}
static unsigned
-map_add(struct tu_descriptor_map *map, int set, int binding)
+map_add(struct tu_descriptor_map *map, int set, int binding, int value)
{
unsigned index;
for (index = 0; index < map->num; index++) {
map->set[index] = set;
map->binding[index] = binding;
+ map->value[index] = value;
map->num = MAX2(map->num, index + 1);
return index;
}
static void
-lower_tex_src_to_offset(nir_tex_instr *instr, unsigned src_idx,
- struct tu_shader *shader, bool is_sampler)
+lower_tex_src_to_offset(nir_builder *b, nir_tex_instr *instr, unsigned src_idx,
+ struct tu_shader *shader)
{
- nir_deref_instr *deref =
- nir_instr_as_deref(instr->src[src_idx].src.ssa->parent_instr);
+ nir_ssa_def *index = NULL;
+ unsigned base_index = 0;
+ unsigned array_elements = 1;
+ nir_tex_src *src = &instr->src[src_idx];
+ bool is_sampler = src->src_type == nir_tex_src_sampler_deref;
+
+ /* We compute first the offsets */
+ nir_deref_instr *deref = nir_instr_as_deref(src->src.ssa->parent_instr);
+ while (deref->deref_type != nir_deref_type_var) {
+ assert(deref->parent.is_ssa);
+ nir_deref_instr *parent =
+ nir_instr_as_deref(deref->parent.ssa->parent_instr);
+
+ assert(deref->deref_type == nir_deref_type_array);
+
+ if (nir_src_is_const(deref->arr.index) && index == NULL) {
+ /* We're still building a direct index */
+ base_index += nir_src_as_uint(deref->arr.index) * array_elements;
+ } else {
+ if (index == NULL) {
+ /* We used to be direct but not anymore */
+ index = nir_imm_int(b, base_index);
+ base_index = 0;
+ }
- if (deref->deref_type != nir_deref_type_var) {
- tu_finishme("sampler array");
- return;
+ index = nir_iadd(b, index,
+ nir_imul(b, nir_imm_int(b, array_elements),
+ nir_ssa_for_src(b, deref->arr.index, 1)));
+ }
+
+ array_elements *= glsl_get_length(parent->type);
+
+ deref = parent;
}
+ if (index)
+ index = nir_umin(b, index, nir_imm_int(b, array_elements - 1));
+
+ /* We have the offsets, we apply them, rewriting the source or removing
+ * instr if needed
+ */
+ if (index) {
+ nir_instr_rewrite_src(&instr->instr, &src->src,
+ nir_src_for_ssa(index));
+
+ src->src_type = is_sampler ?
+ nir_tex_src_sampler_offset :
+ nir_tex_src_texture_offset;
+
+ instr->texture_array_size = array_elements;
+ } else {
+ nir_tex_instr_remove_src(instr, src_idx);
+ }
+
+ if (array_elements > 1)
+ tu_finishme("texture/sampler array");
+
if (is_sampler) {
instr->sampler_index = map_add(&shader->sampler_map,
deref->var->data.descriptor_set,
- deref->var->data.binding);
+ deref->var->data.binding,
+ 0);
+ instr->sampler_index += base_index;
} else {
instr->texture_index = map_add(&shader->texture_map,
deref->var->data.descriptor_set,
- deref->var->data.binding);
- instr->texture_array_size = 1;
+ deref->var->data.binding,
+ deref->var->data.index);
+ instr->texture_index += base_index;
+ instr->texture_array_size = array_elements;
}
-
- nir_tex_instr_remove_src(instr, src_idx);
}
static bool
-lower_sampler(nir_tex_instr *instr, struct tu_shader *shader)
+lower_sampler(nir_builder *b, nir_tex_instr *instr, struct tu_shader *shader)
{
int texture_idx =
nir_tex_instr_src_index(instr, nir_tex_src_texture_deref);
if (texture_idx >= 0)
- lower_tex_src_to_offset(instr, texture_idx, shader, false);
+ lower_tex_src_to_offset(b, instr, texture_idx, shader);
int sampler_idx =
nir_tex_instr_src_index(instr, nir_tex_src_sampler_deref);
if (sampler_idx >= 0)
- lower_tex_src_to_offset(instr, sampler_idx, shader, true);
+ lower_tex_src_to_offset(b, instr, sampler_idx, shader);
if (texture_idx < 0 && sampler_idx < 0)
return false;
lower_intrinsic(nir_builder *b, nir_intrinsic_instr *instr,
struct tu_shader *shader)
{
+ /* TODO: remove this when layered rendering is implemented */
+ if (instr->intrinsic == nir_intrinsic_load_layer_id) {
+ nir_ssa_def_rewrite_uses(&instr->dest.ssa,
+ nir_src_for_ssa(nir_imm_int(b, 0)));
+ nir_instr_remove(&instr->instr);
+ return true;
+ }
+
+ if (instr->intrinsic == nir_intrinsic_load_push_constant) {
+ /* note: ir3 wants load_ubo, not load_uniform */
+ assert(nir_intrinsic_base(instr) == 0);
+
+ nir_intrinsic_instr *load =
+ nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_ubo);
+ load->num_components = instr->num_components;
+ load->src[0] = nir_src_for_ssa(nir_imm_int(b, 0));
+ load->src[1] = instr->src[0];
+ nir_ssa_dest_init(&load->instr, &load->dest,
+ load->num_components, instr->dest.ssa.bit_size,
+ instr->dest.ssa.name);
+ nir_builder_instr_insert(b, &load->instr);
+ nir_ssa_def_rewrite_uses(&instr->dest.ssa, nir_src_for_ssa(&load->dest.ssa));
+
+ nir_instr_remove(&instr->instr);
+
+ return true;
+ }
+
if (instr->intrinsic != nir_intrinsic_vulkan_resource_index)
return false;
nir_const_value *const_val = nir_src_as_const_value(instr->src[0]);
- if (!const_val || const_val->u32 != 0) {
+ if (!const_val || const_val->u32 != 0)
tu_finishme("non-zero vulkan_resource_index array index");
- return false;
- }
- if (nir_intrinsic_desc_type(instr) != VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER) {
- tu_finishme("non-ubo vulkan_resource_index");
- return false;
- }
- unsigned index = map_add(&shader->ubo_map,
- nir_intrinsic_desc_set(instr),
- nir_intrinsic_binding(instr));
+ unsigned set = nir_intrinsic_desc_set(instr);
+ unsigned binding = nir_intrinsic_binding(instr);
+ unsigned index = 0;
+
+ switch (nir_intrinsic_desc_type(instr)) {
+ case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
+ case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
+ /* skip index 0 which is used for push constants */
+ index = map_add(&shader->ubo_map, set, binding, 0) + 1;
+ break;
+ case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+ case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
+ index = map_add(&shader->ssbo_map, set, binding, 0);
+ break;
+ default:
+ tu_finishme("unsupported desc_type for vulkan_resource_index");
+ break;
+ }
- b->cursor = nir_before_instr(&instr->instr);
- /* skip index 0 because ir3 treats it differently */
nir_ssa_def_rewrite_uses(&instr->dest.ssa,
- nir_src_for_ssa(nir_imm_int(b, index + 1)));
+ nir_src_for_ssa(nir_imm_int(b, index)));
nir_instr_remove(&instr->instr);
return true;
nir_foreach_block(block, impl) {
nir_foreach_instr_safe(instr, block) {
+ b.cursor = nir_before_instr(instr);
switch (instr->type) {
case nir_instr_type_tex:
- progress |= lower_sampler(nir_instr_as_tex(instr), shader);
+ progress |= lower_sampler(&b, nir_instr_as_tex(instr), shader);
break;
case nir_instr_type_intrinsic:
progress |= lower_intrinsic(&b, nir_instr_as_intrinsic(instr), shader);
nir_print_shader(nir, stderr);
}
- /* TODO what needs to happen? */
+ /* multi step inlining procedure */
+ NIR_PASS_V(nir, nir_lower_constant_initializers, nir_var_function_temp);
+ NIR_PASS_V(nir, nir_lower_returns);
+ NIR_PASS_V(nir, nir_inline_functions);
+ NIR_PASS_V(nir, nir_opt_deref);
+ foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
+ if (!func->is_entrypoint)
+ exec_node_remove(&func->node);
+ }
+ assert(exec_list_length(&nir->functions) == 1);
+ NIR_PASS_V(nir, nir_lower_constant_initializers, ~nir_var_function_temp);
+
+ /* Split member structs. We do this before lower_io_to_temporaries so that
+ * it doesn't lower system values to temporaries by accident.
+ */
+ NIR_PASS_V(nir, nir_split_var_copies);
+ NIR_PASS_V(nir, nir_split_per_member_structs);
+
+ NIR_PASS_V(nir, nir_remove_dead_variables,
+ nir_var_shader_in | nir_var_shader_out | nir_var_system_value | nir_var_mem_shared);
+
+ NIR_PASS_V(nir, nir_propagate_invariant);
+
+ NIR_PASS_V(nir, nir_lower_io_to_temporaries, nir_shader_get_entrypoint(nir), true, true);
+
+ NIR_PASS_V(nir, nir_lower_global_vars_to_local);
+ NIR_PASS_V(nir, nir_split_var_copies);
+ NIR_PASS_V(nir, nir_lower_var_copies);
+
+ NIR_PASS_V(nir, nir_opt_copy_prop_vars);
+ NIR_PASS_V(nir, nir_opt_combine_stores, nir_var_all);
+
+ /* ir3 doesn't support indirect input/output */
+ NIR_PASS_V(nir, nir_lower_indirect_derefs, nir_var_shader_in | nir_var_shader_out);
switch (stage) {
case MESA_SHADER_VERTEX:
break;
}
- nir_assign_var_locations(&nir->inputs, &nir->num_inputs,
- ir3_glsl_type_size);
- nir_assign_var_locations(&nir->outputs, &nir->num_outputs,
- ir3_glsl_type_size);
- nir_assign_var_locations(&nir->uniforms, &nir->num_uniforms,
- ir3_glsl_type_size);
-
- NIR_PASS_V(nir, nir_opt_copy_prop_vars);
+ nir_assign_io_var_locations(&nir->inputs, &nir->num_inputs, stage);
+ nir_assign_io_var_locations(&nir->outputs, &nir->num_outputs, stage);
NIR_PASS_V(nir, nir_lower_system_values);
NIR_PASS_V(nir, nir_lower_frexp);
+ if (stage == MESA_SHADER_FRAGMENT)
+ NIR_PASS_V(nir, nir_lower_input_attachments, true);
+
NIR_PASS_V(nir, tu_lower_io, shader);
NIR_PASS_V(nir, nir_lower_io, nir_var_all, ir3_glsl_type_size, 0);
+ if (stage == MESA_SHADER_FRAGMENT) {
+ /* NOTE: lower load_barycentric_at_sample first, since it
+ * produces load_barycentric_at_offset:
+ */
+ NIR_PASS_V(nir, ir3_nir_lower_load_barycentric_at_sample);
+ NIR_PASS_V(nir, ir3_nir_lower_load_barycentric_at_offset);
+
+ NIR_PASS_V(nir, ir3_nir_move_varying_inputs);
+ }
+
+ NIR_PASS_V(nir, nir_lower_io_arrays_to_elements_no_indirects, false);
+
nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
+ /* num_uniforms only used by ir3 for size of ubo 0 (push constants) */
+ nir->num_uniforms = MAX_PUSH_CONSTANTS_SIZE / 16;
+
shader->ir3_shader.compiler = dev->compiler;
shader->ir3_shader.type = stage;
shader->ir3_shader.nir = nir;
*options = (struct tu_shader_compile_options) {
/* TODO ir3_key */
- .optimize = !(pipeline_info->flags &
- VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT),
+ /* TODO: VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT
+ * some optimizations need to happen otherwise shader might not compile
+ */
+ .optimize = true,
.include_binning_pass = true,
};
}