1 /*
2 * Copyright © 2016 Intel Corporation
3 * Copyright © 2020 Valve Corporation
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 DEALINGS
22 * IN THE SOFTWARE.
23 */
24
25 #include "nir_builder.h"
26 #include "nir_control_flow.h"
27
28 /**
29 * This file implements an optimization for multiview. Some GPU's have a
30 * special mode which allows the vertex shader (or last stage in the geometry
31 * pipeline) to create multiple primitives in different layers of the
32 * framebuffer at once by writing multiple copies of gl_Position. The
33 * assumption is that in most uses of multiview, the only use of gl_ViewIndex
34 * is to change the position to implement the parallax effect, and other
35 * varyings will be the same between the different views. We put the body of
36 * the original vertex shader in a loop, writing to a different copy of
37 * gl_Position each loop iteration, and then let other optimizations clean up
38 * the mess.
39 */
40
41 static bool
shader_writes_to_memory(nir_shader * shader)42 shader_writes_to_memory(nir_shader *shader)
43 {
44 /* With multiview, we would need to ensure that memory writes happen either
45 * once or once per view. Since combination of multiview and memory writes
46 * is not expected, we'll just skip this optimization in this case.
47 */
48
49 nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);
50
51 nir_foreach_block(block, entrypoint) {
52 nir_foreach_instr(instr, block) {
53 if (instr->type != nir_instr_type_intrinsic)
54 continue;
55 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
56
57 switch (intrin->intrinsic) {
58 case nir_intrinsic_deref_atomic:
59 case nir_intrinsic_deref_atomic_swap:
60 case nir_intrinsic_store_ssbo:
61 case nir_intrinsic_ssbo_atomic:
62 case nir_intrinsic_ssbo_atomic_swap:
63 case nir_intrinsic_store_shared:
64 case nir_intrinsic_store_shared2_amd:
65 case nir_intrinsic_shared_atomic:
66 case nir_intrinsic_shared_atomic_swap:
67 case nir_intrinsic_task_payload_atomic:
68 case nir_intrinsic_task_payload_atomic_swap:
69 case nir_intrinsic_image_deref_store:
70 case nir_intrinsic_image_deref_atomic:
71 case nir_intrinsic_image_deref_atomic_swap:
72 return true;
73
74 default:
75 /* Keep walking. */
76 break;
77 }
78 }
79 }
80
81 return false;
82 }
83
84 bool
nir_shader_uses_view_index(nir_shader * shader)85 nir_shader_uses_view_index(nir_shader *shader)
86 {
87 nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);
88
89 nir_foreach_block(block, entrypoint) {
90 nir_foreach_instr(instr, block) {
91 if (instr->type != nir_instr_type_intrinsic)
92 continue;
93
94 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
95 if (intrin->intrinsic == nir_intrinsic_load_view_index)
96 return true;
97 }
98 }
99
100 return false;
101 }
102
103 static bool
shader_only_position_uses_view_index(nir_shader * shader)104 shader_only_position_uses_view_index(nir_shader *shader)
105 {
106 nir_shader *shader_no_position = nir_shader_clone(NULL, shader);
107 nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader_no_position);
108
109 /* Remove the store position from a cloned shader. */
110 nir_foreach_block(block, entrypoint) {
111 nir_foreach_instr_safe(instr, block) {
112 if (instr->type != nir_instr_type_intrinsic)
113 continue;
114
115 nir_intrinsic_instr *store = nir_instr_as_intrinsic(instr);
116 if (store->intrinsic != nir_intrinsic_store_deref)
117 continue;
118
119 nir_variable *var = nir_intrinsic_get_var(store, 0);
120 if (var->data.location != VARYING_SLOT_POS)
121 continue;
122
123 nir_instr_remove(&store->instr);
124 }
125 }
126
127 /* Clean up shader so unused load_view_index intrinsics are removed. */
128 bool progress;
129 do {
130 progress = false;
131 progress |= nir_opt_dead_cf(shader_no_position);
132
133 /* Peephole select will drop if-blocks that have then and else empty,
134 * which will remove the usage of an SSA in the condition.
135 */
136 progress |= nir_opt_peephole_select(shader_no_position, 0, false, false);
137
138 progress |= nir_opt_dce(shader_no_position);
139 } while (progress);
140
141 bool uses_view_index = nir_shader_uses_view_index(shader_no_position);
142
143 ralloc_free(shader_no_position);
144 return !uses_view_index;
145 }
146
147 /* Return true if it's safe to call nir_lower_multiview() on this vertex
148 * shader. Note that this only handles driver-agnostic checks, i.e. things
149 * which would make nir_lower_multiview() incorrect. Any driver-specific
150 * checks, e.g. for sufficient varying space or performance considerations,
151 * should be handled in the driver.
152 *
153 * Note that we don't handle the more complex checks needed for lowering
154 * pipelines with geometry or tessellation shaders.
155 */
156
157 bool
nir_can_lower_multiview(nir_shader * shader)158 nir_can_lower_multiview(nir_shader *shader)
159 {
160 bool writes_position = false;
161 nir_foreach_shader_out_variable(var, shader) {
162 if (var->data.location == VARYING_SLOT_POS) {
163 writes_position = true;
164 break;
165 }
166 }
167
168 /* Don't bother handling this edge case. */
169 if (!writes_position)
170 return false;
171
172 return !shader_writes_to_memory(shader) &&
173 shader_only_position_uses_view_index(shader);
174 }
175
176 /**
177 * The lowering. Call with the last active geometry stage.
178 */
179
180 bool
nir_lower_multiview(nir_shader * shader,uint32_t view_mask)181 nir_lower_multiview(nir_shader *shader, uint32_t view_mask)
182 {
183 assert(shader->info.stage != MESA_SHADER_FRAGMENT);
184 int view_count = util_bitcount(view_mask);
185
186 nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);
187
188 /* Update position to refer to an array. */
189 nir_variable *pos_var = NULL;
190 nir_foreach_shader_out_variable(var, shader) {
191 if (var->data.location == VARYING_SLOT_POS) {
192 assert(var->type == glsl_vec4_type());
193 var->type = glsl_array_type(glsl_vec4_type(), view_count, 0);
194 var->data.per_view = true;
195 shader->info.per_view_outputs |= VARYING_BIT_POS;
196 pos_var = var;
197 break;
198 }
199 }
200
201 assert(pos_var);
202
203 nir_cf_list body;
204 nir_cf_list_extract(&body, &entrypoint->body);
205
206 nir_builder b = nir_builder_at(nir_after_impl(entrypoint));
207
208 /* Loop Index will go from 0 to view_count. */
209 nir_variable *loop_index_var =
210 nir_local_variable_create(entrypoint, glsl_uint_type(), "loop_index");
211 nir_deref_instr *loop_index_deref = nir_build_deref_var(&b, loop_index_var);
212 nir_store_deref(&b, loop_index_deref, nir_imm_int(&b, 0), 1);
213
214 /* Array of view index values that are active in the loop. Note that the
215 * loop index only matches the view index if there are no gaps in the
216 * view_mask.
217 */
218 nir_variable *view_index_var = nir_local_variable_create(
219 entrypoint, glsl_array_type(glsl_uint_type(), view_count, 0), "view_index");
220 nir_deref_instr *view_index_deref = nir_build_deref_var(&b, view_index_var);
221 {
222 int array_position = 0;
223 uint32_t view_mask_temp = view_mask;
224 while (view_mask_temp) {
225 uint32_t view_index = u_bit_scan(&view_mask_temp);
226 nir_store_deref(&b, nir_build_deref_array_imm(&b, view_index_deref, array_position),
227 nir_imm_int(&b, view_index), 1);
228 array_position++;
229 }
230 }
231
232 /* Create the equivalent of
233 *
234 * while (true):
235 * if (loop_index >= view_count):
236 * break
237 *
238 * view_index = active_indices[loop_index]
239 * pos_deref = &pos[loop_index]
240 *
241 * # Placeholder for the body to be reinserted.
242 *
243 * loop_index += 1
244 *
245 * Later both `view_index` and `pos_deref` will be used to rewrite the
246 * original shader body.
247 */
248
249 nir_loop *loop = nir_push_loop(&b);
250
251 nir_def *loop_index = nir_load_deref(&b, loop_index_deref);
252 nir_def *cmp = nir_ige_imm(&b, loop_index, view_count);
253 nir_if *loop_check = nir_push_if(&b, cmp);
254 nir_jump(&b, nir_jump_break);
255 nir_pop_if(&b, loop_check);
256
257 nir_def *view_index =
258 nir_load_deref(&b, nir_build_deref_array(&b, view_index_deref, loop_index));
259 nir_deref_instr *pos_deref =
260 nir_build_deref_array(&b, nir_build_deref_var(&b, pos_var), loop_index);
261
262 nir_store_deref(&b, loop_index_deref, nir_iadd_imm(&b, loop_index, 1), 1);
263 nir_pop_loop(&b, loop);
264
265 /* Reinsert the body. */
266 b.cursor = nir_after_instr(&pos_deref->instr);
267 nir_cf_reinsert(&body, b.cursor);
268
269 nir_foreach_block(block, entrypoint) {
270 nir_foreach_instr_safe(instr, block) {
271 if (instr->type != nir_instr_type_intrinsic)
272 continue;
273
274 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
275
276 switch (intrin->intrinsic) {
277 case nir_intrinsic_load_view_index: {
278 nir_def_rewrite_uses(&intrin->def, view_index);
279 break;
280 }
281
282 case nir_intrinsic_store_deref: {
283 nir_variable *var = nir_intrinsic_get_var(intrin, 0);
284 if (var == pos_var) {
285 nir_deref_instr *old_deref = nir_src_as_deref(intrin->src[0]);
286
287 nir_src_rewrite(&intrin->src[0], &pos_deref->def);
288
289 /* Remove old deref since it has the wrong type. */
290 nir_deref_instr_remove_if_unused(old_deref);
291 }
292 break;
293 }
294
295 case nir_intrinsic_load_deref:
296 if (nir_intrinsic_get_var(intrin, 0) == pos_var) {
297 unreachable("Should have lowered I/O to temporaries "
298 "so no load_deref on position output is expected.");
299 }
300 break;
301
302 case nir_intrinsic_copy_deref:
303 unreachable("Should have lowered copy_derefs at this point");
304 break;
305
306 default:
307 /* Do nothing. */
308 break;
309 }
310 }
311 }
312
313 nir_metadata_preserve(entrypoint, nir_metadata_none);
314 return true;
315 }
316