small changes

ROCm · Aug 22, 2024 · ef368c5 · ef368c5
1 parent 84a0333
commit ef368c5
Show file tree

Hide file tree

Showing 4 changed files with 8 additions and 4 deletions.
diff --git a/xla/debug_options_flags.cc b/xla/debug_options_flags.cc
@@ -49,7 +49,7 @@ DebugOptions DefaultDebugOptionsIgnoringFlags() {
   opts.set_xla_llvm_enable_invariant_load_metadata(true);
   opts.set_xla_llvm_disable_expensive_passes(false);
   opts.set_xla_backend_optimization_level(3);
-  opts.set_xla_gpu_autotune_level(4);
+  opts.set_xla_gpu_autotune_level(5);
   opts.set_xla_gpu_autotune_max_solutions(0);
   opts.set_xla_cpu_multi_thread_eigen(true);
   opts.set_xla_gpu_cuda_data_dir("./cuda_sdk_lib");

diff --git a/xla/service/gpu/buffer_comparator.cc b/xla/service/gpu/buffer_comparator.cc
@@ -104,6 +104,10 @@ static absl::StatusOr<bool> DeviceCompare(
   return result == 0;
 }
 
+// Host side comparison code that does the same thing, but reports some of the
+// differences as well. It only print logs for debugging.
+//
+// Returns true if no differences were seen, false otherwise.
 template <typename ElementType, typename ComparisonType>
 static absl::StatusOr<bool> HostCompare(const ComparisonParams& params) {
   int64_t n = params.current.size() / sizeof(ElementType);
@@ -195,7 +199,7 @@ absl::StatusOr<bool> BufferComparator::CompareEqual(
            "fp8_e5m2_comparison", buffer_comparator::fp8_e5m2_comparison(),
            params);
 #endif  // GOOGLE_CUDA
-#if TENSORFLOW_USE_ROCM && TF_ROCM_VERSION >= 60300
+#if TENSORFLOW_USE_ROCM && TF_ROCM_VERSION >= 60200
     case xla::F8E4M3FNUZ:
       return CompareEqualParameterized<tsl::float8_e4m3fnuz, float>(
           "fp8_e4m3fnuz_comparison",

diff --git a/xla/service/gpu/buffer_comparator.h b/xla/service/gpu/buffer_comparator.h
@@ -51,7 +51,7 @@ class BufferComparator {
                                     se::DeviceMemoryBase expected) const;
  private:
   Shape shape_;
-  float relative_tol_;  // relative tolerance for comparison
+  double relative_tol_;  // relative tolerance for comparison
   bool verbose_;          // whether to print out error message on mismatch
 };
 

diff --git a/xla/service/gpu/gemm_algorithm_picker.cc b/xla/service/gpu/gemm_algorithm_picker.cc
@@ -343,7 +343,7 @@ class GemmAutotuner {
           comparator.CompareEqual(stream_, /*current=*/OutputBuffer(),
                                     /*expected=*/reference_buffer));
       if (!outputs_match) {
-        LOG(ERROR) << "Results mismatch between different GEMM algorithms. "
+        LOG(WARNING) << "Results mismatch between different GEMM algorithms. "
                    << "This is likely a bug/unexpected loss of precision.";
         CHECK(!autotune_config_.should_crash_on_check_failure());