1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
|
/*
* Copyright © 2015 Red Hat
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*
* Authors:
* Rob Clark <robclark@freedesktop.org>
*/
#include "nir.h"
#include "nir_builder.h"
/* ported from LLVM's AMDGPUTargetLowering::LowerUDIVREM */
static nir_def *
emit_udiv(nir_builder *bld, nir_def *numer, nir_def *denom, bool modulo)
{
nir_def *rcp = nir_frcp(bld, nir_u2f32(bld, denom));
rcp = nir_f2u32(bld, nir_fmul_imm(bld, rcp, 4294966784.0));
nir_def *neg_rcp_times_denom =
nir_imul(bld, rcp, nir_ineg(bld, denom));
rcp = nir_iadd(bld, rcp, nir_umul_high(bld, rcp, neg_rcp_times_denom));
/* Get initial estimate for quotient/remainder, then refine the estimate
* in two iterations after */
nir_def *quotient = nir_umul_high(bld, numer, rcp);
nir_def *num_s_remainder = nir_imul(bld, quotient, denom);
nir_def *remainder = nir_isub(bld, numer, num_s_remainder);
/* First refinement step */
nir_def *remainder_ge_den = nir_uge(bld, remainder, denom);
if (!modulo) {
quotient = nir_bcsel(bld, remainder_ge_den,
nir_iadd_imm(bld, quotient, 1), quotient);
}
remainder = nir_bcsel(bld, remainder_ge_den,
nir_isub(bld, remainder, denom), remainder);
/* Second refinement step */
remainder_ge_den = nir_uge(bld, remainder, denom);
if (modulo) {
return nir_bcsel(bld, remainder_ge_den, nir_isub(bld, remainder, denom),
remainder);
} else {
return nir_bcsel(bld, remainder_ge_den, nir_iadd_imm(bld, quotient, 1),
quotient);
}
}
/* ported from LLVM's AMDGPUTargetLowering::LowerSDIVREM */
static nir_def *
emit_idiv(nir_builder *bld, nir_def *numer, nir_def *denom, nir_op op)
{
nir_def *lhs = nir_iabs(bld, numer);
nir_def *rhs = nir_iabs(bld, denom);
if (op == nir_op_idiv) {
/* We want (numer < 0) ^ (denom < 0). This is the XOR of the sign bits,
* and since XOR is bitwise, that's the sign bit of the XOR.
*/
nir_def *d_sign = nir_ilt_imm(bld, nir_ixor(bld, numer, denom), 0);
nir_def *res = emit_udiv(bld, lhs, rhs, false);
return nir_bcsel(bld, d_sign, nir_ineg(bld, res), res);
} else {
nir_def *lh_sign = nir_ilt_imm(bld, numer, 0);
nir_def *rh_sign = nir_ilt_imm(bld, denom, 0);
nir_def *res = emit_udiv(bld, lhs, rhs, true);
res = nir_bcsel(bld, lh_sign, nir_ineg(bld, res), res);
if (op == nir_op_imod) {
nir_def *cond = nir_ieq_imm(bld, res, 0);
cond = nir_ior(bld, nir_ieq(bld, lh_sign, rh_sign), cond);
res = nir_bcsel(bld, cond, res, nir_iadd(bld, res, denom));
}
return res;
}
}
static nir_def *
convert_instr_small(nir_builder *b, nir_op op,
nir_def *numer, nir_def *denom,
const nir_lower_idiv_options *options)
{
unsigned sz = numer->bit_size;
nir_alu_type int_type = nir_op_infos[op].output_type | sz;
nir_alu_type float_type = nir_type_float | (options->allow_fp16 ? sz * 2 : 32);
nir_def *p = nir_type_convert(b, numer, int_type, float_type, nir_rounding_mode_undef);
nir_def *q = nir_type_convert(b, denom, int_type, float_type, nir_rounding_mode_undef);
/* Take 1/q but offset mantissa by 1 to correct for rounding. This is
* needed for correct results and has been checked exhaustively for
* all pairs of 16-bit integers */
nir_def *rcp = nir_iadd_imm(b, nir_frcp(b, q), 1);
/* Divide by multiplying by adjusted reciprocal */
nir_def *res = nir_fmul(b, p, rcp);
/* Convert back to integer space with rounding inferred by type */
res = nir_type_convert(b, res, float_type, int_type, nir_rounding_mode_undef);
/* Get remainder given the quotient */
if (op == nir_op_umod || op == nir_op_imod || op == nir_op_irem)
res = nir_isub(b, numer, nir_imul(b, denom, res));
/* Adjust for sign, see constant folding definition */
if (op == nir_op_imod) {
nir_def *zero = nir_imm_zero(b, 1, sz);
nir_def *diff_sign =
nir_ine(b, nir_ige(b, numer, zero), nir_ige(b, denom, zero));
nir_def *adjust = nir_iand(b, diff_sign, nir_ine(b, res, zero));
res = nir_iadd(b, res, nir_bcsel(b, adjust, denom, zero));
}
return res;
}
static nir_def *
lower_idiv(nir_builder *b, nir_instr *instr, void *_data)
{
const nir_lower_idiv_options *options = _data;
nir_alu_instr *alu = nir_instr_as_alu(instr);
nir_def *numer = nir_ssa_for_alu_src(b, alu, 0);
nir_def *denom = nir_ssa_for_alu_src(b, alu, 1);
b->exact = true;
if (numer->bit_size < 32)
return convert_instr_small(b, alu->op, numer, denom, options);
else if (alu->op == nir_op_udiv || alu->op == nir_op_umod)
return emit_udiv(b, numer, denom, alu->op == nir_op_umod);
else
return emit_idiv(b, numer, denom, alu->op);
}
static bool
inst_is_idiv(const nir_instr *instr, UNUSED const void *_state)
{
if (instr->type != nir_instr_type_alu)
return false;
nir_alu_instr *alu = nir_instr_as_alu(instr);
if (alu->def.bit_size > 32)
return false;
switch (alu->op) {
case nir_op_idiv:
case nir_op_udiv:
case nir_op_imod:
case nir_op_umod:
case nir_op_irem:
return true;
default:
return false;
}
}
bool
nir_lower_idiv(nir_shader *shader, const nir_lower_idiv_options *options)
{
return nir_shader_lower_instructions(shader,
inst_is_idiv,
lower_idiv,
(void *)options);
}
|
