1/*
2 * Copyright © 2014-2015 Broadcom
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#include "nir.h"
25#include "nir_builder.h"
26
27/** @file nir_lower_alu_to_scalar.c
28 *
29 * Replaces nir_alu_instr operations with more than one channel used in the
30 * arguments with individual per-channel operations.
31 */
32
33static void
34nir_alu_ssa_dest_init(nir_alu_instr *instr, unsigned num_components,
35                      unsigned bit_size)
36{
37   nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components,
38                     bit_size, NULL);
39   instr->dest.write_mask = (1 << num_components) - 1;
40}
41
42static void
43lower_reduction(nir_alu_instr *instr, nir_op chan_op, nir_op merge_op,
44                nir_builder *builder)
45{
46   unsigned num_components = nir_op_infos[instr->op].input_sizes[0];
47
48   nir_ssa_def *last = NULL;
49   for (unsigned i = 0; i < num_components; i++) {
50      nir_alu_instr *chan = nir_alu_instr_create(builder->shader, chan_op);
51      nir_alu_ssa_dest_init(chan, 1, instr->dest.dest.ssa.bit_size);
52      nir_alu_src_copy(&chan->src[0], &instr->src[0], chan);
53      chan->src[0].swizzle[0] = chan->src[0].swizzle[i];
54      if (nir_op_infos[chan_op].num_inputs > 1) {
55         assert(nir_op_infos[chan_op].num_inputs == 2);
56         nir_alu_src_copy(&chan->src[1], &instr->src[1], chan);
57         chan->src[1].swizzle[0] = chan->src[1].swizzle[i];
58      }
59      chan->exact = instr->exact;
60
61      nir_builder_instr_insert(builder, &chan->instr);
62
63      if (i == 0) {
64         last = &chan->dest.dest.ssa;
65      } else {
66         last = nir_build_alu(builder, merge_op,
67                              last, &chan->dest.dest.ssa, NULL, NULL);
68      }
69   }
70
71   assert(instr->dest.write_mask == 1);
72   nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(last));
73   nir_instr_remove(&instr->instr);
74}
75
76static bool
77lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b)
78{
79   unsigned num_src = nir_op_infos[instr->op].num_inputs;
80   unsigned i, chan;
81
82   assert(instr->dest.dest.is_ssa);
83   assert(instr->dest.write_mask != 0);
84
85   b->cursor = nir_before_instr(&instr->instr);
86   b->exact = instr->exact;
87
88#define LOWER_REDUCTION(name, chan, merge) \
89   case name##2: \
90   case name##3: \
91   case name##4: \
92      lower_reduction(instr, chan, merge, b); \
93      return true;
94
95   switch (instr->op) {
96   case nir_op_vec4:
97   case nir_op_vec3:
98   case nir_op_vec2:
99   case nir_op_cube_face_coord:
100   case nir_op_cube_face_index:
101      /* We don't need to scalarize these ops, they're the ones generated to
102       * group up outputs into a value that can be SSAed.
103       */
104      return false;
105
106   case nir_op_pack_half_2x16:
107      if (!b->shader->options->lower_pack_half_2x16)
108         return false;
109
110      nir_ssa_def *src_vec2 = nir_ssa_for_alu_src(b, instr, 0);
111
112      nir_ssa_def *val =
113         nir_pack_half_2x16_split(b, nir_channel(b, src_vec2, 0),
114                                     nir_channel(b, src_vec2, 1));
115
116      nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val));
117      nir_instr_remove(&instr->instr);
118      return true;
119
120   case nir_op_unpack_unorm_4x8:
121   case nir_op_unpack_snorm_4x8:
122   case nir_op_unpack_unorm_2x16:
123   case nir_op_unpack_snorm_2x16:
124      /* There is no scalar version of these ops, unless we were to break it
125       * down to bitshifts and math (which is definitely not intended).
126       */
127      return false;
128
129   case nir_op_unpack_half_2x16: {
130      if (!b->shader->options->lower_unpack_half_2x16)
131         return false;
132
133      nir_ssa_def *packed = nir_ssa_for_alu_src(b, instr, 0);
134
135      nir_ssa_def *comps[2];
136      comps[0] = nir_unpack_half_2x16_split_x(b, packed);
137      comps[1] = nir_unpack_half_2x16_split_y(b, packed);
138      nir_ssa_def *vec = nir_vec(b, comps, 2);
139
140      nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(vec));
141      nir_instr_remove(&instr->instr);
142      return true;
143   }
144
145   case nir_op_pack_uvec2_to_uint: {
146      assert(b->shader->options->lower_pack_snorm_2x16 ||
147             b->shader->options->lower_pack_unorm_2x16);
148
149      nir_ssa_def *word = nir_extract_u16(b, nir_ssa_for_alu_src(b, instr, 0),
150                                             nir_imm_int(b, 0));
151      nir_ssa_def *val =
152         nir_ior(b, nir_ishl(b, nir_channel(b, word, 1), nir_imm_int(b, 16)),
153                                nir_channel(b, word, 0));
154
155      nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val));
156      nir_instr_remove(&instr->instr);
157      break;
158   }
159
160   case nir_op_pack_uvec4_to_uint: {
161      assert(b->shader->options->lower_pack_snorm_4x8 ||
162             b->shader->options->lower_pack_unorm_4x8);
163
164      nir_ssa_def *byte = nir_extract_u8(b, nir_ssa_for_alu_src(b, instr, 0),
165                                            nir_imm_int(b, 0));
166      nir_ssa_def *val =
167         nir_ior(b, nir_ior(b, nir_ishl(b, nir_channel(b, byte, 3), nir_imm_int(b, 24)),
168                               nir_ishl(b, nir_channel(b, byte, 2), nir_imm_int(b, 16))),
169                    nir_ior(b, nir_ishl(b, nir_channel(b, byte, 1), nir_imm_int(b, 8)),
170                               nir_channel(b, byte, 0)));
171
172      nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val));
173      nir_instr_remove(&instr->instr);
174      break;
175   }
176
177   case nir_op_fdph: {
178      nir_ssa_def *src0_vec = nir_ssa_for_alu_src(b, instr, 0);
179      nir_ssa_def *src1_vec = nir_ssa_for_alu_src(b, instr, 1);
180
181      nir_ssa_def *sum[4];
182      for (unsigned i = 0; i < 3; i++) {
183         sum[i] = nir_fmul(b, nir_channel(b, src0_vec, i),
184                              nir_channel(b, src1_vec, i));
185      }
186      sum[3] = nir_channel(b, src1_vec, 3);
187
188      nir_ssa_def *val = nir_fadd(b, nir_fadd(b, sum[0], sum[1]),
189                                     nir_fadd(b, sum[2], sum[3]));
190
191      nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val));
192      nir_instr_remove(&instr->instr);
193      return true;
194   }
195
196   case nir_op_unpack_64_2x32:
197   case nir_op_unpack_32_2x16:
198      return false;
199
200      LOWER_REDUCTION(nir_op_fdot, nir_op_fmul, nir_op_fadd);
201      LOWER_REDUCTION(nir_op_ball_fequal, nir_op_feq, nir_op_iand);
202      LOWER_REDUCTION(nir_op_ball_iequal, nir_op_ieq, nir_op_iand);
203      LOWER_REDUCTION(nir_op_bany_fnequal, nir_op_fne, nir_op_ior);
204      LOWER_REDUCTION(nir_op_bany_inequal, nir_op_ine, nir_op_ior);
205      LOWER_REDUCTION(nir_op_b32all_fequal, nir_op_feq32, nir_op_iand);
206      LOWER_REDUCTION(nir_op_b32all_iequal, nir_op_ieq32, nir_op_iand);
207      LOWER_REDUCTION(nir_op_b32any_fnequal, nir_op_fne32, nir_op_ior);
208      LOWER_REDUCTION(nir_op_b32any_inequal, nir_op_ine32, nir_op_ior);
209      LOWER_REDUCTION(nir_op_fall_equal, nir_op_seq, nir_op_fand);
210      LOWER_REDUCTION(nir_op_fany_nequal, nir_op_sne, nir_op_for);
211
212   default:
213      break;
214   }
215
216   if (instr->dest.dest.ssa.num_components == 1)
217      return false;
218
219   unsigned num_components = instr->dest.dest.ssa.num_components;
220   nir_ssa_def *comps[NIR_MAX_VEC_COMPONENTS] = { NULL };
221
222   for (chan = 0; chan < NIR_MAX_VEC_COMPONENTS; chan++) {
223      if (!(instr->dest.write_mask & (1 << chan)))
224         continue;
225
226      nir_alu_instr *lower = nir_alu_instr_create(b->shader, instr->op);
227      for (i = 0; i < num_src; i++) {
228         /* We only handle same-size-as-dest (input_sizes[] == 0) or scalar
229          * args (input_sizes[] == 1).
230          */
231         assert(nir_op_infos[instr->op].input_sizes[i] < 2);
232         unsigned src_chan = (nir_op_infos[instr->op].input_sizes[i] == 1 ?
233                              0 : chan);
234
235         nir_alu_src_copy(&lower->src[i], &instr->src[i], lower);
236         for (int j = 0; j < NIR_MAX_VEC_COMPONENTS; j++)
237            lower->src[i].swizzle[j] = instr->src[i].swizzle[src_chan];
238      }
239
240      nir_alu_ssa_dest_init(lower, 1, instr->dest.dest.ssa.bit_size);
241      lower->dest.saturate = instr->dest.saturate;
242      comps[chan] = &lower->dest.dest.ssa;
243      lower->exact = instr->exact;
244
245      nir_builder_instr_insert(b, &lower->instr);
246   }
247
248   nir_ssa_def *vec = nir_vec(b, comps, num_components);
249
250   nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(vec));
251
252   nir_instr_remove(&instr->instr);
253   return true;
254}
255
256static bool
257nir_lower_alu_to_scalar_impl(nir_function_impl *impl)
258{
259   nir_builder builder;
260   nir_builder_init(&builder, impl);
261   bool progress = false;
262
263   nir_foreach_block(block, impl) {
264      nir_foreach_instr_safe(instr, block) {
265         if (instr->type == nir_instr_type_alu) {
266            progress = lower_alu_instr_scalar(nir_instr_as_alu(instr),
267                                              &builder) || progress;
268         }
269      }
270   }
271
272   nir_metadata_preserve(impl, nir_metadata_block_index |
273                               nir_metadata_dominance);
274
275   return progress;
276}
277
278bool
279nir_lower_alu_to_scalar(nir_shader *shader)
280{
281   bool progress = false;
282
283   nir_foreach_function(function, shader) {
284      if (function->impl)
285         progress = nir_lower_alu_to_scalar_impl(function->impl) || progress;
286   }
287
288   return progress;
289}
290