[libc-commits] [libc] [libc] DRAFT: Add Generic and NVPTX Sin Benchmark (PR #99795)

[Joseph Huber via libc-commits]

================
@@ -105,8 +106,34 @@ void print_results(Benchmark *b) {
       << " stddev (num threads: " << num_threads << ")\n";
 }
 
+// We want our random values to be approximately
+// |real value| <= 2^(max_exponent) * (1 + (random 52 bits) * 2^-52) <
+// 2^(max_exponent + 1)
+// The largest integer that can be stored in a double is 2^53
+static constexpr int MAX_EXPONENT = 52;
+
+static double get_rand() {
+  using FPBits = LIBC_NAMESPACE::fputil::FPBits<double>;
+  uint64_t bits = LIBC_NAMESPACE::rand();
+  double scale = 0.5 + MAX_EXPONENT / 2048.0;
+  FPBits fp(bits);
+  fp.set_biased_exponent(
+      static_cast<uint32_t>(fp.get_biased_exponent() * scale));
+  return fp.get_val();
+}
+
+static void init_random_input() {
+  LIBC_NAMESPACE::srand(LIBC_NAMESPACE::gpu::processor_clock());
+  for (int i = 0; i < RANDOM_INPUT_SIZE; i++) {
+    random_input[i] = get_rand();
+  }
+}
+
----------------
jhuber6 wrote:

I figured this would be handled in the benchmark itself, like we do it before the latency calculation.

https://github.com/llvm/llvm-project/pull/99795