diff --git a/code/.gitignore b/code/.gitignore
index 2cc89621..72da920f 100644
--- a/code/.gitignore
+++ b/code/.gitignore
@@ -6,3 +6,4 @@
 !*.sh
 !*.py
 !*.html
+!*.csv
diff --git a/code/Makefile b/code/Makefile
index 925d0493..39cb42ef 100644
--- a/code/Makefile
+++ b/code/Makefile
@@ -17,4 +17,10 @@ clean:
 %: %.cpp %.h common.h
 	$(CXX) $(CFLAGS) $< -mlsx -mlasx -o $@
 
+measure: measure.cpp measure.h
+	$(CXX) -O2 $< -o $@
+
+run-measure: measure
+	./measure -p
+
 .SUFFIXES:
diff --git a/code/measure.cpp b/code/measure.cpp
new file mode 100644
index 00000000..d59675a2
--- /dev/null
+++ b/code/measure.cpp
@@ -0,0 +1,264 @@
+#include <assert.h>
+#include <linux/perf_event.h>
+#include <map>
+#include <sched.h>
+#include <set>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <string>
+#include <sys/syscall.h>
+#include <sys/time.h>
+#include <unistd.h>
+#include <utility>
+#include <vector>
+
+// learned from lmbench lat_mem_rd
+#define FIVE(X) X X X X X
+#define TEN(X) FIVE(X) FIVE(X)
+#define FIFTY(X) TEN(X) TEN(X) TEN(X) TEN(X) TEN(X)
+#define HUNDRED(X) FIFTY(X) FIFTY(X)
+#define THOUSAND(X) HUNDRED(TEN(X))
+
+// get time or cycles
+// unit: ns or cycle
+uint64_t get_time();
+// prioritize cycle over time
+void setup_time_or_cycles();
+uint64_t get_time_or_cycles();
+
+void bind_to_core();
+
+// perf related
+void setup_perf_cycles();
+uint64_t perf_read_cycles();
+void setup_perf_instructions();
+uint64_t perf_read_instructions();
+void setup_perf_llc_misses();
+uint64_t perf_read_llc_misses();
+void setup_perf_llc_loads();
+uint64_t perf_read_llc_loads();
+
+int N = 100000;
+
+#define INSTR_TEST(NAME, INST, ...)                                            \
+  void test_##NAME(int n) {                                                    \
+    for (int i = 0; i < n; i++) {                                              \
+      asm volatile(".align 4\n" THOUSAND(INST) : : : __VA_ARGS__);             \
+    }                                                                          \
+  }
+
+#include "measure.h"
+
+#undef INSTR_TEST
+
+struct InstrTest {
+  const char *name;
+  const char *inst;
+  void (*test)(int);
+};
+
+#define INSTR_TEST(NAME, INST, ...)                                            \
+  InstrTest{.name = #NAME, .inst = #INST, .test = test_##NAME},
+
+std::vector<InstrTest> tests = {
+#include "measure.h"
+};
+
+#undef INSTR_TEST
+
+struct InstrInfo {
+  std::set<double> latency;
+  double throughput;
+};
+
+uint64_t get_time() {
+  struct timeval tv = {};
+  gettimeofday(&tv, nullptr);
+  return (uint64_t)tv.tv_sec * 1000000000 + (uint64_t)tv.tv_usec * 1000;
+}
+
+uint64_t perf_read_common(int fd) {
+  uint64_t counter;
+  int res = read(fd, &counter, sizeof(counter));
+  assert(res == sizeof(counter));
+  return counter;
+}
+
+int setup_perf_common(uint32_t type, uint64_t config) {
+  struct perf_event_attr *attr =
+      (struct perf_event_attr *)malloc(sizeof(struct perf_event_attr));
+  memset(attr, 0, sizeof(struct perf_event_attr));
+  attr->type = type;
+  attr->size = sizeof(struct perf_event_attr);
+  attr->config = config;
+  attr->disabled = 0;
+  attr->pinned = 1;
+  attr->inherit = 1;
+  attr->exclude_kernel = 1;
+  int fd = syscall(SYS_perf_event_open, attr, 0, -1, -1, 0);
+  if (fd < 0) {
+    perror("perf_event_open");
+    fprintf(stderr, "try: sudo sysctl kernel.perf_event_paranoid=2");
+    exit(1);
+  }
+  return fd;
+}
+
+int perf_fd_cycles = -1;
+
+uint64_t perf_read_cycles() { return perf_read_common(perf_fd_cycles); }
+
+void setup_perf_cycles() {
+  perf_fd_cycles =
+      setup_perf_common(PERF_TYPE_HARDWARE, PERF_COUNT_HW_CPU_CYCLES);
+  if (perf_fd_cycles >= 0) {
+    printf("Using PMU to count cycles\n");
+  }
+}
+
+int perf_fd_instructions = -1;
+
+uint64_t perf_read_instructions() {
+  return perf_read_common(perf_fd_instructions);
+}
+
+void setup_perf_instructions() {
+  perf_fd_instructions =
+      setup_perf_common(PERF_TYPE_HARDWARE, PERF_COUNT_HW_INSTRUCTIONS);
+}
+
+int perf_fd_llc_misses = -1;
+int perf_fd_llc_loads = -1;
+
+uint64_t perf_read_llc_misses() { return perf_read_common(perf_fd_llc_misses); }
+
+uint64_t perf_read_llc_loads() { return perf_read_common(perf_fd_llc_loads); }
+
+void setup_perf_llc_misses() {
+  perf_fd_llc_misses = setup_perf_common(
+      PERF_TYPE_HW_CACHE, (PERF_COUNT_HW_CACHE_LL) |
+                              (PERF_COUNT_HW_CACHE_OP_READ << 8) |
+                              (PERF_COUNT_HW_CACHE_RESULT_MISS << 16));
+}
+
+void setup_perf_llc_loads() {
+  perf_fd_llc_loads = setup_perf_common(
+      PERF_TYPE_HW_CACHE, (PERF_COUNT_HW_CACHE_LL) |
+                              (PERF_COUNT_HW_CACHE_OP_READ << 8) |
+                              (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16));
+}
+
+void setup_time_or_cycles() { setup_perf_cycles(); }
+
+uint64_t get_time_or_cycles() {
+  if (perf_fd_cycles >= 0) {
+    // cycle
+    return perf_read_cycles();
+  } else {
+    // time
+    return get_time();
+  }
+}
+
+// bind to core 0
+void bind_to_core() {
+  cpu_set_t set;
+  CPU_ZERO(&set);
+  CPU_SET(0, &set);
+  int res = sched_setaffinity(0, sizeof(set), &set);
+  if (res == 0) {
+    printf("Pinned to cpu 0\n");
+  }
+}
+
+int main(int argc, char *argv[]) {
+  bool perf = false;
+
+  int opt;
+  while ((opt = getopt(argc, argv, "n:p")) != -1) {
+    switch (opt) {
+    case 'n':
+      sscanf(optarg, "%d", &N);
+      break;
+    case 'p':
+      perf = true;
+      break;
+    default:
+      fprintf(stderr, "Usage: %s [-p]\n", argv[0]);
+      exit(EXIT_FAILURE);
+    }
+  }
+
+  bind_to_core();
+  if (perf) {
+    setup_time_or_cycles();
+  }
+
+  // calibrate unit cycle time
+  uint64_t begin = get_time_or_cycles();
+  test_unit(N);
+  uint64_t unit_elapsed = get_time_or_cycles() - begin;
+
+  std::map<std::string, InstrInfo> info;
+
+  for (auto it : tests) {
+    std::string name = it.name;
+    if (name == "unit") {
+      continue;
+    }
+
+    begin = get_time_or_cycles();
+    it.test(N);
+    uint64_t elapsed = get_time_or_cycles() - begin;
+    double cycles = (double)elapsed / unit_elapsed;
+
+    std::string base_name;
+    size_t tp_index = name.find("_tp");
+    if (tp_index != std::string::npos) {
+      base_name = name.substr(0, tp_index);
+      printf("%s: throughput 1/%.2lf=%.2lf instructions\n", it.name, cycles,
+             1.0 / cycles);
+      info[base_name].throughput = 1.0 / cycles;
+    } else {
+      base_name = name;
+
+      size_t last_underscore_index = name.find_last_of("_");
+      if (last_underscore_index != std::string::npos) {
+        char *end = NULL;
+        const char *start = &name.c_str()[last_underscore_index + 1];
+        strtol(start, &end, 10);
+
+        if (start != end) {
+          // strip suffix _2 etc
+          base_name = name.substr(0, last_underscore_index);
+        }
+      }
+
+      // round to 0.01
+      cycles = (double)(long)(cycles * 100 + 0.5) / 100.0;
+      printf("%s: latency %.2lf cycles\n", it.name, cycles);
+      info[base_name].latency.insert(cycles);
+    }
+  }
+
+  FILE *fp = fopen("measure.csv", "w");
+  assert(fp);
+  fprintf(fp, "name,latency,throughput(cpi)\n");
+  for (auto pair : info) {
+    std::string latency;
+    auto entry = pair.second;
+    for (auto lat : entry.latency) {
+      char buffer[32];
+      std::sprintf(buffer, "%.2lf", lat);
+      if (!latency.empty()) {
+        latency += "/";
+      }
+      latency += buffer;
+    }
+    fprintf(fp, "%s,%s,%.2lf\n", pair.first.c_str(), latency.c_str(),
+            entry.throughput);
+  }
+  printf("Result written to measure.csv\n");
+}
diff --git a/code/measure.csv b/code/measure.csv
new file mode 100644
index 00000000..02dc5330
--- /dev/null
+++ b/code/measure.csv
@@ -0,0 +1,9 @@
+name,latency,throughput(cpi)
+vfmadd_s,5.00,2.00
+vfmsub_s,5.00,2.00
+vfnmadd_s,5.00,2.00
+vfnmsub_s,5.00,2.00
+xvfmadd_s,5.00,2.00
+xvfmsub_s,5.00,2.00
+xvfnmadd_s,5.00,2.00
+xvfnmsub_s,5.00,2.00
diff --git a/code/measure.h b/code/measure.h
new file mode 100644
index 00000000..35d2f61f
--- /dev/null
+++ b/code/measure.h
@@ -0,0 +1,33 @@
+INSTR_TEST(unit, "add.w $r12, $r1, $r12\n", "r12")
+INSTR_TEST(vfmadd_s_1, "vfmadd.s $vr0, $vr0, $vr2, $vr3\n")
+INSTR_TEST(vfmadd_s_2, "vfmadd.s $vr0, $vr1, $vr0, $vr3\n")
+INSTR_TEST(vfmadd_s_3, "vfmadd.s $vr0, $vr1, $vr2, $vr0\n")
+INSTR_TEST(vfmadd_s_tp, "vfmadd.s $vr0, $vr1, $vr2, $vr3\n")
+INSTR_TEST(vfmsub_s_1, "vfmsub.s $vr0, $vr0, $vr2, $vr3\n")
+INSTR_TEST(vfmsub_s_2, "vfmsub.s $vr0, $vr1, $vr0, $vr3\n")
+INSTR_TEST(vfmsub_s_3, "vfmsub.s $vr0, $vr1, $vr2, $vr0\n")
+INSTR_TEST(vfmsub_s_tp, "vfmsub.s $vr0, $vr1, $vr2, $vr3\n")
+INSTR_TEST(vfnmadd_s_1, "vfnmadd.s $vr0, $vr0, $vr2, $vr3\n")
+INSTR_TEST(vfnmadd_s_2, "vfnmadd.s $vr0, $vr1, $vr0, $vr3\n")
+INSTR_TEST(vfnmadd_s_3, "vfnmadd.s $vr0, $vr1, $vr2, $vr0\n")
+INSTR_TEST(vfnmadd_s_tp, "vfnmadd.s $vr0, $vr1, $vr2, $vr3\n")
+INSTR_TEST(vfnmsub_s_1, "vfnmsub.s $vr0, $vr0, $vr2, $vr3\n")
+INSTR_TEST(vfnmsub_s_2, "vfnmsub.s $vr0, $vr1, $vr0, $vr3\n")
+INSTR_TEST(vfnmsub_s_3, "vfnmsub.s $vr0, $vr1, $vr2, $vr0\n")
+INSTR_TEST(vfnmsub_s_tp, "vfnmsub.s $vr0, $vr1, $vr2, $vr3\n")
+INSTR_TEST(xvfmadd_s_1, "xvfmadd.s $xr0, $xr0, $xr2, $xr3\n")
+INSTR_TEST(xvfmadd_s_2, "xvfmadd.s $xr0, $xr1, $xr0, $xr3\n")
+INSTR_TEST(xvfmadd_s_3, "xvfmadd.s $xr0, $xr1, $xr2, $xr0\n")
+INSTR_TEST(xvfmadd_s_tp, "xvfmadd.s $xr0, $xr1, $xr2, $xr3\n")
+INSTR_TEST(xvfmsub_s_1, "xvfmsub.s $xr0, $xr0, $xr2, $xr3\n")
+INSTR_TEST(xvfmsub_s_2, "xvfmsub.s $xr0, $xr1, $xr0, $xr3\n")
+INSTR_TEST(xvfmsub_s_3, "xvfmsub.s $xr0, $xr1, $xr2, $xr0\n")
+INSTR_TEST(xvfmsub_s_tp, "xvfmsub.s $xr0, $xr1, $xr2, $xr3\n")
+INSTR_TEST(xvfnmadd_s_1, "xvfnmadd.s $xr0, $xr0, $xr2, $xr3\n")
+INSTR_TEST(xvfnmadd_s_2, "xvfnmadd.s $xr0, $xr1, $xr0, $xr3\n")
+INSTR_TEST(xvfnmadd_s_3, "xvfnmadd.s $xr0, $xr1, $xr2, $xr0\n")
+INSTR_TEST(xvfnmadd_s_tp, "xvfnmadd.s $xr0, $xr1, $xr2, $xr3\n")
+INSTR_TEST(xvfnmsub_s_1, "xvfnmsub.s $xr0, $xr0, $xr2, $xr3\n")
+INSTR_TEST(xvfnmsub_s_2, "xvfnmsub.s $xr0, $xr1, $xr0, $xr3\n")
+INSTR_TEST(xvfnmsub_s_3, "xvfnmsub.s $xr0, $xr1, $xr2, $xr0\n")
+INSTR_TEST(xvfnmsub_s_tp, "xvfnmsub.s $xr0, $xr1, $xr2, $xr3\n")
diff --git a/code/vsetallnez_b.cpp b/code/vsetallnez_b.cpp
index 3a4b5807..45b0a28c 100644
--- a/code/vsetallnez_b.cpp
+++ b/code/vsetallnez_b.cpp
@@ -4,15 +4,15 @@ void test() {
   __m128i a = {0, 0};
   assert(!__lsx_bnz_b(a));
 
-  a = __m128i {1, 0};
+  a = __m128i{1, 0};
   assert(!__lsx_bnz_b(a));
 
-  a = __m128i {0, 1};
+  a = __m128i{0, 1};
   assert(!__lsx_bnz_b(a));
 
-  a = __m128i {0x1111111111111111, 0x1111111111111111};
+  a = __m128i{0x1111111111111111, 0x1111111111111111};
   assert(__lsx_bnz_b(a));
 
-  a = __m128i {0x1111111111111100, 0x1111111111111111};
+  a = __m128i{0x1111111111111100, 0x1111111111111111};
   assert(!__lsx_bnz_b(a));
 }
\ No newline at end of file
diff --git a/code/vsetallnez_d.cpp b/code/vsetallnez_d.cpp
index 17af4b4b..82f623f7 100644
--- a/code/vsetallnez_d.cpp
+++ b/code/vsetallnez_d.cpp
@@ -4,15 +4,15 @@ void test() {
   __m128i a = {0, 0};
   assert(!__lsx_bnz_d(a));
 
-  a = __m128i {1, 0};
+  a = __m128i{1, 0};
   assert(!__lsx_bnz_d(a));
 
-  a = __m128i {0, 1};
+  a = __m128i{0, 1};
   assert(!__lsx_bnz_d(a));
 
-  a = __m128i {0x1111111111111111, 0x1111111111111111};
+  a = __m128i{0x1111111111111111, 0x1111111111111111};
   assert(__lsx_bnz_d(a));
 
-  a = __m128i {0x0, 0x1111111111111111};
+  a = __m128i{0x0, 0x1111111111111111};
   assert(!__lsx_bnz_d(a));
 }
\ No newline at end of file
diff --git a/code/vsetallnez_h.cpp b/code/vsetallnez_h.cpp
index d1524475..68cd64c1 100644
--- a/code/vsetallnez_h.cpp
+++ b/code/vsetallnez_h.cpp
@@ -4,15 +4,15 @@ void test() {
   __m128i a = {0, 0};
   assert(!__lsx_bnz_h(a));
 
-  a = __m128i {1, 0};
+  a = __m128i{1, 0};
   assert(!__lsx_bnz_h(a));
 
-  a = __m128i {0, 1};
+  a = __m128i{0, 1};
   assert(!__lsx_bnz_h(a));
 
-  a = __m128i {0x1111111111111111, 0x1111111111111111};
+  a = __m128i{0x1111111111111111, 0x1111111111111111};
   assert(__lsx_bnz_h(a));
 
-  a = __m128i {0x1111111111110000, 0x1111111111111111};
+  a = __m128i{0x1111111111110000, 0x1111111111111111};
   assert(!__lsx_bnz_h(a));
 }
\ No newline at end of file
diff --git a/code/vsetallnez_w.cpp b/code/vsetallnez_w.cpp
index 80e38c62..54fed88a 100644
--- a/code/vsetallnez_w.cpp
+++ b/code/vsetallnez_w.cpp
@@ -4,15 +4,15 @@ void test() {
   __m128i a = {0, 0};
   assert(!__lsx_bnz_w(a));
 
-  a = __m128i {1, 0};
+  a = __m128i{1, 0};
   assert(!__lsx_bnz_w(a));
 
-  a = __m128i {0, 1};
+  a = __m128i{0, 1};
   assert(!__lsx_bnz_w(a));
 
-  a = __m128i {0x1111111111111111, 0x1111111111111111};
+  a = __m128i{0x1111111111111111, 0x1111111111111111};
   assert(__lsx_bnz_w(a));
 
-  a = __m128i {0x1111111100000000, 0x1111111111111111};
+  a = __m128i{0x1111111100000000, 0x1111111111111111};
   assert(!__lsx_bnz_w(a));
 }
\ No newline at end of file
diff --git a/code/vsetanyeqz_b.cpp b/code/vsetanyeqz_b.cpp
index df2dd2c1..1523ae3a 100644
--- a/code/vsetanyeqz_b.cpp
+++ b/code/vsetanyeqz_b.cpp
@@ -4,15 +4,15 @@ void test() {
   __m128i a = {0, 0};
   assert(__lsx_bz_b(a));
 
-  a = __m128i {1, 0};
+  a = __m128i{1, 0};
   assert(__lsx_bz_b(a));
 
-  a = __m128i {0, 1};
+  a = __m128i{0, 1};
   assert(__lsx_bz_b(a));
 
-  a = __m128i {0x1111111111111111, 0x1111111111111111};
+  a = __m128i{0x1111111111111111, 0x1111111111111111};
   assert(!__lsx_bz_b(a));
 
-  a = __m128i {0x1111111111111100, 0x1111111111111111};
+  a = __m128i{0x1111111111111100, 0x1111111111111111};
   assert(__lsx_bz_b(a));
 }
\ No newline at end of file
diff --git a/code/vsetanyeqz_d.cpp b/code/vsetanyeqz_d.cpp
index efdf3c37..3b1eaae7 100644
--- a/code/vsetanyeqz_d.cpp
+++ b/code/vsetanyeqz_d.cpp
@@ -4,15 +4,15 @@ void test() {
   __m128i a = {0, 0};
   assert(__lsx_bz_d(a));
 
-  a = __m128i {1, 0};
+  a = __m128i{1, 0};
   assert(__lsx_bz_d(a));
 
-  a = __m128i {0, 1};
+  a = __m128i{0, 1};
   assert(__lsx_bz_d(a));
 
-  a = __m128i {0x1111111111111111, 0x1111111111111111};
+  a = __m128i{0x1111111111111111, 0x1111111111111111};
   assert(!__lsx_bz_d(a));
 
-  a = __m128i {0x0, 0x1111111111111111};
+  a = __m128i{0x0, 0x1111111111111111};
   assert(__lsx_bz_d(a));
 }
\ No newline at end of file
diff --git a/code/vsetanyeqz_h.cpp b/code/vsetanyeqz_h.cpp
index 883e00d7..5e772e64 100644
--- a/code/vsetanyeqz_h.cpp
+++ b/code/vsetanyeqz_h.cpp
@@ -4,15 +4,15 @@ void test() {
   __m128i a = {0, 0};
   assert(__lsx_bz_h(a));
 
-  a = __m128i {1, 0};
+  a = __m128i{1, 0};
   assert(__lsx_bz_h(a));
 
-  a = __m128i {0, 1};
+  a = __m128i{0, 1};
   assert(__lsx_bz_h(a));
 
-  a = __m128i {0x1111111111111111, 0x1111111111111111};
+  a = __m128i{0x1111111111111111, 0x1111111111111111};
   assert(!__lsx_bz_h(a));
 
-  a = __m128i {0x1111111111110000, 0x1111111111111111};
+  a = __m128i{0x1111111111110000, 0x1111111111111111};
   assert(__lsx_bz_h(a));
 }
\ No newline at end of file
diff --git a/code/vsetanyeqz_w.cpp b/code/vsetanyeqz_w.cpp
index 444f1e8e..de1d136b 100644
--- a/code/vsetanyeqz_w.cpp
+++ b/code/vsetanyeqz_w.cpp
@@ -4,15 +4,15 @@ void test() {
   __m128i a = {0, 0};
   assert(__lsx_bz_w(a));
 
-  a = __m128i {1, 0};
+  a = __m128i{1, 0};
   assert(__lsx_bz_w(a));
 
-  a = __m128i {0, 1};
+  a = __m128i{0, 1};
   assert(__lsx_bz_b(a));
 
-  a = __m128i {0x1111111111111111, 0x1111111111111111};
+  a = __m128i{0x1111111111111111, 0x1111111111111111};
   assert(!__lsx_bz_w(a));
 
-  a = __m128i {0x1111111100000000, 0x1111111111111111};
+  a = __m128i{0x1111111100000000, 0x1111111111111111};
   assert(__lsx_bz_w(a));
 }
\ No newline at end of file
diff --git a/code/vseteqz_v.cpp b/code/vseteqz_v.cpp
index 43d25367..f6ec8b88 100644
--- a/code/vseteqz_v.cpp
+++ b/code/vseteqz_v.cpp
@@ -4,9 +4,9 @@ void test() {
   __m128i a = {0, 0};
   assert(__lsx_bz_v(a));
 
-  a = __m128i {1, 0};
+  a = __m128i{1, 0};
   assert(!__lsx_bz_v(a));
 
-  a = __m128i {0, 1};
+  a = __m128i{0, 1};
   assert(!__lsx_bz_v(a));
 }
\ No newline at end of file
diff --git a/code/vsetnez_v.cpp b/code/vsetnez_v.cpp
index 6aa0bdc0..3eb88701 100644
--- a/code/vsetnez_v.cpp
+++ b/code/vsetnez_v.cpp
@@ -4,9 +4,9 @@ void test() {
   __m128i a = {0, 0};
   assert(!__lsx_bnz_v(a));
 
-  a = __m128i {1, 0};
+  a = __m128i{1, 0};
   assert(__lsx_bnz_v(a));
 
-  a = __m128i {0, 1};
+  a = __m128i{0, 1};
   assert(__lsx_bnz_v(a));
 }
\ No newline at end of file
diff --git a/gen_measure.py b/gen_measure.py
new file mode 100644
index 00000000..98f2a22c
--- /dev/null
+++ b/gen_measure.py
@@ -0,0 +1,50 @@
+import glob
+
+# find known insts from binutils-gdb
+with open('code/measure.h', 'w') as f:
+    # measure unit time
+    print('INSTR_TEST(unit, "add.w $r12, $r1, $r12\\n", "r12")', file=f)
+
+    for line in open("../binutils-gdb/opcodes/loongarch-opc.c"):
+        line = line.strip()
+        if line.startswith("{") and line.endswith("},"):
+            parts = line.split(",")
+            if '"' in parts[2]:
+                name = parts[2].split('"')[1]
+                fmt = line.split('"')[3]
+
+                if name.startswith("v") or name.startswith("xv"):
+                    print("Processing", name, fmt)
+
+                    # latency test
+                    # at least one op depends on vd
+                    fmt_parts = fmt.split(",")
+                    for depend_i in range(1, len(fmt_parts)):
+                        ops = []
+                        for i, part in enumerate(fmt_parts):
+                            if part.startswith("v"):
+                                if i == depend_i:
+                                    # vd is always vr0
+                                    ops.append("$vr0")
+                                else:
+                                    ops.append(f"$vr{i}")
+                            elif part.startswith("x"):
+                                if i == depend_i:
+                                    # vd is always vr0
+                                    ops.append("$xr0")
+                                else:
+                                    ops.append(f"$xr{i}")
+                        print(f'INSTR_TEST({name.replace(".", "_")}_{depend_i}, "{name} {", ".join(ops)}\\n")', file=f)
+
+                    # throughput test
+                    # no dependency
+                    ops = []
+                    for i, part in enumerate(fmt_parts):
+                        if part.startswith("v"):
+                            ops.append(f"$vr{i}")
+                        elif part.startswith("x"):
+                            ops.append(f"$xr{i}")
+                    print(f'INSTR_TEST({name.replace(".", "_")}_tp, "{name} {", ".join(ops)}\\n")', file=f)
+
+                if "fcmp" in name:
+                    break
\ No newline at end of file