1 /*
2 * Copyright © 2020 Valve Corporation
3 * SPDX-License-Identifier: MIT
4 */
5
6 #include "tu_shader.h"
7
8 #include "nir_builder.h"
9
10 #include "tu_device.h"
11
12 /* Some a6xx variants cannot support a non-contiguous multiview mask. Instead,
13 * inside the shader something like this needs to be inserted:
14 *
15 * gl_Position = ((1ull << gl_ViewIndex) & view_mask) ? gl_Position : vec4(0.);
16 *
17 * Scan backwards until we find the gl_Position write (there should only be
18 * one).
19 */
20 static bool
lower_multiview_mask(nir_shader * nir,uint32_t * mask)21 lower_multiview_mask(nir_shader *nir, uint32_t *mask)
22 {
23 nir_function_impl *impl = nir_shader_get_entrypoint(nir);
24
25 if (util_is_power_of_two_or_zero(*mask + 1)) {
26 nir_metadata_preserve(impl, nir_metadata_all);
27 return false;
28 }
29
30 nir_builder b = nir_builder_create(impl);
31
32 uint32_t old_mask = *mask;
33 *mask = BIT(util_logbase2(old_mask) + 1) - 1;
34
35 nir_foreach_block_reverse(block, impl) {
36 nir_foreach_instr_reverse(instr, block) {
37 if (instr->type != nir_instr_type_intrinsic)
38 continue;
39
40 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
41 if (intrin->intrinsic != nir_intrinsic_store_deref)
42 continue;
43
44 nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
45 if (!nir_deref_mode_is(deref, nir_var_shader_out))
46 continue;
47
48 nir_variable *var = nir_deref_instr_get_variable(deref);
49 if (var->data.location != VARYING_SLOT_POS)
50 continue;
51
52 nir_def *orig_src = intrin->src[1].ssa;
53 b.cursor = nir_before_instr(instr);
54
55 /* ((1ull << gl_ViewIndex) & mask) != 0 */
56 nir_def *cmp =
57 nir_i2b(&b, nir_iand(&b, nir_imm_int(&b, old_mask),
58 nir_ishl(&b, nir_imm_int(&b, 1),
59 nir_load_view_index(&b))));
60
61 nir_def *src = nir_bcsel(&b, cmp, orig_src, nir_imm_float(&b, 0.));
62 nir_src_rewrite(&intrin->src[1], src);
63
64 nir_metadata_preserve(impl, nir_metadata_block_index |
65 nir_metadata_dominance);
66 return true;
67 }
68 }
69
70 nir_metadata_preserve(impl, nir_metadata_all);
71 return false;
72 }
73
74 bool
tu_nir_lower_multiview(nir_shader * nir,uint32_t mask,struct tu_device * dev)75 tu_nir_lower_multiview(nir_shader *nir, uint32_t mask, struct tu_device *dev)
76 {
77 bool progress = false;
78
79 if (!dev->physical_device->info->a6xx.supports_multiview_mask)
80 NIR_PASS(progress, nir, lower_multiview_mask, &mask);
81
82 unsigned num_views = util_logbase2(mask) + 1;
83
84 /* Blob doesn't apply multipos optimization starting from 11 views
85 * even on a650, however in practice, with the limit of 16 views,
86 * tests pass on a640/a650 and fail on a630.
87 */
88 unsigned max_views_for_multipos =
89 dev->physical_device->info->a6xx.supports_multiview_mask ? 16 : 10;
90
91 /* Speculatively assign output locations so that we know num_outputs. We
92 * will assign output locations for real after this pass.
93 */
94 unsigned num_outputs;
95 nir_assign_io_var_locations(nir, nir_var_shader_out, &num_outputs, MESA_SHADER_VERTEX);
96
97 /* In addition to the generic checks done by NIR, check that we don't
98 * overflow VPC with the extra copies of gl_Position.
99 */
100 if (!TU_DEBUG(NOMULTIPOS) &&
101 num_views <= max_views_for_multipos && num_outputs + (num_views - 1) <= 32 &&
102 nir_can_lower_multiview(nir)) {
103 /* It appears that the multiview mask is ignored when multi-position
104 * output is enabled, so we have to write 0 to inactive views ourselves.
105 */
106 NIR_PASS(progress, nir, lower_multiview_mask, &mask);
107
108 NIR_PASS_V(nir, nir_lower_multiview, mask);
109 progress = true;
110 }
111
112 return progress;
113 }
114
115