void (*debug_output)(const char *msg,
void *debug_output_data),
void *debug_output_data,
- int program_id, int variant_id)
+ int program_id, int variant_id,
+ bool fallback_scheduler)
{
struct v3d_compile *c = rzalloc(NULL, struct v3d_compile);
c->debug_output = debug_output;
c->debug_output_data = debug_output_data;
c->compilation_result = V3D_COMPILATION_SUCCEEDED;
+ c->fallback_scheduler = fallback_scheduler;
s = nir_shader_clone(c, s);
c->s = s;
}
/* CS textures may not have return_size reflecting the shadow state. */
- nir_foreach_variable(var, &c->s->uniforms) {
+ nir_foreach_uniform_variable(var, c->s) {
const struct glsl_type *type = glsl_without_array(var->type);
unsigned array_len = MAX2(glsl_get_length(var->type), 1);
used_outputs[comp] |= 1ull << slot;
}
NIR_PASS_V(c->s, nir_remove_unused_io_vars,
- &c->s->outputs, used_outputs, NULL); /* demotes to globals */
+ nir_var_shader_out, used_outputs, NULL); /* demotes to globals */
NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
v3d_optimize_nir(c->s);
NIR_PASS_V(c->s, nir_remove_dead_variables, nir_var_shader_in, NULL);
NIR_PASS_V(c->s, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
type_size_vec4,
(nir_lower_io_options)0);
- /* clean up nir_lower_io's deref_var remains */
+ /* clean up nir_lower_io's deref_var remains and do a constant folding pass
+ * on the code it generated.
+ */
NIR_PASS_V(c->s, nir_opt_dce);
+ NIR_PASS_V(c->s, nir_opt_constant_folding);
}
static void
used_outputs[comp] |= 1ull << slot;
}
NIR_PASS_V(c->s, nir_remove_unused_io_vars,
- &c->s->outputs, used_outputs, NULL); /* demotes to globals */
+ nir_var_shader_out, used_outputs, NULL); /* demotes to globals */
NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
v3d_optimize_nir(c->s);
NIR_PASS_V(c->s, nir_remove_dead_variables, nir_var_shader_in, NULL);
static void
v3d_fixup_fs_output_types(struct v3d_compile *c)
{
- nir_foreach_variable(var, &c->s->outputs) {
+ nir_foreach_shader_out_variable(var, c->s) {
uint32_t mask = 0;
switch (var->data.location) {
false, NULL);
}
+ /* In OpenGL the fragment shader can't read gl_ClipDistance[], but
+ * Vulkan allows it, in which case the SPIR-V compiler will declare
+ * VARING_SLOT_CLIP_DIST0 as compact array variable. Pass true as
+ * the last parameter to always operate with a compact array in both
+ * OpenGL and Vulkan so we do't have to care about the API we
+ * are using.
+ */
if (c->key->ucp_enables)
- NIR_PASS_V(c->s, nir_lower_clip_fs, c->key->ucp_enables,
- false);
+ NIR_PASS_V(c->s, nir_lower_clip_fs, c->key->ucp_enables, true);
/* Note: FS input scalarizing must happen after
* nir_lower_two_sided_color, which only handles a vec4 at a time.
return false;
}
-uint64_t *v3d_compile(const struct v3d_compiler *compiler,
- struct v3d_key *key,
- struct v3d_prog_data **out_prog_data,
- nir_shader *s,
- void (*debug_output)(const char *msg,
- void *debug_output_data),
- void *debug_output_data,
- int program_id, int variant_id,
- uint32_t *final_assembly_size)
+static void
+v3d_attempt_compile(struct v3d_compile *c)
{
- struct v3d_prog_data *prog_data;
- struct v3d_compile *c = vir_compile_init(compiler, key, s,
- debug_output, debug_output_data,
- program_id, variant_id);
-
switch (c->s->info.stage) {
case MESA_SHADER_VERTEX:
- c->vs_key = (struct v3d_vs_key *)key;
- prog_data = rzalloc_size(NULL, sizeof(struct v3d_vs_prog_data));
+ c->vs_key = (struct v3d_vs_key *) c->key;
break;
case MESA_SHADER_GEOMETRY:
- c->gs_key = (struct v3d_gs_key *)key;
- prog_data = rzalloc_size(NULL, sizeof(struct v3d_gs_prog_data));
+ c->gs_key = (struct v3d_gs_key *) c->key;
break;
case MESA_SHADER_FRAGMENT:
- c->fs_key = (struct v3d_fs_key *)key;
- prog_data = rzalloc_size(NULL, sizeof(struct v3d_fs_prog_data));
+ c->fs_key = (struct v3d_fs_key *) c->key;
break;
case MESA_SHADER_COMPUTE:
- prog_data = rzalloc_size(NULL,
- sizeof(struct v3d_compute_prog_data));
break;
default:
unreachable("unsupported shader stage");
}
-
switch (c->s->info.stage) {
case MESA_SHADER_VERTEX:
v3d_nir_lower_vs_early(c);
~((1 << MESA_SHADER_FRAGMENT) |
(1 << MESA_SHADER_GEOMETRY))),
+ .fallback = c->fallback_scheduler,
+
.intrinsic_cb = v3d_intrinsic_dependency_cb,
.intrinsic_cb_data = c,
};
NIR_PASS_V(c->s, nir_schedule, &schedule_options);
v3d_nir_to_vir(c);
+}
+
+uint32_t
+v3d_prog_data_size(gl_shader_stage stage)
+{
+ static const int prog_data_size[] = {
+ [MESA_SHADER_VERTEX] = sizeof(struct v3d_vs_prog_data),
+ [MESA_SHADER_GEOMETRY] = sizeof(struct v3d_gs_prog_data),
+ [MESA_SHADER_FRAGMENT] = sizeof(struct v3d_fs_prog_data),
+ [MESA_SHADER_COMPUTE] = sizeof(struct v3d_compute_prog_data),
+ };
+
+ assert(stage >= 0 &&
+ stage < ARRAY_SIZE(prog_data_size) &&
+ prog_data_size[stage]);
+
+ return prog_data_size[stage];
+}
+
+uint64_t *v3d_compile(const struct v3d_compiler *compiler,
+ struct v3d_key *key,
+ struct v3d_prog_data **out_prog_data,
+ nir_shader *s,
+ void (*debug_output)(const char *msg,
+ void *debug_output_data),
+ void *debug_output_data,
+ int program_id, int variant_id,
+ uint32_t *final_assembly_size)
+{
+ struct v3d_compile *c;
+
+ for (int i = 0; true; i++) {
+ c = vir_compile_init(compiler, key, s,
+ debug_output, debug_output_data,
+ program_id, variant_id,
+ i > 0 /* fallback_scheduler */);
+
+ v3d_attempt_compile(c);
+
+ if (i > 0 ||
+ c->compilation_result !=
+ V3D_COMPILATION_FAILED_REGISTER_ALLOCATION)
+ break;
+
+ char *debug_msg;
+ int ret = asprintf(&debug_msg,
+ "Using fallback scheduler for %s",
+ vir_get_stage_name(c));
+
+ if (ret >= 0) {
+ if (unlikely(V3D_DEBUG & V3D_DEBUG_PERF))
+ fprintf(stderr, "%s\n", debug_msg);
+
+ c->debug_output(debug_msg, c->debug_output_data);
+ free(debug_msg);
+ }
+
+ vir_compile_destroy(c);
+ }
+
+ struct v3d_prog_data *prog_data;
+
+ prog_data = rzalloc_size(NULL, v3d_prog_data_size(c->s->info.stage));
v3d_set_prog_data(c, prog_data);
projects / mesa.git / blobdiff