0
|
1 /* Copyright (C) 2009 Free Software Foundation, Inc.
|
|
2
|
|
3 This file is free software; you can redistribute it and/or modify it under
|
|
4 the terms of the GNU General Public License as published by the Free
|
|
5 Software Foundation; either version 3 of the License, or (at your option)
|
|
6 any later version.
|
|
7
|
|
8 This file is distributed in the hope that it will be useful, but WITHOUT
|
|
9 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
|
10 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
11 for more details.
|
|
12
|
|
13 Under Section 7 of GPL version 3, you are granted additional
|
|
14 permissions described in the GCC Runtime Library Exception, version
|
|
15 3.1, as published by the Free Software Foundation.
|
|
16
|
|
17 You should have received a copy of the GNU General Public License and
|
|
18 a copy of the GCC Runtime Library Exception along with this program;
|
|
19 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
|
20 <http://www.gnu.org/licenses/>. */
|
|
21
|
|
22 #include <spu_intrinsics.h>
|
|
23
|
|
24 vector double __divv2df3 (vector double a_in, vector double b_in);
|
|
25
|
|
26 /* __divv2df3 divides the vector dividend a by the vector divisor b and
|
|
27 returns the resulting vector quotient. Maximum error about 0.5 ulp
|
|
28 over entire double range including denorms, compared to true result
|
|
29 in round-to-nearest rounding mode. Handles Inf or NaN operands and
|
|
30 results correctly. */
|
|
31
|
|
32 vector double
|
|
33 __divv2df3 (vector double a_in, vector double b_in)
|
|
34 {
|
|
35 /* Variables */
|
|
36 vec_int4 exp, exp_bias;
|
|
37 vec_uint4 no_underflow, overflow;
|
|
38 vec_float4 mant_bf, inv_bf;
|
|
39 vec_ullong2 exp_a, exp_b;
|
|
40 vec_ullong2 a_nan, a_zero, a_inf, a_denorm, a_denorm0;
|
|
41 vec_ullong2 b_nan, b_zero, b_inf, b_denorm, b_denorm0;
|
|
42 vec_ullong2 nan;
|
|
43 vec_uint4 a_exp, b_exp;
|
|
44 vec_ullong2 a_mant_0, b_mant_0;
|
|
45 vec_ullong2 a_exp_1s, b_exp_1s;
|
|
46 vec_ullong2 sign_exp_mask;
|
|
47
|
|
48 vec_double2 a, b;
|
|
49 vec_double2 mant_a, mant_b, inv_b, q0, q1, q2, mult;
|
|
50
|
|
51 /* Constants */
|
|
52 vec_uint4 exp_mask_u32 = spu_splats((unsigned int)0x7FF00000);
|
|
53 vec_uchar16 splat_hi = (vec_uchar16){0,1,2,3, 0,1,2,3, 8, 9,10,11, 8,9,10,11};
|
|
54 vec_uchar16 swap_32 = (vec_uchar16){4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11};
|
|
55 vec_ullong2 exp_mask = spu_splats(0x7FF0000000000000ULL);
|
|
56 vec_ullong2 sign_mask = spu_splats(0x8000000000000000ULL);
|
|
57 vec_float4 onef = spu_splats(1.0f);
|
|
58 vec_double2 one = spu_splats(1.0);
|
|
59 vec_double2 exp_53 = (vec_double2)spu_splats(0x0350000000000000ULL);
|
|
60
|
|
61 sign_exp_mask = spu_or(sign_mask, exp_mask);
|
|
62
|
|
63 /* Extract the floating point components from each of the operands including
|
|
64 * exponent and mantissa.
|
|
65 */
|
|
66 a_exp = (vec_uint4)spu_and((vec_uint4)a_in, exp_mask_u32);
|
|
67 a_exp = spu_shuffle(a_exp, a_exp, splat_hi);
|
|
68 b_exp = (vec_uint4)spu_and((vec_uint4)b_in, exp_mask_u32);
|
|
69 b_exp = spu_shuffle(b_exp, b_exp, splat_hi);
|
|
70
|
|
71 a_mant_0 = (vec_ullong2)spu_cmpeq((vec_uint4)spu_andc((vec_ullong2)a_in, sign_exp_mask), 0);
|
|
72 a_mant_0 = spu_and(a_mant_0, spu_shuffle(a_mant_0, a_mant_0, swap_32));
|
|
73
|
|
74 b_mant_0 = (vec_ullong2)spu_cmpeq((vec_uint4)spu_andc((vec_ullong2)b_in, sign_exp_mask), 0);
|
|
75 b_mant_0 = spu_and(b_mant_0, spu_shuffle(b_mant_0, b_mant_0, swap_32));
|
|
76
|
|
77 a_exp_1s = (vec_ullong2)spu_cmpeq(a_exp, exp_mask_u32);
|
|
78 b_exp_1s = (vec_ullong2)spu_cmpeq(b_exp, exp_mask_u32);
|
|
79
|
|
80 /* Identify all possible special values that must be accomodated including:
|
|
81 * +-denorm, +-0, +-infinity, and NaNs.
|
|
82 */
|
|
83 a_denorm0= (vec_ullong2)spu_cmpeq(a_exp, 0);
|
|
84 a_nan = spu_andc(a_exp_1s, a_mant_0);
|
|
85 a_zero = spu_and (a_denorm0, a_mant_0);
|
|
86 a_inf = spu_and (a_exp_1s, a_mant_0);
|
|
87 a_denorm = spu_andc(a_denorm0, a_zero);
|
|
88
|
|
89 b_denorm0= (vec_ullong2)spu_cmpeq(b_exp, 0);
|
|
90 b_nan = spu_andc(b_exp_1s, b_mant_0);
|
|
91 b_zero = spu_and (b_denorm0, b_mant_0);
|
|
92 b_inf = spu_and (b_exp_1s, b_mant_0);
|
|
93 b_denorm = spu_andc(b_denorm0, b_zero);
|
|
94
|
|
95 /* Scale denorm inputs to into normalized numbers by conditionally scaling the
|
|
96 * input parameters.
|
|
97 */
|
|
98 a = spu_sub(spu_or(a_in, exp_53), spu_sel(exp_53, a_in, sign_mask));
|
|
99 a = spu_sel(a_in, a, a_denorm);
|
|
100
|
|
101 b = spu_sub(spu_or(b_in, exp_53), spu_sel(exp_53, b_in, sign_mask));
|
|
102 b = spu_sel(b_in, b, b_denorm);
|
|
103
|
|
104 /* Extract the divisor and dividend exponent and force parameters into the signed
|
|
105 * range [1.0,2.0) or [-1.0,2.0).
|
|
106 */
|
|
107 exp_a = spu_and((vec_ullong2)a, exp_mask);
|
|
108 exp_b = spu_and((vec_ullong2)b, exp_mask);
|
|
109
|
|
110 mant_a = spu_sel(a, one, (vec_ullong2)exp_mask);
|
|
111 mant_b = spu_sel(b, one, (vec_ullong2)exp_mask);
|
|
112
|
|
113 /* Approximate the single reciprocal of b by using
|
|
114 * the single precision reciprocal estimate followed by one
|
|
115 * single precision iteration of Newton-Raphson.
|
|
116 */
|
|
117 mant_bf = spu_roundtf(mant_b);
|
|
118 inv_bf = spu_re(mant_bf);
|
|
119 inv_bf = spu_madd(spu_nmsub(mant_bf, inv_bf, onef), inv_bf, inv_bf);
|
|
120
|
|
121 /* Perform 2 more Newton-Raphson iterations in double precision. The
|
|
122 * result (q1) is in the range (0.5, 2.0).
|
|
123 */
|
|
124 inv_b = spu_extend(inv_bf);
|
|
125 inv_b = spu_madd(spu_nmsub(mant_b, inv_b, one), inv_b, inv_b);
|
|
126 q0 = spu_mul(mant_a, inv_b);
|
|
127 q1 = spu_madd(spu_nmsub(mant_b, q0, mant_a), inv_b, q0);
|
|
128
|
|
129 /* Determine the exponent correction factor that must be applied
|
|
130 * to q1 by taking into account the exponent of the normalized inputs
|
|
131 * and the scale factors that were applied to normalize them.
|
|
132 */
|
|
133 exp = spu_rlmaska(spu_sub((vec_int4)exp_a, (vec_int4)exp_b), -20);
|
|
134 exp = spu_add(exp, (vec_int4)spu_add(spu_and((vec_int4)a_denorm, -0x34), spu_and((vec_int4)b_denorm, 0x34)));
|
|
135
|
|
136 /* Bias the quotient exponent depending on the sign of the exponent correction
|
|
137 * factor so that a single multiplier will ensure the entire double precision
|
|
138 * domain (including denorms) can be achieved.
|
|
139 *
|
|
140 * exp bias q1 adjust exp
|
|
141 * ===== ======== ==========
|
|
142 * positive 2^+65 -65
|
|
143 * negative 2^-64 +64
|
|
144 */
|
|
145 exp_bias = spu_xor(spu_rlmaska(exp, -31), 64);
|
|
146 exp = spu_sub(exp, exp_bias);
|
|
147
|
|
148 q1 = spu_sel(q1, (vec_double2)spu_add((vec_int4)q1, spu_sl(exp_bias, 20)), exp_mask);
|
|
149
|
|
150 /* Compute a multiplier (mult) to applied to the quotient (q1) to produce the
|
|
151 * expected result. On overflow, clamp the multiplier to the maximum non-infinite
|
|
152 * number in case the rounding mode is not round-to-nearest.
|
|
153 */
|
|
154 exp = spu_add(exp, 0x3FF);
|
|
155 no_underflow = spu_cmpgt(exp, 0);
|
|
156 overflow = spu_cmpgt(exp, 0x7FE);
|
|
157 exp = spu_and(spu_sl(exp, 20), (vec_int4)no_underflow);
|
|
158 exp = spu_and(exp, (vec_int4)exp_mask);
|
|
159
|
|
160 mult = spu_sel((vec_double2)exp, (vec_double2)(spu_add((vec_uint4)exp_mask, -1)), (vec_ullong2)overflow);
|
|
161
|
|
162 /* Handle special value conditions. These include:
|
|
163 *
|
|
164 * 1) IF either operand is a NaN OR both operands are 0 or INFINITY THEN a NaN
|
|
165 * results.
|
|
166 * 2) ELSE IF the dividend is an INFINITY OR the divisor is 0 THEN a INFINITY results.
|
|
167 * 3) ELSE IF the dividend is 0 OR the divisor is INFINITY THEN a 0 results.
|
|
168 */
|
|
169 mult = spu_andc(mult, (vec_double2)spu_or(a_zero, b_inf));
|
|
170 mult = spu_sel(mult, (vec_double2)exp_mask, spu_or(a_inf, b_zero));
|
|
171
|
|
172 nan = spu_or(a_nan, b_nan);
|
|
173 nan = spu_or(nan, spu_and(a_zero, b_zero));
|
|
174 nan = spu_or(nan, spu_and(a_inf, b_inf));
|
|
175
|
|
176 mult = spu_or(mult, (vec_double2)nan);
|
|
177
|
|
178 /* Scale the final quotient */
|
|
179
|
|
180 q2 = spu_mul(q1, mult);
|
|
181
|
|
182 return (q2);
|
|
183 }
|
|
184
|
|
185
|
|
186 /* We use the same function for vector and scalar division. Provide the
|
|
187 scalar entry point as an alias. */
|
|
188 double __divdf3 (double a, double b)
|
|
189 __attribute__ ((__alias__ ("__divv2df3")));
|
|
190
|
|
191 /* Some toolchain builds used the __fast_divdf3 name for this helper function.
|
|
192 Provide this as another alternate entry point for compatibility. */
|
|
193 double __fast_divdf3 (double a, double b)
|
|
194 __attribute__ ((__alias__ ("__divv2df3")));
|
|
195
|