1 /* SPDX-License-Identifier: MIT */
2
3 #include "nir.h"
4 #include "nir_builder.h"
5
6 static bool
var_is_inline_sampler(const nir_variable * var)7 var_is_inline_sampler(const nir_variable *var)
8 {
9 if (var->data.mode != nir_var_uniform)
10 return false;
11
12 return glsl_type_is_sampler(var->type) &&
13 var->data.sampler.is_inline_sampler;
14 }
15
16 static bool
inline_sampler_vars_equal(const nir_variable * a,const nir_variable * b)17 inline_sampler_vars_equal(const nir_variable *a, const nir_variable *b)
18 {
19 assert(var_is_inline_sampler(a) && var_is_inline_sampler(b));
20
21 if (a == b)
22 return true;
23
24 return a->data.sampler.addressing_mode == b->data.sampler.addressing_mode &&
25 a->data.sampler.normalized_coordinates == b->data.sampler.normalized_coordinates &&
26 a->data.sampler.filter_mode == b->data.sampler.filter_mode;
27 }
28
29 static nir_variable *
find_identical_inline_sampler(nir_shader * nir,struct exec_list * inline_samplers,nir_variable * sampler)30 find_identical_inline_sampler(nir_shader *nir,
31 struct exec_list *inline_samplers,
32 nir_variable *sampler)
33 {
34 nir_foreach_variable_in_list(var, inline_samplers) {
35 if (inline_sampler_vars_equal(var, sampler))
36 return var;
37 }
38
39 nir_foreach_uniform_variable(var, nir) {
40 if (!var_is_inline_sampler(var) ||
41 !inline_sampler_vars_equal(var, sampler))
42 continue;
43
44 exec_node_remove(&var->node);
45 exec_list_push_tail(inline_samplers, &var->node);
46 return var;
47 }
48 unreachable("Should have at least found the input sampler");
49 }
50
51 static bool
nir_dedup_inline_samplers_instr(nir_builder * b,nir_instr * instr,void * cb_data)52 nir_dedup_inline_samplers_instr(nir_builder *b,
53 nir_instr *instr,
54 void *cb_data)
55 {
56 struct exec_list *inline_samplers = cb_data;
57
58 if (instr->type != nir_instr_type_deref)
59 return false;
60
61 nir_deref_instr *deref = nir_instr_as_deref(instr);
62 if (deref->deref_type != nir_deref_type_var)
63 return false;
64
65 nir_variable *sampler = nir_deref_instr_get_variable(deref);
66 if (!var_is_inline_sampler(sampler))
67 return false;
68
69 nir_variable *replacement =
70 find_identical_inline_sampler(b->shader, inline_samplers, sampler);
71 deref->var = replacement;
72 return true;
73 }
74
75 /** De-duplicates inline sampler variables
76 *
77 * Any dead or redundant inline sampler variables are removed any live inline
78 * sampler variables are placed at the end of the variables list.
79 */
80 bool
nir_dedup_inline_samplers(nir_shader * nir)81 nir_dedup_inline_samplers(nir_shader *nir)
82 {
83 struct exec_list inline_samplers;
84 exec_list_make_empty(&inline_samplers);
85
86 nir_shader_instructions_pass(nir, nir_dedup_inline_samplers_instr,
87 nir_metadata_control_flow,
88 &inline_samplers);
89
90 /* If we found any inline samplers in the instructions pass, they'll now be
91 * in the inline_samplers list.
92 */
93 bool progress = !exec_list_is_empty(&inline_samplers);
94
95 /* Remove any dead samplers */
96 nir_foreach_uniform_variable_safe(var, nir) {
97 if (var_is_inline_sampler(var)) {
98 exec_node_remove(&var->node);
99 progress = true;
100 }
101 }
102
103 exec_node_insert_list_after(exec_list_get_tail(&nir->variables),
104 &inline_samplers);
105
106 return progress;
107 }
108
109 bool
nir_lower_cl_images(nir_shader * shader,bool lower_image_derefs,bool lower_sampler_derefs)110 nir_lower_cl_images(nir_shader *shader, bool lower_image_derefs, bool lower_sampler_derefs)
111 {
112 nir_function_impl *impl = nir_shader_get_entrypoint(shader);
113
114 ASSERTED int last_loc = -1;
115 int num_rd_images = 0, num_wr_images = 0;
116
117 BITSET_ZERO(shader->info.image_buffers);
118 BITSET_ZERO(shader->info.msaa_images);
119 nir_foreach_variable_with_modes(var, shader, nir_var_image | nir_var_uniform) {
120 if (!glsl_type_is_image(var->type) && !glsl_type_is_texture(var->type))
121 continue;
122
123 /* Assume they come in order */
124 assert(var->data.location > last_loc);
125 last_loc = var->data.location;
126
127 assert(glsl_type_is_image(var->type) || var->data.access & ACCESS_NON_WRITEABLE);
128 if (var->data.access & ACCESS_NON_WRITEABLE)
129 var->data.driver_location = num_rd_images++;
130 else
131 var->data.driver_location = num_wr_images++;
132 var->data.binding = var->data.driver_location;
133
134 switch (glsl_get_sampler_dim(var->type)) {
135 case GLSL_SAMPLER_DIM_BUF:
136 BITSET_SET(shader->info.image_buffers, var->data.binding);
137 break;
138 case GLSL_SAMPLER_DIM_MS:
139 BITSET_SET(shader->info.msaa_images, var->data.binding);
140 break;
141 default:
142 break;
143 }
144 }
145 shader->info.num_textures = num_rd_images;
146 BITSET_ZERO(shader->info.textures_used);
147 if (num_rd_images)
148 BITSET_SET_RANGE(shader->info.textures_used, 0, num_rd_images - 1);
149
150 BITSET_ZERO(shader->info.images_used);
151 if (num_wr_images)
152 BITSET_SET_RANGE(shader->info.images_used, 0, num_wr_images - 1);
153 shader->info.num_images = num_wr_images;
154
155 last_loc = -1;
156 int num_samplers = 0;
157 nir_foreach_uniform_variable(var, shader) {
158 if (var->type == glsl_bare_sampler_type()) {
159 /* Assume they come in order */
160 assert(var->data.location > last_loc);
161 last_loc = var->data.location;
162 var->data.driver_location = num_samplers++;
163 var->data.binding = var->data.driver_location;
164 } else {
165 /* CL shouldn't have any sampled images */
166 assert(!glsl_type_is_sampler(var->type));
167 }
168 }
169 BITSET_ZERO(shader->info.samplers_used);
170 if (num_samplers)
171 BITSET_SET_RANGE(shader->info.samplers_used, 0, num_samplers - 1);
172
173 nir_builder b = nir_builder_create(impl);
174
175 bool progress = false;
176 nir_foreach_block_reverse(block, impl) {
177 nir_foreach_instr_reverse_safe(instr, block) {
178 switch (instr->type) {
179 case nir_instr_type_deref: {
180 nir_deref_instr *deref = nir_instr_as_deref(instr);
181 if (deref->deref_type != nir_deref_type_var)
182 break;
183
184 if (!glsl_type_is_image(deref->type) &&
185 !glsl_type_is_texture(deref->type) &&
186 !glsl_type_is_sampler(deref->type))
187 break;
188
189 if (!lower_image_derefs && glsl_type_is_image(deref->type))
190 break;
191
192 if (!lower_sampler_derefs &&
193 (glsl_type_is_sampler(deref->type) || glsl_type_is_texture(deref->type)))
194 break;
195
196 b.cursor = nir_instr_remove(&deref->instr);
197 nir_def *loc =
198 nir_imm_intN_t(&b, deref->var->data.driver_location,
199 deref->def.bit_size);
200 nir_def_rewrite_uses(&deref->def, loc);
201 progress = true;
202 break;
203 }
204
205 case nir_instr_type_tex: {
206 if (!lower_sampler_derefs)
207 break;
208
209 nir_tex_instr *tex = nir_instr_as_tex(instr);
210 unsigned count = 0;
211 for (unsigned i = 0; i < tex->num_srcs; i++) {
212 if (tex->src[i].src_type == nir_tex_src_texture_deref ||
213 tex->src[i].src_type == nir_tex_src_sampler_deref) {
214 nir_deref_instr *deref = nir_src_as_deref(tex->src[i].src);
215 if (deref->deref_type == nir_deref_type_var) {
216 /* In this case, we know the actual variable */
217 if (tex->src[i].src_type == nir_tex_src_texture_deref)
218 tex->texture_index = deref->var->data.driver_location;
219 else
220 tex->sampler_index = deref->var->data.driver_location;
221 /* This source gets discarded */
222 nir_instr_clear_src(&tex->instr, &tex->src[i].src);
223 continue;
224 } else {
225 b.cursor = nir_before_instr(&tex->instr);
226 /* Back-ends expect a 32-bit thing, not 64-bit */
227 nir_def *offset = nir_u2u32(&b, tex->src[i].src.ssa);
228 if (tex->src[i].src_type == nir_tex_src_texture_deref)
229 tex->src[count].src_type = nir_tex_src_texture_offset;
230 else
231 tex->src[count].src_type = nir_tex_src_sampler_offset;
232 nir_src_rewrite(&tex->src[count].src, offset);
233 }
234 } else {
235 /* If we've removed a source, move this one down */
236 if (count != i) {
237 assert(count < i);
238 tex->src[count].src_type = tex->src[i].src_type;
239 nir_instr_move_src(&tex->instr, &tex->src[count].src,
240 &tex->src[i].src);
241 }
242 }
243 count++;
244 }
245 tex->num_srcs = count;
246 progress = true;
247 break;
248 }
249
250 case nir_instr_type_intrinsic: {
251 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
252 switch (intrin->intrinsic) {
253 case nir_intrinsic_image_deref_load:
254 case nir_intrinsic_image_deref_store:
255 case nir_intrinsic_image_deref_atomic:
256 case nir_intrinsic_image_deref_atomic_swap:
257 case nir_intrinsic_image_deref_size:
258 case nir_intrinsic_image_deref_samples: {
259
260 /* CL_DEPTH image loads return a scalar value */
261 if (intrin->intrinsic == nir_intrinsic_image_deref_load &&
262 intrin->def.num_components == 1) {
263 b.cursor = nir_after_instr(&intrin->instr);
264 intrin->num_components = 4;
265 intrin->def.num_components = 4;
266 nir_def *scalar = nir_channel(&b, &intrin->def, 0);
267 nir_def_rewrite_uses_after(&intrin->def, scalar, scalar->parent_instr);
268 progress = true;
269 }
270
271 if (!lower_image_derefs)
272 break;
273
274 b.cursor = nir_before_instr(&intrin->instr);
275 /* Back-ends expect a 32-bit thing, not 64-bit */
276 nir_def *offset = nir_u2u32(&b, intrin->src[0].ssa);
277 nir_rewrite_image_intrinsic(intrin, offset, false);
278 progress = true;
279 break;
280 }
281
282 default:
283 break;
284 }
285 break;
286 }
287
288 default:
289 break;
290 }
291 }
292 }
293
294 if (progress) {
295 nir_metadata_preserve(impl, nir_metadata_control_flow);
296 } else {
297 nir_metadata_preserve(impl, nir_metadata_all);
298 }
299
300 return progress;
301 }
302