• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2018 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  */
23 
24 /* Imported from:
25  *   https://raw.githubusercontent.com/ridiculousfish/libdivide/master/divide_by_constants_codegen_reference.c
26  * Paper:
27  *   http://ridiculousfish.com/files/faster_unsigned_division_by_constants.pdf
28  *
29  * The author, ridiculous_fish, wrote:
30  *
31  *  ''Reference implementations of computing and using the "magic number"
32  *    approach to dividing by constants, including codegen instructions.
33  *    The unsigned division incorporates the "round down" optimization per
34  *    ridiculous_fish.
35  *
36  *    This is free and unencumbered software. Any copyright is dedicated
37  *    to the Public Domain.''
38  */
39 
40 #include "fast_idiv_by_const.h"
41 #include "u_math.h"
42 #include <limits.h>
43 #include <assert.h>
44 
45 struct util_fast_udiv_info
util_compute_fast_udiv_info(uint64_t D,unsigned num_bits,unsigned UINT_BITS)46 util_compute_fast_udiv_info(uint64_t D, unsigned num_bits, unsigned UINT_BITS)
47 {
48    /* The numerator must fit in a uint64_t */
49    assert(num_bits > 0 && num_bits <= UINT_BITS);
50    assert(D != 0);
51 
52    /* The eventual result */
53    struct util_fast_udiv_info result;
54 
55    if (util_is_power_of_two_or_zero64(D)) {
56       unsigned div_shift = util_logbase2_64(D);
57 
58       if (div_shift) {
59          /* Dividing by a power of two. */
60          result.multiplier = 1ull << (UINT_BITS - div_shift);
61          result.pre_shift = 0;
62          result.post_shift = 0;
63          result.increment = 0;
64          return result;
65       } else {
66          /* Dividing by 1. */
67          /* Assuming: floor((num + 1) * (2^32 - 1) / 2^32) = num */
68          result.multiplier = UINT_BITS == 64 ? UINT64_MAX :
69                                                (1ull << UINT_BITS) - 1;
70          result.pre_shift = 0;
71          result.post_shift = 0;
72          result.increment = 1;
73          return result;
74       }
75    }
76 
77    /* The extra shift implicit in the difference between UINT_BITS and num_bits
78     */
79    const unsigned extra_shift = UINT_BITS - num_bits;
80 
81    /* The initial power of 2 is one less than the first one that can possibly
82     * work.
83     */
84    const uint64_t initial_power_of_2 = (uint64_t)1 << (UINT_BITS-1);
85 
86    /* The remainder and quotient of our power of 2 divided by d */
87    uint64_t quotient = initial_power_of_2 / D;
88    uint64_t remainder = initial_power_of_2 % D;
89 
90    /* ceil(log_2 D) */
91    unsigned ceil_log_2_D;
92 
93    /* The magic info for the variant "round down" algorithm */
94    uint64_t down_multiplier = 0;
95    unsigned down_exponent = 0;
96    int has_magic_down = 0;
97 
98    /* Compute ceil(log_2 D) */
99    ceil_log_2_D = 0;
100    uint64_t tmp;
101    for (tmp = D; tmp > 0; tmp >>= 1)
102       ceil_log_2_D += 1;
103 
104 
105    /* Begin a loop that increments the exponent, until we find a power of 2
106     * that works.
107     */
108    unsigned exponent;
109    for (exponent = 0; ; exponent++) {
110       /* Quotient and remainder is from previous exponent; compute it for this
111        * exponent.
112        */
113       if (remainder >= D - remainder) {
114          /* Doubling remainder will wrap around D */
115          quotient = quotient * 2 + 1;
116          remainder = remainder * 2 - D;
117       } else {
118          /* Remainder will not wrap */
119          quotient = quotient * 2;
120          remainder = remainder * 2;
121       }
122 
123       /* We're done if this exponent works for the round_up algorithm.
124        * Note that exponent may be larger than the maximum shift supported,
125        * so the check for >= ceil_log_2_D is critical.
126        */
127       if ((exponent + extra_shift >= ceil_log_2_D) ||
128           (D - remainder) <= ((uint64_t)1 << (exponent + extra_shift)))
129          break;
130 
131       /* Set magic_down if we have not set it yet and this exponent works for
132        * the round_down algorithm
133        */
134       if (!has_magic_down &&
135           remainder <= ((uint64_t)1 << (exponent + extra_shift))) {
136          has_magic_down = 1;
137          down_multiplier = quotient;
138          down_exponent = exponent;
139       }
140    }
141 
142    if (exponent < ceil_log_2_D) {
143       /* magic_up is efficient */
144       result.multiplier = quotient + 1;
145       result.pre_shift = 0;
146       result.post_shift = exponent;
147       result.increment = 0;
148    } else if (D & 1) {
149       /* Odd divisor, so use magic_down, which must have been set */
150       assert(has_magic_down);
151       result.multiplier = down_multiplier;
152       result.pre_shift = 0;
153       result.post_shift = down_exponent;
154       result.increment = 1;
155    } else {
156       /* Even divisor, so use a prefix-shifted dividend */
157       unsigned pre_shift = 0;
158       uint64_t shifted_D = D;
159       while ((shifted_D & 1) == 0) {
160          shifted_D >>= 1;
161          pre_shift += 1;
162       }
163       result = util_compute_fast_udiv_info(shifted_D, num_bits - pre_shift,
164                                            UINT_BITS);
165       /* expect no increment or pre_shift in this path */
166       assert(result.increment == 0 && result.pre_shift == 0);
167       result.pre_shift = pre_shift;
168    }
169    return result;
170 }
171 
172 static inline int64_t
sign_extend(int64_t x,unsigned SINT_BITS)173 sign_extend(int64_t x, unsigned SINT_BITS)
174 {
175    return (int64_t)((uint64_t)x << (64 - SINT_BITS)) >> (64 - SINT_BITS);
176 }
177 
178 struct util_fast_sdiv_info
util_compute_fast_sdiv_info(int64_t D,unsigned SINT_BITS)179 util_compute_fast_sdiv_info(int64_t D, unsigned SINT_BITS)
180 {
181    /* D must not be zero. */
182    assert(D != 0);
183    /* The result is not correct for these divisors. */
184    assert(D != 1 && D != -1);
185 
186    /* Our result */
187    struct util_fast_sdiv_info result;
188 
189    /* Absolute value of D (we know D is not the most negative value since
190     * that's a power of 2)
191     */
192    const uint64_t abs_d = (D < 0 ? -D : D);
193 
194    /* The initial power of 2 is one less than the first one that can possibly
195     * work */
196    /* "two31" in Warren */
197    unsigned exponent = SINT_BITS - 1;
198    const uint64_t initial_power_of_2 = (uint64_t)1 << exponent;
199 
200    /* Compute the absolute value of our "test numerator,"
201     * which is the largest dividend whose remainder with d is d-1.
202     * This is called anc in Warren.
203     */
204    const uint64_t tmp = initial_power_of_2 + (D < 0);
205    const uint64_t abs_test_numer = tmp - 1 - tmp % abs_d;
206 
207    /* Initialize our quotients and remainders (q1, r1, q2, r2 in Warren) */
208    uint64_t quotient1 = initial_power_of_2 / abs_test_numer;
209    uint64_t remainder1 = initial_power_of_2 % abs_test_numer;
210    uint64_t quotient2 = initial_power_of_2 / abs_d;
211    uint64_t remainder2 = initial_power_of_2 % abs_d;
212    uint64_t delta;
213 
214    /* Begin our loop */
215    do {
216       /* Update the exponent */
217       exponent++;
218 
219       /* Update quotient1 and remainder1 */
220       quotient1 *= 2;
221       remainder1 *= 2;
222       if (remainder1 >= abs_test_numer) {
223          quotient1 += 1;
224          remainder1 -= abs_test_numer;
225       }
226 
227       /* Update quotient2 and remainder2 */
228       quotient2 *= 2;
229       remainder2 *= 2;
230       if (remainder2 >= abs_d) {
231          quotient2 += 1;
232          remainder2 -= abs_d;
233       }
234 
235       /* Keep going as long as (2**exponent) / abs_d <= delta */
236       delta = abs_d - remainder2;
237    } while (quotient1 < delta || (quotient1 == delta && remainder1 == 0));
238 
239    result.multiplier = sign_extend(quotient2 + 1, SINT_BITS);
240    if (D < 0) result.multiplier = -result.multiplier;
241    result.shift = exponent - SINT_BITS;
242    return result;
243 }
244