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
41static bool
42shader_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_store_shared2_amd:
81         case nir_intrinsic_shared_atomic_add:
82         case nir_intrinsic_shared_atomic_imin:
83         case nir_intrinsic_shared_atomic_umin:
84         case nir_intrinsic_shared_atomic_imax:
85         case nir_intrinsic_shared_atomic_umax:
86         case nir_intrinsic_shared_atomic_and:
87         case nir_intrinsic_shared_atomic_or:
88         case nir_intrinsic_shared_atomic_xor:
89         case nir_intrinsic_shared_atomic_exchange:
90         case nir_intrinsic_shared_atomic_comp_swap:
91         case nir_intrinsic_task_payload_atomic_add:
92         case nir_intrinsic_task_payload_atomic_imin:
93         case nir_intrinsic_task_payload_atomic_umin:
94         case nir_intrinsic_task_payload_atomic_imax:
95         case nir_intrinsic_task_payload_atomic_umax:
96         case nir_intrinsic_task_payload_atomic_and:
97         case nir_intrinsic_task_payload_atomic_or:
98         case nir_intrinsic_task_payload_atomic_xor:
99         case nir_intrinsic_task_payload_atomic_exchange:
100         case nir_intrinsic_task_payload_atomic_comp_swap:
101         case nir_intrinsic_task_payload_atomic_fadd:
102         case nir_intrinsic_task_payload_atomic_fmin:
103         case nir_intrinsic_task_payload_atomic_fmax:
104         case nir_intrinsic_task_payload_atomic_fcomp_swap:
105         case nir_intrinsic_image_deref_store:
106         case nir_intrinsic_image_deref_atomic_add:
107         case nir_intrinsic_image_deref_atomic_fadd:
108         case nir_intrinsic_image_deref_atomic_umin:
109         case nir_intrinsic_image_deref_atomic_umax:
110         case nir_intrinsic_image_deref_atomic_imin:
111         case nir_intrinsic_image_deref_atomic_imax:
112         case nir_intrinsic_image_deref_atomic_fmin:
113         case nir_intrinsic_image_deref_atomic_fmax:
114         case nir_intrinsic_image_deref_atomic_and:
115         case nir_intrinsic_image_deref_atomic_or:
116         case nir_intrinsic_image_deref_atomic_xor:
117         case nir_intrinsic_image_deref_atomic_exchange:
118         case nir_intrinsic_image_deref_atomic_comp_swap:
119            return true;
120
121         default:
122            /* Keep walking. */
123            break;
124         }
125      }
126   }
127
128   return false;
129}
130
131bool
132nir_shader_uses_view_index(nir_shader *shader)
133{
134   nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);
135
136   nir_foreach_block(block, entrypoint) {
137      nir_foreach_instr(instr, block) {
138         if (instr->type != nir_instr_type_intrinsic)
139            continue;
140
141         nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
142         if (intrin->intrinsic == nir_intrinsic_load_view_index)
143            return true;
144      }
145   }
146
147   return false;
148}
149
150static bool
151shader_only_position_uses_view_index(nir_shader *shader)
152{
153   nir_shader *shader_no_position = nir_shader_clone(NULL, shader);
154   nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader_no_position);
155
156   /* Remove the store position from a cloned shader. */
157   nir_foreach_block(block, entrypoint) {
158      nir_foreach_instr_safe(instr, block) {
159         if (instr->type != nir_instr_type_intrinsic)
160            continue;
161
162         nir_intrinsic_instr *store = nir_instr_as_intrinsic(instr);
163         if (store->intrinsic != nir_intrinsic_store_deref)
164            continue;
165
166         nir_variable *var = nir_intrinsic_get_var(store, 0);
167         if (var->data.location != VARYING_SLOT_POS)
168            continue;
169
170         nir_instr_remove(&store->instr);
171      }
172   }
173
174   /* Clean up shader so unused load_view_index intrinsics are removed. */
175   bool progress;
176   do {
177      progress = false;
178      progress |= nir_opt_dead_cf(shader_no_position);
179
180      /* Peephole select will drop if-blocks that have then and else empty,
181       * which will remove the usage of an SSA in the condition.
182       */
183      progress |= nir_opt_peephole_select(shader_no_position, 0, false, false);
184
185      progress |= nir_opt_dce(shader_no_position);
186   } while (progress);
187
188   bool uses_view_index = nir_shader_uses_view_index(shader_no_position);
189
190   ralloc_free(shader_no_position);
191   return !uses_view_index;
192}
193
194/* Return true if it's safe to call nir_lower_multiview() on this vertex
195 * shader. Note that this only handles driver-agnostic checks, i.e. things
196 * which would make nir_lower_multiview() incorrect. Any driver-specific
197 * checks, e.g. for sufficient varying space or performance considerations,
198 * should be handled in the driver.
199 *
200 * Note that we don't handle the more complex checks needed for lowering
201 * pipelines with geometry or tessellation shaders.
202 */
203
204bool
205nir_can_lower_multiview(nir_shader *shader)
206{
207   bool writes_position = false;
208   nir_foreach_shader_out_variable(var, shader) {
209      if (var->data.location == VARYING_SLOT_POS) {
210         writes_position = true;
211         break;
212      }
213   }
214
215   /* Don't bother handling this edge case. */
216   if (!writes_position)
217      return false;
218
219   return !shader_writes_to_memory(shader) &&
220          shader_only_position_uses_view_index(shader);
221}
222
223/**
224 * The lowering. Call with the last active geometry stage.
225 */
226
227bool
228nir_lower_multiview(nir_shader *shader, uint32_t view_mask)
229{
230   assert(shader->info.stage != MESA_SHADER_FRAGMENT);
231   int view_count = util_bitcount(view_mask);
232
233   nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);
234
235   /* Update position to refer to an array. */
236   nir_variable *pos_var = NULL;
237   nir_foreach_shader_out_variable(var, shader) {
238      if (var->data.location == VARYING_SLOT_POS) {
239         assert(var->type == glsl_vec4_type());
240         var->type = glsl_array_type(glsl_vec4_type(), view_count, 0);
241         var->data.per_view = true;
242         pos_var = var;
243         break;
244      }
245   }
246
247   assert(pos_var);
248
249   nir_cf_list body;
250   nir_cf_list_extract(&body, &entrypoint->body);
251
252   nir_builder b;
253   nir_builder_init(&b, entrypoint);
254   b.cursor = nir_after_cf_list(&entrypoint->body);
255
256   /* Loop Index will go from 0 to view_count. */
257   nir_variable *loop_index_var =
258      nir_local_variable_create(entrypoint, glsl_uint_type(), "loop_index");
259   nir_deref_instr *loop_index_deref = nir_build_deref_var(&b, loop_index_var);
260   nir_store_deref(&b, loop_index_deref, nir_imm_int(&b, 0), 1);
261
262   /* Array of view index values that are active in the loop.  Note that the
263    * loop index only matches the view index if there are no gaps in the
264    * view_mask.
265    */
266   nir_variable *view_index_var = nir_local_variable_create(
267      entrypoint, glsl_array_type(glsl_uint_type(), view_count, 0), "view_index");
268   nir_deref_instr *view_index_deref = nir_build_deref_var(&b, view_index_var);
269   {
270      int array_position = 0;
271      uint32_t view_mask_temp = view_mask;
272      while (view_mask_temp) {
273         uint32_t view_index = u_bit_scan(&view_mask_temp);
274         nir_store_deref(&b, nir_build_deref_array_imm(&b, view_index_deref, array_position),
275                         nir_imm_int(&b, view_index), 1);
276         array_position++;
277      }
278   }
279
280   /* Create the equivalent of
281    *
282    *    while (true):
283    *       if (loop_index >= view_count):
284    *          break
285    *
286    *       view_index = active_indices[loop_index]
287    *       pos_deref = &pos[loop_index]
288    *
289    *       # Placeholder for the body to be reinserted.
290    *
291    *       loop_index += 1
292    *
293    * Later both `view_index` and `pos_deref` will be used to rewrite the
294    * original shader body.
295    */
296
297   nir_loop* loop = nir_push_loop(&b);
298
299   nir_ssa_def *loop_index = nir_load_deref(&b, loop_index_deref);
300   nir_ssa_def *cmp = nir_ige(&b, loop_index, nir_imm_int(&b, view_count));
301   nir_if *loop_check = nir_push_if(&b, cmp);
302   nir_jump(&b, nir_jump_break);
303   nir_pop_if(&b, loop_check);
304
305   nir_ssa_def *view_index =
306      nir_load_deref(&b, nir_build_deref_array(&b, view_index_deref, loop_index));
307   nir_deref_instr *pos_deref =
308      nir_build_deref_array(&b, nir_build_deref_var(&b, pos_var), loop_index);
309
310   nir_store_deref(&b, loop_index_deref, nir_iadd_imm(&b, loop_index, 1), 1);
311   nir_pop_loop(&b, loop);
312
313   /* Reinsert the body. */
314   b.cursor = nir_after_instr(&pos_deref->instr);
315   nir_cf_reinsert(&body, b.cursor);
316
317   nir_foreach_block(block, entrypoint) {
318      nir_foreach_instr_safe(instr, block) {
319         if (instr->type != nir_instr_type_intrinsic)
320            continue;
321
322         nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
323
324         switch (intrin->intrinsic) {
325         case nir_intrinsic_load_view_index: {
326            assert(intrin->dest.is_ssa);
327            nir_ssa_def_rewrite_uses(&intrin->dest.ssa, view_index);
328            break;
329         }
330
331         case nir_intrinsic_store_deref: {
332            nir_variable *var = nir_intrinsic_get_var(intrin, 0);
333            if (var == pos_var) {
334               nir_deref_instr *old_deref = nir_src_as_deref(intrin->src[0]);
335
336               nir_instr_rewrite_src(instr, &intrin->src[0],
337                                     nir_src_for_ssa(&pos_deref->dest.ssa));
338
339               /* Remove old deref since it has the wrong type. */
340               nir_deref_instr_remove_if_unused(old_deref);
341            }
342            break;
343         }
344
345         case nir_intrinsic_load_deref:
346            if (nir_intrinsic_get_var(intrin, 0) == pos_var) {
347               unreachable("Should have lowered I/O to temporaries "
348                           "so no load_deref on position output is expected.");
349            }
350            break;
351
352         case nir_intrinsic_copy_deref:
353            unreachable("Should have lowered copy_derefs at this point");
354            break;
355
356         default:
357            /* Do nothing. */
358            break;
359         }
360      }
361   }
362
363   nir_metadata_preserve(entrypoint, nir_metadata_none);
364   return true;
365}
366
367