Package io.milvus.common.utils

Class Float16Utils

  • public class Float16Utils
extends Object

    • Constructor Detail

      • Float16Utils

        public Float16Utils()

    • Method Detail

      • floatToBf16

        public static short floatToBf16​(float input)
        Converts a float32 into bf16. May not produce correct values for subnormal floats.

        This method is copied from microsoft ONNX Runtime: https://github.com/microsoft/onnxruntime/blob/main/java/src/main/jvm/ai/onnxruntime/platform/Fp16Conversions.java

        Parameters:
        input - a standard float32 value which will be converted to a bfloat16 value
        Returns:
        a short value to store the bfloat16 value

      • bf16ToFloat

        public static float bf16ToFloat​(short input)
        Upcasts a bf16 value stored in a short into a float32 value.

        This method is copied from microsoft ONNX Runtime: https://github.com/microsoft/onnxruntime/blob/main/java/src/main/jvm/ai/onnxruntime/platform/Fp16Conversions.java

        Parameters:
        input - a bfloat16 value which will be converted to a float32 value
        Returns:
        a float32 value converted from a bfloat16

      • floatToFp16

        public static short floatToFp16​(float input)
        Rounds a float32 value to a fp16 stored in a short.

        This method is copied from microsoft ONNX Runtime: https://github.com/microsoft/onnxruntime/blob/main/java/src/main/jvm/ai/onnxruntime/platform/Fp16Conversions.java

        Parameters:
        input - a standard float32 value which will be converted to a float16 value
        Returns:
        a short value to store the float16 value

      • fp16ToFloat

        public static float fp16ToFloat​(short input)
        Upcasts a fp16 value stored in a short to a float32 value.

        This method is copied from microsoft ONNX Runtime: https://github.com/microsoft/onnxruntime/blob/main/java/src/main/jvm/ai/onnxruntime/platform/Fp16Conversions.java

        Parameters:
        input - a float16 value which will be converted to a float32 value
        Returns:
        a float32 value converted from a float16 value

      • f32VectorToBf16Buffer

        public static ByteBuffer f32VectorToBf16Buffer​(List<Float> vector)
        Rounds a float32 vector to bf16 values, and stores into a ByteBuffer.
        Parameters:
        vector - a float32 vector
        Returns:
        ByteBuffer the vector is converted to bfloat16 values and stored into a ByteBuffer

      • fp16BufferToVector

        public static List<Float> fp16BufferToVector​(ByteBuffer buf)
        Converts a ByteBuffer to fp16 vector upcasts to float32 array.
        Parameters:
        buf - a buffer to store a float16 vector
        Returns:
        List of Float a float32 vector

      • f32VectorToFp16Buffer

        public static ByteBuffer f32VectorToFp16Buffer​(List<Float> vector)
        Rounds a float32 vector to fp16 values, and stores into a ByteBuffer.
        Parameters:
        vector - a float32 vector
        Returns:
        ByteBuffer the vector is converted to float16 values and stored in a ByteBuffer

      • bf16BufferToVector

        public static List<Float> bf16BufferToVector​(ByteBuffer buf)
        Converts a ByteBuffer to bf16 vector upcasts to float32 array.
        Parameters:
        buf - a buffer to store a bfloat16 vector
        Returns:
        List of Float the vector is converted to float32 values

      • f16VectorToBuffer

        public static ByteBuffer f16VectorToBuffer​(List<Short> vector)
        Stores a fp16/bf16 vector into a ByteBuffer.
        Parameters:
        vector - a float16 vector stored in a list of Short
        Returns:
        ByteBuffer a buffer to store the float16 vector

      • bufferToF16Vector

        public static List<Short> bufferToF16Vector​(ByteBuffer buf)
        Converts a ByteBuffer to a fp16/bf16 vector stored in short array.
        Parameters:
        buf - a buffer to store a float16 vector
        Returns:
        List of Short the vector is converted to a list of Short, each Short value is a float16 value