xref: /aosp_15_r20/external/mesa3d/src/compiler/nir/nir_lower_alu.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright © 2010 Intel Corporation
3  * Copyright © 2018 Broadcom
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22  * DEALINGS IN THE SOFTWARE.
23  */
24 
25 #include "nir.h"
26 #include "nir_builder.h"
27 
28 /** nir_lower_alu.c
29  *
30  * NIR's home for miscellaneous ALU operation lowering implementations.
31  *
32  * Most NIR ALU lowering occurs in nir_opt_algebraic.py, since it's generally
33  * easy to write them there.  However, if terms appear multiple times in the
34  * lowered code, it can get very verbose and cause a lot of work for CSE, so
35  * it may end up being easier to write out in C code.
36  *
37  * The shader must be in SSA for this pass.
38  */
39 
40 static bool
lower_alu_instr(nir_builder * b,nir_alu_instr * instr,UNUSED void * cb_data)41 lower_alu_instr(nir_builder *b, nir_alu_instr *instr, UNUSED void *cb_data)
42 {
43    nir_def *lowered = NULL;
44 
45    b->cursor = nir_before_instr(&instr->instr);
46    b->exact = instr->exact;
47    b->fp_fast_math = instr->fp_fast_math;
48 
49    switch (instr->op) {
50    case nir_op_bitfield_reverse:
51       if (b->shader->options->lower_bitfield_reverse) {
52          /* For more details, see:
53           *
54           * http://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
55           */
56          nir_def *c1 = nir_imm_int(b, 1);
57          nir_def *c2 = nir_imm_int(b, 2);
58          nir_def *c4 = nir_imm_int(b, 4);
59          nir_def *c8 = nir_imm_int(b, 8);
60          nir_def *c16 = nir_imm_int(b, 16);
61          nir_def *c33333333 = nir_imm_int(b, 0x33333333);
62          nir_def *c55555555 = nir_imm_int(b, 0x55555555);
63          nir_def *c0f0f0f0f = nir_imm_int(b, 0x0f0f0f0f);
64          nir_def *c00ff00ff = nir_imm_int(b, 0x00ff00ff);
65 
66          lowered = nir_ssa_for_alu_src(b, instr, 0);
67 
68          /* Swap odd and even bits. */
69          lowered = nir_ior(b,
70                            nir_iand(b, nir_ushr(b, lowered, c1), c55555555),
71                            nir_ishl(b, nir_iand(b, lowered, c55555555), c1));
72 
73          /* Swap consecutive pairs. */
74          lowered = nir_ior(b,
75                            nir_iand(b, nir_ushr(b, lowered, c2), c33333333),
76                            nir_ishl(b, nir_iand(b, lowered, c33333333), c2));
77 
78          /* Swap nibbles. */
79          lowered = nir_ior(b,
80                            nir_iand(b, nir_ushr(b, lowered, c4), c0f0f0f0f),
81                            nir_ishl(b, nir_iand(b, lowered, c0f0f0f0f), c4));
82 
83          /* Swap bytes. */
84          lowered = nir_ior(b,
85                            nir_iand(b, nir_ushr(b, lowered, c8), c00ff00ff),
86                            nir_ishl(b, nir_iand(b, lowered, c00ff00ff), c8));
87 
88          lowered = nir_ior(b,
89                            nir_ushr(b, lowered, c16),
90                            nir_ishl(b, lowered, c16));
91       }
92       break;
93 
94    case nir_op_bit_count:
95       if (b->shader->options->lower_bit_count) {
96          /* For more details, see:
97           *
98           * http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
99           */
100 
101          lowered = nir_ssa_for_alu_src(b, instr, 0);
102          unsigned bit_size = lowered->bit_size;
103 
104          nir_def *c1 = nir_imm_int(b, 1);
105          nir_def *c2 = nir_imm_int(b, 2);
106          nir_def *c4 = nir_imm_int(b, 4);
107          nir_def *cshift = nir_imm_int(b, bit_size - 8);
108          nir_def *c33333333 = nir_imm_intN_t(b, 0x33333333, bit_size);
109          nir_def *c55555555 = nir_imm_intN_t(b, 0x55555555, bit_size);
110          nir_def *c0f0f0f0f = nir_imm_intN_t(b, 0x0f0f0f0f, bit_size);
111          nir_def *c01010101 = nir_imm_intN_t(b, 0x01010101, bit_size);
112 
113 
114          lowered = nir_isub(b, lowered,
115                             nir_iand(b, nir_ushr(b, lowered, c1), c55555555));
116 
117          lowered = nir_iadd(b,
118                             nir_iand(b, lowered, c33333333),
119                             nir_iand(b, nir_ushr(b, lowered, c2), c33333333));
120 
121          lowered = nir_ushr(b,
122                             nir_imul(b,
123                                      nir_iand(b,
124                                               nir_iadd(b,
125                                                        lowered,
126                                                        nir_ushr(b, lowered, c4)),
127                                               c0f0f0f0f),
128                                      c01010101),
129                             cshift);
130 
131          lowered = nir_u2u32(b, lowered);
132       }
133       break;
134 
135    case nir_op_imul_high:
136    case nir_op_umul_high:
137       if (b->shader->options->lower_mul_high) {
138          nir_def *src0 = nir_ssa_for_alu_src(b, instr, 0);
139          nir_def *src1 = nir_ssa_for_alu_src(b, instr, 1);
140          if (src0->bit_size < 32) {
141             /* Just do the math in 32-bit space and shift the result */
142             nir_alu_type base_type = nir_op_infos[instr->op].output_type;
143 
144             nir_def *src0_32 = nir_type_convert(b, src0, base_type, base_type | 32, nir_rounding_mode_undef);
145             nir_def *src1_32 = nir_type_convert(b, src1, base_type, base_type | 32, nir_rounding_mode_undef);
146             nir_def *dest_32 = nir_imul(b, src0_32, src1_32);
147             nir_def *dest_shifted = nir_ishr_imm(b, dest_32, src0->bit_size);
148             lowered = nir_type_convert(b, dest_shifted, base_type, base_type | src0->bit_size, nir_rounding_mode_undef);
149          } else {
150             nir_def *cshift = nir_imm_int(b, src0->bit_size / 2);
151             nir_def *cmask = nir_imm_intN_t(b, (1ull << (src0->bit_size / 2)) - 1, src0->bit_size);
152             nir_def *different_signs = NULL;
153             if (instr->op == nir_op_imul_high) {
154                nir_def *c0 = nir_imm_intN_t(b, 0, src0->bit_size);
155                different_signs = nir_ixor(b,
156                                           nir_ilt(b, src0, c0),
157                                           nir_ilt(b, src1, c0));
158                src0 = nir_iabs(b, src0);
159                src1 = nir_iabs(b, src1);
160             }
161 
162             /*   ABCD
163              * * EFGH
164              * ======
165              * (GH * CD) + (GH * AB) << 16 + (EF * CD) << 16 + (EF * AB) << 32
166              *
167              * Start by splitting into the 4 multiplies.
168              */
169             nir_def *src0l = nir_iand(b, src0, cmask);
170             nir_def *src1l = nir_iand(b, src1, cmask);
171             nir_def *src0h = nir_ushr(b, src0, cshift);
172             nir_def *src1h = nir_ushr(b, src1, cshift);
173 
174             nir_def *lo = nir_imul(b, src0l, src1l);
175             nir_def *m1 = nir_imul(b, src0l, src1h);
176             nir_def *m2 = nir_imul(b, src0h, src1l);
177             nir_def *hi = nir_imul(b, src0h, src1h);
178 
179             nir_def *tmp;
180 
181             tmp = nir_ishl(b, m1, cshift);
182             hi = nir_iadd(b, hi, nir_uadd_carry(b, lo, tmp));
183             lo = nir_iadd(b, lo, tmp);
184             hi = nir_iadd(b, hi, nir_ushr(b, m1, cshift));
185 
186             tmp = nir_ishl(b, m2, cshift);
187             hi = nir_iadd(b, hi, nir_uadd_carry(b, lo, tmp));
188             lo = nir_iadd(b, lo, tmp);
189             hi = nir_iadd(b, hi, nir_ushr(b, m2, cshift));
190 
191             if (instr->op == nir_op_imul_high) {
192                /* For channels where different_signs is set we have to perform a
193                 * 64-bit negation.  This is *not* the same as just negating the
194                 * high 32-bits.  Consider -3 * 2.  The high 32-bits is 0, but the
195                 * desired result is -1, not -0!  Recall -x == ~x + 1.
196                 */
197                nir_def *c1 = nir_imm_intN_t(b, 1, src0->bit_size);
198                hi = nir_bcsel(b, different_signs,
199                               nir_iadd(b,
200                                        nir_inot(b, hi),
201                                        nir_uadd_carry(b, nir_inot(b, lo), c1)),
202                               hi);
203             }
204 
205             lowered = hi;
206          }
207       }
208       break;
209 
210    case nir_op_fmin:
211    case nir_op_fmax: {
212       if (!b->shader->options->lower_fminmax_signed_zero ||
213           !nir_alu_instr_is_signed_zero_preserve(instr))
214          break;
215 
216       nir_def *s0 = nir_ssa_for_alu_src(b, instr, 0);
217       nir_def *s1 = nir_ssa_for_alu_src(b, instr, 1);
218 
219       bool max = instr->op == nir_op_fmax;
220       nir_def *iminmax = max ? nir_imax(b, s0, s1) : nir_imin(b, s0, s1);
221 
222       /* Lower the fmin/fmax to a no_signed_zero fmin/fmax. This ensures that
223        * nir_lower_alu is idempotent, and allows the backend to implement
224        * soundly the no_signed_zero subset of fmin/fmax.
225        */
226       b->fp_fast_math &= ~FLOAT_CONTROLS_SIGNED_ZERO_PRESERVE;
227       nir_def *fminmax = max ? nir_fmax(b, s0, s1) : nir_fmin(b, s0, s1);
228       b->fp_fast_math = instr->fp_fast_math;
229 
230       lowered = nir_bcsel(b, nir_feq(b, s0, s1), iminmax, fminmax);
231       break;
232    }
233 
234    default:
235       break;
236    }
237 
238    if (lowered) {
239       nir_def_replace(&instr->def, lowered);
240       return true;
241    } else {
242       return false;
243    }
244 }
245 
246 bool
nir_lower_alu(nir_shader * shader)247 nir_lower_alu(nir_shader *shader)
248 {
249    if (!shader->options->lower_bitfield_reverse &&
250        !shader->options->lower_bit_count &&
251        !shader->options->lower_mul_high &&
252        !shader->options->lower_fminmax_signed_zero)
253       return false;
254 
255    return nir_shader_alu_pass(shader, lower_alu_instr,
256                               nir_metadata_control_flow, NULL);
257 }
258