1 /*
2 * Copyright © 2015 Intel Corporation
3 * Copyright © 2022 Collabora, LTD
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 "vk_nir.h"
26
27 #include "compiler/nir/nir_xfb_info.h"
28 #include "compiler/spirv/nir_spirv.h"
29 #include "vk_device.h"
30 #include "vk_log.h"
31 #include "vk_physical_device.h"
32 #include "vk_util.h"
33
34 #define SPIR_V_MAGIC_NUMBER 0x07230203
35
36 uint32_t
vk_spirv_version(const uint32_t * spirv_data,size_t spirv_size_B)37 vk_spirv_version(const uint32_t *spirv_data, size_t spirv_size_B)
38 {
39 assert(spirv_size_B >= 8);
40 assert(spirv_data[0] == SPIR_V_MAGIC_NUMBER);
41 return spirv_data[1];
42 }
43
44 static void
spirv_nir_debug(void * private_data,enum nir_spirv_debug_level level,size_t spirv_offset,const char * message)45 spirv_nir_debug(void *private_data,
46 enum nir_spirv_debug_level level,
47 size_t spirv_offset,
48 const char *message)
49 {
50 const struct vk_object_base *log_obj = private_data;
51
52 switch (level) {
53 case NIR_SPIRV_DEBUG_LEVEL_INFO:
54 //vk_logi(VK_LOG_OBJS(log_obj), "SPIR-V offset %lu: %s",
55 // (unsigned long) spirv_offset, message);
56 break;
57 case NIR_SPIRV_DEBUG_LEVEL_WARNING:
58 vk_logw(VK_LOG_OBJS(log_obj), "SPIR-V offset %lu: %s",
59 (unsigned long) spirv_offset, message);
60 break;
61 case NIR_SPIRV_DEBUG_LEVEL_ERROR:
62 vk_loge(VK_LOG_OBJS(log_obj), "SPIR-V offset %lu: %s",
63 (unsigned long) spirv_offset, message);
64 break;
65 default:
66 break;
67 }
68 }
69
70 bool
nir_vk_is_not_xfb_output(nir_variable * var,void * data)71 nir_vk_is_not_xfb_output(nir_variable *var, void *data)
72 {
73 if (var->data.mode != nir_var_shader_out)
74 return true;
75
76 /* From the Vulkan 1.3.259 spec:
77 *
78 * VUID-StandaloneSpirv-Offset-04716
79 *
80 * "Only variables or block members in the output interface decorated
81 * with Offset can be captured for transform feedback, and those
82 * variables or block members must also be decorated with XfbBuffer
83 * and XfbStride, or inherit XfbBuffer and XfbStride decorations from
84 * a block containing them"
85 *
86 * glslang generates gl_PerVertex builtins when they are not declared,
87 * enabled XFB should not prevent them from being DCE'd.
88 *
89 * The logic should match nir_gather_xfb_info_with_varyings
90 */
91
92 if (!var->data.explicit_xfb_buffer)
93 return true;
94
95 bool is_array_block = var->interface_type != NULL &&
96 glsl_type_is_array(var->type) &&
97 glsl_without_array(var->type) == var->interface_type;
98
99 if (!is_array_block) {
100 return !var->data.explicit_offset;
101 } else {
102 /* For array of blocks we have to check each element */
103 unsigned aoa_size = glsl_get_aoa_size(var->type);
104 const struct glsl_type *itype = var->interface_type;
105 unsigned nfields = glsl_get_length(itype);
106 for (unsigned b = 0; b < aoa_size; b++) {
107 for (unsigned f = 0; f < nfields; f++) {
108 if (glsl_get_struct_field_offset(itype, f) >= 0)
109 return false;
110 }
111 }
112
113 return true;
114 }
115 }
116
117 nir_shader *
vk_spirv_to_nir(struct vk_device * device,const uint32_t * spirv_data,size_t spirv_size_B,gl_shader_stage stage,const char * entrypoint_name,enum gl_subgroup_size subgroup_size,const VkSpecializationInfo * spec_info,const struct spirv_to_nir_options * spirv_options,const struct nir_shader_compiler_options * nir_options,bool internal,void * mem_ctx)118 vk_spirv_to_nir(struct vk_device *device,
119 const uint32_t *spirv_data, size_t spirv_size_B,
120 gl_shader_stage stage, const char *entrypoint_name,
121 enum gl_subgroup_size subgroup_size,
122 const VkSpecializationInfo *spec_info,
123 const struct spirv_to_nir_options *spirv_options,
124 const struct nir_shader_compiler_options *nir_options,
125 bool internal,
126 void *mem_ctx)
127 {
128 assert(spirv_size_B >= 4 && spirv_size_B % 4 == 0);
129 assert(spirv_data[0] == SPIR_V_MAGIC_NUMBER);
130
131 const struct spirv_capabilities spirv_caps =
132 vk_physical_device_get_spirv_capabilities(device->physical);
133
134 struct spirv_to_nir_options spirv_options_local = *spirv_options;
135 spirv_options_local.capabilities = &spirv_caps;
136 spirv_options_local.debug.func = spirv_nir_debug;
137 spirv_options_local.debug.private_data = (void *)device;
138 spirv_options_local.subgroup_size = subgroup_size;
139
140 uint32_t num_spec_entries = 0;
141 struct nir_spirv_specialization *spec_entries =
142 vk_spec_info_to_nir_spirv(spec_info, &num_spec_entries);
143
144 nir_shader *nir = spirv_to_nir(spirv_data, spirv_size_B / 4,
145 spec_entries, num_spec_entries,
146 stage, entrypoint_name,
147 &spirv_options_local, nir_options);
148 free(spec_entries);
149
150 if (nir == NULL)
151 return NULL;
152
153 assert(nir->info.stage == stage);
154 nir_validate_shader(nir, "after spirv_to_nir");
155 nir_validate_ssa_dominance(nir, "after spirv_to_nir");
156 if (mem_ctx != NULL)
157 ralloc_steal(mem_ctx, nir);
158
159 nir->info.internal = internal;
160
161 /* We have to lower away local constant initializers right before we
162 * inline functions. That way they get properly initialized at the top
163 * of the function and not at the top of its caller.
164 */
165 NIR_PASS_V(nir, nir_lower_variable_initializers, nir_var_function_temp);
166 NIR_PASS_V(nir, nir_lower_returns);
167 NIR_PASS_V(nir, nir_inline_functions);
168 NIR_PASS_V(nir, nir_copy_prop);
169 NIR_PASS_V(nir, nir_opt_deref);
170
171 /* Pick off the single entrypoint that we want */
172 nir_remove_non_entrypoints(nir);
173
174 /* Now that we've deleted all but the main function, we can go ahead and
175 * lower the rest of the constant initializers. We do this here so that
176 * nir_remove_dead_variables and split_per_member_structs below see the
177 * corresponding stores.
178 */
179 NIR_PASS_V(nir, nir_lower_variable_initializers, ~0);
180
181 /* Split member structs. We do this before lower_io_to_temporaries so that
182 * it doesn't lower system values to temporaries by accident.
183 */
184 NIR_PASS_V(nir, nir_split_var_copies);
185 NIR_PASS_V(nir, nir_split_per_member_structs);
186
187 nir_remove_dead_variables_options dead_vars_opts = {
188 .can_remove_var = nir_vk_is_not_xfb_output,
189 };
190 NIR_PASS_V(nir, nir_remove_dead_variables,
191 nir_var_shader_in | nir_var_shader_out | nir_var_system_value |
192 nir_var_shader_call_data | nir_var_ray_hit_attrib,
193 &dead_vars_opts);
194
195 /* This needs to happen after remove_dead_vars because GLSLang likes to
196 * insert dead clip/cull vars and we don't want to clip/cull based on
197 * uninitialized garbage.
198 */
199 NIR_PASS_V(nir, nir_lower_clip_cull_distance_arrays);
200
201 if (nir->info.stage == MESA_SHADER_VERTEX ||
202 nir->info.stage == MESA_SHADER_TESS_EVAL ||
203 nir->info.stage == MESA_SHADER_GEOMETRY)
204 NIR_PASS_V(nir, nir_shader_gather_xfb_info);
205
206 NIR_PASS_V(nir, nir_propagate_invariant, false);
207
208 return nir;
209 }
210