Add convertion function between binary float16 and float32 in softfloat

softfloat/bf16_to_f32.c

@@ -0,0 +1,80 @@
+
+/*============================================================================
+
+This C source file is part of the SoftFloat IEEE Floating-Point Arithmetic
+Package, Release 3d, by John R. Hauser.
+
+Copyright 2011, 2012, 2013, 2014, 2015, 2016, 2017 The Regents of the
+University of California.  All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+ 1. Redistributions of source code must retain the above copyright notice,
+    this list of conditions, and the following disclaimer.
+
+ 2. Redistributions in binary form must reproduce the above copyright notice,
+    this list of conditions, and the following disclaimer in the documentation
+    and/or other materials provided with the distribution.
+
+ 3. Neither the name of the University nor the names of its contributors may
+    be used to endorse or promote products derived from this software without
+    specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS "AS IS", AND ANY
+EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE, ARE
+DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR ANY
+DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+=============================================================================*/
+
+#include <stdbool.h>
+#include <stdint.h>
+#include "platform.h"
+#include "internals.h"
+#include "specialize.h"
+#include "softfloat.h"
+
+float32_t bf16_to_f32( bfloat16_t a )
+{
+    union ui16_f16 uA;
+    uint_fast16_t uiA;
+    bool sign;
+    int_fast16_t exp;
+    uint_fast16_t frac;
+    struct commonNaN commonNaN;
+    uint_fast32_t uiZ;
+    union ui32_f32 uZ;
+
+    /*------------------------------------------------------------------------
+    *------------------------------------------------------------------------*/
+    uA.f = a;
+    uiA = uA.ui;
+    sign = signBF16UI( uiA );
+    exp  = expBF16UI( uiA );
+    frac = fracBF16UI( uiA );
+    /*------------------------------------------------------------------------
+    *------------------------------------------------------------------------*/
+    if ( exp == 0xFF ) {
+        if ( frac ) {
+            softfloat_bf16UIToCommonNaN( uiA, &commonNaN );
+            uiZ = softfloat_commonNaNToF32UI( &commonNaN );
+        } else {
+            uiZ = packToF32UI( sign, 0xFF, 0 );
+        }
+        goto uiZ;
+    }
+    /*------------------------------------------------------------------------
+    *------------------------------------------------------------------------*/
+    uiZ = packToF32UI( sign, exp, (uint_fast32_t) frac<<16 );
+ uiZ:
+    uZ.ui = uiZ;
+    return uZ.f;
+
+}

softfloat/f32_to_bf16.c

@@ -0,0 +1,92 @@
+
+/*============================================================================
+
+This C source file is part of the SoftFloat IEEE Floating-Point Arithmetic
+Package, Release 3d, by John R. Hauser.
+
+Copyright 2011, 2012, 2013, 2014, 2015 The Regents of the University of
+California.  All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+ 1. Redistributions of source code must retain the above copyright notice,
+    this list of conditions, and the following disclaimer.
+
+ 2. Redistributions in binary form must reproduce the above copyright notice,
+    this list of conditions, and the following disclaimer in the documentation
+    and/or other materials provided with the distribution.
+
+ 3. Neither the name of the University nor the names of its contributors may
+    be used to endorse or promote products derived from this software without
+    specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS "AS IS", AND ANY
+EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE, ARE
+DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR ANY
+DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+=============================================================================*/
+
+#include <stdbool.h>
+#include <stdint.h>
+#include "platform.h"
+#include "internals.h"
+#include "specialize.h"
+#include "softfloat.h"
+
+bfloat16_t f32_to_bf16( float32_t a )
+{
+    union ui32_f32 uA;
+    uint_fast32_t uiA;
+    bool sign;
+    int_fast16_t exp;
+    uint_fast32_t frac;
+    struct commonNaN commonNaN;
+    struct exp16_sig32 normExpSig;
+    uint_fast16_t uiZ, frac16;
+    union ui16_f16 uZ;
+
+    /*------------------------------------------------------------------------
+    *------------------------------------------------------------------------*/
+    uA.f = a;
+    uiA = uA.ui;
+    sign = signF32UI( uiA );
+    exp  = expF32UI( uiA );
+    frac = fracF32UI( uiA );
+    /*------------------------------------------------------------------------
+    *------------------------------------------------------------------------*/
+    if ( exp == 0xFF ) {
+        if ( frac ) {
+            softfloat_f32UIToCommonNaN( uiA, &commonNaN );
+            uiZ = softfloat_commonNaNToBF16UI( &commonNaN );
+        } else {
+            uiZ = packToBF16UI( sign, 0xFF, 0 );
+        }
+        goto uiZ;
+    }
+    /*------------------------------------------------------------------------
+    *------------------------------------------------------------------------*/
+    if ( ! (exp | frac) ) {
+        uiZ = packToBF16UI( sign, 0, 0 );
+        goto uiZ;
+    } else if ( !exp ) {
+        normExpSig = softfloat_normSubnormalF32Sig( frac );
+        exp = normExpSig.exp;
+        frac = normExpSig.sig;
+    }
+    frac16 = frac>>9 | ((frac & 0x1FF) != 0);
+    /*------------------------------------------------------------------------
+    *------------------------------------------------------------------------*/
+    return softfloat_roundPackToBF16( sign, exp - 1, frac16 | 0x4000 );
+ uiZ:
+    uZ.ui = uiZ;
+    return uZ.f;
+
+}

softfloat/internals.h

@@ -89,6 +89,11 @@ int_fast64_t softfloat_roundMToI64( bool, uint32_t *, uint_fast8_t, bool );
 #define fracF16UI( a ) ((a) & 0x03FF)
 #define packToF16UI( sign, exp, sig ) (((uint16_t) (sign)<<15) + ((uint16_t) (exp)<<10) + (sig))
 
+#define signBF16UI( a ) ((bool) ((uint16_t) (a)>>15))
+#define expBF16UI( a ) ((int_fast16_t) ((a)>>7) & 0xFF)
+#define fracBF16UI( a ) ((a) & 0x07F)
+#define packToBF16UI( sign, exp, sig ) (((uint16_t) (sign)<<15) + ((uint16_t) (exp)<<7) + (sig))
+
 #define isNaNF16UI( a ) (((~(a) & 0x7C00) == 0) && ((a) & 0x03FF))
 
 struct exp8_sig16 { int_fast8_t exp; uint_fast16_t sig; };
@@ -103,6 +108,7 @@ float16_t
  softfloat_mulAddF16(
      uint_fast16_t, uint_fast16_t, uint_fast16_t, uint_fast8_t );
 
+bfloat16_t softfloat_roundPackToBF16( bool, int_fast16_t, uint_fast16_t );
 /*----------------------------------------------------------------------------
 *----------------------------------------------------------------------------*/
 #define signF32UI( a ) ((bool) ((uint32_t) (a)>>31))

softfloat/s_roundPackToBF16.c

@@ -0,0 +1,113 @@
+
+/*============================================================================
+
+This C source file is part of the SoftFloat IEEE Floating-Point Arithmetic
+Package, Release 3d, by John R. Hauser.
+
+Copyright 2011, 2012, 2013, 2014, 2015, 2017 The Regents of the University of
+California.  All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+ 1. Redistributions of source code must retain the above copyright notice,
+    this list of conditions, and the following disclaimer.
+
+ 2. Redistributions in binary form must reproduce the above copyright notice,
+    this list of conditions, and the following disclaimer in the documentation
+    and/or other materials provided with the distribution.
+
+ 3. Neither the name of the University nor the names of its contributors may
+    be used to endorse or promote products derived from this software without
+    specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS "AS IS", AND ANY
+EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE, ARE
+DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR ANY
+DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+=============================================================================*/
+
+#include <stdbool.h>
+#include <stdint.h>
+#include "platform.h"
+#include "internals.h"
+#include "softfloat.h"
+
+bfloat16_t
+ softfloat_roundPackToBF16( bool sign, int_fast16_t exp, uint_fast16_t sig )
+{
+    uint_fast8_t roundingMode;
+    bool roundNearEven;
+    uint_fast8_t roundIncrement, roundBits;
+    bool isTiny;
+    uint_fast16_t uiZ;
+    union ui16_f16 uZ;
+
+    /*------------------------------------------------------------------------
+    *------------------------------------------------------------------------*/
+    roundingMode = softfloat_roundingMode;
+    roundNearEven = (roundingMode == softfloat_round_near_even);
+    roundIncrement = 0x40;
+    if ( ! roundNearEven && (roundingMode != softfloat_round_near_maxMag) ) {
+        roundIncrement =
+            (roundingMode
+                 == (sign ? softfloat_round_min : softfloat_round_max))
+                ? 0x7F
+                : 0;
+    }
+    roundBits = sig & 0x7F;
+    /*------------------------------------------------------------------------
+    *------------------------------------------------------------------------*/
+    if ( 0xFD <= (unsigned int) exp ) {
+        if ( exp < 0 ) {
+            /*----------------------------------------------------------------
+            *----------------------------------------------------------------*/
+            isTiny =
+                (softfloat_detectTininess == softfloat_tininess_beforeRounding)
+                    || (exp < -1) || (sig + roundIncrement < 0x8000);
+            sig = softfloat_shiftRightJam32( sig, -exp );
+            exp = 0;
+            roundBits = sig & 0x7F;
+            if ( isTiny && roundBits ) {
+                softfloat_raiseFlags( softfloat_flag_underflow );
+            }
+        } else if ( (0xFD < exp) || (0x8000 <= sig + roundIncrement) ) {
+            /*----------------------------------------------------------------
+            *----------------------------------------------------------------*/
+            softfloat_raiseFlags(
+                softfloat_flag_overflow | softfloat_flag_inexact );
+            uiZ = packToBF16UI( sign, 0xFF, 0 ) - ! roundIncrement;
+            goto uiZ;
+        }
+    }
+    /*------------------------------------------------------------------------
+    *------------------------------------------------------------------------*/
+    sig = (sig + roundIncrement)>>7;
+    if ( roundBits ) {
+        softfloat_exceptionFlags |= softfloat_flag_inexact;
+#ifdef SOFTFLOAT_ROUND_ODD
+        if ( roundingMode == softfloat_round_odd ) {
+            sig |= 1;
+            goto packReturn;
+        }
+#endif
+    }
+    sig &= ~(uint_fast16_t) (! (roundBits ^ 0x40) & roundNearEven);
+    if ( ! sig ) exp = 0;
+    /*------------------------------------------------------------------------
+    *------------------------------------------------------------------------*/
+ packReturn:
+    uiZ = packToBF16UI( sign, exp, sig );
+ uiZ:
+    uZ.ui = uiZ;
+    return uZ.f;
+
+}
+

softfloat/softfloat.h

@@ -154,6 +154,7 @@ uint_fast64_t f16_to_ui64_r_minMag( float16_t, bool );
 int_fast32_t f16_to_i32_r_minMag( float16_t, bool );
 int_fast64_t f16_to_i64_r_minMag( float16_t, bool );
 float32_t f16_to_f32( float16_t );
+float32_t bf16_to_f32( bfloat16_t );
 float64_t f16_to_f64( float16_t );
 #ifdef SOFTFLOAT_FAST_INT64
 extFloat80_t f16_to_extF80( float16_t );
@@ -196,6 +197,7 @@ uint_fast64_t f32_to_ui64_r_minMag( float32_t, bool );
 int_fast32_t f32_to_i32_r_minMag( float32_t, bool );
 int_fast64_t f32_to_i64_r_minMag( float32_t, bool );
 float16_t f32_to_f16( float32_t );
+bfloat16_t f32_to_bf16( float32_t );
 float64_t f32_to_f64( float32_t );
 #ifdef SOFTFLOAT_FAST_INT64
 extFloat80_t f32_to_extF80( float32_t );

softfloat/softfloat.mk.in

@@ -45,6 +45,7 @@ softfloat_c_srcs = \
 	f16_sqrt.c \
 	f16_sub.c \
 	f16_to_f128.c \
+	bf16_to_f32.c \
 	f16_to_f32.c \
 	f16_to_f64.c \
 	f16_to_i8.c \
@@ -76,6 +77,7 @@ softfloat_c_srcs = \
 	f32_sqrt.c \
 	f32_sub.c \
 	f32_to_f128.c \
+	f32_to_bf16.c \
 	f32_to_f16.c \
 	f32_to_f64.c \
 	f32_to_i16.c \
@@ -181,6 +183,7 @@ softfloat_c_srcs = \
 	s_roundMToUI64.c \
 	s_roundPackMToI64.c \
 	s_roundPackMToUI64.c \
+	s_roundPackToBF16.c \
 	s_roundPackToF128.c \
 	s_roundPackToF16.c \
 	s_roundPackToF32.c \

softfloat/softfloat_types.h

@@ -48,6 +48,7 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 | (typically 'float' and 'double', and possibly 'long double').
 *----------------------------------------------------------------------------*/
 typedef struct { uint16_t v; } float16_t;
+typedef float16_t bfloat16_t;
 typedef struct { uint32_t v; } float32_t;
 typedef struct { uint64_t v; } float64_t;
 typedef struct { uint64_t v[2]; } float128_t;

softfloat/specialize.h

@@ -98,6 +98,11 @@ struct commonNaN { char _unused; };
 *----------------------------------------------------------------------------*/
 #define defaultNaNF16UI 0x7E00
 
+/*----------------------------------------------------------------------------
+| The bit pattern for a default generated binary 16-bit floating-point NaN.
+*----------------------------------------------------------------------------*/
+#define defaultNaNBF16UI 0x7FC0
+
 /*----------------------------------------------------------------------------
 | Returns true when 16-bit unsigned integer `uiA' has the bit pattern of a
 | 16-bit floating-point signaling NaN.
@@ -113,6 +118,20 @@ struct commonNaN { char _unused; };
 *----------------------------------------------------------------------------*/
 #define softfloat_f16UIToCommonNaN( uiA, zPtr ) if ( ! ((uiA) & 0x0200) ) (void) (zPtr), softfloat_raiseFlags( softfloat_flag_invalid )
 
+/*----------------------------------------------------------------------------
+| Assuming `uiA' has the bit pattern of a binary 16-bit floating-point NaN, converts
+| this NaN to the common NaN form, and stores the resulting common NaN at the
+| location pointed to by `zPtr'.  If the NaN is a signaling NaN, the invalid
+| exception is raised.
+*----------------------------------------------------------------------------*/
+#define softfloat_bf16UIToCommonNaN( uiA, zPtr ) if ( ! ((uiA) & 0x040) ) (void) (zPtr), softfloat_raiseFlags( softfloat_flag_invalid )
+
+/*----------------------------------------------------------------------------
+| Converts the common NaN pointed to by `aPtr' into a binary 16-bit floating-point
+| NaN, and returns the bit pattern of this value as an unsigned integer.
+*----------------------------------------------------------------------------*/
+#define softfloat_commonNaNToBF16UI( aPtr ) ((uint_fast16_t) defaultNaNBF16UI)
+
 /*----------------------------------------------------------------------------
 | Converts the common NaN pointed to by `aPtr' into a 16-bit floating-point
 | NaN, and returns the bit pattern of this value as an unsigned integer.