1 /*
2 * Copyright © 2017 Red Hat
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 * Authors:
24 * Rob Clark <robclark@freedesktop.org>
25 */
26
27 #include "nir.h"
28 #include "nir_builder.h"
29
30 /*
31 * Remap atomic counters to SSBOs, starting from the shader's next SSBO slot
32 * (info.num_ssbos).
33 */
34
35 static nir_deref_instr *
deref_offset_var(nir_builder * b,unsigned binding,unsigned offset_align_state)36 deref_offset_var(nir_builder *b, unsigned binding, unsigned offset_align_state)
37 {
38 nir_foreach_uniform_variable(var, b->shader) {
39 if (var->num_state_slots != 1)
40 continue;
41 if (var->state_slots[0].tokens[0] == offset_align_state &&
42 var->state_slots[0].tokens[1] == binding)
43 return nir_build_deref_var(b, var);
44 }
45
46 nir_variable *var = nir_variable_create(b->shader, nir_var_uniform, glsl_uint_type(), "offset");
47 var->state_slots = ralloc_array(var, nir_state_slot, 1);
48 var->state_slots[0].tokens[0] = offset_align_state;
49 var->state_slots[0].tokens[1] = binding;
50 var->num_state_slots = 1;
51 var->data.how_declared = nir_var_hidden;
52 b->shader->num_uniforms++;
53 return nir_build_deref_var(b, var);
54 }
55
56 static bool
lower_instr(nir_intrinsic_instr * instr,unsigned ssbo_offset,nir_builder * b,unsigned offset_align_state)57 lower_instr(nir_intrinsic_instr *instr, unsigned ssbo_offset, nir_builder *b, unsigned offset_align_state)
58 {
59 nir_intrinsic_op op;
60
61 b->cursor = nir_before_instr(&instr->instr);
62
63 switch (instr->intrinsic) {
64 case nir_intrinsic_memory_barrier_atomic_counter:
65 /* Atomic counters are now SSBOs so memoryBarrierAtomicCounter() is now
66 * memoryBarrierBuffer().
67 */
68 instr->intrinsic = nir_intrinsic_memory_barrier_buffer;
69 return true;
70
71 case nir_intrinsic_atomic_counter_inc:
72 case nir_intrinsic_atomic_counter_add:
73 case nir_intrinsic_atomic_counter_pre_dec:
74 case nir_intrinsic_atomic_counter_post_dec:
75 /* inc and dec get remapped to add: */
76 op = nir_intrinsic_ssbo_atomic_add;
77 break;
78 case nir_intrinsic_atomic_counter_read:
79 op = nir_intrinsic_load_ssbo;
80 break;
81 case nir_intrinsic_atomic_counter_min:
82 op = nir_intrinsic_ssbo_atomic_umin;
83 break;
84 case nir_intrinsic_atomic_counter_max:
85 op = nir_intrinsic_ssbo_atomic_umax;
86 break;
87 case nir_intrinsic_atomic_counter_and:
88 op = nir_intrinsic_ssbo_atomic_and;
89 break;
90 case nir_intrinsic_atomic_counter_or:
91 op = nir_intrinsic_ssbo_atomic_or;
92 break;
93 case nir_intrinsic_atomic_counter_xor:
94 op = nir_intrinsic_ssbo_atomic_xor;
95 break;
96 case nir_intrinsic_atomic_counter_exchange:
97 op = nir_intrinsic_ssbo_atomic_exchange;
98 break;
99 case nir_intrinsic_atomic_counter_comp_swap:
100 op = nir_intrinsic_ssbo_atomic_comp_swap;
101 break;
102 default:
103 return false;
104 }
105
106 nir_ssa_def *buffer = nir_imm_int(b, ssbo_offset + nir_intrinsic_base(instr));
107 nir_ssa_def *temp = NULL;
108
109 nir_ssa_def *offset_load = NULL;
110 if (offset_align_state) {
111 nir_deref_instr *deref_offset = deref_offset_var(b, nir_intrinsic_base(instr), offset_align_state);
112 offset_load = nir_load_deref(b, deref_offset);
113 }
114 nir_intrinsic_instr *new_instr =
115 nir_intrinsic_instr_create(b->shader, op);
116
117 /* a couple instructions need special handling since they don't map
118 * 1:1 with ssbo atomics
119 */
120 switch (instr->intrinsic) {
121 case nir_intrinsic_atomic_counter_inc:
122 /* remapped to ssbo_atomic_add: { buffer_idx, offset, +1 } */
123 temp = nir_imm_int(b, +1);
124 new_instr->src[0] = nir_src_for_ssa(buffer);
125 nir_src_copy(&new_instr->src[1], &instr->src[0]);
126 new_instr->src[2] = nir_src_for_ssa(temp);
127 break;
128 case nir_intrinsic_atomic_counter_pre_dec:
129 case nir_intrinsic_atomic_counter_post_dec:
130 /* remapped to ssbo_atomic_add: { buffer_idx, offset, -1 } */
131 /* NOTE semantic difference so we adjust the return value below */
132 temp = nir_imm_int(b, -1);
133 new_instr->src[0] = nir_src_for_ssa(buffer);
134 nir_src_copy(&new_instr->src[1], &instr->src[0]);
135 new_instr->src[2] = nir_src_for_ssa(temp);
136 break;
137 case nir_intrinsic_atomic_counter_read:
138 /* remapped to load_ssbo: { buffer_idx, offset } */
139 new_instr->src[0] = nir_src_for_ssa(buffer);
140 nir_src_copy(&new_instr->src[1], &instr->src[0]);
141 break;
142 default:
143 /* remapped to ssbo_atomic_x: { buffer_idx, offset, data, (compare)? } */
144 new_instr->src[0] = nir_src_for_ssa(buffer);
145 nir_src_copy(&new_instr->src[1], &instr->src[0]);
146 nir_src_copy(&new_instr->src[2], &instr->src[1]);
147 if (op == nir_intrinsic_ssbo_atomic_comp_swap ||
148 op == nir_intrinsic_ssbo_atomic_fcomp_swap)
149 nir_src_copy(&new_instr->src[3], &instr->src[2]);
150 break;
151 }
152
153 if (offset_load)
154 new_instr->src[1].ssa = nir_iadd(b, new_instr->src[1].ssa, offset_load);
155
156 if (new_instr->intrinsic == nir_intrinsic_load_ssbo) {
157 nir_intrinsic_set_align(new_instr, 4, 0);
158
159 /* we could be replacing an intrinsic with fixed # of dest
160 * num_components with one that has variable number. So
161 * best to take this from the dest:
162 */
163 new_instr->num_components = instr->dest.ssa.num_components;
164 }
165
166 nir_ssa_dest_init(&new_instr->instr, &new_instr->dest,
167 instr->dest.ssa.num_components,
168 instr->dest.ssa.bit_size, NULL);
169 nir_instr_insert_before(&instr->instr, &new_instr->instr);
170 nir_instr_remove(&instr->instr);
171
172 if (instr->intrinsic == nir_intrinsic_atomic_counter_pre_dec) {
173 b->cursor = nir_after_instr(&new_instr->instr);
174 nir_ssa_def *result = nir_iadd(b, &new_instr->dest.ssa, temp);
175 nir_ssa_def_rewrite_uses(&instr->dest.ssa, result);
176 } else {
177 nir_ssa_def_rewrite_uses(&instr->dest.ssa, &new_instr->dest.ssa);
178 }
179
180 return true;
181 }
182
183 static bool
is_atomic_uint(const struct glsl_type * type)184 is_atomic_uint(const struct glsl_type *type)
185 {
186 if (glsl_get_base_type(type) == GLSL_TYPE_ARRAY)
187 return is_atomic_uint(glsl_get_array_element(type));
188 return glsl_get_base_type(type) == GLSL_TYPE_ATOMIC_UINT;
189 }
190
191 bool
nir_lower_atomics_to_ssbo(nir_shader * shader,unsigned offset_align_state)192 nir_lower_atomics_to_ssbo(nir_shader *shader, unsigned offset_align_state)
193 {
194 unsigned ssbo_offset = shader->info.num_ssbos;
195 bool progress = false;
196
197 nir_foreach_function(function, shader) {
198 if (function->impl) {
199 nir_builder builder;
200 nir_builder_init(&builder, function->impl);
201 nir_foreach_block(block, function->impl) {
202 nir_foreach_instr_safe(instr, block) {
203 if (instr->type == nir_instr_type_intrinsic)
204 progress |= lower_instr(nir_instr_as_intrinsic(instr),
205 ssbo_offset, &builder, offset_align_state);
206 }
207 }
208
209 nir_metadata_preserve(function->impl, nir_metadata_block_index |
210 nir_metadata_dominance);
211 }
212 }
213
214 if (progress) {
215 /* replace atomic_uint uniforms with ssbo's: */
216 unsigned replaced = 0;
217 nir_foreach_uniform_variable_safe(var, shader) {
218 if (is_atomic_uint(var->type)) {
219 exec_node_remove(&var->node);
220
221 if (replaced & (1 << var->data.binding))
222 continue;
223
224 nir_variable *ssbo;
225 char name[16];
226
227 /* A length of 0 is used to denote unsized arrays */
228 const struct glsl_type *type = glsl_array_type(glsl_uint_type(), 0, 0);
229
230 snprintf(name, sizeof(name), "counter%d", var->data.binding);
231
232 ssbo = nir_variable_create(shader, nir_var_mem_ssbo, type, name);
233 ssbo->data.binding = ssbo_offset + var->data.binding;
234 ssbo->data.explicit_binding = var->data.explicit_binding;
235
236 /* We can't use num_abos, because it only represents the number of
237 * active atomic counters, and currently unlike SSBO's they aren't
238 * compacted so num_abos actually isn't a bound on the index passed
239 * to nir_intrinsic_atomic_counter_*. e.g. if we have a single atomic
240 * counter declared like:
241 *
242 * layout(binding=1) atomic_uint counter0;
243 *
244 * then when we lower accesses to it the atomic_counter_* intrinsics
245 * will have 1 as the index but num_abos will still be 1.
246 */
247 shader->info.num_ssbos = MAX2(shader->info.num_ssbos,
248 ssbo->data.binding + 1);
249
250 struct glsl_struct_field field = {
251 .type = type,
252 .name = "counters",
253 .location = -1,
254 };
255
256 ssbo->interface_type =
257 glsl_interface_type(&field, 1, GLSL_INTERFACE_PACKING_STD430,
258 false, "counters");
259
260 replaced |= (1 << var->data.binding);
261 }
262 }
263
264 shader->info.num_abos = 0;
265 }
266
267 return progress;
268 }
269
270