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_control_flow.h"
26 #include "nir_builder.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_add:
59 case nir_intrinsic_deref_atomic_imin:
60 case nir_intrinsic_deref_atomic_umin:
61 case nir_intrinsic_deref_atomic_imax:
62 case nir_intrinsic_deref_atomic_umax:
63 case nir_intrinsic_deref_atomic_and:
64 case nir_intrinsic_deref_atomic_or:
65 case nir_intrinsic_deref_atomic_xor:
66 case nir_intrinsic_deref_atomic_exchange:
67 case nir_intrinsic_deref_atomic_comp_swap:
68 case nir_intrinsic_store_ssbo:
69 case nir_intrinsic_ssbo_atomic_add:
70 case nir_intrinsic_ssbo_atomic_imin:
71 case nir_intrinsic_ssbo_atomic_umin:
72 case nir_intrinsic_ssbo_atomic_imax:
73 case nir_intrinsic_ssbo_atomic_umax:
74 case nir_intrinsic_ssbo_atomic_and:
75 case nir_intrinsic_ssbo_atomic_or:
76 case nir_intrinsic_ssbo_atomic_xor:
77 case nir_intrinsic_ssbo_atomic_exchange:
78 case nir_intrinsic_ssbo_atomic_comp_swap:
79 case nir_intrinsic_store_shared:
80 case nir_intrinsic_shared_atomic_add:
81 case nir_intrinsic_shared_atomic_imin:
82 case nir_intrinsic_shared_atomic_umin:
83 case nir_intrinsic_shared_atomic_imax:
84 case nir_intrinsic_shared_atomic_umax:
85 case nir_intrinsic_shared_atomic_and:
86 case nir_intrinsic_shared_atomic_or:
87 case nir_intrinsic_shared_atomic_xor:
88 case nir_intrinsic_shared_atomic_exchange:
89 case nir_intrinsic_shared_atomic_comp_swap:
90 case nir_intrinsic_image_deref_store:
91 case nir_intrinsic_image_deref_atomic_add:
92 case nir_intrinsic_image_deref_atomic_umin:
93 case nir_intrinsic_image_deref_atomic_umax:
94 case nir_intrinsic_image_deref_atomic_imin:
95 case nir_intrinsic_image_deref_atomic_imax:
96 case nir_intrinsic_image_deref_atomic_and:
97 case nir_intrinsic_image_deref_atomic_or:
98 case nir_intrinsic_image_deref_atomic_xor:
99 case nir_intrinsic_image_deref_atomic_exchange:
100 case nir_intrinsic_image_deref_atomic_comp_swap:
101 return true;
102
103 default:
104 /* Keep walking. */
105 break;
106 }
107 }
108 }
109
110 return false;
111 }
112
113 bool
nir_shader_uses_view_index(nir_shader * shader)114 nir_shader_uses_view_index(nir_shader *shader)
115 {
116 nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);
117
118 nir_foreach_block(block, entrypoint) {
119 nir_foreach_instr(instr, block) {
120 if (instr->type != nir_instr_type_intrinsic)
121 continue;
122
123 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
124 if (intrin->intrinsic == nir_intrinsic_load_view_index)
125 return true;
126 }
127 }
128
129 return false;
130 }
131
132 static bool
shader_only_position_uses_view_index(nir_shader * shader)133 shader_only_position_uses_view_index(nir_shader *shader)
134 {
135 nir_shader *shader_no_position = nir_shader_clone(NULL, shader);
136 nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader_no_position);
137
138 /* Remove the store position from a cloned shader. */
139 nir_foreach_block(block, entrypoint) {
140 nir_foreach_instr_safe(instr, block) {
141 if (instr->type != nir_instr_type_intrinsic)
142 continue;
143
144 nir_intrinsic_instr *store = nir_instr_as_intrinsic(instr);
145 if (store->intrinsic != nir_intrinsic_store_deref)
146 continue;
147
148 nir_variable *var = nir_intrinsic_get_var(store, 0);
149 if (var->data.location != VARYING_SLOT_POS)
150 continue;
151
152 nir_instr_remove(&store->instr);
153 }
154 }
155
156 /* Clean up shader so unused load_view_index intrinsics are removed. */
157 bool progress;
158 do {
159 progress = false;
160 progress |= nir_opt_dead_cf(shader_no_position);
161
162 /* Peephole select will drop if-blocks that have then and else empty,
163 * which will remove the usage of an SSA in the condition.
164 */
165 progress |= nir_opt_peephole_select(shader_no_position, 0, false, false);
166
167 progress |= nir_opt_dce(shader_no_position);
168 } while (progress);
169
170 bool uses_view_index = nir_shader_uses_view_index(shader_no_position);
171
172 ralloc_free(shader_no_position);
173 return !uses_view_index;
174 }
175
176 /* Return true if it's safe to call nir_lower_multiview() on this vertex
177 * shader. Note that this only handles driver-agnostic checks, i.e. things
178 * which would make nir_lower_multiview() incorrect. Any driver-specific
179 * checks, e.g. for sufficient varying space or performance considerations,
180 * should be handled in the driver.
181 *
182 * Note that we don't handle the more complex checks needed for lowering
183 * pipelines with geometry or tessellation shaders.
184 */
185
186 bool
nir_can_lower_multiview(nir_shader * shader)187 nir_can_lower_multiview(nir_shader *shader)
188 {
189 bool writes_position = false;
190 nir_foreach_shader_out_variable(var, shader) {
191 if (var->data.location == VARYING_SLOT_POS) {
192 writes_position = true;
193 break;
194 }
195 }
196
197 /* Don't bother handling this edge case. */
198 if (!writes_position)
199 return false;
200
201 return !shader_writes_to_memory(shader) &&
202 shader_only_position_uses_view_index(shader);
203 }
204
205 /**
206 * The lowering. Call with the last active geometry stage.
207 */
208
209 bool
nir_lower_multiview(nir_shader * shader,uint32_t view_mask)210 nir_lower_multiview(nir_shader *shader, uint32_t view_mask)
211 {
212 assert(shader->info.stage != MESA_SHADER_FRAGMENT);
213 int view_count = util_bitcount(view_mask);
214
215 nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);
216
217 /* Update position to refer to an array. */
218 nir_variable *pos_var = NULL;
219 nir_foreach_shader_out_variable(var, shader) {
220 if (var->data.location == VARYING_SLOT_POS) {
221 assert(var->type == glsl_vec4_type());
222 var->type = glsl_array_type(glsl_vec4_type(), view_count, 0);
223 var->data.per_view = true;
224 pos_var = var;
225 break;
226 }
227 }
228
229 assert(pos_var);
230
231 nir_cf_list body;
232 nir_cf_list_extract(&body, &entrypoint->body);
233
234 nir_builder b;
235 nir_builder_init(&b, entrypoint);
236 b.cursor = nir_after_cf_list(&entrypoint->body);
237
238 /* Loop Index will go from 0 to view_count. */
239 nir_variable *loop_index_var =
240 nir_local_variable_create(entrypoint, glsl_uint_type(), "loop_index");
241 nir_deref_instr *loop_index_deref = nir_build_deref_var(&b, loop_index_var);
242 nir_store_deref(&b, loop_index_deref, nir_imm_int(&b, 0), 1);
243
244 /* Array of view index values that are active in the loop. Note that the
245 * loop index only matches the view index if there are no gaps in the
246 * view_mask.
247 */
248 nir_variable *view_index_var = nir_local_variable_create(
249 entrypoint, glsl_array_type(glsl_uint_type(), view_count, 0), "view_index");
250 nir_deref_instr *view_index_deref = nir_build_deref_var(&b, view_index_var);
251 {
252 int array_position = 0;
253 uint32_t view_mask_temp = view_mask;
254 while (view_mask_temp) {
255 uint32_t view_index = u_bit_scan(&view_mask_temp);
256 nir_store_deref(&b, nir_build_deref_array_imm(&b, view_index_deref, array_position),
257 nir_imm_int(&b, view_index), 1);
258 array_position++;
259 }
260 }
261
262 /* Create the equivalent of
263 *
264 * while (true):
265 * if (loop_index >= view_count):
266 * break
267 *
268 * view_index = active_indices[loop_index]
269 * pos_deref = &pos[loop_index]
270 *
271 * # Placeholder for the body to be reinserted.
272 *
273 * loop_index += 1
274 *
275 * Later both `view_index` and `pos_deref` will be used to rewrite the
276 * original shader body.
277 */
278
279 nir_loop* loop = nir_push_loop(&b);
280
281 nir_ssa_def *loop_index = nir_load_deref(&b, loop_index_deref);
282 nir_ssa_def *cmp = nir_ige(&b, loop_index, nir_imm_int(&b, view_count));
283 nir_if *loop_check = nir_push_if(&b, cmp);
284 nir_jump(&b, nir_jump_break);
285 nir_pop_if(&b, loop_check);
286
287 nir_ssa_def *view_index =
288 nir_load_deref(&b, nir_build_deref_array(&b, view_index_deref, loop_index));
289 nir_deref_instr *pos_deref =
290 nir_build_deref_array(&b, nir_build_deref_var(&b, pos_var), loop_index);
291
292 nir_store_deref(&b, loop_index_deref, nir_iadd_imm(&b, loop_index, 1), 1);
293 nir_pop_loop(&b, loop);
294
295 /* Reinsert the body. */
296 b.cursor = nir_after_instr(&pos_deref->instr);
297 nir_cf_reinsert(&body, b.cursor);
298
299 nir_foreach_block(block, entrypoint) {
300 nir_foreach_instr_safe(instr, block) {
301 if (instr->type != nir_instr_type_intrinsic)
302 continue;
303
304 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
305
306 switch (intrin->intrinsic) {
307 case nir_intrinsic_load_view_index: {
308 assert(intrin->dest.is_ssa);
309 nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(view_index));
310 break;
311 }
312
313 case nir_intrinsic_store_deref: {
314 nir_variable *var = nir_intrinsic_get_var(intrin, 0);
315 if (var == pos_var) {
316 nir_deref_instr *old_deref = nir_src_as_deref(intrin->src[0]);
317
318 nir_instr_rewrite_src(instr, &intrin->src[0],
319 nir_src_for_ssa(&pos_deref->dest.ssa));
320
321 /* Remove old deref since it has the wrong type. */
322 nir_deref_instr_remove_if_unused(old_deref);
323 }
324 break;
325 }
326
327 case nir_intrinsic_load_deref:
328 if (nir_intrinsic_get_var(intrin, 0) == pos_var) {
329 unreachable("Should have lowered I/O to temporaries "
330 "so no load_deref on position output is expected.");
331 }
332 break;
333
334 case nir_intrinsic_copy_deref:
335 unreachable("Should have lowered copy_derefs at this point");
336 break;
337
338 default:
339 /* Do nothing. */
340 break;
341 }
342 }
343 }
344
345 nir_metadata_preserve(entrypoint, nir_metadata_none);
346 return true;
347 }
348
349