mont_mulf.c revision 7c478bd95313f5f23a4c958a745db2134aa03244
/*
* CDDL HEADER START
*
* The contents of this file are subject to the terms of the
* Common Development and Distribution License, Version 1.0 only
* (the "License"). You may not use this file except in compliance
* with the License.
*
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
* or http://www.opensolaris.org/os/licensing.
* See the License for the specific language governing permissions
* and limitations under the License.
*
* When distributing Covered Code, include this CDDL HEADER in each
* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
* If applicable, add the following below this CDDL HEADER, with the
* fields enclosed by brackets "[]" replaced with your own identifying
* information: Portions Copyright [yyyy] [name of copyright owner]
*
* CDDL HEADER END
*/
/*
* Copyright 2005 Sun Microsystems, Inc. All rights reserved.
* Use is subject to license terms.
*/
#pragma ident "%Z%%M% %I% %E% SMI"
/*
* If compiled without -DRF_INLINE_MACROS then needs -lm at link time
* If compiled with -DRF_INLINE_MACROS then needs conv.il at compile time
* (i.e. cc <compileer_flags> -DRF_INLINE_MACROS conv.il mont_mulf.c )
*/
#include <sys/types.h>
#include <math.h>
static const double TwoTo16 = 65536.0;
static const double TwoToMinus16 = 1.0/65536.0;
static const double Zero = 0.0;
static const double TwoTo32 = 65536.0 * 65536.0;
static const double TwoToMinus32 = 1.0 / (65536.0 * 65536.0);
#ifdef RF_INLINE_MACROS
double upper32(double);
double lower32(double, double);
double mod(double, double, double);
#else
static double
upper32(double x)
{
return (floor(x * TwoToMinus32));
}
static double
lower32(double x, double y)
{
return (x - TwoTo32 * floor(x * TwoToMinus32));
}
static double
mod(double x, double oneoverm, double m)
{
return (x - m * floor(x * oneoverm));
}
#endif
static void
cleanup(double *dt, int from, int tlen)
{
int i;
double tmp, tmp1, x, x1;
tmp = tmp1 = Zero;
for (i = 2 * from; i < 2 * tlen; i += 2) {
x = dt[i];
x1 = dt[i + 1];
dt[i] = lower32(x, Zero) + tmp;
dt[i + 1] = lower32(x1, Zero) + tmp1;
tmp = upper32(x);
tmp1 = upper32(x1);
}
}
void
conv_d16_to_i32(uint32_t *i32, double *d16, int64_t *tmp, int ilen)
{
int i;
int64_t t, t1, /* using int64_t and not uint64_t */
a, b, c, d; /* because more efficient code is */
/* generated this way, and there */
/* is no overflow */
t1 = 0;
a = (int64_t)d16[0];
b = (int64_t)d16[1];
for (i = 0; i < ilen - 1; i++) {
c = (int64_t)d16[2 * i + 2];
t1 += a & 0xffffffff;
t = (a >> 32);
d = (int64_t)d16[2 * i + 3];
t1 += (b & 0xffff) << 16;
t += (b >> 16) + (t1 >> 32);
i32[i] = t1 & 0xffffffff;
t1 = t;
a = c;
b = d;
}
t1 += a & 0xffffffff;
t = (a >> 32);
t1 += (b & 0xffff) << 16;
i32[i] = t1 & 0xffffffff;
}
void
conv_i32_to_d32(double *d32, uint32_t *i32, int len)
{
int i;
#pragma pipeloop(0)
for (i = 0; i < len; i++)
d32[i] = (double)(i32[i]);
}
void
conv_i32_to_d16(double *d16, uint32_t *i32, int len)
{
int i;
uint32_t a;
#pragma pipeloop(0)
for (i = 0; i < len; i++) {
a = i32[i];
d16[2 * i] = (double)(a & 0xffff);
d16[2 * i + 1] = (double)(a >> 16);
}
}
#ifdef RF_INLINE_MACROS
void
i16_to_d16_and_d32x4(const double *, /* 1/(2^16) */
const double *, /* 2^16 */
const double *, /* 0 */
double *, /* result16 */
double *, /* result32 */
float *); /* source - should be unsigned int* */
/* converted to float* */
#else
static void
i16_to_d16_and_d32x4(const double *dummy1, /* 1/(2^16) */
const double *dummy2, /* 2^16 */
const double *dummy3, /* 0 */
double *result16,
double *result32,
float *src) /* source - should be unsigned int* */
/* converted to float* */
{
uint32_t *i32;
uint32_t a, b, c, d;
i32 = (uint32_t *)src;
a = i32[0];
b = i32[1];
c = i32[2];
d = i32[3];
result16[0] = (double)(a & 0xffff);
result16[1] = (double)(a >> 16);
result32[0] = (double)a;
result16[2] = (double)(b & 0xffff);
result16[3] = (double)(b >> 16);
result32[1] = (double)b;
result16[4] = (double)(c & 0xffff);
result16[5] = (double)(c >> 16);
result32[2] = (double)c;
result16[6] = (double)(d & 0xffff);
result16[7] = (double)(d >> 16);
result32[3] = (double)d;
}
#endif
void
conv_i32_to_d32_and_d16(double *d32, double *d16, uint32_t *i32, int len)
{
int i;
uint32_t a;
#pragma pipeloop(0)
for (i = 0; i < len - 3; i += 4) {
i16_to_d16_and_d32x4(&TwoToMinus16, &TwoTo16, &Zero,
&(d16[2*i]), &(d32[i]),
(float *)(&(i32[i])));
}
for (; i < len; i++) {
a = i32[i];
d32[i] = (double)(i32[i]);
d16[2 * i] = (double)(a & 0xffff);
d16[2 * i + 1] = (double)(a >> 16);
}
}
static void
adjust_montf_result(uint32_t *i32, uint32_t *nint, int len)
{
int64_t acc;
int i;
if (i32[len] > 0)
i = -1;
else {
for (i = len - 1; i >= 0; i--) {
if (i32[i] != nint[i]) break;
}
}
if ((i < 0) || (i32[i] > nint[i])) {
acc = 0;
for (i = 0; i < len; i++) {
acc = acc + (uint64_t)(i32[i]) - (uint64_t)(nint[i]);
i32[i] = acc & 0xffffffff;
acc = acc >> 32;
}
}
}
/*
* the lengths of the input arrays should be at least the following:
* result[nlen+1], dm1[nlen], dm2[2*nlen+1], dt[4*nlen+2], dn[nlen], nint[nlen]
* all of them should be different from one another
*/
void mont_mulf_noconv(uint32_t *result,
double *dm1, double *dm2, double *dt,
double *dn, uint32_t *nint,
int nlen, double dn0)
{
int i, j, jj;
double digit, m2j, a, b;
double *pdm1, *pdm2, *pdn, *pdtj, pdn_0, pdm1_0;
pdm1 = &(dm1[0]);
pdm2 = &(dm2[0]);
pdn = &(dn[0]);
pdm2[2 * nlen] = Zero;
if (nlen != 16) {
for (i = 0; i < 4 * nlen + 2; i++)
dt[i] = Zero;
a = dt[0] = pdm1[0] * pdm2[0];
digit = mod(lower32(a, Zero) * dn0, TwoToMinus16, TwoTo16);
pdtj = &(dt[0]);
for (j = jj = 0; j < 2 * nlen; j++, jj++, pdtj++) {
m2j = pdm2[j];
a = pdtj[0] + pdn[0] * digit;
b = pdtj[1] + pdm1[0] * pdm2[j + 1] + a * TwoToMinus16;
pdtj[1] = b;
#pragma pipeloop(0)
for (i = 1; i < nlen; i++) {
pdtj[2 * i] += pdm1[i] * m2j + pdn[i] * digit;
}
if (jj == 30) {
cleanup(dt, j / 2 + 1, 2 * nlen + 1);
jj = 0;
}
digit = mod(lower32(b, Zero) * dn0,
TwoToMinus16, TwoTo16);
}
} else {
a = dt[0] = pdm1[0] * pdm2[0];
dt[65] = dt[64] = dt[63] = dt[62] = dt[61] = dt[60] =
dt[59] = dt[58] = dt[57] = dt[56] = dt[55] =
dt[54] = dt[53] = dt[52] = dt[51] = dt[50] =
dt[49] = dt[48] = dt[47] = dt[46] = dt[45] =
dt[44] = dt[43] = dt[42] = dt[41] = dt[40] =
dt[39] = dt[38] = dt[37] = dt[36] = dt[35] =
dt[34] = dt[33] = dt[32] = dt[31] = dt[30] =
dt[29] = dt[28] = dt[27] = dt[26] = dt[25] =
dt[24] = dt[23] = dt[22] = dt[21] = dt[20] =
dt[19] = dt[18] = dt[17] = dt[16] = dt[15] =
dt[14] = dt[13] = dt[12] = dt[11] = dt[10] =
dt[9] = dt[8] = dt[7] = dt[6] = dt[5] = dt[4] =
dt[3] = dt[2] = dt[1] = Zero;
pdn_0 = pdn[0];
pdm1_0 = pdm1[0];
digit = mod(lower32(a, Zero) * dn0, TwoToMinus16, TwoTo16);
pdtj = &(dt[0]);
for (j = 0; j < 32; j++, pdtj++) {
m2j = pdm2[j];
a = pdtj[0] + pdn_0 * digit;
b = pdtj[1] + pdm1_0 * pdm2[j + 1] + a * TwoToMinus16;
pdtj[1] = b;
pdtj[2] += pdm1[1] *m2j + pdn[1] * digit;
pdtj[4] += pdm1[2] *m2j + pdn[2] * digit;
pdtj[6] += pdm1[3] *m2j + pdn[3] * digit;
pdtj[8] += pdm1[4] *m2j + pdn[4] * digit;
pdtj[10] += pdm1[5] *m2j + pdn[5] * digit;
pdtj[12] += pdm1[6] *m2j + pdn[6] * digit;
pdtj[14] += pdm1[7] *m2j + pdn[7] * digit;
pdtj[16] += pdm1[8] *m2j + pdn[8] * digit;
pdtj[18] += pdm1[9] *m2j + pdn[9] * digit;
pdtj[20] += pdm1[10] *m2j + pdn[10] * digit;
pdtj[22] += pdm1[11] *m2j + pdn[11] * digit;
pdtj[24] += pdm1[12] *m2j + pdn[12] * digit;
pdtj[26] += pdm1[13] *m2j + pdn[13] * digit;
pdtj[28] += pdm1[14] *m2j + pdn[14] * digit;
pdtj[30] += pdm1[15] *m2j + pdn[15] * digit;
/* no need for cleenup, cannot overflow */
digit = mod(lower32(b, Zero) * dn0,
TwoToMinus16, TwoTo16);
}
}
conv_d16_to_i32(result, dt + 2 * nlen, (int64_t *)dt, nlen + 1);
adjust_montf_result(result, nint, nlen);
}