Format conversion method of floating point number and device employing the same

A format conversion method of an operand expressed as a double-precision floating point number having an exponent E and a mantissa M to a single-precision floating point number, comprising the steps of:

• obtaining a first exponent e.sub.1 by subtracting an exponent bias of the double-precision floating point number from the exponent E of the operand and by adding an exponent bias of the single-precision floating point number thereto;
• (a) when the first exponent e.sub.1 is a negative value,
• obtaining a first mantissa m.sub.2 by shifting the mantissa M of the operand to right by an amount obtained by adding 1 to an absolute value of the first exponent e.sub.1, and
• making a second mantissa m.sub.3 obtained by rounding the first mantissa m.sub.2 to a predetermined bit length a mantissa of the single-precision floating point number to be obtained, and making 0 an exponent of the single-precision floating point number to be obtained in order to obtain a denormalized conversion result;
• (b) when the first exponent e.sub.1 is 0,
• obtaining a third mantissa m.sub.4 by shifting the mantissa M of the operand to right by 1 bit,
• obtaining a fourth mantissa m.sub.5 by rounding the third mantissa m.sub.4 to the predetermined bit length,
• making 0 and the fourth mantissa m.sub.5 respectively an exponent and a mantissa of the single-precision floating point number to be obtained in order to obtain a denormalized conversion result when no carry is caused by rounding the mantissa m.sub.4, and
• making a second exponent e.sub.3 obtained by adding 1 to the first exponent e.sub.1 and the fourth mantissa m.sub.5 respectively an exponent and a mantissa of the single-precision floating point number to be obtained in order to obtain a normalized conversion result when a carry is caused by rounding the third mantissa m.sub.4 ;
• and
• (c) when the first exponent e.sub.1 is a positive value,
• obtaining a fifth mantissa m.sub.6 by rounding the mantissa M of the operand to the predetermined bit length,
• making the first exponent e.sub.1 and the fifth mantissa m.sub.6 respectively an exponent and a mantissa of the single-precision floating point number to be obtained in order to obtain a normalized conversion result when no carry is caused by rounding the mantissa M of the operand, and
• making a third exponent e.sub.4 obtained by adding 1 to the first exponent e.sub.1 and a sixth mantissa m.sub.7 obtained by shifting the fifth mantissa m.sub.6 to right by 1 bit respectively an exponent and a mantissa of the single-precision floating point number to be obtained in order to obtain a normalized conversion result when a carry is caused by rounding the mantissa M of the operand.

A format converting device for converting an operand expressed as a double-precision floating point number having an exponent E and a mantissa M to a single-precision floating point number, comprising:

• an exponent evaluating unit having functions of obtaining a first exponent e.sub.1 to be an exponent of the single-precision floating point number by subtracting an exponent bias of the double-precision floating point number from the exponent E of the operand and by adding an exponent bias of the single-precision floating point number thereto and of outputting the first exponent e.sub.1 and a numerical value .vertline.e.sub.1 .vertline.+1 obtained by adding 1 to an absolute value of the first exponent e.sub.1 ;
• and
• a first shifter having a function of outputting a numerical value to be a mantissa of the single-precision floating point number obtained by shifting the mantissa M of the operand to right by an amount specified with the numerical value .vertline.e.sub.1 .vertline.+1 outputted from said exponent evaluating unit when the first exponent e.sub.1 is a negative value, outputting a numerical value to be a mantissa of the single-precision floating point number obtained by shifting the mantissa M of the operand to right by 1 bit when the first exponent e.sub.1 is 0, and outputting a numerical value identical with the mantissa M of the operand to be a mantissa of the single-precision floating point number when the first exponent e.sub.1 is a positive value.

A format conversion method of a fixed point number as a integer type operand having a most significant bit expressing a sign and an integer part of a first bit length, wherein a negative value is expressed by 2's complement, to a normalized floating point number having an exponent .epsilon. to which a bias is added and a mantissa .mu. of a second bit length which is shorter than the first bit length, comprising the steps of:

• judging whether the operand is a positive value or a negative value based on the most significant bit of the operand;
• (a) when the operand is a positive value,
• obtaining a distance between the most significant bit of the operand and a bit which is 1 and nearest the most significant bit as a required left shift amount LSA,
• obtaining a first mantissa m.sub.1 by shifting the operand to left by the left shift amount LSA,
• obtaining a first exponent e.sub.1 by subtracting the left shift amount LSA from the first bit length of the integer part of the operand and by adding the exponent bias of the floating point number thereto, and
• when the left shift amount LSA is greater than or equal to a threshold obtained by subtracting the second bit length from the number of bits of the operand including the most significant bit, making the first exponent e.sub.1 and the first mantissa m.sub.1 respectively the exponent .epsilon. and the mantissa .mu. of the floating point number to be obtained,
• when the left shift amount LSA is smaller than the threshold, obtaining a second mantissa m.sub.2 by rounding the first mantissa m.sub.1 to the second bit length, and making the first exponent e.sub.1 and the second mantissa m.sub.2 respectively the exponent .epsilon. and the mantissa .mu. of the floating point number to be obtained when no carry is caused by rounding the first mantissa m.sub.1, making a second exponent e.sub.2 obtained by adding 1 to the first exponent e.sub.1 and a third mantissa m.sub.3 obtained by shifting the second mantissa m.sub.2 to right by 1 bit respectively the exponent .epsilon. and the mantissa .mu. of the floating point number to be obtained when a carry is caused by rounding the first mantissa m.sub.1 ;
• and
• (b) when the operand is a negative value,
• obtaining a distance between the most significant bit of the operand and a bit which is 0 and nearest the most significant bit as a required left shift amount LSA,
• obtaining a fourth mantissa m.sub.4 by shifting the operand to left by the left shift amount LSA,
• obtaining a third exponent e.sub.3 by subtracting the left shift amount LSA from the first bit length of the integer part of the operand and by adding the exponent bias of the floating point number thereto,
• obtaining a fifth mantissa m.sub.5 by subtracting the fourth mantissa m.sub.4 from 0, and
• obtaining a sixth mantissa m.sub.6 by rounding the fifth mantissa m.sub.5 to the second bit length, and making the third exponent e.sub.3 and the sixth mantissa m.sub.6 respectively the exponent .epsilon. and the mantissa .mu. of the floating point number to be obtained when no carry is caused by rounding the fifth mantissa m.sub.5, making a fourth exponent e.sub.4 obtained by adding 1 to the third exponent e.sub.3 and a seventh mantissa m.sub.7 obtained by shifting the sixth mantissa m.sub.6 to right by 1 bit respectively the exponent .epsilon. and the mantissa .mu. of the floating point number to be obtained when a carry is caused by rounding the fifth mantissa m.sub.5.

A format converting device for converting an operand a total bit number of which is 32 having a most significant bit expressing a sign and an integer part of 31-bit length, wherein a negative value is expressed by 2's complement, to a normalized single-precision floating point number having an exponent to which a bias is added and a fraction of 23-bit length, comprising:

• an advancing 1 detector having a function of detecting a bit which is 1 and nearest the most significant bit of the operand when the operand is a positive value, detecting a bit which is 1 and nearest a most significant bit of data obtained by inverting all bits of the operand, respectively, to output a distance between the detected bit and the most significant bit as a required left shift amount LSA;
• a first shifter for outputting a numerical value obtained by shifting the operand to left by the left shift amount LSA;
• an exponent evaluating unit for outputting, as an exponent of the single-precision floating point number to be obtained, a numerical value obtained by subtracting the left shift amount LSA from 31 and by adding the exponent bias of the single-precision floating point number thereto;
• an adding/subtracting unit having a function of receiving an shifted output of said first shifter and 0 and outputting, as an intermediate data, the shifted output plus 0 when the operand is a positive value, a numerical value obtained by subtracting the shifted output from 0 when the operand is a negative value, respectively;
• a rounding unit for outputting, as a corrected mantissa of the single-precision floating point number to be obtained, a numerical value obtained by rounding the intermediate data outputted from said adding/subtracting unit to a predetermined bit length when the operand is a negative value or when the operand is a positive value and the left shift amount LSA is smaller than 8;
• a second shifter for outputting, as a further corrected mantissa of the single-precision floating point number to be obtained, a numerical value obtained by shifting the numerical value outputted from said rounding unit to right by 1 bit when a carry is caused by rounding in said rounding unit;
• and
• an exponent correcting unit for outputting, as a corrected exponent of the single-precision floating point number to be obtained, a numerical value obtained by adding 1 to the numerical value outputted from said exponent evaluating unit when a carry is caused by rounding in said rounding unit.