nir_shader *shader;
nir_builder builder;
- VkShaderStage stage;
struct anv_pipeline_layout *layout;
uint32_t indices_start;
unsigned block_idx_src;
switch (intrin->intrinsic) {
- case nir_intrinsic_load_ubo_vk:
- case nir_intrinsic_load_ubo_vk_indirect:
- case nir_intrinsic_load_ssbo_vk:
- case nir_intrinsic_load_ssbo_vk_indirect:
+ case nir_intrinsic_load_ubo:
+ case nir_intrinsic_load_ssbo:
block_idx_src = 0;
break;
- case nir_intrinsic_store_ssbo_vk:
- case nir_intrinsic_store_ssbo_vk_indirect:
+ case nir_intrinsic_store_ssbo:
block_idx_src = 1;
break;
default:
continue; /* the loop */
}
- unsigned set = intrin->const_index[0];
- unsigned binding = intrin->const_index[1];
+ nir_instr *res_instr = intrin->src[block_idx_src].ssa->parent_instr;
+ assert(res_instr->type == nir_instr_type_intrinsic);
+ nir_intrinsic_instr *res_intrin = nir_instr_as_intrinsic(res_instr);
+ assert(res_intrin->intrinsic == nir_intrinsic_vulkan_resource_index);
+
+ unsigned set = res_intrin->const_index[0];
+ unsigned binding = res_intrin->const_index[1];
set_layout = state->layout->set[set].layout;
if (set_layout->binding[binding].dynamic_offset_index < 0)
b->cursor = nir_before_instr(&intrin->instr);
- int indirect_src;
- switch (intrin->intrinsic) {
- case nir_intrinsic_load_ubo_vk_indirect:
- case nir_intrinsic_load_ssbo_vk_indirect:
- indirect_src = 1;
- break;
- case nir_intrinsic_store_ssbo_vk_indirect:
- indirect_src = 2;
- break;
- default:
- indirect_src = -1;
- break;
- }
-
/* First, we need to generate the uniform load for the buffer offset */
uint32_t index = state->layout->set[set].dynamic_offset_start +
set_layout->binding[binding].dynamic_offset_index;
- nir_const_value *const_arr_idx =
- nir_src_as_const_value(intrin->src[block_idx_src]);
-
- nir_intrinsic_op offset_load_op;
- if (const_arr_idx)
- offset_load_op = nir_intrinsic_load_uniform;
- else
- offset_load_op = nir_intrinsic_load_uniform_indirect;
-
nir_intrinsic_instr *offset_load =
- nir_intrinsic_instr_create(state->shader, offset_load_op);
- offset_load->num_components = 1;
- offset_load->const_index[0] = state->indices_start + index;
-
- if (const_arr_idx) {
- offset_load->const_index[1] = const_arr_idx->u[0];
- } else {
- offset_load->const_index[1] = 0;
- nir_src_copy(&offset_load->src[0], &intrin->src[0], &intrin->instr);
- }
+ nir_intrinsic_instr_create(state->shader, nir_intrinsic_load_uniform);
+ offset_load->num_components = 2;
+ offset_load->const_index[0] = state->indices_start + index * 8;
+ offset_load->src[0] = nir_src_for_ssa(nir_imul(b, res_intrin->src[0].ssa,
+ nir_imm_int(b, 8)));
- nir_ssa_dest_init(&offset_load->instr, &offset_load->dest, 1, NULL);
+ nir_ssa_dest_init(&offset_load->instr, &offset_load->dest, 2, NULL);
nir_builder_instr_insert(b, &offset_load->instr);
- nir_ssa_def *offset = &offset_load->dest.ssa;
- if (indirect_src >= 0) {
- assert(intrin->src[indirect_src].is_ssa);
- offset = nir_iadd(b, intrin->src[indirect_src].ssa, offset);
- }
-
- /* Now we can modify the load/store intrinsic */
-
- if (indirect_src < 0) {
- /* The original intrinsic is not an indirect variant. We need to
- * create a new one and copy the old data over first.
- */
-
- nir_intrinsic_op indirect_op;
- switch (intrin->intrinsic) {
- case nir_intrinsic_load_ubo_vk:
- indirect_op = nir_intrinsic_load_ubo_vk_indirect;
- break;
- case nir_intrinsic_load_ssbo_vk:
- indirect_op = nir_intrinsic_load_ssbo_vk_indirect;
- break;
- case nir_intrinsic_store_ssbo_vk:
- indirect_op = nir_intrinsic_store_ssbo_vk_indirect;
- break;
- default:
- unreachable("Invalid direct load/store intrinsic");
- }
-
- nir_intrinsic_instr *copy =
- nir_intrinsic_instr_create(state->shader, indirect_op);
- copy->num_components = intrin->num_components;
-
- for (unsigned i = 0; i < 4; i++)
- copy->const_index[i] = intrin->const_index[i];
-
- /* The indirect is always the last source */
- indirect_src = nir_intrinsic_infos[intrin->intrinsic].num_srcs;
-
- for (unsigned i = 0; i < (unsigned)indirect_src; i++)
- nir_src_copy(©->src[i], &intrin->src[i], ©->instr);
-
- copy->src[indirect_src] = nir_src_for_ssa(offset);
- nir_ssa_dest_init(©->instr, ©->dest,
- intrin->dest.ssa.num_components,
- intrin->dest.ssa.name);
- nir_builder_instr_insert(b, ©->instr);
+ nir_src *offset_src = nir_get_io_offset_src(intrin);
+ nir_ssa_def *new_offset = nir_iadd(b, offset_src->ssa,
+ &offset_load->dest.ssa);
+
+ /* In order to avoid out-of-bounds access, we predicate */
+ nir_ssa_def *pred = nir_uge(b, nir_channel(b, &offset_load->dest.ssa, 1),
+ offset_src->ssa);
+ nir_if *if_stmt = nir_if_create(b->shader);
+ if_stmt->condition = nir_src_for_ssa(pred);
+ nir_cf_node_insert(b->cursor, &if_stmt->cf_node);
+
+ nir_instr_remove(&intrin->instr);
+ *offset_src = nir_src_for_ssa(new_offset);
+ nir_instr_insert_after_cf_list(&if_stmt->then_list, &intrin->instr);
+
+ if (intrin->intrinsic != nir_intrinsic_store_ssbo) {
+ /* It's a load, we need a phi node */
+ nir_phi_instr *phi = nir_phi_instr_create(b->shader);
+ nir_ssa_dest_init(&phi->instr, &phi->dest,
+ intrin->num_components, NULL);
+
+ nir_phi_src *src1 = ralloc(phi, nir_phi_src);
+ struct exec_node *tnode = exec_list_get_tail(&if_stmt->then_list);
+ src1->pred = exec_node_data(nir_block, tnode, cf_node.node);
+ src1->src = nir_src_for_ssa(&intrin->dest.ssa);
+ exec_list_push_tail(&phi->srcs, &src1->node);
+
+ b->cursor = nir_after_cf_list(&if_stmt->else_list);
+ nir_ssa_def *zero = nir_build_imm(b, intrin->num_components,
+ (nir_const_value) { .u = { 0, 0, 0, 0 } });
+
+ nir_phi_src *src2 = ralloc(phi, nir_phi_src);
+ struct exec_node *enode = exec_list_get_tail(&if_stmt->else_list);
+ src2->pred = exec_node_data(nir_block, enode, cf_node.node);
+ src2->src = nir_src_for_ssa(zero);
+ exec_list_push_tail(&phi->srcs, &src2->node);
assert(intrin->dest.is_ssa);
nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
- nir_src_for_ssa(©->dest.ssa));
+ nir_src_for_ssa(&phi->dest.ssa));
- nir_instr_remove(&intrin->instr);
- } else {
- /* It's already indirect, so we can just rewrite the one source */
- nir_instr_rewrite_src(&intrin->instr, &intrin->src[indirect_src],
- nir_src_for_ssa(offset));
+ nir_instr_insert_after_cf(&if_stmt->cf_node, &phi->instr);
}
}
{
struct apply_dynamic_offsets_state state = {
.shader = shader,
- .stage = anv_vk_shader_stage_for_mesa_stage(shader->stage),
.layout = pipeline->layout,
.indices_start = shader->num_uniforms,
};
- if (!state.layout || !state.layout->stage[state.stage].has_dynamic_offsets)
+ if (!state.layout || !state.layout->stage[shader->stage].has_dynamic_offsets)
return;
- nir_foreach_overload(shader, overload) {
- if (overload->impl) {
- nir_builder_init(&state.builder, overload->impl);
- nir_foreach_block(overload->impl, apply_dynamic_offsets_block, &state);
- nir_metadata_preserve(overload->impl, nir_metadata_block_index |
+ nir_foreach_function(shader, function) {
+ if (function->impl) {
+ nir_builder_init(&state.builder, function->impl);
+ nir_foreach_block(function->impl, apply_dynamic_offsets_block, &state);
+ nir_metadata_preserve(function->impl, nir_metadata_block_index |
nir_metadata_dominance);
}
}
struct anv_push_constants *null_data = NULL;
- for (unsigned i = 0; i < MAX_DYNAMIC_BUFFERS; i++)
- prog_data->param[i + shader->num_uniforms] =
- (const gl_constant_value *)&null_data->dynamic_offsets[i];
+ for (unsigned i = 0; i < MAX_DYNAMIC_BUFFERS; i++) {
+ prog_data->param[i * 2 + shader->num_uniforms] =
+ (const union gl_constant_value *)&null_data->dynamic[i].offset;
+ prog_data->param[i * 2 + 1 + shader->num_uniforms] =
+ (const union gl_constant_value *)&null_data->dynamic[i].range;
+ }
- shader->num_uniforms += MAX_DYNAMIC_BUFFERS;
+ shader->num_uniforms += MAX_DYNAMIC_BUFFERS * 8;
}
projects / mesa.git / blobdiff