
+    tfp = new VerilatedVcdC;
+#endif
+    dut_ptr->trace(tfp, 99);	// Trace 99 levels of hierarchy
+    if (args.wave_path != NULL) {
+      tfp->open(args.wave_path);
+    }
+    else {
+      time_t now = time(NULL);
+      tfp->open(waveform_filename(now));	// Open the dump file
+    }
+  }
+#endif
+
   // init core
   reset_ncycles(10);
 
@@ -348,26 +372,6 @@ Emulator::Emulator(int argc, const char *argv[]):
   init_db(args.dump_db, (args.select_db != NULL), args.select_db);
 #endif
 
-#if VM_TRACE == 1
-  enable_waveform = args.enable_waveform && !args.enable_fork;
-  if (enable_waveform) {
-    Verilated::traceEverOn(true);	// Verilator must compute traced signals
-#ifdef ENABLE_FST
-    tfp = new VerilatedFstC;
-#else
-    tfp = new VerilatedVcdC;
-#endif
-    dut_ptr->trace(tfp, 99);	// Trace 99 levels of hierarchy
-    if (args.wave_path != NULL) {
-      tfp->open(args.wave_path);
-    }
-    else {
-      time_t now = time(NULL);
-      tfp->open(waveform_filename(now));	// Open the dump file
-    }
-  }
-#endif
-
 #ifdef VM_SAVABLE
   snapshot_slot = new VerilatedSaveMem[2];
   if (args.snapshot_path != NULL) {
@@ -440,7 +444,7 @@ Emulator::Emulator(int argc, const char *argv[]):
 Emulator::~Emulator() {
   // Simulation ends here, do clean up & display jobs
 #if VM_TRACE == 1
-  if (enable_waveform) tfp->close();
+  if (args.enable_waveform) tfp->close();
 #endif
 
 #if VM_COVERAGE == 1
@@ -528,16 +532,30 @@ inline void Emulator::reset_ncycles(size_t cycles) {
 #ifdef COVERAGE_PORT_RESET
     dut_ptr->coverage_reset = dut_ptr->reset;
 #endif // COVERAGE_PORT_RESET
-    dut_ptr->clock = 0;
+    dut_ptr->clock = 1;
 #ifdef COVERAGE_PORT_CLOCK
     dut_ptr->coverage_clock = dut_ptr->clock;
 #endif // COVERAGE_PORT_CLOCK
     dut_ptr->eval();
-    dut_ptr->clock = 1;
+
+#if VM_TRACE == 1
+  if (args.enable_waveform && args.enable_waveform_full && args.log_begin == 0) {
+    tfp->dump(2 * i);
+  }
+#endif
+
+    dut_ptr->clock = 0;
 #ifdef COVERAGE_PORT_CLOCK
     dut_ptr->coverage_clock = dut_ptr->clock;
 #endif // COVERAGE_PORT_CLOCK
     dut_ptr->eval();
+
+#if VM_TRACE == 1
+  if (args.enable_waveform && args.enable_waveform_full && args.log_begin == 0) {
+    tfp->dump(2 * i + 1);
+  }
+#endif
+
     dut_ptr->reset = 0;
 #ifdef COVERAGE_PORT_RESET
     dut_ptr->coverage_reset = dut_ptr->reset;
@@ -550,35 +568,32 @@ inline void Emulator::single_cycle() {
     goto end_single_cycle;
   }
 
-  dut_ptr->clock = 0;
+  dut_ptr->clock = 1;
 #ifdef COVERAGE_PORT_CLOCK
   dut_ptr->coverage_clock = dut_ptr->clock;
 #endif // COVERAGE_PORT_CLOCK
   dut_ptr->eval();
 
 #if VM_TRACE == 1
-  if (enable_waveform) {
+  if (args.enable_waveform) {
 #ifndef CONFIG_NO_DIFFTEST
-    auto trap = difftest[0]->get_trap_event();
-    uint64_t cycle = trap->cycleCnt;
-    uint64_t begin = dut_ptr->io_logCtrl_log_begin;
-    uint64_t end   = dut_ptr->io_logCtrl_log_end;
-    bool in_range  = (begin <= cycle) && (cycle <= end);
+    uint64_t cycle = difftest[0]->get_trap_event()->cycleCnt;
 #else
-    bool in_range = true;
-    static uint64_t cycle = 0;
+    static uint64_t cycle = -1UL;
     cycle++;
 #endif
-    if (in_range || force_dump_wave) { tfp->dump(cycle); }
+    bool in_range  = (args.log_begin <= cycle) && (cycle <= args.log_end);
+    if (in_range || force_dump_wave) {
+      if (args.enable_waveform_full) {
+        tfp->dump(20 + 2*cycle);
+      }
+      else {
+        tfp->dump(cycle);
+      }
+    }
   }
 #endif
 
-  dut_ptr->clock = 1;
-#ifdef COVERAGE_PORT_CLOCK
-  dut_ptr->coverage_clock = dut_ptr->clock;
-#endif // COVERAGE_PORT_CLOCK
-  dut_ptr->eval();
-
 #ifdef WITH_DRAMSIM3
   dramsim3_step();
 #endif
@@ -592,6 +607,27 @@ inline void Emulator::single_cycle() {
     dut_ptr->io_uart_in_ch = uart_getc();
   }
 
+  dut_ptr->clock = 0;
+#ifdef COVERAGE_PORT_CLOCK
+  dut_ptr->coverage_clock = dut_ptr->clock;
+#endif // COVERAGE_PORT_CLOCK
+  dut_ptr->eval();
+
+#if VM_TRACE == 1
+  if (args.enable_waveform && args.enable_waveform_full) {
+#ifndef CONFIG_NO_DIFFTEST
+    uint64_t cycle = difftest[0]->get_trap_event()->cycleCnt;
+#else
+    static uint64_t cycle = -1UL;
+    cycle++;
+#endif
+    bool in_range  = (args.log_begin <= cycle) && (cycle <= args.log_end);
+    if (in_range || force_dump_wave) {
+      tfp->dump(21 + 2*cycle);
+    }
+  }
+#endif
+
 end_single_cycle:
 #ifndef CONFIG_NO_DIFFTEST
   if (args.trace_name) {
@@ -1068,7 +1104,7 @@ void Emulator::fork_child_init() {
   tfp->open(cycle_wavefile(cycles, now));
   // override output range config, force dump wave
   force_dump_wave = true;
-  enable_waveform = true;
+  args.enable_waveform = true;
 #endif
 #ifndef CONFIG_NO_DIFFTEST
 #ifdef ENABLE_SIMULATOR_DEBUG_INFO
diff --git a/src/test/csrc/verilator/emu.h b/src/test/csrc/verilator/emu.h
index c513c0f6a..a28912a98 100644
--- a/src/test/csrc/verilator/emu.h
+++ b/src/test/csrc/verilator/emu.h
@@ -64,6 +64,7 @@ struct EmuArgs {
   const char *footprints_name = nullptr;
   const char *linearized_name = nullptr;
   bool enable_waveform = false;
+  bool enable_waveform_full = false;
   bool enable_ref_trace = false;
   bool enable_commit_trace = false;
   bool enable_snapshot = true;
@@ -98,7 +99,6 @@ class Emulator final : public DUT {
 #else
   VerilatedVcdC* tfp;
 #endif
-  bool enable_waveform;
   bool force_dump_wave = false;
 #ifdef VM_SAVABLE
   VerilatedSaveMem *snapshot_slot = nullptr;