1/*
2 * Copyright © 2016 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24#include "anv_nir.h"
25#include "nir/nir_builder.h"
26#include "util/debug.h"
27
28/**
29 * This file implements the lowering required for VK_KHR_multiview.
30 *
31 * When possible, Primitive Replication is used and the shader is modified to
32 * make gl_Position an array and fill it with values for each view.
33 *
34 * Otherwise we implement multiview using instanced rendering.  The number of
35 * instances in each draw call is multiplied by the number of views in the
36 * subpass.  Then, in the shader, we divide gl_InstanceId by the number of
37 * views and use gl_InstanceId % view_count to compute the actual ViewIndex.
38 */
39
40struct lower_multiview_state {
41   nir_builder builder;
42
43   uint32_t view_mask;
44
45   nir_ssa_def *instance_id;
46   nir_ssa_def *view_index;
47};
48
49static nir_ssa_def *
50build_instance_id(struct lower_multiview_state *state)
51{
52   assert(state->builder.shader->info.stage == MESA_SHADER_VERTEX);
53
54   if (state->instance_id == NULL) {
55      nir_builder *b = &state->builder;
56
57      b->cursor = nir_before_block(nir_start_block(b->impl));
58
59      /* We use instancing for implementing multiview.  The actual instance id
60       * is given by dividing instance_id by the number of views in this
61       * subpass.
62       */
63      state->instance_id =
64         nir_idiv(b, nir_load_instance_id(b),
65                     nir_imm_int(b, util_bitcount(state->view_mask)));
66   }
67
68   return state->instance_id;
69}
70
71static nir_ssa_def *
72build_view_index(struct lower_multiview_state *state)
73{
74   if (state->view_index == NULL) {
75      nir_builder *b = &state->builder;
76
77      b->cursor = nir_before_block(nir_start_block(b->impl));
78
79      assert(state->view_mask != 0);
80      if (util_bitcount(state->view_mask) == 1) {
81         /* Set the view index directly. */
82         state->view_index = nir_imm_int(b, ffs(state->view_mask) - 1);
83      } else if (state->builder.shader->info.stage == MESA_SHADER_VERTEX) {
84         /* We only support 16 viewports */
85         assert((state->view_mask & 0xffff0000) == 0);
86
87         /* We use instancing for implementing multiview.  The compacted view
88          * id is given by instance_id % view_count.  We then have to convert
89          * that to an actual view id.
90          */
91         nir_ssa_def *compacted =
92            nir_umod(b, nir_load_instance_id(b),
93                        nir_imm_int(b, util_bitcount(state->view_mask)));
94
95         if (util_is_power_of_two_or_zero(state->view_mask + 1)) {
96            /* If we have a full view mask, then compacted is what we want */
97            state->view_index = compacted;
98         } else {
99            /* Now we define a map from compacted view index to the actual
100             * view index that's based on the view_mask.  The map is given by
101             * 16 nibbles, each of which is a value from 0 to 15.
102             */
103            uint64_t remap = 0;
104            uint32_t i = 0;
105            u_foreach_bit(bit, state->view_mask) {
106               assert(bit < 16);
107               remap |= (uint64_t)bit << (i++ * 4);
108            }
109
110            nir_ssa_def *shift = nir_imul(b, compacted, nir_imm_int(b, 4));
111
112            /* One of these days, when we have int64 everywhere, this will be
113             * easier.
114             */
115            nir_ssa_def *shifted;
116            if (remap <= UINT32_MAX) {
117               shifted = nir_ushr(b, nir_imm_int(b, remap), shift);
118            } else {
119               nir_ssa_def *shifted_low =
120                  nir_ushr(b, nir_imm_int(b, remap), shift);
121               nir_ssa_def *shifted_high =
122                  nir_ushr(b, nir_imm_int(b, remap >> 32),
123                              nir_isub(b, shift, nir_imm_int(b, 32)));
124               shifted = nir_bcsel(b, nir_ilt(b, shift, nir_imm_int(b, 32)),
125                                      shifted_low, shifted_high);
126            }
127            state->view_index = nir_iand(b, shifted, nir_imm_int(b, 0xf));
128         }
129      } else {
130         const struct glsl_type *type = glsl_int_type();
131         if (b->shader->info.stage == MESA_SHADER_TESS_CTRL ||
132             b->shader->info.stage == MESA_SHADER_GEOMETRY)
133            type = glsl_array_type(type, 1, 0);
134
135         nir_variable *idx_var =
136            nir_variable_create(b->shader, nir_var_shader_in,
137                                type, "view index");
138         idx_var->data.location = VARYING_SLOT_VIEW_INDEX;
139         if (b->shader->info.stage == MESA_SHADER_FRAGMENT)
140            idx_var->data.interpolation = INTERP_MODE_FLAT;
141
142         nir_deref_instr *deref = nir_build_deref_var(b, idx_var);
143         if (glsl_type_is_array(type))
144            deref = nir_build_deref_array_imm(b, deref, 0);
145
146         state->view_index = nir_load_deref(b, deref);
147      }
148   }
149
150   return state->view_index;
151}
152
153static bool
154is_load_view_index(const nir_instr *instr, const void *data)
155{
156   return instr->type == nir_instr_type_intrinsic &&
157          nir_instr_as_intrinsic(instr)->intrinsic == nir_intrinsic_load_view_index;
158}
159
160static nir_ssa_def *
161replace_load_view_index_with_zero(struct nir_builder *b,
162                                  nir_instr *instr, void *data)
163{
164   assert(is_load_view_index(instr, data));
165   return nir_imm_zero(b, 1, 32);
166}
167
168bool
169anv_nir_lower_multiview(nir_shader *shader,
170                        struct anv_graphics_pipeline *pipeline)
171{
172   assert(shader->info.stage != MESA_SHADER_COMPUTE);
173   uint32_t view_mask = pipeline->view_mask;
174
175   /* If multiview isn't enabled, just lower the ViewIndex builtin to zero. */
176   if (view_mask == 0) {
177      return nir_shader_lower_instructions(shader, is_load_view_index,
178                                           replace_load_view_index_with_zero, NULL);
179   }
180
181   /* This pass assumes a single entrypoint */
182   nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);
183
184   /* Primitive Replication allows a shader to write different positions for
185    * each view in the same execution. If only the position depends on the
186    * view, then it is possible to use the feature instead of instancing to
187    * implement multiview.
188    */
189   if (pipeline->use_primitive_replication) {
190      if (shader->info.stage == MESA_SHADER_FRAGMENT)
191         return false;
192
193      bool progress = nir_lower_multiview(shader, pipeline->view_mask);
194
195      if (progress) {
196         nir_builder b;
197         nir_builder_init(&b, entrypoint);
198         b.cursor = nir_before_cf_list(&entrypoint->body);
199
200         /* Fill Layer ID with zero. Replication will use that as base to
201          * apply the RTAI offsets.
202          */
203         nir_variable *layer_id_out =
204            nir_variable_create(shader, nir_var_shader_out,
205                                glsl_int_type(), "layer ID");
206         layer_id_out->data.location = VARYING_SLOT_LAYER;
207         nir_store_var(&b, layer_id_out, nir_imm_zero(&b, 1, 32), 0x1);
208      }
209
210      return progress;
211   }
212
213   struct lower_multiview_state state = {
214      .view_mask = view_mask,
215   };
216
217   nir_builder_init(&state.builder, entrypoint);
218
219   bool progress = false;
220   nir_foreach_block(block, entrypoint) {
221      nir_foreach_instr_safe(instr, block) {
222         if (instr->type != nir_instr_type_intrinsic)
223            continue;
224
225         nir_intrinsic_instr *load = nir_instr_as_intrinsic(instr);
226
227         if (load->intrinsic != nir_intrinsic_load_instance_id &&
228             load->intrinsic != nir_intrinsic_load_view_index)
229            continue;
230
231         assert(load->dest.is_ssa);
232
233         nir_ssa_def *value;
234         if (load->intrinsic == nir_intrinsic_load_instance_id) {
235            value = build_instance_id(&state);
236         } else {
237            assert(load->intrinsic == nir_intrinsic_load_view_index);
238            value = build_view_index(&state);
239         }
240
241         nir_ssa_def_rewrite_uses(&load->dest.ssa, value);
242
243         nir_instr_remove(&load->instr);
244         progress = true;
245      }
246   }
247
248   /* The view index is available in all stages but the instance id is only
249    * available in the VS.  If it's not a fragment shader, we need to pass
250    * the view index on to the next stage.
251    */
252   if (shader->info.stage != MESA_SHADER_FRAGMENT) {
253      nir_ssa_def *view_index = build_view_index(&state);
254
255      nir_builder *b = &state.builder;
256
257      assert(view_index->parent_instr->block == nir_start_block(entrypoint));
258      b->cursor = nir_after_instr(view_index->parent_instr);
259
260      /* Unless there is only one possible view index (that would be set
261       * directly), pass it to the next stage. */
262      if (util_bitcount(state.view_mask) != 1) {
263         nir_variable *view_index_out =
264            nir_variable_create(shader, nir_var_shader_out,
265                                glsl_int_type(), "view index");
266         view_index_out->data.location = VARYING_SLOT_VIEW_INDEX;
267         nir_store_var(b, view_index_out, view_index, 0x1);
268      }
269
270      nir_variable *layer_id_out =
271         nir_variable_create(shader, nir_var_shader_out,
272                             glsl_int_type(), "layer ID");
273      layer_id_out->data.location = VARYING_SLOT_LAYER;
274      nir_store_var(b, layer_id_out, view_index, 0x1);
275
276      progress = true;
277   }
278
279   if (progress) {
280      nir_metadata_preserve(entrypoint, nir_metadata_block_index |
281                                        nir_metadata_dominance);
282   } else {
283      nir_metadata_preserve(entrypoint, nir_metadata_all);
284   }
285
286   return progress;
287}
288
289bool
290anv_check_for_primitive_replication(nir_shader **shaders,
291                                    struct anv_graphics_pipeline *pipeline)
292{
293   assert(pipeline->base.device->info.ver >= 12);
294
295   static int primitive_replication_max_views = -1;
296   if (primitive_replication_max_views < 0) {
297      /* TODO: Figure out why we are not getting same benefits for larger than
298       * 2 views.  For now use Primitive Replication just for the 2-view case
299       * by default.
300       */
301      const unsigned default_max_views = 2;
302
303      primitive_replication_max_views =
304         MIN2(MAX_VIEWS_FOR_PRIMITIVE_REPLICATION,
305              env_var_as_unsigned("ANV_PRIMITIVE_REPLICATION_MAX_VIEWS",
306                                  default_max_views));
307   }
308
309   /* TODO: We should be able to support replication at 'geometry' stages
310    * later than Vertex.  In that case only the last stage can refer to
311    * gl_ViewIndex.
312    */
313   if (pipeline->active_stages != (VK_SHADER_STAGE_VERTEX_BIT |
314                                   VK_SHADER_STAGE_FRAGMENT_BIT)) {
315      return false;
316   }
317
318   uint32_t view_mask = pipeline->view_mask;
319   int view_count = util_bitcount(view_mask);
320   if (view_count == 1 || view_count > primitive_replication_max_views)
321      return false;
322
323   /* We can't access the view index in the fragment shader. */
324   if (nir_shader_uses_view_index(shaders[MESA_SHADER_FRAGMENT]))
325      return false;
326
327   return nir_can_lower_multiview(shaders[MESA_SHADER_VERTEX]);
328}
329