-
Notifications
You must be signed in to change notification settings - Fork 2.9k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Fix NaN propagation for float16 min and max operators #22161
base: main
Are you sure you want to change the base?
Conversation
@@ -360,6 +366,16 @@ __device__ __inline__ double _Min(double a, double b) { | |||
return (isnan(a) || isnan(b)) ? std::numeric_limits<double>::quiet_NaN() : ( a < b ? a : b ); | |||
} | |||
|
|||
template <> | |||
__device__ __inline__ half _Min(half a, half b) { | |||
return ISNAN_HALF(a) ? a : (ISNAN_HALF(b) ? b : (a < b ? a : b)); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
May use __hmin_nan(a, b) for _Min and __hmax_nan(a, b) for _Max, when a/b is half or bfloat16 (casted to __nv_bfloat16). See https://docs.nvidia.com/cuda/archive/12.1.0/cuda-math-api/group__CUDA__MATH____HALF__COMPARISON.html#group__CUDA__MATH____HALF__COMPARISON_1g9752fee573b47368538ab19495ab9623
Difference is that those functions return canonical NaN, and more consistent with float/double (see line 366).
if (is_min) { | ||
output_vec_map = input_1_vec_map.min(static_cast<Eigen::half>(per_iter_bh.ScalarInput0<MLFloat16>())); | ||
output_vec_map = input_1_vec_map.template min<Eigen::PropagateNaN>( | ||
static_cast<Eigen::half>(per_iter_bh.ScalarInput0<MLFloat16>())); | ||
} else { | ||
output_vec_map = input_1_vec_map.max(static_cast<Eigen::half>(per_iter_bh.ScalarInput0<MLFloat16>())); | ||
output_vec_map = input_1_vec_map.template max<Eigen::PropagateNaN>( | ||
static_cast<Eigen::half>(per_iter_bh.ScalarInput0<MLFloat16>())); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
How about logic like
Eigen::half scalar_input = static_cast<Eigen::half>(per_iter_bh.ScalarInput0<MLFloat16>());
scalar_input = std::isnan(static_cast<float>(scalar_input)) ? Eigen::NumTraits<Eigen::half>::quiet_NaN() : scalar_input;
if (is_min) {
output_vec_map = input_1_vec_map.isNaN().select(
Eigen::NumTraits<Eigen::half>::quiet_NaN(),
input_1_vec_map.template min<Eigen::PropagateNaN>(scalar_input)
);
} else {
output_vec_map = input_1_vec_map.isNaN().select(
Eigen::NumTraits<Eigen::half>::quiet_NaN(),
input_1_vec_map.template max<Eigen::PropagateNaN>(scalar_input)
);
}
@@ -790,9 +794,9 @@ static Status MinMaxMLFloat16(const OpKernel& inst, OpKernelContext* context) { | |||
EigenVectorArrayMap<Eigen::half> output_vec_map(output, num_elements); | |||
|
|||
if (is_min) { | |||
output_vec_map = input_0_vec_map.min(input_1_vec_map); | |||
output_vec_map = input_0_vec_map.template min<Eigen::PropagateNaN>(input_1_vec_map); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
How about logic like
output_vec_map = (input_0_vec_map.isNaN() || input_1_vec_map.isNaN()).select(
Eigen::NumTraits<Eigen::half>::quiet_NaN(),
input_0_vec_map.template min<Eigen::PropagateNaN>(input_1_vec_map)
);
Description
This makes min and max with NaN for either operand always return NaN for float16 data, matching the behaviour of float and double.
The behaviour for floats and doubles was previously fixed for the CPU provider in #21492 and the CUDA provider in #19984, but these PRs didn't fix the behaviour for float16 due to tests causing asan errors. The memory access violations with float16 data have now been fixed in #22135, so this PR is a follow up to make float16 min and max behave the same as float and double for both the CPU and CUDA providers now that we can add tests for this.
Motivation and Context
Relevant previous issues (not float16 specific):
clip
behaviour with NaN values is different between GPU and CPU onnx inference onnx/onnx#6003