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