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_fadd:
93 case nir_intrinsic_image_deref_atomic_umin:
94 case nir_intrinsic_image_deref_atomic_umax:
95 case nir_intrinsic_image_deref_atomic_imin:
96 case nir_intrinsic_image_deref_atomic_imax:
97 case nir_intrinsic_image_deref_atomic_fmin:
98 case nir_intrinsic_image_deref_atomic_fmax:
99 case nir_intrinsic_image_deref_atomic_and:
100 case nir_intrinsic_image_deref_atomic_or:
101 case nir_intrinsic_image_deref_atomic_xor:
102 case nir_intrinsic_image_deref_atomic_exchange:
103 case nir_intrinsic_image_deref_atomic_comp_swap:
104 return true;
105
106 default:
107 /* Keep walking. */
108 break;
109 }
110 }
111 }
112
113 return false;
114 }
115
116 bool
nir_shader_uses_view_index(nir_shader * shader)117 nir_shader_uses_view_index(nir_shader *shader)
118 {
119 nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);
120
121 nir_foreach_block(block, entrypoint) {
122 nir_foreach_instr(instr, block) {
123 if (instr->type != nir_instr_type_intrinsic)
124 continue;
125
126 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
127 if (intrin->intrinsic == nir_intrinsic_load_view_index)
128 return true;
129 }
130 }
131
132 return false;
133 }
134
135 static bool
shader_only_position_uses_view_index(nir_shader * shader)136 shader_only_position_uses_view_index(nir_shader *shader)
137 {
138 nir_shader *shader_no_position = nir_shader_clone(NULL, shader);
139 nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader_no_position);
140
141 /* Remove the store position from a cloned shader. */
142 nir_foreach_block(block, entrypoint) {
143 nir_foreach_instr_safe(instr, block) {
144 if (instr->type != nir_instr_type_intrinsic)
145 continue;
146
147 nir_intrinsic_instr *store = nir_instr_as_intrinsic(instr);
148 if (store->intrinsic != nir_intrinsic_store_deref)
149 continue;
150
151 nir_variable *var = nir_intrinsic_get_var(store, 0);
152 if (var->data.location != VARYING_SLOT_POS)
153 continue;
154
155 nir_instr_remove(&store->instr);
156 }
157 }
158
159 /* Clean up shader so unused load_view_index intrinsics are removed. */
160 bool progress;
161 do {
162 progress = false;
163 progress |= nir_opt_dead_cf(shader_no_position);
164
165 /* Peephole select will drop if-blocks that have then and else empty,
166 * which will remove the usage of an SSA in the condition.
167 */
168 progress |= nir_opt_peephole_select(shader_no_position, 0, false, false);
169
170 progress |= nir_opt_dce(shader_no_position);
171 } while (progress);
172
173 bool uses_view_index = nir_shader_uses_view_index(shader_no_position);
174
175 ralloc_free(shader_no_position);
176 return !uses_view_index;
177 }
178
179 /* Return true if it's safe to call nir_lower_multiview() on this vertex
180 * shader. Note that this only handles driver-agnostic checks, i.e. things
181 * which would make nir_lower_multiview() incorrect. Any driver-specific
182 * checks, e.g. for sufficient varying space or performance considerations,
183 * should be handled in the driver.
184 *
185 * Note that we don't handle the more complex checks needed for lowering
186 * pipelines with geometry or tessellation shaders.
187 */
188
189 bool
nir_can_lower_multiview(nir_shader * shader)190 nir_can_lower_multiview(nir_shader *shader)
191 {
192 bool writes_position = false;
193 nir_foreach_shader_out_variable(var, shader) {
194 if (var->data.location == VARYING_SLOT_POS) {
195 writes_position = true;
196 break;
197 }
198 }
199
200 /* Don't bother handling this edge case. */
201 if (!writes_position)
202 return false;
203
204 return !shader_writes_to_memory(shader) &&
205 shader_only_position_uses_view_index(shader);
206 }
207
208 /**
209 * The lowering. Call with the last active geometry stage.
210 */
211
212 bool
nir_lower_multiview(nir_shader * shader,uint32_t view_mask)213 nir_lower_multiview(nir_shader *shader, uint32_t view_mask)
214 {
215 assert(shader->info.stage != MESA_SHADER_FRAGMENT);
216 int view_count = util_bitcount(view_mask);
217
218 nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);
219
220 /* Update position to refer to an array. */
221 nir_variable *pos_var = NULL;
222 nir_foreach_shader_out_variable(var, shader) {
223 if (var->data.location == VARYING_SLOT_POS) {
224 assert(var->type == glsl_vec4_type());
225 var->type = glsl_array_type(glsl_vec4_type(), view_count, 0);
226 var->data.per_view = true;
227 pos_var = var;
228 break;
229 }
230 }
231
232 assert(pos_var);
233
234 nir_cf_list body;
235 nir_cf_list_extract(&body, &entrypoint->body);
236
237 nir_builder b;
238 nir_builder_init(&b, entrypoint);
239 b.cursor = nir_after_cf_list(&entrypoint->body);
240
241 /* Loop Index will go from 0 to view_count. */
242 nir_variable *loop_index_var =
243 nir_local_variable_create(entrypoint, glsl_uint_type(), "loop_index");
244 nir_deref_instr *loop_index_deref = nir_build_deref_var(&b, loop_index_var);
245 nir_store_deref(&b, loop_index_deref, nir_imm_int(&b, 0), 1);
246
247 /* Array of view index values that are active in the loop. Note that the
248 * loop index only matches the view index if there are no gaps in the
249 * view_mask.
250 */
251 nir_variable *view_index_var = nir_local_variable_create(
252 entrypoint, glsl_array_type(glsl_uint_type(), view_count, 0), "view_index");
253 nir_deref_instr *view_index_deref = nir_build_deref_var(&b, view_index_var);
254 {
255 int array_position = 0;
256 uint32_t view_mask_temp = view_mask;
257 while (view_mask_temp) {
258 uint32_t view_index = u_bit_scan(&view_mask_temp);
259 nir_store_deref(&b, nir_build_deref_array_imm(&b, view_index_deref, array_position),
260 nir_imm_int(&b, view_index), 1);
261 array_position++;
262 }
263 }
264
265 /* Create the equivalent of
266 *
267 * while (true):
268 * if (loop_index >= view_count):
269 * break
270 *
271 * view_index = active_indices[loop_index]
272 * pos_deref = &pos[loop_index]
273 *
274 * # Placeholder for the body to be reinserted.
275 *
276 * loop_index += 1
277 *
278 * Later both `view_index` and `pos_deref` will be used to rewrite the
279 * original shader body.
280 */
281
282 nir_loop* loop = nir_push_loop(&b);
283
284 nir_ssa_def *loop_index = nir_load_deref(&b, loop_index_deref);
285 nir_ssa_def *cmp = nir_ige(&b, loop_index, nir_imm_int(&b, view_count));
286 nir_if *loop_check = nir_push_if(&b, cmp);
287 nir_jump(&b, nir_jump_break);
288 nir_pop_if(&b, loop_check);
289
290 nir_ssa_def *view_index =
291 nir_load_deref(&b, nir_build_deref_array(&b, view_index_deref, loop_index));
292 nir_deref_instr *pos_deref =
293 nir_build_deref_array(&b, nir_build_deref_var(&b, pos_var), loop_index);
294
295 nir_store_deref(&b, loop_index_deref, nir_iadd_imm(&b, loop_index, 1), 1);
296 nir_pop_loop(&b, loop);
297
298 /* Reinsert the body. */
299 b.cursor = nir_after_instr(&pos_deref->instr);
300 nir_cf_reinsert(&body, b.cursor);
301
302 nir_foreach_block(block, entrypoint) {
303 nir_foreach_instr_safe(instr, block) {
304 if (instr->type != nir_instr_type_intrinsic)
305 continue;
306
307 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
308
309 switch (intrin->intrinsic) {
310 case nir_intrinsic_load_view_index: {
311 assert(intrin->dest.is_ssa);
312 nir_ssa_def_rewrite_uses(&intrin->dest.ssa, view_index);
313 break;
314 }
315
316 case nir_intrinsic_store_deref: {
317 nir_variable *var = nir_intrinsic_get_var(intrin, 0);
318 if (var == pos_var) {
319 nir_deref_instr *old_deref = nir_src_as_deref(intrin->src[0]);
320
321 nir_instr_rewrite_src(instr, &intrin->src[0],
322 nir_src_for_ssa(&pos_deref->dest.ssa));
323
324 /* Remove old deref since it has the wrong type. */
325 nir_deref_instr_remove_if_unused(old_deref);
326 }
327 break;
328 }
329
330 case nir_intrinsic_load_deref:
331 if (nir_intrinsic_get_var(intrin, 0) == pos_var) {
332 unreachable("Should have lowered I/O to temporaries "
333 "so no load_deref on position output is expected.");
334 }
335 break;
336
337 case nir_intrinsic_copy_deref:
338 unreachable("Should have lowered copy_derefs at this point");
339 break;
340
341 default:
342 /* Do nothing. */
343 break;
344 }
345 }
346 }
347
348 nir_metadata_preserve(entrypoint, nir_metadata_none);
349 return true;
350 }
351
352